The

Binary Additive Problems

Translations of Mathematica/ Monographs Volume 4

THE DISPERSION METHOD

IN BINARY ADDITIVE
PROBLEMS

by

Ju. V. linnik

AMERICAN MATHEMATICAL SOCIETY

PROVIDENCE, RHODE ISLAND
1963
 JJJ1CIIEPCHOHHhifi METO.U
B BHHAPHbIX A.IJ..IJ.HTHBHbIX
3A.IJ,A1.JAX

IO. B. JIHHHHK

H3,z:r;aTem.crao Jiemrnrpa.z:r;cKoro
YHHBepcHTeTa

JleHHHrpa.z:r; 1961

Translated from the Russian by S. Schuur

Publication aided by grant NSF-GN 57

from the
NATIONAL SCIENCE FOUNDATION

Library of Congress Card Number 63-15660

Copyright © 1963 by the American Mathematical Society

All rights reserved. No portion of this book may be reproduced

without the written permission of the publisher

Printed in the United States of America

 PREFACE

This book is devoted to the systematic application of elementary concepts of

probability theory-the concepts of dispersion, covariance, and the Cebysev in-
equality-to the solution of a number of binary additive problems. Supplemented
by appropriate purely arithmetic ideas, such an approach permits the solution of
several binary additive problems which are inaccessible to the known methods of
modern additive number theory (the latter being applicable mainly to ternary
additive problems).
It should be remarked that such an application of probabilistic concepts to
number theory has little in common with the probabilistic theory of numbers, a
systematic exposition of which may be found in [8]. Roughly speaking, proba-
bilistic number theory can be understood as a particular form of the theory of
probability without the continuity axiom of Kolmogorov; in this theory the
central concept is that of the independence of events relating to different primes.
The dispersion method, on the other hand, simply takes for use in additive
number theory the simplest concepts of the theory of probability concerning a
finite field of elementary events. Roughly speaking, it can be understood as the
beginning of a correlation theory of binary additive problems. In one of its
parts (the construction of the fundamental inequality for the dispersion; cf.
(0.2.6)) it closely resembles Vinogradov's method for the estimation of double
trigonometric sums.
The ideas of such a correlation theory, discussed in detail in the Introduction,
are not complicated, but its application to concrete binary additive problems
involves a great deal of rather cumbersome computation (the calculation of the
dispersion of the number of solutions) and requires a number of deep arithmetic
lemmas. All these lemmas have been separated into the auxiliary Chapters I
and II, which contain their enunciations and references to the literature, and
sometimes detailed proofs as well (for new lemmas). Chapters III, VII, and VIII
contain detailed solutions of three well-known problems: the additive divisor
problem (for the case k 1 = 2 and any kJ, a problem of Hardy and Littlewood,
and a problem of Titchmarsh concerning divisors and primes. In Chapter IV
we give the application of the dispersion method to a generalization of a well-
known theorem of Kloosterman. In Chapter V a binary additive problem of
some generality, involving a binary quadratic form, is discussed briefly; here the
computation had to be omitted because of its unwieldiness. It should be re-
marked that the results of Chapter V cannot at present be deduced from the
iii
IV PREFACE

extended Riemann hypothesis, while the results of Chapter VIII would follow
from this hypothesis.
Chapter VI treats in detail several analogues of a Hardy-Littlewood equation.
Chapter IX outlines the solution of the generalized Hardy-Littlewood equation,
n = <p(x, y) + p, where <p(x, y) is a binary quadratic form. Chapter X contains
supplementary remarks and various conditional theorems, following from the
dispersion method.
In preparing the manuscript for publication I received a great deal of help
from B. M. Bredihin, 0. I. Rumyanceva, V. V. Vifoyakova and M. E. I'ina, and
find it a pleasant duty to express my gratitude to them.
 NOTATION

K 0, Ki. K2 , ••• ; C0 , C1 , ••• are "sufficiently large" constants; the choice of

each depends on the choice of the preceding ones.
In each chapter the numbering of these and the following constants begins
anew. Sometimes this is also done in individual sections; the reader is informed
of this where it occurs.
a0 , a 1 , ••• ; c0 , c1 , ••• are positive constants, usually "not very large."
e0 , e1 , • . • ; rJo, r1i, ... ; ~0 , ~1 , ••• are "sufficiently small" positive constants.
B denotes a bounded function of all variables introduced in a given chapter;
B is not always the same. In essence it is O(l) (Ordnung), but it emphasizes
uniformity with respect to possible parameters.
B 0 is the same, but depends on C, so that for different C, !Bel can have
different upper bounds.
p denotes a prime, p' a quasiprime (cf. § 5, Chapter I).
pP II m (the notation of Hasse) denotes that the pth power of p divides m
exactly, i.e. m = 0 (mod pP); m ¢. 0 (mod pP+l).
cp(q) is the Euler function; x(m) denotes a Dirichlet character.
-r(m) is the number of divisors of m; (mi. m 2) denotes the greatest common
divisor.
The abbreviation n.s.e. is used throughout for "number of the solutions of a
given equation."

v
 TABLE OF CONTENTS
PREFACE iii

NOTATION v

INTRODUCTION. GENERAL SCHEME OF THE DISPERSION METHOD

1. Binary and ternary problems 1
2. Outline of the dispersion method in the simplest case 5
3. Different versions of the dispersion method. The covariance of the
number of solutions 8
4. Extension of the concept of covariance of the number of solutions.
Coherent numbers 15
5. A variant on the dispersion method, including the application of
covariance, in the absence of sufficiently many coherent numbers 17
6. Some particular cases of the application of the dispersion method 19

CHAPTER I. LEMMAS ON THE NUMBER OF DIVISORS, ON PRIME NUMBERS

AND TRIGONOMETRIC SUMS
1. Lemmas on the number of divisors 23
2. Lemmas on the distribution of numbers m with certain restrictions
on the value of Q(m) 27
3. Lemmas proved by the method of the sieve of Eratosthenes 27
4. Lemmas on certain sums containing Dirichlet characters 29
5. Quasiprime numbers 30
6. Lemmas on the distribution of primes and of certain products in
progressions 31
7. Lemmas orr trigonometric sums 34

CHAPTER II. LEMMAS FROM THE ANALYTIC THEORY OF QUADRATIC FORMS

1. Fundamental equations of the dispersion method for the divisor
problem 36
2. Lemma on the fundamental equations 37
3. Analysis of the fundamental congruence 40
4. Application of the method of Vinogradov 43
5. An asymptotic formula for the number of solutions when (x, z) = 1 46
6. The case (x, z) > 1 48
vii
viii TABLE OF CONTENTS

7. Comparison of the results for coherent numbers 50

8. Lemma for the Hardy-Littlewood problem 52
9. Analysis of the fundamental congruence. Another lemma 54
10. Lemma on a certain indefinite quaternary form 55
11. Preparation for the application of Vinogradov's method 58
12. Application of the method of I. M. Vinogradov 59
13. Summation of the basic series 61
14. A quaternary indefinite form of more general type 62

CHAPTER III. THE ADDITIVE DIVISOR PROBLEM

1. Statement of the problem and several results 64

2. Application of the dispersion method 66
3. Preparatory transformations of the equation 67
4. The dispersion of the difference between solutions for coherent
numbers 72
5. Application of the Cebysev inequality 73
6. Reduction of the additive divisor problem to a ternary additive
problem 75
7. Further transformations. Estimation of the error term 79
8. The final formula for the number of solutions 81
9. Application of the dispersion method to other cases of the additive
divisor problem · 82

CHAPTER IV. APPLICATION OF THE DISPERSION METHOD TO A GENERALIZA-

TION OF A THEOREM OF KLOOSTERMAN

1. Statement of the problem 84

2. Preparatory transformations of equation (4.1.5) 86

CHAPTER v. ON A CERTAIN BINARY EQUATION OF GENERAL TYPE

INVOLVING A QUADRATIC FORM

1. Statement of the problem 91

2. On one particular case of equation (5.1.1) 92
3. The final formula for the number of solutions of (5.2.1) 94

CHAPTER VI. ANALOGUES OF A HARDY-LITTLEWOOD EQUATION

1. Statement of the problem 95

'2. Reduction of the equations to a form convenient for the application
of the dispersion method 96
3. The case k = 6 98
4. Transformation of equation (6.1.1) 99
 TABLE OF CONTENTS ix

5. Determination of A(n, D) 99
6. The dispersion of the number of solutions. Calculation of V1 102
7. Calculation of the product over primes in V1 106
8. Calculation of V3 107
9. Calculation of the products over primes in V3 109
10. Calculation of V2 112
11. Calculation of the product over primes in V2 115
12. Application of the Cebysev inequality 116
13. Derivation of the asymptotic expression for the solutions 117

CHAPTER VII. A PROBLEM OF HARDY AND LITTLEWOOD

1. On an equation of Hardy and Littlewood 119

2. Transformation of the Hardy-Littlewood equation 120
3. The equations Y~, k < 5 121
4. Investigation of Qk(n) 123
5. The nonirregular intervals 125
6. Transformation of the number of solutions 126
7. The final formula for Qk2'(n) 129
8. Investigation of Q1°>(n) 130
9. Investigation of LB 132
10. The covariance of the solutions of the auxiliary equations 134
11. Investigation of the class An 136
12. Estimation of LE 137
13. Estimation of L 2 139
14. Estimation of Ln 141
15. Preparation for the estimation of Le 145
16. Investigation of the fundamental congruences 148
17. Further examination of the fundamental congruences 149
18. Continuation of the examination of the fundamental congruences 151
19. Estimation of Le 152
20. Investigation of LA 154
21. Estimation of the auxiliary sum 157
22. Estimation of the error term 160
23. Final calculation of Qk(n) 161
24. Completion of the proof 163

CHAPTER VIII. A PROBLEM OF TITCHMARSH CONCERNING DIVISORS AND

PRIMES

1. Statement of the problem 164

2. Application of the dispersion method 165
3. The equations Y~ for k < 5 167
x TABLE OF CONTENTS

4. Investigation of Q<~>(n). Coherent numbers 168

5. Investigation of ~B 170
6. The differences between the values of ~..:t and between the values
of ~ 0 172
7. Reduction of the binary additive problem to a ternary one 173
8. Solution of the ternary problem 174

CHAPTER IX. OUTLINE OF THE SOLUTION OF A GENERALIZATION OF THE

HARDY-LITTLEWOOD EQUATION

1. Statement of the problem 176

2. Outline of the solution 177

CHAPTER X. SUPPLEMENT

1. On simplifying the calculation of the dispersion 179

2. Certain deductions from the application of the dispersion method to
the binary Goldbach problem and its analogues 180
3. On the equation n = p + x 2 and its analogues 180

BIBLIOGRAPHY 183
 INTRODUCTION*

GENERAL SCHEME OF THE DISPERSION METHOD

1. Binary and Ternary Problems

During the 1920's and 1930's Hardy, Littlewood and Vinogradov developed a
general method in analytic number theory which made it possible to prove
asymptotic formulas for many additive problems. Vinogradov's discovery of
some of the fundamental properties of trigonometric sums opened the way to the
solution of a whole class of additive problems, a class which we shall call ternary.
In order to explain the characteristic features of such problems, we outline
briefly the classical proof of the Goldbach-Vinogradov theorem. Here we are
concerned with the equation
Pi + P2 +Pa = N, (0.1.1)
where Pi (i < 3) are primes and N is a large odd number. We introduce the sum
S(fJ) = I exp 27Ti1Jp,
1',,;;,N

where p runs through primes. Using NO to denote the number of solutions of an

equation inside the parentheses and writing

Q(N) = N(p1 + P2 + Pa = N),

we have
Q(N) = f (S(fJ))a exp (-27TiN1J) d#. (0.1.2)

Divide the interval [O, 1] into two sets, 11 and its complement / 2, where 11
consists of those 1} which can be exceptionally well approximated by rational
fractions, and forms a set of intervals of small total measure, which can easily be
described explicitly. Then, with an obvious notation,

Q(N) = r +r .
J11 J1.
(0.1.3)

* Translator's note: For advice and assistance with the final version of this Introduction
the translator expresses her grateful acknowledgment to Professor H. Davenport. She also
wishes to thank the author, Professor Ju. V. Linnik, for his encouragement of the project
and for certain slight corrections of his original text.
1
2 INTRODUCTION

An asymptotic approximation for S11 is found by using known results on the

distribution of the primes p =:;;; N into arithmetic progressions qx + /, where
(/, q) = 1 and 0 =:;;; I < q. Here q may increase with N, but not faster than a
fixed (but arbitrarily large) power of ln N. The integral over 11 furnishes the
main term in the asymptotic formula.
The integral over 12 is estimated in the following very characteristic manner:

f12
IS(&)l 3 d& .::;;; sup IS(&)I
~el2 Jof IS(&)l
1
2 d&

=sup IS(&)I S(O), (0.1.4)

~I2

where S(O) = TT(N) = O(N/ln N) is the number of primes not exceeding N.

In order to obtain an asymptotic formula for Q(N), it suffices to prove, for
& E 12 , an estimate of the form

IS(&)I .::;;; S(O) , where g(N) ~ (ln N)1+eo (0.1.5)

g(N)
and s0 is an arbitrary positive constant.
It will be seen that the main part of the work in deriving a sufficiently good
estimate forJ12 is performed by Parseval's equation, which is a simple analytic
property of the trigonometric polynomial S(&). For this we need two sums S(&)
in order to have (S(&)) 2 • After this, there remains the problem (and it is a
difficult problem) of obtaining the modest improvement (0.1.5) on the trivial
estimate for the sum S(&). Thus in this work the presence of three terms is
essential, and the problems to which the method is applicable are of ternary
character.
The scheme just outlined is suited to the solution of ternary problems of the
type
1X+,8+y=N. (0.1.6)
Here IX and ,8 run through sets of numbers which are sufficiently dense in the
interval [l, N], comprising (say) more than N/(ln N) 0 numbers, where c is a
positive constant, and sufficiently well distributed in arithmetic progressions
qx + I, where q may increase slowly with N. The sequence of numbers y may be
comparatively sparse, for instance ~y.;;N 1 may be approximately a power of
ln N; but the fractional parts {&y}, for & E 12 , must be distributed with some
regularity, in order that the sum
,L exp 2TTi&y
y.;;N

may satisfy a sufficiently good estimate.

Equation (0.1.6) represents a typical ternary problem. For its solution we
require that IX and ,8 belong to "dense" sequences and that the sequence y, though
Sec. 1] BINARY AND TERNARY PROBLEMS 3

it may be sparse, has a certain "diophantine smoothness" in the sense just

described.
In this connection it should be noted that (0.1.4) can be extended as follows: if
S1 (#) =! exp 2?Ti1X#; S2(#) =! exp 2'7Tip#; S3({}) =! exp 2?Tiy1J,
a~N P~N 1~N
then

~.!sup ISa(#)I (S1(0) + S2(0)).

2 ~!2
Some problems which might seem at first sight to be of a binary character
can really be formulated as ternary problems; this applies, for example, to
Cudakov's result [27] that almost all even numbers (asymptotically) are sums of
two primes. Let N be a sufficiently large number, and let al> a 2 , ••• , an be those
even numbers ~N that are not sums of two primes. Let g(N) be a monotonic
function, increasing arbitrarily slowly, and suppose that for some infinite
sequence of N we have n ;;;;,. N/g(N). Having chosen one such N, sufficiently
large, we form the equation
-ai +Pi + P2 = 0, (0.1.7)
where Pi and P2 are primes not exceeding N and ai ~ N. If we prove the
solvability of (0.1.7) for a particular choice of the function g(N), it will follow
that for N;;;;,. N 0 we have n < N/g(N), which is the desired result. Equation
(0.1.7) gives a ternary additive problem of the type (0.1.6), in which -ai takes
the place of IX (and in this connection it is important that n;;;;,. N/g(N)),Pi takes
the place of p, and P2 that of y. It should be observed that in the calculation of
f 11 it suffices to make use of the good distribution of Pi· p 2 among arithmetic
progressions; the distribution of the ai, which is entirely unknown, is not needed.
An additive problem involving more than two terms, say
b1 + b2 + ... +bk= N, (0.1.8)
where the bi run independently through certain sequences of numbers, may be
regarded as a ternary problem if it is possible to combine sums of terms bi into
new terms in such a way as to produce an equation of the type (0.1.6), where
repetition of numbers is allowed. For this it is important that the distinct
numbers IX and the distinct numbers p should constitute sufficiently dense
sequences, say
! 1 > c(e)Ni-•, ! 1 > c(e)Ni-•, (0.1.9)
a~N P~N

for any e > 0, where c(e) is a constant depending on e.

From this point of view, the classical equation
x2 + y2 + z2 + t2 = N (0.1.10)
4 INTRODUCTION

does not make a ternary problem, because it is not possible to combine the terms
so as to satisfy (0.1.9) for ex and f3 and have a third term y.
Problems such as this, where it is impossible to distinguish three terms which
run independently through three sequences, at least two of which are sufficiently
dense, will be called binary problems.
The exact formula for the number of solutions of (0.1.10) was found over a
century ago by Jacobi, with the help of the theory of theta-functions. For this,
certain special properties of squares are essential. In 1926, by combining the
theory of the transformation of the theta-functions with the Hardy-Littlewood-
Vinogradov method, Kloosterman obtained an asymptotic expression for the
number of solutions of the equation
ax 2 + by2 + cz2 + dt 2 = N, (0.1.11)
where a, b, c, dare given positive integers (cf. [46]).
The more general problem of the equation
Q(x, y, z, t) = N, (0.1.12)
where Q is a positive definite quaternary quadratic form, was solved by Tarta-
kovskii [24] in 1929, in the sense that an asymptotic expression was found for the
number of solutions. Eichler [33] in 1954 used an entirely different method, and
obtained an error term in the asymptotic formula which is essentially best
possible. (The solvability of (0.1.12) for sufficiently large N, when the appropri-
ate necessary conditions are satisfied, can be proved by elementary methods.)
Quadratic forms in four variables or in three variables (as regards the latter,
see several papers by Malysev, e.g. [18; 19]) furnish us with a few examples of
binary problems for which an asymptotic expression can be found for the number
of solutions. As a well-known example of a binary problem for which only an
upper bound and a lower bound for the number of solutions have been found, we
may mention the equation
(0.1.13)
with "almost prime" numbers P 1 , P 2 • The path to the solution of this problem
was opened up by Viggo Brun [30] in 1919. The best result so far known is that
(0.1.13) is solvable for all large even N with the condition that P 1 , P 2 each have
at most three prime factors [3; 4].
In 1948 Renyi [50] established the solvability of the equation
p + P = N, (0.1.14)
where pis prime and P "almost prime," and N is sufficiently large.
Special methods are needed for the solution of all the binary problems
mentioned above, even if one does not aim at obtaining an asymptotic formula
for the number of solutions. 1
1 It may be noted that the discovery of the appropriate asymptotic formula for (0.1.13)

would substantially facilitate the solution of the binary Goldbach problem.

Sec. 2] DISPERSION METHOD IN THE SIMPLEST CASE 5

2. Outline of the Dispersion Method in the Simplest Case

Let {cp} be any sequence of positive integers, possibly with repetitions; let D'
run through any set of distinct positive integers in the interval [D1 , D 1 + D 2] =
(D); let y run (independently of D') through any set of positive integers in the
interval [Y0, Yo+ Y(,] = (Y); and let n be an integer greater than
(D1 + D 2)(Y0 + 0). Y

The simplest subject for the application of the dispersion method is the
binary problem presented by the equation
n = <p + D'y (0.2.1)
We shall seek conditions under which we can derive an asymptotic formula for
the number of solutions of the equation or congruence (abbreviated to n.s.e.)
(0.2.1).
Let U(m) = ,Lqi=m 1. Then n.s.e. (0.2.1) is given by
_L _L U(n - D'Y). (0.2.2)
D'e(D) ve(v)

Let DE (D) be a given integer, not necessarily in the set D', and consider the
equation
n = <p +DY. (0.2.3)
Suppose we have some heuristic reason for expecting that the n.s.e. (0.2.3)
may be given asymptotically by an expression
A(n, D) (0.2.4)
formed from elementary functions, in order to permit certain simple calculations
which will be described later. Such a heuristic reason might result from an
application of the Hardy-Littlewood-Vinogradov method without an estimate
for the error term, or from a similar application of the sieve of Eratosthenes, or
from the method of least squares (see below).
The fundamental concept of the dispersion method is to consider the dis-
persion, or variance, of the n.s.e. (0.2.1) relative to the assumed asymptotic
value A(n, D). By this is meant the expression

V' = _L ( _L U(n - D'Y) - A(n, D'))2. (0.2.5)

D'e(Dl ve(v)

It is important that (0.2.5) is a double sum. Following one ofVinogradov's basic

ideas for the estimation of double sums, we replace the summation over D' by a
sum over all integers DE (D), and obtain

V' "' V = _L
D1 ,,;;;D.;;;D1+D2
( _L
ve(vl
U(n - DY) - A(n, D)) 2
• (0.2.6)
6 INTRODUCTION

It may be remarked that we could attempt to find A(n, D) by the method of

least squares, i.e. by choosing it so as to minimize V.
Removing the brackets in (0.2.6), we obtain

(0.2.7)
where
2
Vi = L ( 1 U(n - Dv) ) , (0.2.8)
D1 .;;;D.;;;D1+D2 ve(v)

V2 = L A(n, D) L U(n - Dv), (0.2.9)

D1 .;;;D.;;;D1+D2 ve(v)

Va = L
D1.;;;D.;;;D1+D2
(A(n, D)) 2 • (0.2.10)

Now we must seek asymptotic expressions for Vi, V2, Va which may allow us to
deduce a nontrivial upper bound for V.
It should be possible to calculate the sum Va asymptotically in a straight-
forward way. If the numbers v are not too large and the numbers D are not too
small, for instance if v0 + v0 ~ n112--" and D 2 > n112+' (where e is a positive
constant), we can approximate asymptotically to V2 as follows. We write

V2 = ! L
ve(v) D1 .;;;D.;;;D1+D2
A(n, D)U(n - Dv). (0.2.11)

We observe that ~n 1 ,;;;;:V,;;;;n 1 +n 2 U(n - Dv) is the number of solutions of the

congruence <p = n (mod v), with restrictions on the variable <p. The presence of
the simple factor A(n, D) Should not unduly complicate the asymptotic calcula-
tion of V2 •
The computation of V1 is the most difficult task. In (0.2.8) we separate the
term
L (U(n - f)v)) 2 ,
D1.;;;D.;;;D1+D2

for which a crude upper bound can be taken; there remains the sum
! L U(n - Dv 1 )U(n - Dv 2 ). (0.2.12)
D1 .;;;D.;;;D1+D2 vi. v2e(v)
v1 :iCv2

To compute (0.2.12), we need to take in turn each pair v1 , v2 and consider the
total number of solutions of the system of equations

(0.2.13)

Here <p1 and <p 2 run independently through the sequence {cp} and (n - <p 2)/v1 =
= D, any integer in (D).
(n - <p1)/v 2
Sec. 2] DISPERSION METHOD IN THE SIMPLEST CASE 7

We divide the pairs vi, v2 into classes according to the value of their g.c.d.
(vi, v2)= o. We shall show that for each given pair with divisor o equations
(0.2.13) are equivalent to the single equation

(0.2.14)
subject to the conditions

n - 'Pi n - 'Pi
is an integer, -- E (D). (0.2.15)
'J12

Note that when o = 1 the first of the conditions (0.2.15) is always satisfied as
a consequence of (0.2.14); this will emerge from what follows. It is obvious that
if Cfli, cp 2 satisfy (0.2.13) for some D, then they satisfy (0.2.14) and (0.2.15), and
the numbers of solutions (counting Cfli and cp 2 with repetitions if necessary) are
the same. Conversely, suppose Cfli, cp 2 satisfy (0.2.14) and (0.2.15). Putting
'Vi = ov1, 'V2 = ov2, we have (vi, v2) = l and

(0.2.16)
Hence (n - <fJi)vi = (n - cp 2)v2, whence n - Cfli = 0 (mod v2) and n - cp 2 = 0
(mod vi). If o= 1, this in itself implies that (n - Cfli)/v 2 is an integer. If o > 1,
the condition (0.2.16) and the factthat (n - Cfli)/v 2 = (n - cp 2)/vi give a unique D
for which (0.2.13) holds. The mapping (<pi, cp 2) ___.. (D, Cfli, <p 2) is single-valued,
and our assertion is proved. Equation (0.2.14) and the conditions (0.2.15) can of
course be written in the more symmetric form

(0.2.17)

(0.2.18)

This equation is fundamental for the application of the dispersion method to

equation (0.2.1 ). A successful asymptotic evaluation of the number of solutions
leads to an asymptotic expression for Vi, and hence for V. Here the ratio Vi/V3
is of primary importance; if it is sufficiently small, we obtain an asymptotic
solution of the binary problem (0.2.1) under fairly general conditions. From
now on we shall suppose that the set of numbers D' is sufficiently dense, namely
that the number of numbers D' is at least D 2/(ln n)K 1 , where Ki and later K 2, ...
denote constants. The numbers v;, on the other hand, can form a very sparse
sequence. Now suppose we have succeeded in treating (0.2.17) and (0.2.18)
satisfactorily, and so have found an asymptotic formula for Vi, after which
we discover that V/V3 (and hence also V'/V3 ) is small. To be precise, suppose
that
(0.2.19)
8 INTRODUCTION

If K 2 is sufficiently large, we obtain immediately the solvability of the equa-

tion (0.2. l). For if it were not solvable, we should have U(n - D' Y) = 0 for all
admissible values of D' and Y. Then it would follow from (0.2.5) and (0.2.6) that
V ~ V' = L (A(n, D'))2 • (0.2.20)
D'e(Dl

Because of the density of the set D' assumed above, this will contradict (0.2.19) if
A(n, D) is a function of D which changes sufficiently slowly. Hence the result.
Furthermore (0.2.19) makes it possible to obtain an asymptotic formula for
the number of solutions of (0.2.1). To do this, it is enough to appeal to the
classical argument of Cebysev, which is used to deduce the law of large numbers
from an estimate of the dispersion of a sum of independent random variables.
We can, for example, pick out from the numbers D' occurring in the sum (0.2.5)
those for which

I L U(n -
ve(vl
D'Y) - A(n, D') I~ A(n~ D').
ln n
(0.2.21)

The estimate (0.2.19) implies that the number of such D' is comparatively
small. For these values of D', it suffices to determine a crude upper bound,
uniform with respect to D', for the sum I:.e<•> U(n - D'Y). After this, it is not
difficult to deduce that the n.s.e. (0.2.1), which is
L L
D'e(D) ve(v)
U(n - D'Y),

differs by a comparatively small amount from

L A(n, D')
D'e(Dl

and this gives the desired asymptotic evaluation.

3. Different Versions of the Dispersion Method.

The Covariance of the Number of Solutions
The basic equation (0.2.1) was discussed above on the supposition that the
numbers D' and Y are confined to a rectangular region:
(0.3.1)
Sometimes we have to consider a hyperbolic region of variation for D' and Y:
D'y ~n. (0.3.2)
In many cases such a region can be replaced by a union of rectangles in such a
way as to introduce only a permissible error into the number of solutions of
(0.2.1). Here it is essential that the condition Yo + Y~ ~ n 112-•o be maintained;
in other words, the numbers Y must not be too large. We can permit repetitions
Sec. 3] DIFFERENT VERSIONS OF THE DISPERSION METHOD 9

in the numbers D' on condition that the mean square of the number of repeti-
tions is not too large, and we can also permit repetitions in the numbers v.
We proceed to outline the application of the dispersion method under such
conditions. Let T 0(D') denote the number of repetitions of D', and suppose, to
be specific, that
L h(D'))2 < D 2(ln n)K•. (0.3.3)
D'e(D)

Divide the numbers D' into two sets A and B, putting in A those for which
T0 (D') < (In n)< 213>Ka, (0.3.4)
and in B those for which
To(D') > (In n)<2/3)K,_
By (0.3.3) the number of numbers D' in B is less than
D2(ln n)<-112>Ka, (0.3.5)
so if K 3 is sufficiently large there are few such numbers. We can expect that their
contribution to the sum (0.2.2), namely

L L U(n - D'v), (0.3.6)

D'eB ve(v)

will be comparatively small. Here, of course, an appropriate upper bound is

needed. When we form the dispersion of the number of solutions for which D'
is in A, we obviously get

V' L (L
= D'eA ve(v}
U(n - D'v) - A(n, D')) 2

< (In n)< 2! 3>Ka L ( L

D"eA ve(v)
U(n - D"v) - A(n, D"))
2
. (0.3.7)

Here D" runs through the set of numbers A, but without repetition, so we can
apply the arguments of § 1.
Repetitions in the numbers v are also permissible, as is clear from the argu-
ments of§ 1.
We obtained the fundamental inequality (0.2.6) by replacing the set ofnumbers
D' by the set of all positive integers in the interval [D1 , D 1 + D 2]. Sometimes it is
advantageous to construct instead an inequality in which the set D' is replaced by
all the numbers of some reasonably simple system in [D1 , D 1 + D 2]. For example,
ifthe numbers D' are contained in some arithmetic progression (e.g. if they are
all odd), then in forming the inequality (0.2.6) we can allow D to range over all
terms of this progression in the interval [D1 , D 1 + D 2]. The effect will be that in
the conditions (0.2.15), under which the equation (0.2.14) is to be solved, we
must require in addition that (n - <p1)/v2 (and therefore also (n - cp 2)/v1) lies in
the arithmetic progression. In some problems, in which the numbers D' are
10 INTRODUCTION

primes and the inequality (0.2.6) is too crude, it is useful to construct a similar
inequality, replacing the primes by quasiprimes-an "almost prime shell" (this
concept will be defined in § 5 of Chapter I).
Note that the dispersion method is of course applicable also to equations of
the type
n = <p - D'v, (0.3.8)
e.g., the equation <p - D' v = 1 or 2. In the formulas (0.2.2), (0.2.5) and later,
one replaces U(n - D'v) by U(n + D'v). The n.s.e. (0.3.8) is finite because of
the finiteness of the sets of numbers D' and v.
It is also possible to treat certain generalizations of the basic equation (0.2.1 ).
Let {3(v) be a function defined on a set of numbers v, which assumes only positive
integral values. We consider the equation
n = <p + D'v + {3(v) (0.3.9)
and outline the application of the dispersion method to it. The n.s.e. (0.3.9) is
given by
L L U(n - D'v - {3(v)).
D'e(D) ve(v)

Suppose that by heuristic considerations (as in § 2) we obtain an expression

A(n, D) as a likely approximation to the number of solutions of
n = <p + Dv + {3(v) (D.3.10)
for given D. The dispersion of the number of solutions of (0.3.9) from the
expected number A(n, D) is given by

V' = L (L
D'e(D) ve(v)
U(n - D'v - {3(v)) - A(n, D')) 2
• (0.3.11)

Again we construct the fundamental inequality of the type (0.2.6), replacing the
set of numbers D' by all numbers of the interval (D):

V' < V = Di .;;D.;;Di+D2

! (!
ve(v)
U(n - Dv - {3(v)) - A(n, D))
2
• (0.3.12)

As in§ 2 we put V = Vi - 2V2 + V3 , where V3 has the form (0.2.10) and

2
Vi = L (
L U(n - Dv - {3(v)) ) , (0.3.13)
Di ,,;;;D,,;;;Di+D2 ve(vl

V2 = ! A(n, D) ! U(n - Dv - {3(v)). (0.3.14)

Di ,,;;;D,,;;;Di+D2 VE(v)

Observe that I:ni,,;;;D.;;Di+n 2U(n - Dv - {3(v)) is the number of solutions of

the congruence <p = n - {3(v) (mod v) under the appropriate restrictions on
the variable <p; here A(n, D) should be a simple function, the presence of which
Sec. 3] DIFFERENT VERSIONS OF THE DISPERSION METHOD 11

as a factor should not affect significantly the calculation of an asymptotic expres-

sion (see § 2).
We turn to the estimation of V1 • Separate from it the term
z
D1.;;D.;;D1+D2
(U(n - Dv - /3(v))) 2,

for which it suffices to have an upper bound. There remains the sum
z z
Di .;;D.;;D1 +D2 •1.V2E(•)
U(n - Dv 1 - /3(v 1)) U(n - Dv 2 - /3(v2)), (0.3.15)
Vl ¢vz

which is equal to the total number of solutions of the system of equations

n - Dv1 - /3(v1) = <p2,
(0.3.16)
n - Dv 2 - /3(v2) = <pi·
Here (vi. v2) runs through pairs of distinct numbers from the set v; <p1 and <p2 run
independently through numbers of the set {<p}; D is any integer in the interval
[D1 , D 1 + D 2], so that

n - <p 2 - /3(v 1) E (D), n - <pi - /3(v2) E (D).

V1 V2
As in § 2 we divide the pairs (v1 , v2) into classes according to the value of g.c.d.
(v1, 1'2) = c5.
We shall show that for given v1 , v2 (and hence c5) the number of solutions
(D, <p1, <p 2) of the system (0.3.16) is the same as the number of solutions (<p1 , <p 2)
of the single equation
V1<p1 - V2<p2 = n(v1 - v2} + /3(v1)v2 - /3(v-z)v1, (0.3.17)
subject to the conditions
n - <p1 - /3(v 2) = 0 (mod v2), (0.3.18)

n - <p 1 - /3(v2) E (D). (0.3.19)

V2
The fundamental equation (0.3.17) could alternatively be written in the symmetric
form
(0.3.20)
subject to the same conditions (0.3.18) and (0.3.19); note that in (0.3.19) one
can interchange the suffixes 1 and 2.
As in § 2 we observe that for given v1 and v 2 , each solution (D, <p1 , <p 2) of the
system (0.3.16) under the conditions applicable gives a solution ( <p1, cp 2) of (0.3.20)
with the same vi. v2 , and the conditions (0.3.18) and (0.3.19) are obviously
satisfied. Thus the number of triplets (D, cp 1 , cp2) does not exceed the number of
12 INTRODUCTION

pairs <pi, <p 2 which satisfy (0.3.20) with the conditions (0.3.18), (0.3.19). We shall
prove that it is also not less than this number. Put Yi = ov1, v2 = ov2, so that
(v:[, v2) = 1 ; then from (0.3.20) we have, for each solution <pi, <p 2 ,

v;_(n - <pi - /3(v2)) = v;(n - <p 2 - {3(vi)).

Hence n - <pi - f3(v 2) = 0 (mod v2) and n - <p 2 - {3(vi) =
0 (mod v:[). Thus if
o= 1 the condition (0.3.18) is automatically satisfied; if > 1 the number o
(n - <pi - f3(v 2))/v 2 is an integer, and by (0.3.20) it is equal to (n - <p 2 - f3(vi))/vi.
Let the common value be D, then the triplet (D, <pi, <p 2) gives a solution of(0.3.16)
under the conditions applicable, and D is uniquely determined by <pt> <p 2 • This
proves the assertion.
If we can obtain a satisfactory asymptotic approximation for the number of
solutions of the equation (0.3.17) subject to (0.3.18), (0.3.19), and if we find that
V, and consequently also V', is small compared with V 3 , then we can deduce an
asymptotic formula for the number of solutions of (0.3.10).
Together with the dispersion of the number of solutions of an equation of the
type (0.2.1) or (0.3.10), it is often convenient to use a concept which, by analogy
with ideas of probability theory, may be called the covariance of the number of
solutions.
Suppose we are given two equations of the type (0.2.1):
n = <p + D'v, (0.3.21)
n='lf+ D'v, (0.3.22)

where D' runs through a given set of numbers in the interval [Di, Di + D 2], v
runs through a set of numbers in the interval [v0 , v0 + v0] and <p, 'If run through
certain sequences of positive integers. Sometimes it is necessary to compare the
numbers of solutions of (0.3.21) and (0.3.22) in a situation where there is reason
to expect that their difference is relatively small, and where we are not interested
in asymptotic approximations to the numbers themselves. For an example of
such a situation, see § 4 of Chapter III.
Let Ui(m) = Lqi=m 1, U2(m) = L>p=m 1. Construct the dispersion of the
difference between the numbers of solutions:

V' = L (L Ui(n - D'v) - I U 2(n - D'v)) •

2
(0.3.23)
D'e(D) ve(v) ve(v)

As before we replace V' by V > V', given by

(0.3.24)

We have
(0.3.25)
Sec. 3] DIFFERENT VERSIONS OF THE DISPERSION METHOD 13

where
(0.3.26)

(0.3.27)

(0.3.28)

The expressions V1 and V3 are treated as before. It is natural to call V2 the

covariance of the numbers of solutions of (0.3.21) and (0.3.22). If we discard the
expression

for which it suffices to give a crude upper bound, there remains the sum
z z
Di ,,;;;;D,,;;;;D1+D2 v,,v2e(v)
U1(n - Dv1)U 2(n - Dv 2), (0.3.29)

which is equal to the total number of solutions of the equations

n - Dv1 = <p,
(0.3.30)
n - Dv 2 = tp,
subject to the condition
n - <p
D = --E(D). (0.3.31)
'1'1

We see, as before, that for given vl> v2 (v1 ~ v2) this number of solutions is
equal to the number of solutions (<p, tp) of the equation
(0.3.32)
subject to the conditions
n - <p = 0 (mod v1), (0.3.33)

n - <p E (D). (0.3.34)

'1'1

If (v1 , v2) = 1 the condition (0.3.33) is automatically satisfied. If it is possible to

obtain an asymptotic approximation for the number of solutions of (0.3.32)
under the relevant conditions, and if we ascertain that V1 - 2 V2 + V3 is
sufficiently small, we may obtain a satisfactory estimate for the dispersion of the
difference between the numbers of solutions of our two equations, and show that
the difference is comparatively small. It is an advantage of this procedure that
we do not need to compute the corresponding asymptotic expression A(n, D).
14 INTRODUCTION

We can generalize the question by considering, instead of (0.3.21) and

(0.3.22), two equations with different left-hand sides:
n1 = r:p + D'v, (0.3.35)
n2 = tp + D'v. (0.3.36)
In this connection, however, as will be seen later, it is advisable to postulate that
if n1 and n2 have small prime divisors, these coincide. The ideas indicated here
can be useful when the calculation of an appropriate A(n, D) is difficult but there
is some means of computing the total number of solutions of n = r:p + D' v when
n runs through some set of numbers. If, by application of the dispersion method,
we are able to conclude that the numbers of solutions for individual n are nearly
equal, then we can obtain an asymptotic evaluation of the number of solutions of
our equations for individual n. Later we shall have examples of such a situation.
We now go on to discuss the application of the dispersion method to a system
of equations.
Let {~~} and {r:p~} be two sequences of positive integers depending on an
integral vector parameter (as, for example, the values of two quadratic forms
r:p'(x, y, z, t) and r:p"(x, y, z, t) with g = I=(x, y, z, t)). We consider the system
of equations
n' = r:pg + D'v'
(0.3.37)
n" = r:p'i + D"v",
where D' and D" run independently through certain sets of numbers in the
interval [DI> D1 + D 2], and v' and v" run independently through certain sets of
numbers in the interval [v 0 , v0 + v0]. We put
U(m', m") = L l.
tp~=m';tp~=m"

Let D.' and D." be any integers of the interval [D1 , D1 + D 2]; for given D.', D." we
consider the system of equations
n' = r:pg + D.'v',
n" = r:p'i + D."v",
and we suppose that an expression A(n', n", D.', D.") has been found, by heuristic
considerations, for the expected number of solutions of the system. We write the
dispersion of the number of solutions of the system of equations (0.3.37) in the
form
V' = L ( L U(n' - D'v', n" - D"v") - A(n', n", D.', /J.")) 2 • (0.3.38)
D',D" v',v"
Then
V' ~ V = L ( L U(n' -D.'v', n" -D."v")-A(n', n",/J.',/J.")) 2• (0.3.39)
D1~.6.'~D1+D2 v',v"
Di.;;;&" .;;;D1+D2
Sec. 4] COHERENT NUMBERS 15

Here, too, the main problem in finding an upper bound for V' lies in determining
the behavior of the sum
V1 = !
D1<.6.'~D1+D2
(v',v"
! U(n' - !:J..'v', n" - !:J.."v")) 2 • (0.3.40)
D1 .;;a".;;D,+D2
By arguments similar to those given above, we see that after removing from
(0.3.40) the terms for which it is enough to give a crude upper bound, we shall
arrive at the system of equations
(vi. - v~)n' = vi.<pi.s - v~<p;q
(0.3.41)
(v~
- v~)n" = v~ <p~s -
v~ <p~s'
where vi :;f v2, and v1, v2 are numbers. of the set {v'}, and similarly for v'{, v2, and
<J?is' <t>2s run independently through the sequence {<p£} and similarly for <p'{;, <t>2s'
and the following conditions are satisfied:
n' - <t>is = 0 (mod v;); n" - <p~s = 0 (mod v~),
(0.3.42)

Similar arguments apply to a system of any number of equations of the type

(0.3.37). We do not dwell on this in detail, since we shall not be concerned with
systems of equations of the type (0.3.37) in the present monograph. We shall
limit ourselves in what follows to some remarks on the representation of a set of
numbers by a system of quadratic forms.

4. Extension of the Concept of Covariance of the Number of Solutions.

Coherent Numbers
In § 3 we considered the covariance of the numbers of solutions of equation
(0.3.21) and (0.3.22), which have the same non the left. Sometimes it is useful to
consider the covariance of the numbers of solutions of equations of the form
n1 = <p + D'v, (0.4.1)
n2 = <p + D' v, (0.4.2)
where in both equations D' and v run independently through the sets of numbers
considered earlier. If U(m) = Lip=m 1, then in analogy with (0.3.23) we shall
call the expression
V' = L ( ve(v)
D'e(D)
L U(n 1 - D'v) - L U(n
ve(v)
2 - D'v))
2
(0.4.3)

the dispersion of the difference between the numbers of solutions of (0.4.1) and
(0.4.2). As usual, we replace V' by V ~ V', where

V = L
D1 .;;D.;;D1+D 2
( ve(v)
L U(n1 - Dv) - L
ve(v)
U(n 2 - Dv)) •
2
(0.4.4)
16 INTRODUCTION

We write V = Vi - 2V2 + Va, where

2
Vi = L L U(n(
1 - Dv) ) , (0.4.5)
Di .;;;D.;;Di+D2 ve(v)
2
Va = L (
L U(n 2 - Dv) ) , (0.4.6)
Di .;;D.;;Di+D2 ve(v)

V2 = L
Di .;;D.;;Di+D2
(L ve(v)
U(n1 - Dv) L
ve(v)
U(n 2 - Dv)). (0.4.7)

We shall call V2 the covariance of the numbers of solutions of our equations.

By the arguments of § 3, the asymptotic determination of V1 reduces to deter-
mining the number of solutions of
(0.4.8)
subject to
(0.4.9)

Similarly the computation of Va leads to the equation

V1<p1 - V2<p2 = n2(v1 - V2) (0.4.10)
with the conditions
(0.4.11)

Finally, in the same way, the computation of V2 leads to

(0.4.12)
subject to
(0.4.13)

If (v1 , v2) = I, the congruence conditions are automatically satisfied.

We shall consider the total number of solutions of (0.4.8), (0.4.10), (0.4.12),
taking all pairs v1, v 2 such that v1 :;If v2 and vi E (v) (i = 1, 2).
In many particular cases, for example when <p = <p(x, y) runs through the
values of a quadratic form or of a form of higher degree in either one or two
variables, a curious phenomenon occurs: the number Vi - 2 V2 + Va turns out
to be comparatively small provided that n1 , n2 have the same small divisors.
More precisely, if
(0.4.14)
and if g(n) is a suitable monotonic function, say g(n) = (ln n)K•, then I V1 - 2 V2 +
Val will be small in comparison with Vi provided that, for any b ~ g(n), either
b I n 1 and b I n 2, or b 1 n 1 and b 1 n 2 (so that the small divisors of n 1 and n 2
coincide). Numbers n1 and n 2 with this property will be said to be coherent. If,
Sec. 5] VARIANT ON THE DISPERSION METHOD 17

for coherent n 1 and n 2 , the dispersion of the difference between the numbers of
solutions behaves in the manner just stated, we observe that V' will be com-
paratively small. Hence, using considerations related to Cebysev's inequality
like those described earlier, we can conclude in many cases that asymptotically
the numbers of solutions of (0.4.1) and (0.4.2) will be nearly equal. If this
conclusion holds for any pair of coherent numbers, we consider the equation
<p + D'v - n' = 0, (0.4.15)
in which n' runs through some sufficiently large set M of numbers which are
coherent in pairs. Then (0.4.15) represents in effect a ternary problem, which
can be solved easily enough by Vinogradov's method [7], since this equation in-
volves the product D 111, which gives rise to the exponential sum~ D',v exp 2TTirxD' v.
Having found an asymptotic expression for the n.s.e. (0.4.15), on dividing it by M
we obtain an approximation to the number of solutions of (0.4.1) or (0.4.2). Such
an approach has the advantage of avoiding the need to compute an approxima-
tion A(n, D) for the expected number of solutions, and thus avoiding the rather
cumbersome calculations which so arise.
The same arguments can, of course, be applied to the equation <p - D' v = n.
Note that, for the equation <p - D'v = 1, the corresponding equations with
coherent numbers will be, for example, <p - D'v = p, where pis a sufficiently
large prime.

5. A Variant on the Dispersion Method, Including the Application

of Covariance, in the Absence of Sufficiently Many Coherent Numbers
Sometimes the set of numbers, coherent with a given number and not exceed-
ing n, turns out to be very thin. For example, if n = Pib ... Pm (the product of
all primes up to Pm), and g(n) is taken to be the function (In n) 0 , where c > 1, the
numbers coherent with n will be simply the multiples of n; so that of the
numbers not exceeding n the only one coherent with n is n itself. If such a
situation arises, we can apply the basic variant of the dispersion method de-
scribed in § 2. If, when doing this, the estimation of A(n, D) and the computa-
tions related to it turn out to be cumbersome, the following variant of the
dispersion method with the application of covariance may prove to be less so.
Choose a sufficiently large constant C0 ; let g(n) = (In n) 00 • Let n = (ITn)n',
where rrn consists of the prime power factors of n of primes p ~ g(n). If
(0.5.1)
where e > 0 is sufficiently small, we can apply the method described in § 4,
taking as the set of coherent numbers the numbers n" = (ITn)p, where the p's are
primes satisfying
n
n - - -0- ~ ITnP ~ n (0.5.2)
(Inn) 1
18 INTRODUCTION

Ci being suitably chosen. But iflln > n', then we represent the number D' E (D)
in equations (0.4.1) and (0.4.2) in the form
D'=Q'D", (0.5.3)
where Q' consists of the powers of primes p,.;;; g(n) and D" has no prime factors
,.;;; g(n). If
Q' ~ exp (ln ln n) 2 , (0.5.4)
and the set of numbers D' E (D) is sufficiently dense (contains at leastD 2/(ln nfs
numbers) then those solutions for which D' is of the form (0.5.3) can in many
cases be neglected. Therefore we shall suppose that
Q' < exp (ln In n) 2• (0.5.5)
Let n" run through the set of integers of the interval

[n - (ln :)01, nJ. (0.5.6)

We examine the numbers Q' satisfying (0.5.5) and for each of them the numbers
of the form (0.5.3). If we consider

! U(n" - Q' D"v), (0.5.7)

ve(v)

we can expect that for each D' = Q' D", this will be approximately of the form
A(n", Q' D") = p(n", Q')A 0(n, Di, D 2), (0.5.8)
where p(n", Q') consists of "arithmetic" factors, determined by n" and Q', and
A 0(n, Di, D 2 ) depends on "dimensional" or "geometrical" factors, determined
only by the size of n and the dimensions of the interval (D) = [Di, Di + D 2 ].
This suggests an examination of the dispersion of the difference between the
numbers of solutions in the form

V~. = ! (p(n", Q') ! U(n - Q' D"v) - p(n, Q') ! U(n" - Q' D"v))
2
•
Q'D"e(D) ve(v) ve(v)
(0.5.9)
Let D run through numbers which have no prime factor p < g(n). Replacing
the numbers D" by the numbers D, we obtain

Vi;/·,.;;; VQ' = !
Q'l>e(D>
(p(n", Q') !
ve(v)
U(n - Q' Dv) - p(n, Q') !
ve(v)
U(n" - Q' Dv)) •
2

(0.5.10)
We have VQ' = ViQ' - 2V2Q' + v3Q'• where
ViQ' = (p(n", Q'))2 ! (! U(n - Q' Dv))2 (0.5.11)
D1~Q'l>~D1+D2 ve(v)
Sec. 6] APPLICATION OF THE DISPERSION METHOD 19

with similar expressions for V30 , and V20 , (cf. (0.4.5) and (0.4.6)). As earlier,
the asymptotic evaluation of these quantities leads to the consideration of three
equations:
v1 cp 1 - v2 cp 2 = n(v1 - v2), n - cp1 = 0 (mod v2),
(0.5.12)
n - cp 1 =
0 (mod Q') ( n = l, ~'(/Ji, Ilg)
where we have put Ilg = Ilp.;:;;g<n>P•
v1 cp 1 - v 2 <p 2 = v1 n - v2 n", n - cp 1 = 0 (mod v2),
n - <fJ1 = 0 (mod Q') (n Q'<J?l' rrg) = l,
(0.5.13)

v1cp1 - v 2 <p 2 = n"(v1 - v2), n" - cp1 = 0 (mod v2),

(0.5.14)
n" - <J?1 = 0 (mod Q') (n ~' <J?i 'rrg) = 1.
In many cases, the total numbers of solutions of our three equations when
V; (v), 'Jl1 ;;if 'J12 are such that the quantity vl - 2V2 + Va turns out to be
E
comparatively small. In this case the dispersion Vi5· is small so that, applying
the analogue of Cebysev's inequality, we can conclude that
L
Q'D"e(D>
p(n, Q') L
ve(v)
U(n" - Q' D"v)

is approximately equal to
L - p(n", Q') L U(n - Q' D"v).
Q'D"e(D) ve(v)

Thus we can obtain an asymptotic expression for

L
VE(V)
U(n - Q' D"v),
by means of
p(n, Q') L 1 U(n" - Q' D"v), (0.5.15)
(n"l p(n", Q')
where the numbers n" run through the numbers of the interval (0.5.6). Such an
expression is not difficult to calculate.

6. Some Particular Cases of the Application of the Dispersion Method

The dispersion method is well adapted to the investigation of an equation
like (0.2.1) when the sequence {cp} consists of the values of a quadratic form
<p = cp(x, y) = ax2 + bxy + cy 2 ,
definite or indefinite, i.e. when the equation is of the form
n = cp(x,y) + D'v. (0.6.1)
20 INTRODUCTION

If cp(x, y) is indefinite, it is necessary to impose the natural limitations on the

absolute values of the variables x, y.
There is no great difficulty in finding the proposed asymptotic expression
A(n, D) and in approximating to the sums V2 and V3 , given in (0.2.9) and (0.2.10).
The fundamental equation (0.2.14) takes the form
v1cp(x, y) - v2cp(z, t) = n(v1 - v 2) (0.6.2)
with the supplementary conditions (0.2.15). Here x,y, z, t are independent
variables.
For given v1 , v2 (v1 =;if. v2) we have the problem of representing the number
n(v1 - v2) by the indefinite quaternary quadratic form v1cp(x, y) - v2cp(z, t), sub-
ject to the appropriate limitations. A characteristic feature here is the large size
of the coefficients v1 , v2 which causes difficulty in determining asymptotically the
number of solutions of (0.6.2).
A detailed treatment of the equation (0.6.2) on the basis of Kloosterman's
method [46], using the latest estimates for Kloosterman sums (see [57; 31; 19]),
leads to a satisfactory asymptotic formula for the number of solutions, provided
that
Yi < nO.Ol (i = 1, 2). (0.6.3)
If the form cp(x, y) has only one class in its genus, for example, if cp(x, y) =
x 2 + y 2 or x 2 - 2y 2 or xy, the equation (0.6.2) can be treated fairly simply, using
the elements of Vinogradov's method. In this case, using again the latest
estimates for Kloosterman sums, one finds that a satisfactory.· asymptotic
formula for the number of solutions of (0.6.2) can be obtained under the less
stringent condition
y . .;;;:: n 116- ' (i = 1 2) (0.6.4)
i....:::::::: ' '

where sis an arbitrarily small positive constant.

The largely arbitrary character of D' and'v in equation (0.2.l) allows us to
include under that equation the more difficult problem of
n = <p + p, (0.6.5)
where p is prime. We shall outline two ways of doing this.
The first of these follows directly from well-known arguments of Vinogradov
[7] concerning the estimation of trigonometric sums involving primes. Let N(-)
denote the number of solutions of the equation placed inside the parentheses.
Let p run through the primes in the interval [.Jn, n]; let u run through all
numbers which have no prime factor greater than .Jn;' then
N(n = <p + p) = ,L µ(u)N(n = <p + uv) - N(n = <p + 1). (0.6.6)
uv::::;;;n

Hereµ denotes the Mo bi us function, and v runs through all integers < n/u. If u
<
is not too large, say u n'X, where ot is a suitable positive constant, we can obtain
Sec. 6] APPLICATION OF THE DISPERSION METHOD 21

a satisfactory approximation to N(n = cp + uv) by considering the number of

solutions of the congruence <p = n (mod u) under appropriate limitations.
We divide the remaining u, which are in the interval (n(I., n], into two sets,
according as µ(u) = + 1 or -1. Consider one of these sets, and the correspond-
ing sum
± .2 N(n = <p + uv). (0.6.7)
uv::;;;;;n

All prime factors of u are < ,J-;;. Denote the smallest prime factor of u by v.
Then uv = D'v, where D' is the complementary factor to v, and (0.6.7) becomes
the sum
± .2 N(n = <p + D'v). (0.6.8)
D',v

The numbers D', v in this sum are not entirely independent, since D' < n/v
and D' must be of the form uv/v, where all prime factors of u are greater than v.
These difficulties, however, are easily eliminated, and the equation n = <p + D' v
can be reduced to a form which can be treated by the dispersion method.
In this treatment, however, the numbers v appearing in the fundamental
equation of the dispersion method, namely
(0.6.9)
may turn out to be so large as to make the determination of the asymptotic
number of solutions too difficult. In view of this, we shall examine the equation
n = <p +p (0.6.10)
and reduce it to an equation of the type (0.2.1) by another method.
LetP = P(n) be some suitably chosen function of n and let QP denote the set
of all integers whose prime factors are all greater than P. Put
'p(s) =TI (1 - P-T 1 (s > 1); 'p(s) = 1 + T(s).
'P>P

Here T(s) = ~m'enp (m')-•. Whens> s0 , we have jT(s)J < 1, and we have the
expansion
In 'p(s) = T(s) - (T(s))2 + (T(s))s + ... + (-1?-1 (T(s)t + .... (0.6.11)
2 3 k
Let
(0.6.12)

so that Tfc(m) is the n.s.e. x1x2 ... xfc = m, Xi E QP (i < k). Since

where A 1(n) = 1 when n =pm and 0 otherwise, it follows from (0.6.11) that,
form> P,
22 INTRODUCTION

A 1(m) = ri(m) - !r~(m) + !-r~(m) - ir~(m) + .... (0.6.13)

The series terminates with ((-1y-1/r)r;(m) for some sufficiently larger.
Let Q'(n) be the n.s.e.
(0.6.14)
where m >
2, and let Q(n) be the n.s.e. (0.6.10) under the condition P <p .;;;; n.
Then from (0.6.13) we find that
r1 ( ll-1
Q(n) =! - ! r~(n - <p) - Q'(n) (0.6.15)
k=l k <p<;;n
or
r1 ( - l)k-1
Q(n) =L N(n = <p + xix~ . .. x~) - Q'(n). (0.6.16)
k=l k
Here r 1 is an integer, obviously.;;;; ln n/ln P, and Xi E Op; and it suffices to take a
crude upper bound for Q'(n). Now equation (0.6.10) has been reduced to a
number of equations of the form
(0.6.17)
We need to find an asymptotic expression for the number of solutions of each of
them, with a sufficiently small error term.
We denote by Yfc the kth equation in (0.6.17). To reduce Yfc to the form
(0.2.1), we write xJ.x2 ... xfc = D'v, where vis the least prime factor of xJ. ... xfc.
Then the equation Yfc takes the form
n = <p + D'v,
where plainly
p < v.;;;; nllk. (0.6.18)
Here the variables D' and v will not be independent. We have D' .;;;; n/v, and
further all prime factors of D' are greater than v; also the numbers D' and v may
have repetitions. However, these are not vital objections, and the equation
permits the application of the dispersion method.
Of course the same ideas are equally applicable to the equation n = <p - p.
 CHAPTER I

LEMMAS ON THE NUMBER OF DIVISORS, ON PR™E

NUMBERS AND TRIGONOMETRIC SUMS

This chapter is of an auxiliary character relative to the dispersion method;

therefore, as was mentioned in the Preface, most of the lemmas are given here
without proof, but with references to works where the proofs can be found.

1. Lemmas on the Number of Divisors

By Tim) we denote the number of ways m can be written as a product of k
factors, counting order; in particular, T 2(m) = T(m) is the number of divisors of
m. In what follows, k is a constant and m is arbitrarily large.
LEMMA 1.1.1.
(1.1.1)
The proof is well known. 1
LEMMA 1.1.2. Let k and I be constants. Then
L (Tk(m))1 = Bn(ln n)k'- 1• (1.1.2)
m~n

For a proof see K. K. Mardfanisvili's paper [22]; particular cases of this

result are simply treated in another paper by the same author [21].
LEMMA 1.1.3. Let x be a large number, D ~ x 1-", where IX is a positive
constant, 0 < < IX !; (/, D) = 1. Then the following estimate holds:

! T(Dm+l)x~ln~IJ(1-.!). (1.1.3)
Dm+i~o: D D PID , P
A proof can be found in paper [6] of Vinogradov and Linnik.
LEMMA 1.1.4 (GENERALIZATION OF LEMMA 1.1.3). Under the same conditions
as before, if k is a constant, then
! (T(Dm + l))kx-x ( ln-x II ( 1 - 1))2k-l.
- (1.1.4)
Dm+i~o: D D PID P
1 Here and in what follows B is a bounded number, but not always the same one.
23
24 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

If k 1 and k 2 are constants, then under the same conditions

L (Tk 1(Dm + l))k• = B(k1, k 2) ~(In x)aCk,,ko>, (1.1.5)

Dm+z,;;;;., D
where a(k 1 , k 2) is a constant which depends only on k 1 and k 2 •
(1.1.4) is obtained by a trivial generalization of (1.1.3), as is indicated in [6].
(l.1.5) can be obtained from (1.1.4) by elementary arguments.
LEMMA 1.1.5. Under the same conditions, let x 1 < x; x - x 1 ;;;., x 1-cx1 2,
where ex is the constant in the hypotheses of Lemma 1.1.3. Then we have

L (Tk 1(Dm + l))k• = B(k1 , k 2) x - x 1 (1n x)a(k,,k2>. (1.1.6)

z1<;;Dm+!<;;z D
For the particular case k 1 = 2, k 2 = 1 we have the more precise estimate

x-x
L T(Dm + l) = B - -1 lnx. (1.1.7)
z1,;;;;Dm+i<;;x D
The proof is similar to that of (1.1.5).
Lemmas on numbers with small prime divisors play an important part in
what follows.
LEMMA 1.1.6. Denote by F(x, z, q) the number of numbers <x which are
relatively prime to q (q <
x) and have no prime divisors > z.
1°. Let xll• < z < x; then

F(x, z, q) = Bx rr(1 _l).

Plq p
(1.1.8)

2°. Let ln x < z < x I•; 1 ex = In z/ln x; then

F(x,z,q)<c1 xII(1-l)exp[-l(1n!+lnln!) +!+ 20 ],

Plq p ex ex ex ex ex In (l/cx)
(1.1.9)
where 101 <1.
3°. Let z < ln x; then
F(x, z, q) = B ln x exp 7T(z). (1.1.10)

For a proof, see paper [3] by Vinogradov.

LEMMA 1.1.7. For any sufficiently large number n we have:

! l = B 0 (ln n)- 0 . (1.1.11)

oln ~
o >exp (In In n) 2
Sec. I] LEMMAS ON THE NUMBER OF DIVISORS 25

PROOF. Let n0 In be the product of the distinct prime divisors of n. Any

divisoro Inhastheformo = 01m2, where 01 j n0 • Leto= o1m2 > exp(lnlnn) 2 = n1 •
Ifm 2 > ni12 , then obviously an estimate for~ (1/o), taken over such divisors ofn,
isBn;- 115 = Bc(lnn)- 0 . Ifwesupposethatm2 < ni'2 ,theno1 > ni' 2 • From this
we easily derive
(1.1.12)

Thus, replacing n1 by ni1 2 , one can suppose that n is square free. Let
n = n0 = Pib ... Pt; let qi be the ith prime (so that q1 = 2, q2 = 3, ...).
Put n2 = q1q2 ••• qt= 2 · 3 · 5 · 7 ... qt,;,;;; n0 • Clearly

(1.1.13)

Further, by the law of large numbers, In n2 = ~i~t In qi ,..._,qt, so that

qt < 2 In n2when n2 is sufficiently large. Divide the interval [nii 2 , n2 ] into
intervals of the form [N, 2N], where n2 ~ N ~ ni1 2 , and consider

(1.1.14)

The numbers Oare the prime divisors qi ,;,;;; 2 Inn = 2ni/lnlnn ,;,;;; 2Nfilnlnn. To
estimate the sum in(l.1.12), apply Lemma 1.1.6, puttingx = 2N1 , z = 2Nf_flnlnn.
Here £X = In z/In x < 3/ln Inn. By (1.1.9) we obtain the estimate
B exp -l(ln In n)(ln In Inn)= B(In n)-(I/ 5)lrilnlnn (1.1.15)
for (1.1.4). Summing over the intervals [N1 , 2N1 ] (obviously B In n is an estimate
for the number of these), we obtain (1.1.11).
LEMMA 1.1.8.
,L Tio) = B exp -(In In n) 2 , (1.1.16)
oln 0
o> exp (In lnn) 4

here k is a given constant.

PROOF. As in the proof of the previous lemma, putting n1 = exp (In In n) 4
and n0 equal to the product of the distinct prime divisors of n, and using Lemma
1.1.2, we reduce the problem to estimating

.L 'Tk(o) . (1.1.17)
oln 0 0
o>nl'"
26 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

Divide the interval [ni' 2, n0 ] into intervals of the form [N1 , 2Ni]. For one of
these intervals we have

112 112
= .!!...(
Ni
L
ms;;.2N1
('rk(m)) 2) ( L
olno
1) . (1.1.18)
6e[N"2N 1 ]

Applying Lemma 1.1.6 with x = 2N 1, z = 2ni/(lnlnn) 3 < 2Nr'(lnlnn) 3,

IX <3/(ln ln n) 3 we find that

L 1 = BN1 exp - .!. (ln In n)3 • (1.1.19)

oln0 2
6e[N1 ,2N1]

If we now apply Lemma 1.1.2 to (1.1.18), we arrive at (1.1.16).

LEMMA 1.1.9. Suppose that from the interval I= [I, n] an interval of length
mes (I') ;;.. n1- ' 0 , where s0 > 0 is sufficiently small, has been separated; we
consider the numbers q of this interval which have no prime divisors exceeding
Mn = exp (ln ln n)3 ' 2 •
The number of such numbers in I' does not exceed
B 0 mes (l')(ln n)-0 . (1.1.20)

PROOF. The numbers in question are of the form q1q2 , where .J~J Mn <
q1 < .J~, because all their prime factors are less than or equal to q1 . For given q1
the number of such numbers in I' does not exceed

B mes (J').
ql '
none of the prime factors of q 1 are greater than M,.. Applying Lemma 1.1.6, we
obtain (1.1.20).
We shall need one more elementary lemma on the number of divisors. Let
r~(m) = ~"'i···"k=m;zi>I 1, i.e., the number of ways that the number m can be
written as a product of k factors, none of which is equal to 1. Thus if m has
fewer thank prime factors, then r~(m) = 0.
LEMMA 1.1.10. Let m be square free, and suppose that it has at least k prime
divisors; suppose p -r m. Then

r~(pm) = k(r~(m) + r~_ 1 (m)). (1.1.21)

The lemma follows immediately from elementary combinatorial considera-
tions.
Sec. 3] LEMMAS PROVED BY THE SIEVE OF ERATOSTHENES 27

2. Lemmas on the Distribution of Numbers m with Certain

Restrictions on the Value of Q(m)
Let !1(m) be the number of prime divisors of m, counting multiplicity. This
section contains various lemmas which are due, for the most part, to Erdos and
Hooley; that they are useful for binary additive problems is made clear in a
paper of Hooley [43].
Let ~0 > 0 be a sufficiently small given constant; let y be a sufficiently large
number; t <fl< 1 < {3 < f; y 1 =exp (Iny)"•.
LEMMA 1.2.1.
(1.2.1)

and
.L
m.;,y
-m1 = B(ln y)YP, (1.2.2)
Cl(m);;.pinlny-1
where
Ye= c - c Inc (1.2.3)
and ~ 1 = ~ 1 (~ 0)---+ 0 as ~o---+ 0.
This lemma is a trivial modification of Lemma 7 of Hooley's paper [43, pp.
200-201] and is proved in the same fashion.
LEMMA 1.2.2.

m.;,y
L l=B_:!'._.
ln 2 y
(1.2.4)
Q(m)>10Inlny

This is one of Hooley's lemmas [43, p. 205, formula (45)].

3. Lemmas Proved by the Method of the Sieve of Eratosthenes

We shall need several lemmas which can be proved by a more or less straight·
forward application of the sieve of Eratosthenes. To obtain upper bounds (the
only kind of estimate we shall need) it is easiest to apply the sieve of Eratosthenes
in the form given it by A. Selberg [53] (see also a detailed exposition by Ozigova
[23]).
1.3.1 (BRUN AND TITCHMARSH). Let D < x 1-Eo; s0
LEMMA > 0; (/, D) = 1;
0 <I< D; 1T(X, /) = ~p==l(mod D);p.;,x 1; then
x
7r(X, D, l) = B - - - (1.3.1)
rp(D) In x
This lemma can be obtained by a straightforward application of the sieve
method. It appears for the first time in the literature in a paper of Titchmarsh
[55] and it is given in a more complete form (with numerical estimates for
B = B(c:0)) by Culanovskii [25].
28 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

LEMMA 1.3.2 (GENERALIZATION OF A THEOREM OF ERDOS). Let n be a

sufficiently large number; let d, r be integers such that (d, r) = I and dr ~ O.In;
let f3 > 0 (a constant). Then an estimate for the number of solutions of
n = dQ 1 + rQ 2 , (1.3.2)
where all prime divisors of Q1 and Q2 are greater than nP, is

B .!!_ ( 1 + 1 2
) (In In n) 2 (in In _!!_) • (1.3.3)
rd (In (n/rd)) 2 ({3 In n) 2 rd
This generalizes a theorem of Erdos [34]. For a proof, see paper [5] of
Vinogradov.
The next lemmas have to do with the quadratic form x 2 + y 2 • As is well
known,
I 1=4Ixip). (1.3.4)
x 2+u 2=m plm
Consider the equation
n = ~2 + 'Y/2 + Q, (1.3.5)
where all prime divisors of Qare greater than n/3 ({3 > 0 is a constant). To get an
upper bound for the number of its solutions, we eliminate from the numbers
n - ~ 2 - 'YJ 2 those which are divisible by primes p < nY, where y > 0, y ~ f3 is a
suitable constant. In this connection we use the following table of the numbers
of solutions of
~ 2 + 'YJ 2 = n (mod p•), (1.3.6)
where ~, 'YJ run through c-omplete residue systems (mod p); p is odd and p" II n,
p > 0:
For p >- s
p = 1(mod4) p = -1 (mod4)

Q(n, p = p•(s + 1 -
8) ;) Q(n, p8 ) = p•-l (s odd),

Q(n, p") = p• (seven).

For p'< s
p =1(mod4) p =-1(mod4)
1) s odd
p odd (p + l)pP(l - ; ) 0,

2) s odd
p even
(p + l)pP ( 1 - ;) pP(l + ;),
3) seven
p odd
(p + l)pP ( 1 - ;) 0,

4) seven
p even (p + l)pP(l - ; ) pP(l + ;).
Sec. 4] CERTAIN SUMS CONTAINING DIRICHLET CHARACTERS 29

The eliminating is accomplished by the standard sieve method (see [5; 6]).
In this connection it is necessary to take into account those values of p for which
(1.3.6) is not solvable. We cite the final result.
LEMMA 1.3.3. An estimate for the number of solutions of equation (1.3.5) is
B ~II (p :-- l)(p - xlP)). (1.3.7)
Inn i:>ln P - P - XiP)
In Chapter VI we shall need to estimate the number of solutions of
~2 + 'Y/2 + Q = n, (1.3.8)
where all prime divisors p I Q obey one of the conditions:
exp (In In n) < p <exp (In n)"
2 p > nP,0; (1.3.9)
where f3 > 0 is a constant; oc 0 > 0 is a small constant. To find an upper bound
for the n.s.e. (1.3.8) we eliminate from the numbers n - ~ 2 - ri 2 those divisible
by primes p <exp (In In n) 2 and also by primes p under the condition:
exp (In n)"o < p < nY; y > 0 is a constant. Using the above table and Lemma
l.1.7, after a standard application of the sieve method we arrive at the following
result:
LEMMA 1.3.4. An estimate for the number of solutions of equation (1.3.8)
under the given conditions is
B n OCo In Inn II (p - l)(p - xip))
2 . (1.3.10)
Inn Pin p - p + xlp)

4. Lemmas on Certain Sums Containing Dirichlet Characters

Let n be a large integer; n1 =exp (In n)" 1 • Consider a system of numbers
d, 11 , 12 such that
n 112 n~ 1 d- 1 < I; < n 112n 1d-1 (i = 1, 2); (11, 12) = 1, (1.4.1)
n11s < d < n112n11. (1.4.2)
Let c5 be any divisor of the number dl1l2 • For given c'>, consider the sum

! x!(d)xll1)~ll2), (1.4.3)
<p(dl1l2b-)
where the summation is over 11 , 12 , d under the above conditions.
LEMMA 1.4.1. An estimate for (1.4.3) is
Bb(ln n1) 6 • (I.4.4)
This lemma is a trivial modification of the estimate (54) in Hooley's paper [43].
30 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

5. Quasiprime Numbers
Numbers which have only a restricted number of prime factors are sometimes
called almost prime; such numbers will be encountered often in the following
chapters. Also, we shall often use the term "quasiprime numbers." Consider
the sequence of intervals [I, n] where n ~ oo. Let Mn be a monotone function
of n such that
Mn= O(n') (1.5.1)
for any e > 0. Then the quasiprimes of [I, n] with exclusion up to Mn are those
numbers whose prime divisors are all greater than Mn.
In the following chapters we shall use quasiprime numbers with exclusion up
to Mn= exp (In In n) 3 12 ; while in Chapter VII, when using ideas of Hooley
(see [43]) we take quasiprimes with Mn= nl/(lnln n>2.
Let Ln = LiMn) be the number of quasiprimes in [I, n]. A standard
application of the sieve of Eratosthenes gives
LEMMA 1.5.1.
LnX_n__ (1.5.2)
lnMn
Next, an immediate consequence of Lemma 1.1.6 is
LEMMA 1.5.2. For Mn = exp (In In n)312 ,
Ln =n .L µ(m) + B 0 n(ln n)- 0 , (1.5.3)
m..;;exp(lnlnn) 4 m
me An
where An is the set of numbers whose prime divisors are all less than Mn·
As Hooley observed [43], quasiprime numbers with exclusion up to Mn ~
n1/(ln In n>2 are "well-distributed" in arithmetic progressions having a very large
e
common difference D: D = O(n6 ) for any constant < 1. For prime numbers,
which, roughly speaking, can be considered as quasiprimes with Mn = ,,J-;;,, it is
not possible to derive anything of this kind for e ;;.i. t, even from the extended
Riemann hypothesis. We state Hooley's basic lemma [43] on this property of
quasiprime numbers.
Let x = nl/(ln In n) 2 '• p = TI p~x p . Let the positive integer t = npa (the
canonical factorization). Define a function/(m) = fn(m) as follows:
f(m) = g(m) + h(m),
where

{~
if mis a prime ~ x,
g(m) =
otherwise,
if(m,P) =I,
h(m) = G otherwise.
Sec. 6] LEMMAS ON THE DISTRIBUTION OF PRIMES 31

For given t let

t(l) = II pa; !(2) = II pa. (1.5.4)
:1>lt,p,.;;:z; :1>lt,p>:z;

Let y <n; D = O(n 6) (0 < () < 1; () is a constant), 0 < l < D; /Ul and /(2)
are defined with the aid of (1.5.4).
LEMMA 1.5.3.
- 1-L(n)y + B-n- if(lu>, D) = 1,
<:p(D) D Ins n
I f(m) = {
m=s;;;;y n
m==l(modD) B--- if(l(l), D) > 1,
D Ins n
where
L (n) = B (In In n)2
Inn
and depends only on n. We remark that by (1.5.3), L(n) X (In In n) 2/ln n.
The proof of this lemma is given in Hooley's paper [43, pp. 197-198].
Quasiprime numbers are very useful for the solution of the problem of Hardy
and Littlewood (see Chapter VII). They are also useful for the solution of binary
problems of the form (0.2.3) where the D' are prime numbers and the sequence of
11 is very sparse, for example, of equations of the form n = i:p(x, y) + Pi2m, where
<
p 1 n1-"', 2m < n"'; ex: < 0 (cf. the author's paper [11]). Here in forming the
dispersion V' of type (0.2.5) it is expedient to apply, in place of the crude in-
equality (0.2.6), one in which D runs through only a "quasiprime shell D'"
(quasiprime numbers with exclusion, e.g., up to M,. = n1 /(lnin n> 2) instead of
through all numbers of the interval [D1 , D 1 + D 2 ]. Then we need to solve the
basic equations (0.2.13) for quasiprime D. This can be accomplished by putting
such conditions on n - <:p1 and n - <:p 2 as are necessary for carrying out an ele-
mentary sieve operation consistent with the determination of an asymptotic
approximation to (0.2.14). However, the computation becomes very cumber-
some.

6. Lemmas on the Distribution of Primes and of Certain Products

in Progressions
First of all we shall need theorems on the distribution of prime numbers in a
segment of an arithmetic progression qm + l; (/, q) = 1; 0 < l < q; qm + l x <
for large x and q increasing with x. Consider the zeros of all L series L(s, x) for
the modulus q; x is a Dirichlet character (mod q). The character xis said to be
exceptional if the series L(s, x) has a zero in the interval (1 - c/In q, 1) (the
exceptional zero being {3), where c > 0 is some constant less than ih· It is known
32 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

that only real characters can be exceptional, and that there can only be one
exceptional zero f3 for given q.
Let 7T(x, q, I) be the number of primes in the indicated segment of the
progression; Ea= 0 if q does not have an exceptional character; Ea= I if it
does; in this case let i be the exceptional character and f3 the exceptional zero.
Then the following lemma holds:
LEMMA 1.6.1.

7T(X, q, l) = -1 f"'-dt + Ea L-(l) f"' -xP-1 dx + Bx exp -(a .Jln x). 1 (1.6.1)
cp(q) 2ln t cp(q) In x 2

The constant a1 can be effectively computed.

This lemma is a rephrasing of Theorem 50 of Cudakov's book [26, pp.
167-168].
LEMMA 1.6.2 (SIEGEL). In formula (1.6.1) we have
1 - (3 > c(e)q-•, (1.6.2)
where e > 0 is an arbitrary constant, while c( e) depends on e. As is well known,
for given e the constant c(e) cannot be computed at the present time; it is not
effective.
The proof of this well-known theorem is given, for example, in Cudakov's
book. [26, pp. 145-166].
An immediate consequence of Lemmas 1.6.2 and 1.6. I is
LEMMA 1.6.3. Let C > 0 be any constant; q < (In x)°; then

7T(X, q, l) = - 1 f"' -dt (1 + B0(ln x)0 ). (1.6.3)

cp(q) 2 In t
Unfortunately, the constant B 0 here is not effective. In Chapter VII, for
large x (the number n takes the place of x there) one has to consider segments of
progressions qm + I<
x with q <exp (ln In x) 4 •
Let us consider such moduli q. Those for which a formula like (1.6.3) holds
will be called non exceptional; the others, exceptional.
LEMMA 1.6.4. Exceptional moduli are divisible by numbers r 1 , r2 , • •• , r1
which appear among the numbers q less often than the geometric progression
w0 , w5, wg, ... , where w0 > I is any constant. Thus
rb > wg for b > K(roo); (1.6.4)
furthermore,
r1 > (ln x)° for x > x 0(C). (1.6.5)
Inequality (1.6.5) follows immediately from Lemma 1.6.3. Let Xrk and Xrk+l
be two exceptional primitive characters,
(ln x)° < rk < rk+ 1 <exp (ln In x)4 ; (1.6.6)
Sec. 6) LEMMAS ON THE DISTRIBUTION OF PRIMES 33

let {J(rk) and {J(rk+i) be the corresponding exceptional zeros. By Theorem 43 of

[26, p. 125) we have
(1.6.7)
where {J = min ({J(rk), f3(rk+l)).
If rk+ 1 < w 0rk, then at least one of the two inequalities
1 -
/3·Tk > 2-1
a2 . 1
'
/3
- Tk+l
> -21
a2 (1.6.8)
n rk n rk
is satisfied, and for at least one of the moduli rk, rk+l (1.6.3) holds, contradicting
our assumption. This proves the lemma.
Now we come to a lemma which plays a fundamental part in deriving the
asymptotic expression in the problem of Hardy and Littlewood (Chapter VII).
This lemma concerns the distribution of numbers of the form x 1x 2 ••• xk for
k < 6 in segments of arithmetic progressions having a common difference which
is in some sense close to the square root of the length of the segment. Lemmas of
this sort are closely related to estimates of sums of the form l:xmodq IL(s, x)I\
where the summation is over all characters mod q. In paper [15) the author
gives theorems of this form for k < 4 (which establish the solvability of the
Hardy-Littlewood equation n = p + g2 + rJ 2). In paper [16) "total" distribu-
tion laws are derived; we shall state them here.
Let x > x 0 ; consider a collection of different moduli D such that
(1.6.9)
where
(1.6.10)

Here g0 > 0 is an arbitrarily small given number. With respect to each of

these moduli D we are given a residue In for which (ID, D) = I, 0 <In< D,
and a number xn such that
x
- - b <xn<x, (1.6.11)
(Jn x) 0
where b0 > I is an arbitrary constant. Consider the congruence
x 1x 2 ••• xk =In (mod D); x 1x 2 ••• xk < xn, (1.6.12)
where k <
6; let M 0(xn, D) be (for given k) the number of solutions of (1.6.12).
Then we have
LEMMA < 6,
1.6.5. For k

! M(xn, D) = ~ ! -1- [ xi:, (L(s, x<f/)? ds

<Dl 2m <n> cp(D) ~cp 0 s
D
+ Bx ___g exp -(Jn In x) 50, (1.6.13)
D1
where cp 0
is the circle Is - 11 < I/In Di.
34 DIVISORS, PRIME NUMBERS AND TRIGONOMETRIC SUMS [CHAP. I

The proof of this lemma is rather cumbersome. A detailed exposition is given

in paper [16]. Only the cases k ..;;; 5 are used in Chapter VII.

7. Lemmas on Trigonometric Sums

In order to estimate the number of fractional parts of certain functions, vary-
ing in definite intervals, we shall apply a method of Vinogradov.
Let u(x) be a continuous function of an integral argument x, which varies in
an interval I. We are interested in the number Q(µ) of fractional parts {u(x)}
which fall into a prescribed intervalµ = [ix, ,8], 0 < ix < ,8 < I. In many cases
this number can be very precisely approximated by Vinogradov's upper and
lower functions, 'P(z) and °!(z), using the formula
L J!'(u(x)) ..;;; Q(µ) ..;;; L 'Y(u(x)). (1.7.1)
xel xel
The functions 'r(z) and 'P(z) can be chosen in various ways. In [7], Vino-
gradov gives a very important form of these functions: in Fourier series with
period I. By expressing 'P(z) and 'r(z) in this form, the problem can be reduced
to trigonometric sums.
LEMMA 1.7.1 (VINOGRADOV). Let r > 0 be an integer; 0 < l:!.. < 0.5;
l:!.. ..;;; ,8 - ix ..;;; 1 - l:!... Then there exists a periodic function '¥(x) with period 1
such that:
I. '¥(x) = 1 for x E [ix + 0.5l:!.., ,8 - 0.5l:!..],
2. 0..;;; '¥(x)..;;; lfor x E [ix - 0.5l:!.., ix+ 0.5l:!..] andx E [,8 - 0.5l:!.., ,8 + 0.5l:!..],
3. '¥(x) = 0 on [O, l] outside of [ix - 0.5l:!.., ,8 + 0.5l:!..],
4. 'J!(x) has the Fourier expansion:
00

'¥(x) = ,8 - ix + .l' am exp 27Timx

m=-CO

(the accent denotes the exclusion of the value m = 0). Here

laml < 7Tm

-4 ; laml < 4(,8 - ix); laml < 7Tm
-4 ( --r )r· (1.7.2)
7Tml:!..
For a proof see Vinogradov's book [7, pp. 260-261].
LEMMA I. 7.2 (VINOGRADOV). Let f (x) be an integer-valuedfunction, q > 1 an
integer, Ta subset of the set of numbers 0, 1, ... , q - I. Then
27Ti 1 27Tist
q-l
L
teT
exp -f(t) = - L L F. exp -
q q s=O t e T q
, (1.7.3)
where
q-l 27Ti
F. = L exp - (f(x) - sx). (1.7.4)
o:=O q
Sec. 7] LEMMAS ON TRIGONOMETRIC SUMS 35

This lemma is proved by interchanging the summations over s and t and then
directly calculating the right-hand side. If T is the interval [q1 , q2], where
[q1, q2] c [O, q], then it follows immediately from (1.7.3) and (1.7.4) that

I L
t e[q,, qz]
exp 27Ti
-f(t)
q
I= B IF/ In q, (1.7.5)
where
IF/ = max IF.I. (1.7.6)
s
Of the estimates for trigonometric sums we shall need most of all Andre
Weil's very recent estimates for Kloosterman sums.
LEMMA 1.7.3. Let q > 1, xx' = 1 (mod q) for (x, q) = 1; let a and b be
integers. Then

L exp 27Ti (ax + bx') = Bq112T(q) min ((a, q)112, (b, q)112). (1.7.7)
(:c,q)=l q
For a proof see A. V. Malysev's paper [19].
The combination of (1.7.7) and (1.7.5) leads immediately to the lemma:
LEMMA 1.7.4. In the notation of Lemma 1.7.3,for 0 < q1 ~ q we have
L exp 27Tibx' = Bq112T(q) In q(b, q)112. (1.7.8)
z.;;;q, q
 CHAPTER II

LEMMAS FROM THE ANALYTIC THEORY

OF QUADRATIC FORMS

1. Fundamental Equations of the Dispersion Method for the Divisor Problem

In the divisor problem (see Chapter III) the sequence {<p} consists of the
numbers xy ~ n, x ~ 1, y ~ 1. We take as the set of coherent numbers the
number 1 and all primes for which
(2.1.1)
The numbf(rs v1 , v2 are all primes such that
vi ~ n116-• 1 (i = 1, 2). (2.1.2)
When the problem is formulated in this way, the fundamental equations of the
dispersion method take the form
(2.1.3)
under the additional condition
(2.1.4)

(2.1.5)
where
(2.1.6)

(2.1.7)
where
(2.1.8)

Here v1 =;t. v2 ; n1 is any one of the numbers (2.1.1).

Since the numbers vi are primes, congruential conditions do not appear (cf.
§ 4 of the Introduction). We shall assume that the numbers vi lie in the interval

[ v0 , v0 + Vo K
(In n) 1
J = (v 0 ). (2.1.9)
36
 LEMMA ON THE FUNDAMENTAL EQUATIONS 37

Here
(2.1.10)
We shall fix the constant K 1 later. We assume that the interval (D) is of the
form
(2.1.11)
where
K < K1 . (2.1.12)
0 1000
The conditions xy ~ n and zt ~ n are not taken into account here.

2. Lemma on t~e Fundamental Equations

V1 denotes n.s.e. (2.1.3) under conditions (2.1.4); V3 is n.s.e. (2.1.5) under
conditions (2.1.6); V2 is n.s.e. (2.1. 7) under conditions (2.1.8).
LEMMA 2.2.1.
(2.2.1)
where
Vo
'JI~=---=-- (2.2.2)
(In n)K1
and K2 is suitably chosen according to K1 ; K1 and K2 can be arbitrarily large.
We proceed with the proof of the lemma; consider equations (2.1.3), (2.1.5)
and (2.1.7), putting
(2.2.3)
where (n1' n2) can be one of the pairs of numbers (1, I); (n 1 , n1); (n 1 , 1). The pair
(v1 , v2) is assumed fixed.
The additional conditions have the form

(2.2.4)

Note that because of the above conditions, M ~ 0. We introduce one

additional condition: x ~ y, z ~ t. The number of solutions where x = y,
z = tis small; a trivial computation shows that it does not exceed
(2.2.5)
Taking this into account, we shall consider only those solutions for which
x < y, z < t. (2.2.6)
The number of solutions of (2.2.3) under the additional conditions (2.2.4)
will be 4 times smaller than the number of solutions of the original equation
38 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

(and of course, with conditions (2.2.6), to within an error of (2.2.5)). From now
on we shall assume conditions (2.2.6).
In what follows the value of
b = g.c.d. (x, z) (2.2.7)
will be important.
Clearly b JM. First we suppose b = 1. Then v1 ':;!= z. For if v1 J z, then
111 J v2n2 ; by the conditions on 111 and v2, 111 J n2, which is impossible in view of
(2.1.2) and (2.1.1). Since v1 is prime, (v1 , z) = 1. Now from (2.2.3) we find
y = Mv~x'(mod v2z), (2.2.8)
where v:tx' is the inverse of v1x mod v2z.
Further, from (2.2.4), (xy - n1)/v2 E [D1, D 1 + D 2 ];

or
(2.2.9)

where we have put n12 = n1 + v2 D1 •

Let us see what additional conditions are necessary to make congruence
(2.2.8) equivalent to the original equation (2.2.3) with the conditions on it. By
(2.1.8), (2.1.11), (2.1.12), and (2.1.1),
xy ~ 2n; zt ~ 2n
follows from the fundamental equation. Combining this with (2.2.6) we have
0 < x < J1n; 0 < z < J1n. (2.2.10)
Let x and z be a pair of relatively prime numbers satisfying conditions
(2.2.10). From (2.2.8) we find
v1(xy - n 1) = v2(zt - n 2), (2.2.11)
where tis an integer. From (2.2.9) it follows that zt - n2 > 0, and since z > 0,
t > 0. Further, t > z, or, from (2.2.11),

t = v1(xy - n1) + v2n2 > z,

'Jl2Z

(2.2.12)

To this we must add the condition

y> x. (2.2.13)
Sec. 2] LEMMA ON THE FUNDAMENTAL EQUATIONS 39

Thus congruence (2.2.8) with conditions (2.2.9), (2.2.12) and (2.2.13) is

equivalent to the original equation. However, conditions (2.2.12) and (2.2.13)
are in an inconvenient form for computation. We shall replace them by other
conditions at the expense of a certain permissible error in the asymptotic
approximation for the number of solutions. Instead of (2.2.12), consider the
equivalent condition
xy - n
V1 1 + n2 > z.2 (2.2.14)
'llz

By (2.2.9) we have: (xy - n1)/v2 E [DI> D1 + D 2 ]. Replace (2.2.12) by

D 1 v1 > z 2, z < .J D 1v1 • (2.2.15)
lf(2.2.14) does not imply that conditions (2.2.15) are satisfied, then
v1(D1 + D 2 ) + n2 > z 2 ; v1 D1 <z 2, (2.2.16)
whence
(v1 D 1) 112 < z < (v (D + D + n
1 1 2) 2) 112 • (2.2.17)
Hence, considering (2.1.1 ),

where
(2.2.18)
by (2.1.9) and (2.1.12).
Let us estimate the number of solutions of (2.2.3) under the conditions:
xy < <
2n, zt 2n; z is given by formula (2.2.17); g.c.d. (x, z) = <5 = 1. In this
case, if lz is suitably chosen,

xy = l.(mod v2z); (2.2.19)

Here
v 2z = BnI/6+112 = Bn213. (2.2.20)
In view of this, by Lemma 1.1.5 (cf. (1.1.7)) we have: an estimate for the number
of pairs (x, y) subject to condition (2.2.17) for given z is

B Dv 2 In n = B D 2 In n . (2.2.21)
z
Summing this estimate over values of z satisfying conditions (2.2.17) and (2.2.18)
we obtain the expression
(2.2.22)
Thus with accuracy to within such an error we can replace condition (2.2.12) by
the simpler condition (2.2.15).
40 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

We turn to condition (2.2.13). It can be written in the form

(2.2.23)

We introduce the simpler condition:

(2.2.24)

If (2.2.23) does not imply this condition, then obviously

(v2 D1) 112 ~ x < (n 1 + v2(D1 + D2))112• (2.2.25)

This condition is completely analogous to (2.2.17); arguing with respect to the
number of values of x under such a condition and the corresponding number of
solutions of equation (2.2.3) as we did with respect to the variable z after formula
(2.2.17), we come to the conclusion that the replacement of (2.2.13) by the simpler
condition (2.2.25) creates an error of
(2.2.26)
Thus when (x, z) = 1, taking the above errors into account, we may study
congruence (2.2.8) under the conditions

[ n12
YE -, n12 + v2D2]
x x (2.2.27)
z < (D 1v1) 112 , x < (D 1v2) 112 , g.c.d. (x, z) = 1
instead of the original equation.
In what follows we shall refer to (2.2.8) as the fundamental congruence.

3. Analysis of the Fundamental Congruence

Now let us examine the fundamental congruence (2.2.8) under conditions
(2.2.27). We see that y runs through integer values in the interval (2.2.9), which
is oflength v2 D 2 /x. If this length is significantly greater than v2 z, then it is easy to
determine an asymptotic approximation for n.s.e. (2.2.8) for given x, z. Suppose
that
(2.3.1)

For the number of solutions of (2.2.8) under conditions (2.2.27) we have the
obvious expression
D2 + B, (2.3.2)
xz
Sec. 3] ANALYSIS OF THE FUNDAMENTAL CONGRUENCE 41

which is nontrivial under conditions (2.3.1). Under the conditions

z < (D1111)112, x < (D1P2)112,
(2.3.3)
g.c.d. (x, z) = 1, xz < D n-• 2 3,

the set of which we denote by Uc 2 . 3. 3» we sum (2.3.2) over all permissible x and z,
and obtain the expression
D2 _L _!_ + BD n-••, 2 (2.3.4)
U12.a.a1 XZ
where e4 = !e3 •
Now consider the case when xz < D 2n-• 3 is violated, so that the following
conditions hold:
z < (D1111)1'2, x < (D1P2)112,
(2.3.5)
g.c.d. (x, z) = 1, xz D 2n-•3• >
From this we have x > D2n-• 3 /z >
tD2n-112-• 3 provided n is large enough
(which we shall assume).
Turning to (2.1.12) and (2.1.10) we see that D1 > !n 51s+•i1 2, so that x >
!nsts-112+•1-•3 (ln n)-K•. If we suppose that e3 has been chosen so that

(2.3.6)
then we have
x > !n1/s+•,. (2.3.7)
For each given x under condition (2.3.5) and given z, y must vary in the
interval (2.2.27). Here x is in the interval

[D2;-•3, (D1112)112J. (2.3.8)

If it happens that the left number in brackets is larger than the right one, the
interval is considered empty.
Let x 0 be any value of x in the interval (2.3.8); consider the interval
(2.3.9)
where e5 will be determined later on. For each x of (2.3.8), the interval of varia-
tion of y, given by (2.2.27), will not differ from the interval

[ n12 , n12 + P2D2] (2.3.10)

Xo Xo

by more than two intervals of total length

n -•s
B -Xo = B -Xo
n -•s • (2.3.11)
Xo X
42 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

We shall show that the number of solutions for which y is in the indicated
intervals Gf length (2.3.11) is negligible. In equation (2.2.3), let x be in (2.3.9);
and suppose that y, for given x, is in the indicated intervals. For prescribed x, y
belongs to a definite residue class mod 112 (it is easily seen that (x,11 2) = 1), soy
receives
(2.3.12)

values. Then the pair (x, y) has

n -2•s
B -Xo (2.3.13)
112

values. To each such value corresponds B,n' values of zt. Choosing e = le5 , we
obtain
B ~ x-ca12l•s
0
_ BD x-o.4•sx•s
- 2 0 0 (2.3.14)
112

solutions. As we shall see later, such an error can be neglected, and for x in
(2.3.9) we can suppose that y is in the interval (2.3.10), which we denote by lv,xo.
Turning to the fundamental congruence (2.2.8), denote by Ev,xo that part of lv,xo
into which the length 112x fits the maximal integral number ex 0 ~ 0 of times (if
there is such a part); call the remaining interval Fv,xo' so that
(2.3.15)
Put
mes Fy xo
--~·-" =µ (2.3.16)
1l2Z
(whereµ depends on x and z). If
µ < - n- n-•• , (2.3.17)
1l2ZX0

where c; 6 > 0 is some constant (which we shall fix later on), then solutions for
which y is in FY,xo and x is in the interval (2.3.9) can be neglected. Indeed, an
estimate for the number of pairs (x, y) corresponding to fixed z is

B -n x-•s
0 n
-•• ,
1l2Z

and the number of corresponding pairs for permissible z is

B !!.. x 0' 5 n-•• In n.

112

For a given pair x,y, the pair z, t cannot have more than B,n• values; putting
e = -le6 , we find the total estimate

(2.3.18)
Sec. 4] APPLICATION OF THE METHOD OF VINOGRADOV 43

which, as we shall see, is permissible. Thus we shall suppose that

(2.3.19)

4. Application of the Method of Vinogradov

For each possible value of x in its interval (2.3.9) y takes on exactly e,,0
values of Ev,,
• 0
; thus it will not be difficult to determine the number of solutions
with such values. On the other hand, we shall apply Vinogradov's method (see
§ 7, Chapter I) to estimate the solutions wheny E Fv.zo· Wheny E Fv,zo we must
have

o~ {L} ~µ.
V2Z
(2.4.1)

To compute the number Q(z) of values of y under condition (2.4.1), as x runs

through (2.3.9), form the upper and lower functions of Vinogradov, 'P(u) and
'f(u) (cf.§ 7, Chapter I), choosing ix = 0, {J = µ, !::i' = µn-•&, r = 100. We have

L':l"(f(x)) ~ Q(z) ~ L"if·(f(x)), (2.4.2)

(z) V2Z (z) V2Z

where/(x) = Mv]_x' (mod v2z) and (x) denotes the interval (2.3.9). Consider the
behavior of 'Y(u) (th~ upper function 'P(u) is treated analogously). We have
(cf. Lemma 1.7.1)
Mµ
!'(k) = µ + ! L (am + ibm) exp 21Timk
m=l
Mµ
+ i L (am - ibm) exp (-27Timk) + Bµn-• 6• (2.4.3)
m=l
Here
(2.4.4)
and when m ~ Mw
(2.4.5)
Finding an asymptotic approximation for "2:.'Y(f (x)/v2z) reduces to estimating
the sums: -

Here x runs through the incomplete residue system (2.3.9), (x, v2z) = 1. By
(1.7.8) (Lemma 1.7.4) we have
(2.4.6)
44 ANALYTIC THEORY OF QUADRATIC FORMS (CHAP. II

We have M = Y1n1 - Y2n2. By definition of Yi and ni, g.c.d. (M, Y1) = (M, Y2) =
(Y1 n1 , Y2) = 1. Furthermore, as is easily seen,
(Mm, Y2Z) ~ (M, Y2z)(m, Y2Z) = (M, z)(m, Y2Z). (2.4.7)
In view of this,
L (Mm, Y2z)112 ~ (M, z)112 L (m, Y2Z)112_ (2.4.8)
m~Mµ m~M 1,

Then

L
m,,;;Mµ
(m, Y2z)112 = B L
61-.z
15 112 L
m==O(mod 6)
1 =BMµ L 15112 =
61-.z
BMµr(Y 2 z). (2.4.9)
m:::;;;;,Mµ

Using (2.4.6), (2.4.8), (2.4.9) and (2.4.5) in (2.4.3), we find that for given z,

,L ~
(x)
(!(x)) (.L 1)µ(1 + Bn-••) + R.,
Y2Z
=
(x)
(2.4.10)
where
Rz = B(Y 2z)112(M, z)11 2(r(Y2z)) 2 ln nn 2•• = B(Y 2z)112(M, z)112 n 3" 6 (2.4.11)
(cf. Lemma 1.1.1).
There is a similar relation for the upper function 'P(u). Using (2.4.2) we find
that

Q(z) = µ(1 + Bn-••) (.L 1) + Rz.

(x)
(2.4.12)
By (2.3.19),
µ n
>--n- ••.
Y2ZX0

x E (x) runs through values which are relatively prime to z; therefore,

1-es
_Ll>~. (2.4.13)
<x> Inn
and thus
µ L 1 > !:._ xi)•sn-2••. (2.4.14)
(x) Y2Z

But by (2.3.3), z = B.J-;;, and by (2.1.2), Y 2 < n112-t1. Hence

µ L 1 > C2n11a+•1-••x~·s, (2.4.15)
(x)

(y 2z)112 = Bn 113-• 112• (2.4.16)

Note that it is just here that we see the reason for the constant tin formula
(2.1.2). If values Yi > n 1 16 were allowed, (2.4.16) could be larger than (2.4.15),
i.e., the error term in formula (2.4.12) could be larger than the main term, and we
would not be able to carry out the asymptotic calculation.
Sec. 4] APPLICATION OF THE METHOD OF VINOGRADOV 45

We shall suppose that e5 + es < e1 /10; then, comparing (2.4.15) and (2.4.16)
we obtain

(2.4.17)
Taking this into account in (2.4.12) and observing that es< e1/10, we find that
Q(z) = µ(1 + Bn-••(M, z)112) ,L 1, (2.4.18)
x eiz
(x,z)=l

where I., is the interval in which x varies (2.3.9) (previously x E (x) meant that
x EI., and (x, z) = 1).
The quantity Q(z) corresponded to fixed z, x EI.,, (x, z) = I,y E Fy,., 0 If
under the same conditions on x and z, y E EY,"'o' then clearly to each value x EI.,
correspond exactly e., 0 values of y, so that for y E IY,"'o for all permissible values
x EI., we find that the number of pairs (x, y) is

e., 0 ,L 1 + Q(z). (2.4.19)

xe I~
(x,z)=l

Note that e.,0 + µ = mes ly ., /v2z or, by (2.3.10),

' 0

D2
e.,. +µ = - . (2.4.20)
XoZ

Thus the number of pairs (x, y) of the indicated type is expressed by the
formula
(2.4.21)

We separate the numbers z satisfying condition (2.3.3) into two classes, A1

and A2 • Class A 1 consists of the numbers z for which
(M, z) < n•oflO (2.4.22)
and class A2 of the remaining z, for which
(M, z) > n••/10. (2.4.23)
We shall show that there are comparatively few solutions of the fundamental
equation (2.2.3) where z E A2• To do this it is easiest to estimate the "combined"
number of solutions of (2.2.3), when v1 -:;&:. v2 and the pairs v1 , v2 run through the
full set of values determined for them by formula (2.1.9). Let (M, z) =a> e6 /10.
To estimate the combined number of solutions under these conditions, it is
sufficient to estimate the n.s.e.
M
X- Y=-,
a
46 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

where X = 111xy/a. ~ 2v0n/a; Y = 112zt/a ~ 2v0 n/a, and X and Y each run
through B.n• values. Clearly

B, Von n2• ! ! = Bvonn-•sllO = BD211~2n-•s/21 (2.4.24)

<1 6/Me <1
u>ne6/to

is such an estimate if we choose 2e smaller than e6 /10. Such an error, as we shall

see, is negligible. Thus, we can restrict ourselves to z E A 1 and assume that
(M, z) < n•s/ 10 • For such values of z, considering thatµ ~ min (1, BD 2 /zx0) =
µ(z, x 0) and taking into account (2.4.21), we obtain the asymptotic expression

! 1) (D 2 + Bµ(z, x 0)n-•1) ,
( a:eiz e7 = 0.9e6 (2.4.25)
X 0Z
(a:,•)=1

for the corresponding number of solutions.

5. An Asymptotic Formula for the Number of Solutions when (x, z) = 1

We can now give a combined asymptotic formula for the number of solutions
of the fundamental equation (2.2.3) where all permissible values of (111 , 112) are
taken and (x, z) = 1. First we collect all the main terms we have obtained, and
then all the error terms. For each permissible pair (111 , 112) the equation is solved
under conditions (2.2.27); then the numbers of solutions are summed over
(111 , 112). This is done with z fixed, under the conditions applicable. Appealing to
conditions (2.3.1), i.e., assuming that x < D 2 n-•a/z; we have the quantity

D2 ! _!_ + BD n-••
2 12 (2.5.1)
u, .. a.si xz
from (2.3.2).
Next, when x > D 2n-•a/z, we divide up the interval of variation of x,
obtained from (2.3.9), in the following manner:
Let x 0 denote its left end point; consider the interval I., = [x0 , x 0 + x5-•s].
If it fits entirely into the interval of variation of x, then, putting x 0 = x0 + x5-•s,
we consider the following interval I., = [x 0, x 0 + (x0)1-•s], and so on. If an
interval is obtained whose right end falls outside of the prescribed region for x,
then, as we shall show later, the number of solutions corresponding to it can be
neglected. For those intervals I., which fit entirely into the interval of variation
of x, we have for given 111 , 112 , by (2.4.25),

! 1) (D 2 + Bµ(z, x 0 )n-•
( a:eiz 1) (2.5.2)
x 0z
(a:,z)=l

solutions. Then expressions (2.5.1) and (2.5.2) are summed over 111 , 112 •
Sec. 5] AN ASYMPTOTIC FORMULA 47

Now let us consider the errors in the combined formula. We assume that
K1 > 100 K0 •The error (2.2.5) gives
(2.5.3)
in the combined formula. The error (2.2.5), after summing over v1 and v2 , gives
(2.5.4)
In view of the fact that x 0 > tn1 13 (cf. (2.3.7)), the error (2.3.13), after summing
over v1 and v2 , gives
(2.5.5)
solutions. (2.3.18) introduces a similar error. The error (2.4.24) is already given
for the total number of solutions. It has the form
(2.5.6)
Consider also the behavior of the error terms (2.5.1) and (2.5.2) in the com-
bined formula. Summing the error in (2.5.1) over v1 and v2 we find an error of
(2.5.7)
Finally, consider the error term in (2.5.2) (which is valid only for z E A1).
For given x 0 , sum the error first over z E A 1 ; zx0 < D 2 ; v1 , v2 ; this gives an
error of
Bx~-·· D2 n-•1v~2.
Xo

Now we take in turn the intervals (2.3.9) of variation of x, defined by x 0 •

There are Bx~/ such intervals in [x0 , 2x0 ] and they give a total error of

For each value of z in the corresponding interval of variation of x there are

B Inn abutting intervals of the form [x0 , 2x0 ]. Thus for z
E A1 the total error in
the number of solutions is
(2.5.8)
Now consider z E A1 ; zx0 ~ D 2 • Here for given z the error takes the form
1-•sD 2 1
B Xo
n-• 7 -_ BD 2x-•s
0 n
-•1 _ .
x 0z Z

Summing over x 0 , vl> v2 , and z we again obtain the result (2.5.8). Note that
if an interval of the form [x0 , x 0 + x/j-•s] does not fit into I., then the number of
solutions corresponding to it can be neglected. We shall obtain a rough com-
bined estimate for the number of such solutions of (2.2.3): considering that for
48 ANALYTIC THEORY OF QUADRATIC FORMS (CHAP. II

each value of x in the indicated interval, y <

2n/x0 , and that for each x, y the
pair (z, t) takes on B,n• values, the combined number of solutions is
(2.5.9)
if 2e < ie5 •
For given z < (D v 1 1) 112, we denote by U< 2.s.io> the set of conditions
x < (D1v2) 112 ; (x, z) = 1; xz < D2n-• 3 (2.5.10)
and by U< 2.s.n> the set of conditions
x < (D1 v2)1 12 ; (x, z) = 1; xz ~ D 2 n-••. (2.5.11)
We also introduce the conditions U< 2.5 .12>
a= (M, z 1) < n'•; z < (D1v1) 112. (2.5.12)
Then the total number of solutions of (2.2.3) takes the form

4D2 L Lz ( Lz _!._ + L L -1-) + BD 2v~2(ln n)-K 4• (2.5.13)

•1.V2 XZ Ix xelx XoZ
Uc2.s .12i Uc 2.s.10> Uc2.s.11>

6. The Case (x, z) > 1

We now turn to the case (x, z) > 1 in equation (2.2.3).
From the very beginning we shall consider only those solutions which satisfy
condition (2.2.6). Note that always c'J IM.
Put x = x 1 c'J; z = z1 c'J; M = M 1 c'J;
then we have
(2.6.1)

Here XiY < 2n/c'J; z 1 t < 2n/c'J; (xl> z1) = 1, x 1 < y; z1 < t.
We first examine the case
(2.6.2)
For this case it is sufficient to give a rough estimate of the total number of
solutions of (2.6.1). Note that (v1, v2 , M 1) = l. For given v1 , v2 we replace
(2.6.l) by
(2.6.3)

where X <2n/c'J; Y <

2n/c'J the numbers X and Y run through T(X) and T(Y)
values, respectively. N.s.e. (2.6.3), as is easily seen, does not exceed
BL T(x' + M')T(x'),
(o:')

where the summation is over an interval of length BD2 /c'J.

Sec. 6] THE CASE (x, z) >1 49

By Lemma 1.1.4 (cf. Lemma 1.1.5) and the inequality labl ~ !(a2 + b2), an
estimate for the last expression is

B D2 (In n)a<2,2>.
b
Summing over all values v1 ¥:- v2 such that v1 n1 - v2n2 = 0 (mod b) we obtain
the combined estimate
D 12
B 2:0 (In n )a<2,2>. (2.6.4)
b
Assuming
K5 > 10a(3.2)
and summing over b >(In n)K•, we obtain the combined estimate

(2.6.5)
in the total number of solutions, which will be a permissible error. Now we
shall suppose that
(2.6.6)
and analyze equation (2.6.1) under these conditions.
Here (xl> z1) = 1; therefore there is nothing essentially djfferent in the
treatment of this equation by the methods of the preceding paragraphs. Now
n/b takes the place of the parameter n, and M 1 the place of M. Condition (2.2.4)
becomes

so that for given x 1

(2.6.7)

Thus D2 /b takes the place of the number D 2 ; D1 /b, that of D 1 ; nJ./b takes the
place of n1 • As is easily seen, the errors (2.2.22) and (2.2.24) are changed so
that n/b takes the place of the factor n. Then in the computations of§ 3 we need
to put D2/b and D1 /b in place of D 2 and D 1 everywhere, and n/b in place of n.
In § 4 we spoke of the importance of the fact that vi < n1 16-• 1 (e1 > 0). Re-
placing n by n/b, we have

Vi< (~f6-' 1 bl/6-s1 < (~rs-s\ln n)Ks/6. (2.6.8)

It is easy to check that this does not violate the important relation (2.4.17).
The classes A1 and A2 of numbers z can be left unchanged for numbers z1 ; the
corresponding conclusions are preserved. The conclusions of § 5 are preserved
50 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

with the appropriate replacements, and we arrive at a formula which is anal-

ogous to (2.5.13). For given z 1 ~ (D1v1/0)112 we introduce the following sets of
conditions:
u<2.6.9):
X1 < (D
___!'II
0 2 '
2• r (x1, Z1) = 1, xz
1 1
<D2n-•s·
0 ' (2.6.9)

u<2.6.10>:
X1 < (D01 v2 2 '• r (X1, Z1) = 1, x z ~
1 1 ?
D2
0
n-•s.
' (2.6.10)

u<2.6.11>:
a= (M1, Z1) < n' 6, Z1 ~ r
(D01 '111 2. (2.6.11)

o
Then for given under condition (2.6.6) we obtain the following expression for
the total n.s.e. (2.6.1):

4 D2 !' !
0 "1."2 z1
( !
o: 1
_l_
X 1Z 1
+!
I, 1
! _l_) +
o: 1eI, 1 XoZ1
B D2v~2 (In n)-K'.
0
(2.6.12)
U12.&.11> U12.s.10> U12.s.1•>

Here x 0 is chosen in the interval I., 1 as it was chosen before in I.,. The accent (')
in the outer sum in (2.6.12) denotes that the summation is over only those for o
which
(2.6.13)
Formula (2.5.13) is a particular case of that given for = 1. To obtain the o
total combined n.s.e. (2.2..3) we need to sum over all ~ (ln n)K•. In doing this o
all constants K1 , • . . , Ks must be chosen sufficiently large; in particular,
Ks > 10a(3.2) (cf. (2.6.4)). Furthermore, we shall assume that K 4 > 100K1 •

7. Comparison of the Results for Coherent Numbers

Let (n1 , n 2) be one of the three pairs of coherent numbers described in § 2.
We would like to compare the numbers of solutions obtained by summation in
formula (2.6.12) for each of the given pairs.
Conditions Uc 2.6.9» Uc 2.6.10> and U< 2.6.11 > bring in the dependence of the
intervals of summation for z1 and x 1 on v1 and v2 • Let us consider whether it is
possible to replace them by conditions independent of v1 and v2 :
Uc2.7.1>:
1 r 2 x z < D2 n-•s
X1 < (D
~Yo ' (Xi. Z1) = 1, 1 1 0 ' (2.7.1)

Uc2.1.2>:
X1 < (D Yor2'-t (X1, Z1) = 1, X1Z1 ;;;., D2n-•s' (2.7.2)

Uc2.1.a>:
a= (M1 , z 1) < n•&; Z1 ~ (D
~Yo
1 2• r (2.7.3)
Sec. 7] COMPARISON OF THE RESULTS FOR COHERENT NUMBERS 51

For fixed v1 , v2 , CJ the replacement of the conditions in (2.6.12) by these

conditions produces the following error:

= B ln n ln ~ = B(ln n)-K1+1, (2.7.4)

Vo

in view of (2.1.9).
The corresponding difference for the other sums in (2.6.12) also gives an
error of (2.7.4).
Summing over v1 and v2 , taking condition (2.1.9) into consideration, we
obtain the error in the total number of solutions (for given CJ). Recalling that
v0 = v0/(ln n)K 1 (cf. (2.1.1)) we get an error of

(2.7.5)
Therefore we can write

instead of formula (2.6.12). Here the inner summation no longer depends on v1

and v2 ; the condition on the outer summation is

(2.7.7)

This circumstance enables us to carry out the summation with respect to "'i. v2 •
The numbers v1 and v2 run through the interval (2.1.9). By Lemma 1.6.3, the
number of them in this interval is

Li (v 0 + v~) - Li (v 0) + B (ln "'n)° K , (2.7.8)

6

where K 6 is an arbitrarily large constant. By definition of coherent numbers

n1, n2 , for each given v2 , v1 runs through a progression mod CJ with one forbidden
value, v1 = v2 ; by Lemma 1.6.3, the number of such v1 is

Li (v 0 + v~) - Li (vo) +B 'V~

(2.7.9)
qi( CJ) CJ(ln n) Ks
52 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

Note that it is important here that b < (In n)Ks and v ;;.,, exp .Jinn. Putting
this into (2.7.6), we obtain

4 ~(Li (v 0 + v~) - Li (v0)) 2

bp(b)

X L ( L - 1- + L L - 1-) + B D 2v~2 (In n)-K112• (2.7.10)

U •1 U "'1 X1Z1 lx 1 :t:1 Elx 1 XoZ 1 b
c2.1.a> c2.1 .1) Uc 2 •7 • 2 >

It is very important that the main term of (2.7.10), containing sums over x1
and z1 , no longer depends on the pair n1, n2 of coherent numbers. For all three
pairs of coherent numbers given in § 2, with b fixed, the main term of (2.7.10)
has the same value.
Let W(n 1, n2) be the total number of solutions of the fundamental equation
(2.2.3) for one of the three pairs of coherent numbers given in § 2, and let
W(ni, nz.) be the same thing for another of these three pairs. Then we have

W(n 1 , n 2) - W(n~, n~) = BD 2 v~2(ln n)-K512 L l

ci.;;(lnn)K• b

(2.7.11)
Lemma 2.2.1 of§ 2 follows immediately.

8. Lemma for the Hardy-Littlewood Problem

In Chapter VII we shall need one lemma about an equation of form (2.2.3).
Let n be a sufficiently large integer; the numbering of all constants begins anew.
nt n1 = exp (ln 1, (2.8.1)

xy < n; .J;,nJ. < x < ..J;,n 1 1, (2.8.2)

z runs through the same values as x; zt < n,

exp (In n)°' < v < ni/6-•• (i = 1, 2),

2 0 (2.8.3)
v0 n Di
v~ = (In n)K 1
; v0(ln n)Ko
< Di < -vn ; 0
D2 = (In n)K 1
;
D = [DI> Di + D 2], vi E (v 0) = [v0 - v~, v0], (i = 1, 2);
vi are primes. (2.8.4)
We consider three equations Yi, Y2 , Y3 of the form

V1XY - V2Zt = n(v1 -· V2), (2.8.5)

Sec. 8] LEMMA FOR THE HARDY-LITTLEWOOD PROBLEM 53

where
(2.8.6)

Here in equation Y1 , x z== 1 (mod 4); in equation Y3 , x =z =-1 (mod 4);

in Y2 , x = 1, z = -1 (mod 4).
LEMMA 2.8.1. The total numbers of solutions of equations Y1, Y2 and Y3, for
v1 =;if. v2 , differ from each other by an amount
(2.8.7)
where K4 is arbitrarily large for sufficiently large K1 .
In the proof we shall basically follow §§ 2-6 ohhis chapter. Let b = (x, z).
I
Put M = n(v1 - v2); b M. We can suppose that v1 1 z, for if v1 z, then v1 In; I
considering (2.8.3) the number of such solutions of our equations is estimated by
(2.8.7), as is easily seen. Following (2.2.8) we introduce the relations
(2.8.8)

(2.8.9)

where we have put n12 = n - v2(D1 + D 2); we have xy ~ n. From con-

gruence (2.8.8)
(2.8.10)

where tis an integer. We do not require here that y > x, z > t, but the following
conditions must be satisfied:
(2.8.11)

zt ~ n; 1 < t ~ -n
z
(of course, z =;if. 0, for z is odd). Thus it must be that

1 ~ v2 n - v1(n - xy) ~ ~. (2.8.12)

'Jl2Z Z

Replace the quantity (n - xy)/v 2 here by D1 ; then conditions (2.8.12) become

l ~ n - v1D1 ~ ~ . (2.8.13)
z z
The second inequality is trivially satisfied, and the first gives
54 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

which is also satisfied along with conditions (2.8.11), assumed earlier, because
of the conditions on Di. In view of (2.8.6), arguing as in § 2 after formula
(2.2.11), we see that with an error of (2.8. 7) in the total number of solutions, we
can replace (2.8.12) by (2.8.13), i.e., put no conditions besides (2.8.11) on z, t.

9. Analysis of the Fundamental Congruence. Another Lemma

Let (ri. T 2) be one of the four pairs of numbers (1, 1); (-1, 1); (1, -1);
(-1, -1). We shall suppose that z =
T2 (mod 4) and x =
Ti (mod 4). Put
x = 4xi +Ti; x' = (4xi +Ti)' (mod v2z) so that, because of (2.8.8),
(2.9.1)
For each x, y varies in interval (2.8.9). Following§ 3, we divide the interval
of variation of x into intervals of form (2.3.9)
(2.9.2)
which we denote by I., 0 •
Further, as in§ 3 (cf. the arguments following formula (2.3.7)) we show that
with a permissible error in the total number of solutions we can suppose that as x
varies in interval (2.9.2), y varies in

[ n12 , ni2 + v2D2], (2.9.3)

Xo Xo
which we denote by IY,xo~ Then we argue as in§ 3 after formula (2.3.15), which
we can also write here in the same notation.
We can continue as in § 3 after this formula, and then as in §§ 4-5; here
z = 1 (mod T2) does not change, but x = 4xi +Ti, which does not change the
estimate of the trigonometric sums. Because of this we can follow the previous
arguments right up to (2.5.2) which in this case is replaced by

!( !
4 XElzo
1) (DXoZ
2 + Bµ(z, x )n-••).
0 (2.9.4)
(x,z)=l

If we replace (Ti. T2) by (T{, T2) and subtract the corresponding formulas
(2.9.4), then we obtain the difference
Bµ(z, x 0 )n-•• ! 1. (2.9.5)
xelz 0

The summation of (2.9.5) over z and x 0 , according to what was said in§ 5,
gives an error of form (2.8.7).
The case (x, z) > 1 is treated as in § 6.
In Chapter VIII we shall need one more lemma, which is entirely analogous
to Lemma 2.8.1. Let the pair of numbers (ni, n2) run through the values (1, I);
(p', p'); (1, p'), where p' ~ ni-• 1 is a quasi prime number with exclusion up to
Sec. 10] LEMMA ON A CERTAIN INDEFINITE QUATERNARY FORM 55

Mn= exp (In In n)at 2 (see§ 5, Chapter I). Let vi obey condition (2.8.5); xy, zt
run independently through identical systems of values under condition (2.8.2).
Consider three equations YI> Y2 , Ya of the form

(2.9.6)
with

(2.9.7)

where Di and D 2 are defined in § 8.

LEMMA 2.9.1. The total numbers of solutions of equations Yi, Y2 and Ya
differ from each other by an amount
(2.9.8)
where K5 is arbitrarily large if Ki is sufficiently large.
The proof of this lemma is not significantly different from the proof of
Lemma 2.8.1 and is carried out in the same way.

10. Lemma on a Certain Indefinite Quaternary Form

In Chapter VI, and partly in Chapter VII, we shall need a certain lemma on
an indefinite quaternary form of the type v1(x2 + y 2) - v2(z 2 + t 2). In this
section the numbering of constants begins again.
Let n, Di and D 2 be as in the preceding section;

exp (In nl 0 < v0 < n 116 exp -(.Jinn), (2.10.1)

let vi be the primes defined by (2.8.4); Yi¥- v2 ,
2;. <(In n) 0 , (2.10.2)
where C is an arbitrarily large prescribed constant.
We consider the equation
v1 2;.(.;2 + r/) - v22\.;' 2 + r/ =2) n(v1 - v2). (2.10.3)

Here .;2 + 'Y/2 and .;'2 + 'Y/' 2 are odd;

n - 2;.($2
--~---'--'- E
+ 'Y}2) (D). (2.10.4)
Y2

In this connection, if n is even we shall suppose that A. = 0, and that always

(2.10.5)
56 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

LEMMA 2.10.1. The total number of solutions of equation (2.10.3) under the
above conditions is given by

(2.10.6)

Note that condition (2.10.5) is necessary for the solvability of equation

(2.10.3). For $2 + n2 = $' 2 + n' 2 = 1(mod4), so that from (2.10.3) we have

2'{v1 - v2) = n(v1 - v2) (mod 2.i.+ 2).

Ifniseven,il = Oandv1 - v 2 =
O(mod4); ifnisodd, obviouslyv1 - v2 =
O(mod 2-42).
Put M 1 = n(v1 - v2)/2\ so that (2.10.3) is equivalent to equation

(2.10.7)
Puttirig
U(m) = L 1=4Lxip) = 4 L xix), (2.10.8)
g2 +~ 2 =m p Im XY=m

we see that the number of solutions of (2.10. 7) is equal to the sum of the quanti-
ties U(mJU(m 2) over all solutions of the equation
n(v1 - v 2)
V1 m1 - v2m2 :::;::: ;. = Mi (2.10.9)
2
under the conditions
(2.10.10)
We see from (2.10.8) that it is necessary to find

16 L xix)xiz) (2.10.11)
over all solutions of the equation
V1XY - V2Zt = M1, (2.10.12)
n12 < 2"xy < n12 + v2D2; n12 ·= n - v2D1 - v2 D 2 • (2.10.13)
Here xy and zt are odd. We can suppose that

xy =zt =1 (mod4), (2.10.14)

otherwise U(m 1)U(m2) = 0. Note that ·because of our conditions, n(v1 - v2) =
0 (mod 4) and from xy =
1 (mod 4) it follows automatically that zt 1 (mod 4). =
Order the numbers x, y, z and tin (2.10.11) as follows:

x < y, z < t. (2.10.15)

Sec. 10] LEMMA ON A CERTAIN INDEFINITE QUATERNARY FORM 57

The number of solutions where the equal sign is encountered in (2.10.15) and
the sum (2.10.11) over them is small: the trivial estimate of the number of such
solutions is
B(e)n11 2+• (2.10.16)
(see§ 2).
In the remaining solutions the equal sign does not occur; assuming condition
(2.10.15), we shall later only have to quadruple the sum obtained (2.10.11). For
by (2.10.14), xix)xiz) does not change when x is replaced by y and z by t.
In what follows, the value of
c5 = g.c.d. (x, z) (2.10.17)

is important. Clearly, <5 j n(v1 - v2). We suppose first ofall that c5 = 1. Further-
more, we suppose v1 1 z; otherwise the total number of corresponding solutions
is estimated by (2.8.7) (see § 8)
y = M1 v~x' (mod v2 z), (2.10.18)
where v~x' is the inverse of v1x (mod v2z). Then

(2.10.19)

by (2.10.9). Let e1 > 0 be given and suppose that the following is satisfied:
,,,;~2 > v2z exp (In n)113 ; xz < D 2 exp -(In n)113T"'. (2.10.20)

In this case congruence (2.10.18) and the congruence

y = x(mod4) (2.10.21)

are solvable. We also need to take into account conditions (2.10.15). We do

this as in § 2.
The condition t > z gives

t = v1(n - 2"xy) + v 2 n > z. (2.10.22)

Here (n - 2"xy)/v2 E (D). Replace condition (2.10.22) by

v1D1 + n > z; z < (n + v1D1)112. (2.10.23)

z
This can be replaced by the still simpler condition

z < (n + VoD1)1/2. (2.10.24)

58 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

Arguing as in § 2 we see that the replacement of condition (2.10.22) by

(2.10.24) gives rise to an error of
12
BD 2 'V~;. (ln n)-K 2, (2.10.25)
2
in the total number of solutions, where K2 is arbitrarily large together with K1
(see § 8, where K1 enters into the definition of 'Vo and D 2).
Consider the condition y > x, which we write in the form
n - 2;.xy 2;.x 2
--~ <n - (2.10.26)
Replace this by
n - 2;.x 2
- - - >Di,
'V2
i.e.,
x < ((n - P 2 D 1)2-;.)112 • (2.10.27)
We replace this last condition by the simpler one
x < ((n - (P0 - P~)D 1)T,,.)1 12 • (2.10.28)
Again, reasoning as in § 2, we see that the replacement of (2.10.26) by
(2.10.28) gives a total error of (2.10.25).

11. Preparation for the Application of Vinogradov's Method

We shall seek an asymptotic approximation to the sum (2.10.11) for the
solutions of equation (2.10.9) under conditions (2.10.·28) and (2.10.24). First of
all consider the pairs (x, z): (x, z) = 1 undercondition(2.10.20). For each such
pair (x, z) the number y~ in accordance with (2.10.19) and (2'.10.18), runs through

(2.11.1)
values. Here
(2.11.2)

We sum (2.11.1) with a coefficient of xix)xiz). By (2.10.20) we obtain

~~ 2 2 xix)xiz) + BD r,,.-
2
2 exp -Hin n)113 • (2.11.3)
2 zz.;;;2-?.D 2 exp -clnn) 1 / 3 XZ
cz;z)

Here (x; z) indicates summation under all previous conditions on x and z, in

particular, the condition (x, z) = 1. Now consider the case when all these
conditions are observed except for (2.10.20), so that
xz > D 2 2-;. exp -(ln n)1 13 • (2.11.4)
Sec. 12] APPLICATION OF THE METHOD OF I. M. VINOGRADOV 59

From (2.10.28), (2.11.4) and the definition of the numbers vi we have

z > x-1 D 22-;. exp -(In n)113 > n113 exp i(ln n)112• (2.11.5)
In addition, for the same reasons
x > z-1 D 2r;. exp -(In n)113 > n113 exp t(ln n)112• (2.11.6)
Then we proceed as in § 9. We divide the interval of variation of x into
intervals of the form
Ix 0 = [x 0 , x 0 + x 0 exp -Hin n)]1 12. (2.11.7)
The last intervals can be neglected with a permissible error in the total number of
solutions (see§ 3). If x E Ix0 , then in accordance with (2.10.19) we can suppose
that y varies in the interval
fy xo = [ ~12 ' n12 ~ 1l2D2] (2.ll.8)
' 2 x0 2 x0
with a permissible error in the total number of solutions (cf. § 3).
Consider the case y = x = I (mod 4). Let y = 4y1 + 1 ; x = 4x1 + 1; then
y 1 =: M 1(4v1)'(4x1 + I)' - 4' (mod v2z). (2.11.9)
Denote by Ev.xo that part of the interval Iv.xo into which v2z fits an integral
number of times (if there is such a part); denote the remaining interval by Fv.xo'
so that
(2.11.10)
The ratio of the lengths Ev,xo and v2 z is an integer, which we shall denote by ex0
For each possible value of x in its interval, y takes on exactly ex 0 values of Ev.xo
(z is fixed). Further, letµ denote the ratio of the lengths Fv.xo and v2z (it depends
on x and z). If
µ < _n_ exp -l· (ln n)112, 0 (2.11.11)
Xo'V2Z

then it is not difficult to show that at the expense of a permissible error we can
discard a solution where y runs through intervals Fv,xo with such a small µ, and z
ranges over permissible values (cf. § 3). Therefore we shall assume that

µ ....._
,,,,_ -n- exp - 1
1- 0
(I n n )112 .
Xo'V2Z

12. Application of the Method of I. M. Vinogradov

For y E Fv,xo we must have
(2.12.1)
60 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

where we have puty2 = f(x 1) = M1(4v1)'(4x1 + l)' - 4' (mod v2z); { } denotes
the fractional part. In order to estimate the number of y 2 as x runs through its
interval, we proceed exactly as in § 4: we construct upper and lower Vinogradov
functions o/ and 'f to count the fractional parts falling in the interval [O, µ]. If
Q(z) is the desired quantity, then

(2.12.2)

We consider the behavior of 1f"; the second function is treated analogously.

In the lemma ofVinogradov in the form given in§ 7 of Chapter I (Lemma 1.7.1)
we put
tX = 0, /3 == µ, 1::1' =µexp -( 1-'-o(ln n) 112 ), r = 100; (2.12.3)
00 00

T(x) = µ' + -} L (am + ibm) exp 21Timx + -} L (am - ibm) exp (21Timx).
m=l m=l
Here
+ Bµ exp -l-5 (ln n) 112•
µ' = µ (2.12.4)
The series can be broken off form > µ- 1 exp -(l/20)(ln n)11 2 with a permissible
error; here we can suppose that
lam ± ibml = Bµ. (2.12.5)
The problem reduces to finding an estimate for

m ¥=- 0, Jml < µ- 1 exp -l0 (ln n)112 = M,,. (2.12.6)

Then we proceed as in § 4. We apply estimate (2.4.6) and, arguing as we
did in the derivation of formula (2.4.18) with the quantity n-•, replaced by
exp -a;(ln n)1 ' 2 , we find after computations analogous to formulas (2.4.7)-
(2.4.18) an analogue of (2.4.18)
Q(z) = µ(l + B exp -a1(ln n)112(M, z)112) L 1, (2.12.7)
where the summation is over x1 , observing all conditions which have been im-
posed on x (in particular, (x, z) = 1; x = z = 1 (mod 4), and so on).
We had y = =
x 1 (mod 4). If x =
-1 (mod 4), then we must have
y x = = -1 (mod 4). For such x we obtain the same sum as in (2.12.7) with
x =
1 (mod 4) replaced by x =
-1 (mod 4). Clearly the corresponding Q(z)
(for x = 1 and for x =
-1 (mod 4)) will not differ by more than

Bµ( ! 1) exp -alln n)112(M, z)112• (2.12.8)

(.,,zl=l
rr;E[z
Sec. 13] SUMMATION OF THE BASIC SERIES 61

We now sum the expression xix)xiz) for given z and x E /,,: (x, z) = 1, x
odd. Here xiz) is fixed, and by the above, the sum over xix) gives the value
(2.12.8). Then for y E lv,zo the corresponding sum is
Be,,0 ( L
(z.z)=l
1) exp -(In n) 112, (2.12.9)
:xelz

as can be seen by an elementary computation. Altogether, (2.12.8) and (2.12.9)

give the estimate
B(e.,0 + µ)(cz~= 1 1) exp -a 2(ln n)112(M, z)112 • (2.12.10)
ZE[z

We then argue as in§ 5 in summing the error term in (2.5.2); summing over z
and vi, we easily obtain the total error
BD 2v~2 exp -a 3(ln n)112 . (2.12.11)

13. Summation of the Basic Series

o
We have studied the case = 1. Now let > 1. We have on(v1 - v2). o!
If < exp (ln ln n) 2, then, as before, the corresponding solutions can be dis-
o
carded at the expense of a permissible error. Let o >exp (ln ln n) 2 ; x = x 1o;
z = z 1o; (x1 , z1) = 1. Here o is odd. We obtain the equation

V1X1Y - 'P2Z1t = -M . (2.13.1)

0
In this connection, plainly xixz) = xix1z1), and the equation can be treated like
the old one, with M replaced by M/o and Di by DJo (i = 1, 2). We return to the
case o = 1. Consider (2.11.13). We need to sum the series

!
(z,z) xz
(2.13.2)
zz.;;2-i.D 2 exp -(lnn)l/3

observing the conditions which have been imposed on x and z. Put D0 =

2-.i.D2 exp -(ln n)1 13 • Our series can be summed in an elementary fashion:
consider the region 1 < x < .J D 0 ; l < z < .J D 0 and additional regions, the
sums over which give small remainders. As a result of the summation we get

L xix)xiz) = ~ (L(l, X4))2 + Bn-0.01 = ~- + Bn-0.01. (2.13.3)

zz<;;Do XZ 7T
(z,z)=l

We still need to take into account a factor of 16 with xix)xiz), a factor of

D 2/2.i.+2 with (2.13.2), and the ordering x < y, z < t, which forces us to quad-
ruple the result; we find
(2.13.4)
62 ANALYTIC THEORY OF QUADRATIC FORMS [CHAP. II

bringing in all the error terms and summing them over

L 8 ~ 2 (1 + B 0 /ln n)- 0 i). (2.13.5)

Vi ,Cv 2 ;v;E(v 0 ) 2
vi =v 2 (mod 2-'+2)

Now let b > I, b <exp (In In n)2 • By the above, we have the expression

L 8~2 (1 + Bci(ln n)-Ci) (2.13.6)

Vi ,Cv 2 ;v;E(v 0) 2 b
•i =v 2 (mod 2-'+o)

for the total number of representations. Then

! ! = B 0 (ln n)- 0 , (2.13.7)

olMi b
6 >exp (In In n) 2
by Lemma I. I. 7.
Formulas (2.13.6) and (2.13.7) give the result (2.10.6).

14. A Quaternary Indefinite Form of More General Type

Let rp(x, y) = ax2 + bxy + cy2 be a positive binary quadratic form with
discriminant -d = b2 - 4ac < 0, d being free of squares with the possible
exception of the number 4. Let n be a sufficiently large integer; let v1 , v2 be
primes. or identical powers of primes such that
vi E (v) = [vo - v~, vo]; v~ = Vo K '
(Inn) i

exp (In n )"'o < Vo < no.01.

We consider the equation
v1rp(~, 'Y/) - v2rp(~', 'Y/ 1
) = M, (2.14.1)

IMI < v0 n; b0 = (M, 2d) <(In n)"'i. (2.14.2)

Here the following requirements are imposed on rp(~, 'YJ). Let D 1 and D 2 be
defined as in § 8. Then the condition

n - rp(~, 'YJ) E (D) = [D 1, D 1 + D 2] (2.14.3)

'J/2
must be satisfied.
LEMMA 2.14.1. Under the above conditions the number of solutions Q(M, v1, v2)
of equation (2.14.1) is given by the formula
Q(M, v1, v2) = G0 D 2S(n, v1, v2) + B0 D 2(ln n)- 0 , (2.14.4)
Sec. 14] A QUATERNARY INDEFINITE FORM OF MORE GENERAL TYPE 63

where G0 > 0 is a constant;

S(n, v1 , v2) = !
q..;;exp(lnlnn) 4
q- 4 !
(a,q)=l
S 1(a, q)Sla, q) exp - 27TiaM,
q
(2.14.5)

where

(2.14.6)

S 2(a, q) ""
= ..:.. exp - -27Tia V2<p(~, ri). (2.14.7)
~.'lmodq q
The proof of Lemma 2.14.1 can be accomplished by Kloosterman's method
(cf. [46]) but since it is rather cumbersome we shall not give it here. A complete
proof is given in [20].
LEMMA 2.14.2. Let <p = ax2 + bxy + cy2 be a positive primitive binary form
with discriminant -d < 0. Let m > 0, (m, d) = 1; then the number of repre-
sentations of m by primitive forms of the same discriminant is

'P'(m) = w!xic5), (2.14.8)

olm
where xic5) = ( ~d) is the Kronecker character; w = 2 when d < 4, w = 4
when d = 4, w = 6 when d = 3.
For a proof of this, see, for example, E. Landau [47, pp. 144-147].
LEMMA 2.14.3. The number of classes h offorms <p of the above type is given
by the formula of Dirichlet

w)d
h = - L(l, Xd). (2.14.9)
27T
For a proof see, for example, [47].
 CHAPTER III

THE ADDITIVE DMSOR PROBLEM

1. Statement of the Problem and Several Results

The additive divisor problem consists in finding an asymptotic formula for
expressions like
L
Tk 1(m)Tk.(m + a)
m~n
(3.1.1)
and
L Tk (m)Tk (n -
m<n
1 2 m). (3.1.2)

Here Tiin) is the number of factorizations of the number m into k factors,

counting order (with regard to the notation, cf. the section "Notation"). Thus

where the xi are positive integers. In particular, T2(m) = T(m) is the number of
divisors of m.
In what follows, a > 0 is a fixed number and n is an integer -Ho.
This problem has been investigated by a number of mathematicians, princi-
pally by the British. It was first studied by A. Ingham [44) in 1927 for the case
k1 = 2, k 2 = 2, i.e., for the number of divisors. He established the formulas

L T(m)T(m + a),....., 6 2 0'_1 (a)n(ln n)2 , (3.1.3)

m~n '1T

6
L T(m)T(n - m},....., 2 0'1(n)(ln n) 2 , (3.1.4)
m<n 7T

where O'a_{n) = ~din d"".

The singular case a = 0, i.e. an asymptotic approximation for ~m.;;n (T(m)) 2
can be reduced to a straightforward application of the theory of the Riemann
'-function; it was considered by Ramanujan [49) in 1915. Here one obtains
1
L (T(m)) 2 ,.....,
2 n(ln n)3 •
m:s:;;n 'TT

We see that here the right-hand side contains a factor of(ln n)3 i.nstead of(ln n) 2 •
This is proof of the sharp distinction between the case a = 0 and the general
case a> 0.
64
 STATEMENT OF THE PROBLEM AND SEVERAL RESULTS 65

In 1930 Ingham's result was improved by T. Estermann [36; 37] who obtained
asymptotic series in place of the right-hand sides of (3.1.3) and (3.1.4). For the
case (3.1.4), Estermann's result is
2 2-r
L T(m)T(n - m) = n L (In nY.Lar,;a~Un) + Bn718(1n n)2314a_ 31in). (3.1.5)
m<n -o ~O

In 1942 Titchmarsh [55] considered the problem for the case k 1 = 2, k 2 = 3.

His results were of a heuristic nature; he did not aim at rigorously substantiating
an asymptotic approximation. In 1950 Bellman [28] published an asymptotic
expression for the cases k 1 = 2, k 2 = 3 and k 1 = 3, k 2 = 2 of(3.1.1); however,
the reasoning in this paper is incorrect.
In 1957 Hooley [42] gave the first derivation of asymptotic expressions for
(3.1.1) and (3.1.2) in the cases k 1 = 2, k 2 = 3 and k 1 = 3, k 2 = 2. Hooley's
formulas are
.L T3(m)T(m
m..;;n
+ a) l= !A(a)n ln3 n + O(n(ln n In In n)) 2, (3.1.6)
m~nT(m)T 3(m +a) j
L T3(n - m)T(m) = !A1(n)n ln3 n + Bn (1n n In Inn lJn
m..;;n
L T(l) lIn 1) 2• (3.1.7)
Here A(~) and A 1(n) are expressions of rather complicated structure, given in
detail in [42]. The expression A(a) in the simplest case a = 1 is

A(l) = _lµ(~) L <p(d) . (3.1.8)

n=I n aln d
Note that (3.1.8) can be written in the form
00
A(l)=!-2µ(n) ITP 1)2) .
( 1- ( 1 - - (3.1.9)
n=I n pJn p
The divisor problem is a binary additive problem. That is, (3 .1.1) corresponds
to the equation
X1X2 · · · xk, - Yi.Y2 · · · Yk, = a, (3.1.10)
where X; and Y; are positive integers, and
x 1x 2 ... xk, < n. (3.1.11)
Analogously, (3.1.2) corresponds to the equation
x 1x 2 ..• xk, + Yi.Y2 ... Yk, = n. (3.1.12)
Solution of the binary problems (3.1.10) and (3.1.12) for any integer values
of k 1 and k 2 appears to be extraordinarily difficult. It can be observed that using
the author's method, described in papers [10] and [11], and conditionally
accepting the Riemann hypothesis only for the Riemann '-function, it is possible
66 THE ADDITIVE DIVISOR PROBLEM (CHAP. III

to obtain a satisfactory asymptotic formula for the number of solutions of the

inequality
Jx1x2 ••• xk, + YtYz . .. yk 2 - nl <
hno (3.1.13)
where hn = B(ln n)0 (C > 0). Inequality (3.1.13) can be considered as an
approximation to equation (3.1.12). Equation (3.1.12) has some interesting
applications. In particular, it is related to certain classical problems in the theory
of Dirichlet characters. We shall indicate briefly its relation to the well-known
problem ofVinogradov on an estimate for the size of the least positive quadratic
nonresidue (mod p).
Let numbers s > 0and17 > 0 be given. In equation (3.1.12) let k1 = k 2 = k,
and suppose that for given k > 1 and N > N0(k) we can determine an asymp-
totic formula or at least a satisfactory lower bound for the number of solutions
of equation (3.1.12) for all numbers n < N, with BN1-ri possible exceptions (such
exceptions can depend on k). Let p be a prime such that p = -1 (mod 4);
choose N so that Nri 12 < p < 2NY1 12 , and consider the equation

By the above, we shall have a satisfactory lower bound for the number of solu-
tions for at least one n1 . On the other hand, it is easy to give an upper bound for
the number of solutions having X; > p• or Y; > p• for at least one indexj.
If n is sufficiently large in comparison with s, then the number of such solu-
tions makes up a comparatively small part of the total. Having a satisfactory
estimate of the latter, we see that there exist solutions for which X; < p•, Y; < p•
for all j < k. Take one such solution and form the corresponding Legendre
symbols (XiXz~· ·xk) and (YtYz;· ·Yk). By our equation and the fact that
p = -1 (mod 4), we see that at least one of these symbols has the value -1,
whence it follows immediately that the least positive nonresidue mod p does not
exceed p•.
2. Application of the Dispersion Method
The author has applied the dispersion method to the additive divisor problem
in papers [12] and [13]. Only the results and a short outline of the computations
were given in these papers. In this chapter a detailed exposition will be given of
the much simpler scheme of computations resulting when the concept of
covariance is applied.
The dispersion method enables us to find an asymptotic expression for
(3.1.1) and (3.1.2) and to derive an asymptotic series of type (3.1.5) for the cases:
k1 = 2, k 2 is any positive integer;
(3.2.1)
k1 is any positive integer, k 2 = 2.
Sec. 3] PREPARATORY TRANSFORMATIONS OF THE EQUATION 67

Here, however, the error term is worse than in the particular case (3.1.5);
namely, it is of the form B,n(ln n)'. At the end of this chapter we shall also
consider the application of the dispersion method in the general case-that of
any values of k1 and k 2 •
Since the resulting formulas are very cumbersome, we shall give a detailed
discussion only for the simplest particular case of equation (3.1.1): k 2 = 2,
k1 = k > 2 is arbitrary, a = 1; and we shall limit ourselves to the derivation
of an asymptotic formula, without derivation of an asymptotic series. We shall
only give a general outline of the derivation of the latter and of the general case
of equations (3.1.1) and (3.1.2) with k1 and k 2 as in (3.2.1). Close examination
of this outline shows that there are no essential difficulties in obtaining the
corresponding general result; it is just that much cumbersome computation is
required. Here we derive the following result in detail (cf. the author's papers
[10-12]). 1
THEOREM 3.2.1. For any k ~ 2,
,L T(m + l)Tk(m) = k! AkSkn(ln n)k + Bn(ln n)k- (1n In nl,
1 (3.2.2)
m::s;;;;n
where
(3.2.3)

S = lim 1 (3.2.4)
k Y-+co (In Yt- 1

To derive an asymptotic series of type (3.1.5) one needs to make a certain

calculation, which is not complicated, but is cumbersome in the sense of the
computations involved; this is done with the help of formulas (3.6.14) and
(3.6.15), given below in§ 6. Brief comment on this subject appears at the end
of§ 8. As was already remarked, the dispersion method as applied here gives an
error of B,n(In n)' in the asymptotic series.

3. Preparatory Transformations of the Equation

:E111 ,.;;n-iT(m + I)Tim) is equal to the number of solutions of
xy - x 1 x 2 ••• xk = 1, (3.3.1)
where x, y, x1, •.. , x,, are positive integers, and
x 1x 2 ••• xk ~ n.
With the application of the dispersion method in mind, we put cp = xy (of r ..1urse,
repetitions of the values of cp are permissible). We should like to represent the
numbers x1x 2 ••• xk in the form D' v, where v ~ n116-'1, but is not too small.
1 By an oversight, the error term given in [12] is of unsubstantiated precision.
68 THE ADDITIVE DIVISOR P~OBLEM [CHAP. III

Rewrite (3.3.1) in the form

xy- X= I, (3.3.2)
where X < n and each value of Xis counted Tk(X) times. We now show that we
can neglect the number of solutions of (3.3.2) where Tk(X) is extremely large.
Let Ain) be the class of numbers X n for which <
Tk(X) < (ln n)r(k); r(k) = 4a(k, 2),
cf. (1.1.5).
LEMMA 3.3.1. An estimate for the number of solutions of (3.3.2) for which
XE Ak(n) is
(3.3.3)
For the proof, we observe that if x, y, Xis a solution of equation (3.3.2), then
X = xy - 1, TiX) = Tixy - l); we can consider only those solutions where
x< y, so that x <.Jn+ 1,, and then double the number of these solutions.
Further, by Lemma 1.1.4, cf. (1.1.5), we have
2 2 T~(xy - 1) = Bn(ln nt<k, 2>. (3.3.4)
o:<:;;V n+l o:y<:;;n+l

For those solutions where XE A,«n), we have Tixy - 1) > (ln n)'<k>, so that,
by (3.3.4),
( 2_ 2' Tk(xy - l))(ln n)'<k>+i = Bn(ln n)'<k>,
x~ V' n+l xy~n-Fl.
where the accent denotes that the summation is taken over X = xy - 1 E Ain).
Hence Lemma 3.3.1. We can now suppose that in equation (3.3.2), XE Ain).
Now we subdivide the numbers XE Ain) into classes n<u and n< 2 >. The
class n< 1 > will consist of numbers X having a prime divisor 11 such' that
exp .Jln n < < n11
11 &-• 1 ; (3.3.5)
the remaining XE Ain) go into n< 2 >. We now prove that solutions for which
XE n< 2> can be neglected.
LEMMA 3.3.2. An estimate for the number of solutions of (3.3.2) for which
XE n< 2 > is
B,n(ln n)'. (3.3.6)
All the prime divisors 11 of XE are either< exp (ln
n< 2> or> n1 t&-•1. n)11 2
Consider first the latter case. In this case X has no more than 6 prime factors,
so X = Pis "almost prime" and TiX) = B. Hence the number of correspond-
ing solutions of (3.3.2) is
B 2 2' 1, (3.3.7)
:r<:;;vn+l X=l(rnod:r)
X:s;;;n
Sec. 3] PREPARATORY TRANSFORMATIONS OF THE EQUATION 69

where the accent indicates that almost prime X =Pare taken. By Lemma 1.3.1
we have ~~=l(modz);X.;;;n 1 = Bn/cp(x) ln n. Summing over x we obtain
(3.3.6). -
Now consider those X, all prime divisors of which are less than exp .Jin n.
First of all let us estimate
1' T(X + 1) = 1 T(m + l)p(m), (3.3.8)
m~n

where the summation is over these X and p(m) = 1 if m = X; p(m) = 0 other-

wise. For (3.3.8) we have the obvious estimate

B Ct,;+1(T(m))2r2 c~/(m)r2.
By Lemma 1.1.6, where we need to take x = n, z =exp .Jinn, ex= (ln n)-112,
we find, cf. (1.1.9): ~m.;;;n p(m) = Bn exp -(.Jln n); by Lemma 1.1.2,
~m.;;;n+i(T(m)) 2 = Bn(ln n) 3. In view of this,

1' T(X + 1) = Bn exp - t.Jln n. (3.3.9)

Then each X in (3.3.8) must be repeated TiX) times, but since XE Ain),
TiX) ~ (ln nYk. Choosing C = 3rk and summing over x ~.Jn+ 1, we find an
estimate for the required number of solutions
Bn(ln n)-rk.
Together with (3.3.9) this completes the proof of Lemma 3.3.2.
Now we shall assume XE nm. The fundamental equation (3.3.1) can be
rewritten as
xy- vD' = 1, (3.3.10)
where vis a prime satisfying condition (3.3.5) and the number vD' is repeated
TivD') times. We now see to it that in each vD' there are no repeated factors v.
If there are, then X = vD' = 0 (mod v2); it is not difficult to show that such
solutions of (3.3.10) can be neglected. An estimate for the number of them is
B(ln nYk 1 1
(•) X= O(modv 2 )
T(X + 1). (3.3.11)

Since exp 2.J1n n ~ v2 ~ n11s-2e1, then by Lemma 1.1.3, an estimate for (3.3.11)
is
n -
B(ln nrk+l 12 = Bn exp (-t.Jln n), (3.3.12)
(v) 'V

and we can neglect such solutions. Thus we can suppose that there are no
repeated factors v in the numbers vD'. Obviously, the number of such factors in
X = v D' does not exceed
r =.Jinn. (3.3.13)
70 THE ADDITIVE DIVISOR PROBLEM [CHAP. III

Separate the numbers belonging to nm into classes D 1 a where a ~ r, according

to the number a of their factors v, and consider the numbers Xin a definite class
Dia· For X = vD' E Dia the number of representations in the form X = vD' is
obviously equal to a. Further, each XE D 1 a must be counted Tk(X) times.
Clearly, g.c.d. (v, D') = 1 and TivD') = Th)TiD') = kTk(D'). Moreover, if
vD' E D 1 a, this means that D' has exactly a - 1 different prime factors of the
type of v, and TiD') ~ (ln nYk/k. Conversely, any number X = vD', where v
is a prime satisfying condition (3.3.5) and D' satisfies the above conditions, is a
member of the class Dia· We shall call numbers D' of this form numbers of the
class DJ.a· Now divide the solutions of equation (3.3.2), where XE nm, into
classes of equations; Qa will denote the numbers of corresponding solutions.
Thus,
Qa = N(xy - X = 1 XE D 1a), I (3.3.14)
where Xis counted TiX) times. Because of the above,

Qa = ~ N(xy - D'v = 1 D' E I D~a), (3.3.15)

a
where each D' is counted Tk(D') times. Here D' v ~ n, D' ~ n/v; D' does not in
any other way depend on v. In order to make D' and v completely independent,
divide the interval (3.3.5) of variation of the numbers v into intervals of the form

(v) = [v 0 , v0 + v~]; 'JI~=

Yo K ; K 1 > 1000. (3.3.16)
(ln n) 1
Beginning with v0 = exp .Jinn, these intervals must abut; the last (extreme
right) interval may be incomplete.
Consider the equation
xy - D'v = 1, D' E D~a (3.3.17)
and those of its solutions for which vis in (v). Here D' E DJ.a has a - 1 prime
factors satisfying condition (3.3.5); TiD') ~ .J(ln n)'k/k and D' ~ n/v. Let us
replace the last condition by
D' ~ !!._ (3.3.18)
Yo
and estimate the error in the number of solutions caused by this. Here we need
to consider solutions for which
__
n_ ~ D' ~ !!._ (3.3.19)
,-..::::::::: ~ '
Yo+ Yo Yo
so that D' = n/v0 + ~. ~ = B(n/v5)v0. An estimate for the total number of
solutions with such values of D' for all values a = 1, 2, ... , r is
B(ln np L L !' 1, (3.3.20)

where the accent on the inner sum means that D' is subject to condition (3.3.19).
Sec. 3) PREPARATORY TRANSFORMATIONS OF THE EQUATION 71

This expression has the estimate

B(ln nY' ! -nv'0

2
= Bn(ln nr+1(ln n)-2K 1 = Bn(ln n)-< 3l 2>K 1. (3.3.21)
x.;;;Vn+l 'V~X
Summing over all intervals (v), the number of which is estimated by B(ln n)K 1+1,
we obtain the total error in the number of solutions
(3.3.22)
Thus in equations (3.3.17) when a = 1, 2, ... , r we can suppose, for v E (v),
that D' ~ n/v0 , D' E D:la and does not depend on v. Now it is convenient to
divide the interval [1, n/v0 ] of variation of D' into intervals of the form

D _ Di (3.3.23)
2 - (In n)K2'

where K2 is a sufficiently large constant and

(3.3.24)

when K 0 > 100.

We now obtain the equation
xy - D'v = 1, (3.3.25)
with conditions
v E (v); D' E (D); D' E D~a• (3.3.26)
which is in a very convenient form for the application of the dispersion method.
However, we still need to estimate the error arising from solutions where v
belongs to an incomplete interval of the form (v), and those where for given
v E (v), D' ~ D 0 • It is convenient to do this for the total number of solutions,
a= I, 2, ... , r.
As was noted earlier, only the extreme right interval of values of v can be
an incomplete interval of form (v), so
(3.3.27)
For an equation of form (3.3.25), where D' ~ n/v, D' E Dia• we obtain as an
estimate for the number of solutions here
B(ln nr !' ! -r(D'v + 1), (3.3.28)
v D'v~n

where the accent denotes summation over v in interval (3.3.27). Applying

Lemma 1.1.3 we find that an estimate for (3.3.26) is
(3.3.29)
72 THE ADDITIVE DIVISOR P.ROBLEM [CHAP. III

Now let Y E (Y) and D' < n/Y0(ln n)K0 , D' E DI.a, then the corresponding
number of solutions is estimated as in formulas (3.3.20) and (3.3.21) with 2K1
replaced by K1 + K 0 ; the final result is

Bn(ln nrKo/3 (3.3.30)

instead of (3.3.22).
Thus, with a permissible error in the number of solutions, we have reduced
the fundamental equation (3.3.1) to the collection of equations (3.3.25) with
conditions (3.3.26) for a = 1, 2, ... , r.

4. The Dispersion of the Difference Between Solutions for Coherent

Numbers
We introduce the system (2.1.1) of numbers coherent to the number 1(see§2
of Chapter II). Putting <p = xy; U(m) = L<p=m 1 = T(m), form the dispersion
of the difference between the numbers of solutions for equations (3.3.25), under
conditions (3.3.26), and for
xy- D Y = n2 1
(3.4.1)

under the same conditions (3.3.26), where n2 is one of the numbers (2.1.1).
Recalling that D' E D~a' so that TiD') < (ln nYk/k, we obtain for this dispersion,
comparing (0.4.2),

V' < (In nr Z (

D'e(D)
z U(l + D'Y) - z U(n
ve(v) ve(v)
2 + D'y))2 · (3.4.2)

Then we follow § 4 of the Introduction, and for Y1 ¥:- Y 2i Yi E (Y) We come to

the three basic equations of the dispersion method, written in the form (2.2.3)
with conditions (2.2.4). In order to reduce the problem of estimating V' to one
of solving these equations, we also need to estimate
(In nr 1 1. (U(n1 + DY))
Di,,;;;D,,;;;D1+D2 ve(v)
2, (3.4.3)

where n1 = 1 or n1 = n2 • By Lemma 1.1.5 (cf. formula (1.1.5)), we have for

given YE (Y), Ln 1 ,,;;;n,,;;;n 1 +n 2 (T(n1 + DY))2 = BD2(ln n)a< 2·2>. So an estimate for
(3.4.3) is
(3.4.4)
Now we apply Lemma 2.2.l (formula (2.2.1)). Together with (3.4.4) this
gives the estimate
(3.4.5)
for V' if K1 > 12rk, which we shall assume.
 v v
Sec. 5] APPLICATION OF THE CEBYSEV INEQUALITY 73

5. Application of the Cebysev Inequality

We have found that the dispersion of the difference between the solutions is

V' = L (L U(n1 + D'v) - I U(n 2 + D'v))2 = BD 2v~2 (ln n)-< 1t4>K,_

D'e(D) ve(v) ve(v)
(3.5.1)
Here D' E D~a; Di~ D' ~Di+ D2 ; each D' is counted -rk(D') times.
Now divide the numbers D' E Dia over which the sum is taken into two
classes, Ei and E 2 •
The class Ei consists of numbers D' for which

II
ve(vl
U(ni + D'v) - I
ve(v)
U(n 2 + D'v) I~ (ln v~n) K
3
, (3.5.2)

and the class E 2 of the remaining numbers, for which

II
ve(v)
U(n 1 + D'v) - I
ve(v)
U(n 2 + D'v) I> (ln v~ n)
K
3
, (3.5.3)
where
(3.5.4)
(We assume that Ki > 160.)
Now divide the numbers D' E E 2 into subsets E 21 (t = 1, 2, ... ) such that
D' E E21 if

vf(t + l)(ln nrK· ~ II ve(•)

U(ni + D'v) - I
ve(v)
U(n2 + D'v) I~ V6t(ln n)-K•.
(3.5.5)
Since by Lemma I.I.I, U(ni - D'v) = B,n•, clearly t ~ B,n•.
Now, using the obvious analogue of the Cebysev inequality, we estimate the
number of D' E E 21 • From (3.5.1) we have
N(D' E E21) = BD2v~2(ln n)-(l/ 4 >K 1 (v~r 2 t- 2 (ln n)2Ka
= BD2(ln n)-c11s>K1 ~. (3.5.6)
t
Here Ki~ 1000K0 (cf. (2.1.2)); t ~ B,n•.
Now let us estimate the sum

I
D'eE 21
II ve(v)
U(n 1 + D'v) - I ve(v)
U(n 2 + D'v) I· (3.5.7)

By (3.5.5) and (3.5.6) an estimate for (3.5.7) is

BD2v~(ln nr(ln n)-(l/S)K, l . (3.5.8)

t
 74 THE ADDITIVE DIVISOR PROBLEM (CHAP. Ill

Here we take into account the fact that each D' is repeated -rk(D') times.
Summing over t ~ B,n•, we obtain the estimate
(3.5.9)
for the sum
I z U(n
ve(v)
1 2 U(n2 + D'v) I·
+ D'v) - ve(v) (3.5.10)

Then it follows immediately from the definition of £ 1 that

2
D'eE1
I2 ve(v)
U(n1 + D'v) - ve(v) I
2 U(n2 + D'v) = BD2v~(ln nrK• 12. (3.5.11)

In both sums D' is counted -rk(D') times; K 3 > 2rk.

Thus, combining (3.5.10) and (3.5.11), we find

2
D1<;;;D'<;;;D'1+D2
-rk(D') I2
ve(v)
U(n1 2 U(n 2 +
+ D'v) - ve(v) D'v) I= BD 2v~(ln n)-K"12.
D'eD1a
(3.5.12)
Note that the information supplied by (3.5.12) for a given value of a may be
trivial; for example, when a = 1 the set D~a contains relatively few elements D'
and (3.5.12) can be obtained from these considerations. But for a = 1, 2, ... , r
(3.5.12) becomes deeper. Taking into account (3.3.15), (3.3.14) and (3.3.13), we
obtain from (3.5.12)
2 -rk(D'v)-r(D'v + 1) - 2 -rk(D'v)-r(D'v +n 2)
D1<;;;D'<;;;D1+D2 D1.;;;D'<;;;D1+D2
vD'eD('); ve(v) vD'eD('); ve(v)
= BD 2v~(ln nrKal3. (3.5.13)
We remark that the function B has a uniform estimate IBI < C0 for all n2 = p
given by (2.1.1).
Let v E (v); our intervals [D1, D 1 + D 2] were parts of the interval [1, n/v0 ].
Writing (3.5.13) for all such intervals, taking (3.3.24) into consideration, and
combining, we find
Z -riD'v)-r(D'v + 1) - Z -riD'v)'r(D'v + n 2)
D'<;;;n/vo D'~nlvo
vD'eD('); ve(v) vD'eD('); ve(v)
= BD 1 v~(ln nrKal 3• (3.5.14)
The interval (3.3.5) of variation of v contains B(v0/v0) Inn intervals of type (v).
Next, we can extend the summation over D', taking 1 ~ D' ~ n/v for given v.
The total error arising from such a replacement is given by (3.3.21) and (3.3.22).
If we replace equation (3.3.17) by a similar equation in which n2 stands in the
place of 1 on the right-hand side, the estimates of the errors are obviously
preserved. Granting this, we find, with the help of (3.5.14),
2 -riX)-r(X + 1) - 2 -rk(X)-r(X + n 2 ) = Bn(ln n)-2 • (3.5.15)
XeD(') XeD(')
Sec. 6] REDUCTION OF THE ADDITIVE DIVISOR PROBLEM 75

Now we apply Lemmas 3.3.1 and 3.3.2 which, of course, are also valid when
(3.3.2) is replaced by a similar equation with n 2 on the right-hand side instead
of 1. Putting (3.3.3) and (3.3.6) together with (3.5.15) we find the fundamental
result
L
Tk(X}r(X + 1) -
x::s;;;;n
L
Tk(X)T(X + n2) = Bn,
X::s;;;;n
(3.5.16)

where IBI is uniformly bounded for all coherent numbers n2•

6. Reduction of the Additive Divisor Problem to a Ternary Additive

Problem
(3.5.16) enables us to reduce the additive divisor problem to a ternary
additive problem. By (2.1.1) n can be any quasiprime p' with exclusion up to
Mn= exp (ln ln n)312 such that
!nl-•o :e;; p' :e;; nl-•o (3.6.1)
(cf. § 5, Chapter I).
Let L(h) be the number of such quasiprimes. Write the ternary additive
problem
(3.6.2)
with the conditions
x 1 x 2 ••• xk :e;; n; p' is quasiprime;
(3.6.3)
Let Q(n) be the number of solutions (x, y, x 1 ••. xk, p') of this problem. By
(3.5.16) the number of solutions of the fundamental equation (3.3.1) is given by
the formula
Q(n) + Bn. (3.6.4)
L(n)
Thus the problem is reduced to a ternary one (3.6.2). This problem can be solved
by the method of trigonometric sums, there being no difficulties other than
computational ones. However, by making use of properties of the quasiprime
numbers p' we can do this in a much shorter and more elementary way, which
will also be useful in what follows.
Let Qn be the set of all integers whose prime divisors are all :e;; Mn· Further,
let I' be interval (3.6.3) in which our quasiprimes p' must lie. Then by definition
of p', the n.s.e. (3.6.2) has the form
L µ(q)N(xy - x 1 x 2 • . . xk = q; Iq; EI'). (3.6.5)
a::s;;;;no
qeO,.

It will be more convenient for us to have xy and x1 x 2 ••• xk free of geometric

relations of the form q; EI'; this introduces a permissible error into the number
76 THE ADDITIVE DIVISOR PROBLEM [CHAP. III

of solutions. We can also do this in the first equation (3.6.2). Divide the interval
[I, n] into H = [n 2 ' 0 ] intervals oflength n/H; call these / 1 , / 2, ••• , IH. Of these
intervals let the intervals I H 1 , I H 1+1, ••• , / 2H 1 differ in sum by not more than
2n/H from interval (3.6.3); !"will denote the system of those intervals.
Now we shall consider only those solutions of (3.6.2) for which
xy E Ia, x1x 2 ••• xk E /b; p' E /",

a - b = c, (3.6.6)
where c is one of the numbers H 1 - 3, H 1 - I, ... , 2H1 + 3.
We shall prove that with a permissible error we can restrict ourselves to these
solutions. To see this, consider first of all the condition: p' E 10 c /". If it is
violated, then the number of values of p' does not exceed 2n/Hand in that case
the number of solutions of (3.6.2) is obviously
n2
B - n•
•H
= Bnnon" 012' n0 = n 1-• 0, (3.6.7)

which is negligible. Thus, let p' E !"; let xy E Ia, x 1x 2 .•• xk E /b and suppose
that (3.6.4) is not satisfied. Then we have
Ixy - x 1x 2 ••• xk I> n1-•o + -n ,
H
or
lxy - X1X2 .•• xkl < !nl-•o - _!: '
H
which contradicts equation (3.6.2) with conditions (3.6.3).
Thus we shall consider values of xy which are in Ia, and x 1x 2 ••• xk in-
dependently running through the system of numbers of lb under condition
(3.6.4); if we collect all such solutions and restrict ourselves to them, then the
permissible error in the total number is
(3.6.8)
Let Q(Ia, lb) be the number of solutions of equation (3.6.2) when xy E Ia,
x1x 2 ••• xk E lb. Then we have (cf. (3.6.5)),
(3.6.9)

where the accent means that it is necessary to sum in Ia, lb. First of all we need
to terminate the summation in (3.6.5) and (3.6.9)
n1 = exp (In In n)4 (3.6.10)
at the expense of a permissible error in the number of solutions. In fact, at the
expense of an error of Bn(ln n)- 1 k 12 in the total number of solutions of (3.6.2)
Sec. 6] REDUCTION OF THE ADDITIVE DIVISOR PROBLEM 77

we can suppose that each of xy and x 1 x 2 ... xk is repeated not more than
(In n)'k times. Under these conditions we estimate
2
n1 <<t~2no
2 N(xy -
a,b
x1x 2 ••• xk = q~), (3.6.11)
qeO,.

where a and b run through numbers under condition (3.6.6.). This number of
solutions does not exceed
(In n)2rk 2
n1 <a:s;;2n 0
2 N(X - Y= 0 (modq) IX - YI< 2n 0 ) I
qeOn

= B(ln n)2rknn 0 2 ! = B 0 nn 0(ln n)- 0 , (3.6.12)

n <q,.;;2no q
qeOn
by Lemma 1.1.6.
Thus, we can consider
Q0(Ia, 10) = 2
qEOn
µ(q)N(xy - X 1X 2 ••• xk = q~ xy I E Ia; X 1 ••• xk E / 0) (3.6.13)
q:=;;;;n1

instead of (3.6.9). It is convenient to rewrite the expression N(xy - x 1x 2 ••• xk =

I
q~ xy E Ia; x 1 ... xk E / 0) in (3.6.13), using trigonometric sums. Suppose that
for integer a,
S1(a, q) = 2 exp 27Ti ~ xy, (3.6.14)
ll:'VE]a q

S 2(a, q) = 2 exp 27Ti ~-x 1 ... xk. (3.6.15)

'1:1 • • • ZkElo q
Thus,
S 1(0, q) = S 1(0, 1) = 2 1; S 2(0, q) = S 2(0, 1) = 2 1.
ll:YEla Z1 • • • ll:kElo
In that case
X1X2 ••• Xk E/ 0)

1 q-1
= - 2 S 1(a', q)Sz(-a', q). (3.6.16)
q a'=O
Now note that for q1 q, (al> q1) I = 1,
S1(a1, q1) =22 exp 27Ti a 1 xy + Bn112. (3.6.17)
ll:'VEla q1
:c<11

Next, the sum S 2(a1 , q1) can obviously be represented in the form
S2(a 1 , q1 ) = k! 2 exp 27Ti ai x 1 •.. xk
"'1···"'kElo ql
~1< •.. <Zk

+ 2 Ck1 2 exp 27Ti al X1 ••• Xk,

(k1) 11:1..;; ••• ..;;a:k ql
78 THE ADDITIVE DIVISOR PROBLEM [CHAP. III

where k 1 runs through B values, Ck 1 = B, and there are k 1 equal signs in the
relations x1 ~ . • . ~ xk. lfin each such sum x 1 . • . xk-l :;5 0 (mod q1), then,
clearly, an estimate for the corresponding sum is B'J:,x 1 .:;;••• ~xk_ 1 l, where it is neces-
sary to take the appropriate number of equal signs among the X;. The last esti-
mate is of the form B,n1 - 1tk+•. Thus we have

(3.6.18)

where in summing the appropriate number of equal signs among the X; is taken.
Further, we find without difficulty, since µ(q) = 0,

+B -riq1) L 1, (3.6.19)
q 1 ln n x1~ ••• ~"'k

from which, considering (3.6.18),

eQ -rk(q 1) ln q 1
S2(a 1, q;,) = --! S 2(0, 1) + B S 2(0, 1), (3.6.20)
q1 q 1 ln n
where we have put
e
--':!=IT (i- (1--
l)k-1) . (3.6.21)
ql pjql p
In this connection, it is easy to see that

(3.6.22)

In particular, it is obvious and immediate that

(3.6.23)

Thus (3.6.16) becomes

1 1 q.{q 1)eQ
-L L S1(a1, q1)S2( -a1, qi) =- L 2 1 S 1(0, l)S2(0, 1)
q Q1IQ <a1,Q1l=l q Q1IQ qi
Sec. 7] ESTIMATION OF THE ERROR TERM 79

7. Further Transformations. Estimation of the Error Term

Let us return to the fundamental formula (3.6.13) and compute the errors
introduced into the summation by the second term on the right-hand side of
(3.6.24). We obtain the error

BSi(O, l)S 2(0, l) ln ln n ,L ! ,L Tk(q 1) ln q1

ln n q,,,;n1 q q1Jq q1

+ BS1(0, l)S2(0, 1) - 1-
2
,L ! ,L 7 k(qi)T(qi) In q 1 • (3.7.1)
(Inn) q,,,;n1 q q1Jq q1

Using Lemma 1.1.2 and formula (3.6.10) we find the estimate

BS1(0, l)S 2(0, 1) (In In n)K• (3.7.2)

In n
for (3.7.1). Thus, following (3.6.13) we have

(3.7.3)
Now let us examine the sum

(3.7.4)

It can be rewritten as

= .L µ(q 1~eq1 II
q1,,,;n1 ql :1>!'11
(1 - ~) P
.L µ(~').
(q',q1)=l q
(3.7.5)
0'1enn q'q1~n1

Consider those terms of the sum for which

q1 > (In n)10.
Obviously an estimate for the sum corresponding to such terms is
B(ln n)-8 •
Now let us consider the terms where q1 ~ (ln n) 10 • For such terms look at
the inner sum in (3.7.5) for fixed q1 •
80 THE ADDITIVE DIVISOR PROBLEM (CHAP. III

We have
I I µ(;'tJ) = !' µ(m), (3.7.6)
olql q'o,,;;;nl q (J m,,;;;nl m
(q' ,ql)=l

where the accent (') indicates that the summation is over elements of Qn.
Further, it is clear that for fJ I q1 ,

I' µ<;'tJ) = µ(tJ) I µ(:')

q'o,,;;;n1;<q' ,q 1 l=lq CJ (J q',,;;;n/o q
(q' ,(11)=1

An estimate for the last sum, by (3.6.10), (3.7.6) and Lemma 1.1.6, is
B(ln n)-10 • (3.7.8)
Thus from .(3.7.7) we find

II (1 - .!.) I' µ(:') = I' µ(m) + B-r(q )(1n n)-

1 8• (3.7.9)
Pl <1 1 p q',,;;;n 1 q m,,;;;nl m
(q' ,ql) =1

Putting this in (3.7.5), we obtain the expression

(3.7.10)

The inner series in (3.7.10) can be extended to oo and taken over all q; by
Lemma 1.1.6, this gives

(3.7.11)
Put
Ak = ~ µ(q;eq > 'YJk > 0. (3.7.12)
q=I q
After this (3.7.3) gives

QoUa, lb)= ( I' µ(m))s1(0, l)S2(0, l)Ak

m~nl m
+ BS1(0, l)SlO, 1) (In In n)K' , (3.7.13)
In n
where the accent in the summation means that m E Qn.
Sec. 8] THE FINAL FORMULA FOR THE NUMBER OF SOLUTIONS 81

8. The Final Formula for the Number of Solutions

We now need to add together the numbers of solutions of (3.7.13) over all
intervals la, lb under condition (3.6.6). By (3.7.13), it is sufficient to sum the
corresponding quantities

S1(0, l)S 2(0, 1) = ( li.1) ., 1

• :~i.e1bl. (3.8.1)

By definition of Ia and lb, when we sum (3.8.1) over Ia and lb we obtain an

expression of the form
L 1+B L 1,
0:11-0:1 .•• o:i.el'
(3.8.2)
10:11-0:1 ••• o:kl.;;;&n/H
xy<.n ::z:11<.n

where as before I' is the interval (3.6.3). The second term in (3.8.2) is easily
estimated as
(3.8.3)
and can be thrown out with a permissible error in the number of solutions.
Finally,
0:11-0:1
xy<.n
~. o:i.el'
1 = no/2.;;;o:y-t •.. °'k.;;;no 1 =
z.y~n
cl. ·~k.;;;n )On 1 n + B)no. <3·8.4)

Further, elementary considerations give

! 1 = k! (Jn n)k-l nSk, (3.8.5)
X1 . . . Xk<.n
where
-
sk -Yh-+moo (ln
. 1
Y)k-1
JJ ...
dyl ... dyk-1 (3.8.6)
l~Y1 <.v2<. .. ·. <.Yk-1<.Y IYk
Y1 · · · Yk-1

Turning to (3. 7.13) and substituting into it the expressions we have obtained,
we observe further that by Lemma 1.5.2,

n0 !' µ(m) = L(n) + BL(~)

, (3.8.7)
m ln n
m.;;;n1

where L(n) (see § 5, Chapter I) is the number of quasiprimes (3.6.1). By (3.6.4)

we find from (3.7.5) that
(3.8.8)
A remark about the derivation of an asymptotic series of type (3.1.5): we
must reason as before but with a more careful evaluation of the sums (3.6.14) and
(3.6.15). To do this it is sufficient, for example, to introduce sums over an entire
reduced residue system mod q.
<1> 27Tia 27Tia
Saq L exp--
= x,ymodq q
xy; <2>
Saq
"" exp -
= x;modq
"-" q
- x 1 • . • xk,
82 THE ADDITIVE DIVISOR PROBLEM (CHAP. III

and compare (3.6.14) and (3.6.15) with the approximate expressions

s~~ L
zye(l/q")Ia
1 and s~:> L
z 1 .•. zke(l/qk)Jb
1, (3.8.9)

where (I/qk)/b denotes the homothetic contraction of lb by a factor of qk (similarly

for (1/q2)/a).

9. Application of the Dispersion Method to Other Cases of the Additive

Divisor Problem
We have worked out in detail an asymptotic formula for ~m~nT(m + a)Tim)
in the case a = 1. Note that the same asymptotic formula is obtained when
a = -1. This can be seen by examining the work of the preceding sections.
If lal > 1, then as a system of numbers coherent to a (cf. § 1, Chapter II) one
needs to take the numbers ap', where p' is quasiprime, !n1-'o p < <
n1- '0 , with
e0 > 0 sufficiently small. Then one proceeds essentially as before. In applying
the dispersion method to the question of an asymptotic formula for
L T(n - m)Tk(m),
m~n-1

the system of numbers coherent to n can be constructed as in § 5 of the Intro-

duction, if condition (0.5.1) is satisfied. If it is not, then it is necessary to apply
the basic variant of the dispersion method described in § 2 of the Introduction,
or the variant which uses covariance, described in § 5 of the Introduction.
After rather unwieldy computations, which in principle do not differ signifi-
cantly from the work of this chapter, we can obtain asymptotic expressions
analogous to (3.2.2) and asymptotic series analogous to (3.1.5) with an error of
B,n(ln n)'.
If in (3.1.1) and (3.1.2) k 1 > 2 and k 2 > 2, application of the dispersion
method gives only conditional results for the present.
To be specific, consider the case a= 1 in (3.1.1); let 3 < k 1 < k 2 • The
fundamental equation
(3.9.1)

by the same arguments as in§§ 2 and 3, reduces to a number of equations of the

form
(3.9.2)
where 'VE (y); D' E (D), D' E Dia• in the notation of§ 3. The quasiprimes p' in
the interval [!n1- '0 , n1- '0 ] can again serve as the system of coherent numbers.
The three fundamental equations of the dispersion method (2.2.3) will have the
form
(3.9.3)
Sec. 9] APPLICATION OF THE DISPERSION METHOD 83

where the pair (ni. n2) of coherent numbers takes the values (1, 1), (n1 , n1), (ni. 1),
and the following condition is satisfied:

(3.9.4)

To proceed according to the outline of the dispersion method (described in

§§ 2 and 3 of the Introduction) we need to find the expected number of solutions
A(n, D), which does not present significant difficulties (although cumbersome
computation is required); in this case we obtain only one fundamental equation
for the dispersion method (for given vi. v2 ; v1 ¥- v2):

V1X1X2 ••• xk1 -

- 1 ·,
V2Z1Z2 ••• zk1 -
X1X2 ••• Xk1 - 1 E (D) • (3.9.5)
V2

Since k1 ~ 3, we do not at present have a way of finding an asymptotic

approximation for this equation. However we observe a characteristic feature:
the new equation (3.9.5) is similar to the original (3.9.1) and involves only the
number k1 ~ k 2 • If we could find an asymptotic approximation for the total
number of solutions of (3.9.5) (for v1 ¥- v2 ; vi E (v)), then we would have an
asymptotic approximation for the fundamental equation (3.9.1) for any value of
k2 ~ k1.
 CHAPTER IV

APPLICATION OF THE DISPERSION METHOD TO A

GENERALIZATION OF A THEOREM OF KLOOSTERMAN

l, Statement of the Problem

In 1926 H. Kloosterman [46], by supplementing the Hardy-Littlewood
method with particular ideas from the theory of theta-functions, obtained an
asymptotic formula for the number Q(n) of solutions of the binary problem
ax2 + by2 + cz2 + dt 2 = n. (4.1.1)
The left-hand side here is a definite quadratic form (a, b, c, dare positive). The
number n and the determinant of the form, ti = abed, are assumed relatively
prime. This formula was
Q(n) = ;K nS(n) + B,n 11118+<,
where
S(n) = L (!!.)t-1, (4.1.2)
tin t
( ~) is the Kronecker symbol.
Kloosterman's results were generalized by Tartakovskii [24], who con-
structed a general theory of the representation of large numbers by positive
quadratic forms F(x1 , ••• , x.) with s ~ 4 variables. Kloosterman's result for
(4.1.1) was obtained as a special case of the equation Q(x, y, z, t) = n, where
Q(x, y, z, t) is a positive quaternary quadratic form. These results were further
developed by A. V. Malysev, [18; 19], who applied the latest estimates for
trigonometric sums (given by A. Weil), and by other authors.
In 1954 M. Eichler [33] developed a new method in the analytic theory of
quadratic forms. This method is remarkable in that in the asymptotic formual
for the n.s.e. Q(x, y; z, t) = n, the error term is of the form B,n11 2+•, which is
essentially the best possible.
Aside from such generalizations, it is natural to look at equation (4.1.1) from
another point of view, and consider generalizations in another direction.
Equation (4.1.1) can be written in the form
<p1(x, y) + <p2(z, t) = n, (4.1.3)
84
 STATEMENT OF THE PROBLEM 85

where <p1(x, y) = ax2 + by2 and <p2(z, t) = cz 2 + dt 2 are positive binary quad-
ratic forms. Then one can consider arbitrary binary forms, <p1(x, y) and <p2(x, y),
with determinants (-ab) and (-cd) respectively, which, of course, leads to a
generalization of the preceding type.
Let ab and cdbe square-free: µ(ab) '16 0, µ(cd) '16 0. Consider the quadratic
fields k(.J -ab) and k(.J - cd), and classes of ideals 11 and 12 in them, such that
<p1(x,y) runs through the values of the norms of integer elements of k(.J-ab)
which are in 11 , and <p2(x, y) runs through the values of the norms of integer
elements of the field k(J-cd) which are in 12 • Then equation (4.1.3) can be
rewritten in the form
N(ii) + N(b) = n, (4.1.4)

where ii E 11 , b E 12 are elements of integral ideals of the respective fields.

If ab and cd are not square-free, (4.1.4) is not equivalent to (4.1.3). However,
the class of equations (4.1.4) can be extended so as to include those cases too.
For this it is sufficient to require that the elements of the ideals ii and b, expressed
in terms 9f the bases w 1 , w 2 , w~, w; of the respective ideals as x 1 w 1 + x 2 w 2 ,
x~w~ + x;w; (x;, x~ are rational integers), obey the additional condition: X; and
x~ must lie in prescribed arithmetic progressions. For example, for the equation
a2(x 2 + y 2) + c2(z 2 + t 2) = n we can consider the numbers x + yi and z + ti in
identical fields k 1 = k 2 = k(~) under the conditions: x = y = 0 (mod a);
z == t 0 (mod c). One can also consider ideals in "orders" (Ordnungen)
instead of in maximal rings of integers of fields; then forms with determinants
which are not square-free can also be admitted. But such an approach is more
complicated.
Just the same, equation (4.1.4) is only another way of writing (4.1.3). How-
ever, equation (4.1.4) may be studied without the assumption that k 1 and k 2 are
quadratic.
Let us consider equation (4.1.4), where ii E 11 andbE12 ; 11 is a class of ideals
of some algebraic number field k 1 , and 12 is a class of ideals of some other field k 2
(of course, k 1 can be equal to k 2 ).
Let d1 and d2 be the discriminants of the respective fields, and let (n, d1d2) = 1.
Thus stated, the problem is similar to the additive divisor problem (cf. (3.1.2)).
The dispersion method enables us to find an asymptotic approximation for
(4.1.4) for the case when one of the fields k 1 and k 2 is quadratic. We shall
suppose, for example, that k 1 is a quadratic field of discriminant d1 , and k 2 is a
field of arbitrary degree and discriminant d2 ; 11 and 12 are given classes of ideals
in these fields, and (n, 2d1d2 ) = I. Then, using the dispersion method, one can
obtain an asymptotic approximation for the number of solutions of (4.1.4) as
n ~ oo, with an error of Bn(ln ln n) 2/ln n. The computations are very com-
plicated for a number n of general form; they are simpler for an n which does
86 APPLICATION OF THE DISPERSION METHOD [CHAP. IV

not have small prime divisors. In that case, an asymptotic approximation for
(4.1.4) is
(ln In n) 2
Q(n) = G(kl> k 2 , n)n + Bn ,
In n
where G(k1 , k 2, n) is the singular series of the problem. For each given pair
(kl> k 2), where k 1 is quadratic and k 2 is any algebraic number field, one can
ascertain in a finite number of steps whether the main term is larger than the
error term; in the general case, though, it is even difficult to determine whether
G(k1 , k 2 , n) is bounded away from 0. In view of the fact that the result is so in-
complete and the computations so unwieldy, we shall not give the corresponding
asymptotic formula. However, we shall discuss the reduction of the fundamental
equation ( 4.1.4) to a form convenient for the dispersion method, since this is not
trivial. We suppose that k 1 is the imaginary quadratic field, and write the
fundamental equation in the form
cp(x, y) + N(ii) = n, (4.1.5)
where <p(x, y) is a positive primitive quadratic form of discriminant d, and N(ii)
is the norm of a number of the ideal of a given class of a given field k 2 •

2. Preparatory Transformations of Equation (4.1.5)

We shall bring equation (4.1.5) to a form convenient for the dispersion
method, proceeding as we did in § 3 with respect to equation (3.3.1). Consider
the set of norms of elements of a given ideal class I (a= I, 2, ... , h; his the
class number of the field k). For suitable k = k(k), we have
!
N(a)=m
1 = Brk(m), (4.2.1)

and, consequently (cf. Lemma 1.1.2),

! 1 = Bx(ln·x)1c-1 . (4.2.2)
N<a>.;;.x

Let the prime p0 exactly divide the number N(ii) (p0 II N(ii), cf. section on
notation). Then p 0 = N(TT0 ), where TT0 is an ideal of degree 1. This ideal is not
uniquely defined; there can be several different ideals with this property:
<
TT01 , TT02 , • • • , TTog (g a 0). The ideal TT0 exactly divides ii. In counting the num-
ber of repetitions of a given N(ii), we must take all these ideals into consideration.
Let N(ii) = p0 m 0 , (p0 , m0) = I; then if p = N(TT0 ), m 0 = N(ij), where ij is an
ideal and TT0ij E /.
Here if ii E /,then q- E //;1, 0
where/"'0 is the class of TT0 • Let an interval (v) be
given; consider the primes v E (v), ii E /,and the numbers N(ii) = vm0 < x such
that (v, m0) = 1. In order to count these N(ii) with the proper number of
Sec. 2] PREPARATORY TRANSFORMATIONS '87

repetitions, consider all classes of ideals / 1 , / 2, ••• , lh. Let P. be the system of
primes P E (P); P = N(TT .) ; 7T s E / 8 • For each one of these P, we take an ideal ij,
relatively prime to P, belonging to the class 11; 1 , such that N(ij) < x/p0 (if 7T s and
ij exist). Taking in turn all s, 'V E (P), and ij, we obtain all numbers N(ii) = Pm0
of the required form with the proper number of repetitions.
Let X run through the set of numbers of the form X = N(ii), ii E /, with the
proper number of repetitions; rewrite equation (4.1.5) as
<p(x,y) + X = n. (4.2.3)
Here X < n; each value of Xis counted BTk(X)times, in accordance with (4.2.1).
As in§ 3, Chapter III, we introduce the set Ak(n) of numbers X < n for which
(4.2.4)
Now note that
~ 1 = BT(m) = B ~ 1. (4.2.5)
'l'(:t,?l)=m :t11=m

Because of this, arguing as in the proof of Lemma 3.3.1, we obtain an analogue

of Lemma 3.3.l: an estimate for the number of solutions of (4.2.3) for which
XE Ak(n) is
(4.2.6)
Next, as in§ 3, Chapter III, separate the XE Ain) into classes D< 1 > and D< 2>.
The class Dm consists of the numbers X whose prime divisors 'V satisfy

exp ln n < P < n1 ; 0 < y < 0.01, (4.2.7)

10 ln ln n
and the class D< 2 > consists of the remaining XE Ain).
LEMMA 4.2.1. The number of solutions of (4.2.3) for which XE D< 2> is

n
B-(ln ln n) 2 • (4.2.8)
ln n
All the prime divisors of XE D< 2 > are either< exp (ln n/Kln ln n) or> nY.
Consider first of all the numbers of the first type. For them we can argue as in
the derivation of formula (3.3.9), using Lemma 1.1.6, (4.2.4), and (4.2.5). We
find that the corresponding number of solutions of (4.1.5) is

B-n-.
(ln n) 2
For numbers of the second type, TiX) = B; for these we need more precise
estimates than in § 3 of Chapter III. The number of corresponding solutions
will be
BN(<p(x,y) + P = n), (4.2.9)
88 APPLICATION OF THE DISPERSION METHOD [CHAP. IV

where P is an "almost prime" number, all prime divisors of which are greater
than nY. An analogue of Lemma 1.3.2 gives the estimate

B ..!.:__(In In n) 2 (4.2.10)
In n
for expression (4.2.9). This proves Lemma4.2.1. We now suppose that XE D<ll,
so that X = vD'. Let p(X) be the number ofrepetitions of the numberX = N(ii).
If v 1 D', then p(vD') = p(v)p(D'). Arguing as in the derivation of formula
(3.3.12), we see that one can neglect those solutions X which have a repeated
factor v of the type we are considering.
Plainly, the number of divisors of type v of the number X does not exceed
r < 10 In Inn. (4.2.11)

Each divisor v = N(TT,), where TT sis a prime ideal of degree 1 ; as we already have
seen, TT 8 is not determined uniquely, so vis counted the appropriate number of
times. Let a be the number of divisors of type v of the number X; a< r. Thus
X = v1 v2 ••• vaD", where D" has no divisor of this kind; here vi= N(TT 8 .),
where si is the index of the class of ideals I •. to which TT•· belongs. Write a as' a
sum of numbers ' '
(4.2.12)

Here a 1 , a2 , ••• , ah are the numbers of numbers v; for which v; = N(TT), TT EI;;
(of course, repetitions of the same v; are possible for different subdivisions
(4.2.12)). With each such subdivision (4.2.12) we associate a class of numbers X

(4.2.13)
Here, as was mentioned, the same X can belong to different classes (4.2.13);
it will be repeated exactly the necessary number of times. Obviously, the number
of classes ( 4.2.13) does not exceed ha and, by (4.2.10), does not exceed
(4.2.14)

Consider XE D~~>, ... ,ah• where a 1 + ... + ah = a > 0. We have the repre-
sentation

where for a; > 0, TT;i E I; (i <

a;); ij belongs to a definite ideal class and is
counted a definite number of times. Taking in turn all possible classes (4.2.13),
we obtain the necessary number of repetitions of X. Note that N(ij) must not
have divisors of type v. We can write the number X as

X=vD', (4.2.16)
Sec. 2] PREPARATORY TRANSFORMATIONS 89

where v = N(TT;i) (for a1 > 0) is counted once, and D' is obtained from (4.2.15)
when the factors v are separated off. Then each needed representation (4.2.16)
is repeated a times. Hence the n.s.e. (4.2.3) is equal to

L -1 ""
(l)k
N(ip(x, y) + vD' = n), (4.2.17)
ai. ... ,a11, a X eDa1,a2 .... ,a11.

where D' < n/v; D' is of the form described above and is repeated N(ij) times,
where ij is the ideal obtained in the manner described earlier. We see that the
boundaries of the variation of D' depend on v; v must not divide D'; the
composition of D' is determined by the class 11 such that v = N(TTu). We are
interested in making D' as much as possible independent of v. We shall keep the
condition D' < n/v; the condition v -r D' can be dropped with a permissible
error in the number of solutions, as was made clear above. Next we divide up
the v = N(TT;i) according to the value of j (with a1 > O); let v<i> denote the
numbers v with identical j (of course, there can·be equalities among the numbers
v<fl for different classes (4.2.13)). Then for fixedj, only the boundary of variation
of D' will depend on v<fl. Separate the equations
ip(x,y) + vD' = n (4.2.18)

into classes of equations of the form

+ v<fl D' = n,
ip(x, y) (4.2.19)

where D' is a number of the described form, D' < nf'llU>, but the values of the D'
and the number of them do not otherwise depend on v< 1>.
Now we proceed as in § 3 of Chapter III after the derivation of formula
(3.3.15). We divide the interval (4.2.7) into intervals of variation of v< 1> of the
form
(v)=[v0 ,v0 +v~]; v~= VoK· (4.2.20)
(Inn) 1
with a sufficiently large value of K1• Consider v<i> E (v) in (4.2.19). Using
(4.2.20), (4.2.4), and (4.2.14), and arguing as in § 3 of Chapter III after the
derivation of (3.3.15), we find that in an equation of form (4.2.19), when v E (v)
we can suppose that D' <
n/v0 , that is, the boundary of its variation does not
depend on v<i>; the error in the total number of solutions produced thereby will
be permissible.
Further, as in § 3 of Chapter III after (3.3.27), we find that the last incomplete
interval of the form (v) can be discarded with a permissible error in the number
of solutions. Thus we come to equations of form (4.2.19), where D' is a number
of the form described above with the indicated number of repetitions, and
D' < n/v0 ; D' does not depend on v<i>. We can now apply the dispersion
90 APPLICATION OF THE DISPERSION METHOD

method. In this connection the distribution of the primes vU> is important;

these numbers are the norms of ideals of a prescribed class, in arithmetic
progressions with a common difference which can increase as·some power of the
logarithm of the length of the interval in which the ,,u> lie. Here a generalization
of a theorem of Siegel, given by R. Brauer (cf. [29, Theorem 2, p. 247]), plays a
fundamental role.
 CHAPTER V

ON A CERTAIN BINARY EQUATION OF GENERAL TYPE

INVOLVING A QUADRATIC FORM

1. Statement of the Problem

Let <p(x, y) = ax2 + bxy + cy2 be a positive primitive binary quadratic form
of discriminant -d = b2 - 4ac < 0, d being free of squares with the possible
exception of the number 4.
Consider the binary problem
(5.1.1)
Here g is a given prime (or g = l); a 1, a 2 are given numbers; the numbers D'
and v are independent; D' ~ n1-Y; vK ~ nY; y is a constant,
0.01~y>0. (5.1.2)
We shall suppose that v runs through primes, and that the numbers D' run
through some system of numbers such that there are at least

(5.1.3)

of them in the interval [I, n1-Y). Without loss of generality we can suppose that
K1 ~ 1000. If p(D') is the number of repetitions of D',

p(D') ~(In n)K 2; K 2 < ~~ (5.1.4)

may be admitted.
Further, we shall assume some regularity in the distribution of the numbers
D' in arithmetic progressions with a small common difference. To facilitate the
later derivations, we make two different assumptions.
Let K 3 ~ IOOK1 ; q ~(In n)K•.
I. 2
D'=<(modq)
1= !(
q
2
D'e(DJ
1) (1 + (In Bn)Ka) (5.1.5)
D'e(D)
for any remainders mod~.

II. 2
D'::g(modq)
1= ~(
<p(q)
2
D'e(D)
1) (1 + (In Bn)K2), (5.1.6)
D'e(DJ
where E~ = 1 if(~, q) = I, and = 0 otherwise.
91
92 A BINARY EQUATION INVOLVING A QUADRATIC FORM [CHAP. V

We remark that with the dispersion method it is possible to obtain an

asymptotic expression for the number of solutions of (5.1.1) under more general
assumptions than these. Assumption I corresponds to the behavior of the
positive integers, and II to the behavior of the primes. A supposition about the
difference of the behavior of D' = ; (mod q) for(;, q) = 1 and(;, q) > 1 would
be more general.
Further, it would be expedient to single out a limited number of exceptional
numbers q1 (j = 1, 2, ... , c0), relative to which some special behavior of D' is
assumed. However, for simplicity's sake we shall only make Assumptions I
and II. It may be observed that the numbers"' can also be made to run through
rather arbitrary sequence.
Supposing that (n, a1a2d) = 1, we can obtain an asymptotic approximation
for n.s.e. (5.1.1) using the dispersion method. However, in the general case very
unwieldy computations are required. They are carried out in their entirety in
Chapter VI for the particular case a 1 = a2 = l, g = 1; D' and 'V prime; in this
case no restrictions are imposed on the variation of D' and "'(this is necessary
for the treatment in Chapter VII of the Hardy-Littlewood equation). In this
chapter we limit ourselves to the construction of the function A(n, D) (cf.
(0.2.4)), and to the statement of the final result, for one characteristic particular
case of equation (5.1.1).

2. On One Particular Case of Equation (5.1.1)

We shall consider the particular case of(5.1.1) wheng = 2, a 1 = 1, a 2 = 2d;
D' runs through primes. Thus we shall examine the equation
<p(x,y) + 2dD'y2 = n. (5.2.1)
In order to simplify the construction of A(n, D) and the form of the final
formula, we assume also that n has no prime divisor <
(ln n)K• for sufficiently
large K4 •
We now construct A(n, D), the expected number of solutions of the equation
<p(x, y) + 2dDY2 = n, (5.2.2)
<
where Dis fixed and y2 nY; "'runs through primes. By the assumption about n,
(n, 2d) = I. If (m, 2d) = 1, then, as is well known (cf. Lemma 2.14.3), the
number ofrepresentations of m by the system of binary forms <p with discriminant
-dis
w1 (- ~) = w 1xib), (5.2.3)
olm c5 oln
where w = w(d) is described in Lemma 2.14.2. Leth be the class number of the
forms <p, so that (cf. Lemma 2.14.3)
h = w.Jd L(l, Xd). (5.2.4)
27T
Sec. 2] ON ONE PARTICULAR CASE OF EQUATION (5.1.l) 93

It is natural to seek A(n, D) in the form of an approximation to the expression

~2 2 .xio), (5.2.5)
h r•> oln-2dD•

where (v) is the interval of values of v. Here it is important that if o! n - 2dDv 2 ,

then (o, 2d) = 1, by the conditions on n.
Because of (5.2.4), (5.2.5) can be replaced by

27T 2 2 xio). (5.2.6)

.Jd L(1, Xd) (v) clln-2dDv 2

o
Note that may take on values ~ n. For the construction of A(n, D), we shall
o
assume at first that the of (5.2.6) is less than or equal to (ln n)c, because then
the distribution of the numbers v ~ nY1 2 in progressions is well known. Let
o = p~ip~· . .. p~· ~ (ln n)°. For C < K 4 the congruence
2dDv 2 = n(mod p~i) (5.2.7)
is solvable if and only if
(p;, 2dD) = 1 (5.2.8)
and

(2d:n) = +1. (5.2.9)

Under condition (5.2.8) the number of solutions of the congruence

is
2dDe =n(mod p~i) (5.2.10)

1 + (2d:n). (5.2.11)

If in (5.2.7) v runs through the primes ~ nY12, the number of which we denote by
L2 = L 2(n), (5.2.12)
then the main term of the number of solutions of (5.2.7) is

~(l + (2dDn))· (5.2.13)

<p(p~·) P;
Thus, when
(o, 2dD) =I (5.2.14)
the main term of the number of solutions of the congruence
2dDv2 ::: n(mod o) (5.2.15)
IS

~IT (l + (2dDn))· (5.2.16)

<p(O) Piel p
94 A BINARY EQUATION INVOLVING A QUADRATIC FORM

Proceeding from this, we can take

A(n, D) = 27TL2 I xa(c5)rr( 1 + (2dDn)). (5.2.17)

.Jd L(l, Xa) Co,2dDl=l q:i(c5) p
We may then further transform (5.2.17) so as to express it in terms of an ab-
solutely convergent series. To do this it is sufficient to introduce the product

1 TI (1 xiP)) IT (1 xiP))
L(l, Xa) ii12aD p i>f2aD p
(see Chapter VI, wher:e detailed computations are carried out for the case
q:i = x 2 + y2, g = 1 and n of general form).

3. The Final Formula for the Number of Solutions of (5.2.1)

Using (5.2.17), we can foi:m V3 and V2 and obtain asymptotic approximations
for them (cf. (0.2.9)). Next, using Lemma 2.14.1, we find after rather unwieldy
computation, omitted here, that the dispersion V' (cf. (0.2.5)) is small, whence, by
the analogue of the Cebysev inequality, we arrive at a final formula for the
number of solutions Q(n) of equation (5.2.1). We state it here. Let L 1 be the
number of primes D' ~ n1-Y; then

(5.3.1)
where
Aa = rr(1 +
p
xa(p) )
p(p - 1)
(5.3.2)

and C is an arbitrarily large constant.

In the general case, when n has small prime divisors, the formula for Q(n) is
more complicated.
Note that at the present time neither this result nor a solution of the general
equation (5.1.1) can be deduced from the extended Riemann hypothesis.
 CHAPTER VI

ANALOGUES OF A HARDY-LITTLEWOOD EQUATION

1. Statement of the Problem

This chapter is a continuation of Chapter V and in essence treats the particu-
lar case of (5.1.1) where cp(x, y) = x 2 + y 2 • But here we put no limitations on n,
the number to be represented; proceeding according to the basic scheme in § 2
of the Introduction, we carry out the construction of A(n, D) in detail.
We shall examine two equations, which are analogues of the equation of
Hardy and Littlewood to be studied in Chapter VII. We consider an asymptotic
formula for the number Q 1(n) of solutions of
n = x2 + y2 + PJ.P2, (6.1.1)
where Pi. p 2 run through all primes such that
Pi> exp (ln ln n) 2 (i = 1, 2). 1
The second equation is of a more complicated nature and is auxiliary to the
solution of the Hardy-Littlewood problem (cf. Chapter VII).
Assume that n, the number to be represented, is sufficiently large;

P =exp ( In n In In In n) , (6.1.2)
K In Inn
where K ~ 100 is a sufficiently large number. Let QP be the set of all numbers
whose prime divisors are all greater than P. Consider the equation Y;
n = x~x; . .. X~ + e + r/; x; E Qp, (6.1.3)
where k ~ 6. Clearly,
6 ~ k ~
[K In In n].
71 = (6.1.4)
ln In Inn
The purpose of this chapter is to prove the following two theorems.
THEOREM 6.1.1. The number Q1(n) of solutions of equation (6.1.1) has the
asymptotic expression
( ) "-'TT A on In Inn
Q in IT (p - l)(p - X4(p))
(6.1.5)
2 '
Inn Pin p - p + xiP)
1 This condition was mistakenly omitted in the author's paper [14].
95
96 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

where

Ao=II(1+ XlP) ). (6.1.6)

p p(p - 1)
Note that in (6.1.5) (in shortened notation)

II = B In In n; (II)-1 = B In In n, (6.1.7)
Pin Pin

which gives an idea of the behavior of Q(n) as n---+ oo.

Let Q~(n) denote the number of solutions of (6.1.3) for given k under condi-
tion (6.1.4). Further, let
L~(n) = L 1. (6.1.8)
z'1· ·-~~~n
x'ieO:p

THEOREM 6.1.2. When k ;;;i. 7, k ~ r 1, the number of solutions of equation

(6.1.3) has the form
Q~(n) = 1TAo II (p :- l)(p - X (p)) L~(n) + Bn(ln n)-
4 500 • (6.1.9)
Pin p - p + xlP)
(Instead of (In n)-500 we could have put B 0 (ln n)- 0 for any C > 0.)
When k = 6,
Q~(n) = 1TAo II (p - l)(p - xlp)) L~(n) + B n(ln In ~)2. (6.1.10)
Pin P2 - p + xlP) (In n)3 2

2. Reduction of the Equations to a Form Convenient for the

Application of the Dispersion Method
We shall now bring equations (6.1.1) and (6.1.3) to a form to which the
dispersion method may conveniently be applied. This is not difficult to do for
(6.1.1); but for (6.1.3) the transformations are more tedious, especially for the
case k = 6. Let us begin with (6.1.3). For given k, (6.l.3) has the form
n = X' + ~ 2 + YJ2, (6.2.1)
where X' = Xix~ . .. xfc must be counted Tfc(X') times. Here Tfc(X') = ~x~ ...:r~=X'l;
x[ E Op (i = 1, 2, ... , k). We give an estimate for Tfc(X'). Plainly,
(6.2.2)
With a permissible error in the number of solutions (namely, an error of the
form B 0 n(ln n)- 0 ) we can suppose that in (6.2.1) X' is square-free. For all of its
prime divisors are at least as large as P, and therefore the n.s.e. (6.2.1), when
µ(X') = 0, is (cf. Lemmas 1.l.3 and 1.1.l)
B(ln n) 2K ! ! T(n - p 2 ~) = Bn exp (In n)112 • (6.2.3)
p,;;;,p,;;;,vn p2;,;;;,n-1
Sec. 2] REDUCTION OF EQUATIONS 97

Now consider the minimal prime divisor 11 of each X'. The obvious inequalities
p < 11 ~ nl!k ~ nits (6.2.4)
will be very important in what follows. Thus, each of our numbers X' is
repr~sentable in the form
X' = 11D'; D' ~ ~. (6.2.5)
'II

Now let us see what the D' are for given 11 and how many times they are
repeated. We see that each prime factor of D' must be greater than 11; D' E Op;
µ(D') ':;/= O; D' must have at least (k - 1) prime divisors. Let -r'fc(D') be the
number of repetitions of the number D'. Then -r'fc(D') = -rfc(D'v), but it does not
depend on 11. That is, by Lemma 1.1.10,
-r~(D'11) = k{-r~(D') + Tk_ 1 (D')) = -r;(D'). (6.2.6)
Thus the dependence of D' on 11 consists only in that D' ~ n/11 and that all
prime divisors of D' are larger than 11. To do away with this dependence with a
permissible error in the number of solutions, we divide the interval (6.2.4) of 11
into subintervals
(11): (110 - 'II~, 110]; 'II~= 'llo K (6.2.7)
(Inn) 1
(the constant Ki will be fixed later on), beginning the division from the right end
of interval (6.2.4).
We shall begin with the case k > 7, and then consider the special case k = 6.
When k > 7, 11 ~ n11 7 • Note that here the extreme left interval (6.2.7) is in-
complete. We shall prove that it can be thrown out with a permissible error in
thenumberofsolutions. Considertheequationn = 11D' + g2 +17 2,whereD'is
repeated -r'fc(D') times, v E [P, P/(ln n)K1-i]; let us estimate the number of its
solutions. An estimate for this (cf. (1.1.7)) is
B(ln n) 2K 2 -r(n - vD) = B(ln n) 2K+ 2 2 ~, (6.2.8)
D<nlv (v) 'II

where v belongs to the indicated interval.

If Ki> 4K, we shall assume that (6.2.7) gives
Bn(ln n)-<1/4JK1. (6.2.9)
Now we show that the condition that all prime divisors of D' be greater than
v can be replaced by the requirement that all prime divisors of D' be greater than
110, at the expense of a permissible error. Suppose that D' obeys the second but
not the first of these conditions. Then D' has a prime factor p in an interval
(6.2.7). The corresponding equation has the form n = g2 + 172 + vpD", and an
estimate for the number of its solutions is
B(ln n) 2K 2 -r(n - vpD) = B(ln n) 2K.; (v~) 2 = Bn(ln n) 2K- 2K 1•

qD<n ~
98 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

Summing this estimate over all intervals (11), the number of which is obviously
B(ln n)Ki+i, we obtain the total estimate
Bn(ln n)-<112>K1. (6.2.10)
We can replace the condition D' <
n/11 by D' <
n/110 in the same fashion.
That is, if n/11 ~ D' ~ n/110 , then we have the following estimate for the number
of solutions with 11 E (11):
12
B(ln n) 2K 1 T(n - 11D) = Bn(ln n)2K+l ~.
n/(v 0 -v 0>;;.n•;;.nf•o 11~

Summing this estimate over all (11), we obtain

11 12 11
Bn(ln n) 2K+i-%- ~ ln n = Bn(ln n)-< 112 >K 1, (6.2.11)
111 Yo
which is a permissible error in the number of solutions.

3. The Case k = 6
In the case k = 6 we need to separate from the interval of values of 11,
p < 11 <nits, (6.3.1)
one part which hinders subsequent calculations. This part is
n116 exp -(ln n )112 < 11 < n116. (6.3.2)

Consider the class A' of numbers X' of the form (6.2.5), where 11 belongs
to the interval (6.3.2). All the other prime divisors of X' are at least as large
as 11, so X' has < 6 prime divisors, one of which obeys (6.3.2), and Ti(X') =
T6(X') = B. Let us calculate the n.s.e. (6.2.1) for such numbers. This reduces
to finding the n.s.e.
;z + 'Y/2 + vy = n, (6.3.3)

where y has <5 prime divisors. By Lemma 1.3.2, we obtain the number of
solutions B(n/11)(ln In n) 2/ln n for a given prime 11 under condition (6.3.2).
Summing over 11 under condition (6.3.2), we obtain the total estimate

(6.3.4)

Such an error in the total number of solutions will be permissible, so we shall

suppose that fork = 6, the complement of (6.3.2) in (6.3.1) is the basic interval
of variation. Of course, all the results of§ 2 remain valid for a solution where 11
is in this interval.
Sec. 5] DETERMINATION OF A(n, D) 99

4. Transformation of Equation (6.1.1)

We can suppose that p 1 < Pz in (6.1.1) and then double the number of solu-
tions atthe expense of a permissible error. PutPi = v;p 2 = D'; then we obtain
the equation
n = ~2 + r/ + v D'; v ~ n11 2; D' > v. (6.4.1)
Let oc0 be a given, arbitrarily small constant. By Lemma 1.3.3, the n.s.e. (6.4.1),
on the condition that exp (In In n) 2 ~ v ~ exp (Inn)"•, does not exceed

B ...!!:__ OCo In Inn II (p ~ l)(p - X4(p)). (6.4.2)

Inn Pin p - p + xiP)
Further, by the same lemma, the n.s.e. (6.4.1), on the condition that v > n117,
does not exceed
B ...!!:__ II(p ~ l)(p - XiP)). (6.4.3)
Inn pJn p - p +
X4(p)
We shall consider the error (6.4.2) (and hence (6.4.3)) permissible in solving
(6.1.1). Thus in (6.4.1) we can suppose

exp (In nt" ~ v < n117 ; (6.4.4)

with this condition, the arguments of§ 2 are fully applicable to (6.4.1).

5. Determination of A(n, D)
Let us now turn to equation (6.1.3); then the examination of(6.1.1) will not
differ substantially, but we shall still return to it. On the basis of the results of
§§ 2 and 3, we have reduced the problem to the study of
n = ~2 + r; 2 +'PD', (6.5.1)
where

'I'' = Vo . p ~ 'I' ~ n11s exp -(In n)112

o (In n)K1 ' """ o """ '
(6.5.2)
D' is repeated T~(D') = B(ln n) 2K times; D' ~ !!:.. .
Vo
D' varies independently of v; all the prime divisors of D' are greater than v0 ,
and they are all different.
Let L 1 be the number of primes v E (v).
We shall apply the dispersion method to equation (6.5.1) in accordance with
the fundamental lemma of§ 2 of the Introduction, which includes the determina-
tion of A(n, D). In (6.5.1) the number 'PD' is odd, and ifn is even, then ~2 + r; 2 is
100 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

odd. We begin with the case n even. Let D be a prescribed odd number,
D ~ n/v0 • Consider the equation
n = e + 'Y} + vD,
2 2 (6.5.3)
where e + 'YJ
2 2 is odd. The n.s.e. (6.5.3), as is well known, is

4 !xiq)
g,,;n n-Dv=O(modg)
2 L (6.5.4)

For the heuristic determination of A(n, D), we terminate the summation over q
with q > p 1 ' 3 , and for q ~ p 1 ' 3 we proceed as if we knew that the number of v
e
such that v E (v), v = (mod q), (e, q) = 1, were given by a formula with main
term
~ (6.5.5)
cp(q)
It may be observed that in doing this we are acting as if we were certain of the
truth of the extended Riemann hypothesis. But in finding A(n, D) any heuristic
means are applicable, since a choice of A(n, D) will be justified only if the dis-
persion V' (cf. (0.2.6)) is small, permitting us to proceed with the dispersion
method.
Next we formally extend to infinity the series obtained, containing q, and
take the resulting expression as A(n, D). Let us examine the construction of
A(n, D) more closely. Let
n = 2"pf•p:• ... pf' (µ > 0) (6.5.6)
be the canonical factorization of n. For each Pi (i ~ t), let tl.i ~ 0 be such that
p~iii D. (6.5.7)
The set of numbers {tl.1, tl. 2, ••• , tl.t} enters into the definition of D. Consider
the congruences
n - Dv = 0 (modqP) (p > 0); qP < p 113 (6.5.8)
for prime q. We can suppose q is odd, because n - Dv is. If (Dn, q) = 1, then
compare the expression
(6.5.9)
with (6.5.8).
I
Ifq D, q -r n, then (6.5.8) has no solutions. Ifq D, q I
then, in accordance In,
with (6.5.6) and (6.5.7), let q =pi, tl.i > 0.
Let tl.i < pi; then when p ~ tl.i, compare (6.5.8) with the number Li_; if
p > tl.i, the congruence has no solutions.
If tl.i > Pi• then when p ~ Pi• compare the congruence with Li; for p > Pi
there will be no solutions.
Sec. 5] DETERMINATION OF A(n, D) 101

For example, let Pi= Ai. For p ~ Pi• compare our congruence with the
number L1, and for p > Pi• with the number

Li - Li (6.5.10)
<p(qP-Pi) - qp-p;-i(q _ 1) .

Multiply (6.5.9) by xiqP) = (xiq))P, sum from p = 1 to p = oo, and insert a

factor of L1 • After an elementary computation we obtain

Li(1-xiq))-i(1
q
+ xiq) ).
q(q - 1)
(6.5.11)

I
For the case q D, q -r n, only the term L1 will be involved in the formation of
A(n, D); it is convenient to write L 1 in the form

L
i
= Li (1 - xiq))-i(1
q
+ xiq) 1))<1 + ~)
q q - ( q '
(6.5.12)
where
1 + ~ = (q - l)(q - xiq)) = 1_ qxiq) (6.5.13)
q q 2 - q + xiq) q2 - q + xiq)

Next, for Ai < p, in accordance with the above remarks, the number q is
compared to the expression
(6.5.14)
for Ai > Pi it is compared to

and for Ai = Pi• to

Li(l + xiq) + ... + x~'(q)); (6.5.15)

Li ( 1 + x4(q) + ... + x~'(q) + x~•+i(q) + x~'+ 2(q) + ...)

\ q - 1 q(q - 1)

= Li (1 + x4(q) + ... + x~'(q) + x~'+i(q) · q ) .

(q - l)(q - xiq))
(6.5.16)

It is convenient to insert the trivial factor

1= (1 - x4(q))-i(1
q
+ xiq) )<1 + ~q),
q(q - 1)
where 1 + ~q is defined in (6.5.13).
Now observe that

:: =
4
rr(1 -
q>2
xiq))-i,
q
Ao=
P>2
IT (1 + X4(q) )
q(q - 1)
(6.5.17)
102 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

(cf. (6.1.6)). To obtain the expression for A(n, D), we multiply together Li and
the coefficients of Li in (6.5.11), (6.5.12), (6.5.14)-(6.5.16), depending on the type
of prime q, over all primes. Thus we suppose
A(n, D) = 7TAcL1 II (1 + ~ 1,) II (1 + ~P)(l + 17(pi, Ai)), (6.5.18)
PID p;ln
p-f'n
where
I+n<pi,ai)
1 + x4(pi) + ... + xNPi) for A; < p;,
= { 1 + xiP;) + ... + xNP;) for A;> p;, (6.5.19)
. P·XP;+l(p-)
1 + xiP;) + ... + x~;(p;) + ' 4 ' for A; = P;·
(p; - l)(p; - XiP;))
Thus the construction of A(n, D) for even n is complete. We see that it is rather
unwieldy even for the present case of the simplest kind of quadratic form,
<p = x2 + y2.
Now we consider the case n odd. In this case ~2 + 'YJ 2 in (6.5.1) must be even.
For prescribed A = 1, 2, 3, consider
n = (e + 'Y] )2;,. + vD, e + 'Y]
2 2 odd, (6.5.20)
instead of (6.5.3); here Dis odd; v and D satisfy the conditions introduced
earlier. For this case we introduce
Al(n' D) = A(n,;,.
2 D)
(6.5.21)

in place of A(n, D) (cf. (6.5.18)).

6. The Dispersion of the Number of Solutions. Calculation of V1

Following § 2 of the Introduction, form the dispersion V' of the form (0.2.5)
for equation (6.5.1) (with n even). The summation in the fundamental inequality
(0.2.6) is taken over all odd D E [D1, D1 + D 2]; the number ofrepetitions of Dis
estimated by (6.2.2).
According to (0.2.7), we need to obtain an asymptotic approximation for
V1 , V2 and V3 , and then compare V with V3 •
Let us seek an estimate for the sum of terms corresponding to vi = v2 in the
expression for Vi (cf. (0.2.8)). Applying Lemma 1.1.4, we get (cf. (1.1.5))
L L (U(n -
D 1 .;;;D.,;;D 1 +D 2 •E(v)
Dv)) 2 = B L
vE(v)
L
D, .;;;D.,;;D 1 +D2
(r(n - Dv)) 2
Dodd

= BD 2v~(In nr< 2 •2 >

= BDh~)2 exp -(In In n) 2 , (6.6.1)

Sec. 6] DISPERSION OF THE NUMBER OF SOLUTIONS 103

which will be a permissible error term in the estimate for the dispersion. The
fundamental equation of the dispersion method (cf. (0.2.14)) for vi -:;6 v2,
(viv2, n) = 1,
vi(x2 + y 2) - v2(z 2 + t 2) = n('1'i - v2);
x2 + y 2 = z 2 + t 2 = 1 (mod 2) (6.6.2)
automatically guarantees that (n - (x2 + y 2))/v2 is an integer which, obviously,
is odd; it must belong to (D).
Note that since x 2 + y 2 and z 2 + t 2 are odd, the left-hand side of (6.6.2) is
= vi - v2 (mod 4), whence (n - l)(vi - v2) = 0 (mod 4), and since n is even,
vi - v2 = 0 (mod 4). By § 3, vi ~ n116 exp -(ln n)112 . In view of this, by
Lemma 2.10.1 of Chapter II, the number of solutions of (6.6.2) is equal to

8D2 L ~ + B 0 Dlln n)- 0 . (6.6.3)

c5J n(v1-•2> <5
6odd

We have discussed the particular case of even n. If n is odd, we consider

equation (6.5.20) for given A.. Here the basic sequence {qi} = {~ 2 + 172} (~2 + 172
odd) is replaced by {qi} = {2A(~2 + 172)} (~2 + 172 odd), so that instead of (6.6.2)
we obtain
v12\x2 + y2) - v22\z 2 + t 2) = n(v1 - v2);
x2 + y 2 = z 2 + t 2 = 1 (mod 2) (6.6.4)
under the appropriate conditions on (n - 2\x2 + y2))/v2. From (6.6.4)
we find that 2\vi - v2) = n(vi - v2) (mod 2A+ 2), and since n is odd, Yi - v2 =
O (mod 2A+2). From (6.6.4),

(6.6.5)

where the right-hand side is divisible by 4 and (n - 2A(x2 + y 2))/v2 E (D).

We now show that for sufficiently large C and

(6.6.6)
the number of solutions of(6.6.5) can be neglected with a permissible error in the
number of solutions. For given vh v2, the n.s.e. (6.6.5) is estimated by
B~ T(X}r(Y), where viX - v2 Y = n(vi - v2)/2A; (n - 2AX)/v2 E (D). Here Xis
in an arithmetic progression with common difference v2. Further,
B~ T(X)T(Y) = B~ T 2(X) + B~ T 2(Y).
I
Apply Lemma 1.1.5, taking into account the fact that 2A vi - v2, Yi -:;6 v2, so
that 2A < nits. Hence we obtain
B D 2 (In n)4< 2, 2 > (6.6.7)
2-<
104 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION (CHAP. VI

as an estimate for the required number of solutions. Summing this estimate over
all v1 , v2 such that 'I'; E (v), v1 ":/:- v2, v1 - v2 = 0 (mod 2;.), we obtain

Summing this expression over all A under condition (6.6.6), we obtain the total
estimate
(6.6.8)
If
2;. .,;;; (ln n) 0 (6.6.9)
(the opposite of (6.6.6)), then we again apply Lemma 2.10.1 to equation (6.6.5),
taking into consideration the fact that v;.s;;; n116 exp -(ln n)11 2. The number of
solutions of(6.6.5) under the relevant conditions, in particular, for (vi. v2 , n) = 1,
is
(6.6.10)

Now we need to sum (6.6.3) and (6.6.10) over the relevant values ofv1 and v2 •
We begin with (6.6.3) (son will be even). Let the canonical factorization of n be
given by (6.5.6). Formulas (6.6.3) and (6.6.10) are not valid for 1 v In v In;
or 2
since v; > exp (ln n)"o (cf. (6.5.2) and (6.4.4)), there will be B ln n such values of
"';· When we sum over v; in (6.6.3) and (6.6.10), these values contribute a total
error of the form (6.6.8), as is easily seen. On the other hand, when v; In,
the
number of solutions of (6.6.2) under the corresponding conditions, by Lemma
1.1.4, does not exceed BD2(ln n)a<2.2>, and the total number of solutions over all
v; In, on the basis of the above, does not exceed

(6.6.11)
an analogous estimate holds in the case of equation (6.6.5). Thus in (6.6.3) and
(6.6.10) we can sum over all v1 , v2 ; v1 ":/:- v2, v; E (v), v1 - v2 = 0 (mod 4) in the
case (6.6.3), and v1 - v2 = 0 (mod 2.i.+ 2) in the general case.
Put M = (v1 - v2 )n. By Lemma 1.1.7,

~ ! = B 0 (ln n)- 0 . (6.6.12)

olM c5
o> exp(lnlnn) 3

Therefore (6.6.10) (for A~ 0, i.e., including (6.6.3)) can be written in the form

(6.6.13)
Sec. 6] DISPERSION OF THE NUMBER OF SOLUTIONS 105

Let <\ <exp (In ln n)3 be an odd number. Consider the number of pairs
v1 , v2 such that vi E (v); v1 -:;if. v2 ;

(6.6.14)

If Cl1 < (ln n)°1 (C1 is an arbitrary constant), then the number of pairs (6.6.14) is

;. L~ (1 + Bo10n n)-01) (6.6.15)

2 +l<p(Cl1)

(the well-known noneffectiveness of this formula can be circumvented at the

expense of complicating the arguments, cf. Vinogradov [7, pp. 321-322]).
Further, by Lemma 1.3.1, when (In n)° 1 < Cl1 <exp -(In In n) 2 , we have the
estimate
BL~
(6.6.16)
2Hl<p(b1)

for the number of pairs in (6.6.14). If Cl (which henceforth is odd) divides

n(v 1 - v2), then Cl= Cl0 Cl1' where (ell> n) = 1, Cl1 J v1 - v2 ; and if Cl0 = pfip~; . ..
pf•', then v1 - v2 = 0 (mod Pi'), where <1; = max (0, Pi - Pi). Let Cl < (In n)° 1.
Then to the Cl given above correspond
L2
2_;._+_1<p_(-~-)-(--=a11_)--(-a')
U1 <p P1 • . • <p pt
(1 + Ba10n n)-01)
pairs v1 , v2 , which contributes an amount
SD L 2
22;.+1 (Cl) ( a1; i
<p i <p P1 · • ·
(
<p Pt
a')Cl Cl (1
i o
+ Ba,(In•n)-01) (6.6.17)

to the sum (6.6.13). Hence we see that if we-extend to oo the series arising from
the main terms of (6.6.17), and replace the sum in (6.6.13) by this series, an error
of

(6.6.18)

results (as before, C1 is an arbitrary constant).

Such an expression can be written as a product of factors which correspond to
odd primes; this has the form

8D 2 L~ ( 00
1 ) 00
1 (6.6.19)
22Hl L Pm <p( Pm) II P(=O
II m=O
'.l>fn
L Pia·'<p ( Pi'a)' '.l>iln
p>2

where <Ji = max (0, Pi - p;). This form is more convenient for computation.
106 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION (CHAP. VI

7. Calculation of the Product over Primes in V1

We have
8 p>2 (1 -p21)- .
-=II
7T2
(6.7.1)

Next, it is easy to show that

(6.7.2)

(6.7.3)
Further

Thus,

(1 - 1)- Pi'=O
p~
! ' 1 = 1- 1/pf!i+l(
co

pf' <p(p'f')
• 1 - 1)
1 - 1/pi
- +---
1
Pi'+ (pi - 1) p~ 1

= (pf•+l - 1)(pl - 1) + .1
pf'+ 2(Pi - 1) pf'+ (Pi - 1)
1

_ P~ - Pl + 1 (pf ,+i - l)(pl - 1) + Pi

(6.7.4)
- 2(
Pi Pi _ j)
. Pi(
Pi Pi3 _
Pi2 + 1)
Thus, (6.7.3) can be rewritten as
7T2D2v~2 pa - p2 +1 (pP+l - l)(p2 - 1) + p
(6.7.5)
22A+l II
p>2 p2( p - 1) p>2 II p P( p3 - p2 + 1) .
pPJln

Note that if p = 0 in pP II n, then for such p, the second product in (6.7.5)

gives a factor of 1. Taking into account the errors given in formulas (6.6.3),
(6.6.8), (6.6.10), (6.6.11), (6.6.13), (6.6.15), (6.6.16) and (6.6.18), we finally
obtain for given A. ;;;i: 0, 2A ~ (ln n)c:

7T2D2Li pa - p2 + 1 (pP+t - l)(p2 - 1) + p

Vi= - - II~-~~-~~
22Hl p> 2 p2(p _ 1) p> 2 pP(p3 _ p2 + 1)
pPlln
+ Bc D 2Li(ln n)-c1,
1 (6.7.6)
where C1 is arbitrarily large if C is sufficiently large.
Sec.8 CALCULATION OF Va 107

8. Calculation of V 3
Now we need to obtain an asymptotic approximation for V3 and show that
(6.8.1)
We have
(6.8.2)

Here A(n, D) is given by (6.5.18) for the case A = 0, and by (6.5.21) for general
A ;;;;,: 0. Redenoting A1(n, D) in (6.5.21) as A(n, D), we obtain

A(n, D) = 1TA 0 L~ IT (1 + $1,) IT (1 + $i>)(1 + ri(Pi• ~i)) (6.8.3)

2 i>ID i>ln
i>fn

(cf. (6.5.6); (6.5.7)). Hence

(A(n, D)) 2 = 1T 2A~L~ 1;.

2 IT (1 + 2$i> + $;) IT (1 + $i>;} 2(1 + ri(Pi• ~i)) 2 • (6.8.4)
2 i>ID i>d n
pfn

We need to sum (6.8.4) over the odd values DE (D) = [D1 , D1 + D 2]. Before
doing this, divide these D into categories {~1 , ~ 2 , ••• , ~ 1 } (cf. (6.5.7)); here
pt• II D, pP• II n. From (6.8.4) it is evident that in each category {~i. ~ 2 , ••• , ~ 1 },
only ITPID; P-tnC 1 + 2$i> + $!) depends on D; therefore we consider
(6.8.5)

From (6.5.13) we see that

2
l$j)I <;; - . (6.8.6)
p
Putting
(6.8.7)
we see that
2 4
IEi>I <;; -p + -p2 . (6.8.8)

The number of DE {~ 1 • ~2 , ••• , ~ 1 } is obviously

D
__!
2
1
IT -.
i=1 pt•
1(1 - -1) + Pi
BT(n). (6.8.9)

From (6.5.19) we see that

(pf' II n). (6.8.10)

108 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

Given an arbitrarily small constant ~0 , then under the condition

t
II Pt'> nso (6.8.11)
i=l

we deduce from (6.8.8), (6.8.9) and (6.8.10) that summation of (6.8.4) over
DE {~ 1 , ••• , ~t} gives the result
(6.8.12)

Summing (6.8.12) over the categories {~ 1 , ••• , ~t}, the number of which,
clearly, does not exceed T(D) = B.n•, we obtain the estimate

B L2 D2 n-so/3 (6.8.13)
so 1 22.1. •

In view of this we may suppose that

t
II
i=O
Pt'< r/ 0• (6.8.14)

Let q be odd and square-free, (q, n) = I. Put Eq = Ilpfq E'P, so that

IE I ~ T(q) . IEql ~ T(q) (6.8.15)
q """'
q '
q """'
q2 •

The summation of (6.8.5) reduces to the summation of

!
(q); q,,;;;n
Eq
D =O(mod q)
! 1. (6.8.16)
De{a., ... , a,}
This last sum is equal to

!
(q);q..;;;n
(D
Eq ~II-
t
2q
1( P;1) + BT(n)).
1- -
Pt' i=1
(6.8.17)

By (6.8.15), if we terminate the summation in (6.8.17) with q > (Inn)° (C is

any given constant), we obtain the permissible error

B D2 (In n)-c12. (6.8.18)

c p~i ... Pt'
Further,
D~12
! 0 EqT(n) =B 4 4 . (6.8.19)
(q);q,,;;;(lnn) P11·. ·Pt'
Assuming these errors, we replace (6.8.17) by the expression

D2 IT -4.(1 - .!.)Pi
2 i=l Pi '
!
(q); n,,;;; (Inn)
o Eq .
q
(6.8.20)
Sec.9] CALCULATION OF THE PRODUCT OVER PRIMES IN V3 109

By (6.8.15),

(6.8.21)
Next
L Eq = IT (1 + 2 ~" + ~).
q
(q) pfn p p
p>2

Thus (6.8.20) can be rewritten as

D22 IT Pi:,(1 -
i=l
l) IT (1 + 2 ~" + ~;) + BcD2 (~~ n)-c1:
Pi 21-tn P P Pi · · · Pt
1
• (6.8.22)
,,>2

Here condition {6.8.14) was assumed. Ifwe negate this hypothesis, then (6.8.22)
has the trivial bound
BD 2 4
1 4
.
,
IT Pi4, > n~o . (6.8.23)
P1 1 ···Pt'
Now put (6.8.22) in (6.8.2), add up the errors (6.8.12), (6.8.13), (6.8.18),
(6.8.19) and (6.8.23), and sum over all systems {~ 1 , . . . , ~ 1 }. We obtain

V:3 = ""
,(...,
(A(n ' D))
•
2 1T2A~ D 2L12 IT
= 22.<+1 (1 + 2 ~" + ~!) IT (1 + ~"•.)2
D1<;;;D,.;;D1+D2 pfn p p P;ln
Dodd p>2

x ! (1 + 'YJ(Pi• ~i))2 Pi1 (1 - l)

4;=o Pi
+ Bc 2 ~! 1 (In
4, 1 n)-C 1 (6.8.24)

( C1 can be made arbitrarily large provided C is large eno~gh).

9. Calculation of the Product over Primes in V3

We have A 0 =
IT
"
(1 + p(pXiP)- 1)) = IT (1 + p(pXiP)- 1)) .
1>>2

Next, it is immediate that

(l + XiP) ) 2 ( 1 + 2~21 + ~!)

p(p - 1) p p
- (1 - 2x4(q) + q )(q2 - q + xiq))2
- q2 - q + xiq) (p2 - q + xiq))2 q2(q - 1)2
= (q2 - q)2 + q - 1 - q3 - q2 + 1
(6.9.1)
q2(q - 1)2 - q2(q - 1) .
110 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION (CHAP. VI

Thus the first term on the right-hand side of (6.8.24) can be rewritten in the
form
2 + 1
D 12 II p - p IIl>
2 3
7T
(6.9.2)
22.Hl 21'0 i»2 p2( p - 1)
where
II1 =II c1 + ~'/);)2 2 (1 - ~ ..!..A; c1 + 'YJ cp.,. ~-))2
..!.)
£.., • . (6.9.3)
'Piln 1 + 2~'/))Pi + ~'IJ;/P, Pi 11,=o Pi
To simplify the notation, we shall drop the index i from the term of II 1 which
corresponds to the prime pi, and also from ~i (and from Pi in formula (6.5.19)).
By (6.9.1) and (6.5.13) we have
(1 + ~'/))2(1 - (1/p)) = p2(p - 1) (p - 1)2(p - xiP)) 2( 1 _ .!)
1 + U'IJ/p + ~!IP Pa - P2 + 1 p2(p - 1)2 P
_ (p - 1)2(p - x4(p)) 2
(6.9.4)
- p(p3 - p2 + 1)
We need to multiply this expression by the sum
00 1 2
!
A=O
&(1
p
+ 'YJ(p,~)). (6.9.5)

II
In studying this sum, we take into consideration the fact that pP n, p > 0, and
distinguish three cases: 1°. xip) = -1, p odd; 2°. xip) = -1, p even;
3°. xip) = +I. Let us begin with case 1°. By definition (cf. (6.5.19)), we have:
When~< p,
o when ~ is odd,
1+11(p, ~) ={
I when~ is even.
When~> p, 1 + 11(p, ~) = 0 (p odd). When~= p, 1 + 11(p, ~) = p/(p2 - 1).
Therefore, in case 1°, (6.9.5) is equal to

1 + ( -p- ) 2 -1= 1 - 1/pP+l + ( -p- ) 2 -1 .

1 + ... + -
1 +1- + -
p2 p4 pp-1 p2 _ 1 pP 1 _ 1/p2 p2 _ 1 pP
(6.9.6)
Multiplying by (6.9.4), which has the form (p 2 - 1)2 /p(p3 - p 2 + 1) since
x4(p) = -1, we find for case 1° that the corresponding term of the product
(6.9.3) is equal to
(pP+l - l)(p2 - 1) + p
(6.9.7)
pP(p3 _ p2 + 1)
Now consider case 2°: x4(p) = -1, pis even. Here when~ < p, 1 + 11(p, ~)
behaves as before; when ~ > p, 1 + 1J(p, M = I. When ~ = p, 1 + 1J(p, M =
1 - p/(p2 - 1) (cf. (6.5.19)).
Sec. 9] CALCULATION OF THE PRODUCT OVER PRIMES IN V3 111

In this case (6.9.5) is equal to

l - l/pP
1 - 1/p2
+ (l _ _ 2
p_)
_!_
p2 - 1 pP
+ _1_ 1
pP+l 1 - ljp •
This expression must be multiplied by (p2 - 1)2/p(p3 - p2 + 1), which gives,
after a simple calculation,
(pP+l _ l)(p2 _ 1) + p
(6.9.8)
pP(p3 _ p2 + 1)
Case 3°: x4(p) = +I. In this case (6.9.4) has the form
(p - 1)4
(6.9.9)
p(p3 - p2 + 1) .
Here when 1 ~ D. < p, we have 1 + 'Y}(p, D.) = D. + I. When D. > p, 1 +
ri(p, D.) = p + I. When D. = p,
1 + 'Y/(P, D.) = p + 1+ p 2 (6.9.10)
(p - 1)
Calculation of (6.9.5) gives
p-l(D.+1)2 (p+1)2 ( p )21
1 pd + d=p+l 1 pd + p+l+ (p - 1)2 -pP
CX)

d=O

((p + l)(p - 1)2 + p)2 (p + 1)2 22 32 p2

= (p _ 1)4pp pP(p _ 1)
+ 1 + - + - + ...
p p2
+-
pp-1
- ((p + l)(p - 1)2 + p)2 + (p + 1)2(p - 1)3
- (p - 1)4pp
+ (p _ l)4(pP + 22pp-1 + 32pp-2 + ... + P2Po)
(p - 1)4pp
p 2 + (p + l)(p - l)((p + 1)2(p - 1)2 + 2p0 )
(p - 1)4pp (p - 1)4pP
+ (p + 1)2(p _ 1)2 + (p _ l)(pP + 22pp-2 + • .. + p2p)
(p - 1)4pp
where the final equality is obtained by adding and subtractingp2 in the numera-
tor. A simple calculation gives us the expression
p(pP+l _ l)(p2 _ 1) + p2
(p - 1)4pp
Multiplying this by (6.9.9), we get
(pP+l - l)(p2 - 1) + p
(6.9.11)
pP(p3 _ p2 + 1)
112 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION (CHAP. VI

Substituting the results obtained into (6.9.2), we find, by (6.8.24), that

7T2 2 p3 _ p2 +1 (pP+l _ 1)(p2 _ 1) + p
V3 = -2 D2L1 IT IT~-~~-~~
2 .l.+l i»2 p 2(p - 1) pPlln pP(p3 - p 2 + 1)

+ Ba1 2~!1 (ln nra1. (6.9.12)

Comparing this with V1 (cf. (6.7.6)), we obtain (6.8.1).

10. Calculation of V2
We now need to calculate V2 and show that
V2 = V1 + Ba1D 2v~2(ln n)-a1. (6.10.1)
For equation (6.5.20),
L U(n - Dv)A(n, D). (6.10.2)
De<Dl
Dodd

Here U(m) = ~ 2.1.«2+ 71•>=m 1, where g2 + 172 is odd.

Consider once again the categories {~ 1 , ~ 2 , ••• , ~t} of values of D. Using
(6.8.3) we find
V2 = 7TA0L1r.1. ! . IT (1 + gp)(l + 11(Pi• ~i))
41, . .. , .At :Pil n

x L L IT Cl+ gP)U(n - Dv). (6.10.3)

ve(vl De{&1•···• &tl PID
pfn

We seek an asymptotic approximation for the inner sum in (6.10.3) for a given
category {~ 1 , ••• , ~t} of values of D.
For q square-free and relatively prime to n, put gq = IlP!a gp· By (6.8.6),

(6.10.4)

If (q) is the totality of these numbers, then the inner double sum in (6.10.3)
can be rewritten as
L L gq L U(n - Dv). (6.10.5)
v e(v) q e(q);q,,.;n De {&1, •.• ,4tl
D= O(modq)

Since (q,p~ 1 ... ph = 1, the inner sum in (6.10.5) reduces to

L U(n - dpt 1 ••• Pttqv), (6.10.6)
dE(d)

where dis odd, (d, p~1 . .. p~t) = 1, dp~1 . .. p~tq E (D).

Sec. 10] CALCULATION OF V2 113

The totality of such values of d is denoted by (d). Thus we must have

p~ 1 . .. pf•qv <
n. We shall prove that with a permissible error we can suppose
that
(6.10.7)
Suppose that at least one of these inequalities is violated. (6.10.6) is the n.s.e.
2" (~2 + rJ2) + dp~ 1 ... pf•qn = v.
Here we shall assume condition (6.6.9): 2• < (ln nf. Then the n.s.e. of our
equation does not exceed

~ T ( n - d p 141
B £.. ... ~ 1 -- B ,n•
Pt41 qv) -- B ,n• £.. 4
D2 4 (6.10.8)
de(d) dE(d) P1 1 ·Pt 'q
• •

If q > n< 0 , then because of(6.10.4), summing in (6.10.5) over all suchq, we obtain
the estimate BD2 Li_n-< 0 t 2/p~ 1 ... pf•, which gives the estimate BLiD2 n-< 013p~1
... pf• in (6.10.3) and, summing over all categories {d1, . . . , dt}, taking into
account the coefficients (1 + 'YJ(p, d)), we get B(D2/2")Lin-<o1 4 _ If p~ 1 . .. pf• >
n<o, then, summing (6.10.8) with coefficients of(l + 'YJ(p, d)), we find the estimate
BD2n-<onLi_, and then, summing over v, we again arrive at the previous estimate:

B D; L~n-se/4. (6.10.9)
2
Thus we can assume condition (6.10.7). Let us consider the sum (6.10.6).
Here at first put d 2 = d 3 = ... = dt = 0. We have, for DE {d1 },
L U(n - dpt 1qv) = L U(n - xpt 1qv) - L U(n - x 1 pt 1 +1qv), (6.10.10)
dE(d) (z) (z1)

where D1 < xp~ 1 qv < D 1 + D 2 ; D 1 < XiP~ 1+Iqv < D1 + D 2 ; x, x1 run through
the integers.
Consider the corresponding equation:

2\~ 2 + 'YJ 2) + xpt 1qv = n; ~2 + 'YJ 2 odd. (6.10.11)

Here (6.10.11) is satisfied. At the expense of a permissible error in the number

of solutions we can assume that (v, p~1q) = 1. For if v p~ 1 , then v = p 1, v n, I I
and the n.s.e. (6.10.11) is

~T
B £., (n- - xv4 1q ) D 2 lnn
= B-4-1 - . (6.10.12)
(o:) V V1 q
<
Summing overq n<o and over v in (6.10.5), we obtain the result: BD2(ln n)2/v~ 1 .
Summing over d 1 ~ 1 in (6.10.3), we obtain the error

B D2L1(ln n)a (6.10.13)

2" '
114 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

which, of course, can be admitted. If v q, then a simple calculation gives I

(6.10.13) again.
Now consider the congruence

(6.10.14)
Since p~ 1 , q, 'II are odd and relatively prime, the total number of solutions of this
congruence is equal to
Q(n, p~ 1 ) IT Q(n, p),
Plqv

where Q(n, r) is the total number of solutions of $2 + rJ2 = n (mod r) (cf. the
table of § 3, Chapter I).
Similar arguments may be applied to the second term on the right side of
(6.10.10). Because of (6.10.7) and the fact that v ~ n1 16, we find, after an ele-
mentary calculation, that
D D -s.
! U(n - xpt 1qv) = .i. ; 2
2 24 Q(n, p~ 1 ) IT Q(n, p) + B 2 "+1 2q~p 24 '11 . (6.10.15)
(z) 2+qp 1 '11 p]qv 1

An analogous expression is obtained for the second term of (6.10.10). In the

general case DE {~ 1 , . . . , ~ 1 }, assuming (6.10.7), and arguing in a similar
fashion, we obtain (cf. (6.10.5)),

! U(n - DP) = :_~ 22 IT Q(n, p)

De(A,, ... ,A,) 2 q 'JJ pfqv
D=O(modq)

D -h
+B 2n . (6.10.16)
2-'+1q2piA1 ... p~At'JJ

Now on the basis of the table of § 3, Chapter I, we have the expression

7T D2 IT
2.t+l q plqv
(1 - xiP))
4
ITt
.
i=l
[-1
Pi
2A;
Q(n
'
~')
Pi
- _1_
Pi
2A;+2
Q(n
'
A;+l)J
Pi

(6.10.17)

for (6.10.16). Then 1 - xiv)/v can be replaced by 1 with the permissible error
B/v = B exp - (ln n)<Y. 0 • Taking (6.10.4) into account, we see that, after multi-
plying (6.10.17) by $q and summing over q and v, we obtain the error

B D 2L 1 exp - t(ln nt"0 •

(6.10.18)
2.i.+1p~' ... p~·
Sec. 11] CALCULATION OF THE PRODUCTS OVER PRIMES IN V2 115

11. Calculation of the Products over Primes in V2

Now let us consider the series

! ~(1
(q) q
- xiq)).
q
(6.11.1)

where (q) is a set of numbers which are square-free and relatively prime to 2n.
By (6.5.13) (where q is assumed to be prime) (6.11.1) is equal to

II (1 + g1'(1 _XiP))) = II p3 - p2 + 1 . (6.11.2)

(p,2n)=l p p (1J,2n)=l p(p2 - p - xip))
Substitute this in (6.10.3) and sum over {~i. ... , ~t}· If we now extend the
summation over {~1 , . . • , ~t} to infinity, considering that the first term of
(6.10.17) and (6.10.18) does not depend on v, we obtain

X [ _1
2•.
Q( n, Pi4;) - _1 _ Q( n, Pi4;+1)]
2•. 2
+ B C1 D 2Vo'2(In n)-0 ' 1 (6.11.3)
Pia, Pia,+

where C1 is an arbitrarily large constant. Next we observe that

and by the definition of A 0 , the product over primes on the right side of (6.11.3)
can be rewritten as

II Pa;- P2 + 1 II ;~CPi -:- 1)

1'>2 p (p - 1) 1><ln Pi - Pi + 1
(1 - xiPi))c1
Pi
+ rJ(Pi• ~i))

x I [(Pi!
4;=o
4; Q(n, Ph - 2~;+ 2 Q(n, Pt;+l))J.
Pi
(6.11.4)

Now we must make an elementary calculation, using the table in § 3 of

Chapter I and formula (6.5.19).
Let pi= p;pP II n; p > 0. As before, we distinguish three disjoint cases (the
only possible ones): 1°. x 4(p) = -1, p odd; 2°. xip) = -1, p even; 3°. xip) =
+ 1. The computation is analogous to that of § 9 and entirely elementary.
( 6.11.4) is found to be equal to

p3 _ p2 +1 (pP+l _ 1)(p2 _ 1) +p
II
1'>2 p2(p - 1) II
1JPJln pP(p3 - p2 + 1) · (6.11.5)
116 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION [CHAP. VI

Substituting this in (6.11.3), we obtain

7T2 2 p3 _ p2 + 1 (pP+l _ l)(p2 _ 1) + p
V2 = -2 D2Ll IT IT ~--~-~-~
2 "+1 i:»2 p2(p - 1) pPJln pP(p 3 - p 2 + 1)
+ Ba1 D2v~2(ln nra1. (6.11.6)
This proves (6.10.1).

12. Application of the Cebysev Inequality

Taking into account (6.7.6), (6.8.1) and (6.11.6), we find that the dispersion
V' of the number of solutions of
n = 2;.(~2 + r/) + vD', ~2 + r/ odd, 2;. ~(In n)a, (6.12.1)

where v and D' run through the values described in § 3, has the form
V' = Ba1D2L: (In n)-a1+2K. (6.12.2)

Now we can proceed as in § 6 of Chapter IV. Take C1 ~ 200K in (6.12.2). Let

Qr (r = 1, 2, ... ) be the number of D' E (D) such that

L~ (r + l)(ln n)-2°K ~ I !
2 ve(vl
U(n - D'v) - A(n, D') I
(6.12.3)
Clearly
(6.12.4)
Here each value of D' is repeated B(ln n) 2K times (cf. § 2). From (6.12.2),

(6.12.5)

Now let us seek an estimate for

!'
D'e(D)
I! ve(v)
U(n - D'v) - A(n, D') I· (6.12.6)

where the accent on the outer sum indicates that it is over D' which correspond
to the given r. By (6.12.3) and (6.12.5), it does not exceed

QrL1(r + l)(ln nr20K = B~2 L1 (In n)-10K.

2 r
Summing over r > 1, considering (6.12.4), we obtain the estimate
BD2£i(ln n)-6oK. (6.12.7)
Sec. 13] ASYMPTOTIC EXPRESSION FOR THE SOLUTIONS 117

For the remaining case r = 1, we have

IL
ve(v)
U(n - D'v) - A(n, D') I= BL~ (ln 2
n)- 2°K. (6.12.8)

13. Derivation of the Asymptotic Expression for the Solutions

By the results of§ 12, we have

L L U(n -
D'e(D) ve(v)
D'v) = L
D'e(D)
A(n, D')

+ B( L
D'e!D>
1) L1 (ln
2'-
nr20K + BD2 Li (ln nrsoK.
2'-
(6.13.1)

The error in (6.13.1) can be written in the form

BD2 L~ (In n)-1sK. (6.13.2)

2
It remains to calculate I:De(DJ A(n, D'). All the prime divisors of D' are at least
as large as v0 ~exp (In n)"0 ; µ(D') =;6 0. Turning to (6.5.18), we see that, in
accordance with (6.5.19), i17(p;, ~;)I < 4 here.
Consider the corresponding expression

L II (1 + $p) II (1 + 11(p, ~)). (6.13.3)

D'e<D) PID Pin
pfn pi!.lln

It can be rewritten in the form

(6.13.4)
where
YP = $p + 17(p, ~) + $p17(p, ~). (6.13.5)
Putting Eq =TI* yp for q square-free, we find that when q > 1,
Eq L 1 = BD2(ln n)2K Eq. (6.13.6)
D'e(D) q
D'=O (IDOdq)

Here all prime divisors of q are greater than v0 • If 17(p, ~) = 0, then IYPI < 2/p
by (6.8.6); if 17(p, ~) =;6 0, then p Jn; ~ = I,p > v0 • Taking this into con-
sideration, with the help of Lemma 1.1.8 we easily obtain
B D2(ln n)-ioK
L
D'e!D)
1= 2'" . (6.13.7)
D'"'Ocmodq)
Thus (6.13.3) gives
L 1 + B~2 (In n)-10K. (6.13.8)
D'e(D) 2
118 ANALOGUES OF A HARDY-LITTLEWOOD EQUATION (CHAP. VI

Hence

1 A(n, D') = 7rA0L1( 1 1) IT (1 + $1') + B ~2 Li(ln n)- °K, 1

D'e(D) D'e(D) 1>ln 2

or, on the basis of(6.5.13) and (6.13.1),

1 1
D'eW) VE(v)
U(n - D'v) = 71" A; II (p 2-
2 nl 1'
l)(p - xip))
p - p + xiP)
L D'e(D)
1( 1 1)

+ B ~2 L1(1n n)-10K. (6.13.9)

2
Let us return to equation (6.5.1). By (6.13.9), for odd n, summing over
A.= 1, 2, ... , we find that the n.s.e. (6.5.1) is

7rAo II (p ~ l)(p - xip)) L1( 1 1) + BD2L1(1n n)-1°K. (6.13.10)

1>ln P - P + XiP) D'e<D>

For even n, such a form for the number of solutions is obtained without summa-
tion over A.. Taking into account the errors determined in § 2, and summing over
the intervals (v) and (D), we obtain

7rAo IT (p ;- l)(p - xip)) L~(n) + Bn(ln n)-sK, (6.13.11)

1'fn p - p + z4(p)
for the number of solutions of equation (6.1.3) when k ~ 7. Here L~(n) =
Lz 1 ' •••z,.'<;:;n;z';eOp 1 (cf. (6.1.8); note that we are supposing K ~ 100).
The estimate (6.3.4) is very important for the case k = 6. In this case we get

7rAo IT (p - l)(p - xip)) L~(n) + B n(ln ln n)2 (6.13.12)

1'171 P2 - p + x4(p) (ln n) 312
for the n.s.e. (6.1.3).
This proves Theorem 6.1.2. We shall need it in Chapter VII.
Now let us turn to Theorem 6.1.1. Consider(6.4.1). If exp (lnn)" 0 <'II~ n1 17,
then (6.13.10) gives

7rAo IT (p ~ l)(p - XiP)) ( 1 1 + B(ln n)-sK)

1>Jn P - P + XiP) <v:1>>
V'.P~n

for the number of solutions, where 'II runs through that interval. With
errors of (6.4.2) and (6.4.3), we can replace L(v;p);vp=n 1by2Lp1 p2 .;;n;p 1 .;;-v;i 1,..._,
n In ln n/ln n here.
Since cx0 is arbitrarily small, we obtain formula (6.1.5). We remark that the
asymptotic expression (6.1.5) has a poor error term. As will be seen in Chapter
VII, it is possible to obtain the error term Bn(ln n)-i.028 •
 CHAPTER VII

A PROBLEM OF HARDY AND LITTLEWOOD

1. On an Equation of Hardy and Littlewood

In their famous paper [40] of 1923, Hardy and Littlewood consider the
equation
n = p + ~2 + r;2 (7.1.1)
and on the basis of purely heuristic considerations propose the following
asymptotic formula for the number Q(n) of its solutions:

Q( n) ,._, 7TAo ~ II (p -:- 1)(p - XiP)) ' (7.1.2)

ln n PJn p - p + X4(p)
where
Ao = II ( 1 + XiP) ). (7.1.3)
p p(p - 1)
Hardy and Littlewood point out that with the extended Riemann hypothesis one
can prove that "almost all" numbers have the asymptotic formula (7.1.2). This
is not difficult to do at the present time if one uses modern results on the distribu-
tion of primes in segments of arithmetic progressions. In 1928, Stanley [54]
proved it for "almost all" numbers, using the extended Riemann hypothesis.
The papers of Chowla [32], Estermann [38], and Halberstam [39] are related to
this paper of Stanley's. In 1957 the important paper [43] of Hooley appeared, in
which is given a conditional derivation of (7 .1.2) based on the extended Riemann
hypothesis. Besides this, [43] is interesting in that it introduces an important
type of number, which we considered in Chapter I (cf.§ 5, Chapter I) and called
the "quasiprime shell" of almost prime numbers.
The author [15] applied the dispersion method to demonstrate the solvability
of equation (7.1.1), and obtained an inequality for Q(n) which preserves 97.9 %
of the asymptotic formula (7.1.2). In [16], on the basis of [15], a proof is given of
the asymptotic formula
Q(n) = 7TAo 2!_ II (p -:- l)(p - xip)) + R(n), (7.1.4)
ln n Pin P - P + XiP)
where
n
R(n) = B (ln n)i.02s. (7.1.5)

119
120 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Note that TI'.Pln = B In Inn; (II'.P 1n)-1 = B In Inn in formula (7.1.4). In this
chapter we shall essentially reproduce papers [16] and [15] in a slightly different
form.

2. Transformation of the Hardy-Littlewood Equation

We shall make use of§ 6 of the Introduction. Equation (7.1.1) is of the form
>
(0.6.10), where <p = ~ 2 + ri 2 • Further, form 2, N (n = ~ 2 + ri 2 +pm) = B,n112+•.
As in (6.1.2), choose
P =exp ( In In In In ,n n) (7.2.1)
K 0 In Inn
where K 0 > 100 is a sufficiently large (by as much as is needed later) constant.
Let Qk(n) be the number of solutions of the equation Yk (cf. (6.1.3))

n = X~ • •• X~ + e + 'Yj 2 ; x; E Qp. (7.2.2)

Here (cf. (6.1.4))
6 ~ k ~ ri = [Koln lnln lnInn].
n
(7.2.3)
By (0.6.16),

Q(n) = Q~(n) - iQ~(n) + lQ~(n) - tQ~(n) + tQ~(n)

r1 ( l)k-1
- i-Q~(n) +k~7 - k Q~(n) + B,n11 2+•. (7.2.4)

The numbers of solutions Q~(n), Q;(n), ... , Q;Jn) of equations Y~, Y;, ... , Y;,
were considered in Chapter VI. From (6.1.9), (6.1.10) and (7.2.3), keeping the
notation of (6.1.8), we get

I (-it~l Qk'(n) = 7TAo IT(p ~ l)(p - xip))

k=s k '.Pin P - P XiP) +
x ~ (-1t-1 Lin)+ B n(ln In n)2 (7.2.5)
k=6 k (In n )312
We cannot apply the dispersion method directly to the five remaining equations,
Y{, Y2, ... , Y5, since it does not give precise enough error terms. Since we
expect the main term of the asymptotic formula for the solutions of(7.l.l) to be
of the form (7.1.2), the error terms in the asymptotic approximations for Q~(n)
should be, for example, of the form Bn/(ln n)l+'o (-r0 > 0 a constant). It is not
difficult to obtain such error terms for Y{, Y~, Y~ if one uses Weil's estimates for
trigonometric sums (see Chapter I). Equations Y~ and, in particular, Y~ present
much greater difficulty. We shall treat Y{, ... , Y~ by a single method.
Sec. 3] THE EQUATIONS Y~, k <5 121

3. The Equations Y~, k <5

Let us consider the equations Y{, Y~, Y~, Y~, Y~. We shall prove the asymp-
totic formulas:
Q~(n) = 1TLin)Ao II(p-:- l)(p - xip)) + Rk(n), (7.3.1)
"In P - P + XiP)
where
Rk(n) = B n
(In n)I+ro
, k <5 (7.3.2)

(To ;;;i. 0.028 is a constant; A0 = II., (1 + xip)/p(p - 1))).

Let k = 1, 2, 3, 4, 5. By means of an elementary sieve, the equation Y~:
x~ ... x~ + ~ 2 + 'f/ 2 = n, xj E Op (7.3.3)
can be reduced to equations of the form
dx1 ... xk + ~2 + 'YJ2 = n,
where xi runs through all integers.
Let AP denote the set of odd numbers, all of the prime divisors of which are
less than P. Denote by W(n, q1 , .•• , qk) the n.s.e.
ql . . . qkxI ... xk + e + 'fJ 2 = n, qI ... qk E Ap, (7.3.4)
where xi runs through all odd numbers. Obviously,
Q~(n) = 2 µ(q1) .. . µ(qk)W(n, qi ... qk). (7.3.5)
q1 . .. qk~n
ql ... qkEAp

Put q1 ••• qk = d. For given k < 5, we separate equations (7.3.4) into two
classes:
CLASS I. A1: d < nr, where = 0.001,T
CLASS II. An: nr < d < n.
We shall show that the equations of class An can be neglected. We shall
prove the lemma:
LEMMA 7.3.1.
(7.3.6)

where K 1 = K 1(K0)--+ oo as K 0 --+ oo (d = q1 ••• qk E Ap).

Observing that ~<2+~2=m 1< 4T(m), we find that
2 W(n, q qk) < 4 2 T(n -
1 •.• dx)Tix). (7.3.7)
nT <q 1 ... qk~n nT <d~n
deAp
x..;;nld
122 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Let (d) denote the totality of numbers d satisfying the above conditions.
Then (7.3.7) cannot exceed

(7.3.8)
By Lemma 1.1.2,
2 h(x)) 2 = B(ln n)a 1 2 ~ = B(ln nt n 2 Tim) = Bn(ln n)a 1 2• (7.3.9)
dxo;;;n dE(d) d m,,;;n m
de(d)

Now let us estimate the sum

(7.3.10)

Put S = S 1 + S2 , where
S1 = 2
n-r :s;;,;;d~n0.98
; S2 = 2
n0.98::;:;;;d~n
(7.3.11)
dE(d) dE(d)

We examine S2 first. For each d E Ap satisfying the indicated conditions,

separate off that product of its smallest prime factors such that it is represented
in the form d = did2 , where ,/;;JP < di < .J~, and where we take the largest
possible d1 for d. We have
S2 < 2 (T(n - did2x)) 2 = 2 (T(n - did2x)) 2Tidid 2). (7.3.12)
d1,d2,a:: dva2.x
d1d2E(d) d1d2E(d)

In the last sum di d2 is taken without repetitions; the repetitions are accounted
for by the factor Tidid2 ). Further, Tidid2) = BTk(di)Tid2 ).
Therefore, (7.3.12) does not exceed
2
d 1 m~n
T(n - dimMdi)Tim). (7.3.13)
d 1 E(d)

Here .J~p-i <di< .J~; di is taken without repetition; di E Ap. We estimate

the sum (7.3.13) by

(7.3.14)
Next,
1
2 (Tim)) 2 = Bn(In nt 3 2-. (7.3.15)
dim.;;n di d

Here .J~p-i <di< .J~; di E Ap.

Next, from Lemma 1.1.5 we deduce the estimate

2 T2(n - dim)T2(di) = Bn(In nts 2 T2(di) = Bn(In n)a 3• (7.3.16)

~m.;;n ~ di
Sec. 4] INVESTIGATION OF Q~(n) 123

Let us consider l:d 1 (l/d1). We shall show that an estimate for it is

,L .!:._ B(ln n)-K 2 • (7.3.17)
nP,;;;;d1 <2nP d1
d 1 EAp

Here K 2 = K2(K0)--+ oo as K0 --+ oo, and p,;;;;; 1, p > -r is some constant. We

use Lemma 1.1.6 of Chapter I.
The number of numbers ,;;;;; x, all prime divisors of which are .;;;;; z .;;;;; ,J-;, is
estimated by
Bx exp (- l (in l + In In l) + l + _Q_), (7.3.18)
°' °' °' °' a:ln11 ..
where a:= ln z/ln x; IQI ,;;;;; 1.
In our case z = P =exp (Inn In In In n/Kln ln n); x = 2nP.
l = In x > In n K 0 In In n = K 0 ln In n ,
a; In z In n In In In n In In In n
In (l/a:) > (l/a:) In In ln n when n > n0 = n0 (K0).
Applying (7.3.18) we now obtain the estimate (7.3.17), where we can take
K 3 = iK0 • From (7.3.17) we immediately obtain

,L _! = B(ln n)-K3 ; K 3 -- K2 . (7.3.19)

n d1
7 ,;;;;d,_<n 2
After this (7.3.15) may be estimated as
Bn(ln n)-K• (7.3.20)
and combining (7.3.12)-(7.3.16), we obtain, finally,
S2 = Bn(ln nrK•. (7.3.21)
The estimation of S 1 is similar, but even easier: here nT ,;;;;; d,;;;;; n°· 98 and d
itself takes the place of d1. We have finally
S1 = Bn(ln nrKs. (7.3.22)
From (7.3.7), considering (7.3.9) and (7.3.8), we obtain the desired estimate
(7.3.6).
4. Investigation of Q~(n)

Thus it is sufficient to study only q1 ••. qk .;;;;; nT = n°· 001 in the formula (7.3.5)
for Qfc(n).
We are interested in the sum
(7.4.1)

Let a:0 = 10-10 ; d0 =exp (In n)"o.

124 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Let Q12>(n) denote the part of the sum (7.4.1) which corresponds to numbers
q1q2 ••• qk such that
(7.4.2)

and denote the part of (7.4.1) for which q1q2 ••• qk < d0 by Q1°>(n). We have

(7.4.3)

The study of Q1°>(n) is the more difficult, and we shall postpone it. We now
examine Q12 >(n) with the help of the dispersion method.
Divide Q12>(n) into two sums: Q12·+>(n) where µ(q 1)µ(q 2) ••• µ(qk) = +l,
and Qk2·-l(n), where µ(q1)µ(q2) . .. µ(qk) = -1. We shall seek asymptotic
approximations for Qk2 .+>(n) and Qk2 ·->(n) separately, beginning with Q12·+>(n);
Q12 ·->(n) will be investigated in a completely analogous fashion. Qf·+>(n) is the
n.s.e.
(7.4.4)

< <
where d0 q1q2 ••• qk nT; µ(q 1)µ(qJ ... µ(qk) = + 1, qi E Ap.
Divide the basic interval [d0 , nT] into parts of the form I 0 : [tn0 , n0 ] (combining
the extreme left parts). We know that q1 •.. qk E Ap, and therefore when n0 is
large (for example, when it is close to n the number of q1 . . . qk E I 0 is small.
7
),

Separate off those segments I 0 where the number of these numbers is >
n0 /(ln n)K•. In the remaining-"irregular"-intervals I~, the number of q1q2
... qk is Bn/(ln n)K•. Let us estimate the n.s.e. (7.4.4) for q1q2 ••• qk EI~.
For given I~, it has the bound

B ! T(n - q1q2 . .. qkx)Tk(x)

ql · . . OkX~n
ql • • • qkElo'

where K7 > K 6 /2 provided K 6 was chosen larger than 2a4 • Summing (7.4.5) over
all the bad intervals I~, we obtain an error of
Bn(ln nrK• (7.4.6)
in the corresponding n.s.e.
Sec. 5] THE NONIRREGULAR INTERVALS 125

5. The Nonirregular Intervals

Now let us turn to the intervals I 0 which are not irregular. Separate from the
numbers q1q2 ••• qk E I 0 those which are divisible by squares of primes p such
that
p ~ exp (In n t° 110 (7.5.1)
(of course, p < n Using Lemma 1.1.9, it is easy to obtain an estimate for the
7
).

number of these numbers in I 0 : an estimate, for example, is

Bn0 exp -(In In n) 2, (7.5.2)
and, as before, we see that the number of solutions corresponding to such
q1q2 ... qk can be neglected with an error of (7.4.6). Thus with such an error we
shall consider only those of the numbers q1q2 ••• qk which are not divisible by
squares of primes satisfying condition (7.5.1).
Separate the remaining q1 . . . qk E I 0 into classes Ar according to the number
of their prime factors of the form (7.5.1); obviously, r <Inn. Any number
q1 ... qk E Ar can be written in the form
(7.5.3)
where p satisfies (7.5.1), and this can be done in exactly krTipq) ways; in other
words, in writing q1 ••. qk = pq, the number q must be counted krTiq) times.
Thus the n.s.e.
q1q2 · · · qkx1X2 ... xk + $2 + 'Yj 2 = n, (7.5.4)
on condition that q1 ••• qk E I 0 , q1 ••• q,, E Ar, is equal to

~ N(pqx 1x 2 ••• xk + $2 + 'Yj 2 =

n), (7.5.5)
r
where N here (and henceforth) denotes the number of solutions of the equation
in parentheses; q is counted f(q) < Tk(q) times; pq E I 0;pq E Ap;p satisfies
(7 .5.1 ), and in the factorization pq = q1 ••. qk, µ(q 1) . . • µ(qk) = 1.1
Having in mind the application of the dispersion method to (7.5.4) we
rename p as v; the place of D' is taken by D' = qx1 ... xk. Divide the interval
of variation of v
exp (In n)"'0110 < v < n 0 < n' (7.5.6)
into intervals (v0): [v0 , v0 + v0/(ln n)K•]; divide the interval of D' into intervals
(D0): [D1 , D1 + D 1(1n n)-K•]; put D 2 = D 1(1n n)-K•, v~ = v0(1n n)-K•.
As before, we see that we can neglect variations in the boundaries and
introduce independent intervals for v and D',
After establishing independent intervals for the variation of v and D', we can
suppose that they vary altogether independently. D' = qx1 ... xk; X; runs
1 This is why f (q) < -rk(q) instead of being equal to -rk(q).
126 A PROBLEM OF HARDY AND LITTLEWOOD (CHAP. VII

through the integers; q runs through only those numbers q E Ap which have
exactly r - I prime factors satisfying condition (7.5.1), and which are further-
more such that in the representation vq1 ••• qk, µ(q 1) ••• µ(qk) = I. As v varies
in (v0), the boundaries of variation of the q can change, but none of their other
properties change.
The numbers v run through all primes in (v0).
We shall apply the dispersion method to the equation
n = g2 + 'Yj 2 + D'v; D' E (D 0 ); v E (v) 0 • (7.5.7)
For given D' we get (cf. (6.5.18)),
A(n, D') = 7TL1A 0 II (1 + gP) II (1 + gP)(l + 'YJ(P;, ~;)), (7.5.8)
P/D' p;jn

where Li is the number of primes in (v0), and the symbols gPi and 'YJ(p;, ~;) are
explained in (6.5.13) and (6.5.19). Here n is assumed even; for odd n, as in
Chapter VI, we need to study
n = (g2 + 'Y/2)2" + D'v; g2 + 'YJ 2 odd (7.5.9)
<
instead of (7.5.7) (here 2" (ln n) 0 ). Then we must multiply (7.5.8) by 2". For
the time being we shall consider even n.
The dispersion method can be applied to (7.5.7) just as it was in Chapter VI
for the analogous equation. The only difference is that here there may be
considerably more repetitions of the odd numbers D'. Because of this we need to
apply the method of§ 2, Chapter IV. Let p(D') be the number of repetitions of
D' in equation (7.5.7). Let p(D') > (ln n)Kio. On the basis of Lemma 1.1.5, we
deduce that an estimate for the number of solutions of (7.5.7) corresponding to
such D' is
(7.5.10)
Here we can suppose that K9 > 100K10 • For the remaining solutions, p(D') <
(ln n)Kio. The dispersion method can be applied to them as in Chapter VI. Then
we find that the number of solutions of (7.5.7) is
I A(n, D') + BD 2L1(ln n)-K 2, 1 (7.5.11)
D'e<Do>
where A(n, D') is given by formula (7.5.8). For even n this is obtained by
applying the dispersion method to equation (7 .5.7), and for odd n, by applying
it to (7.5.9) and summing over .il such that 2" (ln n)°. <
6. Transformation of the Number of Solutions
Expanding (7.5.8) with respect to all primes, putting A 0 =
IIP (1 + xip)/p(p - I)), 7T = 4(7T/4) = 4IIP (1 - xip)/p)-1, we obtain
00

A(n, D') = 4L1 I y(r, n, D'), (7.6.1)

r=l
Sec. 6] TRANSFORMATION OF THE NUMBER OF SOLUTIONS 127

where y(r, n, D') results from the multiplication together of all quantities in
(7.5.8) which correspond to powers of primes dividing r.
Let
r0 = exp (ln ln n) 4 • (7.6.2)
Consider
! ! y(r, n, D'). (7.6.3)
D'e(Do> r>ro

Repeating the arguments applied in Chapter VI, §§ 8-10, in making analo-

gous summations, we see that an estimate for (7.6.3) is
BD'(ln n)-K1 s, (7.6.4)

so that the number of solutions of (7.5.11) can be written in the form

ro
L L14 ! y(r, n, D') + BD2v~(ln nrKia. (7.6.5)
D'e(Do> r=l

Now Mus turn to § 5 of Chapter VI, where the quantity A(n, D) was intro-
duced. From the discussion there it is clear that when r < (ln n}°, the expression
£iy(r, n, D'), obtained from the expansion of the products over primes in
A(n, D'), is the product of 4xir) by the main term of the number of solutions of
the congruence n - D'v = 0 (mod r) for fixed D' and v E (v0). Note that here
v0 >exp (ln n)"'•'10 ; ln v0 > (ln n)"'•' 10 • By Lemma 1.6.3, for any C > 0 and r <
(ln n)°, the n.s.e. n - D'v = 0 (mod r), v E (v0), has the main term £iy(r, n, D')
and an error term of the form

(7.6.6)
Hence
y'
L 1 y(r, n, D') = xh)N(vD' = n (mod r); v E (v0)) + B .....!! (ln n)K s,
1 (7.6.7)
r
if r < (ln n)Ku.
For the purposes of later computations we should like to obtain formulas
similar to (7.6.7) for
(ln n)Ku < r < r0 = exp (ln ln n) 4 • (7.6.8)

Under condition (7.6.8), the main term, as before, has the form £iy(r, n, D');
the error term, however, has the form B(v0/r)(ln n)-K1s only when among the
I
primitive real characters xP(m) for p r there are no exceptional characters, the
corresponding Dirichlet series L(s, xp) of which have exceptional (Siegel) real
zeros aP,
<Ip > 1 - .!l.L . (7.6.9)
ln p
128 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Such a situation is described in § 6 of Chapter I. As was explained there, when r

satisfies (7.6.8) we shall have formula (7.6.7) only if r is not divisible by excep-
tional moduli p. But if r is divisible by one of these exceptional moduli, then we
need to replace (7.6.7) by the trivial formula
BL1 y(r, n, D') = N(YD' =n (mod r); YE (Y0)). (7.6.10)
By Lemma 1.6.4, the exceptional moduli p are distributed more sparsely than the
geometric progression UJo, w~, wi, ... (w 0 > 1).
Consider the exceptional moduli in the interval [(ln n)-Ku, r 0]; let these be
r1' r2 , ••• , r 1 ; obviously, t = B(ln ln n)4 • Numbers r of the same interval which
are divisible by some r; are dangerous for us. These are numbers of the form
r = r;r'. Summing estimate (7.6.10) over such r, we easily find the total estimate
y'
B _g (In nt". (7.6.11)
r;
Since r; ;;;.i. (In n)Ku and the r; are distributed less frequently than a geometric
progression, summation of (7.6.11) over r; gives
BY~(in n)-K' 5 (7.6.12)
and summation over D' E (D 0) (counting repetitions) gives
BY~D 2 (ln nrK 16 • (7.6.13)
Hence we can write the n.s.e. for (7.5.7):
4 L xir)N(YD' := n (mod r); YE (Y0); D' E (D0))
r~ro
+ B length (Y0) length (D0)(1n n)-K 16 • (7.6.14)
Here D' is repeated the appropriate number of times.
Now sum equations (7.6.7) over all classes A. of numbers YD' (s ln n), <
over all intervals (Y0), (D0 ), then over all intervals / 0, /~. Then by formula
(7.6.14) and all the previous estimates, we .can write the n.s.e. for the equation
thus obtained, (7.4.4):
Qi ,+>(n) = 4 !
2 xir)N(q1q2 . .. qkx1x 2 . .. xk = n (mod r),
r:::;;;ro
q1q2 · · · qkx1X2 . .. xk < n)
+ Bn(in n)-K 17 • (7.6.15)
Here
d0 < q1q2 ... qk < nr; f1,(q1)µ(q 2) ••. µ(qk) = +1. (7.6.16)
Quite analogously,
Qi ·->(n) =
2 -4 ! xir)N(q1 ... qkx1 ... xk = n (mod r); q1 ••• qk < n)
r~ro

+ Bn(in n)-K 17 • (7.6.17)

Sec. 7] THE FINAL FORMULA FOR Qk2l(n) 129

Here
do< q1 · · · qk < n'; µ(q1)µ(q2) • · · µ(qk) = -1. (7.6.18)
Thus, returning to (7.4.3), we find that
Q~2 >(n) = 4 ! xk) ! µ(q1)µ(q2). · · µ(qk).
r~ro do<l'11 ... qk~n-r

N(q1 ... qkxl ... xk = n (mod r); ql . .. qkxl .. . xk < n) + Bn(In nrK 18 •

(7.6.19)

7. The Final Formula for Q~2 >(n)

Since it would facilitate later computations, we should like to extend the
sum over q1 ••• qk in (7.6.19) up to n. We shall prove that this produces a
permissible error.
LEMMA 7.7.1.
! xir)
4

N(q1 ... qkx1 . .. xk = n (mod r); q1 ... qkx1 .. . xk < n) = Bn(ln n)-K 19 •

(7.7.1)
Here it is assumed that q1q2 ... qk E Ap.
When n' < q1 . . . qk < n, we have
N(q1 ... qkx 1 ... xk = n (mod r); q1 ... qkx1 ... xk < n)
B Tk(q)Tk(x). (7.7.2)
qx=n(modr)
qx~n,qEAp

For any q E Ap, q > n', one can obtain the factorization q = q1q2, where
n'l2/p < < ..rrzP.
q1 Further,
Tk(q) = Tiq1q2) < Tk(q1)Tiq2),
Tiq)Tix) < Tk(q1)Tiq2), TiX) < Tiq1)T!(q2x).
Put q2 x = y; then for given r the right-hand side of (7.7.2) can be written as
(7.7.3)

<
Here r r0 • Let (q, r) = c5; then the sum in (7.7.3) has the estimate (see
Lemma 1.1.4)
(7.7.4)
130 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Let q1 = q~ <5; since <5 J r, r ~ r0 , and q1 > nT 12 , q~ > nT 14• Further, Tk(q1) ~
Tk(q;)Tk(b), and for given b, (7.7.4) can be written as

B(ln n)a 6 ~ Tk(b) Tk(qi). (7.7.5)

r q{
Further, q' E Ap. Hence by (7.3.17), when T/4 ~ p ~ 1,
! 1 = BnP (In nrK•o, (7.7.6)
nP~q~:s;,;;2nP

where K20 = K 20(K) -+ oo with K. Hence

! 'TiqD ~ ( ! 1)1'2( ! 'T:(m)·)112

nP:s;;;q~~2nP nP~q~:s;,;;2nP m:s;,;;2nP

= BnP(ln n)-K 21 • (7.7.7)

This makes it possible to sum (7.7.5) over q~ and obtain the estimate

B(ln n)a·~ 'Tk(<5)(ln nrK = 21 Bn 'Tk(b) (In nrK••. (7.7.8)

r r
In (7.7.8) we shall sum over <5 r; this gives us I
Bn 7k(r) (In n)-K 22 . (7.7.9)
r
Summing over r ~ r0 , we obtain
Bn(In nrK 22+1, (7.7.10)
which gives us the desired estimate (7.7.1). Thus we can write

Qk >(n) = 4 !
2
r:s;;;ro
X4(r) L
do<Q1 ... O'k:s;;;n
µ(q1)µ(q2) · · · µ(qk) ·

N(q1 ••• qkx 1 ..• xk = n (mod r); q1 ••• qkx1 ... xk ~ n)

+ Bn(ln nrK.. , ql ... qk EAp. (7.7.11)

8. Investigation of Qk0>(n)
Let us return to formula (7.4.3). We were able to obtain an asymptotic
expression for Qk2>(n); now we shall study

Qk0>(n) = ! µ(q 1) ••• µ(qk)W(n, q1 • • • qk), (7.8.1)

ql · •• qk.;;;do
q;eAp

where W(n, q1 ••• qk) is the n.s.e.

q1q2 · · · q~1X2 · · · xk + ~2 + 'f/ 2 = n, (7.8.2)
where q1 ••• qk ~ d0 (thus q1 ••• qk E Ap).
Sec. 8] INVESTIGATION OF QJ.0>(n) 131

Since ~~·+77•=m 1 = 4~p(m x4(p), we can write

Q~O)(n) = 4 L µ(q1) • • • µ(qk) L Xip). (7.8.3)
<11 • •• <lk.;;;do p(na1 ••• <lk"'1 ••• °'k
Put
n1 = exp (In n)"1, (7.8.4)
where cx1 = cx1(oto) will be specified later.
Divide Q~0 >(n) into 3 sums: ~A' ~B' and ~c (here we apply a variant on an
idea of Hooley [43]):
(7.8.5)
where
1 '·
n-q 1 •• • P,.0: 1 .•• "'k =O (mod p)

(7.8.6)
LB= 4 ! µ(q1)µ(q2) .. . µ(qk) ! xlP) 1;
q1 ••• qk.;;;d0 v';;-n1 1 .;;;p.;;;v'nn1
(7.8.7)
!c=4 ! µ(q1)µ(q2) ... µ(qk) ! x4(p) ! 1.
v';n 1<p<n
(7.8.8)
We need to put the sum ~c in a more convenient form. Putting
q1 · · · qkxl · .. xk = n - pl..,
we get
!c=4
1- !
q, .•. q,.o:, . ··"°k,,;;n-.i.v'nn1
- 1) ·
(7.8.9)
q 1 . . . q,.o: 1 .•• zk=n+.<(mod4?.)

Put Q = q1q2 ••• qk; X = x1x 2 ••• xk; QX is odd.

The congruences
QX= n -A.(mod41..) (7.8.10)
and
QX = n + A. (mod 41..) (7.8.11)
will be considered together. We shall show that they are either both solvable or
both unsolvable. For (Q, 41..) = (Q, A.); if (Q, A.) In - A., then (Q, A.) In+ A.,
and vice versa. Putting Q/(Q, 41..) = Q/(Q, A.)= Q0 , (7.8.10) and (7.8.11) are
reduced to the congruences

Qo
X=-- n-1..(mod4--
- (Q, A.)
A. )
(Q, A.) '
(7.8.12)

n+A.( A. ) (7.8.13)
Q0 X = (Q, A.) mod 4 (Q, A.) .
132 A PROBLEM OF HARDY AND LITTLEWOOD {CHAP. VII

Here n - A. and n + A. are odd, because Q0 X is odd. Multiplying through in

these congruences by Q0 1 (mod 4A.j(Q, A.)), we obtain

x =Q0 ~Q~A.; (mod 4 (Q~ A.)),

1 (7.8.14)

X =Q- n + A. (mod 4 _A._).

1 (7.8.15)
(Q, .?.)
0 (Q, A.)
Another remark which will be important later: the congruences (7.8.14) and
(7.8.15) are of one kind in the sense that

( (Q,
~ 4 __J:__)
A.) ' (Q, A.)
-- (~
(Q, A.) '
4-A. )
(Q, A.) .
(7.8.16)

Since (n - A.)/(Q, A.) and (n + A.)/(Q, A.) are odd, this is immediately verified.
(7.8.14) and (7.8.15) are to be solved with the restriction that

QX = q1 ••• qkx1 ... xk ~ n - A.·/;;n. (7.8.17)

We shall show that we can discard those values of A. ~ .J~n1 1 for which
- n
n - A. .Jnn1 ~ K . (7.8.18)
(Inn) 24
An estimate for the n.s.e. n - q1 •• • q~1 ... xk = A.p; q1 •• • q,,,xi ... xk ~ n(lnn)-K2
is
T1o(n - A.p) = B T1 0(n - x)T(x) = Bn(ln n)-K•s,

(7.8.19)
by a lemma of Chapter I.
Hence we can suppose

n - A .J~nl > n(ln nrK••. (7.8.20)

9. Investigation of !B
We shall leave ~..t and ~c for the time being and work on ~B· This sum will
be treated by a combination of the dispersion method and the remarkable ideas
of Hooley (see [43]).
We shall need the following lemma.
LEMMA 7.9.1.
"" -B n (7.9.1)
£..B - (In n )I+ro '
where
To> 0.028. (7.9.2)
Sec. 9] INVESTIGATION OF :l:B 133

The proof of this lemma is complicated and requires a number of auxiliary

constructions.
Turning to (7.8.7), we divide :l:B into two sums:
!B = !~) + !~» (7.9.3)
where in :l:~+> the numbers q1q2 . .. qk run through the totality Q<+> of products
of k numbers such that µ(q1)µ(q2) ... µ(qk) = + 1; q1 ... qk ~ d0; and in :l:~- >
the numbers q1q2 ... qk run through the totality of products of k numbers such
that µ(q1)µ(q2) . .. µ(qk) = -1. We shall consider :l:k+l; the sum :l:~-> may
be treated in an analogous fashion.
We have
!~> = 4(r<+>(n) - r<->(n)),
where r<+>(n) is the n.s.e.
qiq2 · · . q,;K1X2 •.. xk + xy = n (7.9.5)
under the conditions
xy ~ n; .J~n1 1 ~ x ~ .J~n 1 ; x = 1(mod4),
(7.9.6)
qlq2 · · · qk E Q<+>;
and r<->(n) is the n.s.e.
(7.9.7)
under the conditions
xy ~ n; .J~n1 1 ~ x ~ .J~n 1 ; x = -1(mod4),
(7.9.8)
qlq2 · · · qk E QH; qlq2 · · · qk ~do.
Foreachofequations(7.9.5)and(7.9.7), weseparatethenumbersx1 •• • xk = X
into two classes:
CLASS I. Ar: X has at least one prime factor v such that

(7.9.9)
where oc2 < 10-10 •
CLASS II. Au: all prime factors of X are either <d1 or >n1 f 6--eo.
We shall apply a slightly modified version of the dispersion method to the
case XE Ar. Subdivide the numbers XE Ar into s ~ Inn systems A•>, having
exactly s prime factors of the form (7.9.9). Then X in A•> can be written in
exactly s ways in the form
X= vq, (7.9.10)
where v satisfies (7.9.9) and q is to be counted sTiq) times. When XE A•>,
equations (7.9.5) and (7.9.7) can be written in the form
xy + D'v = n, (7.9.11)
134 A PROBLEM OF HARDY AND LITTLEWOOD (CHAP. VII

where D' = q1 ••• qkq is counted the appropriate number of times, D' runs
through the system of numbers established above (with repetitions), and
D' ~ n/v.
Furthermore, x and q1 ••• qk satisfy conditions (7.9.6). An analogous equa-
tion replaces (7.9.7). Next, by finding the usual upper bounds, we see that we can
neglect the solutions of (7.9.11) furnished by D' for which -rk(D') > (In n)K26,
and keep only those D' which are counted no more than
B(ln n)K 26 (7.9.12)
times in equation (7.9.11).
Let us make further preparations for the application of the dispersion method
to equation (7.9.11) and the equation analogous to it. Divide the interval of
variation of v into intervals (v0): [v0 , v0 + v0(ln n)-K21]; correspondingly, divide
theintervalofvariationof D'intointervals(D0): [D0 , D0 + D 0(lnn)-K21]; with
the usual error in the total number of solutions, we can limit ourselves to these
intervals, and consider v E (v0) and D' E (D0) independent.

10. The Covariance of the Solutions of the Auxiliary Equations

Let (X) denote the interval of variation of x in (7.9.6) and (7.9.8), and ( Y},,
the interval of variation of y, which depends on x. Let
U 1(m) = 2 l, (7.10.1)
m=::t11
ze(X); 11e(Yl,,
:r=l 1mod4)

U 2(m) = 2 1. (7.10.2)
m=:i:11
ze(X); 11e(Y),,
:r=-l(mod4)

The dispersion method will be applied here in the form described in § 3 of the
Introduction, where the concept of the covai-iance of the number of solutions is
applied (cf. (0.3.26)-(0.3.36)). Form the dispersion of the difference between the
numbers of solutions (cf. (0.3.26))

V= L (ve(vo>
<Do>
L U 1(n - Dv) - U 2(n - Dv)) 2
,
(7.10.3)

where D runs through all odd numbers of the interval (D 0 ), once each; v runs
through the primes of (v0). We have

V= L (ve(vol
(D 0 l
L U1(n - Dv))
2
L (ve(vol
+ <Do> L U 2(n - Dv))
2

-2 2 (ve<vol
<Do>
2 U1(n - Dv)) ( 2
ve<vol
U 2(n - Dv)). (7.10.4)
Sec. 10] SOLUTIONS OF THE AUXILIARY EQUATIONS 135

Designate the three sums appearing in(7.10.4) as Vi, V3 , and -2V2, respec-
tively. In§ 3 of the Introduction, V2 was called the covariance of the numbers of
solutions of two equations of the form (7.9.11). By the usual arguments of the
dispersion method, we establish that with an error of
B 2 2 [(U1(n - Dv)) 2 + (U2(n - Dv)) 2] = Bv~D 2(1n n)a•
ve(vol <Do> (7.10.5)
we can equate V1 to the n.s.e.
(7.10.6)
under the conditions
XE(X); ZE(X); yE(Y).,; !E(Y)z;
x =1 (mod4); z =1 (mod4), (7.10.7)
where the pairs (x,y) and (z, t) vary independently; (n - xy)/v 2 E (D).
Next, with the same error, V3 is the number of solutions of (7 .10.6) under the
conditions
x E (X); z E (X); y E (Y).,; t E (Y)z;
x =-1 (mod 4); z =-1 (mod 4), (7.10.8)
and the pairs (x, y) and (z, t) vary independently.
With the same error, V2 is the n.s.e. (7.10.6) when
XE(X); ZE(X); yE(Y).,; !E(Y)z;
x = 1 (mod 4); z = -1 (mod 4). (7.10.9)
The last n.s.e. must be doubled and subtracted from the sum of the preceding
ones. By Lemma 2.8.1, we get as a result
V = BD 2 LWn nrK 28 • (7.10.10)
Here D 2 = D 0(1n n)-K2•; L 1 = ~ve<vo> 1; K28 can be made arbitrarily large if K 27
is sufficiently large. Applying the Cebysev inequality in the usual way (see§ 12,
Chapter VI), we find that
_2 U1(n - D'v) - _2 U 2(n - D'v) = BD 2L1(1n n)-K 29 . (7.10.11)
(v,D') (v,D')

Such is the estimate of the difference between the n.s.e. of two equations of the
form (7.9.11). Adding over As>, s <; ~n n, and the intervals (v0) and (D0), we
obtain the estimate
Bn(ln nrKso (7.10.12)
for the difference between the n.s.e. (7.9.5) and (7.9.7) under conditions (7.9.6)
and (7.9.8). This gives the estimate (7.10.12) for the part ~~+>, where X =
x 1x 2 ••• xk E A1 •
136 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

11. Investigation of the Class An

Now let us consider equations (7.9.5) and (7.9.7) when X = x 1x 2 ••• xk E An.
Any such X can be represented in the form
X=up', (7.11.1)
where u consists of prime factors p < d1 , and p' has all its prime factors >n1 16-'o
(consequently, there are no more than 6 such factors).
Subdivide An into two classes as follows:
AW are those numbers of the form (7.11.1), where u <exp (ln d1) 4 =
exp (ln n) 4"'• = d2 ;
Ail are those numbers (7.11.1), where u > exp (ln d1) 4 = d2 •
The numbers XE Aijl have factors u > exp (ln d1) 4 , consisting of very small
prime divisors p < d1 • Hence, applying Lemma 1.1.6, we find by our usual
methods that the contribution of such numbers to the number of solutions of
equations (7.9.5) and (7.9.7) can be estimated as
Bn(ln n)-Kso. (7.11.2)
Now we can restrict ourselves to the consideration of numbers x1 ••• xk =
XEA<lf.
We shall find an estimate for the difference between the n.s.e. (7.9.5) and
(7.9.7) for these numbers by closely following the remarkable method of Hooley
(cf. [43] and Lemmas 1.4.4, 1.5.3).
Let
(7.11.3)

F(m) = ! xil). (7.11.4)

ilm
.Y;;n,-1 ,,-,;i,,-,;v';;n,
Let us examine the remaining part of "I:,~+ l:
!<tr = L L F(n - ql ... qkxl . .. xk). (7.11.5)
q, ••• q•,,-,;d0 "'' ••• "'• e Au m
µ(q1) ••• µ(q,l=l
Plainly,
L Tk(u) IF(n - q1 .•• qkup')J. (7.11.6)
0'1 • • · O'k~do

Here u runs through all numbers <d2 (not only those whose prime divisors <d1);
while p' runs through those numbers, each of them exactly once, whose prime
divisors are all >n1 16-'o.
Put
(7.11.7)
Sec. 12] ESTIMATION OF :EE 137

By (7.11.6),

!ii"">'= B ! T 10 (u) IF(n - up')I =B ! (u) IF(n - up')I

T 10
u~d 3 u~d 3 ,p':s;;;;nlu
p<;;nlu D(n-up')7"'0
= B( L
up':s;;;;_n; u~d3
1)112( L ho(u))2F2(n - up'))112.
up':::;;;;n
(7.11.8)
D(n-up')7"'0 u<;;da
Let
LD = up'~n;
L u:::;d3
1, (7.11.9)
D(n-up')7"'0
LE= ! T~ 0 (u)F2 (n - up'), (7.11.10)
up~n; u~d3

thus
(7.11.11)

12. Estimation of LE
First of all, we shall seek an estimate for :EE; we use Hooley's method.
Put nu= n/u, and introduce the following parameters for each u.,;;;; d3 :
Let x U = n1ton1nnuJ2.
U '
p U = II P~Xu P.
For each integer t = IIp"', put t = 1< 1 >1< 2 >, where
t(l) =II pa; 1( 2 ) =II pa; Pu= II P·
Pl t pit p<;;"'"
2'::s;;;xu 2'>Xu

For each integer Yu < nu, let

JuCvu) = gu(Yu) + hu(Yu); gu(vu) -
_ {1 if Yu is prime, Yu .,;;;; Xu,
0 otherwise,
_ {1 if (vu, Pu) = 1,
hu(vu) -
0 otherwise.
Thusfu(vu) equals 1 when Yu= p' (for given u).
We apply Lemma 1.5.3. According to this lemma, the hypotheses of which,
as can be verified without difficulty, are satisfied, we have

- 1-A(n )y + B~ if (a(l), h) = 1,
""'
£.,
'u(Yu)
J•
= { <p(h) u n In• n
(7.12.1)
Vu~11 n
v,.=a(modk) B--u- if (a(l), h) > l.
h ln 5 n
Here A(nu) depends only on nu and has the estimate
A(nu) = B (ln In n)2 (7.12.2)
In n
138 A PROBLEM OF HARDY AND LITTLEWOOD (CHAP. VII

Turning to (7.11.10), we have

LE ~ u::::;;a
L 'T~o(u) L vu,<:,n/u
F 2(n - uvu)f(vu)
3

= L 'T~o(u) L xil~)xil~)fuCvu). (7.12.3)

u::::;;d3 l1'm1'=Z2'm2'=n-uvu
v';:,,.1- 1<!1',!2'< v':;;n1
For given I{ and /~, the values of n - v uu, for given u, are = 0 (mod [!{, !~]).
Let(!{,/~) = d; then I{ = d/1, /~ = d/2, (/1, /2) = 1, [!{, /~] = d/1/2.
We introduce the following conditions and notation (capital letters in
parentheses) :
(H): {(11, 12) = 1}; (Li): {n1;2n;_-1d-1 ~Ii~ n1;2n1d-1}.
We have
LE = u.:s;;;d3
L = 'Tio(u) L
l1Z2d2m=n-uvu
x!(d)xil1)xil2)fu(vu). (7.12.4)
(L1),(L2),(H)
For given u ~ d3 , divide the inner sum into
"" -- d~l/8
kl "" and k2
"" -- "" ·'
d ""1/8
~n <n
in ~1,
(7.12.5)

We can use Lemma 1.5.3 under these conditions. Consider the progression
uvu = n (mod d/1/2). Let (u, d/1/2) = b for given d/1/2. We have
11 = (L1), (L,l,
L <H> x!Cd)xil1)xil2) Uvu =n
L
(mod d~ 1 I,)
!uCvu).
'
(7.12.6)
d~nU 2 uvu::::;;;n
If b 1 n, then the inner slim is empty. If b J n, and (d/1 12 /b, nU>/b) = 1, a
condition which we shall denote by (K6), then the inner sum is equal to
A(nu)(n/u)(l/cp(d/1/2b-1)) + B(n/u ln5 n)(b/d/1/2). Hence

Li = L A(nu) ~ L x!(d)xili)~~(l!!) + B ___n_ L - 1- (7.12.7)

<o> u <L1>.<L.>; cp(dl1l 2b ) u ln 5 n <L1>.<L.>;<H> dl 1l 2
(H),(Ki;) d;.-n118

Here b J (u, d/1 /J, and b J n.

The last sum is easily estimated (for given b) as
B-n-. (7.12.8)
u ln 2 n
For given b, divide the first sum into two sums: ~3 and ~4 •

La = L x!(d)xil1)xil2) (7.12.9)
(£1),(L,),(K),(Ki;> <p(dl1l2b-1)
nl/8 < d<nll•n1 -1

x!< d)xi l1)xi 12) (7.12.10)

cp(dl1l2b-1 )
Sec. 13] ESTIMATION OF ~2 139

Then by Lemma 1.4.1,

(7.12.11)
The sum ~ 4 is estimated trivially by replacing each term by its absolute value.
The result is
(7.12.12)
Thus for given o, combining (7.12.7)-(7.12.11), we obtain the estimate
B n-o -
In In-n (l nn )
1
6
. (7.12.13)
u Inn
Now by summing over oI u and then over u, we get
Bn ln ln n (ln n1)s(ln d3)''8 = Bn (In n )"'3 ' (7.12.14)
ln n ln n
where
(7.12.15)

13. Estimation of ,L2

Now let us seek an estimate for ~2. Here d < n118 • We use the formula

1 if (ll, 12) = 1,
rt~/.t(t) =
{
0 if (ll, 12) > 1.
•f=!2

We introduce the conditions

n1;2n-1 n112n
(R): __1_ < r < __ 1;
dt dt
n1;2n-1 nlf2n
(S): _ _ 1 < s <--1;
dt dt
(D): d < n118 ;
(DT): d < n118 , t < n 118 .
For given u we have (cf.§ 12):

L
!2 = rst2dm=n-uv 11
µ(t)x!(t)x!(d)xir)xis)fu(vu) = L,18 + t;;;;;,:nL,18 •
t<n
(7.13.1)
(R),(8),(D)

Let ~ 5 = ~t<n'/8; ~ 6 = ~t;;.n•/8· Then

Ls = L
(~),(SJ,(DT)
µ(t)x!(t)x:(d)xir)xh)fu(vu), (7.13.2)
rst dm=n-uvu

Ls = <Rl,(S>.W>
L µ(t)x!(t)x!(d)xir)xis)fu(vu)· (7.13.3)
t';:!;m 118 ,rst2 dm=m-uvu
140 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

In 1:5,
(7.13.4)
Hence
!s = ! µ(t)x!Ct)x!Cd)xir) ! xis)fuC'Pu)
rt 2 dm<n 314 n1 uvu=n-rst 2dm
CR);(DT) CS)

= Brt dm<!.,.
2
n ni
(R),(DTJ
I L
U\lu=n-rst dm
2

111<:;;vu<112
xis)/uC'Pu)I, (7.13.5)

where

Y2 =max [
n - n 112rtmn-1 1
, 1 ,
J (7.13.6)
u
_
Yi-max [
n - n1/2rtmn 2 + 1, 1J. (7.13.7)
u
If we put A= rt 2 dm, then U'Pu = n - AS, and the inner sum in (7.13.5)
becomes
L fu('Pu) - L fu('Pu). (7.13.8)
u•u =n-l(mod 4l) uv. =n+l(mod 4l)
h<:;;~<~ h<:;;~<~

(7.13.8) involves the congruences

U'Pu = n - A (mod 4A), (7.13.9)
U'Pu = n + A(mod 4A). (7.13.10)
Further, u is odd; therefore, (u, 4A) = (u, A).
If (u, A) In - I
A, then (u, A) n + A, and vice versa. Consider the conditions

(~A~u;' 4 ((A~u)f) = l, (~A~u;' 4 ((A~u)f) = 1. ( 7.l3.ll)

Each of them is equivalent to the condition: ((n - A)/(A, u), 2(Aj(A, u))<O) = l;
therefore they are themselves equivalent. By (7.12.1), the right-hand side of
(7.13.8) gives
Y2 - Y1 A(nu) _ Y2 - Y1 A(nu) + B nu(A, u) =B n(A, u) ,
qi(4A/(A,u)) qi(4A/(A,u)) Aln 5 n Au(ln 5 n)
if
n --A2(-A-)<
(-
1
>) -1 (7.13.12)
(A, u) ' (A, u) -
and
B n(A, u) (7.13.13)
Au ln 5 n'
Sec. 14] ESTIMATION OF ~D 141

if (7.13.12) is not satisfied. Then (7.13.5) gives

!s=B ! (A.,u) n =B(A.,n)_n_. (7.13.14)

r,d,m1t 2..;;n u rdmt 2 ln 5 n u ln 5 n
The trivial bound, the sum of the absolute values of its terms, is taken for ~6 :

!s = Bno.9. (7.13.15)
Thus, from (7.13.14) and (7.13.15) we obtain

!2 = B __!!:__ (A., u). (7.13.16)

ln 2 n u
I
Summing over (A., u) u and over u ~ d3 with a coefficient of T 10(u), we
obtain the estimate
B-n-(ln n)'7'4 = B--n-.
19
(7.13.17)
ln 2 n (In n) ·
Combining this with (7.12.14) and considering (7.12.4), we obtain finally
(Inn)"•
!E=Bn--. (7.13.18)
In n

14. Estimation of !D
Now let us seek an estimate for ~D (cf. (7.11.9)}

kn= !
up'~n
1. (7.14.1)
u..;;d3
D(n-up')*O
Put p" = up'. Then

!n~ ! D(n - p") + 1. (7.14.2)

P .. ~n p"~n
O(n-p")..;;cx ln ln n O(n-p") >ex ln ln n

Denote the first of the sums in (7.14.2) by ~D,• the second by ~n 2 ; il(m) is
the number of prime divisors of m; rx E [l, U will be fixed later.
Introduce the region of summation
1/2
(L) !!__ < I < n112 n1,
n1
n112
(M) - - < m < n 112n1'
A 1 ln 2 n
(P) Im= n - p".
142 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

If we have (P), then in I: D l , either 0.(/) < (rx/2) In Inn, or O.(m) < (rx/2) In Inn,
or both.
1+ 1·
'
(7.14.3)
(L),(Pl (L),(p)
n(l)<;;(o:/2) ln lnn Q(m)<;;(o:/2) ln ln n

in the second sum Im= n - p" < n; I> n112/n 1 ; m < n1 ' 2n1 • Thus

2D1 < 2
!L),(p)
1+ 2.
!Ml,!Pl
2
l <n112 n 1
1. (7.14.4)
O(l)<;;(o:/2)lnlnn O(m)<;;(o:/2)ln Inn m<;;n112 /n 1 ln2 n

Let these three sums be denoted by I: 1 , I: 2 and L 3 • Then I: 1 = O(I: 2). Further,

23 = o(+).
In n
(7.14.5)
Consider I: 2 •
22= (M) P"=n(modm)
1=
(Ml up•=n(modm)
1. (7.14.6)
O(m)<;;(o:/2)lnlnn p"<;;n O(m) <;;(0:/2) In Inn up' <;;n

Next we examine
1. (7.14.7)
up' =n(mod m)
up'~n

Let (u, m) = b. If the sum (7.14.7) is not empty, then b In and

u ,
~p =n ( mod -;5m) ; p, ....._/' n .
~ ~ (7.14.8)

We can estimate the number of solutions of (7.14.8) using the well-known

theorem of Brun and Titchmarsh (cf. Lemma 1.3.1); it is equal to

B !!_ _1___1_ = B n In In ·n l i::;::: B n In Inn (m, u~. (7 . 14 _9)

u Inn cp(m/b) In n u m In n mu
For given b = (m, u), b u, u I <d 3, we take all m, u. We have
(m, u) b 1
~~ =~~-= ~~ (7.14.10)
mu m 1 bu 1 b m1 u1c5
Further, O.(m 1) > (rx/2) In Inn; n1 ' 2/n 1 b ln 2 n <m <n 1 1 ' 2n1/b. Put n1 ' 2/b = nij' 2 ;
n0 = n/.J°"J.We have the condition (M~):

(7.14.11)
By Lemma 1.2.1,
2 _!_ = B(ln n)1'°'12 -1+"'5, (7.14.12)
<M'o> m1
0(m 1 )<;;(o:/2) In Inn
where rx 5 = rx 5( rx0, a 2) ---+ 0 with a 0 and a 2, and y"' = a - a In a.
Sec. 14] ESTIMATION OF L,D 143

Now, summing over u1c5, we find that

1 = B !
!- T(u) = B(ln n)"•. (7.14.13)
U1c5 U< da U

Thus, after summing (7.14.9) over c5 Ju, we obtain

B __!!_(In n)1'a.l2 - 1(ln n)"", (7.14.14)
In n
where 1X6 ----+- 0 with IX; (i ~ 5). Further, L. 1 = B L. 2 •
Thus,
!Di= B ~(In nfa.1c 1(Jn n)"•. (7.14.15)
In n
Now let us seek an estimate for L. D 2 • We have
(7.14.16)
up'~n
O(n-Up')>a.lnln n
where
!4= ! 1, !s = ! 1.
m<n up'~n
O(m)>lOln Inn a.Jn Inn< O(n-up').;;lOJnlnn

The first sum is (cf. Lemma 1.2.2)

n
B - - (In ny•. (7.14.17)
ln 2 n
There remains
l. (7.14.18)
up'~n
a.Jn Inn< O(n-up').;;10 In Inn

Let R,, be the set of numbers m, all prime factors of which are less than
n 1t2oinlnn. Then, puttingp" =up',

1+ l. (7.14.19)
meRn n-v"ERn.
O(m).;;lOlnlnn O(n-p") >a. In Inn

We shall call the first of these sums L.6 , the second, L. 7 •

If il(m) < IO In Inn and m ER,,, then m ~ n1 ' 2 and
!6 = O(n1;2). (7.14.20)
Now let us consider L. 7 • Here n - p" ER,,, Q(n - p") > IX In Inn.
n - p" = rp"', p"' > n1120 In Inn, Q(n) > IX In Inn - 1,
(7.14.21)
· e
r < n1-1/20 In Inn, p ,,, pnm.
144 A ·PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Hence
!1 <r<:.n1-1/2oin1nn
! !
p"'<n
i. (7.14.22)
O(r) >In In n-1 n-1!=r1!

Now we apply a generalization of a lemma of Erdos (Lemma 1.3.2).1 Accord-

ingly, when r < ne-°"'Inn and u < n'1o,

L 1 = B .!!. (In In n)
3 = B _.!!. (In In n)3 (7.14.23)
n-uP'=rP ru ln (n/ru)
2 ru ln 2 n ·
Summing over u, we obtain the estimate
n (Inn)«'
B---. (7.14.24)
r ln 2 n
Hence
1
!
r~n r
O(r) >ex In In n-1

= B __!!..._ (In n )"1 (ln nY" = B _!::.._ (In n y..-1(ln n )"1 •

ln 2 n Inn
(7.14.25)
Thus,
(7.14.26)

Comparing this with (7.14.15) and choosing a= e/2, so that Ycx = Ycx/ 2 =
!e ln 2,
(7.14.27)
where
e
y = 1 - - In 2 > 0.0570, ix8 = max (ix6 , ix 7). (7.14.28)
2
Thus, y/2 > 0.0285.
From (7.14.27), (7.13.18), and (7.11.11), we obtain

L~>' = B _!::.._(In nrY12(ln n)"", (7.14.29)

In n
where ix 9 = (ix3 + ix8)/2. Together with (7.11.2), this gives

L~) = B _!::.._(In n)-Y12(ln nr. (7.14.30)

In n
1 Some small necessary changes have been made in the lemma.
Sec. 15] PREPARATION FOR THE ESTIMATION OF ~C 145

Entirely analogously,

_Li1> = B ~ (ln n)-1 t 2(ln n)"9, (7.14.31)

In n
whence we finally obtain (7.9.1), completing the proof of Lemma 7.9.1.

15. Preparation for the Estimation of Le

Now we come to the asymptotic evaluation of ~ 0 ; we use (7.8.9). Here
Lemma 1.6.5, on the number of solutions of
(7.15.1)
plays a central part.
If n did not have small prime divisors, then Lemma 1.6.5 would suffice for
our later purposes. In the general case, however, we have to consider the n.s.e.
(7.5.1) with (ID, D) > 1.
Consider moduli D for which (cf. (1.6.10))
D1 ~ D ~ D1 + D2 ;
D2 = Di 20 ; D 1 + D 2 ~ x 112 exp(-lnx)~ 0 • (7.15.2)
(In D 1)
Pick out all of these D, with given residues ID, such that
(ID, D) = p';_' · · • Pf' (7.15.3)
is fixed. Let
(7.15.4)

The original congruence is equivalent to

X1 ••• xk =- z<ll
D
(mod D(l))., X1 ••• Xk ./
~ X. (7.15.5)
(ID, D)
In the particular case (ID, D) = pµ (a prime power;µ> 0). Letp8 ; II xi;xi = p8'X:;
p { x: (i = 1, 2, ... 'k). By (7.15.3), S1 + ... + sk ~ µ; put (J = S1 + ... +
sk - µ ~ O;
p• 1+ ... +sk-µx~ ... x~ = z_g> (mod v< ».1 (7.15.6)
I
If p Dm, then, clearly, s 1 + ... + sk = µ. Let p 1' nm; (J ~ 0. The congru-
ence (7.15.6) is equivalent to

(7.15.7)

(where p-fJ is taken mod Dm symbolically).

146 A PROBLEM OF HARDY AND LITTLEWOOD (CHAP. VII

For functions of two variables M(~, 'Y)), we introduce the ring of operators Lq,
depending on a positive integral parameter q, such that Lq Lq = Lq q ;
LqM(~, 'I")= M(~q-1, 'Y)q- 1 ). Put!~>:= /l}>p-P(mod DUl); X= xp-µ; 1M(X, /'};>)1:!::
N(x1 . . xk = /<_J> (mod D< 1l); x 1 ... xk ~ X). If (q, DUl) = 1, then LqM(X, i'J>) =
M(Xq- 1 , /';)lq-1), where q-1 is taken mod D(l). Then the n.s.e. (7 .15. 7) is

LpP(l - LptM(~' 1D).

pµ pµ
(7.15.8)

For the case p -r D(l), the total number of solutions of (7 .15.6) for any s1, ... , sk
such that s 1 + ... + sk > µ, is

L Lpsi+ ... +srµ(l -Lp)kM(Xµ' 1~)· (7.15.9)

•1+ ... +s.;;;.µ P P
I
Now let p D(l). Then s1 + ... + sk = µ, p -r !£>, and (7.15.6) becomes:
~~ ... x~ = !£> (mod D<1l). Here it automatically follows from x 1 ... xk =
i''f> (mod D< 1l) that p -r x1 ••. xk, for p -r !)/. Thus, in this case it is sufficient to
.a.11..e the congruence

x 1 ... xk = lW (mod v<1l); x 1 ... xk ~ ~,

pµ
(7.15.10)
and the total a.s.e. (7.15.10) is

N(s1 + ... + sk = µ)M (-pµX -pµ

[D) . (7.15.11)

For fixed D and (ID, D) = pt1 ... Pi' (cf. (7.15.3)) suppose that we have

P1 1' D (l) , P2 1' D(l) , · · · , Pti 1' v< , Pt1+1 ID(l) , · · · , Pt ID(l) ·
0
(7.15.12)
Let Lp,µ denote the operator
LP,µ = L Lpsi+ ... +sk-µ(l - LP)k. (7.15.13)
s1+ .. . +sk~P.
Then the n.s.e. (7.15.5) with fixed Dis

L:1>1.µ1 ... Lp, 1.µ, 1Nlµt1+1) ... Nlµt+l)MCD~ D, 1w), c1.15.14)

where Niµi) = n.s.e. s1 + ... + sk = µi.
The product of operators LP lt µ 1 ••• LP ti, µ t1 in (7.15.14) reduces to a linear
combination of operators of the form Lq. Let
q< 0 > =exp (In In x) 4 • (7.15.15)
Using Lemma 1.1.8, and arguing as in§ 6 ofChaptc: VI, we discover that if
we expand the operator expressions in (7.15.14) in terms of the operators Lq, and
discaril those Lq for which q > q< 0 l, then the error does not exceed

(7.15.16)
Sec. 15] PREPARATION FOR THE ESTIMATION OF ~O 147

Now let us turn to the basic lemma, Lemma 1.6.5. We consider those of the
moduli D satisfying (7.15.2) for which

(lD, D) = it 1 • • • Pft · (7.15.17)

In this case the moduli D(l) = Df(ID, D) obey the condition

D1 < Dc1> < D1 + D2 (7.15.18)

Pi 1 • • • Pft Pi 1 • • • Pft
Now separate from the numbers p 1 . . . Pt the numbers Pa ... Pa , t 1 < t:
in all, 2i subsets of n:umbers; then consider those Dc 1 > sati~fying ~1c,ndition
(7.15.18) such that Pa 1' D(l); ... ; Pa 1' D(ll; Pi D(l) when j ¥- cxi (i = 1,
t
I
2, ... , t 1). Denote these1 D (1) by Da(1)1' ... "" . Now let (cf. (7.15.2))
• t1

D1 + D2 < x112 exp - 2(ln x)'"· (7.15.19)

Then

D1 + D2 < ( x )1 ' 2 exp - 2(ln x)'"· (7.15.20)

Pi 1 • • • Pit it Pit
1
• • •

Therefore, for given (ID, D) = pt1 ... p/t, the fundamental lemma can be
applied in the form

M(P µl
x
pµt'
D(l)) ~
1 • • • t

= -1
2TTi
"" - -
£.. 1-
m<11 l_ ip( D c1>)
f (Pi1 ....
c Po
· x
p~t
)s -1 c·L (s, X m) )k d s
s
(0)
D

+B x !? 2 exp - (In ln x) 50 • (7.15.21)

Pi 1 • • • Pij D1
Here D(l) runs through values subject to cqndition (7.15.18). If La is an operator
with index q < qC 0> (for all values of D(l) under consideration, (q, D(ll) = 1)),
then qD(l) <
q<o> D<ll <
(x/qpt1 ... pf't)1 ' 2 exp - (ln x)'•; hence the funda-
mental Lemma 1.6.5, is again applicable, and we find that

+B x 02 exp - (In ln x) 50• (7.15.22)

qpi 1 ···Pt D1
148 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

16. Investigation of the Fundamental Congruences

Now let us go on to the fundamental congruences
QX = n - A. (mod 4.A.) (7.16.1)
and
QX = n + A. (mod 4.A.), (7.16.2)
where
Q = ql · · · qk; X = X1 • • • Xk; A..;; Jn n11 ;
x ..;; n - A.Jn n1; (7.16.3)
we shall study them together. We have: (Q, 4A.) = (Q, A.). Let Q/(Q, A.) = Q0 •
The congruences n - A = 0 (mod (Q, A.)) and n + A. = 0 (mod (Q, A.)) are
satisfied or are not satisfied at the same time. Suppose that they are satisfied.
Then (7.16.1) and (7.16.2) are equivalent to the congruences

(7.16.4)
and
(7.16.5)

Heren - A.andn +A.are odd(because Q0 Xis); A.1 = A.j(Q, A.). Then(Q 0 , 4.A.1) = 1.
Multiplying through by Q0 1 (mod 4.A.1), we obtain

(7.16.6)
and
(7.16.7)

We rewrite (7.16.6) and (7.16.7) in the form

X = !1(mod 4.A.1) (7.16.8)
and
(7.16.9)
where 11 and!{ are odd; (II> 4.A.1) = (!{, 4.A.1),

X ,,.;; n - A.J~ n1 .
(7.16.10)
(Q, A.)Qo
Note too that
I
U1, 4.A.1) n - A. and
whence
U1, 4.A.1) I n; (7.16.11)
Sec. 17] EXA.MINAJ'ION OF THE FUNDAMENTAL CONGRUENCES 149

17. Further Examination of the Fundamental Congruences

We shall apply the results of§§ 15 and 16 to the study of the number of
solutions of (7.16.8) and (7.16.9) under condition (7.16.10). Put

X;. = n - :;...J~n 1 . (7.17.1)

(Q, A.)Qo
Consider the odd divisors {J of n, and for given Q = q1 ••• qk and (Q, A.),
choose those A.1 for which
(7.17.2)

l '1
I
wherep"' 1 A.1 , ••• ,p"' 1 A.1 ;p; A.1 (j7Jf cx1 ; i = 1, ... , t 1), i.e. numbers of the
form ).(1)
a:l, 0 a:tl
""'
E D(l)
CXl" • "a:tl
(cf. § 15).
The n.s.e. (7.16.8) under condition (7.16.10) is

II Ii )
Niµ;)M ((l 4i ) '(l 4i ) .
X;. (7.17.3)
L1'cx1µ"'1 ••• LJJcx, µex, .
1 1 1 #-ai 1' Jl..1 b Al

The n.s.e. (7.16.9) with condition (7.16.10) is

L 1'cx1µ"'1 .. . L 1'cx, µ"'' II
. N( k µ,
-)M(~ -14i
(l l' 4i/Ll) '(l ~-).)
(7.17.4)
l l Ffaa; l' /Ll
Next, let
(7.17.5)
Further (cf. (7.8.4)),

A < .Jn n11 = .Jn exp - (In n )"'1, (7.17.6)

We shall suppose that cx1 is chosen so that
(7.17.7)
Further (see § 4)
Q < d 0 =exp (In n)"' 0 , (7.17.8)

n - :;...J~ n1 > n(ln n)-K24 (7.17.9)

X;. = (Q, A) (Q, A) '
on the basis of (7.8.20). Thus
). < .Jx.1. exp - t(ln n)"\ (7.17.10)
and, consequently,
).' < .Jx.1..1. exp - t(ln n)"' 1 ; (7.17.11)
in this connection, compare inequality (7.15.20). We also note that since
A.< .J~ n11,
X.1..1. > nl/2-0.001. (7.17.12)
150 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Putting
_1_1__ 1 .
- 2• (7.17.13)
(ll, 4A.1)
we obtain the expressions
(7.17.14)
and
(7.17.15)
for (7.17.3) and (7.17.4).
Write the operator iri (7.17.14) and (7.17.15) in the form 'I:. a(q)Lq. Following
the usual arguments, and applying Lemma 1.1.8, we find that we can throw out
q ~ q< 0 > =exp (ln ln n) 4 with a permissible error; that is, (7.17.14) and (7.17.15)
become
L a(q)LqM(xu, 12) + Bxu
q<;;q(O) 4A
(T(A.')r exp - !(In ln n) 2 (7.17.16)
1

and

where
a(q) = B(T(q))aaa~-t1. (7.17.18)
Here
(1 2 , 4A.') = (I~, 4A.') = 1, (7.17.19)
a;-t 1 ..;;; (T(A'))a10 • (7.17.20)
The congruences X =/ 2 (mod 4A') and X =
I~ (mod 4-A') automatically require
that Xbe odd, for (/2 /~, 4A.') = 1. We can apply the fundamental Lemma 1.6.5,
considering that q ..;;; q< 0 l = exp (ln ln n) 4 in formulas (7 .17. lti) and (7 .17 .17), and
assuming that
r ~~ (7.17.21)
':>O-.... 100.

a 0appears in the hypotheses of Lemma 1.6.5 as $0 .)

All of our arguments have concerned a given c5 In,
c5 = pi 1 • • • Pi', and
numbersA.1ofthetype A.~~l···"-• (cf.(7.17.2)). For such numbers, A.1/pti .. .pti =
1
A.' is also divisible by
Ila,, ... ,a, = IT
1
j:;l=rx.i
P;, (7.17.22)

and not by the remaining p;. Separate off those of the numbers (7.17.22) for
which
ITai, ... ,a,, ~ exp (ln ln n)5 (7.17.23)
(if there are any), and leave those for which
Ila,,. .. ,a,, <exp (In In n) 5• (7.17.24)
Sec. 18] EXAMINATION OF THE FUNDAMENTAL CONGRUENCES 151

18. Continuation of the Examination of the Fundamental Congruences

I
Given pt1 ... pf• n, and Pa l , ... , Pa •1 , we sum formulas (7 .17 .16) and
(7 .17 .17) over all A. 1 of the type A.~~ .... a, . At first we shall assume condition
l
(7.17.24), i.e., that Pa, ... ,pa are small. In formulas (7.17.16) and (7.17.17),
l '1
for our A.',
(q, 4A') = 1. (7.18.1)
Further
LqM(x;.;., 12) = M(X;;., l q- 2
1), (7.18.2)

where /2q-1 is taken mod 4A'. Since the values of q are comparatively small,
Lemma 1.6.5 is applicable. In (7.17.16), for given (Q, A.) and Q, we sum over
A.1 = A.~~~ .... a, in the interval of values of Ai. which we shall denote by Im:
1

m = 0, 1, 2, ... (7.18.3)

(the last narrow interval begins at 1).

By Lemma 1.6.5, for given q,

J.1=J.(l)
z LqM(x;.;., 12)
Elm
a:1, ... 'a:tl

(7.18.4)

where (for given (Q, A.), and considering (7.15.12))

n
Rm= B exp - t(ln ln n)50 • (7.18.5)
(Q, A.)Q0Pi 1 • • • P~'q
By (7.17.1), (7.17.5), and (7.17.9),
n(ln n)-K 24 n
(Q , /1.,)Q 0P1µ1 · · · Ptµ, ~ x).). ~ (Q , /1.,)Q 0P1µ1 · · · Pt'
µ
(7.18.6)

Furthermore, the number of A.1 = A.~~~ ... , a, 1 E Im is

~ .J~ n11(ln nr1
? C1 2m µ1 µ,
P1 · · · Pt
ITai. ... , a, (Q
1
1)
' II.

and

(7.18.7)
152 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

for A' = /..1 /pr 1 ••• p/ must be divisible by TI" 1, . . . , "t1 . Therefore, in (7.18.5)
1

we have
n (In n) Kz
= B x_.1.=.1.'-'--~--
2 4
B
(Q, /..)QoPi1 · · · Pf'q q

(7.18.8)

Because of(7.17.24) and (7.18.8), we now have

Rm = B exp - -Hin In n) 50 L xu (7.18.9)

.l.1=.l.(l) Elm 4/..' q
av···, cct 1

as an estimate for (7.18.5).

Summing (7.18.9) over all C1t 1 , ••• , Cltt,, and then over Im, we obtain

B exp - i(ln In n)50 L X.1..1. • (7.18.10)

(J.1) 4/..' q

Here Q and ( Q, /..) were fixed. As can be seen from formulas (7 .17 .16) and
(7.18.4), we need to multiply this estimate by a(q) = B(r(q))as(T(/..'W 1 • (cf.
(7.17.18)-(7.17.20)), and sum over q < qco>. For given Q and (Q, /..),we obtain
the estimate
B exp - !(In In n)50 Xu (T(A')r 11 . (7.18.11)
4/..'
Summing over all Q = q1 ... qk < d0 and /.., we easily obtain the estimate
Bn exp - !(In In n)5°. (7.18.12)

19. Estimation of Le
We have summed (7.18.5) and obtained the estimate (7.18.12) under the
condition (7.17.24). With condition (7.17.23) our arguments do not go through,
because the factor n"1· ... '"t1 in (7 .18.8) may be too large. In that case we
replace LqM(x;..1., 12) by the trivial estimate
B X;.;.(ln n)a1
(7.19.1)
qA'
which can be derived from an estimate of the n.s.e.
X1 ••• Xk =: [Cl) (mod 4/..1), X1 ... Xk < X;.. (7.19.2)
Sec. 19] ESTIMATION OF ~a 153

Multiplying (7.19.1) by a(q) and summing over q..:;; q< 0 >, we obtain, as before, in
the derivation of (7.18.11), the estimate

(7.19.3)

Summing over A., (Q, A.), Q, and taking into account the fact that A.' =0
(mod Ilcx 1, ... , "'t ), we easily obtain
1

Bn ! (Q, A.) (r(A.))a 12(1n n)au. (7.19.4)

;.<vii QA.
Put (Q, A.)= Q1, Q = Q'Q1; A.= A.'Q1. Then (7.19.4) gives

Bn "" (In n)a13(r(A.'W1•(r(Q1))a12 = B n(ln n)a1•ai4t1.

£.., (7 .19.5)
'
Q', Qi.J.'~v';;
J. Q1 =o (mod Ilai. ... ,a11 )
Q'A.'Q1 rr"I>····"· 1

It remains to find an estimate for

!
din
(7.19.6)
d>exp (ln lnn)s

By Lemma 1.1.8, an estimate for this is

B 0 (1n nr 0 , (7.19.7)
where C is arbitrarily large; as usual, Ba is bounded by a constant which depends
only on C. Taking C sufficiently large in comparison with a 13 and a 14 , we obtain
the estimate
Bn(ln n)-K31 (7.19.8)
for (7.19.5).
Thus by (7.19.8) and (7.8.12), summation of (7.19.5) gives
Bn(ln n)-K••. (7.19.9)
Returning to (7.17.16), we sum the second terms over Q, (Q, A.) and A.. By
the usual methods, we obtain as a result

Bn exp - l \ (ln ln n) 2, (7.19.10)

which, together with (7.19.9), gives the estimate

Bn(ln nrK• 2 • (7.19.11)

Exactly the same thing is obtained in the study of (7.17.17). The first terms on
the right-hand side of (7 .18.4) coincide for (7 .17 .16) and (7 .17 .17), so that when
(7.17.17) is subtracted from (7.17.16) and the result is summed over all Q, (Q, A.)
and A., we get
Bn(ln n)-K•• (7.19.12)
154 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

as an estimate for the sum of the absolute values of the differences between
(7.17.17) and (7.17.16). Hence, returning to (7.8.9), we obtain
(7.19.13)

20. Investigation of LA
Returning to (7.8.6), we now seek an asymptotic expression for ~A·
Put
r 0 = exp (In ln n) 4 , (7.20.1)

(7.20.3)

In both slims qi ... qi?Ci .. : xk < n. Then

L.A = LA1 +L.A.- (7.20.4)

Let tis take an upper bound for ~A• Let us seek an estimate for

L xiP) L 1 (7.20.5)
r0 <p,,;;,v-;./n 1 n-Q:i:1. ··"'k"' O(modp)
Q:C1 • • · Xk::::;;;n

for fixed Q = q1 • . . qk (wij:h the appropriate number of-repetitions).

A necessary condition for the solvability of the congruence
n - Qx1 ••• xk = 0 (mod p) (7.20.6)
is
CQ, p) In. (7.20.7)
Putting
n n· Q Q· p -p (7.20.8)
(Q, p) = i' (Q, p) = 1, (Q, p) - 1,

we reduce (7.20.6) to the congruence

Q1x1 ... xk = n1 (mod Pi), Q1x1 ... xk < n1. (7.20.9)
Here (Qi, Pi) = I. Further let

---
ni - n.
2, (7.20.10)
(n1, Pi)
Then (7.20.9) is equivalent to

Xi ... xk < ni = n2 (7.20.11)

(n1, Pi) Q1(ni, Pi) Qi.
Sec. 20] INVESTIGATION OF ~A 155

We can apply the results of§ 15 here (cf. (7.15.14)). Observe that p can be
assumed odd, in view of the presence of x4(p) in (7 .20.5); p2 j p is also odd.
The numbers x 1 , ••• , xk were assumed to be odd. Hence we need to multiply
the operator in (7.15.14) by a factor of (1 - L 2)k. Following§ 15, we put
(7.20.12)
where
(7.20.13)

Then the n.s.e. (7.20.11) is

(1 - L2lLPa. µa. . . . Lp« µa.
l l ti
IT Niµ;)M(n 2 , n2 ),
tl ;-:l=a.i Ql Ql
(7.20.14)
where
(7.20.15)
(thus the function M($, 'Y/) depends implicitly on p 2).
We must apply the Fundamental Lemma 1.6.5 to the quantities
M(n 2/ Qi. n 2 / Q1). Here we have

(7.20.16)
Since
p~ .Jn nl 1 = .Jn exp -(In n)"1, (7.20.17)

p2 ~~exp -t(ln n)" 1, (7.20.18)

and because of(7.17.21), Lemma 1.6.5 is applicable. Here, just as in§ 17, we can
suppose that in the expansion of the operator product in (7.20.14) in terms of the
operators Lm, it is sufficient to consider only the values
m ~ m< 0 > =exp (In ln n) 4 , (7.20.19)
the error, for given Q and p, being

B ..!!.!... exp - !(ln ln n) 2 • (7.20.20)

Q1P2
Following (7.20.13) and § 18, let us make the necessary division of the
interval
(7.20.21)
with respect to the possible values of the odd numbers p. For given Q and (Q, p),
consider the numbers p1 = p/(Q, p). Introduce the intervals//{

((Q,V~n-1
~~~l=---
v~n-1]
p)2Hl' (Q, p) 2g
l ' (7.20.22)
156 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

where the extreme left interval must begin at

min (1;(~op))· (7.20.23)

Fix (ni, Pi) = pf1 ... Pi' (cf. (7.20.12)), and denote the Pi with given (ni, Pi) by
p~~>, ... , ex, if they obey condition (7.20.19) (p 2 = Pi/(ni, Pi)). As in§ 17, we
i
observe that when Pi = p~~>, . __ , ex, , (7.20.14) equals
1

(7.20.24)

where
(7.20.25)
Here
(7.20.26)

LmM(n 2 , n 2 ) = N(xi ... xk =Qin m-i(mod p

2
2 ); Xi ... xk ~ ...!!J._). (7.20.27)
Qi Qi Qim
Since m ~ exp (In In n) 4 , and because of (7.20.23), we can apply the funda-
mental lemma. For fixed m, we multiply (7.20.27) in (7.20.24) by xip), and
sum over p~~>, ... , ex, Elg (Q, (Q, p) and (ni, Pi) are fixed).
1
We can apply Lemma 1.6.5 separately for the cases p = 1 (mod 4) and
p = -1 (mod 4), and subtract the results. The number of Pi = p~~>, ... , ex, E lg
1
is (cf. (7.18.7))

and
(7.20.28)

Thus (cf. § 19)

P1=P(i)
L xiP)LmM
elg
(nQi
2 ' n2 )
Qi
«1, ••. ,«t1

X IIa1, ... ,a, 1 ( ""

£,,,
- (1-)exp
)
- l(ln In n) 50• (7.20.29)
P1=P~l) a elg <p P2
l• .•. ' t1
Sec. 21] ESTIMATION OF THE AUXILIARY SUM 157

First sum the first term on the right-hand side of (7.20.29) with respect to
cxl> ... , cx 11 and lg. We obtain, for fixed Q, (Q, p) and (ni, Pi),

(7.20.30)

21. Estimation of the Auxiliary Sum

We need to multiply (7.20.30) by a(m) and sum over all possible values of Q,
(Q, p), (ni, Pi), cxi, · · · , ext 1 , m. -
Given Q, let p run through the odd numbers of [p 0 , n! 1 ] with a given .Jn
value of (Q, p). With respect to this value of (Q, p) we define

(7.21.1)

Further, (ni. Pi) = pt 1 ••• pf', and the subdivision cx1 , ••• , cx 11 , which gen-
erates the number Ilcx 1 , ••• , "'n' is fixed; p 2 = 0 (mod Ilcx 1 , ••• , cx,J
Thus for given (Q, p), all the prime factors of m divide (ni, Pi) = (n/(Q, p),
p/(Q, p)). Therefore with (Q, p) we also need to fix (ni, Pi), i.e., consider p such
that (Q, p) = QP and (ni, p1 ) = nP are fixed.
Suppose this is the case; then m consists only of primes dividing (ni, Pi),
n2 = n/(Q, p)(ni, Pi); Qi = Q/(Q, p) is constant; in (7.20.30) only (xip)/9:{p2))
=
x (L(s, x0))1' varies for Is - 11 ~ p 0 ; p 2 0 (mod Ilcx 1 , . . . , "'• 1) ; P; 1' P2 (j :;C ct;).
Further L(s, Xo) = ~(s)IIPIP 2 (l - 1/p'); (L(s, Xo))k = ({(s))kIJPIP.(1 - ljp•)k.
Thus we need to sum

xiP) II
cp(p/(n, p)) :i>l(p/(n,p))
(i - l)k = xiP) (n, p) II (1 - 1/p't
p8 p :i>l(p/(n,p)) 1 - ljp
(7.21.2)

over p; (Q, p)(n/(Q, p), p/(Q, p)) = (n, p), (Q, p) fixed, (n/(Q, p), p/(Q, p)) fixed,
with the indicated conditions for p 2 • If (Q, p) and (n/(Q, p), p/(Q, p)) are fixed,
then so is their product (n, p).
First of all let (Q, p) = 1, (n, p) = 1, (p, Qn) = 1, m = 1; sECp0 , p0 = 1/lnn;
there are no additional conditions on p 2 • Consider

2 xip)II(1 - P~t. (7.21.3)

(p);p;;,ro p PIP 1 - P
(p,Qn)=l

This is a segment of the series formed by the product

II
p-f"Qn
(i + xiP)
p"
+ xip + ... )<1 -
2

p 2"
) 1/p•)k.
1 - ljp
(7.21.4)
158 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. Vil

To compute (7.21.3), let PI> p 2 , ••• , Pt denote the primes dividing Qn, and
take
L x4(p) II<1 - p-•)
Vn/n 1 ~p~r0 p PIP 1- p -l

_ _ L xiPP1) II (1 - P~t + ... _ .... (7.21.5)

Vn/n 1 ~pp 1 ~r 0 PP1 PIP 1 - P
We shall show later that

(7.21.6)

If Pcx1Pcx2 • • • Pcx, < r612 , then, by (7.21.6), the corresponding terms of (7.21.5)
have the estimate Br0°· 001 /p"' 1 ... Pcxs ; if, on the other hand, Pcx 1 ... Pcxs >
exp t(ln In n)4 , then the same terms are estimated as B/p"' 1 • • • Pcx.- The sum of
these last terms, which divide Qn and are greater than exp i(ln In n) 4 =r6' 2 , gives
B exp - Hin In n) 2 , and the preceding terms give in total Br0 °· 0005 ; altogether
the sum is B exp - t(In In n)2. The estimate (7.21.6) follows easily from the fact
that
II (1 + X4(;) + xi;,2) + ... )<~ - P~t = U(z)L(z, x4),
p p p - p
where U(z) is regular and bounded both above and below in the region a > 0.9.
Thus, for the case (p, Qn) = 1, (7.21.3) has the estimate
B exp - t(In In n) 2 • (7.21.7)
Now let (Q, p) and Q0 be fixed; (Q, p) Jn; and let (n/(Q, p), p/(Q, p)) = '51
be given. Then

P = P1(Q, p); (P1' (Q~ p)) = 1, (7.21.8)

P1 = P2c51; (P2; (Q, np) c5) = 1. (7.21.9)

Thus,

( P2' _JL_ n ) = 1 (7.21.10)

(Q, p) (Q, p) c51
(we are assuming the conditions given earlier for p 2).
Consider the term (n 2/mQ 1)' in formula (7.20.30). We have

n1 n1
n2= - - =
(n1, P1)
(n/(Q, p), p/(Q, p))
n n n2 n
= --- ; --=---- (7.21.11)
(Q, p)(n/(Q, p), p/(Q, p)) (n, p) mQ 1 (n, p)mQ 1
Sec. 21] ESTIMATION OF THE AUXILIARY SUM 159

Since on Cp0 , Res= 1 + 2/ln n, this term introduces a factor of Bn/(n, p)mQ 1
into the sum. This must be taken into account in summing (7.21.2). In this
summation, we divide the terms in the following manner: (n, p) >exp t(ln ln n) 4
and (n, p) ~ exp !(ln ln n) 4 • In the first case, consider the sum of the series
(7.21.2) when r0 ~ p ~ .J-;;/n1 • Here the term

(n, p) ! x4(p) IT (1 - p-•)k

(7.21.12)
(n, p)" p Pl<Pt<n,p)) 1 - p- 1

is singled out. Here r0 ~ p ~ .J-;;/n1; p > (n, p) = (Q, p)(n1, p1) (both num-
bers are fixed, which determines the region of values of p).
First let
(7.21.13)

(7.21.12) always gives the estimate B In n/(n, p), and counting the contribution
of Bn/mQ1 [from (7.21.11), from the number of subdivisions ix1 , • • • , ixt 1 ,
which does not exceed T((n 1, p1)), and from the fact that Q1 = Q/(Q, p)], we
obtain
Bn Inn T(n1, P1) = Bn In nT((n1, P1)). (7.21.14)
mQ 1(n, p) mQ(n1, Pi)

Now we introduce the coefficient a(m) = B(T(mW 11 T((n1 , p 1 W 1 • and sum

B n In n(T(m))417T((n1 , p 1 ))4 2 •
(7.21.15)
mQ(n1, P1)
over m for fixed (Q, p); over (n 1 , p1 ) >exp t(ln ln n) 4 ; and then over (Q, p).
Then (7.21.15) obtains the estimate

Bn Inn ! (T(m))a 17T(Q) ! (T(c5))4•• = Bn exp - !(In ln n) 2

m,Q<;;vn o In c5
o>exp(l/2)(1n ln n) 4

(7.21.16)
(cf. Lemma 1.1.8).
Now let (n 1 , p1 ) ~ exp -!(ln ln n) 4, so that

IT ~exp !(In ln n)4 •

If (Q, p) ~ exp !(ln ln n) 4 , then the sum of the series (7.21.12) under such
conditions gives, summing over p1 ,

B
--ro -o.ooos/4. (7.21.17)
(n, p)
160 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Finally, after summation over the subdivisions, we get the estimate

B n-r((ni, Pi)) r;;-o.ooos/4 = B nT((ni, Pi)) exp - a2iOn In n)2, (7.21.18)
mQ(n1' Pi) mQ(ni, Pi)
so that, in summing over m, (Q, p) I Q, (ni, Pi) In, we need to take

Bn (T(m)t11(T(ni, Pi)r•oT(Qt•o exp (a22 - In In n)2. (7.21.19)

mQ(ni, Pi)
Summing over m, Q, and (ni, p1 ), we obtain the estimate
Bn exp - a 23(ln In n) 2 • (7.21.20)
Now let
(ni, Pi)< exp Wn In n) 4 , (Q, p) >exp !(In In n)4 •
Then separate off (Q, p) in summing (7.21.2), and replace (7.21.2) by the
estimate B In n/(ni, Pi)(Q, p). Using (7.21.11), this gives the estimate
n Inn
B------
(ni, Pi)mQi(Q, p)
Summing over (Q, p) >exp !(In In n) 4 , since (Q, p) In, we obtain the estimate
n Inn 2
B exp - Wn In n) . (7.21.21)
(ni, Pi)mQi
Summing over (Qi, Pi) In, Qi I Q, and m with a coefficient of
(T(mW•1 (T((ni, Pi)))a20,
we obtain the estimate
Bn exp - a 2i(ln In n)2. (7.21.22)

22. Estimation of the Error Term

Let us sum the error term (7.20.20) over p 2 , Q, (Q, p) and (n1' Pi). We have
(cf. (7.20.7)-(7.20.10))
n 2 = - - - n_ _ _ . p _ p (7.22.1)
' 2-
Qi Qi(Q, p)(ni, Pi) (Q, p)(ni, Pi)
For fixed Q, (Q, p), (ni, Pi) and, hence, fixed Qi = Q/(Q, p), we sum (7.20.10)
over p 2, supposing Pi = p~~~ ... , a 11 • We obtain

(7.22.2)

Summing over Qi< d0 , (Q, p) < d0 , (n1' Pi) In, we obtain the error
Bn exp - t(ln In n) 2 • (7.22.3)
Sec. 23] FINAL CALCULATION OF Q~(n) 161

We still need to obtain an estimate for the sum of the second terms on the
right-hand side of (7.20.29), after multiplying it by a(m), which has the estimate
(7.20.6). As in § 17, we distinguish the two cases

ITcz 1, ... ·"ti ~ exp (In ln n) 5 (7.22.4)

and
Ila1, ... ,a11 < exp (In ln n)5 • (7.22.5)
When (7.22.5) holds, the second term on the right-hand side of (7.20.9) gives

~~n n~n
B - - exp - !(In ln n) 50 = B exp - !(In ln n) 50 . (7.22.6)
mQ1 mQ1(Q, p)(n1, Pi)

Multiplyingbya(m) = B(r(m))"1°ai"0t1 = B(T(m))"1° exp a 11 (ln ln n) 5 (considering

(7.22.5)), and summing over Q1 ~ d0 , (Q, p) ~ d0 , (ni, p1 ) and then over all In,
cx1 , ••• , cxt 1 and lg, we obtain

Bn exp -i(ln ln n) 50 . (7.22.7)

Now suppose that we have (7.22.4) instead of (7.22.5). For such combina-
tions of Q, (Q, p), (n 1 , p 1) = p';_ 1 ... pf' and cx1 , ••• , cxt 1, we given the trivial
upper bound for LmM(n 2 /Qi, n2/Q1) (cf. (7.20.27)):

B ___!1_ (ln n)" 12 • (7.22.8)

Q1mP2
Multiplying by a(m) = .B(T(m))"11(T(p2))"10, and summing over p2 =0
(mod Ilcx 1, ... , °'t ) and CXi, ••• , CXt 1, We Obtain
1 .

B ! n(T(m))" 11 (r(p 2))"1" = Bn exp - t(ln ln n) 5 (7.22.9)

p 2 =0(modITcxv····"'•i>.P 2 .;;;v';;: Q1 (Q, p)(n 1, p1)mp2
Qi,(Q, p)(nv P1),cx1•··· ·"'•1

(cf.§ 19; (7.19.5)).

The same estimate is obtained for the suin of the absolute values of the terms
of (7 .20.30) for p 2 = 0 (mod Ilcx 1, .... ex)• under_ condition (7 .22.4). Therefore we
can take them into account in summing (7.20.30) in § 20.
Combining estimates (7.20.30), (7.22.3), (7.22.7), and (7.22.9), we obtain
(7.22.10)

23. Final Calculation of Q~(n)

Now we come to the derivation of the final form for Q~(n) (cf. §§ 3, 4). By
(7.3.5), (7.3.6), and (7.4.1), we have

Q~(n) = Q~1 >(n) + Bn(ln n)-K 31 • (7.23.1)

162 A PROBLEM OF HARDY AND LITTLEWOOD [CHAP. VII

Furthermore, by (7.4.3),

Qki>(n) = Qko>(n) + Qk2>(n), (7.23.2)

the expression for Q.l,2>(n) being given by (7.6.19).
Further, by (7.8.5),
Qk0 >(n) = LA + LB + Le· (7.23.3)
By (7.9.1), (7.9.2), and (7.22.10),

Qk0 >(n) = LAi + B (Inn)n +ro 1 , -r 0 > 0.028. (7.23.4)

Thus, from (7.23.1)-(7.23.4),

Q~(n) = LA1 + Qk2>(n) + B (Inn)n i+ • (7.23.5)

ro

Considering (7.6.15), (7.6.16), and (7.7.1),

Q~(n) = 4 L xir) L µ(qi) . .. µ(qk)

r~ro ql . . . qk~n

+B n (7.23.6)
(In n)l+ro

Here q1 •.• qkx1 .•. xk is odd, qi E Ap., Then from (7.23.6) we find that

Q~(n) = 4 L xir)N(x~ ... x~ = n (mod r); x~ ... x~ :e;;; n) +B n1 ,

~ ~aj~
(7.23.7)
where, as before, x~ E ilp.
For the further transformation of (7.23.7), we need to apply the remarks of
§ 6 about exceptional characters and "Siegel zeros" of the corresponding
Dirichlet series. When r :e;;; (Inn) K32, we obtain, by standard methods (cf. § 6
and Lemma 1.6.3),
N(x~ . .. x~ = n (mod r), x~ . .. x~ :e;;; n) = L~(n) + R(n, r), (7.23.8)
<p(r)
where
R(n, r) = B ~(In n)-Kaa. (7.23.9)
r
If
(7.23.10)
Sec. 24] COMPLETION OF THE PROOF 163

then (7.23.9) holds, provided r is not divisible by any of the exceptional moduli
r 1, r2 , ••• , rt (cf. § 6); if r is divisible by any of them, we need to take the
trivial bound
B!!. (7.23.11)
r
for (7.23.8).
Summing (7.23.11) over r, divisible by ri. r2 , ••• or rt, we obtain, as in§ 6,
the total estimate
Bn(ln n)-K34 • (7.23.12)
If (r, n) > l, then the left-hand side of (7.23.8) is equal to 0. In view of
what was said above, (7.23.7) can be rewritten as

Q~(n) = 4L~(n) 2 xir) +B n i+ro (7.23.13)

r.;;;r 0 cp(r) (ln n)
(r,n)=l
Extending this series from r + I to oo, we easily find, for k ~ 5, that

Q~(n) = 4L~(n) I
r=l cp(r)
xh) + B n
(ln n)l+ro
(7.23.14)
(r,n)=l
This sum immediately reduces to the form

Qk(n) = 7TLk(n)Ao II (p - l)(p - xip)) +B n ' (7.23.15)

pJn p2 - p + x4(p) (ln n)l+ro
where, as before, A 0 = Ilj)(l + xip)/p(p - 1)).

24. Completion of the Proof

Substituting (7.23.15), with k = 1, 2, 3, 4, 5, into (7.2.4), and using (7.2.5),
we obtain
Q(n) = 7rAo II (p - l)(p - xip))! (-l)k-1 L~(n) +B n (7.24.1)
pJn P2 - P + XiP) k=l k (ln n)1 •028
Further, by (6.6.13), we have
r1 ( l)k-1 n n
L - L~(n) = L 1 + B,n 112+• = - + B -2- . (7.24.2)
k=l k p,,;;;n ln n ln n
Substituting this into (7.24.1), we obtain the fundamental formulas (7.1.4)
and (7.1.5).
 CHAPTER VIII

A PROBLEM OF TITCHMARSH CONCERNING DIVISORS

AND PRIMES

1. Statement of the Problem

The problem of divisors and primes was proposed by E. C. Titchmarsh in
1930 [55). It consists in finding an asymptotic expression for

p~n
! r(p - 1) (8.1.1)

as n ~ oo; I is fixed. A natural variant of this is the problem of obtaining an

asymptotic expression for
L
r(n - p).
p:s;;;n-1
(8.1.2)

In paper [55), Titchmarsh established the fundamental theorem: if the extended

Riemann hypothesis is correct, then (8.1.1) has the asymptotic expression
ln ln n
L r(-p ~ l) = E(l)n + Bn - - (8.1.3)
p,;;;;n ln n
where£(/) depends only on/. Further, using the sieve method ofViggo Brun, he
established that (8.1.3) has the form E(l)O(n). Titchmarsh derives a lower
bound for (8.1.1), but here his arguments contain certain inaccuracies.
Later, Erdos studied this problem (for the case l = 1) (paper [34); related
questions are treated in his paper [35)). In [41), Hazelgrove considers, in
particular, a lower bound for (8.1.1), obtaining the result A(l)n/(ln ln n). K.
Prachar also discusses similar questions in his monograph [48, pp. 164--168).
It should be remarked that it was apparently not noticed that
L r(p -
p~n
1) >en

can be derived without difficulty from theorems of the "large sieve" (Ju. V.
Linnik [9]) type, developed by Renyi [50-52).
However, application of the dispersion method allows us to find exact
asymptotic formulas for (8.1.1) and (8.1.2).
Obviously, (8.1.1) is the n.s.e.
p - xy = l; xy <n - I, (8.1.4)
164
 APPLICATION OF THE DISPERSION METHOD 165

and (8.1.2) is the n.s.e.

n = p + xy; p <n - 1. (8.1.5)

Hence it is clear that (8.1.1) and (8.1.2) are analogues of.the Hardy-Littlewood
equation, and an asymptotic expression for them can be found by means of the
dispersion method. In this connection, the computation is, on the whole,
substantially simpler than the corresponding computation of Chapter VII.
In this chapter we shall consider only the case I= 1 of formula (8.1.1):
in that case there exists a convenient system of coherent numbers (see § 6 of
Chapter Ill), permitting considerable simplification of the computation. The
calculation of an asymptotic formula for (8.1.1) and (8.1.2) in the general case is
essentially the same, but it is more unwieldy.

2. Application of the Dispersion Method

Suppose I = 1 and, writing n instead of n - 1, consider the equation

p - xy = 1; xy < n. (8.2.1)
THEOREM 8.2.1. As n ~ oo,

,L T(p - 1) = 315 ~~ 3 ) n + R(n), (8.2.2)

p<;;n 27T
where
R(n) = Bn(ln nr0· 999 • (8.2.3)
In the study of equation (8.2.1), we shall at first follow § 2 of Chapter VII,
keeping the notation introduced there. We introduce the numbers P (cf. (7.2.1))
and r 1 (cf. (7.2.3)), and consider the auxiliary equations Yfc (k r 1): <
x{ x~ . .. x~ - xy = 1; xy < n. (8.2.4)
If Q(n) is the n.s.e. (8.2.1), and Q'(n) is the n.s.e. (8.2.4), then we have a
relation like (7.2.4). Again it is necessary to separate the equations Yfc for
6 < <k r 1 , to which the dispersion method can be applied directly, and the
equations Yfc (k <6) which require a special approach, similar to that of Chapter
VII but not as complicated. First let us consider Yfc with 6 k r1 ; we begin < <
by following §§ 2-4 of Chapter VI. Here Yfc has the form

x~ ... x~ - xy = 1; xy < n.
Taking the system of coherent numbers p" defined below (cf. § 4, formula
(8.4.3)) and letting n~ denote 1 or p", form the equation

x~ ... x~ - xy = n{; xy < n. (8.2.5)

166 A PROBLEM OF TITCHMARSH (CHAP. VIII

By the usual methods we reduce (8.2.5) to a number of equations of the form

D'v - xy = n{; xy < n;
v E [v 0 - v~,v 0 ]; (8.2.6)

Dz = D1 ; v~ = Vo
(ln n)K1 (ln n)K1
Here the number of repetitions of D' in (8.2.6) is B(ln n) 2x, where K is the
<:onstant in the definition of P(cf. (6.1.2) and (6.2.21)). Now form the dispersion
of the difference of the solutions

V' = ,L ( L U(D'v - n{) - L U(D'v - n~))2 , (8.2.7)

D e(Do) ve(vo) ve(vo)

for a pair of coherent numbers n~, n~ (cf. (0.4.2)). As in § 4 of the Introduction

and § 2 of Chapter II, we come to three fundamental equations of the form
(2.2.3). The part of the dispersion obtained for v1 = v2 is easily found to have
the estimate BD 2 v~2(ln nrx• (K2 is arbitrarily large). On the basis of Lemma
2.2.1, we conclude that
(8.2.8)
where K 3 = K3(K1) can be made arbitrarily large with K1 • From (8.2.8), applying
the Cebysev inequality as in§ 5 of Chapter III, we find that the difference between
the numbers of solutions of (8.2.6) for right-hand sides of n~ and n~ is
BD 2v~(ln nrK4 • (8.2.9)
Summing over the intervals (D 0 ) and (v0), we obtain the estimate
Bn(ln nrKs (8.2.10)
for the corresponding difference between the numbers of solutions. Taking into
account the largest possible error in the number of solutions, which can occur
when k = 6 and is given by (6.3.4), we get

Q~(n, ni) - Q~(n, n~) = Bn(ln n)~ 2 fork~ 7, (8.2.11)

Q~(n, n~) - Q~(n, n~) = B n 312 (ln ln n) 2, (8.2.12)

(ln n)
where Q~(n, n~) is the n.s.e. (8.2.4) with n~ on the right-hand side instead of 1;
Q~(n, n~) is defined analogously. Hence we have

r1 ( -1)"-1 n
L (Q~(n, ni.) - Q~(n, n~)) =B 312 (In ln n)4. (8.2.13)
k=6 k (ln n)
Sec. 3] THE EQUATIONS Y~ FOR <5
k 167

3. The Equations Y~ for k < 5

In examining equations Y~ (k < 5), we shall at first follow the reasoning of

§§ 3-5, exactly, keeping the notation of those sections. By means of an elemen-
tary sieve (cf. § 3, Chapter VII), we reduce the Y~ to the equations
dx1 ... xk - xy = 1; xy < n. (8.3.1)
Here the xi run through all the integers. If W(n, q1 ••• qk) is the n.s.e.
q1 ••• qkx 1 ••• xk - xy = 1; xy < n; q1 ••• qk E Ap,
then (cf. (7.3.6))
2 W(n, q 1 ••• qk) = Bn(ln n)-x 6• (8.3.2)
nT<q1···qk,;;;.n

Thus it is sufficient to study only that part of the formula for Q~(n) where
q1 qk
••• < nT = n°· 001, i.e., the sum
(8.3.3)

Introducing the numbers cx:0 = 10-10 and d0 = exp (ln n)"- 0, we let

Qk0 >(n) = qi .. 2
• qk,;;;.do
; Qk2>(n) = do <qi ..2• qk,;;;.nT '
so that
(8.3.4)
The sum Q~2 > may be studied with the help of the dispersion method.
Following § 4 of Chapter VII, we isolate the "bad" (irregular) intervals 10 =
[!n0 ,n0],where the number of q1 , ••• , qk is less than n0/(ln n)K•, and discard them.
Then we consider the intervals which are not irregular (cf. § 5 of Chapter VII);
with an error of (7.5.2), which gives rise to a permissible error in the asymptotic
expression for the total number of solutions, we reduce the problem of calcu-
lating Qf>(n) (following § 5 of Chapter VII and keeping the same notation) to
two equations, corresponding to the two possible signs of µ(q 1 ) ••• µ(qk), of the
form
D'v - xy = 1; D' E (D 0); v E (v0). (8.3.5)

Here D' = qx 1 ••• xk, q is counted/(q) <

Tiq) times; x 1 , ••• , xk run through
the integers.
Form equations of the same type (8.3.5) for the coherent numbers n~ (see
(8.4.3)):
D'v - xy = ni.; D' E(D 0); (8.3.6)
We proceed as in § 9 of Chapter VII. The irregular intervals (D0) with a small
number of D' are thrown away, with a permissible error in the number of
solutions (see § 4 of Chapter VII).
168 A PROBLEM OF TITCHMARSH (CHAP. VIII

Supposing U(m) = :l:xy=m 1, form the dispersion of the difference between

the solutions for a pair of coherent numbers n~, n~ (cf. (0.4.2)):

V' L (ve(vL
= D'e(D U(D'v - nD - !
•E(v 0 )
U(D'v - n~))2 •
0) 0)

Here it can be assumed, with a permissible error in the number of solutions, that
an estimate for the number _of repetitions of D' is

(8.3.7)
(see § 9 of Chapter VII). In accordance with § 4 of the Introduction and § 2 of
Chapter II, we arrive at three fundamental equations of the form (8.2.3). Then
we repeat the arguments of§ 2 and obtain the following estimate for the corre-
sponding difference between the number of solutions:

Bn(In n)-K 8 • (8.3.8)

These arguments are correct for both cases: µ(q1) ... µ(qk) = + 1 and µ(q1) ...
µ(qk) = -1.
Letting W(n, n~, q1 . . . qk) denote the n.s.e. q1 .•. qkx1 ••• xk - xy = n~;

xy < n;
(8.3.9)

so that
(8.3.10)
we get
(8.3.11)

4. Investigation of Q~0>(n). Coherent numbers

Let us return to formula (8.3.4) and examine Q~0 >(n). We have

Qk<o>(n) = °""
£.. µ (q1) · · · µ (qk)W( n, q1 · · · qk) ' (8.4.1)
qi· · · qk.;;ao
q;EAp

where W(n, q1 ••• qk) is the n.s.e.

q1 ... q~1 ... xk - xy = 1, xy < n, (8.4.2)

q1 ••• qk being < d0 (thus q 1 ••• qk E Ap).

Sec. 4] INVESTIGATION OF Qk0>(n). COHERENT NUMBERS 169

Again we introduce a system of numbers coherent to 1, in the form of quasi-

primes p" with exclusion up to Mn= exp (ln ln n) 312 and
(8.4.3)
(cf. (3.6.1)); e1 > 0 is a small constant, to be defined more precisely later.
Along with (8.4.2), consider the equation
q1 ... qkx 1 ... xk - xy = n~; xy ~ n,
where n~ = p", as well as n~ = 1. The n.s.e. (8.4.2) is denoted by Q~0>(n, n~), so
that Qk0 >(n) = Qk0 >(n, 1).
As in (7.8.5), (7.8.6), and (7.8.7), but without the factor xip), we suppose
that
(8.4.4)
where
!A= 2 2 1; (8.4.5)
p,,;;V;;nc 1 q 1 .•. qk°'l ••• °'k"' ni' (mod p)
Q1 ••• O'kX1 ••• xk:s:;;; n+n1'

!B = 2 µ(q1) · · · µ(qk) 2 2 1; (8.4.6)

Q1 ... qk,,;;do V;;n1- 1 ,,;;p,,;;V;;n1 Q1 ••• QkZ1 ••. °'k"' ni'(IDOd P)

!a = 2 µ(q1) · · · µ(qk) 2 2 l. (8.4.7)

Q1 • • • qk,,;;do V;;n1<P<n Q1 ••. Qk°'l ••• °'k"' ni'(IDOd P)

We need to put the sum ~a in a more convenient form (cf. (7.8.9)). Put
q1 ... q~1••• xk = n - pJ... Then

!a = 2 µ(q1) · · · µ(qk) 2 2 1. (8.4.8)

0'1 .•. O'k~do A.~ v';n1- 1 0'1 ... O'kXI ... Xk~n-A ,l;;n1+n1'
Q1 ... qk°'l • •• zk::n 1 •(mod.:!)

As in § 8, Chapter VII (cf. (7.8.20)), we can suppose that

n - ;,..J~n 1 + n~ > n(ln n)-K 1 0.

By definition of then~, which are either equal to 1 or top", it is easy to see that,
with an error in the number of solutions of
Bnl-so (8.4.9)
we can assume that the limits of summation over q1 ••• qkx1 ... x,. in (8.4.5) and
(8.4.6) are independent of n~ and equal to n; so that
q1 ••• qkx1 ... xk ~ n, (8.4.10)

while ,in (8.4. 7) they are equal to n - J.. .J~n 1 , so that

q1 ... qkxl •.. xk ~ n - J.. .J~n 1 . (8.4.11)
With the same error we can suppose that xy ~ n.
170 A PROBLEM OF TITCHMARSH [CHAP. VIII

5. Investigation of LB
First of all let us take up I:B; we follow§ 9 of Chapter VII. By analogy with
formulas (7.9.4)-(7.9.8), we put
~ - ~(+) ~(-)
,,t.,B- ,,t.,B - ,,t.,B ' (8.5.1)
where I;~+> is n.s.e.
(8.5.2)
with conditions

q1 · · · qk E Q<+>; q1 ... qk <d 0; .J~n1 1 < x < .J~n 1 ; xy <n, (8.5.3)

and I:~-> is defined analogously.
Next, as in § 9 of Chapter VII, we separate the numbers x 1 ••• xk = X into
two classes, Ar and Arr. When XE Ar, equation (8.5.2) can be treated by the
dispersion method. Following § 9 of Chapter VII, we reduce equation (8.5.2) in
this case to a number of equations of the form typical for the dispersion method
D'v - xy = n~, (8.5.4)
where D' E (D0); v E (v0) (see§ 9, Chapter VII); here as usual, irregular inter-
vals, containing few D', are discarded (see § 4, Chapter VII).
Let U(m) = I:., 11 =m 1 with the conditions given before for x andy, andform
the dispersion of the difference between the solutions for a pair of coherent
numbers (n~, n;) (cf. (0.4.2))
V' = L (L U(D'v - n~) - L U(D'v - n~)) 2 ;
D'e(D0 ) ve!v 0 ) ve!vo)

here we can assume, with a permissible error in the number of solutions, that an
estimate for the number of repetitions of D' is

(8.5.5)
(see § 9, Chapter VII).
As in § 4 of the Introduction and § 2, Chapter II, we come to three funda-
mental equations of the form (2.2.3), where xy is subject to conditions (8.5.3),
and zt runs independently through the same values. Here the terms of the
dispersion, obtained when v1 = v2, are easily found to have the estimate
BD 2 v~2(ln n)-K 12 , (8.5.6)
where K12 is an arbitrarily large constant; D 2 , v~, as usual, are the lengths of the
intervals of variation of D' and v, respectively.
By Lemma 2.9.1 of Chapter II, we conclude that
(8.5.7)
where K13 can be made as large as wewanttogetherwith(l/lnlnn)ln(D1 /D 2).
Sec. 5] INVESTIGATION OF l;B 171

From (8.5.7), applying the Cebysev inequality, which can be done by following
§ 5, Chapter III, almost word for word, we find that the difference between the
n.s.e. (8.5.4) with coherent numbers ni and n; on the right-hand side is
BD 2 v~(ln n)-K". (8.5.8)
Summing over the intervals (D 0) and (v0), we find that the difference between
the numbers of solutions of (8.5.2) with two coherent right sides ni and n; is
Bn(ln n)-K 15 • (8.5.9)
We have the very same result for the equations associated with 1:~->.
Next we come to the case of XE Au in equation (8.5.2). Following § 11 of
Chapter VII, we put Au = AP + AW, observing that the solutions where
XE AW are negligible (cf. (7.11.2)). But if XE AHl, then X has the form
(7 .11.1 ), where k < d2 , and all the prime divisors of p' are greater than n116-" 0 •
Then equation (8.5.2) takes the form
q 1 • . . qkup' - xy = n~ (8.5.10)
with conditions (8.5.3).
Here q1 ... qkup' < d0d2 <exp (ln n)"'o+4"'2
o
Leto = (q1 ••• qkup', x). If > 1, then In~, so that o o> Mn. Instead of
(8.5.10), we obtain the equation

q1 · · · qkup' n~
at (8 5 11)
0 -by=b, ..
where q1 •.• qkup' < 2n. By Lemma 1.1.5, the n.s.e. (8.5.11) for given x =0
(mod o) is
n n
B(ln nr- = B(ln n)a 1 - • (8.5.12)
X X10

Summing this estimate over x, we obtain the estimate

B(ln n)a1+1 ~.
0
Then summing over fJ ni. I o> Mm by Lemma 1.1.8, we obtain the estimate
Bn exp - i(ln ln n)413 (8.5.13)
o
for the number of solutions of (8.5.11) with > 1. With such an error we can
suppose that (q1 ••• qkup', x) = 1. In that case, according to Lemma 1.3.1 (the
Brun-Titchmarsh Theorem), the n.s.e. (8.5.10) has the estimate

B----n_ __ (8.5.14)
q 1 . • . qkucp(x) ln n
for given x.
172 A PROBLEM OF TITCHMARSH (CHAP. VIII

Summing (8.5.14) under the conditions on x (8.5.3), Vfe obtain the estimate

B n In n1 (8.5.15)
q1 ... qku Inn
Considering (7.8.4) and summing this estimate over q1 ••. qk, u, we obtain
the estimate
Bn- (In n)o.001. (8.5.16)
In n
The same estimate is obtained for the n.s.e. of the equation associated with L)i"'")
for XE Au.
By (8.5.9), (8.5.13), 'and (8.5.16), the difference between the expressions LB
(cf. (8.4.6)) for two values of the coherent numbers n~, in particular, for n~ = 1
and n~ = p", has the form
Bn- (In n)o.001. (8.5.17)
In n

6. The Differences Between the Values of !A and Between the Values of !c

We shall now prove similar facts about the differences between the values of
LA and Le. Here, as we saw in § 4, we can suppose that q1 ••• qkx1 ... xk n in <
<
LA• and q 1 ••. q,JC1 ••• xk n- A. .J-;,n 1 in Le, with a permissible error in the
total number of solutions.
First consider Le. In the congruence
q1 ... qkx1 ..• xk = ni. (mod A.) (8.6.1)
we can assume b = (n 1 , A.) = 1 with a permissible error in the number of solu-
tions. If b = (n 1 , A.)> L then b >Mm and arguing as in the preceding section
in the derivation of (8.5.13), we obtain the same estimate, (8.5.13), for the error
in the number of solutions.
Thus we suppose that (n 1 , A.) = l, whence (q1 ••• qk, A.) = -1. Here
q1 ••• qkx1 ... xk < n - A. .J-;,n1. We shall. apply Lemma 1.6.5, which is basic
for this chapter.
A takes the place of Din this lemma; (n - ~ .J-;,n1)/(q1 ••• qk) takes the place
of x D; the hypotheses of the lemma are satisfied, since A < .J~n1 1 and because
of the definition of n1 • It is important that ix 1 ;;;i. 10ix0 (cf. (7 .17 .7) and § 4).
By the above, with .a permissible error in the number of solutions we can
assume (ni. A.) = l. Then we deduce immediately from the Fundamental Lemma
1.6.5 (cf. (1.6.13)) that

X (L(s, x~0 l)? ds + Rq, (8.6.2)

Sec. 7] REDUCTION OF THE BINARY ADDITIVE PROBLEM 173

where
Rq =B n exp - (In In n) 25 L~. (8.6.3)
ql ... qk (.l.) <p(A.)
For any pair of coherent numbers n~, n; the main terms of (8.6.2) coincide, so
that the differences between the sums on the left-hand side of (8.6.2) have the
estimate (8.6.3). Summing this estimate over Jl, q1 ••• qk we find that the
difference be~ween the values of ~c for two coherent numbers n~ and n; is
Bn exp - (In In n) 2 • (8.6.4)
An entirely analogous statement can be made regarding the difference
between two values of ~A.
Thus for any two coherent numbers n~, n;,
(In n)o.001
Qi0\n, n1) - Qi0 l(n, n2) = Bn (8.6.5)
In n

7. Reduction of the Binary Additive Problem to a Ternary One

Taking (8.3.11) into consideration, we obtain from (8.3.4), when k ,.;;; 5,
( In n)o.001
Q<k1 >(n , n')
1
- Q<k 1 \n , n')
2
= Bn ~~-
Inn (8.7.1)

Ifwe replace W(n, q1 ••• qk) by W(n, n~, q1 • • • qk) in (8.3.2), then, as is easily
seen, the right-hand side of (8.3.2) remains unchanged. Therefore it follows
from (8.7.1) that when k,.;;; 5,
I I I I (In n)0.001
Qk(n, n 1) - Qin, n 2) = Bn (8.7.2)
In n
Together with (8.2.13), this gives

~ (-1t-1 Q'( ') ~ (-1t-1Q'( ') - B (In n)o.001 (8.7.3)

.::.., k n, n 1 - .::.., k n, n 2 - n ,
k=1 k k=l k In n
from which we immediately obtain
(In n)°"OOl
N(p - xy = I
n~ xy = n) - N(p - xy = n~ xyI = n) = Bn -'-~­
In n
(8.7.4)
Thus we have reduced our problem to a ternary one. Let p" run through all
coherent numbers (8.4.3), the number of which we denote by L(n). Then (cf.
(4.6.6)): if T(n) is the n.s.e.
p - xy - n~ = 0, (8.7.5)
then
L
-r(p - 1) = T(n) + Bn(ln n)-o.999. (8.7.6)
p,,;;n L(n)
174 A PROBLEM OF TITCHMARSH [CHAP. VIII

8. Solution of the Ternary Problem

Here we shall follow § 6 of Chapter III and use the notation adopted there.
Let I' denote the interval (8.4.3) of the quasiprimes p". Subdivide [1, n] into
H= [n 2'1]intervals; letpEla; xyElb; a-b=c(cf.(3.6.6)); letn 0 =tn1-•1
(the length of/'). Then, in the notation of (3.6.9),
Q(Ja, lb) = L µ(q)N(p - xy = q~ I PE Ia, xy E lb). (8.8.1)
qeOn
qe2no

With an error of (3.6.12) in the number of solutions, putting

n2 = exp (ln ln n) 4 , (8.8.2)
one can sum over q only up to n2, and consider
(8.8.3)

where the accent indicates that the summation is only over q E Qn· We introduce
the trigonometric sums (cf. (3.6.14))
S1(a, q) = ""
£.. exp 21Tia
- - xy, (8.8.4)
XYElb q

S 2(a, q) = .L exp 21Ti ~ p. (8.8.5)

pEla q
Then
1 q-1
N(p - xy = q~ p I E Ia, xy E lb) = - L S 1(-a, q)S 2(a, q).
q a=O
(8.8.6)

If (al> q1) = 1, then (cf. (3.6.23))

1
S1(a1, q1) = - S1(0, 1) + B T(q1) ln ql S1(0, 1). (8.8.7)
q1 q1 ln n
When q1 < (ln n)c, (a1, q1) = 1:
S (a q ) =µ(qi) S (0 1) +B S 2(0, l) (8.8.8)
2 I• 1 <p(qi) 2 , o (ln n)° ,
by Lemma 1.6.3.
For q1 > (ln nf, q1 ~ n2, by a well-known lemma of Vinogradov [7],
(8.8.9)
We now have
Q(Ja, lb) = .L' µ(q) L L S1(-a1, q1)S2(a1, qi). (8.8.10)
q<;;;n1 q q, I q (a1, qi)=l
Let
Sec. 8] SOLUTION OF THE TERNARY PROBLEM 175

By the above, when q1 < (ln n)c,

p(qi) = µ(q 1) + BT(q1) In q1 , (8.8.11)
q1 q1 Inn
while for (In n)c < q1 < n2,
(In nrc
p(q1) = Ba <p(q1). (8.8.12)
ql
Now put

(8.8.13)
so that for q1 < (ln n)c,
eq 1 = µ(q 1)q1 + BT(q1) In q1 In In n , (8.8.14)
<p(q1) In n
<
while when (ln n)° < q1 n 2, eq 1 is estimated with the help of (8.8.12).
Following § 9 of Chapter IV (cf. (4.9.2)-(4.9.6)) we obtain

Q(Ia, lb)= S1(0, l)S2(0, 1) I' µ(m) I' µ(q~eq 1

m,;;;;n2 m ql ,;;;;nz ql
+ BS1(0, l)S2(0, l)(ln n)-~. (8.8.15)
Then, by definition of eq 1 ,
Q(Ia, lb)= S1(0, l)S 2(0, 1) I' µ(m) L (µ(q)) 2
m,;;;;n 2 m q,;;;;(lnn)C q<p(q)
+ BS1(0, l)S2(0, l)(ln n)-0 · 999• (8.8.16)
The series in (8.8.16) can be extended to oo. We have (cf. [55])

Go= L (µ(q)) 2
= 315s(3),
q<p(q) 27T4
so that
Q(Ia, lb)= S1(0, l)S2(0, l)G 0 + BS1(0, l)S2(0, l)(ln n)-0 · 999
(where the function Bis uniformly bounded for all Ia, lb).
Summing over all intervals Ia, lb (cf. § 8, Chapter III), we obtain finally
315s(3)
L T(p -
p,;;;;n
1) - - - n
21T4
+ Bn(ln n)_0 · 999, (8.8.17)

which proves Theorem 8.2.1.

 CHAPTER IX

OUTLINE OF THE SOLUTION OF A GENERALIZATION

OF THE HARDY-LITTLEWOOD EQUATION

1. Statement of the Problem

A natural generalization of the Hardy-Littlewood equation (7.1.1) is the
equation
n = <p(x, y) + p (9.1.1)
where <p(x, y) is a binary quadratic form and p is prime. Application of the
dispersion method to this equation permits us to establish the existence of
solutions of (9 .1.1) under the natural necessary conditions, and to give a rather
good lower bound for the number of solutions. However, it does not lead
immediately to an asymptotic formula for the number of solutions, although it
appears that sufficient deepening and sharpening of the arguments might lead to
such a formula.
Here we shall only consider the case when <p(x, y) is a positive primitive
quadratic form, <p(x,y) = ax 2 +hxy + cy2 , with discriminant - d = b 2 - 4ac < 0,
not divisible by squares of odd numbers. We shall assume that it is reduced.
As is well known, the genus to which <p(x, y) belongs is determined by the
values of the symbols
(;); pjd; podd; (-1)1 <"'> (9.1.2)

(cf. [1]). These are the genus conditions for <p(x, y).
Let the number m satisfy the genus conditions

(;) - (~); PI d; p odd; (-l)f<m> = (-1) 1 <"'> (9.1.3)

and the condition

(m, d) = 1. (9.1.4)
Let <pa(x, y) = <p(x, y), <p1 (x, y), ... , <p 1(x, y) be reduced representatives of
the genus of <p(x, y). Then by Lemma.2.14.2, the number ofrepresentations of m
by any forms <p;(x, y) of this genus is
'P'(m) = wz
6lm
(-d),
t5
(9.1.5)
176
 OUTLINE OF THE SOLUTION 177

where w = 2 for -d < -4, w = 4 for -d = -4, and w = 6 for -d = -3.

Let xi <5) denote ( ~ d)
The forms <pi(x, y) of one genus are taken into one another by rational
unimodular transformations (see [45])

Si = (cxi Pi) ,
Yi <5i
where cxi<5i - f3iYi = 1 and cxi, ... , <5i are rational.
Let Si= TJri, where Ti is an integral matrix with det(Ti) = ri; ri is an
integer. It is known (see [45]) that it is possible to choose these ri so that they are
prime to any previously specified number.
Suppose that we are given m satisfying condition (9.1.4) and genus conditions
(9.1.3). Then
m = <pi(x,y) (9.1.6)
holds for some i ~ 1. Further, let <piSi = <po, so that
m = <p0(cxix + {3iy, YiX + <5iy).
Put r = r1 r2 ••• r1• Obviously, the number
mr 2 = <po(cxir2x + {3ir 2y, yir2x + <5ir2y) (9.1.7)
is representable by the form <po(x, y). Note that with mr 2 , any number of the
form mr 2u2 (u any integer) is representable by <po(x, y).

2. Outline of the Solution

Form the new equation
n = u 2r 2m + p, (9.2.1)
where for u 2r 2 a suitable value must be chosen from among its possible values;
in particular, u must be even. In this equation, m must run through integral
values satisfying conditions (9.1.3) and (9.1.4).
With respect to all solutions of this equation, using (9.1.5), form the sum
!'Y(m) = w !xib). (9.2.2)
olm
The sum (9.2.2) is a lower bound for the number of representations (9.1.1).
In order to apply the dispersion method, we put {<p} = {u 2r2m}, where m runs
through numbers satisfying conditions (9.1.3), (9.1.4) and is such that
'Y(m) ::/= 0. (9.2.3)
The primes pin (9.2.1) must be chosen such that p = n (mod u 2r 2), sop lies
in a definite arithmetic progression. We denote this condition asp E (U).
178 GENERALIZATION OF THE HARDY-LITTLEWOOD EQUATION

Equation (9.2.1) reduces to the equations Y; (k <r 1):

(9.2.4)
where m is counted 'l"(m) times.
When 6 k < < r1 we apply the dispersion method to (9.2.4). By the usual
methods the equation is reduced to equations of the form
n = u 2r 2m + D'v; v E (v0); D' E (D 0 ); (9.2.5)
D'v E (U). (9.2.6)
We need to subdivide this last condition into conditions v E (U1), D' E (U2),
where (U1 ), (U2) are the sets of numbers of the corresponding arithmetic
progressions.
The conditions on m reduce to the requirement that m lie in a certain set of
arithmetic progressions, which we denote as
m E(M). (9.2.7)
The expression (9.2.2) is obtained as follows: take all possible factorizations
of mas a product of two factors, m = xy, and consider w ~.,11 =m xix).
An expression for A(n, D) may be found without difficulty. The fundamental
equation of the dispersion method takes the form
u 2r 2(v 1 xy - v2zt) = n(v1 - v2); v2 <n 116 exp - .../Inn, (9.2.8)
where xy E (M); zt E (M). We need to sum the expression
!xaCx)xa(z) (9.2.9)
over all solutions of (9.2.8).
The determination of an asymptotic approximation is completely analogous
to the derivation of Lemma 2.10.1.
One can study the equations Yi in this way when 6 k < <
r 1 . When k 5, <
the method of Chapter VII (§§ 3-24) can be applied to the Yi. Examination of
the work of that chapter shows that the central point of the calculation is the
application of the Fundamental Lemma 1.6.5. It plays the same role here. Here,
in sums of the type ~A' ~B' ~c (cf. (7.8.6), (7.8.7) and (7.8.8)), xa(p) appears in
place of xip); and also the conditions n - q1 ••• qkx1 ... xk
2 2
=
0 (mod u2r 2);
(n - q1 ••• qkx1 . .. xk)/u r E (M) appear; the rest is essentially unchanged. Of
course, the computation indicated here is quite unwieldy.
If n, the number to be represented, has no small prime divisor or has only a
limited number of them, the investigation can be made considerably easier by
introducing coherent numbers.
Note that the above arguments also permit us to deduce the solvability of
equation (9.1.1) for large n from the extended Riemann hypothesis, on the basis
of Hooley's method. But such a result is conditional, while the derivation
proposed here involves no unproven hypothesis.
 CHAPTER X

SUPPLEMENT

1. On Simplifying the Calculation of the Dispersion

In many cases the solution of an equation of the form

n = <p + D'v; D' E(D); v' E(v) (10.1.1)

may be simplified, if n has no small prime divisor, by introducing coherent

numbers and considering the dispersion of the difference and the covariance of
the solutions. This was demonstrated in Chapters III and IV. In the application
of the dispersion method in the general case according to the basic outline of § 2
of the Introduction, it is necessary to carry out the unwieldy computation of
A(n, D) and to compare the separate parts of the dispersion V1 , V2, and V3 • As
we saw in Chapters VI and VII, the Vi have the form

V; = U((D), (v)) IT (1 + y;(p, p)) + Rni (i = 1, 2, 3), (10.1.2)

iiflln
where U((D), (v)) is a function of the intervals (D) and (v), y;(p, p) is usually of
the form Bpfp, and Rn is the error term. The calculation of nP:l[n(l + y;(p, p))
presents the basic difficulty; it requires much unwieldy computation.
I
It is clear that with a permissible error, only Pi n, Pi < (Inn)° are significant.
If n has the form n = pPn', where pis a given prime, p <(In n) 0 and all prime
divisors of n' are greater than (Inn)°, then in (10.1.2) there will be only one
factor 1 + y;(p, p). If it is established that V1 , Vi, V3 agree for numbers of the
form pPn', then it can also be established in the general case, because of the
multiplicativity in the first term of the right side of(l0.1.2). Ifp >(Inn)°, then a
similar examination can be made easier by introducing coherent numbers; with
their help one can establish that the first terms of the right side of(I0.1.2) coincide
for i = 1, 2, 3. Then one can use the fact that the functions 1 + y;(p, p) are
polynomials in p±P and p, and therefore their agreement for a bounded number
of values of p for all p implies that they agree for all values of p. These ideas may
facilitate the calculation of the dispersion.
179
180 SUPPLEMENT [CHAP. X

2. Certain Deductions From the Application of the Dispersion Method

to the Binary Goldbach Problem and its Analogues
The binary Goldbach problem, concerning an asymptotic expression for
n = p + p'; n even, (10.2.1)
is inaccessible to the dispersion method in the form given here.
Following § 2 of Chapter VII and keeping the notation of that section, we
reduce (10.2.1) to r 1 equations of the form
n = p + x~ . .. x~; x; E Op (10.2.2)
and then to equations of the form
n = p + D'v; D' E (D 0); YE (v0). (10.2.3)
A(n, D) and V2 and V3 are found without difficulty. The main difficulty is in
the calculation of V1 , which is asymptotically equal to the number of solutions
of the equation
(10.2.4)
where vi, Pi are primes. We cannot solve this equation.
We remark, however, that when primes are involved in equation (10.2.1),
pairs of primes vi, Pi appear in (10.2.4). Note that one need not require that the
numbers vi in (10.2.4) be prime; they may be of a rather arbitrary nature.
If, instead of equation (10.2.1), we consider an equation of the form
(10.2.5)
where pi, qi are primes, and where the problem is to find an asymptotic approxi-
mation for the solutions (since when k ~ 3, l ~ 3, the equations have been
shown to be solvable; see [3; 4]), then an analogous situation results. Let Pi• q;
run independently over certain intervals of values (here, by the way, the question
of the solvability of(10.2.5) remains open). Then with the help of the dispersion
method, the problem of obtaining an asymptotic approximation for the solutions
of (10.2.5) reduces to one of finding an asymptotic approximation for the solu-
tions of
(10.2.6)
where Pk~ n 1 1\p~ ~ n11k, and the number of prime factors involved is I+ 1.

3. On the Equation n = p + x2 and its Analogues

Consider the equations
n = p + x 2, (10.3.1)
n = P1P2 + x 2, (10.3.2)
Sec. 3] ON THE EQUATION n= p + x2 AND ITS ANALOGUES 181

where in the latter case Pi ~ n1-"'; p 2 ~ n"'; ct. > 0; and p 1 , p 2 are independent.
Putting {'P} = {x2}, we can apply the dispersion method to our equations; how-
ever, it does not lead to final results. Let us consider its application to equation
(10.3.1).
Since {'P} = {x2 } forms a sparse sequence, we need to take D < n1 12-• in the
calculation of A(n, D), the expected value of the expression L:(v);n-Dv=z• 1.
Put JJ = p 2 ~ n"'; v = p 1 ~ n1-"'; ct.< t. Then one can obtain a heuristic
expression A(n, D) and find an asymptotic approximation for V2 and V3 • The
asymptotic evaluation of V1 reduces to the calculation of the n.s.e.
(10.3.3)
to which the problem of an asymptotic approximation for the number of solu-
tions of equation (10.3.2) reduces.
 BIBLIOGRAPHY

[1] B. A. Venkov, Elementary number theory, GONTI, Moscow, 1937.

(Russian)
12] A. I. Vinogradov, On numbers with small prime divisors, Dokl. Akad. Nauk
SSSR 109 (1956), 683-686. (Russian)
f3] - - , Application of '(s) to the sieve of Eratosthenes, Mat. Sb. (N.S.)
41 (83) (1957), 49-80. (Russian1
[4] - - , Application of '(s) to the sieve of Eratosthenes. (Correction),
Mat. Sb. (N.S.) 41 (83) (1957), 415-416. (Russian)
[5] - - , Generalization of a lemma of Erdos, Vestnik Leningrad. Univ.
15 (1960), no. 19, 124-126. (Russian. English summary)
[6] A. I. Vinogradov and Ju. V. Linnik, Estimate of the sum of the number of
divisors in a short segment of an arithmetic progression, Uspehi Mat. Nauk
12 (1957), no. 4 (76), 277-280. (Russian)
[7] I. M. Vinogradov, Selected works, Izdat. Akad. Nauk SSSR, Moscow,
1952. (Russian)
[8] I. ·p. Kubiljus, Probability methods in number theory, Gosudarstv. Izdat.
Polit. Naucn. Lit. Litovsk. SSR, Vilna, 1959. (Russian)
[9] Ju. V. Linnik, "The large sieve", Dokl. Akad. Nauk SSSR 30 (1941),
292-294. (Russian)
[10] - - , Some conditional theorems concerning the binary Goldbach problem,
Izv. Akad. Nauk SSSR Ser. Mat. 16 (1952), 503-520. (Russian)
[11] - - , Addition ofprime numbers with powers of one and the same number,
Mat. Sb. (N.S.) 32 (74) (1953), 3-60. (Russian)
[12] - - , Dispersion of divisors and quadratic for ms in progressions and some
binary additive problems, Dokl. Akad. Nauk SSSR 120 (1958), 960-962.
(Russian)
[13] - - , Divisor problems and related binary additive problems, Proc.
Internat. Congress Math. (Edinburgh, 1958), pp. 313-321, Cambridge
Univ. Press, New York, 1960. (Russian)
[14] - - , Some additive problems, Mat. Sb. (N.S.) 51 (93) (1960), 129-154.
(Russian)
[15] - - , All large numbers are sums of a prime and two squares. (A problem
of Hardy and Littlewood.) I, Mat. Sb. (N.S.) 52 (94) (1960), 661-700.
(Russian)
[16] - - , An asymptotic formula in an additive problem of Hardy-Littlewood,
Izv. Akad. Nauk SSSR Ser. Mat. 24 (1960), 629-706. (Russian)
183
184 BIBLIOGRAPHY

[17] A. V. Malysev, On the representation of numbers by positive ternary quad-

ratic forms, Dokl. Akad. Nauk SSSR 89 (1953), 405-406. (Russian)
[18] - - - , Asymptotic distribution ofpoints with integral coordinates on certain
ellipsoids, Izv. Akad. Nauk SSSR Ser. Mat. 21 (1957), 457-500. (Russian)
[19] - - - , A generalization of Kloosterman sums and their estimates, Vestnik
Leningrad. Univ. 15 (1960), no. 13, 59-75. (Russian)
[20] - - - , The representation of numbers by quadratic forms, Dokt. Diss.
Mat. Meh. Fiz.-Tehn., Leningrad. Gos. Univ., Leningrad. (Russian)
[21] K. K. Mardfanisvili, On the number of divisors, Izv. Tbilisk. Inst. 9 (1929),
291-298. (Russian)
[22] - - - , The estimation of a certain arithmetic sum, Dokl. Akad. Nauk
SSSR 22 (1939), 391-393. (Russian)
[23] E. P. Ozigova, Modification of the method of the "sieve of Eratosthenes"
given by A. Selberg, Uspehi Mat. Nauk 8 (1953), no. 3 (55), 119-124.
(Russian)
[24] V. A. Tartakovskii, Die Gesamtheit der Zahlen die durch eine positive
quadratische Form F(xi. x 2 , ••• , xn) darstellbar sind, Izv. Akad. Nauk
SSSR Ser. Mat. 1929, 111-122. (Russian)
[25] I. V. Culanovskii, Certain estimates connected with a new method of Selberg
in elementary number theory, Dokl. Akad. Nauk SSSR 63 (1948), 491-494.
(Russian)
[26] N. G. Cudakov, Introduction to the theory of Dirichlet's L-functions,
OGIZ, Moscow, 1947. (Russian)
[27] - - - , On the density of the set of even numbers which are not representable
as a sum of two odd primes, Izv. Akad. Nauk SSSR Ser. Mat. 1 (1938),
25-39. (Russian)
[28] R. Bellman, On some divisor sums associated with Diophantine equations,
Quart. J. Math. Oxford Ser. (2) 1 (1950), 136-146.
[29] R. Brauer, On the zeta-functions of algebraic number fields, Amer. J. Math.
69 (1947), 243-250.
[30] V. Brun, Le crible d'Eratosthene et le theoreme de Goldbach, Vid. Skr.
Mat.-Nat. Kl. Kristiania 3 (1920).
[31] L. Carlitz and S. Uchiyama, Bounds for exponential sums, Duke Math. J.
24 (1957), 37-41.
[32] S. Chowla, The representation of a number as a sum of four squares and a
prime, Acta Arith. 1 (1935), 115-122.
[33] M. Eichler, Quaterniire quadratische Formen und die Riemannsche Ver-
mutung fur die Kongruenzzetafunktion, Arch. Math. 5 (1954), 355-366.
[34] P. Erdos, On the normal number of prime factors ofp - 1 and some related
problems concerning Euler's <p-function, Quart. J. Math. Oxford Ser.
6 (1935), 205-213.
[35] - - , On the sum ~i=i d(f(k)), J. London Math. Soc. 27 (1952), 7-15.
 BIBLIOGRAPHY 185

[36] T. Estermann, On the representation of a number as the sum of two products,

Proc. London Math. Soc. (2) 31 (1930), 123-133.
[37] - - - , Uber die Darstellung einer Zahl als Differenz von zwei Produkten,
J. Reine Angew. Math. 164 (1931), 173-182.
[38] - - - , Proof that every large integer is the sum of two primes and a square,
Proc. London Math. Soc. (2) 42 (1937), 501-516.
[39] H. Halberstam, On the representation of large numbers as sums of squares,
higher powers, and primes, Proc. London Math. Soc. (2) 53 (1951),
363-380.
[40] G. H. Hardy and J. E. Littlewood, Some problems of partitio numerorum
III: On the expression of a number as a sum ofprimes, Acta Math. 44 (1923),
1-70.
[41] C. B. Haselgrove, Some theorems in the analytic theory of numbers, J.
London Math. Soc. 26 (1951), 273-277.
[42] C. Hooley, An asymptotic formula in the theory of numbers, Proc. London
Math. Soc. (3) 7 (1957), 396-413.
[43] - - - , On the representation of a number as the sum of two squares and a
prime, Acta Math. 97 (1957), 189-210.
[44] A. Ingham, Some asymptotic formulae in the theory of numbers, J. London
Math. Soc. 2 (1927), 202-208.
[45] B. Jones, The arithmetic theory of quadratic forms, Carus Monographs
Series No. 10, The Mathematical Association of America, Buffalo, N.Y.,
1950 (distributed by Wiley, New York).
[46] H. D. Kloosterman, On the representation of numbers in the form ax2 +
by 2 + cz2 + dt 2 , Acta Math. 49 (1926), 407-464.
[47] E. Landau, Vorlesungen iiber Zahlentheorie, Vol. I. Elementare Zahlen-
theorie, Leipzig, 1927; English transl., Theory of numbers, Chelsea, New
York, 1958.
[48] K. Prachar, Primzahlverteilung, Springer, Berlin, 1957.
[49] S. Ramanujan, Some formulae in the analytic theory of numbers, Messenger
of Math. 45 (1915), 81-84.
[50] A. Renyi, On the representation of an even number as the sum of a single
prime and a single almost-prime number, Izv. Akad. Nauk SSSR Ser. Mat.
12 (1948), 57-78. (Russian)
[51] - - - , Sur un thioreme general de probabilite, Ann. Inst. Fourier
(Grenoble) 1 (1949), 43-52.
[52] - - , On the large sieve of Ju. V. Linnik, Compositio Math. 8 (1950),
68-75.
[53] A. Selberg, On an elementary method in the theory of primes, Norske Yid.
Selsk. Forh. (Trondheim) 19 (1947), no. 18, 64--67.
[54] G. K. Stanley, On the representation of numbers as a sum of squares and
primes, Proc. London Math. Soc. (2) 29 (1928), 122-144.
186 BIBLIOGRAPHY

[55] E. C. Titchmarch, A divisor problem, Rend. Circ. Mat. Palermo 54 (1930),

414-429.
[56] - - , Some problems in the analytic theory of numbers, Quart J. Math.
Oxford Ser. 13 (1942), 129-152.
[57] A. Weil, On some exponential sums, Proc. Nat. Acad. Sci. U.S.A. 34 (1948),
204-207.