This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

Donald J. Newman

Springer

Graduate Texts in Mathematics 177

Editorial Board

S. Axler F.W. Gehring K.A. Ribet

Springer

New York

Berlin

Heidelberg

Barcelona

Hong Kong

London

Milan

Paris

Singapore

Tokyo

Donald J. Newman

Analytic Number Theory

1 3

Donald J. Newman

Professor Emeritus

Temple University

Philadelphia, PA 19122

USA

Editorial Board

S. Axler F.W. Gehring K.A. Ribet

Department of Department of Department of

Mathematics Mathematics Mathematics

San Francisco State University University of Michigan University of California

San Francisco, CA 94132 Ann Arbor, MI 48109 at Berkeley

USA USA Berkeley, CA 94720-3840

USA

Mathematics Subject Classiﬁcation (1991): 11-01, 11N13, 11P05, 11P83

Library of Congress Cataloging-in-Publication Data

Newman, Donald J., 1930–

Analytic number theory / Donald J. Newman.

p. cm. – (Graduate texts in mathematics; 177)

Includes index.

ISBN 0-387-98308-2 (hardcover: alk. paper)

1. Number Theory. I. Title. II. Series.

QA241.N48 1997

512’.73–dc21 97-26431

© 1998 Springer-Verlag New York, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the written

permission of the publisher (Springer-Verlag NewYork, Inc., 175 Fifth Avenue, NewYork, NY10010,

USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with

any formof information storage and retrieval, electronic adaptation, computer software, or by similar or

dissimilar methodology nowknown or hereafter developed is forbidden. The use of general descriptive

names, trade names, trademarks, etc., in this publication, even if the former are not especially identiﬁed,

is not to be taken as a sign that such names, as understood by the Trade Marks and Merchandise Marks

Act, may accordingly be used freely by anyone.

ISBN 0-387-98308-2 Springer-Verlag New York Berlin Heidelburg SPIN 10763456

Contents

Introduction and Dedication vii

I. The Idea of Analytic Number Theory 1

Addition Problems 1

Change Making 2

Crazy Dice 5

Can r(n) be “constant?” 8

A Splitting Problem 8

An Identity of Euler’s 11

Marks on a Ruler 12

Dissection into Arithmetic Progressions 14

II. The Partition Function 17

The Generating Function 18

The Approximation 19

Riemann Sums 20

The Coefﬁcients of q(n) 25

III. The Erd˝ os–Fuchs Theorem 31

Erd˝ os–Fuchs Theorem 35

IV. Sequences without Arithmetic Progressions 41

The Basic Approximation Lemma 42

v

vi Contents

V. The Waring Problem 49

VI. A “Natural” Proof of the Nonvanishing of L-Series 59

VII. Simple Analytic Proof of the Prime Number

Theorem 67

First Proof of the Prime Number Theorem. 70

Second Proof of the Prime Number Theorem. 72

Index 77

Introduction and Dedication

This book is dedicated to Paul Erd˝ os, the greatest mathematician I

have ever known, whom it has been my rare privilege to consider

colleague, collaborator, and dear friend.

I like to think that Erd˝ os, whose mathematics embodied the princi-

ples which have impressed themselves upon me as deﬁning the true

character of mathematics, would have appreciated this little book

and heartily endorsed its philosophy. This book proffers the thesis

that mathematics is actually an easy subject and many of the famous

problems, even those in number theory itself, which have famously

difﬁcult solutions, can be resolved in simple and more direct terms.

There is no doubt a certain presumptuousness in this claim. The

great mathematicians of yesteryear, those working in number the-

ory and related ﬁelds, did not necessarily strive to effect the simple

solution. They may have felt that the status and importance of mathe-

matics as an intellectual discipline entailed, perhaps indeed required,

a weighty solution. Gauss was certainly a wordy master and Euler

another. They belonged to a tradition that undoubtedly revered math-

ematics, but as a discipline at some considerable remove from the

commonplace. In keeping with a more democratic concept of intelli-

gence itself, contemporary mathematics diverges fromthis somewhat

elitist view. The simple approach implies a mathematics generally

available even to those who have not been favored with the natural

endowments, nor the careful cultivation of an Euler or Gauss.

vii

viii Introduction and Dedication

Such an attitude might prove an effective antidote to a generally

declining interest in pure mathematics. But it is not so much as incen-

tive that we proffer what might best be called “the fun and games”

approach to mathematics, but as a revelation of its true nature. The

insistence on simplicity asserts a mathematics that is both “magi-

cal” and coherent. The solution that strives to master these qualities

restores to mathematics that element of adventure that has always

supplied its peculiar excitement. That adventure is intrinsic to even

the most elementary description of analytic number theory.

The initial step in the investigation of a number theoretic item

is the formulation of “the generating function”. This formulation

inevitably moves us away from the designated subject to a consider-

ation of complex variables. Having wandered away fromour subject,

it becomes necessary to effect a return. Toward this end “The Cauchy

Integral” proves to be an indispensable tool. Yet it leads us, inevitably,

further aﬁeld from all the intricacies of contour integration and they,

in turn entail the familiar processes, the deformation and estimation

of these contour integrals.

Retracing our steps we ﬁnd that we have gone fromnumber theory

to function theory, and back again. The journey seems circuitous, yet

in its wake a pattern is revealed that implies a mathematics deeply

inter-connected and cohesive.

I

The Idea of Analytic Number

Theory

The most intriguing thing about Analytic Number Theory (the use of

Analysis, or function theory, in number theory) is its very existence!

How could one use properties of continuous valued functions to de-

termine properties of those most discrete items, the integers. Analytic

functions? What has differentiability got to do with counting? The

astonishment mounts further when we learn that the complex zeros

of a certain analytic function are the basic tools in the investigation

of the primes.

The answer to all this bewilderment is given by the two words

generating functions. Well, there are answers and answers. To those

of us who have witnessed the use of generating functions this is a kind

of answer, but to those of us who haven’t, this is simply a restatement

of the question. Perhaps the best way to understand the use of the

analytic method, or the use of generating functions, is to see it in

action in a number of pertinent examples. So let us take a look at

some of these.

Addition Problems

Questions about addition lend themselves very naturally to the use of

generating functions. The link is the simple observation that adding

m and n is isomorphic to multiplying z

m

and z

n

. Thereby questions

about the addition of integers are transformed into questions about

the multiplication of polynomials or power series. For example, La-

grange’s beautiful theorem that every positive integer is the sum of

1

2 I. The Idea of Analytic Number Theory

four squares becomes the statement that all of the coefﬁcients of the

power series for

1 ÷ z ÷ z

4

÷ · · · ÷ z

n

2

÷ · · ·

4

are positive. How

one proves such a fact about the coefﬁcients of such a power series

is another story, but at least one begins to see how this transition

from integers to analytic functions takes place. But now let’s look at

some addition problems that we can solve completely by the analytic

method.

Change Making

How many ways can one make change of a dollar? The answer is

293, but the problem is both too hard and too easy. Too hard because

the available coins are so many and so diverse. Too easy because it

concerns just one “changee,” a dollar. More ﬁtting to our spirit is the

following problem: Howmany ways can we make change for n if the

coins are 1, 2, and 3? To form the appropriate generating function,

let us write, for [z[ < 1,

1

1 − z

= 1 ÷ z ÷ z

1÷1

÷ z

1÷1÷1

÷ · · · ,

1

1 − z

2

= 1 ÷ z

2

÷ z

2÷2

÷ z

2÷2÷2

÷ · · · ,

1

1 − z

3

= 1 ÷ z

3

÷ z

3÷3

÷ z

3÷3÷3

÷ · · · ,

and multiplying these three equations to get

1

(1 − z)(1 − z

2

)(1 − z

3

)

= (1 ÷ z ÷ z

1÷1

÷ · · ·)(1 ÷ z

2

÷ z

2÷2

÷ · · ·)

(1 ÷ z

3

÷ z

3÷3

÷ · · ·).

Now we ask ourselves: What happens when we multiply out the

right-hand side? We obtain terms like z

1÷1÷1÷1

· z

2

· z

3÷3

. On the one

hand, this term is z

12

, but, on the other hand, it is z

four1

/

s÷one2÷two3

/

s

and doesn’t this exactly correspond to the method of changing the

amount 12 into four 1’s, one 2, and two 3’s? Yes, and in fact we

Change Making 3

see that “every” way of making change (into 1’s, 2’s, and 3’s) for

“every” n will appear in this multiplying out. Thus if we call C(n) the

number of ways of making change for n, then C(n) will be the exact

coefﬁcient of z

n

when the multiplication is effected. (Furthermore

all is rigorous and not just formal, since we have restricted ourselves

to [z[ < 1 wherein convergence is absolute.)

Thus

¸

C(n)z

n

=

1

(1 − z)(1 − z

2

)(1 − z

3

)

, (1)

and the generating function for our unknown quantity C(n) is

produced. Our number theoretic problem has been translated into

a problem about analytic functions, namely, ﬁnding the Taylor

coefﬁcients of the function

1

(1−z)(1−z

2

)(1−z

3

)

.

Fine. Awell deﬁnedanalytic problem, but howtosolve it? We must

resist the temptation to solve this problem by undoing the analysis

which led to its formulation. Thus the thing not to do is expand

1

1−z

,

1

1−z

2

,

1

1−z

3

respectively into

¸

z

a

,

¸

z

2b

,

¸

z

3c

and multiply only to

discover that the coefﬁcient is the number of ways of making change

for n.

The correct answer, in this case, comes from an algebraic tech-

nique that we all learned in calculus, namely partial fractions. Recall

that this leads to terms like

A

(1−αz)

k

for which we know the expan-

sion explicitly (namely,

1

(1−αz)

k

is just a constant times the (k − 1)th

derivative of

1

(1−αz)

=

¸

α

n

z

n

).

Carrying out the algebra, then, leads to the partial fractional

decomposition which we may arrange in the following form:

1

(1 − z)(1 − z

2

)(1 − z

3

)

=

1

6

1

(1 − z)

3

÷

1

4

1

(1 − z)

2

÷

1

4

1

(1 − z

2

)

÷

1

3

1

(1 − z

3

)

.

Thus, since

1

(1 − z)

2

=

d

dz

1

1 − z

=

d

dz

¸

z

n

=

¸

(n ÷ 1)z

n

4 I. The Idea of Analytic Number Theory

and

1

(1 − z)

3

=

d

dz

1

2(1 − z)

2

=

d

dz

¸

n ÷ 1

2

z

n

=

¸

(n ÷ 2)(n ÷ 1)

2

z

n

,

C(n) =

(n ÷ 2)(n ÷ 1)

12

÷

n ÷ 1

4

÷

χ

1

(n)

4

÷

χ

2

(n)

3

(2)

where χ

1

(n) = 1 if 2 [ n and = 0 otherwise; χ

2

(n) = 1 if 3 [ n

and = 0 else. A somewhat cumbersome formula, but one which can

be shortened nicely into

C(n) =

¸

n

2

12

÷

n

2

÷ 1

: (3)

where the terms in the brackets mean the greatest integers.

A nice crisp exact formula, but these are rare. Imagine the mess

that occurs if the coins were the usual coins of the realm, namely 1, 5,

10, 25, 50, (100?). The right thing to ask for then is an “asymptotic”

formula rather than an exact one.

Recall that an asymptotic formula F(n) for a function f (n) is one

for which lim

n→∞

f (n)

F(n)

= 1. In the colorful language of E. Landau,

the relative error in replacing f (n) by F(n) is eventually 0%. At

any rate, we write f (n) ∼ F(n) when this occurs. One famous such

example is Stirling’s formula n! ∼

√

2πn(

n

e

)

n

. (Also note that our

result (3) can be weakened to C(n) ∼

n

2

12

.)

So let us assume quite generally that there are coins a

1

, a

2

, a

3

, . . .,

a

k

, where to avoid trivial congruence considerations we will require

that there be no common divisiors other than 1. In this generality we

ask for an asymptotic formula for the corresponding C(n). As before

we ﬁnd that the generating function is given by

¸

C(n)z

n

=

1

(1 − z

a

1

)(1 − z

a

2

) · · · (1 − z

a

k

)

. (4)

But the next step, explicitly ﬁnding the partial fractional decompo-

sition of this function is the hopeless task. However, let us simply

look for one of the terms in this expansion, the heaviest one. Thus

Crazy Dice 5

at z = 1 the denominator has a k-fold zero and so there will be a

term

c

(1−z)

k

. All the other zeros are roots of unity and, because we

assumed no common divisiors, all will be of order lower than k.

Thus, although the coefﬁcient of the term

c

(1−z)

k

is c

n÷k−1

k−1

, the

coefﬁcients of all other terms

a

(1−ωz)

j

will be aω

j

n÷j

j−1

. Since all of

these j are less than k, the sumtotal of all of these terms is negligible

compared to our heavy term c

n÷k−1

k−1

. In short C(n) ∼ c

n÷k−1

k−1

, or

even simpler,

C(n) ∼ c

n

k−1

(k − 1)!

.

But, what is c? Although we have deftly avoided the necessity of

ﬁnding all of the other terms, we cannot avoid this one (it’s the whole

story!). So let us write

1

(1 − z

a

1

)(1 − z

a

2

) · · · (1 − z

a

k

)

=

c

(1 − z)

k

÷ other terms,

multiply by (1 − z)

k

to get

1 − z

1 − z

a

1

1 − z

1 − z

a

2

· · ·

1 − z

1 − z

a

k

= c ÷ (1 − z)

k

other terms,

and ﬁnally let z → 1. By L’Hˆ opital’s rule, for example,

1−z

1−z

a

i

→

1

a

i

whereas each of the other terms times (1 − z)

k

goes to 0. The ﬁnal

result is c =

1

a

1

a

2

···a

k

, and our ﬁnal asymptotic formula reads

C(n) ∼

n

k−1

a

1

a

2

· · · a

k

(k − 1)!

. (5)

Crazy Dice

An ordinary pair of dice consist of two cubes each numbered 1

through 6. When tossed together there are altogether 36 (equally

likely) outcomes. Thus the sums go from 2 to 12 with varied

numbers of repeats for these possibilities. In terms of our ana-

lytic representation, each die is associated with the polynomial

z ÷ z

2

÷ z

3

÷ z

4

÷ z

5

÷ z

6

. The combined possibilities for the

6 I. The Idea of Analytic Number Theory

sums then are the terms of the product

(z ÷ z

2

÷ z

3

÷ z

4

÷ z

5

÷ z

6

)(z ÷ z

2

÷ z

3

÷ z

4

÷ z

5

÷ z

6

)

= z

2

÷ 2z

3

÷ 3z

4

÷ 4z

5

÷ 5z

6

÷ 6z

7

÷ 5z

8

÷ 4z

9

÷ 3z

10

÷ 2z

11

÷ z

12

The correspondence, for example, says that there are 3 ways for the

10 to show up, the coefﬁcients of z

10

being 3, etc. The question is: Is

there any other way to number these two cubes with positive integers

so as to achieve the very same alternatives?

Analytically, then, the question amounts to the existence of

positive integers, a

1

, . . . , a

6

: b

1

, . . . , b

6

, so that

(z

a

1

÷ · · · ÷ z

a

6

)(z

b

1

÷ · · · ÷ z

b

6

)

= z

2

÷ 2z

3

÷ 3z

4

÷ · · · ÷ 3z

10

÷ 2z

11

÷ z

12

.

These would be the “Crazy Dice” referred to in the title of this sec-

tion. They look totally different from ordinary dice but they produce

exactly the same results!

So, repeating the question, can

(z

a

1

÷ · · · ÷ z

a

6

)(z

b

1

÷ · · · ÷ z

b

6

)

= (z ÷ z

2

÷ z

3

÷ z

4

÷ z

5

÷ z

6

) (6)

(z ÷ z

2

÷ z

3

÷ z

4

÷ z

5

÷ z

6

)?

To analyze this possibility, let us factor completely (over the ratio-

nals) this right-hand side. Thus z ÷z

2

÷z

3

÷z

4

÷z

5

÷z

6

= z

1−z

6

1−z

=

z(1÷z÷z

2

)(1÷z

3

) = z(1÷z÷z

2

)(1÷z)(1−z÷z

2

). We conclude

from(6) that the “a-polynomial” and “b-polynomial” must consist of

these factors. Also there are certain side restrictions. The a’s and b’s

are to be positive and so a z-factor must appear in both polynomials.

The a-polynomial must be 6 at z = 1 and so the (1 ÷z ÷z

2

)(1 ÷z)

factor must appear in it, and similarly in the b-polynomial. All that

is left to distribute are the two factors of 1 −z ÷z

2

. If one apiece are

given to the a- and b-polynomials, then we get ordinary dice. The

only thing left to try is putting both into the a-polynomial.

Crazy Dice 7

This works! We obtain ﬁnally

¸

z

a

= z(1 ÷ z ÷ z

2

)(1 ÷ z)(1 − z ÷ z

2

)

2

= z ÷ z

3

÷ z

4

÷ z

5

÷ z

6

÷ z

8

and

¸

z

b

= z(1 ÷ z ÷ z

2

)(1 ÷ z) = z ÷ 2z

2

÷ 2z

3

÷ z

4

.

Translating back, the crazy dice are 1,3,4,5,6,8 and 1,2,2,3,3,4.

Now we introduce the notion of the representation function. So,

suppose there is a set A of nonnegative integers and that we wish to

express the number of ways in which a given integer n can be written

as the sum of two of them. The trouble is that we must decide on

conventions. Does order count? Can the two summands be equal?

Therefore we introduce three representation functions.

r(n) = #{(a, a

/

) : a, a

/

∈ A, n = a ÷ a

/

}:

So here order counts, and they can be equal;

r

÷

(n) = #{(a, a

/

) : a, a

/

∈ A, a ≤ a

/

, n = a ÷ a

/

},

order doesn’t count, and they can be equal;

r

−

(n) = #{(a, a

/

) : a, a

/

∈ A, a < a

/

, n = a ÷ a

/

},

order doesn’t count, and they can’t be equal. In terms of the generat-

ing function for the set A, namely, A(z) =

¸

a∈A

z

a

, we can express

the generating functions of these representation functions.

The simplest is that of r(n), where obviously

¸

r(n)z

n

= A

2

(z). (7)

To deal with r

−

(n), we must subtract A(z

2

) from A

2

(z) to remove

the case of a = a

/

and then divide by 2 to remove the order. So here

¸

r

−

(n)z

n

=

1

2

[A

2

(z) − A(z

2

)]. (8)

8 I. The Idea of Analytic Number Theory

Finally for r

÷

(n), we must add A(z

2

) to this result to reinstate the

case of a = a

/

, and we obtain

¸

r

÷

(n)z

n

=

1

2

[A

2

(z) ÷ A(z

2

)]. (9)

Can r(n) be “constant?”

Is it possible to design a nontrivial set A, so that, say, r

÷

(n) is the same

for all n? The answer is NO, for we would have to have 0 ∈ A. And

then 1 ∈ A, else r

÷

(1) ,= r

÷

(0). And then 2 / ∈ A, else r

÷

(2) = 2.

And then 3 ∈ A, else r

÷

(3) = 0 (whereas r

÷

(1) = 1), then 4 / ∈ A,

else r

÷

(4) = 2. Continuing in this manner, we ﬁnd 5 ∈ A. But now

we are stymied since now 6 = 1 ÷ 5, 6 = 3 ÷ 3, and r

÷

(6) = 2.

The suspicion arises, though, that this impossibility may just be

a quirk of “small” numbers. Couldn’t A be designed so that, except

for some misbehavior at the beginning, r

÷

(n) = constant?

We will analyze this question by using generating functions. So,

using (9), the question reduces to whether there is an inﬁnite set A

for which

1

2

[A

2

(z) ÷ A(z

2

)] = P(z) ÷

C

1 − z

, (10)

P(z) is a polynomial.

Answer: No. Just look what happens if we let z → (−1)

÷

. Clearly

P(z) and

C

1−z

remain bounded, A

2

(z) remains nonnegative, and

A(z

2

) goes to A(1) = ∞, a contradiction.

A Splitting Problem

Can we split the nonnegative integers in two sets A and B so that

every integer n is expressible in the same number of ways as the

sum of two distinct members of A, as it is as the sum of two distinct

members of B?

If we experiment a bit, before we get down to business, and begin

by placing 0 ∈ A, then 1 ∈ B, else 1 would be expressible as

A Splitting Problem 9

a ÷ a

/

but not as b ÷ b

/

. Next 2 ∈ B, else 2 would be a ÷ a

/

but

not b ÷ b

/

. Next 3 ∈ A, else 3 would not be a ÷ a

/

whereas it

is b ÷ b

/

= 1 ÷ 2. Continuing in this manner, we seem to force

A = {0, 3, 5, 6, 9, · · ·} and B = {1, 2, 4, 7, 8, · · ·}. But the pattern

is not clear, nor is the existence or uniqueness of the desired A, B. We

must turn to generating functions. So observe that we are requiring

by (8) that

1

2

[A

2

(z) − A(z

2

)] =

1

2

[B

2

(z) − B(z

2

)]. (11)

Also, because of the condition that A, B be a splitting of the

nonnegatives, we also have the condition that

A(z) ÷ B(z) =

1

1 − z

. (12)

From (11) we obtain

A

2

(z) − B

2

(z) = A(z

2

) − B(z

2

), (13)

and so, by (12), we conclude that

[A(z) − B(z)] ·

1

1 − z

= A(z

2

) − B(z

2

),

or

A(z) − B(z) = (1 − z)[A(z

2

) − B(z

2

)]. (14)

Now this is a relationship that can be iterated. We see that

A(z

2

) − B(z

2

) = (1 − z

2

)[A(z

4

) − B(z

4

)],

so that continuing gives

A(z) − B(z) = (1 − z)(1 − z

2

)[A(z

4

) − B(z

4

)].

And, if we continue to iterate, we obtain

A(z) −B(z) = (1 −z)(1 −z

2

) · · · (1 −z

2

n−1

)

¸

A(z

2

n

) − B(z

2

n

)

¸

,

(15)

10 I. The Idea of Analytic Number Theory

and so, by letting n → ∞, since A(0) = 1, B(0) = 0, we deduce

that

A(z) − B(z) =

∞

¸

i=0

(1 − z

2

i

). (16)

And this product is easy to “multiply out”. Every term z

n

occurs

uniquely since every n is uniquely the sum of distinct powers of 2.

Indeed z

n

occurs with coefﬁcient ÷1 if nis the sumof an even number

of distinct powers of 2, and it has coefﬁcient −1, otherwise.

We have achieved success! The sets A and B do exist, are unique,

and indeed are given by A = Integers, which are the sum of an even

number of distinct powers of 2, and B = Integers, which are the sum

of an odd number of distinct powers of 2. This is not one of those

problems where, after the answer is exposed, one proclaims, “oh, of

course.” It isn’t really trivial, even in retrospect, why the A and B

have the same r

−

(n), or for that matter, to what this common r

−

(n)

is equal. (See below where it is proved that r

−

(2

2k÷1

− 1) = 0.)

A = Integers with an even number of 1’s in radix 2. Then and

only then

2k÷1

. .. .

111 · · · 1 = 2

2k÷1

− 1

is not the sum of two distinct A’s.

Proof. A sum of two A’s, with no carries has an even number of

1’s (so it won’t give

odd

. .. .

111 · · · 1), else look at the ﬁrst carry. This gives

a 0 digit so, again, it’s not 11 · · · 1.

So r

−

(2

2k÷1

− 1) = 0. We must now show that all other n have

a representation as the sum of two numbers whose numbers of 1

digits are of like parity. First of all if n contains 2k 1’s then it is the

sum of the ﬁrst k and the second k. Secondly if n contains 2k ÷ 1

1’s but also a 0 digit then it is structured as 111 · · ·

. .. .

m

◦A where A

contains 2k ÷ 1 − m 1’s and, say, is of total length L then it can be

expressed as 111 · · · 1

. .. .

m−1

◦ 00 · · · 00

. .. .

2

plus 1A and these two numbers

An Identity of Euler’s 11

have respectively m 1’s and 2k ÷ 2 − m 1’s. These are again of like

parity so we are done.

An Identity of Euler’s

Consider expressing n as the sum of distinct positive integers, i.e.,

where repeats are not allowed. (So For n = 6, we have the expression

1 ÷ 2 ÷ 3 and also 2 ÷ 4, 1 ÷ 5, and just plain 6 alone.)

Also consider expressing n as the sum of positive odd numbers,

but this time where repeats are allowed. (So for n = 6, we get 1 ÷5,

3 ÷ 3, 1 ÷ 1 ÷ 1 ÷ 3, 1 ÷ 1 ÷ 1 ÷ 1 ÷ 1 ÷ 1.) In both cases we

obtained four expressions for 6, and a theorem of Euler’s says that

this is no coincidence, that is, it says the following:

Theorem. The number of ways of expressing n as the sumof distinct

positive integers equals the number of ways of expressing n as the

sum of (not necessarily distinct) odd positive integers.

To prove this theorem we produce two generating functions. The

latter is exactly the “coin changing” function where the coins have

the denominations 1, 3, 5, 7, . . . . This generating function is given

by

1

(1 − z)(1 − z

3

)(1 − z

5

) · · ·

. (17)

The other generating function is not of the coin changing variety

because of the distinctness condition. Amoment’s thought, however,

shows that this generating function is given as the product of 1 ÷ z,

1 ÷z

2

, 1 ÷z

3

, . . . . For, when these are multiplied out, each z

k

factor

occurs at most once. In short, the other generating function is

(1 ÷ z)(1 ÷ z

2

)(1 ÷ z

3

) · · · . (18)

Euler’s theorem in its analytic form is then just the identity

1

(1 − z)(1 − z

3

)(1 − z

5

) · · ·

= (1 ÷ z)(1 ÷ z

2

)(1 ÷ z

3

) · · ·

throughout [z[ < 1. (19)

12 I. The Idea of Analytic Number Theory

Another way of writing (19) is

(1 −z)(1 −z

3

)(1 −z

5

) · · · (1 ÷z)(1 ÷z

2

)(1 ÷z

3

) · · · = 1 (20)

which is the provocative assertion that, when this product is

multiplied out, all of the terms (aside from the 1) cancel each other!

To prove (2) multiply the 1 − z by the 1 ÷ z (to get 1 − z

2

) and

do the same with 1 − z

3

by 1 ÷ z

3

, etc. This gives the new factors

1 − z

2

, 1 − z

6

, 1 − z

10

, · · · and leaves untouched the old factors

1 ÷ z

2

, 1 ÷ z

4

, 1 ÷ z

6

, · · ·. These rearrangements are justiﬁed by

absolute convergence, and so we see that the product in (20), call it

P(z), is equal to

(1 − z

2

)(1 − z

6

)(1 − z

10

) · · · (1 ÷ z

2

)(1 ÷ z

4

) · · ·

which just happens to be P(z

2

)! So P(z) = P(z

2

) which of course

means that there can’t be any terms az

k

, a ,= 0, k ,= 0, in the

expansion of P(z), i.e., P(z) is just its constant term 1, as asserted.

Marks on a Ruler

Suppose that a 6” ruler is marked as usual at 0, 1, 2, 3, 4, 5, 6.

Using this ruler we may of course measure any integral length from

1 through 6. But we don’t need all of these markings to accomplish

these measurements. Thus we can remove the 2, 3, and 5, and the

marks at 0, 1, 4, 6 are sufﬁcient. (The 2 can be measured between 4

and 6, the 3 can be gotten between 1 and 4, and the 5 between 1 and

6.) Since

4

2

**= 6, this is a “perfect” situation. The question suggests
**

itself then, are there any larger perfect values? In short, can there

be integers a

1

< a

2

< · · · < a

n

such that the differences a

i

− a

j

,

i > j, take on all the values 1, 2, 3, . . . ,

n

2

?

If we introduce the usual generating function A(z) =

¸

n

i=1

z

a

i

,

then the differences are exposed, not when we square A(z), but when

we multiply A(z) by A(

1

z

). Thus A(z) · A(

1

z

) =

¸

n

i,j=1

z

a

i

−a

j

and

if we split this (double) sum as i > j, i = j, and i < j, we obtain

A(z) · A

1

z

=

n

¸

i,j=1

i>j

z

a

i

−a

j

÷ n ÷

n

¸

i,j=1

i<j

z

a

i

−a

j

.

Dissection into Arithmetic Progressions 13

Our “perfect ruler,” by hypothesis, then requires that the ﬁrst sum be

equal to

¸

N

k=1

z

k

, N =

n

2

**, and since the last sum is the same as
**

ﬁrst, with

1

z

replacing z, our equation takes the simple form

A(z) · A

1

z

=

N

¸

k=−N

z

k

÷ n − 1, N =

n

2

,

or, summing this geometric series,

A(z) · A

1

z

=

z

N÷1

− z

−N

z − 1

÷ n − 1, N =

n

2

. (21)

In search of a contradiction, we let z lie on the unit circle z = e

iθ

,

so that the left side of (21) becomes simply [A(e

iθ

)[

2

, whereas the

right-hand side is

z

N÷

1

2

− z

−(N÷

1

2

)

z

1

2

− z

−

1

2

÷ n − 1 =

sin(N ÷

1

2

)θ

sin

1

2

θ

÷ n − 1

and (21) reduces to

A(e

iθ

)

2

=

sin

n

2

−n÷1

2

θ

sin

1

2

θ

÷ n − 1. (22)

A contradiction will occur, then, if we pick a θ which makes

sin

n

2

−n÷1

2

θ

sin

1

2

θ

< −(n − 1). (23)

(And we had better assume that n ≥ 5, since we saw the perfect

ruler for n = 4.)

A good choice, then, is to make sin

n

2

−n÷1

2

θ = −1, for exam-

ple by picking θ =

3π

n

2

−n÷1

. In that case sin

θ

2

<

θ

2

,

1

sin

θ

2

>

2

θ

,

−

1

sin

θ

2

< −

2

θ

= −

2n

2

−2n÷2

3π

. and so the requirement (23) follows

from −

2n

2

−2n÷2

3π

< −(n − 1) or 2n

2

− 2n ÷ 2 > 3π(n − 1). But

2n

2

− 2n ÷ 2 − 3π(n − 1) > 2n

2

− 2n ÷ 2 − 10(n − 1) =

2(n − 3)

2

− 6 ≥ 2 · 2

2

− 6 = 2, for n ≥ 5. There are no perfect

rulers!

14 I. The Idea of Analytic Number Theory

Dissection into Arithmetic Progressions

It is easy enough to split the nonnegative integers into arithmetic

progressions. For example they split into the evens and the odds or

into the progressions 2n, 4n ÷ 1, 4n ÷ 3. Indeed there are many

other ways, but all seem to require at least two of the progressions

to have same common difference (the evens and odds both have 2 as

a common difference and the 4n ÷ 1 and 4n ÷ 3 both have 4). So

the question arises Can the positive integers be split into at least two

arithmetic progressions any two of which have a distinct common

difference?

Of course we look to generating functions for the answer. The

progression an ÷ b, n = 0, 1, 2, . . . will be associated with the

function

¸

∞

n=0

z

an÷b

. Thus the dissection into evens and odds cor-

responds to the identity

¸

∞

n=0

z

n

=

¸

∞

n=0

z

2n

÷

¸

∞

n=0

z

2n÷1

, and

the dissection into 2n, 4n ÷ 1, 4n ÷ 3 corresponds to

¸

∞

n=0

z

n

=

¸

∞

n=0

z

2n

÷

¸

∞

n=0

z

4n÷1

÷

¸

∞

n=0

z

4n÷3

, etc. Since each of these series

is geometric, we can express their sums by

¸

∞

n=0

z

an÷b

=

z

b

1−z

a

. Our

question then is exactly whether there can be an identity

1

1 − z

=

z

b

1

1 − z

a

1

÷

z

b

2

1 − z

a

2

÷ · · · ÷

z

b

k

1 − z

a

k

,

1 < a

1

< a

2

< . . . < a

k

. (24)

Well, just as the experiment suggested, there cannot be such a dis-

section, (24) is impossible. To see that (24) does, indeed, lead to a

contradiction, all we need do is let z → e

2πi

a

k

and observe that then

all of the terms in (24) approach ﬁnite limits except the last term

z

b

k

1−z

a

k

which approaches ∞.

Hopefully, then, this chapter has helped take the sting out of the

preposterous notion of using analysis in number theory.

Problems for Chapter I 15

Problems for Chapter I

1. Produce a set A such that r(n) > 0 for all n in 1 ≤ n ≤ N, but

with [A[ ≤

√

4N ÷ 1.

2. Show that every set satisfying the conditions of (1) must have

[A[ ≤

√

N.

3. Showdirectly, with no knowledge of Stirling’s formula, that n! >

(

n

e

)

n

.

II

The Partition Function

One of the simplest, most natural, questions one can ask in arithmetic

is how to determine the number of ways of breaking up a given inte-

ger. That is, we ask about a positive integer n: In howmany ways can

it be written as a ÷ b ÷ c ÷ · · · where a, b, c, . . . are positive inte-

gers? It turns out that there are two distinct questions here, depending

on whether we elect to count the order of the summands. If we do

choose to let the order count, then the problem becomes too simple.

The answer is just 2

n−1

and the proof is just induction. Things are

incredibly different and more complicated if order is not counted!

In this case the number of breakups or “partitions” is 1 for n = 1,

2 for n = 2, 3 for n = 3, 5 for n = 4, 7 for n = 5, e.g., 5 has the

representations 1 ÷1 ÷1 ÷1 ÷1, 2 ÷1 ÷1 ÷1, 3 ÷1 ÷1, 4 ÷1,

5, 3 ÷ 2, 2 ÷ 2 ÷ 1, and no others. Remember such expressions

as 1 ÷ 1 ÷ 2 ÷ 1 are not considered different. The table can be

extended further of course but no apparent pattern emerges. There

is a famous story concerning the search for some kind of pattern in

this table. This is told of Major MacMahon who kept a list of these

partition numbers arranged one under another up into the hundreds.

It suddenly occurred to him that, viewed from a distance, the outline

of the digits seemed to form a parabola! Thus the number of digits

in p(n), the number of partitions of n, is around C

√

n, or p(n) itself

is very roughly e

α

√

n

. The ﬁrst crude assessment of p(n)!

Among other things, however, this does tell us not to expect any

simple answers. Indeedlater researchshowedthat the true asymptotic

formula for p(n) is

e

π

√

2n/3

4

√

3n

, certainly not a formula to be guessed!

17

18 II. The Partition Function

Now we turn to the analytic number theory derivation of this

asymptotic formula.

The Generating Function

To put into sharp focus the fact that order does not count, we may

view p(n) as the number of representations of n as a sum of 1’s and

2’s and 3’s . . . , etc. But this is just the “change making” problem

where coins come in all denominations. The analysis in that problem

extends verbatim to this one, even though we now have an inﬁnite

number of coins, So we obtain

∞

¸

n=0

p(n)z

n

=

∞

¸

k=1

1

1 − z

k

(1)

valid for [z[ < 1, where we understand that p(0) = 1.

Having thus obtained the generating function, we turn to the sec-

ond stage of attack, investigating the function. This is always the

tricky (creative?) part of the process. We know pretty well what kind

of information we desire about p(n): an estimate of its growth, per-

haps even an asymptotic formula if we are lucky. But we don’t know

exactly how this translates to the generating function. To grasp the

connection between the generating function and its coefﬁcients, then,

seems to be the paramount step. How does one go from one to the

other? Mainly how does one go from a function to its coefﬁcients?

It is here that complex numbers really play their most important

role. The point is that there are formulas (for said coefﬁcients). Thus

we learned in calculus that, if f (z) =

¸

a

n

z

n

, then a

n

=

f

(n)

(0)

n!

,

expressing the desired coefﬁcients in terms of high derivatives of the

function. But this a terrible way of getting at the thing. Except for

rare “made up” examples there is very little hope of obtaining the nth

derivative of a given function and even estimating these derivatives

is not a task with very good prospects. Face it, the calculus approach

is a ﬂop.

Cauchy’s theoremgives a different and more promising approach.

Thus, again with f (z) =

¸

a

n

z

n

, this time we have the formula

The Approximation 19

a

n

=

1

2πi

C

f (z)

z

n÷1

dz, an integral rather than a differential operator!

Surely this is a more secure approach, because integral operators are

bounded, and differential operators are not. The price we pay is that

of passing to the complex numbers for our z’s. Not a bad price, is it?

So let us get under way, but armed with the knowledge that the

valuable information about f (z) will help in getting a good approx-

imation to

C

f (z)

z

n÷1

dz. But a glance at the potentially explosive

1

z

n÷1

shows us that C had better stay as far away from the origin as it can,

i.e., it must hug the unit circle. Again, a look at our generating func-

tion

¸

p(n)z

n

shows that it’s biggest when z is positive (since the

coefﬁcients are themselves positive). All in all, we see that we should

seek approximations to our generating function which are good for

[z[ near 1 with special importance attached to those z’s which are

near ÷1.

The Approximation

Starting with (1), F(z) =

¸

∞

k=1

1

1−z

k

, and taking logarithms, we

obtain

log F(z) =

∞

¸

k=1

log

1

1 − z

k

=

∞

¸

k=1

∞

¸

j=1

z

kj

j

=

∞

¸

j=1

1

j

∞

¸

k=1

z

jk

=

∞

¸

j=1

1

j

z

j

1 − z

j

. (2)

Now write z = e

−w

so that +w > 0 and obtain log F(e

−w

) =

¸

∞

k=1

1

k

1

e

kw

−1

. Thus noticing that the expansion of

1

e

x

−1

begins with

1

x

−

1

2

÷ c

1

x ÷ · · · or equivalently (near 0)

1

x

−

e

−x

2

÷ cx ÷ · · ·,

we rewrite this as

log F(e

−w

) =

¸

1

k

1

kw

−

e

−kw

2

÷

¸

1

k

1

e

kw

− 1

−

1

kw

÷

e

−kw

2

(3)

20 II. The Partition Function

=

π

2

6w

÷

1

2

log(1 − e

−w

)

÷

¸

1

k

1

e

kw

− 1

−

1

kw

÷

e

−kw

2

.

The formof this series is very suggestive. Indeed we recognize any

series

¸

1

k

A(kw) =

¸

A(kw)

kw

w as a Riemann sum, approximating

the Riemann integral

∞

0

A(t )

t

dt for small positive w. It should come

as no surprise then, that such series are estimated rather accurately.

So let us review the “Riemann sum story”.

Riemann Sums

Suppose that φ(x) is a positive decreasing function on (0, ∞) and

that h > 0. The Riemann sum

¸

∞

k=1

φ(kh)h is clearly equal to the

area of the union of rectangles and so is bounded by the area under

y = φ(x). Hence

¸

∞

k=1

φ(kh)h ≤

∞

0

φ(x)dx. On the other hand,

the series

¸

∞

k=0

φ(kh)h can be construed as the area of this union of

these rectangles and, as such, exceeds the area under y = φ(x). So

this time we obtain

¸

∞

k=0

φ(kh)h ≥

∞

0

φ(x)dx.

Combining these two inequalities tells us that the Riemann sum

lies within h · φ(0) of the Riemann integral. This is all very nice and

rather accurate but it refers onlytodecreasingfunctions. However, we

may easily remedy this restriction by subtracting two such functions.

Thereby we obtain

∞

¸

k=1

[φ(kh) − ψ(kh)]h −

∞

0

[φ(x) − ψ(x)] < h[φ(0) ÷ ψ(0)].

Calling φ(x) −ψ(x) = F(x) and then observing that φ(0) ÷ψ(0)

is the total variation V of F(x) we have the rather general result

∞

¸

k=1

F(kh)h −

∞

0

F(x) < h · V(F). (4)

To be sure, we have proven this result only for real functions but

in fact it follows for complex ones, by merely applying it to the real

and imaginary parts.

Riemann Sums 21

To modify this result to ﬁt our situation, let us write w = he

iθ

,

h > 0, −π/2 < θ < π/2, and conclude from (4) that

∞

¸

k=1

F(khe

iθ

)h −

∞

0

F(xe

iθ

)dx < h · V

θ

(F)

(V

θ

is the variation along the ray of argument θ), so that

∞

¸

k=1

F(kw)w −

∞

0

F(xe

iθ

)d(xe

iθ

) < w · V

θ

(F).

Furthermore, in our case of an analytic F, this integral is actually

independent of θ. (Simply apply Cauchy’s theorem and observe that

at ∞ F falls off like

1

x

2

). We also may use the formula V

θ

(F) =

∞

0

[F

/

(xe

iθ

)[dx and ﬁnally deduce that

∞

¸

k=1

F(kw)w −

∞

0

F(x)dx < w

∞

0

[F

/

(xe

iθ

)[dx.

Later on we show that

∞

0

1

e

x

− 1

−

1

x

÷

e

−x

2

dx

x

= log

1

√

2π

, (5)

and right now we may note that the (complicated) function

F

/

(xe

iθ

) =

2

x

3

e

3iθ

−

e

−xe

iθ

2x

2

e

2iθ

−

e

−xe

iθ

2xe

iθ

−

1

x

2

e

2iθ

(e

xe

iθ

− 1)

−

e

xe

iθ

xe

iθ

(e

xe

iθ

− 1)

2

is uniformly bounded by

M

(x÷1)

2

in any wedge [θ[ < c < π/2(m ÷

M(c)), so that we obtain

∞

¸

k=1

1

k

1

e

kw

− 1

−

1

kw

÷

e

−kw

2

− log

1

√

2π

< Mw (6)

throughout [ arg w[ < c < π/2.

22 II. The Partition Function

The Approximation. We have prepared the way for the useful ap-

proximation to our generating function. All we need to do is combine

(1), (3), and (6), replace w by log

1

z

, and exponentiate. The result is

∞

¸

k=1

1

1 − z

k

=

1 − z

2π

exp

π

2

6 log

1

z

[1 ÷ O(1 − z)]

in

[1 − z[

1 − [z[

≤ c.

But we perform one more “neatening” operation. Thus log

1

z

is

an eyesore! It isn’t at all analytic in the unit disc, we must replace

it (before anything good can result). So note that, near 1, log

1

z

=

(1 −z) ÷

(1−z)

2

2

÷

(1−z)

3

3

÷· · · = 2

1−z

1÷z

÷O((1 −z)

3

), or

1

log

1

z

=

1

2

1÷z

1−z

÷ O(1 − z). Finally then,

∞

¸

k=1

1

1 − z

k

=

1 − z

2π

exp

π

2

12

1 ÷ z

1 − z

[1 ÷ O(1 − z)] (7)

in

[1 − z[

1 − [z[

≤ c.

This is our basic approximation. It is good near z = 1, which

we have decided is the most important locale. Here we see that

we can replace our generating function by the elementary function

1−z

2π

exp

π

2

12

1÷z

1−z

**whose coefﬁcients should then prove amenable.
**

However, (7) is really of no use away from z = 1, and, since

Cauchy’s theorem requires values of z all along a closed loop sur-

rounding 0, we see that something else must be supplied. Indeed we

will show that, away from 1, everything is negligible by comparison.

Riemann Sums 23

To see this, let us return to (2) and conclude that

log F(z) −

1

1 − z

<

∞

¸

j=2

1

j

[z[

j

1 − [z[

j

<

1

1 − [z[

∞

¸

j=2

1

j

1

j

=

1

1 − [z[

π

2

6

− 1

,

or

F(z) < exp

1

[1 − z[

÷

π

2

6

− 1

1

1 − [z[

, (8)

an estimate which is just what we need. It shows that, away from 1,

where

1

[1−z[

is smaller than

1

1−[z[

, F(z) is rather small.

Thus, for example, we obtain

F(z) < exp

1

[1 − z[

when

[1 − z[

1 − [z[

≥ 3. (9)

Also, in this same region, setting

φ(z) =

1 − z

2π

exp

π

2

12

1 ÷ z

1 − z

=

∞

¸

n=0

q(n)z

n

, (10)

φ(z) <

2

2π

exp

π

2

12

2

1 − z

< exp

π

2

12

2

3(1 − [z[)

so that

φ(z) < exp

1

1 − [z[

when

[1 − z[

1 − [z[

≥ 3. (11)

The Cauchy Integral. Armed with these preparations and the

feeling that the coefﬁcients of the elementaryfunctionφ(z) are acces-

sible, we launch our major Cauchy integral attack. So, to commence

the ﬁring, we write

p(n) − q(n) =

1

2πi

C

F(z) − φ(z)

z

n÷1

dz (12)

24 II. The Partition Function

and we try C a circle near the unit circle, i.e.,

C is [z[ = r, r < 1. (13)

Next we break up C as dictated by our consideration of

[1−z[

1−[z[[

,

namely, into

A is the arc [z[ = r,

[1 − z[

1 − [z[

≤ 3,

and (14)

B is the arc [z[ = r,

[1 − z[

1 − [z[

≥ 3.

So,

p(n) − q(n) (15)

=

1

2πi

A

F(z) − φ(z)

z

n÷1

dz ÷

1

2πi

B

F(z) − φ(z)

z

n÷1

dz,

and if we use (7) on this ﬁrst integral and (9), (11) on this second

integral we derive the following estimates:

1

2πi

A

F(z) − φ(z)

z

n÷1

dz

<

M

/

r

n÷1

(1 − r)

3/2

exp

π

2

6

1

1 − r

the length of A.

(M

/

is the implied constant in the O of (7) when c = 3).

As for the length of A, elementary geometry gives the formula

4r arcsin

√

2(1 − r)

√

r

and this is easily seen to be O(1 − r). We ﬁnally obtain, then,

1

2πi

A

F(z) − φ(z)

z

n÷1

dz

< M

(1 − r)

5/2

.

r

n

exp

π

2

6

1

1 − r

, (16)

where M is an absolute constant.

The Coefﬁcients of q(n) 25

For the second integral,

1

2πi

B

F(z) − φ(z)

z

n÷1

dz <

1

2πr

n÷1

· 2 exp

1

1 − r

· 2πr

=

2

r

n

exp

1

1 − r

.

And this is even smaller than our previous estimate. So combining

the two gives, by (15),

p(n) − q(n) < M

(1 − r)

5/2

r

n

exp

π

2

6

1

1 − r

. (17)

But what is r? Answer: anything we please (as long as 0 < r <

1)! We are masters of the choice, and so we attempt to minimize

the right-hand side. The exact minimum is too complicated but the

approximate one occurs when

1

e

n(r−1)

exp

π

2

6

1

1−r

**is minimized and
**

this occurs when

π

2

6

1

1−r

= n(1 − r), i.e., r = 1 −

π

√

6n

. So we

choose this r and, by so doing, we obtain, from (17), the bound

p(n) = q(n) ÷ O

n

−5/4

e

π

√

n/6

. (18)

The Coefﬁcients of q(n)

The elementary function φ(z) has a rather pleasant deﬁnite integral

representation which will then lead to a handy expression for the

q(n).

If we simply begin with the well-known identity

∞

−∞

e

−t

2

dt =

√

π

and make a linear change of variables (a > 0),

∞

−∞

e

−(at −b)

2

dt =

√

π

a

,

26 II. The Partition Function

or

∞

−∞

e

−a

2

t

2

e

÷2abt

dt =

√

π

a

e

b

2

.

Thus if we set b

2

=

π

2

6

1

1−z

and a

2

= 1 − z (thinking of z as real

([z[ < 1) for now), we obtain

∞

−∞

e

zt

2

e

÷π

√

2

3

t −t

2

dt =

√

π

√

1 − z

exp

π

2

6

1

1 − z

,

which gives, ﬁnally,

φ(z) =

e

−π

2

/12

π

√

2

(1 − z)

∞

−∞

e

zt

2

e

π

√

2

3

t −t

2

dt. (19)

Equating coefﬁcients therefore results in

q(n) =

e

−π

2

/12

π

√

2

∞

−∞

¸

t

2n

n!

−

t

2n−2

(n − 1)!

e

π

√

2/3 t −t

2

dt (20)

the “formula” for q(n) from which we can obtain asymptotics.

Reasoning that the maximumof the integrand occurs near t =

√

n

we change variables by t = s ÷

√

n, and thereby obtain

q(n) = C

n

∞

−∞

K

n

(s)2se

−2

s−

π

2

√

6

2

ds, (21)

where

C

n

=

e

π

√

2n/3

π

√

2n

n

n÷

1

2

e

n

n!

,

K

n

(s) =

1 ÷

s

2

√

n

1 ÷

s

√

n

2

¸

1 ÷

s

√

n

e

−s

√

n

÷

s

2

2n

¸

2n

.

Since K

n

(s) → 1, we see, at least formally, that the above integral

approaches

∞

−∞

2se

−2

s−

π

2

√

6

2

ds =

∞

−∞

u ÷

π

2

√

3

e

−u

2

du,

The Coefﬁcients of q(n) 27

where we have set s =

u

√

2

÷

π

2

√

6

. Furthermore, since ue

−u

2

is

odd, it is equal to

π

2

√

3

∞

−∞

e

−u

2

du =

π

√

π

2

√

3

. Thus (21) formally

becomes

q(n) ∼

e

π

√

2n/3

4

√

3n

√

2πnn

n

e

n

n!

. (22)

And score another one for Stirling’s formula, which in turn

gives

q(n) ∼

e

π

√

2n/3

4

√

3n

, (23)

and our earlier estimate (18) allows us thereby to conclude that

p(n) ∼

e

π

√

2n/3

4

√

3n

. (24)

Success! We have determined the asymptotic formula for p(n)!

Well, almost. We still have two debts outstanding. We must justify

our formal passage to the limit in (21), and we must also prove our

evaluation (5). So ﬁrst we observe that xe

−x

is maximized at x = 1,

so we deduce that

1 ÷

s

√

n

e

−s

√

n

≤ 1 (25)

(using x = (1 ÷

s

√

n

)) and also

1 ÷

s

√

n

e

−s

√

n

≤ e

s

2

2n

(26)

(using x = (1 ÷

s

√

n

)

2

).

Thus using (25) for positive s, by (21),

K

n

(s) ≤ e

s

2

for s ≥ 0,

28 II. The Partition Function

and using (26) for negative s gives us

[K

n

(s)[ ≤ (1 − s)e

s

2

−

2s

√

n

1 ÷

s

√

n

e

−s

√

n

2n−2

≤ (1 − s)e

s

2

−

2s

√

n

e

n−1

n

s

2

= (1 − s)e

2s

2

÷1−(1÷s/

√

n)

2

or

[K

n

(s)[ ≤ (1 − s)e

2s

2

÷1

for s < 0. (28)

Thus (27) and (28) give the bound for our integral in (21) of

2se

s

2

−2

s−

π

2

√

6

2

for s ≥ 0,

and

2s(s − 1)e

1÷π

√

2/3s

for s < 0.

This bound, integrable over (−∞, ∞), gives us the required

dominated convergence, and the passage to the limit is indeed

justiﬁed.

Finally we give the following:

Evaluation of our Integral (5). To achieve this let us ﬁrst note that

as N → ∞our integral is the limit of the integral

∞

0

(1 − e

−Nx

)

1

e

x

− 1

−

1

x

÷

e

−x

2

dx

x

(by dominated convergence, e.g.). But this integral can be split into

∞

0

(1 − e

−Nx

)

1

e

x

− 1

−

1

x

dx

x

÷

∞

0

(1 − e

−Nx

)

e

−x

2x

dx

=

N

¸

k=1

∞

0

e

−kx

1 ÷ x − e

x

x

2

dx ÷

1

2

∞

0

e

−x

− e

−(N÷1)x

x

dx.

Next note that

1 ÷ x − e

x

x

2

= −

1

0

t e

(1−t )x

dt

The Coefﬁcients of q(n) 29

and

e

−x

− e

−(N÷1)x

x

=

N÷1

1

e

−sx

ds.

Hence, by Fubini, we may interchange and obtain, for our expression,

the elementary sum

−

N

¸

k=1

1

0

t

k ÷ t − 1

dt ÷

1

2

N÷1

1

ds

s

=

N

¸

k=1

(k − 1) log

k

k − 1

− 1

÷

1

2

log(N ÷ 1)

=

N

¸

k=1

(k − 1) log k − (k − 1) log(k − 1) − N

÷

1

2

log(N ÷ 1)

= N log N − log N − log(N − 1) − · · · − log 1 − N

÷

1

2

log(N ÷ 1)

= N log N − log N! − N ÷

1

2

log(N ÷ 1).

What luck! This is equal to log

√

N÷1(N/e)

N

N!

and so, by Stirling’s

formula, indeed approaches log

1

√

2π

.

(Stirling’s formula was used twice and hence needn’t have been

used at all! Thus we ended up not needing the fact that C =

√

2π

in the formula n! ∼ C

√

n(n/e)

n

since the C cancels against a C in

the denominator. The n! formula with C instead of

√

2π is a much

simpler result.)

30 II. The Partition Function

Problems for Chapter II

1. Explain the observation that MacMahon made of a parabola when

he viewed the list of the (decimal expansions) of the partition

function.

2. Prove the “simple” fact that, if order counts (e.g., 2 ÷5 is consid-

ered a different partition of 7 than 5 ÷ 2), then the total number

of partitions on n would be 2

n−1

.

3. Explain the approximation “near 1” of log

1

z

as 2

1−z

1÷z

÷ O

(1 −

z)

3

**. Why does this lead to
**

1

log

1

z

=

1

2

1 ÷ z

1 − z

÷ O(1 − z)?

4. Why is the Riemann sum such a good approximation to the in-

tegral when the function is monotone and the increments are

equal?

III

The Erd˝ os–Fuchs Theorem

There has always been some fascination with the possibility of near

constancy of the representation functions r

i

(n) (of I (7), (8) and (9)).

In Chapter I we treated the case of r

÷

(n) and showed that this could

not eventuallybe constant. The fact that r(n) cannot be constant for an

inﬁnite set is really trivial since r(n) is odd for n = 2a, a ∈ A, and

even otherwise. The case of r

−

(n) is more difﬁcult, and we will treat

it in this chapter as an introduction to the analysis in the Erd˝ os–Fuchs

theorem.

The Erd˝ os–Fuchs theoreminvolves the question of just hownearly

constant r(n) can be on average. Historically this all began with the

set A = {n

2

: n ∈ N

0

}, the set of perfect squares, andthe observation

that then

r(0)÷r(1)÷r(2)÷···÷r(n)

n÷1

, the average value, is exactly equal to

1

n÷1

times the number of lattice points in the quarter disc x, y ≥ 0,

x

2

÷ y

2

≤ n. Consideration of the double Riemann integral shows

that this average approaches the area of the unit quarter circle, namely

π/4, and so for this set A,

r(0)÷r(1)÷r(2)÷···÷r(n)

n÷1

→

π

4

(r(n) is on

average equal to the constant π/4.)

The difﬁcult question is howquickly this limit is approached. Thus

fairly simple reasoning shows that

r(0) ÷ r(1) ÷ r(2) ÷ · · · ÷ r(n)

n ÷ 1

=

π

4

÷ O

1

√

n

,

whereas more involved analysis shows that

r(0) ÷ r(1) ÷ r(2) ÷ · · · ÷ r(n)

n ÷ 1

=

π

4

÷ O

1

n

2/3

.

31

32 III. The Erd˝ os–Fuchs Theorem

Very deep arguments have even improved this to o

1

n

2/3

, for ex-

ample, and the conjecture is that it is actually O

1

n

3

4

−

for every

> 0. On the other hand, further difﬁcult arguments show that it is

not O

1

n

3

4

÷

.

Now all of these arguments were made for the very special case

of A = the perfect squares. What a surprise then, when Erd˝ os and

Fuchs showed, by simple analytic number theory, the following:

Theorem. For any set A,

r(0)÷r(1)÷r(2)÷···÷r(n)

n÷1

= C ÷ O

1

n

3

4

÷

is

impossible unless C = 0.

This will be proved in the current chapter, but ﬁrst an appetizer.

We prove that r

−

(n) can’t eventually be constant.

So let us assume that

A

2

(z) − A(z

2

) = P(z) ÷

C

1 − z

, (1)

P is a polynomial, and C is a positive constant. Now look for a con-

tradiction. The simple device of letting z → (−1)

÷

which worked

so nicely for the r

÷

problem, leads nowhere here. The exercises in

Chapter I were, after all, hand picked for their simplicity and involved

only the lightest touch of analysis. Here we encounter a slightly heav-

ier dose. We proceed, namely, by integrating the modulus around a

circle. From (1), we obtain, for 0 ≤ r < 1,

π

−π

[A

2

(re

iθ

)[dθ

≤

π

−π

[A(r

2

e

2iθ

)[dθ ÷

π

−π

[P(re

iθ

)[dθ (2)

÷ C

π

−π

dθ

[1 − re

iθ

[

.

III. The Erd˝ os–Fuchs Theorem 33

Certain estimates are fairly evident. P(z) is a polynomial and so

π

−π

[P(re

iθ

)[dθ ≤ M, (3)

independent of r (0 ≤ r < 1).

We can also estimate the (elliptic) integral

π

−π

dθ

[1−re

iθ

[

=

2

π

0

dθ

[1−re

iθ

[

by the observation that if z is any complex number in

the ﬁrst quadrant, then [z[ ≤ +z ÷ `z. Thus since for 0 ≤ θ ≤ π,

1 − re

iθ

is in the ﬁrst quadrant,

ie

iθ

e

iθ

−r

=

i

1−re

−iθ

also is, and

1

[1−re

−iθ

[

=

ie

iθ

e

iθ

−r

≤ (+ ÷ `)

ie

iθ

e

iθ

−r

. Hence

π

0

dθ

[1 − re

iθ

[

≤ (+ ÷ `)

π

0

ie

iθ

e

iθ

− r

dθ

= (+ ÷ `)

log(e

iθ

− r)

π

0

= (+ ÷ `) log

−

1 ÷ r

1 − r

= π ÷ log

1 ÷ r

1 − r

.

The bound, then, is

π

−π

dθ

[1 − re

iθ

[

≤ 2π ÷ 2 log

1 ÷ r

1 − r

. (4)

The integral

π

−π

[A(re

iθ

)[

2

dθ is a delight. It succumbs to Parseval’s

identity. This is the observation that

π

−π

[

¸

a

n

e

inθ

[

2

dθ =

π

−π

¸

a

n

e

inθ

¸

¨ a

m

e

−imθ

dθ

=

π

−π

¸

m,n

a

n

¨ a

m

e

i(n−m)θ

dθ

=

¸

n,m

a

n

¨ a

m

π

−π

e

i(n−m)θ

dθ

and these integrals all vanish except that, when n = m, they are

equal to 2π. Hence this double sum is 2π

¸

[a

n

[

2

. The derivation is

clearly valid for ﬁnite or absolutely convergent series which covers

34 III. The Erd˝ os–Fuchs Theorem

our case of A(re

iθ

) (but it even holds in much greater “miraculous”

generalities).

At any rate, Parseval’s identity gives us

π

−π

[A(re

iθ

)[

2

dθ = 2π

¸

a∈A

r

2a

= 2πA(r

2

). (5)

The last integral we must cope with is

π

−π

[A(r

2

e

2iθ

)[dθ, and,

unlike integrals of [f [

2

, there is no formula for integrals of [f [. But

there is always the Schwarz inequality

[f [ ≤ (

1 ·

[f [

2

)

1/2

, and

so at least we can get an upper bound for such integrals, again by

Parseval. The conclusion is that

π

−π

[A(r

2

e

2iθ

)[dθ ≤ 2π

A(r

4

). (6)

All four of the integrals in (2) have been spoken for and so, by (2)

through (6), we obtain

A(r

2

) ≤

A(r

4

) ÷

M

2π

÷ C ÷

C

π

log

1 ÷ r

1 − r

. (7)

It is a nuisance that our function A is evaluated at two different

points, but we can alleviate that by the obvious monotonicity of A,

A(r

4

) ≤ A(r

2

), and obtain

A(r

2

) ≤

A(r

2

) ÷ M

/

÷

C

π

log

1 ÷ r

1 − r

. (8)

Is something bounded in terms of its own square root? But if x ≤

√

x ÷a, we obtain (

√

x −

1

2

)

2

≤ a ÷

1

4

,

√

x ≤

a ÷

1

4

÷

1

2

, x ≤

a ÷

1

2

÷

a ÷

1

4

. This yields a pure bound on x. Then

A(r

2

) ≤ M

//

÷

C

π

log

1 ÷ r

1 − r

÷

M

///

÷

C

π

log

1 ÷ r

1 − r

. (9)

But, so what? This says that A(r

2

) grows only at the order of

log

1

1−r

as r → 1

−

, but it doesn’t say that A(r

2

) remains bounded,

does it? Wherein is the hoped contradiction? We must revisit (1)

for this. Thereby we obtain, in turn A

2

(r

2

) − A(r

4

) = P(r

2

) ÷

Erd˝ os–Fuchs Theorem 35

C

1−r

2

, A

2

(r

2

) ≥ P(r

2

) ÷

C

1−r

2

, A

2

(r

2

) ≥ −M ÷

C

1−r

2

, and ﬁnally

A(r

2

) ≥

−M ÷

C

1 − r

2

, (10)

a rate of growth which ﬂatly contradicts (9) and so gives our desired

contradiction.

If this proof seems like just so much sleight of hand, let us ob-

serve what is “really” going on. We ﬁnd ourselves with a set A

whose r

i

(n) is “almost” constant and this means that A

2

(z) ≈

C

1−z

.

On the one hand, this forces A(z) to be large on the positive axis

A(r

2

) >

C

/

√

1−r

2

**, and, on the other hand Parseval says that the
**

integral of [A

2

(z)[ is A(r

2

) and

C

1−z

**(being fairly small except near
**

1) has a small integral, only O(log

1

1−r

). (So A(r

2

) < C

//

log

1

1−r

).

In cruder terms, Parseval tells us that A

2

(z) is large on average,

so it must be large elsewhere than just near z = 1, and so it cannot

really be like

C

1−z

. (Note that the “elsewhere” in the earlier r

÷

(n)

problem was the locale of −1, and so even that argument seems to

be in this spirit.)

So let us turn to the Erd˝ os–Fuchs theorem with the same strategy

in mind, viz., to bound A(r

2

) below by

C

/

√

1−r

2

for obvious reasons

and then to bound it above by Parseval considerations.

Erd˝ os–Fuchs Theorem

We assume the A is a set for which

r(0) ÷ r(1) ÷ · · · ÷ r(n) = C(n ÷ 1) ÷ O(n

α

), C > 0, (11)

and we wish to deduce that α ≥

1

4

. As usual, we introduce the

generating function A(z) =

¸

a∈A

z

a

, so that A

2

(z) =

¸

r(n)z

n

,

and therefore

1

1−z

A

2

(z) =

¸

[r(0) ÷ r(1) ÷ · · · ÷ r(n)]z

n

. Since

¸

(n ÷ 1)z

n

=

1

(1−z)

2

our hypothesis (11) can be written as

1

1 − z

A

2

(z) =

C

(1 − z)

2

÷

∞

¸

n=0

a

n

z

n

, a

n

= O(n

α

),

36 III. The Erd˝ os–Fuchs Theorem

or

A

2

(z) =

C

1 − z

÷ (1 − z)

∞

¸

n=0

a

n

z

n

, a

n

= O(n

α

). (12)

Of course we may assume throughout that α < 1. Thereby (12)

yields the boundM(1−r

2

)

−α−1

for

¸

a

n

r

2n

, sothat we easilyachieve

our ﬁrst goal namely,

A(r

2

) >

C

/

√

1 − r

2

, C

/

> 0. (13)

As for the other goal, the Parseval upper bound on A(r

2

), again

we wish to exploit the fact that A

2

(z) is “near”

C

1−z

, but this takes

some doing. From the look of (12) unlike (1), this “nearness” seems

to occur only where (1 − z)

¸

a

n

z

n

is relatively small, that is, only

in a neighborhood of z = 1. We must “enhance” this locale if we are

to expect anything from the integration, and we do so by multiplying

by a function whose “heft” or largeness is all near z = 1. A handy

such multiplier for us is the function S

2

(z) where

S(z) = 1 ÷ z ÷ z

2

÷ · · · ÷ z

N−1

, N large. (14)

The multiplication of S

2

(z) by (12) yields

[S(z)A(z)]

2

=

CS

2

(z)

1 − z

÷ (1 − z

N

)S(z)

¸

a

n

z

n

, (15)

which gives

[S(z)A(z)[

2

≤

CN

2

[1 − z[

÷ 2[S(z)

¸

a

n

z

n

[, (16)

and integration leads to

π

−π

[S(re

iθ

)A(re

iθ

)[

2

dθ

≤ CN

2

π

−π

dθ

[1 − re

iθ

[

(17)

÷ 2

π

−π

[S(re

iθ

)

¸

a

n

(re

iθ

)

n

[dθ.

Erd˝ os–Fuchs Theorem 37

As before, we will use Parseval on the ﬁrst of these integrals, (4)

on the second, and Schwarz’s inequality together with Parseval on

the third.

So write S(z)A(z) =

¸

c

n

z

n

, and conclude that

π

−π

[S(re

iθ

)

A(re

iθ

)[

2

dθ = 2π

¸

[c

n

[

2

r

2n

. Since the c

n

are integers, [c

n

[

2

=

c

2

n

≥ c

n

andsothis is, furthermore, ≥ 2π

¸

c

n

r

2n

= 2πS(r

2

)A(r

2

).

(The general fact then is that, if F(z) has integral coefﬁcients,

π

−π

[F(re

iθ

)[

2

dθ ≥ 2πF(r

2

).)

Now we introduce a side condition on our parameters r and N

which we shall insist on henceforth namely that

1

1 − r

2

≥ N. (18)

Thus, by (14), S(r

2

) > Nr

2N

≥ N(1−

1

N

)

N

≥ N(1−

1

2

)

2

=

N

4

,

and by (13), A(r

2

) >

C

/

√

1−r

2

, and we conclude that

π

−π

[S(re

iθ

)A(re

iθ

)[

2

dθ >

C

//

N

√

1 − r

2

, C

//

> 0. (19)

Next, (4) gives

CN

2

π

−π

dθ

[1 − re

iθ

[

≤ MN

2

log

e

1 − r

2

(20)

and our last integral satisﬁes

π

−π

S(re

iθ

)

¸

a

n

(re

iθ

)

n

dθ

≤

π

−π

S(re

iθ

)

2

dθ

π

−π

¸

a

n

(re

iθ

)

n

2

dθ

= 2π

¸

k<N

r

2k

¸

[a

n

[

2

r

2n

≤ 2π

√

NM

¸

n

2α

r

2n

.

Applying (13) and (14) again leads ﬁnally to

π

−π

S

re

iθ

¸

a

n

(re

iθ

)

n

dθ ≤

M

√

N

(1 − r

2

)

α÷1/2

. (21)

38 III. The Erd˝ os–Fuchs Theorem

At last, combining (19), (20), and (21) allows the conclusion

C

//

M

≤ N

1 − r

2

log

e

1 − r

2

÷

1

√

N(1 − r

2

)

α

. (22)

Once again we are masters of the parameters (subject to (18)),

and so we elect to choose r, so that N

√

1 − r

2

=

1

√

N(1−r

2

)

α

. Thus

our choice is to make

1

1−r

2

= N

3

2α÷1

and note happily that our side

condition (18) is satisﬁed. Also “plugging” this choice into (22) gives

C

//

M

≤ N

4α−1

4α÷2

(2 ÷ 3 log N). (23)

Well, success is delicious. We certainly see in (23) the fact that

α ≥

1

4

. (If the exponent of N,

4α−1

4α÷2

, were negative then this right-

hand side would go to 0, 2÷3 log N notwithstanding, and (23) would

become false for large N.)

Problems for Chapter III 39

Problems for Chapter III

1. Showthat the number of lattice points in x

2

÷y

2

≤ n

2

, x, y ≥ 0,

is ∼

π

4

n

2

. By the Riemann integral method show that it is, in fact

=

π

4

n

2

÷ O(n).

2. If x is bounded by its own square root (i.e., by

√

x ÷ a), then we

ﬁnd that it has a pure bound. What if x, instead, is bounded by

x

2/3

÷ ax

1/3

÷ b? Does this insure a bound on x?

3. Suppose that a convex closed curve has its curvature bounded by

δ. Show that it must come within 2

√

δ of some lattice point.

4. Produce a convex closed curve with curvature bounded by δ which

doesn’t come within

√

δ

1200

of any lattice point.

IV

Sequences without Arithmetic

Progressions

The gist of the result of Chapter IV is that a sequence of integers

with “positive density” must contain an arithmetic progression (of at

least three distinct terms).

More precisely and in sharper, ﬁnitized form, this is the statement

that, if > 0, then for large enough n, any subset of the nonnegative

integers below n with at least n members must contain three terms

a, b, c where a < b < c and a ÷c = 2b. This is a shock to nobody.

If a set is “fat” enough, it should contain all sorts of patterns. The

shock is that this is so hard to prove.

At any rate we begin with a vastly more general consideration, the

notion of an “afﬁne property” of ﬁnite sets of integers. So let us agree

to call a property P an afﬁne property if it satisﬁes the following two

conditions:

1. For each ﬁxed pair of integers α, β with α ,= 0, the set A(n) has

P if and only if αA(n) ÷ β has P.

2. Any subset of a set, which has P, also has P.

Thus, for example, the property P

A

of not containing any arith-

metic progressions is an afﬁne property. Again the trivial property

P

0

of just being any set is an afﬁne one.

Now we ﬁx an afﬁne property P and consider a largest subset of

the nonnegative integers below n, which has P. (Thus we require

that this set has the most members possible, not just to be maximal.)

There may be several such sets but we choose one of themand denote

it by S(n:P). We also denote the number of elements of this set by

41

42 IV. Sequences without Arithmetic Progressions

f (n:P). So, for example, for the trivial property, f (n:P

0

) = n, and

for P

A

, f (3:P

A

) = 2, f (5:P

A

) = 4.

It follows easily from conditions 1 and 2 that this f (n) is sub-

additive, i.e., f (m ÷ n) ≤ f (m) ÷ f (n). If we recall the fact

that subadditive functions enjoy the property that lim

n→∞

f (n)

n

ex-

ists (in fact lim

n→∞

f (n)

n

= inf

f (n)

n

), we are led to deﬁne C

P

=

lim

n→∞

f (n:P)

n

. This number is a measure of how permissive the

property P is. Thus C

P

0

= 1, because P

0

is totally permissive. The

announced result about progression = free sequences amounts to the

statement that C

P

A

= 0, so that P

A

is, in this sense, totally unper-

missive. At any rate, we always have 0 ≤ C

P

≤ 1, and we may dub

C

P

the permission constant.

The remarkable result proved by Szemer´ edi and then later by

Furstenberg is that, except for P

0

, C

P

is always 0. Their proofs are

both rather complicated, and we shall content ourselves with the case

of P

A

, which was proved by Roth.

The Basic Approximation Lemma

It turns out that the extremal sets S(n:P) all behave very much as

though their elements were chosen at random. For example, we note

that such a set must contain roughly the same number of evens

as odds. Indeed if 2b

1

, 2b

2

, . . . , 2b

k

were its even elements, then

b

1

, b

2

, . . . , b

k

would be a subset of

0,

n

2

**and so we could conclude
**

that k ≤ f

n

2

**. Similarly the population of the odd elements of
**

S would satisfy this same inequality. Since

n

2

∼

1

2

f (n), we con-

clude that both the evens and the odds contain not much more than

half the whole set. Thereby the evens and the odds must be roughly

equinumerous. (Thus, two upper bounds imply the lower bounds.)

Delaying for the moment the precise statement of this “random-

ness,” let us just note how it will prove useful to us with regard to

our arithmetic progression considerations. The point is simply that,

if integers were chosen truly at random with a probability C > 0,

there would automatically be a huge number of arithmetic progres-

The Basic Approximation Lemma 43

sions formed. So we expect that even an approximate randomness

should produce at least one arithmetic progression.

The precise assertion is that of the following lemma.

Lemma.

¸

a∈S(n:P)

z

a

= C

P

¸

k≤n

z

k

÷o(n), uniformly on [z[ = 1.

Remark. In terms of the great Szemer´ edi–Furstenberg result that

C

P

≡ 0 (except for P = P

0

), this is a total triviality. We are proving

what in truth is an empty result. Nevertheless we are not prepared

to give the lengthy and complex proofs of this general theorem, and

so we must prove the Lemma. (We do what we can.) The proof, in

fact, is really just an elaboration of the odds and evens considerations

above.

Proof. The basic strategy is to estimate q

n

(z) =

¸

a∈S

z

a

−

C

P

¸

k<n

z

k

, together with all of its partial sums at every root of

unity of order up to N (N is a parameter to be chosen later). The

point is that, if we have a bound on a polynomial and its partial sums

at a point, then we inherit a bound on that polynomial throughout

an arc around that point. (Thereby we will obtain bounds for arcs

between the roots of unity which will ﬁll up the whole circle.)

Speciﬁcally, we have the identity

p(z)

1 −

z

ζ

=

¸

m<n

p

m

(ζ )

z

ζ

m

÷

p(ξ)

1 −

z

ξ

z

ξ

n

, (1)

for any polynomial p of degree at most n, where the p

m

denote the

partial sums. (This simply records the result of the “long division.”)

From (1) we easily obtain the bound [p(z)[ ≤ [ζ − z[

¸

m<n

[p

m

(ζ )[ ÷ [p(ζ )[, and so we conclude the following:

If all the partial sums are bounded by M at ζ, the polynomial is

bounded by M(n ÷ 1)throughout an arc of length 2 (2)

centered at ζ.

44 IV. Sequences without Arithmetic Progressions

So let α ≤ N be chosen, and let ω be any αth root of unity, i.e.,

ω

α

= 1. To estimate q

m

(ω), let us write it as

α

¸

β=1

ω

β

¸

¸

¸

¸

a∈S

a<m

a≡β(α)

1 − C

P

¸

k<m

k≡β(α)

1

,

and let us note that the ﬁrst inner sum

σ

β

=

¸

a∈S

a<m

a≡β(α)

1

counts the size of a subset of S, which therefore has P which is afﬁne

to a subset of

0,

m

α

, and so has at most f

m

α

elements (where we

write f (x) for f (!x|)).

Thus

q

m

(ω) = −

α

¸

β=1

ω

β

f

m

α

− σ

β

÷

α

¸

β=1

ω

β

¸

f

m

α

− C

P

¸

k<m

k≡β(α)

1

¸

≤

α

¸

β=1

f

m

α

− σ

β

÷

α

¸

β=1

f

m

α

− C

P

¸

m

α

(3)

=

α

¸

β=1

f

m

α

− σ

β

÷

α

¸

β=1

f

m

α

− C

P

m

α

= 2αf

m

α

−

α

¸

β=1

σ

β

− C

P

m.

If we next note that

¸

α

β=1

σ

β

is exactly the number of elements of

S which are below m and so is equal to f (n) minus the number of

elements of S which are ≥ m, we obtain

α

¸

β=1

σ

β

≥ f (n) − f (n − m) ≥ C

P

n − f (n − m). (4)

The Basic Approximation Lemma 45

Substituting (4) in (3) gives

q

m

(ω) < 2α

¸

f

m

α

− C

P

m

α

¸

÷(f (n−m)−C

P

(n−m)). (5)

Now we ﬁnd it useful to replace the function f (x) − C

P

x by its

“monotone majorant” F(x) = max

t ≤x

(f (t )−C

P

t ) andnote that this

F(x) is nondecreasing and satisﬁes F(x) = o(x) since f (x) −C

P

x

satisﬁes the same. So (5) can be replaced by

q

m

(ω) < 2αF

m

α

÷ F(n − m) ≤ 2αF

n

α

÷ F(n) (6)

(a bound independent of m).

So choose n

0

so that x ≥ n

0

implies F(x) ≤ x, and then choose

n

1

so that x ≥ n

1

implies F(x) ≤

n

0

x. From now on we will pick

n ≥ n

1

and also will ﬁx N = [

n

n

0

].

Dirichlet’s theorem

1

on approximation by rationals now tells us

that the totality of arcs surrounding these ω with length 2

2π

α(N÷1)

covers the whole circle. Thus using (2) for q(z), ζ = ω and =

2

2π

α(N÷1)

≤ 2

2πn

0

nα

gives

q(z) < [2αF

n

α

÷ F(n)]

1 ÷ 2π

n

0

α

. (7)

We separate two cases:

Case I: α ≤ n

0

. Here we use F(

n

α

) ≤ F(n) and obtain [2αF(

n

α

) ÷

F(n)](1÷

2πn

0

α

) ≤ (2α÷1)(1÷

2πn

0

α

)F(n) ≤ 3α(1÷

2πn

0

α

)F(n) =

(3α ÷ 6πn

0

)F(n) ≤ (6π ÷ 3)n

0

F(n) ≤ (6π ÷ 3)n

0

n

0

n ≤ 22n.

Case II: α > n

0

. Here [2αF(

n

α

) ÷F(n)](1 ÷

2πn

0

α

) ≤ [2αF(

n

α

) ÷

F(n)](1 ÷ 2π). But still α ≤

n

n

0

, or

n

α

≥ n

0

. So F(

n

α

) ≤

n

α

, and

the above is ≤ (2n ÷ n)(1 ÷ 2π) = (3 ÷ 6π)n < 22n.

In either case Dirichlet’s theorem yields our lemma.

So let P be any afﬁne property, and denote by A = A(n:P) the

number of arithmetic progressions from S(n:P)(where order counts

1

Dirichlet’s theoremcan be proved by considering the powers 1, z, z

2

, · · ·, z

N

for z

any point on the unit circle. Since these are N ÷1 points on the circle, two of them

z

i

, j

j

must be within arc length

2π

N÷1

of one another. This means [ arg z

i−j

[ ≤

2π

N÷1

and calling [i − j[ = α gives the result.

46 IV. Sequences without Arithmetic Progressions

and equality is allowed). We show that

A(n:P) =

C

3

P

2

n

2

÷ o(n

2

). (8)

The proof is by contour integration. If we abbreviate

¸

a∈S

z

a

=

g(z), then we recognize A as the constant term in g(z)g(z)g(z

−2

),

and so we may write

A =

1

2πi

[z[=1

g

2

(z)g(z

−2

)

dz

z

. (9)

Now writing G(z) =

¸

k<n

z

k

, g(z) = C

P

G(z) ÷ q(z) (where q

is “small” by the lemma). If we substitute this in (9), we obtain

C

3

P

1

2πi

[z[=1

G

2

(z)G(z

−2

)

dz

z

plus seven other integrals. Each of these other integrals is the product

of three functions, each a Gor a q, and at least one of them is a q. By

our lemma, then, we may estimate each of these seven integrals by

o(n) times an integral of the product of two functions. Both of these

functions are either a [G[ or a [q[. As such each is estimable by the

Schwarz inequality, Parseval equality techniques. The ﬁnal estimate

for each of these seven integrals, therefore, is o(n)

√

nn = o(n

2

),

and so (9) gives

A = C

3

P

1

2πi

[z[=1

G

2

(z)G(z

−2

)

dz

z

÷ o(n

2

). (10)

But reading (9) for the property P

0

shows that this integral is

simply A(n:P

0

) and it is a simple exercise to show that A(n:P

0

),

the number of triples below n which are in arithmetic progression,

is exactly !

n

2

2

|. Indeed, then (10) reduces to (8). Q.E.D.

All of our discussion thus far has been quite general and is valid

for arbitrary afﬁne properties. We ﬁnally become speciﬁc by letting

P = P

A

, and we easily deduce the following:

Theorem (Roth). C

P

A

= 0.

The Basic Approximation Lemma 47

Proof. By the deﬁnition of P

A

, the only arithmetic progressions

in S(n:P

A

) are the trivial ones, three equal terms, which number is

at most n. Thus A(n:P

A

) ≤ n, and so, by (8), C

3

P

A

n

2

2

÷ o(n

2

) ≤ n.

Therefore C

P

A

= 0.

48 IV. Sequences without Arithmetic Progressions

Problems for Chapter IV

1. Attach a positive rational to each integer from 1 to 12 so that all

A.P.’s with common difference d up to 6 obtain their “correct”

measure

1

d

.

2. Prove that, if we ask for a generalization of this, then we can only

force the correct measure

1

d

for all A.P.’s of common difference

d, by attaching weights onto 1, 2, . . . , n, if d = O(

√

n).

3. If we insist only on approximation, however, show that we can

always attach weights onto 1, 2, . . . , n such that the “measure”

given to every A.P. with common difference ≤ m is within e

−n/m

of

1

a

.

V

The Waring Problem

In a famous letter to Euler, Waring wrote his great conjecture about

sums of powers. Lagrange had already proved his magniﬁcent the-

orem that every positive integer was the sum of four squares, and

Waring guessed that this was not just a property of squares, but that,

in fact, the sum of a ﬁxed number of cubes, fourth powers, ﬁfth pow-

ers, etc., also worked. He guessed that every positive integer was

the sum of 9 cubes, 19 fourth powers, 37 ﬁfth powers, and so forth,

and although no serious guess was made as to how the sequence 4

(squares), 9, 19, 37, . . . went on, he simply stated that it did! That

is what we propose to do in this chapter, just to prove the existence

of the requisite number of the cubes, fourth powers, etc. We do not

attempt to ﬁnd the structure of the 4, 9, 19, . . . , but just to prove its

existence.

So let us ﬁx k and view the kth powers. Our aim, by Schnirel-

mann’s lemmas below, need be only to produce a g = g(k) and an

α = α(k) > 0 such that the sum of g(k) kth powers represents at

least the fraction α(k) of all of the integers.

One of the wonderful things about this approach is that it requires

only upper bounds, despite the fact that Waring’s conjecture seems

to require lower bounds, something seemingly totally impossible

for contour integrals to produce. But the adequate upper bounds are

obtained by the so called Weyl sums given below.

So ﬁrst we turn to our three basic lemmas which will eventually

yield our proof. These are A, the theorem of Dirichlet, B, that of

Schnirelmann, and ﬁnally C, the evaluation of the Weyl sums.

49

50 V. The Waring Problem

A. Theorem (Dirichlet). Given a real x and a positive integer M,

there exists an integer a and a positive number b ≤ M such that

[x −

a

b

[ ≤

1

(M÷1)b

.

Proof. Consider the numbers 0, x, 2x, 3x, . . . , Mx all reduced

(mod 1). Clearly, two of these must be within

1

M÷1

of each other.

If these two differ by bx, then 1 ≤ b ≤ M and bx (mod 1) is, in

magnitude, ≤

1

M÷1

. Next pick an integer a that makes bx −a equal

to bx (mod 1). So [bx −a[ ≤

1

(M÷1)

which means [x −

a

b

[ ≤

1

(M÷1)b

,

as asserted. Q.E.D.

We also point out that this is a best possible result as the choice

x =

1

M÷1

shows for every M. (Again, we may assume that (a, b) =

1 for, if they have a common divisior, this would make the inequality

[b[ ≤ M even truer).

B. Schnirelmann’s Theorem. If S is a set of integers with positive

Schnirelmann density and 0 ∈ S, then every non-negative integer is

the sum of at most k members of S for some k ≥ 1.

Lemma 1. Let S have density α and 0 ∈ S. Then S ⊕S has density

at least 2α − α

2

.

Proof. All the gaps inthe set S are coveredinpart bythe translation

of S by the term of S just before this gap. Hence, at least the fraction

α of this gap gets covered. So from this covering we have density α

from S itself and α times the gaps. Altogether, then, we indeed have

α ÷ α(1 − α) = 2α − α

2

, as claimed.

Lemma 2. If S has density α >

1

2

, then S ⊕ S contains all the

positive integers.

Proof. Fix an integer n which is arbitrary, let A be the subset of

S which lies ≤ n, and let B be the set of all n minus elements of

S. Since A contains more than n/2 elements and B contains at least

n/2 elements, the Pigeonhole principle guarantees that they overlap.

So suppose they overlap at k. Since k ∈ A, we get k ∈ S, and since

V. The Waring Problem 51

k ∈ B, we get n − k ∈ S. These are the two elements of S which

sum to n.

Repeating Lemma 1 j times, then, leads to a summing of 2

j

copies

of S and a density of 1 −(1 −α)

2

j

or more. Since this latter quantity,

for large enough j, will become bigger than

1

2

, Lemma 2 tells us

that 2

j÷1

copies of S give us all the integers, just as Schnirelmann’s

theorem claims. Q.E.D.

C. Evaluation of Weyl Sums. Let b ∈ Z, b ,= 0 and k ≤ N,

P(n) be a polynomial of degree k with real coefﬁcients and leading

coefﬁcient integral and prime to b, and let I be an interval of length

≤ N. Then

¸

n∈I

e

P(n)

b

< N

1÷o(1)

b

−2

1−k

where the bound depends on k.

Here – as usual – we denote e(x) = e

2πix

.

We proceed by induction on k, which represents the degree of

P(n). It is clearly true for k = 1, and generally we may write

S =

¸

n∈I

e

P(n)

b

**and may assume w.l.o.g. that I = {1, 2, 3, . . . , N}. Thereby
**

[S[

2

=

N−1

¸

j=−N÷1

¸

n∈{1,2,...,N}

n∈{j÷1,j÷2,...,j÷N}

e

P(n) − P(n − j)

b

.

This inner sum involves a polynomial of degree (k − 1) but has a

leading coefﬁcient which varies with j. If we count those j which

produce a denominator of d, which of course must divide b, then we

observe that this must appear roughly d times in an interval of length

b. So this number of j in the full interval of length 2N ÷1 is roughly

(2N÷1)

b

d.

52 V. The Waring Problem

The full estimate, then, by the inductive hypothesis is

[S[

2

<

¸

d[b

N

b

dN

1÷o(1)

d

−

1

2

k−2

≤

N

2÷o(1)

b

b

1−

1

2

k−2

¸

d[b

1

< N

2÷o(1)

b

−

1

2

k−2

b

o(1)

.

So we obtain

S < N

1÷o(1)

b

−

1

2

k−1

,

and the induction is complete.

Now we continue as follows:

Lemma 3. Let k > 1 be a ﬁxed integer. There exists a C

1

such that,

for any positive integers N, a, b with (a, b) = 1,

N

¸

n=1

e

a

b

n

k

≤ C

1

N

1÷o(1)

b

−2

1−k

.

Our endpoint will be the following:

Theorem. If, for each positive integer s, we write

r

s

(n) =

¸

n

k

1

÷···÷n

k

s

=n

n

i

≥0

1,

then there exists g and C such that r

g

(n) ≤ Cn

g/k−1

for all n > 0.

The previously cited notions of Schnirelmann allow deducing, the

full Waring result from this theorem:

There exists a G for which r

G

(n) > 0 for all n > 0.

To prove our theorem, since

r

s

(n) =

1

0

¸

¸

m≤n

1/k

e(xm

k

)

¸

s

e(−nx)dx,

V. The Waring Problem 53

it sufﬁces to prove that there exists g and C for which

1

0

N

¸

n=1

e(xn

k

)

g

dx ≤ CN

g−k

for all n > 0. (1)

First some parenthetical remarks about this inequality. Suppose it is

known to hold for some C

0

and g

0

. Then, since [

¸

N

n=1

e(xn

k

)[ ≤ N,

it persists for C

0

and any g ≥ g

0

. Thus (1) is a property of large g’s,

in other words, it is purely a “magnitude property.” Again, (1) is a

best possible inequality in that, for each g, there exists a c > 0 such

that

1

0

N

¸

n=1

e(xn

k

)

g

dx > cN

g−k

for all n > 0. (2)

To see this, note that

¸

N

n=1

e(xn

k

) has a derivative bounded by

2πN

k÷1

. Hence, in the interval (0,

1

4πN

k

),

N

¸

n=1

e(xn

k

)

≥ N − 2πN

k÷1

1

4πN

k

=

N

2

,

and so (2) follows with c =

1

4π2

g

.

The remainder of our paper, then, will be devoted to the derivation

of (1) from Lemma 3. Henceforth k is ﬁxed. Denote by I

a,b,N

the

x-interval [x −

a

b

[ ≤

1

bN

k−1/2

, and call J = N

k

[x −

a

b

[, j = [J],

where a, b, N, j are integers satisfying N > 0, b > 0, 0 ≤ a < b,

(a, b) = 1, b ≤ N

k−

1

2

.

By Dirichlet’s theorem, these intervals cover (0, 1). Our main tool

is the following lemma:

Lemma 4. There exists > 0 and C

2

such that, throughout any

interval I

a,b,N

,

N

¸

n=1

e(xn

k

)

≤

C

2

N

(b ÷ j)

.

54 V. The Waring Problem

Proof. This is almost trivial if b > N

2/3

, for, since the derivative

of [

¸

N

n=1

e(xn

k

)[ is bounded by 2πN

k÷1

,

N

¸

n=1

e(xn

k

)

≤

N

¸

n=1

e(

a

b

n

k

)

÷

x −

a

b

2πN

k÷1

≤

N

1÷o(1)

b

1

2

k−1

÷

2πN

3/2

b

≤

N

1÷o(1)

b

1

2

k−1

÷

2πN

b

1/4

,

by C, which gives the result, since j = 0 automatically. Assume

therefore that b ≤ N

2/3

, and note the following two simple facts (A)

and (B). For details see [K. Knopp, Theory and Application of Inﬁnite

Series, Blackie &Sons, Glasgow, 1946.] and [G. P´ olya und G. Szeg¨ o,

Aufgaben und Lehrs¨ atze aus der Analysis, Dover Publications, New

York 1945, Vol. 1, Part II, p. 37]. Q.E.D.

(A) If M is the maximum of the moduli of the partial sums

¸

m

n=1

a

n

,

V the total variation of f (t ) in 0 ≤ t ≤ N, and M

/

the maximum

of the modulus of f (t ) in 0 ≤ t ≤ N, then

N

¸

n=1

a

n

f (n)

≤ M(V ÷ M

/

).

(B) If V is the total variation of f (t ) in 0 ≤ t ≤ N, then

N

¸

n=1

f (n) −

N

0

f (t )dt

≤ V.

Now write α =

1

b

¸

b

n=1

e(

a

b

n

k

) and

N

¸

n=1

e(xn

k

) = S

1

÷ αS

2

, (3)

where

S

1

=

N

¸

n=1

¸

e

a

b

n

k

− α

¸

e

¸

x −

a

b

n

k

¸

,

S

2

=

N

¸

n=1

e

¸

x −

a

b

n

k

¸

.

V. The Waring Problem 55

We apply (A) to S

1

. To do so, we note that

m

¸

n=1

¸

e

a

b

n

k

− α

¸

=

0 ÷

¸

b[m/b]<n≤m

¸

e

a

b

n

k

− α

¸

≤ (1 ÷ [α[)b ≤ 2b.

Also, the total variation of e[(x −

a

b

)t

k

] is equal to 2π[x −

a

b

[N

k

≤

2π

√

N

b

, whereas M

/

= 1. The result is

[S

1

[ ≤ 4π

√

N ÷ 2b ≤ 5πN

2/3

. (4)

Next we apply (B) to S

2

and obtain

[S

2

[ ≤

N

0

e

¸

x −

a

b

t

k

¸

dt

÷

2π

√

N

b

. (5)

Since

∞

0

e(u

k

)du converges we get

N

0

e

¸

x −

a

b

t

k

¸

dt

=

N

J

1/k

J

1/k

0

e(u

k

)du

≤

NC

3

J

1/k

.

Combining this with (5) gives

[αS

2

[ ≤

C

4

N[α[

(1 ÷ j)

1/k

÷ 2π

√

N. (6)

Nowif we apply Lemma 3 to the case N = b, we obtain [α[ ≤

C

1

b

δ

,

δ = 2

1−k

, and by (3) the addition of (4) and (6) gives

N

¸

n=1

e(xn

k

)

≤

C

5

N

b

δ

(1 ÷ j)

1/k

÷ 7πN

2/3

≤

C

5

N

b

δ

(1 ÷ j)

1/k

÷

C

6

N

(b ÷ j)

1/2

.

Since j ≤

√

N andb ≤ N

2/3

, the choice C

2

= C

5

÷C

6

÷C

1

÷2π,

= min

δ,

1

k

,

1

4

**completes the proof.
**

56 V. The Waring Problem

Proof of (1). Choose g ≥

4

**, given as above. By Lemma 4, since
**

the length of each I

a,b,N

is at most 2N

−k

,

I

a,b,N

N

¸

n=1

e(xn

k

)

g

dx ≤

C

7

N

g

(b ÷ j)

4

1

N

k

.

Summing over all a, b, j gives the estimate

C

7

N

g−k

¸

b,j

b

(b ÷ j)

4

≤ CN

g−k

since

¸

∞

b=1

¸

∞

j=0

1

(b÷j)

3

< ∞, and the proof is complete.

Problems for Chapter V 57

Problems for Chapter V

1. If we permit polynomials with arbitrary complex coefﬁcients and

ask the “Waring” problem for polynomials, then show that x is

not the sum of 2 cubes, but it is the sum of 3 cubes.

2. Show that every polynomial is the sum of 3 cubes.

3. Show, in general, that the polynomial x is “pivotal,” that is if x is

the sum of g nth powers, then every polynomial is the sum of g

nth powers.

4. Show that if max(z, b) > 2c, where c is the degree of R(x), then

P

a

÷ Q

b

= R is unsolvable.

5. Show that the constant polynomial 1 can be written as the sum of

√

4n ÷ 1 nth powers of nonconstant polynomials.

VI

A “Natural” Proof of the

Nonvanishing of L-Series

Rather than the usual adjectives of “elementary” (meaning not in-

volving complex variables) or “simple” (meaning not having too

many steps) which refer to proofs, we introduce a newone, “natural.”

This term, which is just as undeﬁnable as the others, is introduced to

mean not having any ad hoc constructions or brilliancies. A“natural”

proof, then, is one which proves itself, one available to the “common

mathematician in the streets.”

A perfect example of such a proof and one central to our whole

construction is the theorem of Pringsheim and Landau. Here the cru-

cial observation is that a series of positive terms (convergent or not)

can be rearranged at will. Addition remains a commutative operation

when the terms are positive. This is a sumof a set of quantities rather

than the sum of a sequence of them.

The precise statement of the Pringsheim–Landau theorem is that,

for a Dirichlet series with nonnegative coefﬁcients, the real boundary

point of its convergence region must be a singularity.

Indeed this statement proves itself through the observation that

n

a−z

=

¸

k

(a−z)

k

k!

(log n)

k

is a power series in (a − z) with non-

negative coefﬁcients. Thus the (unique) power series for

¸

a

n

n

−z

=

¸

a

n

n

−a

· n

a−z

has nonnegative coefﬁcients in powers of (a − z).

So let b be the real boundary point of the convergence region of

¸

a

n

n

−z

, and suppose that b is a regular point and that b < a. Thus

the power series in (a − z) continues to converge a bit to the left

of b and, by rearranging terms, the Dirichlet series converges there

also, contradicting the meaning of b. A“natural” proof of a “natural”

59

60 VI. A “Natural” Proof of the Nonvanishing of L-Series

theorem follows, one with a very nice corollary which we record for

future use.

(1) If a Dirichlet series with nonnegative coefﬁcients represents a

function which is (can be continuedtobe) entire, thenit is everywhere

convergent.

Our ultimate aim is to prove that the L-series have no zeros on

the line +z = 1. This is the nonvanishing of the L-series that we

referred to in the chapter title. So let us begin with the simplest of

all L-series, the ζ -function, ζ(z) =

¸

1

n

z

. Our proof, in fact, was

noticed by Narasimhan and is as follows: Assume, par contraire, that

ζ(z) had a zero at 1÷ia, a real. Then (sic!) the function ζ(z)ζ(z÷ia)

would be entire. (See the appendix, page no. 63).

The only trouble points could be at z = 1 or at z = 1 −ia where

one of the factors has a pole, but these are then cancelled by the other

factor, which, by our assumption, has a zero.

A bizarre conclusion, perhaps, that the Dirichlet series ζ(z)ζ(z ÷

ia) is entire. But how to get a contradiction? Surely there is no hint

fromits coefﬁcients, they aren’t even real. Anatural step then would

be to make them real by multiplying by the conjugate coefﬁcient

function, ζ(z)ζ(z − ia), which of course is also entire. We are led,

then, to form ζ

2

(z)ζ(z ÷ ia)ζ(z − ia).

This function is entire and has real coefﬁcients, but are they pos-

itive? (We want them to be so that we can use (1).) Since these are

complicated coefﬁcients dependent on sums of complex powers of

divisiors, we pass to the logarithm, 2 log ζ(z) ÷ log ζ(z ÷ ia) ÷

log ζ(z − ia), which, by Euler’s factorization of the ζ -function, has

simple coefﬁcients. A dangerous route, passing to the logarithm, be-

cause this surely destroys our everywhere analyticity. Nevertheless

let us brazen forth (faint heart fair maiden never won).

By Euler’s factorization, 2 log ζ(z) ÷ log ζ(z ÷ ia) ÷ log ζ(z −

ia) =

¸

p

2 log

1

1−p

−z

÷log

1

1−p

−z−ia

÷log

1

1−p

−z÷ia

=

¸

p,v

1

vp

vz

(2 ÷ p

−iva

÷ p

÷iva

), and indeed these coefﬁcients are nonnega-

tive! The dangerous route is now reversed by exponentiating. We

return to our entire function while preserving the nonnegativity of

the coefﬁcients. All in all, then,

VI. A “Natural” Proof of the Nonvanishing of L-Series 61

(2) ζ

2

(z)ζ(z ÷ia)ζ(z −ia) is an entire Dirichlet series with nonneg-

ative coefﬁcients. Combining this with (1) implies the unbelievable

fact that

(3) the Dirichlet series for ζ

2

(z)ζ(z ÷ ia)ζ(z − ia) is everywhere

convergent.

The falsity of (3) can be established in may ways, especially if

we recall that the coefﬁcients are all nonnegative. For example, the

subseries corresponding to n = power of 2 is exactly equal to

1

(1−2

−z

)

2

·

1

1−2

−z−ia

·

1

1−2

−z÷ia

which exceeds

1

(1−2

−z

)

2

·

1

4

along the

nonnegative (real) axis and thereby guarantees divergence at z =

0. Q.E.D.

And so we have the promised natural proof of the nonvanishing of

the ζ -function which can then lead to the natural proof of the Prime

Number Theorem. We must turn to the general L-series which holds

the germ of the proof of the Prime Progression Theorem. Dirichlet

pointed out that the natural way to treat these progressions is not

one progression at a time but all of the pertinent progressions of a

given modulus simultaneously, for this leads to the underlying group

and hence to its dual group, the group of characters. Let us look, for

example, at the modulus 10. The pertinent progressions are 10k ÷1,

10k ÷ 3, 10k ÷ 7,10k ÷ 9, so that the group is the multiplicative

group of 1,3,7,9 (mod 10). The characters are

χ

1

: χ

1

(1) = 1, χ

1

(3) = 1, χ

1

(7) = 1, χ

1

(9) = 1,

χ

3

: χ

3

(1) = 1, χ

3

(3) = 1, χ

3

(7) = 1, χ

3

(9) = 1,

χ

7

: χ

7

(1) = 1, χ

7

(3) = 1, χ

7

(7) = 1, χ

7

(9) = 1,

χ

9

: χ

9

(1) = 1, χ

9

(3) = 1, χ

9

(7) = 1, χ

9

(9) = 1,

and so the L-series are

L

1

(z) =

¸

p≡1

1

1 − p

−z

¸

p≡3

1

1 − p

−z

¸

p≡7

1

1 − p

−z

¸

p≡9

1

1 − p

−z

,

L

3

(z) =

¸

p≡1

1

1 − p

−z

¸

p≡3

1

1 − ip

−z

¸

p≡7

1

1 ÷ ip

−z

¸

p≡9

1

1 ÷ p

−z

,

62 VI. A “Natural” Proof of the Nonvanishing of L-Series

L

7

(z) =

¸

p≡1

1

1 − p

−z

¸

p≡3

1

1 ÷ ip

−z

¸

p≡7

1

1 − ip

−z

¸

p≡9

1

1 ÷ p

−z

,

and

L

9

(z) =

¸

p≡1

1

1 − p

−z

¸

p≡3

1

1 ÷ p

−z

¸

p≡7

1

1 ÷ p

−z

¸

p≡9

1

1 − p

−z

.

(Here +z > 1 to insure convergence and the subscripting of the

characters is used to reﬂect the isomorphism of the dual group and

the original group.)

The generating function for the primes in the arithmetic pro-

gressions ((mod 10) in this case) are then linear combinations of

the logarithms of these L-series. And so indeed the crux is the

nonvanishing of these L-series.

What could be more natural or more in the spirit of Dirichlet, but

to prove these separate nonvanishings altogether? So we are led to

take the product of all the L-series! (Landau uses the same device to

prove nonvanishing of the L-series at point 1.)

The result is the Dirichlet series

Z(z) =

¸

p≡1

1

(1 − p

−z

)

4

¸

p≡3

1

(1 − p

−4z

)

¸

p≡7

1

(1 − p

−4z

)

¸

p≡9

1

(1 − p

−2z

)

2

,

and the problemreduces to showing that Z(z) is zero-free on +z = 1.

Of course, this is equivalent to showing that

¸

p≡1

1

1−p

−z

is zero-

free on +z = 1, which seems, at ﬁrst glance, to be a more attractive

form of the problem. This is misleading, however, and we are bet-

ter off with Z(z), which is the product of L-series and is an entire

function except possibly for a simple pole at z = 1. (See the

appendix.)

Guided by the special cases let us turn to the general one. So let A

be a positive integer, and denote by G

A

the multiplicative group of

residue classes (mod A) which are prime to A. Set h = φ(A), and

denote the group elements by 1 = n

1

, n

2

, . . . , n

h

. Denote the dual

group of G

A

by

ˆ

G

A

and its elements by χ

1

, χ

n

2

, . . . , χ

n

h

arranged

VI. A “Natural” Proof of the Nonvanishing of L-Series 63

so that n

i

↔ χ

n

i

is an isomorphism of G and

ˆ

G. Next, for +z >

1, write L

n

i

(z) =

¸

n

j

¸

p≡n

j

1

1−χ

n

i

(n

j

)p

−z

and ﬁnally set Z(z) =

¸

n

i

L

n

i

(z). As in the case A = 10, elementary algebra leads to

Z(z) =

¸

n

j

¸

p≡n

j

1

(1−p

−h

j

z

)

h/h

j

, where h

j

is the order of the group

element n

j

.

As before, Z(z) is entire except possibly for a simple pole at z = 1,

andwe seeka proof that Z(1÷ia) ,= 0 for real a. Soagainwe assume

Z(1 ÷ ia) = 0, form Z

2

(z)Z(z ÷ ia)Z(z − ia), and conclude that

it is entire. We note that its logarithm and hence that it itself has

nonnegative coefﬁcients so that (1) is applicable.

So, with dazzling speed, we see that a zero of any L-series would

lead to the everywhere convergence of the Dirichlet series (with

nonnegative coefﬁcients) Z

2

(z)Z(z ÷ ia)Z(z − ia).

The end game (ﬁnal contradiction) is also as before although 2

may not be among the primes in the resultant product, and we may

have to take some other prime π. Nonetheless again we see that the

subseries of powers of π diverges at z = 0 which gives us our QED.

Appendix. A proof that the L-series are everywhere analytic func-

tions with the exception of the principal L-series, L

1

at the single

point z = 1, which is a simple pole.

Lemma. For any θ in [0,1), deﬁne f (z) =

¸

∞

n=1

1

(n−θ)

z

−

1

z−1

for

+z > 1. Then f (z) is continuable to an entire function.

Proof. Since, for +z > 1,

∞

0

e

−nt

e

θt

t

z−1

dt =

1

(n−θ)

z

∞

0

e

−t

t

z−1

dt =

(z)

(n−θ)

z

, by summing, we get

¸

1

(n − θ)

z

=

1

(z)

∞

0

e

θt

e

t

− 1

t

z−1

dt

or

¸

1

(n − θ)

z

−

1

z − 1

=

1

(z)

∞

0

e

θt

e

t

− 1

−

e

−t

t

t

z−1

dt.

64 VI. A “Natural” Proof of the Nonvanishing of L-Series

Since

e

θt

e

t

−1

−

e

−t

t

is analytic and has integrable derivatives on [0, ∞),

we may integrate by parts repeatedly and thereby get

¸

1

(n − θ)

z

−

1

z − 1

=

1

(z ÷ k)

∞

0

−

d

dt

k

e

θt

e

t

− 1

−

e

−t

t

t

z÷k−1

dt.

This gives continuation to +z > −k, and, since k is arbitrary, the

continuation is to the entire plane.

Problems for Chapter VI 65

Problems for Chapter VI

1. Prove, by elementary methods, that there are inﬁnitely many

primes not ending in the digit 1.

2. Prove that there are inﬁnitely many primes p for which neither

p ÷ 2 nor p − 2 is prime.

3. Prove that at least 1/6 of the integers are not expressible as the

sum of 3 squares.

4. Prove that (z) has no zeros in the whole plane, although, it has

poles.

5. Suppose δ(x) decreases to 0 as x → ∞. Produce an ε(x) which

goes to 0 at ∞but for which δ(xε(x)) = o(ε(x)).

VII

Simple Analytic Proof of the

Prime Number Theorem

The magniﬁcent Prime Number Theorem has received much atten-

tion and many proofs throughout the past century. If we ignore the

(beautiful) elementary proofs of Erd˝ os and Selberg and focus on the

analytic ones, we ﬁnd that they all have some drawbacks. The origi-

nal proofs of Hadamard and de la Vall´ ee Poussin were based, to be

sure, on the nonvanishing of ζ(z) in +z ≥ 1, but they also required

annoying estimates of ζ(z) at ∞, because the formulas for the coef-

ﬁcients of the Dirichlet series involve integrals over inﬁnite contours

(unlike the situation for power series) and so effective evaluation

requires estimates at ∞.

The more modern proofs, due to Wiener and Ikehara (and also

Heins) get around the necessity of estimating at ∞ and are indeed

based only on the appropriate nonvanishing of ζ(z), but they are

tied to certain results of Fourier transforms. We propose to return

to contour integral methods to avoid Fourier analysis and also to

use ﬁnite contours to avoid estimates at ∞. Of course certain errors

are introduced thereby, but the point is that these can be effectively

minimized by elementary arguments.

So let us begin with the well-known fact about the ζ -function (see

Chapter 6, page 60–61)

(z − 1)ζ(z) is analytic and zero-free throughout +z ≥ 1. (1)

This will be assumed throughout and will allow us to give our proof

of the Prime Number Theorem.

67

68 VII. Simple Analytic Proof of the Prime Number Theorem

In fact we give two proofs. This ﬁrst one is the shorter and

simpler of the two, but we pay a price in that we obtain one of

Landau’s equivalent forms of the theorem rather than the standard

form π(N) ∼ N/ log N. Our second proof is a more direct assault

on π(N) but is somewhat more intricate than the ﬁrst. Here we ﬁnd

some of Tchebychev’s elementary ideas very useful.

Basically our novelty consists in using a modiﬁed contour integral,

f (z)N

z

1

z

÷

z

R

2

dz,

rather than the classical one,

C

f (z)N

z

z

−1

dz. The method is rather

ﬂexible, and we could use it to directly obtain π(N) by choosing

f (z) = log ζ(z). We prefer, however, to derive both proofs from the

following convergence theorem. Actually, this theorem dates back

to Ingham, but his proof is ´ a la Fourier analysis and is much more

complicated than the contour integral method we now give.

Theorem. Suppose [a

n

[ ≤ 1, and form the series

¸

a

n

n

−z

which

clearly converges to an analytic function F(z) for +z > 1. If, in

fact, F(z) is analytic throughout +z ≥ 1, then

¸

a

n

n

−z

converges

throughout +z ≥ 1.

Proof of the convergence theorem. Fix a w in +w ≥ 1.

Thus F(z ÷ w) is analytic in +z ≥ 0. We choose an R ≥ 1 and

determine δ = δ(R) > 0, δ ≤

1

2

and an M = M(R) so that

F(z ÷ w) is analytic and bounded by M in − δ ≤ +z, [z[ ≤ R.

(2)

Now form the counterclockwise contour bounded by the arc [z[ =

R, +z > −δ, and the segment +z = −δ, [z[ ≤ R. Also denote by

A and B, respectively, the parts of in the right and left half planes.

By the residue theorem,

2πiF(w) =

F(z ÷ w)N

z

1

z

÷

z

R

2

dz. (3)

Now on A, F(z ÷ w) is equal to its series, and we split this into

its partial sum S

N

(z ÷ w) and remainder r

N

(z ÷ w). Again by the

VII. Simple Analytic Proof of the Prime Number Theorem 69

residue theorem,

A

S

N

(z ÷ w)N

z

1

z

÷

z

R

2

dz

= 2πiS

N

(w) −

−A

S

N

(z ÷ w)N

z

1

z

÷

z

R

2

dz,

with −A as usual denoting the reﬂection of A through the origin.

Thus, changing z to −z, this can be written as

A

S

N

(z ÷ w)N

z

1

z

÷

z

R

2

dz

= 2πiS

N

(w) −

A

S

N

(w − z)N

−z

1

z

÷

z

R

2

dz. (4)

Combining (3) and (4) gives

2πi[F(w) − S

N

(w)]

=

A

¸

r

N

(z ÷ w)N

z

−

S

N

(w − z)

N

z

¸

1

z

÷

z

R

2

dz (5)

÷

B

F(z ÷ w)N

z

1

z

÷

z

R

2

dz,

and, to estimate these integrals, we record the following (here as

usual we write +z = x, and we use the notation α < β to mean

simply that [α[ ≤ [β[):

1

z

÷

z

R

2

=

2x

R

2

along [z[ = R (in particular on A), (6)

1

z

÷

z

R

2

<

1

δ

1 ÷

[z[

2

R

2

=

2

δ

on the line +z = −δ,

[z[ ≤ R, (7)

r

N

(z ÷ w) <

∞

¸

n=N÷1

1

n

x÷1

≤

∞

N

dn

n

x÷1

=

1

xN

x

, (8)

and

S

N

(w − z) <

N

¸

n=1

n

x−1

≤ N

x−1

÷

N

0

n

x−1

dn

70 VII. Simple Analytic Proof of the Prime Number Theorem

= N

x

1

N

÷

1

x

. (9)

By (6), (8), (9), on A,

¸

r

N

(z ÷ w)N

z

−

S

N

(w − z)

N

z

¸

1

z

÷

z

R

2

<

1

x

÷

1

x

÷

1

N

2x

R

2

≤

4

R

2

÷

2

RN

,

and so, by the “maximum times length” estimate (M–L formula) for

integrals, we obtain

A

¸

r

N

(z ÷ w)N

z

−

S

N

(w − z)

N

z

¸

1

z

÷

z

R

2

dz <

4π

R

÷

2π

N

.

(10)

Next, by (2), (6), and (7), we obtain

B

F(z ÷ w)N

z

1

z

÷

z

R

2

dz

<

R

−R

M · N

−δ

2

δ

dy ÷ 2M

0

−δ

n

x

2[x[

R

2

3

2

dx (11)

≤

4MR

δN

δ

÷

6M

R

2

log

2

N

.

Inserting the estimates (10) and (11) into (5) gives

F(w) − S

N

(w) <

2

R

÷

1

N

÷

MR

δN

δ

÷

M

R

2

log

2

N

,

and, if we ﬁx R = 3/, we note that this right-hand side is < for

all large N. We have veriﬁed the very deﬁnition of convergence!

First Proof of the Prime Number Theorem.

Following Landau, we will show that the convergence of

¸

n

µ(n)

n

(as given above) implies the PNT. Indeed all we need about this

convergent series is the simple corollary that

¸

n≤N

µ(n) = o(N).

Expressing everything in terms of the ζ -function, then, we have

established the fact that

1

ζ(z)

has coefﬁcients which go to 0 on average.

First Proof of the Prime Number Theorem. 71

The PNT is equivalent to the fact that the average of the coefﬁcients

of

ζ

/

ζ

(z) is equal to 1. For simply note that

−

ζ

/

ζ

(z) = −

d

dz

log ζ(z) = −

d

dz

log

¸

p

1

1 − p

−z

=

¸

p

d

dz

log

1 − p

−z

=

¸

p

p

−z

log p

1 − p

−z

=

¸

p

log p

p

−z

− 1

.

This last series is the same as

¸

(n)

n

z

where (n) is log p whenever

nis a power of p, p any prime, and 0 otherwise. So indeed the average

of these coefﬁcients is

1

N

¸

n≤N

(n) whose limit being 1 is exactly

the Prime Number Theorem.

In short, we want the average value of the coefﬁcients of −

ζ

/

ζ

(z) −

ζ(z) to approach 0. Writing this function as

1

ζ

(z)[−ζ

/

(z) − ζ(z)] =

¸

µ(n)

n

z

¸

¸

log n

n

z

−

¸

d(n)

n

z

¸

,

we may write this average (of the ﬁrst N terms)as

1

N

¸

ab≤N

µ(a)[log b − d(b)]

=

1

N

¸

ab≤N

µ(a)[log b − d(b) ÷ 2γ ] −

2γ

N

,

where 2γ is chosen as the constant for which

K

¸

b=1

[log b − d(b) ÷ 2γ ]

becomes O(

√

K).

Now we use the Landau corollary that

¸

n≤N

µ(n) = o(N) to

conclude that

1

N

¸

n≤N

µ(n) < δ(N),

72 VII. Simple Analytic Proof of the Prime Number Theorem

where δ(N) tends to 0, and our trick is to pick a function w(N) which

approaches ∞but such that w(N)δ

N

w(N)

approaches 0.

This done, we may conclude that

¸

n≤N

(n) = N ÷ O

¸

N

√

w(N)

¸

÷ O

¸

Nw(N)δ

¸

N

w(N)

¸¸

= N ÷ o(N),

and the proof is complete.

Second Proof of the Prime Number Theorem.

In this section, we begin with Tchebychev’s observation that

¸

p≤n

log p

p

− log n is bounded, (12)

which he derived in a direct elementary way from the prime

factorization on n!

The point is that the Prime Number Theoremis easily derived from

¸

p≤n

log p

p

− log n converges to a limit, (13)

by a simple summation by parts, which we leave to the reader. Nev-

ertheless the transition from (12) to (13) is not a simple one, and we

turn to this now.

So, for +z > 1, form the function

f (z) =

∞

¸

n=1

1

n

z

¸

p≤n

log p

p

=

¸

p

log p

p

¸

¸

n≥p

1

n

z

¸

.

Now

¸

n≥p

1

n

z

=

1

(z − 1)p

z−1

÷ z

∞

p

1 − {t }

t

z÷1

dt

=

p

(z − 1)

1

p

z

− 1

÷ A

p

(z)

**Second Proof of the Prime Number Theorem. 73
**

where A

p

(z) is analytic for +z > 0 and is bounded by

1

p

x

(p

x

− 1)

÷

[z(z − 1)[

xp

x÷1

.

Hence,

f (z) =

1

z − 1

¸

¸

¸

p

log p

p

z

− 1

÷ A(z)

,

where A(z) is analytic for +z >

1

2

by the Weierstrass M-test.

By Euler’s factorization formula, however, we recognize that

¸

p

log p

p

z

− 1

=

−d

dz

log ζ(z),

and so we deduce, by (1), that f (z) is analytic in +z ≥ 1 except for

a double pole with principal part 1/(z − 1)

2

÷ c/(z − 1) at z = 1.

Thus if we set

F(z) = f (z) ÷ ζ

/

(z) − cζ(z) =

¸

n

a

n

n

z

where

a

n

=

¸

p≤n

log p

p

− log n − c, (14)

we deduce that F(z) is analytic in +z ≥ 1.

From (12) and our convergence theorem, then, we conclude that

¸

a

n

n

converges,

and fromthis and the fact, from(14), that a

n

÷log nis nondecreasing,

we proceed to prove a

n

→ 0.

By applying the Cauchy criterion we ﬁnd that, for N large,

N(1÷)

¸

N

a

n

n

≤

2

(15)

and

N

¸

N(1−)

a

n

n

≥ −

2

. (16)

74 VII. Simple Analytic Proof of the Prime Number Theorem

In the range N to N(1 ÷ ), by (14), a

n

≥ a

N

÷ log(N/n) ≥

a

N

− . So

¸

N(1÷)

N

a

n

/n ≥ (a

N

− )

¸

N(1÷)

N

1/n, and (15) yields

a

N

< ÷

2

¸

N(1÷)

N

1

n

< ÷

2

N/N(1 ÷ )

= 2 ÷

2

. (17)

Similarly in [N(1 −), N], a

n

≤ a

N

÷log(N/n) ≤ a

N

÷/(1 −

), so that

N

¸

N(1−)

a

n

n

≤

a

N

÷

1 −

N

¸

N(1−)

1

n

,

and (16) gives

a

N

≥

−

1 −

−

2

¸

N

N(1−)

1

n

≥

−

1 −

−

2

N/N

=

2

− 2

1 −

.

(18)

Taken together, (17) and (18) establish that a

N

→ 0, and so (13) is

proved.

Problems for Chapter VII 75

Problems for Chapter VII

1. Given that

¸

a

n

n

converges, prove that

¸

N

n=1

a

n

= o(N).

2. Given that

¸

a

n

n

converges and that a

n

− a

n−1

>

−1

n

, prove that

a

n

→ 0.

3. Show that d(n), the number of divisors of n, is O(n

ε

) for every

positive ε.

4. In fact, show that d(n) < n

1

log log n

.

Index

Addition problems, 1–2

Afﬁne property, 41

Analytic functions, L-series as,

63

Analytic method, 1

Analytic number theory, 1–14

Analytic proof of Prime Number

Theorem, 65–71

Approximation lemma, basic,

42–47

Arithmetic progressions, 41

dissection into, 14

sequences without, 41–47

Asymptotic formula, 4

Basic approximation lemma,

42–47

Cauchy criterion, 71

Cauchy integral, 23–24

Cauchy’s theorem, 18–19

Change making, 2–5

Commutative operation, 59

Complex numbers, 18

Contour integral, modiﬁed, 66

Contour integration, 46

Contours

ﬁnite, 65

inﬁnite, 65

Convergence theorem, 66

proof of, 66–68

Crazy dice, 5–8

Dice, crazy, 5–8

Dirichlet series, 59–60, 62

Dirichlet theorem, 45, 50

Dissection into arithmetic

progressions, 14

Elliptic integral, 33

Entire functions, 60

Erd˝ os, Paul, vii

Erd˝ os-Fuchs theorem, 31, 35–38

Euler’s factorization, 60

Euler’s factorization formula, 71

Euler’s theorem, 11–12

Evens and odds, dissection into,

14

Extremal sets, 42

Finite contours, 65

Fourier analysis, 65

Generating functions, 1

of asymptotic formulas, 18–19

of representation functions, 7

Inﬁnite contours, 65

Integers, 1

breaking up, 17

nonnegative, splitting, 8–10

L-series

as analytic functions, 63

general, 61–62

77

78 Index

nonvanishing of, see

Nonvanishing of L-series

zero of any, 63

Lagrange theorem, 49

Landau corollary, 69

L’Hˆ opital’s rule, 5

“Magnitude property,” 53

Mathematics, vii

“Monotone majorant,” 45

“Natural” proof, 59

of nonvanishing of L-series,

59–63

Nonnegative integers, splitting,

8–10

Nonvanishing of L-series, 60

“natural” proof of, 59–63

Odds and evens, dissection into,

14

Parseval upper bound, 36

Parseval’s identity, 33–34

Partial fractional decomposition,

3–4

Partition function, 17–29

Permission constant, 42

Pigeonhole principle, 50

PNT, see Prime Number Theorem

Prime Number Theorem (PNT),

65

analytic proof of, 65–71

ﬁrst proof of, 68–70

second proof of, 70–72

Pringsheim-Landau Theorem, 59

Progressions, arithmetic, see

Arithmetic progressions

q(n), coefﬁcients of, 25–29

Relative error, 4

Representation functions, 7

generating functions of, 7

near constancy of, 31

Riemann integral, 20

double, 31

Riemann sums, 20–25

Roth Theorem, 46–47

Rulers, marks on, 12–13

Schnirelmann’s Theorem, 50–51

Schwarz inequality, 34

Sequences without arithmetic

progressions, 41–47

Splitting problem, 8–10

Stirling’s formula, 4, 27, 29

Szemer´ edi-Furstenberg result, 43

Taylor coefﬁcients, 3

Tchebychev’s observation, 70

Unit circle, 13

Waring problem, 49–56

Weyl sums, 51–52

Graduate Texts in Mathematics

177

Editorial Board S. Axler F.W. Gehring K.A. Ribet

Springer

New York Berlin Heidelberg Barcelona Hong Kong London Milan Paris Singapore Tokyo

Donald J. Newman

Analytic Number Theory

13

Donald J. Newman Professor Emeritus Temple University Philadelphia, PA 19122 USA Editorial Board S. Axler Department of Mathematics San Francisco State University San Francisco, CA 94132 USA F.W. Gehring Department of Mathematics University of Michigan Ann Arbor, MI 48109 USA K.A. Ribet Department of Mathematics University of California at Berkeley Berkeley, CA 94720-3840 USA

Mathematics Subject Classiﬁcation (1991): 11-01, 11N13, 11P05, 11P83

Library of Congress Cataloging-in-Publication Data Newman, Donald J., 1930– Analytic number theory / Donald J. Newman. p. cm. – (Graduate texts in mathematics; 177) Includes index. ISBN 0-387-98308-2 (hardcover: alk. paper) 1. Number Theory. I. Title. II. Series. QA241.N48 1997 512’.73–dc21

97-26431

© 1998 Springer-Verlag New York, Inc. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use of general descriptive names, trade names, trademarks, etc., in this publication, even if the former are not especially identiﬁed, is not to be taken as a sign that such names, as understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone. ISBN 0-387-98308-2 Springer-Verlag New York Berlin Heidelburg SPIN 10763456

The Partition Function The Generating Function The Approximation Riemann Sums The Coefﬁcients of q(n) 17 18 19 20 25 III. The Erd˝ s–Fuchs Theorem o Erd˝ s–Fuchs Theorem o 31 35 IV. Sequences without Arithmetic Progressions The Basic Approximation Lemma 41 42 v .Contents Introduction and Dedication vii I. The Idea of Analytic Number Theory Addition Problems Change Making Crazy Dice Can r(n) be “constant?” A Splitting Problem An Identity of Euler’s Marks on a Ruler Dissection into Arithmetic Progressions 1 1 2 5 8 8 11 12 14 II.

The Waring Problem VI. Simple Analytic Proof of the Prime Number Theorem First Proof of the Prime Number Theorem. A “Natural” Proof of the Nonvanishing of L-Series VII. Index 49 59 67 70 72 77 .vi Contents V. Second Proof of the Prime Number Theorem.

They belonged to a tradition that undoubtedly revered mathematics. but as a discipline at some considerable remove from the commonplace. Gauss was certainly a wordy master and Euler another. collaborator. I like to think that Erd˝ s. a weighty solution. the greatest mathematician I o have ever known. those working in number theory and related ﬁelds.Introduction and Dedication This book is dedicated to Paul Erd˝ s. They may have felt that the status and importance of mathematics as an intellectual discipline entailed. did not necessarily strive to effect the simple solution. and dear friend. nor the careful cultivation of an Euler or Gauss. In keeping with a more democratic concept of intelligence itself. The great mathematicians of yesteryear. contemporary mathematics diverges from this somewhat elitist view. even those in number theory itself. can be resolved in simple and more direct terms. The simple approach implies a mathematics generally available even to those who have not been favored with the natural endowments. There is no doubt a certain presumptuousness in this claim. would have appreciated this little book and heartily endorsed its philosophy. whose mathematics embodied the princio ples which have impressed themselves upon me as deﬁning the true character of mathematics. vii . which have famously difﬁcult solutions. whom it has been my rare privilege to consider colleague. perhaps indeed required. This book proffers the thesis that mathematics is actually an easy subject and many of the famous problems.

but as a revelation of its true nature. . That adventure is intrinsic to even the most elementary description of analytic number theory. and back again. The insistence on simplicity asserts a mathematics that is both “magical” and coherent. in turn entail the familiar processes. further aﬁeld from all the intricacies of contour integration and they. yet in its wake a pattern is revealed that implies a mathematics deeply inter-connected and cohesive. This formulation inevitably moves us away from the designated subject to a consideration of complex variables. Toward this end “The Cauchy Integral” proves to be an indispensable tool. Having wandered away from our subject.viii Introduction and Dedication Such an attitude might prove an effective antidote to a generally declining interest in pure mathematics. the deformation and estimation of these contour integrals. But it is not so much as incentive that we proffer what might best be called “the fun and games” approach to mathematics. The initial step in the investigation of a number theoretic item is the formulation of “the generating function”. Yet it leads us. inevitably. The solution that strives to master these qualities restores to mathematics that element of adventure that has always supplied its peculiar excitement. Retracing our steps we ﬁnd that we have gone from number theory to function theory. The journey seems circuitous. it becomes necessary to effect a return.

in number theory) is its very existence! How could one use properties of continuous valued functions to determine properties of those most discrete items. this is simply a restatement of the question. For example. Thereby questions about the addition of integers are transformed into questions about the multiplication of polynomials or power series. there are answers and answers.I The Idea of Analytic Number Theory The most intriguing thing about Analytic Number Theory (the use of Analysis. but to those of us who haven’t. or the use of generating functions. So let us take a look at some of these. or function theory. is to see it in action in a number of pertinent examples. The link is the simple observation that adding m and n is isomorphic to multiplying zm and zn . Addition Problems Questions about addition lend themselves very naturally to the use of generating functions. Lagrange’s beautiful theorem that every positive integer is the sum of 1 . Perhaps the best way to understand the use of the analytic method. To those of us who have witnessed the use of generating functions this is a kind of answer. Analytic functions? What has differentiability got to do with counting? The astonishment mounts further when we learn that the complex zeros of a certain analytic function are the basic tools in the investigation of the primes. the integers. The answer to all this bewilderment is given by the two words generating functions. Well.

3 1−z and multiplying these three equations to get 1 (1 − z)(1 − z2 )(1 − z3 ) (1 + z + z1+1 + · · ·)(1 + z2 + z2+2 + · · ·) × (1 + z3 + z3+3 + · · ·).” a dollar. but at least one begins to see how this transition from integers to analytic functions takes place. and two 3’s? Yes. for |z| < 1. but the problem is both too hard and too easy. let us write. and 3? To form the appropriate generating function. it is zfour1 s+one2+two3 s and doesn’t this exactly correspond to the method of changing the amount 12 into four 1’s. 1−z 1 1 + z2 + z2+2 + z2+2+2 + · · · . 2 Change Making How many ways can one make change of a dollar? The answer is 293.2 I. How one proves such a fact about the coefﬁcients of such a power series is another story. On the one hand. But now let’s look at some addition problems that we can solve completely by the analytic method. this term is z12 . More ﬁtting to our spirit is the following problem: How many ways can we make change for n if the coins are 1. 2. Now we ask ourselves: What happens when we multiply out the right-hand side? We obtain terms like z1+1+1+1 · z2 · z3+3 . and in fact we . but. on the other hand. The Idea of Analytic Number Theory 4 four squares becomes the statement that all of the coefﬁcients of the power series for 1 + z + z4 + · · · + zn + · · · are positive. Too hard because the available coins are so many and so diverse. one 2. 1 1 + z + z1+1 + z1+1+1 + · · · . 2 1−z 1 1 + z3 + z3+3 + z3+3+3 + · · · . Too easy because it concerns just one “changee.

namely. comes from an algebraic technique that we all learned in calculus. The correct answer. Recall A that this leads to terms like (1−αz)k for which we know the expan1 sion explicitly (namely. 2’s. then. respectively into z . z . since we have restricted ourselves to |z| < 1 wherein convergence is absolute. since 1 (1 − z)2 d 1 dz 1 − z d dz zn (n + 1)zn . Our number theoretic problem has been translated into a problem about analytic functions. 1 1 a 2b 3c . and 3’s) for “every” n will appear in this multiplying out. Fine. z and multiply only to 1−z2 1−z3 discover that the coefﬁcient is the number of ways of making change for n. A well deﬁned analytic problem. Thus the thing not to do is expand 1−z . (Furthermore all is rigorous and not just formal. Thus if we call C(n) the number of ways of making change for n. Carrying out the algebra.Change Making 3 see that “every” way of making change (into 1’s. in this case. (1−αz)k is just a constant times the (k − 1)th 1 derivative of (1−αz) α n zn ). 6 (1 − z)3 4 (1 − z)2 4 (1 − z2 ) 3 (1 − z3 ) Thus. (1 − z)(1 − z2 )(1 − z3 ) (1) and the generating function for our unknown quantity C(n) is produced. then C(n) will be the exact coefﬁcient of zn when the multiplication is effected. ﬁnding the Taylor 1 coefﬁcients of the function (1−z)(1−z2 )(1−z3 ) . leads to the partial fractional decomposition which we may arrange in the following form: 1 (1 − z)(1 − z2 )(1 − z3 ) 1 1 1 1 1 1 1 1 + + + . but how to solve it? We must resist the temptation to solve this problem by undoing the analysis 1 which led to its formulation. namely partial fractions.) Thus C(n)zn 1 .

2 (n + 2)(n + 1) n+1 χ1 (n) χ2 (n) C(n) + + + (2) 12 4 4 3 1 if 2 | n and 0 otherwise. The Idea of Analytic Number Theory and 1 n+1 n d d z 2 dz 2(1 − z) dz 2 (n + 2)(n + 1) n z . 25. (Also note that our e n2 result (3) can be weakened to C(n) ∼ 12 . we write f (n) ∼ F (n) when this occurs. but these are rare. Imagine the mess that occurs if the coins were the usual coins of the realm. A somewhat cumbersome formula.4 I. At any rate. One famous such √ example is Stirling’s formula n! ∼ 2π n( n )n . Thus . However. In this generality we ask for an asymptotic formula for the corresponding C(n). a2 . a3 . . Landau. 50. 12 2 (3) where the terms in the brackets mean the greatest integers. The right thing to ask for then is an “asymptotic” formula rather than an exact one. let us simply look for one of the terms in this expansion. the heaviest one. namely 1. . . 10. Recall that an asymptotic formula F (n) for a function f (n) is one f 1.. explicitly ﬁnding the partial fractional decomposition of this function is the hopeless task. A nice crisp exact formula. ak . where to avoid trivial congruence considerations we will require that there be no common divisiors other than 1. χ2 (n) 1 if 3 | n where χ1 (n) and 0 else. 5.) So let us assume quite generally that there are coins a1 . (100?). In the colorful language of E. for which limn→∞ F (n) (n) the relative error in replacing f (n) by F (n) is eventually 0%. but one which can be shortened nicely into 1 (1 − z)3 C(n) n n2 + +1 . (1 − za1 )(1 − za2 ) · · · (1 − zak ) (4) But the next step. As before we ﬁnd that the generating function is given by C(n)zn 1 .

because we assumed no common divisiors. a1 a2 · · · ak (k − 1)! (5) Crazy Dice An ordinary pair of dice consist of two cubes each numbered 1 through 6. (1 − z)k multiply by (1 − z)k to get 1−z 1−z 1−z ··· 1 − z a1 1 − z a2 1 − z ak c + (1 − z)k × other terms. In short C(n) ∼ c n+k−1 . although the coefﬁcient of the term (1−z)k is c n+k−1 . The combined possibilities for the . The ﬁnal 1 result is c . 1−z 1 and ﬁnally let z → 1. c Thus. In terms of our analytic representation. for example. all will be of order lower than k. C(n) ∼ c nk−1 . So let us write (1 − za1 )(1 1 − za2 ) · · · (1 − zak ) c + other terms.Crazy Dice 5 at z 1 the denominator has a k-fold zero and so there will be a c term (1−z)k . By L’Hˆ pital’s rule. When tossed together there are altogether 36 (equally likely) outcomes. the sum total of all of these terms is negligible compared to our heavy term c n+k−1 . or k−1 k−1 even simpler. 1−zai → ai o whereas each of the other terms times (1 − z)k goes to 0. and our ﬁnal asymptotic formula reads a1 a2 ···ak C(n) ∼ nk−1 . All the other zeros are roots of unity and. (k − 1)! But. each die is associated with the polynomial z + z2 + z3 + z4 + z5 + z6 . we cannot avoid this one (it’s the whole story!). Thus the sums go from 2 to 12 with varied numbers of repeats for these possibilities. the k−1 a coefﬁcients of all other terms (1−ωz)j will be aωj n+j . Since all of j −1 these j are less than k. what is c? Although we have deftly avoided the necessity of ﬁnding all of the other terms.

Also there are certain side restrictions. so that (za1 + · · · + za6 )(zb1 + · · · + zb6 ) z2 + 2z3 + 3z4 + · · · + 3z10 + 2z11 + z12 . The only thing left to try is putting both into the a-polynomial. b6 . . b1 . We conclude from (6) that the “a-polynomial” and “b-polynomial” must consist of these factors. and similarly in the b-polynomial. (6) . All that is left to distribute are the two factors of 1 − z + z2 . then we get ordinary dice. . The Idea of Analytic Number Theory sums then are the terms of the product (z + z2 + z3 + z4 + z5 + z6 )(z + z2 + z3 + z4 + z5 + z6 ) z2 + 2z3 + 3z4 + 4z5 + 5z6 + 6z7 + 5z8 + 4z9 + 3z10 + 2z11 + z12 The correspondence. Thus z + z2 + z3 + z4 + z5 + z6 1−z z(1+z +z2 )(1+z3 ) z(1+z +z2 )(1+z)(1−z +z2 ). . repeating the question. If one apiece are given to the a. says that there are 3 ways for the 10 to show up. . The a-polynomial must be 6 at z 1 and so the (1 + z + z2 )(1 + z) factor must appear in it. .6 I. etc. a6 . The question is: Is there any other way to number these two cubes with positive integers so as to achieve the very same alternatives? Analytically. .and b-polynomials. then. let us factor completely (over the ratio6 z 1−z nals) this right-hand side. for example. The a’s and b’s are to be positive and so a z-factor must appear in both polynomials. the coefﬁcients of z10 being 3. . They look totally different from ordinary dice but they produce exactly the same results! So. These would be the “Crazy Dice” referred to in the title of this section. a1 . . can (za1 + · · · + za6 )(zb1 + · · · + zb6 ) (z + z2 + z3 + z4 + z5 + z6 ) × (z + z2 + z3 + z4 + z5 + z6 )? To analyze this possibility. the question amounts to the existence of positive integers.

n a + a }. Now we introduce the notion of the representation function. Does order count? Can the two summands be equal? Therefore we introduce three representation functions.6. order doesn’t count. n a + a }. and they can be equal. namely. where obviously r(n)zn A2 (z). a ) : a. r(n) #{(a.4. So here order counts. In terms of the generata ing function for the set A. a ∈ A.Crazy Dice 7 This works! We obtain ﬁnally za z(1 + z + z2 )(1 + z)(1 − z + z2 )2 z + z3 + z4 + z5 + z6 + z8 and zb z(1 + z + z2 )(1 + z) z + 2z2 + 2z3 + z4 .4. The simplest is that of r(n).3. So. a < a .2. A(z) a∈A z . a ∈ A. we must subtract A(z2 ) from A2 (z) to remove the case of a a and then divide by 2 to remove the order. Translating back. and they can’t be equal. suppose there is a set A of nonnegative integers and that we wish to express the number of ways in which a given integer n can be written as the sum of two of them.5. the crazy dice are 1.3. r+ (n) #{(a. So here r− (n)zn 1 2 [A (z) − A(z2 )].3. and they can be equal. n a + a }. The trouble is that we must decide on conventions. 2 (8) .8 and 1. (7) To deal with r− (n). a ) : a. a ≤ a . a ∈ A. r− (n) #{(a. a ) : a.2. order doesn’t count. we can express the generating functions of these representation functions.

and r+ (6) 2. we must add A(z2 ) to this result to reinstate the case of a a . But now we are stymied since now 6 1 + 5. using (9). else r+ (4) 2. r+ (n) constant? We will analyze this question by using generating functions. The suspicion arises. So. so that. A2 (z) remains nonnegative. though. Continuing in this manner. else 1 would be expressible as . And then 2 ∈ A. (10) 2 1−z P (z) is a polynomial. as it is as the sum of two distinct members of B? If we experiment a bit. And r+ (0). a contradiction. and we obtain 1 2 (9) [A (z) + A(z2 )]. 6 3 + 3. r+ (n) is the same for all n? The answer is NO. else r+ (1) / And then 3 ∈ A. we ﬁnd 5 ∈ A. then 4 ∈ A. and A(z2 ) goes to A(1) ∞. The Idea of Analytic Number Theory Finally for r+ (n). the question reduces to whether there is an inﬁnite set A for which C 1 2 [A (z) + A(z2 )] P (z) + . for we would have to have 0 ∈ A. Answer: No. and begin by placing 0 ∈ A. then 1 ∈ A. except for some misbehavior at the beginning. r+ (n)zn 2 Can r(n) be “constant?” Is it possible to design a nontrivial set A. Couldn’t A be designed so that. before we get down to business. then 1 ∈ B.8 I. Clearly C P (z) and 1−z remain bounded. A Splitting Problem Can we split the nonnegative integers in two sets A and B so that every integer n is expressible in the same number of ways as the sum of two distinct members of A. else r+ (3) 0 (whereas r+ (1) 1). else r+ (2) / 2. that this impossibility may just be a quirk of “small” numbers. say. Just look what happens if we let z → (−1)+ .

else 2 would be a + a but not b + b . 5. But the pattern is not clear. (14) 1 1−z A(z2 ) − B(z2 ).A Splitting Problem 9 a + a but not as b + b . if we continue to iterate. 4. And. Next 3 ∈ A. A(z2 ) − B(z2 ). · · ·}. nor is the existence or uniqueness of the desired A. B. 8. (1 − z2 )[A(z4 ) − B(z4 )]. 6. So observe that we are requiring by (8) that 1 2 [A (z) − A(z2 )] 2 1 2 [B (z) − B(z2 )]. · · ·} and B {1. by (12). (13) 1 . we seem to force A {0. else 3 would not be a + a whereas it is b + b 1 + 2. 1−z (12) Now this is a relationship that can be iterated. (15) n−1 n n . we also have the condition that A(z) + B(z) From (11) we obtain A2 (z) − B 2 (z) and so. We see that A(z2 ) − B(z2 ) so that continuing gives A(z) − B(z) (1 − z)(1 − z2 )[A(z4 ) − B(z4 )]. B be a splitting of the nonnegatives. because of the condition that A. we obtain A(z) − B(z) (1 − z)(1 − z2 ) · · · (1 − z2 ) A(z2 ) − B(z2 ) . 2 (11) Also. 7. 2. 3. we conclude that [A(z) − B(z)] · or A(z) − B(z) (1 − z)[A(z2 ) − B(z2 )]. Continuing in this manner. Next 2 ∈ B. We must turn to generating functions. 9.

otherwise. and it has coefﬁcient −1. Then and only then 2k+1 111 · · · 1 is not the sum of two distinct A’s. since A(0) that A(z) − B(z) ∞ i 0 1. We have achieved success! The sets A and B do exist. “oh. again. This gives a 0 digit so. And this product is easy to “multiply out”. B(0) i 0. 22k+1 − 1 Proof. why the A and B have the same r− (n). we deduce (16) (1 − z2 ). to what this common r− (n) is equal. and indeed are given by A Integers. are unique. Indeed zn occurs with coefﬁcient +1 if n is the sum of an even number of distinct powers of 2. We must now show that all other n have So r− (22k+1 − 1) a representation as the sum of two numbers whose numbers of 1 digits are of like parity. which are the sum of an odd number of distinct powers of 2. of course.” It isn’t really trivial. and B Integers. First of all if n contains 2k 1’s then it is the sum of the ﬁrst k and the second k. it’s not 11 · · · 1. or for that matter. which are the sum of an even number of distinct powers of 2. A sum of two A’s. even in retrospect. (See below where it is proved that r− (22k+1 − 1) 0. with no carries has an even number of odd 1’s (so it won’t give 111 · · · 1). Every term zn occurs uniquely since every n is uniquely the sum of distinct powers of 2. This is not one of those problems where. after the answer is exposed. The Idea of Analytic Number Theory and so. one proclaims. else look at the ﬁrst carry.10 I. say. Secondly if n contains 2k + 1 1’s but also a 0 digit then it is structured as 111 · · · ◦A where A contains 2k + 1 − m 1’s and. by letting n → ∞. is of total length L then it can be expressed as 111 · · · 1 ◦ 00 · · · 00 plus 1A and these two numbers m−1 2 m . 0.) A Integers with an even number of 1’s in radix 2.

where repeats are not allowed. For. To prove this theorem we produce two generating functions. An Identity of Euler’s Consider expressing n as the sum of distinct positive integers. however. 7. each zk factor occurs at most once. 1 + z2 . . 1 + 5. but this time where repeats are allowed. shows that this generating function is given as the product of 1 + z. that is. 3. 1 + 1 + 1 + 1 + 1 + 1.. . A moment’s thought. (So for n 6. (So For n 6. and a theorem of Euler’s says that this is no coincidence. the other generating function is (1 + z)(1 + z2 )(1 + z3 ) · · · .) In both cases we obtained four expressions for 6. . 1 + 1 + 1 + 3. . . it says the following: Theorem. 5. 3 + 3. This generating function is given by 1 . we get 1 + 5. 1 + z3 . The number of ways of expressing n as the sum of distinct positive integers equals the number of ways of expressing n as the sum of (not necessarily distinct) odd positive integers. (1 + z)(1 + z2 )(1 + z3 ) · · · (19) (18) . when these are multiplied out. These are again of like parity so we are done.) Also consider expressing n as the sum of positive odd numbers. . we have the expression 1 + 2 + 3 and also 2 + 4. i. In short. .e. Euler’s theorem in its analytic form is then just the identity 1 (1 − z)(1 − z3 )(1 − z5 ) · · · throughout |z| < 1.An Identity of Euler’s 11 have respectively m 1’s and 2k + 2 − m 1’s. (1 − z)(1 − z3 )(1 − z5 ) · · · (17) The other generating function is not of the coin changing variety because of the distinctness condition. . and just plain 6 alone. The latter is exactly the “coin changing” function where the coins have the denominations 1.

when this product is multiplied out. can there be integers a1 < a2 < · · · < an such that the differences ai − aj . 5. a expansion of P (z). k 0. 3. then the differences are exposed.j 1 i>j z ai −aj +n+ n i. 1 + z4 . all of the terms (aside from the 1) cancel each other! To prove (2) multiply the 1 − z by the 1 + z (to get 1 − z2 ) and do the same with 1 − z3 by 1 + z3 . this is a “perfect” situation. . take on all the values 1. P (z) is just its constant term 1. Thus we can remove the 2.. are there any larger perfect values? In short. and the marks at 0. not when we square A(z). The question suggests 6. is equal to (1 − z2 )(1 − z6 )(1 − z10 ) · · · (1 + z2 )(1 + z4 ) · · · P (z2 ) which of course which just happens to be P (z2 )! So P (z) 0. These rearrangements are justiﬁed by absolute convergence. This gives the new factors 1 − z2 . Using this ruler we may of course measure any integral length from 1 through 6. 1 − z6 . we obtain A(z) · A 1 z n i. . · · ·. as asserted. (The 2 can be measured between 4 and 6. 4. Thus A(z) · A( 1 ) i. 1. call it P (z). and i < j . 2. 3. 1 − z10 . But we don’t need all of these markings to accomplish these measurements. 3.) Since 2 itself then. n ? 2 n ai If we introduce the usual generating function A(z) i 1z . 2. · · · and leaves untouched the old factors 1 + z2 . in the means that there can’t be any terms azk . and so we see that the product in (20). . . The Idea of Analytic Number Theory Another way of writing (19) is (1 − z)(1 − z3 )(1 − z5 ) · · · (1 + z)(1 + z2 )(1 + z3 ) · · · 1 (20) which is the provocative assertion that. 1 + z6 . . i. the 3 can be gotten between 1 and 4.j 1 i<j zai −aj . etc. i j . 6 are sufﬁcient. and 5.j 1 z z z if we split this (double) sum as i > j . Marks on a Ruler Suppose that a 6” ruler is marked as usual at 0. i > j . but when n ai −aj and we multiply A(z) by A( 1 ). 4. 1.e. and the 5 between 1 and 4 6. 6.12 I.

Dissection into Arithmetic Progressions 13 Our “perfect ruler. and since the last sum is the same as equal to N 1 zk . N k 2 1 ﬁrst. But 2 2n − 2n + 2 − 3π(n − 1) > 2n2 − 2n + 2 − 10(n − 1) 2. sin1 θ > θ . A(z) · A 1 z zN +1 − z−N + n − 1. N n . z−1 N n . then. summing this geometric series. for n ≥ 5.) 2 −1. if we pick a θ which makes sin n2 −n+1 θ 2 1 sin 2 θ < −(n − 1). 2 or. (23) (And we had better assume that n ≥ 5.” by hypothesis. then requires that the ﬁrst sum be n . since we saw the perfect ruler for n 4. There are no perfect 2(n − 3)2 − 6 ≥ 2 · 22 − 6 rulers! 2 < −θ 2n2 −2n+2 3π − 2n2 −2n+2 3π 2 . we let z lie on the unit circle z eiθ . and so the requirement (23) follows . (22) A contradiction will occur. In that case sin θ < θ . 2 2 n2 −n+1 − 1 sin θ 2 from − < −(n − 1) or 2n2 − 2n + 2 > 3π(n − 1). for examA good choice. 2 (21) In search of a contradiction. whereas the right-hand side is zN + 2 − z−(N + 2 ) 1 1 z − z− 1 2 1 2 +n−1 sin(N + 1 sin 2 θ 1 2 )θ +n−1 and (21) reduces to A(e ) iθ 2 sin n2 −n+1 θ 2 1 sin 2 θ + n − 1. is to make sin n −n+1 θ 2 3π 2 ple by picking θ . our equation takes the simple form A(z) · A 1 z N k −N zk + n − 1. then. iθ 2 so that the left side of (21) becomes simply |A(e )| . with z replacing z.

4n + 1.14 I. indeed. Indeed there are many other ways. just as the experiment suggested. The Idea of Analytic Number Theory Dissection into Arithmetic Progressions It is easy enough to split the nonnegative integers into arithmetic progressions. (24) Well. and responds to the identity ∞ 0 zn n n 0z n the dissection into 2n. 1. . this chapter has helped take the sting out of the preposterous notion of using analysis in number theory. . 2. n 0. The progression an + b. Thus the dissection into evens and odds corn ∞ 2n + ∞ 0 z2n+1 . we can express their sums by ∞ 0 zan+b n 1−za question then is exactly whether there can be an identity z b1 z b2 z bk 1 + + ··· + . So the question arises Can the positive integers be split into at least two arithmetic progressions any two of which have a distinct common difference? Of course we look to generating functions for the answer. Our is geometric. all we need do is let z → e ak and observe that then all of the terms in (24) approach ﬁnite limits except the last term z bk which approaches ∞. etc. To see that (24) does. Since each of these series n 0z + n 0z n zb . 4n + 3 corresponds to ∞ 0 zn n ∞ ∞ 2n 4n+1 + ∞ 0 z4n+3 . 1−zak Hopefully. (24) is impossible. there cannot be such a dissection. For example they split into the evens and the odds or into the progressions 2n. 4n + 1. but all seem to require at least two of the progressions to have same common difference (the evens and odds both have 2 as a common difference and the 4n + 1 and 4n + 3 both have 4). 1−z 1 − z a1 1 − z a2 1 − z ak 1 < a1 < a2 < . . . then. 4n + 3. . lead to a 2π i contradiction. < ak . will be associated with the function ∞ 0 zan+b . .

2. e . Show that every set satisfying the conditions of (1) must have √ |A| ≤ N . with no knowledge of Stirling’s formula. but √ with |A| ≤ 4N + 1. Produce a set A such that r(n) > 0 for all n in 1 ≤ n ≤ N.Problems for Chapter I 15 Problems for Chapter I 1. Show directly. 3. that n! > ( n )n .

5. then the problem becomes too simple. There is a famous story concerning the search for some kind of pattern in this table. are positive integers? It turns out that there are two distinct questions here. viewed from a distance. 4 + 1. The ﬁrst crude assessment of p(n)! Among other things. 2 for n 2. e. b. questions one can ask in arithmetic is how to determine the number of ways of breaking up a given integer. 3 + 2. or p(n) itself √ is very roughly eα n . 5 for n 4. 7 for n 5. The answer is just 2n−1 and the proof is just induction. depending on whether we elect to count the order of the summands. 3 for n 3. the number of partitions of n. . is around C n. . and no others. this does tell us not to expect any simple answers. Things are incredibly different and more complicated if order is not counted! In this case the number of breakups or “partitions” is 1 for n 1. 3 + 1 + 1. That is. 5 has the representations 1 + 1 + 1 + 1 + 1. If we do choose to let the order count. 2 + 2 + 1. It suddenly occurred to him that. 2 + 1 + 1 + 1. This is told of Major MacMahon who kept a list of these partition numbers arranged one under another up into the hundreds. Remember such expressions as 1 + 1 + 2 + 1 are not considered different. however. The table can be extended further of course but no apparent pattern emerges. .II The Partition Function One of the simplest.. most natural. we ask about a positive integer n: In how many ways can it be written as a + b + c + · · · where a.g. the outline of the digits seemed to form a parabola! Thus the number of digits √ in p(n). Indeed later research showed that the true asymptotic √ π 2n/3 formula for p(n) is e 4√3n . c. certainly not a formula to be guessed! 17 .

How does one go from one to the other? Mainly how does one go from a function to its coefﬁcients? It is here that complex numbers really play their most important role. again with f (z) an zn . The Generating Function To put into sharp focus the fact that order does not count. then an . The analysis in that problem extends verbatim to this one. This is always the tricky (creative?) part of the process. So we obtain ∞ n 0 p(n)zn ∞ k 1 1 1 − zk (1) valid for |z| < 1. The Partition Function Now we turn to the analytic number theory derivation of this asymptotic formula. . But we don’t know exactly how this translates to the generating function. Except for rare “made up” examples there is very little hope of obtaining the nth derivative of a given function and even estimating these derivatives is not a task with very good prospects. where we understand that p(0) 1. if f (z) an zn . even though we now have an inﬁnite number of coins. we turn to the second stage of attack. seems to be the paramount step. Having thus obtained the generating function. this time we have the formula . Thus. n! expressing the desired coefﬁcients in terms of high derivatives of the function. But this a terrible way of getting at the thing. we may view p(n) as the number of representations of n as a sum of 1’s and 2’s and 3’s . But this is just the “change making” problem where coins come in all denominations. We know pretty well what kind of information we desire about p(n): an estimate of its growth. Thus f (n) (0) we learned in calculus that. perhaps even an asymptotic formula if we are lucky. investigating the function. The point is that there are formulas (for said coefﬁcients). To grasp the connection between the generating function and its coefﬁcients. Cauchy’s theorem gives a different and more promising approach. Face it. etc. .18 II. then. . the calculus approach is a ﬂop.

we ∞ ∞ 1 log 1 − zk 1 j ∞ k 1 k 1 j 1 ∞ j 1 zkj j (2) z jk 1 zj . but armed with the knowledge that the valuable information about f (z) will help in getting a good approx(z) 1 imation to C fn+1 dz. and taking logarithms. because integral operators are bounded. Again. is it? So let us get under way.e. The price we pay is that of passing to the complex numbers for our z’s. j 1 − zj Now write z e−w so that w > 0 and obtain log F (e−w ) ∞ 1 1 1 k 1 k ekw −1 . All in all. Thus noticing that the expansion of ex −1 begins with −x 1 1 1 − 2 + c1 x + · · · or equivalently (near 0) x − e 2 + cx + · · ·. we see that we should seek approximations to our generating function which are good for |z| near 1 with special importance attached to those z’s which are near +1. a look at our generating function p(n)zn shows that it’s biggest when z is positive (since the coefﬁcients are themselves positive). But a glance at the potentially explosive zn+1 z shows us that C had better stay as far away from the origin as it can. Not a bad price. F (z) obtain log F (z) ∞ k 1 ∞ j 1 ∞ 1 k 1 1−zk . and differential operators are not. it must hug the unit circle. The Approximation Starting with (1).The Approximation 19 f (z) 1 an dz. i. x we rewrite this as log F (e−w ) + 1 k 1 k 1 e−kw − kw 2 1 1 e−kw − + ekw − 1 kw 2 (3) . an integral rather than a differential operator! 2π i C zn+1 Surely this is a more secure approach..

. However. ∞) and that h > 0. (4) To be sure. by merely applying it to the real and imaginary parts. The Riemann sum ∞ 1 φ(kh)h is clearly equal to the k area of the union of rectangles and so is bounded by the area under ∞ y φ(x). Riemann Sums Suppose that φ(x) is a positive decreasing function on (0. Combining these two inequalities tells us that the Riemann sum lies within h · φ(0) of the Riemann integral. Thereby we obtain ∞ k 1 [φ(kh) − ψ(kh)]h − 0 ∞ [φ(x) − ψ(x)] h[φ(0) + ψ(0)]. Indeed we recognize any A(kw) 1 A(kw) w as a Riemann sum. The form of this series is very suggestive. Calling φ(x) − ψ(x) F (x) and then observing that φ(0) + ψ(0) is the total variation V of F (x) we have the rather general result ∞ k 1 F (kh)h − 0 ∞ F (x) h · V (F ). exceeds the area under y φ(x). approximating series k kw ∞ the Riemann integral 0 A(t) dt for small positive w. Hence ∞ 1 φ(kh)h ≤ 0 φ(x)dx. This is all very nice and rather accurate but it refers only to decreasing functions. k the series ∞ 0 φ(kh)h can be construed as the area of this union of k these rectangles and. we have proven this result only for real functions but in fact it follows for complex ones.20 II. that such series are estimated rather accurately. So ∞ ∞ this time we obtain k 0 φ(kh)h ≥ 0 φ(x)dx. On the other hand. The Partition Function 1 π2 + log(1 − e−w ) 6w 2 1 1 1 e−kw + − + k ekw − 1 kw 2 . we may easily remedy this restriction by subtracting two such functions. It should come t as no surprise then. as such. So let us review the “Riemann sum story”.

in our case of an analytic F . F (kheiθ )h − 0 ∞ F (xeiθ )dx h · Vθ (F ) (Vθ is the variation along the ray of argument θ). Later on we show that ∞ 0 1 1 e−x − + ex − 1 x 2 dx x 1 log √ . We also may use the formula Vθ (F ) ∞ |F (xeiθ )|dx and ﬁnally deduce that 0 ∞ k 1 F (kw)w − 0 ∞ F (x)dx w 0 ∞ |F (xeiθ )|dx. (Simply apply Cauchy’s theorem and observe that at ∞ F falls off like x12 ). 2π (5) and right now we may note that the (complicated) function F (xe ) iθ 2 e−xe e−xe − − x 3 e3iθ 2x 2 e2iθ 2xeiθ iθ exe 1 − − 2 2iθ xeiθ x e (e − 1) xeiθ (exeiθ − 1)2 M (x+1)2 iθ iθ is uniformly bounded by M(c)). so that ∞ k 1 F (kw)w − 0 ∞ F (xeiθ )d(xeiθ ) w · Vθ (F ).Riemann Sums 21 To modify this result to ﬁt our situation. this integral is actually independent of θ . and conclude from (4) that ∞ k 1 heiθ . let us write w h > 0. −π/2 < θ < π/2. . so that we obtain ∞ k 1 in any wedge |θ | < c < π/2(m + 1 k 1 e−kw 1 + − ekw − 1 kw 2 1 − log √ 2π Mw (6) throughout | arg w| < c < π/2. Furthermore.

everything is negligible by comparison. (7) is really of no use away from z 1. 12 1−z However. So note that. near 1. (3). ∞ k 1 1 z 1 1 − zk 1−z exp 2π in π2 1 + z 12 1 − z [1 + O(1 − z)] (7) |1 − z| ≤ c.22 II. 1 − |z| (1 − z) + 1 1+z 2 1−z But we perform one more “neatening” operation. away from 1. or 1+z 1 log + O(1 − z). Here we see that we can replace our generating function by the elementary function exp π 1+z whose coefﬁcients should then prove amenable. we see that something else must be supplied. All we need to do is combine (1). Finally then. 1 − |z| This is our basic approximation. since Cauchy’s theorem requires values of z all along a closed loop surrounding 0. We have prepared the way for the useful approximation to our generating function. which we have decided is the most important locale. and (6). It is good near z 1. we must replace it (before anything good can result). replace w by log 1 . and exponentiate. and. Indeed we will show that. log 1 z (1−z)2 2 + (1−z)3 3 +··· 2 1−z + O((1 − z)3 ). 1−z 2π 2 . Thus log 1 is z an eyesore! It isn’t at all analytic in the unit disc. The Partition Function The Approximation. The result is z ∞ k 1 1 1 − zk 1−z exp 2π in π2 6 log 1 z [1 + O(1 − z)] |1 − z| ≤ c.

6 π2 −1 6 1 1 − |z| . to commence the ﬁring. we obtain F (z) exp 1 |1 − z| when |1 − z| ≥ 3. So. 1 1 where |1−z| is smaller than 1−|z| . we write p(n) − q(n) 1 2π i C F (z) − φ(z) dz zn+1 (12) . F (z) is rather small. (8) an estimate which is just what we need. Thus. π2 2 12 3(1 − |z|) (10) exp The Cauchy Integral. in this same region. 1 − |z| (11) 1−z exp 2π 2 exp 2π π2 1 + z 12 1 − z π2 2 12 1 − z q(n)zn . Armed with these preparations and the feeling that the coefﬁcients of the elementary function φ(z) are accessible. let us return to (2) and conclude that log F (z) − 1 1−z ∞ j 2 1 |z|j j 1 − |z|j 1 1 − |z| ∞ j 2 1 1 j j 1 1 − |z| or F (z) exp 1 + |1 − z| π2 −1 .Riemann Sums 23 To see this. away from 1. we launch our major Cauchy integral attack. It shows that. for example. 1 − |z| ∞ n 0 (9) Also. setting φ(z) φ(z) so that φ(z) exp 1 1 − |z| when |1 − z| ≥ 3.

elementary geometry gives the formula √ 2(1 − r) 4r arcsin √ r and this is easily seen to be O(1 − r). 1 − |z| is the arc |z| r. then. 1 2π i A F (z) − φ(z) dz zn+1 (1 − r)5/2 .. zn+1 and if we use (7) on this ﬁrst integral and (9). i. 1−|z|| Next we break up C as dictated by our consideration of namely.24 II. C is |z| r. |1 − z| ≥ 3. (13) |1−z| . We ﬁnally obtain. The Partition Function and we try C a circle near the unit circle. . As for the length of A. p(n) − q(n) F (z) − φ(z) 1 1 dz + 2π i A zn+1 2π i is the arc |z| r. (16) M where M is an absolute constant. into A and B So. |1 − z| ≤ 3.e. (11) on this second integral we derive the following estimates: 1 2π i A F (z) − φ(z) dz zn+1 π2 1 6 1−r × the length of A. exp rn π2 1 6 1−r . M (1 − r)3/2 exp n+1 r (M is the implied constant in the O of (7) when c 3). 1 − |z| (14) (15) B F (z) − φ(z) dz. r < 1.

n r 1−r · 2π r B And this is even smaller than our previous estimate. i. p(n) − q(n) M (1 − r)5/2 exp rn π2 1 6 1−r . (17) But what is r? Answer: anything we please (as long as 0 < r < 1)! We are masters of the choice. If we simply begin with the well-known identity ∞ −∞ e−t dt 2 √ π and make a linear change of variables (a > 0). by (15). So combining the two gives. e dt a −∞ . we obtain. from (17). the bound 2 p(n) q(n) + O n−5/4 eπ √ n/6 . 1 2π i F (z) − φ(z) dz zn+1 1 1 · 2 exp 2π r n+1 1−r 1 2 exp . r 1 − √π . So we 6n choose this r and.e. √ ∞ π −(at−b)2 .The Coefﬁcients of q(n) 25 For the second integral.. The exact minimum is too complicated but the 2 1 1 approximate one occurs when en(r−1) exp π6 1−r is minimized and 1 this occurs when π6 1−r n(1 − r). (18) The Coefﬁcients of q(n) The elementary function φ(z) has a rather pleasant deﬁnite integral representation which will then lead to a handy expression for the q(n). by so doing. and so we attempt to minimize the right-hand side.

and thereby obtain q(n) where Cn Kn (s) eπ Cn √ ∞ −∞ Kn (s)2se −2 s− π √ 2 6 2 ds. that the above integral approaches ∞ 2se −∞ −2 s− π √ 2 6 2 ds ∞ −∞ π u+ √ 2 3 e−u du. Since Kn (s) → 1. we see. we obtain √ ∞ √2 2 π π2 1 zt 2 +π t−t 3 e e dt . √ n Reasoning that the maximum of the integrand occurs near t √ we change variables by t s + n. (21) nn+ 2 . The Partition Function ∞ −∞ or e 2 −a 2 t 2 +2abt e dt √ π b2 e . a 1 π and a 2 1 − z (thinking of z as real Thus if we set b2 6 1−z (|z| < 1) for now). ﬁnally. 2 .26 II. √ π 2n en n! s 1 + 2 √n 2n/3 1 1+ s √ n 2 s 1+ √ n e −s √ n + s2 2n 2n . φ(z) e−π /12 √ (1 − z) π 2 2 2 ∞ −∞ ezt eπ 2 √2 3 t−t 2 dt. (19) Equating coefﬁcients therefore results in q(n) e−π /12 √ π 2 ∞ −∞ t 2n−2 t 2n − n! (n − 1)! eπ √ 2/3 t−t 2 dt (20) the “formula” for q(n) from which we can obtain asymptotics. exp √ 6 1−z 1−z −∞ which gives. at least formally.

The Coefﬁcients of q(n) 27 2 where we have set s odd. We still have two debts outstanding. √ 4 3n √ π √ 2n/3 (23) and our earlier estimate (18) allows us thereby to conclude that p(n) ∼ eπ 2n/3 . Thus using (25) for positive s. since ue−u is √ π π √ 2 3 . √ 4 3n (24) Success! We have determined the asymptotic formula for p(n)! Well. Thus (21) formally √ 2π nnn q(n) ∼ √ . 2 6 2 2 ∞ e−u du −∞ Furthermore. it is equal to becomes 2 3 π √ u π √ + √ . en n! 4 3n eπ 2n/3 √ (22) And score another one for Stirling’s formula. . and we must also prove our 1. Kn (s) ≤ es 2 for s ≥ 0. which in turn gives q(n) ∼ e . almost. by (21). So ﬁrst we observe that xe−x is maximized at x so we deduce that s 1+ √ n (using x (1 + s √ n e −s √ n ≤1 (25) )) and also −s s s2 √ 1 + √ e n ≤ e 2n n (26) s (using x (1 + √n )2 ). We must justify our formal passage to the limit in (21). evaluation (5).

x Next note that 1 + x − ex x2 − 0 te(1−t)x dt .28 II. The Partition Function and using (26) for negative s gives us |Kn (s)| ≤ (1 − s)e ≤ (1 − s)e or |Kn (s)| ≤ (1 − s)e2s 2 2 +1 s2− 2s √ n −s s √ 1+ √ e n n 2n−2 s2− 2s √ n (1 − s)e2s 2 +1−(1+s/√n)2 e n−1 2 n s for s < 0. and the passage to the limit is indeed justiﬁed. √ 2/3s for This bound. But this integral can be split into ∞ 0 N k 1 0 ∞ (1 − e−N x ) e −kx 1 1 − x − 1 e x x (1 − e−N x ) 1+x−e 1 dx + 2 x 2 1 e−x − e−(N +1)x dx. gives us the required dominated convergence.). e.g. s < 0. ∞). integrable over (−∞. (28) Thus (27) and (28) give the bound for our integral in (21) of 2se and 2s(s − 1)e1+π s 2 −2 s− π √ 2 6 for s ≥ 0. To achieve this let us ﬁrst note that as N → ∞ our integral is the limit of the integral ∞ 0 (1 − e−N x ) 1 e−x 1 − + ex − 1 x 2 dx + x ∞ 0 ∞ 0 dx x e−x dx 2x (by dominated convergence. Finally we give the following: Evaluation of our Integral (5).

we may interchange and obtain. for our expression. 2 N N log N − log N! − N + √ What luck! This is equal to log N +1(N/e) and so. 2π (Stirling’s formula was used twice and hence needn’t have √ been 2π used at all! Thus we ended up not needing the fact that C √ in the formula n! ∼ C n(n/e)n since the C cancels against a C in √ the denominator. by Stirling’s N! formula. Hence. indeed approaches log √1 .) . The n! formula with C instead of 2π is a much simpler result. the elementary sum − N k 1 0 N k 1 N k 1 1 t 1 dt + k+t −1 2 (k − 1) log N +1 1 ds s 1 log(N + 1) 2 k k−1 −1 + (k − 1) log k − (k − 1) log(k − 1) − N 1 log(N + 1) 2 N log N − log N − log(N − 1) − · · · − log 1 − N + + 1 log(N + 1) 2 1 log(N + 1). by Fubini.The Coefﬁcients of q(n) 29 and e−x − e−(N +1)x x N +1 1 e−sx ds.

3.. Explain the approximation “near 1” of log z)3 . 2 + 5 is considered a different partition of 7 than 5 + 2). then the total number of partitions on n would be 2n−1 . Why is the Riemann sum such a good approximation to the integral when the function is monotone and the increments are equal? . if order counts (e. The Partition Function Problems for Chapter II 1. Explain the observation that MacMahon made of a parabola when he viewed the list of the (decimal expansions) of the partition function. Why does this lead to 1 log 1 z 1 z as 2 1−z + O (1 − 1+z 1 1+z + O(1 − z)? 2 1−z 4. 2.30 II. Prove the “simple” fact that.g.

The Erd˝ s–Fuchs theorem involves the question of just how nearly o constant r(n) can be on average. Consideration of the double Riemann integral shows that this average approaches the area of the unit quarter circle. The fact that r(n) cannot be constant for an inﬁnite set is really trivial since r(n) is odd for n 2a. the set of perfect squares. a ∈ A. the average value. namely π/4.III The Erd˝ s–Fuchs Theorem o There has always been some fascination with the possibility of near constancy of the representation functions ri (n) (of I (7). and even otherwise. The case of r− (n) is more difﬁcult. is exactly equal to n+1 1 times the number of lattice points in the quarter disc x. y ≥ 0. and we will treat it in this chapter as an introduction to the analysis in the Erd˝ s–Fuchs o theorem. Historically this all began with the set A {n2 : n ∈ N0 }. 31 π +O 4 1 √ n . (8) and (9)). In Chapter I we treated the case of r+ (n) and showed that this could not eventually be constant. r(0)+r(1)+r(2)+···+r(n) → π (r(n) is on n+1 4 average equal to the constant π/4. n+1 x 2 + y 2 ≤ n. Thus fairly simple reasoning shows that r(0) + r(1) + r(2) + · · · + r(n) n+1 whereas more involved analysis shows that r(0) + r(1) + r(2) + · · · + r(n) n+1 π +O 4 1 n2/3 . . and the observation that then r(0)+r(1)+r(2)+···+r(n) .) The difﬁcult question is how quickly this limit is approached. and so for this set A.

Here we encounter a slightly heavier dose. impossible unless C 0. π −π |A2 (reiθ )|dθ π −π ≤ |A(r 2 e2iθ )|dθ + π −π π −π |P (reiθ )|dθ (2) +C dθ . when Erd˝ s and o Fuchs showed. leads nowhere here. by simple analytic number theory. We prove that r− (n) can’t eventually be constant. and the conjecture is that it is actually O not O 1 n 3 + 4 . hand picked for their simplicity and involved only the lightest touch of analysis. and C is a positive constant. namely. For any set A. From (1). |1 − reiθ | . r(0)+r(1)+r(2)+···+r(n) n+1 C+O 1 n 3 + 4 is This will be proved in the current chapter. 1−z (1) P is a polynomial. We proceed.32 III. The Erd˝ s–Fuchs Theorem o 1 n2/3 1 n 3 − 4 Very deep arguments have even improved this to o ample. but ﬁrst an appetizer. Now look for a contradiction. What a surprise then. Now all of these arguments were made for the very special case of A the perfect squares. On the other hand. The simple device of letting z → (−1)+ which worked so nicely for the r+ problem. by integrating the modulus around a circle. further difﬁcult arguments show that it is . after all. the following: Theorem. we obtain. for exfor every > 0. The exercises in Chapter I were. So let us assume that A2 (z) − A(z2 ) P (z) + C . for 0 ≤ r < 1.

π dθ We can also estimate the (elliptic) integral −π |1−reiθ | π dθ 2 0 |1−reiθ | by the observation that if z is any complex number in the ﬁrst quadrant. then. when n m. and 1 − reiθ is in the ﬁrst quadrant. This is the observation that π −π | an e inθ 2 | dθ π −π π −π m. The Erd˝ s–Fuchs Theorem o 33 Certain estimates are fairly evident. Hence this double sum is 2π |an |2 . iθ i also is. is π −π π dθ ≤ 2π + 2 log |1 − reiθ | 1+r 1−r . The derivation is clearly valid for ﬁnite or absolutely convergent series which covers . then |z| ≤ z + z. they are equal to 2π . (4) The integral −π |A(reiθ )|2 dθ is a delight. (3) independent of r (0 ≤ r < 1). Thus since for 0 ≤ θ ≤ π .n an einθ am e−imθ dθ ¯ an am ei(n−m)θ dθ ¯ π −π an am ¯ n. It succumbs to Parseval’s identity. P (z) is a polynomial and so π −π |P (reiθ )|dθ ≤ M.m ei(n−m)θ dθ and these integrals all vanish except that. π + log The bound.III. eie−r iθ 1−re−iθ 1 |1−re−iθ | ieiθ eiθ −r π 0 ≤( + ) ieiθ eiθ −r . Hence π 0 dθ ≤( |1 − reiθ | ( ( + ) ieiθ dθ eiθ − r π 0 + ) log(eiθ − r) + ) log − 1+r 1−r 1+r 1−r .

but we can alleviate that by the obvious monotonicity of A. and so at least we can get an upper bound for such integrals. (6) All four of the integrals in (2) have been spoken for and so. The Erd˝ s–Fuchs Theorem o our case of A(reiθ ) (but it even holds in much greater “miraculous” generalities). by (2) through (6). x ≤ 1 2 a+ + a+ 1 4 . does it? Wherein is the hoped contradiction? We must revisit (1) P (r 2 ) + for this. so what? This says that A(r 2 ) grows only at the order of 1 log 1−r as r → 1− . The conclusion is that π −π |A(r 2 e2iθ )|dθ ≤ 2π A(r 4 ). This yields a pure bound on x. A(r 4 ) ≤ A(r 2 ). again by Parseval. Then 1+r 1−r + M + C log π 1+r . we obtain A(r 2 ) ≤ A(r 4 ) + C M +C+ log 2π π 1+r 1−r . (7) It is a nuisance that our function A is evaluated at two different points. Parseval’s identity gives us π −π |A(reiθ )|2 dθ 2π a∈A r 2a π 2πA(r 2 ).34 III. but it doesn’t say that A(r 2 ) remains bounded. (8) Is something bounded in terms of its own square root? But if x ≤ √ 1 1 √ 1 1 x + a. At any rate. (9) 1−r A(r 2 ) ≤ M + C log π But. (5) The last integral we must cope with is −π |A(r 2 e2iθ )|dθ . unlike integrals of |f |2 . there is no formula for integrals of |f |. we obtain ( x − 2 )2 ≤ a + 4 . in turn A2 (r 2 ) − A(r 4 ) . and obtain A(r 2 ) ≤ √ A(r 2 ) + M + C log π 1+r 1−r . Thereby we obtain. and. But there is always the Schwarz inequality |f | ≤ ( 1 · |f |2 )1/2 . x ≤ a + 4 + 2 .

. a∈A z . We ﬁnd ourselves with a set A C whose ri (n) is “almost” constant and this means that A2 (z) ≈ 1−z .Erd˝ s–Fuchs Theorem o C 1−r 2 35 . A2 (r 2 ) ≥ −M + −M + C .) So let us turn to the Erd˝ s–Fuchs theorem with the same strategy o in mind. Since 1 n our hypothesis (11) can be written as (n + 1)z (1−z)2 1 A2 (z) 1−z C + (1 − z)2 ∞ n 0 an zn . 1 − r2 C 1−r 2 . on the other hand Parseval says that the C 1−z integral of |A2 (z)| is A(r 2 ) and (being fairly small except near 1 1−r 1 1) has a small integral. and. viz. Parseval tells us that A2 (z) is large on average. If this proof seems like just so much sleight of hand. an O(nα ). C > 0. so it must be large elsewhere than just near z 1. A2 (r 2 ) ≥ P (r 2 ) + C 1−r 2 .. (Note that the “elsewhere” in the earlier r+ (n) problem was the locale of −1. this forces A(z) to be large on the positive axis A(r 2 ) > √C 1−r 2 . Erd˝ s–Fuchs Theorem o We assume the A is a set for which r(0) + r(1) + · · · + r(n) C(n + 1) + O(nα ). only O(log ). and so it cannot C really be like 1−z . so that A (z) 1 2 n and therefore 1−z A (z) [r(0) + r(1) + · · · + r(n)]z . let us observe what is “really” going on. In cruder terms. and ﬁnally (10) A(r 2 ) ≥ a rate of growth which ﬂatly contradicts (9) and so gives our desired contradiction. (11) 1 and we wish to deduce that α ≥ 4 . and so even that argument seems to be in this spirit. we introduce the a 2 generating function A(z) r(n)zn . As usual. (So A(r 2 ) < C log 1−r ). On the one hand. to bound A(r 2 ) below by √C 2 for obvious reasons 1−r and then to bound it above by Parseval considerations.

C . but this takes some doing. (17) −π |S(reiθ ) . the Parseval upper bound on A(r 2 ). this “nearness” seems to occur only where (1 − z) an zn is relatively small. that is. We must “enhance” this locale if we are to expect anything from the integration. an O(nα ). (15) an zn |. (12) Of course we may assume throughout that α < 1. (13) As for the other goal. Thereby (12) yields the bound M(1−r 2 )−α−1 for an r 2n . and we do so by multiplying by a function whose “heft” or largeness is all near z 1. A(r 2 ) > √ 1 − r2 C > 0. The Erd˝ s–Fuchs Theorem o ∞ n 0 or A (z) 2 C + (1 − z) 1−z an zn . only in a neighborhood of z 1. again C we wish to exploit the fact that A2 (z) is “near” 1−z . N large. so that we easily achieve our ﬁrst goal namely. (16) |S(reiθ )A(reiθ )|2 dθ π −π π ≤ CN 2 +2 dθ |1 − reiθ | an (reiθ )n |dθ. (14) The multiplication of S 2 (z) by (12) yields [S(z)A(z)] which gives |S(z)A(z)|2 ≤ and integration leads to π −π 2 CS 2 (z) + (1 − zN )S(z) 1−z CN 2 + 2|S(z) |1 − z| an zn . From the look of (12) unlike (1).36 III. A handy such multiplier for us is the function S 2 (z) where S(z) 1 + z + z2 + · · · + zN −1 .

Since the cn are integers. and conclude that −π |S(reiθ ) A(reiθ )|2 dθ 2π |cn |2 r 2n . by (14). (4) on the second. (19) Next. 1 − r2 π 2 −π C > 0. S(r 2 ) > Nr 2N ≥ N(1− N )N ≥ N(1− 2 )2 and by (13). A(r 2 ) > √C 2 . (1 − r 2 )α+1/2 iθ n (21) . 2 √ |an |2 r 2n ≤ 2π NM k<N Applying (13) and (14) again leads ﬁnally to π −π S re iθ √ M N an (re ) dθ ≤ . (4) gives CN e dθ ≤ MN 2 log iθ | |1 − re 1 − r2 (20) and our last integral satisﬁes π −π S(reiθ ) π −π an (reiθ )n dθ S(reiθ ) dθ r 2k 2 π −π ≤ 2π an (reiθ )n dθ n2α r 2n . furthermore. we will use Parseval on the ﬁrst of these integrals. π |F (reiθ )|2 dθ ≥ 2πF (r 2 ). ≥ 2π cn r 2n 2π S(r 2 )A(r 2 ). π So write S(z)A(z) cn zn . and we conclude that 1−r π −π (18) N 4 .Erd˝ s–Fuchs Theorem o 37 As before. C N |S(reiθ )A(reiθ )|2 dθ > √ . |cn |2 2 cn ≥ cn and so this is. if F (z) has integral coefﬁcients. (The general fact then is that. 1 − r2 1 1 Thus. and Schwarz’s inequality together with Parseval on the third.) −π Now we introduce a side condition on our parameters r and N which we shall insist on henceforth namely that 1 ≥ N.

were negative then this righthand side would go to 0. and (21) allows the conclusion C e 1 ≤ N 1 − r 2 log + √ . √ √ 1 . Also “plugging” this choice into (22) gives 3 C 4α−1 ≤ N 4α+2 (2 + 3 log N). success is delicious. (23) M Well. (20). We certainly see in (23) the fact that 1 4α−1 α ≥ 4 . 2 +3 log N notwithstanding. 4α+2 .38 III. Thus and so we elect to choose r. and (23) would become false for large N . so that N 1 − r 2 N(1−r 2 )α 1 N 2α+1 and note happily that our side our choice is to make 1−r 2 condition (18) is satisﬁed. M 1 − r2 N (1 − r 2 )α (22) Once again we are masters of the parameters (subject to (18)). The Erd˝ s–Fuchs Theorem o At last.) . combining (19). (If the exponent of N .

by x + a). y ≥ 0. If x is bounded by its own square root (i. Suppose that a convex closed curve√ its curvature bounded by has δ. x. . then we ﬁnd that it has a pure bound. 4. is ∼ π n2 .Problems for Chapter III 39 Problems for Chapter III 1. in fact 4 π 2 n + O(n). 4 √ 2. is bounded by x 2/3 + ax 1/3 + b? Does this insure a bound on x? 3. Show that the number of lattice points in x 2 + y 2 ≤ n2 .e. By the Riemann integral method show that it is. Produce a convex closed curve with curvature bounded by δ which √ δ doesn’t come within 1200 of any lattice point. instead.. What if x. Show that it must come within 2 δ of some lattice point.

Again the trivial property P0 of just being any set is an afﬁne one. the property PA of not containing any arithmetic progressions is an afﬁne property. We also denote the number of elements of this set by 41 . ﬁnitized form. the set A(n) has P if and only if αA(n) + β has P .IV Sequences without Arithmetic Progressions The gist of the result of Chapter IV is that a sequence of integers with “positive density” must contain an arithmetic progression (of at least three distinct terms). any subset of the nonnegative integers below n with at least n members must contain three terms a. If a set is “fat” enough. (Thus we require that this set has the most members possible. Any subset of a set. the notion of an “afﬁne property” of ﬁnite sets of integers.) There may be several such sets but we choose one of them and denote it by S(n. c where a < b < c and a + c 2b. The shock is that this is so hard to prove. b. So let us agree to call a property P an afﬁne property if it satisﬁes the following two conditions: 1. then for large enough n. For each ﬁxed pair of integers α. More precisely and in sharper. also has P . At any rate we begin with a vastly more general consideration. Thus. This is a shock to nobody. not just to be maximal.P ). for example. which has P . it should contain all sorts of patterns. this is the statement that. Now we ﬁx an afﬁne property P and consider a largest subset of the nonnegative integers below n. which has P . if > 0. 2. β with α 0.

n and so we could conclude 2 that k ≤ f n . we note that such a set must contain roughly the same number of evens as odds.PA ) 2. we are led to deﬁne CP n n limn→∞ f (n.42 IV. f (5. (Thus. for example. The property P is. we con2 clude that both the evens and the odds contain not much more than half the whole set. f (n. .P ). . f (m + n) ≤ f (m) + f (n). and for PA . Thus CP0 announced result about progression = free sequences amounts to the 0.. b2 . bk would be a subset of 0. because P0 is totally permissive. Thereby the evens and the odds must be roughly equinumerous. so that PA is.PA ) 4. in this sense.” let us just note how it will prove useful to us with regard to our arithmetic progression considerations. At any rate. Indeed if 2b1 . Similarly the population of the odd elements of 2 1 S would satisfy this same inequality. It follows easily from conditions 1 and 2 that this f (n) is subadditive. we always have 0 ≤ CP ≤ 1. 2b2 .P ) . . except for P0 .P0 ) n.e. If we recall the fact that subadditive functions enjoy the property that limn→∞ f (n) exn ists (in fact limn→∞ f (n) inf f (n) ).) Delaying for the moment the precise statement of this “randomness. Since n ∼ 2 f (n). For example. then b1 . totally unperstatement that CPA missive. . if integers were chosen truly at random with a probability C > 0. This number is a measure of how permissive the n 1. and we shall content ourselves with the case of PA . . f (3. . The remarkable result proved by Szemer´ di and then later by e Furstenberg is that. . which was proved by Roth. for the trivial property. So. two upper bounds imply the lower bounds. there would automatically be a huge number of arithmetic progres- . The point is simply that. and we may dub CP the permission constant. i. 2bk were its even elements. . The Basic Approximation Lemma It turns out that the extremal sets S(n. Sequences without Arithmetic Progressions f (n. Their proofs are both rather complicated.P ) all behave very much as though their elements were chosen at random. CP is always 0.

In terms of the great Szemer´ di–Furstenberg result that e CP ≡ 0 (except for P P0 ). (1) for any polynomial p of degree at most n.”) From (1) we easily obtain the bound |p(z)| ≤ |ζ − z| m<n |pm (ζ )| + |p(ζ )|. Nevertheless we are not prepared to give the lengthy and complex proofs of this general theorem. then we inherit a bound on that polynomial throughout an arc around that point. and so we conclude the following: If all the partial sums are bounded by M at ζ.) Speciﬁcally. a∈S(n. Lemma. is really just an elaboration of the odds and evens considerations above. where the pm denote the partial sums. uniformly on |z| 1. we have the identity p(z) z 1− ζ pm (ζ ) m<n z ζ m p(ξ ) + z 1− ξ z ξ n .The Basic Approximation Lemma 43 sions formed. if we have a bound on a polynomial and its partial sums at a point. We are proving what in truth is an empty result. and so we must prove the Lemma. So we expect that even an approximate randomness should produce at least one arithmetic progression. together with all of its partial sums at every root of unity of order up to N (N is a parameter to be chosen later). (This simply records the result of the “long division.P ) za CP k≤n zk + o(n). in fact.) The proof. (2) . a Proof. The point is that. Remark. The basic strategy is to estimate qn (z) a∈S z − CP k<n zk . (We do what we can. this is a total triviality. the polynomial is bounded by M(n + 1)throughout an arc of length 2 centered at ζ. (Thereby we will obtain bounds for arcs between the roots of unity which will ﬁll up the whole circle. The precise assertion is that of the following lemma.

To estimate qm (ω). we obtain α β 1 σβ ≥ f (n) − f (n − m) ≥ CP n − f (n − m).e. Sequences without Arithmetic Progressions So let α ≤ N be chosen. (4) . and let us note that the ﬁrst inner sum σβ a∈S a<m a≡β(α) 1 counts the size of a subset of S. and so has at most f m elements (where we α α write f (x) for f ( x )). If we next note that α 1 σβ is exactly the number of elements of β S which are below m and so is equal to f (n) minus the number of elements of S which are ≥ m. m . let us write it as ω α β 1 ωβ 1 − CP a∈S a<m a≡β(α) k<m k≡β(α) 1 . 1. and let ω be any αth root of unity. i. which therefore has P which is afﬁne to a subset of 0..44 α IV. Thus qm (ω) − + α β 1 α β 1 α β 1 α β 1 ωβ f ωβ f m α m α − m α m α − σβ − CP k<m k≡β(α) 1 m α f m α m α m α ≤ f f m α − σβ + − σβ α β 1 α β 1 f α − CP − CP (3) + β 1 2αf σβ − CP m.

Since these are N + 1 points on the circle. · · ·. z2 . (7) We separate two cases: n n Case I: α ≤ n0 . n n Case II: α > n0 . So F ( α ) ≤ α . j j must be within arc length N +1 of one another. two of them 2π 2π zi . or α ≥ n0 . and denote by A A(n. But still α ≤ n0 . zN for z any point on the unit circle. Here we use F ( α ) ≤ F (n) and obtain [2αF ( α ) + F (n)](1+ 2π n0 ) ≤ (2α +1)(1+ 2π n0 )F (n) ≤ 3α(1+ 2π n0 )F (n) α α α (3α + 6π n0 )F (n) ≤ (6π + 3)n0 F (n) ≤ (6π + 3)n0 n0 n ≤ 22 n. From now on we will pick n [ n0 ].The Basic Approximation Lemma 45 Substituting (4) in (3) gives qm (ω) 2α f m α − CP m +(f (n−m)−CP (n−m)). z. Here [2αF ( α ) + F (n)](1 + 2π n0 ) ≤ [2αF ( α ) + α n n n n F (n)](1 + 2π ). (5) α Now we ﬁnd it useful to replace the function f (x) − CP x by its “monotone majorant” F (x) maxt≤x (f (t)−CP t) and note that this F (x) is nondecreasing and satisﬁes F (x) o(x) since f (x) − CP x satisﬁes the same. So (5) can be replaced by qm (ω) 2αF m α + F (n − m) ≤ 2αF n α + F (n) (6) (a bound independent of m). ζ ω and 2π n0 2π 2 α(N +1) ≤ 2 nα gives q(z) [2αF n α + F (n)] 1 + 2π n0 α . So choose n0 so that x ≥ n0 implies F (x) ≤ x. This means | arg zi−j | ≤ N +1 and calling |i − j | α gives the result. and then choose n1 so that x ≥ n1 implies F (x) ≤ n0 x. Thus using (2) for q(z). 1 . and the above is ≤ (2 n + n)(1 + 2π ) (3 + 6π) n < 22 n.P )(where order counts Dirichlet’s theorem can be proved by considering the powers 1.P ) the number of arithmetic progressions from S(n. So let P be any afﬁne property. n ≥ n1 and also will ﬁx N Dirichlet’s theorem1 on approximation by rationals now tells us 2π that the totality of arcs surrounding these ω with length 2 α(N +1) covers the whole circle. In either case Dirichlet’s theorem yields our lemma.

The ﬁnal estimate √ for each of these seven integrals. Parseval equality techniques.P0 ) and it is a simple exercise to show that A(n. 2 All of our discussion thus far has been quite general and is valid for arbitrary afﬁne properties. each a G or a q. g(z) is “small” by the lemma).P ) 3 CP 2 n + o(n2 ). CPA 0. and we easily deduce the following: Theorem (Roth). Indeed. . Each of these other integrals is the product of three functions. and at least one of them is a q. If we abbreviate a∈S za g(z). z (9) k Now writing G(z) CP G(z) + q(z) (where q k<n z . then.E. 2 (8) The proof is by contour integration. then (10) reduces to (8). We ﬁnally become speciﬁc by letting P PA . By our lemma.46 IV. is o(n) nn o(n2 ). we obtain 3 CP 1 2πi |z| 1 G2 (z)G(z−2 ) dz z plus seven other integrals. the number of triples below n which are in arithmetic progression. then we recognize A as the constant term in g(z)g(z)g(z−2 ). z (10) But reading (9) for the property P0 shows that this integral is simply A(n. As such each is estimable by the Schwarz inequality. and so we may write A 1 2πi |z| 1 g 2 (z)g(z−2 ) dz . and so (9) gives A 3 CP 1 2πi |z| 1 G2 (z)G(z−2 ) dz + o(n2 ). 2 Q. We show that A(n. If we substitute this in (9).D. therefore. Both of these functions are either a |G| or a |q|. Sequences without Arithmetic Progressions and equality is allowed). we may estimate each of these seven integrals by o(n) times an integral of the product of two functions. is exactly n .P0 ).

Therefore CPA . which number is 2 3 at most n. and so. Thus A(n.The Basic Approximation Lemma 47 Proof. three equal terms. by (8). the only arithmetic progressions in S(n.PA ) ≤ n. 2 0.PA ) are the trivial ones. By the deﬁnition of PA . CPA n + o(n2 ) ≤ n.

n such that the “measure” given to every A. if d O( n). 3. . . . .’s with common difference d up to 6 obtain their “correct” 1 measure d . by attaching weights onto 1. 2. n.48 IV. show that we can always attach weights onto 1. however. Prove that. 2. Attach a positive rational to each integer from 1 to 12 so that all A. . .P.P. Sequences without Arithmetic Progressions Problems for Chapter IV 1. .’s of common difference √ d. If we insist only on approximation. if we ask for a generalization of this. 2. with common difference ≤ m is within e−n/m 1 of a . .P. . then we can only 1 force the correct measure d for all A.

. 9. he simply stated that it did! That is what we propose to do in this chapter. just to prove the existence of the requisite number of the cubes. but just to prove its existence. 37 ﬁfth powers. Our aim. ﬁfth powers. despite the fact that Waring’s conjecture seems to require lower bounds. the theorem of Dirichlet. and so forth. B. went on. that of Schnirelmann. . 9. etc. and although no serious guess was made as to how the sequence 4 (squares). 19. . These are A. fourth powers. . 49 . So let us ﬁx k and view the kth powers. etc. the evaluation of the Weyl sums. by Schnirelmann’s lemmas below. So ﬁrst we turn to our three basic lemmas which will eventually yield our proof. . 37. We do not attempt to ﬁnd the structure of the 4. One of the wonderful things about this approach is that it requires only upper bounds. 19. fourth powers. . and Waring guessed that this was not just a property of squares. the sum of a ﬁxed number of cubes.V The Waring Problem In a famous letter to Euler. Lagrange had already proved his magniﬁcent theorem that every positive integer was the sum of four squares. . in fact. Waring wrote his great conjecture about sums of powers. and ﬁnally C. but that. also worked. . need be only to produce a g g(k) and an α α(k) > 0 such that the sum of g(k) kth powers represents at least the fraction α(k) of all of the integers. He guessed that every positive integer was the sum of 9 cubes. But the adequate upper bounds are obtained by the so called Weyl sums given below. something seemingly totally impossible for contour integrals to produce. 19 fourth powers.

Fix an integer n which is arbitrary. Clearly. Altogether. If S has density α > positive integers.E. . we get k ∈ S. then S ⊕ S contains all the Proof. . Then S ⊕ S has density at least 2α − α 2 . two of these must be within M+1 of each other. All the gaps in the set S are covered in part by the translation of S by the term of S just before this gap. then. let A be the subset of S which lies ≤ n. Mx all reduced 1 (mod 1). we indeed have α + α(1 − α) 2α − α 2 . 1 2 . Next pick an integer a that makes bx − a equal 1 1 to bx (mod 1). Since A contains more than n/2 elements and B contains at least n/2 elements. The Waring Problem A. b) x M+1 1 for. then 1 ≤ b ≤ M and bx (mod 1) is.D. 3x. Lemma 2. ≤ M+1 . Consider the numbers 0. b Proof. in 1 magnitude. . Theorem (Dirichlet). B. as claimed. Q. We also point out that this is a best possible result as the choice 1 shows for every M. Given a real x and a positive integer M. So suppose they overlap at k. b as asserted. this would make the inequality |b| ≤ M even truer). If these two differ by bx. Let S have density α and 0 ∈ S. the Pigeonhole principle guarantees that they overlap. if they have a common divisior. 2x. then every non-negative integer is the sum of at most k members of S for some k ≥ 1. Since k ∈ A. and let B be the set of all n minus elements of S. Lemma 1. (Again. Hence. If S is a set of integers with positive Schnirelmann density and 0 ∈ S. at least the fraction α of this gap gets covered. and since . Schnirelmann’s Theorem. there exists an integer a and a positive number b ≤ M such that 1 |x − a | ≤ (M+1)b . So |bx − a| ≤ (M+1) which means |x − a | ≤ (M+1)b .50 V. Proof. we may assume that (a. x. So from this covering we have density α from S itself and α times the gaps. .

P (n) be a polynomial of degree k with real coefﬁcients and leading coefﬁcient integral and prime to b. we get n − k ∈ S. . If we count those j which produce a denominator of d.l. that I |S| N −1 2 j −N +1 n∈{1.D. N }. will become bigger than 2 .g. Since this latter quantity. 2. Here – as usual – we denote e(x) e2π ix . So this number of j in the full interval of length 2N + 1 is roughly (2N+1) d. leads to a summing of 2j copies j of S and a density of 1 − (1 − α)2 or more..V. C.N } n∈{j +1.. Lemma 2 tells us that 2j +1 copies of S give us all the integers. b 0 and k ≤ N. and generally we may write S n∈I e P (n) b {1.E.. just as Schnirelmann’s theorem claims. Thereby e P (n) − P (n − j ) b .2.j +N } This inner sum involves a polynomial of degree (k − 1) but has a leading coefﬁcient which varies with j . . These are the two elements of S which sum to n. 1 for large enough j . and may assume w. 3... Let b ∈ Z.. . Then e n∈I P (n) b N 1+o(1) b−2 1−k where the bound depends on k. Evaluation of Weyl Sums.o. b .. The Waring Problem 51 k ∈ B. Q.. which of course must divide b. Repeating Lemma 1 j times. and let I be an interval of length ≤ N. . which represents the degree of P (n). It is clearly true for k 1. then.j +2. We proceed by induction on k. then we observe that this must appear roughly d times in an interval of length b.

The previously cited notions of Schnirelmann allow deducing. Let k > 1 be a ﬁxed integer. There exists a C1 such that. b with (a. the full Waring result from this theorem: There exists a G for which rG (n) > 0 for all n > 0. 1−k Our endpoint will be the following: Theorem.52 V. for each positive integer s. for any positive integers N. To prove our theorem. N n 1 e a k n b ≤ C1 N 1+o(1) b−2 . Now we continue as follows: Lemma 3. 1 and the induction is complete. by the inductive hypothesis is |S| 2 d|b 1 1 N N 2+o(1) 1− k−2 1+o(1) − 2k−2 dN b 2 d ≤ b b 1 1 d|b N 2+o(1) b− 2k−2 bo(1) . So we obtain S N 1+o(1) b− 2k−1 . we write rs (n) nk +···+nk n s 1 ni ≥0 1. then. The Waring Problem The full estimate. then there exists g and C such that rg (n) ≤ Cng/k−1 for all n > 0. k . If. b) 1. since 1 s rs (n) 0 m≤n1/k e(xm ) e(−nx)dx. a.

note that N 1 e(xnk ) has a derivative bounded by n 2π N k+1 . b. N. Thus (1) is a property of large g’s.b. There exists interval Ia. and so (2) follows with c 4π 2g The remainder of our paper. there exists a c > 0 such that 1 0 N n 1 e(xn ) dx > cN g−k k g for all n > 0. (2) To see this. n it persists for C0 and any g ≥ g0 . N n 1 e(xnk ) ≥ N − 2π N k+1 1 4π N k N . (1) is a best possible inequality in that.V. then. Denote by Ia. j [J ]. By Dirichlet’s theorem. Hence. Then. (b + j ) e(xnk ) ≤ . j are integers satisfying N > 0. (1) First some parenthetical remarks about this inequality. throughout any C2 N .b. 4π1N k ). b > 0. 1). it is purely a “magnitude property. Henceforth k is ﬁxed. 2 1 . these intervals cover (0. b) 1. 1 (a.N the 1 x-interval |x − a | ≤ bN k−1/2 . N n 1 > 0 and C2 such that. in the interval (0. and call J N k |x − a |. b b where a.” Again. since | N 1 e(xnk )| ≤ N. will be devoted to the derivation of (1) from Lemma 3. Our main tool is the following lemma: Lemma 4. 0 ≤ a < b. The Waring Problem 53 it sufﬁces to prove that there exists g and C for which 1 0 N n 1 e(xn ) dx ≤ CN g−k k g for all n > 0. for each g. b ≤ N k− 2 .N . in other words. Suppose it is known to hold for some C0 and g0 .

New a York 1945. Part II. Theory and Application of Inﬁnite Series. Knopp. (A) If M is the maximum of the moduli of the partial sums m 1 an . Blackie & Sons.D. 1946. Q. Dover Publications.54 V. since the derivative of | N 1 e(xnk )| is bounded by 2π N k+1 . Now write α 1 b e( a nk ) and b S1 + αS2 . n N n 1 e(xnk ) ≤ ≤ N n 1 a a e( nk ) + x − 2π N k+1 b b + N 1+o(1) 2π N 3/2 2π N ≤ + 1/4 . then N n 1 N 1+o(1) an f (n) ≤ M(V + M ). Vol.] and [G. and note the following two simple facts (A) and (B). 1. 1 1 b b b 2k−1 b 2k−1 by C. The Waring Problem Proof. p. For details see [K.E. o o Aufgaben und Lehrs¨ tze aus der Analysis. which gives the result. 37]. since j 0 automatically. P´ lya und G. Assume 2/3 therefore that b ≤ N . This is almost trivial if b > N 2/3 . then N n 1 f (n) − 0 b n 1 N n 1 N f (t)dt ≤ V . Szeg¨ . n V the total variation of f (t) in 0 ≤ t ≤ N. (B) If V is the total variation of f (t) in 0 ≤ t ≤ N. (3) e(xnk ) where S1 S2 N n 1 N n 1 e e a k n b x− a b −α e nk . and M the maximum of the modulus of f (t) in 0 ≤ t ≤ N. for. x− a b nk . Glasgow. .

(1 + j )1/k (6) C1 bδ Now if we apply Lemma 3 to the case N b. we obtain |α| ≤ 1−k δ 2 . the total variation of e[(x − a )t k ] is equal to 2π |x − a |N k ≤ b b √ 2π N . The Waring Problem 55 We apply (A) to S1 . The result is b √ |S1 | ≤ 4π N + 2b ≤ 5π N 2/3 . whereas M 1. the choice C2 1 1 min δ. (4) Next we apply (B) to S2 and obtain |S2 | ≤ 0 N e a x− b t k √ 2π N . J 1/k Combining this with (5) gives |αS2 | ≤ √ C4 N|α| + 2π N. we note that m n 1 e a k n b −α 0+ b[m/b]<n≤m e a k n b −α ≤ (1 + |α|)b ≤ 2b.V. √ Since j ≤ N and b ≤ N 2/3 . bδ (1 + j )1/k (b + j )1/2 C5 +C6 +C1 +2π . dt + b (5) Since N 0 ∞ 0 e(uk )du converges we get a x− b t k e dt N J 1/k J 1/k 0 e(uk )du ≤ NC3 . k . and by (3) the addition of (4) and (6) gives N n 1 . e(xnk ) ≤ ≤ bδ (1 C5 N + 7πN 2/3 + j )1/k C5 N C6 N + . 4 completes the proof. To do so. . Also.

since the length of each Ia. N Ia.j b ≤ CN g−k (b + j )4 since ∞ b 1 ∞ 1 j 0 (b+j )3 < ∞. and the proof is complete.56 V. By Lemma 4. b.N is at most 2N −k . .N n 1 C7 N g 1 e(xn ) dx ≤ . j gives the estimate C7 N g−k b.b. The Waring Problem Proof of (1). Choose g ≥ 4 .b. given as above. (b + j )4 N k k g Summing over all a.

Problems for Chapter V 57 Problems for Chapter V 1. 2. then show that x is not the sum of 2 cubes. 4. Show that every polynomial is the sum of 3 cubes. then R is unsolvable. . then every polynomial is the sum of g nth powers. b) > 2c. Show that the constant polynomial 1 can be written as the sum of √ 4n + 1 nth powers of nonconstant polynomials.” that is if x is the sum of g nth powers. P a + Qb 5. where c is the degree of R(x). in general. Show. Show that if max(z. that the polynomial x is “pivotal. 3. If we permit polynomials with arbitrary complex coefﬁcients and ask the “Waring” problem for polynomials. but it is the sum of 3 cubes.

one available to the “common mathematician in the streets. “natural. for a Dirichlet series with nonnegative coefﬁcients. This is a sum of a set of quantities rather than the sum of a sequence of them.” A perfect example of such a proof and one central to our whole construction is the theorem of Pringsheim and Landau. Thus the power series in (a − z) continues to converge a bit to the left of b and. A “natural” proof of a “natural” 59 .VI A “Natural” Proof of the Nonvanishing of L-Series Rather than the usual adjectives of “elementary” (meaning not involving complex variables) or “simple” (meaning not having too many steps) which refer to proofs. So let b be the real boundary point of the convergence region of an n−z . Indeed this statement proves itself through the observation that (a−z)k a−z (log n)k is a power series in (a − z) with nonn k k! negative coefﬁcients. we introduce a new one. which is just as undeﬁnable as the others. the real boundary point of its convergence region must be a singularity. by rearranging terms.” This term. The precise statement of the Pringsheim–Landau theorem is that. and suppose that b is a regular point and that b < a. Thus the (unique) power series for an n−z an n−a · na−z has nonnegative coefﬁcients in powers of (a − z). the Dirichlet series converges there also. Here the crucial observation is that a series of positive terms (convergent or not) can be rearranged at will. Addition remains a commutative operation when the terms are positive. then. is one which proves itself. is introduced to mean not having any ad hoc constructions or brilliancies. A “natural” proof. contradicting the meaning of b.

So let us begin with the simplest of 1 . has simple coefﬁcients. by our assumption. but these are then cancelled by the other factor. then it is everywhere convergent. then. A “Natural” Proof of the Nonvanishing of L-Series theorem follows. All in all. was all L-series. We return to our entire function while preserving the nonnegativity of the coefﬁcients. which. Our ultimate aim is to prove that the L-series have no zeros on the line z 1. (1) If a Dirichlet series with nonnegative coefﬁcients represents a function which is (can be continued to be) entire. to form ζ 2 (z)ζ (z + ia)ζ (z − ia). A natural step then would be to make them real by multiplying by the conjugate coefﬁcient function. The only trouble points could be at z 1 or at z 1 − ia where one of the factors has a pole. which. the ζ -function. which of course is also entire. because this surely destroys our everywhere analyticity. Then (sic!) the function ζ (z)ζ (z +ia) would be entire. A dangerous route. has a zero. A bizarre conclusion. ζ (z) nz noticed by Narasimhan and is as follows: Assume. that the Dirichlet series ζ (z)ζ (z + ia) is entire. This is the nonvanishing of the L-series that we referred to in the chapter title. 2 log ζ (z) + log ζ (z + ia) + log ζ (z − 1 1 1 1 ia) p 2 log 1−p −z + log 1−p −z−ia + log 1−p −z+ia p. a real. Our proof. . 2 log ζ (z) + log ζ (z + ia) + log ζ (z − ia).60 VI. We are led. we pass to the logarithm. passing to the logarithm. by Euler’s factorization of the ζ -function. page no. perhaps.) Since these are complicated coefﬁcients dependent on sums of complex powers of divisiors. then. that ζ (z) had a zero at 1+ia. By Euler’s factorization. one with a very nice corollary which we record for future use.v vp vz (2 + p −iva + p +iva ). par contraire. and indeed these coefﬁcients are nonnegative! The dangerous route is now reversed by exponentiating. in fact. but are they positive? (We want them to be so that we can use (1). ζ (z)ζ (z − ia). they aren’t even real. This function is entire and has real coefﬁcients. 63). But how to get a contradiction? Surely there is no hint from its coefﬁcients. Nevertheless let us brazen forth (faint heart fair maiden never won). (See the appendix.

especially if we recall that the coefﬁcients are all nonnegative.E.D. χ7 (7) 1. A “Natural” Proof of the Nonvanishing of L-Series 61 (2) ζ 2 (z)ζ (z + ia)ζ (z − ia) is an entire Dirichlet series with nonnegative coefﬁcients. at the modulus 10. 1. The characters are χ1 : χ1 (1) χ3 : χ3 (1) χ7 : χ7 (1) χ9 : χ9 (1) 1. Q.VI. χ9 (3) 1.3. Let us look. for this leads to the underlying group and hence to its dual group. and so the L-series are L1 (z) p≡1 1 1 − p−z 1 1 − p−z p≡3 1 1 − p−z 1 1 − ip−z p≡7 1 1 − p−z p≡9 1 .7. 1. The pertinent progressions are 10k + 1. χ1 (3) 1. the subseries corresponding to n power of 2 is exactly equal to 1 1 1 · 1−21 · 1−21 which exceeds (1−2−z )2 · 4 along the −z−ia −z+ia (1−2−z )2 nonnegative (real) axis and thereby guarantees divergence at z 0. χ9 (7) 1. χ7 (3) 1. 1. so that the group is the multiplicative group of 1. 1 − p−z 1 .10k + 9. χ1 (7) 1. χ9 (9) 1. The falsity of (3) can be established in may ways. χ1 (9) 1.9 (mod 10). for example. 10k + 7. For example. 1 + p−z L3 (z) p≡1 p≡3 p≡7 1 1 + ip−z p≡9 . χ3 (3) 1. And so we have the promised natural proof of the nonvanishing of the ζ -function which can then lead to the natural proof of the Prime Number Theorem. χ3 (7) 1. 10k + 3. Combining this with (1) implies the unbelievable fact that (3) the Dirichlet series for ζ 2 (z)ζ (z + ia)ζ (z − ia) is everywhere convergent. the group of characters. χ7 (9) 1. We must turn to the general L-series which holds the germ of the proof of the Prime Progression Theorem. Dirichlet pointed out that the natural way to treat these progressions is not one progression at a time but all of the pertinent progressions of a given modulus simultaneously. χ3 (9) 1.

A “Natural” Proof of the Nonvanishing of L-Series L7 (z) p≡1 1 1 − p−z p≡3 1 1 + ip−z p≡7 1 1 − ip−z p≡9 1 . which seems.) The result is the Dirichlet series 1 1 Z(z) −z )4 (1 − p (1 − p−4z ) p≡1 p≡3 × p≡7 1 (1 − p−4z ) p≡9 1 . (1 − p−2z )2 and the problem reduces to showing that Z(z) is zero-free on z 1. 1 + p−z and L9 (z) p≡1 1 1 − p−z p≡3 1 1 + p−z p≡7 1 1 + p−z p≡9 1 . which is the product of L-series and is an entire function except possibly for a simple pole at z 1. 1 − p−z (Here z > 1 to insure convergence and the subscripting of the characters is used to reﬂect the isomorphism of the dual group and the original group.) The generating function for the primes in the arithmetic progressions ((mod 10) in this case) are then linear combinations of the logarithms of these L-series. 1 Of course.62 VI. Set h φ(A). χnh arranged .) Guided by the special cases let us turn to the general one. at ﬁrst glance. What could be more natural or more in the spirit of Dirichlet. . . nh . to be a more attractive form of the problem. . n2 . . this is equivalent to showing that p≡1 1−p−z is zerofree on z 1. and we are better off with Z(z). And so indeed the crux is the nonvanishing of these L-series. however. This is misleading. but to prove these separate nonvanishings altogether? So we are led to take the product of all the L-series! (Landau uses the same device to prove nonvanishing of the L-series at point 1. . . and denote the group elements by 1 n1 . (See the appendix. . and denote by GA the multiplicative group of residue classes (mod A) which are prime to A. So let A be a positive integer. χn2 . Denote the dual ˆ group of GA by GA and its elements by χ1 . .

and conclude that it is entire. ∞ 1 Lemma. The end game (ﬁnal contradiction) is also as before although 2 may not be among the primes in the resultant product. For any θ in [0.VI. A proof that the L-series are everywhere analytic functions with the exception of the principal L-series. form Z 2 (z)Z(z + ia)Z(z − ia). where hj is the order of the group element nj . A “Natural” Proof of the Nonvanishing of L-Series 63 ˆ so that ni ↔ χni is an isomorphism of G and G. write Lni (z) nj p≡nj 1−χni (nj )p −z and ﬁnally set Z(z) 10. Since. L1 at the single point z 1. with dazzling speed. So. which is a simple pole. So again we assume Z(1 + ia) 0. Z(z) is entire except possibly for a simple pole at z 1. by summing. . elementary algebra leads to ni Lni (z). Nonetheless again we see that the subseries of powers of π diverges at z 0 which gives us our QED. for z > 1 1. As before. Appendix. Next. As in the case A 1 Z(z) nj p≡nj (1−p −hj z )h/ hj . we see that a zero of any L-series would lead to the everywhere convergence of the Dirichlet series (with nonnegative coefﬁcients) Z 2 (z)Z(z + ia)Z(z − ia). and we may have to take some other prime π. We note that its logarithm and hence that it itself has nonnegative coefﬁcients so that (1) is applicable. deﬁne f (z) n 1 (n−θ )z − z > 1. Then f (z) is continuable to an entire function. 0 e−nt eθ t t z−1 dt (z) t z−1 dt . and we seek a proof that Z(1+ia) 0 for real a. for z > 1.1). 1 z−1 for Proof. we get (n−θ )z 1 (n − θ)z or 1 1 − (n − θ)z z−1 1 (z) ∞ 0 ∞ 1 (n−θ )z ∞ 0 e−t × 1 (z) ∞ 0 eθ t × t z−1 dt t − 1 e e−t eθ t − et − 1 t t z−1 dt.

and. t we may integrate by parts repeatedly and thereby get 1 1 − (n − θ)z z−1 1 (z + k) ∞ 0 d − dt k e−t eθ t − et − 1 t t z+k−1 dt.64 VI. This gives continuation to z > −k. . A “Natural” Proof of the Nonvanishing of L-Series θt −t Since ee−1 − e t is analytic and has integrable derivatives on [0. ∞). since k is arbitrary. the continuation is to the entire plane.

. 4. Prove that (z) has no zeros in the whole plane. Suppose δ(x) decreases to 0 as x → ∞. Prove. it has poles. that there are inﬁnitely many primes not ending in the digit 1. Prove that there are inﬁnitely many primes p for which neither p + 2 nor p − 2 is prime. although. Produce an ε(x) which goes to 0 at ∞ but for which δ(xε(x)) o(ε(x)). by elementary methods.Problems for Chapter VI 65 Problems for Chapter VI 1. 3. Prove that at least 1/6 of the integers are not expressible as the sum of 3 squares. 2. 5.

but they also required annoying estimates of ζ (z) at ∞. but the point is that these can be effectively minimized by elementary arguments. due to Wiener and Ikehara (and also Heins) get around the necessity of estimating at ∞ and are indeed based only on the appropriate nonvanishing of ζ (z). page 60–61) (z − 1)ζ (z) is analytic and zero-free throughout z ≥ 1. Of course certain errors are introduced thereby. (1) This will be assumed throughout and will allow us to give our proof of the Prime Number Theorem. but they are tied to certain results of Fourier transforms. on the nonvanishing of ζ (z) in z ≥ 1. We propose to return to contour integral methods to avoid Fourier analysis and also to use ﬁnite contours to avoid estimates at ∞. to be e sure. If we ignore the (beautiful) elementary proofs of Erd˝ s and Selberg and focus on the o analytic ones. we ﬁnd that they all have some drawbacks. because the formulas for the coefﬁcients of the Dirichlet series involve integrals over inﬁnite contours (unlike the situation for power series) and so effective evaluation requires estimates at ∞. 67 .VII Simple Analytic Proof of the Prime Number Theorem The magniﬁcent Prime Number Theorem has received much attention and many proofs throughout the past century. So let us begin with the well-known fact about the ζ -function (see Chapter 6. The original proofs of Hadamard and de la Vall´ e Poussin were based. The more modern proofs.

68

VII. Simple Analytic Proof of the Prime Number Theorem

In fact we give two proofs. This ﬁrst one is the shorter and simpler of the two, but we pay a price in that we obtain one of Landau’s equivalent forms of the theorem rather than the standard form π(N ) ∼ N/ log N. Our second proof is a more direct assault on π(N ) but is somewhat more intricate than the ﬁrst. Here we ﬁnd some of Tchebychev’s elementary ideas very useful. Basically our novelty consists in using a modiﬁed contour integral, f (z)N z 1 z + 2 z R dz,

rather than the classical one, C f (z)N z z−1 dz. The method is rather ﬂexible, and we could use it to directly obtain π(N) by choosing f (z) log ζ (z). We prefer, however, to derive both proofs from the following convergence theorem. Actually, this theorem dates back ´ to Ingham, but his proof is a la Fourier analysis and is much more complicated than the contour integral method we now give. an n−z which Theorem. Suppose |an | ≤ 1, and form the series clearly converges to an analytic function F (z) for z > 1. If, in fact, F (z) is analytic throughout z ≥ 1, then an n−z converges throughout z ≥ 1.

**Proof of the convergence theorem. Fix a w in w ≥ 1. Thus F (z + w) is analytic in z ≥ 0. We choose an R ≥ 1 and 1 determine δ δ(R) > 0, δ ≤ 2 and an M M(R) so that
**

z, |z| ≤ R. (2) Now form the counterclockwise contour bounded by the arc |z| R, z > −δ, and the segment z −δ, |z| ≤ R. Also denote by A and B, respectively, the parts of in the right and left half planes. By the residue theorem, 2π iF (w) F (z + w)N z 1 z + 2 z R dz. (3) F (z + w) is analytic and bounded by M in − δ ≤

Now on A, F (z + w) is equal to its series, and we split this into its partial sum SN (z + w) and remainder rN (z + w). Again by the

VII. Simple Analytic Proof of the Prime Number Theorem

69

residue theorem,

A

SN (z + w)N z 2π iSN (w) −

z 1 + 2 z R

−A

dz z 1 + 2 z R dz,

SN (z + w)N z

**with −A as usual denoting the reﬂection of A through the origin. Thus, changing z to −z, this can be written as
**

A

SN (z + w)N z 2π iSN (w) −

z 1 + 2 z R

A

dz z 1 + 2 z R dz. (4)

SN (w − z)N −z

**Combining (3) and (4) gives 2π i[F (w) − SN (w)]
**

A

rN (z + w)N z − F (z + w)N z

+

B

SN (w − z) z 1 + 2 Nz z R z 1 + 2 dz, z R

dz (5)

and, to estimate these integrals, we record the following (here as usual we write z x, and we use the notation α β to mean simply that |α| ≤ |β|): z 1 + 2 z R z 1 + 2 z R 2x along |z| R2 1 δ

∞ n N +1

**R (in particular on A), 2 on the line z δ
**

∞ N

(6) −δ, (7)

1+

|z|2 R2 1

|z| ≤ R, rN (z + w) and SN (w − z)

N n 1

nx+1

≤

dn nx+1

N 0

1 , xN x

(8)

nx−1 ≤ N x−1 +

nx−1 dn

70

VII. Simple Analytic Proof of the Prime Number Theorem

Nx By (6), (8), (9), on A, rN (z + w)N z −

1 1 + N x

.

(9)

SN (w − z) z 1 + 2 z N z R 1 1 2x 4 2 1 + + , ≤ 2 + 2 x x N R R RN

**and so, by the “maximum times length” estimate (M–L formula) for integrals, we obtain
**

A

rN (z + w)N z −

SN (w − z) Nz

z 1 + 2 z R

dz

2π 4π + . R N (10)

**Next, by (2), (6), and (7), we obtain
**

B

F (z + w)N z

R

z 1 + 2 z R

dz

0 −δ

2 M · N −δ dy + 2M δ −R 4MR 6M . ≤ + δN δ R 2 log2 N

nx

2|x| 3 dx R2 2

(11)

Inserting the estimates (10) and (11) into (5) gives F (w) − SN (w) 1 MR 2 M , + + + δ R N δN R 2 log2 N

and, if we ﬁx R 3/ , we note that this right-hand side is < for all large N . We have veriﬁed the very deﬁnition of convergence!

**First Proof of the Prime Number Theorem.
**

Following Landau, we will show that the convergence of n µ(n) n (as given above) implies the PNT. Indeed all we need about this convergent series is the simple corollary that n≤N µ(n) o(N). Expressing everything in terms of the ζ -function, then, we have 1 established the fact that ζ (z) has coefﬁcients which go to 0 on average.

First Proof of the Prime Number Theorem.

71

The PNT is equivalent to the fact that the average of the coefﬁcients of ζζ (z) is equal to 1. For simply note that − ζ (z) ζ − d log ζ (z) dz − d log dz 1 1 − p−z p−z log p 1 − p−z

p

p

d log 1 − p−z dz log p . p−z − 1

p

p

(n) where (n) is log p whenever This last series is the same as nz n is a power of p, p any prime, and 0 otherwise. So indeed the average 1 (n) whose limit being 1 is exactly of these coefﬁcients is N n≤N the Prime Number Theorem. In short, we want the average value of the coefﬁcients of − ζζ (z) − ζ (z) to approach 0. Writing this function as

1 (z)[−ζ (z) − ζ (z)] ζ 1 N

µ(n) nz

log n − nz

d(n) , nz

**we may write this average (of the ﬁrst N terms)as µ(a)[log b − d(b)]
**

ab≤N

1 N

µ(a)[log b − d(b) + 2γ ] −

ab≤N

2γ , N

**where 2γ is chosen as the constant for which
**

K b 1

[log b − d(b) + 2γ ]

√ becomes O( K). Now we use the Landau corollary that n≤N µ(n) conclude that 1 µ(n) δ(N), N n≤N

o(N) to

Nevertheless the transition from (12) to (13) is not a simple one. In this section. Simple Analytic Proof of the Prime Number Theorem where δ(N ) tends to 0. nz 1 nz 1 +z (z − 1)pz−1 p (z − 1) pz p 1 − {t} dt t z+1 1 + Ap (z) −1 . Second Proof of the Prime Number Theorem. which we leave to the reader. for z > 1. and we turn to this now.72 VII. we begin with Tchebychev’s observation that log p − log n p is bounded. (12) p≤n which he derived in a direct elementary way from the prime factorization on n! The point is that the Prime Number Theorem is easily derived from log p − log n p converges to a limit. This done. So. and the proof is complete. we may conclude that (n) n≤N N +O √ N w(N) + O Nw(N)δ N w(N) N + o(N). form the function f (z) Now n≥p ∞ n 1 1 nz p≤n log p p p log p p ∞ n≥p 1 . (13) p≤n by a simple summation by parts. and our trick is to pick a function w(N) which N approaches ∞ but such that w(N)δ w(N ) approaches 0.

by (1). n and from this and the fact. we recognize that p log p pz − 1 −d log ζ (z). By Euler’s factorization formula. however. N (1+ ) N an ≤ n 2 (15) and N N (1− ) an ≥ − 2. then. dz and so we deduce. xpx+1 log p + A(z) . that an + log n is nondecreasing. a double pole with principal part 1/(z − 1)2 + c/(z − 1) at z Thus if we set an F (z) f (z) + ζ (z) − cζ (z) nz n where an p≤n log p − log n − c. for N large. that f (z) is analytic in z ≥ 1 except for 1. n (16) . By applying the Cauchy criterion we ﬁnd that. we proceed to prove an → 0. 73 where Ap (z) is analytic for z > 0 and is bounded by 1 px (px − 1) Hence. from (14). p (14) we deduce that F (z) is analytic in z ≥ 1. we conclude that an converges. f (z) 1 z−1 p + |z(z − 1)| . pz − 1 1 where A(z) is analytic for z > 2 by the Weierstrass M-test. From (12) and our convergence theorem.Second Proof of the Prime Number Theorem.

(17) and (18) establish that aN → 0. so that N N (1− ) an ≤ n aN + N 1− − 1− N (1− ) 1 . N ]. by (14). an ≤ aN + log(N/n) ≤ aN + /(1 − ). aN ≥ − 1− − 2 ≥ − 2 2 . and so (13) is proved. N 1 N /N 1− N (1− ) n (18) Taken together. So N (1+ ) an /n ≥ (aN − ) N (1+ ) 1/n. an ≥ aN + log(N/n) ≥ aN − . n and (16) gives −2 .74 VII. Simple Analytic Proof of the Prime Number Theorem In the range N to N(1 + ). and (15) yields N N aN + 2 N (1+ ) 1 N n + 2 N /N(1 + ) 2 + 2 . (17) Similarly in [N (1 − ).

In fact. an n an n converges. −1 . 1 . Given that an → 0. is O(nε ) for every positive ε. prove that N n 1 an o(N).Problems for Chapter VII 75 Problems for Chapter VII 1. Show that d(n). n converges and that an − an−1 > prove that 3. show that d(n) n log log n . Given that 2. the number of divisors of n. 4.

65 Generating functions. 23–24 Cauchy’s theorem. 66 Contour integration. 66–68 Crazy dice. 71 Euler’s theorem. 59 Complex numbers. dissection into. 63 Analytic method. 14 sequences without. 42–47 Cauchy criterion. 11–12 Evens and odds. Paul. basic. 5–8 Dirichlet series. 1 of asymptotic formulas. 4 Basic approximation lemma. 2–5 Commutative operation. 66 proof of. 18 Contour integral. 17 nonnegative. 1 breaking up. 42–47 Arithmetic progressions. 65 Convergence theorem. 14 Extremal sets. 41 dissection into. crazy. 62 Dirichlet theorem. 1 Analytic number theory. modiﬁed. 5–8 Dice. 18–19 of representation functions. 60 Euler’s factorization formula. 1–2 Afﬁne property. vii o Erd˝ s-Fuchs theorem. 63 general. 65 Fourier analysis. 46 Contours ﬁnite. 41–47 Asymptotic formula. 60 Erd˝ s. L-series as. 18–19 Change making. 35–38 o Euler’s factorization. 1–14 Analytic proof of Prime Number Theorem. 33 Entire functions. 50 Dissection into arithmetic progressions. 7 Inﬁnite contours. 45. 41 Analytic functions. 65 Integers.Index Addition problems. 61–62 77 . 71 Cauchy integral. 42 Finite contours. 59–60. 14 Elliptic integral. 31. splitting. 65–71 Approximation lemma. 65 inﬁnite. 8–10 L-series as analytic functions.

78 Index ﬁrst proof of. 31 Riemann sums.” 45 “Natural” proof. 65–71 Relative error. 70–72 Pringsheim-Landau Theorem.” 53 Mathematics. 59–63 Nonnegative integers. see Arithmetic progressions q(n). 33–34 Partial fractional decomposition. 50–51 Schwarz inequality. 17–29 Permission constant. 51–52 . 50 PNT. 49–56 Weyl sums. 3–4 Partition function. marks on. see Prime Number Theorem Prime Number Theorem (PNT). 25–29 nonvanishing of. 8–10 Stirling’s formula. 8–10 Nonvanishing of L-series. 63 Lagrange theorem. 43 e Taylor coefﬁcients. 29 Szemer´ di-Furstenberg result. dissection into. 59 of nonvanishing of L-series. 68–70 second proof of. 4 Representation functions. 34 Sequences without arithmetic progressions. see Nonvanishing of L-series zero of any. 59–63 Odds and evens. vii “Monotone majorant. 42 Pigeonhole principle. 69 L’Hˆ pital’s rule. 27. 59 Progressions. arithmetic. coefﬁcients of. 20 double. 60 “natural” proof of. 49 Landau corollary. 7 near constancy of. splitting. 65 analytic proof of. 31 Riemann integral. 3 Tchebychev’s observation. 12–13 Schnirelmann’s Theorem. 14 Parseval upper bound. 4. 13 Waring problem. 7 generating functions of. 41–47 Splitting problem. 46–47 Rulers. 70 Unit circle. 36 Parseval’s identity. 5 o “Magnitude property. 20–25 Roth Theorem.

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd