164 - Additive Number Theory PDF

Graduate Texts
in Mathematics
Melvyn B. Nathanson
Additive
Number Theory
The Classical Bases
Springer
Graduate Texts in Mathematics
164
Editorial Board
S. Axler F.W. Gehring P.R. Halmos
Springer
New York
Berlin
Heidelberg
Barcelona
Budapest
Hong Kong
London
Milan
Paris
Santa Clara
Singapore
Tokyo
Graduate Texts in Mathematics
TAKEUTI/ZARING. Introduction to Axiomatic Set
35 WFRmFR. Banach Algebras and Several
Theory. 2nd ed.

2 OXTOBY. Measure and Category. 2nd ed.
36 KELLEY/NAMIOKA El AL. Linear Topological
3 SCHAEFER. Topological Vector Spaces.

4 HILTONISTAMMBACH. A Course in Homological
Spaces.
37 MONK. Mathematical Logic.
Algebra.
5 MAC LANE. Categories for the Working
Mathematician.
38 GRAUERT/FRrrZSCHE. Severa! Complex

Variables.
6 HUGHES/PIPER. Projective Planes.
40 KEMENY/SNELL/KJ APP. Denumerable Markov
7 SFRRE. A Course in Arithmetic.

8 TAKEUn2ARuNG. Axiomatic Set Theory.
Chains. 2nd ed.

41 APOSTOL. Modular Functions and Dirichlet
Series in Number Theory. 2nd ed.
42 SERRE. Linear Representations of Finite Groups.
43 GILLMAN/JERISON. Rings of Continuous
Functions.
44 KazmiG. Elementary Algebraic Geometry.
45 LOEvE. Probability Theory 1. 4th ed.
9 HUMPHREYS. Introduction to Lie Algebras and

Representation Theory.
10 COHEN. A Course in Simple Homotopy Theory.
I I CONWAY. Functions of One Complex Variable
1. 2nd ed.
12 BEMs. Advanced Mathematical Analysis.

13 ANDERsoN/FuLLEtt. Rings and Categories of
Modules. 2nd ed.
14 GOLUBrrSKY/GUILLEMIN. Stable Mappings and
Their Singularities.
15 BERBERIAN. Lectures in Functional Analysis
and Operator Theory.
16 WiNTER.'Ihe Structure of Fields.
17 ROSENBLATT. Random Processes. 2nd ed.
18 HALMOS. Measure Theory.
19 HALMOS. A Hilbert Space Problem Book. 2nd

ed.
20 HUSEMOLLER. Fibre Bundles. 3rd ed.
21 HUMPHREYS. Linear Algebraic Groups.
22 BARNES/MACK. An Algebraic Introduction to
Mathematical Logic.
23 GREUB. Linear Algebra. 4th ed.
24 HOMES. Geometric Functional Analysis and
Its Applications.
25 HFwrrr/STROMBERG. Real and Abstract
Analysis.
26 MANES. Algebraic Theories.
27 KELLEY. General Topology.
28 ZAIUSKI/SAMUEt.. Commutative Algebra. Vol.1.
29 ZARISKI/SAMUE. Commutative Algebra. Vol.11.
30 JACOBSON. Lectures in Abstract Algebra I. Basic
Concepts.
31 JACOBSON. Lectures in Abstract Algebra II.
Linear Algebra.
32 JACOBSON. Lectures in Abstract Algebra 111.
Theory of Fields and Galois Theory.
33 HIRSCH. Differential Topology.
34 SPrrzER. Principles of Random Walk. 2nd ed.
Complex Variables. 2nd ed.
39 ARVtsoN. An Invitation to C'-Algcbras.
46 LoEva. Probability Theory 11. 4th ed.
47 Mots.. Geometric Topology in Dimensions 2

and 3.
48 SACHS/WI. General Relativity for
Mathematicians.
49 GRut BERG/WEiR. Linear Geometry. 2nd ed.
50 EDWARDS. Fermat's Last Theorem.
51 KitNGENBERO. A Course in Differential
Geometry.
52 HARTSHORNE. Algebraic Geometry.
53 MANIN. A Course in Mathematical Logic.
54 GRAVER/WATKINS. Combinatorics with
Emphasis on the Theory of Graphs.
55 BROWN/PEARCY. Introduction to Operator
Theory I: Elements of Functional Analysis.
56 MASSEY. Algebraic Topology: An Introduction.
57 CRowt;u/Fox. Introduction to Knot Theory.

58 Koat.nz. p-adic Numbers, p-adic Analysis,
and Zeta-Functions. 2nd ed.
59 LANG. Cyclotomic Fields.
60 ARNOLD. Mathematical Methods in Classical

Mechanics. 2nd ed.
61 WHr EHEAD. Elements of Homotopy Theory.
62 KARGAPOLOV/MERLZJAKOV. Fundamentals of
the Theory of Groups.

63 BOLLOBAS. Graph Theory.
64 EDWARDS. Fourier Series. Vol. 1. 2nd ed.
65 WEDS. Differential Analysis on Complex

Manifolds. 2nd ed.
continued after index
Melvyn B. Nathanson
Additive Number Theory

The Classical Bases
Springer
Melvyn B. Nathanson
Department of Mathematics
Lehman College of the
City University of New York
250 Bedford Park Boulevard West
Bronx, NY 10468-1589 USA
Editorial Board
S. Axler
F.W. Gehring
Department of
Mathematics
Michigan State University
East Lansing, MI 48824
USA
Department of
Mathematics
University of Michigan
Ann Arbor, MI 48109
USA
P.R. Halmos
Department of
Mathematics
Santa Clara University
Santa Clara, CA 95053
USA
Mathematics Subject Classifications (1991): 11-01, 11P05, I IP32
Library of Congress Cataloging-in-Publication Data

Nathanson, Melvyn B. (Melvyn Bernard), 1944Additive number theory:the classical bases/Melvyn B.
Nathanson.
cm. - (Graduate texts in mathematics;164)
p.
Includes bibliographical references and index.
ISBN 0-387-94656-X (hardcover:alk. paper)
1. Number theory. 1. Title. II. Series.
QA241.N347 1996
512'.72-dc20
96-11745
Printed on acid-free paper.
C 1996 Melvyn B. Nathanson

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.
Production managed by Hal Henglein; manufacturing supervised by Jeffrey Taub.

Camera-ready copy prepared from the author's LaTeX files.
Printed and bound by R.R. Donnelley & Sons, Harrisonburg, VA.
Printed in the United States of America.
987654321
ISBN 0-387-94656-X Springer-Verlag New York Berlin Heidelberg SPIN 10490794
To Marjorie
Preface
[Hilbert's] style has not the terseness of many of our modern authors
in mathematics, which is based on the assumption that printer's labor
and paper are costly but the reader's effort and time are not.
H. Weyl [ 1431
The purpose of this book is to describe the classical problems in additive number
theory and to introduce the circle method and the sieve method, which are the
basic analytical and combinatorial tools used to attack these problems. This book
is intended for students who want to learn additive number theory, not for experts
who already know it. For this reason, proofs include many "unnecessary" and
"obvious" steps; this is by design.
The archetypical theorem in additive number theory is due to Lagrange: Every
nonnegative integer is the sum of four squares. In general, the set A of nonnegative
integers is called an additive basis of order h if every nonnegative integer can be
written as the sum of h not necessarily distinct elements of A. Lagrange's theorem
is the statement that the squares are a basis of order four. The set A is called a
basis of finite order if A is a basis of order h for some positive integer h. Additive
number theory is in large part the study of bases of finite order. The classical bases
are the squares, cubes, and higher powers; the polygonal numbers; and the prime
numbers. The classical questions associated with these bases are Waring's problem
and the Goldbach conjecture.
Waring's problem is to prove that, for every k > 2, the nonnegative kth powers
form a basis of finite order. We prove several results connected with Waring's
problem, including Hilbert's theorem that every nonnegative integer is the sum of
viii
Preface
a bounded number of kth powers, and the Hardy-Littlewood asymptotic formula

for the number of representations of an integer as the sum of .c positive kth powers.
Goldbach conjectured that every even positive integer is the sum of at most
two prime numbers. We prove three of the most important results on the Goldbach conjecture: Shnirel'man's theorem that the primes are a basis of finite order,
Vinogradov's theorem that every sufficiently large odd number is the sum of three
primes, and Chen's theorem that every sufficently large even integer is the sum of
a prime and a number that is a product of at most two primes.
Many unsolved problems remain. The Goldbach conjecture has not been proved.
There is no proof of the conjecture that every sufficiently large integer is the sum
of four nonnegative cubes, nor can we obtain a good upper bound for the least
number s of nonnegative kth powers such that every sufficiently large integer
is the sum of s kth powers. It is possible that neither the circle method nor the
sieve method is powerful enough to solve these problems and that completely
new mathematical ideas will be necessary, but certainly there will be no progress
without an understanding of the classical methods.
The prerequisites for this book are undergraduate courses in number theory and
real analysis. The appendix contains some theorems about arithmetic functions
that are not necessarily part of a first course in elementary number theory. In a
few places (for example, Linnik's theorem on sums of seven cubes, Vinogradov's
theorem on sums of three primes, and Chen's theorem on sums of a prime and an
almost prime), we use results about the distribution of prime numbers in arithmetic
progressions. These results can be found in Davenport's Multiplicative Number
Theory [ 19].
Additive number theory is a deep and beautiful part of mathematics, but for
too long it has been obscure and mysterious, the domain of a small number of
specialists, who have often been specialists only in their own small part of additive
number theory. This is the first of several books on additive number theory. I hope
that these books will demonstrate the richness and coherence of the subject and
that they will encourage renewed interest in the field.
I have taught additive number theory at Southern Illinois University at Carbondale. Rutgers University-New Brunswick, and the City University of New York
Graduate Center, and I am grateful to the students and colleagues who participated
in my graduate courses and seminars. I also wish to thank Henryk Iwaniec, from
whom I learned the linear sieve and the proof of Chen's theorem.
This work was supported in part by grants from the PSC-CUNY Research Award
Program and the National Security Agency Mathematical Sciences Program.
I would very much like to receive comments or corrections from readers of this
book. My e-mail addresses are nathansn@alpha.lehman.cuny.edu and nathanson@
worldnet.att.net. A list of errata will be available on my homepage at http://www.
lehman.cuny.edu or http://math.lehman.cuny.edu/nathanson.
Melvyn B. Nathanson
Maplewood, New Jersey
May 1, 1996
Contents
Preface
vii
Notation and conventions
xiii
Waring's problem
Sums of polygons
1.1
Polygonal numbers . . . . . . .
1.2 Lagrange's theorem . . . . . .
1.3 Quadratic forms . . . . . . . .
1.4 Ternary quadratic forms . . . .
1.5 Sums of three squares . . . . .
1.6 Thin sets of squares . . . . . .
1.7 The polygonal number theorem
1.8 Notes . . . . . . . . . . . . . .
1.9 Exercises . . . . . . . . . . . .
. .
3
4
. .
. .
. .
7
12
17
. .
24
. .
. .
37
37
38
44
49
71
72
75
77
86
94
Waring's problem for cubes

2.1
2.2
2.3
2.4
2.5
2.6
3
Sums of cubes . . . . . . . . . .
The Wieferich-Kempner theorem
Linnik's theorem . . . . . . . . .
Sums of two cubes
. . . . . .
Notes . . . . . . .
. . . . . .
Exercises . . . . . . . . . . . . .
..
..
The Hilbert-Waring theorem

3.1
3.2
3.3
3.4
Polynomial identities and a conjecture of Hurwitz

Hermite polynomials and Hilbert's identity . . .
A proof by induction . . . . . . . . . . . . . . .
Notes . . . . . . . . . . . . . . . . . . . . . . .
27
33
34
75
Contents
3.5
Exercises
Weyl's inequality
4.1
4.2
4.3
4.4
4.5
4.6
4.7
..
Tools . . . .
. .
Difference operators
94
. .
. .
. .
97
97
99
. .
111
118
118
. .
. .
121
. .
. .
124
125
127
129
133
137
146
147
147
Easier Waring's problem . . . . . .

Fractional parts . . . . . . . . . . .
Weyl's inequality and Hua's lemma
Notes . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . .
5 The Hardy-Littlewood asymptotic formula

The circle method . . . . . . . . . .
5.2 Waring's problem for k II I . . . . .
5.3 The Hardy-Littlewood decomposition
5.4 The minor arcs . . . . . . . . . . . .
5.5 The major arcs . . . . . . . . . . . .
5.6 The singular integral . . . . . . . . .
5.7 The singular series . . . . . . . . . .
5.8 Conclusion . . . . . . . . . . . . . .
5.9 Notes . . . . . . . . . . . . . . . .
5.10 Exercises . . . . . . . . . . . . . . .
5.1
..
102
103
121
..
. .
II The Goldbach conjecture

6
Elementary estimates for primes

6.1
6.2
6.3
6.4
6.5
6.6
7
Euclid's theorem . . . . . . . .
Chebyshev's theorem . . . . .
Mertens's theorems . . . . . .
Brun's method and twin primes
Notes . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . .
151
.
. .
151
. .
. .
. .
. .
153
158
167
173
174
. .
177
177
. .
. .
. .
. .
The Shnirel'man-Goldbach theorem

7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
The Goldbach conjecture . . . . . .

The Selberg sieve . . . . . . . . .
Applications of the sieve . . . . . .
Shnirel'man density . . . . . . . .
Romanov's theorem . . . . . . . .
Covering congruences . . . . . . .
Notes . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . .
..
178
186
191
195
199
204
208
208
Contents
Sums of three primes

8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
211
Vinogradov's theorem . . . . . . . . . .
The singular series . . . . . . . . . . . .
Decomposition into major and minor arcs
The integral over the major arcs . . . . .
An exponential sum over primes . . . . .
Proof of the asymptotic formula . . . . .
Notes . . . . . . . . . . . . . . . . . . .
Exercise . . . . . . . . . . . . . . . . .
9 The linear sieve

9.1
9.2
9.3
9.4
9.5
9.6
9.7
xi
..
. . . . .
A general sieve . . . . .
Construction of a combinatorial sieve
Approximations . . . . . . . . . . .
The Jurkat-Richert theorem . . . . .
Differential-difference equations . . .
Notes . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . .
. .
211
. .
. .
. .
212
213
215
220
227
230
230
231
231
238
244
251
10 Chen's theorem
259
267
267
271
10.1 Primes and almost primes . . .

10.2 Weights . . . . . . . . . . . .
10.3 Prolegomena to sieving . . . .
10.4 A lower bound for S(A, P, z)
10.5 An upper bound for S(Aq, P, z)
10.6 An upper bound for S(B, P, y)
10.7 A bilinear form inequality . . .
10.8 Conclusion . . . . . . . . . . .
10.9 Notes . . . . . . . . . . . . . .
271
272
275
279
. .
. .
281
. .
. .
. .
286
292
297
298
III Appendix
Arithmetic functions
301
The ring of arithmetic functions

A.2 Sums and integrals . . . . . . .
A.1
A.3
Multiplicative functions
A.4 The divisor function .

A.5 The Euler rp-function
A.6 The Mobius function .
A.7 Ramanujan sums . . .
A.8
Infinite products
A.9 Notes
. .
. .
. .
. .
. .
A.10 Exercises .
301
303
308
310
314
317
320
323
327
327
xii
Contents
Bibliography
331
Index
341
Theorems, lemmas, and corollaries are numbered consecutively in each chapter

and in the Appendix. For example, Lemma 2.1 is the first lemma in Chapter 2 and
Theorem A.2 is the second theorem in the Appendix.
The lowercase letter p denotes a prime number.
We adhere to the usual convention that the empty sum (the sum containing no
terms) is equal to zero and the empty product is equal to one.
Let f be any real or complex-valued function, and let g be a positive function.
The functions f and g can be functions of a real variable x or arithmetic functions
defined only on the positive integers. We write
f - O(g)
or
f<< g
or
g>> f
if there exists a constant c > 0 such that

If W1
cg(x)
for all x in the domain of f . The constant c is called the implied constant. We
write
f <<a.b.... g
if there exists a constant c > 0 that depends on a, b, ... such that

If W1 _< cg(x)
xiv
for all x in the domain of f . We write
f - o(g)
if
lim
(x)
x-*oo g(x)
- 0.
The function f is asymptotic to g, denoted
f-g.
if
lim f (x) - 1.
xg(x)
The real-valued function f is increasing on the interval I if f (xt) < f (x2) for all
x1, x2 E I with x1 < x2. Similarly, the real-valued function f is decreasing on
the interval I if f (XI) > f(x2) for all xI, x2 E I with xt < x2. The function f is
monotonic on the interval I if it is either increasing on I or decreasing on 1.
We use the following notation for exponential functions:
exp(x)-e'
and
e(x) - exp(27rix) - e2i".

The following notation is standard:
Z
the integers 0, f1, t2, .. .
R
the real numbers
R"
n-dimensional Euclidean space
Z"
the integer lattice in R'
C
the complex numbers
the absolute value of the complex number z
IzI
J3z
the real part of the complex number z
the imaginary part of the complex number z
z
[x]
the integer part of the real number x,
that is, the integer uniquely determined
(X)
Ilx II
(a1.....
[a1.... , a ]
IXI
hA
by the inequality [x] < x < [x] + 1.

the fractional part of the real number x,
that is, {x} - x - [x] E [0, 1).
the distance from the real number x
to the nearest integer, that is,
Ilxll - min (Ix - nI : n E Z} - min ({x), I - {x}) E [0, 1/2].
the greatest common divisor of the integers a1, ... , a
the least common multiple of the integers a1, ... , a
the cardinality of the set X
the h-fold sumset, consisting of all sums of h elements of A
Part I
Waring's problem
Sums of polygons
Imo propositionem pulcherrimam et maxime generalem nos primi deteximus: nempe omnem numerum vel esse triangulum vex ex duobus
aut tribus triangulis compositum: esse quadratum vel ex duobus aut
tribus aut quatuorquadratis compositum: esse pentagonum vel ex duobus, tribus, quatuor aut quinque pentagonis compositum; et sic dein-
ceps in infinitum, in hexagonis, heptagonis polygonis quibuslibet,

enuntianda videlicet pro numero angulorum generali et mirabili propostione. Ejus autem demonstrationem, quae ex multis variis et abstrusissimis numerorum mysteriis derivatur, hic apponere non licet.... 1
P. Fermat [39, page 303]
I have discovered a most beautiful theorem of the greatest generality: Every number
is a triangular number or the sum of two or three triangular numbers; every number is a
square or the sum of two, three, or four squares; every number is a pentagonal number or
the sum of two, three, four, or five pentagonal numbers; and so on for hexagonal numbers,
heptagonal numbers, and all other polygonal numbers. The precise statement of this very
beautiful and general theorem depends on the number of the angles. The theorem is based
on the most diverse and abstruse mysteries of numbers, but I am not able to include the
proof here....
1.1
1.
Sums of polygons
Polygonal numbers
Polygonal numbers are nonnegative integers constructed geometrically from the

regular polygons. The triangular numbers, or triangles, count the number of points
in the triangular array
The sequence of triangles is 0, 1, 3, 6, 10, 15. ... .

Similarly, the square numbers count the number of points in the square array
The sequence of squares is 0, 1, 4, 9, 16, 25, ... .

The pentagonal numbers count the number of points in the pentagonal array
The sequence of pentagonal numbers is 0, 1, 5, 12, 22, 35, .... There is a similar
sequence of m-gonal numbers corresponding to every regular polygon with m
sides.
Algebraically, for every m > 1, the kth polygonal number of order m+2, denoted
pm(k), is the sum of the first k terms of the arithmetic progression with initial value
1 and difference m, that is,
pm(k)-
1)m+1)
mk(k-1)+k.
2
This is a quadratic polynomial in k. The triangular numbers are the numbers
P (k)
k(k + 1)
2
1.2
Lagrange's theorem
the squares are the numbers
p2(k) - k2,
the pentagonal numbers are the numbers
(k) -
k(3k - 1)
and so on. This notation is awkward but traditional.

The epigraph to this chapter is one of the famous notes that Fermat wrote in
the margin of his copy of Diophantus's Arithmetica. Fermat claims that, for every
m > 1, every nonnegative integer can be written as the sum of m + 2 polygonal
numbers of order m + 2. This was proved by Cauchy in 1813. The goal of this
chapter is to prove Cauchy's polygonal number theorem. We shall also prove the
related result of Legendre that, for every in > 3, every sufficiently large integer is
the sum of five polygonal numbers of order m + 2.
1.2
Lagrange's theorem
We first prove the polygonal number theorem for squares. This theorem of Lagrange is the most important result in additive number theory.
Theorem 1.1 (Lagrange) Every nonnegative integer is the sum of four squares.
Proof. It is easy to check the formal polynomial identity
(X1 +X2+X3 f X2)(YI +Y2+Y3+y4)-Zj+z2+Z3+Z4,
(1.1)
where
ZI
XI Y1 +X2Y2+X3Y3+X4y4
Z2
X1 Y2 - X2Y1 - X3 Y4 + X4Y3
Z3
X1 Y3 - X3Y1 + X2Y4 - X4Y2
Z4
X1 Y4 - X4Y1 - X2Y3 + X3Y2
(1.2)
This implies that if two numbers are both sums of four squares, then their product
is also the sum of four squares. Every nonnegative integer is the product of primes,
so it suffices to prove that every prime number is the sum of four squares. Since
2- 12 + 12 + 02 + 02, we consider only odd primes p.
The set of squares
(a2 I a - 0, 1,
... , (p - l)/2)
represents (p + 1)/2 distinct congruence classes modulo p. Similarly, the set of

integers
(-b2- 1 I b-0, 1,...,(p- 1)/2)
1.
Sums of polygons
represents (p + 1)/2 distinct congruence classes modulo p. Since there are only
p different congruence classes modulo p, by the pigeonhole principle there must
exist integers a and b such that 0 < a, b < (p - 1)/2 and
a2 = -b2 - 1
(mod p),
that is,
a2 + b2 + 1
(mod p).
Let a2+b2+1 - np.Then

z
_2
p<np-a2+b'`+1'`+0'`<2(p
and so
1) +1 < 2 +1 <p2,
I<n<p.
Let m be the least positive integer such that mp is the sum of four squares. Then
there exist integers x1, x2, x3, x4 such that
mp-x +x2+x3+x42
and
1 <m <n < p.

We must show that m - 1.
Suppose not. Then I < m < p. Choose integers yj such that
y; - xi (mod m)
and
-m/2 < y, < m/2
fori - 1,...,4.Then
Y
+ y? + y2 + y4
(mod m)
x 2 + x2 + x3 + x4 - mp =- 0
and
mr-Y2 +Y2+Y3+Y2
for some nonnegative integer r. If r - 0, then y; - 0 for all i and each x2 is divisible
by m2. It follows that mp is divisible by m2, and so p is divisible by m. This is
impossible, since p is prime and I < m < p. Therefore. r > I and

mr - Y2 +Y2 +Y2 +Y2 < 4(m/2)2 - m2.
Moreover, r - m if and only if m is even and y, - m/2 for all i. In this case,
x; = m/2 (mod m) for all i, and so .r? _- (m/2)2 (mod m2) and
mp - xi +x2 +x2 +x4
4(m/2)2 - m2 = 0
(mod m2).
1.3
Quadratic forms
This implies that p is divisible by m, which is absurd. Therefore,
I <r <m.
Applying the polynomial identity (1.1), we obtain
m2rp - (mp)(mr)
_ (xi + xZ + x3 + x4)(Yi + Yz + Ys + Y2)
2
2
Z21+Z2+Z3+
Z4,
where the zi are defined by equations (1.2). Since xi - yi (mod m), these
equations imply that zi = 0 (mod m) for i - 1, ... , 4. Let wi - zi/m. Then
w t , ... , W4 are integers and
rp=w2+w2+w3+w2i,
which contradicts the minimality of m. Therefore, m = I and the prime p is the
sum of four squares. This completes the proof of Lagrange's theorem.
A set of integers is called a basis of order h if every nonnegative integer can be
written as the sum of h not necessarily distinct elements of the set. A set of integers
is called a basis of finite order if the set is a basis of order h for some h. Lagrange's
theorem states that the set of squares is a basis of order four. Since 7 cannot be
written as the sum of three squares, it follows that the squares do not form a basis
of order three. The central problem in additive number theory is to determine if a
given set of integers is a basis of finite order. Lagrange's theorem gives the first
example of a natural and important set of integers that is a basis. In this sense, it
is the archetypical theorem in additive number theory. Everything in this book is a
generalization of Lagrange's theorem. We shall prove that the polygonal numbers,
the cubes and higher powers, and the primes are all bases of finite order. These are
the classical bases in additive number theory.
1.3
Quadratic forms
Let A - (ai.j) be an m x n matrix with integer coefficients. In this chapter, we

shall only consider matrices with integer coefficients. Let AT denote the transpose
of the matrix A, that is, AT - (aT j> is then x m matrix such that
T
ai,j a Ilia
for i - 1, ... , n and j - 1, ... , m. Then (AT )T - A for every m x n matrix A,

and (AB)T
BT AT for any pair of matrices A and B such that the number of
columns of A is equal to the number of rows of B.
Let M,, (Z) be the ring of n x n matrices. A matrix A E M, (Z) is symmetric if
AT - A. If A is a symmetric matrix and U is any matrix in
then UT AU is
also symmetric, since
(UT AU)T = UT AT (UT )T = UT AU.
1.
Sums of polygons
Let SLn(Z) denote the group of n x n matrices of determinant 1. This group acts
as follows: If A E Mn(Z) and U E SL,,(Z), we define
on the ring
A - U - UTAU.
This is a group action, since
and B in M (Z) are equivalent, denoted

A
B,
if A and B lie in the same orbit of the group action, that is, if B - A U -UTAU
for some U E SLn(Z). It is easy to check that this is an equivalence relation. Since
det(U) - I for all U E SLn(Z), it follows that
det(A U) - det(UT AU) - det(UT) det(A) det(U) - det(A)

for all A E Mn(Z), and so the group action preserves determinants. Also, if A is
symmetric, then A U is also symmetric. Thus, for any integer d, the group action
partitions the set of symmetric n x n matrices of determinant d into equivalence
classes.
To every n x n symmetric matrix A - (a,.j) we associate the quadratic form FA
defined by
nn
FA(xl,...,xn)-EEai
nn
i-I j-1
This is a homogeneous function of degree two in the n variables x1, ..., x,,. For
example, if 1n is the n x n identity matrix, then the associated quadratic form is
Ft,(x1,...,xn) -x, +xZ +..-+X2 .

Let x denote the n x I matrix (or column vector)
X1
xxn
We can write the quadratic form in matrix notation as follows:
FA(xI,...,xn)-xTAx.
The discriminant of the quadratic form FA is the determinant of the matrix A. Let
A and B be n x n symmetric matrices, and let FA and FB be their corresponding
quadratic forms. We say that these forms are equivalent, denoted
FA
FB,
1.3
Quadratic forms
B. Equivalence of quadratic forms is an

if the matrices are equivalent, that is, if A
equivalence relation, and equivalent quadratic forms have the same discriminant.
The quadratic f o r m F A represents the integer N if there exist integers x 3 . . .
. .
such that
FA(xj....,
N.
FB, then A
B and there exists a matrix U E
A - B U - UT BU. It follows that
If FA
such that
FA(x) - XT Ax - XT UT BUx - (Ux)T B(Ux) - FB(Ux).

Thus, if the quadratic form FA represents the integer N, then every form equivalent
to FA also represents N. Since equivalence of quadratic forms is an equivalence
relation, it follows that any two quadratic forms in the same equivalence class
represent exactly the same set of integers. Lagrange's theorem implies that, for
n > 4, any form equivalent to the form x2 +
+ xR represents all nonnegative
integers.
I for all
FA(x1, ... ,
f (0, ... , 0). Every form equivalent to a positive-definite quadratic
The quadratic form F A is called p o s i t i v e - d e f i n i t e if
(x1.... ,
form is positive-definite.
A quadratic form in two variables is called a binary quadratic form. A quadratic
form in three variables is called a ternary quadratic form. For binary and ternary
quadratic forms, we shall prove that there is only one equivalence class of positivedefinite forms of discriminant 1. We begin with binary forms.
Lemma 1.1 Let

at2
a1.2
a2.2
be a 2 x 2 symmetric matrix, and let

FA(x1, x2) - a1.1x +2a1.2x1x2 +a2,2x2
be the associated quadratic form. The binary quadratic form FA is positive-definite

if and only if
a1.1>1
and the discriminant d satisfies
d - det(A) -a,, ja2.2
- ai.2 > 1.
Proof. If the form FA is positive-definite, then
FA(l,0)-aj_1 > 1
and
FA(-a1.2, a1.1) - al.lal.2 - 201 j0 .2 +ai.1a2.2

2
- a,.1 (aj.1a2.2 - a1.2)
- a1.1d > 1,
10
I.
Sums of polygons
and sod > 1. Conversely, if a,,1 > 1 and d > 1, then

aI., FA(xi, x2) -(a I. I xj +a3 2x2)2 +dx2 > 0,
and FA(xi, x2) - 0 if and only if (x1, x2) - (0, 0). This completes the proof.
Lemma 1.2 Every equivalence class of positive-definite binary quadratic forms

of discriminant d contains at least one form
FA(xI, x2) - at.Ix + 2a1 2xIx2 +a2.2x2
for which
21ai.21
a,., < .
Proof. Let FB(x1, x2) - bi,1 x + 2bi,2xlx2 +b2,2x2 be a positive-definite quadratic form, where
B- bl.i b1.2
b1.2
b2.2
is the 2 x 2 symmetric matrix associated with F. Let a,., be the smallest positive
integer represented by F. Then there exist integers rl, r2 such that
F(rt,r2)-a,.,.
If the positive integer h divides both r, and r2, then, by the homogeneity of the
form and the minimality of a,,3, we have
F(ri/h, rz/h) =
ai.i
F(r1. r2)
h2
a,.,
h2
and so h - 1. Therefore, (r1, r2) - 1 and there exist integers si and s2 such that
I - rIs2 - r2si - ry(s2 + r2t) - r2(si + rat)

for all integers t. Then
U
r,
r2
S1 +r31
s2 + r2t
E SL2(Z)
for all t E Z. Let
A - UTBU
F(r3, r2)
a',.2 + F(ri, r2)t
a1.i
a1.2
a3.2
a2.2
ai 2 + F(rj, r2)t
F(s, + ri t, s2 + r2t )
1.3
Quadratic forms
II
where
a1.2 - bI., rlsI + b1.2(r1s2 + r2s1) + b2,2r2s2
a1.2 - a1.2 +al.It

a2.2
- F(s1 +r1t, s2 +r2t) > al.l
since (s) + r)t, s2 + r2t) 71 (0, 0) for all t E Z, and al.l is the smallest positive
number represented by the form F. Since {a', 2 +al,lt : t c- Z} is a congruence
class modulo al,l, we can choose t so that
Ial.21 - lai.2+al.It I < a21
Then A -
B, and the form FB is equivalent to the form FA(x1, x2) - al,lx +
2a1.2XIx2 +a2,2x2, where

21a1.21 < al.l
a2.2.
If d is the discriminant of the form, then

2
d-al.la2.2-a1.2,
and the inequality
al.1 <a1.1a2.2-d+aj.2 <d+
2
a`41
implies that
3a 2
l.l
<d
or, equivalently,
al.l
?^
a.
This completes the proof.
Theorem 1.2 Every positive-definite binary quadratic form of discriminant 1 is

equivalent to the form xi + x2.
Proof. Let F be a positive-definite binary quadratic form of discriminant 1. By
Lemma 1.2, the form F is equivalent to a formal, Ix i +2a 1.2x) x2 +a2,2x2 for which
2
21al.21:s al,1<7<2.
Since al,) > 1, we must have al,) - 1. This implies that a1.2 - 0. Since the
discriminant is 1, we have
a2.2 - a1.1a2.2 - ai.2 " 1.
Thus, the form F is equivalent to x2 + x2. This completes the proof.
12
1.
Sums of polygons
1.4 Ternary quadratic forms

We shall now prove an analogous result for positive-definite ternary quadratic
forms.
Lemma 1.3 Let
A-
f al.l
a1.2
a1.3
a1.2
a2.2
a2,3
a1.3
a2.3
a3.3
be a 3 x 3 symmetric matrix, and let FA be the corresponding ternary quadratic

form. Let d be the discriminant of FA. Then
(1.3)
a1, FA(XI,X2,X3)-(a1.lx1
where GA is the binary quadratic form corresponding to the matrix
al.la2.2 - ai2
,
al.la2.3 - a1,2a1 .3
al.la2.3 - a1A
.2a1.3
(1.4)
a,,la3,3 -a 21,3
and GA. has discriminant al,ld. If FA is positive-definite, then GA. is Positivedefinite. Moreover, the form FA is positive-definite if and only if the following three
determinants are positive:

al.l - det(a1.1) > 1,
d' - det
a1,1
a1.2
a1.2
a2.2
and
d - det(A) > 1.
Proof. We obtain identities (1.3) and (1.4) as well as the discriminant of GA
by straightforward calculation.
If FA is positive-definite, then
FA(1,0,0)-al.l > 1.
If GA.(x2, x3) < 0 for some integers x2, x3, then
x3) < 0. Let x1 - -(a1.2x2 + a1.3x3). Then
a.1
l
al.)X3)
al,lxl +a1,2a1,1X2+a1.3a1,1x3 =0,
and so
al.I FA(xl, at.lx2, al.lx3)

- (611.1x1 + a1.2a1.1X2 +a1.3a1.1X3)2 +
al,Ix3)
2
- al.l
< 0.
x3)
a1.1x3)
1.4
Ternary quadratic forms
13
Since FA is positive-definite, it follows that x2 = X3 - 0, and so the binary form

GA is also positive-definite. By Lemma 1.1, the leading coefficient of GA. is
positive, that is,
z
d' al.la2.2-a1.2-1+
and also the discriminant of GA is positive, hence
d-det(A)> 1.
This proves that if FA is positive-definite, then the integers a,.,, d', and d are
positive.
Conversely, if these three numbers are positive, then Lemma 1.1 implies that
the binary form GA is positive-definite. If FA(xl, x2, x3) - 0, then it follows from
identity (1.3) that
x3) - 0
and
al.lxl +a1.2X2 +a1.3X3 - 0.
The first equation implies that x2 e x3 - 0, and the second equation implies that
x, - 0. Therefore, the form FA is positive-definite.
Lemma 1.4 Let B - (b, 1) be a 3 x 3 symmetric matrix such that the ternary
quadratic form FB is positive-definite. Let GB, be the unique positive-definite
binary quadratic form such that
bl.IFB(Yl, Y2, Y3) -' (b1.1 Y1 +b1.2Y2 +b1.3Y3)2
Y3).
For any matrix V' a (v . f) E SL2(Z), let
A*-(V*)TB.V`
(1.5)
and let G A. be the positive-definite binary quadratic form corresponding to the

symmetric matrix AS and equivalent to the form G H.. For any integers r and s, let
Vr.s - (v11) s
vi.1
v11.2
V2
E SL3(Z)
(1.6)
v2 2
and
(1.7)
Ar.s = VT
r.s BV,.s a (ai.j).
Let FA,, be the corresponding ternary quadratic form. Then a,., - b,

a,., FA,, (xI , X2, X3) -(a I. I xl +a 1.2X2 +a 13X3
)2
_1
and
+ GA (X2, X3),
where the matrix A* defined by (1.5) is independent of r and s.
14
1.
Sums of polygons
Proof. Since vl,l - I and V2.1 - V3,1 - 0, it follows from the matrix equation (1.7) that
3
al.j -
1 vl kbk,i Vi. j -
bl.i vi. j
Vk, I bk.i vi. j
k-I i-1
k-1 i-1
i-1
and soal,I -b1,1. Let

xl
x-
Yl
V,.sx - Y -
and
x2
Y2
x3
Y3
3
Yi -
E vi.jxj
j-1
Y2 - v2, IXl + v2,2X2 + V2,3x3 - Vj 1X2 + V* 1.2X3

y3 - v3, 1X1 + V3,2X2 + V3.3X3 - V2 1X2 + VZ 2X3.
Y.Y2
X.
and
Y3
-1
X}
3
Then
It follows that
Y3) -
X3)-
Moreover,
3
b1.1y1 + bl.2Y2 + b1.3Y3 -
i-I
bl.i E vi.jxj
j-1
(bl.1V.J) xj
j-1 i-1
- al,lx1 +a1.2x2 +a1,3x3
Since
FA,., (X I, x2, x3)
_XT A,,x - (Vr,.,X)T B(V,,:x) - YT BY - FB(Y1, Y2, Y,).
it follows that
(al.lxl +al,2x2 +a1.3x3)2 + GA...(x2, X3)
-a1,1FA,.,(XI,X2,X3)
- bl, l FA,., (XI, x2, x3)
- b1.I FB(yj, y2, y3)
- (bl.Iyl +b1.2Y2 +b13y3)2 + GB.(Y2, y3)

-(al, Ix, +al,2X2 +al.3X3)2 + GA' (X2, x3),
Ternary quadratic forms
1.4
15
and so
GA-(x2, x3) - GA;, (x2, X3)
for all integers r and s. This completes the proof.

Lemma 1.5 Let u 1.1, u2,1, and u3,1 be integers such that
(u I.1,U2.1,U3.1)-1.
Then there exist six integers u1 1 for i - 1, 2, 3 and j - 2, 3 such that the matrix
U - (u;.j) E SL3(Z), that is, det(U) - 1.
Proof. Let (u 1. 1, u2.1) - a. Choose integers u 1.2 and U2.2 such that
U1.1U2.2 - U2.1U1 1,2 - a.
Since (a. u3,1) - (u 1.1 , U2.1, U3. 1) - 1, we can choose integers U3.3 and b such that
au3.3 - bu3.1 - 1.
Let
U1.3 -
U3.2 - 0.
Then the matrix
U - (ui.j) -
U1.1
U 1..2
U21
u2.2
U3.1
U3.3
has integer coefficients and determinant 1. This completes the proof.
Lemma 1.6 Every equivalence class of positive-definite ternary quadratic forms

of discriminant d contains at least one form _3J-1 ai,1xixj for which
2max (1al.zl,1a1 .31)
a1,1 < 3 ' d.
Proof. Let F be a positive-definite ternary quadratic form of determinant d, and

let C be the corresponding 3 x 3 symmetric matrix. Let al_ 1 be the smallest positive
integer represented by F. Then there exist integers u 1, 1, u2,1, and 143,1 such that
F(u1.1.U2.1,u3.1)-al.1.
If (u 1.1 ,142.1,143.1) - It, then the form F also represents a,. I 1h 2, and so, by the
minimality ofa,,1,wehave (ul,l, U2. 1, u3.1) - 1. By Lemma 1.5,thereexistintegers
ui,j for i - 1, 2, 3 and j - 2, 3 such that the matrix U - (ui.j) E SL3(Z). Let
B - UTCU - (bi.j).
16
1.
Sums of polygons
Then F is equivalent to the form FB, and
bl.l - a,,1
is also the smallest integer represented by FB. By Lemma 1.3,
aI.1 FB (xI , x2, x3) - (bl.lxl + b1.2x2 + b1.3x3)2 + GB. (X2, x3),
where G8. (x2, x3) is a positive-definite binary quadratic form of determinant a 1, Id.
By Lemma 1.2, the form
x3) is equivalent to a binary form
2
GM(x2, x3) - a1.1x2 +a1.2x2x3 +a2.2x3

*
such that
ai.l <
al.ld.
Choose V' E SL2(Z) such that At - (V')T B` V'. Let r, s E Z, and let Vr.s E
SL3(Z) be the matrix defined by (1.6) in Lemma 1.4. Let
A - VT BVT,s - (a,j).
(1.8)
Note that the integer in the upper left corner of the matrix is still a1,I, the smallest
positive integer represented by any form in the equivalence class of F, and that,
by Lemma 1.3,
a,*,, - al.1a2.2 - ai.2.
Finally, it follows from (1.8) that
al.2 - a1.1r + bl.201.1 + b1.3vs.1
and
al.3 - a1.ls + bl.2vi.2 + b1.3vz.2
Therefore, we can choose r such that

2
1a1.21 5 all
and choose s such that
a1I
z
Ial.31 _<
Since
a1.1
FA(0, 1,0)-a2.2,
we have
a i.I - al.1a2.2
z
-al.la2.2-al.2+a1.2
a
' a1.1 +a1.2

2
a1.ld + a41.
1.5
Sums of three squares
17
This implies that
(2
\f/
ai.l
al. id
or, equivalently,
al.l <
3'd.
Theorem 13 Every positive-definite ternary quadratic form of discriminant I is

equivalent to the form x + x2 + x3.
Proof. Let F be a positive-definite ternary quadratic form of discriminant 1. By
Lemma 1.6, the form F is equivalent to a form FA s E a;.1x; x1 for which
0 < 2max (Ia1.21, Ial.31) < al j <
This implies that a1.2 - a13 - 0. Since d f 0, it follows that al., f 0 and so
al. I - 1. Therefore,
1
where the 2 x 2 matrix
0
0
a2.2
a2.3
a2.3
a3.3
a2,2
a2.3
a2.3
a3.3
has determinant 1. By Theorem 1.2, there exists a matrix
U.
142.2
u23
142.3
143.3
E SL2(Z)
such that (U')T A'U' is the 2 x 2 identity matrix 12. Let

1
U=
0
0
u2.2
U23
142.3
143.3
Then U7A U is the 3 x 3 identity matrix 13. This completes the proof.
1.5
In this section, we determine the integers that can be written as the sum of three
squares. The proof uses the fact that a number is the sum of three squares if
and only if it can be represented by some positive-definite ternary quadratic form
of discriminant 1, together with two important theorems of elementary number
18
1.
Sums of polygons
theory: Gauss's law of quadratic reciprocity and Dirichlet's theorem on primes in

arithmetic progressions.
The statement that a is a quadratic residue tnodulo in means that there exist
integers x and y such that x2 - a - ym. If p is prime and (a, p) = 1, then the
Legendre symbol (P) is defined by (p) - 1 if a is a quadratic residue modulo p
and 01) - -1 if a is not a quadratic residue modulo p. By quadratic reciprocity.
if p and q are distinct odd primes, then (e) - () if p - 1

(mod 4), and (Q)
if p - 1
(mod 4) or q - 1
(P) if p =- q =- 3 (mod 4). Also,
1 if and only
(mod 4), and (2) - 1 if and only if p - 1 or 7 (mod 8).
Lemma 1.7 Let n > 2. If there exists a positive integer d' such that -d' is a
quadratic residue modulo d'n - 1, then n can be represented as the sum of three
squares.
Proof. If -d' is a quadratic residue modulo d'n - 1, then there exist integers
a1.2 and a,., such that
a2 + d' - a,.1(d'n - 1) - al.la2.2,

where
a2,2-d'n-1>2d'-1> 1
and so
a,., > 1.
Equivalently,
d' - al.1a2.2 - a1.2.
The symmetric matrix
A-
a,.,
a1,2
a1,2
a2.2
has determinant
det(A) - (a,. 1a2,2 - a2,,2)n - a2,2 - d'n - a2.2 - 1.

By Lemma 1.3, the quadratic form FA corresponding to the matrix A is positive.
Moreover, FA has discriminant 1 and represents n, since F., (0.0. 1) - it. By

Theorem 1.3, the form x2 + x2 + x3 must also represent n. This completes the
proof.
Lemma 1.8 If n is a positive integer and n

represented as the sum of three squares.
(mod 4), then n can he
1.5
19
Proof. Since (4n, n - 1) - 1, it follows from Dirichlet's theorem that the arithmetic progression {4nj + n - I : j - 1, 2, ...} contains infinitely many primes.
Choose j > I such that
p-4nj+n- I -(4j+1)n- I
is prime. Let d' - 4j + 1. Since n - 2 (mod 4), we have
p - d'n - 1 - 1 (mod 4).

By Lemma 1.7, it suffices to prove that -d' is a quadratic residue modulo p. Let
d'
-flqik,
y, Id'
where the q; are the distinct primes dividing d'. Then
p-d'n-1--1 (mod q;)

for all i, and
d'
fl
(-1)k'
(mod 4).
9, 3 imo4 4)
Therefore,
9i W,
p,.J Imal 4I
By quadratic reciprocity we have

1
since p - 1
(mod 4), and
d')
(-1) (d')
p, aJ (mal )
Ok
20
1.
Sums of polygons
Lemma 1.9 If n is a positive integer such that n = 1, 3, or 5
(mod 8), then it
can be represented as the sum of three squares.
Proof. Clearly, I is a sum of three nonnegative squares. Let n > 2. Let
c_
If n - 1 or 3
ifn =- 1
ifn = 3 (mod 8)
ifn - 5
(mod 8)
(mod 8).
(mod 8), then
cn - 1
2
=1
(mod 4).
If n - 5 (mod 8), then

cn - 1
2
In all three cases,
(4n.
= 3 (mod 4).
cn-1)2
1.
By Dirichlet's theorem, there exists a prime number p of the form
p-4nj+
cn -
for some positive integer j. Let
d'=8j+c.
Then
2p=(8j+c)n-I-d'n-I.
By Lemma 1.7, it suffices to prove that -d' is a quadratic residue modulo 2p.
If -d' is a quadratic residue modulo p, then there exists an integer xo such that
(xo + p)2 + d' _= xo + d' _- 0
(mod p).
Let x = xo if xo is odd, and let x - xo + p if xo is even. Then x is odd and x` +

is even. Since
x2 + d' - 0 (mod 2)
and
x 2 + d' = 0
(mod p),
it follows that
x2 + d' - 0 (mod 2p).

Therefore, it suffices to prove that -d' is a quadratic residue modulo p.
1.5
21
Let
Id,
v,
be the factorization of the odd integer d' into a product of powers of distinct odd
primes q; . Since
2p - -1
(mod d'),
2p - -1
(mod q;)
it follows that
and
(p,q,)-1
for every prime qj that divides d'.
If n - 1 or 3 (mod 8), then p = 1
(mod 4) and
CP)CPI/CPI/
PI )
rJ Qj
k;
q, Id'
k
qF1
,ld
If n =_ 5
(P
(mod 8), then p - 3 (mod 4) and d' - 3 (mod 8). From the
factorization of d', we obtain
d, -
k.
k,
Q,
V,V
Y,.I
q, .3 (.441
Inod 4)
fj
Q;
(-1)k
Y Ids
y, r3
I,n.d 4)
-1
(mod 4)
and so
Y; Is'
q,.3 (m j4)
It follows from quadratic reciprocity that
C P/)(PI)CP
--(P)
(mod 4)
22
1.
Sums of polygons
()k
\P11
j
q,
q,.3
(0803 4)
(mod 4)
()k
n
q,.3 (mod 4)
(noJ 4)
4, .1
9,.I
k,
n
qi.3
(n).1 4)
9i 3 (mid 4)
)moJ 4)
k,
qj Id'
(P)
In both cases,
(-1)k..
F1
(--AR)
91.5.7"'d,
Therefore, -d' is a quadratic residue modulo 2p - d'n - 1 if
E ki - 0
(mod 2).
9, Id'
9i .57 In ,i R)
This is what we shall prove. We have

d' -
qi I"
q,.1 (mc, x)
II
qk
q' Id'
(mod x)
V, Id'
(0801x)
3k,
9i Id'
05 (mod R)
H
41.3.5 (mN x)
3k'
F1
9",
41.5.7
11
fl
qi.7 (mud 8)
(-1)k'
9, Wr
q,.7 ingd x)
(-1)ki
qI
4iW
94.5
(-3)k'
11
qk,
qk
9,.3
)nnd 8)
9, 4'
fl
(mod 8).
(..d8)
If n - i or 5 (mod 8), then c - 3 and
d'-8j+3-3 (mod 8).
(mod 8)
1.5
This implies that
k,-1
(mod2)
E k; - 0
(mod 2).
23
V, .5,5 (mod $
and
Wr
q .5.7 Imgl 81
If n
(mod 8), then c - I and
d' - 8i + 1
It follows that
E k; - 0
(mod 8).
(mod 2)
a,.5.5 rood a(
and
E k; - 0
(mod 2).
Of4'
q,.5.7 (mad 8)
This completes the proof

Theorem 1.4 (Gauss) A positive integer N can be represented as the sum of three
squares if and only if N is not of the form
N - 4(8k + 7).
Proof. Since
x2 = 0, 1, or 4 (mod 8)
for every integer x, it follows that a sum of three squares can never be congruent to
7 modulo 8. If the integer 4m is the sum of three squares, then there exist integers
X1, x2, x3 such that
4m - xi +x2 +x3.
This is possible only if x1, x2, x3 are all even, and so
m=(2)2+\212+`212.
Therefore, 4m is the sum of three squares if and only if m is the sum of three
squares. This proves that no integer of the form 4(8k + 7) can be the sum of three
squares.
Every positive integer N can be written uniquely in the form N - 4m, where
in 2 (mod 4) or m = 1, 3, 5, or 7 (mod 8). By Lemma 1.8 and Lemma 1.9,
the positive integer N is the sum of three squares unless m =_ 7 (mod 8). This
completes the proof.
Theorem 1.5 If N is a positive integer such that N - 3 (mod 8), then N is the
sum of three odd squares.
24
1.
Sums of polygons
Proof. Recall that x2 - 0, 1, or 4 (mod 8) for every integer x. If N =- 3

(mod 8) is a sum of three squares, then each of the squares must be congruent to
I modulo 8, and so each of the squares must be odd. This completes the proof.
1.6
Thin sets of squares
If A is a finite set of nonnegative integers such that every integer from 0 to N can
be written as the sum of h elements of A, with repetitions allowed, then A is called
a basis of order h for N. A simple counting argument shows that if A is a basis of
order h for N, then A cannot be too small.
Theorem 1.6 Let h > 2. There exists a positive constant c - c(h) such that, if A
is a basis of order h for N, then
JAI > cN'1h
Proof. Let Al I- k. If A is a basis of order h for N, then each of the integers

0, 1, ... , N is a sum of h elements of A, with repetitions allowed. The number of
combinations of h elements, with repetitions allowed, of a set of cardinality k is
the binomial coefficient (k'hTherefore,
N+1 <
k+h-1
h
)-
k(k+1). .(k+h-l)
h!
c'kh
- h!
for some constant c' > 0 and all k, and so
IAI -k>
(\ h!N
cN /
-cN'1h.

Since the squares form a basis of order 4, it follows that for every N > 0 the set
QN of all squares up to N is a basis of order 4 for N. Moreover,
IQNI-1+[N'/2] > N'/2.

This is much larger than cN1/4, which is a lower bound for the thinnest possible
basis of order 4. It is natural to ask if for every N there exists a set AN of squares
that is a basis of order 4 for N and satisfies
AN
lim
N-+,c NU2
_0.
The answer is provided by the following theorem.
Theorem 1.7 (Choi-Erd6s-Nathanson) For every N > 2, there exists a set AN

of squares such that AN is a basis of order 4 for N and
JANI
< (l0 2) N'"3 log N.

g
1.6
Thin sets of squares
25
Proof. The sets A2 = A3 = {0, 1) and A4 = A5 = 10, 1,4) satisfy the

requirements of the theorem. Therefore, we can assume that N > 6.
We begin with a simple remark. By Theorem 1.4, if e is a nonnegative integer
and f - I or 2 (mod 4), then f is the sum of three squares. Since the square of
an even integer is 0 (mod 4) and the square of an odd integer is I (mod 4), it
follows that if m 0- 0 (mod 4) and a is any positive integer such that a2 < m,
then either m - a2 is the sum of three squares or m - (a - 1)2 is the sum of three
squares.
For N > 6, we let AN consist of the squares of all nonnegative integers up to
2N". Then
IA(v)l < 2N"I3 + 1.
Let AN) consist of the squares of all integers of the form

[k1/2N1/31
or
[k1/2N1/3]
- 1,
where
4<k<N113.
Then
IA(2), < 2(N"3 - 3) - 2N 1'3 -6.

N
Let
(1)
(2)
AN
AM - AN U AN
Then
IAN)i <4N"3.
Since AN) contains all the squares up to 4N213, it follows from Lagrange's theorem
that every nonnegative integer up to 4N213 is the sum of four squares belonging to
AN
A(O)
Let m be an integer such that
4N2/3 < m < N

and
m # 0 (mod 4).
We shall prove that there exists an integer ao E AN) such that
0<m-a2<4N213
and m - ao is the sum of three squares. Since
4<
N213
<
N1/3,
it follows that
4<krN2/3I<N113
26
1.
Sums of polygons
Let
a - [k1/2N 1/31 .
Then a2 E AN1, (a - 1)2 E
121
a2 < kN2/3 < m < (k +
1)N2/3,
and
a > k1/2N1/3 - 1.
It follows from our initial remark that either m - a2 or m - (a - 1)2 is the sum
of three squares. Choose ao E ((a - 1)2, a2) c AN) such that m - ao is a sum of
three squares. Since 4 < 3N116 for N > 6, we have
0<m-a2
<m -ao
< m-(a-1)2
< (k + 1)N2/3 - (k'/2 N 113 - 2)2
< (k + 1)N2/3 - kN2/3 + 4k1/2N1/3
- N2/3+4k1/2N1/3
< N2/3+4N1/2
< 4N2'3,
and so m -ao is the sum of three squares belonging to A 1. Therefore, if 0 < m < N
N
and m # 0 (mod 4), then m is the sum of four squares belonging to A(,'.
Let
- - log N4
and aEAN)}.
ll
JJJJJJ
Then A N is a set of squares and
IANI < (
11
g4 +I)
(2loNIog4
N 1j
<
) 4N1/3 = (1 4 ) N113logN.
2
Let n E [0, N}. If n # 0 (mod 4), then n is the sum of four squares belonging
to AN) C AN. If n
0 (mod 4), then n - 4'm, where m 0 0 (mod 4) and
0 < i < log N/ log 4. Then
m-aj+a2+aj+a2,
where a,, a2, a3, a4 E A N1, and so
n - 4'm - (2'a1)2 + (2'a2)2 + (2'a3)2 + (2'a4)2
is a sum of four squares belonging to AN. This completes the proof.
1.7
1.7
The polygonal number theorem
27
We begin by proving Gauss's theorem that the triangles form a basis of order three.
Equivalently, as Gauss wrote in his journal on July 10, 1796,
num -A+A+A.
ETPHKA!
Theorem 1.8 (Gauss) Every nonnegative integer is the sum of three triangles.
Proof. The triangular numbers are integers of the form k(k + 1)/2. Let N > 1.
By Theorem 1.5, the integer 8N + 3 is the sum of three odd squares, and so there
exist nonnegative integers k,, k2, k3 such that
8N + 3 - (2k, + 1)2 + (2k2 + 1)2 + (2k3 + 1)2
- 4(ki +k, +k2 +k2+k3 +k3)+3.

Therefore,
N -
ki(k, + 1)
k2(k2 + 1)
k3(k3 + 1)

Lagrange's theorem (Theorem 1.1) is the polygonal number theorem for squares,
and Gauss's theorem is the polygonal number theorem for triangles. We shall now
prove the theorem for polygonal numbers of order m + 2 for all m > 3. It is easy
to check the polygonal number theorem for small values of N/m. Recall that the
kth polygonal number of order m + 2 is
pm (k) -
mk(k - 1)
2
+ k.
The first six polygonal numbers are
(0) - 0
P(1) = 1
p
p (4) - 6m + 4
pm(5)- 10m+5.
If k, , ..., k, are positive integers, then, for r - 0, 1, ..., m + 2 - s, the numbers
of the form
(1)
(1.9)
are an interval of m + 3 - s consecutive integers, each of which is a sum of exactly

m + 2 polygonal numbers. Here is a short table of representations of integers as
sums of m + 2 polygonal numbers of order m + 2. The first column expresses the
28
1.
Sums of polygons
integer as a sum of polygonal numbers in the form (1.9), and the next two columns
give the smallest and largest integers that the expression represents.
m+2
m+2
2m+3
2m+4
3m +4
3m+3
4m+4
4m +5
5m +5
4m+8
5m+6
5m+7
6m+4
6m+4
7m+5
7m+6
7m+8
8m+8
9m +7
8m+6
8m+7
9m+7
10m +7
pm(5)+rpm(1)
10m +5
l lm +6
pm(5)+pm(2)+rpm(1)
Ilm+7 12m+7
rpm(1)
pm(2)+rpm(1)
2pm(2)+rpm(])
pm(3)+rpm(1)
pm(3)+ pm(2)+rpm(I)
4pm(2)+rpm(1)
pm(3)+2pm(2)+rp,,,(1)
pm(4)+rpm(1)
pm(4)+ pm(2)+rpm(1)
2pm(3)+Pm(2)
pm(4)+2pm(2)+rpm(1)
pm(4)+ pm(3)+rpr(1)
This table gives explicit polygonal number representations for all integers up to
12m + 7. It is not difficult to extend this computation. Pepin [95) and Dickson [23]
published tables of representations of N as a sum of m + 2 polygonal numbers
of order m + 2 for all m > 3 and N < 120m. Therefore, it suffices to prove the
polygonal number theorem for N > 120m.
We need the following lemmas.
Lemma 1.10 Let m > 3 and N > 2m. Let L denote the length of the interval
!- [
2+
6N
-3, 3+
8N
-8
Then
L>4
if N > 108m
and
L>fm if1>3andN>7f2m3.
Proof. This is a straightforward computation. Let
x-N/m>2
and
eoaf-6
We see that
if and only if
L- 8x-8- 6x-3+6>f
8x-8> 6.x-3+to,
1.7
29
or, after squaring both sides and rearranging,
2x-120 -5>2Po 6x-3.

Squaring and rearranging again, we obtain
4x (x - (7Po + 5)) + (Pp + 5)z + 12Po > 0.

This inequality certainly holds if
\z
x>7Po+5-7(P-6)
Therefore,
L>P
+5.
N>7(P-6)
if
+5.
Since
7(4-6)
+5-107.86...,
it follows that L > 4 if N > 108m. Since

>7(t-6)2
+5
7P2
for t > 3, it follows that L > P if f > 3 and N/m > 7Pz. Therefore, if t > 3 and
N > 7Pzm3, then L > Pm. This completes the proof.
Lemma 1.11 Let m > 3 and N > 2m. Let a, b, and r be nonnegative integers
s uch tha t
0<r <m
and
N - 2(a-b)+b+r.
(1.10)
Consider the open interval
I-(I+ V6N-3, 3+ V8N-8).

If
bEI,
then
b2
<4a
1.11)
and
3a <bz+2b+4.
(1.12)
30
1.
Sums of polygons
Proof. From equation (1.10), we have
a-I1-?m /
Ib+2(N-r
\\
By the quadratic formula,
)b
b2- 4a- b2 -4I
1- m/ -8
\
-r<0
)
if
/
0<b <21 1-
21 2
?l+
m
41 1-
J
m
+8
( N-r
\m
/J
Ifb E 1, then
0<
<3 + 8N-8
<2 (1
- 2) + r8(
m
<21
-r
m
/
1
?m ) +
1-
41
\\
21
mJ
+8(N-r
This proves (1.11).

Again by the quadratic formula,
(6(N_r)
b2+2b+4-3a-b2-(1- )b_
4)>0
\\
if
b>
(-)+/(-- 3)Z+6
2
(N-r)-4.
\\
If b E I, then
b>
6N
J +--4
312
M
m)2
+6(Nm_r)-4.
This proves inequality (1.12).

The following result is sometimes called Cauchy's lemma.
1.7
31
Lemma 1.12 Let a and b be odd positive integers such that
b2 < 4a
and
3a < b2 +2b+4.
Then there exist nonnegative integers s, t, u, v such that
a-s2+t2+U2+U2
(1.13)
b-s+t+u+v.
(1.14)
and
Proof. Since a and b are odd, it follows that 4a - b2 = 3
(mod 8). By
Theorem 1.5, there exist odd positive integers x > y > z such that
4a -b2 -x2+y2+z2.
We canchoose the sign of z so that b + x + y z = 0
s, t, u, u as follows:
s=
t-
(mod 4). Define integers
b+x +y z
4
b+x
b+x - y:F z
-s=
2
4
b - x+y:F z
b+y
2 -s=
4
bz
b-x-yz
-S
These numbers satisfy equations (1.13) and (1.14) and
s>t>u>v.
We must show that v > 0. By Exercise 8, the maximum value of x + y + z subject
to the constraint x2 + y2 + z2 - 4a - b2 is 12 --3b2. Also, the inequality
3a < b2 + 2b + 4 implies that 12a - 3b2 
b-x-y-z
4
>
-l.
Since v is an integer, we must have v > 0. This completes the proof.

The following result is a strong form of Cauchy's polygonal number theorem.
Theorem 1.9 (Cauchy) If m > 4 and N > 108m, then N can be written as the
sum of m + I polygonal numbers of order m + 2, at most four of which are different
from 0 or 1. If N > 324, then N can be written as the sum of five pentagonal
numbers, at least one of which is 0 or 1.
32
1.
Sums of polygons
Proof. By Lemma 1.10, the length of the interval
I2+ 6N-3,
3+ 8m -8
is greater than 4 since N > 108m, and so I contains four consecutive integers
and, consequently, two consecutive odd numbers b, and b2. If m > 4, the set of
numbers of the form b+r, where b e {b,, b2} and r E {0, 1, ... , m - 3}, contains
a complete set of representatives of the congruence classes modulo m, and so we
can choose b E {b,, b2} c I and r E 10, 1, ... , m - 31 such that
N - b + r (mod m).
Then
a-2rN-b-rl+b-II-
2)b+2I
m
\ N-r1
in
(1.15)
is an odd positive integer, and
N - m2(a-b)+b+r.
By Lemma 1.11, since b E 1, we have
b2 < 4a
and
3a < b2 + 2b + 4.
By Lemma 1.12, there exist nonnegative integers s, t, u, v such that
a - s 2 + t2
+ U2 + V 2
and
b-s+t+u+v.
Therefore,
N - M2(a-b)+b+r
- 2 (s2-s+t2-t+u2-u+v2-v)+(s+t+u+v)+r
= Pm(S)+ Pm(t) + Pm(U)+ pm(v)+r.
Since 0 < r < m - 3 and since 0 and I are polygonal numbers of order m + 2 for
every m, we obtain Cauchy's theorem form > 4, that is, for polygonal numbers of
order at least six. To obtain the result for pentagonal numbers, that is, for m - 3,
we consider numbers of the form b, + r and b2 + r, where b, , b2 are consecutive
odd integers in the interval 1, and r - 0 or 1.
1.8
Notes
33
Theorem 1.10 (Legendre) Let m > 3 and N > 28m3. If M is odd, then N is the
sum of four polygonal numbers of order m + 2. If m is even, then N is the sum of
five polygonal numbers of order m + 2, at least one of which is 0 or 1.
Proof. By Lemma 1.10, the length of the interval I is greater than 2m, so I
contains m consecutive odd numbers. If m is odd, these form a complete set of
representatives of the congruence classes modulo m, so N = b (mod m) for
some odd integer b E 1. Let r - 0 and define a by formula (1.15). Then
m
Na 2(a-b)+b,
and it follows from Lemma 1.11 and Lemma 1.12 that N is the sum of four
polygonal numbers of order m + 2.
If m is even and N is odd, then N = b (mod m) for some odd integer b E I

and N is the sum of four polygonal numbers of order m + 2. If m is even and N is
even, then N - 1 - b (mod m) for some odd integer b E 1 and N is the sum of
a 1. This completes
five polygonal numbers of order m + 2, one of which is
the proof.
A set of integers is called an asymptotic basis of order h if every sufficiently
large integer can be written as the sum of h not necessarily distinct elements of
the set. Legendre's theorem shows that if m > 3 and m is odd, then the polygonal
numbers of order m + 2 form an asymptotic basis of order 4, and if m > 4 and m
is even, then the polygonal numbers of order m + 2 form an asymptotic basis of
order 5.
1.8
Notes
Polygonal numbers go back at least as far as Pythagoras. They are discussed at

length by Diophantus in his book Arithmetica and in a separate essay On polygonal
numbers. An excellent reference is Diophantus of Alexandria: A Study in the

History of Greek Algebra, by T. L. Heath [53]. Dickson's History of the Theory of
Numbers [22, Vol. II, Ch.] ] provides a detailed history of polygonal numbers and
sums of squares.
There are many different proofs of Lagrange's theorem that every nonnegative
integer is the sum of four squares. For a proof using the geometry of numbers, see
Nathanson [93]. There is a vast literature concerned with the number of representations of an integer as the sum of s squares. Extensive treatments of these matters can
be found in the monographs of Grosswald [43], Knopp [74], and Rademacher [98].
Liouville discovered an important and powerful elementary method that produces
many of the same results (see Dickson [22, Vol. II, Ch. 11 ] or Uspensky and
Heaslet [122]).
Legendre and Gauss determined the numbers that can be represented as the sum
of three squares. See Dickson [22, Vol. 11] for historical references. In this chapter,
I followed the beautiful exposition of Landau [78]. There is also a nice proof by
34
1.
Sums of polygons
Weil [140] that every positive integer congruent to 3 (mod 8) is the sung of three
odd squares.
Cauchy [9] published the first proof of the polygonal number theorem. Legendre's theorem that the polygonal numbers of order m form an asymptotic basis of
order 4 or 5 appears in [80, Vol. 2, pp. 331 -356[. In this chapter 1 gave a simple
proof of Nathanson [91, 92], which is based on Pepin 195].
Theorem 1.7 is due to Choi, Erdo"s, and Nathanson [ 131. Using a probabilistic
result of Erdos and Nathanson [36], Zollner 11521 has proved the existence of a
basis of order 4 for N consisting of < N11411 squares. It is not known if the r can
be removed from this inequality. Nathanson [89], Spencer 1118], Wirsing 11451.
and Zollner [151] proved the existence of "thin" subsets of the squares that are
bases of order 4 for the set of all nonneaatitie integers.
1.9
Exercises
1. Let m > 2. Show that the polygonal numbers of order m +2 call be '.riucy)
in terms of the triangular numbers as follows:
p,,,(k) - mpt(k)+k
for all k > 0.
2. (Nicomachus, 100 A.D.) Prove that the sum of two consecutive triangular
numbers is a square. Prove that the sum of the nth square and the (n - l)-tit
triangular number is the nth pentagonal number.
3. Let v(2) be the smallest number such that every integer N can he written in
the form
N - fxi f...fx?(2).
Prove that v(2) - 3. This is called the easier Waring's problem for.Nrtuares.
Hint: Use the identities
2x+1-(x+1)2-x2
and
2x-(x+ 1)2-x2- 12.

4. Prove that if m is the sum of two squares and n is the sum of two squares.
then mn is the sum of two squares. Hint: Use the polynomial identity
(x1 +x2)(y1 +yi) - (xtyt +x2y2)2 +(xty2 - x2yt)2.

5. (Nathanson [881) Prove that there does not exist a. polynomial identity of the
form
(x; +x2+x3)(y1 +yi+y3)-z; +z2+z3,

where Zr, Z2, Z3 are polynomials in x1, x2, x3, yl, y2, y3 with integral coefficients.
1.9
Exercises
35
6. Prove that Theorem 1.4 implies Lagrange's theorem (Theorem 1.1).

7. Prove that the set of triangular numbers is not a basis of order 2.
8. Let S2 - ((x, y, z) E R3 : x2 + y2 + z2 - 1). Prove that
{x + y + z : (x, y, z) E S2} - [-v, '].

9. Let
FA(XI,...,Xn)- Eai.jxixj
i, j-I
and
FB(x1,...,xn)- Ebi.jxixj
i. j-1
be quadratic forms in n variables such that

FA(X],...,xn) - Fa(xl,...,Xn)
for all x1,...,x E Z. Prove that ai,j - b,j foralli.j - 1,...,n.

10. Let A be an n x n symmetric matrix, and let FA be the corresponding
quadratic form. Let
U - (ui.j)
and
B - UT AU - (bi.i)
Prove that
bj.j
- FA(UI.j, u2.j, ... , Un.j)
for j - 1,...,n.
11. ForN> 1,let k-[']and

A ={0, 1,...,k- 1)U{k,2k,...,(k- 1)k}.
Show that A is a basis of order 2 for N such that
JAI <2i/N+1.
12. Let h > 2, k > 2, and
h-1
A-{0}UU{aiki :ai -1,...,k- 1}.

i-0
Prove that A is a basis of order h for kh - 1 and
JAI <h(k-1)+1.
36
1.
Sums of polygons
13. (Raikov [99], StOhr [119]) Let h > 2 and N > 2h. Let A be the set
constructed in the preceding exercise with
k - [NI/h] + 1.
Prove that A is a basis of order h for N such that
IAA < hN'Ih + 1.
2
Omnis integer numerus vel est cubus; vel e duobus, tribus, 4,5,6,7,8,
vel novem cubus compositus: est etiam quadratoquadratus; vel e duobus, tribus &c. usque ad novemdecim compositus &sic deinceps.'
E. Waring [138]
2.1
Sums of cubes
In his book Meditationes Algebraicae, published in 1770, Edward Waring stated

without proof that every nonnegative integer is the sum of four squares, nine cubes,
19 fourth powers, and so on. Waring's problem is to prove that, for every k > 2,
the set of nonnegative kth powers is a basis of finite order.
Waring's problem for cubes is to prove that every nonnegative integer is the
sum of a bounded number of nonnegative cubes. The least such number is denoted
g(3). Wieferich and Kempner proved that g(3) - 9, and so the cubes are a basis
of order nine. This is clearly best possible, since there are integers, such as 23 and
239, that cannot be written as sums of eight cubes.
Immediately after Wieferich published his theorem, Landau observed that, in
fact, only finitely many positive integers actually require nine cubes, that is, every
I Every positive integer is either a cube or the sum of 2,3,4,5,6,7,8, or 9 cubes; similarly,
every integer is either a fourth power, or the sum of 2, 3, ... , or 19 fourth powers; and so
on.
2.
38
sufficiently large integer is the sum of eight cubes. Indeed, 23 and 239 are the
only positive integers that cannot be written as sums of eight nonnegative cubes.
A set of integers is called an asymptotic basis of order It if every sufficiently large
integer can be written as the sum of exactly It elements of the set. Thus, Landau's
theorem states that the cubes are an asymptotic basis of order eight. l..ater, Linnik
proved that only finitely many integers require eight cubes, so every sufficiently
large integer is the sum of seven cubes, that is. the cubes are an asymptotic basis of
order seven. On the other hand, an examination of congruences modulo 9 shows
that there are infinitely many positive integers that cannot be written as sums of
three cubes.
Let G(3) denote the smallest integer h such that the cubes are an asymptotic
basis of order h, that is, such that every sufficiently large positive integer can be
written as the sum of h nonnegative cubes. Then
4<G(3)<7.
To determine the exact value of G(3) is a major unsolved problem of additive
number theory. It is known that almost all positive integers are sums of four cubes,
and it is possible that G(3) - 4.
The principal results of this chapter are the theorems of Wiefcrich-Kcmpner

and of Linnik. Because of the mystery surrounding sums of few cubes, we also
include a section about sums of two cubes. We shall prove that there are integers
with arbitrarily many representations as the sum of two nonnegative cubes, but
that almost all numbers that can be written in at least one way as the sum of two
nonnegative cubes have essentially only one such representation.
2.2
The proof that g(3) - 9 requires four lemmas.

Lemma 2.1 Let A and m be nonnegative integers such that m < A2 and m can
be written as the sum of three squares. Then
6A(A2 + m)
is a sum of six nonnegative cubes.
Proof. Let mt, m2, m3 be nonnegative integers such that
m-m2+m2+m3.
Then
0<mi
<A
for i - 1, 2, 3, and
3
((A + m; )3 + (A - M,)3).
6A(A2 + m) ' 6A(A2 + m2 +M2 + m3) i-1
2.2
39
Lemma 2.2 Let t > 1. For every odd integer w, there is an odd integer b such
that
w - b3 (mod 2).
Proof. If b is odd and w - b3
integers such that
(mod 2'), then w is odd. Let b1 and b2 be odd
bZ
(mod 2').
Then 2' divides

bz -b - (b2 - b,)(b2 + b2bt +b 2).
i
Since b2 + b2bt + b, is odd, it follows that 2' divides b2 - bl, that is,
b, - b2 (mod 2').
This means that if bi and b2 are odd integers such that
0<b1 <b2<2',
then
bi 0 b2
(mod 2'),
and so every odd integer is congruent to a cube modulo 2'. This completes the
proof.
Lemma 2.3 If
r > 10648 - 223
then there exists an integer d E [0, 22] and an integer m that is a sum of three
squares such that
r=d3+6m.
Proof. If the nonnegative integer m is not the sum of three squares, then there
exist nonnegative integers s and t such that
m - 4'(8t + 7),
and so
6m = 6.45 (8t + 7)
0
72
42
90
(mod 96)
(mod 96)
(mod 96)
(mod 96)
It follows that if m is a positive integer and
6m - h (mod 96)
ifs > 2
if s a 1
ifs = 0 and t is even
ifs = O and t is odd.
40
2.
for some
h E N - 16, 12, 18, 24, 30, 36, 48, 54, 60, 66, 78, 841,
then m is the sum of three squares. The following table lists, for various h E N
and
d E D= (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 17, 18,22),

the least nonnegative residue in the congruence class
d3 + h
(mod 96).
The elements of N are listed in the top row, and the elements of D are listed in the
column on the left.
6
12
18
36
48
54
60
66
78
84
12
18
24
24
30
30
36
13
19
25
31
37
54
55
60
48
49
78
79
14
26
32
63
68
87
86
51
56
75
62
45
38
57
44
33
20
39
76
84
85
92
82
47
88
53
94
16
9
46
52
59
65
77
22
83
66
67
74
93
34
95
11
17
4 70
35
42
41
6
7
73
91
80
10
50
69
58
11
13
14
15
17
29
18
0
40
22
72
21
90
43
2
27
64
23
71
81
61
28
89
15
10
Every congruence class modulo 96 appears in this table. Since 0 < d < 22 for
all d E D, it follows that if r > 223, then there exists an integer d E D such that
r - d3 is nonnegative and r - d3 - h (mod 96) for some h E W. Therefore,
r - d3 - 6m, where m is the sum of three squares. This completes the proof.
Lemma 2.4 If 1 < N < 40, 000, then

(i) N is a sum of nine nonnegative cubes;
(ii) if N f 23 or 239, then N is a sum of eight nonnegative cubes;
2.2
41
(iii) if N f 23 or 239 and if N is not one of the following fifteen numbers:

15
22
50
114
167
175
186
231
303
364
212
420
238
454
428
then N is a sum of seven nonnegative cubes;
(iv) if N > 8042, then N is a sum of six nonnegative cubes.

Proof. Lets (N) denote the least integer h such that N is the sum of h nonnegative
cubes. Von Sterneck computed s(N) for all N up to 40,000. The four statements in
the lemma are obtained by examining von Sterneck's list of values of s(N). Using a
computer, one can quickly verify (and extend) von Sterneck's list (see Exercise 8).
Theorem 2.1 (Wieferich-Kempner) Every nonnegative integer is the sum of nine

nonnegative cubes.
Proof. We shall first prove the theorem for integers
N > 810.
Let
n - [N1/31 .
Then
210 < n < 2.
8`'+1
There exists an integer k > 3 such that
8.8 31 < N < 8
g3(k+1).
Let
N; - N - P.
For i - 1.... , n we have

di - N,_, - Ni - i3 - (i - 1)3 - 3i2 - 3i + 1
<3i2<3N2/3 <
82F+3
Choose i so that
N,+, < 8 .83k < N,.

Then i > 1. Since k > 3, we have
N = N-n3
<(n+1)3-n3-1
- 3n2 + 3n
< 6n2
<3,
< 8.8 3k
821'+3
42
2.
Therefore, i < n - 1. It follows that
Ni < Ni_i - (N,_1 - N;)+(N1 - Nl)+N;+1

- di + di+i + Ni+t
< 3 .8*+3 + 8.83k
< I I .83k
Since N_1 - Ni - d, is odd, exactly one of the integers Ni and N,_1 is odd. Choose
a E {i - 1, i) such that Na - N - a3 is odd. By Lemma 2.2, there is an odd integer
b E [ 1, 8k - I] such that
N - a3 = b3
(mod 8k).
Then
7,83" -8.83k -83k < N -a3 -b3 < Na < 11 83k

and
N-a3-b3-8kq,
where
7 -82k < q < 11
Let
824 .
r-q-6.82k.
Then
223 <86<82k <r
<5.82".
It follows from Lemma 2.3 that r can be written in the form
r-d3+6m,
where 0 < d < 22 and m is a sum of three squares. Let
A - 8".
Then
zk
m<6<56 <A2.
Let
c - 2'd.
Then
N-a3+b3+8kq
-a3+b3+8k(6.82k +r)
- a3 + b3 +8 k (6.82' + d3 + 6M)
- a3 + b3 + (2kd)3 +8 k (6 - 82" + 6m)

- a3 + b3 + c3 + 6A(A2 + m).
2.2
43
By Lemma 2.1, 6A(A2 +m) is a sum of six nonnegative cubes, so N is the sum of
nine nonnegative cubes.
Now let
40,000<N< 810.
Then
a - [(N - 10,000)"] > 30, 000113
> 31,
so
d-(a+1)3-a3-3a2+3a+1
<4a2
<4N2/3.
Therefore,
N - (a + 1)3 < 10, 000 < N - a3 - N - (a + 1)3 +d < 10,000+4N 2/3.

If N -a3 < 40, 000, then N -a3 is a sum of six nonnegative cubes by Lemma 2.4.
If N - a3 > 40, 000, then we choose the integer
b - [(N
- a3 - 1O,000)] > 31,
and obtain
N-a3-(b+ 1)3 < 10,000<N-a3-b3 < 10, 000 + 4(N - a 3)2/3.

If N -a 3 - b3 < 40, 000, then N - a3 - b3 is a sum of six nonnegative cubes by
Lemma 2.4. If N - a3 - b3 > 40, 000, then we choose the integer
c - [(N - a3 - b3 - 10, 000)1/3] > 31
and obtain
N-a3-b3-(c+1)3
< 10, 000
< N - a3 - b3 - C3
< 10, 000 + 4 (N - a3 -
b3)213
< 10, 000 + 4 (10, 000 + 4 (10, 000 + 4N2/3)2/3 12/3

10,000+4(10,000+4(10,000+4(810)2/3)2/3) 2/3
<
//
< 20, 000.
Thus, if 40, 000 < N < 810, then there exist three nonnegative integers a, b, and
c such that
10,000< N-a3-b3-C3 <40,000.

By Lemma 2.4, N - a3 - b3 - C3 is the sum of six nonnegative cubes. This
44
2.
2.3
Linnik's theorem
Let G(3) denote the smallest integers such that every sufficiently large integer is
the sum of s nonnegative cubes.
Theorem 2.2 If N = 14 (mod 9), then N is not the sum ofthree integral cubes.
In particular,
G(3) > 4.
Proof. Since every integer, positive or negative, is congruent to 0, 1. or 1

modulo 9, it follows that every sum of three cubes belongs to one of the seven
congruence classes, 0, 1, 2, 3 (mod 9). Therefore, if N = !.4 (mod 9).
then N cannot be the sum of three cubes, so G(3) > 4.
Lemma 2.5 Let n be a positive integer. If there exist distinct primes p, q, r such
that
p-=q=-r=--l
(mod 6),
r<q<1.02r,
(2.2)
4pq <n<pq ,
3
18
18
(2.3)
(mod q6),
2n = p3q'8 (mod r6),
n = 3p (mod 6p),
4n
(2.1)
p3r18
(2.4)
(2.5)
(2.6)
then n is the sum of six positive integral cubes.
Proof. It follows from (2.2) and (2.3) that
p3(4g18+2r18) < 6p3g18

< 8n
< 8p3g18
< p3(4g18+4(1.02r)'8)
< p3(4g18 + 8r'8).
Thus,
p3(4g18+2r18) < 8n < p3(4g18+8r18).
(2.7)
Congruences (2.6), (2.4), and (2.5) imply that
8n = 2p3r18 = p3(4g18 + 2r'8) + 18pg6r6 (mod q6),

8n = 4p3g18 = p3(4gIII +2r18)+ 18pg6r6 (mod r6),
8n
0 = p3(4q 8 + 2r 18) + 18pg6r6 (mod p),
so
8n = p3(4q 18 + 2r18) + 18pg6r6
(mod pg6r6).
(2.8)
2.3
Linnik's theorem
45
It follows from (2.1) and (2.6) that
n=3p=-3-3 (mod6),
so
8n - 24 (mod 48).
(2.9)
By (2.1), the primes p, q, r are odd; hence
p2 = q2 = r2
== 1
(mod 8)
and
p3 (2,q18 + r 18) + 9 pg6r6 m (2 + I) p
+ P = 4P = 4 (mod 8).
Therefore,
p3(4q'8 + 2r'8) + 18 pg6r6 = 8
Similarly, since p = q = r - -1
(mod 16).
(mod 3), we have
p3(4q's + 2r18) + 18pg6r6 = 0 (mod 3)

so
p3(4q'8 + 2r'8) + 18pg6r6 = 24
(mod 48).
(2.10)
Since (pqr, 48) - 1, we can combine (2.8), (2.9), and (2.10) to obtain
8n = p3(4q'8 + 2r18) + 18 pg6r6
(mod 48pg6r6).
Therefore, there exists an integer u such that

8n - p3(4q 18 + 2r'8) + 18 pg6r6 + 48pg6r6u
p3(4q 18 + 2r 18) + 6pg6r6(8u + 3).
It follows from (2.7) that
0 < 6pg6r6(8u + 3) < 6p3r1s

so
0 < 8u + 3 <
By Theorem 1.5,
p2q-6r'2.
8u+3-x2+y2+ z2,
where x, y, z are odd positive integers less than pq-3r6, that is,
max{q3x, q3y, q3z} < pr 6.
(2.11)
Therefore,
8n - p3(4q 18 + 2r 18) + 6pg6r6(x2 + y2 + z2)
(pq 6 + r3x)3 + (Pq6 - r3x)3 + (pq 6 + r'Y)3

+(pq 6 r3Y)3 + (pr6 + g3Z)3 + (pr6 g3z)3.
2.
46
Since each of the six integers p, q, r, x, y, z is odd, it follows that each of the
six cubes in the preceding expression is even. Moreover, each of these cubes is
positive, since, by (2.2) and (2.11),
0 < r3x < q3x < pr6 < pq6,
0 < r3y < q3y < Pr6 < Pq6,
and
0 < q3z < pr6.

Therefore,
n=
(P6+r3x)3+(P6 -r3x\(pq6+r3y)2
(pq6-r3y)3+CPr6+qgz` 3+Cpr6-q3`` 3
2
J\I
is a sum of six positive cubes.

Theorem 2.3 (Linnik) Every sufficiently large integer is the sum of seven positive
cubes, that is,
G(3) < 7.
Proof. Let k and e be integers such that k > I and (k, e) = 1. We define the
Chebyshev function for the arithmetic progression e modulo k by
t9(x; k, e) _ E log p.
The Siegel-Walfisz theorem states that for any A > 0 and for all x > 1,
iO(x; k, e) _
Pk)
+ O C (log x)A I
(2.12)
/I
where (p(k) is the Euler p-function, and the implied constant depends only on A.
It follows that, for any S > 0,
0((1 +S)x;k, e) - O(x;k, e) _ 8X +0 C (log X)A

X
Let k - 6, e - -1, S = 1150, and x = (50/51)(log N )2. For any integer N > 2,
log P
,50;slNk{, )2 yc,b, ,2
y.-1
Una16)
0((logN)2;6, -1)
=
(log N)2
102
+O
- 0((50/51)(logN)2;6, -1)
(log N)2
log N)^
C (log
2.3
Linnik's theorem
47
Since
E logp<Elogp<logN,
p--I p'"
(
pIN
6)
it follows that, for N sufficiently large, there must exist at least two prime numbers,
q and r, such that
qr -- I
(mod6),
(q, N) - (r, N) - 1,
and
50 (log N)2 < r < q < (log N)2 <
50 - 1.02r.
The multiplicative group of congruence classes relatively prime to q6 is cyclic of
order ip(g6) - q5(q - 1). Since q - -1
(mod 6), it follows that ((p(q6).3)- 1,

so every integer relatively prime to q6 is a cubic residue modulo q6. Similarly,
every integer relatively prime to r6 is a cubic residue modulo r6. Since
(2Nr, q) - (2Nq, r) - 1,
there exist integers u and v such that
(u, q) - (v, r) - 1,
4N - u3r18
(mod q6),
2N =- v 3q 18
(mod r6).
and
The numbers 6, q6, and r6 are pairwise relatively prime. By the Chinese remainder
theorem, there exists an integer a such that
e
(mod q6)
(mod r6),
e=-1
(mod6).
Then
4N
e3r18
(mod q6)
2N
e3g18
(mod r6).
and
Let
k - 6g6r6
Then
(k, e) - (6g6r6, e) -1.

Let
x-N1/3q-6.
48
2.
Since q < (log N)2, we have, for N sufficiently large,
log x - 3 log N - 6 log q > 3 log N - 12 log log N > 4 log N

and
k - 6g6r6 < 6(log N)24 < 6(4 log X)24 << (log X)24.
By the Siegel-Walfisz theorem with A - 25 and S - 1/50,
0((51150)x-, k, t) - t9(x; k, e)
50,p(k)
50k
>>
+0 ( (log x) 2s )
+ O ( (log X)25
(log X)21
+0
\(log x)2s)
> 0.
Therefore, if N is sufficiently large, there exists a prime p such that
x<p<5o -1.02r
and
p=C
(mod 6g6r6).
The primes p, q, r are distinct because (qr, e) - 1. Since p = -1
(mod 3),
every integer is a cubic residue modulo 6p, and there exists an integer s such that
s3=N-3p (mod6p).
By the Chinese remainder theorem, there exists t such that
t3
N - 3p (mod 6p),
t = 0 (mod g2r2),
and
1 < t < 6pg2r2.

Let
n-N-t3
Then
4n - 4N - 4t3
2n - 2N - 213
4N
13r3 = p3r1 s
2N
e3q is
p3q 18
(mod q6),
(mod r6),
n-N-t3_3p (mod6p).
Finally,
n - N - t 3 <N-x3gis <p3g18
2.4
Sums of two cubes
49
and
n-N-t3
> x3q'8 - 216p3g6r6
> (1 .02)-3p3q'8 - 216p3q'2)q6_216)p3ql2
3
_ 4p3g1s
+
\\(1.02)3
-4
> 3p3gis
for N sufficiently large. Thus, the integer n = N - t3 and the primes p, q, r satisfy
conditions (2.1)-(2.5) of Lemma 2.5, so N - t3 is a sum of six positive cubes.
Since t is positive, we see that N is a sum of seven positive cubes. This proves
Linnik's theorem.
2.4
Sums of two cubes
The subject of this book is additive bases. The generic theorem states that a certain
classical sequence of integers, such as the cubes, has the property that every nonnegative integer, or every sufficiently large integer, can be written as the sum of
a bounded number of terms of the sequence. In this section, we diverge from this
theme to study sums of two cubes. 2 This is important for several reasons. First, it
is part of the unsolved problem of determining G(3), the order of the set of cubes
as an asymptotic basis and, in particular, the conjecture that every sufficiently large
integer is the sum of four cubes. Second, the equation
N - x3 + y3
(2.13)
is an elliptic curve. If r3.2(N) denotes the number of representations of the integer

N as the sum of two positive cubes, then r3.2(N) counts the number of integral
points with positive coordinates that lie on this curve. Counting the number of
integral points on a curve is a deep and difficult problem in arithmetic geometry,
and the study of sums of two cubes is an important special case.
If N - x; + y3 and.r ' y, then N - y3 + x3 is another representation of N as a
sum of two cubes. We call two representations
N-xi+yi-x2+y2
essentially distinct if (x,, y, 17( {x2, Y2 }. Note that N has two essentially distinct
representations if and only if r3.2(N) > 3.
2This section can be omitted on the first reading.
50
2.
Here are some examples. The smallest number that has two essentially distinct
representations as the sum of two positive cubes is 1729. The representations are
1729- 13+123=93+103.
These give four positive integral points on the curve
1729 - x3 +Y3 ,
so
r3,2(1729) - 4.
The smallest number that has three essentially distinct representations as the sum
of two positive cubes is 87,539,319. The representations are
87539319 - 1673 + 4363
- 2283 + 4233
- 2553 + 4143.
The cubes in these equations are not relatively prime, because
(228, 423) - (255, 414) - 3.

The smallest number that has three essentially distinct representations as the sum
of two relatively prime positive cubes is 15,170,835,645. The representations are
15, 170, 835, 645 - 24683 + 5173
- 24563 + 7093
= 21523 + 17333.
The smallest number that has four essentially distinct representations as the sum
of two positive cubes is 6,963,472,309,248. The representations are
6, 963, 472, 309, 248 - 24213 + 19,083'
- 54363 + 18, 9483
-10,2003+ 18, 0723

- 13, 3223 + 16, 6303.
It is an unsolved problem to find an integer N that has four essentially distinct

representations as the sum of two positive cubes that are relatively prime.
In this section, we shall prove three theorems on sums of two cubes. The first is
Fermat's result that there are integers with arbitrarily many representations as the
sum of two positive cubes, that is,
lim sup r32(N) - oo.
N- oo
2.4
Sums of two cubes
51
Next we shall prove a theorem of Erd6s and Mahler. Let C2(n) be the number of
integers up to n that can be represented as the sum of two positive cubes. Since
the number of positive cubes up ton is n1/3, it follows that C2(n) is at most
Erd6s and Mahler proved that this is the correct order of magnitude for C2(n), that
n2/3.
is,
>> n2/3.
C2(n)
However, numbers with two or more essentially distinct representations as sums

of two cubes are rare. Erdos observed that the number CZ(n) of integers up to n
that have at least two essentially distinct representations as the sum of two cubes
is o(n2/3). More precisely, we shall prove a theorem of Hooley that states that
CZ(n) << n(5/9)+e
This implies that almost every integer that can be written as the sum of two positive
cubes has an essentially unique representation in this form.
Theorem 2.4 (Fermat) For every k > 1, there exists an integer N and k pairwise
disjoint sets of positive integers {xi, y; } such that
N-x3+y3
for i - 1, ... , k. Equivalently,
lim supr3.2(N) - oo.
N-.oc
Proof. The functions

x(x3 +2 y3)
X3 - y3
P x' Y) =
and
Y(2x3 + y3)
g(x, Y) -
x3 - y3
satisfy the polynomial identity

.f (x, Y)3 - g(x, Y)3 - X3 + y3.
If
F(u, v) and
G(u, v) -
(uU33
+ V3
v( U3 3
+ v3
- f (u, -v)
= -g(u, -v),
then
F(u, v)3 + G(u, v)3 - f (U, -v)3 - g(U, -v)3 - U3 + (-v)3 - U3 - v3.
52
2.
Let
0<e<
1.
Let x, and y, be positive rational numbers such that

Yi
XI
<e.
We define
u-f(xi,yi),
v - g(xt, yi)
Then u and v are positive rational numbers such that
U3
- v3 - x + y] > 0.
Moreover,
where p - y,
xi(xI +2y;)
x,
1 +2p3
y, (2x; + yi)
2y,
G+ p3/2)
E (0, 1/4). Since
/x,
1< 1+
p3
it
3p3
-1+ 2+
P3/2
3p3
<
P3
+2,
foll ows t h at
u
x,
3x,p3
3x,
(Yi\3
xi/
3 y2
4(x,) <
3e2
and
-uv > 2y,

-x, > -2e> 2.
I
(2.14)
Next, we define
x2 - F(u, v),
y2 - G(u, v).
Since u > 2v, it follows from the definition of the functions F(u, u) and G(14.
that x2 and y2 are positive rational numbers. Moreover,
x2 + y2
- u3 - v3 - x; + y
Let a - v/u. Then
0<a<2e<1/2
2.4 Sums of two cubes
by (2.14) and
X2
u(u3 - 2v3)
Y2
v(2u3 - v3)
\I-2a3)
2v
u
2v
1 - a3/2
/ 11-2-a3/
3a3
3v
or
2v
2u
2-a3
Since
0<
a
2-a3
<a < 2,
1
it follows that
u
X2
2v
y2
( 2-a
2v
2v
0 <
3ua2
-(
3v
3s
- 4u
- 2u 1\2 - a3
Thus,
x2
Y2
and so
x1
4Y,
x2
Y2
2v
2v
x1
3E
3e2
2y,
< 2E,
xi
->--2e>--2e>->0.
4yi
x2
Y2
4e
8e
This proves that if x, and y, are positive rational numbers such that
0<Y)
<s<1/4,
x)
then there exist positive rational numbers x2 and y2 such that
x2+y2 -x
0<
Y2
x2
< 8E.
and
4x2
xi
Y2
Yt
< 8E.
If 8e < 1/4, then there exist positive rational numbers x3 and x4 such that
x33 + y33 - xZ + yZ ,
53
54
2.

y3
0<
< 82E,
X3
and
4x3
X2
Y3
Y2
< 82E.
Similarly, if k > 2 and

0 < 8k-2E <,
then there exist positive rational numbers x1, yi, x2, Y2, ... , Xk, Yk such that
Xj+y1
_X2+y2_..._Xk+yk
fori-1,...,k,
0<y' <8'-'E
xi
and
4x;+1
xi
< WE
fori - 1,...,k - 1.
Yi I
Yt+1
Let E - 8-k. We shall prove that the k sets {x1, y; } are pairwise disjoint. Since
4az,+j
4j-'xi+j-1
Yi+j
Yi+j-i
<4j-' 8i+j-'E-8' 32j-'e
for j - 1, ... , k - i, it follows that
4j-ix;+j-]
4tz;+1
zi
Yi+1
Yi
Yi+j-i
Yi+j
1
< We E 32j-'
j-1
<8'321E
for I 1, then
xi+t
Xi
Yi+1
Y;
and
-<(4 -1)-_
3x,
Yi
X.
Yi
141x;+1
xi
Yi+1
Yi
It follows that
3 < 8` 321e
yi
(Xi
< 82i-'32162
82ke2
<
- 1,
< 8'321E.
2.4
Sums of two cubes
55
which is absurd. Therefore, (x1, Y1), ... , {xk, yk } are k pairwise disjoint sets of
positive rational numbers. Let d be a common denominator for the 2k numbers x, ,
.... xk., Y1, ..., yk, and let N = (dx, )3 + (dy, )3. Then {d x, , d y, }, ... , {dxk, d yk )
are pairwise disjoint sets of positive integers, and
(dx1)3 + (dyl)3 = (dx2)3 + (dY2)3 = ... - (dxk)3 + (dyk)3 - N,
that is, r3.2(N) ? k. This proves Fermat's theorem.

Next, we shall prove the Erdos-Mahler theorem. This requires four elementary
lemmas.
Lemma 2.6 Let a and b be positive integers such that
a < b.
Let r(a, b) denote the number of pairs (x, y) of integers such that
- x)3 - y3 + (b - y)3
x3 + (a
(2.15)
and
0<x<2
and
0 < y < 2.
(2.16)
Then
r(a, b) <
5a213.
Proof. The function
fa(X)=X3+(a-x)3=3ax2-3a2x+a3
is strictly decreasing for 0 < x < a/2. Let r - r(a, b) > 1. Let (x1, y,),
...,
(Xr, y,) be the distinct solutions of equation (2.15) that satisfy inequalities (2.16),
and let
a
Then
3
4 `fb(2) <fb(Y1)`fa(xl)<fa(0)à3.
and so
a<b<41"3a<2a.
For i= 1, ..., r - I we have
fb(Yi+)) - fa(Xi+1) < fa(Xi) = fb(Yi),
and so
0 <y1
<<y,<2.b
(2.17)
56
2.
Moreover, the point (xi, yi) is a solution of equation (2.15) if and only if (xi, yi )
lies on the hyperbola
//
a(x-2) -b1y-21 -c,

where
C
Fori - 1,...,r,let
b3 - a3
12
> 0.
a
Ui-2-Xi
and
Vi -2-yi.
Then
0 < yr <
< v1 < 2,
and (ui, vi) is a point in the first quadrant of the u v-plane lies on the hyperbola
au2 - bv2 = c.
Since the hyperbola is convex downwards in the first quadrant, it follows that
yi+1 - Vi
vi - vi-1
ui+1 - u;
U1 - ui_1
for i - 2, ... , r - 1, and so the r - I fractions

vi+1 - vi
Yi+1 - Yi
ui+1 - Ui
xi+1 - xi
are distinct for i - 1, ... , r - 1. If r1 is the number of points (xi, yi) such that
xi+1 - xi >
a 1/3
2
then
a1/3r1
< 2'
and so
r <a2/3
Similarly, if r2 is the number of points (xi, yi) such that
al/3
Yi+l - Ys > 2 ,
2.4
Sums of two cubes
57
then
a1/3r2
`<-<a
by (2.17), and so
r2 < 2a 2/3
Let r3 be the number of points (x,, yi) such that
a1/3
xi+1-xi <2
and
a1/3
yi+1-yi< 2
Since the fractions
yi+1 - yi
xi+1 - xi
are distinct, and the numerators and denominators are bounded by a 1/3/2, we have
a1/3
r3 < f
2
a2/3
-4
Therefore,
r(a, b) < r1 + r2 + r3 + l < 3a2/3 + a43 + 1 <
5a2/3.
Lemma 2.7 Let x and y be positive integers, (x, y) = 1. If the prime p f 3 divides
x3 + y3
x+y
then
p=1
(mod 3).
Proof. Let p f3 be a prime such that

x2
- xy + y2 -
x3+ y3
= 0 (mod p).
x+y
If p divides y, then p also divides x, which is impossible because (x, y) - 1.

Therefore, (p, y) - 1. Since
(2x - y)2 +3 Y2 m 0
(mod p),
58
2.
it follows that -3 is a quadratic residue modulo p. Let (o) be the Legendre

symbol. By quadratic reciprocity, we have
(31=1
if and only if p = 1 (mod 3). This completes the proof.

In the proof of the next lemma, we shall use some results from multiplicative
number theory. Let 7r (x, 3, 2) denote the number of primes p < x such that
p - 2 (mod 3). By the prime number theorem for arithmetic progressions,
7r(x; 3, 2) x/(2logx). Moreover, there exists a constant A such that
1:
- 2loglogx+A+O (logx).
P!5.
p.7 )nod 3)
This implies that
- I loglogx - 2loglogx10/11 +O (1J_)
P.1 anal 3)
= 2 log
10
+0 ( log x )
Lemma 2.8 For any positive integer a, let h(a) denote the largest divisor of a
consisting only of primes p - 1 (mod 3), that is,
h(a) -
(2.18)
p.)
pA M
Imul 3)
Let H(x) denote the number of positive integers a up to x such that h(a) < a1110
and a is not divisible by 3. There exists a constant S1 E (0, 1) such that
H(x) > 81x
for allx >2.

Proof. Let Ho(x) denote the number of positive integers a < x of the form
a = pb, where p - 2 (mod 3) is a prime such that p > x10/11, and b is an

integer not divisible by 3. An integer a has at most one representation of this form.
Moreover,
h(a) - h(b) <b-
<x1111 < p1/10 <a1l10.
It follows that every number of the form pb is counted in H(x), and so
Ho(x) 5 H(x).
2.4
Sums of two cubes
59
Also, Ho(2) = H(2) = 1. Let g(x) denote the number of positive integers up to x
not divisible by 3. Then
2x
g(x) > 3 - 1
and
Ho(x)
\P/
io>
P.2 )mW 3)
(2x
)
-1/
<rs.
P=2 )ma13)
Zx
- tr(x, 3 2)
p.2 (ma1 3)
log1011
(logx)
log
11
10
++O
( log x )
>> X.
Lemma 2.9 Let V(d) be the Euler V -function, and let 0 < S < 1. There exists a
constant cl = ci (S) > 0 such that, if n is a positive integer and t > 8n, and if
a, <... <a, <n

are any t positive integers, then
p(a1) > cmnz.
Proof. For any p > 7, we have

2z)P
\1
f`
(p\ (-2)x
P
k-O (\k f1
1-
(p) (-2)k
kk--2,
<1--+1: p
P
k
k
p2k
2k
k-z
E (2)k
k-z
60
2.
<-+
2
P(P - 2)
Since the infinite product
Z
p>7
converges, we have
1 ln!P
\\
Ip
I1C1-p1
//
p7C1-P/
CI--(ln
1
)n/P
> F1
p<7 \
"/P
2 )11
p>7
C21
where
0<C2 < 1.
Since W(d) - d l lp1d (i
-P
and n! > (n/e)", it follows that
fl(i_I)
n
it
cp(d)-JJdfld-I
d-ppd
-n!fl
(I
p<n
Choose C3 > 0 so that

e
<3
Let
m=[2Sn
<
Sn
<m+1.
Suppose that there exists a set D c_ [1, n] such that SDI - m + I and V(d) < c3n
for all d E D. Since co(d) < d < n for all d < n, we have
n
If
flw(d) - rjW(d)jlV(d)
d-I
dal
daP
JAI
J(D
2.4
n
Sums of two cubes
61
<fC3nfln
d-I
d-1
d.D
d(D
< (C3n)m+l nn-nl-I
- Cm+I nn
3
CSn/2nn
<
< /C2n`n
(\ e
f1
which is impossible. It follows that there exist at most m integers in [ 1, n J with

cp(d1) < can. In particular, among the t > Sn integers a;, there must be at least
rSn
Sn
>2
t - m>Sn - I
integers for which W(a1) > can, and so
Sn
C3S
1P(ai)> 2 C3n- 2 n2 -cln 2,

!-1
where c1 - c3S/2. This completes the proof.
Theorem 2.5 (Erdos-Mahler) Let C2(n) denote the number of integers not exceeding n that can be written as the sum of two positive, relatively prime integral
cubes. Then
CZ(n) >> n2'3.
Proof. Let
h(a) -
pk
F1
PA I.'
prI
and let
(mod 3)
a1 <... <a,
<n1/3
be the integers in [1, n1/31 not divisible by 3 such that

h(a;) < a11
/10
Then h(1) - h(2) - 1 and so a1 - 2. By Lemma 2.8, we have
t - H(n1/3) > Sin 1/3.

Let x and y be positive integers such that
x+y-a,
Then
for some i - 1,...,t.
62
2.
Moreover, (x, y) - 1 if and only if (x. ai) - (y. ai) - 1. Therefore, the number
of pairs x, y of positive integers such that x + y - ai, x < y, and (x, y) - I is
v(ai )/2.
Let r(m) denote the number of representations of m in the form
m-x3+y3,
where x and y are relatively prime positive integers such that (x, y) - I and
x + y - ai for some i. Then
n
Ri - 1: r(m) m-l
w(al) > c6n2'3

2 i-2
by Lemma 2.9.
Let R2 be the number of ordered quadruples (x, y, u, v) of positive integers such
that
x3+y3-u3+U3,
ai -x+y < u+v -aj
fori, j E [1,1],
(x, Y)- (u,v)-1,
x<y
u<v.
and
Note that if x3 + y3 - u3 + v3, then x + y - u + v if and only if (x, y} - (u, v}

(Exercise 7). Then
R2-E(r(2)
Let (x, y, u, v) be a quadruple counted in R2. Since
3
h(a`)h(ai)
x + y
- h(aj)h(aj) u + v
and ai and aj are not divisible by 3, it follows from (2.18) that ai / h(ai) and
aj / h(a j) are products of primes p - 2 (mod 3). By Lemma 2.7,
X3 +
(p'
(p' uu+v3)
if p - 2 (mod 3). Therefore,

ai
aj
h(ai)
h(aj)
Fix the integer ai. Since
0<
ai
h(ai )
)h(aj)-a,
<n"
2.4
Sums of two cubes
63
and
a,
9/10
> a.
h(a1)
it follows that
'
n1/3
1 < h(a1) <
9/10
a,
Therefore, to each a; there correspond fewer than

nl/3
9/10
a,
different integers aj. By Lemma 2.6, the number of quadruples (x, y, u, v) such
that x + y - ai and u + v - aj is smaller than 3a2i/3. Therefore, the number R2., of
quadruples (x, y, u, v) such that x + y - a, satisfies
n1/3
3n1/3
R2,1 < 3aZ/3 9/10 - 7/30
a.
ai
and so
< 3n1j3
i7/30
< 3n113(n1/3)23/30
- 3n(2/3)-(7/90)
Let C(n) count the number of integers m up to n of the form m - x3 + y3, where
x and y are relatively prime positive integers. Since
r<I+(r
for all integers r, we have

n
r(m)1
< C2(n)+ R2.
Therefore,
C'2(n) > R1 - R2 > nz/3 - n(2/3)-(7/90) >> n2/3

64
2.
The Erdos-Mahler theorem states that many integers can be written as the sum
of two positive cubes. Hooley showed that very few numbers have two essentially distinct representations in this form. To prove this, we need the following
result of Vaughan-Wooley [130, Lemma 3.5] from the elementary theory of binary
quadratic forms.
Lemma 2.10 Let e > 0. For any nonzero integers D and N, the number of
solutions of the equation
X2-DY2-N
with
max(IXI, IYI) << P

is
<< (DNP),
where the implied constant depends only on e.
Proof. See Hua [63, chapter 11 ] or Landau [78, part 4].
The following lemma on "completing the square" shows how to transform

certain quadratic equations in two variables into Pell's equations.
Lemma 2.11 Let a, b, c be integers such that a f 0 and D - b2 - 41C. =/ 0. Let

(x, y) be a solution of the equation
axe+bxy+cy2+dx+ey+ f -0.
(2.19)
Let
X-Dy-2ae+bd
and
Y-2ax+by+d.
Then (X, Y) is a solution of the equation
X2
- DY2 - N,
where
N - (oaf -d 2 )D + (2ae - bd)2.

Moreover, this map sending (x, y) to (X, Y) is one-to-one.
The number D - b2 - 4ac is called the discriminant of equation (2.19).

Proof. Multiplying equation (2.19) by 4a, we obtain
4a 2x2+4abxy+4acy2+4adx+4aey+4af
- (2ax + by)2 - Dye + 2d(2ax + by) + 2(2ae - bd) y + 4a f
- (2ax + by + d)2
- Dye + 2(2ae - bd)y + (4a f - d2)
- Y2 - Dye + 2(2ae - bd)y + (oaf - d2)

-0,
(2.20)
2.4
Sums of two cubes
65
where
Y -tax+by+d.
Multiplying by -D, we obtain
D2Y'
- 2(2ae - bd)Dy - DY2 - (4af - d2)D
- (Dy - 2ae + bd)2 - DY2 - (oaf - d2)D - (tae - bd)2

- X2 - DY2 - ((4af - d2)D + (2ae - bd)2)
-X2-DY2-N
-0,
where
X-Dy-2ae+bd
and
N =(4af -d2)D+(2ae -bd)2.

The determinant of the affine map that sends (x, y) to (X, Y) is
0
2a
D - -2aD 710
b
since a f 0 and D f 0, and so the map (x, y) H (X, Y) is one-to-one. This

Lemma 2.12 Let P > 2, and let a, b, c, d, e, f be integers such that

P2.
Let D - b2 - 4ac, and define the integer N by (2.20). Let W denote the number
of solutions of the equation
axe+bxy+cy2+dx+ey+ f -0
with max(Ix1, lyI) << P. If a, D, and N are nonzero, then
W << IPI`
for any e > 0, where the implied constant depends only on e.

Proof. By Lemma 2.11, to every solution (x, y) of the quadratic equation (2.19)
there corresponds a solution of the equation
X2-DY2-N,
where
D-b2-4ac<< P4
and
N - (oaf - d2)D + (2ae - bd)2 << P8.
66
2.
Moreover,
XaDy-tae+bd<< P4iyl<< P5
and
Y-2ax+by+d<< P2(IxI+IYl)<< P3
if max(lxI, IYI) << P. It follows from Lemma 2.10 that
W << (DNP5)e << Put << Pe
Theorem 2.6 (Hooley-Wooley) Let D(n) denote the number of integers not exceeding n that have at least two essentially distinct representations as the sum of
two nonnegative integral cubes. Then
D(n) e
nsfs+F
Proof. If N has at least two essentially distinct representations as the sum of

two nonnegative cubes, then there exist integers x1, x2, X3, X4 such that
x1+z2-x3+x4-N
and
0<X3<X1 < X25 X4<N1"3.

For any number P > 2, let S(P) denote the number of solutions of the equation
X
+X2 - X3 +X4
(2.21)
that satisfy
0 < x3 < x1 < x2 < x4 < P.
(2.22)
D(n) < S(n1/3).
(2.23)
Then
If the integers x1, x2, x3, x4 satisfy (2.21) and (2.22), then x1 + x2 f x3 + x4 by
Exercise 7. and so
x1+x2-X3+x4+h,
where
1 < Ihl < 2P.

Let T(P, h) denote the number of solutions of the simultaneous equations
x +X2 -x3+x4
and
XI+x2-x3+x4+h
with
0 <xi <P
fori-1,...,4.
2.4
Sums of two cubes
67
Choose the integer e so that
2f < 2P <
2f+I
Then
S(P) <
T(P,h)
1:51h I <2P
1:
T(P,h)
0x3+x4+h-x1+x2>0.
For fixed h, we can use x1, ... , X4 to define four positive integers uI , u2, u3, and
y as follows:
uI -XI +X2
U2 - XI - X3
U3 - X2 - X3
y-2x4+h,
where
fori-1,2,3
] <ui <2P
and
1 < y < 4P.

Moreover,
U I +U2 + u3 - 2(x1 +X2 -X3) - 2(x4 +h) - y + h

and
h (3y2 + h2) - h (3(2Y4 + h)2 + h2)

h(12x4 + 12x4h + 4h2)
- 4(3x4h + 3x4h2 + h3)
- 4((x4 + h)3 - x4)

- 4((x1 + x2 - x3)3 - x1 - X2 + x3)
12(X x2 + XIx2
x2X3 +X2X3 2
- 12(xI + x2)(xI - x3)(x2 - x3)
- 12u1u2u3.
2
- xj X3
+ x1X32 - 2XIX2X3)
68
2.
Conversely, the numbers u I, u2, u3, and y determine xI , ... , x4 uniquely. It follows
that
T(P, h) < U(P, h),

where U(P, h) denotes the number of solutions of the equations
u1 +u2+u3 - y+h
(2.24)
12u1u2u3 - h(3y2 + h2)
(2.25)
and
in positive integers ui < 2P and y < 4P. If ui - h for some i, say, U3 - h, then
U) +u2-hand
l2uiu2 - 3y2 +h2 - 3u2 +6u1u2 + 3u2+h2.
This implies that
3(u1 - u2)2 + h2 - 0,
which is impossible since h 0. Therefore, ui h for all i - 1, 2, 3. Let

u1, U2, u3, h be a solution of equations (2.24) and (2.25) counted in U(P, h).
Let
(u3,h)-max{(ui,h): i - 1,2,3},
where (a, b) denotes the greatest common divisor of a and b. We define
d3 - (u3, h),
h
d2- u2,d
di -
u,,d2d3-
Then
d3 = max{di, d2, d3}
and d1d2d3 divides h. Let
h
g
and
ui
vi - d'
- did2d3'
for i - 1, 2, 3.
Then
(vi, g) - l and
1 < vi <
2P
di
for i - 1, 2, 3.

l2VjV2U3 - g(3y2 + h2),
(2.26)
2.4
Sums of two cubes
69
and so g divides 12, that is,
fg - 12
for some integer f. Therefore, Jhj - Igdid2d3I < 12d3, and so
d3 IhI113.
(2.27)
gd, d2.
(2.28)
Since u3 f h, it follows that

V3
We can rewrite equation (2.25) in terms of the new variables v;, d;, f, g. Since
h - gdld2d3
and
y - dl vl + d2v2 + d3v3 - h,
we have
12u1u2u3 - fgdld2d3v, V2V3 - f hvi V2V3 - h(3y2 + h2),
and so
fvlv2v3 - 3(dlv1 +d2u2+d3v3 -h)2+h2.
(2.29)
If we fix the integers dl , d2, d3, f, g, v3, then equation (2.29) becomes a quadratic
equation in v1, v2:
3d, v + (6d, d2 - f v3)v1 v2 + 3d2 v2 + 6d, (d3v3 - h)vl

+6d2(d3v3 - h)v2 + 3(d3v3 - h)2 +h2 - 0.
(2.30)
The discriminant of this quadratic is
D - ((6dld2 - f v3)2 - 36d d2
- f2v3 - l2dld2fv3
- f2vs - dld2f2gv3
- f2v3(v3 - dld2g)
f0
by (2.28). Similarly, the integer N defined by (2.20) is nonzero, because
N - (4-3d 2 (3(d3v3 - h)2 +h2) - (6d1(d3v3 - h))2) D

+ (2 .3d2 6d2(d
1
12d1 h2D + (6dl f v3(d3u3 - h))2
12dih2f2v3(v3 - dldzg)+36d1 f2v3ds(v3 - d, d2g))2
- 12dI dsf2u3(v3 - di d2g) ((d, d2g)2 - 3d1d2gv3 +3v3)

- 3d2j d3 f 2 v3(v3 - d, d2g) ((dl d2g)2 + 3 (d l d2g - 2v3) 2)
10.
`2
70
2.
Let W(P, dl, d2, d3, f, g, v3) denote the number of solutions of equation (2.30) in
integers v1, v2 satisfying (2.26). Since the coefficients of this quadratic equation
are all ' P2, it follows from Lemma 2.12 that

W (P, dl, d2, d3, f, g, v3) << P`.
Therefore,
S(P) << log P max
T(P, h)
<H<2"
H<Ihl<2H
U(P,h)
<<Iog P max
I <H<2P
H<Ihl<2H
<< log P max
I <H<2P
H<Ihl<2H fg-12
Sd)J2d3-h
d3>n,o J) A2)
W(P, d1, d2, d3, .1, g, v3)

1
1P;dj
'"l"dld2
<<log P max
I <H<2P
P
H<Ihl<2H fg-12
gd)d2d3-h
15,3_1P/d3
d3>muld1.d21
,3,I djd2
P l+e
<< P` max
I <H<2P
H<Ihl<2H fg-12
d3
1djJ2d3-h
J3>m.a1J .d2 )
<< P1+2P max
1 <H<2P
HIhl<2H
Sd,d2d3d3'.m.adIA2)
d3
Since the number of factorizations of h in the form h - gdld2d3 is << Ihit, and
since
d3IhI113
by (2.27), we have
<<
11<101<2H
SJId2d3''
d3 >m.n(dj.d2)
H<h<2H
h1/3- e
and so
S(P) << PI+2e max
I<H<2P
H23+P
<<
<<
H23+e'
p53+3e
Therefore, by (2.23), we have
D(n) < S(n'/3) << ns/9+e
Theorem 2.7 (Erd6s) Almost all integers that can be represented as the sum of
two positive cubes have essentially only one such representation.
2.5
Notes
71
Proof. This follows immediately from the remark that there are greater than cn2/3
integers that can be represented in at least one way as the sum of two nonnegative
cubes, but there are no more than c'n5/9+F = o(n213) integers that have two or more
essentially distinct representations as the sum of two cubes.
2.5
Notes
Wieferich's proof [ 144] that g(3) - 9 appeared in Mathematische Annalen in 1909.

In the immediately following paper in the same issue of that journal, Landau [75)
proved that G(3) < 8. Dickson [24] showed that 23 and 239 are the only positive integers not representable as the sum of eight nonnegative cubes. An error
in Wieferich's paper was corrected by Kempner [70]. Scholz [ 108] gives a nice
version of the Wieferich-Kempner proof.
Linnik's proof [81J of the theorem that G(3) _< 7 is difficult. Watson [139]
subsequently discovered a different and much more elementary proof of this result,
and it is Watson's proof that is given in this chapter. Dress [25] has a simple proof
that G(3) < 11.

Vaughan [ 1261 obtained an asymptotic formula for r3.8 (n), the number of representations of an integer as the sum of eight cubes. It is an open problem to obtain
an asymptotic formula for the number of representations of an integer as the sum
of seven or fewer cubes.
It is possible that every sufficiently large integer is the sum of four nonnegative
cubes. Let E(x) denote the number of positive integers up to x that cannot be written
as the sum of four positive cubes. Davenport [ 17] proved that E4.3(x) << x29130+F
and so almost all positive integers can be represented as the sum of four positive
cubes. Brudern [6] proved that
E4.3(x)
<<
x37/42+e
There are interesting identities that express a linear polynomial as the sum of
the cubes of four polynomials with integer coefficients. Such identities enable us
to represent the integers in particular congruence classes as sums of four integral cubes. See Mordell [85, 86], Demjanenko [201, and Revoy [101] for such
polynomial identities.
Theorem 2.5 was first proved by Erd6s and Mahler [31, 351. The beautiful
elementary proof given in this chapter is due to Erd6s [31 ]. Similarly, Theorem 2.6
was originally proved by Hooley [57, 58]. The elementary proof presented here is
due to Wooley [ 149]. For an elementary discussion of elliptic curves and sums of
two cubes, see Silverman [ 1151 and Silverman and Tate [ 116, pages 147-1511.
Waring stated in 1770 that g(2) - 4, g(3) - 9, and g(4) - 19. The theorem that
every nonnegative integer is the sum of 19 fourth powers was finally proved in
1992 in joint work of Balasubramanian [2] and Deshouillers and Dress [21 ).
72
2.
2.6
Exercises
1. Prove that
33+43+53 -63
is the only solution in integers of the equation
(x - 3)3 + (X - 2)3 + (X - 1)3 - x3
2. Let s(N) be the smallest number such that N can be written as the sum of
s(N) positive cubes. Compute s(N) f o r N - 1, ... , 100.
3. Prove that s(239) - 9, that is, 239 cannot be written as a sum of eight
nonnegative cubes.
4. Show that none of the following numbers

15
22
175
186
303
364
50
212
420
114
167
231
238
454
428
can be written as a sum of seven nonnegative cubes.
5. Show that none of the following numbers
79, 159.239, 319.399, 479, 559

can be written as a sum of 18 fourth powers.
6. Let v(3) denote the smallest number such that every integer can be written
as the sum or difference of v(3) nonnegative integral cubes.
(a) Prove that

4 < v(3) < g(3).
(b) Prove that
v(3) < 5.
Hint: Use the polynomial identity
6x-(X+1)3+(X-1)3-2x3
and the fact that x - (N - N3)/6 is an integer for every integer N.

It is an unsolved problem to determine whether v(3) - 4 or 5. This is called
the easier Waring's problem for cubes.
7. Let x, y, u, v be positive integers. Prove that if x + y - v + v and x3 + y3

u3 + v3, then (x, y) - (u, v}.
2.6
73
Exercises
8. (Von Sterneck [ 136]) Using a computer, calculate s(n) for n up to 40,000.

Verify the results of Lemma 2.4.
9. (Mahler [821) Prove that I has infinitely many different representations as
the sum of three cubes. Hint: Establish the polynomial identity
(9x4)3 + (3x - 9x4)3 + (1 - 9x3)3 = 1.
(2.31)
Prove that
(9m4)3 + (3mn3
- 9m4)3 + (n4 - 9m3n)3 - n 12.
Let r3.3(N) denote the number of representations of N as the sum of three

nonnegative cubes. Prove that if N = n 12 for some positive integer n, then
r3.3(N) > 9-'13N'112
Note: This is Mahler's counterexample to Hypothesis K of Hardy and Littlewood [49].

10. (Elkies and Kaplansky [27]) Verify the following polynomial identities:
8(x2 + y2 - z3) _ (2x + 2y)2 + (2x
- 2y)2 - (2z)3,
2X +1_(X3-3X2+X)2+(X2-X-1)2- (X2-2x)3,
2(2x+1) -(2X3 -2X2 -x)2 - (2X3 -4X2 - X+ 1)2 -(2X2 - 2x - 1)3,
4(2x+1)=(X3+x+2)2+(x2 - 2x - 1)2-(x2+1)3.
Show that every integer N, positive or negative, can be written uniquely in
the form
N = 842r(2m + 1),
where q > 0, r E 10, 1, 21, and m E Z. Prove that every integer N can be
written in the form
N=a2+b2 -c3,
where a, b, c are integers.
11. Let a be a positive rational number. Consider the equations
a=x3+y3+i3
a =(x+y+z)3 -3(y+z)(z+x)(x+y)
8a = (u + v + w)3 - 24uvw.
Prove that if any one of these equations has a solution in positive rational
numbers, then each of the three equations does.
74
2.
12. Let a be a rational number. Let r be any rational number such that r
0 and
t-72r3
For any rational number w, let
u-
2412
((t + 1)3
and
v-
24t
(t + 1)3
w.
Prove that
-)3
uw
(u + v + w)3 - 24uvw = fSti (r(

t+1
Let w - r(t + 1). Prove that there exist rational numbers .r. v, z such that
and
a-x3+y3+z3.
This proves that every rational number can be written as the sum of three
rational cubes.
13. Let a be a positive rational number. Show that it is possible to choose r in
Exercise 12 so that
a -x3+y3+z3,
where x, y, z are positive rational numbers. This proves that every positive
rational number can be written as the sum of three positive rational cubes.
The Hilbert Waring theorem
Nous ne devons pas douter que ces considerations, qui permettent ainsi
d'obtenir des relations arithmetiques en les faisant sortir d'identites

oil figurent des integrales definies, ne puissent un jour, quand on en
aura bien compris de sens, titre appliquees A des problbmes bien plus
etendus que celui de Waring. I
H. Poincare [961
3.1
Polynomial identities and a conjecture of Hurwitz
Waring's problem for exponent k is to prove that the set of nonnegative integers
is a basis of finite order, that is, to prove that every nonnegative integer can be
written as the sum of a bounded number of kth powers. We denote by g(k) the
smallest number s such that every nonnegative integer is the sum of exactly s kth
powers of nonnegative integers. Waring's problem is to show that g(k) is finite;
Hilbert proved this in 1909. The goal of this chapter is to prove the Hilbert-Waring
theorem: the kth powers are a basis of finite order for every positive integer k.
We have already proved Waring's problem for exponent two (the squares) and
exponent three (the cubes). Other cases of Waring's problem can be deduced from
'We should not doubt that [Hilbert's] method, which makes it possible to obtain arithmetic relations from identities involving definite integrals, might one day, when it is better
understood, be applied to problems far more general than Waring's.
76
3.
these results by means of polynomial identities. Here are three examples. We use
the notation
(XI X2 f ... xh
)k = r
(XI + E2X2 + ... + EhXh )k .
f 1..... Fh'f 1
Theorem 3.1 (Liouville)

1
(x + x2 + x3 + xz)z
(xi + xj)4 +
6 I<i<j4
(x; - xj)4
6 I<i<j<4
is a polynomial identity, and every nonnegative integer is the sum of 53 fourth

powers, that is,
g(4) < 53.

Proof. We begin by observing that
(X I x2)4 = (XI + x2 )4 + (XI - x2 )4 - 2x 1 + 12x 1x2 + 2x2 ,
and so
(xi xj)4 =
(X, - Xj)4
(Xi + xj)4 +
l5i<j<4
1<i<j<4
I<i<j<4
(2X4 + 12x?x2+ 2xI
1<i<j<4
4
=6Ex4+12 Y xIxj
i-I
1<i<j<4
a6 XI
(
+x2+x3+X; )
z
.
This proves Liouville's identity.

Let a be a nonnegative integer. By Lagrange's theorem, a = x2 + x2 + x3 + x4
is the sum of four squares, and so
6a2 = 6 (x 2 + x2 + x3 + x2)
1<i<jt4
(Xi + xj)4 + r (x, _ xj)4

1<i<)<4
is the sum of 12 fourth powers. Every nonnegative integer n can be written in the
form n - 6q + r, where q > 0 and 0 < r < 5. By Lagrange's theorem again, we

have q -a 2 +. + a4, and so 6q - 6a + + 6a2 is the sum of 48 fourth powers.
Since r is the sum of 5 fourth powers, each of them either 04 or 14, it follows that
n is the sum of 53 squares. This completes the proof.
The proofs of the following two results are similar.
Hermite polynomials and Hilbert's identity
3.2
77
Theorem 3.2 (Fleck)

(X, +X2 +x3 +X4)3
1
60
(x, x J)6+
(x1 xJ Xk)b+
30 l<i<j<4
t<i<j<k<4
xir
5 1<i<4
is a polynomial identity, and every nonnegative integer is the sum of a bounded

number of sixth powers.
Theorem 3.3 (Hurwitz)
(x +x2 +x3 +x4)

I
840
1
84
(x1 x2 x3 X4)8s +
(xi Xj)8 + 1
5040 1<i<j<k<<4
(2xi xj xk)
(2x,)6
840 I<<4
1 <i < j <4
is a polynomial identity, and every nonnegative integer is the sum of a bounded

number of eighth powers.
Suppose that
M
(xi + ... +X2)k
a1 (bi.1 x1 + b1.2-r2 + bi3X3 + b;.axa)
(3.1)
i-1
for some positive integer M, integers bi,j , and positive rational numbers ai. Hurwitz
observed that this polynomial identity and Lagrange's theorem immediately imply
that if Waring's problem is true for exponent k, then it is also true for exponent 2k.
Hilbert subsequently proved the existence of polynomial identities of the form (3.1)
for all positive integers k, and he applied it to show that the set of nonnegative
integral kth powers is a basis of finite order for every exponent k. This was the first
proof of Waring's problem. In the next section, we obtain Hilbert's polynomial

identities.
3.2
For n > 0, we define the Hermite polynomial H,, (x) by
Hn(x)-
21)nex.2 dnn / (e_s2l
The first five Hermite polynomials are
Ho(x) - 1
78
3.
HO) = x
1
H2(x) - x2 - 2
H3(x) - x; - 2 x
H4(x) = x 4
- 3x 2 +
Since
HH(x) = (
d (ex ,
21 )n dx
dxn
21) (2x)e12
(e"))
dx"
n
2)"+1
(e-'2) - 2 I -1
= 2xH,,(x) - 2H"+I(x),
n+1
dxn+l
\e-X
the Hermite polynomials satisfy the recurrence relation

Hn+I (x) - x H,, (x) -
H (x ).
(3.2)
It follows that HH(x) is a monic polynomial of degree n with rational coefficients

and that H" (x) is an even polynomial for n even and an odd polynomial for n odd.
Lemma 3.1 The Hermite polynomial H,,(x) has n distinct real zeros.
Proof. This is by induction on n. The lemma is clearly true for n - 0 and n - 1,
since H, (x) - x. Let n > 1, and assume that the lemma is true for n. Then H,, (x)
has n distinct real zeros, and these zeros must be simple. Therefore, there exist
real numbers
hen < ... <VL <Yl
such that
H,,(01)=0
and
H (,8)
for j - 1, ... , n. Since H,, (x) is a monic polynomial of degree n, it follows that
lim H"(x) - oo,
x-.oo
and so
H,,(8I) > 0.
Since the n - I distinct real zeros of the derivative H,(x) are intertwined with the
n zeros of HH(x), it follows that
(-1)f+'H,,'(81) > 0
3.2
79
for j - 1, ..., n. The recurrence relation (3.2) implies that

Hn+I (fig) - tsi Hn (fli) - 2 H,, (+Si )
and so
2 Hn (lli ),
_ f+I
(-I)1 Hit. I(0j) -
Hn(fii) > 0
for j - 1, ... , n. Therefore, for j - 2, ..., n,

has a zero 0, in each open
interval (fif, Pj_1). Since lim.,,, H,,, (x) - oo and H1(fi1) < 0, it follows that
Hn+I(x) has a zero Pi > 01. If n is even, then
0. Since n + I is odd,
Hn+1(x) is a polynomial of odd degree, and so limz-__,o Hn+I (x) - -00. It follows
that
has a zero fin+I < fin. Similarly, if n is odd, H1(Yn) < 0 and the
even polynomial
has a zero fin+1 < Pit. Thus, Hn+1(x) has n + 1 distinct
real zeros. This completes the proof.
Lemma 3.2 Let n > 1 and f (x) be a polynomial of degree at most n - 1. Then
f-C.*
v2 Hit (x)f(x)dx - 0.
Proof. This is by induction on n.) If n - 1, then HH(x) - x and f (x) is constant,

say, f (x) - ao, so
e-x2Hn(x)f(x)dx-aoJ :e-`'xdx-0.
00
00
Now assume that the lemma is true for n, and let f (x) be a polynomial of degree
at most n. Then f'(x) is a polynomial of degree at most n - 1. Integrating by parts,
we obtain
00
2Hn+I(x)f(x)dx
00e
n+1
-\
(_1
- \2
21
-0.
/
1n+1
fcc do+1
xl
f(x)dx
oo dxn+I \e
2
f Oc d n
(e--'2) f'(x)dx
OC dx"
00
f
/ f00
e-z.2Hn(x)f'(x)dx
Lemma 3.3 For n > 0,

Cn -
0C
0o
e-
2 xn dx a j
t
if n is even
if n is odd.
(3.3)
80
3.
Proof. This is by induction on n. For n - 0, we have
e-.r dx
- fir
and so co - 1. For n - 1, the function a-c2x is odd, and so
C*
e-x xdx - 0
and c, - 0. Now let n > 2, and assume that the lemma holds for n - 2. Integrating
by parts, we obtain
00
C-
- nJ-
e-`2x"dx
,p
(n
2 I
-(n22
/ ,7" J00
e-x2x"-2dx
C"-Z-
Ifn is odd, then Cn_2-0 and so c" -0.Ifn is even,
) Cn-2
- (n -1
2
n-1
(n-2)!
- ( 2 ) 2ie-2 ((n - 2)/2)!
c',
n!
2"
(n/2)!.
Lemma 3A Let n > 1, let 01, ... , Yn be n distinct real numbers, and let co, c1,
cnbe the numbers defined by (3.3). The system of linear equations
fork -0, 1,...,n - 1
.fxf -ck
(3.4)
f-1
has a unique solution p l , ... , p". If r(x) is a polynomial of degree at most n - 1,

then
r(ff)pf
J-ooe-s2r(x)dx.
f-1
Proof. The existence and uniqueness of the solution pl, ..., p,, follows immediately from the fact that the determinant of the system of linear equations
,61X1
A
XI
L Xi
+
+
+
X2
Q
,62X2
+--'+
2X2
+"'+
92--1X2
+...+
Xn
CO
YnXn
Cl
RR
Q
AtXn
fl."-IX,,
C2
Cn-1
3.2
81
is the Vandermonde determinant

1
..
f[
...
An
...
A2
01
f2
Let r(x) -
...
P2-I
ti -I
11
I<i<j<n
(fij - fi)
0.
p;;-I
akxk. Then
n-I
+r($j)Pj - 1: 57 akf'Pj
QQ
j-I k-0
n-I
j-1
k-0
akE fijPi
j-1
n-I
akCk
kk-O
nI
ao
f:exLdx
Eak
k-0
-00
- 1 f e-"r(x)dx.
00
Lemma 3.5 Let n > 1, let
be the n distinct real roots of the Hermite polynomial H,,(x), and let pl, ... , p,, be the solution of the system of linear
equations (3.4). Let f (x) be a polynomial of degree at most 2n - 1. Then
1
Tn
QQ
j-I
f 00
f (Yj)Pj =
V 7'
r'.f (x)dx.
e
00
Proof. By the division algorithm for polynomials, there exist polynomials q(x)
and r(x) of degree at most n - I such that
f (x) - HH(x)q(x) + r(x).
j =QQ1,...,n,,Qwehave
Since
f(f31) - H,(Yj)q(f3 )+r(Yj) - r(fi ).

R
and so, by Lemma 3.4 and Lemma 3.2,

n
1: f($j)Pj - 1: r(fj)Pj
j-I
j-I
82
3.
oo
fe_r(x)dx
o0
00
00
e--"HH(x)q(x)dx + 1
]
00
Jp00
rz
e--' r(x)dx
o0
f (z)dx.
Lemma 3.6 Let n > 1, let 01, ... , be the n distinct real roots of the Hermite
polynomial Hn(x), and let p l, ... , pn be the solution of the linear system (3.4).
Then
fori-l.....n.
p;>0
Proof. Since
Hn(X)-FI(x-fij).
j-t
it follows that, for i - I, ... , n,

/
( -fli)
//
fi (x) -
``
- FI(x - YR j )2
111,
1'
is a monic polynomial of degree 2n - 2 such that f (x) ? 0 for all x. Therefore,
JSince f, ($,) > 0 and f ()4j) - 0 for j i i, we have, by Lemma 3.5,

n
f(,5,)pi - E f;(,Oj)pj
j-t
7f
n
e_x2
f;(x)dx
00
> 0.
Lemma 3.7 Let n > 1, and let co, c l, ... , cn_t be the rational numbers defined
by (3.3). There exist pairwise distinct rational numbers 8i , ... , fln and positive
rational numbers p,*, .
p, such that
n
E(,6i)kpi -ck
j-1
fork-0, 1,...,n- 1.
3.2
83
Proof. By Lemma 3.4, for any set of n pairwise distinct real numbers #I, ...
the system of n linear equations in n unknowns
n
fork-0, 1,...,n - 1
fjx1 -cr
j-I
has a unique solution (pl, ..., pn). Let R be the open subset of R consisting
of all points (01, ... ,
such that Pi fj for i j, and let 0 : R -> R be
the function that sends (Yl, .... fn) to (PI, ... , pn). By Cramer's rule for solving
linear equations, we can express each pj as a rational function of l31 , ... , ion, and
so the function
C01, ,A)-(PI, ,Pn)

is continuous. Let R+ be the open subset of R' consisting of all points (x1, ... , x,,)
are then zeros of
such that x; > 0 f o r i - 1 , ... , n. By Lemma 3.6, if NI , ... ,
then (#I, ... , &) E R and

E R+.
Since R+ is an open subset of Rn, it follows that c 1(R+) is an open neighborhood

in R. Since the points with rational coordinates are dense in R, it
of (01, ...
follows that this neighborhood contains a rational point (P', ... , P,*,). Let
(PI+
,OR)ER+.
,Pn)-0(0I+
Since each number p' can be expressed as a rational function with rational coefficients of the rational numbers 0,*, ..., !3,*,, it follows that each of the positive
numbers p' is rational. This completes the proof.
Lemma 3.8 Let n > 1 , let co, c1, ... , cn_I be the numbers defined by (3.3), let
01, ... , fl be n distinct real numbers, and let pl, ... , pn be the solution of the
linear system (3.4). For every positive integer r and for in - 1, 2, .... n - 1,
r
X + ... 2)M/2.
+xra
2
rt
...
\\m
R
pj ... pj. ((sj xl + ... + 0j,x,)
is a polynomial identity.
Proof. The proof is an exercise in algebraic manipulation and the multinomial

theorem. We have
n
E ..
ji-I
Pji ... pj, lflj,xl +... +,Ojxrl
r
A, -_O
1 ..
(0j, xl ) ... (0j, XOM'

1
84
3.
j, -I
n
u,
Jul
...
j1_1
- m!
x (PiI
it
It
m1
Pj)
...xrt
QQu,
j, Pj,
{t,
(filul
j, Pj,
j,-1 ,-1 i'
ji-1
r
= m!
Xu,
n
oi<
F1
Pj
j-1
>n>n
Clt'XU
-m!
F1
i-I
By Lemma 3.3,cm-0if mis odd. If mis odd and
then At
must be odd for some i, and so

n
Pj, ...Pj, /lfij,X1 +...+Pj
E ...
ji-1
Xr1lm
-0.
j,_1
This proves the lemma for odd m. If m is even, then we need only consider partitions of m into even parts At - 2v;. Inserting the expressions for the numbers c
from (3.3), we obtain
It
It
...
Pj,
/Q
,q
\\m
(j,X1 +...+Yj,Xrj
Pj,
j,-1
1,-1
r7
m!
2v;
x;
11
r
(2v; )!!
2v;
(2v;). X
H
i_t 22"+v;! (2v;)!
m!
v, 10
Mr r
-2m E H
r
2v;
v;l
ME r
--!2 ;_1
v,.
>a
2m2
(x2l
v;.
,>a
/ 21 v
(m/2)!
2\" ...`Xr1
v1!...Vr! (X1
ti
Cm
m
- C. (Xl + ... X2)
r
This proves the polynomial identity.
3.2
85
Theorem 3.4 (Hilbert's identity) For every k > 1 and r > 1 there exist an
integer M and positive rational numbers ai and integers b;.jfor i - I__ , M and
j-1,...,rsuch that
AN
(XI +... +xr)k' - Eai (bi.,x1 +...+bi.rX,) 2k
(3.5)
Proof. Choose n > 2k, and let $i ......6h , pl*, ..., p, be the rational numbers
pn
constructed in Lemma 3.7. Then #, . ... .)4h are pairwise distinct and p,*, are positive. We use these numbers in Lemma 3.8 with m - 2k and obtain the
polynomial identity
n
C2k (x1 + ... + Xrl)k s
pi ... pi (0i1 xl +...+PJ Xr) 2k
..
Let q be a common denominator of the n fractions 8,*, ... , f,;. Then qfi is an
integer for all j, and
k
i,-1
1,-,
ps
pr
n
?i
c2kq
(qp*x,+...+qp x.)
2k
is a polynomial identity of Hilbert type. This completes the proof.
Lemma 3.9 Let k > 1. If there exist positive rational numbers a,, ..., am such
that every sufficiently large integer n can be written in the form
m
na E
atYik,
(3.6)
where x1, ... , xM are nonnegative integers, then Waring's problem is true for
exponent k.
Proof. Choose no such that every integer n > no can be represented in the
form (3.6). Let q be the least common denominator of the fractions al, ... , aM.
Then qai E Z for i - 1, ... , M, and qn is a sum of F"f, qa1 nonnegative kth
powers for every n > no. Since every integer N > qno can be written in the form
N - qn + r, where n > no and 0 < r < q - 1, it follows that N can be written as
the sum ofqaj + q - 1 nonnegative kth powers. Clearly, every nonnegative

integer N < qno can be written as the sum of a bounded number of kth powers,
and so Waring's problem holds fork. This completes the proof.
The following notation is due to Stridsberg: Let EM1 aixk be a fixed diagonal
form of degree k with positive rational coefficients a,a , ,.. aM. We write n - E(k)
if there exist nonnegative integers x, , ... , xM such that
m
n =
aixk.
r-I
(3.7)
86
3.
We let >2(k) denote any integer of the form (3.7). Then >2(k) + >2(k) - >2(k)
and >2(2k) - >2(k). Lemma 3.9 can be restated as follows: If n - >2(k) for every
sufficiently large nonnegative integer n, then Waring's problem is true for exponent
k.
Theorem 3.5 If Waring's problem holds fork, then Waring's problem holds for
2k.
Proof. We use Hilbert's identity (3.5) for k with r - 4:

M
(x +...+x4)k -> aj (bj,ixt +...+bi.4x4) 2k

j-t
Let y be a nonnegative integer. By Lagrange's theorem, there exist nonnegative

integers xt , x2, x3, x4 such that
y-x; +x2+x3+x2,
and so
M
2k
Yk -
3.8)
EQj Zj
i-t
where
zj - bj. ix, + ... + bi.4x4

is a nonnegative integer. This means that
Yk - E(2k)
for every nonnegative integer y. If Waring's problem is true for k, then every
nonnegative integer is the sum of a bounded number of kth powers, and so every
nonnegative integer is the sum of a bounded number of numbers of the form >2(2k).
By Lemma 3.9, Waring's problem holds for exponent 2k. This completes the proof.
3.3 A proof by induction

We shall use Hilbert's identity to obtain Waring's problem for all exponents k > 2.
The proof is by induction on k. The starting point is Lagrange's theorem that every
nonnegative integer is the sum of four squares. This is the case where k - 2. We
shall prove that if k > 2 and Waring's problem is true for every exponent less than
k, then it is also true fork.
Lemma 3.10 Let k > 2 and 0 < e < k. There exist positive integers B0.e, Bi.e.
.... Be-1.e depending only on k and e such that
x21Tk-f
f-t
+ E B;.ex1 Tk
j-o
' - E(2k)
3.3
A proof by induction
87
for all integers x and T satisfying
x2 < T.
Proof. We begin with Hilbert's identity for exponent k + f with r - 5:
ht,
2
(x1
+...+x5 ) k+r
- E ai (6;.1x1 +...+6;.5X5) 2k+2f

i-
where the integers Mf and bi.j and the positive rational numbers ai depend only
on k and e. Let U be a nonnegative integer. By Lagrange's theorem, we can write
U-x
+X2+X2+x4
3
for nonnegative integers x1, X2, x3, x4. Let x5 - x. We obtain the polynomial
identity
M,
(x2 + U)k+r -
a; (bix
(3.9)
+c1)2k+2r
i-i
where the numbers Mr, a;, and b; - b,.5 depend only on k and e, and the integers
c; - b, 1 x, +
+ bi.4x4 depend on k, e, and U. Note that 2e < k + e since e < k.
Differentiating the polynomial on the left side of (3.9) 2e times, we obtain (see
Exercise 6)
d21
((X2 + U)k+r\ _
[IX 2f
A,
1x2r(X2
+ U)k-i,
i-O
where the Ai.r are positive integers that depend only on k and e. Differentiating
the polynomial on the right side of (3.9) 2f times, we obtain
d21
M1
ai (b; x + ci )2k+2f
dx2f
M,
1:(2k + I)(2k + 2)...(2k +
ci)2k
i-1
M,
- > a; (b; x + ci )2k

t-1
zk
a y;
where y; - Ib;x + ci I is a nonnegative integer and
a' - (2k + l)(2k + 2) .. (2k + 2e)b2f ai

is a nonnegative rational number depending only on k and f. It follows that, if x
and U are integers and U > 0, then there exist nonnegative integers yi, ... , yhr,
such that
r
Mj
k-i
zk
Ai.tx 2i (X2
i-0
i-t
88
3.
The Hilbert-Waning theorem
Let x and T be nonnegative integers such that x2 < T. Since A1,t is a positive
integer, it follows that x2 < At.eT, and so
U - AT - x2
is a nonnegative integer. With this choice of U, we have
t
E Ai.ex2i(x2 + wk-i - E Ai.ex"(At.eT )k-`
e
i-0
i-0
- r At,lt.e
Ak-ix21T-1
i-0
t
ACk.[ -l+l
t-.Ii-1x2iTk-i
Ai.[AL
i-0
t
where Bt,t - I and
B. x2 Tk-i
.t
Ak-C+1
t.t
i-0
l-i-'
Bi.t - Ai.l At.c
is a positive integer for i - 0, ... , e - 1. Let
a
ai
Then
x2t Tk-t
t-1
k-t+1
M,
+ E B1,1xITk-` - E aiy
i-0
- E(2k).
i-1
Theorem 3.6 (Hilbert-Waring) The set of nonnegative kth powers is a basis of

finite order for every positive integer k.
Proof. This is by induction on k. The case k - I is clear, and the case k - 2

is Theorem 1.1 (Lagrange's theorem). Let k > 3, and suppose that the set of tth
powers is a basis of finite order for every e < k. By Theorem 3.5, the set of (2t)-th
powers is a basis of finite order for e - 1, 2, ..., k - 1. Therefore, there exists an
integer r such that, for every nonnegative integer n and for t - 1, ..., k - 1, the
equation
is solvable in nonnegative integers xl,t, ... , xr.t. (For example, we could
max{g(2t) : e - 1, 2, ... , k - 11.)
Let T > 2. Choose integers Cl,... , Ck_i such that
0<Ct <T fort-1,...,k-1.
let r -
3.3
89
There exist nonnegative integers xj.e for j - 1, ... , r and e - 1, ... , k - 1 such
that
(3.10)
Ck_t.
Then
xj.e<Exj'e < Ck_t<T

j-I
k - 1, and i = 1, ... , f. By Lemma 3.10, there exist
for j
positive integers Bi.t depending only on k and a such that

e-I
xze
j,e
Tk_t
1:(2k) -
+ E B ..tx2'.Tk-i
j.(
(k).
(3.11)
i-0
Summing (3.11) for j - 1, ... , r and using (3.10), we obtain

t-I
Ck-,T k-t+EB1.,Tk
>2x2't,
i-0
j-1
- Ck_fTk-t +Tk-t+1
e-1
EBi.eTf-1
` ExZe
i-0
Ck-tTk_t
j-1
Dk-t+ITk-e+1
- >2(k),
where
t-I
2i
zj_t
Bi.tTe-I-i
Dk-t+1 =
i-0
j-1
for e - 1, ... , k - 1. The integer Dk_t+1 is completely determined by k, f, T, and

Ck_t and is independent of Ck_i for i f e. Let
Then
0 < Ck_tTk-t +
Dk_1+1Tk-e+l
e-I
Ck_eTk-e +
< B*
Bi.eTk-` >
2
j.e
i-O
j-1
T-t+I +rTk +
e)
Tk-i+1
i-I
e-1
rTk+Tk-t+1 >Ti
= B*
1-0
<B*(rTk+
Tk+1
T-1
(r + 2)B*T,
k
90
3.
since T/(T - 1) < 2 for T > 2. Let
Ck-D,-0.
Then
k-I
E (Ck-IT-t + Dk-t.,
Tk-t+I) -
t-I
(Cl + Dt) Tt E(k)

E
t-I
and
k
(Ct + Dt) Tt < (k - 1)(r + 2)B`Tk - E'Tk,
0:5
r-I
where the integer
E' - (k - 1)(r + 2)B'

is determined by k and is independent of T. If we choose
T>E',
then
k
0<>(Ct+D1)Tt <E*Tk
<Tk+I,
t-I
and so the expansion of Ek_I (Cl + DI) T t to base T is of the form

k
1: (Ct+DI)Tt - E,T +...+Ek_,Tk-I +EkTk,
(3.12)
t-I
where
0<E; <T fori-1,...,k-I
and
0<Ek<E'.
In this way, every choice of a (k - 1)-tuple (C,..... Ck_I) of integers in 10,
1 , ... , T - 1) determines another (k - 1)-tuple (E, , ..., Ek _,) of integers in
(0, 1,
... , T - 1). We shall prove that this map of (k - 1)-tuples is bijective.
It suffices to prove it is surjective. Let (E, , ..., Ek_,) be a (k - 1)-tuple of

integers in (0, 1 , ..., T - 1 ). There is a simple algorithm that generates integers C1, C2, .... Ck _ I E (0, 1, ... , T - 1) such that (3.12) is satisfied for some
nonnegative integer Ek < E*. Let C, - E, and I2 - 0. Since D, - 0, we have
(C,+D,)T-E,T+12T2.
The integer C, determines the integer D2. Choose C2 E 10, 1, ..., T - 1) such
that
C2 + D2 + 12 = E2 (mod T).
3.3
91
Then
C2+D2+12-E2+13T
for some integer 13, and
2
Dc, + Dt)Tt -
EtTt + 13T3.
t-1
t-1
The integer C2 determines D3. Choose C3 E 10, 1, ... , T - 1) such that
C3+D3+13 - E3 (mod T).

Then
C3+D3+13a E3+I4T
for some integer 14, and
3
>(Ce + Dt)Tt a
Et T' + 14T4.
l-1
t-1
Let 2 < j < k - 1, and suppose that we have constructed integers Ii and
C1,...,Cj_1 E (0, 1,..., T - 1)

such that
j-1
J-1
1:(Ct + D,) T' -
E, T' + Ij Ti.
t-1
t-1
There exists a unique integer C; E 10, 1, ... , T - 1) such that
Cj +Df+IJ - Ej (mod T).

Then
Cj + Dl + Ij - E1 + 1j+1T
for some integer 1j+1, and
L(Ce + De)Tt
- L EeTt + 1m Tt+I
t-1
t-1
It follows by induction that this procedure generates a unique sequence of integers
C1,C2,...,Ck_1 E (0, 1,..., T - 1) such that

k-1
1:(C, + De)T'
f-1
k-1
E, T' + 1kTk.
f-1
Since Q. - 0 and Ck_1 determines Dk, we have

k
k-I
0 5 1:(Ce + Dt)Tt
f-I
Et T' < E'Tk,
El T' + (Dk + /k)Tk

t-1
f-1
92
3.
where Dk + /k - Ek. Since

k-I
0 < 1: EITI < Tk,

f-1
it follows that
0<Ek <E'
and
k-I
EIT' + E*Tk < (1 + E')Tk < 2E*Tk.
(3.13)
f-1
Recall that
k
EITI
f-1
- j:(Cf + DI)Tt - >2(k).

f-1
Since E' depends only on k and not on T, it follows that
(E' - Ek)Tk - >2(k),

and so
k-1
1: E, T' + E* T' - >2(k)
(3.14)
f-1
for every (k - 1)-tuple (E1, ... , Ek _ 1) of integers Et E 10, 1, ... , T - 1). Choose
the integer To > 5E' so that
4(T + 1)k < 5Tk
for all T > To.
We shall prove that if T > To and if (FO, F1, ..., Fk_1) is any k-tuple of integers
in 10, 1,..., T - 11, then
Fo+FlT+...+FA_1Tk-1+4E'Tk->2(k).
W e use the following trick. Let Eo E (0, 1, ... , T - 1). Applying (3.13) with T + 1
in place of T, we obtain
Eo(T + 1) + E*(T + 1)k < (T + 1)2 + E*(T + 1)k
(1+E*)(T+1)k
2E'(T + 1)k.
(3.15)
Applying (3.14) with T + 1 in place of T, we obtain
Eo(T+1)+E*(T+1)k->2(k).
(3.16)
Adding equations (3.14) and (3.16), we see that for every choice of k integers
Eo, El,.. , Ek-, E
10, 1, ...
,T-
1),
3.3
93
we have
F*-(EST+ +Ek_ITk-I + E*Tk)+(Eo(T+1)+E`(T+1)k)

k-1
-(Eo+E*)+(E,+Eo+kE*)T+E(Ee+()E')Te+2E*Tk
1-2
- E(k).
Moreover, it follows from (3.13) and (3.15) that
0 < F' < 4E*(T + 1)4 < 5E*Tk < Tk+i

since 4(T + 1)k < 5Tk and T > To > 5E*. Given any k integers
Fo, F1,..., Fk_I E {0, 1,...,T- 1},

we can again apply our algorithm (see Exercise 7) to obtain integers Fk and
Eo,El,E2,...,Ek_i E {0, 1,...,T- 1)

such that
Fo+ FIT +... + Fk_lTk-i + FkTk
E(k),
where Ft is an integer that satisfies
0<Fk<5E`.
After the addition of (5E* - Fk)Tk - E(k), we obtain
Fo+FIT
+5E`Tk ->(k)
for all T > To and for all choices of FO, F1, ... , Fk _ i E (0, 1, ... , T - 1). This
proves that n - F(k) if T > To and
5E'Tk < n < (5E' + 1)Tk.

There exists an integer T> > To such that
5E'(T + 1)k < (5E` + 1)Tk
for all T > T3.
Then n - E(k) if T > T, and
5E*TA <n <5E'(T+I)k.
(3.17)
Since every integer n > 5E'Tl satisfies inequality (3.17) for some T > T1, we
have
n-> (k) foralln>5E*T,.
It follows from Lemma 3.9 that Waring's problem holds for exponent k. This
completes the proof of the Hilbert-Waring theorem.
94
3.
3.4
Notes
The polynomial identities in Theorems 3.1, 3.2, and 3.3 are due to l.iouville [79,
pages 112-115], Fleck [40], and Hurwitz [65], respectively. Hurwitz's observations [65] on polynomial identities appeared in 1908.
Hilbert [56] published his proof of Waring's problem in 1909 in a paper dedicated to the memory of Minkowski. The original proof was quickly simplified
by several authors. The proof of Hilbert's identity given in this book is due to
Hausdorff [52], and the inductive argument that allows us to go from exponent k
to exponent k+ 1 is due to Stridsberg [120]. Oppenheim [941 contains an excellent
account of the Hausdorff-Stridsberg proof of Hilbert's theorem. Schmidt [ 105]
introduced a convexity argument to prove Hilbert's identity. This is the argument
that Ellison [28] uses in his excellent survey paper on Waring's problem. Dress [25]
gives a different proof of the Hilbert-Waring theorem that involves a clever application of the easier Waring's problem to avoid induction on the exponent k.
Rieger [102] used Hilbert's method to obtain explicit estimates for g(k).
3.5
Exercises
1. (Euler) Let [x] denote the integer part of x, and let
q-
[(3)']
Prove that
g(k) > 2k + q - 2.
Hint: Consider the number N - q2* - 1.
2. Verify the polynomial identity in Theorem 3.2, and obtain an explicit upper
bound for g(6).
3. Verify the polynomial identity in Theorem 3.3, and obtain an explicit upper
bound for g(8).
4. (Schur) Verify the polynomial identity

22, 680(x1 + zz + x3 +x4)5
-9(2x;)10+180(x; xj)'0+E(2x1 xJ xk)lo

+91(xl x2X3X4)10.
5. Show that every integer of the form 22, 680a5 is the sum of 2316 nonnegative
integral 10th powers.
3.5
Exercises
95
6. Let k, f, and U be integers such that 0 < f < k. Let
f(x) - (x2 +
Show that there exist positive integers A0, A 1, ... , A depending only on k
and t such that
d2c
f -37 Aix2'(x2+U)''.
dx2t
;-o
7. Let k > 1 , T > 2, and D; , E; be integers for i - 0, 1, ... , k - 1. Prove that

there exist unique integers CO, ... , Ck-1 and Ik such that
0<C; <T fori-0,1,...,k-1

and
k-I
k-I
EcT' + IkT'.
1:(Cc + Dc)Tc c-o
<-o
8. This is an exercise in notation: Prove that E(2k) - E(k) but F(k) 71 E(2k).
4
Weyl's inequality
The analytic method of Hardy and Littlewood (sometimes called the

`circle method') was developed for the treatment of additive problems
in the theory of numbers. These are problems which concern the representation of a large number as a sum of numbers of some specified
type. The number of summands may be either fixed or unrestricted; in
the latter case we speak of partition problems. The most famous additive problem is Waring's Problem, where the specified numbers are
kth powers .... The most important single tool for the investigation
of Waring's Problem, and indeed many other problems in the analytic
theory of numbers, is Weyl's inequality.
H. Davenport [18]
4.1
Tools
The purpose of this chapter is to develop some analytical tools that will be needed
to prove the Hardy-Littlewood asymptotic formula for Waring's problem and other
results in additive number theory. The most important of these tools are two inequalities for exponential sums, Weyl's inequality and Hua's lemma. We shall also
introduce partial summation, infinite products, and Euler products.
We begin with the following simple result about approximating real numbers
by rationals with small denominators. Recall that [x] denotes the integer part of
the real number x and that Ix) denotes the fractional part of x.
98
4.
Weyl's inequality
Theorem 4.1 (Dirichlet) Let a and Q be real numbers, Q > 1. There exist
integers a and q such that
1 <q
(a,q)= 1
Q,
and
a--a
Q
Proof. Let N = [Q]. Suppose that (qa) E [0, 1/(N + 1)) for some positive
integer q < N. If a - [qa], then
q
0 < (qa) - qa - [qa] - qa - a <

and so
a--aq
<
-1
<
q(N + 1)
<
N+1'
1
q2.
qQ
Similarly, if {qa} E [N/(N + 1), 1) for some positive integer q < N and if
a-[qa]+1,then
N+1
<
(qa)-qa-a+1<I
implies that
Iqa - al <
and so
a--a
q
N+1
1
q(N+1) < qQ - q2
If
(qa) E
11
LN+1'N+1
for all q - 1, ... , N, then each of the N real numbers {qa) lies in one of the N - 1
intervals
i
i+1
1 ' N 1)
fori - 1.... N - 1.
By Dirichlet's box principle, there exist integers i E [ 1, N -1 ] and qj, q2 E [ 1, N]

such that
I <q, <q2<N
and
i+1
{qia},{q2a}E IN
Let
q - 92 - qj E [l, N - I]
and
a = [q2a] - [q1a].
4.2
99
Then
1
14a - al - I(92a - [92a]) - (4ia - [9ia1)I - I(92a} - (qia}I <
N+1<Q
4.2
The forward difference operator Ad is the linear operator defined on functions f

by the formula
Ad MW a f(x+d) - f(x).
For e > 2, we define the iterated difference operator Ad,.d,_,.....d, by
Od,.d,-,.....d, - 1 d, o Ade-i.....d, t dt 0 Ad2_, o ... p Ode,
For example,
Ad2.d, (f)(X) - Ad, (Ad, (1)) (x)
- (Ads W) (X + d2) - (Ads (f )) (X )
f(x+d2+d1)- f(x+d2)- f(x+d1)+ f(x)

and
Ad3.d2.d,(f)(x) a f(x + d3 + d2 +di) - f(x + d3 + d2)
- f(x +d3 +d1) - f(x +d2 +di)
+f(x+d3)+ f(x+d2)+ f(x+d1)- f(x).

with di - 1 for i s 1, ... , e.
We let Apt) be the iterated difference operator

Then
0(2)(f )(x) - f(x + 2) - 2f(x+ 1) + f(x)

and
0(3)(f)(x)- f (x + 3) - 3f (x + 2) + 3f (x + 1) - f(x).
Lemma 4.1 Let e > 1. Then
L(t )(f)(x) - E(-1)f-j \e/

j-o
f (x +
J)
Proof. This is by induction on e. If the lemma holds for e, then
(X)
100
4.
=o
Weyl's inequality
t(-I)f-j(t)f(x+j)
i-0
_ E(- 1)(-j
Au )(X + j)
j-0
)-o
1je,i i( t
f(x+
)-t
(x+j)
i-0
j)+E(-1)e+i-i(e)f(x+j)
i-0
f(x+e+1)+E(-i)f+i
i-1
((J - 1 )+
\e))f(x+j)+(-1)f+t

We shall compute the polynomial obtained by applying an iterated difference
operator to the power function f (x) - xk.
Lemma 4.2 Let k > I and I < e < k. Let' d.....d, be an iterated dif erem'ce
operator. Then
Ad,....,d,k (x ) = L
diJ ... dl)e x
(4,1)
= dl ... dePk-f(x),
where pk_f(x) is a polynomial of degree k - e and leading rotilicienr k(k 1)

(k - e + 1). If di, ... , di are integers, then pk_e(x) is a pohoomial itirh
integer coefficients.
Proof. This is by induction on t. Fore = 1, we have

(x + di )k - xk
Ad, (xk)
k-1
E ()1'-ixi
k! d),x).
j!jl!
Let I < e < k - 1, and assume that formula (4.1) holds fore. Then
k
Od,.).d,.....d,(x )
Ad,., (Ad,.....d, (x"))
k!
m!Jtt...Je! d)i
.di),Da,.,(xm)
.
4.2
kl
m! j1
!...j
m!
djl...dj(
r
j!je+I!
101
dj(.IXj
e+1
m./I.. .I(>_I
k!
djl ...d!jtdj,.lxj
+I
1
j!j1! ..je!j+I!
dj(dl('Xi
djl
k!
r( j!j1I ... j!je+1 !

Since the multinomial coefficients k!/j! jl! . . . je! are integers, it follows that if
dl, ... , d, are integers, then the polynomial pA-(x) has integer coefficients. This
Lemma 43 Let k > 2. Then

Ad,_,.....d1(X ) = dI ...dk_lk!
d1
x+
Proof. This follows immediately from Lemma 4.2.
Lemma 4.4 Let f > 1 and td,.d,_I..... d1 be an iterated difference operator. Let
f (x) - ax' +
be a polynomial of degree k. Then
Ad...... dl (f)(X) = dl ... d (k(k - 1) ... (k - e +
ifl
1)axA-e + ...
<k and
Od,.d(-I.....dl (J )(X) = 0
if e > k. In particular, if e = k - I and d 1 ... dA -I (0, then
Ad,_,.....d,(f)(X)=dI ...dA_1Max+P
is a polynomial of degree one.
Proof. Let f (x) - Fi_I aix', where aA - a. Since the difference operator A
is linear, it follows that
A
Odr.....dl(f)(X) = E
d,.....d,(X
j-0
dl d
kl
(k
e)!
ax A- +
Lemma 4.5 Let 1 < .t < k. If
-P<dl,...,d,x<P,
then
Ad...... dl(XA) << PA,
where the implied constant depends only on k.
102
4.
Weyl's inequality
Proof. It follows from Lemma 4.2 that

k!
LAdt..... d,(X )I <
pil+...+h+1
J!Jtt... je!
<
k!
-,
Pk
i!J'!... Jf!
(e+1)kpk
(k+ 1)kPk
Pk.
4.3
Easier Waring's problem
Here is a simple application of difference operators.
Waring's problem states that very nonnegative integer can he written as the
sum of a bounded number of nonnegative kth powers. We can ask the following
similar question: Is it true that every integer can he written as the sutra or difference
of a bounded number of kth powers? If the answer is " yes. " then for every k there
exists a smallest integer v(k) such that the equation
n - fxj f x2 ... f xv(k)
(3.2)
has a solution in integers for every integer n. This is called the easier Wuriris,''.'
problem, and it is, indeed, much easier to prove the existence of r(k) than to prove
the existence of g(k). It is still an unsolved problem, however, to determine the
exact value of v(k) for any k > 3.
Theorem 4.2 (Easier Waring's problem) Let k > 2. Then v(k) exists, and
t
v(k) < 2k-t +
Proof. Applying the (k - 1)-st forward difference operator to the polynomial

f (.r) - xk, we obtain from Lemma 4.1 and Lemma 4.3 that
k-1
(-1)k-t_t fk - 1)(x
A('-t)(xk) - k!x +m -
+ e)k,
E-0
where m - (k - 1 ) !('2 ) . In this way, every integer of the form k!x +m can be written
as the sum or difference of at most
-2
( "'
4.4
Fractional pans
103
kth powers of integers. For any integer n, we can choose integers q and r such that
n-m=k!q+r,
where
k!
--k!2 < r <
- 2
Since r is the sum or difference of exactly Ir I kth powers 1k, it follows that n can be
written as the sum of at most 2k-I +k!/2 integers of the form xk. This completes
the proof.
4.4
Fractional parts
Let [a] denote the integer part of the real number a and let (a ) denote the fractional
part of a. Then [a] E Z, {a} E [0, 1), and
a - [a] + {a}.
The distance from the real number a to the nearest integer is denoted
Ila11 - min (In - aI : n E Z) = inf({a}, 1 - {a}).

Then Ila II E [0, 1/21, and
a - nIlall
for some integer n. It follows that
Isin ral =sinnr}ahl

for all real numbers a. The triangle inequality
Ila+fill -< Ila1I+IIf1I
(4.3)
holds for all real numbers a and P (see Exercise 2).

The following two very simple lemmas are at the core of Weyl's inequality for
exponential sums, and Weyl's inequality, in turn, is at the core of our application
of the circle method to Waring's problem. Recall that exp(t) - e' and e(r) _
exp(27rit) - e2"".
Lemma 4.6 If 0 < a < 1/2, then
2a < sin ra < ra.
Proof. Let s(a) - sin ra - 2a. Then s(0) - s(1/2) - 0. If s(a) - 0 for some
a E (0, 1/2), then s'(a) - Jr cos ra - 2 would have at least two zeros in (0, 1/2),
which is impossible because s'(a) decreases monotonically from r - 2 to -2 in
this interval. Since s(1/4) - (J - 1)/2 > 0, it follows that s(a) > 0 for all
a E (0, r/2). This gives the lower bound. The proof of the upper bound is similar.
104
4.
Weyl's inequality
Lemma 4.7 For every real number a and all integers NI <
N2,
N2
L e(an) << min(N2 - N1, Ilall-1).

n-N,+1
Proof. Since le(an)I - 1 for all integers n, we have

N2
N'
L I-N2-N1.
L e(an)
n-N,+l
n-N,+I
If a ' Z, then Ila II > 0 and e(a) f 1. Since the sum is also a geometric progression,
we have
N2-N1-I
N1
L e(an)
e(a(N1 + 1)) E e(a)n
n-N,+1
n-0
e(a(N2 - NI)) - I
e(a) - 1
2
- le(a) - 11
2
le(a/2) - e(-a/2)I
2
12i sinnal
1
I sin ira I
1
sin(7r Ila II)

1
- 211a11
Lemma 4.8 Let a be a real number, and let q and a be integers such that q > 1
and (a, q) - 1.1f

a
a-q
42'
then
1
E
r q/2 Ilarll
Proof. The lemma holds for q - 1,
I<r<q/2
<< q log q.
4.4
Fractional parts
105
Therefore, we can assume that q > 2. For each integer r, there exist integers
s(r) E [0, q/21 and m(r) such that
s(r)
ar
ar - m(r)
Since (a, q) - 1, it follows that s(r) - 0 if and only if r - 0 (mod q), and so
s(r) E [1, q/2] if r E [1, q/2]. Let
a----,
a
q2
where -I < 0 < 1. Then
ar
ar 0'
ar--+---+-,
Or
q2
2q
where
20r
< 101 < 1.
Ilarll -
ar
"'
q +2g11
m(r)
s(r)
0'
+ 2q
29
Ilsr)
q
s(r)
II
s(r)
0'
II
1
2q
2q
2q
Let I < r, < r2 < q/2. We shall show that s(r,) - s(r2) if and only if r, - r2. If
art
are
then
q1 - m(r,) I -
q2
- m(r2))
and so
ar, - fare
(mod q).
106
Weyl's inequality
4.
Since (a, q) - I and 1 < r, < r2
q/2, we have
r2
r1
(mod q)
and so
r, - r2.
It follows that
or
1<r<92
q -1<r<g
-2
s(r)
1<s<q2}
Therefore,
<
_I
s(r)
1
2q
q log q.
Lemma 4.9 Let a be a real number. If

a
q2
where q > I and (a, q) - 1, then for any nonnegative real number V and
nonnegative integer h, we have
q
r-l
Proof. Let
(
min(V
Ila(h9+r)III
--,
a- a +
q
where
<< V+glogq.
q2
-1<0<1.
Then
a(hq+r)-ah+ar+Bh+9r
ah +
ar +
q2
q
[Oh] + (Oh) + Or
q
q
-ah+ar+[Oh]+S(r)
q
q2
4.4
where
Fractional parts
107
Or
-1 < S(r) - {Bh} + - < 2.

q
For each r - 1, ... , q there is a unique integer r' such that

{a(hq + r)} -
ar + [Oh] + S(r)
-r
q
Let
0<t<l--.q
If
t < {a(hq + r)} < t + 1 ,

q
then
qt <ar-qr'+[Oh]+S(r) <qt+1.
This implies that
ar-qr'<qt-[Oh]+]-S(r)<qt-[8h]+2
and
ar - qr' > qt - [Oh] - S(r) > qt - [Oh] - 2.

Thus, ar - qr' lies in the half-open interval J of length 4, where
J - (qt - [Oh] - 2, qt - [Oh] + 2].

This interval contains exactly four distinct integers. If I < r, < r2 < q and
ar, - qr, - are - qr2,

then
ar, = are (mod q).

Since (a, q) - 1, we have
r,=r2
(modq)
and so
r, - r2.
It follows that for any t E [0, (q - 1)/q ], there are at most four integers r E [1, q l
such that
{a(hq + r)} E [t, t + (1 /q)].
We observe that
Ila(hq +r)II E [t, t +(1/q)]

if and only if either
{a(hq +r)} E [t, t +(1/q)]
108
4.
Weyl's inequality
or
1 - {a(hq + r)} E [t, t + (1 /q)].

The latter inclusion is equivalent to
{a(hq + r)} E [t', t' +(I 1q)],

where
0<t'=I- q--t<1- q
1
It follows that for any t E [0, (q - 1)/q ], there are at most eight integers r E [1, q]
for which
IIa(hq + r)II E [t, t +(I /g)].
In particular, if we let J(s) _ [s/q, (s + 1)/q] for s - 0, 1, ..., then

11a(hq + r)II E J(s)
for at most eight r E [1, q].

We apply this fact to estimate the sum
E min
V,
1<r<q
Ila(h4+r)II
If 11a(hq +r)II E J(0) - [0, 1/q], then we use the inequality

min
(V,
IIa(hq +r)JI
<
If 11a(hq +r)II E J(s) for some s > 1, then we use the inequality
min
V,
IIa(hq + r)II
11a(hq + r)II
- qs
Since IIa(hq + r)II E J(s) for some s < q/2, it follows that
1: min
1<r<q
V,
< 8V+8
Ila(hq+'r)lI
q
s
1<a</2 s
<< V +q log q.
a--a <-,
q2
I
where q > 1 and (a, q) - 1, then for any real number U > I and positive integer
n we have
min(kllaklll (q+U+q)log2gU.
1 <k<U
4.4
Fractional parts
109
Proof. We can write k in the form
k - hq+r,
where
1 <r<q
and
0<h<U.
q
Then
n
S - E min
k, Ilakll I
I<k 1 and
hq+r>hq>
(h + 1)q
orh-0,q/2<r<q, and
hq+r - r>
2
(h+1)q
Therefore,
S<<glogq+ 1:
1:
nun
0<h<U/q I<r<q
Note that
I
(4.4)
(h + 1)q' IIa(hq +r)II
-+1 <U+q <2max(q,U)<2qU.
q
Estimating the inner sum by Lemma 4.9 with V - n/(h + 1)q, we obtain
S<<qlogq+ E E min
05h<Ul9I<r<q
(h+1)q' IIa(hq+r)II /
1 10
4.
Weyl's inequality
<< glogq+ E (+" l)q +giogq I

0<h<U/q
<<
glogq+9
0<h</U/q
h+1
\ q +l Iglogq
JJJ
<<
glogq+glog(9 +11 +Ulogq+qlogq
<<
(n +U+q'log2gU.
9
///
--aq <-,
I
qZ
where q > I and (a, q) - 1, then for any real numbers U and n we have
min
(n,
IlakII)/
1<k<U
(q+U+n+ qn
lmaz{l,logq}.
////
Proof. This is almost exactly the same as the proof of Lemma 4.10. We have
S-
1<k<U
min ('i' Ilakll /
min n,
0<h<U/q I<r<q
glogq+
Ila(hq +r)II
9
1: (n+ 1<s<q/2 S
0<h<U/q
<<qlogq+ E (n+qlogq)
0<h <U/q
<<
glogq+(U +1)(n+glogq)
glogq+Ulogq+n+Un
<<
q
<<
(q+U+n+Un
lmax{1,logq}.
q //
4.5
4.5
111
In this section, we denote by [M, N] the interval of integers m such that M < m <
N. For any real number t, the complex conjugate of e(t) - e2' r is e(t) - e(-t).
Lemma 4.12 Let N1, N2, and N be integers such that NI < N2 and 0 < N2 NI < N. Let f (n) be a real-valued arithmetic function, and let
N2
S(f) - L e(f (n))

n-N, +1
Then
IS(f)12 =
Sd(f)
Idk<N
where
e(Od(f)(n))
Sd(f) nEl(d)
and I (d) is an interval of consecutive integers contained in [N1 + 1, N2].
Proof. For any integer d, let
1(d)-[NI+I -d, N2 -d]n[NI+1,N2].

Squaring the absolute value of the exponential sum, we get
IS(f)12 S(f)S(f )
N2
Ns
L e(f (m)) L e(f (n))

n-N,+1
m-N,+1
N2
N2
1: 1: e(f (m) - f (n))

n-N1+1 m-N1+1
N2
2 Ni: -E
e(f(n+d)- f(n))
n-N,+1 d-N,+l -n
N2-n
N,
1:
1:
e(Ad(f)(n))
n-N,+1 d-Nt+l -n
N2-N1-I
1:
1: e(Ad(f)(n))
d--(N2-N, -1) nE/(d)
e(od(f)(n))
Jdl<- N nER(d)
Sd(f )
Idl<.N
112
4.
Weyl's inequality
Lemma 4.13 Let Nt, N2, N, and a be integers such that t > 1..V .
0 < N2 - N, < N. Let f (n) be a real-valued arithmetic fiinction, and let
and
N2
S(f)-
e(f(n)).
Then
IS(f)12` <
... 1: Sd...... d,(A
(2N)2'-f-1 E
Idrl <N
Id, l<N
where
(-I.til
e ('.d,.....d,(f)(n))
Sd,..... d,(f) ,rEl(d,....,d1)
and I (d f, ... , d l) is an interval of consecutive integers contained in IA. - I .:ti : I.
Proof. This is by induction on e. The case e - 1 is Lemma 4.12. Now a.sumc

that the result is true fore > 1. Using the Cauchy-Schwarz inequality, we obtain
Is(f)12,.,
({s(f)12`) z
2
(2N)2'-t-t
<
...
ISd...... d,(f)I
Id,l<N
Idil<N
)
2
2N
Id,I<N
(2N)2,., _21-2(2N)t
(f)I
Id,I<N
5' ... Y, I Sd,....d,(f )12,

d,I<N
Id,l<N
where Sd,..... d, (f) is an exponential sum of the form (4.5). By Lemma 4. 12. ttir
, ..., d f, there is an interval
each d)
c [N(+1.N2]
1(dt+I,dt,-..,d1)c
such that
ISd...... d, (f)
12
e (Od,.....d, (f)(n))
nE!(di....,dt)
e (Od,.,.d...... d, (f )(n))
Id,., I <N nE l(d,., .d,.....d, )
- E Sd,.i.d,.....d,(f),
and so
IS(f)12,., < (2N)2'.i_(t+))-1
F .., r
Id, I<N
Id,kN Id,.,l<N
Sd,.,.d,..... d,(f)
4.5
Lemma 4.14 Let k > 1. K - 2k-1, and e > 0. Let f(x) - axk +
113
be a
polynomial of degree k with real coefficients. If

N
S(f) -
e(.f (n)),
n_I
then
k!N'-'
NK-k+F
IS(f)IK << NK-I +
min (N, I1mall-1)
where the implied constant depends on k and E.
Proof. Applying Lemma 4.13 with f = k - 1. we obtain
IS(f)IK < (2N)A-k > ...
ISd,
....
d,(
Id, 1I<N
Id,l<N
where
e (Ad,. .....d,(f)(n))
Sd,_,.....d,(.f)
and I(dk_I...., d1) is an interval of integers contained in (1, NJ. Since Ie(t)I - I
for all real t, we have the upper bound
ISd,-......d,(f)I
Ie (Ad,..,....d,(f)(n)) I < N.
By Lemma 4.4, for any nonzero integers d1, ... , dA _ 1, the difference operator
Ad,_1.....d, applied to the polynomial f(x) of degree k produces the linear polynomial
Ad...... d,-,(.f)(x) - dk-I ... dlk!ax +0 _ Ax +.8,

where
X - dk_1...dlk!a
and0ER.Letl(dk_),...,d1)-[NI+I,N2].ByLemma 4.7,
e (td._,.d,_;..... d,(f)(n))
N,
E e(An +,0)
n-N,+1
N2
E e(An)
n-N,+I
I
IIhII
_
IldA-I ... dlk!all
114
4.
Weyl's inequality
It follows that
...dk-Ik!all-').
ISd,-......dJf)I < min(N, IIdI

Therefore,
...
(2N)K-1 E
IS(f)IK <
lSd,- .....djf)l
Id,_,I<N
Id1l<N
< (2N)K-k T ... E min(N, IIdI ...dk-Ik!all-').

Id,-,I<N
Idjl<N
Since there are fewer than (k -
choices of dl, ... , dk - I such that

dl . . . dk-I - 0, and each such choice contributes N to the sum, it follows that
1)(2N)'-2
IS(f)IK < (2N)K-k(k - 1)(2N)1-2N

+(2N )K-1
E ... E min(N, IIdI ...dk-Ik!all-1)

I<Id,-,I<N
I<Idil<N
k(2N)K-1
+2k-1NK-k ` ...
min(N, IIdI ...dk-Ik!aII-')

I<d,<N
1<di<N
N
...
<< NK-1 + NK-k
min (N, Ildl .. dk-Ik!all-1) ,

d,-i-1
di-1
where the implied constant depends only on k. Since
I < d1 ... dk_lk! < k!Nk-1
and the divisor function r(m) satisfies r(m) << m' for every e > 0, it follows
that the number of representations of an integer m in the form d1 . . . dk-Ik! is
<< m` << N`. Therefore,
N
IS(f)IK << NK-I +NK-kr... E min (N,lldk-l...dlk!all-')

dA_i-1
d;-1-1
k!N`-'
<< NK-1 +
NK-k+F
E min (N, Ilmair-I),

m-1
where the implied constant depends on k and E. This completes the proof.
Theorem 4.3 (Weyl's inequality) Let f (x) - axk + . . be a polynomial of degree

k > 2 with real coefficients, and suppose that a has the rational approximation
a/q such that
a-a
<
q2
4.5
115
where q ? I and (a, q) - 1. Let

N
S(f) - E e(f(n))
nil
Let K - 2k-1 and e > 0. Then

(N-' +q-l +N-Aq)l"x
S(f) << NI+E
where the implied constant depends on k and E.
Proof. Since I S(f )I < N, the result is immediate if q > Nk. Thus, we can
assume that
1 <q <Nk,
and so
log q << log N << NE.
By Lemma 4.14, we have

k!NR''
+Nx-k+E r`min (N, I1mall-1).
IS(f)IK << NK-I
M-I

k!N;-'
min (N, Ilmall-1) <<
(q
+k!Nk-1
+N+
k!Nk
m-1
<< q+Nk-'+
max{1, log q)
111
logN
<< Nk(gN-k+N-'+q-1)NE.
Therefore,
IS(f)IK << NK-1 + NK+E (qN-k

<< NK+E
+ N-' +q-1)
(qN-k + N-' +q-') .
Theorem 4.4 Let k > 2, and let a/q be a rational number with q > 1 and
(a, q) - 1. Then
q
S(q, a) -
e(axk/q) << ql-l(x+r
x-1
Proof. Apply Weyl's inequality with f (x) - axk/q and N - q. We obtain

S(q, a) <<
q'+E(q-l +q-k+I)IIK q1-l/K+E
116
4.
Weyl's inequality
Theorem 4.5 Let k > 2. There exists S > 0 with the following property: If N > 2
and a/q is a rational number such that (a, q) - 1 and
NI /2 < q < Nk-I/2,

then
N
e(anklq) << N1-6Proof. Applying Weyl's inequality with f (x) - axk /q, we obtain
(N-I +q-I + N-kq)I/K
S(f) << NI+e
< N]+e (N-1 + N-]/2 +N- 1/2 IIK
)
< NI-I/2K+e
< N'-6
for any S < 1/2K. This completes the proof.
Theorem 4.6 (Hua's lemma) For k > 2, let

N
T(a) - E e(ank).
n-I
Then
da N2'-k+e.
IT(a)12t
Proof. We shall prove by induction on j that

I
IT(a)12' da <<
N21-i+f
Jo
for j - 1, ... , k. The case j - 1 is clear since
J IT(a)Ida -
e(a(mk
M-I n-I
nk))da - N.
Let 1 < j < k - 1, and assume that the result holds for j. Let f(x) - axk. By
Lemma 4.2,
Ad...... di(f)(x) - adi ...dIPk-i(x),

where pk_i(x) is a polynomial of degree k - j with integer coefficients. Applying
Lemma 4.13 with NI - 0, N2 - N, and S(f) - T (a), we obtain

<(2N)2'-i-I r
IT(a)1221
Id11<N
(2N)2'-i-I 1`
Id1l<N
...
r r e(Ad,.....di(f)(n))
Id,I<NnE!(d......d1)
... 1:
1:
Id3l<NnE!(d,.....di)
e (adi ... d] Pk-i(n)).
4.5
where I (dj ,
follows that
117
..., d j) is an interval of consecutive integers contained in [ 1, N]. It

IT(a)12,
N2'-'-I > r(d)e(ad),
<
(4.6)
where r(d) is the number of factorizations of d in the form
d-dj...dIpk-j(n)
with Idi I < N and n E I (dj , ..., dl ). Since d << Nk by Lemma 4.5, we have
r(d) << IdlE << NE
ford 0. Since pk_j(X) is a polynomial of degree k - j ? 1, there are at most

k - j integers x such that pk-j - 0, and so
r(0) << Ni.
Similarly, since
IT(a)I21 - T(a)2'
T(-a)2'
N
(e(_axk))k_I
(e(ayL))k_l
r E E ... E e a Ej-IXI - Ej-Iyk

((
N
...
XI-1
i-1
X'-1-1 YI-1
- 1: s(d)e(-ad ),
d
where s(d) is the number of representations of d in the form

j-I
d -
j-I
yk
i-I
XA,
i-i
with 1 <xi,yi < Nfori - 1,...,j - 1. Then
Es(d)-
IT(0)I2,
- N2,
and, by the induction hypothesis,

I
s(0)-
IT(a)12'da
0

I
f IT(a)121*' At - f IT(a)I2' IT(a)12' At

0
N2,-'+E.
4.
118
Weyl's inequality
< N2''-r
N2''
t
r(d')e(ad') Es(d)e(-ad)da
J d
d
' 1: r(d)s(d)
N2'-'-tr(0)s(0) +
N2'-'-1 1: r(d)s(d)
dy0
<<
N2,-i-tNiN21 -i+E +N21 -i-1 NE 1:s(d)

d,o
'-(i+l)+E +
<< N2`*
N2' -i - t
NE N2,
<< N2" -ti+t)+E

4.6
Notes
The material in this chapter is well-known. For the original proofs of \Veyl'.,,
inequality and Hua's lernrna, see 1Veyl [ 1411 and Hua 1621, respectively. Davenport [181,Schmidt [ 106], and Vaughan 11251 are standard and excellent introductions to the circle method in additive number theory.
The easier Waring's problem was introduced by Wright 11501.
4.7
Exercises
1. Prove that
IIxII-II-x11-IIn+x11
for all x E R and n E Z. Let (x) denote the fractional part of x. Graph
f(x) _ (x) + IIxII for0 < x < I.
2. Prove that
Ila+f1I s IlaII+IIxII
for all a, , E R.
3. Let f > 1, and let De denote the iterated difference operator 0 r, r,..,, 1. Prove
that
De(f)(x) _ E(-1)r-j
j-0
\!/ f (x +j).
4. Let Ad ,.._.d, be an iterated difference operator. Find a general formula to

express 1 d......d, (f)(x)
4.7
Exercises
119
5. Let e > 2, let a be a permutation of (1, 2, ... , Z}, and let Ad...... d, be an
iterated difference operator. Prove that
161d,u)..... d,n, ` Ad...... d1.
The Hardy-Littlewood asymptotic

formula
... using essentially the same techniques as Hardy and Littlewood's

but in a different way and introducing certain additional considerations, we shall derive the same result with incomparable brevity and
simplicity.
I. M. Vinogradov [131)
5.1
The circle method
For any positive integers k and s, let rk,,(N) denote the number of representations
of N as the sum of s positive kth powers, that is, the number of s-tuples (x1, ..., x,)
of positive integers such that
N=x
Waring's problem is to prove that every nonnegative integer is the sum of a bounded
number of kth powers. Since 1 = Ik is a kth power, this is equivalent to showing

that
rk.,(N) > 0
for some s and for all sufficiently large integers N. Hilbert gave the first proof of
Waring's problem in 1909. Ten years later, Hardy and Littlewood succeeded in
finding a beautiful asymptotic formula for rk., (N). They proved that for s > so(k),
122
5.
The Hardy-Littlewood asymptotic formula
there exists S - d(s, k) > 0 such that
rk.,(N) - 6(N)l' (i
N(.,1k)-1
kl r (S}
+O(N(`lk)-t-a),
(5.1)
where r(x) is the Gamma function and 6(N) is the "singular series," an arithmetic function that is uniformly bounded above and below by positive constants
depending only on k and s. We shall prove that the asymptotic formula (5.1) holds
for so(k) - 2k + 1.
Hardy and Littlewood used the "circle method" to obtain their result. The idea
at the heart of the circle method is simple. Let A be any set of nonnegative integers.
The generating function for A is
f(z)-Eza
aEA
We can consider f (z) either as a formal power series in z or as the Taylor series
of an analytic function that converges in the open unit disc I z I < 1. In both cases,
Oc
f (z), - E rA.s(N)zN,
IV-o
where rA.,(N) is the number of representations of N as the sum of s elements of

A, that is, the number of solutions of the equation
with
at, a2, ... aS E A.

By Cauchy's theorem, we can recover rA., (N) by integration:
rA..,(N) -
2ni
fN+tj
dz
for any p E (0, 1).

This is the original form of the "circle method" introduced by Hardy, Littlewood,
and Ramanujan in 1918-20. They evaluated the integral by dividing the circle of
integration into two disjoint sets, the "major arcs" and the "minor arcs." In the
classical applications to Waring's problem, the integral over the minor arcs is
negligible, and the integral over the major arcs provides the main term in the
estimate for rA.s(N).
Vinogradov greatly simplified and improved the circle method. He observed
that in order to study rA,s(N), it is possible to replace the power series f (z) with
the polynomial
P(z) - : za.
ncA
c<Y
5.1
The circle method
123
Then
sN
P(Z)' - Erq)(m)zn
M-0
where r, gy(m) is the number of representations of m as the sum of s elements of A

not exceeding N. In particular, since the elements of A are nonnegative, we have
rA gy(m) - rA,., (m) form < N and r(N)(m) - 0 form > sN. If we let
z=e(a)=e2'
then we obtain the trigonometric polynomial
F(a) - p(e(a)) -
e(aa)
and
,N
F(a)s - > rANS(m)e(ma).

,n-0
From the basic orthogonality relation for the functions e(na),

I
e(mce)e(-na)da -
I.
fo
If m - n
ifm 7( n
we obtain
r A,(N) -
F(a)è(-NN)da.
In applications, of course, the hard part is to estimate the integral.

To apply the circle method to Waring's problem, let k > 2 and A be the set of
positive kth powers. Let rk.s(N) denote the number of representations of N as the
sum of s positive kth powers. Let
P - [N Ilk].
Then
P
F(a) -
e(aa) -
e(ank)
and
rk.,(N) - J F(a)Se(-aN)da.
I
0
124
5.
5.2
Waring's problem for k = 1
For k - 1, there is an explicit formula for rt.,(N).
Theorem 5.1 Lets > 1. Then
N'-t
N-1
(N)- s-1 -(s-1)!
+ O (Ns-2)
for all positive integers N.
Proof. Let N > s. We observe that
N into s positive parts if and only if
N - s - (a, - 1)+...+(a,,- 1)
is a decomposition of N into s nonnegative parts. Therefore,
rt.,(N) - Rt.,(N - s),

where R,,,(N) denotes the number of representations of N as the sum of s nonnegative integers.
We shall give two proofs of the theorem. The first is combinatorial. We begin
by computing R,.,,(N) for every nonnegative integer N. Let N - a, + + a, be

a partition into nonnegative integers. Imagine a row of N + s - I boxes. We color
the first a, boxes red, the next box blue, the next a2 boxes red, the next box blue,
and so on. There will be exactly s - I blue boxes. Conversely, if we choose s - 1
of the N + s - 1 boxes and color them blue, and if we color the remaining N boxes
red, then we have a partition of N into s nonnegative parts as follows. Let a, be the
number of red boxes before the first blue box, a2 the number of red boxes between
the first and second blue boxes, and, in general, for j - 2, ... , s - 1, let aj be
the number of red boxes that are between the (j - 1)-st and jth blue boxes. Let
a., be the number of red boxes that come after the last blue box. This establishes a
one-to-one correspondence between the subsets of size s - I of the N +s - I boxes
and the representations of N as the sum of s nonnegative integers. Therefore, the
number of decompositions of N into s nonnegative parts is the binomial coefficient
N+a -1). It follows that
s-
N-1
r,.5(N) - R,.,(N - s) -
s-1
This gives the first proof of the theorem.

There is also a simple analytic proof. The series
1
.f (z) -
zN
N-0
I - z
The Hardy-Littlewood decomposition
5.3
125
converges for I z I < 1, and
f(z)s
RI.s(N)zN.
N-0
We also have
f(z)` -
(1 - z)5
d`
(s - 1)!
dzs-I
d`-' (ZN
dzs-"
(s - 1)!
N-0
r,
E
l\ 1 - z )
N_s+1
N
1)
N
1:(S -1
Therefore,
R1.s(N) -
IN +s - 11.
s-1 J
5.3
The Hardy-Littlewood decomposition
Fork > 2 there is no easy way to compute-or even to estimate-rk.3 (N) for large
N. It was a great achievement of Hardy and Littlewood to obtain an asymptotic
formula for rk,,(N) for all k > 2 and s > so(k). In this chapter, we shall prove the
Hardy-Littlewood asymptotic formula for s > 2k + 1. For N > 2k, let
P - [NI/k]
(5.2)
and
P
F(a) - E e(amk).
(5.3)
in-I
The trigonometric polynomial F(a) is the generating function for representing N

as the sum of kth powers. The basis of the circle method is the simple formula
fI
rk.s(N) -
J0
F(a)è(-Na)da.
(5.4)
126
5.
We cannot compute this integral explicitly in terms of elementary functions. By

carefully estimating the integral, however, we shall derive the Hardy-Littlewood
asymptotic formula.
The first step is to decompose the unit interval [0, 11 into two disjoint sets, called
the major arcs 9)I and the minor arcs in, and to evaluate the integral separately
over both sets. The major arcs will consist of all real numbers a c [0, 11 that can,
in a certain sense, be "well approximated" by rational numbers, and the minor arcs
consist of the numbers a E [0, 11 that cannot be well approximated. Although most
of the mass of the unit interval lies in the minor arcs, it will follow from Weyl's
inequality and Hua's lemma that the integral of f (a)è(-Na) over the minor arcs
is negligible. The integral over the major arcs will factor into the product of two
terms: the "singular integral" J(N) and the "singular series" 5(N). The singular
integral will be evaluated in terms of the Gamma function, and the singular series
will be estimated by elementary number theory.
The major and minor arcs are constructed as follows. Let N > 2''. Then P =
[Nt/k] > 2. Choose
0<v<1/5.
For
1 < q < P,
0<a <q,
and
(a, q) - 1,
we let
MI(q,a)CI E[0,1]:ICI -q
q
ll
and
q
U U 9X(q, a).
1 <q <P,
The interval M(q, a) is called a major arc, and 9A is the set of all major arcs. We
see that
9A(1, 0) - r0, Pk-v 1.
IJ
971(1. 1) _ L1 - Pk
and
TZ(q, a)
Lq
a
q +
pk-"
for q > 2. The major arcs consist of all real numbers a E [0, 1 that are well
approximated by rationals in the sense that they are close, within distance P`k,
to a rational number with denominator no greater than P.
]
5.4
The minor arcs
127
If a E 931(q, a) fl 931(q', a') and a/q -/ a'/q', then Iaq' - a'q I > I and
1
PZ" -qq'
Ia_a
q
q'
< a- -
2
_ PA-V'
which is impossible for P > 2 and k > 2. Therefore, the major arcs 931(q, a) are
pairwise disjoint.
The measure of the set 931(1, 0) U 931(1, 1) is 2Pv-k, and, for every q > 2 and
(a, q) - 1, the measure of the major arc 931(q, a) is 2Pv-k. For every q > 2 there
are exactly p(q) positive integers a such that 1 < a < q and (q, a) - 1. It follows
that the measure of the set 931 of major arcs is
AM) - P2 v E (p(q) < Pk

I <q<P'
<
2
Pk
Pv(Pv + 1) <
2
Eq
1 <q<P"
Pk-3v'
(S.5)
which goes to zero as P goes to infinity.

The set
m-[0, 1]\931
is called the set of minor arcs. This set is a finite union of open intervals and
consists of all a E [0, 11 that are not well approximated by rationals. The measure
of the set of minor arcs is
2
p(m) = 1 - (9f1) > 1 - Pk-3v '

Even though the measure of the set m is large in the sense that it tends to I as P
tends to infinity, we shall prove in the next section that the integral over the minor
arcs contributes only a negligible amount to rk.5(N).
5.4 The minor arcs

We shall now show that the integral over the minor arcs is small.
Theorem 5.2 Let k > 2 ands > 2k + 1. There exists 31 > 0 such that
fm
F(a)Se(-Na)da - 0
(PS-k-a)
where the implied constant depends only on k and s.
128
5.
Proof. By Dirichlet's theorem (Theorem 4.1) with Q - Pk
to every real
number a there corresponds a fraction a/q such that
I < q < Pk-v,
(a, q) - 1
and
a-qI <qpk <min(Pk-v,g2)
If a E m,thena gfit(1,0)U9)t(1, 1), so

I
<a<
Pk-v
Pk-v
and! <a <q - 1.Ifq < P', then

a
a-q
implies that
aE9R(q,a)99)t-[0,1]\m,
which is absurd. Therefore,
PV
<q<
Pk-L'.
Let
K - 2k-1
It follows from Weyl's inequality (Theorem 4.3) with f (x) - axk that
F(a) << P'+` (P-' +q-1 +

<< P1+f
<<
P-kq)'/K
(P-1 + P-v + P-k pk-v)i/K
Pl+e-v/K
Applying Hua's lemma (Theorem 4.6), we obtain
F(a)se(-na)da
1m
fm F (a)s-21 F(a)2Ae(-na)da
fm
F(a)IF(a)12da
)F(a)1,,-2,
< max
Jp
UEm
<<
1F(a)12' da
(Pl+e-v/K))i-2` P2-k+e
Ps-k-61,
where
v(s - 2k)
if e > 0 is chosen sufficiently small. This completes the proof.
5.5
5.5
The major arcs
129
The major arcs
We introduce the auxiliary functions
km'lk-'e(lm)
v($) M-1
and
q
S(q, a) - E e(ark/q)
r-I
We shall prove that if a lies in the major arc M (q, a), then F(a) is the product of
S(q, a)/q and v(a - a/q), plus a small error term. We begin by estimating these
functions.
Clearly, IS(q, a)I < q. By Weyl's inequality (Theorem 4.4), we have
S(q, a) <<
qI-I/K+F
and
S(q, a)
<< q-I/K+F
(5.7)
Lemma 5.1 If I#I < 1/2, then
I-'/k)
v($) << min(P, l
Proof. The function
f(x)- kIxllk-I
is positive, continuous, and decreasing for x > 1. By Lemma A.2, it follows that
N
km'Ik
Iv(f)1
M-I
N
<
k-'x'"k-dx+ f(1)
If I#I < I/N, then P < NIlk < 101-'/k and v(6) << min(P, 1p1-Ilk).
Suppose that I/N < I P I < 1/2. Then 101-'/k << P. Let M = [I0I-'] . Then
M< -I <M+I <N.
130
5.
-ICI-' By
Let U(t) - En .,e(fim). By Lemma 4.7, we have U(t) <<

partial summation (Theorem A.4),
N
E -m'lk-I e(fim) - f(N)U(N) - f(M)U(M) - J U(t).f'(t)dt

N
m-M+I k
MI/k-I
<<
I$1
101-Ilk
<
<< min(P,
Therefore,
m'lk-'e(fim)
v(f)
m-I k
m'l'-'e(fm)
m-M+I k
<< min(P, I$I-'/k).

Lemma 5.2 Let q and a be integers such that I
<-
q < P, 0 <- a < q, and
(a, q) - 1. If CI E TI(q, a), then
F(a) -
(S(a)) v (a -
q/
Proof. Let $ - a - a/q. Then 1,61 < P-k and
F(a) - S(q , a) v(li)

q
e(amk) -
S(q,
- >J a
m-I
a)
m-I
amk ) e(fimk)
9/
m'lk-I e(pm)
m-I k
kmilk-)e(lim)
-S(9 a)
q
ML-I
1: u(m)e(8m),
rn-I
where
e(am/q) - (S(q,
- (S(q, a)/q) k-I m'/k-I
if m is a kth power
otherwise.
a)/q)k-'mI1k-I
We shall estimate the last sum. Let y > 1. Since IS(q, a)I < q, we have
a
e(amklq) I <m<v
e(ark/q)
r-I
m..
mcx{ of
The major arcs
5.5
-S(q,a)f y+0(1))
9
(S(a)) + 0(q).
1. Since v(f6) << P, we have
U(t) _
u(m)
1 <m <r
e(amk/q)
S(q,
q
I<M<1111
= th/k
a)
m'It-t
k
1<m<r
(ick
(S(qa))+o(q)
(S(a))
O(1))
0(q).
E u(m)e(6m) - e(flN)U(N) - 2irif

m-l
0(q) - 2n i#
e(pt)U(t)dt
rN
e(j91)0(q)dt
<< q+IhINq
<< (1 + I$IN)q
<< (1 + P`'-kPk)PV
<< P2v.
Theorem 5.3 Let
E
q
15v:5Q
and
p,-1
J`(N) - j P -i
v($)se(-Ns8)d#.
Let 93Z denote the set of major arcs. Then
f9A
F(a)è(-Na)da a 6(N, P)J*(N) + 0
where 32 - 0 - 5v)/k > 0.
(PS-k-h) ,
131
132
5.
Proof. Let a E 971(q, a) and
f=a--q
a
Let
V
. V(a q a) - S(q,a)v
9
a- a
s(4 a)v(8)
q
Since IS(q, a)I < q, we have I V I << Iv(fl)I << P by Lemma 5.1. Let F Then IFI < P. Since F - V - O(P2v) by Lemma 5.2, it follows that
F` -V` _ (F-V)(F`-)+Fc-2V+...+V`-1)
<<
P21,Ps-i
- Ps-)+2v.
P3`'-4
Since (9J1) <<
by (5.5), it follows that
IP - VÌ da
P3v-kps-)+2v
_ Ps-k-s=,
19)1
where 32 - I - 5v > 0. Therefore,

J931
F(a)'e(-Na)da
V(a, q, a)è(-Na)da + 0 (P`-k-a=)
V (a, q, a)e(-Na)da + O
i<q 2, we have
V(a, q, a)se(-Na)da
1931(q.a)
f/q-P
V(a, q. a)è(-Na)da
V(f +a/q, q, a)se(-N(fl +alq))dp

P.
S(q, a) 1' e(-Na/q) J
S(q, a))'
P -A
v(P)se(-NO)dfl
e(-Na/q)J'(N).
For q - I we have V (a, 1, 0) - v(a) and V (a, 1, 1) - v(a - 1). Therefore,
J91(i.o)
V(a, q, a)è(-Na)da
V(a, q, a)è(-Na)da +
9J1(t. i )
5.6
f
- f 0P
v(a)è(-Na)da +
+
t
v(fl)è(-N8)dj
-P-4
The singular integral
133
v(a - 1)se(-Na)da
fo
P'"'
v(fi)Se(-N,8)d46
- J*(N).
Therefore,
Am
F(a)Se(-Na)da
(S(9.a)1
1<q<P.. a-I
.,ai-i
(P`-k-b,)
- CS(N, P)J`(N) + 0
5.6
Next we consider the integral

1/2
J(N) - f
v(f)è(-c8N)d fi.
1/2
This is called the singular integral for Waring's problem.
Theorem 5.4 There exists 33 > 0 such that
J(N) << P'`-k

and
J*(N) - J(N) + 0
(PS-k-b').
Proof. By Lemma 5.1,

1/2
J(N) << J
min(P, IOI-'1k)3dO
f
f
I /Iv
min(P, IfI-'/k)=dp + f
1IN
/N
1/N
1/2
Psd, + f
1/N
IN
s/kdfi
min(P, I$I-Ilk ).cd$
(5.8)
134
5.
and
J(N) - J'(N) - fp
v(P)'e(-NP)dP
" -'
1,8 1 < 1/2
P-1
1/2
Iv(fi)Isdp
<
1/2
P,."
<<
P-slkdfi
p(k-v)(c/k-1)
- ps-k-b,
where 33 - v(s/k - 1) > 0. This completes the proof.
Lemma 5.3 Let a and P be real numbers such that 0 < 0 < l and cr
N-1
E m6-1(N -
m)a-1 -
r(a)ro)
Na+9-1
+ 0 (Na-1)r(a
+ A)
M-1
where the implied constant depends only on $.
Proof. The function
XB-1(N
g(x) -
X)a-1
is positive and continuous on (0, N), integrable on [0, N], and
f
0
g(x)dx -
xo-1(N
fo
- x)a-dx
Na+Q-1 f tR-(1
- f)a-ldf
J0
Na+#-1
B(a 0)
r(a)r(p)
Net + p)
where B(a, fl) is the Beta function and r(a) is the Gamma function.
If a > 1, then
f'(x)-g(x)(0x
N-x
<0
1
and so g(x) is decreasing on (0, N)\and

N-1
f&(x)dx <
N-1
g(x)
g(x)dx.
fo
m-1
Therefore,
fNg(x)dx
0<
V-1
g(m)
m-1
S. Then
5.6
0
Na-1
If 0 < f < a < 1, then 0 < a +,6 < 2 and g(x) has a local minimum at
(I P)N
E [N/2, N).
c=
2-a-a
Since g(x) is strictly decreasing for x E (0, c), it follows that

c
[c]
L g(m) < f g(x)dx

0
m-1
and
[cl
/ IC)
E g(m) > J
M-1
g(x)dx + g([c])
fCg(x)dx
>
rc
Na-1
> J g(x)dx -
Similarly, since g(x) is increasing for x E (c, N), it follows that

N-1
g(m) <
g(x)dx
Jc
and
N-1
1N-I
g( m) >
g(x)dx + g([c] + 1)
c1+1
m-[c1+1
rN-1
>
>
g(x)dx
Jc
g(x)dx -
NP
Therefore,
N-1
0<
g(x)dx - > g(m) <

o
m-1
N6-1
N'-1
<
2Na-1
135
136
5.
Theorem 5.5 Ifs > 2, then
(k)k)
r
(-1
J(N) = r i +
Proof. Let
Ns/k-1
+0
(N(s-I)/k-1).
1/2
J,(N) - J
t/2
v(f)se(-Nfi)dfl
for s > 1. We shall compute this integral by induction on s. Since
km1/k-Ie(fim).
v(P) m-I
it follows that
N
m,-I
m,-I
v(a)s - k_, E ... 1(m1 ... ms)1/k-Ie((mI +...

and so
N
J5(N) = k-.1 F,
... E(m) ...ms)1/k-1

m,-I
MI-1
sE
Vs
E (ml ... ms) 1/k-t
.1. .,-N
1/2
e((ml +...+m, - N)$)dfi
1/2
I-, 5N
In particular, for s - 2, we apply Lemma 5.3 with a - P - 1/k and obtain

N-1
I l k - 1 1(N
J2(N) - k-2 E
m)1/k'
m-I
(1/k)2r(1/k)2
I'(2/k)
N2/k-1 + O(NI/k-1)
- r(1 + 1/k)2 N2/k-I +
O(NI/k-1)
I'(2/ k)
This proves the result in the case where s - 2.
Ifs > 2 and the theorem holds for s, then
1/2
v(fl)s,1 e(-Nfi)dfl
Js,I (N) - J
1/2
II/z
v(f)v(#)se(-Nf )d9
1/2
I/2 N
mI/k-Ie(Om)v(fi)se(-Nfi)dP
t
/z,r,-1 k
The singular series
5.7
N
137
1/2
111/2
N
v(,)3e(-(N -
m-I
r(1 + l/k)s N-I 1kml/k-I(N
r (s/k )
m)s/k-1
m-I
0N1-1 -1
ml /k-)(N
+O
n!)(`-I)/k-1
Applying Lemma 5.3 to the main term (with a = s/k and fl = 1/k) and the error
term (with a - (s - 1)/k and - 1/k), we obtain
ml /k-1(N
m_1
and
m)s1k-1 , (1/k)r(I/k)r(s/k)N(s+1)/k-1 + 0 (NsIk
r((s+1)/k)
N-1
arI/k-I(N
m)(s-1)1k-I
= O (N31")
,n-1
This gives
Js+I(N)
(1/k)r(1/k)r(s/k) r (1 + I/k)-'
r((s + 1)/k)
r (s/k)
r (I +
l/k)5+1
N(s+1)Ik-1 + D
(Ns/k-1)
N(s+1)/k-I
+O
r((s + 1)/k)
(Ns/k-1)
This completes the induction.
5.7
The singular series
In Theorem 5.3, we introduced the function
t(N, Q) _ E AN (q),
Isq Q
where
AN(q)=F(S9a))se(
Na).
We define the singular series for Waring's problem as the arithmetic function
00
((N) - E AN(q).
q-I
138
5.
Let
0<e<
sK
Since s > 2r + I - 2K + 1, we have

S
K-I-SE >1+K-SC -1+54,

1
where
1
34-K-SE >0.
By (5.7),
q
AN(q) <<
< - 91+64
gslK-S
(5 . 9)
and so the singular series E. AN(q) converges absolutely and uniformly with
respect to N. In particular, there exists a constant c2 - c2(k, s) such that
(5.10)
16(N)I < C2
for all positive integers N. Moreover,
CA(N) - 6(N, P") -
AN (q)
<<
We shall show that CA(N) is a positive real number for all N and that there exists
a positive constant c1 depending only on k and s such that
0<c1 <6(N)<C2
for all positive integers N. The proof is a nice exercise in elementary number
theory. We begin by showing that AN (q) is a multiplicative function of q.
Lemma 5.4 Let (q, r) - 1. Then

S(qr, ar + bq) - S(q, a)S(r, b).
Proof. Since (q, r) - 1, the sets {xr : I < x < q} and {yq : I < y < r) are
complete residue systems modulo q and r, respectively. Because every congruence
class modulo qr can be written uniquely in the form xr + yq, where 1 < x < q
and I < y < r, it follows that
S(gr,ar+bq)-
e
m-l
((ar + bq)mk
qr
The singular series
5.7
139
r,re r(ar+bq)(zr+yq)k1
qr
tte (((aT+bq))
()(xr)t(Y))
l
J to,
A-1-1
qr
,-1
\\(ar+bq)) ((xr)A+(Yq)')
qr
X-1 v-1
EEe(T)e ) (b(Y4)k
r
r-I v-1
(ax)
(byl
- S(q, a)S(r, b).

Lemma 5.5 If (q, r) - 1, then

AN(qr) - AN(q)AN(r),
that is, the function AN(q) is multiplicative.
Proof. If c and qr are relatively prime, then c is congruent modulo qr to a

number of the form ar + bq, where (a, q) - (b, r) - 1. It follows from Lemma 5.4
that
(S(rc)y e - cN\
AN (gr) =
qr
qr
I:
C,
(S(gr,ar+bq))'e ( (ar+bq)N
qr
b-I
m.QI (b.Q)-I
qr
r- N)e(-bN)
Io.Qh i (b .yl-i
1:
1o.4)-I
(S(a))$
e
( aN
q
CS(r,b)15
b-l
-)
r
AN(q)AN(r)
For any positive integer q, we let MN (q) denote the number of solutions of the
congruence
xk
N (modq)
in integers xi such that I < x; < q for i - 1, ... , q.
140
5.
Lemma 5.6 Let s > 2k + 1. For every prime p, the series
x
XN(P)a 1+T, AN(ph)
(5.11)
h-t
converges, and
MN(Ph)
XN(P) - ti eD ph(J-tl
(5.12)
Proof. The convergence of the series (5.11) follows immediately from inequality (5.9). If (a, q) - d, then
)X-1
- E e \ (a/d)xk
q/d
L-1 a \axk
q
a
S(q, a)
(a/d)xk
9/d
d E e\ q/d
d S(q/d, a/d)
X-1
Since
(am)1
q-E
I
if m- 0 (mod )
ai 0
if m
(mod q),
it follows that for any integers xt , ... , x.,
ql
q
a-1
and so
q
..
x(-1
x,-1
e
a-t
ifxk+. +xk-N (mod q)

*N (modq)
5
a(xt +...q
MN(q)
tt 0
JJ
xs - N)-
EE...Ee(a(x1+... xf -N)
q
-t x -1
q
-EEe
1
x -t
axk
(axk
(-aN )
- 1: S(q, a)e (-)l

9
a-1
aq
S(q, a)se
djq
q
( q N)l
l
)N
\- q/dl
- q Y: Y: dss(q/d,a/d)Se r
C9
(,.vw
5.7
1q dlq
q3
4-1
The singular series
141
q/d
q/d /
(S(q/d,a/d))s
a \-(a/d)N
14.91-d
- qs-1 E AN(gld)
dlq
Therefore,
-g1_'MN(q)
EAN(q/d)
dlq
for all q > 1. In particular, for q - ph we have

h
AN(Ph/d) - ph(1-s)MN(Ph)
1 + E AN(pi) j-l
dip"
and so
XN(P) - hlym
(I +
AN(P')
j-1
ph(1-')MN(Ph)
- lira
h-+oo
Lemma5.7 ifs> 2k+1, then

6(N) - F1 XN(P)
(5.13)
Moreover, there exists a constant c2 depending only on k and s such that
0<((N)<c2
for all N, and there exists a prime po depending only on k and s such that
1/2 < J XN(P) < 3/2
(5.14)
P>PO
for all N > 1.

Proof. We proved that ifs > 2k + 1, then
1
AN(q) << ql+64
where S4 depends only on k and s, and so the series Eq AN (q) converges absolutely.
Since the function AN(q) is multiplicative, Theorem A.28 immediately implies the
convergence of the Euler product (5.13). In particular, XN(p)'i 0 for all N and
142
5.
p. Since XN(p) is nonnegative by (5.12), it follows that XN(P) is a positive real

number for all N and p, and so the singular series 6(N) is positive. Again, by (5.9),
O
0<66(N)<E91+a.-c2<00
4-1
and
D
I XN(P) - 11 <
I AN(ph)I <<
h(I+se)
h-1 p
h-1
K
p
I+a+.
Therefore, there exists a constant c depending only on k and s such that
1 - p+a. <Xn(P)<l+
p
4
for all N and p. Inequality (5.14) follows from the convergence of the infinite
products [1p(1
We want to show that 6(N) is bounded away from 0 uniformly for all N. By
inequality (5.14), it suffices to show, for every prime p, that XN(p) is uniformly
bounded away from 0.
Let p be a prime, and let
k - PTko,
where r > 0 and (p, ko) - 1. We define
Y-
r+1 ifp>2
t r+2 ifp-2.
Lemma 5.8 Let m be an integer not divisible by p. If the congruence xt as m

(mod pY) is solvable, then the congruence yk as m (mod ph) is solvable for
every h > y.
Proof. There are two cases. In the fast case, p is an odd prime. For h > y - r+ 1,
we have
(k,1p(Ph)) - (koPT, (p
- 1)ph-1) - (ko, p - 1)pr - (k, W(p")).
The congruence classes modulo ph that are relatively prime to p form a cyclic
group of order rp(ph) - (p - l )ph-1. Let g be a generator of this cyclic group,
that is, a primitive root modulo ph. Then g is also a primitive root modulo pY. Let
xA as m (mod pY). Then (x, p) = 1, and we can choose integers r and u such
that
x = gu
(mod p h )
M = gr
(mod ph).
and
Then
ku
(mod 1p(pY )),
5.7
The singular series
143
and so
r-0
(mod (k, ap(p)')))
and
r = 0 (mod (k,
,p(ph)))-
Therefore, there exists an integer v such that
kv - r
(mod (p(p")).
Let y - g". Then yk - m (mod ph).

In the second case, p - 2 and so m and x are odd. If T - 0, then k is odd.
As y runs through the set of odd congruence classes modulo 2h, so does yk, and
the congruence yk = m (mod 2h) is solvable for all h > 1. If r > 1, then k
is even and m = xk = I (mod 4). Also, xk - (-x)k, and so we can assume
that x - I (mod 4). The congruence classes modulo 2h that are congruent to
I modulo 4 form a cyclic subgroup of order 2h-2, and 5 is a generator of this
subgroup. Choose integers r and u such that
m = 5`
(mod 2h)
and
x - 5" (mod 2h).

Then xk - m (mod 2Y) is equivalent to
ku = r
(mod 2Y-2),
and so r is divisible by (k, 2r) - 2r - (k, 2h-2). It follows that there exists an
integer v such that
kv =- r
(mod
2h-2).
Let y - 5". Then yk - m (mod 2h). This completes the proof.

Lemma 5.9 Let p be prime. If there exist integers a , , ... , a, not all divisible by
p, such that
(modp'),
then
XN(P) > py(I -V) > 0.
Proof. Suppose that aI # 0 (mod p). Let h > y. For each i - 2, ... , s there
exist ph-Y pairwise incongruent integers x, such that
xi = a,
(mod ph).
Since the congruence
x =N-xz-...-xs (mod pY)
144
5.
is solvable with x) - a, # 0 (mod p), it follows from Lemma 5.8 that the
congruence
xi = N - xz - ... - xf
(mod ph).
This implies that

MN(Ph) >
P(11 -Y As - 1)
and so
X,v(P) = hlim
AV(P11)
ph(s-i) > py(..-i) > 0.
Lemma 5.10 Ifs > 2k for k odd or s > 4k fork even. then
XN(P) > PY('-5) > 0.
Proof. By Lemma 5.9, it suffices to prove that the congruence
(modpY)
(5.15)
is solvable in integers a, not all divisible by p. If N is not divisible by p and the

congruence is solvable, then at least one of the integers a; is prime to p. If N is
divisible by p, then it suffices to show that the congruence
ai +
+ak_i + 1k
-N
(mod pY)
has a solution in integers. This is equivalent to solving the congruence
ai +
+ak_)
N-1
(mod
p").
In this case, (N - 1, p) - 1. Therefore, it suffices to prove that, for (N, p) - 1,

the congruence (5.15) is solvable in integers for s > 2k - I if p is odd and for
s>4k-1ifpiseven.
Let p be an odd prime and g be a primitive root modulo pY. The order of g is
p(PY) - (p - l )PY-) = (p - t) pt . Let (m, p) - 1. The integer m is a kth power

residue modulo pY if and only if there exists an integer x such that
xk
- m (mod p1).
Let m - g'
(mod pY). Then m is a kth power residue if and only if there exists
an integer v such that x - g" (mod p)') and
kv - r (mod (p - 1)p`).
Since k - kop` with (ko, p) - 1, it follows that this congruence is solvable if and
only if
r-0
(mod (ko. p - l )pr),
5.7
The singular series
145
and so there are

(P(PY)
p-1
(ko, p - 1)PT
(ko, p - 1)
distinct kth power residues modulo pY. Let s(N) denote the smallest integer s
for which the congruence (5.15) is solvable, and let C(j) denote the set of all
congruence classes N modulo pY such that (N, p) - I and s(N) = j. In particular,
C(l) consists precisely of the kth power residues modulo p}'. If (m, p) - I and
N' - mk N, then s(N') = s(N). It follows that the sets C(j) are closed under
multiplication by kth power residues, and so, if C(j) is nonempty, then IC(j)I
(p - 1)/(ko, p - 1). Let n be the largest integer such that the set C(n) is nonempty.
Let j < n and let N be the smallest integer such that (N, p) - I and s(N) > j.
Since p is an odd prime, it follows that N - i is prime to p for i - I or 2, and
s(N - i) < j.Since N =(N - 1)+lk andN-(N-2)+lk+lk,itfollows that
j + I <s(N) <s(N -i)+2 < j+2

and so s(N - i) - j or j - 1. This implies that no two consecutive sets C(j) are
nonempty f o r j - 1, ... , n, and so the number of nonempty sets C(j) is at least
(n + 1)/2. Since the sets C(j) are pairwise disjoint, it follows that
(P - I)PT _ CO(PY)
IC(j)I >
n+1
2
p-1
(ko, p - I)'
and so
n <<2(ko,P- 1)pt - 1 <2k- 1.

Therefore, s(N) < 2k - I if p is an odd prime and N is prime to p.
Let p - 2. If k is odd, then every odd integer is a kth power residue modulo 2Y,
so s(N) - 1 for all odd integers N. If k is even, then k - 2Tko with r > 1, and
y - r + 2. We can assume that I < N < 2Y - 1. If
s=2Y-I-4.2T-I<4k-1,
then congruence (5.15) can always be solved by choosing a, = I for i = 1, ... , N
and ai - 0 for i = N + 1, .... s. Therefore, s(N) < 4k - I for all odd N. This
Theorem 5.6 There exist positive constants ci = c} (k, s) and c2 - c2(k, s) such
that
cl < 6(N) < c2.
Moreover, for all sufficiently large integers N,
6(N, P') - 6(N) + 0 (P-s)
146
5.
Proof. The only part of the theorem that we have not yet proved is the lower
bound for 6(N). However, we showed that there exists a prime po - po(k, s) such
that
1/2 < fl XN(P) < 3/2
P> PO
for all N > 1. Since
XN(P) > pr(i-s) > 0
for all primes p and all N, it follows that
6(N) _ FIXN(P) >
fl XN(P) ? - fl P YO-S)
= c1 > 0.
P:5N
P-<Po
5.8
Conclusion
We are now ready to prove the Hardy-Littlewood asymptotic formula.
Theorem 5.7 (Hardy-Littlewood) Let k > 2 and s > 2k + 1. Let rk,s(N) denote
the number of representations of N as the sum of s kth powers of positive integers.
There exists 8 - S(k, s) > 0 such that
rk.s(N) - 6(N)I' I 1 + k
1
\s
r kl
N(c/k)-l
O(N(11k)-1-5),
where the implied constant depends only on k and s, and 6(N) is an arithmetic
function such that
c, <66(N)<C2
for all N, where c, and c2 are positive constants that depend only on k and s.
Proof. Let So - min(1, S 1, 82, S3, v84). By Theorems 5.2-5.6, we have
rk,(N) =
f
J
F(a)se(aN)da
F(a)se(-aN)da + fm F(a)se(-ctN)da
0 (ps-k-52) + 0 (ps-k-5'
= 6(N,
_ (6(N) + 0 (P-"54)) (J(N) + 0 (Ps-k-a,)) + 0 (ps-k-52)
+0 (ps-k-5i )
6(N)J(N) + 0 (ps-k-so)
=Ik)
/
6(N)1'
1a
1+
(-1Ns!k-I + 0 (N(c- Olk--1
5.10
+0
Exercises
147
(Ns/k-)-so/k)
-6(N)I'(l+kls
r(k) -I
Ns/k-1 +o(Ns /k 1-b ),
where 3 - So/k. This completes the proof.
5.9
Notes
The circle method was invented by Hardy and Ramanujan [50] to obtain the asymptotic formulafor the partition function p(N), which counts the number of unordered
representations of a positive integer N as the sum of any number of positive integers. The circle method was also applied to study the number of representations of
an integer as a sum of squares. See, for example, Hardy [45], and the particularly
important work of Kloosterman [71, 72, 73].
In a classic series of papers, "Some problems of 'Partitio Numerorum'," Hardy
and Littlewood [47, 48] applied the circle method to Waring's problem. Vinogradov [131, 134, 135] subsequently simplified and strengthened their method.
This chapter gives the classical proof of the Hardy-Littlewood formula for s >
so(k) - 2k + 1. There is a vast literature on applications of the circle method to Waring's problem as well as to other problems in additive number theory. The books
of Davenport [18], Hua [64], Vaughan [125], and Vinogradov [135] are excellent
references.
There have been great technological improvements in the circle method in recent years, particularly by the Anglo-Michigan school (for example, Vaughan and
Wooley [126, 127, 128, 129, 130, 147, 148]). In particular, Wooley [146] proved
that
G(k) < k(logk + log logk + 0(1)).

Another interesting recent result concerns the range of validity of the HardyLittlewood asymptotic formula. Let G(k) denote the smallest integer so such that
the Hardy-Littlewood asymptotic formula (5.1) holds for all s > so. Ford [41]
proved that
G(k) < k2 (log k + log log k + 0(1)).
For other recent developments in the circle method, see Heath-Brown [54, 55].
Hooley [59, 60, 61 ], and Schmidt [ 107).
5.10
Exercises
1. Show that for k - I the Hardy-Littlewood asymptotic formula is consistent

with Theorem 5.1.
148
5.
2. Let k > 2. Show that the number of positive integers not exceeding x that
can be written as the sum of k nonnegative kth powers is x/k! + 0
Show that
(X(k-1)/k1
G(k)>k+1.
Hint: If n < x is a sum of k kth powers, then
n-a
where
0<al <a2<...<akXIlk,
and the number of such expressions is given by a binomial coefficient.
3. Let f (x) be a polynomial of degree k > 2 with integral coefficients, and let
4
Sf(q, a) - E e(af(r)/q)
r-l
Prove that if (q, r) - 1, then
S f(qr, ar + bq) - S f(q, a)S j(r, b).

4. Let Rk..T(N) denote the number of representations of an integer N as the
sum of s nonnegative kth powers. State and prove an asymptotic formula
for Rk,3(N).
Part II
The Goldbach conjecture
6
Brun's method is perhaps our most powerful elementary tool in number theory.
P. Erdos [34]
6.1
Euclid's theorem
Before beginning to study sums of primes, we need some elementary results about
the distribution of prime numbers.
Let s - a + it be a complex number with real part a and imaginary part t. To

every sequence of complex numbers a1, a2, ... is associated the Dirichlet series
00
Wi
F(s) -
ns
If the series F(s) converges absolutely for some complex number so - ao + ito,
then F(s) converges absolutely for all complex numbers s - a + it with Jt(s) -
a > ao - t(so), since

an
ns
_
I
IanI
no
an
Ian
I
nn
ns`
If we let an = 1 for all n > 1, we obtain the Riemann zeta function
(s)
O
n-1
:
1
152
6.
This Dirichlet series converges absolutely for all s with IJI(s) > 1.
Theorem 6.1 Let f (n) be a multiplicative function. If the Dirichlet series
f(n)
F(s) nn--11
ns
converges absolutely for all complex numbers s with It(s) > co, then F(s) can be
represented as the infinite product
F(s) s 11 ( 1 + f (P) + f (PZ) +

PS
J.
J
P2S
If f (n) is completely multiplicative, then
F(s)-fI1- Pf (P)
\\\\\\
This is called the Euler product for F(s).
Proof. If f (n) is multiplicative, then so is f(n)/n. If f (n) is completely mul tiplicative, then so is f (n)/n5. The result follows immediately from Theorem A.28.
Because the Riemann zeta-function converges absolutely for :1t(s) > 1. it

follows from Theorem 6.1 that z(s) has the Euler product
00
(s)-E- -n C1- PSY

//i
n-1
for all s with It(s) > 1, and so (s) f 0 for It(s) > 1. From the Euler product, we
obtain the following analytic proof that there are infinitely many primes.
Theorem 6.2 (Euclid) There are infinitely many primes.
Proof. For 0 < x < 1 we have the Taylor series

00 xn
-log(1 -x)-E -n
n-1
land
l l (l
- p1
'
=-log 1P
00
1: npn()+a)
p n-1
E p1+0, + E E npn('+a)'
0O
n-2
6.2
Chebyshev's theorem
153
Since
<P"
0<E Enpn(l+a)
p
n-2
P(P1-1) <oo,
n-2
it follows that
pi + 0(1).
+a) _
(6.1)
(6.2)
Let 0 < a < 1. Then

r
Jx1 <
a=!
xdx < (l +o) < 1 +t d x =
+1
a
and so
1
0 < log - < log
(1 + a )
! + 1 /I - log Ia + log(] + a)
< log I
\\a
< log
+ a < log 1 + 1.
Therefore,
I
+a)=log-+O(1).
a
(6.3)
Combining (6.2) and (6.3), we obtain

log
= E pig
1
+ O(l)
for 0 < a < 1. If there were only finitely many prime numbers, then the sum on
the right side of this equation remains bounded as a tends to 0, but the logarithm
on the left side of the equation goes to infinity as a tends to 0. This is impossible,
so there must be infinitely many primes.
6.2
Chebyshev's theorem
The simplest prime-counting functions are
r(x)=E1,
p<x
and
154
6.
19(x) and tli(x) are called the Chebyshev functions. Chebyschev proved that the
functions 19(x) and 1/i(x) have order of magnitude x and that ,r(x) has order of
magnitude x/ log x. Before proving this theorem, we need the following lemma
about the unimodality of the sequence of binomial coefficients.
Lemma 6.1 Let n > 1 and 1 < k < n. Then
n
(k
1) < (k) if and only if k < n21
(k
n
(k
(k)
1) >
1) -
(k n)
if and only if k > nZ 11,
if and only if n is odd and k -
Proof. This follows immediately from observing the ratio

ln\
k
(k 1)
(k - 1)!(n - k + 1)!
k!(n - k)!
ki(h-k),
n'
n!
(k-I)'(n-k+1)1
Lemma 6.2 Let n > 1 and N -
n-k+l
k
Then
N <2Z' <2nN.
Proof. Since (2") is the middle, and hence the largest, binomial coefficient in
the expansion of (1 + 1)2, it follows that
< (1 + 1)n - 22
N - (2n
n
22 (n)
k
k_1
(n)
< 2+(2n - 1)2n<

(n )
2n
(2n)
n
2nN.
For any positive integer n, let v,,(n) denote the highest power of p that divides
n. Thus, vp(n) - k if and only if pk 11n. In this case, pk < n and so vp(n) <
log n/ log p.
Lemma 6.3 For every positive integer n,

tbfi n/ bBPI [
00
vP(n!)
-y
k-1
]-
k-1
k
p J
(6.4)
Chebyshev's theorem
6.2
155
Proof. Since vp(mn) - vp(m)vp(n) for all positive integers m and n, we have
00
vp(n!)->Up(m)->>l
F1
"0
m-l
k-I
m-1 Dhi,
k-1
[;].
2>1
This proves the formula.
Theorem 63 (Chebyshev) There exist positive constants c 1 and c2 such that
cIx < t (x) <'/'(x) < n(x)logx < c2x
(6.5)
for all x > 2. Moreover,

lim inf 'fi(x) - lim inf '`(x) - lim inf
X-00
x-00
x
x-00
x
'r(x)log x
x
> log 2
and
lim sup
X-00
0(x)
x
- lim sup (x) - lim sup

x
X-00
log x
7T (X)
X-00
< 4log 2.
Proof. Let x > 2. If pk < x, then k < (log x/ log p), and so
OW )-TlogP
logp -
'G(x)
Erlog x 1logP
P* <x
p<x
p<x
L log P
JJ
< E log x - n(x) log x.

p<x
Therefore,
lim inf
X-00
19(x)
and
hm sup
x-00
z9(x)
X
< lim inf

.1-00
*(x)
x
< lim inf

X-00
n(x)logx
x
ir(x)logx
< lim sup *(x) < lim sup
x - .1-00
X-00
x
Let
0<S<1.
Then
OW
E log p
r a<p<x
(1 - S)logx
a'-J<p<x
(1 - S) (7r (X) - n(x'-a)) logs
(I -S)n(x)log x
-x1_alogx,
156
6.
and so
(1 -S)7r(x)logx
O(X)
It follows that
lim inf
x-.oc
logx
xa
7r (x) log x
0 (x)
> (1 - S) lim inf
x-oo
X
x
This holds for all 6 > 0, and so

fl(X)
lim inf
x- 0c
6(x)
X-.00
lim inf
0(x)
X00
lim sup
X-oc
6(X)
X-00
lim sup
7r (x) logx
> lim inf
> lim sup

x-oo
- lim inf
tfi(X)
X--.00
- lim sup
t/i(X)
7r(x)logx
X
- lim inf
- lim sup
X00
X-.00
2 n
2n(2n - 1)(2n -
(6.6)
X-.00
7r (X) logx
7r(x) logx
(6.7)
Let n > 1, and let

N -
(n
1)
n!
Then N is an integer, since it is a binomial coefficient, and

2zn
-<N<22"
2n by Lemma 6.2. If p is a prime number such that
n<p<2n,
then p divides the numerator but not the denominator of N. Therefore, N is
divisible by the product of all these primes, and so
fl p<N<22n.
n<p<2n
In particular, if r > I and n = 2r-1, then
fl
p<N<22'.
<p<2'
It follows that, for any R > 1,

R
H p -r-1fl2'-1<p$2'
H
p<28
R
p<fl22.
<22"
r-1
Chebyshev's theorem
6.2
157
For any number x > 2, there is an integer R > 1 such that
< x < 2R
2R-1
Then
fj P< H p <
p<x
22R"
< 24X
p<2'
and so
< (4log 2)x.
O(X)p5X
p<X
Thus,
lim sup
rg(x)
< 41og 2.
X-CO
To obtain the lower limit, we use Lemma 6.3 to express N explicitly as a power
of primes:
N - (2n)
- (2n)!
nj2
pop(2+7)-2up(n)'
11
p<2n
where
([]_2[]).
vp(2n)-2vp(n)v
Since [2t] - 2[t]
0 or 1 for all real numbers t, it follows that
vp(2n) - 2vp(n) <
log 2n
log p
By Lemma 6.2,
2z"
2n
<N-
p ,g p
pUp(zn)-zap(n) <
pQn
p:5 2n
< fl 2n - (2n)n(2n)
p :52n
or, equivalently,
7r (2n) log 2n < 2n log 2 - log 2n.
Let n - [x/2]. Then
2n <x <2n+2
and
7r (x) log x > Jr (2n) log 2n > 2n log 2 - log 2n
> (x - 2) log 2 - log x - x log 2 - log x - 21og 2.

It follows that
7r(x)logx
x
> log 2
- logx+2log2
x
158
6.
and so
Jim inf
x-.oo
it (x) log x
> log 2.
x
Since t9(2) > 0, we have #(x) > ctx for some ct > 0 and all x > 2. This
Theorem 6.4 Let p denote the nth prime number. There exist positive constants
c3 and c4 such that
can log n < p,, < CO log n
for aim
2.
Proof. By Chebyshev's inequality (6.5),

c l PR
c2 Pn
log pn
log Pn
and so
C2 1 n log pn < pn < cj n log pn.

Since
logn < log pn,

we have
pn > c2 t n log n - can log n.

For n sufficiently large,
log pn < log n + log log pn + log c

< log n + 2 log log pn
< log n + (1 /2) log pn,

so
log pn < 2 log n

and
pn <cl
In log
p <2cj inlogn.
Therefore, there exists a constant c4 such that p,, < CO log n for all n > 2. This
6.3
Mertens's theorems
In this section, we derive some important results about the distribution of prime
numbers that were originally proved by Mertens.
Lemma 6.4 For any real number x > 1 we have
0 < E log (X) < X.

n<c
6.3
Mertens's theorems
159
Proof. Since the function h(t) - log(x/t) is decreasing on the interval [1, x], it
follows that
log (x) < logx +

n
I<n<x
log (x) dt
J,
- xlogx - J rlogtdt
I
-xlogx-(xlogx-x+l)
< X.

The function A(n), called von Mangoldt's function. is defined by
A(n) - j log p
if n - p'n is a prime power

otherwise.
l0
Then
Ifi(x) - E A(m).
I <n, <x
Theorem 6.5 (Mertens) For any real number x > 1, we have
1: A(n)
n <x
- logx + O(1).
Proof. Let N - [x]. Then
0<Elogx-Nlogx-Elogn-xlogx-logN!+0(logx)<x
n<x
n-1
by Lemma 6.4, and so
log N! -xlogx+O(x).
It follows from Lemma 6.3 and Theorem 6.3 that
log N! - E vp(N)logp
p<N
Oog N11og pl
p<N
rN1logP
k-11
PJ
E [Pk]logP
p <x
[ Pk I log e
--+LnJA(n)
n <x
160
6.
+ 0(1)) A(n)
n
n<x
=x1: A(n)+O
EA(n)
n
n<x
n<x
_X
A(n) + O (*(x))
n
n<x
A(n) + O(x).
=x
n<x
Therefore,
xE A(n) + O(x)-xlogx+O(x)
n<.t
and
A(n)
= logx + O(1).
n<x
Theorem 6.6 (Mertens) For any real number x > 1, we have
1ogp
-log x+O(l).
P<x
Proof. Since
A(n)
0:5
a<x
log P
log P
P<.t
pk
ts,
P_x
logpE k
k-2 P
log P
P'X
P(P -
2 1: log P
P<.t
<
:logn
n2z
n-i
it follows from Theorem 6.5 that
Elogp =E A(n)+O(l)-logx+O(l).
P<x
n<x
Mertens's theorems
6.3
161
Theorem 6.7 There exists a constant b, > 0 such that

1
E p =loglogx+b,+O \log
\
1
p<A
forx>2.
Proof. We can write
log p
EP
p<x
p<x
p log p=
where
u(n)-
'p'
ifn-p
otherwise
(
jl
u(n) f (n),
n<.r
and
f(t) -
log
We define the functions U(t) and g(t) by
U(t) -
u(n) n
log P - log t + g(t).

p<i
Then U(t) - 0 fort < 2 and g(t) - 0(1) by Theorem 6.6. Therefore, the integral
f2' g(t)/(t(log t)2)dt converges absolutely, and
g(t)dt
t(logt)2
logx
Since f (t) is continuous and U(t) is increasing, we can express the sum p<r 1/p
as a Riemann-Stieltjes integral. Note that U(t) - 0 for t < 2. By partial summation, we obtain
p
P<A
- Eu(n)f(n)
n <x
A
I
= 2 +2 f(t)dU(t)
- f(x)U(x) - f x U(t)df(t)
2
r
logx +g(x)
logx
- 1 +O
2
1
(logx
x logs+g(t)dt
t(logt)2
162
6.
ilogtdt+Jt(lo(gt)2dt-J
t(lgogt)2dt+1+O(logx)
- log log x - log log 2 + J
=loglogx+bI+o
\logx
t (log t)2
dt+1+0
(log I x
'
where
61 = 1 - loglog 2 +
t(log t )2 d t.
(6.8)

From the Taylor series for log(] - x), we see that
0<l0g
00
00
1-1
_2
np"
n-2
P(P
P"
It follows from the comparison test that the series
log( 1-
62P
1" v
=1 1
1- 1
P
k_2
kpk
(6.9 )
converges.
Lemma 6.5 Let b, and b2 be the positive numbers defined by (6.8) and (6.9).
Then
b1+b2-y,
where y is Euler's constant.
Proof. Let 0 < a < 1. We define the function F(a) by
F(a)=log C(1+a)-L
p1+o
Y
log1--
P
O
npn(I+a)
1:
n-2
By (6.1) and the Weierstrass M-test, the last series converges uniformly for a > 0
and so represents a continuous function for a > 0. Therefore,
lim F(a) - b2.
(6.10)
We shall find alternative representations for the functions log C(1 + a) and
FP p-I-. Since
a2
a2
1 -a+<e-
<
I
-a+2
2e
Mertens's theorems
6.3
for 0 < a < 1, it follows that
1-2a <
and
2e
2e-a
1+- < 1+
1 - e-Q
a
a
<
I -e-
a
<1-2e
< 1+
2-a
< 1+a.
Therefore,
0 < log es + log (1 - e-) -' < a,

and so
log 1 - log(1
- e-)- + 0(a).
By (6.3), we have
1
logC(I+a)-log-+O(a)
- log(1 - e-)-l + O(a)
00 a-n
+ O(a).
n-I
By Theorem A.5,
-logx+y+01 X)
n<x
for x > 1. Let f (x) - e-x. By partial summation, we have
logc(1+a)-f f(n) +0(a)

n
nfo -i
f(x)dL(x)+ 0(a)
00
-J
L(x)df(x)+O(a)
or fom e- `L(x)dx + O(a).

By Theorem 6.7,
S(x) _ E 1 = log log x + bi + 0 (

P
.r
log x
for x > 2. Let g(x) - x-. Again, by partial summation we have
1,
i.R P
$(P)
P =
00
00
f g(x)dS(x) - - J
S(x)dg(x)
>
163
164
6.
-a
S(x)dx
x14-0
= a fOG e -0xS(e')dx.
Since
S(e-')=1ogx+b,+01 x1)
and
L(x)=logx+y+0(i),
it follows that
L(x)-S(e')-y-b,+0I X
for x > 1. We also have

1
L(x) - S(e') - y - b, + 0
x+1
for 0 < x < 1. Therefore,
F(a)-logc(l+a)- i+0
P
-a/
e-'-'(L(x) - S(e''))dx + 0(a)
J e-0.r(y-b,+0(x+l))dx+O(a)
//11
JJJ0
oo a-0.cdx
roc
-(y-b1)a / e-0'dx+0 a f0
=y-b,+O a
TO exdx
x+l )+0(a}
x+1
Since
00 e-axdx
x+l
I;o a-o.rdx
< J0
00 e-OXdx
x+ 1 + J 1 /0
' dx
e-''d y
+1
- log (i +1)+0(1)
a
JJJ
1 +1
<< log (a
Mertens's theorems
6.3
165
it follows that
F(o) -y-bI+0
l7logl 1 +1))
By (6.10), we have
b2- lim F(o)-y-bi.

Theorem 6.8 (Mertens's formula) For x > 2,
fl
P<x
(1
-P
Y - ey logx + 0(1),
Proof. We begin with two observations. First,

O`
L, kP
p>x k-2
< IS7
P>x P(P - 1)
<
Second, since exp(t) - I + O(t) for tin any bounded interval and 0 (1/ log x) is
bounded for x > 2, it follows that
expl(_L))
-1+0
Therefore,
log
1'
fl(1p<x \
P)
1)
P
O
p<x k-I kP
- 1 +kk
p<x P
p<x k-2
166
6.
- log log x + b1 + 0 I
I + b2 - E E
\ log x ///
p>x k-2 kpk
- loglogx+y+O
log x
'
since b1 + b2 - y by Lemma 6.5, and so
\\\11-
- e}logxexp(O (j_!-_))
Jll
P<X
-e>logx11+01
\\\ log x J lJ
\\\
- e>' logx + 0(1).

This is Mertens's formula.
The following result will be used in Chapter 10 in the proof of Chen's theorem.
Theorem 6.9 For any e > 0, there exists a number u 1 - u 1(E) such that
/I1-II\
<(1+E)log z
P/
<P<Z \\
logu
for any u 1 0 such that
y+S
< 1 +E.
y-S
By Theorem 6.8, we have
I1/
P<x\
ylogx,
and so there exists a number u 1 such that

1
<(y+S)logx
P
P<X
for all x > u 1. Therefore, if u 1 < u < z, we have

)-I
(1
nP<Z
1
Hp<u(1-P)
(y + S) log z
(y - S) log u
<(1+e)
log z
log u
6.4
6.4
167
There is a structural similarity between the twin prime conjecture and the Goldbach
conjecture. The twin prime conjecture states that there exist infinitely many prime
numbers p such that p + 2 is also a prime number or, equivalently, there exist
infinitely many integers k such that k(k + 2) has exactly two prime factors. The
Goldbach conjecture states that every even integer n > 4 can be written as the sum
of two primes or, equivalently, there exists an integer k such that 1 < k < n - 1
and k(n - k) has exactly two prime factors. We begin the study of sieve methods
with a simple proof of the theorem that the twin primes are sparse in the sense that
the sum of the reciprocals of the twin primes converges. This contrasts with the
result (Theorem 6.7) that the sum of the reciprocals of all of the primes diverges
like log log x.
Lemma 6.6 If f > 1 and 0 < m < f, then
(-1)k(k) =(-1)m (em 1 f.

k-0
Proof. This is by induction on m. It is easy to check that the equation is true for
m - 0, 1, 2. If 1 < m < e and the equation holds for m - 1, then

_ (-l)k(e) +(-l)m(t)
E(-I)k(e)
k
k
k-0
k-0
=(-I)m-i(m
1)+(-1)m(E
((e\)
m- (m - ll l /
=(-1),n(em 1).

The following combinatorial inequality, a version of the principle of inclusionexclusion, is the simplest form of the Brun sieve.
Theorem 6.10 (The Brun sieve) Let X be a nonempty, finite set of N objects,
and let P1,..., P, be r different properties that elements of the set X might have.
Let No denote the number of elements of X that have none of these properties.
F o r any subset 1 - (i t , ... , ik) o f ( 1 , 2, ... , r), let N (1) - N (i 1, ... , ik) denote
the number of elements of X that have each of the properties P;, , P,,..., P. Let
N(0) - I XI - N. If m is a nonnegative even integer, then
m
No
(-1)k
k-o
N(1).
Ill-k
(6.11)
168
6.
If m is a nonnegative odd integer, then

m
No > E(-1)k
k-0
N(l).
(6.12)
Ill-k
Proof. Inequalities (6.11) and (6.12) count the elements of X according to the
various properties that each element possesses. We shall calculate how much each
element of X contributes to the left and right sides of these inequalities.
Let x be an element of the set X, and suppose that x has exactly a properties
Pi. If e - 0, then x is counted once in No and once in N(0), but is not counted
in N(1) if I is nonempty. If e > 1, then x is not counted in No. By renumbering
the properties, we can assume that x has the properties P1, P2, ..., P1. Let I c
{ 1 , 2, ... , , . . . , r } . If i E I for some i > e, then x is not counted in N(1). If
1 c (1, 2, ... , e), then x contributes 1 to N(I ). For each k - 0, I, ... , e, there are
exactly () such subsets with I 1I - k. If m > e, then the element x contributes
D-1),
(1)
k-0
to the right sides of the inequalities. If m < f, then x contributes
m
1:(-I)k(ek
k-0
to the right sides of inequalities (6.11) and (6.12). By Lemma 6.6, this contribution
is positive if a is even and negative if f is odd. This completes the proof.
Lemma 6.7 For x > I and for any congruence class a
(mod m), the number

of positive integers not exceeding x that are congruent to a modulo m is x/m + 0,
where 101 < 1.
Proof. If x/m - q E Z, then the set ( I__, qm) contains exactly x/m elements
in every congruence class modulo in.
Suppose that x/m Z. Let [x] and {x} denote the integer and fractional parts
of x, respectively, and let [x] - qm + r, where 0 < r < m. Then
qm <x-qm+r+{x} <qm+(m-1)+0 <(q+l)m,

and so
q <mx- <q+l.
(6.13)
The positive integers up to x can be partitioned into q + I pairwise disjoint sets such
that q of these sets are complete systems of residues modulo m, and the remaining
set is a subset of a complete system of residues modulo m. It follows that there are
either q or q + 1 integers in the congruence class a (mod m). The lemma follows
from inequality (6.13).
6.4
169
Lemma 6.8 Let x > 1, and let pi, , ... , pit be distinct odd primes. Let N(i 1, ... ,
ik) denote the number of positive integers n < x such that
n(n + 2) _- 0
(mod pi,
pit ).
(6.14)
Then
2k X
+24.9,
Pi, ... Pi,

where 101 < 1.
Proof. If p is an odd prime and n(n + 2) = 0
n=0
(mod p), then either
(mod p)
or
n = -2 (mod p).
Moreover, 0 # -2 (mod p) since p > 3. If the integer n satisfies the congruence (6.14), then there exist unique integers u,, ... , uk E {0, -2}
n
n=
(mod pl )
(mod P2)
ul
u2
(6.15)
n = uk
(mod pk)
By the Chinese remainder theorem, for each of the 2k choices of u 1, ... , uk there
exists a unique congruence class a (mod p, . . . pk) such that n is a solution of
the system of congruences (6.15) if and only if
n -a (mod
By Lemma 6.7, this congruence has
x
solutions in positive integers not exceeding x, where I0(a)l < 1. Therefore,
N(i1.... , ik) a
2k x
Pi, ... pi,
+2 k 0
where X91 < 1. This completes the proof.
Theorem 6.11 (Brun) Let 7r2(x) denote the number of primes p not exceeding x
such that p + 2 is also prime. Then
n2(z)
x(loglogx)2
(log x)2
170
6.
Proof. Let 5 < y < x. Let r - pr(y) - 1 denote the number of odd primes
1 ,--- . , Pr. Let 7r2(y, x) denote the
not exceeding y. W e denote these primes by
number of primes p such that y pi for i - 1, ... , r, and
n(n + 2) 0 0 (mod pi)
for all i. Let N0(y, x) denote the number of positive integers n < x such that
n(n + 2) 0 0 (mod pi)
for all i - I-- r. Then

n2(x) < Y + n2(Y, x) < y + No(Y x).
We shall use the Brun sieve to find an upper bound for No(y, x).
Let X be the set of positive integers not exceeding x. For each odd prime
pi < y, we let Pi be the property that n(n+2) is divisible by pi. For any subset I {i 1 , ... , id contained in { 1, ... , r), we let N(I) be the number of integers n E X
such that n(n + 2) is divisible by each of the primes pi, .... , pik or, equivalently,
pik. By Lemma 6.8, we have
such that n(n + 2) is divisible by pi,
2" x
N(I)- N(il,...,ik)-
Pi, ... Pit +2k9.
Let m be an even integer such that 1 < m < r. By inequality (6.11), we have
N(I)
No(Y, x) < 1:(-1)k

/)-k
k-0
< E(-1),
+ 0(2k)k0
Pik
(_2kx
Pi,
(i....,ik)c{1.....r)
m
XY'
k-0 (i,..... it)c(1..... r)
xtE
AI ... Pik
k-0
(-2)k
k-0 (l .....;t)CII.....r) Pi,
... Pit
-x
(r)
+D-1)k0(2k)
(-2)k
+0 E (;)2k).
Pk
(-2)k
k-m+l (;.....it)C(1.....r) Pi,
k-0
We shall estimate these three terms separately. By Theorem 6.8,

r
XEE
k-0 (;,.....it)c(I.....r)
2)'x
'1
ik
2<p<,.
< X F1, (I
2<p<.
x
<<
(log Y)2
6.4
171
Let sk(xl,... , xr) be the elementary symmetric polynomial of degree k in r

variables. For any nonnegative real numbers x1, ... , Xr we have
Sk(X1,...,Xr) -
xii ...Xik
(it.....ik)ct1..... r)
(X1+...+Xr)k
k!
_ (S1(X1,...,Xr))k
k!
< (k)kS1(XI,...,Xr)k
since (k/e)k < k!. Therefore,
(-2)k
Pik
k-m+1 {ii..... ik)c{1..... r} Al
2k
<x E
Pik
k-,n+1 {ii.....ik)c{1.....r) Pit

r
<xT
F 21...(2
\Pi,
k-m+l fit..... ik)c11.....r) \Pit /
r
2)
sX E Sk(2
< X kE Ck) kS
(2
1
Lr
k-m+l
2 )k
Pl .
k CP1
(e)
Pr
+...+
<X
2e
P/
<x r, Ccloglogy)k
k-m+1
(P<.V
k-m+1
where c is an absolute positive constant. If we choose the even integer m so that
m > 2c log logy,

then
r, Cc log logy)k < x

k-m+1
2k <
k-m+1
x
2111
172
6.
Since r is the number of odd primes less than or equal to y, it follows that 2r < Y,
and we get the following estimate for the third term:
M
k-0
k-0
j2k (rk)
< 1:(2r)' << (2r)'" < y"'.
Combining these three estimates, we obtain
n2(x) < Y + (log y)2 + 2"' + Ym < (log y)2 + 2m + y"'.
(6.16)
where the implied constant is absolute, y is any real number satisfying
5<y<x,
(6.17)
and m is any even integer such that
m > 2c log logy.
(6.18)
Let c' = max{2c, (log 2)- }, and let
y=exp`
log x
xW..f
3c' log log x
and
m - 2[c' loglogx].
The number y satisfies conditions (6.17) and (6.18) for x sufficiently large. We
estimate the three terms in (6.16) with these values of y and m. Since
log y s
log x
3c' log log x'
we obtain the main term

x
eel
(log y)2
x (log log X)2
(log
X)2
Next, since c' > (log 2)-l and

m - 2[c' log log x ] > 2c' log log x - 2,
we obtain
2-
4x
4x
=
211 loglog.r(log x)2elog2
4x
< (logx)2'
Finally,
y < y2<' log logx = exp
(2c' log log x log x
3c' log log x
Combining these three estimates, we obtain

X (log log X)2
n2(x) <<
(log x)2
)-
2,3
6.5
Notes
173
Theorem 6.12 (Brun) Let p, , P2.... be the sequence of prime numbers p such
that p + 2 is also prime. Then
O
n-1
pn
pn + 2
5)
11
13
(17
19
< 00.
Proof. Theorem 6.11 implies that

x
n 2(x) -,-
(log X)312
for all x > 2. Therefore,

n - n2(Pn) <<
Pn
(log Pn)3iz
<_
(togPn
n)3/2
for n > 2, and so

1
P,,
n (log n)'/'
It follows that the series

00
n-1-1 Pn
<
r, 1
nn--2
Pn
+ r,
3
nn--2,
n (log n))312
converges. This completes the proof.
6.5
Notes
Dickson [22, vol. I, pp. 421-424] contains a brief account of early results concerning the Goldbach conjecture. Sinisalo [117] has verified the Goldbach conjecture by computer for all even integers up to 4 10". Wang's book Goldbach
Conjecture [ 137] is an anthology of classic papers on this subject.
Brun [7] obtained the first significant result concerning the Goldbach conjecture
in 1920. By means of the combinatorial method known today as the Brun sieve, he
proved that every sufficiently large even integer can be written as the sum of two
integers, each of which is the product of at most nine primes. Brun also obtained
the first nontrivial results concerning the twin prime conjecture. In addition to
Theorem 6.11 and Theorem 6.12, he also proved that there are infinitely many
integers n such that both n and n + 2 are the products of at most 9 primes. The
application of the Brun sieve to the twin prime conjecture follows Landau [78].
By Theorem 6.12, the sum over the reciprocals of the twin primes converges.
The sum of this infinite series is called Brun's constant, its value is estimated to be
174
6.
1.9021604 5 x 10-' (see Shanks-Wrench [ 112] and Brent [5]). It is a difficult

computational problem to determine Brun's constant to high precision. In the
process of trying to improve the estimates for Brun's constant, Nicely discovered
a defect in Intel's Pentium computer chip (see [15]).
A popular game among computational number theorists is to find explicit examples of twin primes. On October 18, 1995, Harvey Dubner announced over the
Internet that p and p + 2 are prime numbers for
p = 570, 918, 348. 105120 _ 1 = 22.33 .7.11 13 .5281 . 101120 _ 1.
The prime p has 5129 digits. This established a new record for the largest twin
prime.
For other elementary results about the distribution of prime numbers, see Ellison
and Ellison [29], Hardy and Wright [51], Ingham [66], and Tenenbaum [121].
Rosen [104] has generalized Mertens's Theorem 6.8 to algebraic number fields.
6.6
Exercises
1. Let n be a positive integer. Prove that
logn = E A(d)
dill
and
(d) logd.
A(n)
dill
2. Let cu(n) denote the number of distinct prime divisors of n. Let it > 2 and
r > 0. Prove that
1: u(d)<0<
(d)
JI,
JI
WJIQr
dAl<2r.1
3. With the notation of Theorem 6.10, prove that

N0=E(-1)'EN(1).
i/i-k
k-0
This formula is often called the inclusion-exclusion principle.
4. Use the inclusion-exclusion principle to prove that

F1/
p1n
where rp(n) is the Euler pp-function.
1)
TA(d)
dIn
6.6
Exercises
175
5. Let (D(x, y) denote the number of positive integers n < x that are not
divisible by any prime p < y. Prove that
4)(x, y) = )<xn 1 -
l + 2"('). << logy

x + 2"0.).
6. Prove that
r
r<p<c
(logx)r
7. Prove that
(log
x-1
= k!x + 0 ((logx)k) .
8. Prove that
exp
(O (logxl)
1+
(logx)
7
Das allgemeine Problem der additiven Zahlentheorie ist die Darstellbarkeit aller natiirlichen Zahlen durch eine beschrankte Anzahl von
Summanden einer gegebenen Folge von naturlichen Zahlen, z. B. der
Primzahlfolge oder der Folge der p-ten Potenzen.'
L. G. Shnirel'man [1141
7.1
The Goldbach conjecture
In a letter to Euler in 1742, Goldbach conjectured that every positive even integer
n > 2 is the sum of two primes. Euler replied that he believed the conjecture
but could not prove it. It is still unproven, but it has been confirmed by computer
calculations for even integers up to 4. 1011.
In 1930, Shnirel'man proved that every integer greater than one is the sum of
a bounded number of primes. This is a great theorem, the first significant result
on the Goldbach conjecture. Shnirel'man used purely combinatorial methods: the
Brun sieve and a theorem about the density of the sum of two sets of integers.
We shall prove Shnirel'man's theorem in this chapter. Instead of the Brun sieve,
however, we shall use a sieve method due to Selberg, which is also completely
'The general problem in additive number theory is the representation of the natural
numbers as the sum of a bounded number of terms from a given sequence of natural numbers,
e.g. the sequence of prime numbers or the sequence of p-th powers.
178
7.
elementary but more elegant and in many cases more powerful than Brun's original
sieve argument.
7.2
The Selberg sieve
Lemma 7.1 (Cauchy-Schwarz inequality) Let a,, ..., a,,, bl, ..., b,, be real
numbers. Then
n
(Easbi)2
< (Ea)
?Eb2
If aj f 0 for some j, then
2)
(Eai&i)2
(Ea)
(Eb)
i-1
i-I
i-I
if and only if there is a real number t such that bi - tai for all i - 1, ... , n.
Proof. Since
(aibj -ajb;)Z
0<
1:5i <j fn
2aiajbibj +ajb?)
(aj
I<i<j<n
n
j-I
i-1
>
i-1
we have
(Ea:tui)2
(Ea),
<
i-I
b2/
Moreover,
(Eaibs)2 v
(tap) (Eb)
if and only if
aibj - ajbi
for all i f j. In this case, if a j ' 0 for some j, let t - b j/aj. Then
bi - b' ai - tai
ajj
for i - 1, ... , n. This completes the proof.
7.2
The Selberg sieve
179
Lemma 7.2 Let a, , ... , a be positive real numbers and b , , ... , b, be any real
numbers. The minimum value of the quadratic form
subject to the linear constraint

(7.1)
is
2)
Tbi
nr=
ai
and this value is attained if and only if

mb,
Yi =a;for all i = 1,...,n.
Proof. Let y, , ... , y be real numbers that satisfy (7.1). By the CauchySchwartz inequality, we have
2
1=
(biY);
2
b,
a; Yr
2 ) ( 11
'
a,
Y'a;Y?
r-t
and so
2
aiY??
bi
=m.
Moreover,
if and only if there exists a real number t such that, for all i = 1, ... , n,
tb,
a;
or, equivalently,
tb;
Y,=-.
a,
180
7.
This implies that

it
l-Ebjy,=t1: b2` -1
,-i
aj
and so
t=m
and
m b;
Y, =-.
a,
Conversely, if y; - mb; /a; for all i, then E"_i b, y; = I and Q(yi, ... , yn) - m.
Theorem 7.1 (Selberg sieve) Let A be a finite sequence of integers, and let JAI
denote the number of terms of the sequence. Let P be a set of primes. For any real
number z > 2, let
P(z)-flp.
peP
The "sieving function"
S(A,P,z)
denotes the number of terms of the sequence A that are not divisible by any prime
p E P such that p < z. For every square free positive integer d, let IAdI denote
the number of terms of the sequence A that are divisible by d. Let g(k) be a
multiplicative function such that
0 < g(p) gl (m)

Then
S(A, P. z) <
G(z) +
E 3dtlr(d)I,
d<:2
A Pt:)
where a(d) is the number of distinct prime divisors of d.

Proof. Since g is a multiplicative function, we have, by Theorem A.7,
g([di, d2))g((di,d2)) = g(dl)g(d2)

for all positive integers d, and d2.
(7.2)
7.2
The Selberg sieve
181
Let z > 2. For every divisor d of P(z), we shall choose a real number A(d)
subject only to the conditions that
and
A(d) - 0
d>z
for all
Since
A(d)
>0
d I (u. P(z ))
for all nonnegative integers a and

z
A(d)l
(di(a.P(z
if (a, P(z)) - 1,
=1
))
it follows that
S(A, T', z)
E 1
alA
Ia.
E E A(d)
aeA
dl(a.P(:))
1: 1:
aEA
E A(d))A(d2)
d,W
d2W
d1iP(;) d2IP(:)
E A(d))A(d2)
uE A
d1.d21 P(z)
A(di)A(d2)IA1d,.d211
d,.d2IP(z)
E A(d))A(d2)(g([di, d2])IAI +r([di, d2]))

dl .d21 P(z)
JAI
d1.d21P(z)
- IAI
g([di, d2])A(dl)A(d2) + E A(d,)A(d2)r([d,, d2])

d,.d21P(z)
1
g((dl, d2))
g(di)A(d1)g(d2)A(d2)
d,.d2IPW
A(d))A(d2)r([di, d2])
d,.d2a1.d,IPl;)
- IAIQ+R,
where
Q-
d,.d2 ;
d .d,IP(z
g((di, d2))
g(di)A(d))g(d2)A(d2)
182
7.
and
R = E X(di)k(d2)r([d,, d2))
dl d2 ac
dl d2, PQ)
Let D be the set of all positive divisors of P(z) that are strictly less than z, that
is,
D-(kIP(z): 1 <k <z).

Then D is a divisor-closed set of square-free integers. If k E D, then 0 < g(k) < 1
since 0 < g(p) 0 and f(kik2) = f(ki)f(k2) if ki,k2 E D and (k1,k2) = 1. By

Mobius inversion (Theorem A.19), we have
1
_ Ef
8(()
(7.4)
(d).
dlk
Then
Q - d1,d2ED
8((di, d2))
E E f(k)g(d1)A(dj)g(d2)A(d2)
d1.d2ED AJdI
A W2
_ >2 f(k)
g(d1)A(di)g(d2).k(d2)
d1.J2eD
kED
Aldl .Ad2
- E f(k)
8(d)A(d)
deD
kED
Ald
- 1 f (k)Y2
kED
where
Yk de D
AIJ
Thus, Q is a quadratic form in the variables Yk.

The set D is finite and divisor-closed. By Mobius inversion (Theorem A.22),
we have
g(d)A(d) _ E a t d f Yk =
AfD
'ilk
\ /
E A(k)Yk
AED
dll
(7.5)
7.2
The Selberg sieve
183
In particular, for d = 1 we obtain
1: u(k)Yk = 1.
(7.6)
kED
We define
F(z) _
f (k)
kED
By Lemma 7.2, the minimum value of the quadratic form
Q= J f (k)Y?
kED
subject to the linear constraint (7.6) is

A(k)2
F(z)'
kE f(k)
(ED f(k)
and this minimum is attained when
(k)
F(z)f (k)
Yk =
We insert these values of Yk into (7.5) to compute,A(d) as follows:

A(d) =
(d)
g(d) tED
p(k)Yk
JIt
(d)
(de)yde
g(d)
JAIPI:)
u(d)
g(d)
lt(df)
<<:/J
(df)
F(z)f(de)
JAIP(:)
(d)
f(d)g(d)F(z)
f(f)
Jt;PI:)
(d)FF,(z)
f(d)g(d)F(z)'
where
Fe(z) =
EM
t
!J
Jt PI_I
In the preceding calculation, we used the fact that if df divides P(z), then d and f
are relatively prime since P(z) is square-free. We shall use this fact again to prove
184
7.
that IA(d)I < 1. Let d be any positive divisor of P(z). Then
F(z)
kEV
f (k)
157
Lid
I
lm <;
lm Pl.)
Ilmd)-t
I
t1df(e)
f(m)
lmIP(:)
Ft
)
IM A- I
1
f(m)
t1d f(f)
em I P(; )
Lid
M)
f (M)
Fd(z) 1
Lid
Fd(z),
f (d)
f(e)
f(d/ )
tad
Fd(Z)
f(d)g(d)
by (7.4), and so
Fd(Z)
f(d)g(d)F(z) -
1.
By Exercise 1, for any square-free integer d there are exactly 3w(d) ordered pairs of
positive integers d1, d2 such that [d1, d2] - d. If d1, d2 < z, then d - [d1, d2] < z2.
If d, and d2 divide P(z), then d - [d1, d2] is a square-free number that also divides
P(z). Therefore,
IRI -
E k(dj)X(d2)r([dj,d2])
d, ,d2 <:
dl.d2IP(:)
E Ir([d,,d2])I
dl,d2 <:
d1.d21P(:)
7.2
The Selberg sieve
185
< > 3`(d)Ir(d)I

d<:2
dIP(:)
and so
<
FJAI
(z) +
S(A, P, z)
3o(dII rdl
d<;
,PP(.)
To obtain the upper bound (7.2) for the sieving function S(A, P, z), it is enough
to prove that F(z) >- G(z). Let gl (k) be a completely multiplicative function such
that
gi(p) - g(p)
for all primes p E P.
By (7.3),
1
F(z) a
keD f(k)
1: g(k) fl(l -
g(P))-'
plk
kEV
Egi(k)fl(l
kED
-g,(P))-'
plk
00
Egl(k)Fj Egl(P)r
plk r-0
kED
- E g, (k) jl
plk r-O
kED
s E gl(k) E00 gi(e)

kED
r-I
p r+plk
gI(kf)
keD
r-)
P11-pIA
DO
kED
iln
p;(-f R)eplk
00
1)
M-1
X)
P14./L)- Plk
gi (m)
PI-PEr
E1
AED
Al.
plm/A-PIA
186
7.
-ren
gl(m)
- G(z),
since, in the last inner sum, we can always choose k to be the "square-free kernel"
of m, that is, the product of the distinct primes dividing m. This completes the
proof of the theorem.
7.3
Applications of the sieve
In this section, we shall obtain an upper bound for the number of representations
of an even integer as the sum of two primes. We also derive an upper bound for the
number of representations of an even integer N as the difference of two primes,
that is, an upper bound for the number of primes p < x such that p + N is also
prime.
Theorem 7.2 Let N be an even integer, and let r(N) denote the number of
representations of N as the sum of two primes. Then
r(N) <<
N
N)z
(log
I,
PIN
p JJJ
where the implied constant is absolute.
Proof. The representation function r(N) counts the number of primes p < N
such that N - p is also prime. Let
a - n(N - n).
Then
A - (ah)n-1
is a finite sequence of integers with I A I - N terms. Let P be the set of all prime
numbers. Let
2<z<vW.
The sieving function S(A, P, z) denotes the number of terms of the sequence A
that are divisible by no prime p < z. If
,IN<n<N-I-N-,
and if a = 0 (mod p) for some prime p < z, then either n or N -n is composite.
This implies that
r(N) < 2,1N + S(A, P, z).
(7.7)
We shall use the Selberg sieve to obtain an upper bound for S(A, P, z). We continue
to use the notation of Theorem 7.1.
7.3
187
Let g(m) be the completely multiplicative function defined by
g(P) -
2/p
if p does not divide N

if p divides N.
i 1/p
(7.8)
Then gI (m) - g(m) for all m. Since N is even, 2 divides N and
0<g(p)<1
for all primes p. Also,
(mod p)
if and only if
n - 0 (mod p) or n = N (mod p).

If p does not divide N, then N 0 0
(mod p) and these two congruences are

distinct. If p divides N, then N - 0 (mod p) and these two congruences are the
same. Let
d-p ..pkgi...gt
be a square-free integer, where the primes p; divide N and the primes qj do not
divide N. Then
t
g(d) -
Since a = 0 (mod d) if and only if a = 0 (mod p) for every prime p dividing

d, it follows from the Chinese remainder theorem that there are exactly 2t pairwise
distinct congruence classes modulo d such that a - 0 (mod d) if and only if n
belongs to one of these 2' classes. Therefore,
IAdI - IAIg(d)+r(d),
where
Ir(d)I < 2t < 2w(d)

By the Selberg sieve,
S(A, P, z) <
+ 5T 3.(d>Ir(d)I,
GJAI
(z)
J<:Z
d (Pin
where
G(z) -
g(m)
M <z
and w(d) is the number of distinct prime divisors of d. Let

t
p;r'
qjfr
(7.9)
188
7.
where the primes pi divide N and the primes qj do not divide N. Then
k
(l
,_1
pi
.,
j_1
qj
:,
2s,+...+.<<
Let dN(m) denote the number of positive divisors of m that are relatively prime to
N. Then
t
dN(m) - d
(') -fl(sj
+1)<fl2`/-2`1+...+,c-.
Therefore,
dN (m)
and so
dN(m)
G(z) - > g(m) >

Since
m <:
M <z
1)-1-
O0
F1 n(1-p
pa+p!N
it follows that
1j(1PIN
1) IG(z)>Edv(m)
P
o0
M<_
_,
P11-PIN
00
EdN(m)
E
M<
mt
,-1
pIr-PIN
00
E dN(m)
M <:
Mlr
00
dN(m)
W
W-1
Mlr
vu./.O pIN
>E
E dN(m).
w<; W
vu.-,I
Let
k
w -
p;
and
k
qj ,
m - fl p;
i-I
j-I
pIN
7.3
189
where the primes pi divide N and the primes qj do not divide N. Since m divides
w, it follows that 0 < r; < u; for all i, 0 < sj < vj for all j, and
k
u,-r,
i-1
e
vJ-S;
i-1
Since every prime divisor of w/m divides N, it follows that no prime qj divides
w/m, and so sj - vj for all j. Therefore,
t
p,r;
m
-1
and
qj
i-1
dN(m) - fl(vj + 1).

1-1
For each integer w, the number of such divisors m is

t
Fl(ui + l).
It follows that for every positive integer w < z, we have
m
i-1
;-1
i-1
E dN(m)- E fl(vj +1)-fl(u;+1)JJ(vj +1)-d(w),

--
FII-/.O pIN
where the divisor function d(w) counts the number of all positive divisors of w.
Let
z-N"8.
From Theorem A.13 we obtain
>> (log z)2 >> (log N)2.
fj (1\ - 1P )-j G(z) > EW d<Z(w)
w
PIN
Equivalently,
1
G(z)
IAI
<<
(logNN)2 F1
PIN
C1
P)
1
(log N)2
<<
(logN)2
C1
PIN
PZ
H C1 + P
PIN
PIN
( 1+-P
7.
190
since the infinite product 11'P.2 (I - p-2) converges.

To find an upper bound for the remainder, we use (7.9) to obtain
R
3"Ad)2",(d) <
3`Xd)lr(d)l <
2
dlp _
6 d).
d<z2
d<:2
dIY(,l
Since
2"Ad) <d
and
6",(d)
db961 log2 < Z2log6/log2,
it follows that
R < E Z2log6/log2 <
2+2 log 6/ log 2 < Z7.2 - N9/10
d<z2
since z - N 1/8. Then
S(A, P, z)
(log N)2 PIN
1+
),N91 10
1+(log N)2 F1
P
and so
r(N) < 2111 + S(A, P, z) <<
(log N)2PIN
U
(1 +
P
Theorem 73 Let N be a positive even integer, and let 7rN (x) denote the number
of primes p up to x such that p + N is also prime. Then
nN(x) << (log X)2
1 (1 + P I ,
PIN
///
where the implied constant is absolute.

Proof. The proof is similar to the proof of Theorem 7.2. It starts as follows. Let
A-{a,, : l <n <x}

be the finite sequence of integers
a - n(n + N).
Then IAl m [x]. Let P be the set of all prime numbers. For any z satisfying
2<z<f,
we let S(A, P, z) denote the number of terms of the sequence A that are divisible
by no prime p < z. If
n>f
7.4
Shnirel'man density
191
and a, - 0 (mod p) for some prime p < z, then either n or n + N is composite.

This implies that
rrN(x) <
+ S(A, P, z).
We again use the Selberg sieve to obtain an upper bound for S(A, P, z). Let
d a pi ... pAgi ... qt

be a square-free integer, where the primes pi divide N and the primes qj do not
divide N. Let I Ad I denote the number of terms of the sequence A that are divisible
by d. For every square-free integer d,
IAaI -
gJAI
(d)
+r(d),
where g(d) is the completely multiplicative function defined by (7.8), and
Ir(d)I < 2' < 2"(d)

Then
S(A, P, z) < G(I) + )T 3-Jd)Ir(d)I,

z
eiPin
where
G(z) -
E g(m)
M<z
'
The proof continues exactly as above.
In the case where N = 2, we obtain the following improvement of Brun's

Theorem 6.11.
Theorem 7.4 Let 7r2(x) denote the number of twin primes up to x. Then
7r (x) 2
7.4
(logx)2'
Shnirel'man density
Let A be a set of integers. For any real number x, let A(x) denote the number of
positive elements of A not exceeding x, that is,
A(x) - E 1.
4c A
IV<,
The function A(x) is called the counting function of the set A. For x > 0 we have
0<A(x)<[x]
192
7.
and so
A(x)
x
0<
<1.
The Shnirel'man density of the set A, denoted a(A), is defined by
a(A) =
inf
A(n)
n-1,2.3....
Clearly,
0<a(A)<1
for every set A of integers. If a(A) - a, then
A(n) > an
for all n= 1,2,3,.... If I gA,then A(1)-0and so a(A)=0.

If A contains every positive integer, then A(n) = n for all n > I and so Cr (A) = 1.
If m A for some m > l ,then A(m) <m- land

a(A)<A(m)<1-
1 <1.
m
Thus, a(A) = I if and only if A contains every positive integer.

If A and B are sets of integers, the sumset A + B is the set consisting of all
integers of the form a + b, where a E A and b E B. If A 1, ... , A,, are h sets of
integers, then
Al+A2
denotes the set of all integers of the form a1 + a2 +
+ as,, where a, E A; for
i=1,2,...,h.IfA;=Afori=1,2,...,h,welet
h times
The set A is called a basis of order h if hA contains every nonnegative integer, that
is, if every nonnegative integer can be represented as the sum of h not necessarily
distinct elements of A. The set A is called a basis of finite order if A is a basis of
order h for some h > 1.

Shnirel'man density is an important additive measure of the size of a set of
integers. In particular, the set A is a basis of order h if and only if a(hA) = 1, and
the set A is a basis of finite order if and only if a(hA) = I for some h > 1.
Shnirel'man made the simple but extraordinarily powerful discovery that if A
is a set of integers that contains 0 and has positive Shnirel'man density, then A is
a basis of finite order.
Lemma 7.3 Let A and B be sets of integers such that 0 c A, 0 E B. If n > 0 and
A(n) + B(n) > n, then n E A + B.
7.4 Shnirel'man density
193
Proof. If n E A, then n= n+ 0 E A+ B. Similarly, if n E B, then n- 0+ n E

A + B.
Suppose that n ' A U B. Define sets A' and B' by
A'-{n-a:a E A, 1 <a <n - 1}

and
B'- (b:bEB,I <b<n-1).

Then IA'I - A(n) since n A, and IB'I - B(n) since n ' B. Moreover,
A'UB'C[1,n-1].
Since
IA'I + IB'I - A(n) + B(n) >- n,

it follows that
A'f1B'./0.
Therefore, n- a- b for some a E A and b E B, and son - a+ b E A+ B.
Lemma 7.4 Let A and B be sets of integers such that 0 E A and 0 E B. If
a(A) +a(B) > 1, then n E A + B for every nonnegative integer n.
Proof. Let a(A) - a and a(B) - 0. If n > 0, then
A(n)+B(n) ? (a+)4)n > n,

and Lemma 7.3 implies that n E A + B.
Lemma 7.5 Let A be a set of integers such that 0 E A and a(A) > 1/2. Then A
is a basis of order 2.
Proof. This follows immediately from Lemma 7.4 with A - B.
Theorem 7.5 (Shnirel'man) Let A and B be sets of integers such that 0 E A and
0 E B.Leta(A)-aanda(B)=P. Then
a(A + B) > a +,8 - afi.
(7.10)
Proof. Let n > 1. Let ao - O and let
I <a, <... <ak <n

be the k - A(n) positive elements of A that do not exceed n. Since 0 E B, it
follows that a, -a,+0 E A+Bfori - I....,k.Fori -0,...,k- 1, let

I <b, <...<br, <ar+i - ai - I
be the r, - B(a,+1 - a; - 1) positive elements of B less than a;+j - a,. Then
a, <a,+b1 <... <a,+br, <ai+l
194
7.
and
ai +bj E A+B
forj=1,...,ri. Let
I <bi <... <br4 <n - ak
be the rA = B(n - ak) positive elements of B not exceeding n - aR. Then
ak <ak+bi <... <ak+brl <n

and
aA+bj E A + B
forj = 1, ... , rk. It follows that

(A + B)(n) > A(n) + E B(ai+i - ai - 1) + B(n - ak)
i-o
A(n) + P E(ai+i - ai - 1) + P(n - ak )

i-o
A(n)+PE(ai+, -ai)+fl(n - ak) - fik

i-o
= A(n)+fn - fk
= A(n)+$n -,A(n)
(I - f)A(n)+$n
> (I -8)an+j9n
(a + P - afi)n
and so
(A+B)(n) >
of +P -a$.
n
Therefore,
a(A+B)n-1.2....
inf (A+B)(n) >a+ap.
n

Inequality (7.10) can be expressed as follows:
1 -a(A+B) <(1 -a(A))(1 -a(B)).
(7.11)
The following theorem generalizes this inequality to the sum of any finite number
of sets of integers.
Theorem 7.6 Let h > 1 , and let A1, ..., A,, be sets of integers such that 0 E Ai
fori = I__ h. Then

1 - a(A1 +... + A,,) < fl(1 - a(Ai))
7.5
195
Proof. This is by induction on h. Let a(Ai) - a, for i - 1, ..., h. For It - 1,

there is nothing to prove, and for h - 2 it is inequality (7.11).
Let h > 3, and assume that the theorem holds for h - 1. Let A 1, ... , Ah be It
sets of integers such that 0 E Ai for all i. Let B - A2 + + Ah. It follows from
the induction hypothesis that
h
<jj(1-a(Ai)),
1 -a(B)= I
i-2
and so
1 -a(Ai+B)
(1 - a(Aj))(1 - a(B))
h
< (1 -a(AI))fl(1 -a(Ai))

i-2
fl(1 - a(A1))
i-i
Theorem 7.7 (Shnirel'man) Let A be a set of integers such that 0 E A and

a(A) > 0. Then A is a basis of finite order.
Proof. Leta(A) - a > 0.Then 0 < 1 -a < 1, and so
0<(1-a)t<1/2
for some integer t > 1. By Theorem 7.6,
1 - a(LA) < (1 - a(A))` - (1 - a)t < 1/2,

and so
a(LA) > 1/2.

Let h - 2e. It follows from Lemma 7.5 that the set CA is a basis of order 2, and so
A is a basis of order 2e = h. This completes the proof.
7.5
We shall apply Shnirel'man's criterion for a set of integers to be a basis of finite

order to prove that every integer greater than one is a sum of a bounded number of
primes. We begin by proving that the set consisting of 0, 1, and the numbers that
can be represented as the sum of two primes has positive Shnirel'man density. To
do this, we need estimates for the average number of representations of an integer
as the sum of two primes.
196
7.
The Shnirel'man--Goldbach theorem
Lemma 7.6 Let r(N) denote the number of representations of the integer N as
the sum of two primes. Then
x2
r(N) >>
(log x )z
N <x
Proof. If p and q are primes such that p, q < x/2, then p + q < x. Therefore,
r(N)
n(x/2)' >>
>>
(log(x/2))2
N<r
(logx)2
by Chebyshev's theorem (Theorem 6.3).
Lemma 7.7 Let r(N) denote the number of representations of N as the sum of
two primes. Then
xs
IIr(N)2
<<
(logx)a
N<x
Proof. By Theorem 7.2, if N is even, then

r(N) <<
1+
(log N)2 pH
(log N)2 d` d
This inequality also holds for odd integers, since an odd integer N can be written
as the sum of two primes if and only if N - 2 is prime, in which case r(N) - 2.
In the following calculation, we use the fact that
didz
[d1, d2] - (d i, d2)
> (did2)I12.
Then
(log N)4
x2
(logx)a
N2
dIN
(:1)
N<x
dIN d J
(log X)4 N<xd'ivd=N didz
<
x2
(logx)a d,,d2<x did2
El
di I-V d2 N
x2
El
(logx)a d12 x didz
N5
Idl,d2II N"
7.5
197
z'-
(logx)4
d, d2<<
dld2 (dl, d2]
x3
x)4
(log
d,.d2<x dl 2d2
2
x3
E dd/2 )
1
(logx)4
d<x
X3
<<
(log x)4
Theorem 7.8 The set

A - (0, 1) U (p + q : p, q primes)
has positive Shnirel'man density.
Proof. Let r(N) denote the number of representations of N as the sum of two
primes. By the Cauchy-Schwarz inequality, we have
(r(N))2
N<x
< Ev<.1 EN<x

r(N)2 < A(x) N<x r(N)2.
By Lemma 7.6 and Lemma 7.7,
1 (I:N<x r(N))2
A(x)
X
X F-N<.c r(N)2
1 (ion
x
ci
(log
1.
This means that there exists a number c, > 0 such that A(x) > clx for all x > X.
Since 1 belongs to the set A, it follows that there exists a number c2 > 0 such that
A(x) > c2x for 1 < x < xo. Therefore, A(x) > min(cl, c2)x for all x > 1, and so
the Shnirel'man density of A is positive. This completes the proof.
Theorem 7.9 (Goldbach-Shnirel'man) Every integer greater than one is the sum
of a bounded number of primes.
Proof. We have shown that the set
A = (0, 1) U (p + q : p, q
primes)
198
7.
has positive Shnirel'man density. By Theorem 7.7, there exists an integer h such
that every nonnegative integer is the sum of exactly h elements of A. Let N > 2.
Then N - 2 > 0, so for some integers k and l with k + f < h there exist t pairs
of primes pi, qj such that
N-2- I+ +1+(pt+qt)+...+(pt+qt)
k
Let k - 2m + r, where r - 0 or 1. If r - 0, then
+(p1+q1)+...+(pt+qt)
N-2
m+1
Ifr-1,then
N-2 +
In both cases, N is a sum of
21+m+1 <3h
primes. This completes the proof.
Theorem 7.10 Let Q be a set of primes that contains a positive proportion of the
primes, that is,
Q(x) > 07r(x)
for some 8 > 0 and all sufficiently large x. Then every sufficiently lure inte,;er is
the sum of a bounded number of primes belonging to Q.
Proof. We shall first show that the set
A(Q) - {0, 11 U (p + q : p, q E Q}
has positive Shnirel'man density. Let r(N) denote the number of representations
of N as the sum of two primes, and let r(,(N) denote the number of representations
of N as the sum of two primes belonging to Q. Then
E rQ(N) ? (Q(x/2))2 > (9n(x))2 >>

N <x
x2
(logx)2
By Lemma 7.7,
rQ(N)2 <
N<x
N<x
x
r(N)2 << (log X)4
It follows exactly as in the proof of Theorem 7.8 that the set A(Q) has positive
Shnirel'man density. Therefore, A(Q) is a basis of finite order. It follows that there
7.6
Romanov's theorem
199
exists a number h, such that every nonnegative integer is the sum of h, elements
of Q U {0, 1).
Choose two primes p,, p2 E Q. By Exercise 3, there exists an integer no no(p,, p2) such that every integer n > no can be written in the form
n - eI(n)P1 +e2(n)p2,
where e,(n) and e2(n) are nonnegative integers. Let
h2 - max{e, (n) + e2(n) : n - no, ... , no + h, },

and let
h-h,+h2.
If N > no, then N -no can be written as the sum of at most h, elements of QU 11),
that is,
N - no - I+ +
k
where
k+e < h,.

Then
no + k - eI(n)P1 +e2(n)p2,
where e, (n) + e2(n) < h2, and so
f,(n)P1 +e2(n)P2+Pi, +... Pi

is a sum of
e+e,(n)+e2(n) < h, +h2 -h

primes belonging to the set Q. This completes the proof.
7.6 Romanov's theorem

Let a be an integer, a > 2. We investigate how many numbers N up to x can be
written in the form
N - p + ak,
(7.12)
where p is a prime and k is a positive integer. Let r(N) be the number of representations of N in this form. Since the number of positive powers of a up to x is
<< log x and the number of primes up to x is 7r(x) << x/ log x, it follows that
Er(N)-I{p+ak<x}I<< IogxI
N<x
I-x.
\logx JJ)
200
7.
Let
A=(p+ak: pprime andk> 11,

and let A(x) be the counting function of the set A. In this section, we shall prove
a remarkable theorem of Romanov that the lower asymptotic density of the set A
is positive, that is, there exists a constant c > 0 such that
A(x) > cx
for all sufficiently large x. This means that a positive proportion of the natural
numbers can be represented in the form (7.12).
Lemma 7.8 Let a be an integer, a > 2. For every integer d > 1 such that
(a, d) - 1, let e(d) denote the exponent of a modulo d, that is, the smallest integer
such that
I
a`(d)
(mod d).
Then the series
converges.
Proof. If (a, d) - 1 and e(d) - k, then
ak - 1 (mod d),
and sod divides ak - 1. Since ak - I has only finitely many divisors, it follows
that there are only finitely many numbers d such that e(d) - k. For x > 2, let
D-D(x)-fl(ak
- 1),
k<r
and let n - w(D) be the number of distinct prime divisors of D. Let
E(x) - F,
k<x
N2(drl
The number d appears in this double sum at most once, and if d appears, then d
divides ak - I for some k < x, so d divides D. It follows that
-n1\ <H(l+\
n
E(x) < E
din
u2(dhl
pID
(I+
pill
,_1
where p1, P2..... pn are the first n prime numbers. Since
2n -2 &XD) < D - fJ (ak - 1 <

k<x
flak < ax(x+l)J2 < ax2

k<r
7.6
Romanov's theorem
201
it follows that
2
x2
By Chebyshev (Theorem 6.4),
log p << logn << log x,

and so, by Mertens's formula (Theorem 6.8),
(1+ }
E(x) <<
p//
PSP.
1 \-]
p
By partial summation,
.Id.r
k<x
E(t)dt
E(x) +
Ek
ta.drl
t2
N2w.1
<<
logx+
logtdt
t2
<< 1,
and so the series

00
-k
al
21J.1
e-I
wall-I
21d.1
de(d)
converges. This completes the proof.
Lemma 7.9 Let a be an integer, a > 2, and let r(N) denote the number of
N-p+ak
where p is a prime and k is a positive integer. Then
1: r(N)2 << X.
N<,x
Proof. Since r(N)2 is equal to the number of quadruples (pi, p2, k1, k2) such
that
Pi +a k, - P2 +a k2 = N,
202
7.
it follows that > N,, r(N)2 is equal to the number of quadruples (p,, P2. k1, k2)
such that
P1+ak' -p2+ak2 <x.

This does not exceed the number of solutions of the equation
Pz - Pt -a
-a
k,
k2
with pl, P2 < x and k1, k2 < logx/ logo.

Choose positive integers k, f k2, and let
h = ak' - ak2
Then h is a nonzero, even integer. The number of solutions of the equation
P2 - P1 - ak'
- ak2 - h
with P1, P2 < x is at most the number of primes p, < x such that p, + h is also
prime. By Theorem 7.3, this is
nh(x) (log X)2 1-7h 1 + pX
If k2 > k, , then
h _ aki lake-k,
- t)
and
11(1+ 1) _
(I+IP )
P l+
pla
n (I+-P'plate
pl(a',
1)
P1
(+....i
P1
) PI(a t2 h
1+1
<<
pI(a+2
where the implied constant depends on a. Similarly, if k, > k2, then
h - -ak2 (ake-k2 - 1)
and
n(1+')
Plh
(1+'-) s
pl(aì-h.-1)
PI(at2-Aji_1)
Finally, if k2 - k, , the number of solutions of the equation
p2-
p,-ak2-aki-0
(1+ 11.
P
7.6
Romanov's theorem
with p1, p2 < x and 1 < k2 < log x / log a is

Jr (x) log x
log a
<<
x.
It follows that
fl
r(N)2 << x+2 E
k1pl(n+z
Nx
_I)
<<x+logx E f (+--)
p
pI(n`-1)
1<k<
x +logx
I<k< lot
Nld l'-I
To estimate the last term, we observe that
if and only if
if and only if
Then
E r(N)2 << x+logx

N :5x
1<k<1
dl(+-1)
02(d)-l
- x+logx
1d)-I
- x+logx
E
y,z(dl-I
r(d)lk
log x
< x + log x
de(d)loga
(.d)-I
<< x +
(log x)' 2
(..d )-I
<< x
since the infinite series converges by Lemma 7.8.
de(d)
203
204
7.
Lemma 7.10 Let a be an integer, a > 2, and let r(N) denote the number of
N = p+ak
where p is a prime and k is a positive integer. Then
1: r(N) >> x.
N <x
Proof. If p < x/2 and ak < x/2, then p + ak < x, so
1: r(N) >> n(x/2) log(x/2) >> x.

N <x
Theorem 7.11 (Romanov) Let a be an integer, a > 2. Let
A=(p+ak : pprime andk> 1),

and let A(x) be the counting function of the set A. There exists a constant c > 0
such that
A(x) > cx
for all sufficiently large x.
Proof. We use the Cauchy-Schwarz inequality. By Lemma 7.10 and Lemma 7.9,
there exist positive numbers ci and c2 such that, for x sufficiently large,
2
(cIx)2 < E r(N)

(N <x
< A(x)
r(N)2 < c2xA(x)

N <x
and so
A(x) > cx.
7.7
Covering congruences
Choosing a - 2 in Romanov's theorem, we see that a positive proportion of the

natural numbers can be written in the form p + 2k. The only even numbers of this
form are 2 + 2k, and they constitute a very sparse subset of the even integers, a
subset of density zero, so almost all of the integers of the form p + 2k are odd.
We shall prove that there exists an infinite arithmetic progression of odd natural
numbers, none of which can be written in the form p + 2k . To do this, we introduce
the concept of covering congruences for the integers.
7.7
205
Let
<mt
1 <m, <m2
be a strictly increasing finite sequence of integers, and let a,, ... , at be any integers. Then the f congruence classes ai (mod mi) form a system of covering
congruences if, for every integer k, there exists at least one i such that
k - ai
(mod mi ).
This means that the congruence classes ai

sense that
(7.13)
(mod m;) cover the integers in the
Z-U(kEZ:k-a; (modmj)).
i-t
It is an essential part of the definition of covering congruences that the moduli
mj are pairwise distinct integers greater than one. Here is a simple example of a
system of covering congruences.
Lemma 7.11 The six congruences
(mod 2)
(mod 3)
(mod 4)
(mod 8)
(mod 12)
23
(mod 24)
0
0
1
form a set of covering congruences.

Proof. First, we show that each of the 24 integers 0, 1, ... , 23 satisfies at least
one of these six congruences. Every even integer k satisfies k = 0 (mod 2). For
odd integers, we have
1-I (mod4)
30 (mod 3)
5
(mod 4)
7=7 (mod 12)
90 (mod 3)
11.3 (mod 8)
13
(mod 4)
15
(mod 3)
17
(mod 4)
19=7 (mod 12)
21m0 (mod3)
23 = 23
(mod 24).
206
7.
For every integer k, there is a unique integer r E {0, 1, ... , 23) such that
k-r
(mod 24).
r =_ ai
(mod m,).
Choose i so that
where a, (mod mi) is one of our six congruences. Each of the six moduli 2, 3,
4, 6, 12, and 24 divides 24, so m, divides 24 and
k=_r
(modmi).
Therefore,
k - ai (mod mi).
Theorem 7.12 (Erd6s) There exists an infinite arithmetic progression of odd

positive integers, none of which is of the form p + 2'.
Proof. We shall use the system of covering congruences ai (mod mi) constructed in Lemma 7.11. For each of the six moduli mi in this system, we choose
distinct primes pi such that
2'"' = 1
(mod pi),
22
(mod 3)
as follows:
1
212 - 1
(mod 7)
(mod 5)
(mod 17)
(mod 13)
224 = 1
(mod 241).
23
24
28 = 1
Let
e-max(pi}-241
and
m - 2t .3.7.5.17. 13.241.
By the Chinese remainder theorem, there exists a unique congruence class r
(mod m) such that r - I
(mod 2t) and r = 2
This means that
r
r
(mod 2f )
(mod 3)
r2 (mod7)
(mod pi) for i - 1, ... , 6.
7.7
21
23
(mod 5)
(mod 17)
27
(mod 13)
r = 223
207
(mod 241),
where the exponents in the powers of 2 are the least nonnegative residues ai in
the six congruence classes in the system of covering congruences. Since r is odd
and the modulus m is even, it follows that every integer in the congruence class r
(mod m) is odd.
Let N be an integer in the congruence class r (mod m) such that
N>21+f.
Let k be a positive integer such that 2k < N. There is a congruence class ai
(mod mi) in the system of covering congruences such that
k - ai
(mod mi )
so k - a; +miui for some integer u;. Since
2"'' - I
(mod pi),
we have
2k - 2, 2m, mi =- 2`
(mod pi).
Since
N-r
(mod pi)
and
r - 2' (mod pi),

it follows that
N - r =- 2' = 2k
(mod pi),
and so
N-2k+piv
for some positive integer v. If k < e, then
piv-N-2k > N-2t >e-max{pi}> pi

for i - 1, ... , 6, and so v > 1. If k > e, then
N - 2k - N - 1 (mod 2t)
and so
piv-N-2k-1+21w>2t >f> pi
and v > 1. In both cases, N - 2k is composite. This completes the proof.
208
7.
7.8
Notes
Shnirel'man's fundamental paper was published first in Russian [113] and then
expanded and published in German [ 114]. By Shnirel'man's constant we mean the
smallest number h such that every integer greater than one is the sum of at most
h primes. Using the Brun sieve, Shnirel'man proved that this constant is finite.
The best estimate for Shnirel'man's constant is due to Ramare [100], who has
proved that every even integer is the sum of at most six primes. It follows that
Shnirel'man's constant is at most seven. The Goldbach conjecture implies that
Shirel'man's constant is three.
In this chapter, I use the Selberg sieve instead of the Brun sieve to prove the
Goldbach-Shnirel'man theorem. See Hua [63] for a nice account of this approach.
Landau [76, 77] gives Shnirel'man's original method. Theorem 7.10, the generalization of the Goldbach-Shnirel'man theorem to dense subsets of the primes, is
due to Nathanson [90].
Selberg introduced his sieve in a beautiful short paper [1091. I use Selberg's
original proof of the sieve inequality (7.2). See Selberg's Collected Papers[ 110,
111 ] for his papers on sieve theory. Prachar [97] contains a nice exposition of the
Selberg sieve, with many applications. The standard references on sieve methods
are the monographs of Halberstam and Richert [44] and Motohashi [87].
Romanov's theorem appears in the paper [103]. Romanov also proved that, for
a fixed exponent k, the set of integers of the form p + nA has positive density.
The proof of Theorem 7.8 of Romanov's theorem was simplified by Erdos and
Turan [30] and Erd6s [33].
Erd6s [32] invented covering congruences and used them to construct the infinite
arithmetic progression of odd positive integers not of the form p+2A, as described
in Theorem 7.12. Crocker [ 16] proved that there exists an infinite set of odd positive
integers that cannot be represented as the sum of a prime and two positive powers
of 2. Crocker's set is sparse. It is an open problem to determine if there exists an
infinite arithmetic progression of odd positive integers not of the form p+2A' +2A2.
There are many unsolved problems concerning covering congruences. It is not
known, for example, whether there exists a system of covering congruences alI of
whose moduli are odd. Nor is it known whether, for any number M, there exists
a system of covering congruences all of whose moduli are greater than M. The
best result is due to Choi [12], who proved that there exists a system of covering
congruences with smallest modulus 20.
7.9
Exercises
1. Prove that for any square-free integer d there are exactly 3`00) pairs of
positive integers d1, d2 such that [d1, d2] = d.
7.9
Exercises
209
2. Let w(n) denote the number of distinct prime divisors of n. Let n > 2 and
r > 0. Prove that
E t(d) `- 0 <
u(d).
3. Let ai and a2 be relatively prime positive integers. Prove that there exists an
integer no = no(aI, a2) such that every integer n > no can be written in the
form
n = e1(n)a1 + e2(n)a2
for some nonnegative integers fi(n), e2(n).
4. Construct a system of covering congruences whose moduli are 2, 3, 4, 6,

and 12.
5. Let us call an integer n exceptional if n - 2A is prime for all positive integers
k < log n/ log 2. Find all exceptional numbers up to 105. Erdos [32] has
written that "it seems likely that 105 is the largest exceptional integer."
6. Let Jai
(mod rn;)
Prove that
i = 1, ... , k} be a system of covering congruences.

1
The method which I discovered in 1937 for estimating sums over

primes permits, in the first instance, the evaluation of an estimate for
the simplest of such sums, i.e. a sum of the type:
e2' "'.
p<N
This estimate in combination with the previously known theorems

concerning the distribution of primes in arithmetic progressions ...
paved the way for establishing unconditionally the asymptotic formula of Hardy and Littlewood in the Goldbach ternary representation
problem.
I. M. Vinogradov [ 135, page 3651
8.1
Vinogradov's theorem
Vinogradov proved that every sufficiently large odd integer is the sum of three
primes. In addition, he obtained an asymptotic formula for the number of representations of an odd integer as the sum of three prime numbers. Vinogradov's
theorem is one of the great results in additive prime number theory. The principal ingredients of the proof are the circle method and an estimate of a certain
exponential sum over prime numbers.
212
8.
The counting function for the number of representations of an odd integer N as

the sum of three primes is
r(N) - E
1.
pi +pl+p?-N
The following is Vinogradov's asymptotic formula for r(N).

Theorem 8.1 (Vinogradov) There exists an arithmeticfunction 6(N) and positive
constants c, and c2 such that
c, < CA(N) < C2
for all sufficiently large odd integers N, and

g2
r(N) - 6(N)
( 1+0
N)3
2(1
(lo8l
gg
NN
The arithmetic function CA(N) is called the `singular series for the ternary
Goldbach problem.
8.2
The singular series
We begin by studying the arithmetic function

00
6(N) - E u(gkg(N)
(
P It
q-1
(8.1)
)3
where
q
e(aN/q)
cq(N) -
is Ramanujan's sum (A.2). The function 6(N) is called the singular series for the
ternary Goldbach problem.
Theorem 8.2 The singular series CA(N) converges absolutely and uniformly in
N and has the Euler product
6(N)=11(1+
(p-1)3
}pU(l - p2-3p+3 I.
1
There exist positive constants c1 and c2 such that
c1 < 6(N) < C2

for all positive integers N. Moreover, for any e > 0,
6(N, Q) = E
q:5 Q
A(q)cg(N)
- CA(N) + 0
tp(q)3
(Q-0-E)),
(8.2)
8.3
213
Proof. Clearly, cq(N) << co(q). By Theorem A.16,
p(q)>q'
for e > 0 and all sufficiently large integers q, and so
p(q)cg(N)
(q)3
(q)2
q2-`
Thus, the singular series converges absolutely and uniformly in N. Moreover,
6(N) - C7(N. Q) <<
<<
q>Q P(q)2
q>Q
ff <<
By Theorem A.24, cq(N) is a multiplicative function of q and
p - I if p divides N
-1
if p does not divide N.
cP(N)
Since the arithmetic function
A(q )cg (N )
p(q )3
is multiplicative in q and (p1) - 0 for j > 2, it follows from Theorem A.28 that
the singular series has the Euler product
M
6(N) =
P"(PJ)cP,(N)
1+
cP(P')3
j-1
1 - cP(N) l
cP(P)3l
I+
pu;
(P - 1)1) pIN
H
1-
-f 1+(p1)3
(p - 1)2
pIN
p2-3p+3
and so there exist positive constants ci and c2 such that
cl < 6(N) < c2

for all positive integers N. This completes the proof.
8.3
As in the proof of the Hardy-Littlewood asymptotic formula for Waring's problem,

we decompose the unit interval [0, 11 into two disjoint sets: the major arcs )71 and
the minor arcs m.
214
8.
L.etB>0and
Q - (log N)B
(8.3)
For
1<q:5 Q,
0<a<q,
and
(a, q) - 1,
the major arc 931(q, a) is the interval consisting of all real numbers at E [0, 11 such
that
If a E W(q, a) fl 931(q, a') and a/q (a'/q', then laq' - a'q I > I and
Q2
I
- a'q I
< qq' < laq'qq'
a-a'
-- a
q
a'
q'
2Q
q'
N'
or, equivalently,
N < 2Q3 = 2(log N)3e

This is impossible for N sufficiently large. Therefore, the major arcs 9R(q, a) are
pairwise disjoint for large N. The set of major arcs is
Q
9N-U U 9(q,a)c[0,1]
(u q)-1
and the set of minor arcs is
m-[0,1]\9Y1.
We consider a weighted sum over the representations of N as a sum of three
primes:
R(N) -
109 P1 log P2 log P3 Pi+P:+P!-N
Vinogradov obtained an asymptotic formula for R(N), from which Theorem 8.1
will follow by an elementary argument. We can use the circle method to express
the representation function R(N) as the integral of a trigonometric polynomial
over the major and minor arcs. Let
F(a) - E(log p)e(pa).

p<N
(8.4)
8.4
The integral over the major arcs
215
This exponential sum over primes is the generating function for R(N), and
log plog plog P3 -
R (N)
pi+pi+p3-N
F(a)3e(-Na)da
rm
J F(a)3e(-Na)da + J F(a)3e(-Na)da.
31
The main term in Vinogradov's theorem will come from the integral over the major
arcs, and the integral over the minor arcs will be negligible.
8.4 The integral over the major arcs

Just as in the Hardy-Littlewood asymptotic formula, the integral over the major
arcs in Vinogradov's theorem is (except for a small error term) the product of the
singular series CA(N) and an integral J(N). In this case, the integral J(N) is very
easy to evaluate.
Lemma 8.1 Let

N
e(mfl).
U(P) _
m-1
Then
J(N) =
f3 - 22 + O(N).
1/2
I/2
Proof. By Theorem 5.1, the number of representations of N as the sum of three

positive integers is
1/2
J(N) -
11/2
u()
N
3e(-N)d-1/2
e((mI +m2+m3 - N)fi)dfi

I/2 nri-I MT-1 In3-1
(Nl)
2
N2
= 2 + O(N).
In the next lemma we shall apply the Siegel-Walfisz theorem on the distribution of prime numbers in arithmetic progressions. A proof can be found in
Davenport [ 19].
216
8.
Theorem 8.3 (Siegel-Walfisz) If q > 1 and (q, a) - 1, then, for any C > 0,
t9(x;q,a)- E
loge'V(q)+O((log
,5,
x) /
,,ed q)
for all x > 2, where the implied constant depends only on C.
Lemma 8.2 Let
Fr(a) -
(log p)e(Pa).
Let B and C be positive real numbers. If 1 < q < Q = (log N)B and (q. a) then
Fr(a/q)- (q)x+O
QN
W(q)
\(IogN)c.
for I < x < N, where the implied constant depends only on B and C.
Proof. Let p - r
(mod q). Then p divides q if and only if (r. q)
1, and so
log p < log q.
E (log P)e(Pa/q) - E(log P)e(Pa/q) <

r_I
(r,q)
P-.
vs=
I I- (0d 4)
P14
Therefore,
F.
( )-
(log P)e (p
r-I p.' P5.

(mad q l
(log p)e
(ra
q
(rArI P- (mod q)
(ra)
PK'
_I
/
/
+ O(logq)
(log P) + O(log Q)
v" (mod q)
(ra) 0(x; q, r) + O(log

q
Q)
x
(ra)/x +0( (logx)C
q
(G(q)
+ O (1)JJlI + O(log Q)
(9) x (_qx
QN
A) + D
#(q)x
(log N)C
since, by Theorem A.24, cq(a) - (a) if (q, a) - 1.
+ O(log Q)
8.4
217
Lemma 8.3 Let B and C be positive real numbers with C > 2B. Ifa E 93t(q, a)
and 0 - a - a/q, then

gz N
F(a)
u(0)+0
(P(q)
((loN)c /
and
GQg2N3
((9 3u(T)3+O
F(a)3
N)c
where the implied constants depend only on B and C.
Proof. Ifa E 9)7(q, a), then a - a/q + P, where Ifi) < Q/N. Let
X(m) 1
if m - p is prime
log p
0
otherwise.
If l < x < N, then
F(a) - (q)
W(q)
u (f) - E log pe(pa) p<N
- E k(m)e(ma) -
lx(q)
V(q)
L e (mp)
m-1
(q) E e (ma)
co(q) m-l
M-1
= E .(m)e (ma +mol _
M-1
E \X(m)e
m_1
a)
\q
A(q)e(m,6)
1-1 `,o(9)
(P(q))
A(x) -
()-(m)e
(ma)
- N(q)1
W(q)
1<M<
A(m)e
1<m<,
(ma )
(q
F, (ql
-O
- -(q)x + O
(P(q)
(_I
(P(q)
G(q)x+0(1)
QN
(log N)c
By partial summation, we obtain
F(a) - k(q)
(q) u (/f) - A(N)e(N)4) - 27rip
<< IA(N)I +
Q2N
<<
(log N)c
fN
A (x)e(xp)dx
IN max{A(x) : 1 < x < N}
218
8.
Clearly, Iu(P)I < N. Since C > 2B, we have

Q2N
(log N)c - (log
N)c-2B
< N,
and the estimate for F(a)3 follows immediately. This completes the proof.
Theorem 8.4 For any positive numbers B, C, and e with C > 2B, the integral
over the major arcs is
F(a)3e(-Na)da = CA(N) 2 +0 (log
log N)c-58 )
1911
where the implied constants depend only on B, C, and e.
Proof. We note that the length of the major arc '91(q, a) is Q/N if q = I and
2Q/N if q > 2. By Lemma 8.3,
F(a)3
4Q
L.
-a
(F.(a)3
e(-Na)da
(q) u
(a - a
G(q)3
e(-Na)da
Q2N3
(log N)c
-0
la.y}I
9
Q3N2
F-
(log N)c
q-Q
Q5N2
)3)
t,,.yrl
q
<
(a
(q.a)
<E E
q<-Q
cqq) u
da
- (log N)c
N2
(log
N)c-5B
< Q/N and
If
q<Q
m.yrl
A(q)
Pcq)
1:
q
=
q<Q
(93 /'alq+QlN
V(q)
alq-Q/N
a
9
)3
e(-Na)da
a)3
u a- q
e(-Na)da
yrl
V(q)
q<Q
ua--
(q.a) (
(q)3
Q/N
e(-Nalq)
b, .y}t
fQ/N
u(1)3e(-N$)dfl
8.4
1: A(q)c,(-N)
IQ/n,
u(p)3e(-Nfl)dp
(P(q)3
q<Q
Q/N
QIN
C7(N, Q) J
Q/N
u(P)3e(-Nfl)dfl.
By Lemma 4.7, if IfI < 1/2, then

u(fl) << IfI-1
and
u2
1/x
)dfl << f/N l u(fl)I3dfl
L IN
Q/N
Similarly,
N2
Q/N
u(#)3e(-NO)dp <<
Q2.
1/x
By Lemma 8.1,
u(fl)3e(-Nf )df Q/N

Q/N
1 /2
u($)3e(-Nf )dfl + O(N2Q-2)
1/2
- 2 +O(N)+OI Q
/N2\
N2
2 +OIQ2I.
By Theorem 8.2,
6(N, Q) - 6(N) + O I
Q I _E
Therefore,
J97t
F(a)3e(-Na)da
Q/N
- 6(N, Q) J Q/N
u(fl)3e(-Nf)dfl
+0
N2
Gog N)c-5B
N2
- C7(N)2
+0 (Q1-E) +0 ((log N)c-se )
N
+0
- 6(N)2x
((IogN)(I-e)B
) + 0 Gog N)c-se)
219
220
8.
8.5
An exponential sum over primes
To estimate the integral over the minor arcs, we shall apply Vinogradov's estimate
for the exponential sum F(a). The proof is based on a combinatorial identity of
Vaughan.
Theorem 8.5 (Vinogradov) If

a
a--
q2'
where a and q are integers such that I < q < N and (a, q) = 1, then
F(a) << I -N2 + N"5 + N 1/2q 1/2) (log N)4 .
The proof is divided into a series of lemmas. The first is an identity involving
arithmetic functions of two variables and truncated sums of the Mobius function.
Lemma 8.4 (Vaughan's identity) For u > 1, let

M,, (k)
(d).
Let ct(k, e) be an arithmetic function of two variables. Then
1: 4)(1, e) + 1: 1: M,I(k)(D(k, e) _
u<t<N
u<k<N ,<t<N
(d)ct(dm, e).
d<u u<t<'v
N
- d m<- Id
Proof. We shall evaluate the sum

N
S - 1: 1: M,,(k)O(k, e)
k-I 11 <t <'4
in two different ways. Since
ifn=1
1: (d)
din
otherwise,
it follows that
1
ifk=1
ifl <k<u.
Therefore,
4)(1, e)+ 1:
S=
u<t<N
1: M,I(k)<D(k, e).
u<k<Nu<t<
8.5
221
On the other hand, interchanging summations and letting k - dm, we obtain

N
S-> E E (d)O(k,e)
dit
k-1
u<I<.v
d<..
- d<u
E Ea4 E
(d)c(k, e)
u<e<
A(d)c(dm, e)
(d)O(dm, e).
d<um<a u<e<T
d<u a<e<j m<re
Lemma 8.5 Let A(e) be the von Mangoldt function. For every real number a,
F(a) - S, - S2 - S3+O(N1/2),
where
S, -
u.(d)A(e)e(adem),
d<N2"5 e<e m<<j
E (d)A(e)e(adem),
S2 -
d<N215 t<N215 m<
to
and
S3 - 1:
1:
MN2 s(k)A(e)e(ake).
k>N215 N2J5<e<N/k
Proof. We apply Vaughan's identity with

u - N2/5
and
4)(k, e) - A(e)e(ake).
The first term in Vaughan's identity is
4'(l, e) -
A(e)e(ae)
N2'5<e<N
u<e<N
- E A(e)e(ae) e-1
A(f)e(ae)
C <N215
E (log p)e(apk) + 0 (N2/5 log N)

pA <N
- E (log p)e(ap) + E (log p)e(apk) + 0 (N2/5 log N)

p<N
y<!v'
222
8.
=F(a)+O E log p
+ O (N2/5 log N)
>2
(1:
- F(a) + O
[-]
rN
log
+ O (N2/S log N)
= F(a) + 0 (7r (N'"2) log N) + 0 (N2/Slog N)
= F(a) + 0 (N 1/2),
since
7r(N1/2) <<
N1/2
log N
by Chebyshev (Theorem 6.3).

The second term in Vaughan's identity is simply
E MN2/5(k)A(f)e(akf) - S3.
N2/5<k<N N2/5<f< i
The third term in Vaughan's identity is
1:
1: E (d)A(e)e(adfm)
d<N215 N2/5<f<m<
e - td
E (d)A(e)e((Ydfm)
_
d<N2/5 f<
m< J
(d)A(I)e(adfm)
d<N2/s f<N215 II1< 1 11
=S, -S2.
In the next three lemmas, we find upper bounds for the sums S1, S2, and S3.
Lemma 8.6 If
a
a-q
- q2'
where 1 < q < N and (a, q) - 1, then

IS, I << I
N+
N2/5 +
ql (log N)2.
Proof. Let u = N2/5. Since F-tir A(P) = log r, we have
S, _
(d)A(P)e(ctdfm)
d <u t5
,,, <
8.5
223
1: 1: (d)A(e)e(adtm)
d<u tm<N/d
_ E E 4(d)e(adr) E A(t)
d<u r<N/d
!(r
EA(d) Y' e(adr)log r
r<N/d
d<u
1: e(adr) log r
<<
r<N/d
d<u
We compute the inner sum by writing the logarithm as an integral and interchanging
summations:
dX
E e(adr) log r = Ee(adr) f

r<N/d
r<N/d
(N/dl
r-2
(N/dI(N/dl
-E
s-2
s-2
dx
- E e(adr) T
S
-1 X
E - e(adr)-dx
X
r-.c
(N/d)
_
s-2
[N/d)
e(adr)
s-I
r-s
dX
X
By Lemma 4.7, the geometric progression inside the integral sign is bounded above
by
(Nldl
e(adr)<< min{ d,IladIr-'r.

r-c
111
and so
E e(ad r) log r << min

r<N/d
(d
\
II ad II -' 1 log N.
JJJ
By Lemma 4.10. we have
Ymin\d ,IIadII-')I\ q
N+N2/5+q)log
d<u
Therefore,
Si <<
d/<u
min (., Iladll-') log N
< I N + N2/5 + q) (log N)2.

q
N.
224
8.
Lemma 8.7 If
a
a--
q2'
where 1 < q < N and (a, q) = 1, then

IS21 <<
N
(+N4/5+q)(1ogN)2.
Proof. If d < N2/5 and f < N2/5, then df < N4/5. Making the substitution
k = d t, we obtain
A(d)A(C)e(adCm)
S2 =
d <N215 1:<N215 m < er
e(akm)
k<N413
ni
Ec(d)A(f)
Yn
N/k
J.r
Since
1: (d)A(f) << E A(f) < E A(f) = logk << log N,

i.dr
A-dt
d.r<.N jl3
ilk
J.r<N2/3
it follows again from Lemma 4.10 that
S2 <<log N E E e(akm)
k<N4," m<N/k
<< E mint
N,
IIak11-'
log N
k<N+/3
<<
IN
-+N4/5+q (logN)2.
\q
Lemma 8.8 If
a
a-q
- q2'
where I < q < N and (a, q) = 1, then

N
IS31 <<
Proof. Let u = N2/5 and
h=
log N
5 log 2
+1.
8.5
225
Then N115 < 2h < 2N)15 and h << log N. If i < h, then 2iu < 2N315 << N. If
N2/5 < e < N/k, then
k < N/f < N3/5 = N'l5u < 2hu,

and so
S3 =
N='5<1<N/k
h
A(e)e(ake)
u<f<N/k
i-1 2'-Iu<k<2'u
h
i-1
where
A(e)e(ake).
S3,i 2'-'u<k<2'u
u<f<N/k
By the Cauchy-Schwarz inequality,

2
IS3.i I2 < E IMu(k)l2
A(e)e(ake)
2'-'u<k<2'u u<l<N/k
21-'u<k<21u
We shall estimate these sums separately.

To estimate the first sum in (8.5), we observe that
IM,,(k)I - E (d)
<El <d(k).
eu
J...
where d(k) is the divisor function. It follows from Theorem A.14 that
d(k)2
IM.(k)12 <
2'-'u<k<2'u
2'-1u<k<21u
<< 2'u (log 2'u)3
<< 2'u(logN)3.
Next, we estimate the second sum in (8.5). We have
2
A(e)e(akf)
2'"'u<k<2'u u<7-N/A
A(e)A(m)e(ak(e - m))
_
2i-1u<k<2'u a<l<N/k a<m<N/k
1:
1:
u<f< 'r u<nt<-'T
e(ak(e - m)),
A(e)A(m)
&E!(t.m)
(8.5)
226
8.
where 1(e, m) is the interval of consecutive integers k such that
2'-lu <k <min(2'u, e, N)

Clearly,
II (e, m) I < 2'-'u,

and so
e(ak(e - m)) <<min
(2'-l
u, II a(e - m)II -') .
kEI(1,m)
Since O < A(t), A(m) < log N for all integers 1,M E [1, N], we have
2
A(e)e(ake)
2'-'u<k<2;u
<<
Lt:5NIk
>2
A(e)A(m)min (2'-'u, Ila(e - m)II
u <1<N/(2'-'u) u<m<N/(2'-'u)
<< (log N)2
>2
>2
min (2'-'u, Ila(e - m)II-')
u<t<N/(2'-1 u) u<m<N/(2'-'u)
Let j - e - m with u < e, m < N/(2'-'u). Then I j I < N/2'-'u, and the number
of representations of an integer j in this form is at most N/2'-' u. By Lemma 4.10,
we have
2
1:
57
A(e)e(ake)
2'-'u<k<2'u Iu<1<N/k
> min (2'-' u, IIajII -')
<< (log N)22'Nu
i <j<N/2'-1u
E min(
<< (log
N,Ilajll-
1<j<N/2'-'u
<< Flu
(N + Z,Nu +q) (log
N)3.
Inserting this into inequality (8.5), we obtain

IS3.1 12 << (2'u(log N)3)
<< N2 (log N)6
Flu
Cl
(N
q- u+
Nu +q (logN)3
q
/
Therefore,
1/z
IS3.1I << N(log N)3 ( ?1/2 +
+
TI 1-5
N1/2
8.6
Proof of the asymptotic formula
227
Since h << log N, we have

S3., << (log N)4 1 -
S3
z + N4/5 +q 1/2 N 1/2

Finally, we obtain Vinogradov's estimate for the exponential sum F(a) by inserting our estimates for the sums S1, S2, and S3 into Lemma 8.5. This completes
the proof of Theorem 8.5.
8.6
We can now estimate the integral over the minor arcs.
Theorem 8.6 For any B > 0, we have
I
Im
F((Y )3e(-aN)da <<
N2
(log
N)(H/z)-s'
where the implied constant depends only on B.
Proof. Let a E m = [0, 11 \ 9)2. By Dirichlet's theorem (Theorem 4.1), for

any real number a there exists a fraction a/q E [0, 1] with I < q < N/Q and
(a, q) = I such that
a--a
<
Q
< min
qN -
Q, 2
Nq
1
If q < Q, then of E 9)t(q, a) c 9)7, which is false. Therefore,
Q<q<
By Theorem 8.5,
F(a) << I qN +N 4/5 +N 1/2 q 1/2 ) (log N)4

<
<<
(log N)a1'-
+ N4/5 + N1/2
(ogN)8)
i/2
(log N)4
N
(log N)(e 2)-4'
Since O(N) = E,,<,, log p << N by Theorem 6.3, we have
f' I F(a)I`da =
0
(log p)2 < log N 1: log p << N log N,
228
8.
and so
I F(a)I3da << sup(IF(a)l : a E m)
I F(a)I2da
(log N 8/2)-4 fo IF(a)I2da

N2
<<
(log N)(B/2)-5
Theorem 8.7 (Vinogradov) Let CA(N) be the singular series for the ternary
Goldbach problem. For all suffciently large odd integers N and for every A > 0,
N2
N2
R(N) - C7(N) 2 +0 ((1OA)'

g
where the implied constant depends only on A.
Proof. It follows from Theorem 8.4 and Theorem 8.6 that, for any positive
numbers B, C. and a with C > 2B,
I
R(N) - f F(a)3e(-Na)da
0
= J F(a)3e(-Na)da + fm F(a)3e(-Na)da
=C7(N)
+0
22 +O ((log N)(1-E)B)
NZ
Gog N)c-sB
) + O Gog
N2
N)(B/2)-5)
where the implied constants depend only on B, C, and E. For any A > 0, let
B=2A+l0andC=A+5B.Lete-I/2.Then
min((] - e)B, C - 5B, (B/2) - 5) - A,
and so
NZ
NZ
R(N) = Ch. (N) 2 +0 ((log
N)'

We can now derive Vinogradov's asymptotic formula for r(N).
Proof of Theorem 8.1. We get an upper bound for R(N) as follows:
R(N) _ E log P1 log P2 109 P3

pi+p:+ps-N
(log N)3
p'+p;+p3-N
_ (log N)3r(N).
8.6
229
For 0 < S < 1/2, let rs(N) denote the number of representations of N in the form
N - pI + p2 + p3 such that p, < N'-8 for some i. Then
r3(N)
<3i1
P1'P2. P3..C
Pi SVi-a
1:-, (
<<
<
PI
:5N' -d
P2+P3-N-Pi
(l)
(
p2
< n(N'-s)n(N)
N2-s
<<
(log N)2'
We can now get a lower bound for R(N):
E log Pi log P2 log P3
R(N)
PI P2.P3-'V
PI.
PZ.P3>At 1-a
(1 - S)3(log N)3
P I P2.P3-N
P1 P21P31.0-9
(I - 3)3 (log N)3(r(N) - rs(N))

N2-s
>> (1 - 3)3 (log N)3 (r(N)
- (log N)2)
Therefore,
(log N)3r(N) < (1 - S)-3 R(N) + (log
IfO<8<1/2,then1/2<1-5<land
I -(I-S)3
0<(I-S)-3-I=
(I - S)3
N)N2-s
- 8(1-(I-5)3) <245.
By Theorem 8.7, R(N) << N2 and so
0 < (log N)3r(N) - R(N) < ((1 - 3)-3 - 1) R(N) + (log

N)N2-b
<< SR(N) + (log

<< SN2//+ (log
= N2 I 6+
N)N2-s
N)N2-s
log N
N3
\\
This inequality holds for all S E (0, 1/2), and the implied constant does not depend
on S. Let
2 log log N
log N
230
8.
Then
S+
log N _ 2 log log N

log N
Ns
log N + (log N)2
logloglogN N
and so
0 < (log N)3 r(N) - R(N) <<
N2 log log N
log N
Let A > 1. By Theorem 8.7,
(log N)3r(N) - R(N) + O

CA(N)
N22+0
N2
C7(N)
(NZ log log N

/log N
)
(logN)A l+
OI
N2 g Ng N
(1+0 (log oloNN

g
Dividing by (log N)3, we obtain
r(N)
N2
C7(N)2(log
N)3
log log N
(1 + O ( log N
))
8.7
Notes
For Vinogradov's original papers, see [ 132, 133]. Vaughan [ 124] greatly simplified
Vinogradov's estimate for the exponential sum F(a) (Theorem 8.5), and it is
Vaughan's proof that is given in this book. There are many good expositions of
Vinogradov's theorem. See, for example, the books of Davenport [ 19], Ellison [29),
Estermann [38], Hua [64],Vaughan [125], and Vinogradov [135].
Vinogradov's theorem implies that almost all positive even integers can be written as the sum of two primes. This was observed independently by Chudakov [ 14],
van der Corput [ 123], and Estermann [37]. Let E denote the set of even integers
greater than two that cannot be written as the sum of two primes. The set E is called
the exceptional set for the Goldbach conjecture. Let E(x) denote the number of
integers in E not exceeding x. The theorem of Chudakov, van der Corput, and
Estermann states that E(x) <<A x/(logx)A for every A > 0. Montgomery and
Vaughan [84] proved that there exists S < 1 such that E(x) << xd. Of course, if
the Goldbach conjecture is true, then E(x) - 0 for all x.
8.8
Exercise
1. Let h > 3. Find an asymptotic formula for the number of representations of

a positive integer N =_ h (mod 2) as a sum of h prime numbers.
The linear sieve
We often apply, consciously or not, some kind of sieve procedure

whenever the subject of investigation is not directly recognizable. We
begin by making a long list of suspects, and then we sort it out gradually by excluding obvious cases with respect to available information.
The process of exclusion itself may yield new data which influences
our decision about what to exclude or include in the next run. When no
clue is provided to drive us further, the process terminates and we are
left with objects which can be examined by other means to determine
their exact identity. These universal ideas were formalized in the context of arithmetic back in the second century B.C. by Eratosthenes,
and are still used today.
H. Iwaniec [68]
9.1
A general sieve
In the next chapter, we shall prove Chen's theorem that every sufficiently large
even integer can be written as the sum of a prime and a number that is the product
of at most two primes. The proof will require more sophisticated sieve estimates
than those obtained from the Selberg sieve in Chapter 7.
We begin by generalizing our concept of a sieve. Let A - {a(n)]', be an

arithmetic function such that
a(n) > 0
for all n
(9.1)
232
9.
The linear sieve
and
IAI - Ta(n) < no.
(9.2)
n-1
Let P be a set of prime numbers and let z be a real number, z > 2. The set 'is
called the sieving range, and the number z is called the sieving level. Let
P(z)-flp.
pEP
P-1
The sieving function is
a(n).
S(A, P, z) (n. P(z))-I
The goal of sieve theory is to obtain "good" upper and lower hounds for this
function.
l'orexarnple. let A be the characteristic function of a finite set of positive integers.
that is, a(n) - I if n is in the set and a(n) - 0 if n is not in the set. Then IA! is
the cardinality of the set. The sieving function S(A. P. ;.) counts the number of
integers in the set that are not divisible by any prime p E P. P < z.. This special
case is exactly the sieving function for which we obtained. in Chapter 7. an upper
bound by means of the Sclberg sieve.
Using the fundamental property of the Mobius function, that
(d) -
(1 * )(rn) _
dim
I if m = 1
0 if m > 1,
where I denotes the arithmetic function such that 1(n) - I for all n > 1, we obtain
Legendre's formula
a(n)
S(A, P, z) -
- Ea(n) E (d)
dl(n.P(z))
- E (d) E a(n)
dIP(:)
din
E (d)IAdl,
diP(:)
where the series
IAdI - Ea(n)
din
converges because of (9.1) and (9.2).

We shall assume that, for every n > 1, we have it multiplicative function g,, (11)
such that
0<gn(p)<I
A general sieve
9.1
233
for every prime p E P. Then

0 < g,, (d) < I
for every integer d that is the product of distinct primes p E P. For such integers
d, the series
1: a(n)gn(d)
it
converges, and we can define the remainder r(d) by
IAdI = E a(n)g,,(d) + r(d).

n
Inserting this into Legendre's formula, we obtain
S(A,P,z)- E A(d)IAdI
dIP(:)
- E u(d)
(a(n);i(d)+r(d))
dIP()
->a(n) E
tt(d)r(d)
dIP(z)
dIP(z)
-1: a(n) fl (1 -gn(P))+

PIP(:)
It
GL(d)r(d)
dIP(:)
a(n)VV(z) + R(z),
n
where
VV(z) _ F1 (I - gn(P))
PIP(:)
and
R(z) - E A(d)r(d).
dIP(z)
If P(z) has a large number of divisors, the remainder term R(z) in Legendre's
formula may be too large to give useful estimates for S(A. P, z). For example, let
A be the characteristic function of the set of all positive integers not exceeding x,
and let P be the set of all prime numbers. Let
gn (d) -
for all n. Then
Vi(z)-fl11
P<z \
for all n > 1. Moreover. for all d > 1,
0< rd
234
9.
The linear sieve
and so
Ir(d)I <
IR(z)I <
2n(z)
dIP(t)
It follows from Legendre's formula that the number of integers up to .r divisible

by no prime less than z is
S(A.P,z)-[x] fl1 1 P<z\
I+O(2"(.))
By Mertens's formula (Theorem 6.8),

1
fl ( 1 - p
e-Y
logz
<
1+O (logz,,
and so the remainder term will be larger than the main term unless z is very small
compared to x.
The sieve idea is to reduce the size of the error term by replacing the Mobius
function with carefully constructed arithmetic functions A'(d) and A (d) such that
A+(1)-A-(I)- 1
(9.4)
(1 *A+)(m)-EA+(d)>0
(9.5)
and, for every m > 2,
dim
and
(1 * A-)(m) - E A-(d) < 0,
(9.6)
dim
Let A+(d) and A-(d) be arithmetic functions that satisfy (9.4), (9.5), and (9.6). If
D is a positive number such that A+(d) - 0 for all d ? D, then the arithmetic
function A+(d) is called an upper bound sieve with support level D . Similarly, if
D is a positive number such that A'(d) - 0 for all d > 1), then the arithmetic
function A-(d) is called a lower hound sieve with support level D.
If P is a set of primes such that A'(d) = 0 whenever d is divisible by a prime not
in P, then A+(d) is called an upper hound sieve with sieving range P. Similarly,
if A-(d) - 0 whenever d is divisible by a prime not in P, then A "(d) is called a
lower bound sieve with sieving range P.
The following result is the basic sieve inequality.
Theorem 9.1 Let A+(d) bean upper bound sieve with sieving range P and support
level D, and let A-(d) be a lower bound sieve with sieving range P and support
level D. Then
oc
00
A-) + R < S(A, P. z) <n-1
a(n)Gn(z, A+) + R+,

n-l
9.1
A general sieve
235
where
G,, (z, ),) = E ,tt(d)gn(d)

dI P(z)
and
R} - > At(d)r(d).
d,P1:
d-n
Proof. Since the arithmetic function a+(d) is supported on the finite set of
integers I < d < D, it follows that the series
>2a(n)
)+(d)
dl(n,P(:))
converges. By conditions (9.4) and (9.5), the inner sum is 1 if (n, P(z)) - 1 and
nonnegative for all n. Therefore,
S(A, P, z) - E a(n)
(n. P(z))-I
1: a(n)
h+(d)
dI(n.P(z))
It
E A+(d) I: a(n)
dI P(:)
din
1 X+(d)lAdl
dIP(z)
E k+(d)
dIP(z)
If
F X+(d)Ea(n)gn(d)+
dIP(z)
L ,X'(d)r(d)
dIP(z)
Ea(n) E a+(d)gn(d)+ 1: A+(d)r(d)

dIP(:)
mvc,
d' 1)
a(n)G,, (z, a+) + R+.
The proof of the lower bound is similar.

The following result shows how to extend the sieving range of upper and lower
bound sieves by any finite set of primes.
Lemma 9.1 Let JAt(d) be upper and lower bound sieves with sieving range PI
and support level D. Let Q be a finite set of primes disjoint from PI, and let Q be
the product of all primes in Q. Every positive integer d can be written uniquely in
the form
d =dld2,
where d) is relatively prime to Q and d2 is a product of primes in Q. Define
),:':(d) - X (d))(d2).
(9.7)
236
9.
The linear sieve
Then the function A*(d) (resp. A - (d )) is an upper bound sieve (resp. lower bound
sieve) with sieving range
P=P, UQ
and support level D Q.
Let g be a multiplicative function, and let
G(z, ).}) _ E A (d)g(d)

dI P(:)
and
Then
G(z,
fit) = G(z, fit) fl (1 - g(q)).

qIQ(:)
Proof. Clearly, A*(1) = A-(1) = 1. Every positive integer m factors uniquely

into a product m = m,m2, where m, is relatively prime to Q and m2 is a product
of primes in Q. We have
D ,'(d) _
E A*(did2)
dint
diJill i d2Un2
Ai(di)1: N-(d2)>0
(1, it?),
d2 Jill 2
since
(d2)
d2 ptt2
ifm2 = 1
ifm2 > 2.
Similarly, if m = m,m2 > 1, then
(d2)<0
dint
thin:,
d2lnt2
since either m2 > I and
1: (d2) = 0,
d,int2
or m2 = 1, which implies that m, > 1, and so
EA(di)<0.
d, lilt,
Thus, the arithmetic functions A}(d) satisfy conditions (9.4), (9.5), and (9.6).
Since ,lt(d) = 0 if d is divisible by some prime not in P, it follows that the

functions At have sieving range P.
Let d = did2, where d, is relatively prime to Q and d2 is a product of primes
in Q. If d = d,d2 > DQ, then either d, > D and X (d,) = 0, or d2 > Q, which
9.1
A general sieve
237
implies that d2 is divisible by the square of some prime q E Q, and so (d2) - 0.

In both cases, ,lt(d) - 0. Therefore, the functions At(d) - 0 have support level
DQ.
Finally, since P(z) - Pi(z)Q(z),
G(z, A}) - E X (d)g(d)

d I P(z)
E E At(d(d2)g(d)d2)
d1 PI(z)d2IQ(z)
E E At(d))g(dt)p(d2)g(d2)
d,IP,(:)d., IQ(z)
E X:I (d))g(d)) E it(d2)g(d2)

d,tQ(z)
d,IP,(z)
- G(z,,li) rl (1
- g(q))
qJQ(:)

Combining Theorem 9.1 and Lemma 9.1, we obtain the following result, which
is an important refinement of the basic sieve inequality.
Theorem 9.2 Let ,lt(d) be upper and lower bound sieves with sieving range Pn
and support level D. Let IAr(d )I < 1 for all d > 1. Let Q be a finite set of primes
disjoint from PI, and let Q be the product of the primes in Q. Let P - Pn U Q.
For each n > 1, let g, (d) be a multiplicative function such that
forallpEP.
Let
G.(z,,lt) _
At(d)gR(d)
dIPi(z)
Then
S(A, P, z) -<00E a(n)Gn(z, ,lI) fl (1 n-)
R(DQ, P, z)
qIQ(z)
and
00
S(A, P, z) > Ea(n)G,, (z, A-I) fl (1 - g,,(q)) - R(DQ, P. z),

n-)
glQ(z)
where
R(DQ, P, z) -
Ir(d)I.
d' P(--)
d-7 1)Q
238
9.
The linear sieve
It often happens in applications that the arithmetic functions g (d) satisfy onesided inequalities of the form
t
fj (I
- gn(P))
log
<K
PEP
"
(logo)
Z
where K > 1 and K > 0 are constants that are independent of n, and the inequality
holds for all n and 1 < u < z. In this case we say the sieve has dimension K. The
case K - I is called the linear sieve. The goal of this chapter is to obtain upper
and lower bounds for the linear sieve that were first proved by Jurkat and Richert
(Theorem 9.7). This is the only sieve inequality that is needed for Chen's theorem.
9.2
In a combinatorial sieve, we reduce the size of the error term in Legendre's formula
by replacing the Mobius function with its truncation to a finite set of positive
integers. This idea goes back to Viggo Brun [7). We construct these truncated
functions in the following theorem.
Theorem 9.3 Let 0 > I and D > 0 be real numbers. Let D` be the set consisting
of I and all square free numbers
d-PIp2...pk
such that
Pk<...<P2<PI <D
and
Pm<(
1/
Pm
PIP2
for all odd integers m. Let D- be the set consisting of 1 and all square free numbers
d -
such that
A <... <P2 <P) < D

and
D
Pm <
I/
PI P2 Pm
for all even integers m. Then the sets D+ and D` are finite sets of square free
positive integers d < D. Let P be a set of primes: and let P(D) denote the product
of all of the primes in P that are less than D. Define the arithmetic functions k+(d)
and X -(d) as follows:
A.+(d)=
J (d) ifd E D+anddIP(D)
l0
otherwise
9.2
239
and
(d) ifd E D- and dIP(D)
X-(d)
otherwise.
I 0
Then A(d) and X-(d) are upper and lower bound sieves with sieving range P and
support level D.
Proof. The condition
D
Pm <
C PI P2
An
is equivalent to
PlP2...pm-1Pm0<D.
E D+. If k is odd, then
d - pI ... Pk-1 Pk < p, ... pk-I
Pk+fl
< D.
If k is even, then k - I is odd. Since pk < pk - , and fi > 1, we have
d-
Pi...Pk-IPk <
Pi... Pk-i <
P1...pk+p < D.
! < d < D for all d E D+.

Similarly, ifd - p, - pk E D- and k > 2, then 1 < d < D. For k - 1, we
have d - p, < D, that is, D- contains all primes strictly less than D. Therefore,
Th e r e f o r e ,
I <d < Dfor all d E D-.

The arithmetic functions A.+(d) and.-(d) are truncations of the Mobius function
(d) to certain subsets of the sets D+ and D-, respectively. Since both sets contain
1, we have
)l+(1) -),-(1) - (1) - 1.

Let m > 2. We must prove that
E A-(d) < 0 < E ;k+ (d).

d1m
(9.8)
dim
Since the functions A(d) are supported on divisors of P(D), we may assume that
m divides P(D). Let w(m) denote the number of distinct prime divisors of m. The
proof is by induction on k - w(m). If k - 1, then m - p < D for some prime

p E P, and so m E V. We have
EX-(d)=g(1)+(P)-0
dini
and
EA+(d)-(1)+h+(P)> 1 - 1 -0.
dim
This proves the lemma in the case k = 1.
240
9.
The linear sieve
Now let k > 1, and assume that inequalities (9.8) hold for all positive integers
m with k distinct prime divisors. If w(m) = k + 1, then we can write m in the form
m=goq,...gk,
where
qk < qk-1 < ... < q, < qo < D,
qo, q, , ... , qx are prime numbers in P, and qo is the greatest prime divisor of m.
Let
m
mi-=q,...qk.
qo
Since m, is a divisor of P(z) with k prime factors, it follows from the induction
hypothesis that
E A (d) < 0 <
dint,
A+(d).
dim,
Every divisor of m is of the form d or qod, where d is a divisor of m, . Therefore,
1: ,L+(d) _ T X+(d) + T, +(qod )

dint,
dim
dint I
r k+(qod)
din,,
1:
l-t (qod )
dlm,
gpdcP'
_-
,,,(d).
dlm,
gpol D'
Similarly,
1: X (d)
(d)
dlmi
gOdED-
dim
If d is a divisor of m, , then
d=p1...p1.
where p, , ... , pj are primes in P such that

pi < ... < pi < qi < qo < D.
Let D, = D/qo > 0, and let D' and D- be the sets of integers constructed from P
and D,. Let 1lt(d) and k (d) be the Mobius function truncated to the sets Di and
Di , respectively. Then qod E D+ if and only if
qo <
qo
9.2
and
goP)
for all even integers m. If
......
P(
Pm
241
An
D )
qo qo
then qod ' D+ and so
p(d)=0
VI IJE D
since the sum is empty. If

qo <
qo
then qod E D+ if and only if d E Di , and

(d)
Jm,
+ilee D'
Jlm
u(d)
1`
d;,,1,
(d) ` 0
JEDI.
by the induction hypothesis. Therefore,
E,L+(d) > 0.
din,
Similarly, qod E D- if and only if d E D', and so
1: u(d)= 1: u(d)= EA1(d)> 0.

J' i
dl,,,,
JEPI
yoJE D
This proves that X+(d) and ,k-(d) are upper and lower bound sieves with sieving
range P and support level D.
Lemma 9.2 Let P be a set of primes, and let g(d) be a multiplicative function
such that
0<g(p)<I
forallpEP.
Let
V(z) = H(1 - g(p)) =E (d)g(d).

piP(:)
PET'
Then V (z) is a decreasing function of z,
0<V(z)<I
for all z, and
g(P)V(P) = V(w) - V(z)
rE r
for all 1 < w < z.
(9.9)
242
9.
The linear sieve
Proof. It follows immediately from the definition that V(z) is decreasing and
V(z) E (0, 1) for all z.
The proof of the combinatorial identity (9.9) is by induction on the number k of
primes p E P that lie in the interval [w, z). If k = 0, then V(w) = V(z) and
1: g(P)V (P) 0.
per
If k > 1, let pI be the largest prime in the interval. Then
E g(P)V(P)- E g(P)V(P)+g(PI)V(PI)
per
per
..<pt:
'SP'PI
- V(w) - V(PI)+g(PI)V(PI)
- V (W) - (1 - g(PI))V (PI )
- V(w) - V(z)Lemma 9.3 Let P be a set of primes. For fi > I and 2 < z < D, let
Ym=Ym(8,D,PI,...,pm)=
I/6
D
CPI...Pr)
Let 0(d) be the upper and lower bound sieves constructed in Theorem 9.3, and
let
G(z, At) - E X: (d)g(d )

dIP(z)
Let
z) _
g(PI ... pn)V
P1...,Pner
)n<P,<... <P(-:
P--V.'-..
(..d2)
Then
00
G(z, k+) = V(z)+ E T, (D, z)
(9.10)
n_.
(rood 2)
and
00
G(z, A-) = V(z) - E T,,(D, z).

,,
0 (mad:)
Tn(D,z)>0
G(z, ), -) < V(z) < G(z,)+).
< log D
log z
then
forn <s-j6.
(9.11)
9.2
243
Proof. It follows from the construction of the sets D1 and the sieves X '(d) that
G(z, A+) >

dip("
dED'
(-1)kgn(PI ... Pk)
pl' <PI :.PfD
P. <t.v.wl
(na42)
and
G(z, X) e
u(d)gn(d)
dl P(:)
dED-
(-1)kgn(P) ... Pk).

P',
P. <). V..0 (mad 2)
We expand the function V(z) to obtain a partition of G(z, X+) as a sum of nonnegative functions:
V(z) -
t(d)g(d)
dlP(:)
(-1)kg(P1 ... Pk)

Pk <.... pl <;
tEP
(-1)kg(P1 ... Pk)

Pk'PI<:. Pi EP
v. <r. v..i
(mad2)
(-1)k9(P1 ... PO
Pk <...<Pi :. Pi lP
3..I
(mod
(-1)kg(Pl ... pk)

pk <. <pl
3..1
.Pi *P
(mad
00
E
..I
e_I
fond 2)
00
+E
..I
(-1)kg(PI ... Pk)
Pk < <PI <:. Pi EP
p. <.. Y. <n,..l i..d2)

,n Pn
E (-1)ng(P ... P.)

,n<pn<;pt-
)m.12)
(-1)k-ng(Pk ... Pn+l)

Pk<<Pn Ii
Pi(P
pie
Pn
..1, Itmd 2)
00
- G(z, ),+) -
F,
..)
n_I
(mad 2)
nSPn<<pl <:. PIEP
p., <,.r.. n.
..1
(mad 2)
00
1:
- G(z, Kl+) n.)
(mod 2)
T,,(D, z),
g(Pj...p)V(Pn)
244
9.
The linear sieve
where
g(Pi...P,,)V(P,,)>- 0.
Therefore,
00
E T,, (1), z) > V(z).
G(z, A') - V(z) +

.,
Similarly,
,.,J 21
G(z, A ) - V (z) - E T,, (D, z)
V (z).
n_i
.M
If
mod 2)
Yn < pn <...<Pt <z,
(9.12)
then
D < Pi ... p,,pp <
that .satisfy inequalz) is a sum over integers p;

Let D - zS. Since
z) - 0 unless c < n + ti. This completes the
ity (9.12), it follows that
proof.
9.3
Approximations
For the rest of this chapter, we shall consider only the case
8-2
in the construction of the sets Dt and the upper and lower hound sieves 0(ci).
Then
D
\Pi ... pm
r
Ym
and the functions
I/2
z) satisfy the following recursion relation.
Lemma 9.4 Let z > 2 and D be real numbers such that
s-
log D
log z
if n is odd,
if n is even.
Then
(9.13)
T1 (D, z) - V (D' I') - V (z).
Let n > 2. If n is even, or if n is odd and s > 3, then
T,,(D, z) p!r
P-
g(P)T,.-t
(D, p l
\P //
(9.14)
9.3
Approximations
245
If n is odd and I < s < 3, then

T, (D, z) _
(Dn , p )
pEP
p<OI' )
Proof. Since
Y1
(9.15)
_ (D/pl)'I2, it follows from Lemma 9.2 that
Ti (D, z) = > g(p1)V(p1)

p1EF
T, g(Pi)V(P1)
PI EP
V(D"3) - V(z).
If n is even, then
T, (D, z) _
g(P) ... pn)V (P,,)
rn_PI <pl c.P, EP

P"P; t
PI
rmrm
vm.:,-.n ,mnl zl
= T g(Pi)
9(P2 ... P,,)V
P, EP
P2
r_rmrm D;Plyz<m-
m-I-I
g(P1)T,,-i
_
P,FP
rl :
I, OJ D
D
-,
Pi)
PI
Let n be odd, n > 3. If p1 < yi = (D/pl)1"2 and p, < z = D"S, then
pi < min (D1/3 D"I.s) =
D'13
if I < s < 3
ifs>3
and the argument proceeds exactly as in the case of even integers n. This completes
the proof.
We shall now construct a sequence of continuous functions ,,(s) that will be

used later to approximate the discrete functions T,, (D, z). For s > 1, let 1Z (s) be
the open convex region of Euclidean space consisting of all points (ti, ... , tn) E Rn
such that
I
s
1,
and
I
if m <nandm - n (mod 2).
(9.16)
246
9.
The linear sieve
For n > 1 and s > 1, we define the function f (s) by the multiple integral
sf,(s)
f ... f.(S)
dt, ... dt,,

(tl ... tn)tn
(9.17)
The function fn(s) is nonnegative, continuous, and decreasing, since Rn(s2) c

Rn(s,) for s, < s2. If fn(s) > 0, then Rn(s) is nonempty, so R,,(s) contains a
point (t,, ... , tn). This point satisfies
<(n+2)t, <
n+2
s
and so
1
(9.18)
< t, < -
n+2
I t f o llows th at
fors > n + 2.
fn(s) - 0
(9.19)
It is easy to compute f, (s) and f2(s). We have f, (s) - 0 fors > 3. For 1 < s < 3,
we have
RI(s) - (1/3, 1/s)

and so
sfi(s) -
J2 -3-s.
r1/s dt,
1/3
(9.20)
t1
Similarly, f2(s) - 0 for s > 4. For 2 < s < 4, we have

R2(s) - (1142):
t,
< t, <
and
and so
1/s
sf2(s) /
(I
dt2 dt,
0/3 t2
\1 -
J1/
J1/4I\
/s
t1
1/a
11
1 - t1
+3- 1)dt,
t,
-s-3log(s- 1)+3log3-4.
The functions fn(s) satisfy the following recursion relation.
Lemma 9.5 Let n > 2. If n is even ands > 2, or if n is odd ands > 3, then
sfn(s) -
00
fn-1(t - 1)dt.
(9.21)
If n is odd and I < s < 3, then
sfn(s) - 3fn(3) - TO fn-1(t - 1)dt.
(9.22)
Approximations
9.3
247
Proof. If n is even and s > 2, or if n is odd and s > 3, then, from (9.18), we
have
A ... d to
- f ... f .(S) (t1 .
sfn (s)
. tn)tn
1/s
f/(n+2)
20'
rl <w<n.l.- Imod 2)
In the inner integral, we make the change of variables
ti -0 -11)ui-I
for i - 2,...,n. Let
1 -t,
Si-
---1,
I
/1
tj
Since tj 1 if n is even and s > 2, and st > 2 if n is odd
ands > 3. We obtain
dt2 ... dtn

0<rw <...<r2 <,,
r2.
(12 ... tn)tn
1-r,
VI<m<n..,
(mod 2)
dul ...dun_1
11)(u1 .
YI<w c.,m-Ir.-1
. un-I)un-1
(mod 2)
du1...du,_,
<Iry1
rl <i. a.m-Ian-1
(u) ... un-1)un-1
(mod 2)
dui...dun_1
I - tj ,f
-1(SI) (u1 ...
s1
-1 -1
,fn-I (sl )
- 1 fn-1
t,
Setting t - 1/t1, we obtain
tj
- J1/:
1/(n+2)
/1
11
tl
fn-,(t - 1)dt
n +2
- fn+2
n_1(t
-J
1)dt
fn_1(t - I )dt,
1)
dt,
11
248
9.
The linear sieve
since f _ I (t - 1) - 0 for t - 1 > (n - 1) + 2 by (9.19).

Let n > 3 be an odd integer. If (t1, ... ,
then tj < 1/s. Also, it
follows from inequality (9.16) with m - I that t) < 1/3, and so t1 < 1 / max(s, 3).
Therefore, if I < s < 3, then
R. (s) - R, (3)
and
...
Sf,(S) -
- f ...
1
(tl ... tn}
to
dtl ... dt
R,(S)
(11 ... In)
d!n
,(3)
- 3fr(3).
We construct the function h(s) for s > 1 as follows:
for 1 < s < 2

for 2 < s < 3
for s > 3.
e-2
h(s) -
e-S
3s"I e'S
(9.23)
It is easy to check (Exercise 8) that
h(s-1)<4h(s)
fors>2.
For s > 2, let
H(s) -
h(t - I)dt.
J5
Both h(s) and H(s) are continuous, positive, and decreasing functions on their
domains. Let
a
2h(3
H(2)
e2H(2)
2)
2e
+3e2
2
O _I
t e-dt.
We can express a in terms of the exponential integral
Ei(x) since
J-oo
e't-I dt
OC
e-'t -'dt - -Ei(-3) - 0.013048... .
We can obtain this number with technology, such as Maple, or without technology,
either by estimating the integral directly or by looking it up in old books, such as
Dwight's Mathematical Tab!es[26, page 107). We find that
a - 0.96068....
(9.24)
9.3
Approximations
249
Lemma 9.6
H(s) < ash(s)
for s > 2
(9.25)
for 1 < s < 3.
(9.26)
and
H(3) < ash(s)

Proof. Ifs > 3, then h(s - 1) < el
and
f OC
H(s) <
el-`dt - el-S -
esh(s)
< ash(s).
For 2 < s < 3, let
Ho(s) - ash(s) - H(s).

We have
s-I-I+(s-2)<es-2,
and so
s)e-S
(1 -
> -e-2.
Then
HH(s) - ah(s) + ash'(s) - H'(s)

- a(l - s)e-s + h(s - 1)
> (1 -
a)e-2
> 0,
and so Ho(s) is increasing for 2 < s < 3. Since
Ho(2) - 0
by the definition of a, it follows that
H(3) < H(s) < ash(s)
for 2 < s < 3.
Let 1 < s < 2. Since a < 1, it follows that h(2) > H(2)/2 and
H(3) - H(2) - erg - H(2) - h(2) < H22) - ah(2) < ash(2) - ash(s).
Lemma 9.7 If n is odd ands > 1, or if n is -seen ands > 2, then
fp(s) <
2e2an-'h(s).
Proof. This is by induction on n. For n - 1, we shall show that
sf, (s) < 2e2sh(s).
250
9.
The linear sieve
For I < s < 3, we have sfj (s) - 3 - s by (9.20). If 1 < s < 2, then h(s) - e-2
and
sf1(s) - 3 - s < 2 - 2e2h(s) < 2e2sh(s).
If 2 < s < 3, then h(s) - e-S and

sfi(s) - 3 - s 3, then f, (s) - 0 and

sfi(s) - 0 < 2e2sh(s).
This proves the case n - 1.

Now let n > 2, and assume that the lemma holds for n - 1. By (9.21) and (9.25),
if n is even and s > 2, or if n is odd and s > 3, then
00
sf"(S) -
Js
f"-i(t - 1)dt
< 2e2an-2
h(t - I)dt
- 2e2ax-2H(s)
< 2e2a"-2ash(s)
< 2e2an-ish(s).
By (9.22) and (9.26), if n is odd and 1 < s < 3, then
Sfn(S) - f 00 fn-,(t - 1)dt
< 2e2an-2
h(t - 1)dt
< 2e2an-2H(3)
< 2e2a"-2ash(s)
< 2e2a"-Ish(s).
Theorem 9.4 For s > 1, the function

00
F(s) - 1 +
fn(s)
,,
Unod b
is continuous and differentiable, and
F(s) - 1 + 0
(e-S)
(9.27)
9.4
The Jurkat-Richert theorem
251
For s > 2, the function

(9.28)
f(s) = 1 n.0 (..A 2)
is continuous and differentiable, and
f(s)=1+0(e-5).
Proof. By Lemma 9.7,
0 < ,,(s) < 2e2a"-lh(s) <
for s > (3 + (-1)" )/2. Therefore,

0C
(s) << e-'`.
"-I
The theorem follows immediately from this inequality.
9.4
From now on, we shall consider only arithmetic functions g(d) that satisfy the
linear sieve inequality (9.29).
Lemma 9.8 Let z > 2 and 1 < w < z. Let P be a set of primes, and let g(d) be
a multiplicative function such that
0<g(p)<1
for all pEP
and
fj (1 - g(p))-' < K
PEP
logz
(9.29)
log u
for some K > 1 and all u such that 1 < u < z. Let
V(z) = r1 (1- g(p)),

PEP
P<:
and let 1 be a continuous, increasing function on the interval [w, z]. Then
E g(p)V(p)4)(p)
P
< (K - 1)V(z)cD(z) - KV(z)J 4)(u)d
Ilogz 1
log u J
252
9.
The linear sieve
Proof. The step function
S(u) -
g(P)V(P)
PEP
is nonnegative and decreasing. By Lemma 9.2 and inequality (9.29),
S(u) - V(u) - V(z)
=I V(u)-1
JV(z)
V(z)
(I - g(P))-I - 1
V(z)
PEP
< (K ogu
g
- 1i
V(z).
Let
W<pk<pk-1<...<pl<z
be all the primes in P that lie in the interval [w, z). Then S(pk) - S(w), S(pl) g(p1)V(p1), and S(u) - 0 for p1 S(P,)cD(Pi+l)

k-1
S(Pk)4)(Pk) + > S(Pi) (4(Pi) - O(Pi+I ))

i-I
S(w)c(w) + S(pk) ('(Pk) - O(w))

k-I
+ E S(Pi) ((NPi) - 4(Pi+l ))

i-1
rA
- S(w)c(w) + J
S(u)dc(u)
Vi
- S(w)4)(w) + J S(u)dfi(u)
Z
- S(z)O(z) -
Ju
(D(u)dS(u)
logz
< (K - 1)V(z)c(z) - KV(z) J I z'(u)d (Iogu)
w
9.4
253
Theorem 9.5 Let z > 2, and let D be a real number such that D > z for n odd
and D > z2 for n even, that is,
s
log D
log z -
if n is odd
if n is even.
Let P be a set of primes, and let g(d) be a multiplicative function such that
0< g(p) < I
for all p E P
and
log z
fl (1 - g(p))- < K log u

pep
for all u such that I < u < z, where the constant K satisfies
1<K<1+-.
200
1
Then
T,, (D, z) < V z) 1 f, (s) + (K - 1)
99
eto-s
(9.30)
100)"
We define the number
r -a+5(K-1)+11e-8
and the functions
h,,(s) - (K - 1)raeloh(s)
(9.31)
for n > 1. Note that

a < r < 0.9607 + 0.0250 + 0.0037 - 0.9894 <
99
.
100
We shall prove that
T,,(D,z) < V(z)(,fn(s)+ha(s)).
(9.32)
This immediately implies (9.30) since h(s) < e for all s > 1.
The proof of (9.32) is by induction on n. Let n - 1. By Lemma 9.3 with $ - 2,
we have T, (D, z) - 0 for s > 3. Since the right side of inequality (9.32) is positive,
it follows that the inequality holds for s > 3. If I < s < 3, then f, (s) - (3/s) - 1
and
T,(D, z) = V(D113) - V(z)
254
9.
The linear sieve
by (9.13). It follows that
T, (D, z)
V(D113)
-1
V (z)
V (z)
(I - g(P)Y' - 1
1 1
!<P<-
< 3Klogz
log D
3K
-1
/ _l)+3
(K-1)
_ (s
< f,(s)+33(K - 1)
< f,(s)+h,(s)
since h(s) > e-3 and r > 11e-8, hence
h,(s) - (K - 1)re10h(s) > (K - 1)11e-' > 3(K - 1).

This proves the lemma for n = 1.
Let n > 2, and assume that the lemma holds for n - 1. For n even and s > 2,
or for n odd and s > 3, we define the function
(log D
(u) =
logu
-1
log D
(logu - 1
for 1 < u < w. The function 1(u) is continuous, positive, and increasing.
Moreover,
1(z) = fee _1(s - 1)+hi_1(s - 1).
It follows from the recursion formula (9.14), the induction hypothesis for n - 1,
and Lemma 9.8 that
(D , P J
T, (D, z)
pen
P.-:
(log D
< E g(P)V (P)
log D
log P
log P
EP
P<z
_ T g(P)V (P)c(P)
PEP
= (K - 1)V(z)c(z) - KV(z) J Ò(u)d (logzl
loguj
= (K -KS (z)
1))
(u)d(l
g gD/
9.4 The Jurkat-Richert theorem
_ (K -
255
1))
+KV(z) f Cc
1))dt,
where the last equation comes from substituting t = log D/ log u in the integral.
By (9.21), we have
x
-K J
I(t - 1)dt = Kf,,(s).
S
Similarly, from the definition of H(s) and (9.25), we have
x
h(t - 1)dt = H(s) < ash(s)
Js
and so
K
S
1)dt < aKh,,_I(s).
Since h(s - I) < 4h(s) for .s > 2, we have

(K - 1)h"_I(s - 1) < 4(K - 1)h"-I(s)
and
(K -
1)
1)
< 8e2(K - 1)ai-2h(s)

=
\al
8e_8
a-'(K -
< 9e-8h, _I(s)
since 0 < a < r and a-I < 9/8. Therefore,

T, (D, z)
<
V(z)
By Lemma 9.7 and definition (9.31), we have

(K - 1)f"(s) < (K - 1)2e2a"-Ih(s) <
and so
8
Since
aK=K-(1-a)K <K-(1-a)=(K-1)+a,
we have
T,, (D, z)
V(z)
<
f(s)+(a+5(K - 1)+Ile-8)h_I(s)
= f,,(s)+rh_I(s)
=
(s)+h(s)
256
9.
The linear sieve
Let n > 3 be odd, and let 1 < s < 3. If z - D", then log DI log z - 3. By the
recursion formula (9.15) and the same argument used above, we obtain
T,(D, z) -
g(P)T,,-1 1 P , P 1
PEP
P, 1,1/3
<
g(P)V(P)4(P)
pfP
< (.fn(3) + h.(3))V (z)

< (.fp(s) +
(z)
since the functions ,,(s) and h(s) are decreasing. This completes the proof.
Theorem 9.6 Let z, D, s, 7', g(d), and K - I + e satisfy the hypotheses of

Theorem 9.5. Let
G(z, A}) - E A (d)g(d).

dIP(z)
Then
G(z, X+) < V(z) (F(s) + se 14-s)

and
G(z, ),-) > V(z) (f (s) - Ee14-1 ,

where F(s) and f (s) are the continuous functions defined by (9.27) and (9.28).
Proof. We note that the sum of the following geometric series satisfies
99 ) ti
00
E
-0
n.0 ,moJ 2,
<51<
( 100
e4.
By (9.10) and Theorem 9.5,
G(z, A+) - V(z)+
< V(z)
f11(s)+ee10-S
1+
(99
00
..I
.., (mW 2)
mod 2)
< V(z) (F(s)+ee14-s)

Similarly, by (9.11) and Theorem 9.5,
00
G(z, A-) - V (z) - E

n.O (maf 2)
z)
9.4
257
ao
> V(z)
Eelo-s
1 -
f,(s) -
,-A :,
> V(z) (f (s) - Eel4-`

Theorem 9.7 (Jurkat-Richert) Let A = {a(n)},_1 be an arithmetic function such

that
a(n)>0 for all n
and
IAI - 00>a(n) < oo.

I,-1
Let P be a set of prime numbers and, for z > 2, let
P(z) - Jl p.
pfP
P':
Let
Oc
S(A, P, z) - E a(n).
,.;(-'10-1
For every n > 1, let g,, (d) be a multiplicative function such that
0<
for all p E P.
(9.33)
Define r(d) by
IAdi L,a(n)
d,w
Let Q be a finite subset of P, and let Q be the product of the primes in Q. Suppose
that, for some a satisfying 0 < e < 1/200, the inequality
fl ( 1 - gn(P))-1 < (1 +E) log
go
of r\Q
(9.34)
sr<.
holds for all n and I z there is the upper bound
S(A, P, z) < (F(s) + Ee14-s)X + R,
(9.35)
and for any D > z2 there is the lower bound
S(A, P, z) > (f (s) - ee14-s)X - R,

where
log D
s -
log z
(9.36)
258
9. The linear sieve
f (s) and F(s) are the continuous functions defined by (9.27) and (9.28),
00
X - Fa(n) fl (I - gn(P)),
M-1
(9.37)
P1 P(-,)
and the remainder term is
R - E Ir(d)I.
dIP(:)
d<DQ
If there is a multiplicative function g(d) such that g (d) = gg(d) for all n. then
X - V(z)IA1,
(9.38)
where
V(z) - H (1 - g(p)).
PIP(:)
Proof. Let P1 - P \ Q. By Theorem 9.3. there exist upper and lower bound
s= I for
alld> 1. We define
kj (d)gn(d)
Gn(z, kt)
sieves k*(d) with sieving range P, and support level D. and with !w` (d)
PI Pi (z)
and
VV(z) - fl (I - gn(P))PI PI(z)
Since P, and Q are disjoint sets of primes, we have
fl (1 - gn(P)) - V,,(z) [1 (I - gn(q))

PIP(:)
gIQ(z)
By Theorem 9.6,
Gn(z, k+) < VV(z)
(F(s)+eel4-S)
and
G,,(z, k) > VV(z) (F(s) -
eel4-s) .
It follows from Theorem 9.2 that

00
S(A, P, z) <- Ea(n)Gn(z, ki) fl (I - gn(q))+ R

n-1
gIQ(z)
00
a(n)VV(z) fl (I - g, (q)) + R
< (F(s) + ee14-s)

n-1
qIQ(:)
00
- (F(s)+se14-s)1: a(n) fl (I -gn(P))+R

n-1
PIP(:)
- (F(s) + ee14-s)X + R.
The lower bound is obtained similarly. This completes the proof.
Differential-difference equations
9.5
9.5
In this section, we shall compute initial values for the functions

00
F(s) - 1 + E f (s)
for s > I
and
00
f(s) - I - E f, (s)
for s > 2.
n.0 (mod 1)
We shall prove that
F(s) -
2e r
for l < s < 3
and
2e1' log(s - 1)
for 2 < s < 4,
s
where y is Euler's constant. We define f (s) - 0 for I < s < 2.
f (s) -
Lemma 9.9
sF(s) - 3F(3)
fort<s<3.
Proof. Let 1 < s < 3. By Lemma 9.5,
3f,(3)
for all odd n > 3.
Since
s+sfi(s)-3
by (9.20), it follows that
M
sF(s) - s +sfi (s) + F

,.i
0_7
tmW 11
OC
-3+
,mod 1,
- 3F(3),
which completes the proof.
Define the constants A and B by
A - sF(s)
for l < s < 3
and
B - 2f (2).
259
260
9.
The linear sieve
Lemma 9.10 The functions F(s) and f (s) are solutions of the system of differential-difference equations
(s F(s))' = f (s - 1)
for s > 3
(sf(s))' = F(s - 1)
fors > 2.
Proof. Let n > 2. By Lemma 9.5, for n odd and s > 3, or for n even and s > 2,
we have
1)dt
r 00
and so
1).
(sf,,(s))'
For s > 3, we have sf1(s) = 0 and so
00
(sF(s))' =
s+
00
fn-I(s - 1)
n_,
n., oo,.i L
00
= 1 -
fn(s - 1)
.a uu,,, 2)
= f(s -1).
Similarly, for s > 2 we have
00
(sf(s))' =
s21
= 1 +
00
1 +
00
=F(.s-1).
1)
1)
9.5
261
Lemma 9.11 For s > 2, let

P(s) - F(s) + f (s)
and
Q(s) - F(s) - f (s).

For s > 3, the functions P(s) and Q(s) are the unique solutions of the differentialdifference equations
sP'(s) = -P(s) + P(s - 1)
(9.39)
and
sQ'(s) = -Q(s) - Q(s - 1)
(9.40)
that satisfy the initial conditions
s P(s) - A + B + A logs - 1)
and
sQ(s)=A-B-Alog(s-1)
for 2 < s < 3. Moreover,
P(s) - 2 + O(e-S)
and
Q(s) - O(e-S).
Proof. Since
sF(s) - A
it follows that
F(s) - s
for l < s < 3,
for I < s < 3
or, equivalently, that
F(s-1)- s-1
A
for2<s<4.
Since (sf (s))' - F(s - 1) for s > 2, it follows that
Aldt
- B+Alog(s - 1)
sf(s)=2f(2)+f2S t
for2<s <4.Since
sF(s)=A
fort <s<3,
it follows that
sP(s)-A+B+Alog(s- 1)
9.41)
sQ(s)-A-B-AIog(s-1)
(9.42)
and
262
9.
The linear sieve
for 2 < s < 3. For s > 3, we have
(s P(s))' - (s F(s))' + (sf (s))' - f(s- 1) + F(s - 1) - P(s - 1),

and so
sP'(s) - -P(s)+ P(s - 1).

similarly,
(sQ(s))' - (sF(s))' - (sf(s))' = f(s - 1) - F(s - 1) - -Q(s - 1)

and so
sQ'(s) - -Q(s) - Q(s - 1).

By Theorem 9.4, we have F(s) - I + O(e-s) and f (s) - 1 + O(e-S), and so
P(s) - 2 + O(e-s) and Q(s) - O(e-S). This completes the proof.
The differential-difference equations (9.39) and (9.40) are of the form
sR'(s) - -aR(s) - bR(s - 1).
(9.43)
Associated with this equation is the adjoint equation

(sr(s))' - ar(s) + br(s + 1).
(9.44)
To every solution R(s) of equation (9.43) and every solution r(s) of equation (9.44),
we associate the function
(R(s), r(s)) = sR(s)r(s) - b f R(x)r(x + 1)dx

for s > 3. Differentiating with respect to s, we obtain
d
ds
(R(s), r(s))
- R(s)r(s) +sR'(s)r(s) +sR(s)r'(s) - bR(s)r(s + 1) +bR(s - 1)r(s)

- (sR'(s) + bR(s - 1))r(s) + (r(s) + sr'(s) - br(s + 1))R(s)
- -a R(s)r(s) +a R(s)r(s)
-0.
Therefore, (R(s), r(s)) is constant for s > 3.

The equation adjoint to (9.40) is
(sq(s))' = q(s) + q(s + 1)
or, equivalently,
sq'(s) - q(s + 1).

This has the solution
q(s)=s- 1.
9.5
263
Clearly,
q(s) ^- s
as s tends to infinity, and
q(1) - 0.
Since Q(s) - O(e-S), it follows that
sQ(s)q(s) - 0 (sze-S) - o(l)
and
Q(x)q(x + 1)dx - 0(1).

1S
Therefore,
lim (Q(s), q(s)) - 0.
5-00
Since (Q(s), q(s)) is constant for s > 3, it follows that
(Q(s), q(s)) - 0
for s > 3. This implies that B - 0, since (x Q(x))' - -(x - 1)-' by (9.42), and
0 - (Q(3), q(3))
3
- 3Q(3)q(3) - f Q(x)q(x + l )dx

i
- 3Q(3)q(3) - f xQ(x)q'(x)dx
i
3
- 3Q(3)q(3) - [xQ(x)q(x)]x;2 +J (xQ(x))'q(x)dx

z
- 2Q(2)q(2) - A
-(A-B)-A
fx3 q(x)- l dx
- B.
Similarly, the equation adjoint to (9.39) is
(sp(s))' - p(s) - p(s + 1)

or, equivalently,
sp'(s) - - p(s + 1).
(9.45)
For s > 0, we introduce the function
p(s) - fo 00 exp(-sx - I (x)) dx,
(9.46)
The linear sieve
9.
264
where
f(1 - e')tdt.
1(x) Since
0<
1 - e-'
t
fort > 0,
<1
we have
0<I(x)<x
forx>0,
and so
exp(-(s + 1)x) < exp (-sx - 1(x)) < exp (-sx).

Therefore, the integral converges for all s > 0, and
-1-
r 30
00
exp(-(s+ 1)x)dx < p(s) < 1
exp(-sx)dx = s
It follows that
sp(s)
as s tends to infinity. Using integration by parts and the observation that
x1'(x) - 1 - e-`
we obtain
.cp'(.c) - - f 00 sx exp (-sx - 1(x)) dx
=
( d exp(-sx)) x exp(-I (x))dx
_ [x exp(-sx - 1(x))]. -
xp(-s.V) I(`
00
-xexp(- I (r))
} (/.f
t!t
= - f exp(-sx)(I - xI'(x))exp(-I(x))dx
0
- f 00 exp(-sx) exp(-x) exp(-1(x))dx

0
_ - f 00 exp(-(s + 1)x - 1(x))dx

0
-p(s + 1).
This proves that p(s) is a solution to the adjoint equation (9.45) for all s > ft
We shall prove that
plI) - e".
We need the following integral representation for Euler's constant:
Y=
f(1 - e')tdt - fe_trIdt
(9.47
9.5
(see Exercise 16 and Gradshteyn and Ryzhik [42, page 956]). Then
f(1 - e`)tdt
I(x) =
f(1 - e')tdt +J (1 - e`)tdt
j(1 - e`)tdt -
+ logx
1
=J (1 -e-')t-Idt
-J
e-`t-Idt+f e-`t-Idt+logx
00
00
=y+ f 3C e-'t-Idt+logx.
It follows that
00
-sp'(s) =
sx exp(-sx - I (x))dx
J0
= e-Y
100
0
s exp (_sx
x
00
00
e-' 1o
- 1a-t_I dt) dx
-u - f e-'t-I dt
exp
du.
/s
For u > 0, we have

00
fe-'t-'dt = 0,
lim
s-.0'
/s
and so
(s+1)
P(l)=Slim
lim sp'(s)
S
0C
= e-Y lim
0'
10
exp C-u
0c
= e-Y 1
0
= e-Y
lim exp
-f
00
e-'t-I dt I du
/s
-u - f
s-+0'
JJJ
00
e-'t_Idt
du
/s
1 exp(-u)du
0
Since P(s) = 2 + O(e-`) and sp(s)
1, it follows that
lim (P(s), p(s)) = lim I sP(s)p(s)+
s-.0c
5-00 \\\
1
s
P(x)p(x + 1)dx
I
2.
265
266
9.
The linear sieve
Since (P(s), p(s)) is constant for s > 3, it follows that
(P(s), p(s)) - 2
for all s > 3. Letting B - 0 in (9.41), we have
sP(s) - A + A log(s - 1)
and
(sP(s))' -
s-1
for 2 < s < 3. Therefore, 2P(2) - A and

2 - (P(3). p(3))
- 3P(3)p(3) + J 3P(x)p(x + 1)dx

z
3
- 3P(3)p(3) - f xP(x)p'(x)dx
z
3
- 3P(3)p(3) - [xP(x)P(x)]s-i + 1 (xP(x))'P(x)dx

z
-2P(2)p(2)+AJ 3 p(x) dx
z x-1
- Ap(2)+A
f2
3 X (x) dx
3
- Ap(2) - A f2 p'(x - 1)dx

- Ap(2) - Ap(2) + Ap(1)
- Ae-Y.
This proves that
A - 2eY.
We can now determine the initial values of F(s) and f (s).
Theorem 9.8
F(s) -
2eY
for l < s <3
and
f(s) -
2eY log(s s
where y is Euler's constant.
for2<s<4,
Proof. Let 2 < s < 3, and let A - 2eY and B - 0 in (9.41) and (9.42). Then
sP(s) - 2eY +2eY log(s - 1)
9.7
Exercises
267
and
sQ(s) = 2e'' - 2e> log(s - 1).

Therefore,
sF(s) =
sP(s) +sQ(s) - 2eY

2
By Lemma 9.9, sF(s) is constant for I < s < 3 and so
for 1 < s < 3.
sF(s) = 2e>'
By Lemma 9.10, we have (sf(s))' = F(s - 1) for s > 2 and so

2e'
'
sf(s)=2f(2)+J F(t - I)dt =
=2e''log(s - 1)
t e
for 2 < s < 4. This completes the proof.
9.6
Notes
The material in this chapter is based on unpublished lecture notes of Henryk

Iwaniec[68]. See Jurkat and Richert [69] for the original proof of Theorem 9.7.
Standard references on sieve methods are the monographs of Halberstam and
Richert [44] and Motohashi [87].
9.7
Exercises
1. Let P be the product of the primes up to T. Prove Legendre's formula
rr(x) - 7r(/)+ I
[x]- F- I - +
Pix
Pi</r
(d)
[d]
Pz
n:<PI<_r[_L_]
P3<t,i<P <f
[P,2P3]
+...
2. Let P be the product of the primes up to fx. Prove Sylvester's formula
p+1=2F(d)[dJ\[dJ+1).
f<P<L
dlP
3. Let A, = {a, (n)} and A2 = {a2(n)} be arithmetic functions such that a, (n) <
a2(n) for all n > 1. Prove that
S(A1, P, z)
S(A2, P, z).
268
9.
The linear sieve
4. Let At - (at (n)) be a nonnegative arithmetic function for t - 1, ... , k, and

let A - {a(n)} be the arithmetic function defined by a(n) -a I (n)+. +ak(n)
for all n. Prove that
k
S(A, P, z) - > S(At, P, z).

f-1
5. Let 2 < w < z. Prove Buchstab's identity:
S(A,P,z)-S(A,P,w)- E S(A,,,P, p).

w<p<z
In particular,
S(A, P, z) - IAI - E S(Ap, P, P)

p<t
6. By iterating the Buchstab identity, prove that, for z, <- z,
S(A,P,z)<IAI- E IA,.I+ E IAp,p2I

Pi <zi
P2<Pi <Zi
1:
S(APiPP, P, P3)
P: <P2 <P[ <Z
7. Let P be a set of primes, and let lt(d) be upper and lower bound sieves with
sieving range P and support level D. Let P, be a subset of P. We define
functions A:': (d) by A (d) - lt(d) if d is divisible only by primes in P,, and
Ai (d) - 0 otherwise. Prove that k} (d) are upper and lower bound sieves
with sieving range P, and support level D.
8. Let h(s) be the function defined by 9.23. Prove that
h(s-1)<4h(s)
fors>2.
9. Use the recurrence relation
sf2(s) -
00
f,(t -1)dt
to prove that
sf2(s)-s-3log(s- 1)+3log3-4
fort<s<4.
10. Prove that
f(x) - xlog
9x
9x - 1
<log
for x > 1. Hint: Show that the function f (x) is decreasing for x > 1.
9.7
Exercises
269
11. Let Q(s) be a continuous function on the interval [ 1, 2]. Prove that there exists a unique continuous function Q(s) defined for all s > I that satisfies this
initial condition and that is a solution of the differential-difference equation
sQ'(s) = -Q(s) - Q(s - 1)

for all s > 2. Hint: For 2 < s <f3, we must have
sQ(s) = -
J2
'
Q(x - 1)dx +2Q(2).
Similarly, for 3 < s < 4, we must have
s Q(s) = - f3' Q(x - 1)dx + 3Q(3).

T he proof proceeds by induction.
12. Let Q(s) be the function defined in Lemma 9.11. Prove that
s(s - 1)Q(s) =
Js1
xQ(x)dx
for all s > 3. Prove that

0 < sQ(s) << s-S.
13. Let Pi and P2 be disjoint sets of prime numbers, and let ff and f2 be
arithmetic functions such that fi (d) 0 only if d is a product of primes
belonging to PI and f2(d) 0 only if d is a product of primes belonging to
P2. Let f = f1 * f2. Prove that
1*f=(1*fi)(1*f2)
14. Let A (d) and Az(d) be upper bound sieves with support levels DI and D2,
respectively, and with disjoint sieving ranges Pi and P2. Let A++(d) be the
convolution of l+ (d) and ?4(d), that is,
+(d) = h1 * X2+(d)
X (di)X2(d2)d-d1 d;
Prove that k+ is an upper bound sieve with support level D = DI D2 and

sieving range PI U P2.
15. Let ), (d) and 4(d) be upper bound sieves with support levels DI and D2,
respectively, and with disjoint sieving ranges Pi and P2, and let A, (d) and
A2 (d) be lower bound sieves with support levels DI and D2, respectively,
and with disjoint sieving ranges Pi and P2. Prove that
;C(d) = AI' *)4(d) - ,l+ *)4(d) + A * AZ (d)

Prove that A is a lower bound sieve with support level D = DI D2 and
sieving range Pi U P2.
270
9.
The linear sieve
16. In the theory of the Gamma function, it is proved that

00
-y - 17'0) -J
e-xlogxdx.
0
From this formula, use integration by parts to obtain (9.47):
j
o
- e`)tdt - je_tCIdt.
10
Chen's theorem
Is it even true that every even n is the sum of 2 primes? To show this
seems to transcend our present mathematical powers.... The prime
numbers remain very elusive fellows.
H. Weyl [142]
10.1
Primes and almost primes
In this chapter, we shall prove one of the most famous results in additive prime
number theory: Chen's theorem that every sufficiently large even integer can be
written as the sum of an odd prime and a number that is either prime or the product
of two primes. An integer that is the product of at most r not necessarily distinct
prime numbers is called an almost prime of order r, denoted P,, and so Chen's
theorem can be written in the form
N - p + P2
for every sufficiently large even integer N. We shall prove not only that every large
even integer N has at least one representation as the sum of a prime and an almost
prime of order two but that there are, in fact, many such representations.
Theorem 10.1(Chen) Let r(N) denote the number of representations of N in the

form
N - p+n,
272
10.
Chen's theorem
where p is an odd prime and n is the product of at most two primes. Then
2N
r(N) >> 6(N)(log
(10.1)
N)2,
where
6(N)n(I_
n>2
p-1
1) F1
(,,
p-2
(10.2)
P>2
The number 6(N) is called the singular series for the Goldbach conjecture.
The proof has two ingredients. The first is the Jurkat-Richert theorem (Theorem 9.7), which gives upper and lower bounds for the linear sieve. The second
is the Bombieri-Vinogradov theorem, which describes the average distribution of
prime numbers in arithmetic progressions. Throughout this chapter, p and q denote
prime numbers.
10.2
Weights
Let N be an even integer, N > 48. We begin by assigning a weight w(n) to every
positive integer n. Let
(10.3)
z - N'18
and
y = N 1/3 .
(10.4)
Then z > 4. We define
w(n)-1- Tk -E
2
2 :<q<,
(10.5)
nI p2p3-e
:API<.<P7eP3
p+Fw
Clearly,
w(n) < 1
for all n, and w(n) - 1 if and only if n is divisible by no prime in the interval [z, y).
Let P be the set of prime numbers that do not divide N. Then 2 P since N is
even. Let
P(z)-11p.
Peg
P-
Let n be a positive integer such that
n < N and (n, N) - (n, P(z)) - 1.

Then n is divisible only by primes p > z that do not divide N. If n - PI P2
Pr Pr+I ' ' ' Pr+s, where
Z<PI <...<Pr <y<Pr+1 <...<pr+s,
10.2
Weights
273
then
N,'3
<N
and so s = 0, 1, or 2. Suppose that w(n) > 0. Since
it follows that r = 0 or 1. If r = I and s = 2, then n = pI P2 p3, where z < pi <

P2 < P3, and so w(n) = 0. Therefore, if w(n) > 0, then either r = 0 and
Y
s = 0, 1, or 2, or r = I and s = 0 or 1. In all of these cases, r + s < 2. Therefore,
if (n, N) = (n, P(z)) = 1 and w(n) > 0, then either n = I or it is an integer of the
form pi or PI P2, where pi and P2 are primes > z that do not divide N.
Consider the set
A={N-p: p<N,pEP).
(10.6)
Then A is a finite set of positive integers, and Aj = 7r(N) - w(N), where w(N)
denotes the number of distinct prime divisors of N. If n = N - p E A and

if (n, N) > 1, then p divides N and so p V P, which is absurd. Therefore,
(n, N) = I for all n E A. We obtain a lower bound for r(N) as follows.
-11,P1
ll.rl PIP:
rl f':?:I
E
E
>
,.A
'EA
"Ell pl rlr2 pl.r.?:
w(n)
rEA
p. rl:lri
nEll.rl . rl rq rl.r:ta
.rr rr:u-l
-A
rr. Pglr-I
Ek
_
21
yl 1
cA
Irr.
rl r?r)' P15r'1
.`-r'I
,rEA
fn.rl;l1-1
y41rr
274
Chen's theorem
10.
We shall express these three sums as sieving functions. If we let A = {a(

be the characteristic function of the finite set A, then the first sum becomes simply
1 - E a(n) - S(A, P, z).

P(:)-I
.EA
I.. ru ))-l
We divide the second sum into two pieces:
Ek= F E 1) + E E(k-1)
.GA
In
.P(U)-1
9c)
/ F
:_9cY
nEA
:VV'
EA
(..P(;)).I
U,.P(U1- 1
qI
The first piece can be expressed as a sieving function as follows: For every prime
q, let A. = (aq(n))' 1 be the arithmetic function defined by
aq (n)
ifn E Aandgln
otherwise.
Since (n,N)-1forallnEA,wehave gePifaq(n)=1, and
nEA
P1:0.1
1-E
aq(n)
:<q<,
C.
E S(Aq, P, z).
z nq <y
k1
It is easy to estimate the second piece. Since z - N'1s
k-2-:
qk
4 and
= (q - 1)2
we have
(k - 1) _ E 00,
z<q<y k-2
REA
i..Pl:))-1
qk I,
(k - 1)
OA
In, PUD-1
q41
k 2
Ip
E EE(k-1)
,<q<y k-2 .<N
<N
qkk
<q < y k-2
N
<q <
z-2
2N
-2
z
N718.
(q - 1)2
Prolegomena to sieving
10.3
275
For the third sum, we let B be the set of all positive integers of the form
N-PIP2P3,
where the primes Pi, P2, p3 satisfy the conditions
z<PI <Y5P25P3
PIP2P3 < N
(P1P2P3, N) - 1.
Let B - (b(n))'i be the characteristic function of the finite set B. An element
of B is a prime p if and only if p < N and N - p - pi p2 p3 E A where

z <_ Pi < Y < P2 < P3 - Therefore,
EE
(..P(W-I
.A
1- E
PIP2P3t`-PI ") P2'P3
PI P2P3eA
-PI <r<P2 Pi
-E1-E1+E1
PEB
Pe8
Pe8
< y+E 1
Pe8
Ply
<Y+ E 1
la Pp)}I
-Y+
b(n)
(n. P(Y))-I
- N1/3 + S(B, P, y).

We now have a lower bound for r(N) in terms of sieving functions.
Theorem 10.2
r(N) > S(A, P, z) - 1

2
S(A,, P, z) - 1 S(B, P, y) - 2N718 - N"3.

z<q<y
We shall obtain a lower bound for S(A, P. z) and upper bounds for Eq S(Aq, P,
z) and S(B, P, y).
10.3
In applying the linear sieve to estimate the three sieving functions, we choose the
multiplicative function
g(d) -
Bp(d)
276
10.
Chen's theorem
for all n > 1. Since N is even, we have 2 ' P and
0<g(p)- p-1 <I

1
forallpEP,
so the functions g(d) satisfy (9.33). To establish inequality (9.34), we apply

Theorem 6.9, which says that there exists a number u I (e) such that
g
fl
(1
u< p<z
< (1 +e/3)
- P)
logu
for any ui(e) u2(f)
p(p - 2)
1+
11
p>u2()
p(p - 2)l)
<1+E
3
since the infinite product converges. Therefore, for
u > uo(e) - max(uI(e), u2(e))

we have
Hg(p))-p 1
<p<z
U5P<z
(p - 1)2
H
1
<p<z p(p - 2) u<p<z
< (I +e/3)2
< (1 +s)
logg u
logg u
Let Q(e) be the set of all primes p < uo(e), and let Q - Tn Q(e). This gives (9.34).
Let Q(e) be the product of the primes in Q(e), and let Q be the product of the
primes in Q. Then Q(e) depends only on e, not on N, and so
Q < Q(e) < log N
(10.7)
for all sufficiently large integers N.
Theorem 10.3 Let N be an even positive integer, and let
V(z) - F1 (1 -g(p))pIP(z)
,.-
in.")
(1
- p-1)
I
Then
e-Y
V(z) - 6(N)logz
(1+O( logN))'
(10.8)
10.3
where
6(N)a
I-
277
11p-1
(p-1)2'IN p-2'
p,2C
Proof. Let
W(z) _ n
1-
2<p<z
Then
V (Z)
1-
W(z)=2p<:
P-1
IY
P-1
PIN
01-
)-'Fl('-
PIN'
PIN
F1p
P>2 P -
Z(I -
PIN
P_.
11)
p-
PIN
Since I - x > e-Z` for 0 < x < (log 2)/2 and 1 - x < e-z for all x, we have
-7
> 1 1 exp
P>:
P-
PIN'
= exp -2 E
PZ: p PIN
> exp
> exp C
= exp
2cv(N)
z-1
-81og N
8 log N
N'!8
>1-8logN
N '!8
Thus,
W(z)
11
2 pP -
(1+0 ()).
i
PIN
To estimate W(z), we see that
pl
P<Z
2f1
p<
P<Z
278
10.
Chen's theorem
11 P(P-2)
-2
2<0<:
(P - 1)2
-2 F1 I-(P-1)2 )
2z
1
1+
I.
-2f 1- (P - 12)n(
p(p - 2)
)
1
P>2
ll)
Since 1 + x < e` z-
n(n - 2)
2(z1
< exp
- 2)
<exp\Z)
2
< I+-.
z
By Mertens's formula (Theorem 6.8), we obtain
W(z)-21I(I-(P
1)2)(1+O(
))n(1
Pz P<,
-2fl(1-(p 11)2) (1+O (z )1logz(1+O(logz)/
p>2
2p>
e-Y
(P- 1)2
logz
(1 +O logN))
Therefore,
V(z)- V(z)W(z)
W(z)
=HP-I
>2
PIN
1- (p-1)2
P-2p>2
(N)logz 1+0
(j-:))
e r
Iogz
1+0
A lower bound for S(A, P, z)
10.4
279
10.4 A lower bound for S(A, P, z)

Theorem 10.4
(e) log3
S(A, P. z) >
+ O(E)
logg N).
Proof. We shall apply the linear sieve and results about the distribution of prime
numbers in arithmetic progressions to obtain a lower bound for the sieving function
S(A, P, z). We use the prime number theorem in the form
n(N)-
logN
(1+0
logN
(---').
JAI - E 1
pt.Y
- 7r (N) - w(N)
- 7r (N) + O(log N)
-logN(1+O(log1
In the Jurkat-Richert theorem, the main term in the lower bound (9.36) is f (s)X,
where
X-V(z)IAI-V(z)logN
(1+0(
log N
and V(z) is defined by (10.8).

The remainder term in the Jurkat-Richert theorem is
R-EIr(d)I,
d. QD
d!PQ1
where
r(d) - IAdI - >a(n)g(d) - I AdI

n
We want to obtain
RK
I AI
(10.9)
N
(log N)3
with D - D(N) as large as possible. We want D large because the function f(s)
in the lower bound of the Jurkat-Richert theorem is an increasing function of
s - log DI log z for 2 < s < 4. We have
IAdI - I:a(n)
dl.
280
10.
Chen's theorem
00
N -PEA
N-p.0 (mod dl
1
PEP
p<N
p.N (,, ,i
1 + O(w(N))
p<,V
p.N (-d d)
_ ,r(N; d, N) + O(log N),

where the term w(N) appears when we include the primes that divide N. Therefore,
r(d) - I AdI -
JAI
cp(d)
- zr(N; d, N) -
7r (N)
cp(d )
+ O(log N)
- 3(N; d, N) + O(log N),

where
S(x;d, a) -n(x;d, a) - 7r
(x)
cp(d)
for x > 2, d > 1, and (d, a) - 1. There are two important results that provide
estimates for S(x; d, a). The Siegel-Walfisz theorem states that
S(x-- d a)
for any positive number A, where the implied constant depends only on A. This
result is useful if the modulus d is not too large, say, d << (log x)'. The BombieriVmogradov theorem tells us about the average distribution of primes in congruence
classes over a large set of moduli. It states that, for every A > 0, there exists a
positive number B(A) such that
1: max IS(x;d,a)I <<
d <D(A) (d.'
(logx)A
for
(logx)a(A)'
We shall apply the Bombieri-Vinogradov theorem with x - a - N and A = 3.
Let
D(3)
log N
N112
(log N)e(3)*I
10.5
An upper bound for S(Aq, P, z)
281
Then D > z2 - N't4. Since Q < Q(e) < log N for N > N(s), we have
N't2
QD < (log N)8(3) - D(3)

and
QD log N <
(logNN)3
N112
(log N)"(3)-'
for N sufficiently large. Therefore,
R - E I r(d)I
d<QD
dIP(z,
1; Ir(d)I
d<QD
(d.N) I
E IS(N; d, N)I + QD log N

d<QD
(d.N}1
<< E IS(N;d, N)I +

d<DO,
N
(log N)3
(d.N}I
N
<< (log N)3
Now we apply the Jurkat-Richert theorem (Theorem 9.7) with z - N'18 and N
sufficiently large. We have
log D
log z
and so
f(s) ` 2e' log(s - 1)

$
1)) log log N

- 4- 8(B(3) -log
N
i eY log 3 +0
log log N
logN
E [3,4]
)a
ey log 3
2
+0(8).
Therefore,
S(A, P, z) > (f (s) - se14-S)X - R
> (f(s)-seI I)V(z)logN (1 +O (1ogN))+O ((logN)3)

>
10.5
NV(z)
(eYIog3
+ O(s)) log N
2
An upper bound for S(Aq, 7', z)
Theorem 10.5
(e}lo6
z<q<y S(A,, P, z) <
O(e)
IoVN)
g
282
10.
Chen's theorem
Proof. We shall apply the Jurkat-Richert theorem again to get an upper bound
for S(Aq, P, z), where q is a prime number such that z < q < y. If n = N - p E A
and q divides both n and N, then q - p, which is impossible since the prime p
does not divide N. Therefore, IAgI - 0 if q divides N, so we can assume that
(q, N) s 1.
Again we choose g(d) - gn(d) - 11V(d) for all n, so inequalities (9.33)
and (9.34) are satisfied. The error term rq(d) is defined by
rq(d) I(Aq)dI -
d)
Let d divide P(z). Since d is a product of primes strictly less than z, it follows that
(q, d) - 1 for every prime number q > z, and so

a (n) -
I (Aq )d I
> a (n) =
I Aqd I
qdln
d.
Then
rq( d ) IAgdI - IAgI
W(d)
- IAgdI -
Al IAl
I
V(qd) + V(qd)
IA g
W(d)
r(qd) - r(q)
cp(d )
where r(qd) and r(q) are error terms of the form (10.9). Let
D
D(4) _
N1/2
log N (log N)8(4)+1
and
Dq - q.
Then Dq > D/z > z. The remainder term for S(Aq, P, z) is
1
Irq(d)I < E Ir(gd)I +r(q)
Rq
d<QDq
d<QDq
d1PQ)
dIPO
d<Qoq
dIPIA
V(d)
From Theorem 9.7, we have the upper bound

S(Aq, P, z) < (F(sq) + ee 14-5q) I Aq I V (z) + Rq,
where
log Dq
Sq
log z
10.5
283
We do not estimate the main term and the remainder term for individual primes q.
Instead, summing over z <- q < y, we obtain
E S(Aq,P,z) <E (F(sq)+se14)IAgIV(z)+R',

<y<
<y
14.E\'1-I
14..5'1-1
where
R'= E Rq
/-
<y<
l4. ,
<
<y<. J.QI)(y
ly..\'}I JI VI:I
Ir(gd)I + E r(q)
J<QI)/y
14..\'I-I
T(d)
OIVI:)
r(q)I ; -
Ir(d')I +
d<NI;:
J1 <Q1)
W(d)
and QD < D(4). Applying the Bombieri-Vinogradov theorem as in the previous

section, we obtain
E Ir(d')I <_ E IS(N;d', N)I + E O(logN)

d'<QIl
J'</IQ
J'<I)Q
10'.5)-I
IJ'.1'/.1
N
(log N)4
<<
Since Y = N 113 < D < Q D for sufficiently large N, we also have

N
Ir(q)I << (log N)4
<y
)q..%
.)
By Theorem A.17,
<< log N
d
G(d)
and so
N
R'
(log N)3
Next, we estimate the main term. We have
log D/q
log z
8log(N112/q)
log N
8(B(4) + 1)loglogN
log N
Since N1/8 = z -< q < y = N1/3, it follows that

4 <
8log(N 2/g) < 3

gN
(10.10)
284
Chen's theorem
10.
and so 1 < sq < 3. By Theorem 9.8, F(s) - 2eY/s for I < s < 3. Therefore,
2eY
F(sq) -
eY log N
4log(N1/2/q) +
s9
and so
4
F(sq)+Ee
log log N
log N
eY log N
(10.11)
4log(N1/2/q)+0 (e).
Also,
JAgI -.r(N; q, N) + O(log N)

- -r(N) + S(N; q, N) + O(log N)
1V(q)
(p(q)ingN
(1+0(logN)/ +3(N;q, N).
Therefore,
(F(sq)+ee14)IAgI
<q<
IV.N/-1
_ :E, ( eY log N
l N (1+0( log N
4log(N lie/q) + O (e)/ (D(q) log N
1
Iq.N,- I
(F(sq)+Ee14)S(N;q, N)
eYN
go(q)log(N1/2/q)
Iq.NFI
+0
log N
q<,
(?(q) log(N1 /2/q)
(q.N)-I
+0
(logN) :" p(q)
+O
S(N;q, N)
<q<
(q.N)-I
Iq.M1 HI
It is not difficult to evaluate these terms. By the Bombieri-Vinogradov theorem

again, we have
N
E
<q< S(N; q, N) By Theorem 6.7, we have
1
:<q <, V(q)

(q.N)-I
:<q <, q qEP
(log N)3
10.5
yE91
= log log y- log log z + 0
\\\
1
log z
83+0(_L)
= log(/)
logz-0(l).
Using this estimate and inequality (10.10), we have

N
log N
log N :" co(q) log(N1/2/q)
(log N)2 :`y' co(q) log(N'1Z/q)

(y. N1
(logNN)2
,<y <,
p(q)
N
<< (log N)2
Therefore,
(F(sq)+ee14)IAgI =
:<y., W(q) log(N1!2/q)
(y %,I
+O
(q.NhI
We note that
1
q-1
co(q)
+O
q2
and
1
q2 log(N112/q)
NE
: q<, g2 log N1 /2/y
6N
= log N
q2
z log N
N7/8
log N
Let
=loglogt+B+0(_!_)
S(t)
v
and
1
log(NI12/t)
eN
log N
285
286
10.
Chen's theorem
The functions S(t) and f (t) are increasing. We shall estimate the sum
1
<Y q log(N1J2/q)
by using integration by parts twice in Riemann-Stieltjes integrals. We have
'
f
q log(N112/q) - J:
I
dS(t)
log(N'/2/t)
f (t)ds(t)
- f (y)SW - f (z)SW - f
, S(t)df (1)
- f (y)(log logy + B) - f (z)(log log z + B)
- f (log logt+B)df(t)
Y
+0 (f (y)1 + 0 1 f Y df (t) )
\\ log z J
\\\
log t
- f ' f(i)dloglogt+O ((10N)2)

1
We compute the integral explicitly by making the change of variable t - N. Then
dt
tlogtlog(N1/2/t)
1/3
I
da
f(t)dloglogt -
logN
f/s a((1/2)-a)
2 log 6
- logN
Therefore,
(F(sq)+ee14)IAgI -
rer
2g6 + O(s)
IoN
and so
E S(Aq, P, z) <
(ey
2g6
+ O(E) 1 oN)
8
<q <y,
10.6
An upper bound for S(B, 1', y)
Theorem 10.6
NV( z)
S(B, P, y) < I c2 +O(S) I log N) + O l (l
g N)3
An upper bound for S(B, P, y)
10.6
287
Proof. Recall that

P3, P1 P2P3 < N, (P1 P2P3, N) - 1).
B- (N - Pi P2P3 : z <_ Pi < Y _< P2
Before estimating the sieving function S(B, P, y), we shall drop the requirement
that (ps, N) - 1 and relax the condition that PIP2P3 < N so that the numbers
pi and P2P3 range over intervals independent of each other. This will produce
a "bilinear form" in pi and p2p3. We shall let the prime pl vary over pairwise
disjoint intervals
e < Pt < (l +e)e,

where t is a number of the form
f-z(l+e)k
such that z < e < y. Then
0
<
k < log(y/z)
log N
1og(1 + e)
e
(10.12)
Let
t3-{N-pIp2P3:z-<Pi <Y<P2<P3,
e < P, < (1 +e)e, fP2P3 < N, (P2P3, N) - l}
(10.13)
and
a - U Be).
e
Then
(10.
Let b(n), bU)(n), and b(n) be the characteristic functions of the sets B, W), and
B, respectively. Since the sets B(t) are pairwise disjoint, we have
IBI -
IB('
and
S(B, P, y) < S(B, P, y) -
S(B(e', P, y).
We shall estimate the sieving function S(B(e), P, y) by using Theorem 9.7 with
the functions
g(d) - g,,(d) -
ep(d)
for all n > 1, and with support level

N 1/2
D -
(log N)A
288
Chen's theorem
10.
Then
IBd"1
1,
v)P2P3.N (m dd)
,5P) "f5P75P3.(5PI q)K
and the error term rd0 is defined by

IB(t1I
d
- I B(OI +r(ej.
Bp(d)
In the next section, we shall prove that

R(c)
N
- r` Irac> <
(log N)4
(10.15)
dl P4+)
With this estimate for the remainder, Theorem 9.7 gives the upper bound
S(B")
where
y) <(F(s)+Ee14)IB(t)1V(Y)+O (l'4)'
S- logD_32+O
loglogN
)E[1,3)
log N
log y
and so, by Theorem 9.8.
-'+
F(s) -
NN
NN
log
log
8

V (Y)
log z
V(z)
logy
8+0
(logN))_
(1+0
(logN)
This gives
S(B(`), P. y)
r
<(
+0(--))('
+0(lo1N))IB(I)IV(z)+
log
2r
<
+O(E)) IB("IV(z)+O
/
((logNN
N
(log N)4
)4).
Summing over the sets B(i), we obtain
S(B, P, y) : > S(B(e), P, y) <
(2
+O(e)) 191V(z)+o
since the number of sets B> is 0(e'' log N) by (10.12).
i l og N)'
An upper bound for S(B, P, y)
10.6
289
Next, we estimate 181. By the prime number theorem,
((I +e)N 1 <
(1 +2e)N
p,P2log(N/P1 P2)
P1P2
for N > N(e). If pI < P2 < p3, and PI P2 P3 < (1 +e)N, then pl p2 < (1 +e)N
and
(1 +e)N
P3 <
P1 P2
IBI < E
:5P1 `P25P]
PI P2Pl'(""N
<
((1+e)N)
P1 P2
;5P
P1
PIP2.(1y)V
<(1+2e)N
<PI <y P1 ,.<P2<((1+e)N/P1)1 "2 P2
log(NI PI P2)
To estimate the inner sum, we introduce the functions
h(t) -
log(N/p1 t)
and
H(u) a J
('V/u)1JS
log(N/ut)d
log log t.
The function h(t) is positive and increasing for 0 < t < N)/pl. Since y - NI/3,
we have (N/y)I/2 - y and so H(y) - 0. Since z - N1/8, we have, with the change
of variable t - N",
llog(H(z)
7/8/t)
= JN'
1
7/16
log N J I /3
=O
olot
da
(7/8)-a
log N
Recall that
S(t)- E 1P =loglogt+B+O().
P<r
Applying integration by parts to the inner sum, we obtain
'<P2 <((1+e)N/Pi )
P2 log(N/PI P2)
290
10.
Chen's theorem
h(P2)
P2
y5P2<((I+e)N1pi)'
-f
((1+c)N/P')"
h(t)dS(t)
((I+e)N/P0,12
-J
h(((1 + e)N/p1)1 /2)
h(t)d log log t + O
logy
C
I
-J
log(N/Pl t)
log(N/
f(NIPI)'/2
- H(p, )+O
d log log t
dloglogt+ O
2
((ION)2)
The error term is obtained as follows. First,

2
h(((1 +s)N/pl)1/2)
log y
log ((I F)P;) log y

2
log
0 +1Ve)T) logy
log
<<
((I+F)) log N1/3
(log N)2
Second, with the change of variable t - (N/p1)'/2s,

((I+e)N/Pi)''2
log(N/Plt)
JN/p,)l2
((1+e)N/p )'12
d log log t
1
t logt log(N/Plr)
JN/p,)I'
dt
ds
s
log ((N/PI )1 /2s) log ((N/PI )I /2s-'

ds
s(log ((N/ p,)'/2) + log s)(log ((N/ pI )I/2) - logs)

r(t+e>'-2
ds
s
((log(N/PI)tn)2 - (logs)2
An upper bound for S(B, P. y)
10.6
(i+f)1i2
ds
(log N)2
-0
291
S
1
(og N)2 )
It follows that the outer sum is
H(Pi)+O
Pi
P<..
Pi (log N)
:<P,
H(Pi)+O(
=:
(log IN)2
Pi
where the error term comes from the fact that
- log log y - log log z + O ((log z)- 1)
-:5PI<Y P1
log(8/3) + O((log N)-1 )

= O(1).
We calculate the main term, as usual, by integration by parts:
H(PI)
:<P,<Y
P1
f
J
H(u)dS(u)
H(u)d loglogu+ O (max(H(z), H(y))1

logy
J
H(u)dloglogu+O
(logIN)2
To evaluate the integral, we make the change of variables t - N and u - NO.

This gives
s
1V'
(N/u)"2
H(u)d log log u

wi
log(N/ut)
JNI-3
(I-)/2
1 /3
fI 1
dad9
afi(l - a -
/3
log(2 - 3,0)
1ogN
/H
d log log td log log u
140 -.8)
dO
log N'
where
r1/3 log(2
c
1/8
- 3)dfl
10(I - )
Therefore,
(1 + O(e))cN
Jill
log N
- 0.363 ....
N
(log N)2
292
10.
Chen's theorem
and
S(B, P, y) <
(2 + 0(e)) I B I V (z) + 0 C (log N)3 )
< (ceY
+ O(s)
10.7
NV(z)
log N
+O
E-1 N
((log N)3 )
A bilinear form inequality
We must still prove inequality (10.15) for the remainder R(t). This will be a
consequence of the following theorem.
Theorem 10.7 Let a(n) be an arithmetic function such that ja(n)j < 1 for all n.
Let A be a positive number, let X > (log y)2A, and let
D*
(XY))12
(log Y)A
Then
max
d<D (a d)-) n<X
a(n)
- "1: 1:
av.dri
a(n)
..om
d.
XY(log XY)2
(log y)A '
(10.16)
Proof. Let (a, d) - 1. By the orthogonality property of Dirichlet characters X

(mod d), we have
rp(d)
x
X(a)X(np) - { 0
(mod d)
if np - a (mod d)
otherwise.
This gives
a(d)
n
a(n)n<X
Z<r.Y
.pw noddi
n<X Z<p<Y ( ) X
1
p(d)
X(a)j:a(n)X(n)
x
(mod d)
The contribution of the principal character Xo

1
X(a)X(nP)
(modd)
n<X
(mod d) to this sum is
L L a(n).
W(d) n<X /<P,)
np.d Nl
X(p)
Z<p<Y
10.7
293
It follows that the left side of (10.16) is bounded above by
E x(P)
E a(n)X(n)
d.V(d),ma)
Z<p<Y
n<X
z'1zp
Every character x (mod d) factors uniquely into the product of a primitive character (mod r) and the principal character (mod s), where rs - d. Therefore,
the sum can be written in the form
1
rs<D' (p(rs)
E x(P)
a(n)X(n)
zlro
.x
(10.17)
z<p<Y
(p.,)-(
:5 E
E a(n)X(n)
<x
E x (P)
Z<p<Y
(p.,)-)
zlzu
where E * denotes the sum over primitive characters (mod r). To obtain the last
inequality, we used the fact that the Euler cp-function satisfies cp(rs) > (p(r)rp(s).
We can estimate the character sum Ep<Y x(p) by means of the Siegel-Walfisz
theorem. We have
E 1
E X(P) < E X(a)

p<Y
P.),
(mod r)
P- (.,d , )
- E X(a)n(Y;r,a)
(mod r)
n(Y)
E
(mod r) x
(a)
co(r) +
(log Y)B
rY
<<
(logY)B
since
a
E X(a)-0
(mod r)
for every nonprincipal character X. Since also
rZ
X (P) <<
p
rY
(log Z) B <<
(log Y)8
it follows by subtraction that
rY
Z<p<Y
X(P) << (logY)B
294
10.
Chen's theorem
If we add the condition (p, s) - 1, we remove at most w(s) << logs << log D'
terms from the character sum and so
E X (p) << (log Y)8 + log D'.
Z<P<Y
Since la(n)J < 1, we also have
E a(n)X(n)
< X.
.<x
Let Do be "small." The inner sum in (10.17), restricted to r < Do, is
L,
a(n)X(n)
.<x
rlr p
(r)
(P.01-I
I..f Y-I
1:
r < D0
rX
V(r)
rY
(log Y)B
1: x (p)
Z<p<Y
+ log D
///
log
(10.18)
<< Do (Iog
The rest of the inner sum in (10.17) ranges over Do < r < D*. We partition this
interval into pairwise disjoint subintervals of the form D*, < r < 2D,*, where
D1 - 2t` Do and 0 < k << log D'. This produces partial sums of the form
1
zlxn
T (r)
E a(n)X(n)
<.r
E x(p)
ZSP<Y
DO -D
<
1* E
1: 'r xlro
(r
}1/2
1: a(n)X(n)
V(r) /
.<x
p0<,<D
(r
1/z
\ co(r)/
1:
Z<P.:r
(Pa)-I
x(p)
By Cauchy's inequality, this sum is bounded above by
DI
D; <r<2D; " /
rrzo
r
(D;r<2D;
W(r)
,iro
E.<xa(n)X(n)
E x(p)
Z<p<)
(P=rI
10.7
295
The large sieve inequality [ 19, page 160] states that
E T r 'x
r<R w(r)
(moe,)
x'x
L+M
<< (R2 + M)
a(n)X(n)
n-L+1
la(n)12
n-L+1
for every arithmetic function a(n). Applying this inequality to each of the factors
in the product, and using the condition that Ia(n)I < 1, we obtain
1
<<
D-
,)
1: a(n)X(n)
R<x
x (n)
z<p<P
(p.,rI
((D'2 + X )X)1 /2 ((D-2 + Y)Y)112
XY\\ 1/2
C\D*2+X+Y+ D 2)
(XD y)-1/2
<< I Dj +X1/2+Y1/2+
(XY)1/2
(D*
<<
+ X112 + Y1/2 +
(XY)1/2)(XY)1/2.
D0
\\
Multiplying this by the number of partial sums, which is O(log D'), and adding
(10.18), we obtain the following upper bound for the left side of (10.16):
L:
d<D-
<
v(d)
1: a(n)X(n) E x(n)
Z:s p<Y
n<X
x "xp
s<D- (P S) r<D-
1: *x
a(n)X(n)
o,be
"xo
X(P)
Z<p<r
log D'
.'E<n (P (S) D*3XY
(log Y)B
I
<
1:
s
-
(i* +
(XD)1/21
112
+ Y1/2 +
(PO
(XY)1/2Iog D*
Da3XY(log D*)2
(log Y)B
+ (D*
(D* + X112 + y112 + (X
Y)1
D*0
/2 I (X Y)1/2(log D*)2.
f
Note that we picked up a factor log D* from the estimate (Theorem A. 17)
<< log D*.
:<n cp(s)
296
10.
Chen's theorem
Choose B - 4A and Do - (log Y)A. Since X > (log Y)2A and Y >> (log Y)'-', it
follows that the left side of (10.16) is
( D'
X Y(log D`)2
(log
<<
XY(log D')2
D'
+ (XY))/2 + X1/2 + Yt/2 + Do*
Y)A
XY)2
<< ( X'/2/2 + Y'/2/2 + (log1Y)A) XY(log

X Y(log XY)2
(log Y)A

We can now derive the upper bound (10.15) for the remainder term
R(t) - E I rd )I,
d< D
dIPl,1
where z < Z < y. From the definition (10.13) of the sets tar, we obtain the
individual error terms
rde)-IB()I- Bp(d) IB(r)I

'p(d)
aP1<,<P2tD)
fPZP7cN.'P2P7.N)-1
(P2P7<N.IPZP7.N}I
ImNd(
We delete some numbers from the second sum by adding the condition that
(pt P2P3, d) - 1. This is equivalent to (pt, d) - 1, since the condition (p_ p:, d) =
I already follows from the fact that d divides P(y). This additional condition
decreases the second term by at most
1
T(d)
< (I + e)N
(p(d)
P,P2P7<IININ
p d.p1 :
< (I + e)Nw(d)
i,V log d
z(p(d)
ntId.Pj . Pt
Let a(n) be the characteristic function of the set of numbers of the Form n = p2 p
where y < p2 < p3 and (p2p3, N) - 1. Then we can write the error term in the
form
r(t)
a(n) -
-1
n<X
z<P<r
a(n) +
P(d) E
1
n<X
op .o (mod dl
(oP.d)-1
where
X - N/t?
Y - min(y, (1 + e)t)
Z - max(z, e)
a-N.
/(Nlogdl
z'p(d) J
10.8
Conclusion
Since f < y, we have
D' =
(XY)1/2
(log Y)A
N'12 min(y/E, (1 + x)112
(log y)A
N't 2
(log N)A
= D.
Similarly,
D` < (XY)1"2 < (Ny)' /2 <
N.
By Theorem 10.7,
Rct =
) a(n) -
d<n n
.,pa
diP'))
O
d
<n
nE
Inod d)
E a(n)
(p.d)-1
N log d
zcp(d )
XY(log XY)2+NlogD'
(log Y)A
d<D (P(d)
N
D')2
<< (log N)A-2 + N7/8(log
N
+ N7"8(log)2
<<
(log N)4
<<
N
(log N)4
if we choose A = 6. This completes the proof.
10.8
Conclusion
We can now prove Theorem 10.1.

Proof. It follows from the formula for V (z) in Theorem 10.3 that
NV(z) _ C7(N) 8e-''N

log N
(log N)2
(i+o(J_)).
log N
297
298
10.
Chen's theorem
Theorem 10.2 gives a lower bound for r(N) in terms of three sieving functions.
Using the estimates for these sieving functions in Theorems 10.4, 10.5, and 10.6,
we obtain
r(N) > S(A, P, z) - 2 E S(Aq, P, z) - 2S(B, P, y) - 2N718 - N'13

-<q <.v
> (2 log 3 - log 6 - c - 0(e))

+O
'N
( (log N)3)
) -
2N718
eYNV(z)
4logN
- N'13
> (2 log 3 - log 6 - c - O(e)) C7(N)
+0
E_1 N
( (log N)3)
) -
2N718
(lo
N)2
(1+0(
log1 N
- N1/3.
Since
2 log 3 - log 6 - c = 0.042... > 0,

we can choose a such that 0 < e < 1/200 and
2 log 3 - log 6 - c - O(e) > 0.

For this fixed value of e, we have
E_1 N
(logN)3
((loN)3)
Then
r(N) ((N)
2N
(log N)2
This completes the proof of Chen's theorem.
10.9
Notes
Chen [10, 1 1 ] announced his theorem in 1966 but did not publish the proof until
1973, apparently because of difficulties arising from the Cultural Revolution in
China. An account of Chen's original proof appears in Halberstam and Richert's
Sieve Methods [44). The proof in this chapter is based on unpublished notes and lectures of Henryk Iwaniec [67]. The argument uses standard results from multiplicative number theory (Dirichlet characters, the large sieve, and the Siegel-Walfisz
and Bombieri-Vinogradov theorems), all of which can be found in Davenport [ 19].
Other good references for these results are the monographs of Montgomery [83]
and Bombieri [3]. For bilinear form inequalities, see Bombieri, Friedlander, and
Iwaniec [4].
Part III
Appendix
A.1
The ring of arithmetic functions
An arithmetic function is a complex-valued function whose domain is the set of

all positive integers. Let f and g be arithmetic functions. The sum f + g is the
arithmetic function defined by
(f + g)(n) - f (n) + g(n).

Addition of arithmetic functions is clearly associative and commutative, and every
arithmetic function f has an inverse - f defined by (- f)(n) - - f (n).

The Dirichlet convolution of the arithmetic functions f and g is defined by
(f * g)(n) - E f(d)g(n/d)
dIn
It is easy to see that Dirichlet convolution is commutative, that is, f * g - g * f ,

and distributes over addition in the following way:
f * (g + h) - f * g + f * h
The following theorem shows that Dirichlet convolution is also associative.
Theorem A.1 If f, g, and h are arithmetic functions, then
f *(g*h)-(f *g)*h.
302
Proof. For any n > 1,
((f * g) * h)(n) = 1:(f * g)(d)h
(n)
dIn
(f * g)(d)h(m)
_
d ni-n
f (k)g(e)h(m)
din-n kt-d
f(k)g(e)h(m)
_
kfn,-n
=>.f(k) E g(e)h(m)
fm-n/k
kIn
>2 f (k) E g(e)h

ti(n/k)
kin
nH
f(k)(g*h)k
k1n
(f * (g * h))(n).
We define the arithmetic function S(n) by
S(n) =
ifn=1,
if n > 2.
Then for any arithmetic function f we have
(f *S)(n)=L,f(d)S(n) -f(n),
dIn
and so the set of complex-valued arithmetic functions forms a commutative ring

with identity element S(n). This ring is an integral domain (Exercise 3).
The product f g of the arithmetic functions f and g is defined by
(f - g)(n) = f(n)g(n)
Let L be the arithmetic function L(n) = logn. Multiplication by L is a derivation
on the ring of arithmetic functions, that is,
(Exercise 11).
A.2
A.2
Sums and integrals
303
Sums and integrals
In number theory, we often need to establish asymptotic formulas or at least good

estimates for sums of the form
Mf(x) - > f(n),

/l <T
where f (n) is an arithmetic function. It is sometimes possible to estimate these

"mean values" by integrals.
Theorem A.2 Let a and b be integers with a < b, and let f (t) be a monotonic
function on the interval [a, b]. Then
b
f (k) -
m in(f (a ), f (b)) <
f (t )dt
max(f (a ), .f (b))
Proof. If f (1) is increasing on [a, b], then

k+I
f(k)
f(t)dt
f
k
fork -a,a+ 1,...,b- land

k
f(k) ? I f(t)dt
-I
fork - a + 1, ... , b. It follows that

b
b-1
k-a
k-a
f(k)f(k)+f(b)
fb
f(t)dt+ f(b)
and
b
b-I
1: f(k)- E f(k)+f (a) ? f f(t)dt+ f(a).

k-a
kro+l
Thus,
b
f(a) <
E f(k) -
fh
f( t)dt < f(b)
Similarly, if f (t) is decreasing, then

b
f(b) <- Ef(k)-
fb
f(t)dt <.f(a).
k-a

From this result, we get a useful estimate for W.
304
Theorem A.3 For any positive integer n, we have

e
(n
< n! < en (e )n
Proof. Since the function f(t) - logs is increasing on the interval [1,n], it
follows from Theorem A.2 that
logn! _Tlogn f logtdt+logn-nlogn-n+I+logn

n
and
log n! >
71
log tdt - n log n - n + 1.
The result follows from exponentiating these two inequalities.

Partial summation is another simple and powerful tool for computing sums in
analysis and number theory.
Theorem A.4 (Partial summation) Let u(n) and f (n) be arithmetic functions.
Define the sum function
U(t) - E u(n).
nv
Let a and b be nonnegative integers with a < b. Then
b-1
E u(n) f (n) - U(b) f (b) - U(a) f (a + 1) -
U(n)(f (n + 1) - f (n)).
n-a+l
n-a+1
Let x and y be real numbers such that 0 < y < x. If f (t) is a function with a
continuous derivative on the interval [y, x], then
1: u(n)f(n) - U(x)f(x) - U(y)f(y) y<n<x
U(t)f'(t)dt.
In particular, if f (t) has a continuous derivative on [I , x], then
u(n)f(n)-U(x)f(x)-f U(t)f'(t)dt.
n <x
Proof. This is a straightforward calculation.

b
1: u(n)f(n)
n-a+I
b
- E (U(n) - U(n - 1)) f (n)

n-a+l
b-I
U(n) f (n) n-a+I
U(n) f (n + l)
n-a
b-I
-U(b)f(b)-U(a)f(a+1)- E U(n)(f(n+1)- f(n)).

n-a+l
A.2
Sums and integrals
305
If the function f (t) is continuously differentiable on [y, x], then

r,I+I
f (n + 1) - f (n) -
f'(t )d t
Jn
and
n+I
U(n) (f (n + 1) - f (n)) - J
U(t) f'(t)dt.
Leta-[y]andb-[x].Then
E u(n)f(n)
y<n<s
b
E u(n)f(n)
n-a+1
b-1
U(n)(f(n+I)- f(n))
=U(b)f(b)-U(a)f(a+1)n-a+1
- U(x)f(b) - U(y)f(a + 1)
-E
n-a+l
n+ l
U(t)f'(t)dt
- U(x)f(x) - U(y)f (y) - U(x)(f(x) - f(b)) - U(y)(f (a + 1) - f(Y))

rb
- Ja+I U(t) f'(t)dt

U(x)f(x) - U(Y)f(Y) - f f U(t)f'(t)dtIf f (t) is continuously differentiable on [ 1, x ], then
E u(n)f(n) - u(1)f(1)+ E u(n)f(n)

n
I<n<a
-u(1)f(1)+U(x)f(x)-U(1)f(1)-J U(t)f'(t)dt
c
U(x) f (x) -
U(t) f'(t)dt.

Here is an application of partial summation. Recall that every real number x can
be written in the form
x - [x] + {x},
where [x] is the integer part of x and {x} is the fractional part of x.
Theorem A.5 Let
y- 1-
f
1
00
it 2 dt.
306
Then0<y<land
n<x
n -logx+y+O(X).
The real number y is called Euler's constant.
Proof. Since 0 < {t) < I for all t, we have
!dt=1,
0< f 00 12)dt<f
00
Ti
and so Euler's constant y is a well-defined real number in the interval (0, 1).
We apply partial summation with u(n) = 1 for all n and f (t) = 11t. Then
U(t) = [t] = t - {t}

and
n<.rn =n<.r
u(n)f(n)
=[X)JX[t)
+
{
t2
}
dt
p.r
dt-
f .r
{ )
t2
- log x + l - J0 {t) dt +
t2
dt
{t) dt
t2
- IX)
x
=logx+y+OC1/.
x

As another application of partial summation, we obtain the Euler sum formula.
Theorem A.6 (Euler sum formula) Let f (t) be a function with a continuous
derivative on [y, x]. Then
v<nr
f (n)f(t)dt+R,
f
R=J r{t}f'(t)dt*(y)f(y)-{x)f(x)= frO(t)f'(t)dt+O(y)f(y)-O(x)f(x),

where
A.2
Sums and integrals
307
Proof. We apply partial summation with a(n) = I for all n. Then A(t) _ [t] _
t - {t}and
E f('t)
_ [x]f(x) - [Y]f(Y) - J [t]f'(t)dt
[x]f(x) - [Ylf(Y) - f
tf'(t)dt +
=[tlf(x)-[Ylf(Y)- (xf(x) - yf(y)

=
{t} f'(t)dt
fl
f(t)dt l +
{t}f'(t)dt
f(t)dt + f r{t} f'(t)dt + {y} f(y) - {x) f(x).

There is a simple expression for partial summation in terms of Riemann-Stieltjes
integrals. If f and g are bounded functions on [y, x] and if

f, gdf also exists and
f fdg +
fdg exists, then
fgd f = f(x)g(x)
- f(Y)g(Y)
This lovely reciprocity law is called integration by parts. (See Apostol 11, chapter
9].) Let u(n) be a nonnegative arithmetic function, and let
U(t) _
u(n).
<1
If f is continuous on [y, x], then
u(n)f(n) = f c f(t)dU(t) = U(x)f(x) - U(y)f (y) - f C U(t)df(t).

If f has a continuous derivative on [y, x], then
U(t)df(t) = f ` U(t)f'(t)dt,
and we recover the formula for partial summation. Similarly, if we let
U(t)=1: u(n)
It <(
and if f is continuous on [y, x], then
u(n)f(n) = U(x)f(x) - U(y)f (y) - f T U(t)df(t).
(A.1)
308
A.3
An arithmetic function f (n) is multiplicative if
f(mn) = f(m)f(n)
whenever m and n are relatively prime positive integers. Since f (1) = f (I
I) =
f(1)2, we have f(1)= I or O. If f(1)=0, then f(n)= f(n 1)= f(n)f(1)=0

for all n > 1. Therefore, if the multiplicative function f is not identically zero,
then f (1) = 1.
If f and g are multiplicative functions, then the Dirichlet convolution f * g is
multiplicative (Exercise 2). An arithmetic function f (n) is completely multiplicative if f (mn) = f (m) f (n) for all positive integers m and n.
Theorem A.7 Let f be a multiplicative function. Then
.f ([m, n]).f ((m, n)) _ f (m)f (n)
Proof. Let pl, ... , pr be the prime numbers that divide m or n. Then
r
M -
pr
and
r
n =
pn,
i-I
where r1, ... , rr, s1, ... , Sr are nonnegative integers. Moreover,
n] _
[m
r
i-1
and
(m, n) =
f1
p min(r,,s, )
i-1
Since
{max(ri, si), min(ri, si)} _ {ri, si }
and since f is multiplicative, it follows that

r
r
.f(pmin(r,.s,))
fl.f(pmax(ri.s,))
f([m, nl)f((m, n)) _

i-1
i-1
=IIf(p;' I-If(p')
i-1
i-1
=f(m)f(n)
A.3
309
The Mdbius function (n) is defined by
ifn=l,
p(n) -
(-I)'
if n is divisible by the square of a prime,

if n is the product of r distinct primes.
Thus, (n) f 0 if and only if n is square-free, and
(n) - (-I)-(n)
for square-free integers n, where w(n) is the number of distinct prime divisors of
n. It is easy to check that the arithmetic function p(n) is multiplicative.
Theorem A.8 Let f be a multiplicative function with f (1) - 1. Then
1: (d)f(d) - fl(1 - f(P))

pin
din
Proof. This is certainly true for n - 1. For n > 1, let n' be the product of the
distinct primes dividing n. Since (d) - 0 if d is not square-free, it follows that
E A(d)f(d) - 1: t(d)f(d) - fl(l - f(P))

din
pin
din*
Theorem A.9 Let f (n) be a multiplicative function. If
lira f(pk)-0
p4-oo
as pk runs through the sequence of all prime powers, then
lim f (n) - 0.
n-+oo
Proof. There exist only finitely many prime powers pk such that If (pk) I
Let
A-
fl
> I.
If (p') 1.
If (PI)I> I
Then A > 1. Let 0 < e < A. There exist only finitely many prime powers pk such
that if (p*)l > c/A. It follows that there are only finitely many integers n such
that
If(p")I >- e/A

for every prime power pk that exactly divides n. Therefore, if n is sufficiently
large, then n is divisible by at least one prime power pk such that If( pk)I < e/A,
and so n can be written in the form
n=
r+s
r+s+!
i-I
i-r+1
i-r+3+1
fl pk, fj pk, fj Pk, ,
310
where P1, ... , pr+.s+t are pairwise distinct prime numbers such that
1_ If(p'I
E/A < If (pi I < I
If(pR'I <e/A
fori = 1,...,r,
fori=r+1,...,r+.s,
fori =r+.s+1,...,r+s+t,
and
Therefore,
r+s
r+.c+t
If(n)I = fl If(p`')I fl If(pA')I fl If(p;')I < A(E/A)' < E.

i-I
i-r+s+l
i-r+I
A.4
The divisor function
The divisor function d(n) counts the number of positive divisors of n. For example,
d(n) = I if and only if n = 1, and d(n) - 2 if and only if n is prime.
Theorem A.10 Let
m=pi'...pr'
be a positive integer, where pI, ... , Pr are distinct primes and kl, ..., kr are
nonnegative integers. Then
If m and n are any positive integers, then
d(mn) < d(m)d(n).

If (m, n) = 1, then
d(mn) = d(m)d(n),
that is, the divisor function is multiplicative.
Proof. Every divisor d of m can be written uniquely in the form

d=pi'...p,
where
0SJ'i <ki
A.4 The divisor function
311
for i = 1, ... , r. Since there are ki + 1 choices of ji for each i - 1, ... , r, it follows
that
d(m)=fl(ki+1).
i-i
Let n be a positive integer, and let
n-pl'...Pr
where e1, ... , er are nonnegative integers. Then
d(n) - fi(ei + 1).

i-i
Since
Pk'+['
mn i-1
and since
ki+t,+1 <(ki+1)(ei+l)
for all nonnegative numbers ki and ei , it follows that
r
d(mn) - fl(ki + Ci + 1) < fl(ki + 1)(ei + 1) - d(m)d(n).

i-1
i-1
If(m,n)- 1,then ki -OorL, =0 foreachi - 1,
r. In this case,
ki+ei+l -(ki+1)(ei+1)
and
r
d(mn) - fl(ki +fi + 1) - fl(ki + 1) fl(ei + 1) - d(m)d(n).

i-1
1-I
filo
1-1
1"'D
Theorem A.11
d(n) << nE
for every e > 0.
Proof. Let f (n) - d(n)/n'. We shall prove that f (n) - o(1). Since the arithmetic
functions d(n) and nE are multiplicative, it follows that f (n) is multiplicative, and
so, by Theorem A.9, it suffices to prove that
lira f (pk) = 0.
312
Since (k + 1)/2 412 is bounded for k > 1, we have
f(Pk) =
d(Pk)
Pke
k+1
PkF
G+1)
PkF1z
<
(k+11
2ke/z J
PkF1z
y/2
< I
Theorem A.12
D(x) = 7d(n) =xlogx+(2y - 1)x+ 0(,fx-).

n<x
Proof. We can interpret the divisor function d(n) and the sum function D(x)
geometrically. In the uv-plane,
d(n)=E1=EI
din
n-uv
counts the number of lattice points (u, v) on the rectangular hyperbola u v - n that
lie in the quadrant u > 0, v > 0. Then D(x) counts the number of lattice points in
this quadrant that lie on or under the hyperbola u v = x, that is, the number of points
(u, v) with positive integral coordinates such that I < u < x and I < v < x/u.
These lattice points can be divided into three pairwise disjoint classes:
1<u<f and 1<v<,,

or
1<u</ and
or
/< u < x
and
%fx- <v<x/u,
1< v< x/u.
The last class consists of the lattice points (u, v) such that
I<v<,fx-
and
lx- <u<x/v.
It follows from Theorem A.5 that
D(x) =
\Lu]
([f] - /1)
A.4 The divisor function
- [`r]2+2 >2 (Lu] I<u<.T
[u]-[.]2
=2I<u<fr
>2
(u
-2
-{
}) -
(,r -
{r})2
-2xI<u<f> -2 >
I:5 u<_f
-2x(logf+y+0(
I1 -x+0(,fx-)
)) -x+0(.)
x
-xlogx+(2y - 1)x+0(f).
Theorem A.13
dnn) = I (logx)2 + 0(log x).
n <x
Proof. It follows from Theorem A.12 that
D(x) - >2d(n)-xlogx+O(x).
n<x
By partial summation, we obtain

d(n)
n
D(x)
x
+J x D(t)dt
t2
I
- x logx + O(x)
+ /' ` t logt + O(t)dt
-logx+0(l)+
fx
I
t2
logtdt+0
I
(r x 1dt'
I
- 2(logx)2+0(log x).
Theorem A.14
>2d(n)2 << x(log x)3.

n<x
Proof. Since d(ab) < d(a)d(b) for all positive integers a and b, we have
d(n)2 - >2d(n) 1: 1
n<x
n<c
n-ab
313
314
E d(ab)
ab<.c
< E d(a)d(b)
ab <.%
_ Y'd(a) Y d(b)
a<.x
b<.i/al
d(a)+O
< x lo gx a<.`
( 11
\a
a<`
<< x(logx)3.
A.5
The Euler gyp-function
Let n > 1. We denote by rp(n) the number of positive integers a < n such that
(a, n) = 1. If a - b (mod n), then (a, n) _ (b, n), and so rp(n) also counts the
number of congruence classes modulo n that are relatively prime to n. This is
exactly the order of the multiplicative group of units in the ring Z/nZ.
Theorem A.15 The arithmetic function V(n) is multiplicative, and
cp(n) - nf I-1
pb+
Proof. Let (m, n) = 1, and let rp(rn) = r and V(n) = s. Let a1, ... , ar and
b, , ... , b, be complete sets of representatives of the congruence classes relatively
prime to m and n, respectively. We shall prove that the rs numbers an + b/m

for i = 1, ... , r and j = 1, ... , s form a complete set of representatives of the
congruence classes relatively prime to mn. If
a;n +b/m - akn +bfn (mod mn),

then
a;n+bfm =atn+bfn (mod n)

and so
b/m
bfm
(mod n).
bjm
bfm
(mod n).
Since (m, n) = 1, we have
A.5
The Euler V--function
315
This implies that j -1. Similarly, we obtain i - k. Thus, the rs integers an +b j m

represent distinct congruence classes modulo mn. If (a;n+b jm, mn) > 1 for some
i and j, then some prime p divides mn and a1 n+b jm. Since (m, n) - 1, the prime p
divides exactly one of m and n. If p divides m, then p divides an, and so p divides
a;. This contradicts the fact that (a;, m) - 1. Therefore, (a;n +bjm, mn) - I for
alliandj.
We shall show that every congruence class relatively prime to mn is of this
form. We note that (m, n) - 1 implies that the r integers a; n form a complete set of
representatives of the congruence classes relatively prime tom, and the s integers
bjm form a complete set of representatives of the congruence classes relatively
prime ton. Let (c, mn) - 1. Then (c, m) - 1, and so
c - ain (mod m)
for some i. Since
(c, n)-(c-ain, n)-1,

it follows that
c - a;n - bjm (modn)

for some j. Therefore,
c - atn + b jm (mod n)
and
c=a;n+bjm (mod m);

hence
c=a-n+bjm (modmn).
Thus,
rp(mn) - rs - rp(m)rp(n).
This proves that V is multiplicative. If p is prime and k > 1, the only integers not
prime to pk are multiples of p, and so
1
c0(Pk) - Pk - Pk_3 - pk
I -
Therefore,
--
w(n) - 11 V(P1) -
pk 1.
Theorem A.16 Let e > 0. Then

nt_`
for all sufficiently large n.
- -P - nl-pin
1
pk
< rp(n) < n
1-
316
Proof. It is clear that (p(n) < n for all n > 1. We shall prove that
nl-F
urn -/ -0.
n-.oo W(n)
Since p/(p - 1) < 2 for every prime number p, we have
pm(I-E)
pm(1-e)
Pm_pm-I
p-1
pm
pmE
pm(I-E)
(p(P'n)
Therefore,
lim
m(t -e)
- 0.
p1-'O0 gp(Pm)
Since the arithmetic function nl -E/V(n) is multiplicative, the result follows from
Theorem A.9.
Theorem A.17
1
<< log X.
nix (p(n)
Proof. Let d` denote the square-free part of d, that is,
d* -flp.
pfd
Then
1
v(n)n pin
'O
inEd'
P/
and so
d-1
1
dd
logx.
The integers of the form dd* are precisely the integers that are square .f-u11 in the
sense that if p divides d, then p2 divides d for every prime p. We have
E dd*
d-t
...1
H(1+
12+
13+
P
P
p
llJ
A.6 The Mobius function

+p(1-
HPP
317
PI)-,
P(p1-
1)
\1 +
A.6
The Mobius function
The fundamental property of the Mobius function is the following.
Theorem A.18
1: (d)-g(n)dIn
ifn-1,
ifn > 2.
Proof. This is clearly true for n - 1. If n > 2, then

k
P;H''
i-1
where k > 1, pi, ... , pk are distinct prime numbers, and r, > 1 for i - 1, ... , k.
Let E' denote a sum over square-free integers. Then
(d) - E'/2(d)
dIn
din
1: (d)
dlpj ...pt
(_ 1)a,cd)
dlpi...pk
E \e/ (- I )1
-0.

We define the arithmetic function
1(n) - 1
for all n. Then Theorem A.18 can be rewritten in the form
A*I-3.
318
A nonempty set D of positive integers is called divisor-closed if whenever n E D

and d divides n, then d E D. For example, the set N of all positive integers and the
set of positive integers less than a fixed number z are examples of divisor-closed
sets. The set of all divisors of a fixed positive integer is divisor-closed. If f and
g are functions defined on a divisor-closed set D, then their Dirichlet convolution
f * g is also defined on D.
Theorem A.19 Let D be a divisor-closed set, and let f (n) be a function defined
for all n E D. If g is the function defined on D by
g(n) - E f(d).
din
then
f(n) -
`j \dl g(d)
din
for all n E D.
Conversely, let g be a function defined on E). If f is the function defined on V
by
l
din
then
f(d)
g(n) din
for all n E D.
Proof. If n E V and d In, then d E D, since V is divisor-closed. Let
g(n)-Ef(d)
din
for n E D. Then
g - f * 1,
and so
( n)
g(d)-(g*A)(n)
din
- ((f * 1) * A)(n)
- (f * (I * A))(n)
(f
s)(n)
f(n).
similarly, if
f(n) -
l
din
l g(d) - (g * ti)(n),
( n)
d)
A.6
The M6bius function
319
then
1: f(d)-(f * 1)(n)
d In
-((g* s)*1)(n)
-(g*(A* 1))(n)
a)(n)
g(n).
(g
Theorem A.20 Let f and g be arithmetic functions. Then
g(n) - > f(d)

dIn
if and only if
f (n) - 1: is
g(d).
Proof. This follows immediately from Theorem A.19 with the divisor-closed
set V equal to the set N of all positive integers.
Theorem A.21 Let f (x) and g(x) be functions defined for all real numbers x
Then
f(x/d)
g(x) d <x
if and only if
f (x) -
(d)g(x/d)
d<x
Proof. Let f be a function defined for all x > 1. If
g(x) - E f (x/d),
d<x
then
1: s(d)g(x/d) - 1: (d) E f (x/dd')

d<x
d<x
d'<.r/d
E (d) f (x/dd')
dd' <x
1: f(x/m)Eis(d)
m<x
-f(x).
The proof in the opposite direction is similar.
dim
320
Theorem A.22 Let D be a finite divisor-closed set, and let f and g be funcrions
defined on D. If
g(n)-1: f(d)
dID
.Id
for all n E V. then
f(n)-EA(n)g(d)
dID
.Id
for all n E D. Conversely, if
f(n)-EA()g(d)
daD
.a
for all n E D, then
g(n) - 1] f(d)
dID
.M
for all n E D.
Proof. This is a straightforward computation:
E
,(.v
.Id
d4v
.N
f(k)
kaD
dIA
(h) E f (k)
+aP
nhED
.AA
N (h) E f(nhl)
nhlED
nhED
1 f (nr) >2 i (h)

nrED
AqV
hj,
f (nr) E a (h)
hIr
nrED
- f(n).
The proof in the opposite direction is similar.
A.7
Ramanujan sums
Let q and n be integers with q > 1. The exponential sum

q
q(n) - > e
la.q}I
an
(-l
q
(A.2)
Ramanujan sums
A.7
321
is called the Ramanujan sum. These sums play an important role in the proof of
Vinogradov's theorem (Chapter 8).
Theorem A.23 The Ramanujan sum cq(n) is a multiplicative function of q, that

is, if (q, q') = 1, then
cgq,(n) = cq(n)cq'(n).
Proof. Since every congruence class relatively prime to qq' can be written
uniquely in the form aq'+a'q with I < a < q, 1 < a' < q', and (a, q) - (a', q')
1, it follows that if (q, q') = 1, then
q
(an)
_=1 e
cq(n)cg
(n)
\a n
fo.y1-I
(a .4M1
qq I
Ia 91-1
to V>-1
94
(a"n
qq'
10 yq hl
= cqq'(n).
Theorem A.24 The Ramanujan sum can be expressed in the form
\91 d.
cq(n) _
In particular, if (q, n) = 1, then
cq(n) = (q)
Proof. Since
d I={
e(tn
fd(n)=>
if d in
if d Vn,
it follows that
q
(In)
1-1
cg(n)= E el\\
11.41-1
e (In)
J
k-1
(d)
dl(k.q)
In
=E/2(d)I:e\q )
dlq
k-1
dl
322
qld
A(d)Ee
f-l
dlq
In
\9/d/
' E A(d)fgld(n)
djq
1: (q/d)fd(n)
dlq
1: A(gld)d
dlq
din
- E A(gld)d.
dl(n.q)
If (q, n) a 1, then cq(n) - (q).

Theorem A.25 The Ramanujan sum can be expressed in the form
A(gl(q,n))co(q)
(P(q/(q, n))
cq(n)
Proof. We define
q' - q/(q, n).

If the prime p divides q but not q', then pI(q, n). It follows from Theorem A.15
that
97(q)
q ]-[tllq(l - 1/P)
97(q')
q'
q,(l - 11p)
(q, n) fl(1 - up)

pIq
vl4,
(q, n) [1 (1 - 1/P)
vl,q.m
vW'
Then
cq(n) - E (d
d
dl(q.n)
(q, n)
(q, n) d
d
1: 14 (q'c) d
cd-(q.n)
,a
(q) A(c)d
rJ-(q.nl
-I
it (q/)
rJ-cq.,,,
cd
Infinite products
A.8
(q')
n) E
-A(9')(9,n)
323
(c)
iQ,
v IC'
_ (9')(P(9)
1G(9')
A.8
Infinite products
This is a brief introduction to infinite products and Euler products.
Let a1, a2, ... , an, ... be a sequence of complex numbers. The nth partial
product of this sequence is the number
n
pn-a1...an-flak.
k-1
If as n tends to infinity, the sequence of nth partial products converges to a limit a
different from zero, then we say that the infinite product fl ak converges and
00
fl ak - Jim
pn - n-+oo
lim k-1
flak - a.
,,-+00
k-1
We say that the infinite product diverges if either the limit of the sequence of partial
products does not exist or the limit exists but is equal to zero. In the latter case, we
say that the infinite product diverges to zero.
Let
ak - 1 +a.
If the infinite product fkOO-1(1 + ak) converges, then ak
-1 for all k. Moreover,
Pk
-1,
lim(l+ak)- Jim
k-.o0 PA-1
k-,oo
and so
lim ak - 0.
k-+oo
Theorem A.26 Let ak > 0 for all k > 1. The infinite product nk- 1(1 + ak)
converges if and only if the infinite series E0 1 ak converges.
Proof. Let s,, - En_t ak be the nth partial sum and let p,, - fl .,1(1 + ak)
be the nth partial product. Since an > 0, the sequences {sn) and
monotonically increasing, and pn > I for all n. Since
1+x <e:
are both
324
for all real numbers x, we have

n
It
0 <Ea,< < fl(l+ak) <flea, - explEak)

k-1
k-1
k-1
k-1
0 < s,, < p < e'".

This inequality implies that the sequence { p,, } converges if and only if the sequence
{s,,} converges. This completes the proof.
We say that the infinite product ]-[,.1(1 + a,,) converges absolutely if the infinite
product
00
fl(1 + Ian l)
/I-1
converges.
Theorem A.27 If the infinite product f.00 I (1 +
converges absolutely, then it
converges.
Proof. Let
p _ f(1 + ak)
k-1
and let
k-1
If the infinite product converges absolutely, then the sequence of partial products
{ P } converges and so the series
00
E(P" - P.-I)
n-2
converges. Since
0 < IPn - Ian-I I

= lanpn-1I
n-I
=
k-1
n-I
Ia,, fl(l+lak1)
k-1
Ia,,IPn-I
= Pn - Pn - I ,
A.8
Infinite products
325
it follows that
00
IPn - P.-I I
n-2
converges, and so
00
(pn - PI)
E(Pn - P.-I) - lim
1:(Pk - Pk-I) - lim
n-+oo k-2
n-+o0
n-2
converges. Thus, the sequence of partial products { pn } converges to some finite

limit.
We must prove that this limit is not zero. Since the infinite product fl(1 +
ak) converges absolutely, it follows from Theorem A.26 that the series F_k-1 lak I
converges, and so the numbers ak converge to zero. Therefore, for all sufficiently
large integers k,
11 +akl > 1/2
and
-ak
21akI
1 + ak
It follows that the series

O0
k-1
I - ak
l-
+ ak
converges, and so the infinite product
k-1
ak
I+ak/
converges absolutely. This implies that the sequence of nth partial products
"
/1 -
ak
+ak)
k"I l +ak
(-1+ ak)
Pn
converges to a finite limit, and so the limit of the sequence {p,} is nonzero.
Therefore, the infinite product ]-[001(1 + ak) converges.
An Euler product is an infinite product over the prime numbers. We denote sums
and products over the primes by r p and fl p, respectively.
Theorem A.28 Let f (n) be a multiplicative function that is not identically zero.
If the series
converges absolutely, then

00
00
+f(Pk)
n-I
k-I
326
If f (n) is completely multiplicative, then

00
E f(n) - fl(1 - f(P))-I.

n-1
Proof. If r0c1 f (n) converges absolutely, then the series

00
ap - E f (PI)
k-t
converges absolutely for every prime p. Also, the series

OQ
E lapl
- E E f(Pk)
k-1
00
If(Pk)I
 0, and choose an integer No such that
I f(n)I < E.
n> No
For every positive integer n, let P(n) denote the greatest prime factor of n. Then
Y-P(n)<N denotes the sum over the integers all of whose prime factors are less
than or equal to N, and EP(n)>N denotes the sum over the integers that have
at least one prime factor strictly greater than N. Since the series Ek"-o f(pk)
converges absolutely for every prime number p, any finite number of these series
can be multiplied together term by term. Let N > No. It follows from the unique
factorization of integers as products of primes that
(1+f(Pk))_
pN
k-1
P(n)<N
and so
00
00
Ef(n)n-I
11
PN
f(n)
"0
k-I
n-1
f(n)
P(n)<N
A.10 Exercises
327
f(n)
P(n)>N
E If(n)I
P(n)>N
E If(n)I
n>N
E If(n)I
n> Na
< E.
Therefore,
(1+f(P')) =
00
f(n)= lim
l+f(P)
00
k-1
If f (n) is completely multiplicative, then f (pk) - f (p)k for all primes p and all
nonnegative integers k. Since f (pk) tends to zero as k tends to infinity, it follows
that If (p) I < 1. Summing the geometric progression, we obtain
l+Ef(Pk)-1+Ef(P)k=
k-1
and so
k-1
I - f (P)
FI(I+E
P
f(Pk)
k-1
Fl(1 - f(P))-1
P
Notes
A.9
All of the material in this chapter is basic elementary number theory. Comprehensive standard references are the books of Hardy and Wright[51 ] and Hua [63].
Cashwell and Everett [8] proved that the ring of arithmetic functions is a unique fac-
torization domain. Hardy's book Ramanujan [46] contains a chapter on Ramanujan's function cq(n) and its connection to the problem of representing numbers as
sums of squares.
A.10
Exercises
1. Prove that
1: (k)d(n/k) - I
kIn
for all n > 1.
328
2. Prove that if f and g are multiplicative functions, then the Dirichlet convolution f * g is multiplicative.
3. Let f and g be arithmetic functions. Prove that if f * g - 0, then either

f - 0 or g - 0. Thus, the ring of arithmetic functions is an integral domain.
4. An arithmetic function f (n) is additive if f (mn) - f (m) + f (n) for all

positive integers m and n such that (m, n) - 1. An arithmetic function f (n)
is completely additive if f (mn) - f (m) + f (n) for all positive integers m
and n. Let n - p"' - - - pk'. We define the arithmetic functions w(n) and 92(n)
as follows. The arithmetic function w(n) counts the number of distinct prime
factors of n:
w(n) - k.
The arithmetic function 92(n) counts the number of prime factors of n with
multiplicities:
f2(n) - r, + - - - + rk.
Prove that w(n) is additive but not completely additive. Prove that St(n) is
completely additive.
5. Let n - p(' - - pk. Liouville's function A(n) is defined by

A(n) - (-1)o(") Prove that A(n) is completely additive.
6. Let f (n) be an arithmetic function. There exists a unique completely multiplicative function f, (n) such that ft (p) - f (p) for all primes p. Show that
t(n) - A(n).
7. Show that the functions (n), fp(n), and oa(n) are not completely multiplicative.
8. Prove that
d(n) < 2n(") < n

for every positive integer n. Prove that if n is square-free, then
d(n) - 2i'(") - 2nt".
9. Prove that
E(d(n))Z >> x(logx)2,

n<x
Hint Apply the Cauchy-Schwarz inequality to F-n<x d(n).
10. Let f be an arithmetic function. Prove that f is invertible in the ring of

arithmetic functions if and only if f (1) - 1.
A.10
Exercises
329
11. Let f and g be arithmetic functions. Define the function L by
L(n) = logn.
Prove that pointwise multiplication by L(n) is a derivation on the ring of
arithmetic functions, that is,
L.(f *g)=(L.f)*g+f *(L.g)

12. Let f and g be arithmetic functions with Dirichlet generating functions F(s)
and G(s), respectively. Prove that F'(s) is the generating function for L f
and that (F(s)G(s))' is the generating function for L (f * g).

13. Prove that
fq(n)
n_1
> cd(n)
(an)
= d19
9
Use Mobius inversion to deduce Theorem A.24 from this identity.

14. Let
ar(n) _ Ed.
dlrr
Prove that
n<or (n)<nlog n+O(n).

Hint: a(n) = Ed,,, n/d.
15. Let (n) be the Mobius function. Prove that
Ann)
n-,
>< fl (l
P
PS
16. Prove that the Dirichlet convolution of arithmetic functions is associative,

that is, if f (n), g(n), and h(n) are arithmetic functions, then
(f*g)*h=f*(g*h).
17. Let L(n) = log n for all n > 1. For any arithmetic function f, define L f
by Lf (n) = L(n) f (n). Prove that L is a derivation on the ring of arithmetic
functions, that is,
L(f *g)=(Lf)*g+f *(Lg)
330
18. Let f, g, and h be arithmetic functions. Prove that
f(d)h(n/d)
g(n) din
if and only if
f(n) -
A(d)g(n/d)h(d).
din
19. Compute
2
F
k_2I (1
k(k+ 1))
20. Show that the infinite product

F1
k-2
converges, but not absolutely.
21. Let 0 < b < I for all n. Prove that if E' I bn converges, then rjoo t (1- bn )
converges.
22. Let 0 < bn < I for all n. Prove that if E', bn diverges, then Fj 1(1 - bn)
diverges to zero.
Bibliography
[I ] T. M. Apostol. Mathematical Analysis. Addison-Wesley, Reading, Mass., 1957.
[2] R. Balasubramanian. On Waring's problem: g(4) :S 20. Hardy-Ramanujanl., 8:140, 1985.
[3] E. Bombieri. Le grand crible dans la theorie analytique des nombres. Number 18 in
Ast6risque. Soci6te Math6matique de France, Paris. 1974.
[4] E. Bombieri, J. B. Friedlander, and H. Iwaniec. Primes in arithmetic progressions to

large moduli. Acta Math., 156:203-251, 1986.
[5] R. P. Brent. Irregularities in the distribution of primes and twin primes. Math.
Comput., 29:43-56, 1975.
[6] J. Brildern. On Waring's problem for cubes. Math. Proc. Cambridge Philos. Soc.,
109:229-256,1991.
[7] V. Brun. Le crible d'Eratosthene et le th6oreme de Goldbach. Skrifter utgit av Videnskapsselskapet i Kristiania, I. Matematisk-Naturvidenskabelig Masse, 1(3):1-36,
1920.
[8] E. D. Cashwell and C. J. Everett. The ring of number-theoretic functions. Pacific J.

Math., 9:975-985, 1959.
[9] A. L. Cauchy. Demonstration du th6or6me g6n6ral de fermat sur les nombres poly-
gones. Mcm. Sci. Math. Phys. Inst. France, 14(l):177-220,1813-1815. Oeuvres(2),

vol. 6, 320-353.
[10] J. Chen. On the representation of a large even integer as the sum of a prime and the
product of at most two primes. Kexue Tongbao, 17:385-386, 1966.
332
[ 11 ]
Bibliography
J. Chen. On the representation of a larger even integer as the sum of a prime and the
product of at most two primes. Sci. Sinica, 16:157-176. 1973.
[ 12] S. L. G. Choi. Covering the set of integers by congruence classes of distinct moduli.
Math. Comput., 25:885-895, 1971.
[ 13] S. L. G. Choi, P. Erd6s, and M. B. Nathanson. Lagrange's theorem with N''3 squares.
Proc. Am. Math. Soc., 79:203-205,1980.
[141 N.G. Chudakov. On the density of the set of even integers which are not representable
as a sum of two odd primes. Izv. Akad. Nauk SSSR, 2:25-40, 1938.
[15] B. Cipra. How number theory got the best of the pentium chip. Science, 267:175,
1995.
[ 16] R. Crocker. On the sum of a prime and two powers of two. Pacific J. Math., 36:103107,1971.
117] H. Davenport. On Waring's problem for cubes. Acta Math., 71:123-143, 1939.
(181 H. Davenport. Analytic Methods for Diophantine Equations and Diophantine

Inequalities. Ann Arbor Publishers, Ann Arbor, 1962.
[19] H. Davenport. Multiplicative Number Theory. Springer-Verlag, New York, 2nd

edition, 1980.
[20] V. A. Dem'yanenko. On sums of four cubes. Izv. Vyssh. Uchebn. Zaved. Mat.,
54(5):64-69, 1966.
[21] J.-M. Deshouillers and F. Dress. Sums of 19 biquadrates: On the representation of

large integers. Annali Scuola Normale Super. Pisa, 19:113-153, 1992.
[22] L. E. Dickson. History of the Theory of Numbers. Carnegie Institute of Washington,
Washington, 1919; reprinted by Chelsea Publishing Company in 1971.
[23) L. E. Dickson. All positive integers are sums of values of a quadratic function of x.
Bull. Am. Math. Soc., 33:713-720, 1927.
[241 L. E. Dickson. All integers except 23 and 239 are sums of eight cubes. Bull. Am.
Math. Soc., 45:588-591, 1939.
[25] F. Dress. Theorie additive des nombres, probleme de waning et theoreme de Hilbert.
Enseign. Math., 18:175-190, 301-302, 1972.
[26] H. B. Dwight. Mathematical Tables. Dover Publications, New York, 3rd edition,
1961.
[27] N. Elkies and I. Kaplansky. Problem 10426. Am. Math. Monthly, 102:70, 1995.
[28] W. J. Ellison. Waring's problem. Am. Math. Monthly, 78:10-36, 1971.
[29] W. J. Ellison and F. Ellison. Prime Numbers. John Wiley & Sons, New York, 1985.
[30] P. Erd6s and P. Turin. Ein zahlentheoretischer Satz. Izv Inst. Math. Mech. Tomsk
State Univ., 1:101-103, 1935.
Bibliography
333
[31] P. Erdos. On the integers of the form xt + y". J. London Math. Soc., 14:250-254,
1939.
[321 P. Erdos. On integers of the form 2{ + p and some related problems. Summa Brasil.
Math., 2:113-123, 1950.
[33) P. Erdos. On some problems of Bellman and a theorem of Romanoff. J. Chinese

Math. Soc. (N.S.), 1:409-421, 1951.
[34] P. Erdos. Some recent advances and current problems in number theory. In Lectures
on Modern Mathematics, volume 3, pages 196-244. Wiley, New York, 1965.
[35] P. Erdos and K. Mahler. On the number of integers which can be represented by a
binary form. J. London Math. Soc.. 13:134-139, 1938.
[361 P. Erdos and M. B. Nathanson. Lagrange's theorem and thin subsequences of squares.
In J. Gani and V. K. Rohatgi, editors, Contributions to Probability, pages 3-9.

Academic Press, New York, 1981.
[37] T. Estermann. On Goldbach's problem: Proof that almost all positive integers are
sums of two primes. Proc. London Math. Soc., 44:307-314, 1938.
[38) T. Estermann. Introduction to Modern Prime Number Theory. Campridge University
Press, Cambridge, England, 1952.
[39] P. Fermat. Oeuvres. Gauthier-Villars et Fils. Paris, 1891.
[40] A. Fleck. Ober die Darstellung ganzer Zahlen als Summen von sechsten Potenzen
ganzer Zahlen. Mat. Annalen, 64:561, 1907.
[41] K. B. Ford. New estimates for mean values of Weyl sums. Int. Math. Res. Not.,
(3):155-171, 1995.
[42] 1. S. Gradshteyn and I. M. Ryzhik. Table of Integrals, Series, and Products. Academic
Press, San Diego, 5th edition, 1994.

[43] E. Grosswald. Representations of Integers as Sums of Squares. Springer-Verlag, New
York, 1985.
[44] H. Halberstam and H.-E. Richert. Sieve Methods. Academic Press, London, 1974.
[45] G. H. Hardy. On the representation of an integer as the sum of any number of squares,
and in particular of five. Trans. Am. Math. Soc.. 21:255-284,1920.
[46] G. H. Hardy. Ramanujan: Twelve Lectures on Subjects Suggested by his Life and
Work. Chelsea Publishing Company, New York, 1959.
[47] G. H. Hardy and J. E. Littlewood. A new solution of Waring's problem. Q. J. Math.,
48:272-293,1919.
[48] G. H. Hardy and J. E. Littlewood. Some problems of "Partitio Numerorum". A new
solution of Waring's problem. Gottingen Nach., pages 33-54, 1920.
334
Bibliography
[49] G. H. Hardy and J. E. Littlewood. Some problems of "Partitio Numerorum": VI.

Further researches in Waring's problem. Mat. Z., 23:1-37, 1925.
[501 G. H. Hardy and S. Ramanujan. Asymptotic formulae in combinatory analysis. Proc.
London Math. Soc., 17:75-115, 1918.
[51 ] G. H. Hardy and E. M. Wright. An Introduction to the Theory of Numbers. Oxford

University Press, Oxford, 5th edition, 1979.
[52] F. Hausdorff. Zur Hilbertschen Losung des Waringschen Problems. Mat. Annalen,
67:301-305, 1909.
[53] T. L. Heath. Diophantus of Alexandria: A Study in the History of Greek Algebra.

Dover Publications, New York, 1964.
[54] D.R. Heath-Brown. Cubic forms in ten variables. Proc. London Math. Soc., 47:225257, 1983.
[551 D.R. Heath-Brown. Weyl's inequality, Hua's inequality, and Waring's problem. J.
London Math. Soc., 38:216-230, 1988.
[561 D. Hilbert. Beweis far die Darstellbarkeit der ganzen zahlen durch eine feste Anzahl
n" Potenzen (Waringsches Problem). Mat. Annalen, 67:281-300, 1909.
[571 C. Hooley. On the representation of a number as a sum of two cubes. Mat. Z., 82:259266, 1963.
[58] C. Hooley. On the numbers that are representable as the sum of two cubes. J. reine
angew. Math., 314:146-173, 1980.
[591 C. Hooley. On nonary cubic forms. J. reine angew. Math., 386:32-98, 1988.
[60] C. Hooley. On nonary cubic forms. J. reine angew. Math., 415:95-165, 1991.
[611 C. Hooley. On nonary cubic forms. J. reine angew. Math., 456:53-63, 1994.
[621 H. L. Hua. On Waring's problem. Q. J. Math., 9:199-202, 1938.

[63] H. L. Hua. Introduction to Number Theory. Springer-Verlag, Berlin, 1982.
[641 L. K. Hua. Additive Theory of Prime Numbers, volume 13 of Translations of

Mathematical Monographs. American Mathematical Society, Providence, R.I., 1965.
(65) A. Hurwitz. Uber die Darstellung der ganzen Zahlen als Summen von n` Potenzen
ganzer Zahlen. Mat. Annalen, 65:424-427, 1908.
[66] A. E. Ingham. The Distribution of Prime Numbers. Number 30 in Cambridge Tracts
in Mathematics and Mathematical Physics. Cambridge University Press, Cambridge,
1932. Reprinted in 1992.
[671 H. Iwaniec. Introduction to the prime number theory. Unpublished lecture notes,
1994.
[68] H. Iwaniec. Sieve methods. Unpublished lecture notes, 1996.
Bibliography
335
[69] W. B. Jurkat and H.-E. Richert. An improvement of Selberg's sieve method. I. Acta
Arith., 11:207-216, 1965.
[70] A. Kempner. Bemerkungen zum Waringschen Problem. Mat. Annalen, 72:387-399,
1912.
[71) H. D. Kloosterman. On the representation of numbers in the form axe+by2+cz2+dt2.

Acra Math., 49:407-464, 1925.
[72] H. D. Kloosterman. On the representation of numbers in the form axe+by2+cz2+dt2.
Proc. London Math. Soc., 25:143-173, 1925.
(73] H. D. Kloosterman. Over het uitdrukken van geheele positieve getallen in den vorm
ax 2 +by2 +cz2 +dt2. Verslag Amsterdam, 34:1011-1015, 1925.
[74] M. I. Knopp. Modular Functions in Analytic Number Theory. Markham Publishing
Co., Chicago, 1970; reprinted by Chelsea in 1994.
(75] E. Landau. Uber eine Anwendung der Primzahltheorie auf das Waringsche Problem
in der elementaren Zahlentheorie. Mat. Annalen, 66:102-106, 1909.
[76] E. Landau. Die Goldbachsche Vermutung and der Schnirelmannsche Satz. Gottinger
Nachrichten, Math. Phys. Klasse, pages 255-276, 1930.
(77] E. Landau. Uber einige neuere Fortschritte der additiven Zahlentheorie. Cambridge
University Press, Cambridge, 1937.
[78] E. Landau. Elementary Number Theory. Chelsea Publishing Company, New York,
1966.
(79] V. A. Lebesgue. Exercices d'Analyse Numl' erique. Paris, 1859.

[80] A: M. Legendre. Thiorie des Nombres. Firmin-Didot, Paris, 3rd edition, 1830.
[81) Yu. V. Linnik. On the representation of large numbers as sums of seven cubes. Mat.
Sbornik NS, 12:218-224, 1943.
[82] K. Mahler. Note on hypothesis K of Hardy and Littlewood. J. London Math. Soc.,
11:136-138, 1936.
[83] H. L. Montgomery. Topics in Multiplicative Number Theory. Number 227 in Lecture
Notes in Mathematics. Springer-Verlag, Berlin, 1971.
[84] H. L. Montgomery and R. C. Vaughan. The exceptional set in Goldbach's problem.

Acta Arith., 27:353-370, 1975.
(85] L. J. Mordell. On the four integer cubes problem. J. London Math. Soc., 11:208-218,
1936.
[86] L. J. Mordell. Diophantine Equations. Academic Press, London, 1969.
[87] Y. Motohashi. Sieve Methods and Prime Number Theory. Tata Institute for
Fundamental Research, Bombay, India, 1983.
336
Bibliography
[88] M. B. Nathanson. Products of sums of powers. Math. Mag., 48:112-113, 1975.

[89] M. B. Nathanson. Waring's problem for sets of density zero. In M. I. Knopp, editor,
Analytic Number Theory, volume 899 of Lecture Notes in Mathematics, pages 301310, Berlin, 1981. Springer-Verlag.
[901 M. B. Nathanson. A generalization of the Goldbach-Shnirel'man theorem. Am. Math.
Monthly, 94:768-771, 1987.
[91 ] M. B. Nathanson. A short proof of Cauchy's polygonal number theorem. Proc. Am.
Math. Soc., 99:22-24, 1987.
1921 M. B. Nathanson. Sums of polygonal numbers. In A.C. Adolphson, J. B. Conrey, A. Ghosh, and R. I. Yager, editors, Analytic Number Theory and Diophanune
Problems, volume 70 of Progress in Mathematics, pages 305-316, Boston, 1987.
Birkhauser.
[93] M. B. Nathanson. Additive Number Theory: Inverse Problems and the Geometry of
Sumsets, volume 165 of Graduate Texts in Mathematics. Springer-Verlag, New York,
1996.
[94] A. Oppenheim. Hilbert's proof of Waring's problem. Messenger Math., 58:153-158,

1928.
[95) T. Pepin. Demonstration du theoreme de Fermat sur les nombres polygones. Atti
Accad. Pont. Nuovi Lincei, 46:119-131, 1892-93.
[96] H. Poincare. Rapport sur Ie prix Bolyai. Acta Math., 35:1-28, 1912.
[97] K. Prachar. Primzahlverteilung. Springer-Verlag, Berlin, 1957.
[98] H. Rademacher. Topics in Analytic Number Theory. Springer-Verlag, New York,

1973.
[99] D.Raikov.OJberdieBasendernaturlichenZahlentreihe.Mat.SbornikN.S.2,44:595597, 1937.
[100] O. Ramare. On Inirel'man's constant. Preprint, 1995.

(101) P. Revoy. Sur les sommes de quatre cubes. Enseignement Math., 29:209-220, 1983.
[102] G. J. Rieger. Zur Hilbertschen Lbsung des Waringschen Problems: Abschatzung von
g(n). Archiv Math., 4:275-281, 1953.
[103] N. Romanov. Ober einige Satze der additiven Zahlentheorie. Mat. Annalen, 109:668678, 1934.
[104] M. Rosen. A generalization of Mertens' theorem. Preprint, 1995.

[105] E. Schmidt. Zum Hilbertschen Beweise des Waringschen Theorems. Mat. Annalen,
74:271-274, 1913.
[106] W. M. Schmidt. Analytische Methoden fir Diophantische Gleichungen. Birkhauser

Verlag, Basel, 1984.
Bibliography
337
[107] W. M. Schmidt. The density of integer points on homogeneous varieties. Acta Math.,
154:243-296,1985.
[108] B. Scholz. Bemerkung zu einem Beweis von W ieferich. Jahrber. Deutsch. Math. Ver.,
58:45-48, 1955.
[109] A. Selberg. On an elementary method in the theory of primes. Norske V id. Selsk.
Forh.,Trondheim, 19(18):64-67, 1947.
[110] A. Selberg. Collected Papers, Volume 1. Springer-Verlag, Berlin, 1989.
[1111 A. Selberg. Collected Papers, Volume 11. Springer-Verlag, Berlin, 1991.
[112] D. Shanks and Jr. J. W. Wrench. Brun's constant. Math. Camp., 28:293-299, 1183,
1974.
[113] L. G. Shnirel'man. On the additive properties of integers. Izv. Donskovo Politekh.

Inst. Novocherkasske, 14:3-27, 1930.
[1141 L. G. Shnirel'man. Ober additive Eigenschaften von Zahlen. Mat. Annalen, 107:649-
690,1933.
[1151 J. H. Silverman. Taxicabs and sums of two cubes. Am. Math. Monthly, 100:331-340,
1993.
[116] J. H. Silverman and J. Tate. Rational Points on Elliptic Curves. Springer-Verlag, New
York, 1992.
[117] M. K. Sinisalo. Checking the Goldbach conjecture up to 4 1011. Math. Comput.,

61:931-934, 1993.
[1181 J. Spencer. Four squares with few squares. Preprint, 1990.
[119] A. Sti hr. Eine Basis h-Ordnung fur die Menge aller naturlichen Zahlen. Mat. Z.,
42:739-743,1937.
[1201 E. Stridsberg. Sur la demonstration de M. Hilbert du th6oreme de Waring. Mat.
Annalen, 72:145-152, 1912.
[1211 G. Tenenbaum. Introduction to Analytic and Probabilistic Number Theory.

Cambridge University Press, Cambridge, 1995.
[122] J. V. Uspensky and M. A. Heaslet. Elementary Number Theory. McGraw-Hill, New
York, 1939.
[123] J. G. van der Corput. Sur 1'hypothr se de Goldbach pour presque tous les nombres
pair. Acta Arith., 2:266-290, 1937.
[124] R. C. Vaughan. Sommes trigonomEtriques sur les nombres premiers. C. R. Acad. Sci.
Paris, Sir. A, 285:981-983, 1977.
[125] R. C. Vaughan. The Hardy-Littlewood Method. Cambridge University Press,

Cambridge, 1981.
338
Bibliography
[126] R. C. Vaughan. On Waring's problem for cubes. J. reine angew. Math., 365:122-170,
1986.
[ 127] R. C. Vaughan. A new iterative method in Waring's problem. Acta Math., 162:1-71,
1989.
[ 1281 R. C. Vaughan. The use in additive number theory of numbers without large prime
factors. Philos. Trans. Royal Soc. London A, 345:363-376, 1993.
[129] R. C. Vaughan and T. D. Wooley. On Waring's problem: some refinements. Proc.

London Math. Soc., 63:35-68, 1991.
[ 130] R. C. Vaughan and T. D. Wooley. Further improvements in Waring's problem. Acta
Math., 174:147-240, 1995.
[1311 1. M. Vinogradov. On Waring's theorem. lzv. Akad. Nauk SSSR, Old. Fiz.-Mat. Nauk,
(4):393-400, 1928. English translation in Selected Works, pages 101-106, SpringerVerlag, Berlin, 1985.
1132)
I. M. Vinogradov. Representation of an odd number as the sum of three primes.

Doklady Akad. Nauk SSSR, 15(6-7):291-294, 1937. English translation in Selected
Works, pages 129-132, Springer-Verlag, Berlin, 1985.
[133] 1. M. Vinogradov. Some theorems concerning the theory of primes. Mat. Sbornik,
2(44):179-195,1937.
[134] 1. M. Vinogradov. The Method of Trigonometric Sums in the Theory of Numbers,
volume 23. Trud. Mat. Inst. Steklov, Moscow, 1947. English translation published
by interscience, New York, 1954.
[135] 1. M. Vinogradov. The Method of Trigonometric Sums in Number Theory. Nauka,
Moscow, 1980. English translation in Selected Works, pages 181-295, SpringerVerlag, Berlin, 1985.
[136] R. D. von Sterneck. Sitzungsber. Akad. Wiss. Wien (Math.), 112, Ila:1627-1666,
1903.
[ 137] Y. Wang. Goldbach Conjecture. World Scientific, Singapore, 1984.
(1381 E. Waring. Meditationes Algebraicae. Cambridge University Press, Cambridge,

1770.
[ 139] G. L. Watson. A proof of the seven cube theorem. J. London Math. Soc., 26:153-156,
1951.
[140] A. Weil. Sur les sommes de trois et quatre carres. Enseign. Math., 20:215-222, 1974.
[141] H. Weyl.UberdieGleichverteilungvonZahlenmodEins.Mat.Annalen,77:313-352,
1913.
[1421 H. Weyl. A half-century of mathematics. Am. Math. Monthly, 58:523-553,

1942. Reprinted in Gesammelte Abhandlungen, volume IV, pages 464-494,
Springer-Verlag, Berlin, 1968.
Bibliography
339
[143) H. Weyl. David Hilbert and his mathematical work. Bull. Am. Math. Soc., 50:612654. 1944. Reprinted in Gesammelte Abhandlungen, volume IV, pages 130-172,
Springer-Verlag, Berlin, 1968.
(1441 A. Wieferich. Beweis des Satzes, daB sich eine jede ganze Zahl als Summe von
hochstens neun positiven Kuben darstellen IaBt. Mat. Annalen, 66:95-101, 1909.
(145] E. Wirsing. Thin subbases. Analysis, 6:285-308, 1986.

11461 T. D. Wooley. Large improvements in Waring's problem. Ann. Math., 135:131-164,
1992.
[147] T. D. Wooley. On Vinogradov's mean value theorem. Mathematika, 39:379-399,

1992.
[148] T. D. Wooley. Breaking classical convexity in Waring's problem: Sums of cubes and
quasi-diagonal behavior. Inventiones Math., 122:421-451, 1995.
[149] T. D. Wooley. Sums of two cubes. Int. Math. Res. Not., (4):181-185, 1995.
[150] E. M. Wright. An easier Waring's problem. J. London Math. Soc., 9:267-272, 1934.
[151) J. Zollner. Der Vier-Quadrate-Satz and ein Problem von Erd6s and Nathanson. PhD
thesis, Johannes Gutenberg-Universitat, Mainz, 1984.
[1521 J. Zollner. Ober eine Vermutung von Choi, Erdos, and Nathanson. Acra Arith.,
45:211-213,1985.
Index
Additive basis, 2
additive function, 328
adjoint equation, 262
almost prime, 271
asymptotic basis, 33
Basis, 2
basis of finite order, 192
binary quadratic form, 2
Brun's constant, 173
Brun's theorem, L73
Cauchy's lemma, 30
Cauchy's theorem, 31
Chebyshev functions, 154
Chen's theorem, 221
Choi-Erdos-Nathanson theorem, 24
circle method, L21
classical bases, 7
completely multiplicative function, 308
counting function, 121
covering congruences, 204
Difference operator, 99
Dirichlet convolution, 301

Dirichlet series, 151
discriminant of a form, 8
divisor-closed set, 318
Easier Waring's problem, 72. 102

equivalent matrices, 8
equivalent quadratic forms, 8
Erd6s-Mahler theorem, 61
Euler products, 325
Euler sum formula, 306
Euler's constant, 306
exceptional set, 230
Goldbach conjecture, 112

Goldbach-Shnirel'man theorem, 127
Hardy-Littlewood asymptotic formula,

L46
Hennite polynomial, 72
Hilbert-Waring theorem, 88
Hooley-Wooley theorem, 66
Hua's lemma, l lfi
342
Index
Implied constant, xiii

inclusion-exclusion principle, 174
infinite product, 323
Jurkat-Richert theorem, 251
Lagrange's theorem, 5
large sieve inequality, 295
Legendre's formula, 232
linear sieve, 231
Linnik's theorem, 46
lower bound sieve, 23.4
Major arcs, 126, 213

minor arcs, 127. 213
multiplicative function, 308
Riemann zeta-function, 151
Selberg sieve, L80

seven cube theorem, 46
Shnirel'man density, 192
Shnirel'man's addition theorem, 193
Shnirel'man's constant, 208
Siegel-Walfisz theorem, 46,216
sieve dimension, 23.8
sieving function, 232
sieving level, 232
sieving range, 232, 234
singular series, 131
sumset, 192
support level, 234
symmetric matrix, 2
Ternary quadratic form, 9

Partial summation, 304
polygonal number theorem, 31
polygonal numbers, 4
positive-definite form, 9
Upper bound sieve, 234
Vinogradov's theorem. 212

Quadratic form, 8
Ramanujan sum, 321

Ramanujan's sum, 212
Ramare's theorem, 208
Waring's problem, 37
well approximated, 12t
Weyl's inequality, 11.4
Wieferich-Kempner theorem, 41
The classical bases in additive number theory are the polygonal

numbers, the squares, cubes. and higher powers, and the primes.
This book contains many of the great theorems in this subject:
Cauchy's polygonal number theorem, Linnik's theorem on sums of
cubes, Hilbert's proof of Waring's problem, the Hardy-Littlewood
asymptotic formula for the number of representations of an integer
as the sum of positive kth powers, Shnirel'man's theorem that every
integer greater than one is the sum of a bounded number of primes,
Vinogradov's theorem on sums of three primes, and Chen's theorem that every sufficiently large even integer is the sum of a prime
and a number that is either prime or the product of two primes.
The book is also an introduction to the circle method and sieve methods, which are the principal tools used to study the classical bases.
The only prerequisites for the book are undergraduate courses in

number theory and analysis.
Additive number theory is one of the oldest and richest areas of
mathematics. This book is the first comprehensive treatment of the
subject in 40 years.
ISBN 0-387-94656-X

164 - Additive Number Theory PDF

Uploaded by

Copyright:

Available Formats

164 - Additive Number Theory PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

164 - Additive Number Theory PDF

Uploaded by

Copyright:

Available Formats

Graduate Texts

Graduate Texts in Mathematics

Graduate Texts in Mathematics

TAKEUTI/ZARING. Introduction to Axiomatic Set

35 WFRmFR. Banach Algebras and Several

Theory. 2nd ed.

36 KELLEY/NAMIOKA El AL. Linear Topological

3 SCHAEFER. Topological Vector Spaces.

38 GRAUERT/FRrrZSCHE. Severa! Complex

6 HUGHES/PIPER. Projective Planes.

40 KEMENY/SNELL/KJ APP. Denumerable Markov

7 SFRRE. A Course in Arithmetic.

Chains. 2nd ed.

9 HUMPHREYS. Introduction to Lie Algebras and

12 BEMs. Advanced Mathematical Analysis.

19 HALMOS. A Hilbert Space Problem Book. 2nd

Complex Variables. 2nd ed.

39 ARVtsoN. An Invitation to C'-Algcbras.

46 LoEva. Probability Theory 11. 4th ed.

47 Mots.. Geometric Topology in Dimensions 2

57 CRowt;u/Fox. Introduction to Knot Theory.

60 ARNOLD. Mathematical Methods in Classical

the Theory of Groups.

64 EDWARDS. Fourier Series. Vol. 1. 2nd ed.

65 WEDS. Differential Analysis on Complex

continued after index

Additive Number Theory

Mathematics Subject Classifications (1991): 11-01, 11P05, I IP32

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

C 1996 Melvyn B. Nathanson

Production managed by Hal Henglein; manufacturing supervised by Jeffrey Taub.

a bounded number of kth powers, and the Hardy-Littlewood asymptotic formula

Notation and conventions

Waring's problem for cubes

The Hilbert-Waring theorem

Polynomial identities and a conjecture of Hurwitz

Easier Waring's problem . . . . . .

5 The Hardy-Littlewood asymptotic formula

II The Goldbach conjecture

Elementary estimates for primes

The Shnirel'man-Goldbach theorem

The Goldbach conjecture . . . . . .

Sums of three primes

9 The linear sieve

10.1 Primes and almost primes . . .

The ring of arithmetic functions

A.4 The divisor function .

Notation and conventions

Theorems, lemmas, and corollaries are numbered consecutively in each chapter

if there exists a constant c > 0 such that

if there exists a constant c > 0 that depends on a, b, ... such that

Notation and conventions

for all x in the domain of f . We write

The function f is asymptotic to g, denoted

e(x) - exp(27rix) - e2i".

by the inequality [x] < x < [x] + 1.

ceps in infinitum, in hexagonis, heptagonis polygonis quibuslibet,

P. Fermat [39, page 303]