PREFACE

TH:E main object of this publication is to render more accessible the papers
of a great mathematician, which in their original form appeared in many
journals over a period of about 50 years. The editors have kept in view
a second object also: that of rendering the work useful to mathematicians
generally by providing introductions to groups of papers, or comments
where appropriate. These editorial additions, while not always systematic
or exhaustive, will (it is hoped) assist the reader to view Hardy’s papers in
proper perspective.
It is this second object which has led the editors to divide th.e pders
into groups (or further into subgroups) in accordance with the nature of
their subject-matter, instead of publishing them in chronological order.
The editors have been very conscious of the difficulty of making such
a classification, which is most acute in those instances in which a paper
that is primarily on one topic has subsequently proved to be of great
importance for another. There are cases in which our allocation of a paper
to one section rather than to another has been in the nature of an arbitrary
choice, but we hope that adequate cross-references are provided. It may
be a matter for regret that our policy has sometimes resulted in distribut#ing
the papers of a series (such as ‘Notes on the Integral Calculus’) among
several sections, but we believe that any arrangement which kept them
together would have been less satisfactory.
We are grateful to Professor Littlewood for his permission to include all
the Hardy-Littlewood papers, and for his approval of our policy of
treating them on the same footing as Hardy’s own papers.
THE EDITORS
 ACKNOWLEDGEMENTS

THE editors are grateful to the following societies and publishers of journals who have
kindly given permission for the reproduction of Hardy’s papers. Details of the sources
are given in the List of ~qers by G. H. Ha&y, which appears at the end of each
volume.
Abbtilungen aw dem Mathematischen Seminar der Urtiversitit Hamburg: Messrs. A. Liebing
(Neudruck Journalfranz), Wiirzburg,
Acta Mathematiba: Institut M&tag&e&r, Sweden.
American Mathematical Monthly: the Mathematical Association of America.
Ann&i della Scuola Normale Euperiore de P&z: The Director of the Scuola Normale Super&e.
Annals of Muthematics: the Editors.
British Association Reports: the British Association for the Advancement of Science.
Bulletin of the American Mathematical Society: the American Mathematical Society.
Bulletin of the Calcutta Mathematical Society: the Calcutta Mathematical Soceity.
Comptes Rendus de l’Acad&mie des Scie+wes: Messrs. Gauthier-Villars, Paris,
Duke Mathematical Joumml: the Editor.
Fundamenta Mathematicae: the Editor.
Journal fiir die reine und angewatite Mathemutik: Messrs. Walter de Gruyter & Co., Berlin.
Journal of the Indian Mathematical Society: the Council of the Indian Mathematical Society.
Jourrzal of the London Mathematical Society: the Council of the London Mathematical Society.
Matemutisk Tidaakrift : Dansk Matematisk Forening, Copenhagen.
Matematiskfyaiske Meddelelser: I>& Kongelige Danske Videnskabernes Selskab, Copenhagen.
Mathematical Gazette: the Mathematical Association.
Mathematische Annalen: Springer-Verlag, Heidelberg.
Mathematische Zeitachrift : Springer-Verlag, Heidelberg.
Mind: the Editor.
Nuchrichten van der G~ellschuft der Wissenschaften zu GtittiNgen: die Akademie der Wissenschaften
zu Giittingen.
Proceedings of the Cambridge Philosophical Society: the Cambridge Philosophical Society.
Proceedings of the London Mathematical Society: the Council of the London Mathematical
Society.
Proceeding8 of the National: Academy of Sciences: the National Academy of Sciences, Washington,
D.C.
Proceeding8 of the Royal Sociely : the Council of the Royal Society.
Resdicmti de2 Circolo Matemutico di Palermo: the Editor.
Tdhoku Mathemuticul Jourruzl: the Editor.
Tranzx&ow of the American Mathematical Society: the American Mathematical Society.
Trawactiow of the Cambridge Phibsophica~ Society: the Cambridge Philosophical Society.
 EDITORIAL NOTE

The work will comprise seven volumes

FOR convenience of reference, papers are numbered according to years,

e.g. 1912, 4. A complete list of Hardy’s
I papers will be found at the end of
this volume (pp. 683-99) and will be reproduced at the end of each
volume. This list is based on that compiled by Titchmarsh (Journal of the
Londm i’kxthematical Society, 25 (1950), 89-101).
The date of publication of a paper, where it differs from the year men-
tioned in the reference number, is given (for the sake of its historical
interest) in the contents list of the volume containing the paper.
Where reference is made, in the corrections or comments, to the pages
of a paper, the numbers used are those of the original pagination and not
the consecutive page numbers of this volume. The joint papers with
Ramanujan are reproduced from The Cdlected Papers of S. Ramanujan
(Cambridge, 1927), and for these the ‘original pagination’ relates to that
volume and not to the first publication in a journal.
 CONTENTSOFVOLUMEI
Godfrey Harold Hardy, by E, c, TITCEMARSH* I
I. DI:OPHANTXNE APPROXIMATION

Introduction 15
1912,4 (with J. E, Littlewood). Some problems of Diophantine approximation. 20
Proceedings of the 5th International: Compe8s of Mathemmticiana, Canbridge, 1912, i. 223-9.
Published 1913.

1914, 2 (with J. E. Littlewood). Some problems of Diophantine approximation. 28

I. The fractional part of nk8.
Acta MatAwmtica, 3 7, 155-91.

1914, 3 (with 5. E. Littlewood). Some problems of Diophantine approximation. 67

II. The trigonometrical series associated with the elliptic &functions.
Acta Mathematics, 37, 193-238.

1916, 3 (with J. E. Littlewood). Some problems of Diophantine approximation: 115

A remarkabIe trigonometrical series.
Proceedings of the National A.cademy of Sciences, 2, 583-6.

1916,9 (with J. E. Littlewood). Some problems of Diophantine approximation: 119

The series 2 e(h,) and the distribution of the points (h, QI).
Proceedings of the National Academy of Sciences, 3, 84-88. Published 1917.

1919, 4. A problem of Diophantine approximation. 124

JuumaL of the Indian Mathematical Society, II, 162-6.

1922,5 (with J. E, Littlewood). Some problems of Diophantine approximation: 130

A further note on the trigonometrical series associated with the elliptic theta-
functions.
Proceedkgs of the Cctw&idge Philosophical Society, 2 1, 1-5.

1922,6 (with J. E. Littlewood). Some problems of Diophantine approximation: 136

The lattice-points of a right-angled triangle.
Proceed&g& oj the Londm Mathematical Society, (2) 20, 35- 36.

1922, 9 (with J. E. Littlewood). Some problems of Diophantine approximation: 159

The lattice-points of a right-angled triangle. (Second memoir.)
Abhandlungeta auo dem Ildathsmatischen bni~~.ar der Hamburgischen Univemitlit, 1, 212-49.
Published 1921.

* Reprinted with slight changes from Obituary Noticea of Fellowaof the Royal Society, 6 (194@),
447-58.
 CONTENTS

1923,3 (with J. E. Littlewood). Some problems of Diophantine approximation: 197

The analytic character of the sum of a Dirichlet’s series considered by Hecke.
Abhandlungen a.us dem Mathematischsn Seminar der Hambwgischen UniversitZit, 3, 57-68.

1923,4 (with J. E. Littlewood). Some problems of Diophantine approximation: 212

The analytic properties of certain Dirichlet’s series associated with the distri-
bution of numbers to modulus unity.
Tramactions of the Cambridge Philosophical Society, 22, 519- 33.

1925, 4 (with J. E. Littlewood). Some problems of Diophantine approximation: 227

An additional note on the trigonometrical series associated with the elliptic
theta-functions,
Acta Mathematics, 47, 189-98. Published 1926.

1930, 3 (with J. E. Littlewood). Some problems of Diophantine approximation: 237

A series of cosecants.
RuEleti~ of the Calwtta Mathematical Society, 20, 251.-66.

1946, I (with J. E. Littlewood). Notes on the theory of series (XXIV) : A curious 253
power series.
Proceedings of the Cambridge Philosophical Society, 42, G-90.

2. ADDITIVE NUMBER THEORY

(a) Combinatory analysis and sums of squares

Introduction 263

1916, 10. Asymptotic formulae in combinatory analysis.

Quatrikme Congrks des Mathe’maticiens Scandinaves, Stockholm, 1916, 45-53. Published 1920. 265

1917, 1 (with S. Ramanujan). Une formule asymptotique pour le nombre des 274
partitions de n,
Contptes RenduS, 164, 35-38.

1917, 4 (with S. Ramanujan). Asymptotic formulae for the distribution of 277

integers of various types.
Proceedings of the London Mathematical Society, (2) 16, 112-32.

1918, 2 (with S. Ramanujan). On the coefficients in the expansions of certain 294

modular functions.
Proceedings of th#e Royal Society, A, 95, 14655.

1918, 5 (with S. Ramanujan). Asymptotic formulae in combinatory analysis. 306

Proceedings of the London Mathematical Society, (2) 17, 7S115.

1918, 10. On the representation of a number as the sum of any number of 340
squares, and in particular of five or seven.
Proceedings of the National Academy of Sciences, 4, 189- 93,

1920, 10. On the representation of a number as the sum of any number of 345
squares, and in particular of five.
Transaction of the Americart Mathematical Society, 21, 255-84.
 CONTENTS

(b) Waring’s Problem

Introduction 377
1920, 2 (with J. E. Littlewood). A new solution of Waring’s Problem. 382
Quarterly Journal of Mathematics, 48, 272-93.

1920, 5 (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’: 405

I. A new solution of Waring’s Problem.
Nachrich ten van der K. Gesellschaft der Wiasens~hujten zu G&tinge%, Math.-phys. Klasse, 1920,
33-54.

1921, I (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’: 427

II. Proof that every large number is the sum of at most 21 biquadrates.
Muthematische Zeitschrift, 9, 14- 27. PubIished 1920.

1922, 4 (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’: 441

IV. The singular series in Waring’s Problem and the value of the number
a(k)
lkthmnutische Zeitsdrift, 12, 161-88.

1925, I (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’; 469

VI. Further researches in Waring’s Problem.
Mathemutische Zeitschkft, 23, 1-37.

1928, 4 (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’: 506

VIII. The number I’(k) in Waring’s Problem.
Proceed&p qf the Lortdon Mathematical Society, (2) 28, 51842.

(c) Goldbach’s Problem

Introduction 533
1919, 1 (with J. E. Littlewood). Note on Messrs. Shah and Wilson’s paper 535
entitled: ‘On an empirical formula connected with Goldbach’s Theorem’.
Proceeding8 of the Cambridge Philosophical Society, 19, 245-54.

1922, 1. Goldbach’s Theorem. 545

Matematisk Tidmkrijt B, 1922, 1-16.

1922, 3 (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’: 561

III. On the expression of a number as a sum of primes.
Acta Mathematics, 44, I-70.

1922, 8 (with J. E. Littlewood). Summation of a certain multiple series. 631

Proceedings of the London Mathematical Society, (2) 20, xxx. Published 1921.

1924, 6 (with J. E. Littlewood). Some problems of ‘Partitio Numerorum’ : 632

V. A further contribution to the study of Goldbach’s Problem.
Proceedings of the Lo&on, Mathematical Society, (2) 22, 46-56. Published 1923.
 CONTENTS

1920, Il. Some famous problems of the Theory of Numbers and in particular 647
Waring’s Problem.

Arrangement of the Volumes. 681

Complete list of Hardy’s mathematical papers, 683

 GODFREY HAROLD HARDY
1877-1947

GODFREY HAROLD HARDY was born on 7 February 1877, at Cranleigh, Surrey.

He was the only son of Isaac Hardy, Art Master, Bursar and House Mast&
of the preparatory branch of Cranleigh School. His mother, Sophia Hardy,
had been Senior Mistress at the Lincoln Training College. Both parents were
extremely able people and mathematically minded, but want of funds had
prevented them from having a university training.
The future professor’s interest in numbers showed itself early. By the time
he was two years old he had persuaded his parents to show him,how to write
down numbers up to millions. When he was taken to church he occupied the
time in factorizing the numbers of the hymns, and all through his life he amused
himself by playing about with the numbers of railwayL carriages, taxi-cabs and
the like.
He and his sister were brought up by enlightened parents in a typical
Victorian nursery, and, as clever children do, he agonized his nurse with long
arguments about the efficacy of prayer and the existence of Santa Claus: ‘Why,
if he gives me things, does he put the price on? My box of tools is marked
3s. 6d.’ The Hardy parents had many theories about education. Their children
had few books, but they had to be good literature. In the nursery G. H., who
was slightly older than his sister, read to her such books as Don Qkxote,
Gulliver’s Travels and Robinson Crusoe. They were never allowed to play with
any toy that was broken and past repair. The nurse gave them some instruction
in reading and writing, but they never had a governess, and on the whole were
left to find things out for themselves.
A minute newspaper written by G. H. at the age of eight was unfortunately
lost in the London blitz. It contained a leading article, a speech by Mr Gladstone,
various tradesmen’s advertisements, and a full report of a cricket match with
complete scores and bowling analysis. He also embarked on writing a history
of England for himself, but with so much detail that he never got beyond the
Anglo-Saxons. Two exquisite little illustrations for this work have survived.
He inherited artistic ability from his father, but it was crushed by bad teaching
at Winchester. He had no interest in music.
As soon as he was old enough G. H. went to Cranleigh School, and by the
time he was twelve he had passed his first public examination with distinctions
in mathematics, Latin and drawing. By this time too he had reached the sixth
form-the Cranleigh standard was at that time very low-so some of his work

Reprinted with slight changes from ObituaryNotices Fellows of the

Royal Society, 6 (1949 ), 447-58.
448 Obituary Nbtices
was sent to Winchester. He was offered a scholarship there on his mathematics
alone, but was considered too young to go that year, and went the following
vear.
Apparently he was never taught mathematics in a class. Mr Clarke, Second
Master at Cranleigh, and Dr Richardson, Head of ‘College’, always coached
him privately. He was never enamoured of public school life. He was grateful
to Winchester for the education it gave him, but the Spartan life in ‘College’
at that time was a great hardship, and he had one very bad illness.
There was some question of his going up to New College, btit his mind was
turned in the direction of Cambridge by a curious incident, which he has
related in A Mathematician’s Apology. He happened to read a highly coloured
novel of Cambridge life called A Fellow of Trinity, by ‘Alan St Aubin’ (Mrs
Frances Marshall), and was fired with the ambition to become, like its hero,
a fellow of Trinity. He went up to Trinity College, Cambridge, as an entrance
scholar in 1896, his tutor being Dr Verrall. He was first coached by Dr Webb,
the stock producer of Senior Wranglers. He was so annoyed by Webb’s methods
that he even considered turning over to history, a love of which had been
implanted in him by Dr Fearon, Headmaster of Winchester. However, his
Director of Studies Sent him to A. E. H. Love, and this, he considered, was
one of the turning points of his life, and the beginning of his career as a ‘real
mathematician’. Love was, of course, primarily an applied mathematician;
but he introduced Hardy to Jordan’s Cours d’AnaZyse, the first volume of which
had been published in 1882, and the third and last in 1887. This must h&e
been Hardy’s first contact with analysis in the modern sense, and he has
described in A Mathematician’s Apology how it opened his eyes to what
mathematics really was.
Hardy was fourth wrangler in 1898, R. W. H. T. Hudson -being Senior
Wrangler, with J. H. Jeans and J, F. Cameron, later Master of Gonville and
Caius, bracketed next. He took Part II of the Tripos in 1900, being placed in
the first division of the first class, Jeans being then below him in the second
division of the first class. In the same year he was elected to a Prize Fellowship
at Trinity, and his early ambition was thus fulfilled. Hardy and Jeans, in that
order, were awarded Smith’s Prizes in 1901.
His life’s work of research had now begun, his first paper apparently being
that in the Messenger of Muthematics, 29, 1900. It is about the evaluation of
some definite integrals, a subject which turned out to be one of his permanent
minor interests, and on which he was still writing in the last year of his
.
hfe
In 1906, when his Prize Fellowship was due to expire, he was put on the
Trinity staff as lecturer in mathematics, a position he continued to hold until
1919. This meant that he had to give six lectures a week. He usually gave two
courses, one on elementary analysis and the other on the theory of functions.
The former included such topics as the implicit function theorem, the theory
of unicursal curves and the integration of functions of one variable. This
was doubtless the origin of his first Cambridge tract, The Integration of Functions
 Godfrey Harold Hardy 449
of a Single V&able. This work is so well known now that it is often for-
gotten that- its systematization was due to Hardy. He also sometimes took small
informal classes on elementary subjects, but he was never a ‘tutor’ in the
Oxford sense.
In 1908 Hardy made a contribution to genetics
-. which seems to be little
known by mathematicians, but which has found its way into textbooks as
‘Hardy’s Law’. There had been some debate about the proportions in which
dominant and recessive Mendelian characters would be transmitted in a large
mixed population. The point was settled by Hardy in a letter to Science. It
involves only some simple algebra, and no doubt he attached little weight to it.
As it happens, the law is of central importance in the study of Rh-blood-groups
and the treatment of haemolytic disease of the newborn. In the Apologr Hardy
wrote, ‘I have never done anything “useful”. No discovery of mine has made,
or is likely to make, directly or indirectly, for good or ill, the least difference
to the amenity of the world.’ It seems that there was at least one exception to
this statement.
He was elected a Fellow of the Royal Society in 1910, and in 1914 the
University of Cambridge recognized his reputation for research, already world-
wide, by giving him the honorary title of Cayley Lecturer.
To this period belongs his well-known book A Course of Pure Mathematics,
first published in 1908, which has since gone through numerous editions and
been translated into several languages. The standard of mathematical rigour
in England at that time was not high, and Hardy set himself to give the ordinary
student a course in which elementary analysis was for the first time done properly.
A Course of Pure Mathematics is hardly a Cows d’Analyse in the sense of the
great French treatises, but so far as it goes it serves a similar purpose. It is to
Hardy and his book that the outlook of present-day English analysts is veryd
largely due.
He also played a large part in the reform of the old Cambridge Mathematical
Tripos Part I, and in the abolition of the publication of the results in order of
merit.
Another turning point in Hardy’s career wits reached about 1912, when -he
began his long collaboration with J. E. Littlewood. There have been other
pairs of mathematicians, such as PhragmPln and Lindeliif, or Whittaker and
Watson, who ‘have joined forces for a particular object, but there is no other
case of such a long and fruitful partnership. They wrote nearly a hundred
-
papers together, besides (with G. P6lya) the book Inepualities.
soon afterwards came his equally successful collaboration with the Indian
mathematician Ramanujan, though this was cut short six years later by
Ramanujan’s early death. An account of this association is given by Hardy in
the introductions to Ramanujan’s collected works and to the book Ramanujan.
In a letter to Hardy in 1913, Ramanujan sent specimens of his work, which
showed that he was a mathematician of the first rank. He came to England in
1914 and remained until 1919. He was largely self-taught, with no knowledge
of modern rigour, but his ‘profound and invincible originality’ called out
 450 Obituary Notices
Hardy’s equal but quite different powers. Hardy said, ‘I owe more to him
than to any one else in the world with tine exception, and my association with
him is the one romantic incident in my life’.
Hardy was a disciple of Bertrand Russell, not only in his interest in mathe-
matical philosophy, but in his political views. He sympathized with Russell’s
anti-war attitude, though he did not go to the lengths which brought Russell
into collision with the authorities. In a little book Bertrand Russell and Trinity,
which he had printed for private circulation in 1942, Hardy has described the
Russell case and the storms that raged over it in Trinity. It was an unhappy
time for those concerned, and one may think that it all would have been better
forgotten. It must have *been with some relief that, in 1919, he heard of his
election to the Savilian Chair of Geometry at Oxford, and migrdted to New
College.
In the informality and friendliness of New College Hardy always felt com-
pletely at home, He was an entertaining talker on a great variety of subjects,
and one sometimes noticed every one in common room waiting to see what he
was going to talk about. Conversation was one of the games which he lov red
to play, and it was not always easy to make out what his real opinions
were.
He played several games well, particularly real tennis, but his great passion
was for cricket. He would read anything on this subject, and talk about it
endlessly. His highest compliment was ‘it is in the Hobbs class’. Even until
1939 he captained the New College Senior CommQn Room side against the
Choir School and other opponents. He liked to recall the only occasion in the
history of the Savilian chairs when one Savilian professor (himself) took the
wicket of the other (H. H. Turner). The paper, ‘A maximal theorem with
function-theoretic applications’, published in Acta Math. 54, and presumably
_ addressed to European mathematicians in general,, contains the sentences,
‘The problem is most easily grasped when stated in the language of cricket . . .
Suppose
I that a batsman plays, in a given season, a given %tock” of innings . . .‘.
A vivid account of Hardy’s affection for cricket and of his life in his later
Cambridge years is given by C. P. Snow, in an article entitled ‘A mathematician
and cricket’, in The Saturday Book, 8th Year.
He liked lecturing, and was an admirable lecturer. His matter, delivery and
hand-writing (a specimen of which appears on the dust-cover of A Muthe-
maticiun’s Apology) were alike fascinating. Though no original geometer, he
fulfilled the conditions of his Oxford chair by lecturing on geometry as well
as on his own subjects. He also lectured occasionally on mathematics for
philosophers, and drew large audiences of Okford philosophers to whom
ordinary mathematics made no appeal. His Rouse Ball lecture’ on this subject,
delivered at Cambridge in 1928, entitled Mathematical Proof, was published
in Mind, 38,
Hardy had singularly little appreciation of science, for one who was suffi-
ciently nearly a scientist to be a Fellow of the Royal Society. In A Mathe-
matitiun’s Apologr he is at some pains to show that real mathematics is useless,
 Godfrey Harold Hardy 4s
or at any rate harmless. He says, ‘It is true that there are braaches of applied
mathematics, such as ballistics and aerodynamics, which have been developed
deliberately for war . w . but none of them has any claim to rank as “real”.
They are indeed repulsively ugly and intolerably dull; even Littlewood could
not make ballistics respectable, and if he could not, who can?’ His views on
this subject were obviously coloured by his hatred of war, but in any case his
whole instinct was for the purest of mathematics. I worked on the theory of
Fourier integrals under his guidance for a good many years before I discovered
for myself that this theory has applications in applied mathematics, if the
solution of certain differential equations can be called ‘applied’. I never heard
him refer to these applications.
Nevertheless, he was a Fellow of the Royal Astronomical Society, which he
jdined in 1918’ in order that he might attend the meetings at which the theory
of relativity was debated by Eddington and Jeans. He even once, in 1930, took
part in a debate on stellar structure, which involved R. H. Fowler’s work on
Emden’s and allied differential equations. On this he made the characteristic
remark that Fowler’s work, being pure mathematics, would still be of interest
long after all the physical theories which had been discussed had become
obsolete. This prophecy has since been very largely fulfilled.
I first came into contact with him when I attended his advanced class at
Oxford in 1920. The subjects which I remember specially as having been
discussed at this class are Fourier series, continued fractions, and differential
geometry, a commentary on R. H. Fowler’s Cambridge tract. Whatever the
subject was, he pursued- it with an eager single-mindedness which the audience
found irresistible. One felt, temporarily at any rate, that nothing else in the
world but the proof of these theorems really mattered. There could have been
no more inspiring director of the work of others.
He, was always at the head of a team of researchers, both colleagues and
students, whom he provided with an inexhaustluble stock of ideas on which to
work. He was an extremely kind-hearted man, who could not bear any of his
pupils to fail in their researches, Many Oxford D.Phil. dissertations must
have owed much to his supervision.
Hardy always referred to God as his personal enemy. This was, of course, a
joke, but there was something real behind it. He took his disbelief in the
doctrines of religion more seriously than most people seem to do. He would
not enter a religious building, even for such a purpose as the election of -a
Warden of New College. The clause in the New College by-laws, enabling a
fellow with a conscientious objection to being present in Chapel to send his
vote to the scrutineers, was put in on his <behalf.
He has been described as- absent-minded, but I never saw any sign of thi8.
If he dined at high table in tennis clothes it was because he liked to do so, not
because he had forgotten what he was wearing. He had a way of passing in the
street people whom he knew well without any sign of recognition, but this
was due to a sort of shyness, or a feeling of the slight absurdity of a repeated
conventional greeting.
452 Obituary Notices
His likes and dislikes, or rather enthusiasms and hates, have been listed as
follows:

Enthusiasms
(i) Cricket and all forms of ball games.
(ii) America, though perhaps he only came into contact with the pleasanter
side of it.
... Scandinavia,
(111 1 its people and its food.
(iv ) Detective stories.
0V Good literature, English and French, especially history and biography.
( vi > Walking and mild climbing, especially in Scotland and Switzerland.
( vii ) Conversation.
l .*

( Vlll ) Odd little paper games, such as making teams of famous people whose
names began with certain combinations of letters or who were
connected with certain countries, towns or colleges. These were
played for hours in hotels or on walks.
( iX ) Female emancipation and the higher education of women (though he
opposed the granting of full membership of the university to Oxford
women).
( X ) The Times cross-word puzzles.
xi
( 1 The sun.
( xii 1 Meticulous orderliness, in everything but dress. He had a large library
and there were piles of papers all about his rooms, but he knew where
everything was and the exact position of each book in the shelves.
l . .

( x111 ) Cats of all ages and types.

Hates
0i Blood sports of all kinds, war, cruelty of all kinds, concentration
camps and other emanations of totalitarian governments.
( ii ) Mechanical gadgets; he would never use a watch or a fountain pen,
and the telephone only under compulsion. He corresponded chiefly
by prepaid telegrams and post cards.
l .*

( 111 > Looking-glasses; he had none in his rooms, and in hotels the first thing
he did in his room was to cover them over with bath-towels.
( iv ) Orthodox religion, though he had several clerical friends.
(V 1 The English climate, except during a hot summer.
( vi > Dogs.
( vii ) Mutton-a relic of his Winchester days, when they had by statute to
eat it five days a week.
l . .

( Vlfl ) Politicians as a class.

( ix ) Any kind of sham, especially mental sham.

He was an extraorornary mrxture or out-or-tne-way mrormatlon and rgnor-

ante. ‘What is a milliner? Would you call the Army and Navy Stores a -.
 Godfrey Harold Hardy 453
milliner’s ?’ ‘No hawking! (this on Brighton front); I shouldn’t have thought
Ehey had to forbid that nowadays.’ In doing a cross-word puzzle: ‘The word
comes to ladders, but the clue is about stockings’.
Returning to his mathematical career, I may refer here to the founding of
the Quarterly Journal of Mathematics (Oxford series). Glaisher, the editor of
the Messenger of ikfuthematics and the old Quarterly Journal, had died in 1928,
and these two periodicals had come to an end. There was an obvious need for
something to replace them, and it was largely due to Hardy that a new series
of the Quarterly Journal was started in Oxford.
The London Mathematical Society occupied a leading place in his affections.
He served on the Council from 1905-1908, joined it again in 1914, and from
that time, except fur two absences of a year each, in 1928-1929 (when he went
to America) and 1934-1935, he was on it continuously until his final retirement
in 1945. He was one of the secretaries from 1917 to 1926, President in 1926-1928
and again for a second term in 1939-1941, and secretary again from then until
1945. In his Presidential address (1928), Prolegomena to a Chapter oiz Inequalities,
he boasted that he had been at every meeting both of the Council and of the
Society, and sat through every word of every paper, since he became secretary
in 1917. He was awarded the Society’s De Morgan medal in 1929.
In 1928-1929 he was Visiting Professor at Princeton and at the California
Institute of Technology, 0. Veblen coming to Oxford in his place. In 1931
E. W, Hobson died, and Hardy returned to Cambridge as his successor in the
Sadleirian chair of Pure Mathematics, becoming again a Fellow of Trinity.
Perhaps the most memorable feature of this period was the Littlewood-
Hardy seminar or ‘conversation class’. This was a model of what such a thing
should be. Mathematicians of all nationalities and ages were encouraged to
hold forth on their own work, and the whole thing was conducted with a
delightful informality that gave ample scope for free discussion after each paper.
The topics dealt with were very varied, and the audience was always amazed
by the sure instinct with which Hardy put his finger on the central point and
started the discussion with some illuminating comment, even when the subject
seemed remote from his own interests.
He also lectured on the calculus of variations, a subject to which he had been
drawn by his work on inequalities.
After his return to Cambridge he was elected to an honorary fellowship at New
College, He held honorary degrees from Athens, Haruard, Manchester, Sofia,
Birmingham, Edinburgh, Marburg, and Oslo. He was awarded a Royal Medal of
the Royal Society in 1920, its Sylvester Medal in 1940, and the Copley Medal, its
highest award, in 1947. He was President of Section A of the British Association
at its Hull meeting in 1922, and of the National Union of Scientific Workers in
19246, He was an honorary member of many of the leading foreign scientific
academies.
Some months before his death he was elected ‘associi ktranger’ of the Paris
Academy of Sciences, a particular honour, since there are only ten of these
from all nations and scientific subjects. He retired from the Sadleirian chair .in
 I

454 Obituary Notices

1942, and died on 1 December 1947, the day on which the Copley Medal was
due to be piesented to him.
He was unmarried. He owed much to his sister, who provided him throughout
his life with the unobtrusive support which such a man needs. Miss Hardy
has supplied most of the personal information contained in this notice.
In addition to the books mentioned above, Hardy wrote three more Cambridge
tracts, Orders of Injcinity (1910), The General Theory of Dirichlet’s Series, with
M. Riesz (1915), and Fourier Series, with W. W. Rogosinski (1944). In 1934
he published Inequalities with J. E. Littlewood and G. P6lya, and in 1938
The Theory of Numbers with E. M. Wright. In 1940 ‘followed Ramanujan, a
collection of lectures or essays suggested by Ramanujan’s work. His last book
was on Divergent Series, and was completed but not published at the time of
his death. His inaugural lecture at Oxford, Some famous problems of the theory
of numbers, and in particular Waring’s problem, was published in 1920. He was
also one of the editors of the collected papers of Ramanujan, which were
published in 1927.
The student of Hardy’s style should also read his obituary notices* of Rama-
nujan, Mittag-Leffler, Bromwich, Paley, Hobson, Landau, W. H. Young, J. R,
Wilton, and that of Glaisher at the end of the Messenger of Muthematics. These
tributes to his late colleagues must have made every mathematician wish that
he could have seen his own career described in the same generous terms.
Hardy was the author, or part author, of more than 300 original papers,
covering almost every kind of analysis, which by their originality and quantity
marked him as one of the leading mathematicians of his time. It is rarely possible
to disentangle his own contributions from those of others. He liked collaboration,
and much of his best work is to be found in joint papers, particularly those
written with Littlewood and with Ramanujan. He used to say that each author
of a joint paper gets much more than half the credit for it, No doubt the bulk
of his work is greatly increased by these collaborations, but he was certainly
the prime mover in much of it. He described himself as a problem-solver, and
did not claim to have introduced any new system of ideas. Nevertheless, if we
may judge by the references to his work in the writings of others, he had a
profound influence on modern mathematics.
When he began research there was probably no one at hand who could
give him the sort of supervision-which he was to give to so many others, and it
was some years before he found a problem of first-rate interest.
His early series of papers on Cauchy’s principal values was overshadowed by
the work of Lebesgue and others who were generalizing the integral in other
directions. Nevertheless it contains some interesting formulae. Perhaps the
most noteworthy are the inversion formulae
m

m 77PJxzydx,
Z-.1
9( >
-
1 mf@,
g(y) = - 77
-P “ydx,
s -
-QI --do
# For references, see p. 700. The notices will be reprinted in volume 7 of these Collected
PapWS.

8
 455
which have come to be known as the formulae of ‘Hilbert transforms’. In later
years he wrote many papers on transforms and inversion formulae of various
kinds. This work lies on the borders of the theory of integral equations; he never
worked on the central Fredholm theory itself, though he lectured on it in his
second period at Cambridge.
We next find him writing on the summation of divergent series, and this
turned out to be one of the permanent interests of his life.
The theorem of Abel, that if
%l=a,+a,+...+ati+ (1)
then
lim &p?=s (2)
X41-0 TI=l
is classical; and in 1897 Tauber had proved a sort of converse, that (2) implies (1)
if the coefficients a, satisfy the condition a,-o(l/n), i.e. nafl=+O. It was easy
to pro ve the corresponding theorem for Cesiro summability, that if
s,+...+s
u*= #-+S
n

and a,= 0(1/n), then again (1) follows. The simple and satisfactory appearance
of the proofs of these theorems gives them an air of finality; but in 1909 Hardy
proved that (3) implies (1) under the less restrictive condition a, = 0( l/n),
i.e. na, is bounded. This result, often referred to as ‘Hardy’s theorem’, was
the first ‘O-Tauberian theorem’, the forerunner of a whole science of such
theorems. It was followed in the next year by Littlewood’s theorem that (2)
implies (1) if a,= 0(1/n). Later the two originators of the theory published a
great deal of work on it together, and the whole matter has now been summed
up in Hardy’s last book.
His first important paper on Fourier series seems to be that in volume 12 of
the Proceedings of the London Mathematical Society (1913,4). The modern theory
of Fourier series, depending
-- on the Lebesgue integral, was then being constructed
by Lebesgue, FejCr, W. H, Young and others, and it was Young’s work that
inspired Hardy particularly. The first theorem in the paper referred to is that
any Fourier series is summable (C, 8) almost everywhere, for any positive 6,
and there are many others in the same order of ideas. Later Hardy and Little-
wood together developed a whole theory of ‘Fourier constants’ or coefficients,
gcneralizing the Young-Hausdorff theorem that if If(# is integrable, where
l<p<2, and an, b, are its Fourier constants, then

is convergent. A typical Hardy-Littlewood theorem is that the integrability of

If(x)141xl”-” over (- rr, rr), where ~2, implies the convergence of C( IanlQ+ lb,(4).
A curious by-product of this analysis is that, if the two conjugate series
C(a, cos nx+b, sin ttx), C(b, cos m--u,, sin nx)
are both Fourier series, then C(la,l+ Ib,l)ln is convergent.
 456 Obituary Notices
The theory of the Riemann zeta-function had begun with the guesses of
Riemann (1860), and the analysis of Hadamard and de la Vallke-Poussin, who
proved the prime-number theorem (1896). The modern theory of the function
had just been started by H. Bohr, Landau and Littlew-ood. The great puzzle
of the theory was the ‘Riemann hypothesis’, that c(s) has all its complex zeros
on the critical line a(s) =+, This presented all workers in the field, as it still
does, with a perpetual challenge, It was Hardy who first gave any sort of answer
to it, with the discovery, in 1914, that c(s) h as at any rate an infinity of zeros
on the critical line. The work was again carried on jointly with Littlewood,
and it was proved that, if N,(T) d enotes the number of complex zeros of c(s)
with real part Q and imaginary part between 0 and T, then N,(r) >AT for
some constant A (the total number of complex zeros in this region being
asymptotic to (T/27~) log T). It is only recently that this result has been surpassed
by A. Selberg, with the proof that N,(T) >AT log T. Hardy used to say that
any one who had a really new idea about the zeta-function must surely prove
the Riemann hypothesis, but Selberg’s work seems to have disproved this.
Another of the main features of the Hardy-Littlewood analysis, the ‘approximate
functional equation’, was discovered later to have been anticipated to a certain
extent by Riemann himself, though the applications which they made of it go
far beyond anything in Riemann.
Another subject to which Hardy made a fundamental contribution was- that
of the lattice-points in a circle. The number R(x) of lattice-points in a circle
of radius dx, i.e. of pairs of integers p, V, such that p2 +v2<x, is roughly
equal to the area TTXof the circle, but closer approximations to R(x) are difficult
to make. It had been proved by Sierpinski that, if
R(x) =rx+P(x),
then P(x) =0(x*), but the true order of P(X) was unknown.
Hardy obtained an exact formula for R(x) as a’ series of Bessel functions.
If x is not an integer this is

where r(n) is the number of solutions in integers of @-+ v2 =-n. If x is an integer,

R(x) must be replaced by R(x)-+(x). Th is ‘exact formula’ is very striking,
but it is not of much use in the problem of the order of P(x). If we could treat
the series as a finite sum, the ordinary asymptotic formula for Bessel functions
would give at once P(X) =0(x*). It is tempting to suppose that at any rate
P(X) =0(x%+“), but nothing approaching this has ever been proved. What
Hardy did prove was that each of the inequalities P(X) >Kx$ P(x) <-Kx*, is,
satisfied, with some K, for some arbitrarily large values of x. The true order of
P(X) therefore lies somewhere between X* and x$ and later research has done
a little, but not much, to narrow this gap.
I must now describe Hardy’s work on partitions, the ‘circle method’ in the
analytic theory of numbers, and his association with Ramanujan. They wrote

10
 Godfrey Harold Hardy 457
five papers together, the most famous being that in volume 17 of the Proceedings
of the London Muthematical Society (1918, 5), a section of which is reproduced
on the dust-cover of A lbthematiciun’s Apology. In this it was shown that
p(n), the number of unrestricted partitions of n, can not only be represented
approximately by an asymptotic formula, but that it can be calculated exactly
for any value of n. The ‘circle method’ on which this depends is, no doubt,
Hardy’s most original creation. It proceeds roughly as follows. The numbers
p(n) are the coefficients in the expansion
1
f(z)=l+ 5p(n)z”=
?Z=1 (l-z)(l-z”)(l-z”)***’

1 x
so that P(n ) =- 277is pxf( nfl ) dz ’
where r is a path enclosing the origin and lying entirely inside the unit circle,
and is taken to be a concentric circle of radius just less than 1. The unit circle
is a line of essential singularities off (x), but certain points are found to have
a particularly strong influence on the integral, and it is from these that a
dominant term is ultimately derived. In the Dirichlet series method for proving
e.g. the prime-number theorem, the dominant term is easily identified, and
almost the whole difficulty lies in showing that it is dominant. In the circle
method this is not so, and a whole apparatus, involving the Farey dissection
of the circle and the linear transformations of elliptic modular functions, is
needed to produce the result. It is all the more astonishing that the analysis
should have been carried to the point at which the exact value of p(n) could be
obtained.
Similar methods were applied later by Hardy and Littlewood to many
other problems, particularly to the Waring problem of the expression of a
number as a sum of given powers, and to problems involving primes, One
such problem which had long defied analysis was Goldbach’s theorem, or
hypothesis, that any even number can be expressed as the sum of two primes.
The success of the circle method in the study ofp(n) suggests a similar approach
to Goldbach’s theorem. Letf(x) now denote EP, wherep runs through primes.
Then {f(z))” =ww,
where rz(n) is the number of ways in which n can be expressed as the sum of
two primes. If we can prove that a(n) is always positive, Goldbach’s theorem
. will follow, but the difficulties prove to be even more formidable than in the
case of partitions. Actually it is a little easier to discuss

where b(n) is the number of ways of expressing n as the sum of three primes.
Hardy and Littlewood showed that, if certain hypotheses of the type of the
Riemann hypothesis are made, then b(n) is ultimately positive, so that any
sufficiently large odd number 12is the sum of three primes. Later Vinogradofi,

11
 458 Obituary Notices
by combining the essential ideas of the circle method with some entirely new
ideas of his own, showed that all unproved hypotheses could be dispensed with.
The whole method is perhaps the most remarkable example that at present
exists of analysis carried through apparently insoluble difficulties to ultimate
success. But the original Goldbach hypothesis still remains unproved..
Hardy had many other interests of which there is no space to speak at length
here: orders of infinity, Diophantine approximation, Bessel functions, in-
equalities.
Hardy’s work has had a profound influence throughout the whole of analysis.
It has resulted in the complete remodelling of some parts of the subject, and
. -. has
enriched other parts with new methods and theories of fundamental importance.
E. C. TITCHMARSH

12
1. DIOPHANTINE APPROXIMATION
 INTRODUCTION TO PAPERS ON DIOPHANTINE
APPROXIMATION

Practically all Hardy’s researches on t<his subject were carried out in collaboration
with Littlewood, the only exception being represented by 1919, 4. Apart from this
paper and 1946, 1, all the papers appeared lnnder the general title Some problems
of Diophantine approximation’, with various subtitles.
The series began with the famous communication to the 1912 Congress. This is
largely a statement of results, with occasional indications of the method of proof.
Re-reading it now, one is surprised to find how many important and fundamental
discoveries had been made by the authors, and how many delicate distinctions be-
tween apparently similar questions they had already perceived. The results announced
in 1912 were published for the most part in 1914,2 and 1914, 3, but some of them not
until 1922, 6.
Hardy and Littlewood were primarily interested in problems of distribution
modulo 1. The simplest such problem is: given a function f(n), can we say that the
values off(n) for n = 1, Z,... are everywhere dense (mod I), i.e. that their fractional
parts are everywhere dense in (0,l) ? In 1914, 2 it was proved that this is the case
if&) is a polynomial with at least one irrational coefficient (other than the constant
term). Results were also proved for the simultaneous distribution of the values of
several polynomials; these results are generalizations of Kronecker’s theorem, which
is itself the particular case when the several polynomials are all of degree 1. More
precisely, Kronecker’s theorem? states that if 1, e,,.., Brnare linearly independent
over the rationals, there exist integers n for which the numbers
d,,..., ne,
are arbitrarily near (mod I) to any WAprescribed numbers. Hardy and Littlewood
proved that the same holds for the WZ~numbers
n”O* (q = I ,..., p; j = l,...,m),
and their method applied in principle to more general polynomials.
Once it is known that the values of a function f(n) are everywhere dense (mod l),
the further question arises of their uniformity of distribution. For this we require that
the frequency with which f (n) falls (mod 1) into any given sub-interval of (0,l) shall
be proportional to the length of that sub-interval, Two other problems which prove
t For some remarks about various results related to Kronecker’s theorem, see the comments on 1914,2.

15
 INTRODUCTION TO PAPERS ON

to be closely related to the uniformity of distribution are those of estimating the Burn

S(N) = 2 {f(n)}, where {t) = t-[t]-4, (1)

n=l
N
and the sum s(N) = 2 @f@j.
?3=1

These various questions form the basic themes of most of the papers, and they are
treated by a variety of methods. This is not-surprising, for they are questions which
lie very much on the borderline between the theory of numbers and analysis, and
can often be approached from either side.
As early as 1912 Hardy and Littlewood had proved the uniformity of distribution
(mod 1) of the values of a polynomial with an irrational coefficient, and had announced
alsa consequence the estimate E(y--
1 +it) ---..-_--
= @lo@) as t +
----.---_-- _SJ The proofs of these results
never appeared, since Weyl’s &moir of 1916 rendered their publication unnecessary
(see 1916, 9). Weyl reduced the question of uniformity of distribution to that of
estimating sums of the type (2), and gave a simple and powerful method for finding
such estimates when f(n) is a polynomial.
In 1914, 2 Hardy and Littlewood also laid the foundations of the ‘metrical’ theory
of Diophantine approximation, in which results are proved to hold for almost all
values of a real parameter, in the senseof Lebesgue measure. Y.-_a.-
,__-
The second big memoir (1914,3) was entirely devoted to the study of the exponen-
tial sum (2) in the particular case

f(n) = Bn2+qin.

It was proved that if 8 is an irrational with bounded partial quotients in its continued
fraction, then s(N) = O(N*), and that this is best possible. Other results were deduced
on other hypotheses concerning 8. The basic principle of the proofs was the so-called
(approximate functional equation of the Q-function’, which enabled them to relate
s(N) to the continued fraction expansion of 8.
One possible application of Diophantine approximation which Hardy and Little-
wood kept in mind was the provision of explicit examples to illustrate general
theorems in the theory of functions or the theory of series, and to show to what
extent they are best possible. Some such applications were given in 1914, 3; but in
1916, 3 other examples which are lessintimately related to Diophantine approxima-
tion were shown to be equally effective.
Two other large memoirs (1922, 6 and 9) were devoted to the triangle problem.
This is the problem of approximating to the number N(v) of points with integral
coordinates in the triangle
x > Q, Y > Q, wx+w’y < 77
as 7;7
-+ 00, where ct), o’ are fixed positive numbers whose ratio 8 = W/W’ is irrational.

16
 DIOPHANTINE APPROXIMATION

It is easily Been that

where

f’ denoting the fractional part of +L Thus &(q) is a sum similar to the sum S(N)
in (1).
Two methods were used in the study of&(q), one elementary and the other ana-
lytical. The elementary method is based on a transformation formula and is on similar
general lines to the method of 1914, 3, It was proved that
4h) = d1) for any irrational 8,
M?) = O(log 7) if 0 has bounded partial quotients,
and that both of these are best possible. The analytical method uses contour integra-
tion and the double zeta-function
5&, a, w, 0’) = 2 2 (a+mw+m’d)-*.
m=Om’=O
For the estimation of 2$(v), the analytical method is no more effective than the
elementary method, and is if anything slightly less powerful. But it led the authors
to a remarkable explicit formula for &(q) in the form of an infinite series; a formula
which can be compared with that of Voronoi for the divisor problem or with that of
Sierpifiski for the circle problem.
The sum 1 {no}, w h’1ch is a particular case of (3), was deeply studied in the years
1922-5 by Hecke, Behnke, and Ostrowski, as well as by Hardy and Littlewood. The
analytical character of the function
$(a) = 2 {d}n-8 (4)
n=l

depends very much on the arithmetical character of 0. When 0 is a quadratic irra-

tional, Hecke proved that 4(s) is meromorphic and specified its poles; and the same
results were found by Hardy and Littlewood by a different method (1923, 3 and 4).
When 8 is any irrational, and A is defined by

(where the (xy are the denominators of the convergents to 0), Hardy and Littlewood
proved that the series (4) is convergent for
8s > X/(A+1)
and that +(s) has the line !Rs = X/(X+1)
as a line of singularities if A > 0.
The paper 1919, 4, by Hardy alone, is not related to the rest of the work, but is
of considerable historical interest. Here Hardy proved the basic property of the so-
called Pisot-Vijayaraghavan numbers. Suppose 8 (> 1) is any algebraic number and
 I~TRODUCTIONTO PAPERS ON
A is any real number except 0. Suppose that
AP+O (modl)
as n -+ 00. Then the conclusion is that 8 is an algebraic integer with the property that
all the algebraic conjugates of 8 (whether real or complex) have absolute values less
than 1, and A is an algebraic number in the field generated by 8.
Further comments are given immediately after the individual papers.
I conclude by listing a few problems in the subject, connected (directly or in-
directly) with the Hardy-Littlewood body of work, which are still unsolved.
(1) Little is known about the order of magnitude of
iv
e277ing0.
2
n=l

as N + 00, where 19is a fixedirrational number of somespecific type; say, with bounded
partial quotients. One easily deduces from Weyl’s inequality that for such 0 the sum
is O(Ng+E), but it is doubtful whether this is the full truth. There is the same problem
for almost all 8. Any new results may well prove to be significant for Waring’s
problem.
(2) The tetrahedron problem, that is, the analogue in three dimensions of the
triangle problem. The bounding plane is now
wx+tdy+w’% = 7,
where 7 -+ 00. It is appropriate to assume that or),w’, U” are linearly independent
over the rationala. It is easily proved that the error term is o(~~), but it is not known
whether this is best possible.
(3) The nature of #(s) in (4) when A = 0 (or more particularly when 0 has bounded
partial quotients), but 8 is not a quadratic irrational. It was conjectured in 1923, 4
that +(s) has 93s= U as a line of singularities, but this has never been proved.
(4) In 1930, 3 it is proved that if 8 = &z+l), ’ where a is an odd integer, then

as N -+ 00. The proof of this remarkable result is curiously indirect; it involves con-
tour integration and the use of Ces&ro means of arbitrarily high order. In the same
paper it is stated that for any quadratic irrational 8, the above sum is

as N + 00, where A(8) = 0 for the special values of 0 just mentioned, but is not
always 0. The problem is to give a simpler and more direct proof of these results.
(5) Littlewood’s problem on simultaneous Diophantine approximation: to prove (if
it is true) that for any real 0, + and any E > 0 there is a positive integer n satisfying

For references, see Davenport, iKzU&zatiA~~, 3 (1956), 131-5.

18
 DIOPHANTINE APPROXIMATION

(6) Khintchine’s problem on uniform distribution. If S is a subset of (0,l) with

measure IS 1 in the sense of Riemann (or Jordan), then the frequency of those n for
which n8 lies in S (mod 1) is ISI, and this holds for every irrational 0. The problem is
to prove (if it is true) that the same holds for aEmostall 6 if S has measure ISI in the
senseof Lebesgue. See Khintchine, Math. 2. 18 (1923), 289-306.
XI. D.
Abbrtkzted titles
In the comments which follow the individual papers, references to Cassels’s Tract, to Hardy
and Wright, and to Koksma are meant to refer to:
J. TV. S. Cassels, An introduction to Diophuntine upproximtion (Cambridge Mathematical Tract
No. 45), Cambridge, 1957.
G. H. Hardy and E. M. Wright, An introduction to the theory of numbers, Clarendon Press,
Oxford, 4th ed., 1960.
J. F. Koksma, Diophantische Approximutionen (Ergebnisse der Math. IV, 4), Springer, Berlin,
1936.

19
 223

SOME PROBLEMS OF DIOPkANTINE APPROXIMATION

BY G. H. HARDY AND J. E. LI~LEWOOD.

1. Let us denote by [x ] and (x) the integral and fractional parts of the real
number x, so that
( X ) =x-[xl, OS(~X)<l.
,
; Let 8 be an irrational numb& and u any number between 0 and 1 (0 included).
; Then it is well known that it is possible ti find a sequence of positive integers
i
% n2, n,, .*’ such that
(d>-,u
asr+oo. Now let f(n) denote a positive increasing function of n, integral when n
is integral, such as
7a,n2, 99, ,.., 2n, 3*, . .. . n!, 28: ***, 2”, A.,
and let f7 denote the value of f(n) for n = n,. The result just stated’suggests the
following question, which. seems to be of considerable interest :-For w?u.ztforms of
f(n) is it true’thdt, for any irrational value of 9, alzd tiny value of c1such that
0 s Q< 1, a sequencen, can be found such that

It is easy to see that, when the increase of f(n) is sufficiently rapid, the result
suggested will not generally be true. Thus, if f(n) = 2n, and 0 is a number which,
when expressed in the binary scale, shows at least k O’s following upon every 1, it is
plain that
(299) < 9 + xk:,
where X& is a number which can be made as small as we please by increasing k;
sufficiently. There is thus an (‘ excluded interval ” of values of a, the length of
which can be made as near to 8 as we please. Iff(n) = 3n we can obtain an excluded
interval whose length is as near Q as we please, and so on, while if f(n) = n! it’ is
(as is well known) possible to choose B so that (n! a) hax a unique limit. Thus
(n ! e) * 0.

2. The first object of this investigation has been to prove the following
theorem :-

Theorem 1 R {f f(n) is a polynomial in n, zuith integral coeficients, thert a

seqwnce cm be founii fw which (,fr8) + a.

1912, 4 (with J. E. LittIewood) Proceeding8 of the 6th Inter-

20 natiml Cmgress of ikfathematicians, Cambridge, 1912, i. 223-0.
 We shall give the proof in the simple case in which
= n2,
f() n
a case ‘which is sufficient to exhibit cIearIy the fundamental ideas of our analysis.
Our argument is based on, the following general principle, which results from the
work of Pringsheim and Lndon on double. sequences and series* :
If fT, 81 +r, 81 "*

are a finite number of functions of the p&dive integral variables r, s ; and ;f

We shall first apply this principle to prove that a sequence 92, can be found
so that
(qe) + 0, (n”Te) + 0
simultaneously. We shall, in the argument which follows, omit the brackets in (n@),
etc., it being understood always that integers are to be ignored.
We can choose a sequence mr so that t1,8 -0. The corresponding values n2,8
are infinite in number, and so have at least one limiting point $ ; f: may be positive
or zero, rational or irrAtiona1. We can (by restricting oumelves to a subsequence of
the &) suppose that

If 1:= 0, we have what we want. If not we write

fr,8 =(% + n8)gr, #T,& =(% + n8pe*

Then lim
lim f?,,8= lim n8e = 0,
gew
r-w 3-w
lim
lim &, 8= Iim (E + na,O)= 2F.
8-w
r-w s-w
Hence, by the general prii&ple,
we can pick out a new sequencepr such that
p,e-+o, pre+2t.
Repeating the argument, with n, +p8 in the place of n, + n,, we are led to a
sequence qr. such that

and it is plain that by proceeding in this way sufficiently often we can arrive at
a sequence n.,1k such that
nr, k @+ 0, nar,k e + &
for any integral value of k,
Now whatever number f is, rational or irrational, we can find a sequence k8
such that

ass+oo. Then

Pringsheim , Sitzu~ngsbmichtf?
der k. b. Akademie der Wiss. xu Ntinchen, WI. 27, p* 101, and Math.
len, vol. 53, pv 289 ; London, Math. Annah, i6id. , p. 322.

21
 SOME PMOl3LEMS OF DIOPHANTINE APPROXIMATION 226

Applying the general principle once more we deduce a sequence of values of 7z for
which (ne) + 0, (n2S) + 0 simultaneously.
When we have proved that there is a sequence n, for which r&0 + 0, it is very
easy to define a sequence +nr, where Y, is an integer depending on j*, which gives
any arbitrary a as a limit, We thus complete the &oof.of Theorem 1 in the case
f’(n) = ?P. An analogous method may be applied in the case of the general power 18~.
As in the course of this proof we obtain a sequence for which

simultaneously, we thus prove the theorem when u = 0 for the general polynomial f(n).
The extension to the case a > 0 may be effected on the same lines as in the case
f(n) =nk, but it is more elegant to complete the proof by means of the theorems of
the next section.
It may be observed that the relation
nB -+ 0
may be satisfied urtiformly for all values of 8, rational or irrational ; that is to say,
given any positive G, a number AT (e) can be found such that

for every B and some n, which depends on E and 8 but is less that m(e). Similar
results may be established for n2B, n38, . . . + The chief interest of this result lies in
the fact that it shows that there must be some function $(n), independent of 8, which
tends to zero as n + a0 and is such that for every 8 there is an infinity of values of n
for which
n20< +(n)*.

3. The following generalisation of the theorem quoted at the beginning of 5 1

was first proved by Kroneckert :-
If 8, (p, $9 .” are my number of linearly independent irrationals (Le. ;f rio
relation -of the type

where a, b, c, . . . are integers, not all zero, holds between 8, +, +, . . ,), and if a, & y, .. .
are any numbers between 0 and 1 (0 incltuded), then a sequence n, can be found
such that
n#+u, n&+3, n&+y, .-
This theorem, together with the results of 0 2, at once suggest the truth of the
following theorem :-

Theorem 9. If 8, +, $, . . . are linearly in.dependen t irrationals, and

4
w, I%* yz3 ..* (Z=l, 2, .*., k)

+ It is we11 known that? in the case of n8, #(n) may be taken to be l/n. No such simple resul& holds
when CL> 0 : exception has to be made of certain aggregates of values of 8, On the other hand, if B is a
Axed ‘irrational, the relation nd + a holds uniformly with respect to CL. A11 these results suggest
numerous generalisations.
t Werkc, vol. 3, pa 31. The theorem has been rediscovered independently bg various authors, e.g. by
Borel, F. Riesz, and Bohr (see for example Borel, Legona mr Zes se’ries divergenta, p. 135, and F. Riesz,
Comptee Rendus, vol, 139, p. 459).
226 G, H. HARDY AND J. E. LITTLEWOOD

k sets of numbers all lying bet?,ueen0 and 1 (0 inclzcded), thert it is 1)ossible to jnd
a sequence of values of n for which

This theorem we prove by means of two inductions, the first from the case of hi sets
ah I% yz3 .a9 to the case of k + I sets in which the numbers of the last set are all zero,
the second from this last case to the general case of Ic + 1 sets. The principles which
we employ do not differ from those used in the proof of the simpler propositions
discussed in. 5 2.
4. The investigations whose results are summarised in the preceding sections
were oiiginally begun with the idea of obtaining further light as to the behaviour of
the series
2 ePi20TTi
1 2 e?Z30d
? . ..
from the point of view of convergence, summability, and so forth. If we write*
Sn 12) = 2 e(p-3)2@&, s,(3) = C eV28Ti, s,,(4) = 2 (- l)Y-1 &B?ri
usn vsn vsn

it is obvious that, if srais any one of sJ2), .. . , then sn= 0 (n). If 0 is rutionul, either
& = 0 (1) or S,&= A, + 0 (I), where A is a constant : the cases may be differentiated
by means oi the well known formulae for “ Gauss’ssums.” Similar remarks apply to
the higher series in which (e.g.) Y2 is replaced by u3, y4, . The results of the l ’
preceding sections have led us to a proof of
Theorem 3. If B is irrational, then S, = o (n) : the same result is true for the
corresponding higher sums.
The argument by which we prove this theorem has a curious and unexpected
application to the theory of the Riemann c-function; it enables us to replace Mellin’s
result c(l +ti)= O(logIti)t by
~(l+t;)=o(logitI).
Theorem 4, Theorem 3 is the best possible theot-emof its kind, that is to say
the o (n) which occurs in it cannot be replaced by 0 (n#), where + is my deJinite
function of n, the samefor all O’s, which tends to #zero as n + 00.
*But although Theorem 3 contains the most that is ‘true for all irrational B’s,it
is possible to prove much more precise results for special classesof 8’s. Here we use
methods of a less elementary (though in reality much easier) type than are required
for Theorem 3, the proof of which is intricate.
In Chap. 3 of his Culcul des Rksidusl M. Lindelijf gives a very elegant proof of
the formula

u The notation is chosen EO as to run parallel with Tannery and MOWEI notation for the Q-functions : n
is not hecessarily an integer.
+ Landau, Hand&h der Lehre even der Verteil~unog der Primzahlen, p. 167.
$ pp. 73 et seq.

23
 SOME PROBLEMS OF f)IOPHANTINE APPROXIMATION 227

of Genocchi and Schaar. Here p and q are integers of which one, is even and the
other odd. By a suitable modification of Lindelijf’s argument, we establish the
formula

where B is an irrational number, which we may suppose to lie between - Z and 1,

A is one of 2, 3, 4, X, a corresponding one of the same numbers, and 0 (1) stands for
a function of n and 8 less in numericA value than an absolute constant.
We observe also that the substitution of 6 + 1 for B merely permutes the indices
2, 3, 4, and that the substitution of - 6 for 8 changes sn into its conjugate. If now
we write t9 in the form of a simple continued fraction
1 1 1
s+G+a,+.,.’
1 1
and put e e
=a,+ l=aa’ ‘m’
we obtain

and so on. We can continue this process until n&9,& . ~-1, when the first term
l n

vanishes, and we are left with an upper limit for Is,&I* the further, study of which
depends merely on an analysis of the continued fraction.
We thus arrive at easy proofs of Theorems 3 and 4 for k = 2. We can also prove

Theorem 5. If the partial quotients a, of the continued fraction for 8 are

limited * then sN(d)
, = 0 (&), In partic;ular this is true if 8 is a qundratik swd,
pure or mixed.
5. The question naturally arises wheth& Theorem 5 is the best possible of its
kind. The answer to this question is given by
Theorem 6 I If 6 is any irrational tiuw&r, it is p~h!Xe to J(illd a constctnt H
and un inJinity of values of 12such that
1s, (8) 1> H dn*
The sameis true ojctill Cesdro’smeansformed front the series.
The attempt to prove this theorem leads us to a problem which is very interesting
in itself, namely that of the behaviour of the modul.ar functions
~qOW, xg”“, r, (- q-1 (f2
as q .tends along a radius vector to an ‘( irrational place ” eenion the unit circle. If
f(q) denotes any one of these functions, it is trivial that
f(q)‘=0 I<1 -‘Id)-%
* This hypothmis may be generalised widely.

24
228 G, H. HARDY AND J. E, LITTLEWOOD

If p tends to a rational place, it is known that f(q) tends to a limit or becomes

definitely infinite of order +. By arguments depending upon the formulae of
transformation of the S-functions, and similar in principle to, though simpler than,
those of 5 4, we prove

Theorem 7 l Wh q tends to any irrational place on the circle of convergeme,

No better result thm this is true irt yenerd. If q + @, where 9 is one of the
irratimzals deJCined in Theorem 5, theyz.

m=o {(l -Iqj)-f].

Further, whatever be the value of 8, we cm $4 a cmstant &l ctnd an i@&y of mlues
of 1q 1, tending to unity, such that

In so far as these results assign upper limits for if(y) I, they could be deduce’d
from our previous theorems. But the remaining results are new, and Theorem 6 is
a corollary of the last of them. Another interesting corollary is

Theorem 8. The series

where 8 is irrationnl, be convergent, sunzmnble by any uf

Cesdro’s 772&2?2S.
On the other hand, if tx > +, these series are each certainly convergent for an
everywhere dense set of values of 0. They are connected with definite integrals of
an interesting type : for example

where - = &k), whenever the series is convergent,

6. We have also considered series of the types 2 (no), 2 @PO), . . . l . It is

convenient to wrife

Arithmetic arguments analogous to those used in proving Theorems 3 and 4

lead to

Theorem 9. If 0 is any irrational mmtber, then s, = u (n). The sume result

holds for the series in urhich v is replaced by v2, Ys, . . , yk, . . . *. Further, this result is
l

the best possible of its Kind.

+ This result, in the case k= 1, has (as was kindly pointed out’to us by Prof. Landau) been given by
Sierpinski (see the Jahrbuch fibber die Fortschritte der Math., 1909, p. 221). Similar results hold for the
function
sfa-[$+a]-4
which reduces to (5) for a = 0.

25
 SOME PROBLEMS OF DIOPHANTINE APPROXfMATION 229

When k = 1, we can obtain more precise results analogous to those of $ 4, 5.

The series
. X (no) behaves, in. many ways, like the series XenW% The r6Ze of the
formula of Genocchi and Schaar is now assumed by Gauss’s formula

where p, p are odd integers. Taking this formula as our starting point we easily
prove Theorem 9 in the case k= I, Further, we obtain

Theorem 10. If 0 is an irrational number of the type dejned in Theorem 5,

then sn = 0 (log n).
This corresponds to Theorem 5. When we come to Theorem 6 the analogy
begins to fail. We are *not able to show that, for every irrational 0 (or even for
every B of the special class of Theorem 5), sn is sometimes effectively of the order of
log n. The class in question includes values of $ for which this is so, but, for any-
thing we have proved to the contrarj3 there may be values of -0 for which So= 0 (1).
And when we consider, instead of sn, the corresponding Ces$ro mean of order 1, this
phenomenon does actually ,occur. While engaged on the attempt to elucidate these
questions we have found a curious result which seems of sufficient interest to be
mentioned separately. It is that ‘y-a TJ,$jj@< !&W >
L
C ivep = +g72+ 0 (1) ,i,j,,+
r I ‘t/ ),*II-“i ,
1,,i&*,4
vsn
for all irrational values of B. When we consider the great irregularity and obscurity
of the behaviour of C (v#], it is not a little surprising that C {vtY)z (and presumably
the corresponding sums with higher even powers) should .behave with such marked
regularity.

7. The exceedingly curious results given by the transformation formulae for

the series 23enagri,C {no] suggest naturally the attempt to find similar formulae for
the higher series. It is possible, by a further modification of Lindelijf’s argument;,
to obtain a relation between the two sums

where K = h/(32/278). The relation thus obtained gives no information about the
first series that is nut trivial. We can however deduce the non-trivial result

Similar remarks apply to the higher series ILPkewi and to the series C ink@),where
k > 1. But it does not seem probable that we can make much prqress on these lines
with any of our main problems.
In conclusion we may say that (with the kind assistance of Dr W. W. Greg,
Librarian of Trinity College, and Mr J.. T. Dufton, of Trinity College) we have
tabulated the values of (KM) for the first 500 values of n, in the cases

e =ko =+31622776..., O=e.

The distribution of these values shows striking irregularities which encourage a

closer scrutiny.
 COMMENTS
This communication to the 1912 Congress? is mainly a summary of the principal results of
1914, 2 and 1914, 3, though the proofs of Theorems 9 and 10 were not published until 1922,
6 and 1922, 9,
0 4, The footnote to Theorem 5 does not mean that the same result can be proved under
a more general hypothesis, but that other hypotheses could be made about 8 which would
imply similar (but weaker) results. Such results were given in 1914, 3 and 1922, 5,
8 6. The statement concerning C(I&}~ is erroneous and was corrected in the last sentence of
1922, 6. The correct form appears as Theorem 11 of 1922, 9.
8 7, For some remarks on more general transformation formulae, such as that for ZeVsgTi
mentioned in the text, see the comments on 1914, 3.
The final sentence ddes not seem to have given rise to any further investigation, and it
would be of interest to know in what senses the fractional parts of n28 are less well distributed
than those of d?.

t Some of the F8SUltS had been briefly communicated to the London Mathematical Society at its
meeting on 8 February 1912 (see Proc. 11 (1912), xxi-xxii).

27
 SOMEPROBLEMSOF DIOPHANTINEAPPROXIMATION.
BY

G. H. HARDY and d. E. LITTLEWOOD,

TRINITY COLLEGP, CAMBRIDGE.

I.
The fractional part of nk8.
1. o - Introduction.

I. 00. Let us denote by [x] and (x) the integral and fractional parts of x,,
so that

Let .r3 be an irrational number, and a any number such thaf o <a < I ,
Then it is well known that it is possible to find a sequence of positive integers
n,, $1
such that
n3,- l m

(I * 001) (wea
as r--too.
It is necessary to insert a few worda of explanation as to the meaning
to be attributed to relations such as (I. oar), here and elsewhere in the paper,
in the particular case in which a = 0. The formula (I. OUI), when a > o, asserts
that, given any positive number E, we can find ru so that

The points (n, 0) may lie on either side of a. But (n, 0) is never negative, and
= o, the formula, if interpreted in the obvious
so, in the particular. case in which ~11
manner, asserts more fhan this, viz. t-hat

28 1914, 2 (with J, E. Littlewood) Acta Mathematics, 37, 15!?%91,

156 G. I& Hardy and J, E. Littlewood.

The obvious interpretation therefore gives rise to a distinction between the value
a= o and other values of CI which would be exceedingly inconvenient in our
subsequent analysis.
These difficulties mat ” be avoided by apreeirig
A that, when CY= o, the formula
(I. OOI) is to be interpreted as mea.ning ‘the set of pints (n, 0) has, as its de
limiting point or points, one or both of the points 1 and O’, that is to say as imp-
ying that, for any r greater than rO, one or other of the inequalities

o+t,O)<e, r-s< (n,8)<1

is satisfied. In the particular case alluded to above, this question of interpreta-

tion happens to be of no importance: our assertion is trui: on either inter-
pretation. But in some of our later theorems the distinction is of vitAal im-
portance.
Now let f(n) denote a positive increasing function of n, integral when n is
integral, such as

The result stated atI the beginning suggests the following * question, which seems
to be of considerable interest: - For what forms of f(n) is it true that, for any
irrational 0, and any value of a such that o ~a < I, a sequence (n,) can be found
~cch that

It is easy to see that when the increase of f(n) is mfficiently rapid the
result suggested will not always be true. Thus if f(n) = P and 0 is a number
which, expressed in the .binary scale, sho.ws at least k o’s following upon every I,
it is plain that

when & is a number which can be made as small as we please by increasing k

sufficiently. There is thus an* Dexcluded intervaln of values of a, the length of
FJrhich can be made as near to + as we please. If f (n) = 3n we can obtain an ex-
cluded interval whose length is’ as near to 3 as we please, and so on; while if
,q
f( ) =n! it is (as is well known) possible to choose 0 so that (n! 0) tends to a
unique limit. Thus (n!e)--o.
At the end of the paper we shall return to the general problem. The im-
mediate object with which this paper was begun, however, was to determine whe-

29
 Some problems of Diopbantine Approximation. 157

ther the relation (I 002) always holds (if 0 is irrational) when f (n) is a po’wer of
l

n, and we shall be for the most part concerned with this special form of f(n).
I . 01. The following generalisatidri of the theorem expressed by (I.@ was
first proved by KRONECKER~

Theorem 1.01. If 0, Fo,, -- . Ornare linearly independent irrationals (i. e. if no

relutiun. of the type

a,8,+a,O,+~.~+a,O,+a,+l=o,
c

where a,,a,,-a,.+1 are integers, not all zero, holds between 0, , 0, ,. - lOm) , and Us,
Q,,
CS~are numbers such that o < ap < I, then a sequence(n,) can be found such that
l ’
l

USr-~. Further, in the speciul case when all the 2s are zero, it is unnecessary to
make any restrictive hypothesis concerning the O’s, or even to suppose them irrationc;tl.
This theorem at Once suggests that the solution of the problem stated at
the end of I: . oo may be generalised as follows.
Theorem 1 .OlL If O,, O,, -- m0, are linearly independent irratiunals, and
the a’s are any numbers such that o -< CY< I, then a sequence (nc) can be foztnd
such that

l KRONE~IWR, Berliner Sitamgsberichte, 11 Dec. 1884; We&e, vol. 3, p. 49.

A number of special cases of the theorem were known before, That in which all the
~1% are zero was given by DIBICHLET (Berliner Sitzungsberichte, 14 April 1842, Werke, vol. 19
p. 635). Who first stated explicitly the special t-heorems in which ti$ = r $e have been unable
to discover. DIRWHLET (1. c.) refers to the simplest as &ngst bekannt, : it is of course an immediate
consequence of the elementary theory of simple continued fractions. See also MINROWSKI,
xDiophantische Approximations, pp. 2, 1. KRONECKER’S general theorem has been rediscovered
independently by several writers. See e. g, BOREL, LeGons SW Ees shies diztsrgentes, pm 135; F. RIESZ,
Comptes Rendus, 29 Aug. 1904. Some of the ideas of which we make moat use are very similar
to, those of the latter paper. It should be added that DIRICHLET’S and KRONECKER’S theorems
are presented by them merely as particular cases of Inore general theorems, which howeve;
represent extensions of the theory in a direction different from that with which we are con-
cerned.
A number of very beautiful applications of KRO~ECKER’S theorem to the theory of the
RIEMANN c-furiction have been made by I-I. BOHR.

30
158 G. H. Hardy and J. E. Littlewood,

Further, if the u’s ure all zero, it is unnecessary to suppose the O’s restricted in
any way.
I +oz. This theorem is the principal result of the paper: it is proved in
section I , 2, The remainder of the paper falls into three parts. The first of
these (section x . I) consists of a discussion and. proof of KRONECKER’S theorem.
We have thought it worth while to devote some space to this for two reasoni.
In the first place our proof of theorem 1.011 proceeds by induction from k to
k + I, and it seems desirable for the s’ake of completeness *to give some account
of the methods by which the theorem is established in the case k= 1. In the
second placemthe theorem for this case possessesan interest and importance rsuffic-
ient to justify any attempt to throw new light upon it; and the ideas involved
in the various proofs which we shall discuss are such as are important in the
further developmentIs of the theory. We believe, moreover, that the proof we
give is considerably simpler than any hitherto published.
The second of t.he remaining parts of the paper (section I, 3) is devoted to
the question of the rapidity with which the numbers (nxOP) in the scheme (I . OII)
tend to their respective limits. Our discussion of the problems of this section ig
very tentative, and the results very incomplete;1 and something of the same
kind may be felt about the paper as a whole. We have not solved the problems
which we attack in this paper with anything like the definiteness with which
we solve those to which our second paper is devoted. The fact is, however, that
the first paper deals with questions which, in spite of their more elementary
appearance, are in reality far more diffimlt. than those of the second. Finally,
the last section (I, 4) contains some results the investigation of which was sug-
gested to us by an interesting theorem proved by F. BERNSTEIN? The disting-
uishing features of these results are that they are concerned with a single irrat-
ional 0 and with sequences which are not of the form (AT), and that they
hold for almost all values of 0, i. e. for all values except those which belong to
an exceptional and unspecified set of measure zero.

I.I- Kroneckefs Theorem.

I: . IO. KRONECKER’Stheorem falls naturally into two cases, according as to
wh ether or not all the. Q’S are zero. We begin by considering the simpler

1 Some of the results that we do obtain, however, are important from the point of view
of applications to the theory of the series 2 enkoi and that of the EIEMANX c-function. It was
in part the poasibility of these applicati&s that led us to the researches whose results are given
in the present paper. The applications themselves will, we hope, be given in a later paper.
p M&h. - Annalen, vol. 71, p. 421.

31
 Some problems of Diophantine Approximation. 159

when all the cr’s are zero. Unlike most

of the tPheorems with which we are con-
cerned, this is not proved by induction, and there is practicalIy no difference
between the cases of one and of sevefal varia.bles. The proof given is DIRJCH-
LET’&

Let 2 denote the number which differs from q bv an integer and which is
such that -+<S+. Then the theorem to be prove: is equivalent to the theo-
rem +hat, given any integers q and N, we can find an 72 not less than N and
such that

Let us first -suppose that N = I. Let_ B be the region in m-dimensional space

for which each coordinate ranges from u to I. Let the range of each coordinate
be divided into g equal parts: R is then divided into Q” parts. Consider now
the Q” + I points ’

(Ye,), (Y8,),--, (v&n); (Y=O, I:, 2, nq”>*

There must be one part of R which contains twu points; let the correspondibg J
values of Y be V, and w,. Then clearly

19/q, I(%- m%II.dq, l ,I(~,-%)~mf(~/q,

and I Yl - %(>I.
We have therefor only to take KA= IY~--Y~I. We observe that we have also

n<p%

a result to which we shd have occasion to return in section I. 3.

If N > I we have only to consider the points (Y NO,), (Y N O,), q= instead ofl

the points (Y&), (vOJ,-.

I . II. We turn now to the case when the ~1% are not all necessarily zero,
In this case the necessity of the hypothesis that the B’s are linearly independent
is obvious, fur the existence of a linear relation between the 8’s would plainly
involve that of a corresponding relation between the a’s; naturally, also, the added
restriction makes the theorem much more difficult than the one just proved.
Our proof proceeds by induction from nz to ry1+ I; it is therefore import’ant
to discuss the case m = 1:. The result for this case may be proved in a variety
of ways, of which we select four which seem to us to be worthy of separate dis-

32
160 G. H. ‘Hardy and J. E. Littlewood.

cussion. These proofs are all simple,, and each h’as special advantages of its own.
It is..important for us to cotisider very carefully the ideas involved in them with
a view to seletit’ing those which lend themselves most readily to generalisation.
For example, it is essential that our proof should make no appeal to the theory
of continued fractions.
(a). The first proof is due to. KRONECKER. It follows from the result of
1 . 10, with m = I, or from the theory of continued fractions, that we can find
an arbitrarily large q such that

(I. III) q 0 - p = d/q.

It is possible to express any integer, and in particular the integer

nearest to q a, in the form

where rt and n, are integers, and 1nl (q/z. From the two equations

we obtain

?t6
q(nB+n,)=-+qa++&
u 2
I4l-F
and so

Or

If we write 2! = n, + q and use (T. III), we see that

lb4 -+wq, ql2 < fv < 34/z;

so that
lW>- a I < 3/y

33
 Some problems of Diophantiue Approximation. 1er

for some value of Y between q/z and 34/z. This evidently establishes the truth of
the theorem.
If we attempt to extend this proof to the case of several variables we find
nothing to correspond to the equation

But KRONECKER’S proof has, as against the proofs we shall now discuss, the very
important advantage of furnishing a definite result as to the order of the ap-
proximation, a point to which we shall return in I .3.
(b). Let E be an arbitrary positive constant. By the result of L , xo, we
can find an n such that o < 0, < E or I - E< 0, < I:, where 0, = (no). Since t) is
irrational, t), is not zero. Let us suppose that o < 0, < E; the argument is sub-
stantitilly the same in the other case. We can find an rye such that

m8,(a<(m + I)@,,

I ml 0 --I<+$;
and so
j(nmO)-al<&,

which proves the theorem.

(c)J Let S denote the set of points (no). X’, its first derived set, is closed.
It is moreover plain that, if a is not a point of S, then neither is (a + n8) nor
(a -- n 8).
The theorem to be proved is clearly equivalent to the iheorem fhat S con-
sists of the cont!inuum @,I). Suppose that this last theorem is false. Then
there is a point a which is not a point of S’, and therefore an interval co&
taining a and containingg no point of S. Consider I, the greatest possible
such interval containing a, 3 The interval obtained by translating I through a
distance 8, any number of times in either direction,” must, by what was said
above, also confain no point of X’. But, the interval thus obtained cannot. over-’
lap with I, for. then I would not be the ,>great,est possible* interval of .its kind.

’ This proof was discovered independently by F. RIE~Z,but, BO far as we know, has

not been published.
a In ittJ interior, in the strict sense,
’ The existence of Buch a #greatest possibler interval ia eaeil$ established by the classicA
argument of DEDERIND.
* Taking the congruent interval in (0, I), This interval may possibly consist of two separ-
ate portions (0, &), and (Ez, I).
162 G. H. Hardy and J, E. Littlewood.

Hence, if we consider a series of [I/S] translations, where d is the length of I,

it is clear that two of the corresponding [#J + I intervals must coincide.
Clearly this can only happen if 0 is rational, which is contrary to our hypothesis.
(d). We argue as before that, if the theorem is false, t*here is an interval
I, of length z E and middle point a, cont.aining no point of X’. By the result
of x.10 we can find n so that, if tl,=(nOj, then o<O,<s or I-&<O,<I.
By the reasoning used in (c) it appears that the interval obtained by trans-
lating I through a distance O,, any number of times in either direction, must
contain no point of S. Rut since each new interva1 overlaps with the preceding
one it is clear that after a certain number of translations we shall have covered
the whole interval o to I: by intervals containing no point of S, and shall thus
have arrived a#t a contradiction.
I: . 12. Let us compare the three last proofs. It is clear that (6) is consid-
erably the simplest, and that (d) appears to contain the essential idea of (b)
together with added difficulties of its own. It appears also that, in point of
simplicity, there is not very much to choose between (c) and (djl and that (c) has
a theoretical advantage over (d) in that it dispenses the assumption of the
theorem for the case CI= o, an assumption which is made not only in (b) and (d),
but also in (a). When, however, we consider the theorem for several variables,
it seems that (b) does not lend itself to direct extension at all, that the com-
plexity of the region corresponding to I in (c) leads to serious difficulties, and
that (d) provides the simplest line of argument. It is accordingly this line of
argument which we shall follow in our discussion of the general case of KRON-
ECKER'S theorem.
I . 13. We pass now to the general case of KROBEOKER’S theorem. We shall
give a proof by induction. For the sake of simplicity of exposition we shall
deduce the theorems for three independent irrationals 0, q, +, from that for
two. It will be obvious that the same proof gives the general induction from n to
$2 + I irrationals.

We wish to show that if we form the set S of points within the cube
o(~<r,o<y<r,o<z,<r, which axe congruent with

then every point of the cube is a point of the first derived set S. It is plain
that, if (CI, Is, y) is not a point of 8, then neither is ((a + no), (a + ny), (y + nq))
nor ((a- n@, (p - nrp), (y -T#)). If now our theorem is not true, there niust exist
a sphere, of centre (a, 8, r) and radius e, which contains 1 no point of S. By
1 Within or upon the boundary.

.
35
 Some problems of Diophantine Approxima Con. 163

the resuPt of I. IO, there is an n such that the distance 6 of ((no), (ncp), (n+))
or (O,, rp,, I/J,) from one of the vertices of the cube is less than e/1/2. Let. us sup-
PUSB, fQr eXM@e, that the vertex &question is the point (o, o, 0). Consider
the .straight line
X -a Y-P 2-Y
(I l 131) -=-=-1
0i [Pl IP1

and the infinite cylinder of radius 6 with this line as aPxis. It is clear that the
finite cylinder C obtained by taking a length 6 on either side of (ol,‘g, y) is en-
tirely . contained in the sphere and therefore contains no point of S. Hence the
cylinder obtained by translating C through (O,, cpl, ql), any number of times in
either direction, also contains no point of S’, so’ that, since each new position of
C overlaps with the preceding, the whole of the infinite cylinder, or rather of
the congruent portions of ‘the cube, is free from points of S.
Let uti now consider the intersections of .the totality of straight’ lines in
the cube, which are congruent with portions of the axis of the cylinder, with
an arbitrary plane x=x,. We shall show that they are everywhere dense in
ttle square in which the plane cuts the cube, whence clearly follows that no
point of the cube is a point of S, and SO a contradiction which establishes the.
theorem.
The intersections (y, x) are congruent with the intersections of the axis
(I. 13x) with

X =x, -#- y, ( Y = ..* t -2, -1, 0, I, 2,4,

and so they are the points congruent with

But, under our hypothesis, (pl /O, and +I /O, are linearly independent irrationals,
and so, by the theorem for two irrationals, this set of points is everywhere
dense in the square, The proof is thus completed.
I s 14. We add two further remarks on the subject of KRONECICER'S theo-
rem, in which, for the sake of simplicity of statement, we confine ourselves to
the case of two linearly independent irrationals 0, cp.
(a) Suppose that. O<CC<I, o<p<r. KRONECKER'S theorem asserts the exis-
tence of a sequence (na) such that (n, 0) - a, (n#rp) I - @. Let us choose a se-
quence of points
(% PJ, (P = I, 2, 3,-m*),

36
164 G. H. Hardy and 6. El. Littlewood.

such that
qL>% a,>a, q4-% /p-+Bm

There is, for any value of p, $ sequence (n,,) such that

as 8-m. From this it is easy to deduce the existence of a sequence (n,) for
which (no?) and (n& tend to the limits CI and /? and are always greater than
those limits, so that the direction of approach to the limit is in each case from
the right hand side.’ Similarly, of course, we can establish the existence of a
sequence giving, for either 0 or rp, either a right-handed or a left-handed ap-
proach to the limit.
If we apply similar reasoning to t-he case in which ~11or @ or both are zero*
we see that, when 0 and c-pare linearly independent irrationals, we may abandon
the convention with respect to. the particular value o which was adopted in I . oo,
and assert thaf there is a sequence for which (~9) - a and (nq) - @, (x and p
having any values between o and I, both values included, and the formulae
having the ordinary interpretation. This result is to be carefully distinguished
from that of I . IO. The latter is, the former is not, true without restriction on
the O’s, as may be seen at once* by considering the case in which rp = -0.
(b). It is easy to deduce from KRONECHER’S theorem a. further theurem,
which may be stated as follows: B if we take any portion y of the square o <x < r,
o --
< y ---
< I, bounded by a finite number of regular eurws, and of area 6; and if we
denote by N, (n) the number bf the points

(W), bfrp)), b = f, 2, ‘. ’ nL

which fall inside y; then

N,(n) w 4%
as n-a.
This result, when compared with the various theorems of this paper, sug-
gests a whole series’ of further theorems. The proofs of these appear’likely to be
very difficult, and we have, up to the present, considered only the case of a
single irrational 0. We have proved that, if N,(n) denotes the number of t#he
points
(W9), (v= I, 2, m n), l

’ The reasoning by which this is established is essentially the same as that of 1 20. n

a This is a known theorem. For a’proof and references see the tract ‘The Riemann
Zeta-function and the Theory of Prime Kumbers’, by Ii. BOHR and J. E. LITTLEWOOD, shortly to
be published in the Cumbridqe Tracts in Muthematics and Muthmatiml Phqsica

37
 Some problems of Diophantine Approximation. 165

which fall inside a segment y of (o, I), of length 6, then iV7 (n) cv 6~. This re-
sult may be compared with that of Theorem 1 .483 at the end of the paper.
But results of this character will find’ a more natural place among our later
investigations than among those of which we are now giving an account.

I l z* - The generalisation of Hronecker’s theorem

I . 20. We proceed now to the proof of theorem 1 . 011. Our argument is
based on the following general principle, which results from the work of PRINTS-
HEIM and LONDON on double sequences and series?
1.20. If

then we cm find a sequence of’ Ms (rlg9, rzlt, -. - Tkn) such that, as n - cro,
_

We shall show that, if this principle is true for all values of ry1and a part-
icular k, tPhen it is true for k + I. As it is plainly true for k = I, we shall
thus have proved it generally.
We shall abbreviate ‘lim lim -0 lim’ into +‘lim’ , or, when there is
l

rl -a r,-00 rk-00 b rk VI p r2 1 l

no danger of confusion, into ‘lim’.

Let .
llm i p (rl 1 rz 3’ *’ rk+l) = fp (rk+l)-
ri 3 rz 3 - ’ rk

Then by hypothesis
fp bk+l) - A,

as rk+l - 00 w Let us choose an integer rk+f 3n, greater than P, for which

1jp (rk+l, w) - ,$I 1 < 2-+--l, (p = 1, 2: l ” ml*

By the principle for k variables, we can find rln, rsn: . . . rkflP .a11 greater than
2?s, and such that
l PRING~~HEIM,Mtinchener Sitzungsbevichte, vol. 27, p. 101, and Math. dnnalen, ~1. 53, p 289;
LONDON, Math. AnnaZen, vol. 63, p. 322.
I We thus obtain a sequence of sets (rln, ~2~;. - rk+l,n), such tht every member
of the nfh set is great&than P and

If p ( h, r2nc ’ rk+&--Ap)<z-n, @=I, z,-m).

This sequence evidently gives us what we want.

An important special case of the principle is the following:
1 ,201. If for all values of t we am find a -sequence nit, n2t,- a, n,t, . - such l

that

r,,bJrt) -* -4, t 7 (13 - I, 29- * ’ m),

us r-00, and if
A pt- A yt (P = I, 2,. .* m),

as t-q then there is a sequence (n,) such that

f&b)--Apr (p= 1, 2,- m),

as s-a.
This is in reality merely a case of the principle that a limiting-point
I of
limit#ing-points is a limiting-pot&.
I . 21. F7e consider first the case in which all the ~2s are zero, and the 0%
are unrestricted. Jn this case the proof is comparatively simple.
Theorem 1.21. There i;~ a sequence (n,) such that, as r - 00

I (nFO,,)-0,

We prove this theorem by induction

(x= I, 2,-k;

from k to k + I: we have seen that it

p-1, 2,-m).

is true when k- I. We suppose then that there is a sequence (& such that

I
(I l 211) (p;OJ- 0, (x = I, 2,-b k; p- I, 2,- m).

The sequence

fias at least1 one limiting point rp,, q2,.4 VF~; hence, by restricting ourselves to a
subsequence selected from the sequence (/la), we can obtain a sequence (u,) such
that, ads s - 00,

b~~p)-% (x -< k); (vk+l

8
Up) - tpP; (p = I, 2, - - a m).

39
 Some problems of Diophantine Approximation. 167

We then have, for *A.(k + I,

where the C’s are constants, X, + X, + 9m*+ XL= x, and xq <k. In virtue of
(I . ZII) we can evaluate at once every repeated limit on the right hand side, and
it is clear that we obtain Iry, or o according as 3c= k + I or 3c< k. It follows
from the general principle 1 20 that we can find a sequence @A), (r= I, z,..*>,
l

srlch that, as r - cro,

Rub, by the’orem 1 .01, we can find a sequence (A,) ‘such that

(n,qp)-o, !p= 1, 2?m);

and we have only to apply the prrnciple 1.201 to obtain the theorem for k + 1.
I , 22, We pass now to the general case when the ~1’s are not all zero. We
have to prove that if O,, O,, . . - 0, are linearly independent irmtionds, there i8
a sequence (n,) such that, us r - GO,

We shall prove this by an induction from k to k + I which proceeds by two

steps.
(i). WC assume the existence, for a particular k, any number m of o’s,
and any corresponding system of LX’S,of a sequence giving: the scheme of limits

n
nB
l . . l . l :

. l

.
nk

and we prove the existence, for any number m of O’s, and any corresponding
By&em of a’s, of a sequence giving the, scheme

40
168 G. Et. Hardy and J, E. Littlewood.

n
n”
.

It will be understood that neither m, nor the O’s, nor the cr’s are necessarily
the same in these two schemes, all of them being arbitrary,
(ii). We then sImw that we can pass from the last- written scheme of limits
t.o the generd scheme in which the elements of the last row also are arbitrary.
I , 23. Proof of the first step. To fix our ideas we shall show that we can ’
pass from a sequence (n,) giving’

to a sequence (m,) giving

It will be clear that the argument is in reality of a perfectly general type,

Suppose we are given aI11 &, pII, &, and that 0, rp, cy’,, (x’, , & , &, are line-
arly independent irrationals. Then by hypothesis we can find a sequence giving
the scheme

Further, the set of points ($8, n:cp) has at least one limiting-point (A, [L), and,
by restricting ourselves to a subsequence of &), we may suppose that we have
also

1 In what follows WV shall omit the brackets in (126,. . .; it ia of court to he under-

stood that integers are to be ignored.

41
 Some problems of Diophantine Approximation. 169

We express all this by saying that we can find a sequence (n,) giving the scheme

The sequence (&), where k,= z n,., gives us t*he scheme,

2a,,I zp’,, 0, Q, 0, 0,
(I l 222) 4 4, 1 0, o 1
4 p’z

8;1, 8~1.

By the general principle 1 .%I, we can find ai sequence (Zr) giving t,he scheme

where ‘lim stands for llm

’ .
rl, r2, - n l r8

Consider the repeated limit

which is easily evaluated with the aid of the table (I . 221). The limit of a term
n& 0 is k: that of a ‘cross-term

is zero, since n&t9 tends to an a’ or a @‘, and ntj a’ and nFj 8’ tend to zero. Thus
we obtain the repeated limit 8k. In all the other repeated limits the cross-terms
give zero in the same way, and we see that the sequence (I,) gives the scheme

81 , 8,~.

Consider now the repeated limits

170 G. 8. Hardy and J. E. Littlewood.

All the cross-terms contribute zero as before, and we obtain the scheme

(8 + 2 m) a’, , (8 + wp’,, o,o, o,o,

(8 + 4mw,, (8 +*4wp’2, 0, 0,
(8 + 8m)a 9 (8 + 8m)p,
Or

6cr’, + (m + I)ZC& 6& + (m + Ij2&, oj o, o, o,

It; is possible, then, to find a sequence giving this scheme. But now, since it is
possible to find a sequence of m’s such that

(m + 1) F-9 w==L @I, 4&, 4&, 8L 8/l),

it follows (in virtue of the firinciple 1.20) that we can find a sequence giving
the scheme
w, $f,, 0, 0, 0, 0,
4&, 4/L 0, 0,
0, 0.

This gives us what we want (and something more) provided it is possible to

choose

This is the case provided 0, y, cr,, &, CI,, & are linearly independent irrationals:
it remains only to show that this restriction on a,, &, IX,, & may be removed.
lt is obvious, in virtue of the principle 1.20, that this may be done provided
we can find a sequence (~11~~&, aZlt, &) such that, for each n, 0, r(l, u17L,/L,
aZn, pztr are linearly independent irrationala, and such that

mn -a,; P In - B 13 azn--a2, P 2n- I”s 2’

l

Now it is easy to see thaf there must be points (aIn, pin, a& &) interior to
the ‘cube’ with (CQ,& , a,, ,&) as centre and of side 2+, and exterior to that

43
 Some problems of Diophantine Approximation. 172

with the same centre and of side z+--l, and such that 0, rp, aIn, /?ln, Olga,/Se*
ztre linearly independent irrationals. By selecting one such point correspondin&
to each value of n we obtain a sequence of- the kind desired?
I . 24, Proof of the second step. Here also we shall consider a special case
l
for simplicity: the argument is redly general. We shall show that we can pass
from a sequence giving the scheme

to one givipg

As in I . 23, we ma-y suppose, without real loss of generality, that 0, rp,

a19/$ are linearly independent irrationals. Let (n, ) be a sequence giving

a2 B
2

2
0 - a,
3
n3 0 0

1 This argument depends ostensibly on ZEERMELO’S ‘Auswahlsprinzip (or WHITEHEAD and

RUSS~XL’S Wultiplicative Axiom’). This difficulty can however be surmounted with ? little trouble.
It should perhaps be observed that we have ignored several. similar. points early in the paper:
in all of these the difficulty is comharatively trivial, and we hav.e only called attention to it
in the present instance because it occurs in a more serious form than is usual in constru@ive
mathematics.
An alternative line of argument from that in the text proceedsas follows. It is easy
to show that if at most a finite nuvber of primes are omitted, any four of the sequence
log 2, log 3, log 5, log 7, log II,* ‘, together with B’and y, form a set of six linearly independent
l

irrationals. Moreover it can he deduced from known resulta concerning the distribution of the’
primes that we can find a sequence (logpa, logqn, logm, logsn), where pn, qti, in, and art are
primes, such that
(bgpts)- a1, (log @> - PI 1 (log ?h) - Q p (log sn! - h-

44
172 G. H. Hardy and J. E. Littlewood.

Then

lim(n, + n.) B = lim (I- ~3, -+ n,O) = a1

r, s r 2

lim (n, -I- n,)V9 = lim (n, ay, -I- n:O) = 4x,,
r, 3 T

]im (n, + 7tJW = lim (b2~a, +nf@ = a,;

r>s T 2

with similar results for q. It follows by the principle 1.20 that there is a se-
quence giving the desired scheme, and the proof of the induction, and therefore
that of the theorem, is completed.

x.3. - The order of the approximation.

I. 30. We have proved that under certain conditions we can find a sequence
(n,) such that

(1’ 301) (n;OP)-+MxP (x=1,2,-k; p=1,2,-m).

There are a number of interesting questions which may be asked with regard to
the rapidity with which the scheme of limits is approached.
The relations (I .301) assert that, is we are given A, there is a function
@(k,m; O,,t?,,*~&; aII,a12,‘-*GLkm; n)l such that

for some n <.a. It is hardly necessary to observe, after the explanations of I . oo,
that this inequality requires & modification when &p = o, which may be express-
ed roughly by saying that olXPis then to be regarded as a two-valued symbol
capable of assuming indifferently the values o and I.
(i) Does Q) necessarily depend on the 0% and U’S: can we for example, find
a D independent of the Q’S? It will be seen that this last question is answered in
the affirmative,
(ii) Can we assert anything concerning the order of Cp qua function of L,’
the variables 0 and a being supposed fixed? The same question may be asked
concerning any @ which is independent
, of the-a’s; it should be observed, more-
over, that the best answer to the latter question does (not necessarily, give the
best answer to the former.

’ For shortneaaWQshall writ9 this d (k,m, 8, all).

 Some problems of Diophantine Approximation. 17.3

Our attempts .to answer these questions have not been successful, and such
results as we have been able to obtain are of a negative character, The ques-
tion then arises as to whether we can obtain more definite results by imposing
restrictions ‘on the O’s or the IX’S, by supposing for example that all the CI’S are
zero, or that the B’s belong to some special class of irrationals.
(iii) The relations (I .301) imply the truth of the following assertion : there
is a function rp (k, m,O, a, n) which tends to infinity with n, and is such th.at

for an infinity of values of n. A series of questions may then be asked concern-

ing rp similar to those which we have stated with reference to a.
I .3 r. We shall begin by proving two theoreIns which are connected with
the questions (i). The first of them deals with the case in which all the a’s
are zero, and it will be convenient to use in its statement; as in I . IO, not the
function (x), but the allied function X.
Theorem 1 31. There is a; function clr(k, m, A), depending onlyL on k, m, and
l

;1, such that

for Some n<Q,.

For suppose that this theorem is false. Then to every r corresponds a set
of O’s, say .8,, &, . . . rl’)m, such that the inequalities

are not all true unless n > r. The set of points (~9, , #., , m. ,.H,,) has at least
l

one limiting point (O,, O,, . m1O,), and by restricting ourselves to a subsequence
of r’s we can make

From this it follows that we can choose a number n, which tends to infinity
with r but so slowly that

Clearly we may suppose that n,~_r, and so we have, for an infinit,y of values
of r, loll5-r and
I?4 G. H. Hardy and J. E. Littlewood.

From (I ,311) and (I .313) it follows .that the inequalities

cannot all be true unless n > n,, and 80, since n,-+a, cannot be true for any
value of n. This contradicts Theorem 1 .Oll.
In the case k = I it is possible ‘to assert much more tban this. It is known,
and is proved in I. IO, that in Ghia case we may take

(1 l 3x4) CD= @] + 1p

This problem, in fact, mavc he regarded as completely solved. When k > I, how-
ever, the case is very different. We have not even succeeded in finding a defin-
ite function @ (A), the same for all O’s, such that

for nA@i It would be not unnatural to suppose that the tibest possible)}
function1 Q) .is less than K A, where K is an atbsolute constant. But we have been
unable to prove this or indeed any definite result as to its order in A.
r ,32. Theorem 1.32. If the 8’s are linearly independent irrutionals, it is
possible to find a fzcnction @ (k, m, 0, A), independent of the a%, such thut

for some n <CD.

That this theorem is true for the special case k = I, m = I, follows from the
argument (a;) in I. II. It is easily proved in the most general case by an argu-
ment resembling, but simpler than, that of I. 31.
If the theorem is untrue, it is possible to find a sequence of sets ( flxP)
(r =I,Z,- ) for which the inequalities of the theorem do not all hold unless n > T.
The sequence of sets has at least one limiting set (Or,& let tis choose r so that

Then clearly the inequalities

cannot all be true unless PZ> r, and so, since r is arbitrarily large, cannot all
be true for any n. This contradicts Theorem 1.011.

’ That is, the function which has, for each value of k, the least possible value. For the
existence of this function it is necessary that the sign < above should not be replaced by <.

47
 Some problems of Diophantine Approximation, 176

I .33* Let us -consider more particularly the ca,se in which k: = x,

The equation (I .314) .suggests that it may in this case be possible to choose
for @ a function of the form

J2 (m, 8, cL)E.

This we believe
to be improbable, but we have not succeeded, even when vz = I,
in obtaining a definite proof. What is certain is that no dorresponding result is
true of the @ of Theorem 1.32. It is impossible to choose a function f2 (m, 0)
independent of k, and a function J,D(m, A) independent of the O’s, in such a way
that the @ of this theorem may be taken ‘to be of the form

This is’shown by the following theorem.!

Theorem 1.33. Let Zy (A) be an arbitrary function of 1 which fends steadily to
infinily wilh A. Then it is possible to find irrational numbers 0 for which the asser-
tion ‘there is a functibn

tD (6, a) = 52 (6) q (A)

such thul, whert ;1 is chosem, the inetjuulily

is satisfied, for every a, by sore n less than @’ is false.

Suppose that the assertion in question is true. Taking a = I/& we see that

(1 ’ 331) o<(nCr)<a/n

for some n less then a,

Let pv14y be the wth convergent to the simple continued fraction

-x + I
+.*.
.a1 a3

which represents 6,. so that pI = I, qI = a,; and let us consider the system of
‘intermediate convergenti’

L Xn proving a result of this negative character we may evid&y confine oursel~~ to the
special case in which m = I.

48
176 Cr. H. Hardy and J. E. Littlewood.

intercalat’ed between .pzn/qzn and pzn+2/qzn + zm These fra&ons are all less than
o and increase ulith r. Also

where &n+2 is the complete quotient. corresponding to tzzn+ 2, and

QlBn+2 = b&+2 q2n+ 1 + q2w

Let
2 qL+2
(1 ’ 333) k n=-r.
Q2n+2-8'

where s is a particular value

of r which we shall fix in a moment. We shall
suppose t~2~+2 large, and s also large, but small in comparison with a2n+2. In
these circumstances 1, will be approximately equal to z QZ,&+1 l

We shall now prove that if

(1 ’ 334) 0 < (Q(j) < 2/n,

then

11 rn335) Q> 92rw

From (I l 334) it follows that there is a fraction P/Q such that

P 2
(1 l 336) o<fj -- Q <n,Q’

On the other hand

2
(1 ’ 337) fl --=-mv
P%s hn 2n,e

If P/Q actually gave a better approximation by defect to 0 than ~2,Jq2~,~, it

would follow at once that & > q2n,s* We may therefore suppose the contrary;.
and then it follows from (I ,336) and (I ,337) that

Hence
0< ‘r)zn,~ Q- q2n,s P < 2 Q2n, A

But

49
 Some problems of Diophantine Approximation. 177

I:- a’2n+2 qzn+ 1 + q2n

il It z >q2,2t1 t
2 a'2,1+2-~

and

$2 72,8 -=~n+5q2n+l<(t~+I)q2ntl~

Hence Bn,s Q- @n,a P is les~~ than s + z, and &d

(1 l 338) pet+&- qzn,s P = 0 (U v< 4 -< 3).

On the other hand
p2n,u 42n,e-9 - q2ng p2n,8-@ = Q;

and so

pm,8 (8 - qwr-0 ) = qzn,u cp -- p2n,s-4

Hence either Q = q2n,s-e, or Q- q2jt,s-I, is divisible by qz<#; and the latter hypo-
.
thesis plainly involves that Q > q2,1,s,
On the other hand, if Q =qq12,8-,, then P = ~2~,~--~? and

p a’zn+a-8 + Q

‘which contradicts (I ,337). Hence in any case Q > qzn+

It is now easy, to complete the proof of the theorem. We have at fortiori
Q> 8. Also, if Gzn +2 is large, and s large, but small in comparison with azn+ 2,
kta will clearly be less than 4 Qzn+ 1 We may suppose for definiteness
l th&
6 = [l/azrr+zl-
We choose a value of (1 such that the inequality

ab+2> W(4 q2n+i):i

is satisfied for an infinit,y of values of n, Then

But if Q, and u fovtiori s, is less than II,, we must have

and this is obviously impossible when n is sufficiently large. This completes the
proof of the the&em.
178 G. H,.I aHardy and J. E, Littlewood.
It should be observed that the success of our argtiment depends entirely on
our initial choice of CL iti such a way that (n 0) is sm.all. It would not be euough
th,at a ahould be 8mEtl1, that is to <say that (n 8) ahould be nearlyI equal to
either o or I: thi8 can of conr~e be secured by choice of an IZ-less than U, CI,
being indeed independent of 8.
I. 34, We turn now for a moment to the questions concerning r/l. Jf we have
found a function @(A) which is continuous and monotonic, the inverse function is
plainly a cp* The converse, however, is not true, and we cannot, from the existence of
a cp of given form, dcaw any conclusion as to the order of 0 for a# values of A.
This is clear from the fact that, fo put it roughly, the existence of ;P ‘aaRer@
an inequality which need only hold very occasionally, and which therefore gives
US information as to the behaviour of QI ody for occasional values of A, Thus
the existence of a U asserts much more than that of the corresponding rp. Since
moreover it will appear (in the third paper of the series) that in applic&ions of
the present theory it is always the properties of ‘I,, and not those of q~, which
are relevant, we are justified in regarding theorems concerning rp as of rather
minor importance. There are, however, one or two results which are worth noticing,
and which are not deductions from the corresponding results concerning Q‘,. It
should be observed that whereas we wish @ to increase as slowly as -possible, we
wish rp to increase as rapidly as possible.
Theorem 1,340, It is possible to choose the ~1% so that pp(m, 0, a, n) increases
with arbitrary rapidity. Moreover the cc’smay lw chosen in un arbitrarily small neigh-
bourhood of any set (q, aa, -. u,).
n

We omit, the proof of this theorem, which is easy.

Theorem 1 341. If k = I a*rtd m = I, then, provided
l only that 19is iwatjonal,
we may take

(a function independent of both 6( and a).

This fallows at once from the argument (a) of I . II. It is natural to snp-
pose that, when wx> I, we may take

where 10(m) depends only on m. But this we have not been able to prove.
A comparison of Theorems 1 33, and 1.341 shows very clearly the differ-
l

mce between theorems involving Q, and those involving q~, and the greater depth
and difficulty of the former.

51
 Some problems of Diophantine Approximatioll. 1’79

I .35. Theorem 1,33 shows that it is ‘hopeless to expect any such Gmple
re&lt concerning @ as is asserted concerning rp in Theorem 1 341. It is howeverl

possible to obtain theorems which involve @ and correspond to Theorem 1.341:

l *

if we suppose that certlain classes of irrationals (as well as the rationals) are
excluded from the range of variation of 0. In the two theorems which follow .it
is supposed that WJ= I and k = I.
Theorem 1.350. Let 8 be confined to the class of. irrdonals whose partidl
qziotients are limited, QCset which is everywhere dense. Then *we may take

tD = a 52(0).

Theorem 1. 351. Let 0 be confined to the class of irratidnals tihose pahd

quotients u n mtisfy, from a certccin value of n onwurds, the inepality

Then we may tuke,

0 = a (log a)1+ d’L2(0)

where 8 ztaany number greater than 8.

The interest of the last theorem lies in the fact that the set in question is
of measure I,~ 80 that we may take @ to be of the form il (logA)1+&J2(0),8
where & is an arbitrarily small positive-number, for almosf all values of 0.
The proofs of these theorems are simple and depend merely on an adapt-
ation of KRoNBcKER’g argumenf reproduced in x SII. Suppose first.that the par-
tial quotients of 0 are* limited. We can choose H so that, when K is assigned,.
there is always
c a denominat.or qrn of a convergent to 0 such that

We take q ==qma It follows from KRONECEER’S argument that there is for any CI
a number Y such that

’ Sy a theorem of BOREL and BERNSTEIN. See BOREI,, Rendicod di Putemto, vol. 27, pa 247,
md Math. Am., vol. 72, p. 578; BERNSTEIN, Math. Ann;, vol. 71, p. 417.
’ It is not difficult to replace k(log A)l+& by 1 log X (log log I)l+&, or by the ~orrespontl-
ing bat snore complicated functiona of th-e logarithmic scale.
 The proof of Theorem 1.361 is very similar. We suppose that

There is a constant 4 such that pm > eg*, r and from these facts it follows easilyI that

q, < A (log n)l+ @

for sufficiently large values of A. The proof may now be completed in the same
manner as that of Theorem 1.350.
It is natural to suppose that these theorems have analogues when m >. I. But
our arguments, depending as they do on the theory of continued fractions, do
not appear to be capable of extension.

x*4* - The general sequence (f(n) 0) ad the particular sequence @‘VI)

1.40.We return now to tke general sequence (f (n) 8): it will be convenient
to write in for f (n). We suppose then that (k,) is an arbitrary increasing sequence
of numbers whose limit is infinity?
It would be natural to attempt to prove that, if 61is irrational and CI is aBy
number such that o < CI < I, a. sequence (n,) can be found such that

but awe saw in I. oo that this statement is certainly false, for example when
=.
;1rr zn or R,=lz?
The result which is in fact true was suggested to us by a theorem of BERN-
STEIN,’ which runs as follows:
If il, is albays m integer, then the set of vrtluesof 0 for which

ia of meuaure zero.
This result, when considered in conjunction with what we have already
proved, at once suggests the following theorem.

.l In the introductory remarks of I .OO we etated our main problem subject- to the restrict-
ion that hn is an integer. No such m&Action, however, ie requited in what folIowE.
q l? BERNSTEIN, Zoc. cit. 3

53
 Some problems of I)iqhautine Apyroxiiuation. 181

Theorem 1.40. The set 01 w1at.m of 19, for which the set of pints (in 8) is
not everywhere dense iti the interval (0, I), is of measure zero.
In other words, the main question asked *in I. oo may be answered affirm-
atively if we make exception of a set of measure zero.
I. 41. The proof will be based upon the following lemma..
Lemma i .41. #upposewthut a finite number of is~terrals are exclu$ed from the
continuum ‘(0, I), and that the length #of the remainder 8 is 1. Let a be any number
betwen o and I, and consider the set T of [A] inter&s of length 811 (6 < I) whose
centres are at the points

Then the length o/ the wmmon prt of S and T is

dl +&I,,
where en-0 cts a---d.
The truth of the lemma is almost obvious. A formal proof may be given
as follows. Let the lengths of the intervals excluded from S be 1, , I,, . . , , ZP. If
now we extend each of these intervals a distance I/Z A at eactl end,e we obtain
a system of p intervals of length

We denote what is left of (0, I) bv S’.

If (a + r)/l falls in S, the whYole of the corresponding interval of T falls iu
S. Hence tie part of S inside T has a length not less than ~,a/& where p i# the
number .of points (a + r)/l in S’. If v,, Y,, , . . , vP are the numbers of these
point8 which fall ‘in the intervala excluded from S, we have

2p-1,

1 It is of course to be understood that1 .an interval, or a part uf m iuterwl, which falk

outside (0, I), is to
be replaced by the congruent interval ineidle,
’ we suppose’ 1 large enough to emwe that. thi8 extension does wt cwae any overlay-
ping. If any part of an extended interval should fall outside (0, I), as wil2 happen if an inter-
val container o or 1, we of course replace t’hiis part by the congruent part of (0, I).

54
182 G, H. Hardy md J, IL Littlewood.

Hence the length in question is greater than

A similar argument, which we may leave to the reader, furnishes a corre-

spending upper limit for the length; and the lemma follows, It is plain that

r. 42. We can now prove the following theorem, which is a generalisation

of BERNSTEIN’S, but is itself contained in Theorem 1.40.
Theorem 1.42. If I is any intervul contuimd a’n (0, I), the set 0 of pints
0 such that no one of the points (A, 0) falh inside I, is of measure zero.
Let a be the centre of I and 6 its length; and let Tm be the set T of -the
lemma, with 1 =A,. If, for anv value of m, 0 falls in Tn, then (A,& falls in
I, and so 8 belongs to the set complementary to 0.
Let
S ,=T,+T,+--t-T,,

and let Zfl be the length of &. Finally let Zfl-+ I as n-m l We have to show
that I= I.
We noIv apply the lemma, taking S to be the set RY complementary to S,,
and T to be Tm. If m is large enough, the length of the common part (RR, Tm)
of S(,, and’ T, is greater than

Any point which belongs either to this set or to S, itself belongs to s~nze S,.
Hence
1 ) I,, + 6 (I - In) - E;
and so

which is impossible unless l- x.

I. 43. We can now complete the proof of Theorem 1.40. Let IV, be the
set of v~~lues of 0 lsuch that some one of the intervals

contains no point (LO). Then E, is of -measure zero, and 80

55
 Some problems of Diophantiae Approximation.

E=E,+E,+E,+......

is of measure z0ru.
If now the set (&J?> is not everywbere dense in (o, r), there ia an interval

i which contains no (LO). We can choose n so that Borne interval (f, T)

falls inside i. Then t? belongs to E, and SO to Em Thus the theorem is estab-
lished.
I. 44.Perhaps the most interesting special sequence falling under the ge:en-
era1t$pe (f w1 is that in which J(n) = CP, where a is a positive integer. When
8 is expressed as“ a decimal in the scale of a, the effect of multiplication by a
is merely to displace the digits. To study the properties of the sequence (a‘nH)
is thereforeequivalent ta studying the distribution of the digits in the expression
of B in the scale of a: it is to this fact that &his form of f(n) owes its pecu]jar
interest.
Let b be one of the possible digits O, I, 2, . ., a - x, and’let ~(n, m) denote
l

the number of decimals of 7c figures whose digits include exactly nz b’s, Then

n!
(1 ’ 441) Phm)=ml . (n-m)r(U-I)?‘-m*
l

We write

(1 ’ 442)

80 that p is. the excess of the number of Vs above the average.

We shall base our investigatjbn on a series of lemmas.
Lemma I . 441. Given any psitive number 6, we can find a p&be numbw
6 such that

where

for 1p I< &n and all sufficiently lurge values of n.

We omit the proof of this lemma, which depends merely on a straightfor-
ward applioation of STIRLIN~S Theorem.
Lemma 1 .442. Given any positive number E, we can find a positive nu7nbcl~
j- such that

56
 184 G. H. Hardy and 6. E. Littlewood.

jur 1p 12 8’32and all suiiicienlly large dues of n.

Suppose, e. g*, it> I en. Then

2

I
a---z--a&
p(n,m+xj z!z! n ---m < . 2 cI’

wbm) b--wm+I) @-I)(*+I_aE) ?

2

and from this it is easy to deduce the truth of the lemma when ji > &n. A sim-
ilar proof applies when !t < - sn.
Lemma 1 i 443. Let c be a positive constant. Then

(1 l 4431)

(1 ’ 4433) lim a-“&+,m)<r.

n-y=,
/I < c I/n

Of these three inequalities the first is plainly a consequence of either the

second or third. It will be, enough t-o pru.ve the second.
We have
(ti- 1)n/u
a-n~p(n,m)=a-n~+a-“~=s,+s2,
P> -c v-?t -c v-n en

say. By Lemma 1.442,

And 1)~
.I Lemma 1.441,

57
 Some problems of Diophantine Approximation. 185

The term of order x/l/k may be ignored. The remainder is less than

-c

which is less than I. Thus the lemma is proved. In a similar manner we can
prove
Lemma 1.444. rf v ifiq a function of n such that u/ I/n - 00, then

whire K ‘depends only on a.

We are now ‘in a position
I*45 to prove our main theorems. We observe.
first that all irrational1 numbers 0 between o and 1, whose decimals have just m b’s
in their first n figures, may be included in a set of intervals whose total length is

a -n P fn.4.
For let t?,, O,, . . . , 8,, where Q =a”, denote the terminating decimals of n fig-
ures. The set of intervals (Or, 0, + a-n) just fills Up the whole interval (0: I).
Among the numbers Or there are ~(n, m) lvhich have just rt2 b’s, which we may
(%a11i= It
319
-- and the set of intervals
521 l ’ ‘J &I;
(&. ES+ aen) fulfils our requirements.
Theorem 1.46. Let 8 be any positive nu.mber. Then the set of nT&ers 0
for which

Let $ denote the complementary set. Any number belonging to s satisfies

for an infinity of values of n, 8’ beingc anya. positive number less than &

l The end points of the intervals will be rationa numbers satisfying the condition. In
what follows we may confine ourselves to irrational values of 8, since the rational values forIn
in any ca80 a set of measure zero,
188 G. 11. Hardy and J, E. Littlewood.

All O’s for which this inequality is true for LI particular n may be enclosed
in a set of intervals whose total length is

0 '451)

We can choose a positive number d” such that

ztrld then choose n, so that the expression (I: .45x) is less than

for n --> n,. To provk the theorem it is enough to show that the result of sum-
ming this expression for n = n,, n, + I, . , . . . . can be made as small as we plea#se
by choice of .n,; and it is obvious that this conclusion cannot be affected by the
presence of the term a:-*. But

I’
-28’va----->o; d
-- 26’ d”
u va
and plainly

can be made as small as we please by choice of n,. Thus the theorem is proved.
Theorem 1.45 includes as a particular case
Theorem 1.451. If nb is the number of h’s in the first n figure& of the ex-
pressiun of (“I as Q decimal in the scale of a, then

for almOst all mlues of 0.

59
 Some problelns of Diophantine Approximation. 187

I .46. Theorem 1.45 shows that the deviation, from the average n/a, of
the number of occurrences of a particular figure b in the first n places, is not
in general of an order materially greater than V& If we were to suppose that
there was a steady deviation from the average (instead of a merely occasional
deviation), we would naturally obtain a more precise result. Thus reasoning
I analogous to, but simpler than, that which led to theorem x .45, leads also to
Theorem 1.46. If rp(n>- 00 with n, then the set of 0’s fur which

is mof measure zero.

This theorem, however, is included in a much more interesting and general
theorem which we shall now proceed ‘to prove, which, to put it roughly, assigns
a lower lim’it for the deviation‘ in either direction.
x . 47. Theorem 1 ; 47. If G is any positive constant, the set of O’s for which,

p(n) > -&,

and the set fur which !c (n) < c I/n, are of measure zero.
Let1

By Lemma 1.443, there is a positive number &, such. that

lima-” j$p(n,m)=~--&,.
n-* p>-c&b
And if c -e
< c, < c,,, it is clear that

ii&-=&7(n,m)=x-&,,,
?I-= p>--c,G
where

Let E, be the set of the theorem. We can enclose E, in a set of intervals

of total length

l lt follows from the elementa of the theory of errora that the ‘most probable error’ is of
order h,
1'88 G, H. Hardy and J. E. Littlewood.

a -n1 &I(?%,, ?n)< I--f&*

2
p> -c v-- 731

CunGder now any one of the

N=Cp(n,m)

intervals of this set, each of which is of length a-*1; and let %= (~3). As 0
ranges in the interval in question,. 5 ranges in the whole interval (0, I).
If tl belkgs to E,,* the corresponding 5 has the property that

,Ll(n’) > - c I/n, + 7%’

for all values of n’; and so, if n’ is large enough compared witb n,,

where
c’ = c (I + z---l).

We may now enclove the E’s in a set of intervals whose total length is.
less than

I
I: -- d c’ ;
2

and therefore w0 may enclose the 8’s which lie in the particular interval under
I
corkdera in a a& of interval6 Whose total length is less than a-*a (I- - & ).
2

If we do this for each of the N intervals, we have enclosed the B’s in a set of
intervals of length less than

Repeating this argument, it is clear that we can enclose the B’s in a set of
intervals of total length less than

61
 Some problems of Diophantine Approximation. 189

the indicea n, being integers which tend to infinity with v, as rapidly as we

please. Plainly &J < c, and so

As this tends to zero as Y- 00J our theorem is proved.

From Theorem 1.47 we can at once deduce
Theorem 1 .471. The set of Ok such that to each 0 corresponds a c for which
p b> > - cl& is of measurezeru.
Let E, denote the set of Theorem 1.47. The set of this theorem is plainly
the sum of the sets E, , E, , E,, . .; and so is of measure zero.
l

I .48. So far we have considered merely the occurrence of a particular

digit b in the decimal which represents 8. But our results are easily extended
so as to give analogous information concerning the. cwurrence of any combin-
ation of digits, Tho method by which this extension is effected is quite simple
in principle, and it will be sufficient to show its working in a special case.
Consider the succession 317 of digits, in the scale of 10. In the scale of
rooo, the number 317 corresponds to a-single digit z; and, if 0 is expressed in
the scale of 1000, it will, by theorem 1.451, be almost always true that the
number nz of occurrences of z, among the first n figures, satisfies the relation

Now the combination 3x7, in the expression of 8 in the scale of IO, will
occur when, and only when, the digit T occurs in the expression of one or other
of the three numbers

8, IO& 1008

in the scale of 1000. Hence it is almost always true that the number of occur-
rences of the combination 317, in the first n digits of the expression of 8 in the
scale of x0, is asymptotically equivalent to

n
looom

We may ROW, without further preface, enunciate the following theorems,

62
190 G. H. Hardy and J. E. Littlewood.

Theorem 1 .48. lt is almost always true thai, .when a number 0 is expressed

in aq scale of notation, the number of occurrences of any digit, or any combination
of digits, is asymptotically equivalent to the average number which might be expec’ted.
Theorem 1.481. It is almost always true that the deviation from the average,
in the first n places, is nut of order exceeding m.
Theorem 1 .482. It is almost always true that the &viation, in both directions,
is sometimes of order exceeding z/G.
Theorem I .483. The nuwiber uf the first n numbers (aV) which fall inside
an interval of length * 6 included in the interual (0: I> is almost always asymptu~icallly
equivalent to 6%
The laat theorem is merely a’ translation of theorem 1.48 intAo different
language, and a corresponding form may of course be given to theorem 1.481
and 1 .482.
1 .4g. Throughout this section (I .4) we have confined ourselves to results
concerning a single irrational f). Some of our theorems, however, have obvious
many-dimensional analogues. It will be sufficient, for the present, to mention the
foliowing, which are generalisations of Theorems 1 I 40 and 1.483 respectively.
The interval (0, I) is now replaced by an m-dimensional ‘square’,
Theorem 1. 49. The set of values (O,, 0, . .. O,), fur which the points
(Lo:, L&;’ l;inOm) are nut everywhere dense in the square, is of measure zero.
Theorem 1.491. The number of the first n points (avt),# uy O,, =-. avO,>, which
fall inside a portion of the square, uf areu 6, is almost always asymptotically equi-
valent to 6 n.
We leave the proofs to the reader. The first theorem may be proved by
an obvious adaptation of the proof of Theorem 1.40, and the second deduced
from Theorem 1. 483 by a process of correlation very similar to that employed
n 1.48.

63
 CORRECTIONS
p. 175, The last displayed formula but one should read:
---1 1 1
a,+ a,+ aa+**-
23. 177. The parenthesis in (1.338) should read: (0 < p < 5).
p* 185, Aemma; 1,444. Under the sign of summation, read: 1~1 > Y.
p. 185, laatfvmula. Read: 1~1~
p. 186, (1.451). Read: 1~1 > Y,,
p. 188. In the second displayed formula, read nfr for n.

COMMENTS
8 1.00. A simpler notation, which has come into use since the time of this paper, renders un-
necessary the use of fractional parts and the special convention when a= 0. Instead of writing

one now writes

-E < ln,iLa < e (modl), or l+J-~ll < E (mod I).
8 1.01, As mentioned in the footnote on p. 157, the form of Kronecker’s theorem which is stated
by Hardy and Littlewood as Theorem 1.01, and generalized by them in Theorem I ,011, is only
a particular case of the original theorem of Kronecker (though the general case is deducible from
it), The original theorem gave necessary and sufficient conditions for a system of linear equations
(not generally homogeneous) to have an arbitrarily good approximate solution in integers. The
condition in Theorem 1.01 that &,..., 6m shall be linearly independent irrationals is a particuIar
case of Kronecker’s condition on the urithmaeticaZ m%k of a system of linear equations. It is
customary now to express the condition in the form: 1, &..., 8,, are linearly independent over
the rationals.
A good account of Kronecker’s theorem, in the form considered by Hardy and Littlewood,
is given in ch. 23 of Hardy and Wright. Here several later proofs are given and their merits
compared. This account supersedes most of 8 1.1 of the present paper,
Theorem 1.011 implies the existence for any E > 0 (under the conditions stated) of infinitely
many integers n such that

I + 0 P --orK,J<e (modl) forl<p<m,lgK<k.

P. Szti~z (Acta Math. Acd Sci. Hwngaricae, 4 (1953), 115-18) has proved that there exists
a number JI(&...,&; k) such that there is an integer m with the property in every interval
(4 t+w.
8 1,14 (a). We have here an early result on one-sided, or asymmetrical, Diophantine approxima-
tion. It will be seen that Hardy and Littlewood realised the important distinction (in the case
of homogeneous approximation, when the Q?~ and BP are zero) between approximation from one
side and from both sides. For an investigation into this question, see C. A. Rogers, Proc. London
Math, sm. (2), 52 (1951), 186-90.
§ 1.14 (b). The two theorems stated here are early results on uniform distribution (mod 1). In
the first, it is to be understood that 1, 8, # are linearly independent, as in Kronecker’s theorem.
The second result includes as a particular case the uniform distribution of V@ (mod 1) for any
irrational 8, This seems to have been first proved by P. Bohl in 1909 in work on perturbation
problems, and independently by Sierpiriski and Weyl at about the same time (see Koksma, pa 92).

64
 In his important memoir of 1916 (Math. Annalen, 77 (1916), 3 13-52) Weyl gave his criterion
for uniform distribution in terms of exponential sums, and this provided a powerful technique
for proving the uniformity of distribution of many sequences. On the relationship between this
memoir and the work of Hardy and Littlewood, see 1916, 9. The method of exponential sums
also enabled Weyl to prove a stronger form of Theorem 1.011 asserting simultaneous uniform
distribution; see Koksma, p. 93, Salz 10.
The Bohr-Littlewood Cambridge tract, announced in footnote 2 on p. 164, never appeared.
But the Bohr-Littlewood manuscript formed the basis for Tit&marsh’s tract on the Riemann
zeta-function and (to some extent) for Ingham’s tract on the distributionof primes. Tit&marsh’s
tract developed later into The theory of the Riemnn xeta-function. (Oxford, 1951), and this
contains the applications of Kronecker’s theorem to the zeta-function mentioned in footnote 1
on p. 158 of the present paper.
More recently, Kronecker’s theorem was the starting-point for an important body of work by
Turk (see his book: Eine neue Methode in der Anal@s und deren Anwendungen, Budapest, 1953).
If 1, Ol,..., 8, are linearly independent, Kronecker’s theorem asserts that for any real al,*.., ‘31, and
any E > 0 there exist integers v for which

1+ 2 e*(@-crl) > m+ 1 -E*

I j=l I
But it is impossible to give any bound for v, or to name any interval in which some v must fall,
unless the hypothesis is strengthened by assuming some quantitative measure for the degree
of linear independence of o,, ,. l , flnt. The general idea of Tur&n’s work is that if one is content with
a much weaker inequality, it is possible (under reasonable conditions) to give good limits for Y
and (what is equally important) to omit the requirement of linear independence. The work
has applications to the zeta-function and to a variety of questions in analysis and in analytic
number theory.
5 1.31. The specific problem (p. 174) of finding a function Q(h) such that for any 8 there is a solu-
tiun of
In2e] < Aa1 (mod 1)
with n < @(h) has been answered most effectively to date by Heilbronn (@art. J. of Muthe 19
(1948), 249-56). He proved that if A is large (as one may suppose) the result holds with
@(A) = h2+-&, where 6 is any fixed positive number. A similar and more general result, but one
that is less precise for this particular problem, had been found by Vinogradov in 1927 (see
Koksma, p, 119). The conjecture, described by Hardy and Littlewood as ‘not unnatural’, that
perhaps one could take D(h) to be proportional to h, is, however, false. For take fl = a/p, where
p is a large prime of the form 4x+ 1 and a is a quadratic non-residue (modp). On choosing h
so that G)(h) = p - 1, we should deduce from the conjecture that there is a quadratic non-residue
(modp) which is bounded independently of 23, and this is obviously false.
5 1.33. The conjecture at the beginning of this section, which Hardy and Littlewood say they
believe to be improbable, is in fact false. This is a particular consequence of results found later
by Blichfeldt (see Koksma, p. 86, Satz 6).
3 1.34. There were earlier results, not mentioned by Hardy and LittJewood, of the same character
as Theorem 1.341, Thus Chebyshev proved in 1866 that there are infinitely many n > 0 for
which < 2/n (modl)
~n&--~~ (1)
(Koksma, footnote on p. 76). This is equivalent to #(n) = @ in the present notation. Minkowski
proved that provided 01is not of the form IZ~&--mm,, there are infinitely many integers n for which

In&al < 4;
-
n
I (modl);

here n is not restricted to positive values. This result is best possible of its kind. The best possible
result when n is restricted to positive values was found by Cassels (Math. Annulen, 127 (1954),
27
288-304); the constant which replaces 2 in (1) above is m, and the proof is difficult.

65
 The conjecture concerning t$(@ when m > 1 is false, as follows from the results of BZichfeldt
mentioned above. There is, in fact, a vita1 difference in problems of non-homogeneous approxima-
tion between the one-dimensional case and the many-dimensional case.
8 1.4. This section is an early contribution to the subject now called ‘metrical’ Diophantine
approximation. Almost the only previous papers were those of Bore1 and Bernstein quoted on
p. 179. For a recent account of the subject, see Cassels’s Tract, ch. 7. The later part of the section,
beginning with 3 1.44, is concerned with ‘normal decimals’, and on this subject the reader should
consult ch. 9 of Hardy and Wright.
The notion of an exceptional set of Lebesgue measure zero can be refined by considering the
Hausdorff measure of the exceptional set. See, for example, Eggleston, Proc. Lolzdolz Mu& Sot.
(2), 54 (1952), 42-93.

66
 SOMEPROBLEMSOF DIOPI-IANTINEAPPROXIMATION.
I

RY

G. H. HARDY and J. E. LITTLEWOOD,

TRZ~ITY COLTAGE, CAMBRI~IQE.

II .

The trigonomet&cal series associated with the elliptic

3-functionsm

2. 0. - hltrodnction.

2. 00. The series I

00
2s b-1)1 2, 1 + 2 2qn2, I + 2 f& (- I)” qn2,
1 1 1

where q = eniT, are convergent when the imaginary part of z is positive, and
represent the elliptic a-functions

When z is a real number x, the series become oscillating trigonometrkal series

which, if we neglect the factor z and the first terms. of the second and third
series, may be written in the forms

l The notation is that of TANNERY and &TX’S Thht-ie des functiim etliptiques. 1% shall
refer to this book as T. and M.

1914, 3 (with J. E. Littlewood ) Acta Mathematics, 37, 193-238. 67

 194 G. H. Hardy and J. E. Littlewood.

These series, the real trigonometrical series formed by taking their real ut’
imaginary parts, and the series derived from them by the introduction of con-
vergence factors, possess many remarkable and interesting properties. It was
the desire to elucidate these properties which originally suggested the researches
whose results are contained in this series of papers, and it is to their study
that the present paper is .devotedJ /
2. or, We shall write v L

It is obvious that, if sn is any one of a,‘,, & &, then

(2. 0x2) I-c n = O(n).

Our object is to obtain more precise information about tell; and we shall begin
by a few remarks about the case in which x is rational. In this case So is
alwnvs of one or other of the forms

o(l), An + O(I),
where A is a constant. It is not difficult to discrimina#te between the different
cases; it will be sufficient to consider the simplest of the three sums, viz. &.
We suppose, as plainly we may do without loss of generality, that x is
positive. Then x is of one or other of the forms

according as the denominator of 5 = Ex is congruent to o, I, 2, or 3 to mod-

2
ulus 4.

1 So~ne of the properties in question are stated shortly in our paper ‘Some problems of
Diophantine Approximation’ published in the Promdings of the fifth Idemational &ngmss of
Mdhematicians, Cambridge, J912.

68
 Some problems of Diophantine Approximation. 195

Now it is easy to verify that

is of the forms

according as 8 co, I, 2, 3 (mod. 4); and from this it follows immediately that
8:: is of the forms

(k I: h i)An + O(r),

ztAn + O(I),

k iAn + O(I)?

in these four cases. Thus, for example, the series

oscillates finitely if x is of the form (zk + I) / (2,~ + I) or ZA / (4 p + 3), and

diverges if x is of the form (zA + X)/Z/L or zA/(4,~ + I)/

2. I. - 0 and o Theorems.

2. IO. We pass to the far more difficult and interesting problems which
arise when x is irrational. The most important and general result which we
have proved in this connexion is that

(2, 101) &a = 0 (n)

for any irrational x. This result may be established by purely elementary

reasoning which can be extended so as to show that such series as

I This result (or rather the analogous result for the sine series) is stated by BROMWTCH,
Ziz$nite Series, p. 485, Ex. IO. We have been unak.Ae to find any complete discussion of the
question, but the necessary materials well be found in DIRJCHLFT-DEDEKIKD, tTorZesungen iiber
Zahlenthm-ie, pp. 285 et aep. See also RIEYANN, Werke, p. 249; GENOCCHI, Atti 6% Twino, vol. IO,

69
 196 G. H. Hardy and J. E. Littlewood.

also possess the same property. We do not propose to include this proof in the
present paper. Although elementary, it is by no means particularly easy; and it
will find a more natura.1 place in a paper dealing with the higher series (2. 102).
In the present paper we shall establish the equation (2. IOI) by argumentIs of a
more transcendental, though really simpler, character, which depend ultimately on
the formulae for the linear transformation of t#he a-functions, and will be found
to give much more precise results for particular classes of values of z.
2. ZL It is very easy to see that, as a rule, the equation (2. 101) must be
very far from expressing the utmost tOhat can be asserted about sm.
It follows from the well known theorem of RIESZ-FISCHER that the series

(2. III)

are FOURIER’S series. Hence, by a theorem of W. H. YOUNQ~, it follows

that they become convergent almost everywhere after the introduction of a
convergence factor 7~~’ (6’> 0). As 8 and 6’ are both arbitrarily small, the
series themselves must converge almost, everywhere. Hence the equation

(2. 112)

must hold for almost all values of X. It is evident that the same argument
may be applied to 8: and st, and to the analogous sums associated with such
series as (2. 102).
If, instead of the series (2, III), we consider the series

cos n”mx sin &TX

(2. 1x3) 2 9 I: 9
nf (log ,,+ nf (log n)fr+6

and use, instead of YOUNG’S theorem, the more precise theorem that any
FOURIER’S series becomes convergent almost everywhere after the introduction of
a convergence factor 1:/log ~8,~ we find that we can replace (2. x12) by ihe’
more precise equation

’ HARDY, ~~mOc.pond. ~afhw Hoc., vol. 12, p. 370. The theorem was ah discovered hde-
pendently by M. R[ESZ.

70
 Some problems of Diophantine Approximation. 197

and it: is evident that, we can obtain still more precise equations by the use of
repeated logarithmic factors. These we need not state explicitly, for none of
them are as precise as those which we shall obtain later in the pa,per. These
latter results have, moreover, a considerable advantage over those enunciated
here, in that the exceptional set of measure zero, for which our equations may
possibly cease to hold, will be precisely defined instead of being, as here,
entirely unspecified. The main interest of the argument sketched here lies in
the fact that it can be extended to series such as (2. IOZ)?
2. rzo. We proceed now to the analysis on which the principal, results of
the paper depend. These are contained, first in the equation (2. XOI), and
secondly in the equation
(2. 1201) s,&=o(l/n),

which we shall prove for extensive classes of values of x.

In Chap. 3 of his CaZcuZ des Rthidus, LINDEL~F gives an extremely elegant
proof of the formula

(2. 1202)

where p and p are positive integers of which one is even and the other odd?
Our first object will be to obtain, by an appropriate modification of LINDEL~F'S
argument, ardogous, though naturally rather less simple, formulae, applicable
to the series &n2nix, where x is irrational, and to the other series which we
are considering.
We shall, however, consider sums of a more general form than those of
which we have spoken hitherto, viz. the sums

l The argument may even be extended to series of the type SAS” where & is not
mcemwily a mdtiple of x; but for this we require a whole series of theorems concerning
Dmmnm’s series.
3 The formula is due to GEN~CCHI and SCHAAR. See LINDET~%, 2, T,, p. 75, for references
to the history of the formda.
 198 G. H. Hardy and J. E. Littlewood

Here x and 0 are positive and less than I, x is irrational, and n is not neces-
sarily an integer. These sums are related to the functions 3, (v, z-), I 1 as &, . .
l l

are related to &(o, z-), . .

l

2. IZI. We consider fhe complex integral

taken round the contour C shown in the figure. We suppose that the points
o, n are in the first instance avoided, as in t.he figure, by small semicircles of
iH /

where P is the sign of CAUCHY’S principal value, and the dashes affixed to the
sign of summation imply that the terms for which Y = o and Y = n are to be
divided by 2.
We shall find it convenient to divide the contour C into two parts G, and
C,, its upper and lower halves, and to consider the integrals along C, and C,

72
 Some problems of Diophantine Approximation. 199

separately. When we attempt to do this a difficulty * arises from the fact that,
owing to the poles of the subject of integration ak x =o and z = n, the two
integrals are not separately comergeM. This difficulty is, however, trivial and
may be avoided by means of a convention.
Suppose that f(x) is a real or complex function of aI real variable x which,
near 3: = CI, is of the form

G
- + y (4,
X-CI

where cp(xj is a function which possesses an absolutely convergent integral across

x = U; and suppose that, except at x = CY, / (x) is continuous in the interval
(a, A), where a<a<A. Then CAUCHY’S principal va,lue

P *j(x)ax
J
a

exists; but f(x) has no integral in any established sense from a to CI or from CI
to A. We shall, however, write

P 9f (x)dz = lim
J e-0
a

A
.
P ‘f(x)dx= llm “f(x)dx + C log E
J e-0 J

and it is clear that, with these conventions, we have

a 4 A

p f(x)dx+P f(x)dx=P f(x>dx.

aJ’ Ja s
a

It is clear, moreover, that a similar convention may be applied to complex

integrals such as tOhose which we are considering; thus

73
200 Cr. H. Hardy and J. E, Littlewood.

(taken along the line o, iH) is to be interpreted as meaning

We may now write (2. 1211) in the form

where xlow C, and C, are each supposed to be described starting from o. In

the first of these two integrals we write

and in the second

2'1
cotnx=-c-----r---*
,-22n’L- I

The two constant terms in these expressions give rise to integrals which may be
taken along the real axis from o to n, instead of along C, and C,; uniting and
transposing these terms we obtain

nI
n
(2. 1213) f ev”nix CO8 2wco -
z ez2nx cos zxd..dx= II -j- I,,
I i
0 0
where
efni= cos 2 x n r')
I 1= P e- 2Zd _I 1 dx 9

Cl
i'

I Z eiY2fiixcos 2 2 7-c0
2 P e2zni, 1 a2 l

1 i
ca

We now write
I e2kzni
= e2zni + e4z3-ck + . .. + e2(k-l)zni + _eZzni
e- zzni _ 1
I

in I,, and
 Some problems of Diophantine Approximation. 201

in I,. If we observe that

12
= 2 ezanix co8 2vm co8 227&h,

I,s

we see that (2. 12x3) may be transformed int.o

k-1 jG
(2. 1214) i]tev2nixcus 2~nt9---2 2 e*2-:ix cos wnx cos 2x~rOdx = K, + K,,
I
0I o i'0
where

e-2 kz,zi

I - e---3rJ-d ax l

2. 122. We shall now suppose that H 4 a,, so that the parts of C, and
C, which are parallel to the axis of x go off to infinity. Tf x = $ + iv, and v
is large and positive, the modulus of the subject of integration in K, is very
nearly equal to

f exp -zq?(k+gx--0) ;
i i

while if x=6--iv, and 7 is again large and positive, the modulus of the sub-
ject of integration in K, is very nearly equal to

From this it follows immediately that, if

(2. 1221) k>nx+O,

the contributions to K, and K, of the parts of C, and C, which we are causing

to tend to infinity will tend to zero.
 202 G, H. Hardy and J. E. Littlewood.

We are now left with two integrals each of which. is composed of two
parts taken along rectilinear contours, and ‘we may write

Of the four rectilinear integrals thus obtained two, viz. the two taken along
the imaginary
, axis, cancel one another. In the other two we write

z = n + it, 2: = n -it

respectively, and then unite the two into a single integral with respect to t;
and when we substitute the result in (2. 1214) we obtain

where

2. 123. We now write

and we proceed to show that

(2. I23I)

uniformly in respect to 0, hy which we imply that there is an absolute con-

stant A such that

76
 Some problems of Diophantine Approximation. 203

for 0<x<1, oQWI,

-- all values of ?i, and all values of k subject to the in-
equality (2. 1221).
We may plainly ignore the factor ien2~~~ in K. The factds in curly brack-
ets is equal to

z(cos znjz0 cash &-CO sinh znxat + i sin anzO sinh 2tzO cash 2nx4.

The factor e--fa”Gi” we separate into its real and imaginary par& When we
multiply these two factors together our integral splits up into four, of which
the integral
ul
,-ant
(2. 1232) cos PACX cash zt~9 sinh znxzt __ e--2& at
I
I i‘

is typical; and it will be sufficient to consider this integral, t!he same arguments
applying to all four.
The function I / (r - e-2nt ) decreases steadily as t increases from I to a0 l

Hence, by the second mean value theorem, the integral (2. 1232) mavY be writ-
ten in the form
T

(2. 1233) A ‘cos tkx cash 2td sinh znxzt @knW,

i
1*,

where A (as always in this part of the paper) denotes an absolute numerical
constant, and T > I. In (2. 1233) we replace the hyperbolic functions by their
expressions in terms of exponentials; and the integral then splits up into four,
of which we need only consider

the arguments which we apply to this integral applying u fortiori to the rest.
The integral (2. 1234) may, bvd another application of the second mean value
theorem, be transformed into
 204 G. II. Hardy and d. E. Littlewood.

Now, if T and I” are any positive numbers whatever, we have

T4 Tt I&

cos t%c.x dt = I cm au2 du;

t/X s
TG

and the integral last written is less in absolute value than an absolute constant.
We have therefore proved the equation (2. 1231), and it follows that

(2. 1236) 2' ev2"ixcos 2 v7c0---2 6*2ai=cos 2YncX CQS zz7EOdz s K' + 0 I/$
I
0

2. 124~ The next step in the proof consists in showing that, in the equat.ion
(2. r236), k may be regarded as capable of variation to an extent O(I) on either
side, that is to say that we may repla,ce k by any other integer k’ lying between
k--A and E + A, without affecting the truth of the equation. That this iS
so if k is increased is obvious from what precedes, as the inequality (2. 1221) is
still satisfied; but when k is decreased an independent proof is required.
We consider separately the effects of such a variation on the two sides of
the equation (2. 1236). As regards the left hand side, it ia plain that our
assertion will be true if

n
I
e2,ix cus 2x7~ dx = 0 v- ;
s
0

uniformly for all values of m and 616,and therefore certainly true if

7a
‘1 -
@isf2X3TiU~~ x I-1
J
0
OV X

’ The ~4 in thig formula ia of cmme not the mme numericaI con&ant ag before.

7%
 Some probIelns of Diophantine Approximation. 205

But

and this expression is evidently of the form desired.

We have now to consider the effect of a variation of k on the right hand
side of (2. 1236). The difference produced by such a variation is plainly of
the form
1
‘I g-2kd - e-2k’nt 1
0 e2dbmfO) at
x __ e-2nt

-_f*
=0(+z0
v 35
Thus finallywe may regard the k; which occurs on either Hide of (z. 1234 a3
capable of variation to an extent# 0 (I),
2. 125. We proceed now to replace the integrals which occur on the 1eft
hand side of (2. 1236) by integrals over the range (0, 00)~ We write

Now consider the integral

r
w
ezznix cos 2 unz co9 2x9~0 dz,

taken round the rectangular contour whose angular points are n, n + N,

79
 206 G. H. Hardy md J. E. Littlewood.

n+N+iH, n+iH. The modulus of the subject of integration is less than

a constant multiple of
~-2J-rq(~x-.v--~).
t

and from this it is easily. deduced that, if

the contributions of the sides (n, + N, n + N + iH) and (?z + ZV + iH, n + iH)
tend to zero as N and H tend to infinity, and so that the second integral
which occurs in our expression for I, may be replaced by one taken along the
line (n, n + ice). In order that this transformation may be legitimate for
Y--O, I...., #--I we must have

(2. IZjI) k’<nx+ 1-0.

It is important to observe that this condition and the condition (2. 1221)
cannot always be satisfied with E = I?; but that the difference between the least
k such that E > nx + 0 and the greatest E’ such that k’< nx + I -0 cannot be
greater
I than I 2
On the assumption that (2. 1251) is satisfied, we have

a0
I
sinh (2 k’--1)7zt
C’ i e(nz-t2)rcix-Znxnt cos 2(n + it)760 dt
sinh zt

say; and so, bearing in mind the results of the analysis of 2. 124,

00
n ’ 72-l
(2. 1252) 7’
2 re y2Jrix
cos ZY720- ,,
22 J et2aixCOS2unz cm zxn0 dx
0 Q 0

’ Lt in these fact8 which render necessary the analysis of 2. 124.

80
 Some problems of Diophantine Approximation. 207

2. 126. We next write

.,r-i!,Jd+Lff,
0 6 1

and we proceed to show that

--
I
L II E 0 v -9
X

SO that L may be replaced by L’ in (2. 1252). The argument is practically the

same as that of 2, 123. We have to consider a number of integrals of which

w
sinh (d- I) rrtdt
(2. 1261) cm t%rx cash ztw9 e--2*xrrt
i sinh at
i

is typical. Writing 2 e-“j / (I - e-2.zt) f or cosech z t, observing that the factor

r/ (I -e-2nt) is monotonic, and using the second mean value theorem as in
2. 123, we arrive at the result desired.
We may accordingly replace L by L’ in (2. 1252). And our next step is
to show that the Z which occurs in this modified form of (2. ~252) may be
regarded as capable of variation to an extent 0 (I). Here again our analysis is
practically the same as some of our previous work (in 2. 124), and there is
therefore no need to insist on itUs details. We may now write (2. 1252) in
the form

n
I
(z r262) ev2nix co3 2 vmo- ezairix COS 2Y7t:Z COS 2X7d dX
I
0

where

(2. 1263)
I
sinh @IL-z) St
i e3-cix(nz-f+2ftXd ~0s 2 (n no at ;
s
sinh 7Gt
0

81
208 G. El. Hardy and J. E. Lit&wood.

and, as the k’s which occur in these equations may all be regarded as capable
of variation to an extent 0 (I), there is no longer any reason to distinguish
between k and k’.
2. 127. Again

- 2 e-2nxJtt sinh (zk- I) nt cm 2(n + it);70

=zcos 2n~8 cash ztzOsinh(znx-2k+ x)rrt

+ 2i sin 2727~0 sinh 2trrO cash (znx- zk + I)&.

We select the value of k for which

--r<mx-2k+1<1;

a#nd the integral (2. 1271) splits up into two, of which it will be sufficient to
consider the first, viz.

1 ainh (znx-zk + 1)7rt

(2. 1272) i cm zn7-50 &X(n”-12)
cash 2~0 dt .
J‘ sinh it

This is of the form

1
0 (1) ‘,-/?Gx sinh crzt
(2, 1273) cash ztd dt
sinh zt ’
0

where a= Iznx- zk + 11. It will be enough to consider the real part of this
integral, the imaginary part being amenable to similar treatment.
The function
sinh amt
(O<or<I)
sinh srtt

decreases steadily from a as t increases from zero. Hence

 Some problems of Diophantine Approximation. 209

-1 sinh crat
cos XXP cash ztzO at cos sxtB cash ztw9 dt
d0 sinhct =Q s
0

= CI cash 2zd cos xxts dt,

z’s

z and z’ denoting positive numbers less than I. Since o<a<x, 050<1, the
first, factor here is of the form O(I); and the second is (cf. 2. 123)
c of the form
I
OV X
5. Hence finally
-
Q--Q=0 :,
v X

and so the left hand side of (2. 1262) is itself of the form 0 5.
X

2. x28. But

Substituting this expression in (z. 1262), and observing that k may now be
supposed to be the integral part of nx, we obtain

Theorem 2. 128 If o<x< I, o<0(1, then

-
where 0 k denotes a function of n, x, and 0 which is in absolute value less

than a constant multiple of

We have omitted the lower limits of summation, and the dashes, which
are now plainly irrelevant.
We can also prove, by arguments of the same character as those of
55 2. 121 et seq.,

l LINDEL~~F, 2. c., pm 44.

83
210 G. H. Hardy and J. E. Littlewood.

Theorem 2. 1281. Under similar mndil ions

It will hardly be necessary for us to exhibit any detai 1s of the proofs,

and we will only remark that the integral

of 2. 121: is replaced by one or other of the integrals

It is on the transformation formulae contained in Theorems 2. 128 and 2. 1281

that all the results of this part of the paper will depend.
2. 13. We have the following system of formulae:

s; (x + 1, 0) = I/i s; (x , 0),

&(-x3 O)= &(x, 0))

(2. 131) q-x, O)= &(x, O),

-.
&(x,O)=v 2,-dzjx
px X
(-;, $) + 0 j/i,
 Some problems of Diophantine Approximation. 211

Here & denotes the conjugate of slz. 11; will be convenient in what follows to
r
write 0 in the equivalent form
x
O(I
1

Now suppose bhat x is expressed in the form of a simple continued fraction

(2. 132) f_+I + A...,

% u, a,
and write

I =-- I
E--
a, + x,’ Xl
X
(2. 133) a, -+ x2’ ’ l ’ ’

01 =_--[I0- 0=0
0
X
--1 [I0-1x ’ 2 Xl Xl ’ l ‘*’

so that
O<Xr<I, O(&<I

for all values of r. Further, let 1, denote an unspecified index chosen from
the numbers 2, 3, 4; and. let LO denote a number whose modulus is unity but
whose exact value will vary from equation to equation.
This being so, we have

sgx, o)=&;x ---I, 0- + Oh)

-
if-X ( x x 1 I/ x

__-coSk1(---q--x,, 8,) + -O(I)

v- X
ns
I/x

- co .1
-P S1 ( -x,, 0,) + O$
l/x nx

cd - .l Q(I)
E- S3 ( Xl, 01) + T’
vz -nx X

Transforming sk (x,, 0,) in the same way, we obtain

212 G. H. Hardy and J. E. Littlewood.

Repeating the argument, we find

(2. 134)

Now
I
Xr<-3
- 1 + Xrfl

x,+1
(2. 135) xr x,+1 -1< - <I
+xr+1 2'

and so xx,..,xr-o as r-00. We may therefore define v by the inequalities

(2. I36) ?&xx,... xv-1 xy < I <n x x1 l l l Xv-l.

This ‘being so, the first of the equatious (2. 134) gives

sA(x, 0) = 0 (?2Vzx,...x,-l)
(2. 137)
I

z/x x, l l . xv-1

and the second gives

(2. 1371) +, @=0(I) &+ & + ‘*’ + vxx

’ x

1”’ y-
1

We have thus two inequalities for st(x, O), the further study of which depends
merely on an analysis of the continued fraction (2, 132). These inequalities,

however, may be simplified. For, by (2. 139, x,x~+~ < 5, and 80

2

I
=
l/xx, . l . xv-1
 Some problems of Diophantine Approximation. 213

K
<
1/2x, l . . X,-l’

Hence (2. 137) may be replaced by

I
(2. 138) sk(x, i9)=O(?dxx,. . . x+1) + 0 ;
v xx, . l . X,-l

and similarly (2. 1371) may be replaced by

I
(2. 1381) Sk (x, 0) = 0
1/zx, . l . x,-lx,’

2. 14. From (z. 138) and (2. 1381) we can very easily deduce the principal
resulfs of fhis parf of the paper.
Theorem 2. 14. We have

%4(x, 0) = o(n)

for any irrational x, and uniformly for all values of 0. In particular, if 0 = O,

we have
= o(n)
s?&

Since nxx, . . .x,-fl_x, the second term on the right hand side of (2. 138)
is of t.he form 0 (I&). And since xx, . . . xVB1-o as V- 00, the first is of the
form Q(n). Thus the theorem is proved.
Theorem 2. 141. If the partial quotients a, in the expression of x as u con-
tinued fraction are limited, then

ST&(X,0) = 0 (VG),

uniformly in respect to 0; and in prticular

$,=0(G).

These results hold, for example, when x is any quadratic surd, pure or mixed.

87
214 G. H. Hardy and J. E. Littlewood.

For, if (zIt < K, xv lies between

I K
K’ K+I
and so
xx,... xy-~xy>xxl*.. x,-,/K>rl(nK).

Using (2. 138x), the result of the theorem follows.

Thearem 2. 142, If us = 0 (ne), then

Theorem 2, 143. If a, = O(e@) s where Q < F log 2, then

2

‘+ Q+E
sn (x, o>= 4 n2 log2 ) 3
for any positive vdue of E.
For
I 1
xx* ,..x,<-~xx,...x,_l<2-2~ 9
n-

where p = Y or p = v - I, according as Y is even or odd.. Hence

ly < (2 + 4 log n
log 2 ’
But
xx,... xv > HY-@xx, . . . X~-~,

where H is a constant,, and so

I
=O@ 6) = *(n: (log nJ2}*
vx x, . . , xv

This proves Theorem 2.142. Similarly, under the conditions of Theorem 2. 143,
we hare

I = ()(eb fi) = 0 (n~+i&+~).

1/5x,. l l xy
 Some problems of Diophantine Approximatioll, 215

2. 15, Suppose now that y(n) is a logarithmico-exponential functioll’

(L-function) of n such that the series

is, to put it roughly, near the boundary between convergence and divergence,
so that bhe increase of rp (n) is rlear to that of n. Then, arguing as in 2, 14,
we see that, if an=O{p(n>),

I: =OVii-gq = 0 v?vp(log?z)’
1/x5,. l . xy

Now it has been proved by BOREL and BERNSTEJN” that the set of values of
x for which
a, = 0 -[y (?a)>

is of measure zero when the series (2. x51) is divergent, and of measure unity
when the series is convergent. Hence we obt.ain
Theorem 2. 16. I/ v(n) is a logarithmico-ex~onenliaE fumdion of n such
that

is converyent, then

Sn = 0 vncp (lug n)

for almost all values of x. In prticuluq if d is positive, then

for almost all values of x.

It was this last result to which reference was made in 2. XI.

1 HARDY, ora of Incfinity, p. 17.

89
 216 G, H. Hardy and 3. E. Littlewood.

2. r6. Suppose that a series 2~” possesses the property that

+ being 8 function which tends steadily to infinity with n; and let (p be a

function which tends steadily to zero as n - 00, and satisfies the condition that

is convergent. Then it follows immediately, by an elementary application uf

ABEL'S transformation, that the series

is convergent, This obvious remark may be utilised to deduce a number of

l

corollaries from some of our theorems. To give one instance only, it follows
from Theorem 2. 15 that the series

is convergent for almost all values of x, and, for any particular x, uniformly
with respect to 0.
A rather more subtle deductlion can be made from Theorem 2. 14. It does

not follow that, because an= o (n),

the serieg 2 5 is convergent; and indeed
n
we shall see later that it is not true that (e. g.) the series

enZ&X
(z. 161)
z n

is convergent for all irrational values of x. But it t’s true that, if sn = o (n),

the series 2 5 is either convergent or not summable by any of CES~RO’S meansl;

?t
and t.his conclusion accordingly holds of the series (2. 16~). Similarly, if x is
such that a, =O(I), the series

1 HARDY and LITTLEWOOD; Pvoc. irmcl. Mzth, SOC., Vol. 11,pm433.

90
 Some problems of Diophantinc Approximation. 217

possesses the same property. We shall see later that it is the second alternatiw
which is true.
2. 17. So far we have dealt with series in which the parameter 0 occurs
in a cosine co8 znz0 or cos (zn- r)&. It is naturally suggested that similar
results should hold for the corresponding series involving sin znxd and
sin (2 n - r)rrO; and this is in fact the case. These series are, ,from the poirlt of
view of the theory of funcbions, of a less elementary character: they a,re not
limiting form8 of series which occur in the theory of elliptic functions. But it
is not difficult to make the necessary modifications in our analysis.
We write

aqx,0)= Ze(v-iil )‘nixsin (2~ -1)7d

aAn

u~(x,O)=~(-I)Ye@nixsin 2vd
v<n
-

Theorem 2. 17. If O<Z<I, OQWX,

-A then

unifurmly in respect to 0,
Let us consider, for example, the second of these equations. We start
from the integral

erznix sin 22~4 745 cot zx dz,

s

and we arrive, by arguments practically the same as those of 2. IZI-2. 127,

at the equation
m -
I
(2. 172) 2 ev2;rix sin 2 2,a0 -2 x i’ ex2~ixcos 2~7-a sin 2xrdl dx-0 v -.
X
0

91
 218 G. H. Wardy and J. E. Littlewood.

The only substantial differences between the reasoning required for the proof of
this equation and those which we used before lie in the facts, first that some
of the signs of the principal value which we then used are now unnecessary,
and secondly that the two integrals dong the axis of imaginaries no longer
cancel one another. These integrals, however, are of the form

w
e-2 kd
e--t2ab sin h z f z 0 p - e-2nt at ’
s
0

and are easily seen to be small when il: is large. They are accordingly without
importance in our argument.
The integrals which occur in (z. 172), unlike the corresponding cosine
integrals,
”
cannot be evaluated in finite form. We have, however,

09
2 &nix cos zvflx sin2mO dx- I(u -t tl)-I(Y-Oo),
I^

where

Now let us consider the integral

s ez~~ix+2~niA dx
(A > 0)

taken round the contour defined by the positive halves of the axes and a circle
of ra#dius R. It is easy to show, by a type of argument familiar in the t#heory
of contour integration, that the contribution of the curved part of the contour
tends to zero as R- 00. Hence we deduce

and so

92
 Some problems of Diophantine Approximation. 219

CQ

cos zx7cAdz + e-t%-cis-2tnA d t ,

.i
0

Again, it is easy to show

e-t%ix-ztnA d t =- P I
A +“Ll) -3 J
ir

where /I = I/Z z. Hence

m

I(Y $O)-I(v--8 =i ez2~ix~~o~ 2(v+ t9)7tx- co3 z(u--O)m> dx

i
6

p:
+ --- P
(2. 175) v+0 u-0 +o ($ 1

From (2. 171)~ (z. 173), and (2. 175) we at once deduce the second equation of
Theorem 2. 17; and the others may be established similarly.
2. 18. From Theorem 2. 17 follow the analogues for t:he sums 0 of those
already established for the sums s. Thus we have
Theorems 2. 18, 2. 181-d. The remits established in Theorerns 2. 14,
2. r4r-3, 2. 15, for series ~Wolving c&rteS, are true also for the corresponding
series involving sines.
2. 19. The preceding results have a very interesting application to the
theory of TAYLOR'S series.
Let

be a power series whose radius of convergence is unity, and let, as usual, 2M (r)
denote the maximum of 1f 1 along a circle of radius r less than 1. Further,
suppose that
M (r) = 0 (I - ~)-a,

93
 220 G. H. Hardy and J. E. Littlewood.

and let

Then it ia known that1

1
-a---

SW r) 2*

Further, it is known that the number ?- occurring in the last formula cannot
2
be replaced by any smaller number, that is to say that, if d is any positive
number, a function f(x) can be found such that the difference between the

orders of g (r) and M (r) is 5 - $ ? But so far as we are aware, no example

2

has been given of a function f(x) such that the orders of g (r) and 2V (r) differ

by as much as 5. We are now in a position to supply such an example.

2

Let

where 5 is an irrational of the type considered in Theorem 2. 141, so that the

partial quotients in its expression as a continued fraction are limited. Then, if
x = re2xi8, we have, by Theorems 2. 141 and 2. 181,

, I
,

uniformly in 0; and from this it follows that

f (2) = f (rtFfl

uniformly in 8. Hence

M(r)=O1/--I-9
I -r
while

l HARDY, QuarterZy JmmaE, Vol. 44, pmMm

* HARDY, 2 c,, p. 156.

94
 Some problems of Diophsntine Approximation. 221

Thus the orders of g(r) and M(r) differ by exactly I* If we consider, instead
2
of j(x), the function

we obtain in the same way an example of 8 function such that

M(r)=O(r-r)-a,

These examples .show that the equation

M(r) = 0 (I -.rj--” (a > 0)

does not involve

1
= 0 (I - r)-a-2;
. go r

a possibility which had before remained open.1

2. Ig. Theorems 2. 14 etc. also enable us to make a number of interesting
inferences as to the behaviour of the modular functions

z (fi-I-
>4
Q 2 Y Znn2, X(-I)“Qnz

as Q tends along a radius vector” to an irrational place eait on the circle of con-
vergence. Thus from Theorem 2. 14 we can easly deduce that, if f(p) denotes
any one of these functions, then
1

J(q) = O(I -1qp;

and from Theorem 2. 141 that, if 5 is an irrational of the class there considered,
then

’ ’ HARI)Y, 1. C,, p. 150.

a Or along any ‘regular path’ which does not touch the circle of convergbnce.

95
 222 G. II. Hardy snd J. F,, Littlewood.

These results are, however, more easily proved by a more direct method, which
enables us at the same time to assign certain lower limits for the tiagnit*ude of
If (dL and to show that Theorems 2. 14 et seq are in a certain sense the best
possible of their kind. It is to the development of this method, which depends
on a direct use of the ordinary formulae for the linear transformation of the
8-functions, that the greater part of the rest of the paper will be devoted.

2. 2. - Q Theorems.

2. 20, We have occupied ourselves, so far, with the determinlttion of cer-

tain upper limits for the magnitude of sums of the type sn* Thus we proved
that sn = o(n) for any irrational x, and that sn = O(G) for an important class
of such irrationals, including for example the class of quadratic surds, But we
have done nothing to show that these results are the best of their kind that
are true. The theorems which follow will show that this is the case,
We shall begin, however, by proving a theorem of a more elementary cha-
racter which involves no appeal to the formulae of the transformation theory.
Theorem 2. 20. Suppose that cp(n) is a positive decreasing function of n,
such that the series &(n) is divergent. Then it is possible to jind irrutionals x
such that the series
2 rp (12)en2,ix

is not convergent. The sume is true of the series

and of the wd and ilnaginary

prts of all these series.
Consider, for example, the real part of the first series. We shall suppose
that, among the convergents &qy to x, there are infinitely many of the form
W4P + I>* Let (4;) be a subsequence selected from the denominators of these
convergents. We are clearly at liberty to suppose that the increase of ay+l,
when compared with that bf any number which depends only on qy and the
function q, is as rapid as we please,
We shall consider the sum

96
 Some problems of Diophantine Approximation. 223

A,&,--
& = 2 rp(n) cos (n%x),
QV

where A, is an integer large compared with qy but small compared with ply+l/qy.
We shall suppose A, so chosen that

Av%’ 3

(2. 201) !p 2 !Pw-+-

QY

(2. 202)

and we shall show that, in these circumstances, Is,l tends to infinity with * Y,
and hence that the series

2 cp (n) cos (n%x)

cannot converge.
We may consider, instead of X,, the sum

s ‘1’ -
S, i’i-,‘(, { cos (ns 7-rx) - cos (nZ a py / c&.

Now

where a’*+1 is the complete quotient correspdnding to the partial quotient av+l,
and &,+I = a’,+1 qv + qv-1; and from this it follows that IS, - S,I is less than a
constant multiple of

and so of

Thus Cry- S:, -0 as v-00, in virtue of (2. 202).

224 n G. H, Hardy and J. E. Littlemood.

We may write X’, in tile form

If in this sum we replace rp(r qy + s) by q~(rqyj, the error introduced is not

greater than

Thus, with an error not greater than qvy(q9,), and a fartiori not greater than
q,cp (I), we can replace SI1, by

Now

Y’(plzt)+ y’(2qw) -I- +a. + (r{(AY-~)qy} >L

and so

Hence
II -- 3 A,!7,,
IS I
- 2 2 $lqn)-+
qv&&, qv *v q 1’

which tends to infinity with Y, in virtue o.f (2. WI). Hence S$, and so S,,,
‘tends to infinity with VU;which proves the theorem.
In particular it, is possible fo find irrati,onal values of x for which the series

z co9 (nkx)
n
? 2
COB(dnx)
n log n ’ ’’’
l

are not convergent.

 Some problems of Kophalrtine Approximation. 225

2. 2x. We shall find it convenient at this stage to introduce a new nota-:

tion. We define the equation

!=WL

where rp is a positive functiori of a variable, which may be integral or con-

tinuous but which tends to a limit, as meaning that, there exists 8 constant H
md a sequence of values of the variable, themselves tendingL to the limit in
question, such that

for each of these values. In other words, f = L! (rp) is the negation of f = o(y).
In the notation of Messrs WHITEHEAD and RUSSELI; we should write

J=fi(q~).= .m(f=o(rp)). Df.

2. zz. We shall now prove the following theorems.

Theorem 2. 22. If x is irrational, then

Theorem 2. 221. If q~ is any posi2ive fmction of n, which ten& to zwo as

n - 00, then it is possible to find irrutionals x such that

These theorems show that the

sfg= o(fiL
f
est.ablished by Theorem 2. 141 for a. particular class of values of X, ca.nnot-pos-
sibly be replaced by any better equation; and that the eqna.tion

% = 0 (n)

of Theorem 2. 14 is t.he best that is true of all irrationals. We shall deduce

these theorems from certain results concerning the elliptic modular functions.
2. 23. We write

99
 226 G. H. Hardy and J. E. Littewood.
1 a
3,(0, T)=z& ( n-2 > l
1

a3(0,~)=I+2&',

a,(~, z.)= I + 2 r((-~)nqt

We suppose that p, /qn is a convergent to

and write

We shall consider a linear transformation

6 tdz
T
=+’
where

In either case ad - bc = I$ = I l

Finally, if a’,+1 is the complete quotient corresponding to a,+l, we write

Q’n+l= da+1 qn + qn-1,

and we take
Y = I: / (qn Q’n+l) *
When
p,+l is even, p,, is odd,

we shall say that the convergents ~~-11 qn+ pn/‘qn form a system of type

100
 Some problems of Diophantino Approximation. 227

There are six possible types of system, viz.

which we number

LO, P, 3O, 4O, 50r 6O.

The following remark is of fundamentsl importance fur our present purpose.

In uny continued fraction whatever, one or other of the systems LQ, P, 50, 6~ must
occw infinitely often. This appears from the fact that the second column in
cases 30 and 40 is 0, 0, and that all COWS in which the first column is 0, 0
fall under 10, z”, 50, or 6O.
2. 24. In ccws x0, z”, 50, or 60 we have

where ctl is &n 8-th root of unity, and $ stands for one or other of a, and *9,.*
Now

Also, if Q = eniT; ive have

= e-d.,
IQ1
where

Y = Q1fi+l > I
= (IjQI*b+$ -t q; y” 2 471 2 ’
Hence

IQ!
<e+ <1/(4-8)<.21,

21&l + 21&l* + -** < 2 (m21) + 2 (* ZI)4 + .’ *

<I-,
2
~.- -
* T. am-l M., Vol. 2, p. 262 (Table XLII).

101
 228 G. H. Hardy and J. E. Littlewood.

From this follows at once

Theorem 2. 24. If q = r@, where x is irrational, then ’

00
I + 2 zI tr2=
US r-1.

From this we can deduce Theorem 2. 22 as a corollary. For if we haid

8,=0(G),
the series

would satisfy .the condition

and 80 we should have

- (r-r) hop

an equation which Theorem 2. 24 shows to be untrue.

Again, let ‘p( I / y) be any function which tends
I to zero with ?/* We have

We choose a value *of x such that, for an infinit-y of values of n corresponding

to one of the favourable’ cases 10, 20, 50, 6O, we have

W-G ‘P(Qndn+d;

102
 Some problems of Diophantille dpproximat.iw 229

this may certainly be secured by supposing that a,+1 is *sufficiently large. We

have then

l~h 4l>~WYrp(dYh
From thi8 we deduce .
Theorem 2. 241. Given any fun&m rp which tends tu zero, it ia pssible to
find irrutional dues of x such that

when q=renL a& F-L.

From Ma theorem Theorem 2. 221 follows a8 a corollary just as Theorem
2. 22 followed from Theorem 2. 24.
2. ~5~ It ia interesting to consider a little mc)rc closely tile case in lqhich
5 ia an irrational for which a, = 0 (I).
Let UH, in&ad of considering only the special value I: / (qn Qln+l) of y,
consider the range R,, defined by

-f-iy<+
n
Qr’+1 -Pn
or
7 <y< = 3
qn &a+1 - - “7qn qL+1

where 7 = q,, / q’,+1. It is clear that, for different valueB of n, these ranges
cover up the whole range of variation of 21. If now y = 5 /(q- &+I), so that
&(r/q, we have

The least valuea of A, correspond to 5~= q, I / 7; and then

q/i+1 > 5
;1L=
q: + q’A+1 2 l

$uppose first that n corresponds to a system of one of ‘the type8 lo, 2O, 5O, 6O.
Then the argument of 2. 24 shows that the absolute value of 3(o, T) lies

between 5 and 3. If 011 the other hand n corresponds to a system of type so

2 2
Or 4Y we have

103
 NOW

a,(% T) =2g: (I + Q2 + Q” + am-),

3 5
and the absolute value of tIhe second factor lies between - and :-.. On the
4 4
other hand A lies between &+I / (& -+ q’A+l) and &+I / z Q~, and a fortio~i between
I
and 5 (K + I), where X is the greatest value of a partial quotient. Hence
z 2
in this case also 1.3(0, T)I lies between fixed positive limits.
Thus, as the ranges R, fill up the whole range of variation of y, we can
determine two constants H,, H, so that

and it is easy to see that the second factor under the radical lies between fixed
positive limits. Hence we obtain
Theorem 2. 25. If q = renix, the partial quotients to x being limited, and
Y - I, then

2. 26. In the preceding discussion, the argument which showed that

A > i was independent of any hypothesis as to the continued fraction. Hence

we have in any
H2

a0 qn - 00. Hence we obtain

Theorem 2. 26. For any irrutional vu& of x, toe have

1 The formula j’X ‘p implies that IfI / y 1ies between fixed positive limits: see HARDY,
Orders of Ikfinity, pp. 2, 5.

104
 Some problems of Diophantine Approximation. 231

This result may of course also be proved as a corollary of Theorem 2. 14, by

reasoning analogous to that used in 2. 24. But the direct proof is none the
less interesting,
2. 27. The argument used in 2. 24, in deducing Theorem 2. 22 from Theo-
rem 2. 24, may be adapted so as to prove an interesting generalisation of the
former theorem. Let us write, as before

and suppose that k /nk is one of CES~RO’S means associated with the series
k:! Sn

k+;
(2. 271) Skn-- n >.

For if this were not so, we should have

From (2. 271) it follotvs that the series 2 un cannot become summable (CIc) on
1
the introduction of a convergence factor n-4 2 And from this we deduce
Theorem 2. 27. The series

z n --n @&Tix a<z

( -2 )

cannot be convergent, or summable by any of CESARO’S means, for any irrational a+.
We need hardly remark that the same is true of

(n-12 >%ix
II nOae
* , 2 (- I)n ?.&-a @nix*

On the other hand, if a > 5, all these series converge prespue partou~ (2. II, 2.16).
2

1 HARDY and LITTTZWOOD, Proc. Zo&. Math. Sot., Vol. II, p 435.

105
 232 G. H. Hardy and J. E. Littlewood.

2. 3. - An application to the theory of trigonometrical series.1

2. 30. The problem of finding

a trigonometrical series whose coefficients
tend to zero, and which converges,
if ever, only for a set of values of the ar-
gument of measure 0, was first formulated by FATOU~ and first solved by
LUSIN. s The results of the earlier part of this paper have led us to a solution
of FAWJ'S problem which seems to us to have considerable advantages over
LUSTN’S. *
We can, in fact, prove the following theorem, which is an extension of
Theorem 2. 27.
Theorem 2. 30. The series

z n-” cos (n2 n x), 2 rra sin (VPfix),

where o < cc< f , are never convergent, or summable by any of CESARO’Smwn,c, for

any irrational value of x?

Considered simply as solutions of FATOTT’B pmblem, these series have, as
against LUSIN’S, two advantages. In the first place, thev are series of a simple,
natural, and elegant aanalytical form. In the second plaie, t.he problem of con-
vergence is solved completely; there is no exceptional set of values of x for
which doubt remains!
2. 31. We proceed to the proof of Theorem 2. 30. This theorem is a
corollary of

t An abstract of the contents of this part of the paper flppeared, under the title ~Tri-
gonometrical Series which Converge Nowhere or Almost Xowherea, in the &co&s of Proceed
ings of the London ilfuth. Sot. for I 3 Febr. I 913 l

p Acta Mathematics, Vol. 30, p. 398.

8 Rendliconti di P&rm, Vol. 32, pa386.
4 The cosine series converges when x is a rational of the form (21-t I) / (2~ + I) or
2A /(4~ + 3), the sine series when =x:in a rational of the form (211+ I) / (2~ + I) or 2A/ (4~ + I)
(see 2. OI), In the abstract referred to above this part of t’he result (which is of course trivial)
was. stated incorrectly.
5 It is only since this paper waswritten that we have become aware of a different solution
given by H. STEINHAUS (Cmptes Rendus de la Soci&t! Scienti,fique de Vumotie, 1912,p* 223).
STEINHAWS also solvea the problem. of convergence for hia series completely; they converge, in
fact, for no values of X. Thus in this respect our examples have no advantage over his; the
advantage; if anywhere, is OTI his side. In respec’t of simplicity etc. our examples have the
advantage over his aa much as over LusIE;‘s.

106
 Some problems of Diophantine Approximation. 233

Theorem 2. 31. If q =reni*, where x is irrational, then, as r - I, both the

real and the imuginary parts of

f (q) =1+2&”
1

Tn fact, when once this theorem has been established, Theorem 2. 30 follows
from it in the same way as Theorem 2. 22 followed from Theorem 2. 24. And
the proof of Theorem 2. 31 is in principle the same as that of Theorem 2. 24,
though naturally more complicated.
Our notation will be the same as in 2. ~3~ We shall prove first that, ilrt
cases P, z”, so, and 6O, we ?imve

0i Ia&, r>i > K$,

(ii)

for all integral values of wt, .K and 6 being positive constants, provided either

(CI1 an+1 > I

UT
(PI an+1 = I:, an+2 -= I.

We shall express this shortly by saying th& IO, 20, 50, 60 are favow&le cases,
except possibly when
an+1 = 1, &L&+2> 1;

a ‘favourable case’ being one in which we can prove the inequalities

We have

(2. 312) %b, 4= No, T)-

A-T%
If a n+l> L
- nQh12~n < e-z < I,
IQ1I_ e 23?
and if a,+1 = I:, an+2== I,,

107
 234 G. 11. Hardy and J. E, Littlewood.
1
I
q nfl

In either case

and so

(2. 313) la(0, T)J& lam9(0, T)I<arctsn~+~.

4
Again
da + bz- It” (qn + i) I qk+1,

--
(2. 314) la + bq-k=z : v&>K$,

(2. 3%) am (a + br)-2 s--~~~~Yc (mod. :7fi) l

\ 1

From (2. 312), (2. 313), (2. 3rq), and (z. 315) it follows, first that the modulus

of 9, (0, z) is greater than a constant multiple of y-f (as has been shown already
under 2. 24), and secondly that

(2. 3x6) (mod.;Y&j,

I
where -7~ denotes a number whose absolute value is less than J- n’. Hence
I
I2 1
XL2
am 8, (o, z) must differ by at least

m ST
-m-z- Y;c
8 I2 24

from any multiple of f 76; and so the cases which we are considering are a.11
2
favourable,
2. 32. We shall now prove that, as n - a0 , fccvourable cases must recur in-
finite1 y often. This will complete the proof of Theorem 2. 31.
We represent the state of affairs, as regards the oddness or evenness of
pn and qn9 in a way which will be made most clear by an example. If every

108
 Some problems of Diophantine Approximation. 235

pn is odd, and qn is alternately odd and even, we represent the continued frac-
tion diagrammatically in the form

00 0 0 o....
0 E 0 E O....
- and so in other cases.
Suppose first t.hat 0 0 occurs infinitely often above. Then one or other
of the systems

must occur infinitely often. If the first, which is system P, either favourable
cases recur infinitely often, or the ensuing partial quotient is always I. We
represent this state of affairs by the symbol

00
OE' I

rn this case our diagram continues

OOE
0 E I 0;

and as is case SO, either favourable cases recur continually, or the next

quotient is also I:, so that we have

OOE
OEO'I I

But then the first four letters represent a system of type z” followed by two

quotients a,,+1 = I, an+2 - I; and this ie a favourable case.

continually, favourable cases recur continually.

We consider next the result of supposing that recurs continua.lly.

This is case 40. If the diagram continues with an 0 above, it must continue
in the form
000
EOE

109
 236 G. H. Hardy and J. E. Littlewood.

and then we can repeat our previous argument. The only alternative is that
it should continue.
OOE
EQO

- and as the last four letters form a system of type GO, the next quotient
must (in the unfavourable case.) be I. Hence we obta.in

OOEO
EOOE' I

The next quotient must also be r:; and so the system -of type 6O gava in reality
a favourable case.
We have thus proved that<, whenever the succession -0 0 recurs continually
above, we obtain an infinity of favourable cases. It only remains to consider
the hypotbesis that pt is altertlately odd and even.
If we have 0 E above, we have one or other of the systems
[E Z)* (E E);
systems 50 and 60. Thus we have a favourable case unless an+1 = x . If the
system is of type 50, we are led to

OEO
OOEI I

- so that the system is favourable. On the other hand, if it is of type 6O,

we are led to .
OEO
E 0 I0.

As the next numerator is even, the next denominator is odd. Hence the next
and we have seen that this case must be favourable. ’
We have now examined all possible hypotheses, and found that they all
involve the continual recurrence of favourable cases. Thus Theorem 2. 31 is
established.
z. 33. -From this theorem we can, as wa8 explained in 2. 31, deduce
Theorem 2. 30 as a corollary. The latter theorem ha.s an intVeresting consequence
which we llave not seen stated explicitly.

110
 Sow problems of Diophantine Approximation.. 237
The series

-2nea (30s (n% X) , 2 n-a sin (nB z x),

I:
where cu< 2, me not bURIEB’s series.

For if they were they would be summable (C I) almost everywhere, by a

theorem of LEBESWE. 1 It follows that trigonometrical series exist, such that

is convergent, for every positive 6,2 which are not, FOWRIER’S series. This
is ,of interest for the following reason. If 2 (aA + bL) is convergent, the series
is the FOURIER’S series of a function whose square ia summable.? Further if
p is any odd integer, and

is convergent, then the function has its (I: + p)-th power sum&bleW4 It
would be natural to suppose that the RIESZ-FISCHEB T&wP~ might ,be capable
of extension in the oppoaite direction. One might expect. for example, to find
that a series for which

is convergent must be the FOURIER’S series of a function whose

I + 5 -th
( 231
power is summable. That thk is not true has been shown by YOUNG, by
means of the series

l Math. Amah, Vol. 61, p. 251. See also LeGon sur Zes at%8 trigonome’triques, p. 94
where however the proof is inaccurate. A FOURIER'IJ series i8 in fact Bumtnable @a), for any
positive 3, almost everywhere (HARDY, Proc. Lo& &ktI~ Sot, Vol. 12 p. 365). That our series
are not FOURXER'~ series when u < f can in Iact be inferred merely from their nonmconver-
2
gence, since to replace ‘ybca by r&-B, where p is any number greater than a, would, if they were
FOURIER'S series, render them convergent almost everywhere (YOUNG, Comptes Rendus, 2~ Dece 1912).
t Or even for which
IanI’ + lb7Ll2
2 (log n)l+b
is cdnvergent.
* This is the 'RIESZ-B'ISCHIER Theorem’.
' 'w. H. bUNG, BOG. hd. &fiZth. h., vol. 12, p. 71:

111
 238 G. EL Hardy and J. E. Littlewood.

~0s nx: + sin nx

z
d (log TX,:

- here p = 3. Our examples however show a good deal more, viz. that as
soon as the z which occurs in the RIESZ-FISCHER Theorem is replaced by any
higher index, the series ceases to be necessarily a, FOURIER'S series at all.
2. 34. There are other classesI of series the theory of which resembles in
many respects that of the series studied in this paper. One such class comprises
such series as
2 cosec nrtx, 2 k dn cosec n7rx

and the corresponding series in which the cosecant is replaced by a cotangent:

these series are limiting forms of q-series such as

Another class comprises the series

and the corresponding series in which (nx) - I is replaced by n& We have

2
proved 3 considerable number of theorems, re blating to these various series, of
which ‘tv~ hope to give a systematic account on some f u t4re occasion.

112
 CORRECTIONS

p. 211 (2.132) md 23. 226 (fourth dbplayed formula). &ad:

1 1
X= - -*a*.
al-l- a,+
p. 231. The fourth root in the first displayed formula should be a square root.

COMMENTS

The basic result of the paper, Theorem 2.128, was later named the approximate functimal
equation. of the theta-fuwtion. It is not, however, of the same character as the approximate
functional equation of the Riemann zeta-function, discovered later by Hardy and Littlewood
(1923, 5). The former is an approximate transformation of one finite sum into another, the
latter is an approximate expression for &+it), when 0 < 0 < I, as the sum of two finite sums.
The original proof of this result, given in the present paper, is unnecessarily elaborate. Simpler
proofs were given later by Hardy and Littlewood themselves (1925, 4), by Mordell (J. London
Math. Sot. 1 (1926), 6%72), and by Wilton (ibid, 2 (X927), 177-80). Model1 uses contour integra-
tion round a parallelogram, but with the vertical sides replaced by sides inclined at an angle
h to the real axis. Wilton uses Poisson’s summation formula

in much the same way as Dirichlet used it in his evaluation of Gauss’s sum. Wilton also gives
a numerical estimate for the constant in the error term.
8 2.10. The transformation formula for a sum of the form
&,iX
2
(see 2.102) expresses this sum in terms of a similar sum with h replaced by K, where
1/k+1/JC = 1. If Ic is a positive integer greater than 2, the transformation does not give any
useful information about the magnitude of the original sum. A very general transformation
formula, applicable to a wide range of sums of the type

was given by van der Corput (Math. Amzakn, 87 (1922), 66-77, and 90 (1923), 1-18). For
a somewhat simpler treatment, see Wilton, J. Lmadon Math. Sot. 9 (1934), 194-201 and 247-54,
For an account of the part played by such formulae in some problems of analytic number theory,
see Rankin, Quart. J. of Math. (2) 6, (1955), 147-53.
For references to later work on sums of the form 2 ellanix, see Koksma, ch, 9, 5 3.
$ 2,120. The exact formula (2.1202) is attributed (following Lindelijf) to Genocchi and Schaar.
But it is an easy deduction from the value of Gauss’s sum, and may well have been known to
Gauss or Dirichlet.
Q 2.13, In the estimate (2.138), the last term can be omitted, by virtue of (2.136).
It may be helpful to restate (2.138) and (2.1381) in the usual symbolism of continued frac-
tions. If
Pn = - 1 -*.*-l 1
an a,+ a,+ an’
the choice of v made in (2.136) is effectively equivalent to
Yv < n < Qv+f (1)
(it is actually equivalent to 4: < n < qi+l, where yi is the ‘complete quotient ‘, as defined in

113
 8 2.22) and instead of (2.138) and (2.1381) we get
8, = O(min(nq,-f, af+& (2)
This renders the deduction of Theorems 2.14, 2.141, 2.142 more immediate.
5 2.14. In Theorem 2.143, the constant log 2 could be replaced, in both occurrences, by
log 9( 3 + &), The inequality xx,... x, < 2-i’,
used by Hardy and Littlewood, could be replaced by a stronger inequality with I/W,, on the
right, where u, is the vth Fibonacci number. This leads to the above improvement.
Theorem 2.143 has the defect that it gives no information if p is large, and in particular it
leaves open the question whether any postulated explicit upper bound for a, as a faction of n
implies some explicit upper bound for s, which is better than o(n). This is in fact true, For
suppose
an = Wxpf (NY
where f (n) is any increasing function. Then
Qn = o(eXpc~ p(n)),
where C1 is a constant and F(n) = f( 1) 4 .. +;f(rz). It follows from ( 1) above that
l

v+l > Q(C&gn),

where G is the function inverse to p. Hence, by (2) and the fact that qV > exp(C,v) always,
we have 8, = OhGV = O(nexp(-Q)) = O(nexp( -C5 G(C210gn))).
This is the desired explicit upper bound.
Further results of the type of this section were given in 1922, 5.
The dependence of 8,(x, 8) on 0 was investigated by Behnke, who proved the curious result
that for any irrational x there is some 6 for which
sn(x,O) = O(d).
See Koksma, p. 111.
§ 2.19. Xn connexion with this application to the theory of functions, see 1.916, 3 and comments,
9 2.3. Again see 1916, 3 and comments.
.

114
 SOME PROBLEMS OF DIOPHANTINE APPRoXIMAT ION: no
A REMARKABLE TRIGONOMETRICAL SERIES

By G. H. Hardy and J. E. Littlewood

TRINITY COLLEGE, CAMBRDGE, ENGLAND
Receioed by the Academy, Augurt 7,Wtb

1. The title of this note is perhaps not very appropriate: we retain

it because the &tents of tde note form a natural sequel to those of
three papers which we have published under same title elsewhere,’
and in particular those of our second paper in the Ada Mathematics.
We there discussed in detail the series
eani* + Z@rirr
I (1 . 1)
c
and other similar series associated with the elliptic Theta-functions,
and used our” results to elucidate a variety of difficult points in the theory

1916,3 (with J. E. Littlewood) Proceedinga ofthe National Academy ofSciwww,2,68S6. 115

 584 MATHEli4ATK.S: HARDY AND LITTLl?WOOD

of Taylor’s series and trigonometrical seriks. We have since discovered

t*hat even simpler and more elegant illustrations may be derived from
the series
mrin log n + 2fhrin
c e (1 2) l

This series behaves, for different values of the parameters CYand 3,

far more regularly than does the series (1 A). To put the matter roughly,
the beha.viour of the series does not, in its most essential features, depend
upon the arithmetic nature oj &.
2. 0ur fundamental formula is
w 01
UPnpn 1 2nin -p-?g O3 C-Y)”
=- -
c log -V
a -W
r P+------
log a >
Y 2 o n! (up+n - 11 (2 1)
.
0 l

Here a > 1, p is real, and R (y) > 0. The formula becomes illusory ’
when p is zero or a negative integer, but the alterations required are of
a trivial character. The formula is easily proved by means of Cauchy’s
Theorem: similar formulae were proved by one of us in a paper pub-
lished in 1907?
We now write y = Q + it, where t > 0, suppose that 0 -3 0, and ap-
proximate to the series of Gamma-functions by means of Stirling’s
Theorem. We thus obtain
Ht -P
e-'"'f(z) = F(a)+&), (2 2) l

log a
where
w
f(z) i C nP-t eain log n f; (2.21)
I

cy - 2T H = (log(2nP1
a)P+4, 2 = re ;B, r = e-au/l , 0 = a log : , (2.22)
= log a’ 0
so that r--, 1 when +Q 0;'

F(a) = 2 ume-cu-tsan; (2.23)

0
and + (a) is of one or other of the forms
1
A + Q (l), 0 log - , 0 (0ff),
( P)
according as ~24, p= 3, or p>+*
3. It is known3 that, if. p >O,

F (u) = 0 (u-~), F (0) = n’(~-)~ (3 1)

l

116
 MATHEMATICS: HARDY ANb UTTLEWOOD 585

when Q -3 0, the second of these formulae meanine that F (e) is rtot

of the form o (C), and the two together that
O<h=limBP(B)< =* (3.2)
These relations all hold uniformly in f. It follows that, if p > 0 and
r = ] x 1--) 1, the jwzction f (2) is exactly oj the order (1 - r)-‘, a& this
unijurmly in 0. Incidentally, of course, it follows that every point of
the unit circle is a singular point: but this is known already.6
The series furnishes an example in which the orders in the unit circle
of the functions j (z) = IZ a,~” and g (z) = 2 ] a, 1zH differ by exactly 3,
the maximum possible.6
When p = 0,f (z) is bounded, but does not tend to a limit when z
approaches any point of the unit circle along a radius vector. We know
of no other example of a function possessing this property. When p < 0,
f (2) is continuous for [ 2 [ S 1.
4. Let
#
s, = kp-) &k log k + Z$dk .
c1 Y (4 . 1)

and suppose first that p > 0. Then it is easy to deduce from the results
of 83 that s, is of the form Q (nP) when 12+ 00. The correspondirig
‘0’ result lies a little deeper: all that can be proved in this manner is7
that s,?=o (nPlogn). But a direct investigation, modelled on that of
the early part of our second paper in the Acta Mathsmatica, shows that
the factor log vzmay be omitted. It shduld be observed that an essential
step in our argument depends on an important lemma due to Landau, 8
according to which
X

IS 1
x7 ,iz 1% (77X) dx < 23 X7++
(4 12)

for X21,. +O,q > 0. We thus tid that s, is, j& every positive v&e
of TV, exactly of the order np, and this wiformly irt 8. The series

(4 3)
l

is never convergent, or summable by any of Cedds mesas.

When p = 0, s, is bounded, but the series is never convergent or
summable. When p C 0 it is convergent; and uniformly in 8.
5. For further applications it is necessary to consider the real and
imaginary parts of our function and series separately, and this is most
easily effected by introducing some restriction as to the value of CY.
Suppose that a is an integer, not of the form 4 k + 1. Thus we may take
a I 2, a = 2~ /log 2. Then the results of 553-4 hold for the real and
 586 MATHEMATICS: HARDY AND LJT7ZEWOOD

imaginary parts of the function or the series, In particular ike series

- c d- cus (an lug n + 2em) ‘(p 20) (5 . 1)

is never convergent or summable for arty value of 8, and is uccurdin~ly not
a Fourier’s series. We thus obtain a solution of what, in our former
paper, we call Fatou’sg problem which combines all the advantages
of those given previously by ,Lusiqg Steinhaus,Q and ourselves.
We can also obtain in this manner exceedingly elegant examples of
continuous nun-differentiable functions. Thus the fwzction

. r(e)=.c sin (092 log fi + 20~2)

?P
UC B < 8 (5.2)
does not possess a finite diflerentiul coefici’ent for any due of 0,

f G. H. Hardy and J. E. Littlewood, Some probkms of Uophantine approximation:

ti) Proc. Fifth Itt, Cmgms Math., Cambridge, 1, 223-229 (1912); (ii) Acta Math., 37, 15%
190 (1914); (iii) IbX, 193-238.
2 G. H. Hardy, On certain oscillating series, Qmwtmty J. Math., 38,269-288 (1907).
8 G, IT. Hardy, Weierstrass’s non-differentiable function, Trans. Amer. Math. Sm., 17,
301-325, (1916).
4 1.c. supa’ (1) (i;i), p. 225.
6 G. N. Watson, The singularities of functions defined by Taylor’s series, Quairmty J.
Math., 42, 41-53 (1911).
6 G, H, Hardy: (i) A theorem concerning Taylor’s series, ibid., 44, 14’1-160 (1913); (ii)
Note iti addition to a theorem on Taylor’s series, IM., 45, 77-84 (1914).
7 Cf. E. Landau, Abschgtzung der Koeffizientensumme einer Poteqzreihe: (i) Arch, Math.
Physik, ser, 3,21,42-50 (1913); (ii) Ibid., 250-255; (iii) Ibid., 24,250-260 (1915).
* E. Landau, Uber die An&l der Gitterpunkte in gewissen Bereichen, GWlager IVach-
richten, 687-771 (p. 7071, (1912).
9 For references see pm232 of our paper (1) (iii).

CORRECTION
The statements at the end of 8 3 and of $4 about the case p = 0 are incorrect:
see the end of 1916, 9.
COMMF,NTS
The interest of this paper is almost entirely analytical, and (as the authors
remark in the first sentence) its inclusion under the general title ‘Some problems
of Diophantine approximation’ is not entirely appropriate.
The substance of Q$ 1-3 is given again, in somewhat more detail, in I&He-
wood’s Lectures on the theory of functions (Oxford, 1944), pp. 99-102.
The substance of $5 4-5, with further results, is given in ch. 5 of Zygmundk
Trigonometric series (2nd ed., Cambridge, 1959). See also the references on
p, 379 of Zygmund.

118
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION:
THE SERIES Z e (A,> AND THE DISTRIBUTION OF
THE POINTS (ha)
By G. H. Hardy and J, E. Littlewood
TRINITY COLLEGE,CAMBRIDCE,ENGLAND
Gmmunicd by E.H. Moore,Deccmbrt5, 1916

1. In our previoui writings on the subject of Diophantine approx-

imation, which we refer to in a short note published in the October
number of these PROCEEDINGS,~ we alluded in several places to a series
of further results which, we hoped, were to form the material for a-third
memoir in the Acta Mathematics. The prosecution of this work was
delayed, in the first instance, by our occupation on a long memoir on
the theory of the Riemann Zeta-function, now in type and shortly to
appear there, and subsequently by other causes; and there is; under
present conditions, little hope of its completion in the immediate future-
The subject has since been reopened by the appearance of work by other
writers,2 and in particular of a very beautiful memoir by Weyl- in the
latest number of the Mathematische Annalen. This paper contains
allusions to our unpublished work: and it seems desirable that we
should make some more definite statement than has appeared hitherto
of our results and the relations in which they stand to Weyl’s.
The main problems which we considered were three.
2. (a) The first problem was that of proving that, if e(x) = e’““’
and
L = afiR + a1 nk-’ + . . . + ak
is a polynomial irt B with at least one irrationul coe$cient, theti

s, = 2e(hk) = o(n).
1
We may plainly suppose that every CYhas been reduced to its residue
to modulus unity: and there is no substantial loss of generality in sup
posing the first coefficient irrational.
This theorem we enunciated first, in the special case in which A,, =
mk, in our communication to the Cambridge Congress, characterising
the proof as ‘intricate.’ In our second memoir in the Acta we discussed
in detail the case K = 2, using a transcendental method which leads to
a whole series of more precise results; and we promised a proof of the
more general theorem in the third memoir of the series. Weyl’s memoir
contains a complete statement and proof, both quite independent of
ours, of the theorem in its most general form.

1916,9 (with J. E. Littlewood) Proceediqp of the National Academy of Sciencea, 3,8&M. 119
 MATHEMATICS: HARDY AND LITTLEWOOD 65

The limitation on the form of X,, which appears in the theorem as we

stated it, was introduced merely for the sake of compactness of expres-
sion and does not correspond to any real simplification of the problem.
Our argument indeed depends upon an induction which compels us to
consider the problem generally. The most comprehensive result which
appears in our analysis is as follows: given any pus&De numbers E md q
we cm determine v (E), N (E, TJ), and a system of intervals j, includirtg all
ratiunals whuse denominators are less than v, und of total length less thun
v, so that ] s, 1 < en fur n > N, all values of QIexterior to the inter&s j,
and all values of al, CY~,. . . . , ak. From this result it follows at
once that sn = o (n) for any particular irrational a, and uniformly in
al, a2, . . . I , ak*
Weyl’s proof and ours are widely daerent, and each, we hope, may
prove to have an interest of its own. The same is true of the deduc-
tion of the formula r (1 + it) = u (log t), made by Weyl as well as by
ourselves.
3. (b) The second principal problem was, to use Weyl’s phraseology,
that of the ‘uniform distribution’ (Gleichverteilung) of the points (A,)
where (x) is the residue of x to modulus unity. Suppose that ntl is the
number of the first n such points which fall within an interval j of
length 6, Then the points are said to be unz’furmly distributed if tij N Snfor
every such interval j. It is plain that a corresponding definition may
be given of uniform distribution of an enumerable sequence of points
in space of any number of dimensions.
That the points (A,) are uniformly distributed when K = I and Q! is
irrational was proved independently by 2?Sierpinski, and Weyl
in 1909-10. The general result (with ‘the same unessential limitation
as to the form of X,) was stated by us in our first paper in the AC&Z.
Our proof, which has never been published, proceeded on the same lines
as that of the thoerem of 82. But Weyl has now established a ‘prin-
ciple’ which renders such a proof entirely unnecessary, and which has
led him to results in this direction far more comprehensive than any of
ours. This ‘principle’ is expressed by the theorem: i;f

%e (mx,) = 0 (n)

jur every positive integral value uj m, then the points (A,) are uniformly
distribtited Gz (0, 1). The proof depends on a simple but ingenious use
of the theory of approximation to arbitrary functions by finite trigo-
nometrical polynomials; and there is a straightforward generalisation
to space of any number of dimensions.

120
 86 MATHEMATKS: HARDY AND LITTLEWUUD

Weyl’s ‘principle’ enables him to deduce, with singular ease and ele-
gance, theorems of ‘uniform distribution’ from theorems of the char-
acter of that of $2, and to generalise them immediately to multidimen-
sional space. It enables him to prove, for example, that the points whose
courdinates are
(np a*) (p = 1, 2, . . . , k ; q = 1, 2, . l w , 1; n = 1, 3, 3, . l l)
where al, 012, . . , al is any set of linearly independent
l l irratioflals,
are urtifu-rmly distributed in the ‘unit cube’ of kl dimensions, All that we
had been able to prove was that the points were everywhere dense in
the cube.
4. (c) Corresponding questions arise in connection with an arbitrary
increasing sequence Xx, Xz, X3, . l Are the points (X, CY),for
.

example, uniformly distributed? The answers to such questions in

general involve an unspecified exceptional set of values of ar of measure
zero, instead of (as when X, = nk) a specified set such as the rationals;
they are, in other words, only ‘almost always’ true.
In our first paper in the Acta we proved quite generally that the set
(A, a) is almost always everywhere dense. The corresponding theorem
of ‘uniform distribution’ we discussed only in one especially interesting
particular case, that in which A, = an, whex’e a is an integer. The
theorem is in this case substantially equivalent to results obtained by
l Borel,4 from the standpoint of the theory of probabilities, and by Faber,6
as a corollary of Lebesgue’s theorem that a rectifiable CUIV~ a tan-b
b. t almos ver Our an-ever contains the first
direct and general discussion of the problem, and leads to results nota-
bly more precise than that of mere uniformity of distribution. These
results were afterwards made the subject of important generalisations
by Fowler,2 whose investigations covers all cases in which X, increases
with tolerable regularity and as fast as an exponential of the type 8’.
Weyl’s ‘principle’ enables him to reduce this problem to a study of .the
series me (X, a)? and leads him to the following theorem, so far the most
general of its kind. If c > 0, 6 > 0, and A, increases by at least c
whenever 1z increases from lz by as much as h (log h)-I-“, then

s, = 2e(X,a) = o(n), (1)

and the points (x, CW)
are uniformly distributed, for almost all values of LY.
In our second,paper in the Acta we stated that the equation could,
in very many cases, be replaced by the much more precise equation
s, = 0 (n’f3 (2)

121
 MATHEMATICS: HA/WY AND LITTLEWOOD 87

for every positive E. The publication of Weyl’s work had led us to a

reexamination of this question and to the following theorems.
A. Ij (i) Xn+[+] -X,+=,
0ii 1 a1 12 + 1 aa 12 + l l l + 1 a, ] 2 = 0 (d+q,

for almost all CW’Sand every positive e.

B. Ij (i) is rephced by (i’) L+[d+q - L+ 00, where 0 < p < 1, thett

(iii) may be replaced by

To these two theorems Weyl’s forms a completing third. It should be

observed that (i) is certainly satisfied if &I - X, 2 c > 0, and in
particular if A, is always integral, and (ii) if a, = O(PZ’), and in partic-
ular if a, =l.
If A, is an integer, and we separate the real and imaginary parts in
the equation (iii), we obtain a theorem concerning. a particular system
of normal orthogonal functions for the interval (0, l), viz., the functions
4 2 cus 2&X, J 2 sin 2Ax. Our argument is then directly exten-
sible to a general orthogonal system, and we are led to a new and inter-
esting proof of Hobson’9 theorem that if &(x> is asy normalorthogonal
system, and z n’ 1cn I 2 is convergent for some positive 6, thevt IZC~+~(x) is
convergent almost everywhere.
JVeyl’s hypothesis concerning 1, asserts, roughly, that the increase
of X, is appreciably more rapid than that of (log n)? It is easy to see
that this hwothesis cannot be capable of much wider generalisation.
For, when A, = log 1z, sn is definitely of order S. It seems probable, too,
that the index + (1 + p) of Theorem B is the correct one.
5. We conclude by correcting an error in our recent note. The re-
sults concerning the special case p = 0 are stated wrongly. It is not
true that, when p = 0, f(z) and sfi are bounded; all that we can assert
1
is that they are of the forms 0 log- and 0 (log M) respectively.
( 1-p’ 1
That j (z) should be bounded would contradict a general theorem of
Fatou,? in virtue of which a bounded function must tend to a limit, for
almost all values of 8, when z = re ie tends to the circle of convergence
along a radius vector, The error has no bearing on the general case.

122
68 MATHEMATICS: HARDY AND LITTLEWOOD

1 Hard?, G, H., and Littlewood, J. E., these PROCEEDINGS, 2, 1916, (583-586).

1 &wick, W. E. H., Mess. Math., Cambridge, 45, 1916, (154-160); Fowler, R. H., Lmdm,
pm. Afatk, Sm., (Ser. 2), 14, 1915, (189-207); Kakeya, S., Tdbku Sci. Rep. Im#. Unio:,
2, 1913, (33-54) and Ibid., 4, 1915, (IOS-109).
3 Weyl, H., Math. Ann., Leipzig, 77, 1916, (313-352) ; see also Gtittingcn Nuchr. Gcs.
Wiss., 1914, (234-244).
4 Borel, E., Pdermo, Red Circ. Mat., 27, 1909, (247-271); see also ‘notes to Borel, E.,
Lgms sur la thbrie des fonctions, 2d. ed,, Paris.
4 Faber, G., Math. Ant, Leipzig, 69, 1910, (372443), especially p. 400.
6 Hobson, E. W.,-London, Proc. Math. Sot., (Ser. 21, 12, 1912, (297-308). *
7 Fatou, P., Actu Math., StacklroEm, 30, 1906, (335AOO), especially p. 349.

CORRECTIONS
p. 85. In the last line but one of the second paragraph, read f( 1 +i+
In the first line of the last paragraph, read k = 1.
13. 86. In the first displayed formula, read rPolq.

COMMENTS
Q 4. It does not appear that proofs of Theorems A and B were ever publkhed. .
For some results, valid for almost all a, when the h, are integers, see Kokmna,
pp. 94-95.

123
 162

A Problem of Diophantine Approximation

BY G. H. HARDY
I have on several occasions found myself faced by the following prob-
lem, which appears to be one of considerable interest and difficulty.
Suppose that a and 8 are positive and a > 1, and that (x) denotes the
difference between x and the integer nearest to x. Then, in what
circutista’mes can it be true that
(a”e) + 0, (1)
when n -+ m,, i.e. that ame= P,+E, (2)
where pn. is an integer and E, -+ 0 ?
The general problem seems, as I said, to be one of great difficulty.
There is, however, one case in which the answer is almost immediate,
namely, that in which a is an algebraic number. We have in fact the
following theorem :
THEOREM A Suppose that a is a real algebraic number greater
l

the root of an irreducible equation

I
k,am+k,am-l+...+lc, = 0, (3)
where I&, lc, ,..., k, are integers. Then, in order that numbers 0 should
exist which satisfy (I), it is necessary and suficient that E, = I, so that
a is an algebraic integer, and that the muduli of all the roots of (31, other
than a itself, should be lessthan 1. The numbers 0 are then all rational in
the corpus of a; and tune)= O(bn), (4)
where b is the numerically largest root of (3, other than a itself.
We have

Ax = ~Pnxn+&Xn =m+4x),
say; and

(5)
where
qn = k,Pn+klPn-l+~+m+kmPn-m,
512 = kgEn+klEn-l+**g+kmEn-,,
p’s and E’S with negative suffixes being regarded as equal to 0. The

124 1919, 4, Journal of the Indian Mathematical Society, 11, 162-6.

 163
left hand side of (5) is a polynomial of degree nz- 1, so that

4n-a = 0 (n > m).

As qn is an integer, and cn + 0, we must have
4n = 0, ln = 0,
from a certain value of n onwards. Thus p(x) and E(X) are rational
functions, with the denominator
E,+k,x+-...+E,xm.
It is evident that p(x) can be expressed in the form

where p(z) is a polynomial and E,, E,,... are integers. From this it
follows, in the first place, that k, = 1, so that a is an algebraic integer.*
Now, let us denote the roots of (3) by
a, a,, a2,***, b,, b 22.‘.3
the a’s having moduli not less than 1, and the b’s moduli less than 1.
Then
P(X) =P(x,+&x+~*x+&+ (6)
-k - kX’

the A’s and B’s being rational in the corpus of a and none of them being
zero, On the other hand E(X), since its coefficients tend to zero, can
have no pole inside or on the circle Ix ] = 1, and so
B ;c
44 = e(x)+ 2 -X9 (7)

where e(x) is a polynomial. From (6) and (7) we obtain

A
Ax = p(x)+e(x)+laxf 2

so that A E e1 Ak = 0, Bk+B;, = 0.
But we have already seen that no A, can vanish. It follows that there
can be no rout of (3) of the type ak.
~+l,x+...+lm-lxm-l
* If
k,+E,x+...+k,x”
is expansible in a Taylor’s series with integral coefficients, k, must be 1.
For a proof of this well-known proposition see, e.g., P. Fatou, ‘Series
trigonometriques et series de Taylor’, Acta Mafhematicct, vol. 30, 1906,
pp. 335-400 (pm 369).

125
 The conditions stated in Theorem A, with reference to the equation
(3), are therefore necessary. Also 8 = A is rational in the corpus of a,
and E,, which is, from a certain value of n onwards, equal to 2 Bib&
is of the form O(P), where b is the numerically greatest root of (3),
other than a itself.
In order to complete the proof of the theorem, it is only necessary
to show that, if (3) satisfies the conditions stated, 8’s exist which satisfy
(1). And this is obvious; for
an-+2 b%
is integral for all values of n so that
(an) = O(bn).
Thus (1) is satisfied when 8 = 1.
2. Theorem A is a special case of a more general theorem:
THEOREM B, Sqpose that aj (j = 1,2,-, r) is an algebraic number,
greater than I in absolute value, and the root of an irreducible equation
k, ,jUF+E, ,jUp-l+m*m+kqj 1 = 0; (8)
that Pj(n) is a polynomial with integral coeficients; and that
+(n, 0) = P,(n)a~~,$-P2(n)a~e2+... i-P,(n)a~e,.
Then, in order that it should be possible to find a system of numbers el, 8,,
**a,8,, all&fferent from zero, for which
0

when n -+ GO,it is necessary and su$icient that k, j = 1 for each value of j,

so that aj is an algebraic integer, and that, am&g the complete system of
roots of all the equations (8), all save a,, a2+.., ar themselvesshuuld be in
absolute value less than 1.‘ Each tIj is then rational in the corpus of the
corresponding aj; and
(#n, 0)) = own),
where c is a constant and b is the numerically greatest among thi roots of
the equations (8), other than a,, a,,..., ar themselves.
The proof of this theorem does not differ in principle from that of
Theorem A, the role of the polynomial
ko+k,x+...+k,xm
being now played by
K(X) = fi (k,j+k~,jx+...+k~,jxrn~)8~+’
j=l ’

126
 165
where + is the degree of Pj(n). The special theorem is, from dur present,
point of view, more interesting than the general one, and I shall there-
fore confine myself to enunciating the latter.
3. With Theorem A should be associated another theorem which is
merely a special case of a theorem already proved by Borel. **
THEOREM C, If a and 8 are pifive, and a > 1, and
(anO> = O(P),
where 0 < b < 1, then a is algebraic.
Combining Theorems A and C we obtain an interesting criterion’for
the transcendentality of a, viz.
THEOREM D. If a > 1, 0 > 0, and
(&) + 0,
then a is algebraic or transcendental according as the equation
(a”e) = O(e-8n)
is or is not true for some positive value of 8.
4. It is interesting to consider in more detail the two simplest cases
of Theorem A.
In the first place, suppose m = 1, so that a is rational. If then (1)
is satisfied, u must be an integer, and anO must be an integer for suffi-
ciently large values of n. Thus the only solutions are given by
e = p/P,
where p is an integer.?
The next case is that in which m = 2, Then the equation satisfied

** E Bore1 ‘Sur une application d’un th6orhme de M. Hadamard’,

Bull& de la ’ Socidte’ Mathe’matique de France, ser. 2, vol. 18, 1894,
pp- 22-25. Borel’s theorem is really (when stated in slightly different
language) that which corresponds to B as C corresponds to A.
The criterion is a simple one: to find an application of it is quite
another matter; and I know of no example of a transcendental a which
satisfies (1) for any ‘value of 0.
t This trivial case of the theorem is of interest in connexion wifh
Weierstrass’s function. [See G. H. Hardy, ‘Weierstrass’s non-differen-
tiable function’, Transactions of the American Mathematical Society, vol.
17, 1916, pp. 301-X]
 166

by a is of the form X2 -mx-n = 0;

and it will easily be verified that, in order that this equation should have
two real roots a and b, such that a > 1, - 1 < 6 < 1, it is necessary
and sufficient that
m 2 1, l-m <n < l+m.
The simplest case is that in which m = 1, n = 1, and
U= ;(&+l), b = +(-lis+l).
The determination of the corresponding values of 0 presents no diffi-
culty. We have in fact
THEOREM E. If a is a real quadratic surd greater than I, and values
of 8 exist
which satisfy ( 1 ), then
U2 -mu-n = 0,
where m > I, 1-m < n < 1 j-m. The correyndinq values of 19are the
numbers

if m is odd, und the nambers

4 U--r
I ’ +&m2+n) 1
if m is even: here r is zero or a positive integer, und p und q halves of in-
tegers; and, when m is odd, p and q are either both integers or buth halves
of odd integers.
When a = &b+l) the simplest values of 8 are
e= 1, 8 = 4s.

128
 COMMENTS
This is the only paper on Diophantine approximation which was written by Hardy
without Littlewood’s collaboration,
In the footnote on p. 163 it is tacitly supposed that &,..., ?cm.have no common
factor. For the result quoted there, see also Pblya and Szeg6, Aufguben ulrzd
Lehmtitze au8 der AnaE@s, Section VIII, Problem 156.
The algebraic integers a (or CL)occurring in Theorem A, that is, those QI > 1
whose algebraic conjugates a’ all satisfy I$ ] < I, have since been called the Pisot-
Vijayaraghavan numbers; though it will be seen from this paper that their most
important property was discovered by Hardy.
Vijayaraghavant gave various further properties, the main one being that
Theorem A still applies if the sequence lw?? has any finite number of limit points
(mod 1) instead of just 0.
Theorem C can be regarded as a first step towards the more precise theorem of
Pisot (1946) that if a0

converges, where a, 8 are real numbers, QI > 1 and 8 # 0, then QI is necessarily

algebraic and of the type defined in Theorem A. For an account of this and other
work, and references, see ch. 8 of Cassels’s Tract. For further references, see Pisot,
Journal ftir Math. 209 (1962), 82-83.

t For references, see J. &o&m J&h. BOG. 33 (IS!%), 252-5.

129
 Some problem of Diophantine approximation: A further note on
the trigonometrical series associated with the elliptic theta-functions.
BY Prof. G. H. HARDY and Mr 3. E. LITTLEWOOD.
[Received 6 July 1921.]

1. This note contains a short addition to a memoir, with a

similar title, published in 1914 in the Acta Mathem&*. In that
memoir we considered the sums
s,2= 5,2(5,e) = C e@- +I' riX CQS(zv - 1) d3,
rsn
~~3 = s,3 (2, e) = C ev2*ix COB21~0,
vgt
s,4= ~~4(x, e) = c ( - 1)’ ev2rix COB
2&,
rp
where x and 0 are real and x irrational?. There is plainly no real
loss of generality in supposing either x or 8 to be positive and less
than unity, if it be understood that 8 may be zero.
Our main results may be stated as follows. We denote by
s, = s, (x, 0) any one of the sumssm2, Q, s,? Then, in the first place,
& = 0 (n) *.**.* .. .. . ...(l-l).
.I... . . . . l

for every irrational x, and uniformly in 8 $ And this equation is

a best possible equation of its kind; there is no function + = + (n),
tending to infinity with n, such that

for every irrational x5.

* G. H. Hardy and J. E. Littlewood, ‘Some probleme of Diophantine Approxi-

mation’, Acta Muthem.atica, vol. 37 (1914), pp. 193-238.
t The second and third sums reproduce one another when 8 +# is written for 8,
They are considered separately for the sake of forma1 symmetry in the analysis.
$ p. 213 (Theorem 2-14). It should be observed that we there use 8, in the
more restricted mme of 8 (x, 0).
§ p. 225 (Theorem Z&l).

130
 2 Prof. War@ and Mr L&lewd, Someproblems
On the other hand much more than (14) is true for special
classes of values of x. In particular, if
1 1
x= G a,+ .*.’
and the partial quotients a, are bounded, then
s, = 0 (6) .,I...... ....* ..I ... .. .l (1*3),
and again uniformly in 8 *. And this resuIt too is a best possible
of iti kind, for
sn = 0 (4;) ..~1~.,~1*1‘1**1.**1....
(1 4)
l

ia false for 8 = 0 and alny irrational XT.

2. There was one obvious gap in our former results. We did
not give any simple criterion fbr distinguishing the classes of
irrationala x for which
42 = 0 (ta=) (2.1) l 1***.11 . . . . . l .*.*.***.*
3

where a is an assigned number between 4 and 1. The theorems of

this character which we provedt were avowedly tentative and
unsatisfactory. We did not even prove that someequation of the
type (24) holds for every algebraic x. It is this gap which we pro-
pose to fill in the present note.
We denote by pnlP;Ra typical convergent to x, taking
PoOP1tlPs
-=- __=- -sp .%
Qo 1’ Ql 4 !I2 a,cx;,+ 1 ' '** ;

and write, as in our former memoir,

I 1
X =- Xl
=-
a, + q ’ a,+ x2 p .“’
a,’ = a,+~,, a,‘=a,+x,, . .. )
Q‘= a,’ q+r+ q+2 *
We shall say that an irritional x is af class E if
. . . ..~.........~...~~ .
Qn+l< A%k (2 2) 1
where A= A (x) is independent of n. We shall use A generally to
denote a number ;of this kind, not the same in different formulae.
If x ia of class E, and k < k;‘, then x is of clam k’. ’
If x is of class k,

* p. 213 (Theoran 2.141).

t p. 226 (Theorem
2.22).
$ pe 214 (Theorems 2d42, %143).

131
 Further

for all positive integral values of p and pmThue a number of class, k

might be defined as one for which (%3) is true. If a, is bounded,
(22) is true with k =I, so that x is of class 1. In particular a
quadratic surd is of cla& 1, and every algebraic number is of
finite class.
We shall now prove
THEOREM A. If x is of class k then

and unifurmly in 8.
In particular we have, as a corollary,
THEOREM B. 1f x is algebraic then s, = 0 (w), for some value
c$
. a less thas 1, and uniformly in 8.
We shall also prove that Theorem A is in a sense the best
theorem .of its kind. This will follow from
THEOREM C. It is possible to choose an x of class k and a 0 so that

3. We require the following lemmas :

1
Lemm I: Qn+l + Xn+l!ln = xxl x2
x l

a.. *

For
Qn+ xn %--I
qn+1j- xn+1qn= (a,,, + 2,+1)!7?8+ G-1 = F+ h-1= x .
n n
1
As q1+ wlo= a, + $1 = 2'

the lemma follows. As an obvious corollary we have

Lemma 2:

4. We can now prove Theorem A. If we choose v so that

nxxl ,.. x,,, xv< 14 nxxl .., xvml **,***I*I***
(4 1)
l

we have*
1
= W n Ir,x, . I x,J,
)+OLl
l

2x1 . . l xv-1

(4.21)
* L.C., p. 213.

132
4 Prof. Hurdy and lk L&lewd, Some prdlems

and ah s.n = .,a.. . a.. . . . , . . . ..(4~22).

i
we write
xX1x9I..Xv-l= 93-j (Ocjs 1).
Then
1
-= au+1 + xv+1 < a+1 + 1 < Au,+~< A -!b+1 < -A @-I,
xv 9 V

by (2e2); and so, by Lemma 2,

1

It follows, from (4*21) and (4022), that

01* s, = O(W), y=Min(l-Qj,ikj) .,.*I.***(4.4)I

Now

Hence in any case

k
.
s, = 0 (Pw) .*.~1~.,.~~~1~*...~...*. (4 5) 3
which proves the theorem.
Theorem B ia an immediate corollary, since an algebraic number
of degree m is of classm - 1 *.

5. The proof of Theorem C also requires only a slight modifica-

tion of our former analysis. We take 8 = 0, and write, as before,
q = @ = @X-my = remi”
(x> 0, y> 0, 0 < r < 1) l *:(54),

%~=9~(0,7)=1+25p l *...*..* . . . . . .
(5 2)
F
l
l

1
Suppose it were true that sN= o (v). Then the series
1 + 2 Cq”” = 1 + 2c en=* r”’ = zu, P
+ By the classical theorem
of Liouville: we, e.g. Bore& Lyons &UT la tit&ie des
fonctiolzs (ed. 2, 1914), p . Z&29, It haa indeed been shown by A, Thue (‘uber
Anniiherungewerte alge pbraiwher Zahlen’, Journal jiir Math., vol. 135 (1909),
pp* 284-305) that an algebraic number of de ree na is of class 4~ + E for every
posifive 6. See Bad, Lepna GUT la th&ie P e la croiasmxe (1910), pp. 164166.
More recently C. Siegel ( ‘ Approximation dgebraisc her Za hlen,’ Math. Zeitschrijt ,
voL 10 (1921), pp. 173-213j has shown that nn algeljraic number of degree ‘112ia of
clams 2&i - 1.
 would satisfy the condition
u m =uo+ul+ . ..+u.=o(mfa),
and .we should have

It is therefore sufficient to show that (5.3) is false for an appro-

priate x; that is to say that

for a sequence of values of y whose limit is zero.

We suppos-e that
%+1> Aqn” “.*~~.*......*...****~.. (5 .5)
for an infinity of values of x, and consider, as on p. 229 of our
former memoir, the range R, of values of y defined by
1 1
-----,sys-.
I
Pn+l !l n

It is sufficient to fix our attention on a single value of y, viz.

y = q/k--l,

which plainly falls within R, when Ic> 1.

We employ (as on p, 226 et seq.) the linear transformation

where the sign is chosen so as to make ad - bc = 1; and here we

make another assumption, viz. that this transformation is one of
the types which (following Tannery and Molk) we denoted by
lo, 20, 50, or 60, and which transform 4, (0, T) into one of the
functions 9s (0, T) or ‘;)4(0, T). It is plain that this may be secured
by an appropriate choxe of x*.
This being so we have, aa on p. 23Ot,

whkh proves the theorem.

* We cannot prove that 8, = 0 (na) is rcezler true for an irrational of cIas8 k; for
it is poaaible that every n for which (5-5) ie true should give rke to a tramformation
of type 3” or 4Y
t The condition a, = 0 (1), used there, is only required in connection with
cases 30 and 40, hera excluded.

-
134
 COMMENTS
This is a supplemntary note to 1914, 3. The results of that paper, on
the order of magnitude of 8,(x, 8), were proved on various suppositions
as to the rate of increase of a,, in the continued fraction for x, as a func-
tion of TL Here irrationals x are classified according to the values of 3c
for which qV+Jql is bounded. This classification has been followed by
later workers (see Koksma, pp. 27-28).
5 3. The work of this section can be greatly simplified by using the
estimate (2) in the comments on 1914, 3. This gives
% = 0 ((rtq,‘))kl(k+l)(q~+~)lI(~+l))
= 0(nW+ll)
if qV+JqF is bounded.
8 4 (joottiote on JIM a). Our knowledge concerning approximation to
algebraic numbers has been completely transformed by Roth’s theorem
of 1955. In the language of Hardy and Littlewood, the theorem implies
that every algebraic number is of class 1 +E for any E > 0. Thus we have
813(x,O) = O(n++E)
if x is algebraic.

135
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION : THF,
LATTICE-POINTS OF A RIGHT-ANGLED TRIANGL’I?

By G. H. HARDY ad J. E. LITTLEWOOP.

[Received March S&h, 1920.-Read April 22ncl, I92O,]

I. In ha24~th.
f . 1. The problem considered in this paper may be stated as follows.
Suppose that or)rend W’ are two positive numbers whose ratio 8 = w/d
iH irrwtional ; and denote, by A the triangle whose sides nl*e the coordinrtte
axes and the line

circle (the problem of Gauss and Sierpinski), and the ~IG!IZ~~~ C# tljfe wt-
~W@W hyperbola (Dirichlet’s divisor problem), both of which have been
the subject of numerous researches during the last ten years. TlX
psrticular problem which we copsicier here has not, so far as we kl- AOW,
been stated quite in this form before. It is however ertdy brought into
cdnnection with another problem which has attracted ti certain amout of
attention, and which has been considered, from varying points of view, by
Lerch,f by Wevl,
L 1 and by ourselves.$ This problem, which we shall call

1922, 6 (with J. E. Littlewood) Proceedings of the Lolndon Muthe-

136 m&d Society, (2) 20, E-36.
lb * G. H. HARDYmd J. E. Lrlrru~woou rAppi ‘22,

P7dde772 B, is REIfollows+ Suppose that, a ususl, [SC] denotea the integral

part of z, and that
(1.12) {z) = x-p]--+
Then whd is the most that cm be said CM to thre ode of mqrzitudc of

(1 . 13) 8 (0, 72) = i { PO1.

v=l

1 . 2. We begin, in 4 2, by proving the formnln jvhich estnhlislles the

connection between Pr&lems A and B, and showy; that the first lxol~lem
is a qnerPsliaed and more symmetrical form bf the second. We prove in
fact that

(1 l 21)

where S(q) is a sum very similar to the sum 1 . 13.

It is trivial that

(1 ,211)

the area of the triangle, together with an error of- the order of the pi-
meter, The second and third terms of (1. 21) occur naturally when we
consider, instead of A, the similar MKI similarly situated triangle whosa
vertex is at (1, 1) instead of the origin ; for the area of this triangle is

--4 71
--- rl 1
2w’ 2w 2w’ + 2’
But no closer approximation than (1 ,211) is in any way tCin1; md,
when 8 is rational, S(q) is effectively of order q,’ so thnt a universal
formula, professing tu be mare precise than (1 , 21X), woul~~ necessarily1
be false.
In 5 3 we deduce transformation formulE for N an& 8, which are
generalisations of a formula given without proof by Lerch, anti which
enable us to study these sums by means of the expression of 8 as a simple
continued fraction. In 5 4 we prove (a) that

for any irrational 0, and (b) that (1 .22) is the most that is tine for every
such irrational. Incidentally we obtain the corresponding results co,;-
cerning Problem B : the first of them at any rate is in this case fftmilinr,

137
 19ao.l 17

In 5 5 we consider more closely cues in which the rate of inorease af the

quotients in the conCinued fraction is comparatively slow, alld in particular
the case in which they are bounded ; and we prove that in this 0888

and that this result too is a best possible result of its kind. There are
naturally analogous results for Problem B ; that corresponding to (1 23) l

was stated as a new theorem in our communication to the Cambridge

congress, but had, as was pointed out to us by Prof. Landau, been given
aJready by Lerch.
Up to this point our argument is entirely elementary, and both methods
and results are of a kind to be found in our ‘previous papers on Diophantine
approximation or in those of other writers. We have therefore aimed at
the maximum of compression and have omitted a good deal of elementary
algebraical calculation. The concluding section (8 6) is more novel. In
it we prove that #Lf8 is ulgebraic then

(1 .24) SM = Oh%
w?we u < 1. This result is unlike any which we have been able to prove
before, and is obtained by entirely diffment methods, based on the proper-
ties of the analytic function
1
(I .25) s”,(s, a, 0, w’) = Fd
WI,n=o (a+mw+~@‘)”

Thicr fun&ion will be recognised as a degenerate form of the CLDouble

Zeta-function ” introduced into analysis by .Dr. Baxn8g.*

2. 1. 1W8 write

(2.11)
[I0
4 = tf +f,
w 7q = [I7q +f’,
where O<f< 1, 0 ,<f”< 1.

+ E. W. Barnes, Ib d memoir on the DotibM3ampoa-fun&iron “, PhiE, Tnzw, lb#. &QC.,

(1)),VOI. 196(WUl), opt 26W87 ; see in psrticularpp. SU-349. Fez a study of some of the
propatios of the degenerate function (for which the ratio W/W’ is real) 888 a, E& .I&wdy, u On
double Fourier seriee, and in particular those which represent the double Zeta-fun&m with
real and incommm~b& pmamekrs “, QwrbrEy Jo74mu,z, ‘Vd, 37 (1906,,pp. 63-79.

138
 Suppose first that there is no lattice-point oln the line (1 b II), 0L: AI?
d the figure. Then the number of lattice-points inside OAB is

Now [-z-J = - [%]-l+Q, where ercis 1 or 0 according as z is or is not

an integer ; and ~0-f’ cannot be an integer, since then q-p~would be
an integral multiple of O’ and there would be a lattice-point on AB. Thua

(2 l 18) If'-pq = - [j&-f']-1 = - @-f'J+ {PO-f '} -4.

Substituting into (2. 12), and using (2 Sll), we obtain, after a little re-
duction

(2 .14)

where
(2.141)
and
(2.142) m = Y{p&f’\.
/L=l

since 9 is bounded, the problem is reduced, substantially, to the dis-

cussion of S(q)).
The preceding argument requires a trifling modification when there is
a lattice-point on AB ; there cannot be more than one, since 8 is irra-
tional. In this cam the sum (2 12) gives N(q)+1
l instead of N(V).
There is one value of p for which p0-f’ is integral, and for this p the
-8 in (2. X8) is chsnged into +. The net result is to leave the final
form& unchanged.

3.1. In order to obtain a formula fur the transformation of N(q) or of

S(q), we employ the familiar device of adding together the number of
lattice-points of the triangles OAB, O’A’B’ of the figure.
If we take new axes O;X, O/Y, aa shown in the figure, it is plain that

x+y=
[ w1
‘J +1, x+y=[51+I;

139
 19

and the equation of &dB, referred to the new axes, is

(3.. 11) w’x+wY = ~+w(l-f)+o’(l-f’) = 23,
my. Repeating the arguments of 5 2, we find, for the number N’(H) of

Y
lattice-points of O’A’B’,

(3 12)
l

where

and

F and F’ being defined by

(8.123) gw = gw +F,
[1 $ = [F]+F’,
0 Q,(F<l, 9,(F’<l.

3.2. We suppose now that o < w’, 0 < 1. A glance at the figure
shows that
[$I = [5]+1-
Substitutipg for H in terms of PJ,frcm (3. ll), we find at once that
(3 .21) F F = 0(1-f).
The same argument shows that

(3 .22) F = --l-fF p1
0

140
20 G. H. HARDYa92d J. E. LITTLEWOOD [April 2%

where J.I is an integer ; it happens that the value of p is not m.ateriaI ho

the argument.
It is also clear from the figum that

(3 w23) NM+WH 1 = pJ[cJE,

xvhere e is zero unless there is a lattice point on AB, and then unity.
Substituting for N(q) and N’(H) from (2.14) and (3 12), using (2 W11), l

(3.11), and (3 .21), and reducing, we obtain, finally,

(3 924) s+sr+~ = -

8 .3. It is important, in view of Problem B, to show that this formula

includes a formula given by Lereh? Suppose then in particular that
d= 1, o= 8 < 1, and kite

(3 31)
l

1
where art is the integral part of *O.
Starting with an arbitrary positive integral 32, we write 920 = M+S,
where M is an integer and O’< d < 1, and take

r = M+l = ne+1-&

Then f’ = 0, F I_ $ (mod I),

by. (2.11) and (3 22) ; and there is no lattice point on AB, so that E = 0.
l

Suppose now that q ia a positive integer and

l--s
ix< 7 < q+1.+

The11 II 1-s
-=+----=n+g+f, f=yj--$P-S
6 f
Also H = 11+1+0(1--f) lies between M+Z and M+S. Hence
M+2 v-1
Sf Jf+1 +$r
(3.32) = 2 z- ;
v=l { 8 I *+{ e 1

* M. ZIerch,zoc.cit.
f It is easy to 881~that (l-8)/8 cannot be ink@.

141
1920.1 f&ME P-E~ OF DIOPHASTLNB APPROXIMATION. 21

and

Also

rJi=l 1 t’= 1

say. And (n+1)8, l o.9 (n+q) 8 have all the integral part X, since
qe < 1-S < (q+1)8. Henae

(3.34) so = k (ne+re-M-*) = E (re+s-g

p=l r=l

Substituting from (8.52), (3.381), (8 33), and :S. 34) into

l (3.241, and
reducing, it will be found that
S(l-S)
(8 .85) s+d = p-7,
which is the formula of Lerch.

We may clearly suppose thab 8 < 1. Suppme bhat

1 1 1
(4 11) e=
a, +ct, +9 +**0'
l

(4. 12)

We have, from (3 .24>,

(4 . 18) S$-P = O(l/@,
the constant of the 0 being independent uf both q and 0.
we write q = wf, so that

5 = ip+-e(b-fJ+i--f’ = fe+o(l),
22 G. He HAHDY and J. E. LITTLEWOUD [April 22,

say ; so that

(4. 14) s = 0(1/O) -sp

Similarly, we have
S1 = 0(1/&)-S,, SQ = 0(1/e,) -sat . . ‘1

where &, Ss, . . . are sums of the type&

S2= Y {&i-f&s $B= Fa {P&-,f8t, l . ‘,

PS”r1 lb=1
so that s a = o(m), $a = o(fee,od, . l e .

It follows that

(4.151) s = 0 (i) +o (&) +**.+o (+-) +we, l ** %-1)

and

(4.152) 8 = 0 (*) +o (+) +...+o ($) +ocfee,. . . e,-mm

1

We ahall require both of these equations.

4.2, We choose v so that

(4 21)l tee 1 .C* 8,-l& < 1 4 fO& l .* B,-1.

It may beverified at once * that &8,+1< & for every s. Henae on the
one hand
(4.22) ee1 l . . 0
U-l = 0(2=+),

and on the other

-iv
(4 .23) $+-&+..& = O(vM+ = 0 (A) .** v-1 = O(vS+[)~
1 F 1

From (4 . 151), (4 .22), and (4 .28), we obtain

(4 . 24) s = o(s2+~)+0(2~q3 = o(f),

aincg Y tenda to infinity with f; asnd the theorem follows from (2.14) and
(4.24).

+ Bea our paper “ $ume problems of Diophantine approdmation (rI) ” [Aeta M&+
~a&a, Vol. 8’1 {1914), pp. 19%230 (p. 212)].

143
 1920.1 Soaa~ PROBLEMB OF DIOPHANTINE APPROXIMATION. 23

4.3. To Theorem Al correspondti, for Problem B, the well known

theorem :
TH~REY ml.-If 9 is irra timtal, tl en

s@, 72) = 5 {+po; = u(72).

p=l

The proof of this theorem irr included in that of Theorem AI. We

have only ho take q = kw’, where k is an iategnr, so that f’ = 0, and to
write & = q/w = k/O, n = [(I9

4.4. THM~REM la.-+ q(q) is my fuozctio~fi of 17which teds steadily

tu inJilzity with ‘I, then there is an irrational 0 sunlit that each of ths
inequalities

is satisjed fur a sequenceof indefimite2y imreasi~zg valuer of q,

Thus Theorem Al is the best possible theorem of its kind.

Making the transformations indicated in 4 . 3, we see at once that it
isr enough to prove
THEOBEM ~%.---lf +( n ) is my fwtction of n wJGch tmds steadily to
infinity w iih n, then there is an irrational 0 such tlmt each of the in-
equalities

is satis$ed focr m in$nity of values of 7t.

To prove thie we use Lerch’s formula (3 35).

l Writing
(4.41) nl = fne] = 7ze-s, 1~a= nlel-sl, ...3 u,+~ = d-h,

(4 . 42) 9 8 = PS8- Ul-&I

- 28L ?

we have

l (4 .43) ~(6, n) = q+-s

(
$, n,
>
= $o-s(Ol, 723 = ~o-~l+s(02, 7tJ

We suppose ar+l even, and exceedingly large in comparison with the pre-
ceding quotients ul, a,, ge+t a,, and take 0)~~= &z7+ Then 7~9-+~ = o and

144
24 CL H. HARDY and J. IL L~ILE~WOOD [April 2’2,

& is pm&ally ;2, so that &(l-&) is certainly greater thin +. Having

fixed nT, we can determine ~t,-~, 7~2, . . . , n,, N from the equations i4.41) ; snd

It i8 then plain that, if a,+1 is sufficiently large in comparison with the

preceding partial quotientt3, ~(8, la) will have the sign of (- l)Y, and

(4 .41)

And, by choosing a 8 for which sufficiently violent increments ia the

order of magnitude of the quotiente occur at an infinity of a@es in the
continued fraction, we can secure the truth of (4 . 41) for an infinity of
valu08 of 92,

5. Results cmcmziug specid classesof irratiw2ds.

THEORRM ~3,-- U~zder tl2e same cmtdition,

s@, n) = O(lQg n).

To prove Theorem 63, we ref‘or~ ta the an@is uf 4 . 1 and 4, 2, but

utie (4,152) instead of (4.151). In this ca8e we have plainly

s z o($-) +o ($) +...+o (i) = Q(v)*

1 V

we have Y = O(Iog &> = O(log q); and the theorem is proved. Theorem ~3
follows. 0$0rdkr~ : third is &he theorem which, 88 WQ expkimd in 1.2, w~a
claimed’ $8 a new theorem in our communication to the Cambridge cton-
grey, but i8 rashly due fo Lerob.
It will easily be verified that, if W8 swwrne

145
 1920.1 SOME PRosLmm OF DIOPIL~NTINIE A~PROX~ATION. 25

W8 obtain an error term of thea order

s = d{ (log ‘I)P+lf.*;-
if we assume uJb= O(@“), where p lies below a certain limit, weiobtain
s = O(f) (a < l).”
As so little is known concerning the order of magnitude of the quotienb
in the continued fractions which express irraficnals of particular types, it
is hardly worth while to go into further detail.

THBOREMB~. -There are values of 6, with btxmded quotimts, suc?~ tlmt

each of the inequalities

Thue Theorems 63 and ~3 are, alsu best possible theorems of

their kiti Ta prove this, it is plainly
& enough to prove Theorem M ;
md this we ehall do by considering the simplest irrational of all, viz.
45-l 1 1 1
2 =T+T+i+...
We writ0

and take the convergent8 to 0 to be

l’hen it is wily vmikd that

146
26 G. H. HARDY md J. E. LITTLWOOD [April 22,

5.3, w8 first tske PZ= qs in the formula (3 31).

l We find without
difficulty that
[PO@] = Qd-11 s = qse-[qge] = $+I,

if 8 in 8v8n, and ‘[@e] F &-1-l, 6 = l--$“+l,

if s ie odd ; and that in either case
‘I#
(5.81) ni= c~ {ve/
v=l
satisfies the equation
(5 .82) a,+a,-1’ = * ( P+l+(-ly+lw+*).

Using this recurrence equation to express 6,0in terms of

we find, after reduction, that

@W2
(5 l 83) as 12*/5 -g(-,),+l&+1+(-1),+1$.
t‘

Sappoee now that

(5 -84)

We can express rc in one snd only one wsy in the-form

92= qc+qs+q&. l 9+qS, = q8-t Ql,
where a, sI, s2, . . . are descending integers differing by at laa8t 9 ; and
Ql
m N = as+ 2 {(q8+)oq*
p=l
Now q,8 differs from an integer by less than does any ,uO. Hence
[@l+ru>q = q,l,+Wl
ana {(Qs+d@t = q8- q,-1+&- [/Al]-+ = (-1) @+l+ {&?
s(8, 42) = a,+(-1)W1&1+8Qp
We now write
&I = as,+q~cb=+qe, = qq+Qg, &a = qQ+Qa,
and SO 011, and repeat the argument. We thus obtain

(5 185) ~(0, ~2) = cs+aB,+aB,+ l l l +rs,

+(-1)“8”+‘Q,+~--1)“1B1+1Q,+...+(-l)s~.-l~k-1+I~.~-

147
1920.-j SOME PROBLBMB
OF DIOPHANTINEAPPROXIMATION. 27

5 . 4. If in (5.35) we substitute the values of the a’s given by (5 . 33),

the first two terms of (5.38) will plainly give a contributiotz boundti ior
all values of s, so that
(5 .41) cF~+bgl+...+rQ = - -J5 ((-1)8+(-1)~~+***+(-1)s1)+0(1).

Again

(5 rn42)

and the Burn of the second terms is numerically less than k, and a forthi
than .s. The sum of the contributions of all such terms to (5 35) ie l

therefore less in absolute value than

s@+l+slP1+l+... = O(1).
These terms, then, may be disregarded. Making this simplifioation, and
substituting from (5 .4,1) and (5 .42) into (5 35), we obtain, -finally,l

(5 l 43) s (0, ?B)= O(l)- 8& ((-l)“+(-lr+...+(-1p)

5 5. This. formula enables US to study the behaviour

l of ~(0, n) for
different forms of n, and in psrticular to prove our theorem. Let us take,
f 01’ example,
s =4k+4, sl=4k, s,==dk--4, . . . . sk:= 4.
Then the right-hand side of (5 .43) becomes

-- se 1 +@+-@-4+**~+-e” +0(l) = (vs$-O(l)

I 9
41/5 + Jz ( 1-P >

where C =&(&o)=-$)#O;

and s(8,7&) ia negative and greater than a constant multiple of 8. Simi-

larly, if we were to take
s = 4k+S, s1 = 4k--1, ...3 So = 3,
28 [April 22,

we should find s (0, $2) to be positive and greBter tkn a constant multiple
of s,, Sitlw s is gr&er than. & coast& multiple of log 31, this BOAT
plefes the proof of Theorems A4 and 84.

5 +6. We should perhaps, before passing to more transcendental in-

vestigations, add a word concerning the case, so far excluded, of a mtimur2
0. It is easy to see that, when 0 is rational, no such results as we have
proved in the irrational case are tlrue : ~(0, 72) is effectively of order ?h, md
the oscillatory
. part of N(q) of order q, Tim, to take a simple case, the
aeries IZ { $p 1 is
-v_-
1
4;
fil ‘
++
g-i-
3+ $-p- ii-+ l ...

and s(#, ?2) - - +42.

In general, for a fixed rational 8 = piq, we have s (8, 11,)- @, where

6 . 1. The substance of our concluding section liee somewhat deeper.

Our goal is to prove

where u < 1.

THBORM ~5.- U72der the same conditions

s(8, 92) = O(na) (a < 1).

We requiresome preliminary lemmas concerning the function-

a 1
(6.11) [&, a, 0, w’) = m c.,L=O.
(a+mo+7&)’ ’
where a, W, and w’ are p&k, snd s = vfit. This fun&& is a de-
generate form of the double Zeta-fnnction of Dr. E. K Bamaa Barnes
considers only the mm in which (as in the theory of elliptic function@ the
ratio 0 = w/d is co?q&x. The se&e (6,. II) de&m the function in the
fir& instance for 6 > 2.

149
 und s= 1, whare it bahmes like
l l 1 w3-w'-2a 1
T
ow s-2' 2&d S-l
respee tidy.
This is proved by Barnes when 0 is complex, and hi8 proof, depending
on the formula

is equally applicable in the case considered here. we should observe that

(-u)S-1 c e@-l)‘w (+‘), where log (-u) has ite principal value, that the
contour of integration is the same as in the well-known Riemann-Hankel
formula for the ordinary Gamma.andl.Zetri .functions, and that the formula
is vaJid for a!1 values of s except positive integral values.

6.22. LBYMA j%- Suppose that 0 <u <.w+d, and that 8 = w]d
is ati algebraic irmtimul. Then there ti a K such that

tW’8 4 mow
m=l ml-’ sin -
d

To prove this form& <westart from the integral (6.211) rend i&grate

150
30 G. H. HARDY and J. E, LITTLEWOOD [April 22,

that the horizontal line8 (3) pass at a distsncs greater than a constant S
from any pie of the subject of integration, and that the loop (1) passes
between the origin and the poles & 2&/w, & 24;’ nearest the &gin.
This being so, it is easy to see that the contributions of the rectilinear
parts of the contour tend to zero when the sides of the r&angle move
away ta infin& and that
& = I? (1 -s) lim E&
where R is a residue of the integrand. A simple calculation shows that
the residties yield the two series required. If 9 = O/O’ is algebraic, we

/ ain'?-l > 7721-5 1 sin 7 1> 37t+,

where c is a constant depending on the degree of the algebraic equation

which defines 8, It follows that the two series of the lemma are abso-
lutely convergent if CT is negative and su%ciently large? We shall
suppose in what follows that the series are absolutely convergent for
Q < -K. The for%ula (6.221) may of course hold in a wider region
than this.

SUppOBe that Cl< r < ba* We may suppose the contour of integra-
tion in (6.211) deformed in such a manner that
I~l=l~rg(-~)l~ib+~~
.at every point of it, and I#Jl=h+h
at all distant points. We have then
1 (-@S--l 1 < k 1u 1” pl < A I,@ @+L) VI,

where A is a number depending on cl and q,

f I-(&-s) 1= O(g-f* It i 1t p-“) = o(e-(~-lwI)~

S, = 0 (e+l ~~~e!~~~,l,~+$utul) = O(eeItt).

+ It is hardly necessary to give fuller details of the’ proof, as the substance of the Iemma
ia contained in the papr of Hardy referred to in the footnote to p. 17.

151
 1920.1 SOME PROBUXS OF DIOPJLANTLNE APPROXI~L~TION. 31

6.24. Lemma y irr required only in order to prove a somewhat deeper

lemms, viz. :

(6 . 242)

Of these relations, (6.241) is obvious, since the series (6 , 11) is

absolutely convergent for 6 > 2 ; and (6 .243) follows from (6.221),
since we ‘have
( 2rnT \
(27r)“-‘lY(l--S)si~ ,I -(iw’-U)+*(l-S)K j = O~iPIt’Il?(l--S)l~
w

uniformly in 1z~, and, of course, 8 similar result in which o snd or)’ are
interchanged. Finally, (6.242) follows from (6.241), (6 248): and the l

well-known theorem of Lindeliif. t Lemma y is used only to show

that the conditions of Lindeliif’s theorem are satisfied.

6 . 25. Our last lemma is of a different character. we write

the numbers ZP (no two of which are equal, since 8 is irrational) being
arranged in order of magnitude. We suppose that f is not equal to any

there exists a number H, indepwdent of T alrtd [, such that

# For explanations concerning the “ p-function ” of a funcfionf(s), defined initially by

a Dirichlet’s series, see G, H. Hardy an-d M. Rieq &‘ The general theory of Dirichlet%
series, ” Chtb-idge MuthmuticaZ Trm@, no, 18, 1915, pp. 14-18.
t Theorem 14 of the tract referred to above.

152
 32 ci. H. HARDT and J. Em LZTTLEWOOD. [April 22,

We have

the right-hand side of (6.252) may be written in the form

say, Now”

G
(6. 253) 2%
T
If we write lP z @P, k = ep, the series becomes

(6 .254)

Now Bohr,t generalieing 8 remit of Landau, 1 has shown that the eerierr
(6.254) is bounded, provided only that

md it ia easy to verify that tbe condition (C) ieJ satkfled by our se&a
Xl-”P z &-% . For
1p+1- ?P= a+m’w+dd-u-mm-nw’ = ho+ko’ = W’ (k+?@,
say, and so, since 8 is algtibtaic and apl+l < &+I?,

E 4‘) > HI;;H;

xp+1- Ap = log 1+ p+;-
( P

153
 38 *

H, mh~rever it occurs, denoting ti positive constant, not of COWHO the

~tne at different occurrences. Thus l3olw’~ conclit;iou icJ mtisfied, tmd
Lemmu E follows from (6 ,253).

6.3. 14% cat1 how prove our thewem, We take T = & where
0 < y < 2. We choose arbitrrtry podtive numbers 8 alxl e, and take
(J= 2+s. . .
1% therl tipply Csuchy’s theorem to the integral

take11 round the rectangle

(CAT, c+iT, +fC+iT, -K-iT),
the eides of which, taken in or&r, we denote by (1), (2), (B), and (4).
U&g Lmlmb U, rv0 obtain

Now
(6 .32) w(f)+0 ($) = W(g)+o@--t6-49
0) =
by Lemma E ; and

by Lenma s. It remaiue to estimate the cotkibutiow3 of the horizontal

sides ; and it is clear9 from Lemk d, that the coutribution of either is of .
the form
O( Max %* TF(“)-’ +‘) = O(Mnx eq),c

.I t*+ KM-d
q=a+;- -1;. y+c (-K\<a<S),
2+K

It is clear that q cannot exceed the greator of its values for. r = - Ii

and c = c, viz, .
-K+(K-#y+% 2+s-y+e*

The possible errw-term arising frum the fir& of t.he vltlues mayI be
&sorbed into that already presimt iu (6 I 38). That ..‘cwrespndiug to

154
 [April 22,

the second, la8 well 11sthat in (6 32), may be absorbed in a single term
l

o(pt6-Ytc ). We have therefore, on collecting our results,

We have still y at our disl~osnl. Taking

I -K+fK++) y = ‘2+6-y,,

2+s+I!c
we o bbaiu y.=
g+K

I
(jvhich is, aa we sul)lmwd,

\
a+&y - w
positive and less than 2), and
&(Q+K)--
$+K
l

Tlh irJ etjual to (l+K)/($+K) < 1 when S = 0, a~1 is therefore 1~s .

than unity if S iu sufIicieot& small. We have therefore

(6 l 35) W(f)= & + +-- s+Olka>,

where rx < 1. In order to obttliu Theorem AS, it is ot$ necesstbry to
attribute to a the yarticulur value o+wI and to replace f by q, Yince W(f)
then becon~es N(q).
Our rrrgumeut rrtrturally yieldti a definite value for a., But it becomes
clear, when we consider ths particular catie of a qtaadrdtic 0, that the
valtie w obtained is, in the light of Theorem AZ, not the best v&e possible+
1!k1 ale &he&ore content to show tllrtt u is in any cue less thtbn unity.

Additioml Note (March 13th, 1921).

We have developd the transcendental method of 5 6 considerably

since this ~~tryer was first communicated to the Society.
SuylJosr: thut h > 0 and
- wk<$> = 2 @-&jkb
w

Then W&=)= & r+i’ [&) ;;$~f;) fB+kds

T’ c-b
if c > 2. We trantiforn~ thie equation by (1) moving brrck the path of
integration to tlw line d = -q < 0, with the a’ppropriste corr&ions for the
residues, (2) substituting for c&s) from (6.221); and (3) integrntiug term

155
 by term. !l!his process caz1 be justified if 0 = W/W’ is algebraic and k and
q ;u*e &oven appropriately, and we obtain an expression for lV&) in the
forti of an absolutely convergeat series.
We then make *use of a lemma which is of some interest in itself,
viz. : if there are constads h > I ard H’> 0 such that

N(d = & - &- - &, +O(qa+‘),

(5) N(v)= ‘&, - 2 - &, + o”+llP+12wo’ Qwm’

156
fib* &l&U!! 1WBiiMS Oh' bIOYHINTlNK Ai'PHO~tItiATfok

Here t-k qh’ ‘*IS irrational and algebraio, and the ssriea i8 to be inter-
preted as meaning
hm
l

-c

We mtly take this opportunity . of correcting th misutatem& in our

commuuicution to the Ctimbricige Cougregv referred to on I). 15. It was
 CORRECTIONS
p. 32, Line above (6.264). Read: 1, = dp.
p. 34, last line before Additional Note. Read: Theorem A3.

COMMENTS
This paper was communicated to the Society at its meeting on 22 April, 1920 (see
Proc. 19, xxiii-xxiv),
0 1.2. The reference, in the sentence following (1.21 I), t&he triangle with vertex
at (1,l) does not seem to be appropriate. In comparing the number of integer points
in a region with the area, the simplest procedure is to put a square of side 1 with its
centre at each integer point in the region. As an approximation to the number of
points, we then get the area of the triangle

This agrees with the main term in (1.21) except for an additive constant.
8 3. For the proofs of the 0 results (Theorems Al, B 1, A3, B3) one could use in place
of the transformation formula (3.24) the simpler formula obtained by counting the
integer points horizontally instead of vertically. In the notation of the paper, this is
s+~f+~ef(l-f) = S*+*f’+#yf’(l-f’), (1)
where 8” = 2 (d-1-f }*
6 g/w’
Thus S* takes the place of S’, defined in (3.122). But for the a results (Theorems
A2, B2, A4, B4) it seems that this formula is not adequate.
The preceding remark is relevant, to the tetrahedron problem, mentioned in the
introduction. There are still two formulae similar to (1) above, but there is no
obvious analogue of (3.24), since it is not possible to put two tetrahedra together
to make a rectangular box, in the manner of the figure on p* 19.
$8 5.2, 5.3. The numbers qO, ql, qz ,... are the Fibonacci numbers. The representation
of n in the form stated after (5.34) is obtained by taking qs fo be the largest, Fibonacci
number not exceeding n, then qsl to be the largest not exceeding P+- %, and so on.
8 6.6. The behaviour of z {@} f or rational 8 depends on the convention adopted for
the vaIue of {t} when t is an integer. In the present paper the value - & is used, in
accordance with ( 1.12). But in general a more appropriate value is 0, so that {t}
is an odd function oft, With this convention, 2 {PO} oscillates finitely for rational 8.
5 6. Although the analytical method of 0 6, which was developed further in 1922, 9,
is a remarkable triumph of technique, it was not necessary for the proofs of
Theorems A5 and B5. More precise results than these were proved elementarily in
the second half of 1922, 9, and had in the meantime been found by Ostrowski, using
a different elementary method,
As regards algebraic values of 8, see the comments on 1922, 5. The effect of Roth’s
theorem is that one can take QIarbitrarily small jn the results of Theorems A5 and BE.

158
,

Some problems of Diophantine approximation:

The lattice-points of a right-angled triangle
(Second memoir)

By G. H. HARDY in Oxford and J. E. LITTLEWOOD in Cambridge

1. ,lntroduction, ’ . 1
1.1 This memoir’ is a sequel to one published recently in the Pro-
l

ceedings of the London Mathematical It contains the proofs of

Society’).
a number of theorems enunciated in an appendix to our former memoir,
together with a considerable amount of additional matter.
The prdblems which we consider have occupied us at intervals
since 1912, when w’e referred to them briefly in a communication to the
Cambridge Congre&), and indicated certain questions which we were
then una.ble to answer. In the meantime they haire attracted the attention .
of Herr HECKE~). and Herr OSTROWSKI~), who’ have dealt with them in
two very beautiful memoirs published recently in this journal, and to
whom we are indebted for this opportunity of publishing our own,
The very remarkable analysis of’ HECKE is Mainly transcendental,
while OSTBOWSXI’s is entirely elementary, and we use both elementary
and transcendental methods. <Our tra.nscehdental method is entirely unlike
HECKE’S, and little need be’said. uas. regards the relations between his
results and ours. The relation6 of ok elementary’work to OSTROWSKI’S
are a good deal closer. Our method, depending as it does on formulae
like those of SYLVESTER and LERCE~), is fundamentally different, but
. the ‘r&ults are to a considerablb extent the same. A detailed analysis

I),& II. E.~DP and J. E. LITTLEWOOD, “Some problemsof Diophantineap-

proximation: The lattice-pointsof a right-angledtriangle”, Pyoc.Londolrz &zth. Hoc. (21,
20 (1921), 15136. We refer to this memoiras I.
3 6 H, HARDY and J. E. LITTLEWOOD, “Some problemsof Diophatitine ap-
proximation”, Proceedings of the Jifth intematiortal congrem of mathematicians, 19.12, 1,
223-229.
3, E. HECRE, “Ober analytischeFunktionenund die Verteilung van Zahlen mod.
Eins”, Hamburg. Math. Abh. I (X921),54-76. We refer to this as IX
3 A. OSTROWSKI, “Bemerkungen &rTheorie derDiophantischenApproximationen”,
ibid., 77-98. We refer to this a3 0.
“) see -3.1.

1922, 9 (with J. E. Littlewood) AbhmdZunge~ a~ dem Mathed-

mhen SemiMr t&w Hamburg&when Univeraitit, 1 , 212-49. 159
 The lattice-points of at’ right-angled triabgte. f-f. L213

of the points of resemblance and difference would occupy a good deal

of space and seems to us unnecessary, though we indicate the theorems
which have been proved by OSTROWSHI as they occur. We should add
one word, however, as to the relative advantages of OSTRUWSKI’S method
and our own. In some parts of the theory the advantage ~~.OSTR~WS~‘S
method Seems to us incontestable; in others there is little between them;
,and in others the advantage seems to lie with ours. It seems to us
desirable to develop the whole theory systematically from our own point
of view; but Where OSTROWSKI’S method is clearly simpler, we content
ourselves with an outline of our demonstrations, suppressing the algebraical
.details of our work and condensing our argument to the limit of intelligi-
bility. In particular we have followed this course in 3.4.
12. All our theorems involve 4an irrational number 8, which we
generally suppose positive, less than 1, a.nd expressid as a simple con-
tinued fraction
@=--- 1 1 1
(1.21)
al+ae+wt-~-*~
We write
(1.22)

and denote the convergents to (1.2fj by

-Pl - 1 -Ps - as
Qi R’ qs -.- alh+l,’ “”

We shall make continual use of the two lemmas which follow,

which are trivial, but very useful, and which seem to ha& escaped
attention.

1
(1.231) eyer+1. l l . ep+a-l< us)

where thsis the s-th ikp~ 0’f FIBONACC_T’S

serie31, %, 3, 5, 8, 13, . . . .
We deduce this from
Lemma 2. We have

1 1
(1.24.)
28el.. . er-l 4 qrL Oer 2. or+ ’
For

160
214 Cl. H. Hardy
, and J. E. Littlewood.

and so
w-t-'Brqr-I = 0g L e ,
1 l , l I
r-l

which \proves the lemma. (

TO deduce Lemma I we observe that (1.24) gives

and that, for given s, q3 is a minimum when al = Q = . . = as = I, in l

which case gs = zcs. Taking now or for 8 we obtain the desired result.
13 We write, as usual, [x] for the integral part of X, and
l .

(1.31) (X > =x - [xl, {x} = x - [x] ,- $ I

Thus (x} is the arithmetical function denoted, by HECKE and C)STR@WSRT,

by R(z) -+. Further we write
(1.32) x E x-x
1
where X is the integer nearest to x, If x is of the form n+ 2, we take
1
xc-.
2
Throughout our argument the letter A (or occasionally B, C, . I .)
denotes a positive constantI. This constant may be absolute, or may depend
upon the. parameters involved in the theorem in question; it will not ’
generally be the sanze constant in successive inequalities. The O’s and o’s
which occur involve constants implicitly. It will generally be obvious on
what, if any, parameters these constants depend.
We ‘shall frequentSly be concerned with conditions of the type

(1.331) nh 1sinn,W j> A In 2 11,

Or
(1.332) nhisinnWI <A (n = nj>,

where h > 1, and the notation implies that the second inequality is satis-
fied for an infinite sequence nl, n2, . ., 92j of values of n. These con-
l

ditions are obviously equivalent to the corresponding conditions in which

sin n 8 n is replaced by n. Further, (1.331) and (1.332) are equivalent to
(1.341)

161
 The lattice-points of it right-angled triangle. II. 215

and these again, by Lemma 2, to

$
(1.351) 1 (r 2 11,
r= r (8 '18 l 1 l l 8r-ll)h-1

1 A
(1.352) = Yj).

7’ P(88I.
CT

* l . 0 t?r-l)h-l
r

It is well-known that a#condition of the type (1.331) is satisfied by every

algebraical 8.
The A’s of these inequalities may be absolute or may depend on 8
and h. If absolute in (1.331) a’nd (1.332), they are absolute in the other
inequalities.

2. The analytic treatment of the triangle problem.

2.1 In this section we continue the study of the %iangle” problem
l

(Problem A of 1) by analytic methods. We denote by N(Q) the number

of lattice-points inside the triangle whose sides are
X =o, $4 = 0, wx+w’~ =+70,

where w and zu’ ase two positive numbers whose ratio 8 =$ is irrationd.
1
We proved in 1 that

(2.112) --7
w -
I 7-1+j ?’ 7
20
-= [-71+fI
WI wr
1

(2,113) (D(q) =

(2.114) Nq) ===cIE” 8 7-f r lm

lLp<-- Y
- =tv

Our problem is the study of U(q), or, since a(q) = O(l), of S(T).
We proved in 1, (ce) that
(2.1-2) U(9) = o(q)
for every irrational 8, (b) that this result is the most that is universally
true, (c) that
(2.13) U(q) = 0 (log q)

162
 216 Q. II, Hardy
I and J. E.Littlewood.
when 8 has bounded quotients, (d) that there ’ are B’S, with bounde’d
quotients, for which each of the inequalities

I (2.14) U(q) 7 A logq,

is satisfied for arbitrarily

U(q) 4 - A log7
large values of 8, and (e) that
I

(2.15) ~ U(q) = O(q9,

where a = a (0) < 1, whenever 8 s&i&es an inequality of the type (1.33i),
and in particular whenever 8 is algebraic. Of these results (a)_(d) were
proved by elementary reasoning. and @) analytically. Our immediate
object is to prov(e more precise’ results in place of (e).
We denote by CS(s) = Se(8, u, w, w’) the analytic function defined,
when the ‘real part -6 of s = d + it is greater than 2,. by the series

where a is positive and ,?gs has its principa.1 value. This function is
a degenerate case of the “double Zeta-function” of BARNES~). Its
principal properties, so far as they are relevant to our investigations,
are summarised in 1,
In this section the R’s, B’s, . . . are in gelieral not absolute but
l

functions of the parameters 8, 72, . . . . .

2.2. Lemma 3. If h > k 2 I U&

(2.21) G]sinn8+7A
-t7m
mz
(2.22) I
s m= c = 0 (log nz)“.
1 721hi sinnB+

It is plain that neither hypothesis nor conclusion is affected if

we replace sin% 8 7t by a. We have therefore

(2;231) nhI.lk>B9), _ T
and, if we define h, by
(2.232) - nhn 1n t?Ik = B,
we have & < h. Consider
’ now the sum
2m.
(2.224) Tm=c !- =jf u..
m nhlnel” ?Ja

I) E. W. &iNES, “A memoir on the double Gamma-function”, Phil. Truws. Ray.

Sue. (A), 196 (1901), 265-387.
2) B is the same constant throughout this sub-section,

163
 The lattice-points of a right-angled triangle. 11. 217

The &ms of Tm for which h, ( 12 - 1 contribute

.
1
02 = 0 (1).
i mn-1
We classify the remaining t,erms as follows. We. choose a positive
integer 7, write
b r =h--l$-r (r=0,1,2,...,q-1)
71
and call a typical term lb of Tm a term of classr if fan1. h, -L Qr +I+ If
zc, is of class r,
/nej”s Bi+.
But
[zik > Bs-~ > 2kBn-‘r
if
br
(2.25) oe4G
and then

‘for all values of s which satisfy (2.25). Hence nb term 2cn+a corresponding
to such a, ‘value of s is a term of class as high as r,
( 1-L)
The number of terms of &lass r is therefore
. 0 m h’ , and their
contribution to Tm is

1 ‘7- br-h+&+l=
h
It follows t.hat

(2.26) T 7n = O(qmt) = O(logm),

since we may take 7 =* [logm].

If now we define Y by D 5: 5 2 D, we have

In
1
u 7n = c = wu+ o(&&)
1 nhlnOik

= o(l)+0 &log$ = 0 (log mY,

( p=1 1
which is
L equivalent to (2.22).
As a corollary we have ’

164
 G, H, Hardy and J. E. Littlewood.

Lemma 4. If (2.21) is satisjed, the series

I
c &+EjsinM@
is cmvergent for every po,sitive E.
The case of most importance is that in which k = 1, when (2.21)
redubes to (1.33.1).
2.31. We proceed to establish an analytical formula for the sqm

The k here has no connection with that of 2.2.

In order #to abbreviate our formlllae we adopt the following
convention. We are often concerned with associated pairs of series,
of the forms 2 0 (m, w, w’) and 2 @( q WI, 20); a’nd we write generally

(2.312) X(w, w’)C @(m, w, w’) +X(w’, w)C @cm, w’, to) =
(X(w, td)C qm, w, d))*.

Such associated pairs of Geries have been considered by various writers,

and in particular by LERCIEI~). We shall als’o sometimes use a similar
notation when there is no summation.
We recall the formulae)
’ I - aj+ +(I - s)n ;fc
o. sin -2mK -Ygw
L cs,a, w, w’) 1 ( w ! 1
c .
(2b313) (2 7+9-V(l- 8) = ?&&=l mw’7r
( r)21ws sin7 i
This formula is valid %&never 0 -C a ( 1~)+ w’ and the two series on
the, right are absolutely convergent. -
2.32. Theorem 1. i!hpp~se that (1.331) is satisfied and that k > ?i- I.
Then ’
(2.321) Wk (8 * V&)-
00 cos (2 $,I+Eea -F 1 n *
1 I k 1
(2n)-k--lr(k+ I) wk 2 ” + O(Ekfl--q)
b ?
?MW’7
i n&=1 mk%intU i
tohere
(2,321l) V&) = &Lgk+2--P,
p=o
I) See, for example, M. LERCH,“Sur une s&ie analogue aux fonctions modulaires”,
Comptes Ben&q 18 April 1904; G, H. HARDY, “On certain series of discontinuous functions
connected with the modular functions”, Quarterly Journd (1904), 93-123; and writings
of RIEMANN and EL J, S. SMITH there referred to.
2) (6.221)of I.

165
 The lattice-points of a right-angled tribgle. II. I 219

w+w’-22
2(k+l)ww’ ’
and q tlze integer such that k + I c Q(: k + 2.
is
I
We suppose for the present that k>h- 2. Since k 7 0, we havel)

(2.322)
C-ii

1
if c 7 2. We choose 6 so that 2 < I$kl,
2 kHi---$+6, and
1
Y-- 2-k + 6 is not. an integer. We have then
I
(2.323) -.-k-qul-h
2
- and
2-7
(2.324) I 52(s) = 0 (I t12 ) = ml@)
uniformly for d > 7 2).
We may therefore apply CAUCHY’S Theorem to the strip y ( ti ( C,

and we obtain

where

p being the largest ihteger such that --JI >r. We may write

(2.327) - u&) =2cp Ek+-,

P=O
1 3
where r = pi-2 is the integer such that k-/-T--dLr<k+2-da

The : index of the last power in &(E) lies between L-/-d

2 ;and $Sd,
’
whereas that in V&) lies between 0 (inclusive) and 1; the form of the
two sums is otherwise the same,
I) G.ELXARDY and M, RIESZ,“The general theory of Dirichlet’s series”, Cumb.
ACut&. Tracts, 18 (19$5), ‘51 (Theorem 40). , &
‘) See I, Lemmas & 8. The series which occur in (2.313) (or (6.221) of I) are
absolutely convergent for d = 7, in virtue of (2.323) and Lemma 4; and the conclusion
then follows from Lemma a.

166
 220 d. H. Hardy and J. E. Lit&wood.

2.33. II-I (2.325) we subditute for & (s) from the formula (2.313), valid
since l- r >h, ;I,nd integrate t&m-by-term. This t&m-by-term integration
, is legitimate because
1 *’
(i 7q-l q1- 8) -
2q

are convergent. We thus obtain

(2.331)
where

(2.3311) JI = -?

JP is conjugate to ,J,, J: a,mdJ: are obtained from J, and Ja by exchanging

and (--uy has its principal value, real when 2c is negative.

The function 0(u) is one of a type whose asymptotic expansions
have been considered by various writers. We h%ve

(2.332) 0(u) =

say, wit’h similar formulae for the derivatives of ID@), which- may be
written down by formal differentia,tion’). ’

‘) The function may be expressed in the form

u3 fl8- P
-2 r(k+Y+m-p) ;
X6=0

herep is the integer such that --p - I < r( --p, As regards the asymptotic expansions
of such functions, see, for example, E. W.BBRNES, “On function,5 defined by simple
types of. hypergeometric series”, Tram. Cmnb. Phil. SW, 20 (1906), 253-279, The
actual result required here is easily proved in a variety of ways.

167
 The lattice-points of a right-angled
k triangle. II. 221

We denote by J1,1, . . , I the results of replacing a, in J1, . . . . by

@I+%; by A,%. 4.. the results of replacing rl, by m3-/- @,: so that we
have four eQuat#ions of the type ,
(2.333) ,J1= Jl,~+Ji,s+
Consider first the sums J~,I, . . . . wit.h I as second suffix. If we
substitul-e CD, for @ in & . , . and combine the results, we obtain, after
l

a straightforward calculation,

which is equal, by (2.313), to

I-1) p+1F2c--u- 1) w+ I) ~]c-p-l

(p-+ I)! r(k-p) - ’

This is of the same form as the general term in (2.326), and the con-
tribution of @I may be accounted for by replacing p, in (2.326), by ~7j- 1.
If we do t.his, the last index in U&) will lie between -%-t-a and
I
and [Jk(E) may become identical with 01’ may Contain
2
extra term; it is in any case of the form V&)-j- O(5k+1-q).
There is also CB, to be considered, but a2 is of lower order than
1
rPl to the extent of a fa,ctor -, and its contribut4ion is accordingly trivial.
mE
We therefore obtain
(2.334) U&)+&,I+&,I+ J:,I+&,I = Vk(E)+fW+l--q)~
Next we ‘consider the sums J1,a, . . . l l Substituting first @ for CD,
we obtain, after reduction,
cos --2wm I I *
i pu’$ &--a
( i -2 kn)
-((2rr)-k-lr(k+1) wlc~ w I
t ~~k+.lsin2!!!!~2
i

And as m4 is of lower order than CD,, by a factor -z,1 its con-

= 0 @+l--q), Thus

168
222 G. II. Hardy and J* E, Littlewood. .

(2.335) Ji,z+ Jz,z+ Ji,,+ J: 2 = fl+ 0 (gk+l--),

where 8 is the se&d, term on the ‘right of (2:321).
Collecting’ our results from (2.331), (2.333), (2.334), and (2.335),
we obtain the result of Theorem 1. At present, however, the theorem
1
is proved only when k > 12--%, and it is necessary to extend this
range to Ioh-1.
2.34. Suppose then that k = h + q > h > h- -$, so that (2.321) is
proved; and let us differentiate formally with respect to 5, and divide
by 1~ We ha.ve’
1 dwk
k T = wk-1,
and
1 av,
--= v&-l+ o(Ek+--) = -&-i+ o(~k-4’),
k dt
where q1 = 4 - 1 is the integer such that k L 4’5 k t I. Finally the
same process, applied to the infinite series, yields the corresponding
series for k- 1, a series which is, by Lemma 4, absolutely and uniformly
convergent. It appears then that we are led back to our original
formula, with k- 1 in place of k, and this is just what we require.
The proof is however insufficient, since we are not entitled to differentmte
the error term d(Ek+l--).
There is no difficulty of principle in completing the proof, but it
is necessary to go back to (2.331). ’ We differentiate this equation, and
substitute for a(u) and its derivative @‘(u’> the approximations given
by (2.332) a.nd the corresponding derived equation. The result is
an absolutely and uniformly convergent series, and the term-by-term
differentiation is thereby justified. JVe have then only to repeat our
previous calculat,ions, in a slightly more complicated form, t*he formulae
which we use being in substance the formal derivatives of those which
we have used already. The final result is the same as before, except
that k is replaced by k - 1, and that the result holds whenever
k-4 =.h-l+q>h- 1. When we restore k in the place of k - 1,
the proof of Theorem 1 is completed.
2.4. From Theorem I we can deduce a pfoof of the equation
numbered (2) in the appendix to our memoir 1. This equation.is not quite
so precise as one which we shall obtain later in an elementary manner,
but it is of some interest to show how it follows from the analytic theory. .
Theorem 2. If (~331) is satisfied the?

(2.41)
for every jgositive &.

169
 The lattice-points of a right-angled triangle. IL , 223
Suppose, in (2.321), that k is the integer such &at -k5 h 4 FE$: 1,
The p of Theorem 1 is now k+ 2, and
(2.42).
wk($) =

-= vk(Q+m+ 0 (+)
sa.y. If. now we suppose 0 K 8< 1, and write generally,

we have
(2.43) nw#l#l=
We consider first AS. Since

we have

= 0 ( 8($)“-1+‘) + O( (+rk-l+-) = 0 (8k+M--~)~

(2.442) _
Finally \
(2.443) ” om = pel%)+ oca”+v) = k! p(g)+ o(p+q),

From (2.43), (2.441), (2.442), (2.443); and (2.4431) we deduce, in

the first place,
(2.45)

170
I 224 G. E. Hardy aad J, E. Littlewood.

I- 1
III (2.45) we take 3 = 5 h, and we obtain ’

( ‘-+),
(2.46) W(E)< V(F)t 0 E
Similarly we have
( I:$,.),
(2.47) w(!:+k@T(~~+O E
II or, on replacing T+ k3 by E
( l-+,,i.
(2.48) W(T) > V(r) + 0 E
Finally, from (2.46) and (2.48) we deduce

(2.49). W@) = v(g)+0 !El-‘;;IfB 1.

Attributing to a the specia.1value w+ w’ and replacing 5 by 7, we

obtain (2.4l)l).
1
2.51. We next prove a theorem which shows that the index 1 - h
of Theorem 2 is the %orrect” one.
Theorem 3. If h 71 and (1.332) is satisfied for U’YLinfinity of values
of n, then each of the ihequalities _
I-
(2.511) u(q) >&I :, U(q)< 34q+
is sattifieaf0r &-bit~ily la9-ge’
valzles 0f i.
Let f(z) be the function defined, when B(z) 71 0, by the equations
e- 2 (W+ #‘)
(2.512) f(x) =~e-x~mw+~wD)=~~-xb = (l-e-““j(l
m,n==l p--l 4
We have
w-j-w’ ws+ ;;;;=f wfB + o(x), .
(2.513) f(x) =1- wwv
2ww’x +
when x is small.
Suppose next that
(2.514) E- 2NZi +a 1
X
d
where $ is small ‘and positive; and vz has one of the values for which
(1.332) is true. Then
I Ie-(w+w’)x17A, . f1-egw’x1<A8,
. ‘) The proof of the theorem is modelled on the argument used by LANDAU, ,,Ober
Diricblets Teiler-Problem”, Mtinchener SiWng&ticMe, 1915, 317~328..
 The lattice-points of it Aright-angled triangle. II, 225
and
-wx --1/(1-e. --W@Pn~ 4 AI/P+ nwzh,
I1 -49 I -+6)“+4e
so that
(2.515) lfWl7 A
aVa2+n-2h'
On the other hand, we have
1

(2,516) f(x) = zCx’p =~pj;e-xudu = xfN(u)e-“?d u,

1 1 2P 0

since N(U) is the number of Z’s which do not exceed et; or

1
(2.517) f( ‘x ) E- 1UW’X” - p&1 + Qw,
where
90
. (2.5171) a(x) = x s lJ(t() e- x” du,
0

Comparing with ‘(2.513), we see that

@(x)+M = we+3wwy- 2iP

(2:518)
12ww’

when ?-+O, and in particular if x = 6 and $--+ 0.

2.52. Now suppose that GI> 0 and
(2521) 3c(2fJ= U(u) + Bu" > 0 I)

for all sufficiently large values of u, say for 242 uo. It follows from
(2.521)? (2.5171), and (2.518) that

say, when a-+0. On the other ha.nd,

. if x is given by (2.514), we have,
by (2.5171) and (2.521),
00 :
(P(x)
= xs0me-““&4lxflPe-du = &(u)e-““du+ O(l),
0 0

*) B and C (unlike A) r&in the same values throughout the argument which
foH0 WS.

172
 226 G. H, Hardy and J. E. Littlewood.

(2.523)

<2 .--.
2nn
WI
C&-l-a<ACn$-l--u

by (2.522). Comparing (2.515), (2.517) and (2.523), we see that

A
ACWF-~-~~ dJ/6”+%-2h f C%“(P’+ + n-2h) 7 Adza.

Taking in pkticulaSr 6 = n--$ we ha,ve

(2.524) C7AnU--a) h-1 0

1 1
Prom (2.524) it follows that a > 1 -h; and if we ta,ke ti = 1 -7 then
C?A or B .>A. Unless these conditions are satisfied, (2.521) cannot be .
true for all sufficiently la.rge values of 14; and therefore the second of
the inequalities (2,511) must be true for arbitrarily large values of 7
and some value of B. The first inequality cali naturally be proved in
.’ ‘a similar manner.
We have supposed he> 1,. so that the critical vake of a is positive.
In this case a less precise form of (2.X8), . vG, m(z) = 0 (z-a), would
have been sufficient .for our aqpment. When h = 1, the critical value
of a is ‘zero. In this case the value of N becomes relevant to the
argument. We must take x(u) = U(u) _ +- M+ B,, and the final conclusion
is that B >A, i. e, that each of

U(q) =-M+A, u(q)< M-A

is true for nrtitmrily q. The conclusion is not entirely

lcwge valuesmmof
trivial, but it is certainly much less interesting, and is nd longer in any
sense a best possible result;
2.61. We proceed next to the proof of the exact formula for X(q)
enunciated at the end of. our former memoi?). ‘. This As, the analogue of
VORONOY’s formula for the number of lattice-points _in the aTea x >O,
p-0, qls q* -.
It is now necessary to consider the exact definition of N(q) when 7
is of the form p 21;-f-’ qtu’ and there is a lattice-point on the boundary
of the triangle, We agree that such a pdint is to be counted as one-half.
I) I, p, 35, formula (5),
 The lattice-points of a right-angled triangle. II. 227

Lemma 5. If ;1 is positive and Y real, then the integral

1 sinvx dX
(2.611)
(cash INA - cus wx) (cash w’;c L cos w’x) x
0 l

is convergent; and it may bewarhated

. by expanding
.. the subj’ectof integration
as a double power-series in emwA and e-“-5 and integrating’ terv-by-term.
The formal result of this process is
00
(2.612) 4 CK) ’ sinvx cospwx cosqw’x
--Rpw+q?u’l
c
sinh WA sinh- w’;t p,q,pEp e x
dX t
J
0
1
where Ed -= - (p = 0), &p= 1 (lp 70). There is at most one term fur
2
which ~&pw-cpw’= -0. This term, if it exists, we remove from the
double series and consider independently, and we denote the modified series
’ by21 Since term-by-term integration is certainly permissible over any
finite range (0, X), it is sufficient fro show that
00
I --1(pw+qw’~
(2.6.13) c wqe s
.-. ,
X

is convergent and. tends to zero when X -+ 00; and this will be so if

’ singz
(2.614) d=
2
F

&P % e
;-Rtpw+qw’)
I
X
- 5. ax < ,&, --

where
(2.615) e = Y-cpw&qwI, .

is less than E for every positive E and sticiently large values of x,

We write
(2.616) ,’
:
say, tih&e H*70. Since

. I1
we have. ‘. _ G
.. ,
(2.617) 014 - A Od A '1
H c e---RPw+qw < -
H
< -&
l

2
P&=-o

if H. is sufficientb large. It is the&fore sufficient for our purpose to

prove that &en H is jixed we can so choose& that

174
228 (3. H. Hardy and J. E. Littlewood.

< -&1
F -~@~~+qw’)
(2,618) AC e
2

the summation &ending ow those values of p and q fey which

(2.619)

2.62. We divide the terms in question into four blocks corresponding

to the four chdices of sign, and esta.blish a corresponding conclusion
1 1
for each of them, with 8 E in place of 2~. If the signs attached to p
and Q are the same, there is nothing to prove; for, as Y is not of the
form pw+pw’ or -pw-qw’, there is no term which satisfies (2.619)
when X is sufficiently la,rge. It is therefore sufficient to cbnsider‘ the
case in which, for example, e = Y-pw + qw’. But the inequalities

H H
n‘r --<V-pw+pw’~x,
X

where H is tied, are only possible when p 7 F, Q 7 F, where g+ 00

when %+ 00, and the number of values of q, corresponding to a given p,
is 1 at most. Hence the. sum extended over such values of p and p
dues not exceed
e-AP 4 A@-“$ c
P-5

which tends to zero when X+ 06; This establishes our conclusion and
completes the proof of the lemma.
2.63. Lemma 6i If 3, is positive and [ real, then the integds
I
,CiX dX
(2.631) 1 1 x .’
sinyzqxsinsw’x 1

in which the ‘puth of integration is a tine parallel to _the real axis, are
cokxrgent; usd tlkr v&&s may be 4&2.&&d by expanding the cosecants.
in &Mr~ of CwA and esw’$ and integrating term by term. 5
The series ‘LoILbe used are different in the two integrals; for

1 , * p++ wdx 2i&P++),yie-(P++A .

1 -’
-- ioe ( ) =-
?C
sib-;;- w x 0
or
1 l
OD -**++ wix

22 ( >

sin 147x 0
9

2 , ..
1 Oe

175
 The lattice-points of a right-angled triangle. II. 229

according as x = E Jr ik or x = 5 - ik Apart from this, the argument

is the same for the two integrals, and it will be sufficient to consider
the second.
Consider first the analogous integral in whichI X
is replaced by
---1 1
The method of evaluation contempl&ted is certainly
x+ir
lzgititiate for tlzis integr&
. - since
cm

1 1 :
is convergeit. We may therefore replace,
I -x by r+z and then

Next I
,a x eCiE sin zw(5+in)sin
I -1,1(5+iq
2
CI CL
3 1 1 = 4e 2.r(coshwA-coswQ(coshw’rZ-cosw%)’
sin zwzsin+x

Working out the imaginary part of the numerator, we find that it is

a sum of constant multiples of. terms of the type sin
Thus Lemma 6 is reduced to Lemma 5.
2.64. We- define a sequence (Bj) as a sequence of values Rj of A! which
tends to infinity. and all of whose members differ, by more t.han A, from
2m76 2n7c
any of the numbers ---, ----Cm, n = 0, 1, 2, I 4 .),
\ WI
Theorem 4. If y>w + wc <so that N(q) PO), then

The sign of sumtiation is to be inttipreted us fullows:

the sum we form
w’R WR
of all thoseterms of the two msociated
, ,swiesfor which m< -- n< -
27G’ 2rr

I eCia:- 1
(2.642) J(Rj) = 93
27G I I
( f siny wx sin -gw!s

1 cix

= 2rr3‘U sin 21 ewx -1

sin 21 W’x dXx 1:’ ,A

176
230 G. II. Hardy and J. E. Littlewood.

the contour of integration being the rectangle (-ii, Rj-ii, BJ+ i R, in),
where A7 0, except that the origin is excluded by a small semi-circle
of radius e, described about it as centre. We make ,j -+ 00, e -+ 0.
The integrals along the imaginary axis vanish identically. The
integral along the side parallel to the imaginary axis tends to zero, That
along the semi-circle tends to the limit

E2

The integrals along the sides parallel to the real axis also tend to limits,
by Lctnma 6. It follows that the sum of the real part,s of the residues
of the Integrand, at poles within the contour, tends to a limit 8; and
we have

(2,643)

say,
.We evalua.te CT,by integra,tion term-by-term, which is shown to be
legitimate by Lemma 6; and we obtain
00
(2.644) J 1 Z c A4 41
21,4=-1
where
--iR+w
(2.6441) ‘(p- 1) e- (JZ-+u+zo’))iz dx
X

and $2 = p w + QZU’. We may add to j$, 4 the corresponding integral

along the line (0, -in), since this vanishes identically, and we may then
deform the path of integration into the real axis. This gives
00
2
h,4= -y sin (q -G)X+sin
0
which is zero if JJ 7 7 and - 2. if 2 4 + Thus we obtain’ from (2.644)

(2.645) J 1=- 221 = -2 N(q).

h
This equation still holds when Q is of the form 23w -/- qzu’, if we adopt
the convention’ stated in 2.61.
Similarly we obtain a series for Jg, in which the typical term
involves the integral

177
 The lattice-points of a right-angled triangle. II. 231

sin(g + S2-w-&)x-sin 0I
0
Thus
(2.646) J e= 07

as is evident a priori, since the value of the integral is independent of

;1 and tends to 0 when 1 4 .=.
From (2.643), (2.645), and (2.646) we deduce

A straightforward calculation shows that

-1 *
COS2mx
pu’q-+wq
1 (-1)” *
(2,648) S=
n4 msin-
mUllIt
msiny
WtW’R
1 l

W
ButI)

Thus we obtain the result of the tdeorem.

3. The sum s(n,0). . -

,
3.1. In’ this section we use elementary methods. We are conkerned
primarily with what, in our former membir, we called Problem B, that
of the order of magnitude of the sum

(3.11) s(n, 0) =ccm 01,

m=l
though sometimes we return to Problem A. I
Lemma 7. If 0 is positive and i~dhai, x 2 0, y 2 0x, and
f (0) = g(0) = 0, then

I) This is easily proved directly by contour integration. See the second memoir
quoted in 2.31.

178
 232 & IT. Hardy and J. E, Littlewood. .

The sum with respect to m is

%h) C (g(m)-gh-W--f([yl>
n--O 2 (g(m)-gtm-1))
m@=n rm&=ryLm>x

=nf&@, (9([+I) -g ([+I)) --f&d) (g ([qq-,([,I))

and the fmt term here is

which gives ‘the result.

In (3.12) take f(u) = g&J E u, and write

I (3.13)

(3.14)
We obtain
p = [x],
a, = %a@) = {?M),

Y = [y] 7 [ex],
.A= Me)

6 = e,-,=
= a?&(+) = (-q,

@[XI--[0x].

(3.15) ’ $ m+$[;] = P,

or, on ,exp&sing [m O] and $

I in terms of a, and ,&, and redutiing,
[

(3,16) .’ . &+& & yp.

1 1

. Lemma 8. If n, n,, n: are positive or zero integers such that

..
(3.17) * 928 =n,+a*=n;+e, OOWl, -l<e<O,
thm *
j3.18) . s(n,e)+s(*l, &)= $-+;-,

(3;19) s(n,e)+s(n;,$J= +
e (1 -4
2e - [ 1
y . .’

The first of these formulae is the special form of (3.16) obtained

by supposing that x is an integer ~2. The second is a simple variant.
Since. $ = .ti& 1, the left hand sides of (3.18) a,nd (3.19) differ only by
{n: o,}, a.nd (3.19) follows’ from (3.18) by simple algebra.
The formula (3.18) is that which, in our former memoir, we attributed’
-to LERCH’). Herr Os~~oywu has pointed out td us that it had (in substance
l) See; I, -p.-20.

179
 The Iattice-points of a right-angled triangle. 11. 233
at any rate) hen found before by SYLVESTER~). SYLVESTER’S formula
is indeed ’ equivalent to the more general formulae (3:15) and (3.16).
General formulae of the type (3.12) appear to 0riginat.e with DIRICHLET,
and the actual formula (3.12) was given, in the spe$al case in which x
is an +integ.er,by HACK@.
With these formulae should be associated the formula (3.24) of our
memoir I.
3.2. Theorem 5. IJ” 8 satisfzes (1.331), and h 7 1, thm

-- 1
(3.21) s(n, 0) =
d
2 h
1
l

This theorem is due to OSTROWSKP).In the less precise form in

1
which l- x is replaced by 1- I,h E, 1‘t’ 1s included in Theorem2, which
was enunciated without proof in our memoir 1 4).- The reading of
OSTROWSKI’S memoir suggested, to us the following theorem, in which
Theorem 5 is included5).
Theorem 6. If -8 satkfies (1.331), und h7 I, then
1
(3.22) = 0 I3l-x 1.
This theorem includes both Theorem 5 and Theorem 2. It is
easily proved by a combination of Lemma 2 with formulae taken from 1.
If (1.331) is true, (1.341) is also true. As in 1, we choose G+-
so that
(3.23) b $801 l l . l tip- l@=l~E8&.*.*@r-1,

where 5 = 4 We have then6)

WI’

(3.24) U(q) = 0

I) 3. J. SYLVESTER, "Sur la fonction E(X)", Comptes Rendus, 50 (1860), 732~734

(CoEZected math. papers, 2, 179480). See also pp. 176, 177, 179 of the same volume of
the collected papers.
“) ,,Uber Summen von grb8ten Ganzen”, Acta Mathemntiw, IO (1887), I-52. See
also J. W.L. GLAISHER, “On certain transformations of Lejeune-Dirichlet’s in the theory
of numbers, and similar theorems”, Quarterly Journal, 43 (1912), 123-142.
3, 2. ca p. 82.
4, 2, c. p. 35, equation (2).
s) Generally, an “0” theorem relating to Problem B is included in the corresponding
theorem relating to Problem A, as is explained in 1, An W” theorem, that is to say,
a theorem which, like Theorem 3, tends in the opposite direction, is on the other hand
more difficult than the corresponding theorem concerning Problem A.
“) p. 22, equation (4.151).

180
234 G. %I. Hardy aad J, E. Littlewood.

‘Write now
(3.25) 08 1 1I I &--I = E-j,

where 0 dj 5 1. Then, by (1.351), E-(h-l)j < A&, and so

It follows that j’ Q-, and

(3.26) s’oe 1 l l ‘I @r--r

Again, _if 1 ( s.( T, we have

by (3.23). Using (1.351), we obtain

From (3.27) it follows that

(3.28)
1 1 I- --
+ + ++<A$ i ( l+s:_i +(HY-Z&-If

&-1 or-2 +”

<-A$ -- “22
100

k=O

by Lemma 1. Finally, from (3.24), (3.26), and (3.28) the theorem follows.
The constants of the argument are not absolute: the theorem is
not true uniformly in h.
3.31. Theorem 7. If h7 1 a& (1.332) is tme,, thm
1
(3.311) ls(m,e>l3-A h

for an inJ%aity uf valuesof n. ’

It should be observed that this theorem ha.s different interpretations,
according as the A’s of (1.332) and (3.311) depend upon 8 and h or are
absolute constants. It is true on either interpretation, but: is a little
harder to prove on the second.
Taking the second interpretation, we may restate what we have
to prove as fdllows:-If h> 1 and

(3,3121) limd+inn876/ SB .

181
 The lattice-points of’ a right-angled trialn& II. 235
,
then
(3.3122) ~s(n,e)~ r.Cnl?, _*

where- C dqpncls only on B. In what follows R’s denote absolute

constants; ‘C’s constants dependiw only on B; the O’s are absolute.
Let E be the upper bound of the numbers z for which

(3.313) +imnzIsinn07rI (B,

Clearly we have Hz h. We proceed to show that there exists

an. h 2 h such that
(3.3141) lim nhi 1sin 928 7zI( B
and
(3.3142) lim(8 @I . . a0Bn- #++&+ . l l ++*=o

n-1

whep n + =. We must distinguish two cases.

Case (iJ: H > 2. In this case we have only to take + .

h i= Max (h, 2).

For, since (3.313) holds both when x = h and when x.= 2 < H,i (3.3141)
is satisfied. Also

qee,....e,-1) (-r+-+
1 1
L... *+- b-11 4Ane-An = o(1).
C 0 01 1

Case (ii): H ( 2. We begin the discussion of this case by showing

that numbers hl and HI exist for which

(3.316) K I = (hl-7l)(~~-l)-(H~~~)>o,

(3.317) .l~md+in128nI SB,

(3.318) lim#+inn@rI = &,

The last of these is an immediate consequence-of H< HI, and we need

only consider the first three, If h = IT we choose Ii, = Jz, and &
greater than H by so little th&t (3.316) is satisfied. If h < ;K we -
choose hl and I& on either side of H, and differing from it by SOlittle

182
236 G. B. Bczrdy and II, E. Littlewood.

&at h(h, and (3.316) is satisfied. It is clear that (3&31?), or (3*3141),

is satided in &her case.
3,32, We denote by K’s positive constants depending only on B, h,
and Hr. It follows ‘from (1.352) and (3,317) that

(3.321)

I
for an infinity of values of Y; and from (1.(351) and‘(3.318) that
1 K
(3.322)
. -< I )H,Ili '
en (0 0 1 L l 0 On-1

for all values of n. Hence, observing that h1( 2, and using Lemma 1,
we have
(3.323)
(0 0 la l . 0 n-l)fbi-

-<~e-“w-r)
l

fee1 ’ ‘e’ erP-f (0 p(n-ml),

or r

(3.324)
(0 8 &pi-i (00 1 . . . 0rgl-l 2-h
-I-(I&-I)(%-hd
1. l .

-z
) (00 _ l..J%-lF
& or

From (3.323) and (3.324) it follows that

which is (3.3142). This completes the discussion of case (ii).

3.33. It is now not difficult to prove Theorem 7. We suppose v
selected so that (3.321) is true, and we write

(3.332) n < 88 ” 8
1 . . l
Y- 1
by (3.321), or
(3.333)

On the other hand we have, by (3.19)l),

l) We have nr+ 1 = n&+-g, where O<,q< 1. It is therefore the second of the
transformation formulae (3.18) and (3.19) to which we appeal.

183
 The lattice-points of a right-angled triangle. Il. 237

for 05~4~; and so

s(n,8) = (-1)ys(12y,~y>+o ++++...++),

i 1 Y- 1

Is(n,e)l7~ls(ll,,H,)l--8(~+~+ *+q.
I Y- 1
But

and so
(3.334) .I,(~,s)l,~-A(~+~+...+~)*
Y 1 Y- 1

,The ratio of the second term on the right to the first is less than

K (WI . . . O,-,)hl-L

by (3.32!), and, this tends to zero as Y -+ 00, by (3.3142). Hence

--
ls(0,m)I>+al-$,clll k,
Y
3.34. It will be useful to observe that the ~2of our argument satisfies
inequalities
(3.341) Acl,+l< Pi4< 4,+lL Qlv-+-I

when v is large. The second and third.of these are immediate consequences
of (3.331), (3.332), and Lemma 2. To prove the first we observe that

A K 1 (h-11)v
n,>n,>--7 >KZ2 7Y (Op<Y)
0,
IO@ 1 l l . e,-Jh’-l

for all sufficiently la.rge values of Y. Hence

%$I I7 %q-1
T&7----- --(l-&)>~(l-~)~
& 0r
and so

184
238 G. H. Hardy and J. E. Littlewood.

3.41 I The proofs of our next two theorems are the most difficult in
the memoir. The results were enunciated without proof in our former
memo?), and discovered and proved independently by OSTROWSKI~), I
We give a proof here based on the formulae (3.18) and (3.19),
We have abbreviated this proof in every possible w8y, and present it-
almost in the form of a sketch; for we recognise hhat the quite different
proof of OSTR~WSKI
. is simpler. It is indeed here that OSTROWSKI~ method
shows to the great&t relative advantage. At the same time our proof
seems to us interesting in itself, and it is essential, if we are to develop
the theory systematically from our own point of view, that this crucial
theorem should appear in its proper place.
Theorem 8. There is a positive A such that
(3.411) ’ (s(n, e)l > Alogn
fuY every irrational 0 and an infinity of valuesof n.
Theorem 9. There is a B = B(K) such that each of the inequalities
(3.412) s(n,@>Blogn, s(n,8)<--Blogn
is true for every 0 for which an =L K and for an in$mity of values of n.
In proving Theorem 8, we may suppose tha.t 8 satisfies (1.331)
for some h: we may take, for example h = 2, in which case

(3.413)

For, if the condition is not satisfied for h = 2, we have, by Theorem 7,

IS@, $)I 7,4 I/n> Alogn

for an infinity of values n”).
3.42. Let
(3.421) a, = 1 (c&-=3),
z a, =

6 6
(3.423) y1 = - d, = $, j/B = cf, = - $t y3 = - a, = -2,
1

‘) p* 36. We raised the question which they answer in our note of 1912.
2, O., pp. 85-92.
3, One of the inherent advantages of OSTROWSKI’S method is that it enables him
to avoid making this distinction.

185
 The lattice-points of El right-angled triangle; II. 239

with corresponding equations in which every stix is increased by

4, 8, 12, . I . .
Further, let m be a large positive integer, and

(3.425) c4m+2 = &m+2(1 -y4m+d, ik--I =’ ---%-&O<‘rcr4m

I x. = + 8,

(3.426) N,8, = x++4+5, (09~4m+2), *I

(3.427) ,N47n+3= 0,

it being understood that’ a zero suffix may always be omitted, so that,

emg*, Yo= y. The C’s .are defined by (3.425), and the equations (3.426)
and (3.42-7) then define Ji:,+z, A?d,+l, 0 = N in turn. It is
. l . N

,not bbvious from the definitions that the N’s ire integers, but it follows .
immediately from them that lV&+2 = 1. If now Nr is an integer, and
we consider congruences to modulus 1, we have

)I NT
&&--- d r-ST -- dr- 5r
0r-l
- Nr-A--l-4-r-L-l d ’
- r- 5r = N r-19
8 r- 1

by (3.427) and (3.426). It follows by induction that eveky NV is integral,

We write
(3.428) * sr = s(N,, 6,).

3.43. We use the following properties of the numbers y, 6, c, AR-

(3.431) O<Y~Yl(Y6Y8<1,

(3.432) Y8 74

(3.433) l-yi>Aa4--al (OG<3),

E =
014 ,4

(3.434)

(with similar results in which every suffix is in&eked by 4, 8, 12, . . .),

.(3.435)
+ lcrl <l-yy, (o(:r(4m+2),

186
 240 U. II. Eardy and J. E. Littlewood.

(3.437) NV21 (0954m+2),

(3.438) N,+= (m-+ 4,

24mS2
(3.439) N-L 08 1 l . . 04mt1'

Of these results, (3.431), (3.432), (3.433) and (3.434) follow from Lemma 1
a,nd the definitions of the y’s; (3,435) is obvious from the definitions
when r = 4 m + 2, and is easily proved generally by induction; and
(3.436) .is an immediate consequence of (3,435).
The results (3.437) and (3.438) follow at once from (3,426) and
(3.436):-we find in fact that Nr-l => NY, and that the sign if equality
iS impossible if r is even, Finally, (3.426) now gives N,<*2N,+l, which
Groves (3.439).
3.44. Lemma 9. If

(3.442) ut = U4t -U4t-t l+u4t+2 -u4t+s -I- 2a4t+4,

then
m-1
(3.443) .2sb, 81 = 2 ut+ ZCqw-U4m+l-+ O(l). ’
t-0

Here, and in the arguments which follow, the O’s and R’s are absolute.
Let yr = (Nr8,) = (NT+1 + a,$- &). Then, by (3.426) and (3.436),
we .have
!m = a,*+ c2r, y2r+-I = 1+ 82+x+ Lr+1*
l3y (3.18)
I <(a+ 5)(1---a--5)
(3.4441) -2(s+s,) ==a+L--
e
S- d’ d(i-a). 5 = u+o(g) 1t
0 +O(‘8 1
. and by (3.19) m
(3.4442) if, + 51
a(s,+Sg)=~*+951~ @~+~~~y~--~J _+T
1= u1+OGd,
since
[-q+] =I,%, +o($ = O(CB)*
Similarly we find
(3.4443) _ 2 (%I+‘sQ) = u%+o(b), ’

(3.4444 2h+d = us-2”a-t O(L). .

 The lattice-points of a right-angled triangle. 11. 241

From (3.4441)-(3.4444) it follows that

(3.445) 2(s 0 - d = uo + OK*)*

We have similar equations in which every sugx is increased by 4, 8, . , , .

Adding them, and using (3.425) and Lemma 1, we obtain

(3.446) c -s47d
2( co =3iJi + O(l).
0
We have also
(3.447) w34 m -s4mf2) = Uglln--U4pn+r+O(549n+2),

and (3.443) follows from (3.446) and (3.447), since ~4 nt+2 = 0( 1) and
54m+2 = O(1).
3.45. Lemma 10. ‘We have
m-1

(3.451) s(N,@7- A + Jyh4t+4 -~4t+d + &94t+2.

t-o t=0

An elementary reduction shows that

(3,452) Ut = 2(y4t-y4t+4) + ~(1-~4tHat+l)o-Yati-1)

1 1
t- y@4t+2Y4t+a(l --4t+1@4t+2Y4t+9) + 2 ~4t+4U-Y4t+d

7 w4t- y4t+4)+A~4t+2-t.-A(a4tf41 a4t+4),

by Lemma 1, (3,423), (3.432), and (3.433). Also

(3.453) Upm-Uqtrp-I= 00+0(~)+0($=)=0(1).

The result follows from (3.452) and (3.453).

3.46. Theorems 8 and 9 _follow easily from Lemma 10.
The number N is a function iV(m, 0) df m and 8 alone. We write

(3.461) N z- = j&t+) (i = 0, 1, 2, 3).

By Lemma 10,

(3.462)
f=O r-l

A A
If a,>= 4, ar--aa,rAu.,r -;8 and if arcE 3, 8,-l 78m Hence
r--l r-1

188
242 G.-H. Hardy and J. E, Littlewood.

(3.462) involves
(3.463)
i=o

where
24’??2$2
(3.4631) %!
= 001. l . 04m+1 ’

If now Nis the greataest of N, N1, NS, A$, we have, by (3.439) and (3.413),

Hence, if s (K, s> is tha one of the sums s(A& &) whose modulus is
greatest, we have
-
>Alog%>~4logN>Abgn;

and G -+ 00. Theorem 8 is therefore true for one of the three nugbers
8,8,, &, OSand therefore, by the fundament.al formula (3.18), it is true for 8.
The deduction of Theorem 9 is more immediate. In this case 8,7 B I),
and so
m-l
s(N,0)7- . A+ ACht+2 7 Bm,
0
while, by (3.439), N< B4m+2. It follows that

s(N,@)7BlogN,

which is one of the desired inequalities. The formula (3.18) then shows
at once that
s (N’? 0):~ -B 1ogN’

for arbitrarily -large v&es of 2~“.

It is not possible, in the general case, to prove two-sided inequalities
of’ the type-s 7 A logn? 4~ -A logn. Herr OSTROWSKI has gone furthel
in this direction; he has shown that s may in fact be bounded on one
side, and has, investigated- the conditions under which this is possible”).
We have xot attempted to apply our own method to this problem, as
we had not considered it before the publication of O~TR~WXKI’S memoir,
and it is dear that a proof on these lines could not be SO simple as his.
I) B denotes throughout a positive number depending only on K.
2, The question ia ieft open in 0. (S. 92’.

189
 The lattice-points of a right-angled triangle. II. 243

3.51. We return for a moment to the t#riangle problem. In this problem

. the analogue of Theorem 9 holds without restriction on 8. To prove
this we require the lemma which follows, which is of course included
in OSTROWSKI% work, where it occupies a ‘more central position. .*
.hLemma 11. We ?~UW
(3.511) s (q,, 0) = O(l)*

Pu P,,l P,,,
It is evident that, if-, -, -, l l l are typical convergents to
Qu Ql h
. Q y,2
8, &, OS,. . ., we have/ ’
Pu = Qu-l,l? P,-1,x = Qv-2,2t ’ l ’ l

If .now we take n = qv in (3.18) or (3.19), we have sl = p, or rzi = pv,

while d or e, whichever is relevant, is’ 0

s(Qy, 0) ++zv-l,l,
-
01) = 0 Qv+lB
1 ( 1 f

&-l,l, 81)+qP,-2,2J 09) =3) 0 Q 1 9

Y,

and so on, Consequently

1 1 1
+
&u) = 0 y----$+- ( ’ Qr, 181
. I l l +

q1,v 0, 1 *
= o(el;b I b, ~,+e,....e,+....+e,+l~=.o(l),
by Lemma 1.
3.52. Theorem 10. There -is us A such that each of the imqualities

is sati+$ed for eve&y 8 and ahitrarily, large values of Q.

It is sufficient’) to prove that ’

for arbitrarily large values of 7, together with nn analogous inequality

involving - A1o.g ,q. By a change in 7, ef ,magnitude .O (1), we ’can
make f’ anything we please. between 0 and 16): It is therefore enough

‘) I, p. 18.
2) I, p. 17. ” .

190
 244 G. B. Hardy and J. E. Littlewood.

to etitablish .the existence of a g k g(n) such that 0 (g < 1 and ’

(3.523) &(n) =2{p e-g} rAlog72

1
for arbitrarily large values of n, with a corresponding-result for -Alog%. ’
I We suppose’ first that (1.331) is s&is&d for some particulak h,
say h= 2. Then .
(3.524) logqY+ls Alog.p,.

There are, by Theorem 8, large. values of n for which .c& of the

inequalities
N ’ a
C{p e} > Alogn, C(r 0) 4 -AJogIn
1 1

is true, say the first. . Then (3.523) is true, with g = 0, for such vglues
of n. Let N be one of these values, as determined in 3.42, -and let
(3.525)
then, by Lemma 11,
%+I %+I N
{pe} =z--Cc-AlogN,
1 1
Qv+l -N
(3.526) ?{~~+(N8)‘3+AlogN.

If now p is the least positive integer for which

04j =pe-(bTe)41
and
(3.527) n = qv+1--w-P, -
A
then p < 8- and ’ _

(3.528) &e-g} q:e.+(Ne)j+ 0(+4logN,

if N is large enot&h. But

r by (3.525) and (3.624), and SO *

‘2{$-g} < -Albgn
1

for an itinity of values of ’ 92.

191
 The lattice-points of a right-angled triangle. JI. - 245

3.53. This proves the theorem when (1.331) is sat&fied for 12 = 2.

When it is not satisfied, much more is true ;. for then (1.332) is satisfied
for I2 = 2, and in this case, by Theorem 3, we have

(3.531) N(q) 7 A-VT, N(T) < -A-G,

each for arbitrarily large values of 7.

The proof of Theorem 3, given in 2.5, was trans cendent al; and it
is worth while to observe that it may ‘also be proved by an argument
like that of 3.5% We have only to suppose that N is the n of 3.3
so that (say]
N a 1 ,L - -

&e)r AN k

Arguing’ ‘just as in (3.52), and making use of (3.341), we establish the

existence of an infinity of values of YZfor ‘which
n -c I
z{po-g} -L --Ail h,
1

and the theorem then follows in the same way.

The Cesikro l’Bl8WIS of -the S8ri8S&e).

3,61. A good deal of additional light is thrown on the behaviour

of s(n, 0) by the study of the corresponding CesBiro mean
1
(3.611.) “(X, e) = - 2(x-m){m 0).
x m--l
The study of c(x, 6) leads us naturally to consider also the sum

(3.612) t(x,e)=~ ({w+&).

c m=l
Lemma 12. We hiw, in the nut&n of 3J,

(3.613)
where
(3.614)

If in (3.12) we take f(u) = us, g(u) = u, express the summa&

in terms of am and b,, and reduce, we obtain
(3.615)

192
246 Gt, II, Hardy and J, E, Littlewood.

where P is a polynomial whose coefAcients are absolute cons&s. If on

the other hand we take f(?l) = U, g(.zt) = u”, we obtain a similar formula
1
in which x, y; 8, e; anz,& are interchanged. When we multiply this
by 8, and subtract from (3.615), the terms in & am and zn& disappear;
and when finally we substitute for& +z& from (3.16), we obtain
\
the result of the lemma.
3.62. We say that 8 belongs to class r(H) 3f

(3,621)

The convergence of the series is equivalent to that of the series

c w-1 4,”
l

Theorem 11. If 8 is of classr(H) theiz

(3.622) t (x, 0) = CqH).
Write -
(3.623) xl = OX, x2 = o~x~,, , , , ah---& = a,,,

By (3.613), we have

Write zl, O1for x, 8, and multiply by 8; 3, Oti for x, 8 and multiply by 8 &;
and so on until the second sum disappea.rs. Adding the resulting equations,
and using (3,621), we obtain the result’).
3.63. Lemma 13. If e b&ngs to CEWr(H), thm

(3,631)
where
(3.632)

We have, from (3,611) and (3.615),

2 y (@@, 0) + a (y, 6,) = 4 +cz,-n(~~~+e~~)-~(~~--~).

1 1
The last term is 0 (H), by Theorem 11, We write p e+ O(1) for y, sub-
stitute for Ca,+z& f rom
- (3.16), and divide by 2~; and a simple
calculation gives the result.
3.64; We say that 8 belongs to class C(K) if an< K, where K >2.
In this case all of 8, &, &,- . . . belong to a class r(H) for which H-C AK.
I) This is fhe revised form of the theorem st&ed incorrectly on p. 229 of our
Cambridge. commnication, and noted a8 incorrect on p. 36 of 1.

193
 The lattice-points of a right-angled triangle. II. 247

Lemma 14. If 8 is of clnss C(K), and Y = Y (x) is de&ted by

(3,641)
then

(3.642.) e, 0) =‘j$-l) c (0,) + 0 (K”).

r=O

We have, by repeated use of Lemma 13, ,

r)
v-l
(3.643) 8(X, e) = (-vd% 6,) +C(--l)‘cw
r=O

+0 -H - I + &+*..f l -
()* ~-8,-,
(x (0 1 1)
--1 1 +
+O( ( .x e2
&+...+88

1
I...
&p-’

Y-2 v-1
0 l

1
But, since a,< K, we have 7 = 0 (K), and the second line of (3.6’43) is
r

0 K (l+8,~1+o,~,o,~, = 0 (K”),
80p l . 8,-,x

by (3.641) and Lemma 1. Similarly the third line of (3.643) is 0 (A?).

Finally
X,+2
a,, 0,) = w,> = 0 8 * .)= O(P),
v v+1
whence the result.
3.65. Theorem 12. There are 8’s of class C(KJ for which

(3.651) 4x, 0) = O(l),

and ot?aers
for which
(3.652) a(x, 0) - Llogx,
AK AK
where L>-- or L<---.
1ogK 1ogK
(i) If
(j------L---1 I
1+ 1+...
then or = 8 for every r, and (3.651) follows directly from (3,642).
(ii) If k = [K] 2 2 and

(g=----
1 1 1 1
k+ I+ k+ l+...’
that so
e+ --18 $I----
1
4
= k-1,

194
248 *G. R. Hardy and J. E, Littlewood.

t&n it follows from (3.642) that

(3.653) a@, 0) = (k-l)&] + O(K') - $(k-1)~

Also
[211
logx =&I& + O(logK),=~Iog ;zr;r+I + mgm
r-o 8,

=([+] ,l)log-++o(l,K&g R+f+-;-k*

Thus we obtain (3.652) with

k-4 AK
L E-
’ 1ogK’
log$ (k + 2 + I/M)

If we exchange 8 and &, we obtain an example in which L has the

opposite sign.
4. Conclusion,
4.1. The proof of Lemma 13 indicates clearly that, if we were to
attempt the construction of a COW&& theory-of the series&? it would be
necessary to construct at the same time a theory of the series2 (a:- &)*
A little further investigation-shows that. we must also consider the series
.l
A.b %( \ 4 a&- ad, *,a I1 the nth term of the pth series being subst a.ntially
12
1
the PthB&noullian function of a,+ 2’ Tkere . are many cur’ious theorems
connected with -these series: we content ourselves with mentioning one.
Theorem 11. If d belongs to a class r(H) (and in pnrticeclar if
a, is hnded) thm the s&Hi (a:-&) is summable (C, I), or by any
1
Ces$ro mean of positive order, to sum -12’
4.2. ‘We conclude by a brief reference to a Herent matter. It is
of considerable interest to determine the largest half-planes in which the
.-
functions .
(4.21) -- I
h (4 =2$, fi Cd Gzai n,la , f&)= &z4a;;an,,..

are regular. HECKE has shown that, when 8 is a quadratic irrational,

fl ts) is merombrphic all ever the plane, and has at most a doubly infinite
system of simple poles, at the points
-2k & 2my7Gi (k, na += 0, 1, 2, . . J
where y is a constant depending on 8,

195
 The lattice-points of a right-angled triangle. II. 249

Our more elementary methods are applicable to problems of this

kind also. Let 2 be defined as the le8st number for which

for every positive E, so that (1,331) is satisfied with h = 1 +A --/-E but

not with h = l-j-1----~. Then we can prove that
(a) fp(s) is ~q92.A~ for

(W d= bP is CGbarrier of singularities for fp (s), except possibly

wAen n = 0, so tf2at
(c) fp (s) is &the; regular for tr > 17 p, or hm a barrier of
singdarities Zo the right of 0 = 1 -p, and in partimlar
(d) fi (s) is eithw regular for 6 7 0 or lzas a barri’w to the right
I
of d=O;
(e) the series for, fP(s) is sutimable by Cesdro’s means for CT7 tip
and in particular
(f) tJze series Ca, PP is commgmt, when A+> 0, throughout the
region of existence of the function fi (8).
The propositions (a) and (e) have been proved independently, and in
a different manner, by Herr BEBNKE~).
The case in which R = 0 is exceptional a.nd more difficult. It would
seem that G = q= 1 -p is still a barrier in all cases except that of
u quadratic 8, but this we have not been able to establish rigorously.
In this last case, finally, our met#hod reveals the existence of
HECK& poles, though it does not render a complete account of them
so readily as that of HECRE himself. This is only natural, as HECKE%
method is so much more special and so much deeper than ours.
In view of the length of the memoir, we confine ourselves here
to the statement of these results, reserving a fuller discussion for
publication elsewhere.
I) II BEHNKE, Ober die Verteilung vbn Irrationalit~ten mod.I. (DieseAbhand-
lungen, I3d.I, vorliegendesHeft,) .

COMMENTS
The paper falls into two separate halves, one (§ 2) analytical in its approach and
the other (§ 3) elementary. Both sections are continuations and developments of
corresponding work in 1922, 6.
For references to other related work, see Koksma, pp. 103-6.

196
 Someproblems of Diophantineapproximation: I
The analytic character of the sum of a DirichlOt’s
se&s consideredby Hacke.
By G, R, HARDY in Oxford and J. E. LITTLl?,WOOD .in Cambridge.

1. It has been shown by HECKJP) that the function J(S) = J(s, 0),
defined when c 7 1”) by the series

- 0 . 11 J(s, 0) = &g2-s,

I
where (1.11) an = dc,fe) =n$-[n+--3 .

and 8 is a real quadratic ‘irrational, is a meromorphic fuqction whose

only singularities are‘ simple poles. we establishes, by means of the
t#heory of the new “Zeta functions” which he has introduced into
*analysis, a formula for J(S) which effects its co&inuatlpn. all over the
plane and exhibits explicitly the- nature of every singularity.
: ‘. In this note we show how it is possible to attain the same end
by an entirely different method. The-formula at which we ultimately
arrive differs fundamentally from HECKE’S in structure, and its relations
to his-are of considerable formal interest. It is remarkable, moreover,
that we are able to establish our formula, for the half-plane Q 7 2, by
elementary methods “); a transcendental a.rgument is required only to ,
prove that it converges all over the plane.
,Our method may’ be applied to any quadratic 8; but -we wish to
exhibit its principle without unnecessary formal complications, and we
therefore limit ourselves to a particular case, as did HECKE in the
memoir to which we h&e referred. We suppose that

1 1 1 / as a
8=--- ---
(12)1 a+ a+ a+ .I...* = 1/ 1+ 4 2 7

where a is an i nteger.
‘) E. HECm, Uber analytische Funktionen und die VerfeiIung van Zahlen
mod. eias, Hamburg. Math. Abh,, I, (1921), 54-76.

3) That is to say, without employing the ideas of the of analytic functioqs.

I

1923, 3 (with J. E. Littlewood) Abkadmgen, mm dem Mathemati-

schen Seminur der Hamburgiscken Ulziversitit, 3, 57-68. 197
 58 Q. II. Hardy and J. E. Littlewood.

2.. If x 2 0, y = 8x, and f(0) = g (0) = 0, we have

CSC([m4) km - g(m - 1)) +R9([;]) vi4 -ff(n - 11)

, 77&2X =

(2 1) l

= fbl)Y ([xl) 4>*

In (2.1) we write

f(u) = 1- f3-y 9(4 = 1 - e-el, where E is real and

(2 2)
l z-1 = c-05,

c being positive. Making x tend .to infinity, qe obtain

= a,&) = m19~
- [n O,] - f = ; - [;] --$,

and we obtain from (2.3), afkr some simple reductions

I
I
$-f El sh,F
2
(? I6) F (cj 8, E) =

4, This is Lemma 7 of our m.emoir in ~1, I of this journal (212-249).

a “) It is only a$ a later
c stage that; we: irntroduc’e the hypothesis (LB).

198
 The analytic ehamctor of the sum of a Dirichlet’s mice. 59

using (2.6) repeatedly, we find

the sequences (cp) ami (Fp) being defined by

1
so that & lies between 2 cl and cl. It’ follows sitiarly that EP lies
1
between.Z +I and ~~-1. Hence EP + a, and the series in (2.7) converges

when continued to infinity, Also F(++l, &+I, &+I) = O(e-cp+d+p+* )

tetids to zero when JI’-+ cro, Thus

1 1 O”
(2 I8) I;‘(c, 0,9 = ph--fZ( 2 v=u - 1Y

3. Subtracting
e-alc
F(c,8,0; f l -e-c
e-c 1
2shp

from both sides of (2.8), dividing by 5, and making 5 -+ 0,“) we obtain

(3 1)
l a (c, e) =
2
1
%ne --= w 6% 0) + x(c, e),

3 There is IIQ difficulty in jWifying this process,

199
 60 Cl. ET.Hardy and J. E. Littlewood.

OcC 1 e- c
where (3.21) w (c, e) = -- -e-c)” -2 1-e-c ’
(1

(3.22) x(c, 6) = Z(-1y fy-+~+

Y=l 0 -e v)(L---e >

and the Q’S are the values of the Z’s when To = 0. We write

(3.3) e = c&, and & is then defined by

(3.4)

1
Since = qc-1+ 0Y-1 qv-2,
801 l ’ ’ l 0y-2

where -IJy is the yfh convergent to 6,‘) the syond equation (3.4) is
Qsp

I% -qF-1 = tLl(QY--2--JBY--I);

and be = B1 - & = al = ql. Hence

(3 I5) bY= qv-1.

4. We multiply (3.1) by 6’ ’ (supposing in the first instance that

G is sufScientlJ7 large) and integrate from c = 0 to c = 00 l We find

(4 1-l
l J(s, 0) =Z$
1
= e5(s--l)-+~(s)+x(s, e),

where (4.2)

5, By Lemmtt2 of our memoiraheady referred to,

200
 The analytic character of the sum of a Dirichlet’s series. 61
00
3 p-‘ly+l

and (4.3) x,(s, e) = & ta-1ac

J0 (1 -e -‘lY > 0
- p+l >

06 I
- 1 e- ml,_,-i -kq,lc c871dc
h, ck=l r(s) s
0
w
1
2 = 52(s, qv-I+ qv, qt’--I? qv>’
=h.Ic=1(hqv-1+kqi)'

cz (s, ct, W, w’) being the double Zeta-function of BARNES*). The integral ’

is convergent if ti P 2 t so that our formal process Is valid. Thus we find

This expression of J(s, 6) as a series of double Zeta-functions is valid

for all 8, It is now that we proceed to specialise.
5. Let us suppose in particular that

(5.1) B=
1
-ui
-- 1
a+
1
a+**--
a2
-V1+ 4-2-e
--a -r t

so that y> 0 and t& = 8 for every Y. An elementary calculation

ev + (- ,>I’-1 e-v
shows that (5.2) Q,,-~ =
8 + e-‘-T ’

(5.3)

8>See p. 216 of our krmer memoir.

201
 62 G. k Hardy and J, E. Littlewood.

where (5.31) P = II+k& Q = h-kc?

It follows that

and so (5.4) J(s, ej = e5(e--1) - j- t (4 + Q (8, e) t

where (5.5) JJ(s, 19)

~=l h,k E-U

The .quadruple series is absolutely convergent if G7 2 l We have

then in fact
a+ 1)
IPTQ-+cA(h+k)-*-‘, ]&I~-d(#-k)z,

where the A’s are constants: and the series may be compared with

We may therefore rearrange the series as we please. Effecting t-he

. summation with respect to Y, we obtain

These equations are valid for a7 2, It will be obsemed that,

up ta this point, our argument is entirely “elementary”,

202
 The analytic character of the sum of a Dirichlet’s seriw. 63

6. We shall sow ‘show that the equation (5.6) gives the analytical
continuation of 2(s, o), and so of J(s, @), all over the plane. We begin
by proving the following lemma.
Lemma. The functzh &(s, O) is regulur all ove7* the plane, mqt
possibly fh simple poles at the points

(6 * 1) s e -z+1,4+2, l L..! 0,1,2.

If D is any bounded domain in the plane of s, from which these points

are excluded, then
(6 02) I zzb, 8) I -= e”l,

where E is my yusitivs number, fur all value uf s in Dj and at1 suf-

jlc&tly large valuesof 1.
We may obviously suppose that ~s3 at all points of D. We have

(6 .3) z= 2&, 0j fcpq &E = ws+lz @*‘- kw)I -

I (h w + kwF)3+E’

if (6.31) w =e, wI=l, d7=, 2.

we write

(6.4) z __ es”’ (8 - 2) (s - 1) mm (8 + 2 - l l 1)
(8 - 4) (S i- 5) l . , l (s - l- 5j 2u-9z?

where 8 7 0; and we prove

(i) that 2 is regular in a strip

(6.5) -AA,-{ G
< a<E 3;

(ii) that lZl-+O when /tI+m, for every fixed value of 2, uniformly
throughout the strip ;

(iii) that (6.6) py<A, (d-3);

(iv) that
(6 7)
l /z/<&@ (d= -AA,l)

for all sufficie@ly large values of 2,

203
 64 G. II. Hardy and J. E. Littlewood.

In these propositions the A’s are appropriately chosen constants

(independent of 6, 8, tid E).
‘7. To ptove (i) we’ observe that

(%l) s(6+l)~oo~(S+z-1)20-"-EZ

=s(s+1)~~*++~-~)~ (hw’-kw)z,
(hw + kw’j8+

=a (hw+kw- = D5,(&w+ w’,w,20’)= D5,(s),

if 6 > 2, D being a diBerentia.1 operator

As ia (s) is regular save fey simple poles at s = 1 and s = 2, the same

is true of DC,(s), ,and so 2 is regular in the strip (6.5).
The truth of (ii) is almost obvious; for Sz(s), and so *Z f is of finite
order in the strip, while -esp tends to zero like C?
To prove (iii) we observe that, when d = 3,

y w(hw’-kw) i 1 -7 I
-d htw+ kw’ (hw + kw’)’ =<z (h w + kwTmg)

O) The ~function,

--w (w’- w x) - tql--ex)

w+wtx - 8+x

decreases steadily from I to - e2 as 2 increases from 0 to CID.

204
 The analytic character of the sum of a Dirichfet’s series. 65

It reknains to prove (iv). It is plain that, if we write

then (7.4) 17 fs) 1 < p--dP 21;1< &-JP,

when (3= - A, I and 2 is suffxiently large.

Again, the number of terms in B is I + 1, and the numerical value
of a coefficient & is less than Ai. Hence, if

we have (7.6) 121 < (I + 1) A; eAcdl’--‘= 51~-c eA+@&

if 1 is sufficiently large.
In order to obtain a,n upper bound for Zy, we use the equation lo)

The two series involved iri (7.7) are absolutely convergent, when
8 il; quadratic, for o*< 0. We have to examine the effect on the series
of the operator l)r. Expanding the num,erators in (7.7), we obtain four
series? two of the type

(7.8) (27+-l T(l -s) sin+(1 -s) X&P-~ cot??+,

lC

and twd of a simpler type (without the cotangent). It will be clear

from our analysis that it is sufficient to consider the series (7.8).

‘9 See p, 218 of our former memoir for this formu18, and for an explanation of,
the notation.

205
 66 G, H. Hardy and J. E, Littlewood.

The external factor is indbpendent of w and PU’, and it is easily

verified that, when s = - A1 I + i 5,

(7 9)
l 1- 8) sin 3‘l (1 - 8) 7r 1 4 (A, z”+ Pyv)
d

8.t Again, it is easily verified that

where the number of terms is less than A;, the coefficients Cj are all
less than ZAi*t, aj and bj are numericaUy less than AI1 2, cjz 0, and
4 . ’ -- 1 1. The effect of the operator Dr on the function (‘7.8) is
thzefore to produce a series which possesses a msjoratit of the form

rnw'n A
But cosec--7 > 2;
m

and the. series. ifi (8.2) is I therefore convekgeht, and less than Aid7 if
d = - A1 I and 2 is sufficiently l<Te. It follows that

if E is sufficiently large.
Taking the maximum of. the right hand side of (8.4) for variation
of t, we find
Al
AJs+te= -+,

Ii) We have in fact

206
 The malytie character of the sumof a Dtichlat’s series.

SQ that

if +t is sufficiently large. This is (iv) of 5 6.

From (i), (ii), (iii) and (iv) of 5 6 it fdh$‘) that

for -&l<a<3.--
- ---
- If now D is the domain of 56, and

- ,416E
< d =< 3
throughuut B, we have in D

if 6 is sufficiently small a.nd 2 sufficiently large. This completes &e

proof of the lemma.
8; The poifits 0, - 1, - 2, . . . wwere excluded from D in the.lemma,
as they are possible singularities of some of the functions 22. A bounced
domain D’ can contain at most a finite number of these points. We
may suppose it to contain some or a11 of the points 0, i I, , - JI, l ..*

but neither of the points 1,2. The function

is plainly regular in D’, and satisfies an inequality analogous to that

satisfied by 2~ in D.
The series (5.6) is uniformly convergent in D’, and gives the
analytic continuation of JJ(a, e), and so of J(s, @), all over the plaw.

‘9) If f(a) is regular and bounded in the strip Q 2 us & and L (0) is the upper
-bound of its modulus for 13= u + it, then log 1; (b> is ‘a, convex function of u, 80 that

‘.L%e bheorem is a variant of Bdamard’er CLthree circle theorem”. See G. Doetsch,

,,Uber die obere Grenze des absoluten Betrages einer analytischen Funktion auf Gemden”;
Xzlh. Zeitachrift, 9 (192U), 237-240.

207
 68 U* II. Hardy and J, E. Littlewood.

It follows that the function ,J(s; 0) is meromorphic, that its poles are
simple, and iha,t they lie at some or all of the points given by

(9.1) or s= ++EC,
1% -ii-

* where 2 = 0, 1,2, . . . . and Y runs through all even or all odd values,
according as E is odd or even.
Our analysis may be extended to the case of ,a general quadratic 8,
but, in view of the existence of HECK&S alternative method, we shall
not attempt to carry out the details of the work. The poles of J(s, 0)
lie, in the general case, at some or all of a series of points on the lines
d = 0,. - 2, - 4, . . , at intervals
l l

27ri
1 ’
logo *

CUP, w-b . I . . , ++*-I) being the periodic part of the continued fraction
for 8 It is easy, when 8 -+
l - l/o, to define 0 in terms of the solutions of
the PELLIAN equation Rg- D.ya = 1 (the unities of the corpus K(1/D)).
We conclude by one remark as to the relation between our
formula (5.6) a.nd HECKE’S formulal’). HECKE’S formula involves series
of the iype

sin i ,s+,(l--a)+-
I 1 ni 1 ( 11 1
7min ,sf,(l-a)y-ni 76’
2logq 2lw 1

It is not necessary to explain the meaning* of all the symbols. The

essential difference between his formula and ours is this, that his
collects into one term all poles on a liiie parallel to the real axis, and
ours all on a line parallel to the imaginary= axis. We have not been
able to find any simple method of passing from one formula to the other.

13) XECKE, 1. C. 63.

208
 COMMENTS

There was a previous brief communication to the London Mathematical Society at

its meeting on 9 February 1922 (see BOG. 21 (1923), xv-xvi).
The essential difference between the method of this paper and that of Hecke is
that Hecke’s work is based on the properties of the zeta-functions corresponding to
his ‘Grijssencharactere’ for a real quadratic field, whereas Hardy and Littlewood use
the periodicity of the continued fraction for a quadratic irrational. The relationship
between the two resulting expressions for the series (1. I>, each of which exhibits
the analytic character of the function, is described at the end of the paper. It seems
to be still the case that no simple method has been given of passing from one ex-
pression to the other.

209
-
 TRANSACTIONS
OF THE

CAMBRIDGE

PHILOSOPHICAL SOCIETY

VOLUME XXII. No. XXVXI. pp. tm-534.

XXVII. SOME PROBLEMS OF DIOPHANTINE APPROXIMAl’ION: THE

ANALYTIC PROPERTIES OF CERTAIN DIRICIHLET’S SERIES
ASSOCIATED WITH THE DISTRIBUTION OF NUMBERS T0
MODULUS UNITY

BY

G. 131.HARDY, M.A.,
SAVILIAN PROFESSOR, UNIVERSITY OF OXFORD

AND

J. E, LITTLEWOOD, M.A,, H
TRINITY COLLEGE, CAMBRIDGE

CAMBRIDGE
AT THE UNIVERSITY PRESS

M.DCCCC.XXXII
 XXVII. Some problems of Diophantine approximation : T/M analytic proper&~
of certair~ Dirichlet’s series associated with the distribution of numbersto

By G. WI. HAR.DY, Trinity College, and J. E, LITTI~EWOOD, Trinity College.

[Received 6 June 1922.1

l*l, The series in question are

where s = Q+it, a, = a, (6) = {no) = n@- [d] - 9, . . . . . . . ,. . . . . . . . .(l*lll) l

0 is irrational, [x] is the integral part of x, and the summation (as always unless the contrary is
stated) extends over positive integral values of n. The general formula for the Lth function is

Ivhere +(x) is defined by

&2 (4 = Pm(x)
+ (- l)m-l&, &n+l(x)=P~+l(x), (O<X<1)...(1~~31)
9(X+r)=+(x) (V132) l *...*.,,...*.....,,..*****.*.***...

and the P’s are Bernoulli’s polynomials*.

The properties of these functions, which are very remarkable, are intimately bound up with
the pr&lem of the distribution of tlhe numbers n0 to nrlodulus 1-f.
1% The properties of the function F1 (s) have already been investigated by fXecke$ when 0
is a quadratic surd. Hecke supposes in particulur that 8 = I/D, where D is free from squared
factors and congruent to 2 or 3 to modulus 4. He shows that in this case F1 (s) is meromorphic,
and that its only possible singularities are simple poles at the points
2mi ,
s z-Q+--. ..*.,.**~.,,.......~......*..,,.*.***.. (1.21)
log7
where g=o,1,2,...;r=...,-1,0,1,2, l ... . . . . . ,,...*....*,,*,,,* (1*211)
and q is a particular unity of the corpus K (I/D). H is method rests upon the theory of the new
( Zeta-functions ’ which he has recently introduced into analysis, and there can be no doubt that
it is the best for the particular problem with which he is concerned.
It is none the less of interest to discuss the function for general values of 8, and by methods
as elementary as possible. When we do this, we find ourselves compelled to treat F1 (s) as the
+ We follow the notation of LindGf (Le calcul des re’sidus memoir) ‘, Hamburg M&h. Ahh., 1 (1922), 212-249.
et se8 application4 d la the’orie des fonction8, 32 et seq.). H. Weyl, ‘fiber die Gleichverteilung von Zahlen ,mod.
The definition of the functions for integral values of x is Eins’; Math. Ann., 77 (1916), 313-352.
immaterial. E, Hecke, 6&r analytische Funktionen und die Ver-
+ In regard to this problem see the following memoirs: teilung van Zahlen mod, Eins’, Hamburg Math. Abh., 1
G. EL Hardy and J. E. Littlewood, ‘Some problems of (1921), 64-76.
Diophantine approximation ’ : (1) Proceedings of the Fifth A. Ostrowski ; (1) ‘Bemerkungen zur Theorie der Dio-
Internation&! Congress of Mathematicians, Cambridge, 1912, phantischen Approximationen ‘,ibid.,77-98; (2, ‘ Zu meiner
1,223-229 ; (2) ‘The fractional part of &?‘, Acta M&h., 37 Note : Bemerkungen u,a.w.‘, ibid., 250-251.
(19141, 155-190; (3) ‘The lattice points of a right-angled H. Behnke, ‘aber die Verteilung von Irrafionalititen
triangle ‘, Proe, London Math. Sm (2), 20 (192X), 15-36 ; mod. l’, ibid,, 252-267.
(4) ‘The lattice points of a right-angled triangle (second $ I& supra.

1923, 4 (with J. E. Littlewood) Traction of the Cambridge

212 Philosophical Society, 22, 519-33.
520 MR HARDY AND MR LITTLEWOOD,
first of the sequence of functions Fk (8). We also find ourselves led to the following classification
of irrationals 8.
We suppose, as we may without 10s~ of generality, that 0 < 0 < 1, and we write

Il( 1.22)
where a,, G, . . . are the partial quotients in the expression of t7 as a simple continued fraction.
We tsaythat 61is 01 cEass X. if x is the lea& number such that
(649, I . . e,,,r+ye, + 0 1*11.1...1~.~~~.*1.*~.~....**.***..* (M.3)
for every positive C, or, what is the same thing, such that

for every positive e. If no such number exists, we say that 8 is of infinite

” class. A quadratic surd
is of class 0, and every algebraic number is of finite class.
Our principal results may be summarised as follows. In the-first place, Fk (s) is repEar for

(194)

in particular, F’(s) is regular for u > h/(1 + x). Thia we prove for k > 1 in 5 2, and for iJc= 1 in 5 3.
There are alternative proofs of this theorem. When E= 1, it may be derived from Theorem 2
of our memoir (Q), or from the sharper Theorem 5, due originally to Ostrowski; but the analysis
of 5 3 is necessary in any case for our further inv,estigations. When k:> I, it has been proved by
Behnke’, by means of the formulae of linear transformation of the Theta-functions, The proof
given here is a good dettl simpler.
If X > 0, the result just stated is final; for then G = ok is u m*nguhr th& for the f tm&h. We
prove this in 5 3. We have no doubt that the line is still singular when X =O, except whm $ is
quadratic, so that the case considered by Hecke is completely exceptional; but this we are unable
to prove.
In 5 4 we consider the question of the convergence or summability of the series (ill), and
show that the regions of convergence or summability are always as extensive as is consistent with
the analytic properties of the functions and the order of magnitude of the coefficients. Some
fheorems concerning convergence have been found already by Behnket. These are included in
oum, which assert the most that can be true.

k
a>flk=i--
1 -I-x’

We have1 4 2nt (x) = (- lY+’ 2(2mP Iz co8 2y7rx (m z 1) # . . . . . . . . . . ..(2*111)

(2wJ0rn v2*
2 (2m + 1v 2 sin 2mx (m z 0). ,.,.,,, .,(g*lq
4m+1(4 = (-- IF+’(2+?&+1 - u’?t&+l
It is therefore sufficient to show that the functions
gk(8) t= 2 ‘9, hk(8) = r] ‘k(;ne), l . . . ...*. . . . ...* l . ..*.(2*12)
l

where +k(x) = 2 “9 (k > I)§, l . . . ..*.......*.*... .( 2921)

are regular for Q > ~;t. We shall discuss only gk (s), observing that our argument remains valid
with a formal change throughout 9 of ln0 into - nB.
+ Behnke, LG., 265-266, t Bebnke, Lcm, 266. 2: Lindeliif, LG., 34, Q We write e (x) for eenQ, following Weyl.

213d
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION. 521

Suppose first that CT> 1. Then

, .~*~.*.*..~.*I~**..**.~..** * I.(2*13)

where x(v)= x (8,e,v)= ?1 eq ..I..***..**.*.*.**....,*.~~*~,

(2931)
This function is an integral function of 8, and its continuation all over the plane is given *by
r (1 -23)
x( V > = ~---_zTi
(- x)8-1 dx, ..*.*.***. ... ...,....(2*14)
where C is a loop enclosing the positive real axis in the clockwise direction, and passing inside
all the poles
X = xwx = ZTi(712-I- Ye) (nt= ... -1, 0, 1, l m.)
of the integrand. We write
x (4 = x (s,R 4 =x(s,u
r (1 - @ 9 G&)= l*j. . . . . . . . . . . . . . . . . ..(295j

2.2. There is one and only one of the numbers X~ whose modulus is less than 7r. We define
a number 6 = S(Y) as follows, If h is the x,~ of least modulus, and 1xrra 1Z irr, we take 6 = 4~.
If Iq& iv, we take S=j1r. We denote by CO the contour formed by the semicircle / x I= 6,
1arg (- a); s &, and the two lines R(x) z 0, 1I(g) I= 6. The distance of any point of CO from
the nearest pole is greater than tr. Hence, if we write
x0(V)= x0(5, 8, V) = & c lcg+zG(- x)8-’ dg, ...*.* .rn...... . ..(2*21)
l

I 0

we have X.(v)=U(/ ~e~xi~(-x)@-lj~dx~)=O(l), l ..~**..~.*.....**~., (2*22)

Gl
uniformly throughout any bounded domain T in the plane of s,
NOW x(v)=&(v) (I%aI~Q4 , a.,,...... .,a..#.a... ...(2+231)
and X(v) = X,(v) + (-x&l (1x,1 < 4~). ..... ...... ....(Z423~)
The series 2 WV)
-P-
is, by (29, uniformly convergent throughout Y’, and its sum is an analytic function regular
throughout T. It hllows, from (2-13),(2Ti), (2*231),and (Z-232), that G&) is regular in any
bounded domain throughout which the series

is uniformly convergent. This is certainly so if the series

PD 1
2
1 vkjii9p---“’
where 2 is the difference between vB and the integer nearest to v@,is uniformly convergent ;
and this series is, by Lemma 3 of our paper (4), uniformly convergent in any half-plane
k k
a21 --1++04---
l+h’
In other words, GE(S) is regular for the values of s specified in the theorem.
It follows from (2%) that gk (8) is ah regular, except perhaps at the poles s = 1, 2, 3, . . . of
I? (I- s)* Of these,s = 2, 3, .., are plainly not poles of gk(8). When s = I, X (Y), and therefore
Gk (s), vanishes. Thus gk (s) is regular also for s--- 1, whiih completes the proof of the theorem.

214
522 MR HARDY AND MP, LITTLEWOOD,

3% The method of 5 2 fails when E = 1, and more intricate analysis is necessary.

LemmczA. If B is irmtional aad positive,x Z 0, y = 6x, rind f(0) = g (0) = 0, the?1

z fcE4) (9 (N -Y (17’1
- l))+n:yg ([;I, (f (n> - f 02 - 1)) =f(CYI) 9 @I)*
msX =.
. . . . . . . ..(3*11)
This is Lenma 7 of our paper (4).

e-BT Qo -mc u E: e-it1 a

then Xe (em -l)+---- s emncl (f&l1 -I)= W; . . . . . . . . . . . . . ..(3*13)
1 - ee5 1 1 -e-t1 1
,441 e-&t e-c e-e e - Cl
where W=W(c, 6, f)= - -e-E)(l-e-~l) - --- ------ - ---- . .(3+131) --. l

(1. l-e-5 l-eBC l-e-411-~-~1

In (3’11) take f(u)=&em@, g(fti)=l -e+@l,
where c = BE + & > 0, and make x -+ crsl
a We obtain

Substituting for [mO] and [n/O] ln

’ t erms of h and &, and making some simple reductions, we
obtain (3.13).
Taking the limit of (3m13) as % -0, we obtain

-, c m
e 3
Z a meBnlc+ a1 e gnc - l)e-‘lC~= w, . . . . . ,... 1-.*. . . . . . . . .(3.14)
l

1 l-e+

where

for all positive values of c, where

qJ= v (q 0) = - Be+ --L’ -1 -- e+ 1 __..--
e+l
----7- ..**.** ..rn..... . ..(3-151)
(1 -e ) 2 leemC-c l-eecl’
The left-hand side of (3%) is 0 (ebc)= o(ce+) if c Z 1, We may therefore suppose c c I.
In (3m14) we may write
e-f@ 1 c
-=--~~+O(d), *...*....,.*..I..........~*...... (396)
1-p c
e/GC- 1 = &C + &&c”+ 0 (c”j. .~..I.*..~~~~*~*..*~* ..,, l * (3.17)
Since c1 > c and I& ) < 1, we have
cP&QJ emncl z 0(&1,-C)
for aI positive integral values ofp and q. Hence the left-hand side of (3%) takes the form

215
 SOME PROBLEMS OF DIOPWANTINE APPROXIMATION. 523

Also, by (316),

+ 0 (ce+)

=-1 - -------
evC1 - 2;
- S$W + O(ceec)...... l . ,.... ..(3*19)

C 1 -edC1 1

Hence (3.14) takes the form (3%).

3*2. Lemma E. If 6 is sujkientky large,

The equation (3%) follows at once from (3-151) by dir& integration. It may be verified
at once that the right-hand side is regular at its only possible singularities, viz. s =I and s-2,

343. In what follows we denote by ]D (a) a finite domain in the plane of s, all of whose points
satisfy Q Z a + 6 > Q[; and by R (s, a) a function regular and bounded in D (a). It is to be under-
stood that the upper bound of such a function depends upon the form of D, and in particular
upon 6, but not up012 8, and that the 0’s which we use are also uniform lvith respect to 0.

Lemma F. If $ (c, 0) = 0 (cqeUC),

whereq z 0, then

-!I- f om* (c, 0) CS-~~C= R (s, - yj. . rnrn. rn. . . . . .. . r. . . ..(3’31)

x (% e) = rl($), l l

For the integral is uniformly oonvergent in D (- q).

LemmaG. The function
Fl (s, 8) + 08E: (8, 19,)+ &d+‘Fg (s + 1, 8,) . . . . . . . . . . . l . . . . . . . . . . . .(3#32)
is regular for Q> - 1.
Supposing first 6 sufficiently large, multiply (3%5) by c”+/~?(s), and integrate from c = U to
c = a0 . The result then follows immediately from Lemmas E and F. We obtain in fact

Fl(s,0)+88F& 8,)+$sWF&+1, e,)=6~(s-1)-~~(~)-~~-~‘(“,~)+R(s,-l),

s-
l . . l .4v33)

and 2, (s) = 21 (s, 0) = 49& - 1) - g (8) - BTzG . l 1 l 1 l . l l . . . . . l l l .(3*34)

is an integral function.
As a corollary, we have

THEQREM2. The function & (s, 0) + tPFl (s, 0,) is ~eyularfor Q> 0,
For EJ (s + I, 8,) is plainly a function R (s, 0).

3’4. We have, from (3*33),

Fl (s, 8) + es& (8, 8,) = zl(s, 0) + R (8, 0). . . . . . . . . .. . . . . ..*....(3~41)
Similarly fi (8, 4) + 4% (8, 82) = 4 ($1 6) + R (8, 0) , ..*.** . . . . . . . . . . . . . ..(3~41~)
and so on generally, From the first ~2such equationswe deduce
F&, 8)+(-y-qee,... 8,,$F&, en)=Q,+*n, l ms.,..m...ms..
(3.42)

216
524 MR HARDY AND MR LITTLEWOOD,

fl+1(-
Y ly(86, ... IY~-~)BR(~,
Yfi= O)*. . ..a. .. . . . . . . ..w.. . . . . . (3.422) l

v=o
We suppose for the moment that Q > 2.
Then i Fl (4 &J I < A
where A is independent of ~2,and s, and the second terln on the left-hand side of (3.42) tends to
zero. Similarly the functions @, and V, tend to

4 (s) = 5 (- 1)’ (00, l . . By-$ &(s, 0 y ) ,*..**1..1..*.*...**.*.., (3’431)

v=Q
Y(s)=
z (-l)~(8e~...e”1~)6:R(S,u), .,.,+..,....a . . . . . . . . . ..(3*432)
v=o
respectively ; and Fl (s, 0) = 4 (s) + y (s) ,.*....*~.*..~,*..*...~~~~*~...*~~..~ (3m44)
for a’> 2. This relation between analytic functions holds throughout any region in which each
of them is regular. The function V (8) is plainly regular for Q > 0, since 19~&+~ < &, We thus
obtain

The study of the singularities of FI (s, O), f or o > 0, is thus reduced to that of the singularities
of 4 (s) in the same region.
3.5. THEOREM 4. If 8 is uf clctss X, then each oj the functions
E: (4 6, 4 (8)
is regdar fur

We observe first that 21 (s) = 0 (P-l) + 0 (I), ..,1.1.*.**,*...*..*.,..**,**.**, (3.51)

uniformly in 0, on any closed curve C lvhich does not pass through either of the points s = i or
s = 2, The series for @ (s) is thus the sum of two series, of the types
r, 0 (00, ..* By& r, o(ee, l *. ov-J@8,u-1

respectively, The first series is uniformly convergent on C if C lies in any half-plane G 2 6 > 0.
The second is convergent if
a+(a-l)X>U,
i.e, if Q > Q,, and is uniformly convergent on Cy if C lies in any half-plane c z C, + 6 > cl. It
follows that @ (s), and therefore FI (s, 01, is regular inside any curve C subject to these conditions,
and therefore for 0 > q.
It remains to show that, when X > 0, c = c1 is a singular line, and it is plainly enough, after
what precedes, to show that the line is singular for

or for X (8). We choose 6 > 0, and divide the V’S into two classes v’, v”, writing v = v’ if
0, < (SO, l l * 0,-y .**I~**~....~**.,****~~*~,**...**..,,.. (3-53)
* R (8, 0) is of course a diflerent function in different terms of this series.

217
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION, 525

and v- v” in the contrary case. In virtue of the definition of X, there are, for every 6, an infinity
of v%
We write x (8) = 2 = s + z = X’ (8) + X” (8). ..,**....*..~~,..*...~**.*. ( 3m54)
v v’ vrr
The series fur X” is absolutely convergent if 0 + (C - 1) (X - 6) > 0 or
x-6
Q>l+h-6’
and the number on the right-hand side is less than ul. Hence X” is regular across the line
u = cl. It is therefore sufficient to prove the line singular for X’.
Suppose that the values of v’ are v,, v,, . . . , vk, .,. , and write
e+= Ml . . . 0,. (3%) l *..*.**~...*.~*11...*...*..*~..*******

Then the series for X’ (s), viz.

is a Dirichlet’s series of the type 2 ale - ‘k3, and

1 1
kk=+, - Xk = -2 logg-- > (x - 6) log &&- e m
log8 ++I 0 9+2 “* 0 vk+l 1 l **

%+1 yk+l- 1

when k-00. It follows, by a theorem of Wennberg *, that the line c = cl is singular for X’,
which completes the proof of the theorem.
We have supposed 0 c h < 00, When x = ~10 the result is stili valid, u = 1 being a singular
line ; and only trivial modifications are needed in the proof. The case h = 0 is much more
difficult, It appears to be true that c = 0 is then a singular line, except in the special case in
which 8 is a quadratic surd ; but we are unable to prove this rigorously. The exceptional case is
that studied by Hecke.
3.6. Suppose in particular that 19is a quadratic surd, The continued fraction for 8 is then
periodic, and we have
0 r+kwa =& (rzp, k=l, 2, . ..).
if p is the number of non-repeated B’s and m the length of the period.
In this case F2 (s, 0) is, by Theorem 1, regular for 0 > - 1. It follows from Lemma G that
each of the functions
FJS, 8)+(-l)~--x(~B,...~p-~~F~(s, Q,
Fl(S, ep> + (- y-l (e,&+l ’ l l &+?n-A8 Fl(& &+m)

is regular for Q > - 1. But the last function is

(1 + (- 1y-l OS) E: (8, 0,).
It follows that Fl (s, BP), and therefore PI (8, a), is regular for c > - 1, except possibly where
1 + (- ry’ @Is= 0,
at which points it may have simple poles. These point8 are the points

where k is an arbitrary odd or arbitrary even integer, according as m is odd or even.

* Wennberg, (Zur Theo&
der Diriohbt’schen Reihen ‘, An+l-Xs>A, X,jnea:.
Inmgural dissertation, Ups&t,1920,
3-7. It hss been See F. Carlson and E. Landau, ‘Neuer Beweia und Verall-
ahown by Carbon and Landau that the remit is true under gsmeineruqp des Fabrphen Liickeasatze~ ‘, Gattinger
the more general oonditions NachTichtm, 1921, 184-188.

218
526 MR HARDY AND MR LITTLEWOOD,

3% There appear to be no doubt of the truth of the following propsi tions :

(ak) Fk(s) ti rep&h for u > uk;;
(W u = uk is a singular line for & (8) whenever X > 0 ;

merowmphic when 6 is quadratic; it8 p&s are they are &uated

adoubly in&de system of points distributed at along th4 lims
a-l-k+ (p=O, 1, .*.);
(ek) &(8, 0) + (- I)k-1 &fk+ E;i (8, 6,) is regzrlar for Q > gk+l - 1; and d = @k+l - 1 is a

and a complete theory of the functions would contain proofs of these propositions in full generality.
Of these propositions we have proved (ak), in 5 2 when k > 1 and in 5 39 when k = 1.
We are unable to prove (ck, in any case. The case in which 8 is quadratic is doubtless best
treated by the deeper methods of Hecke. We have however shown, in 5 3.6, that our method
will accomplish something in the direction indicated by (dk).
There remain the propositions (bk) and (ek), of which, at present, we have proved (b 1) only.
We proceed now to the general proof, The particular case contains most of the leading ideas,
and Eve have cbndensed the general argument wherever the ground is familiar. In what follows
the A’s, O’s, and B (s, a)‘~ depend on X: in addition to the regions D; they are either independent
of 6, as in 5 3a3, or at any rate, when we have to consider a sequence of irrationals 8, &, &, . . . , of
the n in 19~~.

Lemma H, If k > I we ihave, througho,ut D (as),

IF~(s,8)I<AZv-kIvB/~-~--:~+A=AT+A~A~~-k~~$juk-~+~a+A,
This is a straightforward deduction from the results of 8 ZG, 2.2. By (2,231) and (2*232),
~~I,(s)i~~jx(v)h’-” r~(ix,(v)i+aixmlQ-‘)v-k+
Also /X,(V) 1<A, by (2422); and, since A 1YeI< 1x, I< A, we have
1x, y--l < A + A 1a p.
Hence ]t?k(~)I<Axv-kjv6jcr--l+A<AT+A. .*.*...**.*...** l ..**.rnrn(3-71)
Let D’ be the domain obtained by removing from D circles Cl, Cz, . . of radius $8 surrounding
l

such poles 1, 2, , . , of F (1 - S) as fall in D. Then

Ig&)i=/r(l-$)I lGk(S)i<A Z,V-~I&JI~++A<AT+A ... ...... ...(?72)
in D’. On Cl, 1 - 48 s Q s I + 48, and it is easily deduced that
1gk (s)I< A 2 v-~ 1a 1-as+ A < AT + A . . . . , l . l rn l . l . l . . . . . . . . l .(393)
on Cl. The middle term here is independent of G, and glr, (s) is regular for s = 1, so that the
inequalities (3.73) are valid also inside QI. Similarly it may be shown fhat 1gk (s) / < A5? + A
throughout Cz, C3, ..*, and so, by (3q12), throughout D. A similar argument, may be applied fo
?&k(s), and the lemma follows, since Fk (s) is a linear combination of the two functions.

* The important cake is Q < 0. When u2S the second term may be absorbed in the first; the proof will be if
U-C 0 is thought of as the standard case.

219
 SOME PROBLEMS OF DIOPI-TANTINE APPROXIMATION. 527

The left-hand side is less than

A 2 V-~-T ~YII,~“+‘-‘-ts+A<Ar:,-k-l~~n~u-~s+A,
byLemmaH. Lete=~~/(~+1)<6,and~~=X+1+4. Then

where t, = (08, l I l &-Jh-’ = (e . . I &-$+Y l ..***..*.11.*1.1~*...*..** (3%)

If now P,J&nr is the lntth convergent of the continued fraction for 8,,, (394) implies that
Qm< Acl QL1, by Lemma 2 of our paper (4), and therefore that
----
1ue,, 1 > At,v-h.
Hence
1Fk+&++r, &)I < s,V-~-I[A +(A~,,v-~)=-~~)< A + At,~“-Q+-l-h IT-H s A + A&p-+&2 p-k-l-h(uk+l+&
v V

The index of v is’

80 that the series last written is convergent. Hence

IFk+,(s+r, &,)I <A+A(o... &
- l) (*-N (A+@) < A + A (0 . . l @,,) ~-34 iA+?

the result of the lenlma,

3%. We return now to the identity (3*13), and we equilte the coefficients of tk-l/k! in the
Laurent expansions of the two sides. If we define +&) to be unity, then

Hence the coefficient in the first term on the left is

~f+k(crm+&-+k(&)] e-mc=~~k(me)e-‘“-~k(~)~e-mc. . . . . . . . . . . ..(3*81)
The coefficient of Ek$c! in the second term is
k (- 19jk--l2 zc,,eanCI, l . ..*..*.*.**..~..*.......*~...*~~.* (3-82)
d k-1 e-iC
where u,, = IL,, (c) =
0 ( dc --- ---- (eS””
L-e+

Now

say; and it is easily verified that

( a (x) j < A (x > 0). I.,*~..*...**~..~.**.~.*.~~~,~ (3*84)
Summing up from (3m81), (3*82), (3*83), and (3*84), we find that the coefficient of fk-l/k! in the
left-hand side of (3l3) is

220
528 MR HARDY AND MR LITTLEWOOD,

We consider next the co&Gent of Ek-‘/kI in W, the right-hand side of (3*13). Now W is
regular at 5 = 0, and, expanding formally, we have

Collecting the coefficientof Ekal/k!, a.nd equating it tb (8*85), we obtain

Ik+1
d+k (m0) e-MC + (- O)k-l 2 $k (d,) ewncl + C (-
r-1

= 0 (8”) @A + K(c), . . . . . .(3?361)

where

Vk = i (", t&(O) (- @jk" (;)- (-'-) -~{(.Ky2!2)-~~~~}.

r=O r c ec - 1
. . . . . . . ..(3-$62)
1 1 eac
In (3*$62) WC associate a term - - with p and -p , perform some trivial rearrangements
c ec - 1 ec - 1
and reductions *, and obtain the alternative expressions

+ (- l)kkP1
er,_l {(;j”-‘(e& -5) -z,e;/ . . .. . . . ..(3%63)

Or v. = vk, 1+ vk,s + 0 (Ck+‘) e& t .~...**......~~~,~.*~*..., (3.864)

where VQ and &,e are the first and second ferns 0x1 the right-hand side of (3*863)$
We now multiply (3*861) by cB+/r (s), integrate from c = 0 to c = 00 , and obtain
k+l k
Fk (s, 0) + (- l)k-lt?k-l+s Fk (s, 8,) + (- l)k-l r: t?k-l+s+r -- ’ @+ “) Fk+r (s+ cr,8,)
r=l k+r r(S)r(r+ 1)

= $9
fii -J- fa {Vk,l (c) + Vk,,<C)] F-'dc
-k-l)+r(s). 0

= R (s, -k:-l)+Z&, 0), .*~.,*...*,.*.*.....*~**.=.......*~..~.~*.*,, (3*871)

say. By (3.864) and (3*862),
vk L+ vk z= vk + 0 (ck+2e-c) = 0 (Ckf2eaC) (c > 1).

Further, the formulae which ;efine’V k,l and Vk,2show that these functions tend to limits w
c--p 0. It follows that
m( vk,l + vk,2) @--l&
I0
exists, and defines a function of s regular for Q > 0.
Next, we observe that
&j C(~s-r(~l-~)~-ldO=~ris),

+ Observing in paMmaIm that the term for which T= k + I + The expression in curIy braokets in the third term ie
vanishm, since &+I (4) =O. 0 (P).

221
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION. 520

exist and d&e function6 regular for 0 < d < 1. Hence

Zk (9, 8) = Fz A,Ok-vr (8) + P’z (8) ,~*.**.~.*,*.......**~*~.~.~ (3g88)

f=O
this equation giving, moreover, the analytic continuation of & wherever qr and z are regular.
Now + and z do not contain 8; they belong to a well-known ~1~s of integrals; and it may be
shown that they are regular everywhere, except possibly for aimple @es at certain pmitive
integral values of 8*. This being so, and if we suppose positive integral values of 8 to be excluded
from D (- k - 1) by circles of radius 8, we have

;: &$k-r+(s)=R(e, -k-l);
r=O
for only positive powers of 8 occur on the left-hand side, Hence, from(3*871),
Fk (s, 8) + (- fjk-’ dk--l+’ Fk (s, 8,) = 22 (8, -k- 1) + Wz(s)+ &@, &), . . . . ..(3%9)

where

319. We are now in a position to prove that O= uk is a barrier for & (8, 8) when x > 0. The
case h = XI is comparatively trivial, and we suppose that 0 < h < 30 Then l

ok > 1 - k, gk > bk+v

We write 0,~, for 19 in (3*89), multiply by (- I)kn (190, . , . O,+2)k-1t8, and sum as in 5 3%. We then
apply our former argument, lvhich shows on the one hand that
. , &-Jk-1+8 R (s, - I;;- 1)
22 (- l)k’a (e l

is regular fur Q > I - k (except possibly for certain positive integral values of s), and therefore
regular across d = Q; and on the other that
2 (- 1p (0 . . . t9,1-Jk-1+6e;I: z (8)
has CT= gk for a barrier. To complete the proof of (bk) it is sufficient to show that
IS (- l)k?a (0 , . a i9,,-2)k-1+x Q (s, 0,,)
is regular across Q = flk ; and this is true provided that, for some Q’ c ok, the series
X (0 . . , 0nw2)k-1+= 1Q (s, 0 ) i yb I*.*m+.m.**n . . . . . l . ..**.*...*..* . . . . +
(3.91)

ia unifornlly convergent iii the part of D (ak+l) for which c 2 cl, xow every term in (3~$91)
contains 0 to the power k - 1-t c at least ; hence, by Lemma K,
(Q(s, &)( < As,ky;+“{l+ (8...~~1)(a-~S)(~+~)j,
and the general tern- in (3#91) is less than
A (0 I l l e,&-Jk-- + A (0 . l l &-p+J’+ @‘-So (*+*)*

The first index is positive if ok - U’ is small enough, since flk > 3 - k ; and the second i6
k--1+Uk+U&f~)-(Uk- u’)(l+~+6)-~~(~+~)=h+~~k--(a~--‘)(1+~+~)--~~(X+~),
and is positive if 6 and ak - u’ are small enough. This establishes the uniform convergence of
the series (3#91), and SO finally the general result (bk).
There remains(ek); and for the proof of this the material we havs already ii3 sufficient.
We observe that
q+1- 1 s - k-l, Q-r s Qk+l (r 2 1).

* he for example A. Hur witz, ‘ Ueber die Anweudung functionen theoretischen Prinoipm auf gemisse bestimmte
Integrale’, Math. Ann&n, 53 (1900), 220-224.

222
530 Ma HARDY AND MR LITTLEWOOD,

From these facts, and frown (3.871) and (3*88), it follows that the only posiible singularities of
Gk (s) = Fk (s, 8) + (- l)k--i e-t”& (s, e,),
in u> ak+l- i, are the singularities of Zk in this region. These can occur only for positive
integral values of 8. On the other hand (3%71) shows that (when x < 30 ) * Zk is regular in Q z c’,
where P’ < I+ Hence Zk and Gk are regular in c > dkfl - 1.
On the other hand; if )L > 0,
crk;+l- 1 > - k,
nnd ck+r is a strictly decreasing function of r. It follows from (8%71) that Ok and

(- l)“e”+k & s Fk+1 (s + 1, 8,)

are equi-singular in the region Q > Max (- E - 1, OIJE+~ - 1). Since Fksl (s -+ 1, 0,) has a barrier
Q = =k+l - 1 in this region, this line is also a barrier for Ok.
We have therefore proved
THEOREM 5. If X > 0, the Line d = gk is a &$&t?- h&e of Fk (8, 8).
THEOREM 6. If X > 0 and 8 = l/(a, + O,), thea tHe fwcticm
Ii;,@,8)~(- 1)k-188fLFk(S, 0,)
is regdar for c > ok+r- 1 ; med the line Q = ck+.a- 1 is a si~2gula7~ he of+ the fum%m
4* I, We conclude with a brief discussion of the problem of the convergence or summability
of the series &$k (~6) 121~~in the region of existence of the corresponding function Fk (s, 8). It
will be seen that our conclusions may be roughly expressed by saying’that whatever could be
true is true. A Dirichlet’s series cannot be summable outside its half-plane of regularity, and it
cannot be summable (C, r) unless its nth term is of the form o (927 : we shall show that our series
is summable (with least possible order) except when these restrictions apply.
THEOREM7. The seriesI&#&+P is cowwged ifd > ck, c >O; ccnd sumnzubLe (C,- fl+8),
for eve?yp&i& 8, if g > flk, c c 0.
The case x = 00 is trivial, since ok= 1, and we suppose that X < ~0. We may confine our’-
selves also to the case k > 1; for the result for k = 1 is an immediate deduction from the
formula

We shall in fact prove rather more than we have stated, when k > 1, viz. that the series is sum-
mable (C, - 6’ + 6), where C’ = Min (a, I).
When II-> 1, & (nQ) is of the form A (& (PZ@ & I/Q (- &)), where t,b,(~Ze) is the function
(~21). It is therefore sufficient for our purpose to show that the series
r3/& (nB)I 7P, S& (- n 0) 72-y
are summable (c, - Q’ + 6) for Q > ok. Further, it is enough to prove this for real values of 8,
and hence also, since qk (no) and qk (- n0) are then conjugate imaginaries, enough to prove it
for the first series. This we do by a series of lemmas,
+ The case X=a, ie a trivial deduction from (3*89). gularity at most a t3imple pole at 8 = 1, the residue being a
t Thus ZIc is an integra1 function of 8 when XC do, This rational function r (8) of 8, Since r (8) vanishes for every 6
conclusion may be extended aieo to the cafle X= 00 . For our for which 0 -K A c CO, it must be identically zero.
argument shows that, in any case, Xk can have as a sin- $ See OUT paper (Q), Theorem 2.

223
 SoME PROBLEMS OP DIUPHANTTNE APPROXIMATION. 531

42. Let p be a (large) positive integer, and let 0 < # < 1, - 1 < Q< /3 s a + 1 ;

Let A, be the contour L, + L1’ of the figure (in which the sense of description is indicated by
an arrow j, A, be L, + Lk, and A be A, + A,. Finally, let C be the indented rectangle
L,+ L, +M,+N~+N, + Mp
In what follows A’s denote positive constants depending only on E, a, and P, and the O’s have
a corresponding meaning.
2;ennnaL : )S ($) 1< A @-(a+Y
This is a particular case of a known result. In fact the function f(Z) defined, for j z / < 1, by
the series SP,P, has z = 1 for its sole singularity, and
If(z)) < A 11 e zpa+l),
so that / S ( c$) 1 = j f (emi+) j < R 11 I ea+ j -(a+l) < A #--C+l),

This is a particular caseof a very general formula in the theory of reGdues*,

* See Lindeliif, IA, ch. v, $ 53.

224
532 MR HARDY AND MR LITTLEWOOD,

We may suppose that 0 < C#B

;( 8, and we begin by showing that

X(p, +)=J,,Lz+-;)B --g9~+o(T),*.m.***.* ..I. .... ....(4*31)

1 !4
2” - eb9i+ dz + 0 (T), 1..*..~*..***...*.I..*.~...*,. (4~32)
&+L, eati-1
where T = $r@+l) pa-@. In the first place, by the theorem of residues, we have

Now

and so

Hence the contributions of MI and & to the right&d side of (433) are of the form
0 (pa+1 e-49 ) = 0 (T) (p+)8+’ eMAp+ = 0 (T),
since fi + I > (1, so that (&,Y+l e -AM < A. Similarly the straight portions of I&, IV2 contribute

Lastly, the curved portions of 2v, and IV2 contribute 0 (~~-8) = 0 (T). Thus the contour C,
less L, -I- L2, contributes 0 (T), and (MI) is proved.
For X (4) we have
eezrie e-2&+
- p --- dx= P - - 1 dz=O(/m~ewA~+dy)
Ll’+& e2rwi
W) 1 e2Z+i - 1
&+L! i

I 0 e-APcPaoya e-AY* dy) = 0 (e-AP+@‘a-l) = 0 (e-49 (&)@-a T) = 0 (T).

( I0
This is (4m32).
From (4.31) and (4#32) we deduce

z!
NOW
Ii, 1
on L,+L,. Hence the straight portions of L, + L, contribute

0pya.L- AUPdll)=O(~~~u.(~~-,e-A~~d~)=O(2-8)~~~yBe-dy(d~=O(TX
(I 3 P
sincefl- a ~landy/psl. Finally the curved portions contribute

0 (‘) = 0 (/P-q = 0 (T).

P
This completes the proof of Lemma N.
4*4. We can now deduce that x .\I’k(~9) n-c is summable (C, - G’ + 8) for u > ~k;~ We may ’
suppose that Q < 1, since a convergent series, whose general term is 0 (‘]./~a), is summable
(C, - 1 + 6)*. Also Q c 1, since X < 00. Hence we may sll,ppse that ck < 6 = u’ < 1. This being
* 0. H. Hardy and J. E. Littlewood, ‘( Contributias to the tlrithmetio theory of pleriea ‘, Proc. Lodm Math. Sot. (Z),
11 (19121, 411-478 (462, Theorem 37).

225
 SOME PROBLEMS OF DIOPIYIANTINE APPROXIMATION. 533

so, we take OL=-d=-d>--1,/3= - Q’ + 8 = a + 8 > a. We shall furthk suppose, as we may,

that S<a-ukand 6~ J., so that p-a< 1.
Now

C g Y-v ((UB)) -#- ” u -k (S (p, (ue)) - 8((d))}, . . . I .(4D41)

l l l l

pr=l Y=

provided that one of the series on the right converges. But

l~(~,(ue))-s((;e))I=ls(Cc, ve)-8(a)l<A l;;ep”‘p+,
by Lemma N, Also, since- /3 = Q- 6 > t& the series
2 v-k 1VB p-l

is convergent, Hence z v-k] S(p, ye)-s(a)/

U=l
is convergent,and its sum tends to zero (like pa+) as p - ob . It now follows from (4m41) that
2 v-k s ((v0)) .~1~*,~....~.*~..*~*..~~.~.~.,,*~....*..~* (4*42)
converges, and that the series 2 qk (no) n-”
is summable (C, @), i.e. (C, - U’ + S), the sum being given by (4a42). This completes the proof of
Theorem 7.

COMMENTS
This paper is concerned with the convergence and analytic character of X{d}n-3, considered in
1923, 3 when 8 is a quadratic irrational, but now treated for general irrational 8. The problem
stated in the penultimate paragraph of 8 1.2 is one of the unsolved problems mentioned in the
Introduction.

226
 SOMEPROBLEMS
OF DIOPHANTINE APPROXIMATION: AN
ADDITIONAL NOTE ON THE TRIGONOMETRICAL SERIES
ASSOCIATED WITH THE ELLIPTIC THETA-FUNCTIONS.
BY

G. H. HARDY and J. E. LITTLEWOOD

New College Trinity College
OXFORD. CAMBRlDWL

I. In this nute we give &II alternative and more instructive proof of the
fundamental theorem un which our earlier researches in this field1 were based.
The theorem may be dated as follows.

Theorem A. Suppose that

(1. 4 O<XCI, 05851, Cd>1

and

(1.2) S(W)=S(W,x,8)=~~-"9"i"cos 212ne.

0&7i~S-,
The92
-pi1 ai8’
(I* 3) s(w, x, d)- ee x 8(X%, -;, i) -o(k)
v X

u?2iformly in CI) and 6 se

i G. H. EARDY and CT. E, LITTLEWOOD, ‘Some problems of Diopbantine Approximation’,

. Acta mdhemathz, 37 (IgIg), 193-238, and &X k&-i@ Phil. SW., 2X (r923)1 I---$
-- 1
’ That is to q, the abeolnte valne of the left hand side is leas than kc * 2, where A is
an tibeolu te consfant.

1925, 4 (with J. E. Littlewood) Ada; Mathemtka, 47, 189-98. 221

 190 %. H. Hardy and J. E. Littlewood.

earlier proof, which followed the classical lines of the calculus of

Our
residues, as exposed in Lindeliif’s book’, was fairly straightforward, but very long.
Two other proofs have been given recently by VAN DER CORPUT? The proof
which we give here proceeds on lines different from any of these, and seems to
us in some ways the most natural. It has also the advantage of being appli-
cable, in principle at any rate, to the Hum9 associated with any power of a
a theta-function, such as the sum

where Y(12)is the number of representations of n as a sum of two squares.

The proof which we give here owes very much of its comparative simplic-
ity to the criticism of Mr. A. E. INGHAM, to whom we submitted our original
version. In particular Mr. Ingham pointed out to us the usefulness of the
elementary identity (5. z), and we have rewritten the whole of @ 5-7 in accor-
dance with his suggestions.

2, We begin by showing that we may assume certain supplementary

hypotheses without prejudice fo the generality of the theorem.
In the firat place, since we are aiming at & resdt which holds uniformly
in 0, we may suppose that o <8< I . The result for 8-o or 8= I will then
follow by continuity.
Next, we may suppose that

(2. 1) ad-i

is excluded from each of the sets of intervals

' E. LTNDEL~F, Le cakul de8 rrhidt~, 1905.

’ J. G, VAN DEB CORPWT, ‘u&r Summen, die mit den elliptischen 9-Funktionen eusammen-
hiingen’, Math. Annalqn, 87 (I~zz), 66-77, and go (Igzj), I-X%. van der Corput proves a good
deal more than is aaaerted by the theorem, stnd. his proofs appear for this reasonto be more
elaborate than they are. The first proof is based on the theory of Fourier series, while the second
follows lines more like those of our original proof, on which it is (when reduced to its simplest
terms) a considerable improvement.

228
 Some problems of Diophantine Approximation. 191

where Q is an appropriate gositive constant. Here m = o, I , 2, . . . , and any

negative part of any interval is to be discarded as irrelevant.
To prove this, consider the interval I defined by

ogl!.!.<a<M+
[v1
I
X
=M+p.

and j’s, For each i is of length ---= and .is followed by a complementary in-
v X
terval k of length

There may be no complete k inside 1. In this case there is at most one (com-
plete or incomplete) i, and at most

of 1 is inside i’s, Otherwise 1 contains at least one complete k, and the number
of complete or incomplete 2:‘s does not exceed twice the number of complete k’s,
The ratio of the total length of the i’s to that of the k’s is accordingly less than

46
I -29

T
which is less than 8 d if
d’<- ; and then the length of the i’s is not greater
4
than 8 bp. A similar argument applies to the j’s, and the part of I, inside one
interval or another of the two systems, can in no case exceed 16 6~. Our

conclusion follows if S< $6 l 1

Suppose now that Theorem A has beea established for k’s excluded from
the i’s and j-k Any given value of ;1 lies in an 1, and there is therefore
a A subject to our restrictive conditions .and differing from A by less than p.
It is plain that, if we change A into A’, the alteration of the left hand side
of (I. 3) is

1 It would naturally be enBy fo improve on this number if it were necessary,

229
 192 G. ET. Hardv * and J. E. Littlewood.

o(k) + o(i-:,)o(‘)=o(~)~
so that the theorem, if true for A’, is true for ;1.

3. We write

(3 l I
1 f (8)-f (8,e)= I + 22 e-71”fi8~0~2tine,

where s=crt it, a>o. It is well-known1 that

We write also
Elt = T (?a=o), &)I =2 (17>0).

on writing s --ix for s and using (3.2).

We may invert the order of integration and summation; for, when t is large,

= o(m) = O.(lt I),

and

1 See for example

A’ E, LANDAU, H&dbuclb der Lehre van tie?* verteilmg dw &imza/ikn, 277.
2 See G. H. HARDY and M. RIPSZ, The general theory of Dirichlet’s se&a, 50 (Theorem 39).
We require the result of the theorem for absolutely convergent series only.

230
 &XIX problems of Diophantine Approxhation. 193

m5
s t-h&
is convergent. We have thus

(3 . 3)

where

If we suppose now that A=l/w is .non-integral, as plainly we may do without

loss of generality, and differentiate (3. 3) formally with respect to c;t), we obta’in

139 5)

where

(39 51)

This process is certainly legitimate if 2 Jn is uniformly conve?gent in a neigh-

bourhood of the particular value of ct) considered. That this is SO will appear
- incidentally in the sequel.

4. Our main idea is to approximate to J,) , in (3, 5), by the saddle-point

method or ‘method of steepest descents’.
The saddle-points of est8) are given byL

w ISEQ
or

sZ +;I --_I
-el -&‘N,
A

’ See G. N. WATSON, Theory of Bessel Bhctiom, 235, for it general account of the method.

231
 194 G, H. Hardy and J. E. Littlewood.,

say. The curve of ‘zero level’, given by

. % cP(s)=o, has t,he equation

o((T”3- t”--Ny==o,

and consists of the imaginary axis and the circle described un the line joining the
saddle-poipts as diameter. It divides the plane into four regions, the ‘low’ regions,
for which 5%m(s) < O, being the right hand inside and the left hand outside
regions,
We define the path C = CI + C2 by the lines CI and Cz from the point

s=LNto
Infinity
l
through the upper and lower saddle-points respectively. The
2

whole path lies in low ground, except at the saddle-points, and it cannot pass
through S=Y’X, since (owing to the restrictions of § 2)

for any value of ~11. ‘V6fe can deform the path of integration in (3, 5 I) into C,
l1

if we introduce the appropriate correction when this deformation involves crossing

a pole. This is so if and only if x > N, i. e. if 111-8 1 < Rx, and the correction
required is accordingly

which differs from

1
- h7ri ,z
-- i tP
e I 8
2- - r- e x s x269: --) -
I/X ( x x

We thus obtain

where XL differs from JTLin being taken along C; and the proof of Theorem A
is reduced to a proof that

232
 Solne problems of Diophantine Ayproxi~ation. 193

5. It is convenient to write

Lx
X c-=- X >o, l’=I~=wN>o,
AT 6

and to transform J,1-by writing LVs for s. We thus obtain

where the path of integration r= r;

-+ I‘, is C= CI + CL
d reduced in the ratio I : N,
I :
so that .it passes through the points --i, 2, z l

Since

(592) -z I i(8- .;) i(s+i) i(2+ 1)x

s-i s 2s(I +X) - 28(b---X) + s(s--iX)(I-X2)’

we may write (5, I) in the form

R
(5 3)
n -(h;,,1-K,l,2+~n);
5

where

(5. 34

X nY 8--f s2+ I
(5033) L 1L= e ( ) 3 ds.
2Z(I -X”) s (84X)2

6. The integrals K R,1 and K,&,2 are linear combinations of Bessel functions
I I
of orders 2 and- ;, and may accordingly be evaluated as elementary functions?

We have in fact

233
 196 G. H, Hardy and J. E. Littbwmd.

.
i g-237iY, z
5 ,-2.zii.z
K -
%l- 27c(IfX)VY 27t: n Ax+C’

i , G 2nil;‘
e2ntY=- z Se
K n, 2=
If --
2741 -X)VY 27x ;1 nx-5’

Thus the contribution fo 27, of these integrals is

when we combine the positive half of each series with fhe negative half of
the other.
Now 1
O” g--3niR(a-8e) e2titd (ix-@)
2niRe
2 iz+n-e=rre sin (Ax-4) d
-a

where

A=&[A]-’ 2’
and this is

in virtue of the restrictions of 8 2. The second series may be treated in the

Same way, so fhat the total contribution of Kn, 1 and K,l 12 is 0

7. It remains only to discuss the contribution of L,, , It appears at once

from a figure that, un either & 01”&,

(7 ) l I (8-ix1 >A(.l-xl,

I
sz >A,
I
Is*+ I 1 <AIgI

if --g&I and
2 2’

I See, for example, T. J. I’A. BROMWICH, hjhik 8ede8.

234
 Some problem8 of Diophsntiue Approximation. 197

(7. 2 1
I
if at-;, the A’s being absolute constants. Also

is negative except when IT-O, and

(7-3) I(s-+-Ao8 (-fd), +;)<A~ (o<-‘;).

It U~OWS from (7, I) - (7m 3) that

AX AX
IL4 < (I-x)e(I +X) io(e-A”d~~+(I-X)‘(I +X) wl~-4~~~~~
s s
-- 1
--03

AX AI/X AX
<(I-X)yI+xj (I-x)“Y+ 1. xIsy: d
I-

Thus the contribution of L,, is

In these series there are at most four terms in which the denominator is less
than unity, and the ccdribution of these terms is

in virfue of the restrictions on ;1. The contribution of t.he remaining terms is

obviously 0 (I). Thus the total contribution of Ln is 0

235
 198 G. H. Hardy and J. E. Littlewood.

This completes the proof of the theorem. It is only necessary to add one
word in justification of the assumption made provisiona.lly in 5 3, tlhat the series
CJ,, is uniformly convergent in a neighbourhood of any particular value of ctl
under consideration, The series has in fact been decomposed into a number of
parts, all of which have been proved uniformly convergent by direct estimation
of their terms, except those which were summed in 5 6. The uniform conver-
gence of these last series is classical.

COMMENTS
The proof of the approximate functional equation of the theta-function given here is much
simpler than the original proof given in 1914, 3. But it is not entirely self-contained9 in that, the
functional equation for O(z) is assumed. See the references to Mordell and to Wilton in the
comments on 1914, 3.

236
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION ::
A SERIES OF COSECANTS.
BY

G, H, HARDY (OXFORD), AXD 5. E. LITTLEWOOD (CAMBRIDGE)

[Read December 29, 2928.]

1, We consider here someproperties of the series

W) (S) z -L
n sin &r 1 w 2 && I

where 0 is irrational+ The analytical and arithmetical peculiarities of these

series resemble those of the series

s *e”afly z (d- [se]-,)*1

which have been discussed very thoroughly both by ourselves and by

others? The analysis connected with the series (14) is however in some
ways parCcularly simple, and, although many writers, ourselves included,
have considered similar series from time to time, there are some obvious
questiona concerning them which have remained unanswered,
We do not aim primarily at generality. The ca*se in which we shall be
interested particularly is that in which

esa+L 1 = d(cP+l),
2a+ 2ac+*I.

where a is an odd integer. It will however sometimes be obvious that our

arguments a.pply to wider classes of irrationala. In particular a good macy
of our results hold lfor the class 0 of irrationals whose continued fractions
have bounded co-efficienta. O-+L ‘-fiT, -P*-I;,.
T

1 [s] ie the integral part of s:.

s Eapcially Hecke, Mrowaki, asd ‘Bebnke, SW the list of papers at the end,

1930,3 (with J. E. Littlewood) Bdetlirrzof the Calczdtta

Matliemtical
20, 251-66.
So&et yt 237
252 G, H. HARDY AND J, E. LITTLEWOOD

We denote the general term of S by s,, and the Burn of its first TZterms
by S,, and use a similar notation in T and other letters. It is plain that
there is no universal upper bound for S II valid for all irrational 8 ; I]S,l will.
be, for appropriate 0 and rc, larger than any assigned function + (n). lt is
also plain that S cannot converge for any 8, srince ] sin n&r / <A/d for any
8 and an infinity of KL The first question which suggests itself is whether S,
is boutided for any 8, and, if so, whether S is then summable by Ceshro’s or
other means. Our first object is to answer this question by proving Theo-
rem I below.

Theorem I. There are quadratic B for .which S has the following proper&
ties :
(i) S osdhtca finitely ;
(ii) S is nut summabh by any CesW mean ;
(iii) S is summable by Riesa’s logarithmic means of any positive
order.

In particular ull this is true when

where a is an odd integer, and the (Riesxian) sum uf the series ir* then

The theorem shews that the behaviour of the series in these respects is
like that of
I-lsO+l+OtO+0-1$-o+..,

(where the ranks of the non-zero terms are 1, 2, 4, B,...). The most difficult
part of the proof is the proof of (ii, which we defer to $ 3. In 6 2 we prove
the rest of the theorem, taking (i) for granted where it is necessary.

Proof of Theorem I, (ii) and (iii>,

29. It is convenient to work in terms of the series,

(Z-11)

a Here and later A=A (0) & poeitive number depending only UD 0,
Qiue value ia immrtaxi81.
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION 258
(which differ onlg trivially from S and T, becoming S and T when 8 is
replaced by O+l). We shall also have to mnsider the series

(2-13) f (4 =f b10)=

is absoltitely convergent whest the red part 0 of sru+it ir su&iently Eargs ;

and

(2’14) O’y(s,e) +w’ #-I f(%;)

where[, is thedoublezeta-fmctionof Barnes.1

For the proof, seeHardy and Littlewood, 3 (Lemma p, p- 29). There ~0
consider the algebraic case. If 8 belongs to 0 (when of course it is not
generally a@braic) the result holds for u> 1. See Hardy and Littlewood,
Ik (Lemma 3, p. 216).
SupposeDOWin particular that

(2%) e- g-+&+& l a*
= qa”+l)--a,

where a is odd. We t&e

og=J(a9+l)-a, w%~~a*+l)+a,

= (2r)u[o(1-8, &&;d8 O8 d)
(2’16) f0 s ’as(s) sin $A cash {(s-l) log 0’)

239
 254 Or. H. HARDY AND 3. E. LfTTLEW001)

The zeta-function is regular all over the plane, except for simple poles
at s-0 and $=--I, and vanishes when s-2, 4, 6,.., .,* . . . . . . . 1 Bence f(s) is
meromorphio, with poles st s=O, S= -1, and

d+(Z++)“’
log w t

.where 2 ia an integers positive or negative. If now we write

(297) ds)=Y(% Q=fb+l, d+l)=z ,+l ,li, &$

?ii

Lemma 2. If e= J(a’ + I), where a Z’San odd irtteger, then g(s) z’s a
meromorphic funGtien of 8, regular except for pde6 at ss - 1, 8~ -2, id

(2M) (22+ l)xi

6=hi$(J(a8+1)-a)
We may remark in passing that the main result of the lemma, that
g(s) is meromorphio, holds for all quadratic 8. The proof ia in principle the
same &13 that which we have given in the special case, but it ia naturally
more elaborate and (except whefi, ati here, ~11 the partial quotients of the
continued fraction for &are even) it is necessary to treat the two fun&ion8

simultaneously. g

2% The Rieszian mean of )u,, of logarithmic type and order k, i8:

(221) uk(w)=w-k r; 2c,(w-log ny

log nga

l BeeBanlea, pp* 338, 340. That the zeta-function which occura here vanishe8
for Pa, 4*r..m*u follow8 from the regulwify of /(e) at thoere points, and may be vbfiad
directIg from Barnes’ contour in&@.

240
 SOME PROBLEMS OF DIOPHANTINE APPROXIMATION 255
When k=O and ew is an integer, U”(w) reduces to U,. If we suppose
that, aa here, the series g(+~~~n-’ is absolutely convergent for o>O, we
have 1
c+i-

(2’22) id 61 eW”ds.
-&l+k
s,

The zeta-function in (2ml6), and 80 g(s), in of finite order tin the sense of
Lindelijf and Bohr)’ in any ship q ~(t~q? It follows that, if k irJ a
sufficiently large integer, we may deform the path of integration inb the
line (-b-h, 4 +ioc ), where O< b< 1, if we introduce the appropriate
corrections for the residues. We thus obtain.

where R* is the residue of the integrand at the origin and R 1 a typical resi-
due at a pole (2.18).
The residue R* is the coefficient of 8’ in g(s) P’, which is a polynomial
of degree k in w whose leading term is

The series >R 1 is

where G, is the residue of g(s). Since g(s) is of finite order, this series con-
verges, absolutely, and uniformly in W, when I%is suffioientljr large, and is a
periodii: fucction of W, with period 4 log w. Finally the integral in (2*23) is,
for the same reason, the product of w-’ by an absolutely and uniformly
convergent integral. It follows that

IJOthat U is summable, to sum g(O), by Riese 23 means of sufficiently high order.

So far our argument demands no. unproved assumption. If we now assume the

k Hardy and Rieaz, 11, p* 50.

g Hardy and Rieaz, 11, p. 14.
3 Hardy and Littlewood, 3, p. 81.

241
 256 G, H.HARDY AND J.E. LITTLEWOOD

truth of clause (i) of Theorem 1, it follows, by the ‘ convexity theorem ’ for

Rieszian means ,1 that U is summltble, fo sum g(O), by means of any positive
order.
2’ 3. we have thus (subject to otxr provisional assumption) proved
clause (iii) of the theorem. To prove clause (ii) we consider R&z’s
’ arithmetic ’ means, known to be equivaren t to C&WO’s. B The arithmetic
mean of U, of order Ii, is s

($31) U’k’(zu)=~-kS ~,(w-fi)~,

nsw

c+i-
(2a32) U’k+~)= “‘z’i”’
77
s
c-i=
9( s) r(*+
w
lS+ k)
, w’ds,

if w>O, c>O. The characteristic difference between this formula and (2*22)
lies in the absence of the faotor w-k on the righthand side.
It is plain that we may transform (2 48B) as we transformed (2 22). l

But there will be. no factor wok multiplying the series which corresponds to
the series (2 n24) and which is, like that series, periodic in UL It, will follow

that U IW (w) oscillates finitely for sufficiently large k, when ZL’,->oo, and
this will prove (ii).
It is important to observe that we have proved incidentally, and without
assuming t#he truth of (if, that the arithmetic means (and therefore the
Ces$lro means) of sufficiently high order are bounded. As this observation
plays an essential part in the proof of (i), we state it formally in a lemma.

Lemma 3. The Oesiro means of U# of suficiently high order, are bounded,

3.1. We pass to the proof of (i). It now becomes necessary to take

account of the properties of the continued fraction for 8.

1 See Riesz, 16;. Particular c&Begof the theorem had been proved before by oumelvem
and other writers.
9 The only complefe proof of the equivalence is that given (after Rieszj by Hohson,
13, pp. 90-98.
8 Hardy and Riesa, 11, pl 23.
j Hardy and Riesx, 11, p. 51.

242
 SOME PROBICIEMB OF DIOPHANTINE APPROXIMATION 257
Throughout what fotlows 0 and o refer to the limit processn,->oo or
s->oo ; l the constants implied by the O’s depend on 8 (i.e. on a) only.
Also A =A 0) denotes generally a positive number depending only on 8,
c=c(e) a number between 0 and 1 depending only on 8.
W’e suppose that 8 is defined by (Z+ 15), and that

are the convergents to 8. Then p, =qoml t and pI and qI are of opposite

parity. We denote by 2~’ the complete quotient corresponding to the partial
quotient 2a, so that
2af=2a+8= J(d+i)+a,
and write
de+1 =2a’q, +ql-y
Then

(341) q8+1cAq8r qr+lMf+A)q8~ qf8+1w+A)q8+l.

It is familiar that

(342) C H-1- c 8=(-1)’ ,$-c =.(-1.>‘*

!I 8Qlfl 8 wta+r

If Y is an integer less than q,, we have r

for all integers a’. Also

(3.14) cosecv&=0(v)

for all v, and

(3 ’ 15) I cosecv&r 1 >Av

fur an infinity of v (-e.p. ~=9,).

34, We shall make repeated use of the following lemma, the form of
which was suggested to us by Mr. E. C. Titchmarsh.

l 8 ia tt positiveinteger ; the s of 42 does nut appear again.

a Qerrun, 14, p1m*

243
 958 G. H. HARDY AND J. E. LITTLEWOOD

Lemma 4. Suppose that 01, that v<q,+l, md that v isnot a mdtipk

OfQa- Then there is a Positive p(O) less thun 1 such that

(3.21) l-C<Isin-sinVCv&r,7r I <l-f-t:

for all v for which either v0 or VC, di’ers from an thteger by less then 3,
And for a21 v

(3*22) cdsec v&r= O( 1 cosec VC ,7r 1 )..

Write v=rq,+p, where r=[v/4,], so that O<p<q, if r<Zn aqd

o<p<qI-l if r=2a ; and write

VC, =4, +fp 9 ve=fp +f’,

where fp and I$ are integers and 1 fp 1 and [ f’, I do not exceed 1- .I Then
2

The l&hand side is e r hy@iesi numerically less than 3 and the ]a& term
on the righthand side k numerically less t.han 1 /qH < 2, It follows
that ta’ =$ and that

But 1 fp 1 21/q 8a Hence fp and ~ have the same. sign and

it result plainly equivalent to (3-21). Asregards (3*22), this follows from

(3b21) if either fyP or f’, is less thah Q, and is trivial if neither is less than i.

’ I f, 1 may be b* in whiob caee there is ambiguity ; we may agree then to take’fp

244
 SOME PROBLEMS OF DIOPBANTINE APPROXIMATXON 269

qr-1 1
313. Lemma 8. Z1 sin%C,w
=O(q? 1

When v varies from 1 to qI -1, ffi assumes, each once, the valuee

-+x/g, (A=&2,..m, As* q,)

the value +J, occurring, if at all, with one sign only. Hence

2 coseca&,w=O {q/ 1+ 2;' +-1_+.., )=O(q,*),

39 >
which proves the lemma.

Lemma 6 1:

Since q , + ,/q I is bounded, it is enough to prove this when ti=q,. The

term ti=q, contributes O(p,*). The remainder, by Lemmas 4 and 5, con-
tribute
O(S case@ ~&x)=O(q,~)

zP4. Lemma 7. If p and q are coprime integers of opp@te parity, and

B (p, q)=‘i’ (-1)’ cosec -VPr ,

Y=l 4

$Q+q-i 1 1
(3.42) 1-BCp, q)+L B(w)=~ (z=-l)~xa~-l) - ?& 0’
9 P w

1 1
dx.
-4s (a+l)P

It is easily verified that the integrand is bounded for 2~1 and s:- -1.
It followra from Cauchy’s Theorem that the value of the integral is 4i time8
the sum of the residues of the integrand at poles above the real axis. Cal@u.
lating these residues, we obtain (3042)~

Lemma 8, If C, =p I /q I is a convelrgent to 8, then

(3*43)

2 This leppna is not. actually used, and we include it because it ie interghing in ifealf,

245
 260 ct. H. HARDY AND J, E. LITTLEWOOD

We denote the integral in (3942) by J(p, 4). Making some obviona

% elementary transformationa we obtain
Qo
1 I 1 1
sinh w sinh +ta cash a (w/2q)
--I-
-JG

=- 1s0
29
1
sinhg(w/2q)

1
sinh w ljinh + WI-+? I -

1 1 + sinhg(w/2y)
1
4pq- cosh*(w/2q)

where +=p{ q. The integral here is

t-O0 Pw
1 dw
sinhnsinh t#d 3 cwh’ w

when q->oo, p/qb->8. It therefore follows from (3.42) that

when B->oo.
Now
qq,, P,)=B(2aq,,,+q,-,, qr-t)=B(qr-3, h-a)=B(~m qd

rnd so, from (344),

=0(1)90 ( * )+o ( 2 ) +r.rn*rr*

on repeating the argument. Since q, inareases more rapidly than a gee

metrical progressioll, this is 0 (Q.
 We may observe that the results of Lemmas 6 and 8 are in frtct true
for any 8 of 0 Etnd in particular any quadratic 8, Only slight ahanges rare
required in the proof of Lemma, 6, or in that of Lemma 8 when zlll the partial
quotients of 8 are even. But in the general case the proof of Lemma 8
becomesmore complicated became (&Bin the extension of Lemma 2, referred
to at the end of .§ 2.1) it is then nsoessrtryto consider simultaneously Burn8
of two slight,lg different types,

3.5. bmma9. go a;nw

VqB (e)= iE (-1y =QW
1

The contribution of the term v=q, is O(p,*), If O<u< q,, we hsve

8in &r-sin vOr9=0 sin $ I+-0,)

(I I)=O (3

by Lemmaa 4 and 5. The result now follows from Lemma 8.

L0mm810, V.C~)&--k$&
?
=0(n).

This is tbe principal lemma. It has been proved (Lemma 9) whets

n=q,. If Q*<rr<q,,,, w0 u&n write

fi=bp, -i-n,,

since then, repeating the argument, we obtain

v,=~1q,)+o(q,-~)+...=o~q,)=O(~):.

247
 262 & a. HARDY AND J. E, LITTLEWOO~

we writ0

say. We first consider V,, r. We can omit the term v=(r+l)p I with error
Wq,) : in the remaining terms v-q, + p, where O< p <q I For such v we l

have

co880ve?r- co8ecVCI T = 0 sin

VT
~050~verr,cosec &,A
(I %?&i+, ’ I>

by Lemma 4. Hence

+?, +p.) %--l -( qp

+(-1) >; - =O(qA
p=l sin pCp

by Lemmas 5 and 8. If follows from @TL4 and (3453) that

(3’54) v,=o(q,)+v’,l l

In tho last term on the right we have v=bq, +p, where ()<pg,~ and

248
 PROBLEMS
OF DIOPHANTINE APPROXIMATION 263
by Lemma 4. Hence

by Lemma 5, From (3 *54) and (3’55) we deduce (3*51) and so the lemma.

.3m6. Lemma 11. If some Ceshro mean of U is bounded, and

y+2~, + . ..+ny.=O(fi).

then U, i8 bounded.
We writa for conveniencezc, = 0. Then

Hence the difference between the CesAromeansof U, of orders 0 and 1,

is bounded, If fallows tlhat the difference between the means of orders k
and h+ 1 is bounded, and this proves the Lemma.
3’7. We can now complete the proof of the theorem. By Lemma 3, the
Ces&ro means of U, of sufficiently high order, are bounded, By Lemma 10,
the secondhypothesis of Lemma 11 is satisfied, It follows from Lemma 11
that U oscillafes finitely.

Further Results.

4% We add a few remarks about the behaviour of our seriesfor general

quadratic 8 or general 8 of class 0, without attempting b justify all that
we say in detail.
The arguments of § 6 302-305 are extensible, without new difficulties of
principle, to any 8 of o ; the results

(4Il) v, =0(n), IV. =O(nB)

of Lemmas 10 and 6 hold for all such 8. When the quotients of 0 are al1
even, but little elaboration is needed; in the general case there is the
complication alluded to in $ 3% The results (411) are obviously ‘ best
possible’ for any &

249
264 G, EL HARDY AND J. Em SrITTLEWOOD

The situation in regard tS U is a little more complex. It is easy to prove

that

(4.12) U.=O(log 9%)

for till 6 of 0. For some 8, as we have seen, more is true (c, being
bounded), but 1432) is the most that is true even for all quadratic 8, For
quadratic 8, in fact, there are only two possibilities, tlk. U,&=0(I) and

($43) &=A log n + O(1) (A#W*

If is interesting to give an example of the secoCi case,

Recurring to theanalysis of $Zal, let

where a#b. Then

where
O=d( ba >(J(ab+l)- d(d)), 9,= 4
(
;-
>
(q&+1)- d(q)*

we obtain, by analysis sitiilar to that of $2,

The arithmetic mean of ‘D of order k is @en by (%32), with h(b) in play

of g(s). The difference is that there ie now a double pole at the origin.
Otherwise we may argue as in 5 2.3, and we tid that dk)(w) is, for suffi:
.
oietitly la;rge k, of the form

A log w + 0 (I),

J (b-u)r
A
= -12 log { .sr(d+1)- d@a>)
 SOME PROBLEMS OF DIOPHANTIMZ APPROXIMATION 265
It follows (substantially, as in the proof of Lemma 11) that U, itself icr
of khe smneform. If in particdar a ~2, kl, we find

k$I/6-11 A= I
12 log (A- 42) ’

We add in ooncluaion that

for all 8 of 0, a;nd that these results also are the be& possible of their
kind. I

BibEiog.mphicaZNote.

Our own relevant writings have been published in various journala

under the general title LSome Problems of Diophantine Approximation.’
We may ref0r in pdicular te

1. Some Problems of Diophantine Approximation : Proc. Fifth. ItA+

national Con gress (191 Z), 223-229 (a preliminary sketch).

2. The Trigonometrical Series associated with the Elliptic Theta-

functions : &ta, Math,, 37 I 191/A), 193-W [an additional note in PTOC,
Camb. Phil. Sm., 21 (1923), I-51.

3. The Lattice Points of a Right-angled Triangle: Proc. London Math,

rsoc. (2), 20 (1921), 15-36.

4. The Lattice Points of a Right-angled Triangle (seoond memoir) ;

Hamburg Math. Abhandlungm, 1 (1922), 212-249.

5. The Analytic Character of the Sum of A Dirichlet’s Series considered

by Eecke! ibid, 3 (1923), 57-68.

6, The Analytic Properties of Certain Dirichlet’s- Series associated

with the Distribution of Numbers to Modulus Unity: Trans. Camb, Phil.
sot., 22 (1923), m-533,

251
 266 0, H. II[ARDP AND J, E. LITTLEWOOD

The other bmks and memoirs referred to in the text w

7,E. W. Barnes : A Memoir on the ‘Double Gamma-function : Phil.

Truns. Royal SOL (A), 196 (Ml), 265-387.

8. H. Behnke : Ober die Verteilung van ImtioaalitiSten mod, 1:

Hamburg Math. Abhdluttgen, 1 (1922), 252-267.

9. H. Behnke: Zur Theorie der diophantisohen Appro ximcltionen :

ibid, 3 (1924), 261-318 and 4 (1926), 33-46.

PO. R. Cooper: The Behaviour of Certain Series associated with

limiting Cases of Elliptic Theta-functions : Proc. London Math. Soa., 27
(1928), 410-426.

11. G* H, Hardy and M. Riegz : The General Theory of Dirichlet’s

Series: Cumb. Math. Tracts, 18 (1915).

12. E. Hecke : CIber analytische Functionen und die Verteilung von

Zahlen mod. Eins: Hamburg Math. Abhandlungen, 1 (.l922), 34-76.

13. E. W. Hobson: The Theory of Functions of ti Real Variable,

Vol. 2, second edition, 1926.

14. 0. Perron : Die Lehre von den Kettenbrtichen, 1913.

15. M. Riesz : SUP WI thBor&me de let moyenne et sesapplications :

Acta Univ. Hungaricae, 1 (1923), 3-E

Furt.her references to the older literature connected with.series of the

type considered here will be found in a paper by Hardy, ‘ On Certairl Series
of Discontinuous Functions connected with the Modular Fuactions, ’ Quarterly
Journal, 36 (1905), 93-123,

CORRECTIONS
p. 254, footnote (2). Read: en’%P.
p. 258, Line 4 from below. Read: Hence fp and f; have . I .

COMMENTS
As mentioned in the Introduction, it would be desirable to have a simpler and more
direct proof of the results of this paper.

252
 NOTES ON THE THEORY OF SERIES (XXIV):
A CURIOUS POWER-SERIES

BY G. H. HARDY AND J. E. LITTLEWOOD

lifewived 13 November 1945

1. This note originates from a question put to us by Mr W. R. Dean concerning series

of the type
(1.1)
Xn
lza
msin 077sin 2077 .,. sin 724%’
where 0 is irrational *, Series of the type

(1 2)
l
Ca -2.Y

‘asin E&T
are familiar: it is well known, for example, that
Xn
WV c
sin non
may have any radius of convergence from 0 to 1 inclusive, according to the arithmetic
nature of -0. It is natural to ask how this radius is connected with that of
Xn
(14 c
sin Bmsin 287~.. . sin n0ti ’
and, more generally, how those of (14) and (1.2) are connected.
We show here that the radius of convergence of (14) is usually half that of (142).
We prove this by two methods each of which has points of interest. The first, which is
rather oddly indirect, depends on an algebraical identity, the second on arguments
of a type more usual in the theory of Diophantine approximation.
Actually we shall use not (1.3) but
Xn
( 1*5) c --
n sin nBn ’
which has obviously the same radius of convergence.

* See Proc. Cambridge PM. SW, 42 (1946)? 24. The actual series encountered by Dean was
cos 077 cos 2&T l . l cos n&T xn

c
sin877 sin28n...sin(n+ 1)&Z”

1946, 1 (with J, E. Littlewood) Proceedings of theCambridge Philo-

aophical Society, 42, 85-90. 253
 86 G. H. HARDY and J. l3. LITTLEWOOD
2. We begin by proving
THEOREM 1. If the diw of convergenceof (13) is p, where 0 sp 5 I, then thut of
(1.4) is Qp.
WeshowfistthatifIxI<l, Iplcl,and

(24 f(w) = +&)+..., (2-2) F(x,q) = 1+ 5+ x2

- l-q (1-q)(1-q2)+m**J
then
(2.3)
In fact

and so
1
(2. 5) ef(x,*) = (l-x)(1-qx)(&q2r$) ,., = G(x,q),
say. Also G(x, gx) = (1 -x) G(x, q) and, if
Q(x, a) = 1 +c,x+c2x2+ l **,

tjhen 1 +c1qx+c2q”x2+
. . . = (l-x) (1 +c,x+c,x”+ l .*).
Equating coefficients, we find that c,( 1 -qn) = cndl and so G(x, q) = F(x, q)*.
If (lG5) diverges for all x, then (1.4) plainly doesthe same.We may therefore suppose
p > 0; and we shall show that (2.3) is then still true when 1x 1c p and
q = q. = @ie,
so that 141 = 1.
The series f (x, q*) = &+) = ii c (,,-nio)n
- ?hsm ?t7To
has aIs0 radius p. If

then

where A is an absolute oonstantt. Sincef(x, qO)is uniformly convergent for

jxIsp-s<p,
it follows that f(x, q) is uniformly convergent for 1x f =<p - 6,O 2 5 5 1; and that
f (x3a)+f (x340)
.when &+ 1, uniformly for I x 15 p - 8. Hence also
F( x, q) = eR%
d+ ef@,!I&
with the same uniformity.

* The identity (2.3) is naturally not a new one, though we cannot refer to a quite explicit St&t@-
ment of it. The expression of ‘F(x, q) as an intiite product goes back to Euler (see for example
Macmahon, Combinatory an&y&, vol. 2,2), and the transformation in (2-4) is one of a type familiar
in the theory of elliptic functions.

t For /;--gz:~12= 4 sin9 fin0

(1-Y”)B+4psin2nn8’
which does not exceed 16 if p s & or 2. if 5” 2 8. Actually the best &ue of A is 2.

254
 Notes on the theory of series 87
It now follows from a familiar theorem of Weierstrass* that, if
F(x, q) = CC,(q) xn, ef(x*Qo) = lzC,(q,) xn,
then C,(q) + C,,(qJ when &-+ 1; so that
1 1
Cn(qo)= limC,(q) = lim
(1 -q) .*. (1-qy = (l-40) .*. (l-q,“)’
Thus the functions in (%3) remain regular, and the identity valid, for 1x 1<p, when
Q is replaced by qO. In particular, the radius of convergence of F(x, qO)is at least p,
and therefore that of (b4) is at least +p*
3. We have still to prove that the radius of convergense of P(z, aO)does not exceed
p. From this point on we drop the suffix in goand supposethat Q = e2nie.
If F(x, q) has a radius c greater than p, and p < 7 KC, then F(x, q) is regular, and has
at most a finite number of zeros, in. 1x I=<r; and hence

(34 fW)
= 5 PJog(a,-x)+w~
Z-1
where 1C+15 7, the p1are positive integers, and #(x) is regular for 1x 15 r. If a,, a,, .. ,, a,
are the a, with smallest modulus, say a, and
= we--ial (1 ggs, uscc,<2n, q+q if l*l’),
%

then f(x, q) = 5 p,log-(we-i”J-x) + @(x),

I=1
where ti(x) is regular for 1x ] s a, It follows on equating coefficients that
S
wn
-=_ nial + O(P),
CP ze
1-p
z-1

whereO&<l,forrz>l;andsothat

(34 = P+ZQ-+O(Sn),

where P= s;+o, Q = &~;cos~(a,-q+

Z=l 141'
It, is now easy to derive a contradiction. It is plain that (3.2) is impossible if w 2 I,
sincethen the right-hand side is bounded and the left is not. We may therefore suppose
m < 1. Summing (3~2) from n = 1 to n = N, and observing that cc,=+q, we obtain
Iv Wn
WV s N = c- - 4PN+O(1).
n.=l sin2ne7r -
Wn
Hence, first, - = 4P+O(l)<H,
sin2988~
say, for all n. Secondly, since0 is irrational, the numbers lze, taken mod 1, are uniformly
* See for example Harkness and Morley, Introduction to the theory of u&ytic functions, 134-6.
The theorem is stated there in the form in which the coefficients depend on an integral parameter
im which tends to infinity (‘ Weierstrass’s double series theorem ‘).
Alternatively, we mny appeal to the representation of C,(y) as an integral by means of Cauchy’s
theorem,

255
 88 G.H. HARDY and J.E. LITTLEWOOD
distributed in (0,l) *. Hence, if N is large, and 7 independent of n, the number of n
in (1, N) for which sin2n&r < 11is lessthan NC(r), where c(q) + 0 with 7. Thus

for small 7, in contradiction to (3m3).

It follows that the radius of convergence of QC, q) doesnot exceed p, and this com-
pletes the proof of Theorem 1. It will be observed that this is a theorem about all
irrational 8, whereasthose in what follows are theorems about almost all-f-.
4. The radius of (1.3) is 1 for almost all 0, in fact whenever
1
(44 0 = a*+-
a,+ l .*’
23,/q,is the rth convergent to 8, and
P-2) ar+l = qey
for every positive E, a condition satisfied by all but very abnormal 8. It follows that
the radius of convergence of. (14) is 8 for all such 8, so that
lim VI cosec7~0cosec2ne .. . cosecnd 1 = 2,

y-1 n
W) lim - C log 1cosecmd 1 Z log 2 1
n+anm=l
for all such 0.
On the other hand it is easy to show that
1 n
lim - C log 1cosecrn7r01zlog2
n-+ao ?bw=l

for all irrational 8. For, if we define a function fs(x), with period 1, as log cosec7rx in
(&I- 6) and 0 in the rest of (0,1), then fs(x) is Riemann integrable, and therefor@
1 n
,‘Fm ,my8w) =jIf&)dx =~~~dlOg~Ose~~~d~.

Hence lim: 5 log ] coscc7~3~~0

12 lim A- 5 f&(mO)
---nmzl n nz=l
l-6 1
= log cosecnxdx > log cosccnxdx - E = log 2 - C,
s s s 0
for any c and a corresponding S.
Combining our results, we obtain
THEOREM 2. If 0 satkftes (4m2), then
1 n
bw lim - C log 1cosecrnd 1 = log 2;
n+a, n n&=1
and this is true for almost all 8.
* By a well-known theorem of Bohl, Sierpinski and Weyl: see Weyl, Ma%, Annalen, 77 (1916),
313-52 (314), or Hardy and Wright, Introcluction to t7he theory of numbers, 378-81.
t Except Th oorem 5, which refers to a special class of 0.
$ By another theorem of Weyl, substantially equivalent to the theorem of uniform distribution
already quoted.

256
 Notes on the theory of series 89
It is also plain that, when (4.5) is true, the result of Theorem 1 may be extended to
the more general series (1.1) and (1.2).
THEOREM 3. The radius of convergence of (U 1) is half that C$ (14) for almost all 0.
5. It is natural to ask for a more direct proof of Theorem 2, and for more general
results of the type
1
W) Iim 1 $ f(w9) = /(x)~x,
n-+00wn=l s
where f(x) is a function with infinities at x = 0 and x = 1. We suppose throughout
what follows thatf(x) has period 1, so that

where X is the difference of x from the nearest integer, that it is Riemann integrable in
(&I - 6) for every positive 6, and that it increasessteadily to infinity when x + 0 and
x+1. We prove
THEOREM4. If
1
(5-2) f(x) log2I+ log2+x)
X - dx < co
f0 (
then (4m1) is true fur almost aZ1 0.
THEOREM 5. If the partial quotients a, of 0 are bounded, then (51) is true whenever
on the right is jinite.
the integral
We might replace the logarithmic factor in ($2) by

where jk is any positive increasing function for which

a dU
s u@(u) < O”*

We use two lemmas, in which N > 0, h 2 2, and the results are asserted for almost
all 8, viz.
(a) tk number of n for which
NsnsN+h, jnel<h-I,
is O(log2h), uniformly in N;
(b) the number of m for which
lsmsn, ]nBI<h-l,
is O(n log2h/h).
It is plain that (b) is a corollary of (a). To prove (a) we observe that, if it is false, there
are, for large h, numbers n, and n2 such that
h
Nsn,<n,<N+h, = p = n,-n,<-
log2 h ’

But then h>,ulog2hzplog2,x, j!8= 0

and p is large with h, since2 = O(h-I) is small, It is well known that this is false for
almost all O*.
* See for example, Hardy and Wright, 168.

257
 90 G. H. HARDY and J. E. LITTLEWOOD
If nowfs(x) is periodic, isf(x) in (&,I -S), and is 0 in the rest of (0, l), then

s = iif = G+f&W
ln +k+ tf b-4-h(me)>
= sl +-s2,
where c f (mt?) = 1 i T,,
n k-0 2-~-1~~~75j~<2-~s n k=O
say. When 8 is small, f (me) is positive throughout &.
The number of m in Tkkis, by (b),

and one or other of

=
f(m@<f(2-k-1S), f(~@~f(l-2-k-Q),
is true for each such ?n, according to the sign of mB. Hence

<H
= ‘f(x)log2 %Zx+ :J(z)log2 +$x),
-
0 X s
for an appropriate H depending only on 0. Thus
(54 O<S,<E
for 6 s a&) and sufficiently large n.
On the other hand
sl -+/If8(x) dx = [I-‘f (x) ax
when n+-oO, so that
1
(55) f( x v x,<e
s1- 1
I s0 ;
for 6 s &(e) and sufficiently large n; and it follows from (53)-( 55) that

is-~)(x)dx~ c2E
for sufficiently large n.
This proves Theorem 4. The proof of Theorem 5 is similar but simpler, since we can
omit the logarithmic factors throughout the argument.
6. We end with a remark about the actual seriesencountered by Dean. The effect
of the factor (%!)-I is naturally to make p = 0~)for almost all 8 (though it may still be
0 for sufficiently eccentric 8). If we omit this factor and consider
cosen cos2077... cos n&r
(W c sin 6%sin 2Bn . ..sin(n+l)h xn,

then the odd factor in the denominator does not affect the results. The effect of the
cosinesis to replace Jlog ] cosecTX 1dx by Slog 1cot nx 1dx, log 2 by 0, and + by I j so
that (6*1) will have the sameradius of convergence as (l-3).

TRINITY COLLEGE
CAMBRIDGE

258
 CORRECTION

13. 89. In the statement of Theorem 4, for (4.1) read (5.1).

COMMENTS

It is surprising that the problem raised should admit of such a simple answer l Apart from
that, the main interest of the PaPer lies in the variety of types of reasoning that are shown to
be relevant.
Theorem 4 is perhaps deserving of separate statement, being a useful supplement to a classical
result. If j(x) is period .ic with period 1, and is bounded and Riemann integrable in (0, l), it is
well known that

for any irrational 8. Now suppose j is no longer bounded but increases steadily to 00 as x + 0
from the right or from the left. Then the result still holds for almost all 8 provided that

is finite, where #(u) is any positive increasing function for which

O” du
s qw
mnverges.

259
 2. ADDITIVE NUMBER THEORY

(a) Combinatory Analysis and Sums of Squares

 INTRQDUCTION TO PAPERS ON COMBINATORY
ANALYSIS AND SUMS OF SQUARES
It is no accident that of all Hardy’s papers on additive number theory, those on
modular forms come first in chronological order. The general idea is to determine
(either exactly or approximately) an arithmetical function r(n) by means of the
formula 1
r(n) = fi f(x)x-n-1 dx,
n s
where the ‘generating function’
f(x) = 2 r(?2)xn
n=O
is regular in the interior of the unit circle and the integral is taken along a circle
IX I =r<l.
The most interesting case arises when the unit circle is the natural boundary of
the function J?(X). In many important cases the function f (@) is a modular form in
r in the upper haIf-plane, so that the behaviour of f(eri’) as T approaches a rational
number h/k is easily determined from the behaviour off(eniT) as T approaches 0.
The first example of this type which Hardy dealt with, in collaboration with
S. Ramanujan, was the partition function p1L9the number of partitions of n into
positive integers without respect to the order of the terms, but with repetitions
allowed.* The generating function in this case is

fi (l-x”)-l = I-t- g pp,

Ti%=l n-l

which is closely related to the well-known modular ‘discriminant’ :

? I i;1 QC---L
= &T fi (1 wenim7)24a !J4 \
4 T1
m=l
In their first paper on the subject (1917, 4) Hardy and Ramanujan had already
obtained the result

Their method consisted of the application of a Tauberian theorem to f (x) near x = I

only. Though the method can and did lead to fairly general theorems, it could not
produce results as deep as those found in their later work.
In the paper (1918, 5), for which the papers (19 16, 10) and (19 I7, 1) are preliminary
announcements,? Hardy and Ramanujan obtained a rapidly convergent asymptotic
formula giving pn with an error term O(n-4). Twenty years later H. Rademacher
(Proc. London Math. Sot. (2), 43 (1937), 241~54), by a slight but far-reaching modi-
fication of the proof, replaced the asymptotic expansion by a convergent series for pm.
* J. V. Uspensky alao investigated the asymptotic behaviour of pn by means of the contour integral;
see Bull. de Z’Acad. des Sciences de Ruasie, (6) 14 (1920), 199-218.
t An abstract of the paper itself wa8 publ&hed in PTOG, Loch Math. Sm. (a), 16 (1917), xxii.
$ There was also an amouncement in Proc. London Math. Sot. (Z), 17 ( 1918) xxii-xxiv.

263
 COMBINATORY ANALYSIS AND SUMS OF SQUARES

The next paper (1918,2) deals with f+unctionsf(x) based on modular functions which
have poles, but no essential singularities, in the upper half-plane.
The paper (1920, 10) deals with the number r,(n) of representations of n as a sum
of s squares. [ (1918, 10) is a preliminary announcement. $1 The relevant function is
W

@a%
T=
e( ) 2 ?
n=- co

which satisfies the functional equation

(p) = e(-;)*
Hardy and Littlewood had already studied the asymptotic behaviour of e(r),
particularly in (1914, 3).
By the use of the formula
e(T)8 = f r,(n)&nT
n-0
and Cauchy’s integral formula, Hardy obtained an exact expression for r,(n) if
5 < s < 8 and an asymptotic expression if s > 9. Hardy was, of course, aware of
the fact that his results for 5 < s < 8 were not new, though he stresses quite rightly
that his method deals simultaneously with even and odd values of s and presents
a unified approach to the problem for all values of s, though the classical cases s = 4
and especially s = 3 present some special difficulties. On the case s = 3, see T. Ester-
mann, Proc. London Math. SW. (3, 9 (1959), 575-94.
The whole matter was put into a more general context by C. L. Siegel [Alznals of
Matk (2) 36 (1935), 527-6061. He proved not only that Hardy’s formula holds,
mutatis mutandis, asymptotically for the representation of a number by any definite
quadratic form with rational integral coefficients, but that it holds exactly if a system
of forms is considered, which has one representative from each class in a fixed genus.
As the number of classes of positive definite quadratic forms of discriminant 1 is one
if and only if s < 8, it is now apparent why Hardy’s formula holds exactly for $ < 8
only. H. H. .
NOTE
The four papers [1917, 1; 1917, 4; 1918, 2; 1918, 51 written in collaboration with Ramanujan
are reprinted from CoLlected Pupers of S. Rumanujan (Cambridge, 1927).
The reader is referred to Hardy’s own comments given there, and to Chapters 3, 8, 9 of his
Ramnujun (Cambridge, 1940).

264
 ASYMPTOTIC
FORMULAEIN COMBINATORY
ANALYSIS.
BY

G. H. HARDY.
Trinity College, Cambridge (England).

1. The researches of which I propose to give a summary

here are the joint work of tOhe distinguished Indian mathemat-
ician, Mr. S. RAMANUJAN, and myself. They are the result of
an attempt to apply to the principal problems of the theory
of partitions the methods, depending upon the theory of ana-
lytic functions, which have proved so fruitful in the theory of
the distribution of primes and allied branches of the analytic
theory of numbers.
The most interesting functions of the theory of partitions
appear as t.he coefficients in the power-series which represent
certain elliptic modular functions. Thus p(n), the number of
unrestricted partitions of n, is tile coefficient of P in the ex-
pansion of the function

1
j(x) = 1 + &(n)P= (1-2)(l-x*)(1,5y *:
1 l

If we wri.te

where the imaginary part of z is positive, we see that f(x) ia

1916, 10 Qtiatrikme Cmgrks des Mathdmztickm Soadnaves, Stock-

holm, 1916,45-53. 265
 46

substaptiallz the reciprocal. of the modular function called by

TANNERY and MOLK h(z); that in fact

The theory of partitions has, from the time of EWLER on-

wards, been developed from an almost exclusively algebraical
point of view. It consists of an assemblage of formal identi-
ties - many of them, it need hardly be said, of an exceedingly
ingenious and beautiful character. Qf asymlptolic formulae, one
may fairly say, there are none. So true is this, in fact, that
we have been unable to discover in the literature of the sub-
ject any allusion whatever to the question of the order of
magnitude of p(n)*

2. The function p(n) mavL of course be expressed in the

form of an integral

U) E-
1 i( )X
dx
Fw
2ziJ xn+l
by means of CAUCHY’S Theorem, the path of integration en-
closing the origin and lying entirely inside the unit circle. The
idea which dominates this paper is that of obtaining nsymp-
totic formulae for p(n) by a detailed study of the integral (-1).
This idea is an extremely obvious one: it is the idea which
has dominated nine-ten ths of modern research in the analytic
theory of numhers; and it may seem very strange that it should
never have been applied to this particular problem before. Of
this there are no >doubt two explanations. The first is that
the theory of partitions has received its most important deve-
lopments, since its foundation by EULER, at the hands of a
series of mathematicians whose interests have lain primarily in
algebra. The second and more fundamental reason is to be
found in the extreme complexity of the behaviour of the gen-
erating function f(x) near a point of the unit circle.
It is instructive to contrast this problem with the cor-
responding problems which arise for the arithmetical functions
a(n), 8(n), q(n), /l(n), d(n), c. . which have their genesis in
RTEMANN’S Zeta-function and the functions allied to it. In
the latter problems we are dealing with functions defined by

266
 47

DIRICH~[~ET’S series. The study of such functions presents dif-

ficulties far more fundamental than any which confront us in
the theory of the modular functions, These difficulties, how-
ever, relate to the distribution of the zeros of the functions
and their general behaviour at infinity: no difficulties whatever
are occasioned by the crude singularities of the functions in
the finite part of the plane. The single finite singularity of
&), for example, the pole at s = 1, is a singularity of the
simplest possibla character. fit is this pole which gives rise to
the dom inant terms in th&ymptotic formulae for the arith-
meti cal fun&ions associated with c(s)7 To prove such a for-
mula rigorously is often exceedingly “difficult; to determine
precisely the order of the error which it involves is in many
cases a probl.em which still defies the utmost resources of ana-
lysis, But to writ-e down the dominant terms involves, as a
rule, no difficulty more formidable than that of deforming a
path of integration over a pole of tho subject of integration
and calculating the corresponding residue.
In the theory of partitions, on the other hand, we are
dealing with functions which do not exist at all outside t.he
unit circle. Every point of the circle ia an essential singularity
of the function, and no part of the contour of integration can
be deformed in such a manner as to make its contribution
obviously negligible. Everv ” element of the contour requires
special study; and there is no obvious m&hod of writing down
a ‘dominant term’.
The difficulties of the problem appear then, at first sigtlt,
to be very serious. We possess, however, in the formulae of
the theory of the linear transformation of the ellipt.ic functions,
an extremely powerful analytical weapon by means of which
we can study the behaviour of I(X) near any assigned point
of the unit circW It is to an appropriate use of these for-
mulae that the accuracy of our final results, an accuracy which
will, we think, be found to be quite startling, is due.

3. It is very important, in dealing with such a problem

as this, to distinguish clearly the various stages to which we
can progress by arguments of a progressively ‘deeper’ and less

1 See G. H. TARDY and L E. LITTLEWOOD, ‘Some problems of Dio-

phantine approximation (IX : The trigonometrical series associated with the
elliptic Theta-functions)‘, Acta Mathematics, vol. 37, 1914, pp. 193-238, for
applications of the formulae to different but not unrelated problems.

267
 48

elementary character. The earlier results are naturally (so far

as t ‘he particular pro blem is concerned) superseded by the later.
But the more elementary methods are likely to be applicable
.
to other problems in which the more subtle analysis is im-
practicable.
we have attacked this particular problem by a consider-
able number of dif feren t methods, and cannot profess to have
reached any very precise conclusions as to the possibilities of
each. A detailed comparison of the results to which they lead
would moreover expand this summary to a quite unreasonable
length. I shall therefore content myself with referring to
a brief note entitled ‘Une formule asymptotique pour le nombre
des partitions de n’, which appeared in the Comptes Rendus of
2 Jan. 1917, and to a full account of our researches which is
to appear in the Proceedings of the London Mathematical Society;’
aI]d pass on at once to an account of the more powerful me-
thods which give the best resul6s in this particular problem.

4. In order to obtain a good approximation to p(n), we

begin by constructing an auxiliary function P,(X) which is
regular at all points of the unit circle except x = 1 and has
there a singularity of a type as near as possible to that of
f(x). We may then hope to obtain a fairly close approxima-
tion by applying CAUCHY’S Theorem to f(x)- P(x) instead bf
to P(z). For, although every point of the circle is a singular
point of f(x), the point x = 1 is, to put it roughly, much the
heaviest singularity. It foll.ows from the formulae of the trans-
formation-theory that f(x) satisfies the equation

(2)

where
47E2
x’==exp - -1 ;
( log -X )

so that, when x--, 1 by real values, f(x) tends to infinity like

an expon .ential

’ Since this was written, the memoir here referred to haa been published*
See Proc. Lo&m Math* BOG., ser. 2, vol. 17, 1918, pp. 75; 115,
 49
2Jriip
--
On the other hand it may be shown t#bat, when x= re q ,
p and Q being co-prime integers, and r - 1, J(x) tends to in-
finity like an exponential

d ;
exp (6q’(l-r) 1

and if x=re20ni, where 0 is irrational, f(x) can become infinite

at most like an exponential of the type

-(o(&)) rn
5. The function required is

F(x) =

where
t/l(n)
=g (oosh
yn- ‘)9 n

This function may be transformed into an integral by means

of a general formula given by LINDEL~F; and it is then easy
to prove that the ‘principal branch’ of F(x) is regular
I all over
the plane except at x = 1; and that

where

is regular for x = 1. If we compare (2) and (3), and observe

that f(x’) tends to unity with extreme rapidity when x tends
to 1 along any regular path which does not touch the circle
of convergence, we can see at once the very close similarity
between tfle behaviour of f(x) and F(x) inside the unit circle
and in the neighbourhood of x = 1.

269
 50

k!pplying CATJCHY'S Theorem to f(x) - F(x), we obtain

(4) + O(eDvG) t

where D is any number greater than

6. The error term in (4) is of an exponential type, and

may be expected ultimately to increase with very great’rapidit-y.
It was therefore with considerable surprise that we fuund what
exceedingly good results the formula gives for fairly large values
of n. For 32 = 61, 62, 63 it gives 1121538.972, 130Q121.359,
X505535.606, while the correeb values are 1~21505, 1300156,
1505409. The errors 33.972, - 34.641, 36 c 606 are relatively
very small, and alternate in sign.
The next step is naturally to direct our attention to the
singular point of f(x) next in importance after that at x= 1,
viz. that at x = - 1; and to subtract from f(x) a second
auxiliary function, related to this point as F(x) is to x= 1 l

No new difficulty of principle is involved, and we find that

@fi) 9

1
where D is now any number greater than 3 C. It now be-
comes obvious why our earlier approximation gave errurs alter-
nately of excess and of defect.
It is obvious that this process may be repeated indefinitely.
The singularities next in importance are those at x = eini and
J% the next those at x= i and. x = -i; and so on’. The
next two terms in the approximate formula are found to be

and

270
 51

As we proceed further, the complexity of the calculations in-

creases. The auxiliary function associated with the point
zpni
X =e Q involves a certain Z4p-th root of unity, connected
with the linear transformation which must be used in order
to elucidate the behaviour of f(x) near the point; and the
explicit expression of this root in terms of 1) and q, though
known, is somewhat complex. But it is plain that, by taking
a sufficient number of terms, we can find a formula in which
the error is

where v is n fixed, but arbitrarily large, integer.

7. A final question remains. We have still the recourse

of making Y a function of n, that is to say of making the
number of terms in our approximate formula itself a function
of n. In this way we may reasonably hope, at any rate, to
find a formula in which the error is of order less than that of
any exponential of the type ,an; of the order of a power of n,
for example, or even bounded.
When, however, we proceeded to test this hypothesis by
means of numerical data most kindly provided for us by Major
P. A. MACMAHON, we found a correspondence between the real
and the approximate values of such astonishing accuracy as
to lead us to hope for even more. Taking n = 100, we found
that the first six terms of our formula gave

190568944.783
+ 348 872
l

- 2.598
+ .685
-I- + 318
,064
190569291.996,

while

271
 52

p(100) = 190569292;

so that the error after six terms is only . OM. We then pro-
ceeded tmo calculate p(ZOO), and found

3, 972, 998, 993, 185.896

+ 36, 282 w978
-87.555
+ 5.147
+ 1.424
+ 0.071
O*OW
+ 0.043
3, 972, 999, 029, 388.004;

and Major MACRIIARON’S subsequent calculations showed that

p(200) is in fact
3, 972, 999, 029, 388.

These results suggest very forcibly that it is possible t,o

obtain a formula for p(n> which not only exhibits its order of
magnitude and structure but may be used to calculate its exact
value for any value of n. That this in fact is so is shown by
the following theorem.

Theorem. Suppose that

where

for all positive integral values of q; that p is a positive integer

less than and prime to q; that cop,q is a 24 q-tit root of unity,
defined when p is odd by the formula

1 This term vanishes identically.

272
awl when q is odd by the formula

‘tip,*= (+p(-[a(n-l) + & (4j (w-PpI+Pw]7ci) 1

Q
where b-- is the sydml of LEUENDRE and JACOBI, and p’ is any
0
posilive integer szxh that 1 -I- pp’ is divisible by q; that

?472) = 2A,+, + O(?L.-:);

1

so that p(n) is, for all su&cieutly large values of n, the integer
nearest to

The proof of this theorem is naturally intricafe, but it in-

volves no fundamental idea beyond those which I have ex-
plained.

213
 UNE FORMULE ASYMPTOTIQUE POUR LE NOMBRE
DES PARTITIONS DE n

(Cmytes Rendzcs, 2 Jan, 1917)

1. Les divers problhes de la thborie de la partition des nombres ont &6

&udi& surtout par les mathbmaticiens anglais, Cnyley, Sylvester et Mac-
mahont, qui les ont abordbs d’un point de vue purement algbbrique. Ces
suteurs n’y ont fait aucune application des m&hodes de la theorie des
fonctions, de sorte qu’on ne trouve pas, dans la theorie en question, de
formules asymptotiques, telles qu’on en rencontre, par exemple, dans la
thborie des nornbres premiers. 11 nous semble done yue les rksultats que
nous allons faire connake peuvent p&enter quelque nouveautk.

2. Nous nous sommeaoccuph surtout de la fonction p (n), nombre des

partitions de n. On a
1I c10

f(x)=(1-x)(l-x2j(1-~)~~.=~~(njxn (1x1< 0

Nous avons pens&d’abord & faire usage de quelque thdor&me de caractere

3’ccuMien : on dbsigne ainsi les th&or&mes rkiproques du thAor&ne classique
d’Abe1 et de ses g6n6ralisations. A cette catkgorie appartient l’enonc6
suivant :
Soit g (5) = Saenxnune she de puhwancesh coe&ielzts POSITIFS,teh? yu’o?z
clit
A

quand x tend vers m par des valeurs positives. Alors on a

logs,=log(a,+u,-t... +a,)-Z@n),
quand n tend vers E’in&i f.

+ Voir le grand trait6 Combz’nutory Atialysis de M. Y. A. Macmaholl (Cambridge,

1915-16).
$ Nous awns do& des gh&alisations &endues de ce thhr~me dans WI m&moire
qui doit yaraftre dans un autre recueil.

274 1917, I (with S. Rttmanujan) Cbnptes Rends, 164, 35-38.

 240 Une Furmule Asynxptotique
En posant y (x) = (I- x) f (x), on c?l
T2
A =--- ;
6
et nous en tirons
p (n) = ,*JGn) (I + 4 I ..*,..,,,1.~~*...~.~*~.~,.~ (I)
oh E tend vers zh avec l/n.
--L,.-y ,,..
h-e...m-..,U
__.,.__,”-_I-..-3-y.I...---
1--.e.-,...
-..-~--I-‘“.r__U----~-““-
3. Pour pousser l’approximation plus loin, il faut recourir au th&or&me
de Cauchy. Des formules

avec un chemin d’intbgration convenable int&ieur au cercle de rayon un, et

(fournie par la thborie de la transformation linkaire des fonctions elliptiques),

nous avons tire, en premier lieu, la formule vraiment asymptotique

p (n) r~,P (n) =

On a
p (10) = 42, p (20) = 627, p (50) = 204 226, p (80) = 15 796 476 ;
P(10)=48, P (20) = 692, P (50) = 211590, P (80) = 16 606 781.
Les valeurs correspondantes de P (n) :p (n) sont
1945 ; 1904; 1965 ; 1951:
la valeur approximative est toujours en excbs.

4. Mais nous avons abouti plus tard & des r&mltats beaucoup plus satis-
f&ants. Nous consid&ons la fonction

I .,., l ...
(4)

En faisant usage des formules sommatoires que dbmontre 11. E. Lindelijf

dans son beau livre Le c&uZ des rtkidus, on trouve ni&ment que F(x) (on
parle, il va sans dire, de la branche principale) a pour seul point singulier
le point x = 1, et que la fonction

F lx) - &h)
“*J(logk) [exp{ci--I..} - 11

est r&guliAre pour x= 1. On est conduit naturellement & appliquer le

the’or&me de Cauchy & la fonction f(x) - F(x), et l’on trouve

275
 oh lc dksigne un nombre quelconque supbrieur h r/46* L’approximation, pour
des valeurs assez grandes de n, est t&s bonne : on trouve, en effet,
p (61) = 1 121505, p (62) = I300 156, p (63) = 1505 499 ;
Q (61) = 1121539, & (62) = 1300 121, Q (63) = 1505 536,
La valeur approximative est, pour les valeurs suffisamment grandes de n,
alternativement en exci?s et en dbfaut.
5. On peut pousser ces calculs beaucoup plus loin, On forme des fonc-
tions, analogues it F(z), qui prhentent, pour les valeurs
X = - 1, pi, +i, i, -i, e)“i, I., ,
des singularitks d’un type t&s analogue A celles que phente f(x). On
x
soustrnit alors de f ( ) une somme d’un nombre fini convenable de ces fonc-
tions. On trouve ainsi, par exemple,

oh k dhigne un nombre quelconque plus grand que a~&

276
 ASYMPTOTIC FORMULiFi FOR THE DISTRIBUTION
OF INTEGERS OF VARIOUS TYPES*

(Proceedings of the London iWathematical Society, 2, XVI, 1917, 13 2432)

1. Statement of the problem.

1% We denote by g a number of the form
(1%) 2”2zJa3 5”5 . l . pap,

where 2, 3, 5, . . . , p are primes and

(Ml) a,>aj,>,al,>, . . . >,cc,;
and by Q (3) the number of such numbers which do not exceed x: and our
problem is that of determining the order of Q (LC). We prove that

(142)

that is to say that to every positive Ecorresponds an X, = x0(E), such that

(1*121)

for X > 65,. The function Q (z) is thus of higher order than any power of log x,
but of lower order than any power of x.
The interest of the problem is threefold. In the first place the result
itself, and the method by which it is obtained, are curious and interesting in
themselves. Secondly, the method of proof is one which, as we shew at the
end of the paper, may be applied to a whole class of problems in the analytic
theory of numbers: it enables us, for example, to find asymptotic formulae
for the number of partitions of n into positive integers, or into different
positive
.- integers, or into primes. Finally, the classof numbers q includes as
a sub-classthe “highly composite” numbers recently studied by Mr Ramanujan
in an elaborate memoir in these Proceedings$ The problem of determining,
with any precision, the number H(x) of highly composite numbers not
exceeding x appears to be one of extreme difklty. Mr Ramanujan has
proved, by elementary methods, that the order of H(x) is at any rate greater
* This paper was originally communicated under the title rrA problem in the Analytic
Theory of Numbers.”
t Ramanujan, “ Highly Composite Numbers,” Proc. London Math. SW., Ser. 9, Vol, XIV
1915, pp 347-409,

1917, 4 (with S. Rarnanujan) Proceeding8 of the London Mathe-

muGcal Society, (2) 16, 112-32. 277
than that of log x* : but it is still uncertain whether or no the order of H(x)
is greater than that of any power of log x. In order to apply transcendental
methods to this problem, it would be necessary to study the properties of the
function

where h is a highly composite number, and we have not been able to make
any progress in this direction. It is therefore very desirable to study the
distribution of wider classes of numbers which include the highly composite
numbers and possess some at any rate of their characteristic properties. The
simplest and most natural such class is that of the numbers Q; and here
progress is comparatively easy, since the function

possesses a product expression analogous to Euler’s product expression for

5( s ) , viz.

where I, = 2.3.5 . . . pn is the product of the first rz primes.

We have not been able to apply to this problem the methods, depending
on the theory of functions of a complex variable, by which the Prime Number
Theorem was proved. The function @ (s) has the line c = 0t as a line of
essential singularities, and we are not able to obtain suficiently accurate
information concerning the nature of these singularities. Bnt it is easy
enough to determine the behaviour of @ (s) as a fun&on of the real variable s ;
and it proves sufficient for our purpose to determine an asymptotic formula
for @ (s) when s + 0, and then to apply a ‘I Tauberian ” theorem similar to
those proved by Messrs Hardy and Littlewood in a series of papers published
in these Proceedings and elsewhere:.

This CLT,zuberian” theorem is in itself of considerable interest as being

(so far as we are aware) the first such theorem which deals with functions or
sequencestending to infinity more rapidly than any power of the variable.

log Lx J( log log JQ .

+ As great as that of l

.
(log log log P>+

see p. 385 of his memoir.

t We write as usual 8= 0 +it.
1 See, in pa.rticular, Hardy and Littlewood, ClTauberian theorems concerning power
series and Erichlet’s series whose coefficients are positive,” Pwc. London M&L SOC.,
Ser. 2, Vol. XIII, 1914, pp+ 174-191 ; and Wome theorems concerning Dirichlet’s series,”
Xessenpr of M&m2atics, Vol. XLIII, 1914, pp. 134-147.
 DisWibution of Integersof Varz’ozcs
Types 247

2. Elementary results.
29. Let US consider, before proceeding further, what information con-
cerning the order of & (x ) can be obtained by purely elementary methods.
Let
(2U) ln=2.3.5 ,..p,=,S@n~,
where 9 (x) is Tschebyschef’s functiun
9 (x) = 2 logp.
P4X
The class of numbers q is plainly identical with the class of numbers of the
form
(212) llbl l,bz l *a 1n 48I
where bl>, 0, b, > 4)) l . . ) bn 2 0.
Now every b can be expressed in one and only one way in the form
(Z-13) b; = Ci 1
m 2” + Ci m-1 2”-’ 9
+ l l l + Ci
3
0~

where every c is equal GOzero or to unity. We have therefore

T.n
??I Z Ct,.i2j

(2.14) -e i
j-0 = 5 fi &, j 54= 5 @,
Q n(z
i=l 1 j=o dzf j=o

say, where
(2*141) rj = tfl, j I~%j . . . In% j,
Let r denote, generally, a number of the form
(Z-1 5) r = Jl%12Ca l ”
7: %an 1

where every c is zeru or unity : and R (z) the number of such numbers which
do not exceed x. If q < x, we have
To < G 7-12< x, ?q<X, . ..I
The number of possible values of rO, in the formula (Z+l4), cannot therefore
exceed R (8) ; the number of possible values of rl cannot exceed B (x*) ; and
so on, The total number of values of q can therefore not exceed
(Z-16) S (x) = R (x) R (xi) R (x’) .. . R (x2-m),
where = is the largest number such that
(2*161) x2-=
2 2, x >, z2”,
Thus
(2917) Q(4 s 8 (4

2.2. We denote byfand g the largest numbers such that

(2*211) ?f<X,
(Zm212) 11
12"' 1 g\ <x;.

279
 248 Distributioni of integers of Various Typea
It is known* (and may be proved by elementary methods) that constants
~4 and B exist, such that
(2221) 9 (2) 2 Lb (x 2 2),
and
(2.222) pnaBnlogn (n&1>.
We have therefore &f< X

f logf = 0 (To;E),
(2*23) log f = 0 (log log x) ;

and s 9 (pv) G log $9 E p, = 0 (log x),

I 1

E Y lg0 Y = 0 (log x), g2 logy = 0 (log x),

1

(2-24) g=o&g,.
But it is easy to obtain an upper bound for R (x) in terms off and g, The
number of numbers Z,, I,, . . . , not exceeding x, is not greater than f; the
number of products, not exceeding x, of pairs of such numbers, is a fortiori
not greater than if (f - 1) ; and so on. Thus

+ l **,

where the summation need be extended to g terms only, since

11
12”’ &Jg+, > x.
gi forth-i, we have
f f
R(x)41+f+~+...+~<(l+f)B=eg~og(L-i-f)*
l l

Thus
(2*25) R (x) = e&glwf) = ,O fh’(log
MlwN,
by (2%) and (2.24). Finally, since
log $/x log log &c < 8 log x log log x,
it follows from (296) a,nd (2.17) that
1 I
(2*26) Q (x) =eup 0 1 +Q+3+‘..+5 I) &log x log log x)/l
[ i(
= ,o Ih’(log
x loglog“11,

2.3. A lower bound for & (5) may be found as follows. If g is defined as
in 2-2, we have
11
1 2. . . 1, <x < I,& I.. lglg+l.
f See Landau, Handbuch, pp, 79, 83, 214.

280
 @stribution of Integers of Various Types 249
It follows from the analysis of 2.2 that

and 11
12 l **

Thus x < &7”10w~ ;

which is only possible if g is greater than a constant positive multiple of

Now the numbers I,, I,, ., . , I, can be combined in 8 different ways, and
each such combination gives a number q not, greater than GL Thus

(2w31)

where K is a positive constant. From (2%) and (2*31) it follows that there
are positive constants K and L such that

(Z-32) < log Q (x) < L 2/(log x log log x).

The inequalities (2-32) g ive a fairly accurate idea as to the order of

magnitude of & (x). But they are much less precise than the inequalities
(1*121), T o obt ain these requires the use of less elementary methods.

3. The behuviour of @ (s) when s-0 by positive values.

3.1. From the fact, already used in 29, that the class of numbers Q is
identical with the class of numbers of the form (N2), it, follows at once that

(194)

Both series and product are absolutely convergent for u > 0, and
(3.11) log@(s)=+(s) +&+(~s)s-$+(~s)+ ...j
where
(3ml11)
We have also
(3’12)
l-2-8
=-e-+?-k&+2-“a-.~+ 9..
+( s>
1 l \ + 0..
58’-1- -J
c -

281
250 Distribution of htegers of Various Typea
3#2. LEMMA.-If x > 1, s > 0, then

(3-21)

Write x8 = eU : then we have to prove that

(3-22) --1 1 P 1

for all positive values of u ; or (writing w for +u) that

(3.23)

for all positive values of w. But it is easy to prove that the function

is a steadily decreasing function of w, and that its limit when w -0 is & ;

and this establishes the truth of the lemma,

3*3. We have therefore

(3.31) #O s = & - !Pl (4 = &2 - 6 (4 -I- 0 (11,

where
(3m311)

From the second of these inequalities, and (2*221), it follows that

ca e-Asx
A - dx
2 logx

and so that
(3m32) 4 (s) > Aj2*ee; dx + 0 (1).

Qn the other hand there is a positive constant B, such that

%(x)<Bx (x@)*.
Thus

and so
(3.33)

+ Landau, Handbuch, lot. cit.

 Distribution of Integers of Varims Types 251
3a4. hIMA.---lf H is any positive mm&r, then

Given any positive number c, we can choose f: and LX, so that

0 < j, (s) < wIds HcHudu = 0 {Js log (l/s)) = o’(l),

2s
E
O<*$ (s) < 2 HrH”du< 2e,
I 0

j, (s) =jtxHcHudu + o (I),

0 < j, (s) < Ia lkH”du < E;

x ,
and so

< 5~ + o (1) < 6e,

for all suf3Eicientlysmall values of s,

3-5, From (3-32), (3.33)) and the lemma just proved, it follows that

(351)

From this formula we can deduce an asymptotic formula for log @ (s). We
choose IV so that
(3m52)
and we write

say.
In the first place
1+0(l) $1
(3’541) @I (s) = slog (l/s) 1 -?P’

283
 252 Distribution, of Integers of Varkms Types
IIn the place
+ (ns) = 0 l
i ns log (l/ns) 1 )
and log (l/m) > + log (l/s),
if N < n < l/& It follows that a constant K exists such that

(3.542) a*(8)< s logK(l/s) Ncnc 1 KE

- < s log (l/S) ’
n2
Thirdly, 2/s< ns < 1 in CD3
(s), and a constant L exists such that
L
+ @) < 2/s log (l/s) +
Thus
1; us 1 2L
--
(:3m543)
~3(s)~Jslog(l,s)~Yt~ 2/s’
for all sufficiently small values of s.
Finally, in @, (s) lve /lave IZS> 1, and a constant J4 exists such that
$3(ns) < iW2-Y
Thus
(3.544) @4(s) < M c 2-‘“s < SM c 2-n” < &- = 0 (1).
l/s<?&n I/S<?2

From (3*53), (3=541)-(3 544), and ($52) it follows that

(3.55) log QB(4 = 1 [;1+001 (g +,,tpj

s log (l/s)

where
Thus
(3-56)

(3-57) @ (4 =exp
[
{I+ 0 (I)] 6s~o;;l, S-

1
) l

4* A Tmbwian theorem
4% The passage from (39) to (1.12) depends upon a theorem of the
CcTauberian ” type,
THEOREMA. Suppusethat

284
 Bistribution of h&gem of Vur.ious Types %3
(4) A >o, a > 0;
(5) &,e-+ is convergent for s> 0;
(6) ,f(S)=&,e-A~s=exp {l+o(l)}A~-~
[
when s--f 0. Then
A, = a, + a, + l . l + exp [{l + 0 (1)l BXnal(l+aJ(log X,)-B/@+a)],
a, =

where _B= &l(lta) a-a/P+4(1 + ~)l+[Plllta)J

1

We are given that

(491) (l-s)As.(log~)-P < logf(s) < (1 +s> AS-~(logf)-8,

for every positive 6 and all sufficiently small values of s; and we have to
shew that
(4.12)
(1 - E)Bx,a!(l+a) (log X,)-~i(l+a) < log A,
< (1 + s) Bh,Ql+a) (log An)-@l(IsQ),
for every positive E and all sufficiently large values of 12.
In the argument which follows we shall be dealing with three variables,
6, s, and ‘n (or m), the two latter variables being connected by an equation or
by inequalities, and with an auxiliary parameter 6 We shall use the letter 7,
without a suffix, to denote generally a function of 6, s, and n, (or m)*, which
is not the same in different formula, but in all casestends to zero when 8
and s tend to zero and n (or m) to infinity; so that, given any positive E,we
hue
~+7~<“,
for o<s<s,, o<s<s,, n>n,.
We shall use the symbol l;lc to denote a function of &Yonly which tends to
zero with g, so that
O-+Jl-W
if c is small enough. It is to be understood that the choice of a f: to satisfy
certain conditions is in all casesprior to that of 6, S, and ‘22(or m). Finally,
we use the letters 25, K, . tiCto denote positive numbers independent of these
variables and of c,
The secondof the inequalities (4*12) is very easily proved. For
(4931) An e-+ < U, e-Q + aa 62’ + . . . + cc,e-Q
1 -@
<f(s) < exp (t+6)AS--a log- ,
i ( s 11
(4.132) 43 < exp XI
where
1 -I3
($9321) X=(l+S)AS- log- + Ls.
( s)
* 7 may, of course, in some cases be a function of some of these variables only.

285
 254 Distribution/ of Integers of Various Types
We can choose a value of s, corresponding to every large value of m, such
that
(4.14) (I- 6) AUK-~-Q log1 -’ <A, < (1 + S) Aus-‘-~ (log ;)-‘0
( s>
From these inequalities we deduce, by an elementary process of approximation,

(4.151) SD+a)

1
(4*152) i=~logx,<log~< 1 IL+%gh,,
I+a

(4954) x < (1 + rl) Bh,“l(=+a) (log X,)+/(l+“).

We have therefore
(496) log A, < (1 + E) BT@~+~) (log h,)-o/(l+a),
for every positive Eand all sufficiently large values of n”.

4.2. We have
(4%) f (8) = &q.p-hns = CA, (e--Q - &ws)
hn+1
=s:A, e-sxdx=i ao&4 (x) e-sxdx,
1 1An I0
where & (x) is the discontinuous function defined by
csqx)=& (haa= bl+l)t,
so that, by (496),
(4*22) log Jr4 (Lc)< (1 + f) Bxalcl+a)(logx)-~l(l+a,
for every positive Eand all sufficiently large values of x. We have therefore

(4*23) exp (I - 8) A s-a

cexp (l+S)Ara
1 --q
log-
i ( >I s
for every positive 6 and all sufficiently smslfl values of S.
f We use the second inequality (4.12) in the proof of the first, It would be sufficient
for our purpose to begin by proving a result cruder than (4%9, with any constant K on
the right-hand side instead of (1 + E) B, But it is equally easy to obtain the more precise
inequality. Compare the argument in the second of the two papers by Hardy and Little-
wood quoted on p. 114 [pa 246 of this volume] (pp. 143 et q.).
t Compare Hardy and Riesz, (‘ The General Theory of Dirichlet’s Series,” Cumbridge
Tmcts in &xthematics, No. 18, 1915, p. 24.

286
 We define x,, a steadily increasing and continuous function of the con-
tinuous variable z, by the equation
xz = b, + (x - n> (LSl 4,) (n~x~?zfl).
We can then choose tiz, so that

(4*24) 1 = l+cr h,l/W) (log &Jfl/(lfa).

s UB
We shall now shew that the limits of the integral in (423) may be replaced
by (l- 5) A, and (I+ C) hi, where c is an arbitrary positive number less
than unity.
We write.
(4.25)

where H is a constant, in any case greater than 1, and large enough to sntisfy
certain further conditions which will appear in a moment; and we proceed to
shew that J1, Ja, J4, and J6 are negligible in comparison with the exponentials
which occur in (4*23), and so in comparison with J3.
4.3. The integrals J1 and Js are easily disposed of. In the first place we
have
(431)

by (4.22) +. It will be found, by a straightforward calculation, that this

expression is less than

(4*32) (1 + 7) A (1 + a) H-a!(l+a) ra ,

and is therefore certainly negligible if H is sufficiently large.

Thus J1 is negligible. To prove that J5 is negligible we prove first that
sx > 4Bx+~“) (log x)-flI(‘+n),
if x > Hh, and H is large enought, It folbws that

a0
<S e-6S”Z:dx
= 9)
0

and is therefore negligible.

* With 6 in the place of E.
t We suppress the details of the calculation, which is quite straightforward.

287
 256 Distributim of Inhxps -of Various Types
4~4. The integrals Jz and Jd may be discussed in practically the same
way, and we may confine ourselves to the latter.

We have
(4*41)
where
(4*411) q = (1 + 6) &I?++~) (log ,)-@/(I+@ - sx,
The maximum of the function $ occurs for x = x0, where

(4#42) A = (1 + 7) 2 X,IItl+al (log xp+4.

S

From this equation, and (4*24), it plainly results that

(4.43) (1-rl)~m<Xo<(l+r7)hm,
and that x0 falls (when 6 and s are small enough) between (1 - r) A,~ snd
(1 + 5) h?%*
Let us write x=x&E
in J4. Then
q (x) = + (x0) + 4 (1 + 8) BE2 g {Z,“/(l+a) (logx,)-@lrl+a,],
I2
where x0 < x1 < x and GG
fortiori
(1--c)h,<x,< HA,.
It follows that

(4.44) g {x$t1+4 (log xJ-BlilSa) ] < - KX,4+4 -2 (log &J-P/tl+‘r)*

la
On the other hand, an easy calculation shews that

(4-45) (1-q) As- logi + < + (x0) < (1 + 17)AP (lo&)-@ l

S
( >

Thus
(4-46)

Since f: is independent of 6 and s, this inequality shews that J4 is negligible ;

and a similar argument may be applied to J2,

288
 1257
Distribution of liztegem of Vurioua~
Types
4.5, We may therefore replace the inequalities (4m23) by

(4m51)

Since &Z (x) is a steadily increasing function of x, it follows that

(44521)
exP
(4.522)

exP

Or (4’531)
(eTah - e-gs*m) ~4 ((1 - c) A,) < exp (1 + 6) A S- (logy +w} >
(4-532)
1 -p I
pm - e- csAm)&4 {(I + 5)X,}
> exp ((1 - 8) AS- (log s) +X,sb.
J
If we substitute for s, in terms of x,, in the right-hand sides of (4531) and
(4m532), we obtain expressions of the form
exp {(l + q) B&++“) (log &J-91(1M~).
On the other hand &Am - e-Ssll,
is of the form exp {9&#1*) (log &J+j(l*j ).
We have thus
I
(4.541)
A Ij( I- 3) h,J < exp {(I + 775+ 7) Bx$(~+) (log X?n)-~!~l*J},
(4.542)
A {(I + r) A,,) > exp { (1 - 715- 7) Bhmul(l+“j (log &J-Bl(l+~).
Now let Y be any number such that
(4.55) (1 - 0 Lb G L 6 (I+ 0 b
Since A&,-, + I, it is clear that uZZ numbers n from a certain point onwards
will fall among the numbers Y. It follows from (4.541) and (4.542) that

and therefore that, given E, we can choose first c and then n, so that
(4.57) exp {(l - e) Bh,“lcl*) (log A&-~l(lsa)} < A (X,)
< exp {(l + E) Bhnul(l+“) (log h,)-~~~l*)),
for 122~. This completes the proof of the theorem.

289
 258 Dz’stribution of Integers of Various Types
4*6. There is of course a corresponding “Abelian” theorem, which we
content ourselves with enunciating. This theorem is naturally not limited
by the restriction that the coefficients a, are positive.

(3) A >o, O<a<l;

(4) A, = a, + a2 + . . . + a, = exp [{I + 0 (I)1 ALa (log q-q
when n- a0 . Then the series 2 a,e-Q is convergelz t for s > 0, and

f(s) = Cane-Q = exp (1 + o (l)} Bs-+-+) (lug y”-)] ,

where
when s-0.
Th e proof of th is theorem, which is naturally easier than that of the
correlative Tauberian theorem, should present no difficulty to anyone who
has followed the analysis which precedes.
4.7. The simplest and most interesting cases of Theorems A and B are
those in which
x, = 72, p = 0.
It is then convenient to write it’ fur e? We thus obtain
THEOREMC. If A ~0, Otcrc 1, and
lug A, = log (q + a2+ . .. + a,) - Ana,
then the series&,xn is convergmztfor 1LC1c 1, and
log f (z) = log (&x,,z~) - B (I - x)-1(+,
where B = (1 - a) @-CL) A11(1-,,
whelz 1 by real values.
If the coeficients are positive the converseinference is ulso correct. That is
to say, ;$1
A > 0, a > 0,
md log f (x) - A (I - x)-,
then log A, - B na/(l*),
where B = (1 + a) pl(1*) ~l/w4*

5. Application to our problem, and to other problems in the

Theo?-yof Numbers.
5-l. We proved in 3 that
7T2
(3.56)
log @ (‘) - 6s log (l/s) ’

290
 In Theorem A take

xm= log 12, A=$, a=l, /3=1.

Then all the conditions of the theorem are satisfied. And A, is Q (n), the
number of numbers Q not exceeding n, We have therefore

(5.11)
where
(592)

5.2. The method which we have followed in solving this problem is one
capable of many other interesting applications.
Suppose, for example, that R, (R) is the number of whys in which n can
be represented as the sum of= number of rth powers of positive_---A+--
integers? h
We shall prove that -- -

(5.21) log R, (n) - (r + 1)

In particular, if P (n) = R, (n) is the number of partitions of n, then

(5*22) log .P (n) - 7T

2n
3’ Jc >
We need only sketch bhe proof, which is in principle similar to the main
proof of this paper. We have

and SO
(5’23) f (4 = 5 [R, (n) - R, (n - I)1 es”” = $ [&$+
-
1
It is obvious that R, (n) increases with n and that all the coefficients in
f(s) are positive. Again,

(5*24) logf (s)=~log(~~)=~(e-8”+~~~2sv~+...)

-
2

where
(5.241 + (s) = ii2e-V
2

* Thus 28=33+13=3. 23+4. 1 3=2.23+12.l3=23+20.13=28, 13:

and R3 (28) = 5.
The order of the powers is ;~upposed to be indifferent, so that (e.g.) 33+ l3 and 13+ S3 ~0
not reckoned as separate representations.
t & (0) is to be interpreted as zero.
But
(5.25) 4(S)-r(E + l)s-‘/:

when s-0 ; and we can deduce, by an argument similar to that of 3~5,that

(5*26) logf(s) - r (; + 1) 5(3 + 1) s-Ii’*

We now obtain (5*21) by an application of Theorem A, taking

z
)Ln=n,a=-, P=O, A=$+1)5(i+lj.
In a similar manner we can shew that, if 8 (n) is the number of partitions
of n, into diftwent positive integers, so that
IZS (n) emns= (1-t edg)(1+ Czs) (1+ e+) + . . l

1
=
(1 - P) (1 - e-3s) (1 - e+) ... '
then
(S-27) logS(?2)Vr
f ;
Jc >
that if Tr (‘n) is the number of representations of n as the sum of rth powers
of primes, then
(5.28)
log Tr (n>- (r + 1) {F (; + 2) c (; + l)}r”7+1’ PI~P+~) (logn)-“I(T+lj ;

and, in particular, that if T (n) = T1 (n) is the number of partitions of n into

primes, then
(5%1) log T (n) - Jj
2T&%) l

Finally, we can shew that if r and s are positive integers, a > 0, and
O&sl, and
{(l+ ax) (1-t axz) (l+ ax3) . ..I’
(5a291) Fd# (n) x”=
{(l - bx) (1 - bx2) (1 - bti) .a.~“’
then

where
(5.2921)

In particular, if a = 1, b = 1, and r=s, we have

(5*293) ~~(n)x~=(1-2x+21-2x9+...)-:
(5.294) 1% + (n> - T 2/(m).

[Added KU& 2&h, 1917.-Since this paper was written M, G, Valiron

(‘I Sur la croissancedu module maximum des s&es enti&res,” Bulletin de la
 Di8tribution of Integers of Vurious Type8 261
sOci&e’ mathdmatiqw de France, Vol. XLIV, 1916, pp. 45-64) has published a
number of very interesting theorems concerning power-series which are more
or less directly related to ours. M. Valiron considers power-series only, and
his point of view is different from ours, in some respects mure restricted and
in others more general.
He proves in particular that the necessaryand su$icierd conditions that

where M (r) is the m&mum modulz~ off(x) = ~Za,+n for 1s I= r, are that

f or n = np (p = 1, 2, 3, . ..). where nP+&aP+l and +-to as p-,00.

M. Valiron refers to previous, but less general or less precise, results
given by Bore1 (Lepotis sur Eess&es d termes po&fs, 1.902, Ch. V) and by
Wiman (“ Uber dem Zusammenhang zwischen dem Maximal-betrage einer
analytischen Funktion und dem grijssten Gliede der zugehGrigen Taylor’schen
Reihe,” Acta Mathematics, Vol. XXXVII, 1914, pp. 305-326). We may add a
reference to Le Roy, “ Valeurs asymptotiques de certaines s&&s procbdant
suivant les puissancesentSres et positives d’une variable rbelle,” BdZetin des
scienJces
mnthdmatiques, Ser. 2, Vol. XXIV, 1900, pp. 245-268.
We have more recently obtained results concerning P (n), the number of
pwtitions of n, far more precise than (5*22). ]

293
 0N THE COEFFICIENTS IN THE EXPANSIONS
OF CERTAIN MODULAR FUNCTIONS

(Proceedings oJfthe RoyaZ Society, A, WV, 2919, 144-155)

1. A very large proportion of the tnost interesting arithmetical functions

-of the functions, for example, which occur in the theory of partitions, the
theory of the divisors of numbers, or the theory of the representation of
numbers by sums of squares- occur as the coefficients in the expansions of
elliptic modular functions in powers of the variable Q= e? All of these
fu.nctions have a restricted region of existence, the unit circle ( $11= 1 being
a “ natural boundary ” or line of essential singularities. The most imy>ortant
of them, such as the functions*
. (~~/r)l~A = ~~((1 -9”) (I - 4”) .,.j24,
(1.1)
l

(12) %~(0)=1+2q+2q4+2q9+...,

(13) l

are regular inside the unit circle ; and many, such as the functions (PI) and
(1*2), have the additional property of havingL no zeros inside the circle, so that
their reciprocals are also regula3
In a series of recent paperst we have applied a new method to the study
of these arithmetical functions. Our aim has been to express them as series
which exhibit explicitly their order of magnitude, and the genesis of their
irregular variations as n increases. R7efind, for example, for p (n), the number
* We follow, in general, the notation of Tannery and Molk’s l?+Mments de la the’orie
des jonctions eZl@tiques. Tannery and Molk, howover, write 166 in place of the more
usual A.
f (1) G. H, Hardy and S, Ramanujan, “Une forrnule asymptotique pour le nombre
des partitions de n,” Comptes Rendus, January 2, 1917; (2) G. H, Hardy and S.
Ramanujan, “Asymptotic Formulae in Combinatory Analysis,” Proc. London Mu&
XOC,, Ser. 2, Vol. XVII, 1918, pp, 75-115; (3) EL Ramanujan, “On Certain Trigono-
metrical Sums and their Applications in the Theory of Numbers,” Truns. CCCW&.
Phil. XOC., Vol. XXII, 1918, pp. 259-276; (4) G. H. Hardy, “On the Expression of a
Number as the Sum of any Number of Squares, and in Particular of Eve or Seven,”
Proc. lVational Acad. of Sciences, Vol, IV, 1918, pp. 189-193: [and G. H. Hardy, “On
the expression of a number as the sum of any number of squares, and in particular of
five,” Trans. Amsrimn Math. Sot., Vol. XXI, 1920, pp. 255-2841.

1918, 2 (with S. Ramanujrtn) Proceedings of the RoyaE Society, A, 95,

144-55.
of unrestricted partitions of 12, and for ra (n>, the number of representations
of n as the sum of an even number s of squares, the series

where

and our methods enable us to write down similar formulae for a very large
variety of other arithmetical functions.
The study of series such as (1.5) and (1%) raises a number of interesting
problems, some of which appear to be exceedingly difficult. The first purpose
for which they are intended is that of obtaining approximations to the
functions with which they are associated. Sometimes they give also an exact
representation of the functions, and sometimes they do not. Thus the sum
of the series (ln6) is equal to r8 (-n) if s is 4, 6, or 8, but not in any other case.
The series (l-5) enables us, bv stopping after an appropriate number of terms,
to find approximations to p (91) of quite s tling *accuracy; thus six terms of
the series give p (200) = 3972999029388, a number of 13 figures, with an error
of OaOO4, But we have never been able to prove that the sum of the series is
p (n) exactly, nor even that it is convergent.
There is one class of series, of the same general character as (1.5) or (P6),
which lends itself to comparatively simple treatment. These series arise
when thegenerating modular function f(q) or 4 (T) satisfies an equation
L*VJ ; \ i-0 \- _

where 12 is a positive integer, and behaves, inside the unit circle, like a
rational function ; that is to say, possesses no singularities but poles. The
simplest examples of such fr-xnctions are the reciprocals of the functions (1.3)
and (P4). Th e coefficients in their expansions are integral, but possess other-
wise no particular arithmetical interest. The results, however, are very
remarkable from the point of view of approximation ; and it is, in any case,
well worth while, in view of the many arithmetical applications of this type
of series, to study in detail any example in which the results can be obtained
by comparatively simple analysis.
We begin by proving a general theorem (Theorem I) concerning the
expression of a modular function with poles a,g a series of partial fractions.

295
312 On certain Mudular Functions
This series is (as appears in Theorem 2) a “ PoincarA’s series ” : what our
theorem asserts is, in effect, that the sum of a certain PoincaA’s series is the
only function which satisfies certain conditions. It would, no doubt, be
possible to obtain this result as a corollary fruti propositions in the general
theory of autoxnorphic functions ; but we thought it best to give an inde-
pendent proof, which is interesting in itself and demands no knowledge of
this theory.
2. TEEORE:M 1. SU~OS~that
b l

(2 1) f(q) =f bm = 4 (7)
is WgulaP for q= 0, has no singularities save poles within the unit circle, and
satisfies the functional equation
1
(2 2) cp(7)= (a + b7)“$ (ST) = (a+ bT)?y(T),
n being a positive integer and a, b, c, d any integers such that ad - bc = 1. Then
.
(2 3) f (4) = =,
where R is a residue of f (Nkl - 4
at a pole off(x), zy 1Q) < 1; whlile $7 [ q 1> 1 the su/muf the series on the right-
haqkd side of (2-3) is zero,
The proof requires certain geometrical preliminaries.
3. The half-plane I(T) > 0, which corresponds to the inside of the unit
circle in the plane of q, is divided up, by the substitutions of the modular
group, into a series of triangles whose sides are arcs of circles and whose
angles are $T, $T, and 0”. One of these, which is called the fundamental
[email protected]. (P) t, has its vertices at the points p, p2, and ioo , where p = etd, and
its sides are parts of the unit circle 17I= I and the lines R (T) = + ha
Suppose that Fm is the “Farey’s series” of order 132, that is to say the
aggregate of the rational fractions between 0 and 1, whose denominators are
not greater than m, arranged in order of magnitude:, and that h’/lc’ and h/k,
where 0 < h//k' < h/k < 1, are two adjacent terms in the series, We shall
consider what regions in the r-plane correspond to P in the T-plane, when

(3.1)
h-kr
(3*2) T =hw;.
Both of these substitutions belong to the modular group, since hk’ - h’k = 1.
The points i~o, QI -8, in the Z’-plane correspond to h/k, (h + Zh’)/(~c + Zk’),
* It is for many purposes necewxry to divide each triangle into two, whose allglee are
9 V, +, and 0 ; but this further subdivision, is not required for our present purpose. For
the detailed theory of the modular group, see Klein-Fricke, Vorkesungen iiber die Thmrie
der E?Zi’ti~~h i!fodulfunktionen, 189&l 892.
t See Fig. 1.
$ The first and last terms are O/1 and I/L A brief account of the properties of Far+
scriea is given ill $43 of our payer (2),,
(h - Zh’)/(lc - 21~‘) in the T-plane. Thus the lines R (T) = 9, R (!i!‘) = - 8 corre-
spond to semicircles described on the segments
h h-2h’
p m
respectively as diameters, Similarly the upper half of the unit circle corre-
sponds to a semicircle on the segment

Fig. 1.

The polygon P corresponds to the region bounded by these three semicircles.

In particular, the right-hand edge of P correspondsto a circular arc stretching
from h/k (where it cuts the real axis at right angles) to the point
htld
I
(3 3) _-- + -hlc
_-.-+ 4 (hlc’ + h’k) -+- 8; d3
k” + kk’ + lcf2
corresponding to T = p.
Similarly we find that the substitution (3.2) correlates to P a triangle
bounded by semicircles on the segments

In particular, the left-hand edge of P correspondsto a circular arc from h’/k’

to the point (3m3). These two arcs, meeting at the point (3*3), form a con-
tinuous path o, connecting h/k and h//k’, every point of which corresponds,in
virtue of one or other of the substitutions (39) and (3*2), to a point on one
of the rectilinear boundaries of P.
* Fig. 2 illustrates the case in which h/k- -3, K/k’=+. These fractions are adjacent in
Ffi and F6, but not in FT.

297
 Performing a similar construction for every pair of adjacent fractions of
F rn? we obtain a continuous path from 7 = 0 to 7 = 1. This path, and its
reflexion in the imaginary axis, give a continuous path from T = - I to 7 = I,
which we shall denote by a,. To -Iz, corresponds a path in the q-plane,
&,&b,x~ call Hm ; Hm is a closed path, formed entirely by arcs of circles
which cut the unit circle at right angles.

Fig. 2.

The region shaded horizontally corresponds to P for the substitution (3-I ), that shaded
vertically for the substitution (3%). The thickest lines shew the path O; the line of
tnedium thickness shews the semicircle which corresponds (for either substitution) to
the unit semicircle in the plane of Y’. The large incomplete semicircle passes through
r z 1.

Since

the path w from h’/lc’ to h./k is all-ays passing from left to right, and its
length is less than twice that of the semicircle on (U/k’, h/k), i.e., than rrr/lck’.
The total length of II, is lessthan 2~; and, since

the length of H, is less than 27~. Finally, we observe that the maximum
distance of & from the real axis is less than half the maximum distance
between two adjacent terms of &, and so less than 1/2m*. Hence & tends
uniformly to the real axis, and H, to the unit circle, when mea l

4. The function + (7) can have but a finite number of poles in P; we

suppose,for simplicity, that none of them lie ox1the boundary. There is then
a constant K such that If(q) ]< K on the boundary of P.
* See Lemma 4% of our paper (2).
 315
We now consider the integral
1 f-k4 dx
(4 1)
l

ihi I z-q
- -

’
where 1q I< 1 and the contour of integration is Hm*. By Cauchy’s Theorem,
the integral is equal to

where R is a residue of f (x)/(4 - x) at a pole of f(x) inside E&J-. To prove

our theorem, then, we have merely to shew that the integral (49) tends to
zero when m-t w ,
Let ol’ and o1 be the left- and right-hand parts of o, and cl’, cl and r the
corresponding arcs of H,. The length of O, is, as we have seen, less than
&+k’, and that of 53 than $+‘/?&A Further, we have, on cl,

Thus the cont’ribution of & tu the integral is numerically less than

C/(klc’n+l), where C is independent of 112; and the whole integral (49) is
numerically less than

where the summation extends to all pairs of adjacent terms of Fm.

M7hen v is fixed and m > U, the number of terms of Fm whose denominators
are less than Z, is a function of v only, say N(V). If h/k is one of these, and
V/k’ is adjacent to it, Ic + lc’ > ml, and so k’ > m - u, Thus the terms of (42)
in which either k or k’ is less than v contribute less than 8CN (u)/(nz - v).
The remaining terms contribute less than
-45 1
--=- 4c
vn lcr v” l

Hence the sum (4-3) is less than

8CN (v)-- + FC
--“-- -
97X-V vn ’

and it is plain that, by choice of first v and then rrh, this may be made as small
as we please. Thus (4.1) tends to zero and the theorem is proved. It should
be observed that CR must, for the present at any rate, be interpr&ed as
meaning the limit of the sum of terms corresponding to poles inside Hm ; we
have not established the absolute convergence< of the series.
* Strictly speaking, f(x) ia not defined at the points where Hm meets the unit circle,
and we should integrate round a path just inside Hm and proceed to the limit. The point
is trivinl, as f(x), in virtue of the functional equation, tends to zero when we approach
a cusp of HL from inaide.
t We suppose 112 large enough to ensure that x2”- q liea inside H,, .
1 Se0 our paper (Z), 20~. cit.

299
 316 On certain ibdular Functiuns
We supposed that no pole of #I (T) lies on the boundary of I? This restric-
tion, however, is in no way essential; if it is not satisfied, we have only to
select our “ fundamental polygon ” somewhat differently. The theorem is
consequently true independently of any such restriction.
So far we have supposed ] q I< 1. It is plain that, if 1g I> 1, the same
reasoning proves that
l

(4 3) XR=O.

5. Suppose in particular that # (7) has one pole only, and that a simple
pole at T = a, with residue A. The complete system of poles is then given by

(5.1) T = al- = (ad-bc=l).

If a and b are fixed, and (c, d) is one pair of solutions of ud - &= I, the com-
plete system of solutions is (c + KU, d -I- ma), where m is an integer. To each
pair (a, b) correspond an infinity of poles in the plane of 7; but these poles
correspond to two different poles only in the plane of q, viz.,
(52) l

the positive and negative signs corresponding to even and odd values of YIX
respectively. It is to be observed, moreover, that different pairs (cc,b) may
_ give rise to the mm pole q.
The residue of c#(7) for 7 = a is, in virtue of the functional equation (P2),
A
(a + bu)n+s;
and the residue off (q) for g = q is
;

A
(U + 6~)““’
Thus the sum of the terms of our series which correspond to the poles
I is
(52)
TiA PI-q q 27&l q2
(a + btx)n+a q-q q+q ) =(u+ba)n+2m*
We thus obtain :
THEOREM2. If +(T) 12u.sone pole only in P, uiz., a simple pole at T = 61,
with residue A, and ) q I< 1, then

(5.3) f(q)=2riAX ’ --f-

(a + ba)n+Zq2- qa’
c + da
Where q=exp
( >
a-+ga ti;

c, d being any pair of solutions of ad - bc= 1, ana the su7~h9d0~ ddifig

over all pairs a, b, which give rise to distinct values of q. If 1q 1> 1, the sum
of the serieson the right-hand side of (%3) is zero.
 On certain M od/rclur Functions 317
If #I (T) has several poles in P,f(q), of course, will be the sum of a number
of series such as (5.3). Incidentally, we may observe that it now appears that
the series in question are absolutely convergent.
6. As an example, we select the function

say, where GC = q? It is evident that p, is always an integer; the values of

the first 13 coefficients are
p. = 1, pl = 504, p, = 270648, p, = 144912096,
p, = 77599626552, p, = 41553943041744, p, = 22251789971649504,
p7=11915647845248387520, ps=6380729991419236488504,
Pg =3416827666558895485479576,
plo=1829682703808504464920468048,
pll=979779820147442370107345764512,
plS=524663917940510191509934144603104;
so thab pin is a number of 33 figures.
By means of’ the formulae*
93= 2%(e1- eJ2(1 + k2) (1 - QP) (1 - 2h?),

we find that

The value of YI is 6. The poles of f(q) correspond to the values of T which

make K = k2 equal to - 1, 2, or 9. It is easily verifiedt that these values are
given by the general formula

80 that
.
(6 a

‘+ In order to determ”lne
The value of Q is a,+. A we observe that
Z644
+ (1 _ ,4>2 + l mm l

* All the formulae which we quote are givexl in Tannery and Molk’s Tables ; see in
particular Tables XXXVI (8), LXX1 (3), XCVI, CX (3).
t A full account of the problem of finding r when K ia given will be found in Tannery
and Molk, Zoc. cit., Vol. III, ch. 7 (“ On donne A+ ou g2, g3 ; trouver T 011 ml, w3 I’>.
1 It will be observed that in thia case a ia on the boundary of P; see the concluding
remarks of $4, As it happens, r= i lies on that edge of P (the circular edge) which was
not used in the construction of Hm, HO that our analysis is applicable as it stands,

301
 318 On c&ah Mudular Functions
The series in curly brackets is +,he function called by Ramanujan* @1,6andt
1oo8@~,~ = &” - PR,

Here R = 0, so that
lq 27q4
1008Q116= &” = 1 -f- 480@,,,; = I+ 480 i-q+ 1-q+- l

Hence we find that A = iId, 2&A = - 2/C,

where
l

(6 3)

Another expression for C is

.
(6 4)

We have still to consider more closely the values of a, and b, over which
the summation is effected. Let us fix E, and suppose that (a, b) is a pair of
positive solutions of the equation a2 + b2= h. This pair gives rise to a system
of eight solutions, viz.,
(t a, + b), (+, b, f a).
But it is obvious that, if we change the signs of both QI;and b, we do not
affect the aggregate of values of a. Thus we need only consider the four pairs

If a or b is zero, or if a = b, these four pairs reduce to two.

Itl is easily verified that, if (a, b) leads to the pair of poles

T
---
a2+ b2 ’
then (a, - 6) and (6, a) each lead to g = f, 4, where S is the conjugate of q.
Thus, in general (a, b) and the solutions derived front it lead to four distitlct
poles, viz., k q and k & These four reduce to two in two cases,when q is
real, so that q = q, and when q is purely imaginary, so that q = - a. It is

Q S. Ramanujan, V3n Certain Arithmetical Functions,” Tram, &zndj. PM, &c., Vol.
XXII, pp. 159-184 (p. 163).
+ Ramannjan, Zoc. cit., p. 164.
$ Ramanujan, lot. cit., pe 163.

302
l
easy to see that the first case can occur only when k = 1, and the second
when k=2’.
If E=l we take a4, b=O, c=Q, d=l; and q=g=eBr, If k:=Z we
take a = I, b = 1, c = 0, d = I; and q = - 4 = ie-+. The corresponding terms
in our series are
1 1
1 - qaeffl’ Z4(1 +pe*)’

If I;: is greater than 2, and is the sum of two coprime squares a2 and b2, it
gives rise to terms
1 ---- 1 1 1
(a + 6i)” 1 - (q/q)” + (a 1 - (p/g)”
There is, of course, a similar pair of terms corresponding ‘to every other
distinct representation of k as a sum of coprime squares. Thus finally we
obtain the following result :

THEOREM 3. If

f(q) -= T6 1
= 216@g3
and ] q I< 1, then

(6.5)

and S is the conjugate of g, The summation upplies to every pair of c~priltne

positive numbers a and b, such that k = aa+ b22 5, such pairs, however, only
being counted as distinct if they correspond to independen.t representations of
k as a sum of squares. If ] q 1> 1, then the sum of thle series on th,e right-hand
side of (65) is zero.

* When a and b are given, we cax1always choose c and d so that 1ac + bd I< + (as+ ba).
If q is real, we have ad - bc- I and ac +bd-= 0 simultaneously : whence
(a2+ b3) (c2+d2)= 1.
If q is purely imaginary, we have
ad-bc=l, 21ctc+bd1=a2+b2,

whence (C2+d2)3=(Iuc$bdI--2--2)2+1.

This is possible only if 8 +d”= 1 and 1ac + bd I= 1, whence a2+b2= 2.

303
 7. It follows that

say. Here x is the sum of two coprime squares, so that

h = 2a25a513”ls ua17 . . l ,

where a~ is 0 or 1 and 5, 13, 17, l are the primes of the form 4k+ 1 ; and
the first few values of CA(n) are
cl (n) I 1, c2 (~3) = (- I)“, c5(n) = 2 cos (@r + 8 arc tan Z),
cl0 (n> = 2 cos (+T - 8 arc tan 2>, cla(n) = 2 cos (98~ + 8 arc tan 5).
The approximations to the coefficients given by the formula (7 . 1) are
exceedingly remarkable. Dividing by 8 C, and taking n = 0, 1, 2, 3, 6 , and 12,
we find the following results :
(0) 0.944 (1) 505-361 (2) 270616.406
+ 0.059 - P365 -+ 31585
- O-003 . + O-004 + @009
p* = l*OOO p= 504*0UO pa = 270648.000
(3) 144912827*002 (6) 2225178996259245Oa237
- 730m900 + 905705t688
- O*lOl + 2*081
- @OOl - 0,006
p3 = 144912096*000 p6 = 22251789971649504-000
(12) 524663917940510190119197271938395~329
+1390736872662028~140
+ 268@418
+ 0*1.30
- O-014
- Oh003
pla=524663917940510191509934144603104-000
An alternative expression for C is
0 = 962en8”lJ{( 1 - emrT)(I - e-8rr) . . .}I~,
by means of which C may be calculated with great accrlracy? To five places
we have 2/C = 0*94373, which is very nearly equal to 352/373 = 0*94370.

* Gauss, %dx, Vol. III, pp, 418--419, gives the values of various powers of e-r to a
large number of figures.

304
 On certain ilCodular Functions 321
l

It is easy to see directly that pn lies between the coefficients of x” in the

expansions of
1 1 - 7m5x
(1 - 535x) (1 + 317) ’ (1 - 535*5c) (1 + 242) ’
and eu that
(535)n+l- (- 3L)““l 352 (535.5)” + 21(- 24)”
Gp& I
566 373

The function

has very remarkable properties. It is an integral function of x, whose maxi-

mum modulus is less than a constant multiple of e21rW It is equal to pn, an
integer, when x = TC,a positive integer ; and to zero when x = - n. But we
reserve the disc ussion of these peculiarities for some other occasion.

305
 ASYMPTOTIC FORMULE IN COMBINATORY
ANALYSIS*

(Proceedings of the London iifathemutical fiociety, 2, XVII, ‘1918, 75-115)

1. INTR~DWCT~ON AND SUMMARY OF I~ESULTS.

1% The present paper is the outcome of an attempt to apply to the
principal problems of the theory of partitions the methods, depending upon
the theory of analytic functions, which have proved so fruitful in the theory
of the distribution of primes and allied brancheq of the analytic theory of
numbers,
The most interesting functions of the theory of partitions appear as’the
coefficients in the power-series which represent certain elliptic modular
functions. Thus p (n), the number of unrestricted partitions of m, is the
coefficient of xn in the expansion of the function

(loll) f(x) = 1 + 51, (II) xn= t

(l-x)(1-x2)(1-ti)..* l
1

If we write
(1-12) x = q2 = e2=i+,
where the imaginary part of 7 is positive, we see that f(z)
the reciprocal of the modular function called by Tannery
that, in fact,
(193)

The theory of partitions has, from the time of Euler onwards, been
developed from an almost exclusively algebraical point of view. It consists
of an assemblageof formal identities- many of them, it need hardly be said,
of an exceedingly ingenious and beautiful character. Of asyq~otic form&
one may fairly say, there are nones. So true is this, in fact, that we have
* A short abstract of the contents of part of this paper appeared under the title L6Une
formule asymptotique pour le nombre des partitions de n,” in the Gonzptes Rendus, January
and, 1917.
t P, A. MacMahon, Combinakvy Analysis, Vol. II, 1916, p, 3.
1 J. Tannery and J. Molk, Folzctions elli’tipes, Vol. II, 1896, pp. 31 et se9 We shall
follow the nobtion of this work whenever we have to quote formulaa from the theory of
elliptic functions.
s We should mention one exception to this statement, to which our attention was called
by Major MacMahon. The number of partitions of n into parts none of w/&h exceed r is
the coefficient pOr (n) in the series
1
l+~pr(n)x~=(l_,j(l-d)...(l-xr)*
This function has been studied in much detail, for various special values of r, by Cayley,

1918, 5 (with s. Ramanujan) of the Lomdm Mathe -

306 7?mtical society, (2) 17, 75-115.
been unable to discover in the literature of the su bject any allusion whatever
to the question of the order of magnitude of p (n).
12
l

; The function p (12) may, of course, be expressed in the form of an

integral
(1%) =-
1 f( X1dX
P( n ) I
277-i.r x ?I+1
’

by means of Cnuchy’s theorem, the path r enclosing the origin and lying
entirely inside the unit circle. The idea which dominates this paper is that
of obtaining asymptotic form& for p (n) by a detailed study of the integral
(P21). This idea is an extremely obvious one ; it is the idea which has
dominated nine-tenths of modern research in the analytic theory of numbers:
and it may seem very strange that it should never have been applied to this
particular problem before. Of this there are no doubt two explanations. The
fist is that the theory of partitiotls hns received its most important develop-
ments, since its foundation by Euler, at the hands of a seriesof mathematicians
whose interests have lain primarily in algebra. The second and more funda-
mental reason is to be found in the extreme complexity of the behaviour of
the generating function f (x) near a point of the unit circle.
It is instructive to contrast this problem with the corresponding problems
which arise .for the arithmetical functions r (n), 9 (n), $ (n), p(n), d (n), . . .
which have their genesis in Riemann’s Zeta-function and the functions allied

Sylvester, and Glaisher : we may refer in particular to J. J. Sylvester, Wn a discovery in

the theory of partitions,” Quarterly J ournal, Vol. I, 1857, pp. B&--85, and “On the parti-
tion of numbers,” ibid., pp. 14l- 152 (Sylvester’s II%&, Vol. II, pp. 86-89 and 90-99) ;
J. W. L. Glaisher, Wn the number of partitions of a number into a given number of
parts,” Qumterly Journul, V& XL, 1909, pp. 57--143 ; ~CForm~l~ for partitions into given
elementu, derived from Sglvester’a Theorem,” ibid., pp. 275-348; “Formulae for the
number of partitions of a number into the elements 1, 2, 3, , .,, n up to n=9,” ibid.,
Vol. XLI, 1910, pp. 94-112: and further references will be found in MacMahon, Zoc. cil.,
pp. 59-71, and E. Netto, Lehrbuch der Combinattwik, 1902, pp. 146-158. Thus, for
example, the coefficient of 5~ in
1
(l-X)(1 -x2)(1-d)

is PQ(n)=~(n+3)1-~~+Q(-l)n+~cos2~;

as is emily found by separating the function into partial fractions. This function may
also be expressed in the forms
& (n + 3)2 + (* cos +>” - (3 sin $frn)2,
1 +[&n (n+6>]3 (1%(n+3)2l*
where [n] and (n} denote the greatest integer contained in n and the integer nearest to n.
These formulae do, of course, furnish incidentally asymptotic formula for the functions in
question. But they.are, from this point of view, of a very trivial character : the interest
which they posse;sf3is algebraical.

387
 to it. In the latter problems we are dealing with functions defined by
Dirichlet’s series. The study of such functions presents difficulties far more
fundamental than any which confront us in the theory of the modular
!functions. These difficulties, however, relate to the distribution 6f the zeros
of the functions and their general behaviour at infinity : no difficulties what-
ever are occasioned by the crude singularities of the functions in the finite
part of the plane. The single finite singularity of C(S), for example, t&e .~olz

ne precisely the order of the error

which it involves is in many cases a problem which still defies the utmost
resources of analysis. But to write down the dominant terms involves, as
a rule, no difficulty more formidable than that of deforming a path of integra-
tion over a pole of the subject of integration and calculating the corresponding
residue.
In the theory of partitions, on the other hand, we are dealing with
functions which do not exist at all outside the unit circle, Every point of
the circle is an essential singularity of the function, and no part of the contour
of integration can be deformed in such a manner as to make its contribution
obviously negligible. Every element of the contour requires special study ;
and there is no obvious method of writing down a “ dominant term.”
The difficulties of the problem appear then, at first sight, to be very serious,
We possess, however, in the formulae of the theory of the linear transformation
of the elliptic functions, an extremely powerful analytical weapon by means
of which we can study the behaviour off(x) near any assigned point of the
unit circle? It is to an appropriate use of these formulae that the accuracy
of our final results, an accuracy which will, lve think, be found to be quite
startling, is due.
1*3. It is very important, in dealing with such a problem as this, to
distinguish clearly the various stages to which we can progress by arguments
of a progressively ” deeper ” and less elementary character. The earlier results
are naturally (so far as the particular problem is concerned) supersededby
the later. l3ut the more elementary methods are likely to be applicable to
ot,her problems in which the more subtle analysis is impracticable.
We have attacked this particular problem by a considerable number of
different methods, and cannot profess to have reached any very precise con-
clusions as to the possibilities of each. A detailed comparison of the results
* See G. H. Hardy and,J. E, Littlewood, %xne problems of Diophantine approxima-
tion (II : The trigonometrical series aaaociated with the elliptic Theta-functions),‘” Actu
Mathematics,Vol. XXXVIT, 1914, pp. 193-238, for applications of the form& to. different
but not unrelated problems.

308
 Asymptotic Furmuh in Cumbinatory Analysis 279
to which they lead would moreover expand this paper to a quite unreasonable
length, But we have thought it worth while to include a short account of
two of them. The first is quite elementary; it depends only on Euler’s
identity
(1*31)
1 X4

(1 -x) (1 - x2) (1 - ti) . . . = 1+~+(l-x)2(1-x2)2-~~~m

-an identity capable of wide generalisation --and on elementary algebraical
reasoning. By these means we shew, in section 2, that
(F32) eAdn. < p (7%)< eBdn,
where A and B are positive constants, for all sufficiently large values of n+
It follows that
(P33) AJn<logp(n) < Bdn;
and the next qu estion which arises is the question whether a con&a& C exists
such that
(P34) logp (n) - C &I.
We prove that this is so in section 3. Our proof is still, in a sense,Cc ele-
mentary.” It does not appeal to the theory of analytic functions, depending
only on a general arithrnetie theorem concerning infinite series; but this
theorem is of the difficult and delicate type which Mews Hardy and Little-
wood h.ave called “ Tauberian.” The actual theorem req uired was proved by
us in a paper recently printed in these Proceedings*. It .shews that

(1.35) c ,I- 2 *.
\/6 ’
in other words that
(1*36)

where e is small when n is large. This method is one of very wide application.
It may be used, for example, to prove that, if p@j(n) denotes the number of
partitions of vx into perfect s-th powers, then

It is certainly possible to obtain, by means of arguments of this general

character, information about p (7h)more precise thati that furnished by the
formula (1*36). And it is equally possible to prove (l-36) by reasoning of a
more elementary, though more special, character: we have a proof, for example,
based on the identity

+ G. H. Hardy and S. Ramanujan, “Asymptotic formulae for the distribution of

integer8 of various types,” Proc, London Math. i%c., XVI, 1917, pp. 11243%
Ser. 2, vol.

309
 280 Ag-mptutic Formuh in Cbmbinatory Analysis
where 6 (v) is the sum of the divisors of Y, and a process of induction. But
we are at present unable to obtain, by any method which does not depend
upon Cauchy’s theorem, a result as precise as that which we state in the next
paragraph, a result, that is to say, which is ” vraiment asymptotique.”

1*4. Our next step was to replace (P36) by the much more precise
formula

The proof of this formula appears to necessitate the use of much more
powerful machinery, Cauchy’s integral (121) and the functional relation

This formula is merely a statement in a different notation of the relation

between h (7) and h (T), where
c + dT
T u=d=O, b=l, c=-1;
=CC$-b7)

viz h (7) = d(t) h (T)?

It is interesting to observe the correspondence between (1+41) and the

results of numerical computation. Numerical data furnished to us by Major
MacMahon gave the following results: we denote the right-hand side of
(l-41) by = (n).

n P (4 ‘iJ (4 rnlP

10 42 48-104 1*145

20 627 692.385 1 *x04

50 204226 217590*499 X*065

80 15796476 I 16606781 n567 I*061

It will be observed that the progress of m/p towards its limit unity is not
very rapid, and that P - p is always positive and appears to tend rapidly to
infinity.
* Tannery and Molk, Zoc.cit., p- 265 (TableXLV, 5).

310
 1.5. In order to obtain more satisfactory results it is necessary to con-
struct some aukiliary function F(X) which is regular at all points of the unit
circle save s = I, and has there a singularity of a type as near as possible to
that of the singularity of f(x). W e may then hope to obtain a much more
precise approximation by applying Cauchy’s theorem to f - F instead of tocf:
For although every point of the circle is a singular point off, x = 1 is, to put
it’ roughly, much the heaviest singularity. When x-1 by real values, f (cc)
tends to infinity like an exponential

when X = y#Pnif?7 t
p and q being co-prime integers, and r-1, If 1 tends to infinity like an
exponential

X = re2eni,
while, if
where 0 is irrational, If(z) j can become infinite at most like an exponential
of the type

The function required is

(l-51) F(x)
where
(l-52)

(1*53) C = 2+6 = rrd(B>, A, = J(n - &)a

This functiun may be transformed into an integral by means of a general
formula given by Lindelijft ; and it is then easy to prove that the Icprincipal
branch ” of F(z) is regular all over the plane except at x = 1 i; and that

* The statements concerning the “rational” points are corollaries of the formulae of
the transformation theory, and proofs of them are contained in the body of the paper,
The proposition concerning 5rratiouaI” points may be proved by arguments similar to
those used by Hardy and Littlewood in their memoir ,already quoted. It is not needed for
our present purpose. As a matter of fact it is generably true that f(z) + O when B is
irrational, and very nearly as rapidly as J(I -r). It is in reality owing to this that our
final method is so successful.
t E, Lindeliif, Le calcd u?es re'sidus et se5 u+ppZications d h thbrie des folzctions
(Gauthier-Villars, Collection Bore& 1905), p. 111.
1 We speak, o f course, of the principal branch of the function, viz. that represented by
the series (lfil) when x is small. The other branches are singular at the origin.

311
 282 Asymptotic Elxvaul~ in Combinatory Analysis

is regular for x= 1. If we compare (1~42)and (1*54), and observe that f (x;‘)

tends to unity with extreme rapidity when x tends to 1 along any regular
path which does not touch the circle of convergence, we can see at once the
very close similarity between the behaviour of f and F inside the unit circle
and in the neighbourhood of x- 1.

It should be observed that the term - I in (l-52) and (1.54) is-so far as
our present assertions are concerned- otiose : all that we have said remains
true if it is omitted ; the resemblance between the singularities off and F
becomesindeed even closer. The term is inserted merely in order to facilitate
some of our preliminary analysis, and will prove to be without influence on
the final result, ’

Appl$ng Cauchy’s theorem to f - F, we obtain

(l-55) ) = -_.-
PC32 -

where D is any number greater than

1*6, The formula (l-55) is an asymptotic formula of a type far more precise
tlhan that of (1.41). The error term is, however, of an exponentia1 type, and
may be expected ultimately to increase with very great rapidity. It was
therefore with considerable surprise that we found what exceedingly good
results the formula gives for fairly lsrge values of 12. For n = 61, 62, 63 it
gives
1121538*972, 1300121-359, 1505535*606,
while the correct values are
1121505, 1300156, 1505499.

The errors 33*972, - 34+641, 36%06

are relatively very small, and alternate in sign.

The next step is naturally to direct our attention to the singular point of
f(x) next in importance after that at x = 1, viz., that at x = - 1; and to
subtract from f (x ) a second auxiliary function, related to this point as F(x)
is to z = 1. No new difficulty of principle is involved, and we find that

(1’61) + 0 py,

312
 Asymptutk Fummlw in Combinatory Analys& 283
where D is now any number greater than QC. It now becomes obvious why
onr earlier approximation gave errors alternately of excess and of defect*

It is obvious that this process may be repeated indefinitely. The

singularities next in importance are those at x = e#rri and x = P; the next
those at x=i and x- - 4 ; and so on. The next two terms in the approximate
formula are found to be

an d

AS we proceed further, the complexity of the calculations increases. The

auxiliary function associated with the point x = e2pTpilqinvolves a certain
24q-th root of unity, connected with the linear transformation which must
be used in order to elucidate the behaviour of If(x) near the point; and the
explicit expression of this root in terms of ~3and 4, though known, is some-
what complex. But it is plain that, by tlaking a sufficient number of terms,
we can find a formula in which the error is

where, v is a fixed but arbitrarily large integer.

1.7. A final question remains. We have still the resource of making

v a function of 92, that is to say of making the number of terms in our
approximate formula itself a filnction of rz. In this way we may reasonably
hope, at any rate, to find a formula in which the error is of order less than
that of any exponential of the type P; of the order of a power of n,, for
example, or even bounded.
When, however, we proceeded to test this hypothesis by means of the
numerical data most kindly provided for us by Major MacMahon, we found
a correspondence between the real and the approximate values of such
astonishing accura,cy as to lead us to hope for even more. Taking 72= 100,
we found that the first six terms of our formula gave
190568944*783
+ 348.872
- 2598
+ ~685
+ a318
- ,064
190569293*996,
while p (100) = 190569292 ;

313
 so that the error after six terms is only 904. We then proceeded to calculate
p (200), and found
3, 972, 998, 993, 185*896
+ 36, 282m978
- 87#555
+ 5m147
+ I*424
+ 0*071
+ o*ooo*
+-~- 0.043 -
3, -972, 999, 029, 388*004,
and Major MacMahon’s subsequent calculations shetved that JI (ZOO) is, in fact,
3, 972, 999, 029, 388,
These results suggest very forcibly that it is possible to obtain a formula for
p (n), which not only exhibits its order of magnitude and structure, but may
be used to calculate its exact value for any value of n. That this is in fact so
is shewn by the following theorem.

Statement of the mait tl~eow72.

THEOREM. suppose that

where C and X~ are dejned by the equations (P53), for all positive integml
valwes of q ; that p is a positive integer less than and prime to q ; that QJ~,~is
a 24q-th root of unity, defined when p is odd by the formula
(lT21)

and when q is odd by the formula

(P722).
wp,,=(~)“xP~-Ia(p-l)+I?a(q-~) W-p.fPWjrriJ

where (a/b) is the symbol of Legendre and Jacobi-f, and p’ is any positive
integer tiuch that 1 + pp’ is divisible by q ; that
(1.73) A, (n) = IZmP,qenmPrilq;
(PI

and that a is any positive constant, and v the integral pa.rt of u Ja.
* This tern varnishes identically.
t See Tannery and Molk, lot. cit., pp. X04-106, for a con@& set of rules for the cal-
culation of the value of (a/b), which is, of coum, always I or - 1. Wheu both p and p
are odd it is indifferent which formula is adopted.

314
 Aqmptotic Formuh in Combinatory Analysis 285
Then
(1.74) p (4 = fi A,+, + 0 (n-i),
1

so that p (n) is, for all sujiciently large dues of n, the i?zteger nearest to

It should be observed that all the numbers A, are real. A table of A,

from q = 1 to q = 18 is given at the end of the paper (Table II).
The proof of this theorem is given in section 5; section 4 being devoted
to a number of preliminary lemmas. The proof is naturally somewhat
intricate; and we trust that we have arranged it in such a form as to be
readily intelligible. In section 6 we draw attention to one or two questious
which our theorem, in spite of its apparent completeness, still leaves open.
In section 7 we indicate some other problems in combinatory analysis and
the analytic theory of ,numbers to which our method may be applied; and
we conclude by giving sume functional and numerical tables: for the latter
we are indebted to Major MacMahon and Mr H. B. C Darling. To Major
MacMahon in particular we owe many.thanks for the amount of trouble he
has taken over very tedious calculations. It is certain that, without the
encouragement given by the results of these calculations, we should never
have attempted to prove theoretical results at all comparable in precision
with those which we have enunciated.

2. ELEMENTARY PROOF THAT &Jn< p (n)<eBJn FOR SUFFICIENTLY

LARGE VALUES OF n,
2% In this section we give the elementary prpqf of the inequalities
(1432). We prove, in fact, rather more, viz., that positive constants H and K
exist such that
pm)

for n 2 1’. We shall use in our proof only Euler’s formula (P31) and a
debased form of Stirling’s theorem, easily demonstrable by quite elementary
methods : the proposition that
n ! P/r@*
lies between two positive constants for all positive integral vaiues of n.

* Somewhat inferior inequalities, of the type

2’ Ldfll <p (n) <nB 1dn3,
may be proved by entz’rely eIementary reasoning ; by reasoning, that is to say, which
depends only on the arithmetical definition of p(n) and on elementary finite algebra, and
does not presuppose the notion of a limit or the definitions of the logarithmic or eupo-
neatial functions.

315
 2.2, The proof of the first of the two inequalities is slightly the simpler.
It is obvious that if
1
Qr @) xn = (1 - x) (1 - $2) l l l (1 c xr)

so that p, (n) is the number of partitions of r~ into parts not exceeding r, then
(Z-21) pT (n) = p,.4 (n> + p,+ (n - T) -f-p,-, (n - 279+ l l l l

We shall use this equation to prove, by induction, that

(2.22) p, (n) 2 rnr-1,

(r !)”
It is obvious that (2.22) is true for F = 1. Assuming it to be true for r = s,
and using (2#21),we obtain

p,+,(n)>& {TP-I + (42 -s- I),-1 + (n -2s - 2)$-l+ . ..}

.

2s
s!
( ) 2
ns
s(s+l) 1
ns- (n - s - 1)” + (n - s - 1)” - (12- 2s- 2)$

(s+ l)nS
s (s 5 1)
+ I l .

=(s+l) (S!)2={(s+1)!)2*
This proves (2%). N ow p (n) is obviously not lessthan pr (n), whatever the
value of r. Take r = C&z] : then
Cd 1 rPnl H
p t4 ap[hl (n) a2 n -{[&a]!)” ’ -12&lE,

by a simple application of the degenerate form of Stirling’s theorem men-

tioned above.

2.3. The proof of the

. second inequality depends upon Euler’s identity.
If we write

% wxn= (1 -x)2(l - J)2 . l l

(1 +a)2
we have
(2.31) qr (n) = qr+ (n) -I- 2q,+ (n - r) + 3q,-, (~1- 2r) + . 1, l

and
(2.32) p (n) = q1(n - 1) + qz(n - 4) + q3(12- 9) + . . . .
We shall first prove by induction that

(2.33) qr (n>< (n + v-
-. r - 1) ! (r !)”
(2 l

This is obviously true for r,= 1. Assuming it to be true for r = s, and using
(2.31), we obtain
1
q&1(4 < (2s-l)! (s!)2 {(n + s2)2s-1 + 2 (n + s2 - 8 - w1

+ 3 (n + S2- 2s - 2)25--l + . l .}*

316
Now m(na- 1)urn-2
b2< (u+ by”- 2~”-I-(a- b)“,
if rye is a positive integer, and a, b, and cr:- b are positive, while if a - b < 0,
and wx is odd, the term (a - b)” may be omitted. In this inequality write
m=%+l, a=~-+-~~--k~-k (k=O, 1,2, l ..), b=s+l,
and sum with respect to k. We find that
(2~+~)2~(~+1)~{(12+~~)~~-~+2(n+s~-s~l)~~-~+~~~}5(~+~~+~+~)~~+~;
and so
< {n+ (s+ l)*)“+l
qs+l(nk (2s+l)2s(sf 1)2(2s-1)!(s!)2k (2s+l)!((s+1)!}s*
Hence (2.33) is true generally.
It follows from (2.32) that
p(n)=q1(7a-1)+q2(ng4) +l-+ (&Jr-l)!
(qi’ p32r-1

But, using the degenerate form of Stirling’s theorem once more, we find
without difkulty that
1 2srK
(2r-l)!(r!)2c 4r! ’
where K is a constant. Hence

This is the second of the inequalities (291).

3. APPLICATION OF A TAUBERIAN THEOREM TOTHE DETERMINATION

OFTHECOXSTANT c.
3-l. The value of tOheconstant
C = lim log P (n)
dn ’
is most naturally determined by the use of the following theorem.

A
1% 9 (4 - Ix-
when x-1, then
log sn =log(a,+a,+...+a,)~2J(An)
when n+ 00.
This theorem is a special case* of Theorem C in our paper already
referred to.
Now supposethat
1

s(x)=(1-x)f(s)=):(1,(n)-P(n-l)}x~=(I_x2)(~-~~)(1-x~) l . .
l

* Lot. cit., p. 1% (with a = I),

317
 Then %=p(+p(n-l)
is plainly positive. And
(3*11)
1 z1 --N- ti’ 1 a 1 7r2
logg (S) =$ogi--= e
2 - xp 1 Yl-Lc” l-xl 7=-l
when x+ 1 +. Hence
(392) log P (n) =u*+a,+... +upC~n,
where C= 2~/J6=rI/(%), ag in (1*53).
3.2. There is no doubt that it is possible, by “Tauberian ” arguments, to
prove a good deal more abdut p (n) than is asserted by (3-12). The functional
equation satisfied by f (z) shews, for example, that

a relation far more precise than (3*11). From this relation, and the fact that
the coefficients in g (x) are positive, it is certainly possible to deduce more
than (392). But it hardly seems likely that arguients of this character will
lead us to a proof of (1.41). It would be exceedingly interesting to know
exactly how far they will carry us, since the method is comparatively ele-
mentary, and haa a much wider range of application than the more powerful
methods employed later in this paper. We must, however, reserve the dis-
cussion of this question for some future occasion.

4. LEMMAS PRELIMINARYTO THE PROOF OFTHE MAIN THEOREM,

4.1. We proceed now to the proof of our main theorem, The proof is
somewhut intricate, and depends on a number of subsidiary theorems which
we shall state as lemmas.
Lemmas concerning Farey’s series.
421. The Furey’s series of order m is the aggregate of irreducible rational
fractions
p/q W?xq<m),
* This is a special case of much more general theorems : see K. Knopp, Wrenzwerte
van Reihen bei der Anntiherung an die Konvergenzgrenze,” htauguraZ-DimwtutimP Berlin,
1907, pp. 25 et seq. ; K. Knopp, &‘Uber Lambertsche Reihen,” JournaE@4fu~h., Vol. CXLII,
1913, pp, 283- 315; cf. H. Hardy, LCTheorems connected with Abel’s Theorem on the
continuity of power series,” hoc. hadm~ if&ath.&c., Ser. 2, Vol. IV, 1906, pp. 247-265
(pp. 252, 253); Cl. H. Hardy, ‘%ome theorems concerning infinite Iseries,” &th. Ann.,
Vol. LXIV, 1907, pp. 77-94; G. I-I. Hardy, “Note on Lambert’s series,” Proc. London
Math. Sm., Ser. 2, Vol. XIII, 1913, pp. 192-198.
A direct proof is very easy: for
vx~-‘(l~x)<I~xv<Y(l~x),
1 XW 1 xv+l
-c yl%Yw=r_;l;yp
l-s

318
 Asymptotic Form&z in Cumbinatory Analysis 289
arranged in ascending order of magnitude. Thus

LEMMA 4%. If Y/q, p’/q’ are two successive teq*ms of a Furey’s series, thelz
(4.211) p’q - fl.lq8= 1.
This is, of course, a well-known theorem, first observed by Farey and
first proved by Cauchy *. The following exceedingly simple proof is due to
Hurwitzt.
The result is plainly true when m = 1. Let us supposeit true for m = k;
and let pip, p//q’ be two consecutive terms in the series of order k.
Suppose now that p”/q” is a term of the seriesof order k: + 1 which falls
between p/q and p//q’. Let
p’lq - pqff = h > 0, p’p” - p”q’ = /d > 0.
Solving these equations for p”, q”, and observing that p’q -pq’ = 1, we obtain
p” = /&p+ Ap’, q” = #Aq+ xq’.
Consider now the aggregate of fractions
(PP -f- ~PY~c14+-w
where x and r~)are positive integers without common factor. All of these
fractions lie between p/q and p//q’; and all are in their lowest terms, since
a factor common to numerator and denominator would divide
x = q kP + WI - P kl+ w
and 111=Pf bq + w - !i cw -thP’).
Each of them first makes its appearance in the Farey’s series of order
clQ+kq’, and the first of them to make its appearance must be that for
which A = 1, ,U= 1. Hence
p” =p+p’, q” = q -t- q#,
p/f q - pqlf = p’q” -p”q’ = 1,
The lemma is consequently proved by induction.
LEMMA 4*22. Suppose that p/q is a term of the Farey’s series of order. m,
and p”/q”, p’/q’ the adjacelzt terms on the left and the Tight : and let & dewte
the i&em&

* J. Frey, Wn a curious property of vulgar fractiona,” PJLiZ. Msg., Ser. 1, Vol. XLVII,
1816, pp. 385, 386; A. IL Csuchy, ~9kmonatration d’un th&w&nle curieux HUF les nom-
bres,” Exercips de mattimatiques, TM I, 1826, pp. 114-I 16. Cauchy’s proof was fir&
published in the B&et& de Zclt X&&e’ Philomatipu~ in 1816.
t A. Hurwita, Ii Ueber die angenB;herte Darstellung der Zahlon durch rationale Eruche,”
Mak Ann., Vol. XLIV, 1894, pp. 417-436.
$ When p/g ia O/l or l/l, only the part of thie interval inside CO,1) ia to be taken ;
thusj0,1 is 0, l/(m+l) andjIll is t-l/(m+l), 1.

319
 290 Asymptotic Furmukt! in Combinatory AnaZy&
Then (i) the intervals &4 exactly fill up the continuum (0, l), and (ii) the length
of each of the parts into which & is divided by p/q + is greater than 1/2mq
and lessthan llmq.
(i) Since
1 1 l P’q-Pq’ Pr P
-+
q(q+qI) q’(q’+q)=$= -QPI=pp
the intervals just fill up the continuum.
(ii) Since neither q nor q’ exceeds m, and one at least must be lessthan nz,
we have
1 1
q(4) ’ 2nzq ’
Also q + q’ > nz, since otherwise (p +p’)/(q + q’) would be a term in the series
between p/q and p’/q’. Hence
1 1
q (Q + 4’) < T ’
Standard dissectiolz of a circle.
4-23. The following mode of dissection of a circle, based upon Lemma
4.22, is of fundamental importance for our analysis.
Suppose that the circle is defined by
X = Beztie (0 < 0 < 1).

Construct the Farey’s series of order m, and the corresponding intervals &, q.
When these intervals are considered as intervals of variation of 8, and the
two extreme intervals, which correspond to abutting arcs on the circle, are
regarded as constituting a single interval &, the circle is divided into a
number of arcs

where q ranges from 1 to no and 23through the numbers not exceeding and
prime to g-t* We call this dissection of the circle the dissection En&.

Lemmas f ram the theory of the linear transforma tion of the elliptic
modular jknctions.
4% LEMMA 4*31. Supposethat q is a positive integer; that p is a positive
integer not exceeding and prime to q; that p’ is a positive integer such that
I + pp’ is divisible by q; that op,q is defined by the formula (1.721) or (1~722);
th>at
2rz 2prri 2~ 213%
X =exp
(
--+-
Q Q >
? x’=exp --+----
( QZ Q >
1

+ See the preceding footnote [footnote $ of p. 2891.

+ p = 0 occurring with q = t only.

320
 Aymptotic Furmul~ in Combinatory Analysis 291
where the real part of z is positive; and that
1

Then

This lemma is merely a restatemen t in a d&rent notation of well-known

formulae in the transformation theory.
Suppose, for example, that p is odd. If we take

a=p, b=-q, c=- 1 +PP’ 1 d =- p’

3
9
so that ad - bc = 1; and write
f-&l = qa = eznir, r&r = Q2 = e2tiT,

so thst 2 + ia T ,P’I 6
7=9 4’ P P3
then we can easily verify that
c + dr
T =-
a+br-
Also, in the not#ation of Tannery and Molk, we have

and the formula for the linear transformation of h (7) is

{$(a-l)-,i,[u(b-c)+bd(a’d-l)])ri

where d(a + by) has its real part positive *. A little elementary
1 2/(a;-tbr)h(~),

algebra will
shew the equivalence of this result and ours.
The other formula for wPsQ may be verified similarly, but; in this case we
must take
u =- p, b=q, c=--, 1 +PP’ d=p’.

We have included in the Appendix (Table ;) a short table of some values

Of %F or rather of (log oP, J/&.
hMMA 4.32. The function f(x) sati#es the equation
(4-321)
79
6qa log (l/x& ’ (xa’ ‘)’
where
(44322)
f 47T2 2 hi
xP9 Q
= xe-w!f, xrp, q = exp - +p
t q2 1% WP,J 4 i l

* Tannery and Molk, Zoc, cit., pp. 113, 267.

321
 292 Asymptotic Form&z in Combinatwy Analysis
This is an immediate corollary from Lemma 4*3X, since

T 7T2 27r 2p’k

-=
= X’P, Q’
-a- --

12qz 6@log (l/~~,~) ’ x’ = exp qg + q

If ore observe

we see that, if s tends to ez@iq along a radius vector, or indeed any regular
path which does not touch the circle of convergence, the difference

tends to zero with great rapidity. It is on this fact that our analysis is based.

Lemmas concerning the auxiliary fun&n. & (2).

4-41, The auxiliary function Pa (x) is defined by the equation

Fa (x) = ci qa (n) xn,

1

where qa (n) =g cashJ1”- - I,

n

X,= 4(7a - &), a > 0.

LEMMA 4*41. Suppose that a GUI! is wiade along the segment(I, 00) in the
plane of x, Then,Fa (x) is regular at all points inside the region thzcsdefined.
This lemma is an immediate corollary of a general theorem proved by
Lindelijf on pp* 109 et seq.of his CaZcuZdes r&dus*.

The function

satisfies the conditions imposed upon it by Lindelijf, if the number which he

calls a is greater than &; and
l

(4*411) Eh (x) =I”za&z + (2) dz,

a-h

if X = r&Q, 0 < t9c 2w, ti = exp {z (log r + if?)).

442. LEMMA 4*42. Suppose th,at D is the region defined by the in-
equalities
- 7r < - 8, < 8 < 0, < 7r, r. < r, 0 < 9-0< 1,

* Lindelijf gives references to Mellin and Le Roy, who had previously established the
theorem in less general forms.

322
 and that log (I/x) h as its principal value, so tlmt log (l/x) is one-valued, and
its square root two-valued, Iin D. Further, let

that value of the square mot beiy chose)2 which is positive when. 0 < x < 1.
Then

ia regular inside D*.

We observe first that, when 8 has a fixed value between 0 and 27f, the
integral on the right-hand side of (4.411) is uniformly convergent for
&<~<a,. Hence we may take a = +$ in (4411). We thus obtain

where the 2/($ and 2/(- it) which o~eur in *a (& + it) and qa (& - it) are to
be interpreted as &mi& and e-tniJt respectively. We write this in the form

= x, (x)-I-0, (x)-I-0, (x),

say, where X,(x) = ixa’+ aDxi6 qa (& -I- it) dt.
i0
Now, since xit = e-h!, 1x-it 1= eel!,
1 I
the functions 0 are regular throughout the angle of Lemma 4942, And

The form of this integral may be calculated by supposing x and ,U positive,

when we obtain

Hence

ww xu (4 = I/b log W)} xA [exp I4 loga;llz)/ - l] = %a (x>,

and the proof of the lemma is completed.
* Both F,(z) and xa (x) are two-valued in D. The value of F@ (x) contemplated is
naturally that represented by the power-series.

323
 294 Asymptotic Formulae in Cumbinatory Analysis
Lemmas 4-41 and 4.42 shew that x = 1 is the sole finite singularity of
the principal branch of Fa (x).

4m43, LEMMA 4e43. S~ppe that P, &, and A ore positive constants,
8, being less than T. Then
) Fa (4 I< K = K (P, 4, A),
f or ocr& e1~e~2T-e1, O<a<A.
We use K generally to denote a positive number independent of tr: and
of a. We may employ he formula (4m411). It is plain that

where 7 is the imaginary part of z. Hence

4944. LEMMA 4-44. Let. c be a circle whose centre is x = 1, and whose

radius 8 is less than unity. Then
1 Fa (4 - Xa Cc) I < Kaa~

if x lies in c and 0 < a < A, K = K (6, A) being as before independent of x and

of a.
If we refer back to (4-421) and (4*422), we see that it is sufficient to prove
that
10, (x) 1< Ka2, 1a2 (x) ] < Ka2 ;
and we may plainly confine ourselves to the first of these inequalities. We
have

Rejecting the extraneous factor, which is plainly without importance, and

integrating by parts, we obtain

@(x) = Ia @(t) ‘Osha;jit) - ’ dt,

0
ixit log x ST&t e - +i + St

where @ (t) = -
e-&ri+2at- 1 + (e-*&+2rt_1)g’

< KeW It follows that

1tD (t) I< Ke-lrt ;

324
 < Ki,” e-@ (coshg-cosg) dw

6. PROOFOFTHEMAINTHEOREM.
5I. We write

where C = T J+, xp,q= xe-2pnilq; and

(5*12) @(4 =f (4 - 253Fp, q (4
QP
where the summation applies to all values of p not exceeding q and prime
to q, and to all values of q such that
(5-13) 1 <q c v = [a &I,
u being positive and independent of n. If then
(5.14) Ii;7#q (4 = ~~p,q~n~n>
we have
(5.15)

where r is a circle whose centre is the origin and whose radius R is less
than unity. We take
(5.16) J&l-@
7%’
where p also is positive and independent of n.

Our object is to shew that the integral on the right-hand side of (5-15)
is of the form 0 @I*); the constant implied in the 0 will of course be a
function of a and /% It is to be understood throughout that O’s are used in
this sense; 0 (1), for instance, stands for a function oj X, n, p, Q,Q, and @ (or
of some only of these variables) which is less in absolute value than a number
K = K (u, p) independent of x, n, p, and q.
We divide UP the circle F’, by means of the dissection E, of 4.23, into
arcs fPIq each associated with a point IWtiiq ; and we denote by yP,q the arc
of r complementary to &. This being so, we have

= SJ p,q - 2 jp,9)

325
 296

say. We shall prove that each of these sums is of the form 0 (n-1) ; and
we shall begin with the second sum, which only involves the auxiliary
fuuc tions E7.
Proof that cj, I q = 0 (d).
6-21. We have, by Cauchy’s theorem,

where c& consists of the contour LMNM'L' shewn in the figure. Here
L and L’ are the ends of & q, LM and WI;’ are radii vectores, and MNlK’ is
part of a circle I& whose Radius R, is greater than 1. P is the point evtila;
and we supposethat R, is small enough to ensure that all points of LM and
XL’ are at a distance frown P less than -i. The other circle c shewn in the
figure has P as its centre and radius +, We denote LM by -P,q, M’L’ by
plP, q, and AfNM’ by yP,q: and we write

5*22. Suppose first that x lies on yP,q and outside c.. Then, in virtue of
(5.11) and Lemma 4m43,we have
(5221) Fp,q (4 = 0 w-
If on the other hand x lies on Y~,~, but inside c, we have, by (591) and
Lemma 4.44,

326
But, if we recur to the definition of xn (a) in Lemma 4.42, and observe that
a2 a2log (l/Y)
exp 4 log (1/x) = exp 4 [{log (l/?-)y + 82-J< 1
if
reie and r > 1, we seethat
II; 3

(5.223) xp,q (4 = 0 Wd
on the part of vP,* in question. Hence (5.221) holds for all vP,qmIt follows
that
jl,, q= 0 (R-” d9),
(5.224) 2 I
j, q = 0 (RI-” 2 qt) = 0 (ng R,-“) *.
This sum tends to zero more rapidli than any power of n, and is therefore
completely trivial.
The contrihtions of lQ& and cj$,q.
5m231. We must now consider the sums which arise from the integrals
along mPIQand & ; and it is evident that we need consider in detail only
the first of these two lines. We write

sav,
In the first place we have, from (5*222),
l

-4 R, dr -&)$-1),
-Z

U&W = 0 (4

Oq I R rnfl
( . >

since
--n
(5.2312) R 9-n= 1 -- rB = 0 (1) .
( n )
Thus
(5.2313) Xj$,, = 0 {n-l Q<zfdnJ q+) = 0 (n-3).

5*232. In the secondplace we have

It is plain that, if we substitute y for xe-aprrilq,then write x again for y, and

finally substitute for XC/~ its explicit expression as an elementary function,
given in Lemma 4*42,,we obtain
(5.2321)
1 I

jl’ P? Q =O(&)~{E(x)-l],/(log~jx-n-~~dm=O(Jq)
: J9

* Here, and in many passages in our subsequent argument, it is to be remembered

that the number of values of p, corresponding to a given q, is lessthan q, and that the
number of values of Q is of order LJ1z. Thus we have generally
zO(Q8)=0( z p+l) = 0 (da+l)*
q-=Wn)

327
 say, where
(5*23211) E(x) = exp
rr2
i 6y” log (l/x) 1 ’
and the Path of integration is now a line related to x = 1 as P~,~ is t)o
X = ew+ : the line defined by x = M, where R< r < R,, and 8 is fixed and
(by Lemma 4.22) lies between 1/2qv and l/q~.
Integrating J by parts, we find
(5.2322)

( n -2$) J= - [{E(x) -1) &g~~ x-fi+h]I:

-8 J{E(x)-l}
+g2 /E(x) (log;)-tx-n-(Bdx= J1+ J,+&,
say.

5m233. In estimating JI, &, and J3, we must bear the following facts in
mind.
(1) Since j x 12 R, it follows from (5*2312) that 1x I+ = 0 (1) throughout
the range of integration.
(2) Since 1 - R = a/n and l/Zqv < 8 < l/yv, where v = [cx 2/n], we have

log(;)=0(:,I>)
1
when Y = R, and - = 0 (q d?2),
log U/x)
throughout the range of integration.
7T2log (l/r)
(3) Since I E (4 I = exp 6q” [/log (l/r)}” + tl”] ’
E (lz’) is less than 1 in absolute value when r > 1. And, on the part of the
path for which r < 1, it is of the form
1
exPO
a
Y no ( >
= exp 0 (1) = 0 (1).

It is accordinglv u of the form 0 (1) throughout the range of integration.

5934. Thus we have, first
(5.2341)
J1 = 0 (1) 0 (1) 0 (RI-“) + 0 (1) 0 (q-’ c-t) 0 (1) = 0 (q-)x+),
secondly

(5.2342) Jz = 0 (1) 0 (qhf) [z-&i = 0 (q*,n-p),

.

328
 Asymptotic Furmzch in Cumbinatory Analysis 299
and thirdly

(5m2343) J3 = 0 (q-2) 0 (I) 0 (q&a*) 1;’ r+& = 0 (q-*n-‘).

From (5*2341), (5*2342), (5*2343), and (59322), we obtain

J= 0 (g-b -q + 0 (&2-f);
and, from (5*2321), jt/p,q = 0 (n-Q) -I- 0 (yn-f).
Summing, we obtain
(5*2344) Xjt’p,q= 0 (n-9 C q) + 0 (d 2 q2)
!I< WIJ~~ q < 0 (h)
= 0 (n-)) + 0 (12-f) = 0 (n-q.

5,235, From (5#2311), (5*2313), and (5.23441, we obtain

(5.2351) r: jzp,q = 0 (n-f) ;
and in exactly the same way we can prove
(512352) z& = 0 (n-f).
And from (5m212), (5*224), (5-2351), and (5*2352), we obtain, finally,
(5m2353) c j,, q = 0 (12-i).

Proof that EJp q = 0 (n-*>.

5-31. We turn now to the discussion’of

= J1p,q-I-J2p,p+ J3,,p,
say, where PP, Q(X) = 01>,!l

xp,a (4 = XP,Q (4 -t PP,Q (4 = PP, Q w E (XP# !A

E (x) being defined as in (5*23211).

D~wu&I~ of I$ J2p, q and C Jsp ,qrn

6.32, The discussion of these two sums is, after the analysis which
precedes, a simple matter. The arc & is less than a constant multiple of
I/q 2/n; and xvn. = O(1) on Ep,q. Also

lc,P,&) -xP,&) I= 0 (Y 3,
by (5,222) ; and
(5.321) $/qb+J = 0 (q-,0,
since / xp,q I= R = 1 - (B/n>, ] am xp, q 1< l/qv.

329
 Hence J$ I Q= 0 (q - 8, - *),

and Jap 3q = 0 (q-%-‘1,

(5,323) ZJ 3p 1 q = 0 (n-g 2 1) = 0 (xi-)).
4 < 0 (W

Discussion of Z JI,, B l

5m33, From (4.321) and (5m2221), we have

(5.331)

if/z/<l,and dp,q = exp -

47r2 log (1/R)

Now
’ x’p*q’ = exp [- if {[log (l/B)]2 + q 1 ’
tvhere B is the amplitude of x~,~, Also

q” ([log (1/R)]2+ B”) = 0 {&+&)} =0(5)

while log (l/Ii) is greater than a constant multiple of X/n. !l’here is therefore
a positive number 6, less than unity and independent of /n and of q, such that
I4,qk~;
and we may write fit K&4) = 0 ( 14, q I )a
We have therefore

E(~~,,)~(~‘~,q)=o(lxrp,qI-~,o~~~~,q/)=~(/~~,q~~Q)=O(1);
and so, by (5m321),
f (4 - xp,q (4 = 0 wq> 0 1+--q 1) 0 (1) = 0 w9*
t
And hence, as the length of fp,* is of the form 0 (l/q Jn), we obtain
J1P,Q= 0 (q-l n-g>,
(5.332) CJ 1 p, q= O(n-2 C l)= 0 @-a).
q-a (ha)
5-34. From (5*311), (5m322), (5*323), and (5W332), w obtain
(6*341) ZJ p,q = 0 (n-f).

330
 Completion of the prooJ:
5*4. From (5*15), (5*17), (5m2353), and (5*341), we obtain
(541) P( 12 ) -y Yhpqn=O(n-+
433 "
Eut

where

Al: that remains, in urder to complete the proof of the theorem, is to shew
that
cash (Cx,/q) - 1
mal; be replaced by
”
~&%zlq l

and in order to prove this it is only necessary to shew fhat

On differentiating we find that the sum is of the form

q<zdN)
@{o
($)+0(-k)}
=0{in<&9-j=0wa>*
Thus the theorem is proved.

6. ADDITIONAL REMARKS ON THE THEOREM.

6-l. The theorem which we have proved gives information about JI (n)
lvhich is in someways extraordinarily exact, We are for this reason the more
anxious to point out explicitly two respects in which the results of our
analysis are incomplete.
6#21. We have proved that
p (n) = LQs + 0 (n-i),
where the summation extends over the values of 4 specified in the theorem,
for every fixed value of a; that is to say that, when a is given, a number
K = K (01)can be found such that
1p (n) - sA,+, 1~ Kn-t
for every value of n/. It follows that
(6.211) P (n> = (CA*&?L
where {x} denotes the integer nearest to x, for ~2 an,, where n, = 110
(0~)is
a certain function of cy,
The question remains whether we can, by an appropriate choice of a,
secure the truth of (6.211) for all values of n, and not merely for all sufficiently
large values. Our opinion is that this is possible, and that it could be proved
to be possible without any fundamental change in our analysis. Such a proof

331
 302 Asynaptotic Fwmul~ in Combinatory Analysis
would however involve a very careful revision of our argument. It would be
necessary to replace all formulae involving O’s by inequalities, containing only
numbers expressed explicitly as functions of the various parameters employed.
This process would certainly add very considerably to the length and the
complexity of our argument. It is, as it stands, sufficient to prove what is,
from our point of view, of the greatest interest; and we have not thought it
worth while to elaborate it further.

6*22. The second point of incompleteness of our results is of much greater

interest and importance. We have not proved either that the series

%A Q+ q
I
is convergent, or that, if it is convergent, it represents p (n). Nor dcies it seem
likely that our method is one intrinsically capable of proving these results,
if they are true -a point on which we are not prepared to express any definite
opinion.
It should be observed in this connection that we have not even discovered
anything definite concerning the order of magnitude of A, for large values
of 4. We can prove nothing better than the absolutely trivial equation
A*= 0 (Q), On the other hand we cannot assert that A, is, for an infinity of
values of 4, effectively of an order as great as q, or indeed even that it does
not tend to zero (though of course this is most unlikely).
6.3. Our formula directs us, if we wish to obtain the exact value of up(71)
for a large value of n, to take a number of terms of order l/n. The numerical
data suggest that a considerably smaller number of terms will be equally
effective ; and it is easy to see that this conjecture is correct.

Let US write @=4rrI/@)=4C,

and let us supposethat a < 2. Then

p-
[ 1
$ 9

2 A,+,= $0 (q*) 0 (-$ 0 (ecdnlq)= O(Gl Jpecinlq)

I.c+1
Y
= 0 (J-l
Jx ec Jnbdlx: ,
rh F
since dqecNQ decreasessteadily throughout the range of summation*.
Writing 2/n/y for z, we obtain

* The minimum occurs when q is about equalto 2C&

332
It follows
32
that it is enough, when n is sufficiently
Pd
[ 1
log 12
large, to take

terms of the series. It is probably also necessctryto take a number of terms

of order dn/(log 92) ; but it is not possible to prove this rigorously without
a more exact knowledge of the properties of A, than we possess.

6.4. We add a word on certain simple approximate formulae for log p (n)
found empirically by Major MacMahon and by ourselves. Major MacMahon
found that if
(6.41) . log,,p (n) = l(n + a> - %b,
then u% is approximately equal to 2 within the limits of his table of values
of p (n) (Table IV). This suggested to us that we should endeavour to find
more accurate form& of the same type. The most striking that we have
found is
(6-42) logmp (4 = $“-{Il(n + 10) - a,> ;
the mode of variation of a, is shewn in Table III.
In this connection it is interesting to observe that the function
13-“9 (n)
(which ultimately tends to infinity with exponential rapidity) is equal to 9’73
for n = 30000000000.

7. FURTHER APPLICATIONSOF THE METHOD,

7.1. We shall conclude with a few -remarks concerning actual or possible
applications of our method to uther problems in Combinatory Analysis or the
Analytic Theory of Numbers.
The class of problems in which the method gives the most striking results
may be defined as follows. Suppose that Q(n) is the coefficient of AF in the
expansion of F(s), where F (11;)is a function of the form

(7.11)

f(x) being the function considered in this paper, the a’s, b’s,a’s,and 6’s being
positive integers, and the number of factors in numerator and denominator
being finite ; and supposethat 1F(x) 1t ends exponentially to infinity when x
tends in an appropriate manner to some or all of the points esp”i/g. Then our
method may be applied in its full power to the asymptotic study of Q(n), and
yields results very similar to those which we have found concerning p (n).

* Since f(-X) , uYx2>~3

=f(x>fl
the arguments with a negative sign may be eliminated if this ia desired.

333
 304 Asymptotic Forwulce in Colnbinatory Analysis
Thus, if
f(x)=(I+I)(1+x2)(1+x3)
F (4 = ff, “‘=(l-G)(l I - ;)(l-a’) ,
. . l

so that 4 (n) is the number of partitions of n into odd parts, or into unequal
parts *: we find that

q b) = al djl
d Jo [ir 2/(3 (a + +$)}I

The error after [OL&J] terms is of the form 0 (1). We are not in a position to
assert that the exuct value of 4 (fir) can always be obtained from the formula
(though this is probable) ; but the error is certainly I;dunded.

If f (x2) f (4f (4
P(xQm= -If-$jp-=(1+x)(l+x3)(1+~~) l rnej

so that Q(n) is the number of partitions of 12into parts which are both odd
and unequal, then

The error is again bounded (and probably tends to zero), ‘i

Jqx) z-ccf w2 1
If
f(x2) 1-2x+2ti-2x”+...’
q (n) has no very simple arithmetical interpretation ; but the series is none
the less, as the direct reciprocal of a simple %-function, of particular interest.
In this case we find
1 d earln j/3 d e+rEJjb
--

4 (4 = G &?y --& + G cos (Qnr - Q 71-J & & + ’ ” l

The error here is (as in the partition problem) of order 0 (n-t), and the exact
value can always be found from the formula.
7#2. The method may also be applied to products uf the form (741) which
have (to put the matter roughly) no exponential infinities?! In such casesthe
4
approximation is of a much less exact character. On the other hand the
problems of this character are of even greater arithmetical interest.
The standard problem of this category is that of the representation of a
number as the sum of s squares, s being any positive integer odd or event.
We must reserve the application of our method to this problem for another
occasion; but we can indicate the character of our main result as follows.
* Cf. MacMahon, Zoc, cit., 1~ II. We give at the elld of the payer -a table (Table V) of
the values of q (n) up to n = 100, This table way calculated by Mr Darling.
t As is well known, the arithmetical difficulties of the problem are much greater when
s is odd.

334
 If rs (72) is the number of representations of n as the sum of s squares,
we have
F (x) = xrg @) xn = (1 + 2X + 2& + ,, l )g= ${~ff$.-- = ‘IB!$p.
-x
We find that
d
(7m21) 7’s(4 = &)

where cq is a function of q and of n of the same general type as the function

A, of this paper. The series

is absolutely convergent for sufficiently large v&es of s, and the summation

in (7.21) may be regarded indifferently as extended over all values of q or
only over a range I< q < tl &. It should be observed that the series (7.22)
is quite different in form from any of the infinite series which are already
known to occur in connection with this problem.
7.3. There is also a wide range of problems to which our methods are
partly applicable. Suppose, for example, that

F Cxl = xp2 Cot’)

xsb= p -X

so that p2(n) is the number of partitions of n into squares. Then F (LC)is not an
elliptic modular function; it possesses no general transformation theory: and
the full force of our method cannot be applied. We can still, however, apply
someof our preliminary methods, Thus the “Tauberian” argument shewsthat
logp”(n)-2-~3&{~(;)}Q.+.
And although there is no general transformation theory, there is a formula which
enables us to specify the nature of the singularity at it’ = 1. This formula is

f+)=” J(s) exp {$5(-q

x ii (1 - 2e-2TJ W) cos 2r &l/Z) + e+ 4 W~J}.

1

By the use of this formula, in conjunction with Cauchy’s theorem, it; is

certainly possible to obtain much more precise information about p2 (n), and

The corresponding formula for p8 (n), the number of partitions of PZ,into

perfect s-th powers, is .

p”(n.) w (2T) --H8f1)

where

335
 306 Asymptotic Formula &E Cumbinatory Analysis
The series (7421) may be written in the form
d8 &-l r, OS
.2&-q pw+l,
TFJ P*Q Q
where oP, q is always one of the five numbers 0, eg&, e-t&, -@, - Key. When
s is even it begins

It has been proved by Ramanujan that the series gives an emct repre-
sentation of r# (n) when s = 4, 6, 8 ; and by Hardy that this is also true when
S= 3, 5, 7. See Ramanujan, I( On certain trigonometrical sums and their
applications in the Theory of Numbers “; Hardy, “On the expression of a
number as the sum of any number of squares, and in particular of five or
seven f
TABLE I mPIq.
P Q h3 wp,ql=i P Q 1% WP,d”fi P Q
1 1 0 3 11 3122 8 7118
1 2 0 4 tl 3/22 11 -19p
1 3 l/l8 5 I? - 5122 13 - 7118
2 ‘4 - l/l8 6 3t 5122 14 - l/90
1 118 7 T1 - 3122 1 - 29132
3 ii -218 8 1? - 3122 3 - 27132
1 l/5 9 ?I - 5122 5 - 5132
2 1! 0 10 ii2 - 15/22 7 - 3132
3 >9 0 1 55172 9 3132
4 jf - l/5 5 99 - 1172 11 5132
1 6 5118 7 t3 l/72 13 27132
5 11 - s/l8 I1 ii3 - 55172 15 29132
1 7 5114 1 11113 1 - 14117
2 ts 1114 2 t1 4113 2 8117
3 31 - l/14 3 33 l/13 3 5117
4 I? l/14 4 99 -l/13 4 0
5 ?1 - l/M 5 ‘)I 0 5 r/n
6 ii - 5/14 6 ?3 - 4113 6 5117
1 7116 7 31 4113 7 l/17
3 5’) l/l6 8 19 0 8 - s/17
5 ?? - l/l6 9 11 l/13 9 8117
7 ii T/l6 10 3? -l/13 10 -l/17
1 - 14127 11 $f - 4113 11 - 5117
2 ?P 4127 12 ii - 11113 12 - 1117
4 ?1 - 4127 1 13/14 13 0
5 I9 4127 3 If 3114 14 -5117
7 93 - 4127 5 31 3114 15 - 8117
8 16 - 14/27 9 tt - 3114 X6 14115
1 315 11 I? -3114 1 - 20127
3 ?I 0 13 1”5 - 13114 5 Z/27
‘7 t? 0 1 w 7 - 2127
9 1;’ - 315 2 It 7118 11 2127
1 15122 4 13 19/m 13 - 2127
2 39 S/22 7 !9 - 7118 17 20127

336
A4 = 2 CO8 &arr - +>.

4 =2co0(+2v- $) + 2 CO8 Qnn.

4 = 2 cos (&m - &m)*

A7 = 2 cos (Qnlr - i$T) + 2 COS (+?i$T - &W) + 2 CO8 (+m +&p).
AS = 2 COB (&nr - 25 T) + 2 COB ($nr - &“).
A9 = 2 cos (@aa - &@) + 2 COB(@T - #p) + 2 cos (@an +&T).
Am= 2 CO8 ($nr - @) + 2 COB +.

It may be observed that

4 =O(nz 1, 2 (mod 9, A:, =o (nd, 3, 4 (mod 7)),
Aro=O(n= 1, 2 (mod w, dll=o (n+ 2, 3, 6, 7 (mod II)),
AIs= (72~ 2, 3, 5, 7, 9, 10 (mod 13)), Al,=0 (n= 1, 3, 4 (mod rJ)>,
Ale=0 (n=O (mod Z)), A17=0 (nrl, 3, 4, 6, 7, 9, 13, 14 (mod 17I> ;
while A 1, Aa, A 3, 84, A,, AS, A g, Ala, A16, and Al8 never vanish.

TABLE III : Iog,,p (~2)= J# {&n + 10) - a,}.

11 %a %
1 3.317 4m148
3 3*176 4.364
10 3’011. moooO 4,448
30 2.951 300000 4-267
100 3m036 1000000 3.554
300 3.237 3cmom 2.072
1000 3.537 loooooor) - X*188
3000 3.838 3oooom - 6.796
a0

331
 308

TABLE IV*: p (~2).

1..* 1 51... 239943 lOl.** 214481126 EL.. 45060624582
2 2 52.u 281589 102. . 241265379
l 152.e. 49686288421
3::: 3 53... 3299:3 1 103... 271248950 153*.. 54770336324
4... 5 54... 386155 104... 304801365 x54... 60356673280
5 l . . 7 55... 451276 105... 342325709 155.,. 66493182097
6 l ** 11 56.o 526823 106.u 384276336 156.m. 73232243759
7 *.a 15 57... 614154 107... 431149389 157... 80630964769
8 *.a 22 58.u 715220 ‘108.u 483502844 158... 88751778802
9 30 59,.. 831820 109... 541946240 159*.. 97662728555
10::: 42 6@,.. 966467 HO... 60’7163746 160... 107438159466
II,.. 56 6L.. 1121505 ill... 679903203 161... 118159068427
12... 77 62.u 1300156 112... 761002156 162... 129913904637
13... 101 63,., 1505499 113.,. 851376628 163.e. 142798995930
14. .* 135 64... 174I630 114... 952050665 164.u 156919475295
15... 176 6X.. 2012558 ll5... 1064144451 165... 172389800255
16. . . 231 66... 2323520 116.., 1188908248 166... 18933482S579
K.. 297 67.u 2679689 117.*. 132771UO76 167... 207890420102
18... 385 68... 3087735 118... 1482074143 X68,.. 228204732751
19. rn. 490 69.u 3554345 x19... 1653668665 169.o 250438925115
2a.. 627 70,. . 4087968 120... 1844349560 r70... 274768617130
2L.. 792 71... 469iZO5 121... 2056148052 l71..* 301384802048
22... 1002 72.e. 5392783 122... 2291320912 172... 330495499613
23... 1255 73... 6185689 123... 2552338241 P73... 362326859895
24,.. 1575 74... 7089500 124... 2841940500 174... 397125074750
25,.. 1958 75.*. 8118264 125... 3163127352 175... 435157697830
26.*. 2436 76.u 928909 1 126.,, 3519222692 176.u 4i6713857290
27.u 3010 77... 10619863 127, . . 3913864295 177,. . 522115831195
28, ,a 3718 7L 12132164 128.,. 435 1078600 178s.. 571701605655
29.u 4565 79.. . 13848650 129.., 4835271870 179.m 625846753120
30... 5604 m,.. 15796476 130#.. 5371315400 180s.. 684957390936
31... 6842 8L.m 18004327 l-31... 5964539504 18L.. 749474411781
3s.. 8349 82, * 20506255
l 132.v 6620830889 182... 819876908323
33.., 10143 83. l . 23338469 133,*. 7346629512 lE33... 896684817527
34*.. 12310 8L. 26543660 134... 8149040695 184. . 980462880430
l

35... 14883 85... 30167357 135. . . 9035836076 185...107182377-$337

36... 17977 86... 34262962 136., JO015581680 I86...117143d692373
37... 21637 $7... 38887673 137...11097645016 l87...12800110422Gi8
38... 26015 88.., 44108109 138...12292341831. 188...1398341745571
39.,. 31185 89. . 49995925
l 139. .a 13610949895 189. . . 1527273599625
40.. . 37338 90- 56634173 140...15065878135 190. A667727404093
4L.. 44583 91... 64112359 14LJ6670689208 191..,1820701100652
42... 53174 92... 72533807 142.. .18440293320 192. J987276856363
43*.. 63261 93.u 82010177 l43...20390913S757 193,,.2168627lOM69
44... 75175 94. .* 92669720 144.. .22340654445 194...2366022741845
45*.. 89134 95.,.104651419 145. l .24908858009 195. ..2580840212973
46,. .105558 96...118114304 146.m.27517052599 196...2814570987591
47...124754 97.. l 1332:30930 147.,.3038867X978 197 . . .3068829878530
48-147273 98u.150198136 14L33549419497 198,. l 3345365983698
49..*173535 99. l .169229375 149.. .37027355200 199...3646072438125
5O.a.204226 100.. .190369292 150. ..40853235313 200,. .3972999029388

* The numbers in this table were calculated by Major MqcMahon, by means of the
recurrence formulae obtained by equating coefficients in the identity

(1-x-x2+ti+x7-x~~-x16+ ..JZp(n) xn=l.

0

We have verified the table by direct calculation up to n= 158. Our calculation of p (ZOO)
from the asymptotic formula then iseemed to render further verification unnecessary,

338
 TABLE V* : g (~2).
n %a 12 % n CT& n %
1. 1 26... 165 FL. 4097 76... 53250
2::: 1 27... 192 52.. . 4582 77... 58499
3 .*. 2 SK. 222 53*.. 5120 78.m 64234
4... 2 29... 256 54- 5718 79,. . 70488
5... 3 30... 296 55.., 6:378 SO.., 77312
6... 4 31... 340 56,.. 7108 Sl... 84756
7 l 5
.* 32... 390 57:. . 7917 82... 92864
8 1.. 6 33. . . 448 58.v 8808 83.. JO1698
9 8 34... 512 5x.. 9792 84...111322
10::: 10 35... 585 60. l .10880 85...121$92
Il... 12 36... 668 61...12076 86-133184
12.,. 15 37... 7.60 62.. A3394 87,. .145578
13... 18 38... 864 63...14848 88.. .159046
14... 22 39... 982 64...16444 89...173682
15... 27 40...1113 65...18200 913.. .189586
IB... 32 41..,1260 66. ..20132 91.. .206848
17... 38 42...1426 67. .22250
l 92. ..225585
IS... 46 43.. .1610 68...24576 93.. .245920
lg... 54 44...1816 69.. .27130 94.. .267968
20... 64 4L.2048 70. .29927
l 95.e.291874
Zl... 76 46...2304 7L.32992 96.. -317788
22... 89 47n.2590 72.. .I36352 97. ..345856
23. l ,104 48. ..2910 73.. .40026 98...376256
24. .A22 49...3264 74.. .44046 99...409174
25. A42 50.. .3658 75.. .48446 100, . .444793

* We are indebted to Mr Darling for this t&h.

339
 UN THE REPRESENTATION OF A NUMBER AS THE SUM OF ANY
NUMBER OF SQUARES, AND IN PARTICULAR OF
FIVE OR SEVEN
BY G. H, HARDY
TRINTTY COLltEGE, ENGLAND
cAMB2IDGE,

Communicated by E. H. Moore, May 21, 1918

1. The formulae concerning the representation of a number as the sum of

5 or 7 squaresbelong to one of the most unfamiliar and dticult chapters in
the Theory of Numbers, and only one proof of them h&s been given. The
proof dependson the general arithmetic theory of quadratic forms, initiated
by Eisenstein and perfected by Smith and Minkowski. This theory, of
which a systematic account will be found in the fourth volume of Bachmann’s
ZahZe&eorie gives a complete Solution of the problem of any number s of
squaresnot exceeding 8. Beyond s = 8 it fails.
When s is even there is an alternative method, This method, which de-
pends on the theory of the elliptic modular functions, is much simpler in idea
than the method of Smith and Minkowski; and it has another very important
merit, that it can be used-within the limits of human capacity for calcula-
tion-for afly even value of s. Thus Jacobi solved the problem for 2, 4, 6 and
8. In thesecasesthe number of representationscan be expressedin terms of
the divisors of n. Suppose,e.g., that s = 8; and let us write, generally,

1-+ 2 f, (n)f =(1+2q+2q4+ . l l )‘={&(o,T)]s=9’,

where q = e’? Then

and r8 (a) is 16 2 a3if n is odd and 8 z Si - 8 z Sf if fl is even, 6 denoting

340 1918, 10 Proceedings of the National Am&my of Sciences, 4, 189-93.

 190 MATHEMATICS:. G. H. HARDY

b,a divisor of n, & a even, and 61an odd divisor. ,When s exceeds8 the formulae
are less simple, and involve arithmetical functions of a
n&~e.
e. Liouville gave formulae concerning the casess = 10 and s = 12,
and Glaisherl *hasworked out systematically all casesup to s = 18. More
recently important papers on the subject, to which I shall refer later, have
been published by Ramanujan2and Mordell? In the latter paper the whole
subject is exhibited as a corollary of the general theory of modular in-
variants.
The primary object of my own researcheshas been fo deducethe formulae
f or s =Sands= 7 from the theory ofelliptic kftinctions,and so to place the
casesin -which s is odd and even, so far as may be, on the samefooting. The
methods which I use have further important applications, but this is the one
which I wish to emphasizeat the moment. The formulae which I take as my
goal are the formulae

given by Bachmann (pp. 621, 655). H ere n as an odd number not divisible
by any square (so that there is no distinction between primitive and imprimi-
tive representations); m runs through all odd numbers prime to lz; B is 80,
160, 112, or 160, according as 12is congruent to 1, 3, 5 or 7 (mod. 8); and C
is 448, 560,448 or 592 in similar circumstances. These formulae are the cen-
tral formulae of the theory: they involve infinite series,but these seriesare
readily summed in finite terms by the methods of Dirichlet and Cauchy.
With them should be associated-theformula

where A is 24, 16, 24,.or 0: but this formula, as we shall see,stands in some
ways on a different footing.
2. My new proof of the formulae (1) and (2) was arrived at incidentally in the
course of researchesundertaken withadifferent end, that of finding asymptotic
formulae (valid for all values of s) for r,(n) and other arithmetical functions
which present themselves as coefficients in the expansionsof diptic modu-
lar functions. In a paper4shortly to appear in the Proceedingsof the London
Mathematical Society, Mr. Ramanujan and I have developed an exceedingly
powerful method for the solution of problems of this character, and applied
it to the study of p(n), the number of (unrestricted) partitions of n. This
method, when applied to our present problem, introduces the function

341,
 where

and the summation applies to K = 1, 2, 3, . and all positive-values

1

of h less than, of opposite parity to, and prime to k (h = 0 being associated

with k = 1 alone). The coefficient of q* in O,(q) is

(5)
and our method leads to the concltision aat

at any rate for every value of s exceeding 4.

When s is even, F(q) is an elementary function; and (&, R)~is easily exprai-
a
ble in a form which does not involve the ‘Legendre-Jacobi symbol’ ; .
0
The function X&z) is then susceptible of a variety of elementary transforma-
tions and it was shown by Ramanujan, in the second of his two papers quoted.
above, that X,(n) is identicul &h Y&Z) when s = A, 6 or 8. In what follows
I confine myself to the casein which s is odd, merely remarking that my method
(which is entirely unlike that usedby Ramanujan) leadsdirectly to an alterna-
tive proof of his results.
3. When s is odd, F(q) is not an elementary function. But it is not diffi-
cult to prove that

every term on the right hand side having an argument numerically lessthan
~ST. Further, si k: = Si$ &,h; and the first factor can always be expressedin
a simpleform. &pose, to fix our ideas, that s = 5. Then & = (- 1)“k2.
Substituting from this equation and from (7) into (4), and effect’ng someob-
vious simplications, we obtain

where now Ic assumesall values of opposite parity to and prime to k. This

formula may be simplified further by multiplying each sidebY

342
192 MATHEMATICS: G. H. HARDY

f+i+$+ . . l =?r”*

We then find

the summation now extending to k = 0, 1,2, . . . and all lz of oppositeparity

to k. This is our fundamental fbrmula, when s = 5. Two steps remain:
first, to prove the identity of 85(q) and #; secondly, to deduce the formulae of
Smith and Minkowski.
4. The first step presentsno very seriousdifficulty, for it involves nothing
beyond an adaptation of the ideas usedby Mordell in his paper quoted in 01.
We prove first that 86 behaves like 6” in respect to the linear modu-
lar transformations 7 = T + 2, 7 = -l/T; so that e&Y6 is an invariant of
the modular sub-group called by Klein-Fricke and Mordell Pa. Secondly, by
studying the transformation 7 = (a - 1)/T, we prove that 8&P is bounded
in the ‘fundamental polygon’ associatedwith Pa. It then follows that the quo-
tient is a constant which is easily seen to be unity. In all this the only
difficulty arisesfro& the useof certain reciprocity-formulae satisfiedby Gauss’s
sums.
We now transform (9) by effecting the summations with respect to h,
using certain contour integrals of a type common in the work of Lindelijf
and other writers. We thus obtain

(10)

a fundamental identity which contains the whole theory of the representation

of numbers by sumsof 5 squares. The symbolsj and p alone require expla-
nation; j runs through the complete set of least positive residues of 0, 12,
221 - * - ?(k - 1)2 to modulus k, each taken as often as it occurs; and &
is the multiple of k deducted in order to arrive at such a residue. And the
remainder of the work is purely arithmetic al. Picking out the coefficient
of qn, Oveobtain a seriesfor r&z) which is found, after somereduction, to be
equivalent to the seriesgiven by Bachmann.
4. The formulae which correspond to (10) for s = 7 and s = 3 are

343
 MATHEMATICS: G. Et. HARDY

=1+8z (-1))u-i)
g3 k z 2* 1,3,5,. l l j In=0
(mk+j)ff’+’ (12)

1
+ x ~2 2 (-l)m~~(mk+j)fq*‘~ l

2,4,6,... j m==O t

The interpretation of j and p is as before, except that, when k is even, j is a

residue of one of the numbers #, +K + 12, . . , *k + (k - 1)2. These l

identities embody the theory for 7 or 3 squares. It should be noted however,

that the application of my method becomesvery much more dficult when
s = 3, as the double seriesusedare then not absolutelyconvergent;and in fact
the only proof of (12) which I possessconsistsin an identification of the results
which it gives with those already known.
I conclude by a word concerning the casesin which s> 8. Here, when s is
oddt we are on un trodden ground. We have the asymptotic formula (6);
and we can evaluate X,(n) as when s = 5 or 7, thus obtaining a seriesof new
results. But it is no longer to be expected that our results should be exact,
and I have verified that, when s = 9, they are not exact, evw when PZ= 1.

1 Glaisher, J. W. L., Pm. Lodon Math, Sm., (Ser. Z), 5, 1907, (479490).
2 Ramanujan, S., Trans. Camb. Phil. Sot., 22, 1916, (159-184) ; Ibid., (in come of
publication).
3 Mordell, L. J., Quwt. J. Math., 48, 1917, (93-104).
4 Hardy, G. H., and Ramanujan, S., Proc. London Math. SOL, (Set. Z), 17, 1918, (in
course of publi$ation).

CORRECTION
p. 192. Formula (9) is incorrect; see the footnote at the end of 0 3.22 of the next
paper (1920, 10).

344
ON THE REPRESENTATION
OF A NUMBERAS THE SUM OF ANY
NUMBEROFSQUARES,
AND IN PARTICULAROF FIVE*
BY

Cr. H. HARDY

1. INTRODUCTION
1. 1. In a short note published recently in the P r o c e e d i n g s o f t h e
National Academy of Sciences L I sketched the outlines of a
new solution of one of the most interesting and difficult problems. in the
Theory of Numbers, that of determining the number of representations of a
given integer as the sum of five or seven squares. The method which I use
is one of great power and generality, and has been applied by Mr. J. E. Little-
wood, Mr. S. Ramanujan, and myself to the solution of a number of different
problems; and it is probable that, in our previous writings on the subject,*
we have explained sufficiently the general ideas on which it rests. I may
therefore confine myself, for the most part, to filling in the details of my
previous work. I should observe, however, that the method by which I
now sum the “ singular series “, which plays a dominant role in the analysis,
* Presented to the Society, February, 1920.
1 G. H. Hardy, On the expe8sion of a number as the wrn of any number of 8quwe8, and in
particular of fwe or wen, froceedinga of the National Academy of
S c i e n c e s , vol, 4 (1918), pp. 189-193.
f G. H. Hardy and S. Ramanujan: (1) Uneformuk asymptotiqw pour le nom&e des partitions
den, Comptes Rendus, 2 &IL 1917; (2) Asymptotic formulae in Combinator~ Analysis,
Proceedings of the London Mathematical Society, ser. 2, vol. 17
(1918), pp* 75-115; (3) On the we@ients in the expansions of certain modular junctions, P r 0 1
ceedings of the Royal Society, (A),vol.95 (1918),pp. 144-155:
S. Ramanujan, On certain ttigiwwmetrical sums and their appltiions in the theory of num-
bera, Transactions of the Cambridge Philosophical Society,
vol. 22 (1918), pp. 259-276:
G. H, Hardy and J. E. Littlewood: (1) A ww solution of Waring’s Problem, Q u a r t e r 1 y
Journal of Mathematics, vol. 48 (1919), pp, 272-293: (2) Note on Messrs. Shah
and Wilson’s paper entitled On an empirical formula connected with Goldbach’s Theorem, P r o -
ceedings of the Cambridge Philosophical Society, vol. 19 (1919),
pp. 245-254: (3) Some problems of ’ Partitio Numerorum’, (I) A new solution of Waring’s Prob-
lem, Giittinger Nachrichten, 1920: (4) Some problems of ’ Partitio Numerorum’, (II)
Proof that every Earge number is the sum of at most 91 biquadrates, M a t h e m at i s c he 2 e it -
schrift, 1920.
The ,two lastpapers will be published shortIy.

1920, 10 Transactioras of the Am&can .M&ema;t&l So&y, 21,

266-84. 345
 256 G. H, HARDY WY

is quite different from that which I sketched in my former note. The new
method has important applications to a whole series of problems in Com-
binatory Analysis, concerning the representation of numbers by sums of
squares, cubes, kth powers, or primes. It is in the present problem that it
finds its simplest and most elegant application, and it is most instructive to
work this application out in detail,
-<“‘It is well ,known that the solution of the problem is a good deal simpler
when s, the number Of squares in question, does not exceed 8 s If s is 2, 4 1
I
\ 6, or 8, the number of representations may be expressed in finite form by
-- a,?
Tmeans of the real divisors of n; if s is 3,5, or 7, by means of quadratic residues
and non-residues. If a > 8, other and more recondite arithmetical functions
1.
I are involved. In this paper I confine myself to the cases in which s ZZ 8.
\ Among these, those in which a is odd have always been regarded as notably
-“the m&e difficult, and one of my principal objects has been to place them
all upon the same footing. But I generally suppose s = 5 or s = 8, cases
typical of the odd and even casesrespectively.
In Section 2 I construct the singuhr series

where
A1 I ktk =
1,

&, k denoting the Gaussian sum3

k
ef2iihdlk
x 9
j 11
and the summation extending over all positive values of h less than and prime
to k. The seriesmay be written in the form
21”tl VlL
+m> +-@cos (#nn - im)

the zero terms corresponding to k = 2 and k = 6.

2
~8(COS~nk+CoS(~np-SK)} +o+ l **

1 J

In Section 3 I show that, when s = 8 or s = 5, the sum of the singular

seriesis in fact re (n) , the number of representations of n as a sum of s squares.
The meth& used are equally applicable in the casesof 3, 4, 6, or 7 squares;
8 In my former note I denoted a. typical ” rational point ” on the unit circle by ehsilk,
and a typical Gaussian sum by

In this paper I generally use the forms involving a 2. Each notation has special advantages
for particular purposes.

346
 1920] REPRESENTATION’ OF A NUMBER AS THE SUM OF SQUARES 257
t

but the case of two squares is abnormal4 Throughout this section I am very
deeply indebted to a paper by Mr. Mordell, published recently in the Q u a r -
terly Journal of Mathematics? My proof of the identity of the
functions which I call 8% and O8 is in fact based directly on his work. It is
true that Mordell considers only the case in which $ is even; but his argu-
ment is applicable in principle to either case, and was applied by him to the
even case only merely because, at the time when his paper was written, he
had no method for the construction, when s is odd, of the essential ” principal
invariant ” denoted by him by x. It is the construction of this invariant by
a uniform method in all cases, through the medium of the “singular series “,
that is my own principal contribution to the subject.
----Tin Section 4 I show how the singular series may be transformed into a
product, and give general rules for the calculation of the terms of the product.
All the results of this section are independent of the hypothesis s 5 8. In
Section 5 I sum the series when 8 = 8, and obtain Jacobi’s well-known results.
In Section 6 I consider the case 8 = 5, supposing however that n has no squared
factor, so that there is no distinction between primitive and imprimitive
representations; and I obtain results equivalent to those enunciated first by
Eisenstein and proved later by Smith and Minkowski. In Section 7 I con-
sider the general case, and show that the method leads to the more complete
results of Smith. I conclude, in Section 8, by some remarks as to the appli-
cation of the method when a > 8. I do not pursue this subject further because
such applications belong more naturally, either to Mr. Littlewood’s and my
own researches in connection with Waring’s problem, or to Mr. Mordell’s in
connection with the general theory of modular invariants.
It will be noticed that the explicit formulas for the powers of the funda-
mental theta-function, such as the familiar formula

a8 = ( 1 + 2q + 2q4 + l l l )8 = 1 + 16 gq + ++2? + $ + l . .> ,

or the new formulas

do not appear at all in my present analysis.

4 See Mr. Ramanujan’s paper quoted in footnote 2.
6 L. J. Mordell, On the representations of numbers as a sum. of 2r squares, Q u a r t e r 1y
Journal of Mathematics, vol. 48 (1917), pp. 93-104. See also a later paper by
the same author, On the representations of a number as a sum of an odd number of squares,
Transactions of the Cambridge Philosophical Society, vol. 22
(1919), no. 17, pp. 361-372.
6 This is formula (10) of my former note, where the meaning of j and p is explained. See
also p, 360 of Mr. Mordell’s second paper cited above.

347
 258 G. H. HARDY [July .

In the sequel I give references only to isolated results directly required for
the objects of my analysis. It is more convenient to collect here some notes
concerning the older memoirs dealing with the problem.
Jacobi’s classical results concerning 2, 4, 6, or 8 squares are quoted by
Smith on p. 307 of his Report on the Theory of Number8 (Cdeded Papers,
vol. 1). They are contained implicitly in $5 40-42 of the Fundamenta Nova
(pp. 103-115).
Liouville gave formulas relating to the cases of 10 and 12 squares in a
number of short notes in the second series of the J o u r n a 1 d e s m at h 6 -
m a t i q U e s : see in particular vol. 5, p. 143; vol. 6, p. 233; vol. 9, p. 296;
vol. 10, p. 1, These notes appeared between 1860 and 1865.
Later Glaisher, inaseriesof paperspublishedin the Quarterly Journal,
worked out systematically all cases in which s is even and between 2 and 18
inclusive. He has given a short summary of his results in a paper On the
numbers of representations of a number as a sum of 2r squares, where 2r does not
exceed 18, published in the Proceedings of the London Math+
matical Society, ser. 2, vol. 5 (1907), pp. 479-490. This paper con-
tains full references to his more detailed work.
The results for 5 squares (for numbers which have no square divisors)
were stated without proof by Eisenstein on p. 368 of vol. 35 (1847) of
Crelle’s Journal. They were completed by Smith, who stated the
general results at the end of his memoir On the orders and genera of quadratic
forms containing more than three indeterminates (Pro ce e din g s o f t he
R o y a 1 S o ci e t y $ vol. 13 (1864), pp. 199-203, and vol. 16 (1867), pp. 197-
208; ColLected Papers, vol. 1, pp. 412-417, 510-523). No detailed proofs,
however, appeared before the publication of the prize memoirs of Smith
(Me’moire SW lu repr6~entation des nombres PUT des Sommes de cinq carr&,
Memoires present& par divers savants B I’AcadGmie,
vol. 29, no. 1 (1887), ppm l-72; Collecded Papers, vol. 2, pp. 623-680) and
Minkowski (Me’moire GUT la the’orie des forme quadratiques 2Gcoefiients en-
t&es, ibid., no. 2, pp. l-178; Gesammelte mathematische Abhandlungen,
vol. 1, pp. 3-144).
The methods for the summation of the series
f ‘3
i
Y

which is fundamental in the five square problem, and other series of similar
type, are due to l&i&& (Recherches SW divers upplicutions de l’analyse
infinit&male h la the’oke des nombres, Crelle’s Journal, vol. 19 (1839),
pp. 324-369, and vol. 21 (1840), pp. 1-12, 134-155; We&e, vol. 1, pp. 411-497)
and to Cauchy (Me’moire SW la thkrie des nombres, M 6 m o i r e s d e

348
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARE$ 259

l’Acad6mie des Sqiences, vol. 17 (1840), pp. 249-768; especially

Note 12, pp. 665-699).
A systematic account of the whole theory is given by Bachmann in vol. 4
of his ZuhZentheorie. Bachmann works out the case 8 = 7 also in detail.

2. FORMAL CONSTRUCTION OF THE SINGULAR SERIES

.^.- . _- ._._
^__-_-.,_ __r_...-.A. *m--J>
2. 1. I write, as in my former note

(i.ll)f(q) = l+&(n)4”= (1+2q+2q4+ l *)s

1
= (&(O, 7>)6 = P,
where q = ewdrand 3 ( 7) > 0; and I consider the behavior of this function
when q tends radially to a “ rational point ” ezAr’lk upon the unit circle. We
or that k is greater than unity and h positive,
less than k, &d prime to k l

If (2.11)
q I 4eanilk,

so that 0 5 q < 1, q +l, we have

Now

when q + 1 l It follows that

S h. k
(2.12) s- -\r7r k
where
k
(2.121) 2pirrle 1
sh, k = c e
j=l
and
(2.13)

it being understood that, when Sg, k = 0 (as is the case if, and only if, k is of
the form 4m + 2) , this equation is to be understood as meaning

(2.131)

349
 260 G, H. HARDY CJUlY

2.2. The principle of the method is to write down a power-series

(2.21) fh, k(q) = cch, k, nq”,

which, ( a) is as simple and natural as possible and ( b ) behaves as much like

f ( q ) as possible when q + 8Ar’lk; and to endeavor to approximate to the
coefficients inf (q) by means of the sums
(2.22)
It is plain that, in forming these sums, we may ignore values of k of the
form 4m+2.
The appropriate auxiliary function (2.21) is
aa
(2.23) hek(q) = ~(3’s)
where
(2.231)
1
It is in fact well known that
-+a
F8(x)

is regular at x = 1 *a We are thus led to take

(2.24)
and

(2.251) A1I 1f Ak = k-8 c (&, k)l ~-2**~~~,

A
the summation extending over all positive values of h less than and prime to k.
I call the series
(2.26)
the singular series. The process by which it has been constructed is of a purely
formal character. It remains (1) to investigate more rigorously its bearing
on the solution of our problem and (2) to find its sum.

3. PROOFTHAT THE SUM OF THE SINGULAR SERIES, WHEN 8 = 8 OR 8 = 5, IS

THE NUMBER OF REPRESENTATIONS OF n
3.1. Proofthat ps(n) = T&).
3.11. When s is 3, 4, 5, 6, 7, or 8 (but not 2 or any number greater
7 See, for example, E. Lindeliif, Le calcul des re’sidus, p, 139.

350
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 261

than 8) the sum of the singular series gives exactly the number of representa-
tions of n. In this section I prove this when s = I8 and when s = 5. These
cases are perfectly typical, but formally a little simpler than the others.
Suppose first that s = 8. Then
(3.111) @(g) = 1 + Efh, k(Q) = 1 +~~(~)*Fs(Pe-z~~~~‘).
Now
s i,k = qk k4,

where qk: is 1, 0, or 16 according as k’is odd, oddly even, or evenly even.

Also, if x = e-g, we have
F*(x) = c n3xn = x nWny = $(; cosech2 $y) = 6 x (y + iniTi)(,

where rt runs through all integral values. Hence

F8 ( qe--Bhriik ) 1
6k
f c
1
It {2(nk -p h) - kri4’
and
(3.112)

the summation extending over the values of h, k, and n already specified.

If k > 1 J nk + h assumes all values prime to k; if k = I, all values. Thus

(3.113) 08(q) = 1 + C(2h Iilck,)“;

wherenow k = 1, 2, 3, and A assumes all values such that ( h, k ), the

l

highest common factor of h and k, is unity. But this equation may be written
16
(3a114)
e*(q)= 1+ k=l, Fhx,JJh ’ kT)4+ c 3, l ..; f = - k=4, 8, . . . . (h, k)=l (2h - kr)4

= 1+ 1 1
+ x
&=I, 3,.x... (ti, kl=l(2h - krj4 ,=a, 4,.... (h, k)=l (h - kr)4
1
=1++-k7)1s
where now k = 1, 2, 3, -0 and h assumes alI values of oppoa& parity to
and prime to k.
Multiplying both sides of (3.114) by
7r4 1
-=l+;+~+*-,
96
we obtain
(3.1151)

351
 262 G. H. HARDY [July

or
(3.1152)

In (3.1151) k = 1, 2, 3, 0.0, and in (3.1152) k = 0, 1, 2, l .0; in each

equation h assumes aEEvalues of opposite parity to k.
3.12. We now write
(3.121)

and consider the effect on x ( 7) of the modular substitutions

(3.1221) 7’ = 7 =t 2, (3.1222) 7’= -l/r.

It is obvious in the first place, from (3.1151) or (3.1152), that

(3.123) X(7 *2) = x(4*
Again, we may write (3.1151) in the form

where -h and k assume all values of opposite parity except that (as is indicated
by the dash) the value k = 0 is omitted. Thus

Changing h and k into - k and h, we obt.ain

(3.124) x(- l/T) = 74x(T).
Now { 83 ( 0, 7) I8 = f18 = $ ( 7) satisfies the equations

and so it follows, from (3.123) and (3.124), that the function

is invariant for the substitutions (3.122), and therefore for the modular sub-
group which they generate, the group called by Klein-Fricke and Mordell r8.
3.13. The next step in the proof is to show that 1;1(T) is bounded through-
out the Cdfundamental polygon ” Ga associated with the group &. This
region is defined by
7 = x + iy, Ir) z 1, -1Ss~1,
and has only the points T = & 1 in common with the real axis. It is there-
fore sufficient to show that q ( 7) is bounded when 7 approaches one or other

352
1920] REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 263

of these points, say T = 1 l For this purpose, following Mordell, I consider

the effect of the substitution
1
7 = l-7
T
If we write T = X + iY, and suppose that 7 + 1 from inside Gas then
Y + ar and 1Q 1 = lew”] is small. An’d

(3.131)

=f+;T4C’F(k:hT)L k

=; T4Cc (k+‘hT)4’
k h

where now k assumes all integral values and h all odd values.
Write hT = a, tP = rj and sum with respect to k We have l

1 2
- - # cosec2 a?r + 7r4 cosec4 a~;
&+a)‘= 3

and this function, when expanded in powers of r1 begins with the term
#T’ c2 = #7r” Q2”.
Hence, when T = X + iY and Y is large, we have

(3.132) e,(q) = 256T4Q2 + l *.a

But we have also

(3.133) {~3(0,~-~))I=T4{o~~0,T))8-256T4Q2+~~~~
and so
(3.134)

Thus 7 ( 7) is an invariant of & and bounded throughout Ga; and is therefore

necessarily a constant, which is plainly unity.
It follows that
(3.135) a8 = @8((j)
and-so that
(3.136) 6’8(n> = r*(n).

353
 264 G. H. HARDY CJUlY

3.2. Proof that p5 (n) = r5 (n) .

3.21. When s = 5 the proof proceeds on the same lines, but is not quite
so simple. We shall require certain well-known identities which I state as
lemmas.
LEMMA 3.21L8 If h and k are positive integers of opposite parity, then
-
k
e-.Pkxb]h.
(3.2111) J
j=l
and if h and k uw positive integers, and h odd, then

5 ( _ 1 ) j ej’hrdlk =
(3.2112)
j=l
Here 16 = eiTi.
LEMMA 3.213. Suppose that 0 < Y < 1, and that CTand the red part of t are
positive. Then
.
(3.2121) L$!f 2 (m + y )u-l e-2rrt(m+vl= $ v&,
0

where
(t + nip = exp(alog (t + fri>l =exp(alog~t+ni~ +a&
and - +?r<+<+. The formula dill bids for v = 0, if g > 1.
This result is due to Lipschitzg We shall require two special cases.
(i) Suppose that c = 3~ > 1, Y = 0, t = - &, x = P; so that
g(7) >Oand 1x1 < 1. Thenweobtain

(3.21221) I(2 n - ~>;I38 = exp{+dog 1(2n - r)il + *+i}

(ii) Suppose that Y = 3 ( I + 0), 6 = X/K, where K and X are integers

and -K<X<K,andt= -KA. Then

(3.2123)

where P(x) is an ascending power series in x.

3.22. Supposing now that a = 5, we have

(3.221)
‘8 See, for example, G. Landsberg, Zur The&e der Gaussschen Sum ,men und der linearen
Transformation der Thetafunktioxen, J o u r n a 1 f ii r M a t h e m a t i k,, vol. 111 (1893),
pp. 239-253. Both formula3 ar? included in formula (17b), p. 243, of Landsberg’s memoir.
The first is &o proved by Lindeliif, lot. cit., pp. 73-75.
9 R. Lipschitz, Untersuchung der Eigenschaften einer Gattung vun unendlichen Reihen,
Journal fiir Mathematik, vol, 105 (1889),pp. 127456.

354
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 265

and
S;,k = qk k2,

where now qk is 1, 0, or - 4 according as k is odd, oddly even, or evenly

even, Thus
(3.222) 05(q) = l-t- F2 c $ sh, k F5 ( qCzhnilk > .

Substituting from (3.2122), we obtain

S
(3.224) e5w = 1 -t C$mMh&q5,2J

the ranges of summation in these equations being the same as in (3.112) and
(3.113) respectively. The last equation can be expressed in a more con-
venient form by introducing the sum
k-l
c”
(3.225) T h,k = ce3 ‘zhlrilk .
0
In fact (3 224) may be written
s
Wq)=l+ c
k=l, 3, . . . . (h, k)=l zl {(ah -hk:)i]5~2
4 s
-
k=4,
c
3, . . . . (It, k)=f Ti ((2h -hkkr)i)512’
In the first sum k is odd; 2h = H is even and prime to k; and sh, k = TH k l

In the second k = 2K, where K kuns through all even values; h is odd aLd
prime to K; and sh, k = ZT, Ke Effecting these substitutions, and then
replacing W (or K) by h (or k) ; we obtain

( - l)h Th, k
e,(q) = 1-t c-
& {(h - k+j5J2
where now k = 1, 2, 3, l a and h assumesall values of opposite parity and
prime to k ?
Multiplying by
T2 1 1
-=p+;+p+ ‘..P
8
and observing that, if X is odd,

(- l)xh = (- l)h, 6((Xh - Xk+ij5f2 = X3#(h - k+j5[2,

T Ah,hk = XT h, kp
10This is equation (8) of my former paper,

355
 266 G. H. HARDY WY

we obtain
9t2 7r2 (- llh Th, k
(3.226) g%(q) = s+c-- & {(h - kr)i]512’
where h now assumes all values of opposite parity to k .I1
3.23. The discussion now follows the lines of 3.12 and 3.13. We write
1r2
(3.231) p”(q) = XbL
and it is obvious at once that
(3.232) xb *a = x(4*
The discussion of the transformation # = - l/7 requires a little more care,
owing to the presence of many-valued functions under the sign of summation,
It is convenient to begin by including negative values of k.
We write generally
x8 = exp{s lug 1x1 + iama)
where the particular value of amx to be selected has to be fixed’ by special
convention. Thus in ((h - Jc7)ij512, where k > 0, amf(h - k7)i) lies
(as has already been explained) between - +?T and in We now agree l

that, if k is still positive, am ( - k) = ?r, so that -J- k = i AIE, while

am { ( - h + kr ) i) lies between - 8~ d and - +?T. It will easily be verified
that
(3.233) m{( - h + k+)ij512 = &((h - kr)ij512.
Further, we write by definition
(31234) T-n, -k = &i, k l

We know from (3.2111) that, when h and k are both positive,

and it is easy to verify that, with our conventions, we have generally

(3.235)

where E = 1 unless h > 0, k < 0, in which case E = - 1.

3*24. We have, from (3.226), (3.233), and (3.234),

T2 T2 (- 1)”
=3+x( - +--5t2 +;G- ((h T;:)il”l’

l1 This equation takes tile place of (9) of my former papa, which is not printed correctly.
The fist term on the right * is omitted, and k = 0 is included wrongly under the sign of sum-
mation.

356
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 267

where now h and k are any integers other than zero and of opposite parity.
Writing - l/7 for 7 in (3.241), using (3.235), and then replacing h and k by
K and - H, we obtain

(3.242) x: (-5) =~+f(~)-611+4~CI(o~=( {(K-*&.ll,2,

where E is 1 unless Ix and K are both positive, and then - 1, and

am{(K - H/?)i]
lies between - $Tand$rifH<O,between-#nand-$rifH>O.
It may be verified without difficulty that

(3.243)

where 0 < am ( - l/r) < ?r and the value of am{ (H - Kr)i) is fixed in
accordance with our previous conventions. Consider, for example, the case
H > 0, K < 0. In this case

- gr < a = am((K - H/~)ij < - +?T,

O<P = am( - l/7) < 7r,
and
- %?r < y = am{(H - KT)~} <&r.
Thus p + y lies between - #n and $, and, as a! differs from p’+ y by a
multiple of %, we must have a = 0 + y - 27r and

s'2((H - Kr)ij5j2,

in agreement with (3.243). The other possible cases may be treated similarly.
Thus

(3.244) x(-i) t ;+;(i)-s’2

where - 3 ?r < am(+) < &n, 0 < am( - l/r) < ?r* And this equation

357
 268 G. I-I. HARDY WY

This is the same functional equation as is satisfied by a5 l Hence

is an invariant for each of the substitutions (3.122), and so for r3.

3.25. It remains to verify that q ( T) is bounded in G3. As in 3.13, it is
only necessary to consider the neighborhood of T = 1. Putting 7 = I - I/T,
as in 3.13, in (3.241), and then writing h = H - K, k = H, we obtain

(3,251)
-@ { ( -K+H/+i}5/2’
where H assumesall values save 0 and K ill odd values, and
am{(- K+H/+j
lies between - 3 r and 3 r, or between - Q at and - 3 z=, according as
H>OorH<O.
Now
II-1
(3.252) T H-H, H = c ( - 1 )j &-i”Kwr’E = UK, H,
0
say. It follows from (3.2112) that, if h and k are both positive, and h is odd,
and
k

(3.253) Uh, k = y ( - 1 )i e-Ph*W, wh, k = c

8 C3-&)zhNk
1
0 1
then
k
h- 6, h;

and it is easily verified that, if we adopt the same conventions as in 3.23

concerning the meanings of 6 h 1 6 k 9 &, -k, and w-h, -k , we have
12We first restore the terms for which H = 0, and then observe that
(-l)*=-(-l)g.
we have to verify t-hat, with our con?entions,
-&(-I/+-#(-A)--8=1
and
- &( - l/r)-3 = - (i/+-k
these verifications present no difficulty.

358
1920f REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 269

generally

(3.254)
where E = lunlessh<O,k>O,inwhichcasee= -1.
Using this equation in (3X11), we obtain

It is now easy to verify, by arguments similar to those of 3.24, that

{( - K + H/T)i15j2 = eTw5j2{(H - KT)ij512,

where am T lies between 0 and r, while am ( ( H - KT ) i} obeys our previous
conventions. We thus obtain

the summation being now limited only by the fact that K is odd. In virtue
of (3.233), this equation may be written

3.26. The series in (3.256) may be expressed in the form

c (- 1)H~-(i-twwK
K=l, 3, 5,.., {(H B KT)i)5’2m*
Suppose that
(J . -122 1 - hj (mod 2K) (-K<X;<K),
and 6i = Aj/K . Then
H e-(j-~)YhilK E e-H0jTt
c’- 1) (-- 1)

E ((H - KT)ij5j2 = 5 {(H - KT)ii512

may, by (3X23), be expanded as a power series in Q = P, in which the
lowest power of Q is
QK+hj .
The smallest possible values of K + Xj are i ,8, l l l ; and K + Xi = $ involves

3
‘“-j+aE
-K+t(mod2K)
or
‘2
2 -j = (2p - l)K,

359
 270 Cr. H. HARDY WY

where p is an integer, i.e. an equation whose left-hand side is even and whose
right-hand side is odd. Thus K + Xj 2 $; and the left-hand side of (3,256)
is the product of T5j2 Q5j4 by an ascending power-series in Q. But

9a(0,1-+) =$a2co, T)
is the product of P2 Q114by a power-series in Q. It now follows, just as in
3.13, that v ( 7) is bounded, and so is a constant, which is plainly unity.
We have thus established the identity of ed and 68, and so of p8 (n) and
r,(n), when s = 8 and s = 5. The same method may be used for any
value of 8 from 5 to 8 inclusive.13 In order to complete the solution of our
problem, we have to sum the singular series (2.26).

4. GENERAL RULES FOR THE SUMMATXON OF THE SINGULAR SERIES

4.1. The due of &, k.

The known results concerning the value of the Gaussian sum &, k are as
follows.14 We assume that ( h, Jc) = 1.
If k and k’ are prime to one another
(4.11) s h, kk’ = s hk’, k Shk, k’-

We need therefore consider only the cases in which k = 2’ or K = p*, p being

an odd prime.
Ifk=2
(4.121) Sh,k = 0.
If k = 2’ = 22P+1, and p > 0,

(4.122) SA, k =
2~4-1$hri .

If k = 2A = 22p, and p > 0,

(4.123) s h, k = 2~ ( 1 + $) = 2’*+1~0s ihn eihai.
If k = p,

(4.131)

where (h/p) is the well known symbol of Legendre and Jacobi.

--
18When s = 2 or s > 8, the conclusion is false. The cases s = 3 and s = 4 are exceptional,
The conclusion is true, but new difficulties arise in the proof because the series used are not
all absolutely convergent. These difficulties are easily surmounted when s = 4, but are
more serious when s = 3 l

14For proofs of these assertions see the chapter on Gau& sums in the second volume of
Bachmann’s ZahZenUzeorvie, A less complete account is given in Dirichlet-Dedekind, VorZesungen
tiber Zahbntheorie, ed. 4, 1894, pp. 287 et seq,

360
1920-J REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 271

If k = pA = p2fi+l, and p > 0,

(4,132) sh, k = p’Sh,p.

Finally, if k = pA = p2gr and p > 0,

(4.133) s h, k = pp.

These formulas enable us to write down the value of Sa, k for all co-prime
pairs of values of h and k.

The multiplication rule for Ak.

4.2. The first step is to prove that
(4.21) A kk’ = Ak Ak’

whenever k and k’ are prime to one another.

In the formula which defines Ak, viz.
k8 Ak = c (s,, k)’ e-2nhriik,
h assumes all positive values less than and prime to II: us call this set l Let

of values, or any set congruent to this set to modulus k, a k-set. It is easy to

see that if h runs through a k-set, and h’ through a k’-set, then
h = hk’ + h’ k
runs through a kk’-set. For the number of values of h is
cmw = mw,
and it is obvious that all are prime to kk’ and incongruent to modulus kk’.
Thus
(kk’)” Akk’ = c ( sh, kk’)’ e-2nhrilkk’
h

= c ( shkf,k)8 ( Shk, k’)8 e-2nhrilkk’,

by (4.11). But
k k
e23’hk’rijk = C ,ZGk’)‘hri/k = sh k ,
s hk’, k = Il ,
j=l j=l

since jk’ runs through a complete system of residues to modulus k; and

Siddy &, k’ = &‘, k’. Thus
( kk’)* Akkf = c (&, k >” (Sh’, k’)” e-2nhtiikk’
h,h’

= (kk’)” & Ag;

which proves (4.21).

361
 272 43. H. HARDY WY

It follows15that
(4.22) s = Al + A2 + A3 + l l l = 1 + A2 + A3 + l l l = nxp

where
(4.23)
and p runs through all prime values.

Calculution of Azx.
4.3. Suppose first that p = 2. Then the value of Aflh is given by the
following system of rules.
4.31. A2 =
0
l

4.32. If X ia odd and greater than 1,

(4.321) & = 0,
?&?I 1e8S

(4.3221) n = 0 (mod F-3)

and
(4.3222) v -8s 2-cx-3) n - 8 = 0 (mod4),
in which case
(4.3223) Ap = 2--($u-l>
(X-1)@(P-8)‘Ki.

?fx = 3, (4.3221) is 8atisJiedautomatically.

Let X = 2~+1(~ > 0). Then
s A, + =
p+l ethri= p-1 Sh, 89
by (4.122). We write
h=&+h’ (~=0,1,+~~,2~-~-1; h’=1,3,5,7).
Then

= p+f-2) c ( sh,
,
3 )a e-2nhfxi/22p+1 c e-2n=i122~2,
h’ x

which vanishes (in virtue of the summation with respect to x ) unless rz = 0

(mod 2211--2)
.
rf ?t = 22pB2
Y, we have
A2x = p-w (P+ac (Sh’, 8>” e+h’ri.
h’
The sum with respect to h’ is

16We assume that the series and product are absolutely convergent. This is obviously the
caseifs>4,asS~,~=O(~),A~=O(k~-+),andl-as<-1.

362
19201 REPRESENTATIBN OF A NUMBER AS THE SUM OF SQUARES 273

which is 0 or 22sf2e-3 (~-8)ni

according as (4.3222) is not or is satisfied. This
completes the proof of 4.32.
4.33. Ij X is ezlenand greater than 1,
(4.331) A+ = 0
unless
(4.3321) nc 0 (mod Zxm2),
in which case
(4.3322) A,x = ~m-~>O-U cos '(3 yr - a a*) 9
where n = Zhs2Y. If x = 2, the last formula holds in any case,
If X = Zp, we have
s h, 2X
= 2h1 cos $hr eihai = 2p-l Sh,4j

by (4.123). We write
h = 4x + h’ (x = 0, 1, 0.0, F2 - 1; h’ = 1, 3).
Then
A9 = 2-8 (p+I) C ( Sh, p )a e+nhriPp
h

which vanishes (in virtue of the summation with respect to z ) unless n = 0

(mod 22pB2) . But if n = 2*pB2Y, we have
Aqh = 22w-2-8 (p+l) c ( Sh’, 4 )” e-4vh’ri,
la’
and the sum here is

This completes the proof of 4.33. If X = 2, z disappears from the argument

and h and h’ are identical.

Cakxdation of A,x when Q is odd.

4.4. The corresponding results when p is odd are as follows.
4,41. If ~2=#=0 (mod p) then
(4.411) AP =-

(4.412) AP = (s =
- I>,
0 f p-4(8-1)

(4.413) A p= - ( - lp--l) p-4” (8 = q,

(4.414) A p = (- ; p-e-l> 1)--l) (s _ 3);

0
the congruencesfor s referring to modulus 4. But if n = 0 ( mod p ) th,en
(4.415) A p = (p - l)p-4” (s = O),

363
 274 G. H. HARDY LJUlY
(4.416) A = 0 ( 8 = 1,3),
(4.417) A p= (-l;~‘p”(p-l)gi. (ss2).
We have

If s is even, this is

and the sum is equal to - 1 or to p - 1, according as n is not or is a multiple

of p. This leads at once to the results stated for even values of s.
If on the other hand s is odd, we have

which is equal to 0 if TZis a multiple of p, and to

3 b-1) (P--l)’

otherwise. We thus obtain the results stated for odd values of S.

4.42. If X is odd and greater than 1, then
(4.421) A,x = 0
ifn + 0 (mod $-I);
(4.422) A+ = p- N-1) 0-l) d4, ( y )
if n = p”1 vandu+O(modp); and
(4.4231) Aph = (p - l)pA-1-~8x (a = o),
(4.4232) A+ = 0 (s = L3),
(4.4233) A,x = (- l)i(ype l)p+l+x (8 = e>,
ifn = 0 (modp”).
If X = 2~ + 1, p > 0, we have sh, g = pw ah, p J by (4.132). we write

h=px+h’ (~=O,l,*~*,p~-~+; h’=1,2, l ~~,p-1).

Then
ApA = p(c+l) c ( Sh, p)8 e-2nh*olpw+1 l

If TZ+ 0 (mod p2” ) 9the sum with respect to x vanishes, and we obtain (4.421).

364
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 275

If n = p2rcY, we obtain

A+ = p2’L-8(IL+1)
c (&,‘, p)8 e-2vh’r<jp,
tbt
If Y + 0 (mod p), this is p- (p2)F ~4, ( y> J and we obtain (4.422). But if
VS 0 (mod p) we have

and we obtain the equations (4.423).

4.43. If X is even and greater than 1, then

(4.431) Apx = 0
ijn + 0 (mod PA+);
(4.432) Apx = - p++,“~

if n = PA--l y anav + 0 (mod p); ana

(4.433) A,x = (p m 1 )px-l-n”x
;ifn = 0 (modp”).
If X = 2~~ we have Sh, Px = p” , by (4.133). Hence

which is zero if n + 0 (mod pzFB1), and

if n = p2pB1 y; and the sum here is equal to - 1 or p - 1 according as v is

not or is divisible by p.

5. SUMMATION OF THE SINGULAR SERIES WHEN a = 8

5. 1. The formulas of Section 4 enable us to sum the singular series what-
ever the value of s. I take as typical the cases s = 8 and a = 5. I suppose
first that s = 8 and that 12 has no squared factor. We have to determine
the factors xP of (4.22).
In the first place, let p = 2. Then, as 72 is not divisible by 4, we have
Ale= A 32= •~~ = 0, by (4.321) and (4.331); and also Aa = 0, by (4.321),
since Y = 1 and Y - s is not a multiple of 4. If 12is odd, Ad = 0, by (4.331);
but if n is even, 4
A 4 = 2-3 cos 4VT = - *,

by (4.3322). Finally As = 0 in any case, by 4.31. Thus

(5.11) x2 = 1 (n&d), x2 = $(neven).

365
 276 CL He HARDY [July

Next, suppose p odd and ~+n.‘~ Then Ap2 = Ap3 = = 0, by (4.49)

and (4.431), and A, = - p-*, by (4.411). Thus
-4
(5.12) xp= 1 -P (PW*
Finally suppose p odd and ~1 rz. Then A,, = Apr = l
. . = 0, by (4.421)
and (4.431); Apl = ps7, by (4.432); and
A p = CP - 1)P-4,
by (4.415). Thus
(5.13) xp = 1 + (p - l)p-4 - p-7 = (1 + p-“) (1. - jP).
We have therefore
s = x2 n (1 - p-") I1 I(1 +p-3)(1 - p-“)}
P-b Pb

= x2 n (1 - p-“) I-I (1 + p-9 = $2 n (1 + p-“),

Pb Pb
since
H(1 -p-4) =&=$*
If 12is odd,
7r4n3 96
(5.14) m(n) =---4n: 1 + p-3) = l6~3(~~),
r(4) -T pin
where 03 (n) is the sum of the cubes of the divisors of ~2. If ~2is even,
(5.15) p&z) = 16n3(1 - 2-3) I-I (1 + P-3) = 16{a;(n) - cif(n)},
Pb

where a; (n) and a:’ (n) are the sums of the cubes of the even and odd divisors
respectively. These are Jacobi’s well-knowTn results, proved at present,
however, only when n is not divisible by any square.
52 I Proceeding to the general case, suppose that
l

(5.21) n = 2” (az0; a,a’,

Ma
- >O) w’ af . . l

and consider first Ag .

If x = 1, Ag = 0, by 4.31. If X is odd and greater than 1, A~x = 0,
by (4.321), unless Y = n/ZAS3 = 0 (mod 4) , i.e. unless n = 0 ( mod 2AB1),
or unless X 5 CI!+ 1 If this condition is satisfied, and n = 2” N, so that
l

N is odd, we have
2--3(A-1) e-2af1-A NVR~
Ag = 1
by (4.3223); and so

(5.22) Azx = 2-3+-1) ( x < a + 1) , A 2x = - z--3(+1) (A = Q!+l)*

On the other hand, if A is even, Ag = 0, by (4.331), unlessn = 0 ( I;lod 2+2 ) ,

16Following Landau, I write p f 72 for ‘ p is a divisor of n ’ and p f n for ’ p is not a di-
Gsor of 72’ .

366
19201 REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 277

i.e. unless X Z a + 2. If this condition is satisfied we have, by (4.3322),

A 2x = 2--3V-11 cos ( 2afl---h NT) l

The cosine is 1 if X < a + 1, - 1 if X = Q -t 1, and 0 if X = a + 2. Thus

the equations (5.22) still hold for even values of X. We have therefore
(5.23) x2 = 1+0+2-3+2-“+ .I. +2-3@-11 w/J-3a,

the zero term corresponding to X = 1.

Next suppose that p is odd. If 23 is not an ~3, xP = 1 - ps4, as before.
If P =wandX<a+l,n=O(modw”),and
A,h = (w - 1)0-~*-~,
by (4.415), (4.4231), or (4.433). If X = a + 1,

by (4.422), (4.411), and (4.432). And if X > a + 1,

A,x = 0,
by (4.421) and (4.431). Thus
(5.24) xw = 1+ (0 - 1)0+ (a - l)cJ’+ l l l + (0 - +-3a--1 -@a-4

= (1 -u-“>(I +K3+ur6+ l a’ +u-3a).

From (5.23) and (5.24) it follows, as at the end of 5.1, that

1~~3~1+2-3+~..+2-“(“-1~-2-3”)~(1+~-3+***+kj-3”),
PsW =
w
it being understood that the factor in curly brackets is to be replaced by
unity when a = 0; and it is easily verified that the formulas (5.14) and (5.15)
are still correct.

6. SUMMATION OF THE SINGULAR SERIES WHEN 8= 5 AKD n HA4S NO

SQUARED FACTORS

6.1, Suppose next that s = 5 and that n is not divisible by any square,
and first that p = 2. Then Ale = Aa:! = 4 4 = 0, by (4.321) and (4.331). l

And As = 0, by (4.321), unless n = 1 ( mod 4 ) , in which case, by (4.3223),

ThusAs= -~ifn=1(modS),A8=~ifnz5(modW8),andotherwise
As = 0.
Next,
A 4 = 2-312 cos(i n,n - i T),
by (4.3322), so that A4 = - 2 if n = 1 (mod 4) and A4 = $ otherwise.
Finally AZ = 0 in all cases, bv” 4.31 o

367
 278 G. E. EARDY WY

Collecting these results m*e find that

(6.11) fi=+(n=l), xz=$(n=2,3,6,7), xz=p(n=5),

the congruences being to modulus 8.
Ifpisoddandp f n,wehave

AP A ps= A,* = l *+ = 0,

by (4.412), (4.421), and (4.431). If p is odd and p [ n, we have

A P 3 -?
0 A pa=-p -4
9 A Pa= A += l *. so,

by (4.416), (4.432), (4.421), and (4.431). Thus

(6.12) (P t n),

If 7z + 1 (mod4), we have

Also
1
1 -- 4
P
where nz runs through al1 odd values prime to n. Hence finally

(6.13)

If n = 1 ( mod 8 ) the value of ~2 is 8 instead of 8; and if 12= 5 ( mod 8 ) it

is 3 . In these cases the numerical factor 160 must be replaced by 80 and by
112 respectively.
These are the results of Eisenstein, subsequently proved by Smith and
Minkowski by means of the arithmetical theory of quadratic forms. The
series (6.13) is easily summed in a finite form, by methods due to Dirichlet
and to Cauchy. I have nothing to add to this part of the discussion.

7. THE GENERAL CASE WHEN 8 = 5

7.1. So far it has not been necessary to distinguish between one type of
representation and another. At this stage the distinction between 1cprimi-
tive ” and cl imprimitive ” representations becomes of importance.
A representation
n = x; + XJ + x; + x; + xx

368
1920-J REPRESENTATION OF A NUMBER AS THE SUM OF SQUARES 279

is said to be imprimitive if xl, x2, 53 j x4 1 x5 possess a common ,factor, and

primitive in the contrary case. It is plain that imprimitive representations
can exist only when n is divisible by a square. When s = 8 (and the remark
applies equally when s is 2, 4, or 6) the distinction is, for our purposes, irrele-
vant, even when PZis divisible by a square: the formulas (5.14) and (5.15)
are valid in any case. But when s = 5 the distinction is important. It will
be remembered in fact, by anyone familiar with the work of Minkowski and
Smith, that th.e right-hand side of (6.13) represents, in general, not the total
number of representations but the number of primitive representations. Oru
series (2.26), on the other hand, gives the total number of representations;
and its relation to the Smith-Minkowski series must therefore generally be
more intricate than in the simplest case treated in 6.1,
The theorem which I shall prove is as follows:
The sum of the series
c
(7.11) -n312
7r2
where m runa through all odd numbers prime to n, and

C = 8O(n = 0, 1,4), C = 160(n = 2, 3, 6, 7), C = 112(n = 5),

the congruencesbeing to modulus 8 t is TV ( n ) , the number of primitive repre-
sentations of n .
We shall require the following
LEMMA. If17

(7.12) r(n) =

where q2runa through all squared divisors of n, then

(7.13) t&(n) = f(n).
To prove this, suppose first that 12is divisible by p2, but by no other square.
Then
4cn> = r(n) -+(s) = r(n) - r(s) = r(n).

Next, if n is divisible by p2, pr2, and (p~‘)~, where p’ # p, but by no other

square, we have
4(n) = r(n)

= r(n) -

17In what follows I omit the suffix 5 in ~6 (n ) , etc.

369
 280 Cr. I-I. HARDY [July

A similar proof applies if p = p’; and it is plain that a repetition of the argu-
ment leads to a general proof of the lemma.
7.2. Suppose first that r2 is congruent to 2, 3, 6, or 7, so that 12is not
divisible by 4 and ~2 = k, by (6.11). If we write

the value of xP requires reconsideration when p is an e3 and a > 1. Using

the formulas of 4.4, we obtain the’ following results.
lf a = Zb + 1 then
A W
= 0f A w2= ( #- w4 -, A wa= 0, A ,4= ( 0 - l)u7,
...9 A “a-1 = 09 A w2b= ( #- 1) u-3bl Y

A ,!2Wl = 0, A ,2b+2 = - u
-3b-4
1 A w2H8= A ,2bi4 = ’ l m = 0 .

If a = Zb, the values of the A’s, up to A,,,, are as above, but

cd
-3~2
? A ,2bi-2 = A ,2b+3= “* = 0,

where v = LP n . We thus find that

- u-4) (1 - m-3&3)
(1
(7.22) xw =

if a’is odd, and

(7.23) xw ==:(l - w14)(1 - ylbl]l:l + (3+-3b
- c3--3

if a is even.
Suppose now that n = 2” tia W’ ar = w2d, where d has no suuared l l l

factor. Then the odd primes p fall into four classescharacterized as f;llows.
(i) p=pJPf ?z. Inthiscase
n
(7.241) xp = 1+ ; p2.
0
(ii) p = ul, ~3~+ O, w&L In this case
(7.242) x wt = 1 - or4 l

(iii) p = w, +, QJZt a* In this case a is even, say equal to zb, and

(7.243) xw2 = (1 - wr4) 1 + wr3 + l l l + a;=+3 + W’3b t

1

by (7.23).
(iv) p = 03, @3/U, &. In this case a is odd, say equal to 2b + 1 I
and
(7.244) x wt = (1 - UT*) (1 + UC3-pa* + W;3b),

370
19201 REPRESENTATION OF A NUMBER AS THE SUiM OF SQUARES 281

by (7.22). And we have

(7.25) s = %rI xp=-; II XP II XWl I-I Xw2 II xw3 l

7.3. We now multiply out the product (7.25), treating the second factors
of Xw2and X”3 each as a sum of b + 1 separate terms. We thus obtain

(7.311)

X S i!12,if wig = aib2 is the highest power of 02 which divides n; p 5 ba,

if ~2 = uib3+l is the highest power of ~~13which divides n; and 8 is an addi-
tional factor which is equal to 1 unless X = &, and then to

If we denote the product which appears under the sign of summation in

(7.31) by Q~,PJ we have
4,rr2
(7.32) p(n) =- 3 n3J2S =;-n160 3’2rIYpo C%,P = c P&W?
P
say.
Suppose first that X does not, for any ~2, assume its maximum value bz,
so that all the 8’s in Q, P are equal to unity; and write

(7.34) *(n) = Fn3j2 C( i)k2,

so that

*(n) = ~n312nYpo,
le
the product extending over all odd primes which do not divide ;II. Then

where
q2 = n&y-I&p
w2 w3

is a typical square divisor of n, division by which does not eliminate com-

pletely any prime factor of 12.
This transformation would not, as it stands, be valid if X = b2for some ~2,
since there are then certain primes 02 which divide n and not n/d. &It
with each of these primes 02 there is associated an additional factor 6 = ywl ( Y)
in QA,
&b
1 and these factors provide exactly the corrective required. We have
 282 G. IL HARDY WY

therefore in any case px,P = I) ( n/q2) and

T(n) = p(n) =

the summation extending over all square divisors of nI And therefore, by

um, ti (4 = ‘i; ( n ) , the result required.
Our theorem is thus proved when n is congruent to 2, 3, 6, or 7 to modu-
lus 8. In order to prove it when ~2is congruent to 1 or 5, we have only to
write Q or 5 instead of 4 throughout our argument. It is only when n is
divisible by 4 that further discussion is required.
7. 4. We have now
n = 2” N = 2” ua /’ .. . (aZ2).

The value of xP J when p is odd, is the same as before. The value of m may
be calculated by means of the results of 4.3; and we find that
1 1 1 1 1
(7.41) ~2 = I, -4-48-4x2- l m’ -4Fl+F8@,
. . .
l

where a is odd and equal to Zp + 1 or even and equal to 2p.

Let us denote by T* (n) the number of representations of n which are
primitive SOfur as 2 is concerned, that is to say in which x1, x2, x3, x4, and x5
are not all even. It is plain that
n
(7.42)
and that
r(n) =r*(n)+r* a + .**+r*
0 (
12
48 J
>

(7.43) r*(n) =
co n
F -2 ,
Q
where now the summation applies to all odd square divisors of n . Further,
as in 7.1, we can show that ;f
(7.44 r*(n) = x4(;),
where the summation applies to all odd square diwkors of n, then
(7.45) 6(n) = r(n).
Bearing these remarks in mind, we can complete the proof of the theorem
as follows. Since p(n), PCM, differ only in the factor ~2 -and the
l l l

outside power of n, an, . , we have, by (7.41),

l l

1 1 1
- 8l3-1 1 -
-

4 - . l . - 488-2
. + m=i
.

372
19201 REPRESENTATIOPir OF A NUMBER AS THE SUM OF SQUARES 283

But, by (7.42), we have

r*(n) =r(n)-r 0
; =p(n)--p (i) = 4.85($) = 4.8~~‘r(-.$);

and therefore, by our previous results,

r*(n) = 4.8@-lzF

the summation applying to all square divisors of n/4@, or, what is the same
thing, to all odd square divisors of n. Hence, by (7.45),
fi(n) = 2?(n).

This completes the proof of the theorem. It is easily verified that the
results are in complete agreement with those of Smith?

8. CONCLUDING REMARKS

81. . I have assumed, throughout this paper, that a Z 8; and it is well

known that the analogous results when a > 8 are false.
The analysis of the paper breaks down, when s > 8, in one section only,
namely Section 3. We can still form the singular series, and sum it by methods
differing only in detail from those of Sections 4-7. We obtain a simple
function of the divisors of 12 when a is even, a series of the Smith-Minkowski
type when a is odd; and this series can still be summed in terms of the quad-
ratic residues and non-residues of n. We can still prove, moreover, that
the sum of the singular series behaves, in respect to the fundamental trans-
formations of the modular subgroup &, exactly like the appropriate power of
the theta-function 8, and that the function corresponding to 7 ( 7) is an
invariant of the group. What we cannot prove is that 7 ( 7) is bounded;
and the conclusion which would follow from this, namely that 7 ( 7) is constant,
is in fact false.
We have still, however, all the materials for a complete solution of the
problem. But it is necessary to replace the analysis of Section 3 by a more
complex discussion in which we deal not with a single invariant but with a
linear combination of invariants, among which that represented by the sum
of the singular series is the first and most important. And our conclusion
will be that the number of representations of n is the sum of a function of the
types considered in this paper and of a number of other arithmetical functions
defined in a more recondite manner. Some of these functions have already
appeared in the work of Liouville, Glaisher, and Mordell. If I do not pursue
I* See in particular pp. 673 et seq. of the second volume of his. Cotkcted Papers.

373
 284 G. H. HARDY

this subject further, it is because such developments seem to be a part of

Mr. Mordell’s researches rather than of mine.
There is another question which arises more naturally out of my own
researches. The singular series or principal invariant yields in any case ah
asymptotic formula for T&( n ) , valid without restriction on s . But, with the
entry of asymptotic formulas, the peculiar int.erest of squares as such departs,’
and the problem becomes merely a somewhat trivial case of the much larger
problem usually described as Waring’s problem, and so of the investigations
which Mr. Littlewood and I are publishing elsewhere.

374
(b) Waring’s Problem
 INTRODUCTION TO PAPERS ON WARING’S
PROBLEM
Hardy’s papers on Waring’s problem were all written in collaboration with J. E.
Littlewood. In a sense, they followed naturally from his earlier papers on the repre-
sentation of a number as a sum of squares, but the new difficulties which had to be
overcome were severe.
The starting-point of the method is the same. For each integer k: >, 3 the generating
function is defined as
f(x) = 1+2 2 xnk,
T&=1
so that f( x)3= +X=0
2 q&)xn,
where r&z) is the number of representations of n as the sum of s absolute values
of integral kth powers. For even i% this is, of course, the same as the number of
representations

=r&) 1s
by kth powers.

r(n), =2i
Then by Cauchy’s Theorem the number r(n> is given

I?
(f (x))“x-“-1 dx,

where the integral is taken along the circle 1x1 = e-I/Y This circle, x = e-n-1+2mi0, is
divided into ‘Farey arcs’ by first marking all rational points 8 = p/q with
1 < q < ?a==, (PP4) = 1, o<P=q.
Two neighbouring ration& 8 = p/q, 8 = p’/q’ are separated by their mediant
e= w-Pw4+41)~ SO that to each rational p/q there belongs a zone of influence,
bounded by its two neighbouring mediants, These zones are the Farey arcs MPQ’
In the treatment of Waring’s Problem they fall into two classes,major am if q < nli’e,
and minor arcs if nllk < q < d--T
On each major arc Hardy and Littlewood introduced an approximating function
Fp9*w = Cr(slk)(q-lS~,~)g{log(e2”i”‘pX-l))-s’k,
where c = {Zr( I+ l/k)}8{ I1(s/k)}-1,
q--I
sP3,jz = 2 e2’rrihkpfqa
h=o

They were then able to show that on each major arc Mu

q the function f B(x) can
be approximated by F,,(x) with a sufficiently good error teim not to upset the final
result. The difficulties ‘they had to overcome were formidable and their first paper
on the subject (1920, 2) makes impressive reading even today. In one respect they
were perhaps fortunate, H. Weyl’s celebrated paper (Math. Annalen, 77 (1916), 31%
52) on exponential sums had just appeared and the methods developed in it made
the work more manageable.

377
 INTRODUCTION TO PAPERS ON WARING’S PROBLEM

The second difficulty arose on the minor arcs, which do not occur at all in the case
of squares. Whereas Weyl’s inequality was a convenience on the major arcs, it was
and to a large extent still is a necessity on the minor arcs, because the error in the
approximation to f”(x) by J$&$ becomes too large; indeed it is larger than the
principal term.
The third difficulty was to deal with the singular series. Having replaced fS(x) by
Fp v,(x) on all major arcs, Hardy and Littlewood obtained for s > EZk-1 the asymp-
totic formula
r(n) - cn-l+dkS ?

where

is the singular series.

In their first paper (1920,2) they did no more than prove that the series is absolutely
convergent and greater than a positive constant for k = 4, s = 33 and all n, though
they suggested that there would be no serious difficulties in extending this result to
general k: > 3 if s > k2 k--l . This programme was carried through in (1920, 5),* the
first of their famous papers on ‘partitio numerorum’.
In EN. II (1921, I), the asymptotic formula is proved for k = 4, s = 21 and it
is shown that the singular series is uniformly positive. Apart from a detailed study
of the singular series the paper introduces an improvement in minor arc technique,
namely the use of the inequality (m denoting the union of all the minor arcs)
I
m
s 1f (r@io)I28d0 < max ]f (re2nie)j2cg-2) 1f (reznie)I4 de.
einm s
0
For the first factor on the right Weyl’s estimate is available, and the second factor
can easily be estimated by elementary means.
In P.N. IV (1922, 4) the authors prove their asymptotic formula for r(n> for all
s > (k-2)2k-1+5. But the real interest of the paper lies in the study of the singular
series. It is now commonplace to realize that the inequality S > 0 implies, subject
to absolute convergence, that the congruence
x:+.,.+x,” = n (modm)
is soluble for every modulus m; conversely if the congruence is soluble for all 132,then
S > 0. In fact we can be more precise. If in the summation formula for S we
restrict Q to the powers of a fixed prime W, the series turns out to be finite for n # 0
and represents w -0-1) times the number of solutions of the above congruence mod+
for sufficiently large t. Moreover, S equals the product of the restricted sums extended
over all primes W.
This idea, which is fundamental not only for Waring’s Problem, but for all applica-
tions of the Hardy-Littlewood method, appears first in P.N. IV. It is historically
interesting that the important relation between the singular seriesand the congruence
was formulated by the authors as a lemma only (Lemma 2).
* There WBEI a preliminary Eannouncement in Proc. London Math. Sm. (2), 18 (1920), vii-viii.

378
 INTRODUCTION TO PAPERS ON WARING’S PROBLEM

In P.N. VI (1925, I)? the authors break new ground. They prove first that almost
all positive integers are the sum of fifteen fourth powers, and more generally the sum
of
(&k- 1)2k-f+3
non-negative Tcth powers for k = 3 and k: > 5. Secondly they prove that all large
integers are the sum of
(k--2)log z-log k+log(k-2)
(#k- 1)2k-l+k+5+
log k-log(k-1)
I
non-negative kth powers. The idea of the proof is to introduce a sequence of integers
which have a relatively large density and are representable as the sum of few kth
powers. It was the first step in the combination of analytical with elementary
methods, serving as a signpost to the future, and in particular to the work of
Yinogradov.
Finally they introduced in P.N. VI the Hypothesis K, which asserts that the
number of solutions of
n= xt+.*.+x;
is 0(n~) for each E > 0. Under the assumption of Hypothesis K they proved that
their asymptotic formula for r(n) holds for s 2 2k+ I, in particular that each large
integer is a sum of 2k+ 1 non-negative kth powers, if k is not a power of 2. Unfortu-
nately, K. Mahler proved (Journal Londm Math. Xoc. 11 (1936), 136-8) that Hypo-
thesis K does not hold for k = 3 by means of the simple identity
(9x4)3+(3xy3-9x4)3+(y4-9x3y)3 = y12.
Whether the hypothesis is true for k > 4 is still an unanswered question. For the
application to Waring’s Problem it would suffice if the hypothesis held in a weaker
‘mean square’ form.
P.N. VIII ( 1928,4) differs from the preceding papers in posing a purely arithmetical
problem. Let l?(k) be the smallest integer s such that for all n and all primes p, and
all m > 0, the congruence xf+,..+x,” z n (modp”)
has a solution in which not all the x, are G O (modp).
The authors determine five distinct classesof k for which l?(k) > k and calculate
l?(h) explicitly. They also prove that l?(k) = k if 2k+ 1 is a prime, For all other cases
they prove the inequality r(k) < k.
As a conjecture they mention the innocent-looking relation
lim r(k) > 4,
k-a

which has remained unproved. It would be sufficient to find for all large k a prime 7~
such that 77L_ 1 (modI%), 7~< k4/3.
Finally, a footnote should be quoted:
‘We may add that “P.N. 7”, which is still unpublished, contains an application
of our methods to the problem of the order of magnitude of the difference between
t An ahtract appeared in PTOC. London Math. Sot. (2), 23 (1925), xx-xxi,

379
 INTRODUCTION TO PAPERS ON WARING’S PROBLEM

consecutive primes. We prove (subject to our generalized form of -f7ie Riemann

hypothesis) that
'
lxm P,+l--P~ <gJ
n3 logn ‘- 3’
The paper was never published, but R. A. Rankin (Proc. Camb. Phil. XOC. 36 (194Q
X15-66) proved that the lower limit in question does not exceed 3/5 under the
generalized Riemann Hypothesis, whereas P. Erd% (Duke Math. J. 6 (1940), 438-41)
proved without any assumptions that it is less than 1. Later R. A. Rankin (Journal
London Math. Sot. 22 (1947), 226-30) proved that the lower limit is less than 57159.
The subsequent major improvements upon the Hardy-Littlewood results in
Waring’s Problem are due to I. M. Vinogradov. An account of his results is given
in his book, The Method of Trigonometrical Sums in the Theory of Numbers, English
translation by K. F. Roth and A. Davenport, 1954.
He proved first (Annals of M&h. 36 (1935), 395-405) that
G(k) < 6klog Ic+(4+log 216)k,
where G(k) is defined as the smallest s such that
lim r&n) > 0. A simplified version
n+m
was given by H. Heilbronn (Acta Arithmetica, 1 (1935), 212-21). Subsequently
I, M. Vinogradov improved this result progressively in various papers, and in his
book he proves G(k) < k(Nogk+ll).
He also in a series of papers considered the validity of the asymptotic formula for
r&n). The best result given in his book is that the formula holds for
s > loPlog k-1.
L. K. Hua had previously (Quarterly Journal, 9 (1938), 199-202) proved the
asymptotic formula for s > 2k+ 1.
The inequality G(3) < 7 was first proved by U. V. Linnik (Recueil Math. 12 (EW),
218-24) and subsequently a simpler proof was given by G. L. Watson (Journal London
Math. Sot. 26 (1951), 153-6).
H. Davenport proved that almost all numbers are sums of four positive cubes (Acta
Math. 71 (1939), 123-43); that all large integers not = 0 or 15(mod 16) are sums of
fourteen fourth powers (Annals of Math. 40 (1939), 731-47); and that G(5) < 23
and G( 6) < 36 (American J. of Math. 64 (1942), 199-207).
Finally, the problem raised in P.N. VIII, that of improving the upper bound for
r(k), has not received much attention. It was, however, proved by I. Chowla (Proc.
Nut. Acad. Sci. India, sect. A 13 (1943), 195-220) that r(k) < kc, for some c with
0 < c < 1, in all cases in which Hardy and Littlewood had not shown that P(lc) > k.
In recent years there has been renewed interest in the Hardy-Littlewood method,
which has been successfully applied to prove the existence of integral solutions of
homogeneous algebraic equations with integral coefficients in a sufficiently large
number of variables (subject to certain conditions if the degree is even). The proofs
are very sophisticated, requiring a good deal of ‘old-fashioned’ algebraic geometry.

380
 INTRODUCTION TO PAPERS ON WARING’S PROBLEM

For some of these results, and for references, the reader may consult H. Davenport,
Andytic methods for Diophantine equatiuns and Diophantine inequalities, Ann Arbor,
Michigan, 1963, H. H.

381
 A NEW SOLUTION OF WARING’S PROBLEM.

By G. IL HARDY and J.E, LITTLEWOOD.

1. T’ was asserted by Waring* in 1782 that every number

1 ia tile
sum of at most four squares, nine positive
cubes, nineteen fourth powers, and, in general, g (k) positive
k-th powers, where ,9(k) is a number depending upon k done.
Waring advanced no argument of any kind in support of his
assertion ; and there is no re;~or~ to suppose that it rested on
any basis more substantial than the examitlation of a nuder
of particular cases.
If the proposition ‘ every number is the sum of at
most 912positive k-th powers ’ is true for any particular
value of m, it is true u for&i for any larger value. The
number ,9 (Ic) is, by d&&ion, the 8m;rllest value of 3n fol
which- the proposition is true. I’he problem suggested by
Waring th falls ~~atur~lly into two parts. The i-h-at is the
proof of’ the existence ot’ g(k), the secorld the determination
of itf3 actual value a3 a function of k. It iH the tirtit of these
two problems that. has generally been de& bed as T%7~ri~~9’s
l3dxm, ‘i’his problem was solved by Hilbertt it] INN, t’tle
the existence ot’ g(lc) having beet] proved before only when k
is 2, 3, 4, 5, 6, 7, 8, or IO. The second problem is still url-
solved, except wirer1 k is 2 or 3.

* Es Waring, Meditrrtiones AIgehnicae, ed. 3, 1752, pp. 349-350.

t D. Hilbert, ‘ Beweis fiir die Darstellbarkeit der ganzen ZahIen dmxh eine
feRte Anzahl nter Potenzen ( Waringsche . Problem) ‘, mttit~gw Nuchrichten, 1909,
pp. 17-36, md M& Alrttnlen, vol. lxvii,, lW9, pp. 91-300.
$ Twenty-five in the firnt version of his memoir,

1920, 2 (with J. E. Littlewood) Quarterly thurnal Mathematics,

382 48, 272-93.
 2, In this note we yropose to give a short atccaunt of
nnott~r solution of Waring’s Prolhn wlkAl uw IMW dis-
covered recently a,nd wllich proceeds on entirely diffttre~~t lines.
This solution is not, ia my S~HM of the word, elen~entary. It
is based throughout on Caucl~y’s l’heore~n :ind the ordinary
machinery of the theory of ntlalytic fhctions, and has, fro111
beginnitqg to end, no point of contact with Hilbert’s solution.
It migllt seemtllat a I$hly transcendental proof of a theorem
wLclr 1~s already been proved, and that in an entil*ely
elementary manner, is unnecessary. Tlh view, we tlA<, would
rest u ~011a Gap preh ensior1 It seemsto us most desirable and
l

important that W;hg’s Probletn, alld all ainA~r problems of

Combinatory Analysis, should te brougllt illto relation wi tlr
the transce11dental side of the Analytic Theory of’ Nu~nbers.
Further, the method wlkh we follow, and whicfl we describe
shortly as tAe method of Furq dissection, is in many WRJYIthe
mo8t natural alId, in spite of its considerable technical diffi-
culties, the most straightforward method f’or tile discussion of
any problem of this cl&-acter; and it is a mct.hod of great power
and wide scope, applicable to almost any problem colxerning
the decompositiotl of integers into parts ok+a particular kind 11,
and to many against which it is difficult to suggest any other
obvious method of attack.
* F. HausdorE, ‘Znr Elilbertschen Lijsung des Waring&en Problems’, Mcrth,
Atu2dq vol, lxvii., 1909, pp. 30 I-305.
t %. Stridsberg, ‘ Sur la dhmonstration de M. Hilbert du th&Gme de Waring ‘,
Mrrlh. Awden, vol. Irxii., 1912, pp. 145-152. This paper gives references to two
earlier notes by the author in the &*I& f&m Mctte~~til, vol. vi., 1910 and 19 12,
$ R. FLemak, 4 B emerknng zu H errn Stridsbergs Ueweia des W aringschen
Theorems’, Mrth. Awdw, vole lxxii., 1912, pp. 153-156.
5 That is to say, it does not depend upon arguments which involve limiting
processes.
11 8ee G, H, Hardy and S. Ramanujan, ‘Une formule asymptotique pour le
nombre des partitions de tz’, C’omptes Retldus, 2 Jan. 19 17 ; ‘ Asymptotic formulae
in Combinatory Analysis ‘, Proc. London dkth. SW., ser. 2, vol. xvii , 19 18, pp.
75-115 ; ‘ On the cmfficients in the expansions of certain modular fun&o&,
Proc. Royal Sm. (A), vol. xcv., 1918, pp. 144-155 : S. Liaruanujan, ‘ On certain
trigonometrical MHIIS and their applicationa in the analytic theory of r,umbers ‘,
5!iwns. Cad. Phil, Sot., WI. xxii,, 1918, pp. 259-276 : G, II. Hardy, ‘ On the
expression of a number as the sum of any number of squares, and in particular of
five or seven’, Proc i’Vationa2 Acrid. of Scie~~ces (\Vashington), vol. iv., 19 18, pp.
189493 : G. H. Hardy and J. E. Littlewood, i Note OII Meem. Shah aud V\‘ilaon’ti
paper ell titled An anzpiricd fomda connected with Go’ldbnch’s i%%JWJ2 ‘, fhc.
Camb, l’hi/. SOL, vol. xix. (1919), pp. 245-254.

383
 Moreover, when our metllod is applied to this particular
problem, it yields a great deal mnore ht~ can be obtained by
more elementary methods. In particular it enables us to assign
a dehite upper bound, in the form of a f’unction of k, not
indeed for 9 (Q but for allother number G (Ic) which seems in
some way; more fund;lment;tI. ‘l’t~itr number G (Q ia the
least number 112of which it is true that every number fm12 a
cerlcr:i?zj&l mwards is tile sum of at most 712positive k-th
powers. It is obvious that G(k) <g(k), and it would seem
that itI general Q (k) <g (kj; alld tllnt G (k) is more
fundamental than ,9(k) bec:luse its vnlue is less likely to be
determined, in any particuh c:lse, by a number of arithmetical
coincidences, The value of CJ(h) ia not kuowu for arly value
of k save z ; nor has any general upper bound for G (K) yet
been determined. Our method elIaMes UHto prove that

(2 I 1) G(h)52k-1k+ I.
Finally our method yields not only a l>roof that repre-
sentations of 12 in the form required exist, but also asymptotic
fbrmulae for the number of representations.
w e propose to give here a short account of our aualyh.
not complete ill detail, ht full enough to shew cledy the
mrrin ideas on which it is founded. ‘i’l~e complete investigation
will be published later.

3. Suppose first that k is even, and let us denote by

tile number of representations of 32in the form

72= a,‘+ a,l’+...+ a#‘,
where the a’s arc integers, positive, negative, or zero, and
representations which differ only in the order* of the a’s are
reckoned as dihct. Then r, &2)Js the coefficient of x” in
the expansion of the fumAou ’

(3.1) fk*(x)=
t [f,(x)}‘= (1 + 2bk)r= %-k,,(,,.n.
1 0
We shall sometimes omit the suffix K to f or II’, when no
ambiguity results.
It is convenient to conserve the same notatibn when R is
odd. The arithmeti&l interpretation of rkf,(~a) iB tile11 not

384
 quite SO simple, but the function is equally well ad@ed for
our purpose. In particular, wider h is odd or even, Hilbert’s
Theorem may be stated as follows:

* We admit the value 0 for p when q= 1 only.

t These asymptotic formulae are valid when x I) xp,, along my regular path
.which does not touch the unit circle, The behaviour of fk (z) whet1 x tends to an
‘irrational’ \point of the circle is far more complex : for a detailed study of the
case k = ‘2, in which fk(x) is a Theta-function, see G. IL Hardy and J. IL Little-
Mod, ‘ Some problems of Diophnntine Approximation ‘, Actu Mdicmt ica, vol,
xxxvii., 1914, pp. 193-238.

385
 Fa ($5)= ii va-’if = r(n) (log a)-+ G (x),
1

IV 11e1*o c={2++#7r (1+&),

instead of (4.3). We are thns led to replace fk,, (X) by

S
(5.1) $ ha(2) = I + cz Au zF,$,e(Xe-2P”ii4),
( 9 )
and Y&H) by the coefficient of X” in +k,l (x), viz. by

The srmrmtions apply to alI positive values of q, and al1

positive values of p less thnn md prime to 9, p = o being

386
associated with q = I alone; ana tile constfillt term in @I)
is added merely for tile s&e of formal correspondence, Arld
it is of course to be understood that our wMe argument, up
to tllis point, is of a purely form;kl character. It prove3
nothing ; it merely suggests that pk ,(12) rnzly not unreasonably
be expected to i&Al some sort ot” qproxirnation fbr 5J12),
a hope wlkh more rigorous atdysis proves to be well-
founded.
We shall call the series

(5 13)

6. It will evidently be essential to prove that the singular

series (5.3) is convergent for a giveu Ic alld all sufficiently
large values of s. The most obvious wily of effecting this ia
to shew that
Sptq= 0 h%
uniformly in p, c~being a fuuctiotl of k only and 0 62 < 1. It
is well known tLt 8 = 0 (dq) whn k = 2 ; it is riot
difficult to prove, by theP;Pse OCknown results in the thea*y of
tile divisim OF the circle, tllat
Sp) q= Q @+‘)
when k = 3 and Sp,*= 0 (8”)
when k = 4, E being any positive number ; and it appenys
that we m:ky suppose, in gellertil, tllat a is any number greater
tflan (Ic - 1)/h.

where m is an arbitrary integer; and we are unable to prove,

fur these more geaeral sulils, result8 21saccurate as tlrose stated
above for tile specitil sum8 SP4, wlGch seem to be substathally
tile best possible of tth &d. We have not at present
proved any result, collcernirlg tile order of SpQ,,t, which seell]H
likely to embody the beat tl~t cata be stated tGIy.* For some
* We have proved that, if q ie prime,
8 p,q,m = 0 (P”)
uniformly in p and nt.

387
purposes, in particular that of solving the second problem of
6 I, it is of the utmost importance that the best poAble result
nf HIis kiild should be discovered. But for our first purpose,
that of proving Hilbert’s Tl~eorem, not so much is necessarym
All that is required is 50172e result of the form

and such a result ia fortuwtely not difficult to find. We can

in fact p~mw, by tire use ofcertairrl very elegant trarlsforlnations
due to WeyI,* tllat

7. It follows at aIce th:tt the aillgulw series i9 convergent

for sufficiently large values of 5. FurtIler, its &solute value
may be mde, by choice of IX sufficiently large value of a, aa
near to unity as we pleue. !l%e tirst terrrr is unity : let us
write the series, then, in the form
s= I + 8’.
We have I sp,p/p.I e 1
for q 2 2 ; and I x,,,/a I < @rzp9
where p = 1/4K and H depends on k only, for all values of Q.
Choose a value of Y RHCIItllat
w> H-,

aud cnn therefore Le made atssmall a~ we please by chuice of S.

* H. Weyl, ‘ Uber die Oleicltverteilung van Zahlen mod, Eiug ‘, Matk Amah,
vol. Ixxvii., 1916, pp. 313-352, See in particular p. 336, equation (11).
t ‘l’he number of values of p aaaocictted with each value of Q ie + (q), the
number of numbera less than and prime to +
 279
It follows from (5.2) that
(7 1)
l j Pk,#(11)j > ~c’,(6/kl-1
for all sufficiently large values of s attd all values of II,

(8 11)
the cotltotir of itttegration being a circle I? whose cetltt-e is the
origin and whose radius B is less than 1. We take

111order to study the integral @I), we divide the circle

r into a large nuher of srr~all pieces by what we call a Farel/
dissectiulr. It is the use of this tnethod of dissection which ie
the most charactet-istic feature of our analysis.
!lh Rhey)d sen’esof order N is the aggregate of irreducible
rat iorlal h,chtls p/q, whet-e o -<p 5 p 5 1y, arranged in as-
cending order of mng~~itude; it is plain that p> O except
when y = 0, y = I, aud y <p except when y = I, g= 1. ‘I’ht~

is the Farey’s series of order 5, It is known that, if p/q nud

p’/~’ are two successiveterms itt a Farey’s series, tltett
$4 -pqr= 1.
Further, if p/ q is a term in the Farey’s series of order N, and
$‘/q” and p’iq’ are the adjacent terrns on the left and the
rigll t, and jp, Qdenotes the interval

then the intervalsjpk exactly fill up the cothuurH (0, I), and
the length of the t&o parts into which jP,q is divided by p/q
is greater than I /ZqN ntjd less than l/qN.* ‘.l~e essenti;\l
poirlt ir3that each of the ports o’j, 1Qis of order ~/qlv.t
* When p/q is O/l or I/f, J& consists of a single part only, one or other of ;u”/q”
and p’/q’ beirlg non-existerl t.
t For proofs of all these nssertions see the second paper of Hardy mii Rama-
nujan quoted otl p, 278, $§ L21 9 422,

389
 If no~v we suppose that 17 is defined by
x = BezTie (0 ( 8 5 l)t
that the range of variation of B is divided up by hking r~s
points of division all points belonging to the Farey’s series of
order IV, and that the two extreme intervals, ending at B= 0
and 8 = I respectively, are regarded as definirlg a single arc
af the circle, we obtth the dissectiotl of the circle which we
describe m the lhey dissectbn of order N, l’he arc associated
with 21/p we call the I%re~ arc & 4,
We shall find it convenient to whe
I k-l 1
(8 3) -z
Q, - =l-(q -= b -=
k-2 I w b9
k-q I
l

k k ’ k-1
80 tht a = b/(1 + b) nnd b = a/(~ ma). We take N- [PI;
md we shll find that, for the purposes of this problem, the
Farey arcs fall irlto two ches, arcs for which Q< 72a, which
we ml1 m@r arcs, and arcs tbr wlhh 12a < p < dwa, which we
chll 772hm* arcs. ‘i’he behaviour off8 (x) 1~s to be studied by
quite different methods, according as x lies on an arc of one or
he other of these two categories. When ?C= 2, a = I -a,
and the minor arcs disappear: thus one of the characteristic
difficulties of the problem ia abseut in this particular case.

9. We begin by considering the mnjor arcs. Ue charac-

teridicof a majo; arc is tht f.(x) ;k to a certain extent
dominated by the approximative t’uuction (44, viz,

A (5)’ (log $)-=.

We can prove in fact that, on a major arc

(9IIj f(x)=1 t 2w=qbp,+cP I Pd
where
S I -a
(9 I2) * p*=2r(1+a)
I - logx
P ( >
and
(9 3)l @ = W ~p”fE
PiQ
1 7

for every positive value of E. Let us suppose, to fix our ideas,

&t q is not hrge and that .X is real. ‘hen $P Qis, in general,
of order 72a, nud therefore, &Ice K < I, of h&r order than

390
 of Waring’s problem.

1
--I = log --ie=v-ie,
(10.2) l;i-0‘6 ;k A!

an itIterval extellding, on e& side of U,

t ht1 1 / 2qiv id less t hl 1 /piv* ‘lb
uestiull is then

=Zr(l $ a) X,,Y-“+
) 4% Bp,,m$
1
(IV),
1
where

TVllen A= 2, tp (~2) can be ev,zluated in terms of elementary

functio~~s, and we are led bock to a ft-rmilitlr forrr~l:~ irl tile
theory of tile ‘I’lleta-furictiotls, Tile general fr~rrrnd~ (10.4)
may be proved witllout difficulty by tire ntetllods of tile
calculus of residues, as expounded by Li~ldelijt’ in llis work
quoted in Q5,

The fmctimt + (nt).

11, It is now necessary to study tile integral (10.5) iu
golne detail; and a variety of rnetllods ;tt*e av~iMle for tl&
diwussion, We observe tirst tlr;lt % or am X rises, towards
the ends of a ~naior NT, to tile order of m;tgtlitude I/J@
Sirwe q < 32a and iV<- &a, tllis is never of lower order than
l/n, and in geueritl of greater. Ths irl gener;tl tlm Y is
nearly equal to - +r when x is at tile upper end of a ma,ior

391
 (i) Whatever method is followed, the annlysk is a little
aimpIer when k is even? Let us suppose first, to fix our
ideas, that k= 4. We have

where

Usi~~~ BarlIes’ results, we art: led to the conclusion that,

for the range of values of Y linder consideratiou, we have
(NJ) + (nz) = 0 I&Y-k exp (- HPM Y-J) 1;
wlme H is positive atId, together with tile constmt implied
by the 0, independent both of Y and of ~12, aild tile ambiguous
aigrl is to be interpreted irl the most ut~t’;tvourable nlallller;
chosen, thnt is fo say, so that, when x is Ilear the end ot’ tile
rllajor arc in question,

is aa nearly as possible equal to *vr or - 47~. It is clear that

we must take the plus sign WIMI 8 is positive rttld am Y
rlegative, and the minus rrigtl in the opposite case”.
It was by the ailalysis sketched above that we first obtained ’
the formula -(11A)?, on w lk+ the must critical pact of our
proot (wllen k = 4) depends. It would 310doubt be possible
to treat the general case on similar lines, The formal compli-
catiom of t Ile llecessnry antrlysiv would however be considerable,

* So far as the mere proof of Hilbert’s Theorem is concerned, we might without

prejudice suppose k even throughout; for if (e#g.) a number is always expressible
m the sun-~ gf a fixed number of sixth powers, it ie afortiori alwaya expreseille as
the guru of a fixed nulnber of positive cubes. For other purposes it would be very
undesirable that k should be subject to any restriction,
f H, W. Uarnes, L’l’he asymptotic expnnsion of integral functions defined by
hypergeometric series ‘, Pwc. Lmdon With. SOL, aer’, 2, vol, v., pp. 69416.

392
and b = &.
WI& k is odd,
00
$ (712) = Q p- Yt4~+2m?rui du + i e- Yuk-2tlmdi &
I

Each of the inteE;rals on the right m:Iy be analysed by the

method of steepest desceuta, hut there will be two term3 in
* When LCis at the centre of the arc the sign is indiffererrt, since Y is real and
cos+w = CO8 (- @r) = &*
t It would be possible, by using the full force of Barnes’ results, to wplace
(I 1 .I) by a far more accurate formula, but Such refinemerlts would be irrelev>Lnt to
our purpose.
$ See, for example, G. N. Watson, LThe harmonic functions apsociatell with
the parabolic cylinder ‘, Proc. London Jhth. Sot., ser. 2, vol. xvii. 11918), pp, 1 IG-
148, where mnny references to the literature of the mljjt?Ct nre given.
8 There are E- 1 ‘~01s’ : the one with which we are concerned is that which
is real when Y in purely imagihary.

393
 fixed elld of the patI1 of iGegra.tion. It will be f’ouud that
(11.3) $ (m) = 0 (4) -I--0 (972+-I1Yj).
‘he second term is trivial : there is nothing of importance in
ill NW ;trlalysis whh depends uPon the distinction between

Jf k is odd, our upper bound (11.3) for ) $(w) ( cont,zins two

telm~. ‘he second of these, wheu suhtituted into (I 2, I ), will
coM&ute
0 (qK+”[Y 1Znz-“-I) = 0 (QKf&j Y I).
~This contribution is hogether trivial ill compkson with other
error terms which we aMI be compelled to rctrliu. We n+hly
thereI’ore proceed with the argument CMthough LJwere even,
Suppose now hat B> o and
Y= 9’” (v - ;e> = qk p&,
a0 that p=vsec+ 1/(12co4).
Then

pt {- Hef(‘-“J”iml+bY-bJ =-HTI~~+~~+~v ’ (cos~)“cos!P,

where YP= + (I - b) r+&
It is easy to verify that the lxtio co&/cos~ remains between
positive bourlds f’or o <#s $7~ Ttms
1E j = 0 [~XP [- JdtbflW (COS$)ltb]],
where J, like H, is independent of’ 72, q7 or m. Finally, front
(11.2) and (124, we deduce
(12.2) 4p,q= 0 [q"' 1 Yl-@
x 5 &PbJ exp (- Jml'bp-l-~Ilb(coS~)I+Bj].
I

394
 r.
uf Wu3’il2g’s yrobzenz. 283

13, Pn using (12.2) we must distinguish two cases,

(i) Suppose first that 910< I.* ‘ll~en S/v is less than,
and cos$ greater than, a positive constant. Also
f2bq-l-& (nq+Qb >_ I
and J712” ’ q-‘-” I? (Co9$)I+ b> hl”3,
where L is a comtant like H and J. Finally, 1Y/ is greater
thr~ a coustnnt multiple of q’/u. Heme

(ii) Suppose that 12e7 1, Then cos$ is greder than

a constant multiple of’ ~/h+ and
q-1-b?2e (cos~)"~

is greater than a constant multiple of

-1-b f’ @-l-6,
4
and therefore, since 0 < I /412’-~, than a constant multiple of
924+(~--4(~+W = 1 l

Also i Y 1is greater than n constant multiple of qkB. Hence

we obtain an equation of the form

Finnlly, B> I 172, so that we are again led to (HJ).$ Wo

have ths proved equations (W)-(W) of’ 4 9,

TIie behctviuur offk (x) on tik n&or arcs.

14. On a minor arc qk) 12; and 1Y 17 pkln, so that Y is
in general h:ge. Ths the argument of the precedirlg sectbus,
which was based on the assumption that Y is generally smell,
fails in principle.
* Any other po sitive constant might be used instead of 1.
t We assume here that k > 2. In this case K > &kb (the sign of equality
occurring when k = 3, K = 3, b= 4). ‘I’he argument rtiqu&ea modificntiou when k = 2.
$ The argument again requires anodification when k = 2.

395
 It is easy to see, moreover, that tire formulae (9.1)-(9.3),
if true, will- no longer express a, genuine fipproxim:\tion to
f (cc). For
I -L 0 (qK+a(&ny}
!PPY
I 0 (,fqK--l+E)
aK+E
o( 72 1 3
=

since ti < I, so that the term $P,p will be irrelevant, and

we shll have simply
(14.1) f(X) = 0 (7taKSE)*
Now there is nothing to suggest any abnormal rise in the
order of m;\gnitude of j’ (;c), new p:Wicular points of the circle
of convergellce, except ill 80 f;cr as such a rise may be accounted
for by the influeilce of a term 9, *I It is such a term which
dominates f (2) on a major arc : bn a minor WC, where there
is no such domirlating iilfluence, tllcre is HOreason to expect
thntf(x) will exceed the order of the error terms wld afbct
all our auatysis. In short, it is rensont\ble to expect (14.1) to
Md over any mirlor WC; and this expectation is in fact
fulfilled, though an entirely different proof is required.
15. Our proof rests upon the following lemma in the
theory of’ ‘ I)ioplml&e Approximation ‘#
Y na- < g < 7P+ ad p < 71a+c,the7z

dA woi-d of explanation is required as to the use of-’E. ‘I’lle

&ted in ‘4 6. The result is sufficient for oL purpose but

(like the result of fi 6) is prob;lbly not final, and fur this reason
\ve reserve the det:kils of the proof,
Our lemma leads at once to a proof of (14.1). Suppose,
for simplicity, that X is real (so that we are considering the
central point of a minor arc). Then
f(x)=r+2 2 xuk+2 xxuk
l_<lNp t-w
= IQ (xl + 0 (11,

396
 (15.4) 4pp=2r
9 (1-t a) %u (log g,
q
x=x/x, Q= x&Q-i/q, a = 1 /If, H:= (Ii- l)/ I!!= I i 2-k+i,
r

16. Returning to the contour integrctl (&I), we have

the summation extending over all arcs EDQ of the Farey

dissection of order 312 = [P]. We write thi’3 equation in the
form
(16.2) 3’,(77)= q + s2,
where Bl comprises the tertm contributed by major NW and
& those contributed by minor arc3, It foHow at owe horn
(15.3) that
(16.3) Is3 = 0 (12baK+9 ;
but Sl requires cotwidernbly more discussion.
* The value of 18 1 at the end of a minor arc in less than 11 qniBa < 11 T$, since
q> ?P. Hence the liues from xP, q to the ellds of its associated arc make with the
radius vector an angle less than @, so that the whole arc lies within a =gion of
the type employed (e.g. by Stoh) in the ordinary extensiona of Abel’s therlrem.
See for example Bromwich, I@uite $wies, pp. 210412. On major adds the li~ti
referred to are generally nearly pnrttllel to the tangent at xp qv

397
 (16.5) Rpn=
I 0 (J EP,Q
= PPq+ %,*~
say. It is plnin, in the

will be convergent if (S- 1) X > 25 a’,e,if

(16.7) s>2K+l;
a condition we shall henceforth suppose to be mtidied. Ad
80
BpRq= 0 {12s4-1+&},
ZR O( @iZK+t) + 0 (@+-nk--1+E),

(16.8) ‘1;(11)= z T S2= XiPI*+ 0 (?PK+~)+ 0 (123n-nX-l+$

17. We next perform two transformations on ipq8 I

(i) We first replace $!, in iPja, by a function of tile tyPI:
considered it1 $5. Ttle t’unctiom

q= {2r (1 +u)}8(y (logya

* It mustbe remembered that 2 impliea summation with respect to p as well

as p.

398
 289

where F is defined 8s in 0 5, differ by a function regular for

X- I ; and it is easy to see that, in suhtituting one for the
other, we introduce only an error term trivial in comparison
with thone already present in (16.8~.

(ii) We next replace ttle arc of integration EPp by the

complete circle of which it is a part. This proce’sa also is
easily shown to introduce only an error term of lerssimportance
than those present in (16.8).

We have thus replaced X& in (16,8), by

where 0 is defined as in 6 5 and the path of integration ia the

circle II; and this is equal to

18. l’he summation h (17.1) extends over all values of p

which do not exceed #, and all valuea of p lessthan and prime
to p. The series remains convergeM, when p ranges to intinity,
if sh> 2 or s> zK, a condition inciuded in (16.7). Also

We have therefore

The equation (XI) container the proof of Hilbert’s

TheoremI In fact, we have z3eeflin 0 7 that the coefficient of

399
290 Mr. Hwdy and ii&. LiMewood, A new soltt tion

n 8u-1is positive (and as near to hty as wt: please) for every 72,
if only s is large enougll. Also
sa-ah--l esa-1,
and SUKctSU-I

ifs&>1 or s > kK- - 2”-‘k .

It is to be observed that this last condition includes the con-
dition S> ZK+ 1, previously imposed upon s, in all cases
except that in which k= 2.
lt follows that Y, (~2) > O for s >_ G (I$ where G (k) is a
number which depends only on K, and fbr all sufficiently large
values of 31. In other words every laqye mmber 29 tfic ~2~712 of
at most G (k) positive kth powers: and’ f&m th it obviously
follows that every number is the sum of at most 9 (k) positive
kth powers, which is Hilbert’s l’heorem. l

It is plain that we have in reality proved much more thah

Hilbert’s Theorem ; in particular* tllat to every k correspotzds
u c, (k) such that
prp + (“lk)}]’&lk)-1 2
x 8e-Znpxijp
A!&
rk ! I b, -
rp + (s/k)} (P )
for s> G, (k). Here G, (k), like g (k) and G (rE), is a nutnbel
which depends only on k,

On the vdues of g (k) ad G (k).

19. The numbers of ,q (k) arld G (k) we the I~st numbers

such that (1) every rlumber is the NM of at most 9 (k) positive
kth powers ;tud (2) every Ecrrgenumber is the S&I of ;tt most
G (k) positive kth powers, It is ohious tht
Q (k) L g (k).
The preceding analysis est;tblishes the existence of these
nulnbers ; but does not, atsit stands, lead to a dehite value of
either, though it suggests very forcibly that
(19.1) G (It). -< Pk+ 1.a
The known results, fbr the first few valnes ot’ k, are as
follows.*
~~- .---_ -..-..
.-
* See the dissertations of A. 6. Kempner, obey &IS Waringsche ProHem and
cinige VeraElgenlcinerungen, GWingen 1912, and W. is. Baer, Bcitrtige mm
Wtwingschelz A-obknt, G ijttingen 19 13.
 In the first place, it is known that 9 (lc) does not exceed
the numbers given in the following table:
k= 2, 3, 4, 5, 6, 7, 8
9 (k) -e 4, 9, 37, 58, 478, 3806, 31353.*
l’llese m~l;bers t’ur~kh, a fortiorz’, upper bounds for G (I& It
has also been proTed by Landau that G (3) _<8.
The corresponding lower bounds for g (Ic) are
g (k) 24, 9, 19, 37, 73, 143, 279.
Thus g (2) = 4 and g (3) = 9, while there is a wide range of
uncertainty fbr all higher values of k
As regards Q (k), there is even greater uncertainty. It is
known that G(k) > k+ I for all vahes of k. Thus 4_<G (3j18,
but the actual v&e of G (3) is unknown : empirical data
point to 5 or 6. Similarly, nothing is known about G (5)
save that it lies betweerl 6 arId 58. When k is a power of&
a little more is known, viz. that
G (k)>- 4k:
in particular G (4) -> 16, so that G (4) lies between 16 and 37,

20. If s> 23, the error terms in (18.1) are of lower

order than the power of 11which multiplies the singular series.
Or~r analysis therefore makea it appear very probable that the
itlequnlity (19.1) is true. The vtllues of the ri@t-Iland side
of’ the illeqtl;klity are 5, 13, 33, 81, 193, 449, 1025; so that it
lvould give new values for G (k, except when k ia 2, 3, OL’
5. Itt order- to prove it gerlel*i\lly, it is necessary to study
the wi@ar series 111 ucl~ more closely.
The singuk series may be written in the form
8=1 +A,+A,+...+Ap+*.e .
Frr particular values of k arid s, the values of the most important
terms of the seriesmay be calculated mithou t excessive labour,
and upper boutldn assigtted to ttle remainder. The series may
thus be expressed as n sum of terms which oacike according
to simple l;tws, and show in a most illuminating manner the
source of the irregular v;lriations of Txt8(12). Suppose, f’or
example, that k- 4, s = 33. It is pIail; that the most irn-
* The numbers towards the end of the table need not be taken too striouely,
9s they have been obtained by writers anxious otlly to establish ~0~~2~ numericii
result. But a good deal of work haa been devoted to the earlier cam, and in
particular to the cases k = 3 md k = 4.

401
 portant term8 in the series are the fur which 1& Q1 is most
nearly equal to unity : these correspond to 4 = I 6, q’= 8, q =5
and certain associated values of p. When s is as large as 33,
all other terms of’ the series are quite small ; and we find that
rr a,,(11)=cnyl+ 1~054cos(3nP” l&p)
+ *I47 CO3 (&2n- - #T) + (*038) } + 0 (n++E),
where(9038) denotes a number whose rnodu.tus is less than
CM, and
+ = 2 x 32 = 2&l < 223 = 23,
It is easily verified that
1+ la54 cos (&nrr - &7T) + *I47 COS (:?27T - &7T) > ‘1029
80 that 9’ 4&) > 0 for all sufficiently large values of n. It
fuliows th& G(4) -< 33, or in other words that eve)-y large nwmber
is the sum of at most 33 fourth powers,
It may be hoown similarly that
G (6) 3< 193.
In this case it will be found that aZEthe term3 in the singular
series, except the firat, are quite srm~ll: a fact which in its&’
suggests very forcibly that the result, although a considel*able
improvement on what is known, is a long way fi*om the
ultimate truth.

21, Our general proof of (ELI) does not involve any

difficultg of principle, or novelty of idea, otiter ttlan those
which we have explained in Q§3-17. But the completion
of the proof demands a considerable number of’ simple alge-
braical calculations of which it would be difficult to give
a short and intelligible account. We must therefore confine
ourselves for the moment to :I few gene& remarks.
We supposethat s = P%+I; and wllat we have to prove
is hat there is a number 6, independent of 12, suctl that
1X1,6.
It would no doubt be possible to prove this, for any particular
value of k, by an argument proceeding on the lines of Q19.
The general proof, however, is very much simplified by a
preliminary transformation ot’ our series into sin infinite
product.
It is easy to prove that

402
aatkfies tile equatim
whenever q and q’ are prime to one another. From this it
follows tt1at

X =1+dw+&z+d~+r..
The problem of inding a lower bound for S is thue reduced
to that of finding lower bounds for each of the factors xW.
The discussion of x= depends upon the arithmetical relations
between a and k, the nnost troublesome factors being those
f’or which w is a divisor of k or is of the form p&c + I. Suppose,
to fix our ideas, that k is prime. Then the factors in question
correspond to P = k and w = nzk + 1. The other factors are
easily shown to be unimportant.
If P is of the form rnh~+ I, and n is not divisible by a,
Am2, A+, 1.1 are found to be zero, and xlm reducee to the
simple tbrm
xw =l+Aw.
All tl~at is necessary, then, is to find an appropriate upper
bound tbr IA,!, and this C~UIbe done if we know such a bound
f’or 1At& 1. Now it is easy to prove, first that
j sp,w j 5 (k - 1) da,
alld secondly that

Using one or other of these two inequ:klities, the first if a > 4’

itnd ihe second if IT<&, we obtain ir; either &ae an inequality
i’or 1A,1 which proves sufficient for our purpose. If n is
divkible ‘by m, the argument is slightly more Gnplicated, but
in principle the 13alne. Tht: factor xk, corresponding to a=k,
mav be dealt with in a similar manner.
‘i’l~e cokderation ok’ composite k’s does not introduce any
fresh difficulties of a sel-ionsn&ure; but it is hardly possible, in
such a sketch a~ this, to give a coherent account of the various
complications which ’ a& ; atld we must pass thcln by until
we he able to publish ourI proof in full. lhr object in this
paper has been Lerely to gi;e a general account of&r method,
Ml enough to make clear to the reader the fundamental ideas
on which it rests, and the ~ature of the difficulties wlkh arise
and the analysis which is necessary to overcome them.

403
 CORRECTIONS

There is a correction to (10.4) in 1920, 5 (footnote on p. 45).

In the formula for C, above (5.1) on p. 276, the denominator
should read l+/k). Similarly at the top of p* 289 and in the
asymptotic formula for q&Q on p. 290.
 33

Some problems of ‘Partitio Numeronnd ;

I: A new solution of Waring’s Problem.
van
6. H. Hardy in Oxford und tl. E, Littlewood in Cambridge.
Vorgelegt van E, L a TI da u in der Sitzung vom 30, Januar 1920.

Introduction.
1. The present memoir is the first of a series, in which we
propose to develop in detail the new analytic method which we
have found for, the discussion of Waring’s Problem and a number
of allied problems in kdditiver Zablentheorie’. The general lines
of our method, in so far as it concerns Waring’s Problem in parti-
cular, were explained in a recent paper l) in the Quaded Jorm2uZ
of’ Afatlmmtics, This paper contains also full references to the
literature connected with the problem. Our object here is to give
full details of the proofs, up to the p&t at which Hilbert’s famous
theorem, first proved in this journal*) in 1909, emerges as a corol-
lary from our analysis.

Notation and terminology.

2, 1. The following notation will be adhered to throughout
the memoir: -
k, s, +I; p, q, and 112 are positive integers or zero, k > 2, s > 0,
C
O-=P<4r IP, 9) = 1, p>O if q>l;

1920, 5 (with 6. E. Littlewood) Nachrichten won der K. Oesellschaf t

der Wtisenschaften zu Gbttingen, Math. -phys. Klasse, 1920, 33-54. 405
 34 G, H, Hardy and J. E, Littlewood,

1 1
1y = 2it-1I X = 1 s- 1 3 a=-, be----
K lc R-l, 7

X
e(x) = e2nix, e q0 =e-,
( Q1

We denote by r&) the coefficient of x” in (f(x))“t so that

It is evident that rk,# (12)> r,,,(n) if s > s’. If k is even, rk, 8(rr) is
the number of representations of B in the form

where nj is an integer, positive, negative, or mm, anarepresen-

tations which differ in the sign or order of the numbers fij
are reckoned as distinct. If k; is odd, the arithmetical inter-
pretation of 9.k, &) is not quite so simple, W’e have in fact, if
fz > 0,

where Qtr (n) is th e number of representations of n by exactly p

positive k-th powers.
8. 2. It is plain that, if n > 1,
1 If CxY ax1
(2.21) 5 ,SW = m --Yq--
s X
the contour of integration
being the circle r whose centre is the
1
origin and whose radius is R = 1 - -*
%a
Wa divide r into arcs & 41 which we call Farey arcs, as fob
lows. We fdrm the Far@ keries of order N = [d4], the first

406
 Some problems qf Tartitio Numerorum’. 35

0 1
and last terms being i and 1. We suppose that -tP where
¶
q > I, is a term of the series, and 2; and c the adjacent terms
Q 4
on the right and left; and we denote by &rl the interval
I I
P P
--a

4 4 (a +a”> ’ Q + 4(Q + 47 l

The intervals ,iOI t and jl 9I are defined to be ( 0 ’ I&i) ana

1
1 - =, 1 respectively. The intervals &, ‘I just fill up the
(
interval (0, l), and the length
of each of the parts into which
1 I
A4(7 is divided by pG! is not than - and less than -
1~ 3).
2!0 0
If now ths intervals jP , c2 are considered as intervals of variation
0
of -> where f) = arg x, and the two extreme ‘intervals amal-
2x
gamated into one, we obtain the required dissection of r ihto
ar@s &7Q’
Wle call $v a nzt#Ir arc if q ( a l&NW arc if 98’ < 4 ( n? tiaj

The am of r complementary to & we de&e by qP,q.

In considering any particular Farey arc &rlf we shall write

X = Xxp,,l = 1Ye 4 = xe,(u>,

so that X is positive when x lies on the radins vector to the point
xf+‘I’ fina 0; kp,q we have
X = IX 7
where

We shall in general use E to denote a fixed, but arbitrary, posi-

tive number, and R to denote a positive number depending upon 7~
only. Different E’S or A’s will, when necessary, but not always,
be distinguished by suffixes. The symbols 0 and u will be ased
in the manner which is now classical. It is always to be under-
stood, however, that the inequalities or asymptotic relations implied

3) For formal proofs of these well known propositions see G. H. El ar d y

and S. R a m an u j a n , Aspptoti~ ~bmu~cte &+acontbi~twy adysis, Proc, London
Math. SOL, ser. 2, vol. 17 (199, pp. 75-115 ($5 4,21-4.22).

407
 36 G. H. Hardy and J. E, Littlewood,

‘in the use of these symbols are, unless the contrary is stated ‘ex-
plicitly, satisfied uniformly in all parameters other than k and
(when it occurs) E. Thus, if q is defined for ~2 = 2, 3, 4, . . .,

means that to every positive E corresponds an H = H(k, E) such

that I~~<HP?‘+~ for 92 = 2, 3, 4, . . . .

Statement of the mtin theorem.

3. Hilb er t ’ s theorem may be stated as follows.
Theorem A, To evwy k corresponds n +g = g(k) smfi tlmt
r,,,(12)20 for szg and n> 1.
We shall prove
Theorem B. TO WW~ k correspods u G, _I Q, (k) szt?chtld 4,

for s > G,.

Itwill be shown that this series is absolutely convergent for
s > s, = s, (k), and that its sum - which is real, since the terms
(p, q) and (q -p, q) are conjugate - is positive.
From this theorem we shall deduce, first
Theorem C. To every k cowesponds a G = G (1;) ami! c~ti
% = no(k) secchthat rk &2) > 0 for s 2 G, n 2 rz,;
and then Hilberks theorem.
We shall use CJ,G,, G to denote the teas1 numbers for which
these assertions are true.

The singular series.

4, We call the series

s = c (+qeJ- 9223)
PIcl
the silayul~r series. The use of this series, which is the dominatilg
idea of our work, is suggested by the following considerations.
If we write x = XxFIq, we have
P---I
f( X >=-I+2 5 xvke,&9 x--1+2 c e&k”y) 5 X(24+Ir’“*
V=O h =I) I =I,

4) f(n) TV g(n) means that -f + 1 when N + 00.

g

408
 Some problems of ‘Partitio Numerorum’. 37

If now X + 1 by positive values, so that x teds to xpIq along

a radius, then

E c)(l) + ru + 4 log f i(l

P ( )

f(x) ru 2r(l + a)% log & Oa;

!l ( 1
it being understood that, when SP, fl = 0, this formula means
1 -ta

We have also
( ),
f( 5> = 0 log x

where F(X) is regular for X = 1”). Thus

(f(x))” w C(-rFs (X) = FP,rl,

k

Our fd~pdal idea is that of approximating to the coeffi-

cients of f” by a sum formed from those of the various approxi-
mating f-unctions Fp, ‘i. We therefore replace X by x eq(- p),
3, (X) by its expansion as a power series in X, pick oat the
T

5) This well known proposition may be proved as follows. The function

is regular for y = 0 (the integral being uniformly convergent in the neighbour-

hood of y = 0). And if y > 0 myehave

O”-t-”dt = rp - G)r (G)

y-“.
s0 t+Y
‘f’hesc wptiom establish the truth of the proposition for 0 < G < 1. It is
trivial for G = 1; and it may at once he extended to other values of G > 0 by
il i f ferentiation. Incidentally we see (from the integral expression of x ti”-l e-9
- F,, (e-g)) that I;*lr (x) is regular at all points of the unit circle except z = 1.

409
 38 G, H. Hardy and Y. E. Littlewood,

coefficient of xn, and sum with respect to p and p. We thus

obtain

as the function by means of which we are hoping to approximate

to %a (fl>.

Properties of the singular series.

5, 1, Lemma 1. lf /?,y, . . . are integers then

In accordance with our use of the symbol 0, this equation

holds uniformly in fi, y, . . . .
Our proof of this fuClamenta1 lemma is based on certain
transformations due in substance to Weyl”>. Eor the sake of
formal simplicity we consider first the case k = 4. We have
4 =
ITI 4 ~~e*(p(hl-z*)+B(K-z3)+y(h)--z*)+6(h-I)).
9

the summations with respect to h an&Z extending over a complete

set of residues to modulus p. Writing h = I + Ii, ~0 &t&n
a- -
R e,(Qph.,P+ B, I’+ Ch I+ 4) t
ITI 4 1
t 1 1 I
where 33, , Cl& , . . . are quadratic, cubic, a . polynomials l in 1~.
19
with inte&aI Loefficietits, whose precise form is irrelevant to the
argument, Writing h again in place of 2, we have
e =
c (1) e, 4p h, ?T + Bh h’ + Ch ‘I),
lz’l 4
h1 F ( 1 1
where (1) denotes a number whose modulus does not exceed unity.
Applying the familiar inequality of C an c h y and S c h w ar Z,
we obtain
/T,I’zSqC ~e,(4ph$+Bh h’+C h ’
1
h h h1 )I
1”)+ B, 1(h’- I’)+ ch1(h- 1)).

6) H. We y l,, iiber die Gle&mteklung 210~ Zu?&n mod. E&w, Mathem

Annalen, vol. 77 (1916), pp. 313-352.

410
 Some problems of ‘Partitio Numerorumy. 39

Writing 18 = 2 + h,, and repeatixlg the argnment,

where B, h isr a polynomial in h, and Ih, with integral cr>effi-

13 a
dents. Applying the C au c h y - S c h w a r z ineqnality again, and
replacing I once more by h, we have

Finally, writing ?a = 2 + Ir,, we have

It is plain that the argument used here is perfectly general;

ad the resulting inequality in the general case is

5. 2. If (k!, q) = 6, we write k ! = 6k,, Q = a&, SO that

&P, Q) = 1.
The summation with respect to lo,, 7$*,. . ,, h,_, is defi_lled by
0 =f
< h. < q(l5.j s A:- 1), SO that H< gk-** And evidently

where x* is defined by

The number of terms in C’ is O(qkwn), ad in each of them

c - qe Thus
12I -
(5.22) 5r? = 0 (q-).
In C”, C is mro unless H E 0 (mod. Q), when it is go The con-
I

7) The index of the power of q may be at once verified by induction: for

2(2k-l-7Q+k-l = 2”-7c-1.

411
 40 G. El. Hardy and J. E. Littlewood,

gmence is satisfied for at -most F = O(qk-‘) values of H; and

therefore*) for at most 0 (q ’ v2 ’ “) sets (II.,, hp, I l ‘, ;h,_,). Thus

From (5.21), (5. 22), and (5.23) the lemma follows immediately,
As obvious con&&s we have

The term for which y = 0, Q= 1, is 1; write then S = 1+ S’.

We have
-- 1 4-E
SP*Y = -(l--%)+E K
& I= oq( 17
Q

1
where p = -4K l Also

Thus

%!+I3
<-,- 12
V
< v’tY+- p-2
1
5f = l+S’q,

if only s is sufficiently large.

8) See E. Landau, if&v die A~~721 der Gitterpunkie ha gewisom Be-

reiches, GGttinger Nachrichten, 1913, pp. 687-771 (p. 717).

412
 Some problems of ‘Partitio Nunvmrum’. 41

Behaviour of f(x) on a minor arc.

We deduce this from

Lemma 4. If q and t WC wbitrwy fixed positive mmbm, and

We &note generally by E,, Ed, ..J functions of q and c which

have the property stated.
We begin by performing on UP a series of transformations
agalogolrs to those of $ 5.1. In this case, however; we may at
once quote the find result, a~ it is given explicitly by Wey lg).
We have

where H = h, h, . . . l+,,1, and the summation with respect to

h,, h 27 ..I t
h
B-1 is d&d by

ad that with rwpect to Z iS over an unbroken sequence of not

more than p+ 1 values.
We replace this inequdity by
(6.11) 1U/f ( A/Pek c’ j 2 / + AyKek C” 1C / = u’ + r,
t 2 2
2’ and x” being defined as iri 5 5.2. AJI~ we have, in the first
place,
(6.12) Ti’ = 0 (#-k pk-Pp) = 0 cpK-l) z 0 (12(8_ ‘1 a + ‘I).
6, 2. In discussing U” we observe that
F (p+l (k!pHsO (mdq)),
I I

9) We y 1, Zoc. cit., p. 330.

413
 42 G. H. Hardy and J. IS. Littlewood,

Suppose now that k! p H = ;1 (md. q), where 0 5 A c q, and write

-w
ci = UC+ Vi, where il = 0 in Ul and A > 0 in Ui* If I = 0
we must have (in the notation of 5 5.2) H G 0 (mod. Q); and this
k-r
caa happen for at most 2 !-- = values of H (for none
Q
if pkwl(r Q). To ia& of these correspond at most 0 h@- ‘) “)
or 0 (p”“) sets (hi, Aa, . l l &-I)’l Hence

Since y < nQ + ‘, q > j2" -‘, we have .!!- = 0 (la”l) and so

‘1

Next, suppose that A > 0, and that

k!pH E ik!pH z ;1 (md. q),
where H' = hi hi . . . I&-,. Then H'- HE 0 (mod. Q), or H’ = m$+H,
where IP~ is an integer. Since 1Hj a’nd 1H’I are each less than
P k-1, this can happen for at most

2 --yk-’ + 1
Q
different values of IT, ad so for at most
7c- 1 + E#J
0 Max
( ( P
2
tP
E7
)I

4 -1 in 1-t EH
B= z 7 cosec- Q = O(plogq) = oh 1 l

;1=1

The first term gives

,p-w+r,) .
4
In discussing the second term we may suppose yka' c 8; and
we obtain
$2(X-k)f l--Ct+E~~ (K--)a+~,
4 1= 4 1,

414
 Some problems of Tartitio Numerorum’. 43

since q < ?a1- a+ ? Thus U;l and U;, and therefore U”, are of
the form 0 (titK- ‘)*$- ‘)* SO, by (6* 12), is U’ ; and Lemma 4
now follows from (6* 11).
6, 3. III order to dedtice Lemma 3, we write
t( ‘x > x - 1 + 2 2 x u’i+2 c x Vk = -l+f;(x)+f&q,
OSVSOJ
-- v>m

where ~3 = H-i
I$2 I and 5 is an arbitary positive number. Then
I

415
 44 G. H. Ifardy and J. E. Littlewood,

Behaviour of f(k) on & major WC.

1. 1. We now suppose that x lies on a major e~rc, We write

It should be observed that, on a major arc, 1 Yj cannot exceed a

fixed upper bound (and is in general small); for g12 =< 1 ad

Ig”el = 0(-&J = o(g- = O(1).

(7.12) g e,emY(‘+jJk = 2 g 8,$-e- Y"'icos2mz(2~-j)du,

Z--O m =o 0

10) This formula may be proved, by classical methods, in the following way.
The function
. - Y(2 +j)”
fb> = c El+j e ?
Z---W
where E, = 0 for z < 0, has the period I ; and it may be verified at once that
it has a derivative f(j) bounded for 0 < j < 1 and that f (0) = 8 (f( -0) + f(+ 0)).

Hence it is expressible in a Fourier’s series 2 2 E,,$yrn, where

1 ??I=o
1
Ym =
,S 0
cos2nm(u--j)f(?d~du =
I-
cos 2~127~(2 + u -j) ~t+~ e--Y(~+u)kdu

00
O” e-YzP cos2m?r:(u -j) du.
u- e- rZCk cus 29~ (u -j) d@=
&G
s -00 s0

416
 Borne problems of ‘Partitio Numerornm’.

Substituting fronn (7.12) and (7.14) into (7.13), ad changing the

order of summation, we obtain the result of the lemma. It may
be written 11)

a log 3
when Y4 has its principal value, vk e- , the imaginary
part of log Y lyjing between - +a ad +a.

Asymptotic formulae for v(m) and x(m),

Lemma 6. lf x lies on a major arc, so that f Yi is not

exp(-,(YzckT2m~i21))a21 = +-$&+o(lY~~-~-~)
-$(1-b) -+ 7
+ 4 m I YI 4 t

E = exp (- A d+” 1Y I4 cos #J), $J = arg 1.

It is enough to consider the integral in which the ambiguous
sign is negative ;; the other assertion fallows by changing @ intu

so that
I- W d--2mniu) = - Z(v”- kiv).

11) In our paper in the Quarterly Journal tbe term in s;I, (I,m was omitted
in error in the case’ when k is odd: 84, q,m = 0 when k is even. The omistsion
i8ects nune of the results of the paper.
 46 G. H. Hardy and J. E. Littlewood,

When Y is real (and positive) we have

The right hand side is an analytic function of 2 when 8 (2) B 0 ;

the equation (7.21) therefore ‘holds for 8 (Y) > 0,
Let

where c and g are positive. We can choose A, SO that lu,l ( 1

and O<a( z z when c < A,. We shall ultitiately choose c to
be a number depending only on k and satisfying this ct>ndition.
It may be observed, in order to avoid any possible misuder-
stmding, that (x, and. all geometrical elements of the figure ad-
joined, are functions 0f k only.
Since larg 21 = [ - bl~, 1~ i ab, the path of integration may
be changed in the right haGI side of (7.21) from the real axis to

(1)+ (2)+ (9, when (1) is the straight line from 0 to v,, (2) the
straight line from vI to et,, and (3) the continuation of this last
line to infinity j2)* We write then

(7.22) exp (- Z(v” - k i-v>) C&J= I, + I9 + I..*

Srn
0
Consider first II. We have, when v lies on (l),

418
 Some problems of ?artitio Numerorum’,

fR(-Z(v"-7cia)) = [z(~(x;'"-'COS(l~cr-~)-kcos(a-~))
= 121 S((LP cos 76cc- k cos a) cos !F+ (SP sin ka - Tcsin a) sin P).

- ;-nib
When v lies on (2), we have v = v, - y e , where 0 E<y (c, and

16-. i-sib - 16 -+-7cibpj 3

- Z( ck- h-i 9 = -2 ((v+7eiv,) + 0
2 e vf2 y2 0
3 e () y ++- r
the term in y vanishing since vt+ = A This is
- 4p a~-kiv,)-AIZIexp((~n.(l-b)-qnb-~~)i).ye(l+O(y)),
Now

(7.24) ~(-~(zj’-Iriu))I~(-Z(,~~-7civ,))-AIZIy2(1+O(Y))
2 %(- z(v; -kiv3)-A1Z(y2
for 0 ~5y c A,. We choose c = +Min (A,, AJ; (7.24) then holds
when v is on (a), and

419
 48 43. H. Hardy and J. E, Littlewood,

FMly, on (3) we have 2; = V, + ye- @“, where y 2 0, and

- p+bzi 21;“‘ < 4 lr

Iarg (Ze )1
for k 2 3, 3 ( r <Tt k. Hence

!R (- Z(vk - ki v)) ( !H (- Z(t$ - k i vJ - A Ze- i bni vt-” yq).

Thus

(7.26) Is = e- i-bxi O” exp (- z(v”- J;iv))dy

s0

= 0 lexp(- Z(+-k~vq))J OOe--AfzlY”dy

( s0
G O(IZ-+ exp (- Z(vt- kiv,))~),

as before. Now
- z( v; - kiu,) = (k-l)iziv, = -(k-l)lZ]e-ip,
so that
(7.27) a(-Z(v,“-kiiv,)) = -A~Z~cos~
From (7,25), (7.26) and (7.27) it follows that

(7.28) Ia+ Ia = o(lz(-* a*(-Al21 cos F));

and from (7.21), (7.22), (7.23) and (7.28) that

SW0
exp(-(YUA-2maik))dN = 2 “z’ (1, + I, + 13)

Substituting for 2 in terms of Y, ad observing that the ratio

cos ?R Cos * remains superior to a positive bound for all values of
@ between - + g and 4 n, we obtain the result of the lemma.
7. 3. Copbining the results of Lemma 6, we see that if x
lies on a major arc then

420
 Borne problems OF ‘Partitio Nnmerorum’. 49

(7.31) Qp(nz) = 0 (tn- *(I - b, 1Y I- + b E) + 0 (112-k-l1.y I),

(7.32) 1 (at) = & + XI (4

(7.33) x,(m) = 0 (nz ii1 - b, f Y I- “%) + O(?P* 1YI).

Behaviour of f(z) on a major arc (cot2tinued).

8, 1. We have, from (7.15) ad (7.32),
f(x) = %*q+ $,‘I7
when

(8.W !Pp,‘L = ‘zr(l+a)&,,Y” =

and
@P,fI

By (7.31) and (7.33), this is

if q = 1, the sum has the value 0. Hence:

(8. 12) aa = 2 5 .,‘, s;,,,+ O(1) = 0 *x+J nag 1 g+ O(l)

g n2=1 i
= 0 (qx + ‘)*
Clearly dso
(8.13) @t= 0 (*I + ‘>*

421
 50 G. )I. Hardy and 3. E, Littlewood,

8, 2. There remains @I to be consiqered. We have

so that

Hence
(8.21) E = 0 (exp (- A dfb q-l-” d (cos 7~l)++~)),
and it follows from the defini.tion of a1 that

We have 0 =< 8” < A q-9 n-4+5”, and SO

YS VP 1
cos’q = -- >
v2+ t)* VP+ A ¶-= 9Pqa ’ 1 + A q-’ da
1
2 (1 + A) q-2 ,2a = A$ figa’,

!l-labn”(cos q)‘+’ > A q+’ n” (q PZ-~)‘+~= A.

Hence, as 1PI > Aq”0, we.have

From (8.12), (8.13) and (8.22) we have

O(*%+~)+O(n~B*%+E--~,%b) = &1”rS’)
(8. 23) @p,‘f =
since 9’ ( ti.
8, 3. We may summarize our conclusions, as regards the be-
haviour of f(z), as follows: WC Ihave

Proof of Theorem B.
9. 1. Returning now to the integral (2.21), we have
1 If or dx
r,,(w9 = ‘23ti s JT-(f(2))” dx = 2 1
x- se
2si P* rj --F- p

422
 Some problems of ‘Partitio Numerorum’. 51

the summation extending over all arcs gP,(! of the Parey dissection
of order N = [d-“]. We write this in the form

(9.11)

= s,+ s,,
Z!R and m denoting typical major and minor arcs of the dissection.
It follows at once from Lemma 3 that
sux+r I4
(9.12) s e= o( n 1 ).
As regards S, we observe that, on an arc D?,

where

= 0 (Qp,
y)+ 0 (6p,,)Y
say. It is plain that
sax+&
(9. 14) x @p,9 = 4 ?a 1I
Also
9 = O(pJfY~-") = o(*x+E(qkIv_iHI)-u)
--1+~(l)s+F)P)-q*
= d4
From this equation and (8.23) we deduce
dF)
e Pt ‘1 =oti ccx+E*(S-1)(X-l)+E O”

( s -oo +2+ ,*>iw- 1)

1
nx+as-u-11-t (S-1)(X-1)-f-E
C n
4 Q )I
00 q-1
provided s a k + 1. NOW the series 2 C qCs- ” (‘I- ‘) + ’ is
4 =lp=O
convergent if (s - I) (1 - x> > 2 0r s > 2 K + 1, and supposing, as

14) This’ formula is not true uniformly in s, and may therefore be held to
violate our general principle (5 2.2) as to the use of 0 and O. But the formula
falls into line with the general principle so soon as we fix our attention on any
particular s = s(k); and this is sufficient for our argument.

423
 62 G. H. Hardy and J. E. Littlewood,

we shall do henceforth, that this condition is satisfied, we have

= 4 Inax+as--a-l+8
m P#!l 115).
t
and SO, by (9.13) and (9.14),
su%+&)+ Q(,8aX+“8-a--l+E)t
EJ P*!l= 4 ‘yb
(9.15) Y&2)=
-i+s I*
sl+s* = ~I,.,+o(ra”““+‘)+o(~aX~+as-a
9. 2. We now perform two transformations on IP,g.
(i) The functions

and
(2 J-0 + w sp,* 8 g #+f x9 - c IS$,n s
F P,‘l = 4 cm
rw ( ¶ 1n= 1 4 1

differ by a function regular at X = 1 and GCfortiori bounded

near x = 1. Further, it is plaia that

(2I’(l+ cc))”(log -$)* - CFJX) = O(1)

uniformly in p and p. Hence, if we replace vi,,*, in &, by FP,IP

we intSIuce an error

o(q+y) = o(~p-“(‘-X)+e) = O(l),

provided only s(1 - X) > 2, s > 2X, a condition which we have

already assumed to be sati&& This error is plainly trivial in
comparison with the error terms already present in (9.15).
(ii) We next replace the arc &, in , q, by the complete Ip
circle C The error thus introduced is

15) It w3li be useful later (though irrelevant for the purposes of this me-
moir) to point out thsbt this formula is s-till valid when 8 = 2 K+ 1, since then
Q-1 n”
(S-l)(%-l)+E
5 c Q = 0 2 q- l+$ = O@~).
!I =lp=‘() ( p-1 1

The inter& of this lies in the fact that 2 X + 1 = 9 when k = 3.

424
 Some problems I of ‘Partitio Numerorum’. 53 ’

Bat, if X = Bei’, we have, uniformly ifi p kl g,

- -isa
F,m = 0 (11 - xI-“a, = 0
(C (1-B)P+4Rsin’~U
1
1 1

= O((L&+q-y.

If b. is the angular codibe of one end of gPt y, nF), > A $ > A.

Hence our integral is the sum of two, each of which is of the form

Hence the total error inhducd is

() *(SU-1)(1-a) c ps”-l-s(l-x)+”

( PI 4 >
= o n(su-l)(L--a) c psa-s(17X)+~ = o nmx+2a-l+8

( q ( 12” 1 ( 1
g~sa” -f- E
= 4 11

9. 3. There is now no difficulty in completing the proof. In

the first place
1 Ksx) (fx = da+ e, (- n p>,
2ni s r xnel

SO that the first term on the right had side of (9.21) is

The series may be regarded as limited by g ( VP, or as extended

to infinity, indifferently; for the difference is
0 ga-1 c qe80+1fe = o(n4z-~(12a)-s(1 --X)+2-t-g
1
( q > ?P 1

sm+2u-1++
= 4 n ) = o(nsax+$

425
 54 G. H. Har dy and J. E. Li ttl ew ood, Waring’s problem.

If s is sufficiently large, S > !, by Lemma 2, so that the

leading term in (9.21) is effecti& of order dn? The third
term is always df lower order, since m - a -e 0. Finally, the
sec0nd is 0f 10~4~ der if
sax<su-1, sn(1 -x) > 1,
or
s > kK = k2'-'.
10. We have thus proved Theorem B. It is plain that Theo-
rem C is an immediate corollary. To deduce Theorem A we have
only to take g = Max (G, ?Q, since any number less than tid, is
the sum of at most 32, positive 7+th powers.
We have not, in this memoir, determined an explicit upper
bound for any of the numbers yt G,, G. The whole trend of our
analysis is, however, to suggest that the necessary value of G(7c)
does not exceed
k2"-'+ 1;
SO that, for example, every large number is the sum of at most
33 biqu&ates, 81 fifth powers, UP 193 sixth powers II;), It is, in
fact, possible to prove more than this; but the proofs involve a
detailed examination of the singular series which we must post-
pone to our later memoirs. Our second memoir, which will be
concerned particularly with the case k = 4 (a case in some ways
the most delicate and interesting of all), and in which (going
beyond the results which we have indicated) we shall prove that
every large is t32eS211’12 of’ ot 2wOst21 bigzcadrates, will, we
hope, be published shortly in the Matheenatische Zeitschrift.

16) This result is now fur 7; = 4 and X:2 6, but not for k = 5. ‘The best
result we can prove for k = 5 is G (5) 5 53, which is new.

426
Some problems of ,,Partitio numerorum~~: II. Proof that every
large number is the sum of at most 21 biqwdrak.
BY
G. H. Hardy in Oxford and J. E. Littlewood in Cambridge.

1, Intlroduction,
1. I . This memoir is essentially a sequel to one which we published
recently in the Gijttinger Nachrichter?). It could not in any case be
intelligible to a reader unacquainted with our earlier memoir; and we
shall therefore quote formulae from the latter without further explanation.
In the memoir referred to we laid the foundations of our new method
for the solution of Waring’s Problem, carrying our analysis just so far aq
was necessary for the proof of Hilbert’s Theorem, the fundamental exis-
tence theorem for the numbers g(k) and G(k). Here our object is to
find the best possible inequality for the particular 0number G (4). A good
deal of our analysis, however, is valid for a general Ic, and will be useful
to us when we proceed to the corresponding general problem. It will be
found that the special interest of the case k = 4 is quite sufficient to
justify its consideration in a separate memoir,
1.2, It is known that
19<3(4)237, 16 5 Q(4) 2 37,
these inequalities, from left to right, being due to Waring, Wieferich,
Kempner, and Fief erich respectively. For detailed references ee may
refer to the dissertations of Kempner”) and of Baer3). We need men-

1) G. H. Hardy and J.E.Littlewood, Some problems of ‘Partitio numerorum’;

I: A new solution of Waring’s Problem, Gijttinger Nachrichten 1920, S. 33 -554. We
shall refer to this memoir as W. P.
“) A. J. Kempner, Uber das Waringsche Problem und einige Verallgemeine;
rungen, Inaugural-Dissertation, Gijttingen 1912.
“) W. S. Baer, Beit#r&ge zur~ Waringschen Problem, Inaugural- Dissertation,
GSttingen 1913.

1921, 1 (with J. E. Littlewood) Mathemutische Zeitschrift, 9, 14- 27. 427

G. H. Hardy and J. E, Littlewood. Some problems of ,Partitio numerorum”. 1.5

tion only that the deepest ksult, viz. g (4) < 37, was obtained in 1909 by
Wieferioh, whose analysis is a refinement upon that by which Landau,
in 1907, had proved that g (4) 2 38, Here we shall prove nothing con-
cerning g (4); but we shall improve the upper bound for G ( 4) very
notably, by proving
Theorem A: a(4) 5 21.

2, A sharpening of our earlier analysis.

2.1. In 8 9.2. of W. P. we proved that, assuming always
sr2K+1=2E+1,
(2.11) rE.&) = &pJS + O(n8an-t~) + 0 (nsa+ax-a-t+&)
= @&B)+ O(n8ax+E) + O(,~a+@----+E),
where

It will be necessary now to replace the term 0 (n8aXfE) by a terti of

lower order 4).
2.2. It will be found, on an examination of the analysis of W, P.,
that the critical error term 0 (,~a~+~) arises in two places only. All
other errors are of lower order than that of the dominant factor naa-l,
either independently of the value of 8, or at any rate whena s *z 2 K + 1.
The two critical errors arise as follows.
In the first place we have

where m is a typical minor arc of the Farey diaeection.

Secondly, when we consider the corresponding sum connected with
the major arcs, we are confronted by a sum 2 ap, q, where

and !D2 is a typical major arc of the dissection, and we write

nsax+E
2 $4 = O( >.

It will be observed that these two errors arise in exactly the Same
way. The upper bounds are obtained by substituting in the integrals the
crude approximations, f = 0 ( nax+&) on a minor arc and @ = 0 ( naXfE)

4) The formula (2. 11) would lead only to @ (4) <- 33, in itself a new reeult,
16 G. H. Hardy and J. E. Littlewood.

on a major arc. Here we refine on our previous argument by the use

of a single new idea. This idea consists in an appropriate use of a known
result, viz. that the number of positive integral solutions of the equation
Xk + yk = n ie 0( ne) for every k > 1 6), or, as we may express it in
our notation,
rk,2 (n) = 0 (n$
2.3. We have

Hence

Now

1
and so, since R=l---,

IIt
Hence

4 = 0 (~~!f(z)!“if(~)!“-“Id2.1)
m
E ~~s-4~ an-t-e ZJi f(x) I* ldx I) = O(n(8-4~n*+ZQ+~).
O(
7tt
Again, we have

6) For a formal proof of this result see II. Cauer, Neue Anwendungen der
Pfeifferschen Methode zur Abschiitzung zahlentheoretischer Funktionen, Inaugural-
Disaerlation, Gijttingen 1914, S. 38. ‘For k = 2 (when the result includes a fortiori
the corresponding results for 4, 6, . . .) see E: Landau, Ober die Anzahl der Gitter-
punkte in gewisaen Bereichen, Gijttinger Nachrichten, 1912, s. 750.

429
 17

Q P

(k = 3).

2.4. The argument of W. P. showed that

r&n) \ =CWa-IS+ O(d),

where i -< sa - 1 if sax < su - 1, i. e. if 8 > EK = k2? It is now
clear tJhat this result holda if only

(s- 4) ax+2a<s&-1,
i. e. if
s > (k- a)K + 4.
For k = 4 these inequalitles reduce to s > 32 and s > ‘LO respectively,
so that the improvement is very substantial. And if only’ we can estab-
lish the existence of & positive constant CJsuch that

when s > ‘21, we ghall have proved not only Theorem A but the more
precise theorem
Theorem B: ~~,~(n) - Cn!8-3S
(8 2 21).
The proof of this theorem is in fact reduced to a discussion of the
singular series S.

430
 18 G. II. Hardy and J. E. Littlewood.

3. Factorization o’f the singular series.

3.1. The discussion of the singular series is greatly simplified by
the following fundamental lemma.
Lemma I. If

where 3 is prirrd) and

We have

where h and h’ describe conlplete systems of residues to ‘moduli q

and q’. But
+

where Q = h q’ + Tc’q ; and, since (q, q’) = 1, f~ describes a complete

system of residues to modulus qq’. Hence

(3.11)
Next we observe that, if z, describes a complete system of residues
pime to modulus pl, and # a similar system for modulus q’, then
p = pq’ + p’q describes a similar system for modulus q q’. Also

S
Hence

:= S pqrk-G )” (Spgk--l,q’)g e (- ;;,)

P

which proves the lemma.

-- -.-. -
*) The symbol 3t’ is used in this sense down to the end of 5. 2, after which it
is used in the ordinary sense.

431
 Some Problem of ,,Partitio nume~rum”. 19

4, Rules for the calculation of .A=, l

4.1. Lemma 1 i8 true for tiny value of E. The lemma8 wbch follow
are also true generally, provided anly that k is not divisible by z Thus
when k = 4 they hold for n > 2.
The sum An, involves the argument PS, and we might write
A 47Y= A;r, (n). When, as will Bornetimes happen, n is rqlsced by
another argument, this argument will .be shown explicitly..
Lemma 2. If (z, k) = 1, a > 0, 0 < p < E, then
A ,akfp = 0

acceding & n i3 not 0~ is a: multiple of nut.

(1) We have first

._
We write
h E ,pk+,-Q + h’ (O&<n, 0,<ht<nuk+4u-1),
and we obtain

The sum with respect to x vani& unlessh’ ia divisible by n, i. e. ?L = z hi,

where 0 2 h, < JP~+~-~; in this casethe sum is JE. Observing that this
range of variation of hl iti, to modulus +-l)b+p, equivalent to nE-9 de-
scriptions of the range 0 < h, < n(a-l)k+p, we obtain

It should be observed that the preceding argument i8 valid even whe’n

,U= 0. We obtain in fact
(4.11)
and otherwise
(4.12) Sp,nak+p = na(k-l) SP,nCm
(2) We have now

432
20 G. H, Hardy and J. E. Littlewood.

We write
13= mu2 + (pt CP > o>,
where 0 2 a < n uk snd $ ia leea fhsn z jr and not divisible by 7~. We
have then, by (4.12),

A aktsp =gz--
4

The sum with respect to z vanishes unlessn is divisible by JP% If however

n-nab, where v is an integer, we have

4.2. Lemma 3. If (JT, k)I = 1, l-1> 0,

A ,ak -0 - (n+O (modnakB1)),
A nakz-n a(k-8)-l (n ~0 (modnuk-I), n_)rO (mod d)j,
A ,ak= (n- II ’ &L (k--&-1 (n -, 0 (modnak)].
By (4. ll), we have
s p/k &k-l) 1 A ,uk JT--q-g,
= =
P
Writing
13=nx+pl (O(X<nak-l, O<p’<3+
we obtain
A iTak=n-q-~~-~)*
P’ 2
The sum with respect to x is zero unless in is a multiple of z*k-? If
n = @--I 9,, we have

The last Bum is - 1 or n - 1, according as Y is nut or is divisible by Z,

This proves the lemma.
4.3. Lemma 4. If (n,k)=l, lip<k,
= 0
A ,-c/b (ns;zO (modzp-I)),

( I) In the equation

433
 Some problems of ,,Pmtitio numerorud. 21

we write

and we obtain

The sum with respect to z vanishes unlessh’ _I 0 (mod n) , or unlessh’ = nFY,)

where 0 < hl < JTP-Q. In this casethe exponential is unity (since ~1< k ),
and we obtain
S E Jpf+1.
p,--rp

(a> We have thus

A I-T+
II =Kqe(-g*
P
Writing
P =xz+p’ (o~x’<?r~“-l, O*<p’<n)
we obtain

and the sum with respect to x vanishes unless n = 7cp-l Y, where Y is 8n

integer. In this case

A ,zI’”--
--

and the sum is - P or z - I., according as Y is not or is divisible by n.

6. The form of 2, (k===4,n>2).

5.1. We now suppose E = 4, BO that all the results of 5 4 hold for

n>2. Taking first the case ,U= 0, we have

j A,ak 12 nafk-8),
by Lemma 3,
Next, if ,u = 1, we have ;AJ<n and so

IA ,ak+l 1 < natk-8)+1,

by Lemma 2.
Finally, if 1 < p < k, we have

by Lemma 4, and

by Lemma 2.

434
22 G. H. Hardy and J. E. Littlewood.

Thus the terms of x, may be exhibited in the form;)

1 + A, + p-q + [ma-q + l l ’ + [nk-l-]

+ tik-“(Cl] + [z] + CC”] + [n”-a]+. 1 + [2+1-q)

l

+ n-y[I] + [Jr] + [n-j + p-q +. l l + [3+-q)

+ .** ..* . . . . . . . . . . l . . . ..*.

where [x] denotes a number whose modulus is less than x. Hence (pro-
vided only 8 > k) we have

L =l+An+&,
where

Taking now k = 4,s > 20, we have

IB z 1 < n-17+27C-‘7(1+n+Jt-17) < 7r-16 (n-14,

(5.11) L = 1 +A,+[n-14].
5b 2. When we come to consider A,, it is necessary fo distinguish
different cases.
Suppose first that n is of the form 4m + 3, Then the residues of
h” to modulus n are the same as those of Fy‘, and SP. X reduces to an
ordinary Gaussian sum. Thus I SP,X 12 ~‘5 and

1A, 1 < nl+ < n-9,

& = 1+ [q-9] + [n-14] = 1+ [Jr*] (7c = 4m+ 3).

Next, suppose z of the form 4m + 1. Then”j

If z=== 17, j Al7 1 < 17(;)= < &,

(5.21) X17 = 1+ [&I + [17-l4]= I$- [i].

&?1
<n-3*3, ] A,: < n-~*s,

‘) It should be observed that, owing to the vanishing of Aldrlk and A,crk+!s

when rt does not satisfy certain congruence conditions, x,-t is in all cases a finite ‘series;
but this ia irrelevant for our argument.
“> See IS. Weber, Lehrbuch der Algebra, Bd. 1, S. 584. In Weher’s notation,
& is one of the numbers

435
 Some problems or ,Partitio numerorum5 23
and
(5.22) x,= 1 +[n -2’5] + [n-141 = 1 + [x2] (ZZ= 4m +- 12 29).
From Lemma 1, (5. ll), (5.21), and ([5.22), it follows that

8 =X3%53113 (1+ [81)11(1+ [J-l>g 0 + tn-“lb

rc=Pm+lpJ lz=4m+3
-

Thus irz order to estublish OUT conclusion when 8 = 21, it is only ne-
c~sary to show thut
lxa/>~>O~ 1X6+-d, lxls/>~~
5.3. We find by direct calculation that “)

It is however (as we have to consider the case of 13 also) more. convenient

to proceed &s follows. The numbers & are the roots of the equationlo)

(r”+15)a = 2o(r - l)“,

from which,

X5 = 1+ [*291] + [5-l*] = 1+ [13],

[&1>‘7=a.
5.4. Similarly the various values of XP 13 are the roots of
l
*

(1‘”+ 3q3 = 52([ - s>“,

from which

Irl”,< 39+ 6V’13 < 60.7,

p, J=lCl -c 7%
1A,, 1‘( 12 (06)~’ < ,002,
so that
--- ----._--- IX13/>fl*
9) From this point onwards JT is wed in the ordinary sense.
lo) We her, 1. c., p. 584.

436
24 G. H. Hardy and J. E. Littlewood.

The pof of Tiiemems A and I3 is thw educed to a proof that

I&l >Orn
6. Discussion of x2.
6.1. The arguments of 0 3 fail when JZ= 2, and it is necessary to go
back fo the definitions of A,, A,, . . . The first step is to calculate the
sums spt,,r. We find by direct calculation that
spI ‘J= 0, sp , 4 = 2(1+e+i), s, 18 = 4(1+ d’“‘), SpI18== 8 (l+ dPni).
If Y&5,
sp,gy
=&($) h
(o 5 h < 2’)s
Writing

we obtain
s PA v=

and the sum with respect to z is zero unless k’ is even. Suppoaing

I
h = 2 h,, so that 0 <C IJ, < 2wT4, we obtain

Thus
Sp,++p = Zaa S#, ap
6.2. Observing that A, = 0, we write

~54, we have

] A,ru-tz + A,4a+a + A24a+4 1 < 28~2-l’~,

1B 1< 28 cy> < 2-l’ < l00025,

~9= 1 + A, + A, + A,, + [+00025].

6.3. Of the terms A,, A,, A,, the most important is the last. We have

Ala==
c (~~s~~lex~(~~-~~)=~~+~~+$+~,
p=l, 3, ,. l , 16

437
 Some problems of ,Partitio numerorumK, 25

where a, is given by p = 1, 15, 91a by p = 3,‘13, and so on. And

(cos$)~~= - (cos$~= l66535O + [3-W’],

%1 =- (1 l 3307 + [lo-*]) COB(“,x - 7) ;

(cos;+)21= - (COSMIC= l020736+ [lo-“],

2& =(*0415+ [lo-“]) cos(;+ g;

(cos$= - (coB~)21~ [5*10-a],

and so also for (cos ;)=. Thus

tills + ai = [2 40-J l

Similarly we may write

where ‘!$ is given by p = 1, 7, and so on: and

(cos;)~~= - (cosy-r’=.189636 + [lo-“],

I
%
1
Z==- cm3793 + [lo-“]) cos (7 + y).;

(cos
!gl= - jcosg2L [1(-p],
8; = [lo-‘].
Finally

Collecting our results, we may write

&=l- lm3307 cos ( g - y ) + ‘0415 COB(;g + F)

- l3793 cos ( y + y) + 3 [ ~OOOl] + [*0017],

and the total possible error is [ lOO2] L

6.4. We have now to verify that the sum of the first four terms of x0I
is in all cases greater than l 002 . It is easy to see that the least fa-
5n
vourable cases are those in which cos - - Fc has its greatest possible
( 16 >
value, viz, cos Fc , This happens when n z 2, 3 (mod 16). We have then

438
36 G. H. Hardy and J. E. Littlewood.

C 1 - 1’3051 +-- no345 + ‘3504 -+- 3 [ ‘OOOl] + [ -0021

= 0108 + [ loo231 > '0085 = o > 0.
It will easily be verified that, when n has any other residue to
modulus 16, the margin is much greater.

7. Conclusion.
‘7. I. We have now proved Theorem B when s = ‘Ll , and Theorem A
is an obvious corollary. It is not immediately obvious that, if Theorem B
is true for s = 21, it is also true for s > 21+ All our arguments are
valid for s > 21, except those of $5 6. 3 - 6, 4; but the numerical discussion
of these two paragraphs has, strictly, to be repeated for each value of s
in question, Our own calculations refer only to the cases s = 2 1, 31, 33,
in which we have, at various tities, been particularly interested. No
point of principle is involved, and the calculations in other cases may
be left to anyone who may be sufficiently interested in the matter to
make them 11),
It is evident that we may, with the help of the singular series, study
as closely as we wish the variations of ra S (n) as n assumes various
residues to modulus 16, It is clear, for ’ example, that the numbers
lGm+Z and 16m+3 are, to put i-f roughly, less readily expressible
by 21 biquadrates than any other numbers, and something like 200 times
less readily expressible than the numbers 16 m + 10 and 16 nz + 11.
There is no difficulty in applying the methods of this paper to the
proof that
(7.11) G(k)5 (k - 2)2’-‘+5
for any (part&Ear value of L, as for example 3, 5, 6 or 7. We find
thus that G(3) 2 9, C(5)<53, G(6) 2 133, and G(7) =< 325. The
first of these inequalities is not new Ia), and in fact Landau has proved
that G (3) 2 8: but the numbers 53, 133, 325 compare very favourably
with the 58, 478, 3806 at present known. The proof that (7. 11) is true
generally, however, presents certain algebraical difficulties, of complication
rather than of principle, and we must postpone it to a later memoir.
We have not indeed worked out this proof in detail, the analysis which
we possesscarrying us only so far as the less favourable inequality
G(k)lk2E-'+l
ndicated by our earlier researches.
11
) See however’ the following note of H,err Ost rowski.
Is> The accompanying asymptotic formula is of course new.

439
 Some problems of ,,Partitio numerorud. 27

We conclude with OIM final remark. It might well be supposed that

the proof of (7. 11) for (say) k: = 7 or 13 would be more difficult than
for k -. 4. This is not so; the proof for k = 4 ia, in essentials, more
delicate and critical than far any other value of k. The f&t is that
it is only fur k = 4 that our inequality expre~~ something near the
ultimate truth, It is known that Q (4) 2 16, and, the difference be-
tween 16 and 21 is comparatively small : this corresponds fo the facts
that the critkd factor of 0 6 nearly mnishes in the least favourable case,
and that there is a term in xs which is sometimes actually greater than
the leading term 1. When ‘k: is larger, our value ia much too high, and
the singular series tends (for such values of 8 M are contemplated in our
analysis) to be dominated completely by its leading term.

(Eingegangen im Januar 1920.)

440
 Some problems of ‘Pwtitio Numerorum’:
IV. The singular series in Waring’s Problem
and the value of the number G(k).
By
G. H. Hardy in Oxford snd J, E, Littlewood in Cambridge.

1. Introduction.
1, 1, In this memoir we continue the investigations initiated in two
earlier memoirs bearing a similar title, and complete the proof of a]ll the
assertions which they contain l). We Bhall assume throughout that tie
reader is acquainted with the notation and terminology of these memoirs.
l

The fundamental theorem of Hilber t”) wserta the existence of the

numbers g (k) and U (k). In our first memoir we proved that, if
1 K = 2’-l, 1
a=- ,k:’ X=1-ff’ s>2K+1,
then
(1.11) r,, 8 (9i) = C nea--l S + 0 (It’QX+E ),
wbre S is the ‘singtrtar pJeries’

(1. 12)

1) G, H, Hardy eniI LE. Lit&wood, Some problems of ‘Partitio Numermm’:

I. A new solution of Waring’s Problem, Gijttinger Nachrichten 1920, S. 33-M;
II. Proof that every large number ia the sum of at most 21 biqua&rates, Mathe-
mat&he Zeitschrift 9 (192 f), 8. X4-27.
The third memoir nf the series (Some problema of ‘Partitio Num’erorum’ : III. On
the expression of a number ae a sum of primes) will appear shortly in the A&a
Mathematioa. The problems considered in this memoir are of a somewhat different
character. We refer to these memoirn as P. N. 1, P. N, 2, P. X 3.
8, D. Hiibert, Beweis fiir die Darstellbarkeit der gwian Zahlen duroh ein?
feste Anzahl n-ter Potenz411, GGttinger Naohrichtsn 1909, S. 17-36: reprinted w&h
certin ohanges in Mathematisohe Annaten, 67 (1909), S. 251-300.

1922, 4 (with J. E. Littlewood) Matbwtische Zeitschr$t, 12, 161-88. Ml

 162 G. H. Hardy and J. E. Littlewood.

The sum of the series is positive, and indeed greater than k, if s is

sufficiently large; and pax < sa - 1 if s > kK. Thus
(1. 13) ?y, &a> - Cnsa-w,
as n-+m, for all large enough values of s, say for sZG,(k). It is
plain that Hilbert’s theorem follows as a corollary.
Important simplifications of our method have been effected by
Landau”) an,d Wey14). These improvements relate to our treatment of
the ‘major arcs’. In particular Weyl has shown that, if we are concerned
with an existence theorem only, so that it is not important tJo obtain
the best possible upper bound for G (k), the rather difficult analysis
which we used may be replaced by an argument of a much more elementary
character.
We proved nothing in this memoir about the values of G, (k j or,
G [k),
‘. though our analysis suggested very forcibly that
(1.14)
In order to prove this it is necessary to examine the singular series more
closely, and to prove that
(1. 15) S > ir = a(k, s> > 0
fur srs,. This would be sufficient; but in fact, as Herr 0s t ro w ski
has shown”), the truth of (1. 15) for s r s0 will involve
(1. 151) s > CJ= a(k) > 0

t the value of 0 being independent of s) for s > q,. Ina our second me-
moir, however, we effected an improvement in (1, I l), showing that
(1.16)
rfr, s @) = &pa---l S +- 0 ( n(R-44)flx+2a+E)

( a better result if only k > 2). If now we can prove that (1. 15), and
therefore ( 1. 151), is true for s > (k - 2) K + 4, we shall have proved that
(1. 17) G(k)~G,(k)~(k- 2)K+5.
This we proved before when k = 4, in some ways the most interesting
case. It is the general proof of (1. 17) that is our primary object now.

3) E. Landau, Zur Hardy - LittlewoodMhen Lijsung des WaringBchen

Problems, Gijttinger Nachrichten 1921, S. 8G-92.
4, H. Weyl, Bemerkung zur Hardy-Littlewoodschen L6sung des Waring-
s&en Problems, Gijttinger Nachrichten 1922.
6j A. Ostrowski, Bemerkung zur Hardy - Littlewoodachen Liisung dea
Waring&en Problems, Mathematiache Zeitschrift, 9 (1921), S. 28-34. We ret’um
to thig point in $ 6. 3.
 Some problems of ‘Partitio Numerorum’. IV. 163

Our principal the&em then, will be :

Theorem 1. There is a positive number G = o ( k, a) such that S -y 6
for &(k- 2)&j-5, so that
rk, Jn> - CrP-l S
fur all such dues of s. In purticulur, rk s (n) ia. positive for all such
vulue~ of s and all sufficiently large v&e9 of n, so that
G(k’)s(k--2)K+5.
I
1. 2. We have in any event to undertake a detailed examination of
the singular series; and we shall push our analysis a good deal further
than is necessary for our immediate purpose. We do so primarily because
the analysis is interesting in itself. Rut it must be remembered also
that the inequality ( 1. 17) is, in all probability, far short of the actual
truth. It is not unlikely that the order of the error term in (1. N),
which is the obstacle to further progress at present, may before long be
materially reduced. The discussion of the singular series, for values of s
smaller than those contemplated in Theorem 1, will then become of
immediate importance, as every improvement in ( 1. IG) will give a
corresponding improvement in the value of G (k).
It may be useful if we summarise at this stage the existing state of
knowledge as regards the values of g(k) and G (k). This is exhibited in
the following table.
_~------ --- __-- - -
k= 2/3j4151617 18

4 9 I 37 58 ’ 478 ~ 35OG / 31355

b 1
I’ I
4 9 / 19 37 I 73 1 143 1 2’79
I I
4 8 37 ; 58 1 478 1 3806 31353
(5) I (9) 1 21 I 53 / 133 325 773
4 4 1 16 ’ 6 I 91 8 32

The numbers in the first line are the upper bounds for g(k) which have
been obt#ained by elementary arguments, and are due in order to
Lagrange, Wieferich, Wieferich, Baer, Baer, Wieferich, and
Kempner respectively’). Those in the third line are the corresponding
----u
6, The names tire those of the authors who found the actual numbers quoted.
The proofs of ‘Waring’s Theorem ’ for the cases in question arc due to Lagrange,
Maillet, Liouville, Maillet, Fleck, wieferich, and Hurwitz (and Maillet)
respect ively. For detailed references 8ee A. J. Kempner, Uber das Waring&e
Problem und einige Verallgemeinerungen. Inaugural-Dkertation, Gijttingen 191’1,
and W. S. Baer, Beitriige zum Waringschen Problem, Inaugural - Dissertation,
 , H. Hardy Littlewood.
16-i G, and J. E.

upper bounds for G (k), and are identical with the numbers in the first
except when k = 3. The inequality G(3) 2 8 is due to Landau’).
The fourth line contains the upper bounds given by Theorem 1. It
will be observed that the numbers for E = 2 and k = 3 are inferior to
those already known, but that there is a very substantial improvement
for all larger values of k.
The second and fifth lines contain the best known lower bounds fur
g\ kj and G @) respectively. It was observed by Euler, and later by
Bretschneid&), that the number 2Q - 1, where 2 is determined by
3E =2”1+m, 0 < m < 2k,
requires I -+- P - 2 powers; and this observation gives the numbers
tabulated. The numbers 4, 9, 19 are mentioned by War i n g , but there
is nothing to show that he had recognised the general law”).
The numbers in the fifth line are more interesting and require further
explanation. It was proved by HurwitO) and Maillet?) that
G(kj& k+ I
for every k; and in some cases, e. g. for k = 3, 5 and 7, no more than
this is known.
In other cases it is possible to prove a good deal more by the
consideration of simple congruence relations. The simplest case is k = 4.
Every biquadrate is congruent to 0 or to 1 to modulus 16, so that a number
1Gm + 15 requires at least 15 biquadrates. Thus (as was ob;served
by Landau) G(k) 15, and Kempner, considering numbers 16”. 31,
--_.l__l__-
Giittingen 1913. The numbera for k = 7 and k = 8 could no doubt be substlantially
reduced.
Proofa of the existence of g (k), from which tin upper bound for g (k) could be
calculated, have also been given for k= 10 (I. Schur), k = 12 (Kempner) and
k = 14 (Kempner).
7) E. Landau, Uber eine Anwendung dar Primzahltheorie auf das Waringsche
Problem in der elementaren Zah’lentheorie, Mathematische Annalen, 66 (1909)?
s. 102-105.
“) See Kempner, lot. ok, S. U-45.
9) Waring asserts quite explicitly, not merely -that g ( k) exi&, but that
g (2) = 4, 9 (3) = 9, 9 (4) = 19, %t sic deinoepa’. Nothing is known, so far as we are
aware, inconsistent with the view that the numbers in question are the actual values
of g (k) for every k.
I0> A. H u rw it z, Faber die Darstekng der ganzen Zahlen als Summen von
n-ter Potenzen gamer Zahlen, Mathematische Annalen, 65 (1908), S. 424-427.
lL> E. M ail1 et, Sur la dkompoeition d’un entier en une somme de puissances
huit i&mee d’entiera , I3ulletin de la soci&b, mathkmatique de France, 86 [1908),
p. 69--77,

444
 Some problcme of Partitio Nunlerorum’. IV. 16't 3

where @ is kge, improved this inequality to G (4) 1G. He also proved

that G(k)> 4k whenever k is a power of 2. and that 6(6)&g* This
is the ori& of the remaining number8 in ou; table. Again, every ninth
power is congruent to 0, 1, or - 1, to modulus 27, 80 that B number
27 m. + 13 requires at least 13 ninth powers: thus G (9) > 13.
Considerations of this character concerning cubes lead only to the
Hurwitz-Maillet inequality; and when k = 5 or k = 7 the resulting
inequalities are entirely trivial, for any residue to modulus 25 can be
generated by 3 fifth powers, and any residue to modulus 49 by 4 seventh
powers. It will be found that these simple facts have a very ‘interesting
bearing on the structure of our singular series.

2. The formal theory of the singular series.

2. 1. The singular series is absolutely convergent for sufficientFly
large values of 8 ,12) and. is then expressible as an infinite product
(2.11) 8=1 +A,+A,+ l .‘*. =CAQ=&Xs& .., =- r-x,,

where z is a prime and

(,2. 12) 7un =1 -+ A3 + AZ, + . . . =~AJ3)
The sum 31, is a finite sum, for Ann is always zero from a cert,ain value
of iL onwar&14).
The question of the absolute convergence .of the series and product will
be discussed more precisely later. Our immediate object is to determine
the form of the factors xX,
2. 2. We suppose that g = d(;C 2 I), and we denote by ,v(& q, n)
the number of solutions of the congrence

for which
(2.23ij
We write
(2.22) y(q, q, n) = M (q, n> = 2w(q)m’“J
Pinally, we denote by
i 2.23) Won) = ml)
12) P. K 1, S. 40.
13) P. N. 2, S. 18.
I’) P. N, 2, S. 22 (f. n. 7). This will also appear incidentally later (S. 374).
Is) When it ia unnrccsmy to show explicitly the dependence of 3-f on N.
166 G, H. Hardy and J. Es Littlewood.

the number of solutions of (2.21) far which 0 -< xr < q (r g 8>, and for
which not every x is divisible by Z. Such a solution we may call a pi-
milive solution.
IFollowing Landau, we write ”y 1z’ and ’ y $- z’ for cx is divisible
by y ’ and 6x is not divisible by y ‘* We shall find it convenient, moreover,
to have a special notation ta express ‘y’ 1z, yf+l f z’, i. e, ‘y’ is the
highest power of y that divides x’l In these circumstances we shall
write ‘y’J z’,
This being so, the value of xlr is given, in terms of the numbers N,
by the following theorem.
Theorem 2. suppose thut
(2.24) k:> 2, n81k (02 O), (JWI~ (B 2 O??)
and let

(2.2611) B-O (B=O)7

The proof of this theorem rests on a series of lemmas.

2.3. Lemma 1. If
ne i-k (0 2 o), q = 22, r.>8+1 (n>a-j, /i>0+2 (n=2),
(2.31) x E 5 + anj.-8-l,

(2.32) (mod q)*

We have

The terms r = Of1 are those which occur in (2.32).

Suppose then r 2 2. The index of the highest power of it that .
divides r ! is

Id) (n k )‘@1n meane, of course, 3~Bk 1 b JzrB+l)k + n. It a meaning is different from

that of spk 1 n, which would mean fl BEi,, d-+-n.

 Some probkms of ‘Partitio Numerorum?. IV. 167
Hence the r- th term ia divisible by JP, where

c ‘> 0 - z~+Y(rE-O--- l)-d+x~;T+(T- l)(A+-6-2)-z.

If n>2, c-R>;T-22-11. If n=z, nzO+3, and so

c- i;>r--1--2,-l. C IF either ca8e c --A>-I,. arc--kxl. E

U. Lemma 2: ~AxA(f2)- 7~~+-~W(n~~,n~. <’

m
Writing, a8 u~ual,~YJ== &, we have

A, == A,(n) qy) e,(---- np)

P
Q-1
=p-“z 2 e,(p(x:+x,f+ .,. -+x~-n))=q~s~c,(X),
p s~,x,,...,xg=O x1. xp,. l ., x0

where X=$+x:+ . . . +x,“- n and c,(X) is Ramanujan’s sum17)

c,(X) =~c(pX).
P
If A.=l,
P Z n, c,T(x)=-l (x-+X), c,(X)=n-1 (nix),
and
(2Aj A, = 3 -“(C(-l)fCn)=n-Xi-n”+nM(n))=iz’-*M(n)- 1.
x UTIx
If i-d,
c,~(X) / = 0 (d-1 + X), c .z&Xj = -29-l (+11X),

If now

we have

and so

17) See FL Ramanujan, On cerbain trigonometrical sums and their applications

in the theory of numbers, Transactions of the Cambridge Philosophical Society, ‘231
t 1918), pp. 259-276; G. I-I. Hardy, Note on Ramaaujan’a function cp (n), Proceedings
of the Cambridge Philosophical Society, 20 (1921), pp. 263-271; and P. N. 3.

447
168 G. H. Hardy and J, E. Littlewood.

The lemma follows fpom @, 41) and (2.42). As corollaries we have

Lemma 3: ~~2 0.
L em ma 4 : If n is representable in any manner as a sum of s
( positive, negative, or zero) k-th powers, then x3 > 0.
Lemma 3 ip an immediate consequence of Lemma 2. To prove
Lemma 4 we have only to observe that A a= 0 from a certain value
of 1 onward@), and thst, under the hypotheiis of the lemma, M (&j > 0
for every 1.
4>r,. Lemma 5. If ~761k (02 O), thert
4.
(2.51) N &i, m> = n(~t-‘I )(8--1)
N @I, m),

where Q! is defined m in Theurem 2, p 2 p7, and m is wbjtrary.

We may suppose p > v, and write
(2.52) xr = f&+ .rni’-e-l (0 2 [,< m-y 0 2 a,< nH+l)*
Let k=nfik*. Then (k&=1. Also
(2.53) X;G (,“+ ko(lrE,lk--lw~ (modnrl),
by Lemma 1. If now
(2.54)
the congruence
(2.55) \7 k--v
&xXr=m
s equivalent to the pair of congruences

(2.551)

(2.552)

In what follows we take into consideration only primitive solutions

of (2.55) and (2.551). I n such a solution of (2.551) some & say &, is
not &visible by z. This being so, the values of a,, a3, . . , in (2.552)
may be assigned arbitrarily, and then, since (k. F1-;, n) = I, the value
of a, will be determined uniquely to modulus Z. There will therefore
be ~8 possible values of a1 less than ++I, and no (~fl+l)@-l= z(~+~)x--I
seta of a’ 8 associated with every solution of (2.551). That is to say
we have.
(2.56) N @I, Ina>z ,(@+I)-1 Nl,

la) See S. 369, footnote I’).

 Some problems of bPartitio Xurnerururn’. IV. 169
c

where N (n.!‘, m) is the number of primitive solutions of (2.55) and IV1

the number of primitive solutions of (2.55i).
Again, N CM-~, m) is the number of primitive solutions of

(2.57;)
If here we write

and use Lemma 1 and the hypothesis p > Q?, we obtain

(2.58) N (njr--l, mj ==bJ-9 l = dWV1.

f$’fL,.
- I..,II8
From (2.56) and (2.58) we deduce
(2.59) N( n!l, m> = ng-1 N (JT!~-~, m )
and the lemma follows immediately.

Proof of Theorem 2.
2.61. Let I’ be the integer such that ~#k+~ 1n#,,so that 0 5 11c k; let
t 2.611) ;1,=Max(~k$-~+1,liSk+(1));
and suppose that 1) i,,.
We divide the &tions (primitive or imprimitive‘) of

into classesas follows . In t*he first class we put the primitive solutions,
NC R$93) in number; in the second class the solution B in which every x
is divisible by z but not every x by # ; in the third those in which
every x is divisible by 9 but not every x by n”; and so on,
The second class of solutions is correlated with the class of primitive
dutions of

and the number of primitive solutions of (2.133) is plainly xtk-Q R times

the number of similar solutions of

449
 170 G, II. Hardy and J. E. Littlewood.

or ia
,(k--l)o N &k, 2 .
>
Similarly the number of solutions of the (CI+ 1) - th class, where
44 ifi
(2.615)

There are no solutions of any higher class, since (,8 -/- 1) k: > pk + 3’+ 1
and jt@fgl+l f n, Hence, if 1> Bk -f- Y + 1, and so certainly if A 2 i,, .
we have

2. 62. Again, if R.- ak =

> 9, and so certainly if 1> A0 and u 2 p:
we have, by Lemma 5,
-ak
(2.621) d

Making this substitution in (2.616), and multipIying by zr.+8’, we

obtain

(2.622) &(1--d M @A, n)

a=0

If a < ,8 and 43< k, --$ is divisible by 7~v, and we may re-

place it in N by 0. If n>2, p=O+1~2@<;rt@~k:. If 7c=2:
9=8+252e<k unless 8= 0 or 0 = 1 S in which cases 93s 3 2 k.
Hence we may replace every N in (2.622), except that for whicll
a = /3, by 0.
It follows that the right hand side of (2.622), is equal, when 12 &,
to the value for xJ?Tgiven in Theorem 2. It is also independent of R?
and therefore, by Lemma 2, equal to

(2.623) lim J$(~--~~

H (&, n) = x,.
I,+=
This completes the proof of the theorem. We may observe that we
have shown incidentally that
(2.624) A +== 0 (1 3 &).

450
 Some problems of Tartitio Numerorum . IV. 171

3. Some properties of the sums SP,*.

In this section we establish certain properties of the Gaussian

which will be useful for the further study of the singular seriesIi)). We
have not attempted to make the theory complete, though we have deve-
loped it a little further than is absolutely necessary.
We denote by
x = x, = xx(m) (12 x =< h = 43(q),l
the h Dir i c h 1et’s ,character# to modulus q. “O) x1 is the principal character,
and X, is the character conjugate to x,. We shall be concerned only with
the c80e u -= 79, where z > 2 and As Y 1.

It w;ll be convenient to write

q--l
(3.12) &,=&oe,~~k~
j=O

It is plain that, if R 2 k;,

I n
(3.13) sp,q=&,*+A’1 = s;,q+2-l.
edj
3.2. Lemma 6. If (1,q) = 1 then,

Cx,Vh,(m) = 0
x
unless m s I (mod q), in which case the sum is h.
The result is obvious if (m, qj > 1. If (m, q) = 1, we determine xo
tirn the congruence mm’= 1 (mid q). We have the n

YJ) x&Q = z,(l) zx.mt) = XxPo

and
Cz,(lm’) = 0
K

unless Zm’ II 1 or m :: 2, in which case the sum is h.

IQ) What we do is, in effect, to develop from our own point of vkw certain
portions of the theory of the division of the circle (Kreisteilung) It is not unlikely
l

that the substance of our analysis is to be found elsewhere; but it is not altogether
eany to extract, from the classical accounts of the theory, the particular parts which
we require.
fO) A systematic account of the thmry will be found in Landau ‘a H~&tcck, 1
(Zweites Buch) .

451
 172 G. H, Hardy arid J, E. Littlewood.

We write
(2.21) 6 = (h, kj = (p(q), k;) = (+I(, - l), k).
3.3. Lemma 7. There are just 0 ch.aructers 31, which pumess
the property
(3.31) x,” = Xl l

These characters are given by

(3.32) E e t+ x,wI
( -iv >
where p = 0, 1, 2, :, ., d - 1 and z is the index of 1.
We have generally

where y is the index which specifies 3~.‘I) The necessary and sufficient
condition for (3.31) is that kyx E 0 (mod h) for every 2, or that
.- (mod h).
(3.34) kY= 0
From (3.34) we deduce
(3.35) kY -0 [mod$.
-F=
which has the single solution y E 0 to modulus a’
” Thus (S.35) has the rl,
solutions
-- ioh
---. __
y -- fi (e-o, l,..., 0-l)

to modulus h. These are all solutions of (3. 34), and are plainly the
only solutions.
We shall call the characters x’==- x,, which satisfy (s. 31 j the special .
characters. It, is clear that zx, is a special character.
Lemma 8. We have

Lemma 9. Suppuse that q = d (A > 1), und that k 1 fl: - 1, 80

that is= k. mThen
(3.37) Ze,(Zp) = 0,

“I) Landau, S. 401-402.

452
-- Some problems of ‘Partitio Nunmrorud. IV. 173

if = I and the summath

(P?P) is extended over those residues 1 of q
for which 6 1 x,
We denote by
G=g+mn

the primitive root [mod a> to which the indices refer, g being a primitive
root (modn)?)
Suppose first that B = k = 71:- 1. Then the indices of the E’s in
question are
0 1 n-l, Z(Jz-l), .m., (nl-l-l)(n-1).
Suppose that x, and x, are any two of these &--l indices, x, > z,,
and I, and I, the corresponding values of 1. Then

where ,U is an integer, and

GPh- 1 s g’“s - 1 s 0 (modn).

Hence I,--- E,E 0 (mod+ On the other ;and, Zx and I, are incon-
gruent to modulus q, since pd = x,- 2, < zA-l (JE- 1) and G is a pri-
mitive root fur Q. It follows that the Z’s in question are the numbers
of the arithmetical progression

so that

The lemma is therefore proved when S = 7t - I. The extension to

the general case is immediate,, The indices of the Z’s in question are
1lOW
0, ci, 26, r*., n-l, l ... 7+(7z-l)- CE
lir -- 1
and form B arithmetical progressions of the type

A, A+n-I ,..., A-+-(+‘-l)(n-I),

where A is one of 0, 6, 26, , . , z - 1 A 8, The Z’s corresponding to
l

tyhe indices contained in any one of these progressionsform an arithmetical

pregression of difference n, and the sum of the lemma splits up into
,“G
-- 1
8 sums which vanish individually.

32) Landau, Handbuch, S. 394.

453
 174 G. H. Hardy and J. E. Littlewood.

3.4. Lemma 10. We have

the summation with respect to X’ extending over all special characters.

We may plainly restrict 1 to values prime to q . If ( I, q ) = I,
cm, qJ = 1, we have, by Lemma 6,

Hence, if i runs through values less than and prime to q,

The sum with respect to j is zero unless x,, is special, when it is h:

whence the lemma.
Lemma II. If q = d (1 < 12 k) alnd 6 = (h, k), then
? I
(3.42) s P,Q,k = Sp,u,6 l

This is an immediate consequence of Lemma 10. For the right hand

side of (3.41) involves k only in so far as the special characters are fixed
by k, and is therefore unaltered when & is replaced by 6.
Lemma 12. If q- & (1 < A 2 k) and n+ k, then
(3.43) s &I 93
>
P,%k= l

It is plain from (3.13) that what we have to prove is

I
(3.44) S p,u,k = 0 Y

or, by Lemma II,

I
(3.45) S p, $8 =
0
n

By Lemmas 10 and 8, we have

where the last summation is restricted to values of 2 whose indices are

multiplea of 6 ; and this sum is zero, by Lemma 9 14)*

23> T& has been proved already, in a different manner, in P, N. 2, 8.19-21;

but it iH interesting to see how the result arises from our present point of view.
a4) Since S1n I I when z f k.

454
 Some problems of Tartitio Numerorum’. IV. 175

3.5. Lemma 13. If A=$, q-z, and 6=1, then

(3.51)
But if 6 > 1 then
(3.52)
where

and the summation with respect to 3~’ extends over the special characters f I
exclusive of the principal character xl. AIso
(3.54)
We may regard (3.52) as including ( 3. 5f), since its right hand side
disappears when S = 1.
We have, by (3.13) and (3.41),
S P* a k: -1
+ s;dM = 1 +Jp,(~P)x,o --t-~e,(lp)27X,4)~
I 2 x’
where the principal character ia now excluded from the summation with
respect to x’, and I runs from 0 to g - 1. The sum of the first two
terms is
1+c,(P)=~+P(JI)=~*
The third term is
Cx,~(1,)Ce,(zP)xxt(zP~8
x’ 2
Since 1~ runs through the residues of q when I does so, the inner sum
is z,~, whence the result of the lemma.
Finally, to prove (3.54), we have only to observe that, q being
prime, X, is primitive ( eigentZich)23), and
ir,i=V&

4, The behaviour of x72for large values of n.

4. 1. In this section we are concerned with large values of Z, and
may suppose z > k, so that 8 = 0, F = 1. The O’s which occur refer
to the passage of JZ to infinity; the constants which they imply depend
upon k and s, but not upon n.
We suppose that k E> 3.
Lemma 14. We have

455
 176 G. H. Hardy and J. E. Littlewood.

other than. the principal character.

This follows at once from (3.52).
Lemma 15. If $2 1, p=O, then
(4.12J &r = 1 + q&““)a
We suppose first that z+n, so that P= 0. Then
i4.13) x,=1+&
Here We replace A, by the right hand side of (4.11). Any product of
$s is “X and so, when we expand by the multinomial theorem and invert
the order of summation, we obtain

where T is a product of s t’s, x a product of 8 x’s , and the number of

terms in z is O(lj. The inner sum is O(J&) for every x and all
1
values of n in question 21), and so
A A _= O( Jp. (Vjq8* )q z qJ&-q,
which proves the lemma when z + n.
Next suppose z ) in, so that 0 < Y < k l In this case R, = Y + 1 snd

Now t$&. = nM for 299+19, by Lemma 12; and so

Ad = n-s&i. (- n p) = F8CAR (n),
P

This completes the proof of Lemma 15.

If n is fixed, 7t j- n from a certain value of YZonwards. Hence we
obtain
Theorem 3, The singulur series S = 2Aq9 and the product
p=Trx,, are absolutely convergent for 8 2 4, and S = P.

*3 It is - 1 if x is the principal character, and the producb of a x and a t if 1

ie non-principal (and so primitive: Landau, Hundbuch, S. 480).

456
 Some problems of ‘hrtitio NumerorUm’. IV. 177

4.2. Lemma 16. If s > 1 then

(4-21) i+o(n3-11)<~~~il+~~-~+,.,+~~~k-~))(l+o~~-~--)8)).
This is proved already if /3 = 0, and we may suppose ,6 > 0. From
Theorem 2 we have, on the one hand
( 4.22) x ,T(n) C
> d--dN(n, 0),
and on the other
f4.2.3) @> 5 (I+ nk--p + . . . -; ,(B-1)(k-8)) d-N (n, 0)
+ n/~(k-81+1--N(n, n’),

where n’= -E- Since neither 7~ nor n’ is divisible by zk, we have

,Pk’
+-8N(n, 0) = d-8N(n,n) = &j, n1-8N(iz,n’) = &d),

and each of these is, by Lemma 15, of the form 1+ O( &id). Thus
h.21) follows from (4.22) and (4.23 ).
,4s a corollary we have
Lemma 17. If s&k+2 then x,=1+0(+.

5. The numbers y,) F (k) .

5.1. Given k and n, and any positive integer rrz, there are two
possibilities. Either (i) there is a number
(5.11 j h.z = h(k,s,n) > 0
such that
(5.12) x .3!>h.7
=
for s 2 m and all values of n, or (ii) there is no such number. We
define
y, = y (k 4
as the least value of ‘YYL
for which (i> is true, and r(k) by
(5.13) w Z Max Y, .
;r
Further, we define
7,: = y’(k,n).
as the least value of m such that
(5.14) 31,> 0
for s 2 rn and all values of lip.
It is evident that rh 5 y, l

Lemma 18. If x, > 0 for all sufficientJy large values of 92, then
x, > 0 for all values of rz~.
 178 c:. H. Hardy and d. E. Littlewood.

11 proving this Lemma we leave out of account for the moment the
special case k = 4, n = 3. That the result is still true in this case will
appear incidentally later.
It is easy to see that, apart from the exceptional case, q < k:. Thus
if n>2, q=ti+1~2Qd5k.
If ?tI= 2, ek3, then cp=O-$-2<2’sk.
If z = 2, 0 = 0, k is odd and 43 = 2 c 3 2 k.
If ~=2, 0-1, then k is oddly even and. p==3<<sk,
If n=2, 0=2, then q~= 4~69, unless k-4.
Thus 43 < k in every case except that in which k = 4, IT = 2, when
cp-k
Nbw let
n=nP;‘,+n’ (0 < nt < e’).
If rt’ + 0 then /Y = 0 (since Q?< k) and so, by Theorem 2,
X.-rp) = n’~----) N (ny n) = nyf--) N (ny n’) = xx (7~‘).
But xn (n) > 0 for large values of m, and therefore xn (n’) > 0. It
follows that xJc > 0 for all values of n that are not divisible by JC.
Again, if (m, n) = 1, we have, by Theorem 2,
X+Trn) ===L
n~f~--s)N (no, 0),
since q7<k. The left hand side is positive if m is large, and , so
N(ny 0) > 0. Hence, whatever be the value of m (prime to 7~).
xx (nv m) > &I-~)N (STY, 0) > 0.
It follows that xrr i 0 also when n’ = 0, which proves the lemma,
5.2. Lemma 19. The necessary and sufficient condition that
(5.21) N(Ju, n) I;- 0,
for every n, is that s --
> ynd’ Further,
(5.22) u: = y;r
except when k = 4, n = 2, in which case
(5 23) Ya =16, 3’a‘-15.
Leaving aside the exceptional case, so that F < k, let s > 7:. Then
~,(nv)>O. But p-0 when n=zv (since q-&), and so=

Hence

If on the other hand n E/F 0 (mod ZT), then /3 = 0 (since v E< k: ).

Hence
&T = nr(-)N (n’p, n>

458
 Some problems of ‘Partitio Numerorum’. IV. 179

and

Thus 8 > 74 is a sufficient condition that (5.21) should hold for every ~2,
Next, suppose that (5.21) holds for s = 5, and every n. Then it
holda, a forfiori, for s > 8, and every n, and the N’s that occur in
Theorem 2 are both positive. Hence

It follows, first that s 2 rk is both necessary and sufficient ‘for (5,21),

and secondly that s 2 7; involves s > ys, ii e. that $ = y3.
If E=4, n==2, then 2’ = 16 .-Now x 4 is congruent to 0 or to 1
to modulus f 6, according as 5 is even or odd. It follows that N (16, n) > 0
for s 2 16 and every n.; that
N(lG, n> > 0 (1s.tn), N(16, O)= 0
when ~45; and that N(l6, 15j=N(16,0)=0 when8<15. Finallyc
it follows, from Theorem 2, that
x3 > h, (s 2 16), & > 0 (s = 15),
xl( I& 15) ~ 28i;-l~)+4(l-lirN(1T,, 15) = 2-11(fl+1) 8 = '2 i
( 15) 1 )

md
;ys(I#* 15) = 0 (s < 15).
-11 (@+I)
Since 2 +Owhenb-*oc,, these results embody (5. 23). Incidentally
we see that Lemma 18 is still true in the exceptional case.
5.3. Theorem 4: G(k) 2 r(k).
Leaving aside for the moment the exceptional case k = 4,
suppose that s 2 G(k). Then any sufficiently large n is the sum of
s k- th powers, so that xn > o for every 7t and all sufficiently large
values of 72. Hence, by Lemma 18, xx > 0 for every n and every n, so
that s&+ It follows that G (k) 2 yz for every X, which proves the
theorem, apart from the exceptional case. In this case y,,rl= I& and the
result is still true, since G (4) 2 16 a8),
__---_
37) lV(l6,15) = $I’ when S= 15, since each x may have any one of the values
1,3,5, “., 15.
as) The lower bound F for G is associated with the vanishing of the aiagular
series S for s< l’- 1, except when k = 4. When E = 4, r r= 16, and the series is
positire for s =--IS, but assumes arbitrarily small values for suitable values of rt.
It should be observed that our proof (see 3 5. 5 below) that

G(P( 7c--3))~yx=~q (z>2)

(Fortaetzung der Fubote 28 auf nficheter Seite )

459
 180 G. H. Hardy and J. E. Littlewood.

5.4. Lemma 20. Suplpose that ,t”lk, and that 9: is defined CIS
in Theorem 2. Further, suppse that

and
*z- 1
(5.43) d E---.-;
F
so that (3 - 1 and (k,, d) -1. Thelk
(5.44)

We write Q=- JI (1’. We must distinguish the cases ,YF> 2 and JT=..I 2,
ci) If z> 2, y= 8 + 1. We suppose that G is a primitive root
(mod’ Q). We divide the residues to modulus Q into classesas follows.
Consider first the residues n, prime to p, If zt is the index of in,, we have

WI; leas one 01’ other of the d values 0, 1, . . , cl -- 1, and e one or other
l

of the ‘vfo values 0, I, . . , ‘t/~~- 1. The d values of n, with a common

l

e we class together and call the numbers

Cd,” ( e =o, l,*..,v’o-lI);
the class of numbers cc: with a fixed e we call C,“.
Next, consider the residues ni for which YZ~1ni, where 0 < i < p3.
We have
ni II ,“EiNi,
-5
where the N,$ are the ) numbers less than and prime to &l-i .
As G is also a primitive to modulus write

is essentially the same as K em pne r ’ s proof (see pp. 45-46 of his Inaugural-Dkser-
tation ) that
G(29>2~=2?.
=
His proof too fails when k = 4, and he has to appeal to the structure of the particular
number 31.
aR> Wle write # (e) for E ule r ‘,B fun&ion usually denoted by T ( p), aa Q?is ueed
hero in a different sense.

460
 Some problems of Tartitio Numerorum’. IV. 181

where

.mi has again one or other of the values 0: 13 , . , d -- 1, and e cme or l

other of the values 031 f - “7 ‘v/i-- 1. The y+ new classes obtained In

this manner we denote by
c e2 (e-0, l,..., ~‘j-lj,
and a typical member of C% by C$
Finally, the single number 0 is the sole member ccl of a class C,Y.
The ‘otal number of classesinto which the residues are divided is
F-1
‘4’0 + l/J1 + l l . + y+1+ 1 = .-+ + 1 = cn = c.

We may denote the whole system of classes, in the order in which

they have been defined, by C,, C,, . . ., G,, . . ., C,, and a typical member
of C, by c<,.
.
The class bPOconsists of the residues of k- th powers of numbers
x prime to n. For
k c n0& :, ~,“e(“-‘) = &)‘I’o*
d

Also x G’ for some f (since (x, n) = I), and

so that xk is an CI*. Moreover we can choose t so that tk, has an arbi-

trary residue m, to modulus d, since (k,, d) = I, so that every tiO is an xk
Finally, to complete the properties of the classeswhich are imme-
diately relevant, c.1) I belongs to Co, (2) q+xr, where ti0 and u,. are any
members of Co and C, respectively, belongs to C,,, and (3) uOq, where
LI,. is a given member of C,,, can be identified with any member of C,,
by& choice of a,.
Of these properties (1) is obvious. To prove (2) we observe that, if

is an cx,,, since y+ 1ylO. Finally

and we can choose m, so that nim, + mi shall have an arbitrary residue

(mod d), since (z, d) = 1; hence ti,,,c~~
can be identified with any member of C, l

r
5. *I. To prove Lemma 20 it is enough, by Lemma 19, to show that

461
 182 G. Il. Hardy and J. El, Littlewood,

for 8 2 c and every in. And the necessary and sufficient condition for
(5. 51) is that every n should be congruent (mod JP) to the sum of at
most c numbers aO. If any CII, is the Burn of not more than c cl,‘s, then
so, by (2) and (3) of the I&st paragraph, ia every ti,,. In these circum-
stances we shall say that C, is representable, and what we have to prove
is that this is so for aI1 the c values of r,
Suppose that 1 2 c’ 2 c. Then there are at least cf differed dames
representable by not more thm c’ a0 ’ B. For, in the first place, this is
true when c’ = 1, Suppose that it is true for c’ = 8 < c but false for
c’ = F + I, and let C be a typical class representable by c Q’S, and C,.
a c. Then ar belongs to a c’, and therefore, since no new classes become
representable when G is changed to z + _-- 1, ti, + 1 belongs to a c.
Similarly czr+ I + I = ccr+ 2 belongs to a C, and, re;?eating the argument,
cuerv residue (mod Q) belongs to a c, which is a contradiction.
Taking c’ = c we see that c distinct classes, and therefore all residues
(mod ej, are representable by c tiO’s, which proves the lemma, when z > 2.
(ii) There remainsthe casen=2, inwhich v-0+2, e-d-l,
=
c ZF=@. In this case there is nothing to prove, for any residue
(mod Q) is representable by at most Q i’s,
A particularly interesting case is that in which d = 1, E = TG- 1.
In this case
k = nO(n - l)ko,
where k0 is prime to Z. Here
))x=<n~~=nO+’ (n>2), y2=<2F=2e+a ( 7cC 2) .
If n > 2, y, = 7c’p. For

so that 1 is the only a,. Hence N(n~,O).=O if S<M, and Y~&P,

by Lemma 1% In particular
p=71-k+1
if E-W- 1. Thug, y,=5 if k-4, y-=7 if k--L
If ?t = 2, k=2”&. Suppose first’ that 0 > 0. Then

X’90 -=l (mod 2’+‘),

and ao xk I I (mod 2 ‘7, Except when k -= 4 our argument above applies,
and we obtain
y2 = kp” = 20f” (0 > o)*
d
The result still holds when k: = 4, since then y9 = 1G= 2’.

462
 Some problems of ‘Partitio Xunworum’. IV. 183

The argument fails if 8 = 0 (so that E is odd). Here Q= 29 = 4 ;

- I is a k-ic residue (mod 4); and 0, I, CL,3 are all representable by
at most two of the numbera + I. Thus
yz = 2 E p+l (e = 0).

Y
*I. 6. In general it is possible to go a little further than in Lemma 20.
Lemma 21. Suppose that d, /d, where d, > 1. Then
( 5. tx‘i y, 2 Max(d,, c - l)m
Since d, 1n - 1, (5. til ) gives in particular
ye7~Max(n- l,c- 1
in all oasesand,
~~7~M.ax(k- 1, c- I
if 0 > 0.
To prove Lemma 21, suppose that 1 2 c’ 2 c, and let Y(c’) be the
number of classes, other than the classC, (containing the residue 0 only),
that are representable by not more than c’ q$. Then
(5. 62) v(c’+- I)2 Min@(c’) + 1, c - I).
For, if (5. 62) is false Y(C’ -$- 1) = P(c’j -;r: c -- I. Let c be a
typical class of the Y(c’) classes, and C,. a C’, Then, if fir belongs to C ,
al, + I must belong to a C or to C,, since no new classes, other thk
perhaps Cc, are representable by c’ + 1 c$s. If q.+1:10, a,+2
belongs to Co, and therefore to a c’. If ar + 1 belongs to a C, cxr+ 2
must belong to a 6 or to C,. Repeating the argument, we see that
every residue, other than 0, belongs to a c, which is a contradiction.
From (5. t;?) it follows that
v(c - I) 2 c - 1,
so that all residues, 0 perhaps excepted, are representable by at most
c .- I cI()b. It remains to consider the residue 0. Let d = and qd,
Then tf; + 1, since 7~~ < y(p) and G is a primitive root (mod e),

and 0 is representable by d f cto’s, which completes the proof of the

lemma.
Suppose in particular that d1 = d = 2, so that n; > 2 and

463’
 184 G. II. Hardy and J. E. Littlewood.

In this case the C+ are the two numbers -+ 1, and

But
1 1
c--l=2(ne+‘-lj=g(n’l.-l),
so that

Thus in this case also we can determine y, exactly.

5. 7. It is convenient to sum up our results concerning the cases
d = I and d = 2 in a separate lemma.
Lemma 22. If k- n+- Ilk,, where 7~> 2 and k0 is prime
to 7c, them
(5. 71) L = ,@+I,
If k = 29, where 0 > O and k0 is odd, then
(5. 72)
Ii k is odd, then ya=2.
If k=;-nt) 7T-- W o:, where JZ> 2 and k0 is prime to ST, then

(5.73) y~+z~+l - 1).

5. 8. We know that G(E) 2 r(k)= Max yn. Thus, when k is

given, every value of y, gives a lower bound for G(k). These, when
less than k + 2, add nothing to our knowledge of G (k), since G (k) is
always greater than k. There is therefoke a special interest in determining
as systematically as possible all casesin w,hich
Y,>k+lm
Lemma 23. We huve]
(5. 81) r,s+1
dess (CC)k==2’ (ti>i~), IF==~, when, y,=P’=4k,
(p) k=2’3 (0>0), ~=2, when y,=2’+‘=ak.
or (y) k- GE (0>0), where n>2 andeln-1.
In cases(a) and (p) (5. 81) is fa 1se; in case (y) it may be true or false.
We write k=#&k,, as in Lemma 20. If 8-0, 77> 2, then

YIt= CC =E+llk+l,
by Lemma 20. If 8= 0, z-- 2, then y. = 2 by Lemma 22. Thus we -
seed only consider cases in which 8 > 0 ._I

464
 Some problems of ‘Par’titio Xumerorum’. IV. 1X5

8uppose first z > 2. If k, > 1, we have

k+ 1.

Thus (5* 81) is true unless kO= 1, k = JE%, which is case (7~).
Next suppose ZF==~, k===~“k,. If LO> 3, we have

Thus (5. 81) is true unless k, = 1 or 3, cases (EL) and (p) .

The casein which k = 6 is interesting as falling under both (p) and (-3’j.
0
If n=3, k=3.2=--n(n - 1>) ~=-n- 1, d-..l,andy,-3-=!k And
;?2=i 9 x ,. 8.
In case (y), (5. 81) may be true or false. Thus it is true when
k =:3, n=- 3, for then yi3= 4. But it is false when k =---.6, n = 3.
5. 9. We must now collect OUFresults and state them as theorems
concerning I ‘( k ) . We shall say that k is exceptional if it has one of
the forms in <a), (p), or (7) of Lemma 23.
The or em 5. If k is not exceptional, then.
I’(kjh =< k --;- 1.
This is an immediate corollary of Lemma 2tS.
Theorem ti. If 8 ,:‘- 1 then I’(a”> r:-i 2’+‘.
Theorem i. If fj :> 1 then I’i2”3) 1_.z g’-+‘,
Theorem 8. r( 6) =-I 9.
These theorems follow from Lemma 23, when we observe that the
llumbers in question in each case exceed k + 1.
Theorem 9. If n >p2, 8 >- 0, then l“(ne(n - 1)‘)z--d+? This
equality holds also whm 8 = 0, pmuided that k = z - 1 is nut exceptionid.
The second part follows from Theorem 5 and Lemma 22. We mav
therefore suppose 8 > 0. We have already seen that y, = - n e+l, whi&
is greater t,han k $- 1. If JZ~is a prime other than z, y,, 2 k + 1 unless
,71=-=2, d (Z - I) =: 2’l, or n1 =-- 2, @in - 1) = 2” 3, or z 1 >> -2 I y
d+- ij==gG1, where E& - 1.
It is easy to see that the first and third alternatives are impossible,
and that the second can occur only when z = 3, 0 = 1, k = 6. In this
case the result has been proved already; in all other cases we have
Y.73L y, and r(k) =:=y, =: G+k
Theorem 10, If n:r3 2, 0-2 0, then
‘1 1
rt> pH(n --- 1)), f--z -&@+l
A
- 1).
 186 I:. H. Hardy and J. E. Littlewood.

Here yn = $(n*+l - 1), since d = 2. This is greater than k + 1 except

when z = 3, 0 ‘* I, k = 3, when the two numbers are equal. Moreover
i-nO(, - 1) cannot be equal to 2 01, 2 tfr 3, or zfl F, , where z1 += n, 8, ‘Y 0,
‘>
‘“1 1n1 - 1, Hence ynl <- yx and r(k) = yr
Theorem 11. If n>2 and k--z%, where ti> 0, E[F-- f, then
I”(k) 2 Max(y,, k + 1).
It may be verified at once that JXHc’ cannot be of any of the fornls
9, 2@13, np*s,, except when z~=3,8=l,e=2, k--6. In this case
r’(k) = y3 = 9. The result follows from Lemma 23.
Theorem 12. In all cases
r(kj < 4k.
The sigrl of equality occurs if and only if k :: 2 t+ (0
, 2 2 1.
Theorem 1~. Xn all caSea
I’(k) < (k - 8~2’~‘j-5.
This theorem, which is included in Theorem 12 except when k = 3,
iti inserted only because it is what we require for the proof of Theorem 1.
Our actual bounds for r(k) are much better.
When k = 3, I’(3) = 4 < 9 = 1.4 + 5.
It may help to elucidate the results which we have obtained if we shobv
in tabular form the actual values of r’(k) for a number of values of k.
k=s 4 5 6 7 8 9 10 11 12 13 14 15 16 17 W
T(k)-- 4 16 5 9 4 32 13 12 11 16 6 ‘14 13 64 6 27
k =f; 19 20 21 22 23 24 25 26 27 28 29 30 :‘31 32
r(k) y= 4 25 24 23 23 32 10 26 40 29 29 30 5 12H

The values of I’(k) f or k-&4,6, 8,9, 10, 12, 16, l&20, 21, 24,
27 and 28 are given by the actual theorems and lemmas which we have
proved; the determination of the remaining values demands further cal-
cuIations into which we cannot enter here.

6. The behaviour of the singular series when s ZZZ

> F(k).
tj.1. Theorem 15. Suppose that k > 2 and a, = Max( I’(k), 4). Tken
(611) S>O
for a 2 a, and all values of n.
By Lemma 16, we have
x;r> 1 II &-G

466
 cv. r
iSome problems of ‘Partitsio Numerorum . 18,

Hence there is a x0 = n&8) such that

and (ti.11) follows from (6.12) and (6.13).

It is plain that our main purpose is now accomplished; with Theo-
rems 13 and 15, the proof of Theorem 1 is completed.
It is of some interest also to obtain an upper bound for Ft.
Theorem 16. If s>
=I=k: +- 2 then

and t,he result follows immediately.

Theorem 1.7. If sZk>3, then
(6.22) S *;‘_W

where p-7= 1 unless nk i a, and then pG7

= 1 --/-@. It is plain that

where xa(n. As ci(n) = 0 (nr>, the theorem follows.

The interest of this theorem lies in the resulting equation

There is some reason for supposing that

(6.24) rk,,(n> = 0 (nF),
an equation from which very important consequenceswould follow. This
equation would cease to be plausible if (6.23) at any rate were not true.
6.3. In conclusion, we return for a moment the equations ( 1.15)
and il.151). As we remarked before, the equation (1.15) is sufficient for

467
188 G.H. Hardy and J.E. Littlewood. Some problems of Partitio Numerorum. IV.

our present purpose; I but it is interesting to bring the remark of 0 st r ow s k i

into relation with our analysis.
Suppose that
N (nq,,n) C1 1
for every n and for s = so. There is then a primitive solution of

for every n. Consider now the similar congruence in which s, is replaced

bY s > 9,. Of the x’s, the last s - so may then be selected arbitrarily,
and there will be at least one primitive solution of the ensuing congruence
in the first s,,. Hence

It follows that the inequalities which we have used, of the type

\
2 XT >
I ,Y4-8);
may be replaced by inequalities of the type

and our numbers h, = h (E, JZ,s> and 0 = 0 (k, s) by numbers of the type
hn E h(k,n,s,) = h(k,n), and a=o(k,s,)=o(k). It is however unne-
cessary to develop. this remark further at the moment.
We add, finally, that the number r(k) has a simple and interesting
arithmetical interpretation. In fact I’(k) is: the least ntimber m such
thut every arithmetical progression contuins an infinity of numbers which
are sums of m k - th powers.

(Eingegangen am 31. Oktober 1921.)

CORRECTIONS
In the table on p. 186 the value of I?( 30) should be 31; see 1928, 4 (footnote on
p. 540).
In various footnotes (e .g. 14 on p. 165) there are references to other pages of the
present paper. In these, the numbers are incorrect and the correct numbers can be
found by subtracting 204.
On p. 172, for (2.21) read (3.21).
On p, 175, in the formula above the heading of 5 4, for 71~read 7x’
On p. I76 and later, Lemmas 15, 16, 17 need correction: see 0 3 of 1925, 1.
On p* 188, the fmal statement needsmodification when k = 4 ; for this, see 1925, 1
(p. 7) anti 1928, 4.
Some problems of cPartitio Numerorum’ (VI): Further
researches in Waring% Problem’).
BY

G. H. Hardy in Oxford and J. E. Littlewood in Cambridge.

1, Introduction.
1. I. In this memoir we continue the researches of the first, second
and fourth memoirs of the series.
The memoir falls into three parts. In the first (§§ 2 - 6) we are
concerned with properties of (almost all’ positive integers n. We say
that almost all numbers n possessthe property P .if ‘the number of
numbers less than n, for which P is false, is o (n) when iy~is large. Thus
almost all numbers are composite.
We suppose throughout that k > 2. We denote by
w =r*(n)=r, 1Jn)
the coefficient of x1’ in
(1.11) (f(x)y= (1 +a $x-“)T
We write

I) G. H. Hardy and J. E. Littlawood, Some problems of “Partitio Nume-

Forum” : (I) A new solution of Waring’s Problem (Gijttinger Nachrichten, 1920,
33-54); (II) P roof that every Iarge number is the sum of at most 21 biquadrates
(Mathematische Zeitschrift 0 (1921), 14-27); (IV) The singular series in Waring’s
Problem, and the value of the number G (k) (ibidem 12 (19221, 161~188), We
refer to these memoirs as P. N. 1, P. N. 2, P. N. 4,
We shall allso have occasion to refer to the fifth memoir (P. N. 5), viz* ‘A
further contribution to the study of Goldbach’s Problem (Proc, London Math, Sot.
(2) 22 (1923), 46.--56). This memoir, though concerned with a different problem,
contains, in a different setting, several of the essential ideas of our present analysis.

1925, I (with J. E. Littlewood) Mathemattiche Zeitschrift, 23, 1-37. 469

 2 G. H. Hardy and J, E. Littlewood.

where
(1.14) q(n) = /q$eq(- np),
P
(1.15) e(x) = Pis, e, cx> =e;, i) S p#Q= Yeq(phR),
h=O
P = 0 for Q = 1, while otherwise p runs through the numbers less than
and prime to Q; and

Our argument involves a number of parameters, and of letters used

in conventional senses, so that our system of notation requires very
careful explanation.
Our principal variables are n, k, s, E, and 6. Of these, E is an arbi-
trary positive number; 6 is also positive, and both E and 6 are to be
thought of as small. The choice of 6, which has tIo be made so as to
satisfy the varying requirements of our analysis, is always subsequent to
that of k, s, and E.
We use the letters A, B, C, with or without suffixes, to denote
positive numbers whose value varies from one occurrence to another.
When no suffix is used, A is an absolute constant (such as 2); B = B ( k)
is a function of k only ; and C ;- C (k, s) a function of k and s.
Sometimes, however, A, B, C will depend on other parameters, in
which case these parameters will be indicated explicitly by suffixes. Thus
B w?= B(k, E), CE,s== C(k, s, E, 8).
In exponents, we use c instead of C. When c occurs not in an
exponent, our conventions do not apply, and all the variables on which c
depends are shown by suffixes.
The symbols 0, o refer to the passage of n to infinit,y. Thus
f = 0 iv), where f and 93 are functions of n (and in general of k, s or
other parameters also) means that / f/ < My. Here _M is a number of
one of the types A, B, C; and it is to be understood, unless the con-
trary is stated, that M involves all parameters whose occurence is possible
in the nature of the cme. If any parameter is missing, we say that the
result holds uniformly in that parameter.
Similarly f’ = 0 (9) means that 6 -+ u when n --, 00, or that, if q
is any positive number, then ifi<vv for n&t,(q,aJ,...), where
4 p, l
are some of the parameters
’ l
k, s, E, 6 ; and again it is to be
understood, unless trhe contrary is stated, that all possible parametxs
are involved.

470
 Some problems of ‘Partitio Numerorum’ (VI).

1. 2. We prove first
Theorem 1. I/
(1.21)
then
(1.22) &(m))q= &(m) - p(?#= O(TP8a-y.
m=1 m=1

From this we deduce

Theorem 2. If E =+4, then almost all numbers are sums of
($,k - 1) K + 3 non-negative k-th powers ; and in particular of 5 cubes,
id7 fifth powers, 67 sixth powers, 163 seventh powers, 387 eighth powers,
899 ninth powers, and 2051 tenth powers. The num#ber of non-represent-
able numbers is in all cases O(TP~>.
The result for cubes is considerably in advance of anything proved
before, the utmost that is known being that a finite proportion of num-
bers are sums of 7 cubes“)*
The case E = 4 is abnormal, since then (i Ic - 1) K + 3 = 11~ 15,
while the singular series (5 (n) vanishes, for all numbers of appropriately
chosen arithmetical progressions, for any value of s less than 15. In this
we prove
Theorem 3. Almost all numbers are sums uf 15 biquadrates.
In this theorem, 15 is the correct number, and cannot be replaced
by any smaller number. If G, (k) is the smallest value of s for which
almost all numbers are sums of s non-negative E-th powers, then

(1.23) G, (4) = 15,

while C, (3) is either 4 or 5, and G, (5) may be any number from 6 to

27 inclusive.
1.3. In the second part of the memoir (§ 7) we return to G (k),
The best upper bounds known for G i3), G (4), . . . , G (A$, . . . , are 8,
21, 53, 133, ...j (k-2)K+5, ..* .’ Our object is to improve these
numbers (apart from G (3), with which we can do nothing) by combining
the results of the first part with a simple elementary idea, We prove
Theorem 4. I/ k > 3 then

2) That is to say, there is a constant A such that more than An numbers less
than rc are sums of 7 cubes. This result is due to Baer (W. S. Baer, Ober die
Zerlegung der ganzen Zahlen in sieben Kuben, Mathematische Annalen 74 (X913),
511-514).

471
 4 G. H. Hardy and J. E. Littlewood.

(k-ti)logZ-bgk+log(k-2)
(1.32) I = ----- _--. -_ --._ ~--- -__-
5‘k log?+log(k- 1)

In particular, all large numbers are sums of 19 bipuudrates; 41

fz’fth powers, 87 sixth powers, 193 seventh powers, 425 eighth powers,
949 ninth puwers, and 2113 tenth powers.
We can still prove nothing new for cubes ; but the result for bi-
quadrates is particularly interesting. All numbers from a certain number IV4
onwards are sums of 19 biquadrates, and it would be possible to deter-
mine a numerical bound for &* Now g(4) 2 19 (since 79 = d-2’+ 15.1’
actually requires 19 biquadrates ) . The determination of the actual value
of g (4), whether it be in fact 19 or some higher number, is thuti theor-
etically reducible to a problem of computation.
1.4. The third part (5 8) stands on a different footing, since it
depends on an unproved hypothesis, viz.
Hypothesis me number of representations of n by E k- th
K.
powers is 0 ( W ) for every posi,tiue E.
This hypothesis (a separate hypothesis for each value of k) is equi-
valent to
(1.41) r&n) --- O(w).
It is true when k = 2 “), but has not been proved for any higher value
of k. Its consequencesin Waring’s Problem are very striking, and it
is well worth while to investigate them in anticipation of a proof,
We prove first
Theorem 6. I/ Hypothesis K is true, then (1. 22) is true for
szk+l.
From this follows
Theorem 6. If Hypothesis K is true, k + 4, and
(1.42) rl (k) = Max (r( k), k + l),
where r(k) is the least value of s fur which the singular series is always
positive, then almost all numbers are sums of rl (k) non-negalive k-th
puwers. In particular almost all numbers are (on Hypothesis K) sums
of 4 cubes, 6 fifth powers, 9 sixth powers, 8 seventh powers, 32 eighth
powers, X3 ninth powers, and 12 tenth powers. me number of non-re-
presentable numbers is in any cuse 0 @l-c>.
Y--
“) E. Landau, Uber die Lnzahl der Gitterpunkte in gewiesen Bereichen,
Giittinger Nachrichten (1912), 687-7’71 (750).
I ’

472
 Some problemfl of ‘Partitio Numerorum’ (VI).

It was proved by Hurwitzl) that

except when k = 4. Hence, except in this case,

Combining this inequality with .Theorem 6, we obtain

Theorem 7. Ilf Hypthesis K is true, and k += 4, then
(1.43) G, (kj = r;(k).
As r(k) can be calculated, with sufficient labour, for any k, the
problem of G, (k) would be solved completely if Hypothesis X were proved.
Finally we prove
Theorem 8. If Hypothesis K is true, then
(1.44)
2k+ 1; .
from which follows
Theorem 9, If Hypothesis K is true, and
(1.45) r,(k)=Max(r(k), 2k+1), “)
then all large numbers are sums of r9 (k) non-negative k - th puu;ers; and
in particular of 7 cubes, 16 bipuadrates, II fifth powers, 13 sixth powers,
13 seventh powers, 32 eighth powers, 19 ninth pouTem, and 21 tenth
puu~ers.
We show also that when (as is usually the case), r (k) < k + I,
Theorem 9 may be deduced in an elementary manner from Theorem 67.
It can hardly be doubted that the numbers of Theorem 9 are far
closer to t,he realities of the problem than any given before. It is not
too much to say that, if Hypothesis K were proved, Waring’s Problem
would be within measurable distance of a final solution.
- ---_-
“) A. H u r w i t z, Uber die Darstellung der ganzen Zahlen als Summen von n-ten
Potenzenganzer Zahlen, Mathematische Annalen 65 ( 1908), 424-427.
“) See P. N. 4, 179 (Theorem 4). The ground of the inequality G (k) > r(k)
being (except for k = 4) the existence of a forbidden arithmetical progression, the
same inequality holds for GI (k).
6, In fact rg (k) = 2 k + 1 except when k is a power of 2, in which ca’e it is 4k.
‘) This deduction is still possible in the exceptional cases, but the proof requires
elaboration.

473
 6 G. H. Hardy and J. E. Littlewood.

474
 Some problems of ‘Partitio Numerorum’ (VI). 7

It may be useful that we should exhibit, in t;he form of a table, all

that is known about the numbers g(h), G(kj, G, (AT), r(k) for 2~kz510.
We ,may repeat that g (k), G (k j, G, (k) are the least numbers s such
that all, all large, almosf all numbers are sums of s k- th powers; while
r (L) is, except for k = 4, (a) the least, number s for which the singular
series is positive for every n, and (E) the least number s such that every
arithmetical progression contains an infinity of representable numbers.
When k = 4, T(k) = 16, while @) and (b) are true for u’2 15, but for
no lower value of 8. “)

2. Further definitions.
2. 1. We must begin by defining a number of additional symbols.
In our earlier memoirs we used the ‘Fare y dissection’ of order
N- [n l-a]? the ‘major arcs’ ‘$R and ‘minor arcs’ m of the dissection
being defined by q 2 na and q > n” respectively. It was then shown by
Weyl lo) and Landau 11) that our argument could be simplified by using
aK
a different dissection, viz. thaf of order [ n1-a], where ti = K+
- ----1’ Here
we use two dissections, each different from either of those used before.
In the proof of Theorems I- 4 we use the dissection of order
(2.11) N1 zIz=
[
,1-q,

where
(2.12) a, = a, (k, 6) = a - 6
and 6 is positive and small. The choice of 6 will be subsequent to that
of k, S, E, so that S will then become a function of those variables, tend-
ing to zero with E. The major and minor arcs will be defined by
(2.13) q<
C nal, q > nal
and denoted by %JI,, m, .
In the proof of Theorems 5 -9 (those which depend upon Hypo-
thesis K) we use a quite different dissection, viz. that of ordel
(2.14) N3 C
[ 13nl-3U

the major and minor arcs E,, m, being defined by

(2.15) q&P,
E q>na,
9, See P, N. 4, 179, f, n. CH), This point about k = 4 is overlooked on p. 188
(last sentence).
l*) H. W ey 1, Bemerkung iiber die Hardy-Littlewoodschen Wntereuchungen zum
Waringschen Problem, Gijttinger Nachrichten 192 1, 189 - 192,
11) E. Landau, Zum Waringschen Problem, Hilbert Festschrift (1922), 423-451
[Math. Zeitschr. 12 (1921), 219-2471.

475
 8 G. H. Hardy and J. E. Littlewood,

Here ti is a number of the type c, which will be chosen small enough

to fulfil the demands of our analysis.
2. 2. We use t = fp 9q, as hitherto, for a ‘Farey arc’ in general.
We write
- -1+i?p
(2.211) X E Ix I eiV’ =e 98
and (on the arc & J
(2.212) x=&)P, p=;+iB, e==q
Q
We write also
s 1
(2.22) (DE ‘P,,,=2r(l+u)--~~-$.

(2.23) g (2) = 2 m8a-r 2” (1x1 < l),

??S=l
(2.24) F(x)= F(x,k,~)=~~~(m)x~=~m~~~~~~rn)x~,
m=l m-1

(2.25)

Pinally, we write
(2.26) ~,(n)=Km$(n>, G,(n)=O22$(n),
qs !7>y
so that

and
(2.27) Fl(x)=~m~a-lG1(m)x~, F,(~)=zrn8a-~@~(rn)xm,
m=l Ina=
so that
(2.271) F(x) = E: (4 + FQ(x>’
Then
(2.281) Fl = (ii$2 2 (+)8m8a-i(xeq(- p)>” = c Fp,*, 11a)
P,q5 m P,QS
and similarly
(2.282) F a = 2 Fp q-
Pt!l>=V ’
We shall take
(2.29) V = d;
b will be chosen later.
We snppoge always that n > 1.

lla) This notation means that we gum for q 5 Y and for all values of p aB-
sociated with each such q.

476
 Some problem of ‘Partitio Numerorum’ (VI). 9

3. Lemmas concerning the singulrsr series,

3. 1. We have frequently to appeal in what follows to the results
and arguments of P. N. 4, and we must begin by revising slightly a por-
tion of that memoir. An incorrect argument in the proof of Lemma 15,
to which our attention was called by Prof. Landau, affects the results
of Lemmas 15, 16 and 17, and the proofs, though not the results, of
Theorems 15, 16 alrd 17. It happens that the correction is relevant to
the present memoir.
Lem-ma P). We have, for s > 0,
(3.111) y!&(n)1 < Q-H”

(3.112) [q(n) 1< qw8

(3.221) 1x, - l/<CpkG
(3.122) ‘X, - I ] < cpl- kg (P I % P”fnL

Also
(3.14) IX,--1 I< cp’ (Skk-+ a).
Here (3.121) and (3.122) give the corrected form of Lemma 15,
(3.13) that of Lemma 16, and (3.14) that of Lemma 17.
The result (3.111) is correctly proved on p. 176; and (3.112) follows
by the same argument, the only difference being that the ‘inner sum’ is
the trivial 0 (p ) instead of 0 (I$). 13) Th e subsequent argument of the
text is a correct deduction of (3.122) from (3.112), provided we write
O(P 1-h) for qp+q in the two places in which the Iatter Occurs.
The other formula (3.121) is, Of course, merely a convenient restatement
of (3. ill), since Xp =-1$-B, when #n.
If now we substitute, in the proofs of Lemmas 16 and 17, correct
expressions from (3. ll), (3.12) in place of incorrect expressions from
Lemma 15, and leave the arguments otherwise unchanged, we obtain
(3.13) and (3.14).
3.2. The remaining Lemmas and Theorems of P. N. 4 are actually
true, but the proofs of Theorems 15, 16, and 17 have to be revised. We
require Theorem 15 in the sequel-, and restate it.
--
12) Were 8, p, X are the A, ZT, x of I? N.4.
12a) This notation means that (P’)~ is the higheet power of pk which divides n:
pee P. N. 4, 166. The /Y here has of course no connection with that of (2.29).
13) Our mistake there lay in failing to distinguish the two cases.
 10 G.. H. Hardy and J. E. Littlewood.

Lemma 2. If 8.2 Max(r(k), a> then

(3.21) G(n) > c=
The original proof, when corrected, requires the 4 to be replaced
by 5. But the result is still valid when k = 4. la) For if s = r(k) = 4,
k is odd, since otherwise ylr = 2 0+2 > 4. Then, writing

in U1, by (3.121). On the other hand, in L&,

d S,, p is real, S& is po-
sitive or zero, and (since $))n)
!!ip= zSi,pep(- np)= xS,“.& 0.
If now /? = 0 we have Ia)

while if p > 0 we have le)

(3.24) Xp&Xp(p)~l-p-“’
by (3.23), since p = 0 for n = p. From (3.22) to (3.24) we obtain

G>C n x, B(l-p-‘)n(l-p-3)>C;
PiW<C I,tn Pb
since X, is positive when p is fixed and s 2 r(k), and has only a finite
number of distinct values for different values of n.
tj. 3. In the proof of Theorem 16 we have, by (3.13) above,

x,~(1+~-“+~-J+..*)(1~-Cp-~-)<1+C~-~,
and the theorem follows.
In the proof of Theorem 17 we have s 2 k > 3. Hence

G < c 12x, nx, < c mp,

P-h m Pin
since
)(<~+&i-“<l+Q-~;

If p n, we have, by (3.13),

x, < 1+ cp’-;f”Q,,

14) We owe this observation to Prof. Landau.

l6> P. N. 4, 176, (4.14) and the equation five lines lower.
‘3 P . N. 4, 177, (4. 22).

478
 Some problems of ‘Partitio Numerorum’ (VI). 11

where pp = 1 unless $Ik j n, and then pp = 1 + b. Thus

x,< (l+;)op< (l-;)-cQp?

n XP < (&jC JhPm

PlfJ 9%
The first factor here is less than n& for sufficiently large n, and it is
proved in the original argument that the same is true of the second. Hence
G < nE
for sufficiently large n, the result desired.
S. 4. Lemma 3. We have
I&
! _ -2..q_ < B q-a,
I Q

(3.42) pl,(m)j < cql-Bu

for m = 1, 2, 3, . . . li).
It is enough to prove (3.41), (3.42) being an obvious corollary;
and it is enough, in virtue of the multiplication theorem I’) for x?, I Q, to
prove (3.41) (i) when (q, k) = I and (ii) when Q is a product of powers
of primes which divide k.
(i) If p + k, we haveI”)
s p,paWr ’
(3.431) - -- --- I=p --a-l 2 (pakf,rrjwa
ak+,u
:p I
for

and
(3.432)

for d > 0. Write now

(3.441
where Q contains all factors py of q for which r is not congruent to 1
(mod k ), and q1 is a product of different primes. It follows from the
mulfiplication theorem and the inequalities (3.43) that

17) Compare P. N. 5, 49 (Lemma 5). The argument there is simpler.

le) P. N. 2, 18.
Is) P. N. 2, 19-21.
20) Here, and in the formulae which follow, the value8 of the ~‘8, which differ
from formula to formula, are irrelevant.

479
 12 G. H. Hardy and J. E. Littlewood.

so that it is enough tlo prove (3.41) when Q= ql. But, if Q= q1= IQ,
we have 31)
I ~Pdh 1 23
] pi 1 -c G < PI-”

for p1 > B, and so

(3.45)

This is (3.41) for q = ~1 l

(ii) Suppose that po(k, where 8>0, that

A=ak+p>k&8+2, O-c&k,
and

Then 23)

where (P, p) = 1. Hence

The inner sum vanishes unless b’= ph,, when it is pe+l. Hence

and
(3.46)

This equation naturally holds also when A -< k, with a = 0, R= p* From

(3.46) we deduce

(3.47) p+pf =~-“l~,*~ -=-CT,

P
since the last product has at most B factors, all less than B. This
disposes of case (ii), and campIIetes the proof of the lemma.
3.5. Lemma 4. If PI, &, .- . . P$I, is a set of C primes, and Q is
a number formed from these primes only, then
(3.51) &Q-C< CC” (6 > 1).
For

“.‘) P. N. 4, 173,
az) P. N. 4, 166, Lemmsl I.

480
 Some problems of ‘Partitio Numerorum’ (VI). 13
c

3.6. Lemma 6 zx). l%e series C 8, (m) is absolutely convergent

i or m >O, s&Jc-j-1. Also
(3.61) ~l%q(m)I < C~~-c(d(m))S (ek+l),
9>y
(3.62) j G3(m)[ < CY-c(d(m))8 (8 2 k + 0,
(3.63) 1G,(m) 1< cYz-s= (8 2 2E + 1).
The last result is easily proved, for, by (3.42),

! @ (m) ( < C 2 I q,(m) j < C 2 ql-ea< CY~--SO

9>*” or
(since 1 - sa < - 1 if s ‘I Sk).
The proof of (3. Gl), of which (3. G2) is an obvious corollary, is
much more elaborate. We write
(3.64)
where now q’ is the product of all odd factors of q which are prime to k,
and q1 is the product of the primes which divide q’ once only la); so that
(3.65). VI9= ~,,%Jo.

In each term of (3.61), one at least of ql, qa, Q exceeds Ye ; and if

Z;, Z$,, Z8 are the sums of the terms in which (i) ql> Y$ (ii) qs> Y;,
and (iii) Q > Y’, then

3.7. We write q1= LIP,, qO= n.p,” ; and we make three prelimi-
nary observations. First y
(3.71) Irn,l< cq-C< c
for all q and m, by (3,12), and in particular when q is ql, qs, or Q.
Next aa)
9$(m) = 0
2

unless pi-’ I~YL, and a jortiori unless .p,”1m9; so that Sq, (m) = 0 unless
q21m2. From this last remark it follows that not more than d (m3),
and therefore not more than ( d (m)) ‘, different values of q9 can occur
in nun-zero terms of (3.61). Finally

Ba) Compare P. N. 5, 52, Lemma 6.

s4> Thus, if k= 13 and q = 22-3.58~7aJ1, qr= 11, q2= 72, Q= 2’m3-5’.
*y P, N. 2, 20; P. N. 4, 170.

481
 14 G. H. Hardy and J. E. Littlewood.

In particular, taking 5 = +, we have

(3.721) 2 1% I < Cd(m).
!A> 0
We can now write down upper bounds for ZI, Z$, and J$. Consider
first q. Associated with a given ql, we have at Lost (d (m))’ values
of q3 ; and ) !Xq, 1 < C, 1‘%Q1 < CQ -‘. Hence the sum of these terms of
XI is less than
c(d(m))“j~~~/ZI&-c<c(d(m)!2](IlplI,
Q
by (3.51); and
(3.73) zl < C(d(m))“z j%,,j < CV-~ (d(m))‘,
Q1>,j-
by (3.72).
The discussion of ZT is similar, but we use the inequality
I %2 I < %-c- We plainly obtain
(3.74) Z2 < Cv-c(dCm))’ ~\%,,I < Cv-c(d(m))3,
Ql>O
by (3.721). Finally, the discussion of Z3 depends on the inequalities
used for Z;. We have
(3*W 4< qdw)Z,r)%q,j 2 cc
!ll>O
Q >.p+
< C(d(m))3 )7 &-“< Cv-c(d(m))3T
Q>A
by (3.51).
It is clear that ( 3.61) is a consequence of (3.66) and (X 73)-(3.75).

4. The behaviour of f(z) 0x1 a major arc WI.

4.1. Our object in this section is to prove
Lemma 6. On a major arc !Dit,
(4.11) If-~~,&BEd~x+E,
f
so that
(4.12) If - %,A < BeqXfE
if 6 is chosen appropriately as a function of k and E.

482
 Ic
Some problems of ‘Partitio Nnmerorum’ (VI). 13
This lemma may be deduced from the analysis of P. IV. 1, $4 8.1- 8.2.
Every !!& is part of an 92 (since a, < cc and 1 - ce, > 1 - a), SO that
the formulae there given may be used. The sums there called Qp and Qi,
are of the form required, and it is a question only of the m,agnitude
of Q1. In this sum we have now
cos%p> Bq%-“ap)
-1-b
Q n b (~osy)l+~ > B q-l-bnb(pn-al)l+b = B n’l+b’d > Bnd,

1E 1< Be-BmlSbnS.
It follows at once that Q1 is also of the form required, and indeed extre-
mely small.
If this course were adopted, Lemmas i’ - 11, which are required only
for the proof of Lemma 6, would become unnecessary, and our analysis
would be appreciably shortened.
The objection to this procedure is that it involves the reinstatement
of the whole of the analysis of 5s 7 - 8 of P. N. 1, which is eliminated
entirely in L and au ’ s version of the proof. If we are to preserve the
simplifications introduced by Landau and Weyl, an independent proof
of Lemma 6 is essential.
4.2. Lemma 7. I/
(4.21) 02 P
then
(4.22) / hm I =
I
We have 27)

which proves the lemma when nz = 0. And in

Lemma 8. Suppose that Osh<q,

(4.23) x = e-', Y- 4'9, t = $, V- e-y'v+t)"(v + t)'-',

(4.24)
-.
2s)We llae for the moment the notation of P. N. 1, except naturally that, in
conformity with the conventions laid down in 9 1. 1, we write B instead of A.
e7) Landau, 1. c, 226 (Hilfssatz 2).

483
16 G. H. Hardy and J. E. Littlewood.

and that P(v), P,(V), Pz (v), .*. are polynomials in (v) = v - [v],
defined successively by
(4.251) PC v > =v+ l-(v)=v+,P(v)=~gv+ P(v),

(4.252) PI(v) = s” P(w)dw = c,v + PI (v),

0

(4.253) ~~(v)=fP,odw=c,v+P,(u), .. ..
0

so that P(v)= 1 - (v’), P&J)= i(v) -B(v)‘, co = 1, c, = $,

1 and in,
general cr is a function of v only. Then

(4.26) &- =&-l)rc,,~vv”l’dv+ (- I)‘~P@)V1”dv

p=o 0 0

This lemma merely embodies the definition of certain symbols, and

the results of certain formal transformations required for the proof of
Lemma 6.
If we define N(x) by

we have

Writing u=(vq+h)l,qL(v+t)‘, and observing that N(u)=Y(vj,

we obtain

f h = kYrP(v)Vdv = kYjPvVdv + kYfp;(v)VdP,

0 0 0

Z kYs*vVdv - c,kyjiV’dv-kY~=P1(v) V’du.

0 0 0
This is ( 4.26) with Y = I, and it is plain that the general formula follows
from a repetition of the argument.
Lemma 9. If
(4.27) H=W(Y)=W(@y)=$
then
(4.28) 1Yl < $$- < An-”
u?t m,.
 Some problems of ‘Partitio Numerorum’ (VL). 17

For H < n”lk-l< 1 and

I- WI l-aiyj <An~-a=An-“.
H1-a =qn

4.3. Lemma 10. We can choose v. = v. (k, S) so that

Since 1k Y j < A k < B, we may ignore the outside factor. We take

v =vlk, where 35 > 2.
We write
VW r3 e-Y (v+tlk y c
- k ) m “(’ + t)mg
i,m ’ ’
Every differentiation of the exponential factor of V introduces a factor I’,
and a term in Y’ can be generated only by I such differentiations together
with Y - I which do not bear upon the exponential. Hence the term in
YE is a multiple of

(4d
dV
v-z [(v kY)‘( v + l)‘k-” (z+l)],

and is zero unless (k - 1) (2 + 1) 2 Y - Z or

v-k+1
Z>Z -j->v,-1.

That is to say, Y, is the least possible value of 1. Associated with a

given 1 we may have. values of ry1 from 0 to (k - 1) (I + 1) .
It follows that

< B,, c; Y1z(H-“-C H-(z+l)(l-a,-o)

1
Z=Vl

%
< Bv,~H%+< Bv,~d-+ < -n

(by (4.28)), if y1 =v&k,d), i. e. if v=vn(k,6),

485
 18 G. H, Hardy and J. E, Littlewood.

It should be observed that we can aIways replace vI) (E, S) by a

larger number of the same type.
4.4. We understand by & (Y, b) a polynomial in Y and t, whose
degree in either variabIe has an upper bound of the type l3d and whose
coefficients have upper bounds of the same type. Since 1Y 1< A, 0 =< t < 1,
we have always I& ( Y, t) 1< Barn And we understand by R a number
%
whose absolute value is less than %-.
Lemma 11.. We can choose vO= vO(k, S) so that, in addition to
(4.31), we have

(4.41)

(4.42) kYIo = kYjbdv=T(l+u)Y-a+Q(Y,t)+R.

0
We have
e-Ytk = m, (- Ytky + r
c Wt!
.--- ------------ 3
m-0

if m, = m, (k, S) is sufficiently large; and so

edytE = Q(Y, t) + R.

=kYQ(Y,t)(Q(T:t)+R)=Q(Y,t)+R;
while if p = 1

EYI; = kYfuV’dv=-kYfVdv=e-Y@=Q(Y, t)+R,

0 0
Finally, if p = 0,

486
 Some problems of ‘Partitio Numerorum’ (VI). 19
t

4.5. We can now prove Lemma 6. We have

q-1
(4.51) f= --1-+2~e,(phk)~e~“rr’2/=-l++~eq(phk)fh.
h=o m-0 h

It followa from (4.26), (4.31), (4.41) and (4.42) that

(4.52) f,+=r(I+a)Y-“+Q(Y, t)-t-R
when Y=Y~. Combining this with (4.U), we find

(4.53)
where
(4.s41) e=~~+ep+e~, &==--1, ee=Z:e,(phk)R,
h

(4.542) e3 = J&W) Q (y7 f) n

Now

(4.55)
Further,

Q(y7,“,=t:%,
Lm l,xkl
yz($)Y
where I, m, and the coefficients have upper bounds Ba. And

(4.56)

by Lemma 7. Lemma 6 follows from (4.53) to (4.56).

5. Further lemmas preliminary to the proef of Theorems l-4.

5.1. Lemma 12. I/ 1x 1-c I and

then
(5.11) Isi4 --c CI&lg+-8a,

(5.12) 19(z) - r(m)(logyal < c.

These are known results.

Lemma 13. We have

487
 20 G. II. Hardy and J. E. Littlewood.

This is trivial; for, if x= “e,(- p) = e-g, we have v= logi and

by Lemma 12.
Lemma 14. We have
(5.14)
For the sum is less than

and

-2sa 02 #
84
E +28a- 1
--Qt
s ----_
(1 +u2ya
dzc
I__ - <(, @a-1

-m
< A~PQ-~.

Hence, by Lemma 3,

(5.16) siF,!'dy< CnBsa-+'

0

(5.17)
where
aI----- zP( sa- 2) ,

17
) @ is ultimately chosen in three different ways to suit three different argu-
me&.. In each case it is chosen aa a function of k and 8 only. We anticipate
this choice and allow ourseIves to treat p as a c.

488
 . Some problems of ‘Partitio Numerorum’ (VIj. 21
(i) The integral is

2n~(~,(m))am2”“-2~xj~m<Cn-c~mZ”‘-”(d(m))0~x~2m
< cEn-C2m28&-2++y< c$28Q-~iC++,

by Lemma 5 (3.62). The inequality (5.16) follows if we take E to be

half the c which precedes it, and then replace c by 2 c.
(ii) If s 2 2 k + 1, the integral is also less than

2 mzra -2
4--28a 1212~<Cy4-2sant’sa-1c= Cn28a-b~l,
Cr
by Lemma 5 (3.63).
5.2. Our next lemma, which concerns the value of when x Fp,Q

lies on an arc [, *, requires a word of preliminary explanation. The

function Fp Q and ‘the arc & 8 are invariant if P is altered by a multiple
of & , and ‘we obtain the colhplete set of arcs fP, for a given &, if P Q,

runs through any complete set of residues prime to &* In what follows,
p, Q, & being fixed, we select for our set of residues that for which

(and of course P= 0 if & = 1 )* Prom this it follows that

Lemma 17’“). Suppose that x lies on & and that &,Q is a Farey
arc other than F,* Q. l%en
(5.21) fIF~,~'(CN~alpQ-pPj-'a.

We have
(5.22) x s 1x \ e+ = Ixie,(p)e+@= /xieg(P)e-i”,
where
(5.221) -2n($-;)+e,
so that / o iSn+pq< $7~ if N1 2 4. Hence

( 5.23) ~F~,s~=~~(~)8~(~~ie-".')!<CQ-8a~w - 8tX

I 1

by Lemma 12 (5.11) and Lemma 3 (3.41).

If p, IT, and P, & do not correspond to arcs adjacent in the dissection
then
P 4x
n p A- > 22 32p1,
lI2 ( i-
- Q >II-+-- I-
“*) Compare P. N. 5, 52 (Lemma IO).

489
 22 G. H. Hardy and J. E. Littlewood.

and so
Pi A
(5.24) ICC)/>A,;--i=-----;pQ-qP1,
&I QQ
As q 2 N1, (5.21) follows from (5.23) and (5.24).
This argument fails when the arcs are adjacent. In this case
ipQ - qPj == 1. As x is not on &Q,
A _ A..-- :r,Q-tP!j
IuI I >
QN1
‘II QN
1
j

so that the conclusion still follows.

Lemma 1P). If
(5.25) G= r:IFP,Q.,

where the sign C implies summation over those pairs (9, Q) 3 distinct
frmn (p, q>, for which Q 5 v = nb, and s >C k -/- 1, then

where
(5.261) a.~=a(2sa-l)-4p,
and 1 = l(k, s, S) tends to zero with a; so that

(5.27) 2 sG”lde; c Cen28a---1-2+F

4Parl
if d- 6 (k, s, R> is sufjicienfly small.
Since G is a sum of less than 9” terms, we have
--“-
G2<v2z1Fp,J2< CY"Nf8'zjpQ- qPj+
by Lemma 17. Hence

where now the inner summation is defined by q 2 IV--, p Q - q P + 0,

and the outer summation by Q 5 Y. But, if Pq E h (mod Q), where
Osh<Q, we have
~IpQ-qP~e280< ~‘lmQ+him28a30) <%+-2Q-28~~lm~2sa<A,
P m=-m

FO) Compare P. N. 5, 52 (Lemma, 11).

30) The dash denoting that any term for which m Q + k = 0 ie to be omitted.

490
 Some problems of ‘PartiC Numerorum’ (VI). 23
and so
(5.28) ~~G2~dBi<CyZN:d”-1~~~<C~4~~85-11~g(~~+2)
t EQ Q
< Cn@8a-1)(h+k48 logn < ~&28U-1+&tE+t
where Z= (2sa - 1 j 6, This PROWS (5.2s>. To prove (5.27), we chdm 6
so that 1 < E, and then replace 2~ by E,
5.3. Lemma 19. If s~k+l, v=d, and ~-‘<a,,
ther,

Here
(5.33) n. = Min(2/3(sa - I), A& 2 Mk@,, A,),
R, and A9 being &ined by ( 5.18) and (5.261); 1= 1(E, s, S) tends to
xeru with 6; and the second inequality is in each case secured by an
appropriate choice uf 6.
(i) We have

(5.341) IF 1
if qgv, and

(5.342)
if q>v* Hence
(5.35)
P

by Lemmas 13, 15 and 18. As A, < 2p(sa - 1) and A3-<A, d < 2sa - 1,
(5.31) follows*
(ii) The proof of (5.32) is similar, but rather simpler. Since
w a,, q>y on every #I,. Hence
(5.36) zJjF.J2idel 222 S$“jd+t-2Z SG21deI,
1nt q>r ;” 4x> 4:
and the conclusion follows as before.

491
 24 G. H. Hardy and J. E. Lit&wood.

5.4. Lemma 20. If $2 2

(5.41) so=so(k)= (ik- 1)K+2,
and
(5.12)
then
Zsa-1-A4+vkC 2sa-i-R,+&
(5.43) FJ ] f,2”id0/ cr: C,n < CEn ?
ml

the meaning 0,f 1 and of the double inequality being the same as betwe.
It is known31) that
a-;-fE
;fl< B,ti
on ml. Hence

(28 - 4) (a - 2) + 3a = 2sa - 1 -p,

where

which proves the lemma.

5.5. Lemma 21. (i) Ij 8 2 2k + 1 then

(5.511) a6 = Min (I,, 1,> = Min(R,, au),

and the second i?zequulity is secured by choice of 6.

(ii) If skk+l then

(5.521) &= Min(R,, It,‘, = Min (Ad, $),

and the second inequality is secured by choice of 6.

a1) Landau, 1. c. 230 (Hilfamtz 4).

a2) See P. N. 2, 164~; Landau, 1, c. 241.

492
 Some problems of ‘Partifio Numer~rum’ (VI). 25
We have
If”-~“j3<A,jf-~~2”+A,If-~~2/p71*8-2,

(5.53) ~SIf”-~~/eIdeI<AsZ1S!f-~i861de1
% at1
+A,~‘SIf--p,l”Ip,188-21~~I=A.e~+~Be9~
%
say, Now
] f - y/ < Be,dqx+s< B, ¶dn~++EJ< B, ,6n”(“+“)
by Lemma 6, and so
de 1< CE,dn28ax+8G--1+e,
by Lemma 14. Now
2sax + 2a - 12 2sa - 1 - A4 = 2sax -4ax+2a
if 4a x < 1, which is true (with inequality except when k = 3). Hence
(5.54) & < CE,dn28a-1--b~+E.
Again, using Lemma 6 once more, we have

-
Now
] hw
I Q
by Lemma 3, and

(as in the proof of Lemma 15). Hence

(5.55) g3 < cE,a+(8-1)-1 )7 q2x+1-2a(e--1)+&,
qsn”l
and so also, replacing q 2w+eby its upper bound TZ~~++E)
- (5.56) es < Cf,an 2U(8-1)-iid&K+E
2 91--Z(8--)e
q~?e
In order to establish the complete lemma it is enough, after (5.54),
to prove that
(5.57) @*< Qs,a ~28a--1--2afe (s22Ic+1)
and
5!8a--l-
3a
K +E

(5.58) & c &an (8 2 k -+ 1).

Now 2x+1-2a(g-1)<3-4&=-l if s>2k+l, and then

(5.57) follows from (5.55). Finally, we may replace u, by a in (5.56);

493
&J(j G. H. Hardy and J, J1:. Littlewood.

observing then that l-22cc(s-P)<l-2uk=-1Z when s&k+l,

(5. 56) gives
sag-y +& zag---03a
K +’
& -=c CE, 6 n logn < CE,dn 3

which is (5.58).

6. Proofs of Theorems 1- 3.
6. I. Proof of Theorem 1, We have

say. Upper bounds for all these sums have been found in the preceding
lemmas. Thus, when 6 is chosen appropriately, we have
23a-l++-&
(6.121) J1 < C,n (8 2 2k + I>,
(ti.122) J1 < QEn2sa-l-;-g+& p 2 k + l),

by Lemma 21;

(6.13) (s 2 JG+ 1)
by Lemma 19 (5.31);

by Lemma 20;

(6.15) J
4
< C
&n2sa-l-~3+~ (s 2 k + q,
by Lemma 19 (5.32); and

ts* 161) J6 < C+a-1-& (s 2 211: +1>,

(6.162) J5 < Cn2ga-l-Y (8 2 k + I),

 Some problems of ‘Partitio Numerorum’ (VI). 27

by Lemma 16, y being the particular c which occurs in that lemma.

Hence, on the one hand,

(6.17) &G(o (m))” < CE98a-1-I.+e (s22k+l),

tPS=l

(6.172) 2, = 2B(sa - 2), 1,= 42s~ - 1) - q?, 1, = g(s - so),

and on the other

(6.18) (s 2 k -j- l),

where
(6.181) A’= Min(y, A,, A,, 1:) =Min(y, 2B(sa - l), A,,
ti 1,, Ai),

The inequality (6.17) 1l s much better than (6. 18) when it is valid,
that is, when s E
> 2 Ic + 1; and it is essential to the proof of Theorem 4.
But (6. 17) is not valid for s > sOwhen k: = 3, since so= 4, 2 k + 1 = 7 ;
and we therefore use (6.18) in proving Theorem 1.
We note first; that s,>k+l; for if k=3, sO=4, and if k>3,
s,> 2”-‘+2 > k+l. Hence (6.18) holds for s >s,. Also
2~(sa--I>~2aj9, IZ,=a(2sa-l1)--4~>a-4~, &&$
If we take p = i a (and suppose 6 small enough to Secure that /3 < a,),
2a
the numbers on the right hand sides of these inequalities are 2, ia, -* .
All these numbers are c’s (indeed independent of s>, and so is y+ Hence
we may write c for 1’ in (6. 18). If finally we chose E appropriately as
a function of k and 8, we obtain (1.22), and the proof of Theorem 1
is completed.
6.2. Prod of Theorem 2. If
(6.21) S>S,= Max(r(k), 4)

then, by Lemma 2,
(6.22)

Now ($k - 1) K + 3 2 4k 2 r(k) except for k = 3 and k = 4, while

(i k - 1) K + 3 = 5 p- 4 = r(k) for k = 3. Hence (6.22) is true in all
cases considered in Theorem 2.

495
 28 G. H. Hardy and J. E. Littlewood.

If r(m) = 0, then lo(m)/ = e(m)> Crn80-? If then ,u(@ is the

number of non-representable numbers between in (exclusive) and m jin-
elusive), we have

$u(mf> c($2)-p(?2)

and so
p(n) < c228a-2n2-2~ai,(a(m))2< Cnl-c,

by (1.22). The total number of non-representable numbers less than n

is therefore less than

cd-++ ($)1-c+(~)‘-c+.*.)<Cd-c.
This proves Theorem 2.
6.3. Proof of Theorem 3. When k = 4

(;k-l)K+3=11<16=r(k),
and the preceding argument fails. In this case the conclusion of
Theorem 2 is false.
suppose that
s = 15, f>O, -&<rn~n,
and write
m I m’ (16’-+m), m = m” (16’1 m).
Further, suppose that a zero suffix attached to a number implies that it
is not representable by 15 biquadrates, and that p (n), p’(n), p”(n)
are the numbers of the classesm,, mi, m{.
We have 33)
(6.31) p(m’) > A,(m’)’

+n

(6.32) p (n) = #+> i- tLYn ,) < n(A,ra-+ 16-‘).

35) P. N. 4, 179.
=) We we a momentarily for an indicial A (d. e. an absolute c). We may
plainly svppose a < 1,

496
 Some problems of ‘Partitio Numerorum’

If y(n) is the total number of non-representable numbers less than n,

then, by (6.32),

-.“--A,p+ 2.16-’
21-L 1

by choice first of t and then of n, This proves Theorem 3.

7. Proof of Theorem 4,
7. 1. We denote by

N!g(4 = Nk ,@>
the number of distinct numbers not exkeeding n and representable by s
or fewer positive k- th powers, and by
Y&)=Y, ,,(n)=n+l- Nk, Jn>
the number not so representable; and by
#* = % Ck>
the largest number [ such that
N,(n) > C$-&

for 782 no@, 8, E). It is plain that cly exists, does not exceed 1, and
is a non-decreasing function of s. Also, as there are [ na] + 1 k- th
powers which do not exceed n, we have as> a for s > Z 1.
Lemma 22. We have
(7.11) a,2 I- (1 - 24(1- q2 (s 2 2).

First, suppose s = 2. Then (7.11) is a,,I- > 2 a, and is an obvious

deduction from the formulae

we may suppose then that s > 2.

We divide the interval (1, n> into the n = [n&l parts
(l”, 2’), (a”, 3”), + l ., (I?, n>
(of which the last may disappear). We denote the interval beginning at
j” by $, and its length by Zj. Then Zj> kj’-l, except possibly for the
last interval. If then we consider in particular the intervals $ included
in ($ n, $ n), it is clear that there are at least Bna such intervals, and
that each of them is of length greater than Bd-a.

497
 30 G. I-3. Hardy and J. E. Littlewood.

In (0, Zj) lie N,_, (Zj) numbers representable by 8 - 1 k- th powers;

and 80 more than N,-, (B&-a) such numbers. Adding each of these
to p, we obtain more than AT#- 1 (BP+=) dz'stin~t numbers, lying in ‘j,
and representable by s E-th powers, It follows that
N,(n) 2 Bn” dv,-, (Bnl-“),
and therefore that

From this (7. 11) follows by induction.

7.2. For the proof of Theorem 4 we work with a new value of the
number /? of 5 6, viz.
2sa- 1
B Z- 2s ’
observing that the necessarycondition p < a, is satisfied if 29s < 1, which
is certainly true if 6 is sufficiently small. This value of makes
(2sn- -l)(.W-2)
(7.21) d1 E R-
2-
--.- _-. - _--_ b
S

We denote now by Iz= Rs= R, s the number defined by (13.17I j and

(6.172) when 6 has its new valie.
Lemma 23. If

(7.23) v, (nj < c&-A+e l

The proof is similar to that of Theorem 2 (§ 6.2). If m is non-

representable
(a.(m))’ = (r, Cm) - e,(m))2 = (I,)” > CrnzBa-‘,

L~DUM 24. Suppose that 3 satisfies (7.22). !l%en if R, > 1 - a,* t

and therefore certainly if

(7.24) VU - 2 a) (1 - aje’m2 (5’ 2 2)

we have
(7.25) . G(k) Is +a’.

498
 Some problem of ‘Partitio Numerorum’ (VI). 31
In fact, if n is not representable by s + s’ powers, there are at least
Ngf (n) numbers, not exceeding n, which are not representable by s powers.
That is,
Y8(n) 2 Nsf (n), 1 - A, 2 a,’
in contradiction with the hypothesis 1, > 1 - a,‘.
7.3. We are now in a position to prove Theorem 4. The case
E = 4 is exceptional, and we suppose first that k > 4. We shall verify
that numbers s and s’ exist which satisfy the conditions

(7.31) s+d=($k-- 1)K+k+5+[:,1;

i7.32) s 2 r(k), skZk+l, s>s(), s’>- 2;

(7.33) 1, > (I - Za)(l - .y’-“.
We shall prove that these conditions are satisfied by taking

(7.34) s=(~k-l)K+k+l=~O+k-I, g’=[t,]+4.

In the first place, these values of s and s’ satisfy ( i. 31). Next

(7.35) 8, > $k - 2)zE-’ > 2k-3 k>4kzMax(r(k),Zk+l).

_

Thus the first three inequalities (7,32) are satisfied. Also 5, is positive,
since its numerator is
-2
log+>o;
and so the last inequality (7.32) is also satisfied. It remains only to
verify (7.33).
When we substitute the value of s from (7.34) into (7.21) and
(6.172), we find
(2s,+k-2)(s,-k-l) n 20-q 2 get
R1 _I Ra = ___ -.----_---~
k2p+k-l) ’ 4= K ’ 6= l

Of these numbers, the least is R, . For it is obvious that R, < &; and
s,z4k+l if kk5, so that

Hence i = i,, and it is only necessary to verify the inequality

(7.36) 1=1,=2a(k-1)>(1 - 2a)(l -a)[C’1+2.

This is
t$kl+z (k-2) K
’ 2(k--l)

499
 32 G. H. Hardy and J. E, Littlewood.

[[J + 1 > (k-2~1QR2-logk+log(k-2). Elk

(7.37) logk-log(k-1)
which is true. Hence (7.33) is true3&).
74. lIt remains to consider the casek = 4; here [c,] = 2 and we have
to prove that G(4)s 19. The difficulty is that so+ Ic - 1 = 13 < r(k);
and thiEt difficulty cannot be overcome directly bY another choice of s
and s’. There are two alternatives.
(i) We may consider each progression Z (mod 16) separately as
follows. Let y8 I (n> be the number of numbers m G 2, not exceeding n,
and not represintable by s fourth ‘.I.powers;
__ and &it (n) the number of
distinct numbers m E Z’, not exceeding n, and representable by s’ fourth
powers.
It may be shown, on the one hand that
(7.41) p(m); >A (s=so+3=13, m+o, 14,15)
and so
(7.42) v, I l(n> < A&-“+ (qa 14,w;
and on the other that

(7.43) N,), lf (n) > A, nCL&‘--E

,
where a##satisfies (7.11). Suppose now that arbitrarily large numbers
n E 1 are non-representable by 19 fourth powers. We take s = 13 ?s’ = 6,
choosing 2’ so that
z-z’+o,14,15.
Then the argument of Lemma 24 shows that there exist at least A&l 2’(u)
numbers m, less than n, not congruent tm 0, 14 or 15, and not iepre-
sentable by s’ fourth powers. Thus

and so
1 -

But this is false, since calculation gives

The proof of (7. al), however, reopens the question of the arithmetic
of the singular series, which we do not wish to discuss here. We there-
fore adopt a second alternative.

33) It may be that no other of 8 and 8’ Ieade to a

of 8 +.a’.

500
 Swe problem8 of !Pa,rtitio Numerorum’ ( VI). 33
(ii) We take the pair of numbers s I--- 15, 5’ = 4; it happens that
we can, by a special device, suggested to us by Mr. A. E. Ingham, prove that
Y, (n) < A,r~l-~+&
for s = 15, and this proves to hbe sufficient for our purpose.
Let rt (t ) be the number of numbers m for which (i) 16t 1ry1,
(ii)0 < wz< [, (iii) m is not representable a% a sum of 15 fourth
powers. If now in < m < n and 16 +- nz, we have3ja\/

Q(m) > Am: > A&L An&*-l;

and, if also m is non-representable,

Hence by (6.17) 38)

whence
(7.44) z. (n) < A, nkL+& .
Now a number m = lstm is representable as a sum of 15 fourth
powers if and only if m’ is so representable. Hence

and so
y4,15(n) = zrt(n) = 2 z. (w16-‘) < AEn’-i+E,
t=n t=o
by (7.44). Now
(15-2)(15-8) 15-N 1 5
-m-’ 2 = 16’
-4
and
1 - 2a) (1 - ay

if s’ = 4. Our former argument therefore applies to show that

G(k)s15+4=19;
and the proof of Theorem 4 is completed.

Ilblr) This is the epecial cage of (6.31) in which t = 0.

3u) V&d since 15>8=2k.

501
 34 @. H. Hardy and J. E. Littiewood.

8. Proof of Theorems 6 -9.

8.1. In this section we use the Farey dissection of order N,, = [n1-3u],
the major and minor arcs being defined by 4 < na and Q ; na, and a
being a number of type c, which will be chosen small enough for our
purposes. It is plain that, if ti is small enough, every major arc 9X9 is
part of an 9X, and tz fortiori of an ‘92.
We return to the anilysis of 6.1. We take (6. 11), replacing !!I&, m,
by 9X,, m,, and we have to show that each of the terms J.. , Jo, J3, J4, J5
is less” thin C+a-l-c. An examination of the argument shows that,
except in respect of J3, it requires very little alteration.
(i) The term J5 is disposed of by Lemma 16, and no further dis-
cussion is necessary.
(ii) J1, and J4 are disposed of by (6.13) and (6.15), each of which
depends 0; Lemma 19 (and so on Lemmas 17 and 18). The result of
Lemma 17 stands with Ng in place of A$; and (5.28) is replaced by

\‘SG’Id8, < Cv4N;‘a--llog(N,+2)

s
(l-3 u) (2stz--1) + 4p
< Cn logn.
This is plainly of the form required if p is chosen sufficiently small.
The other sums which appear in (5.35) and (5.36) 3.7) present no
difficulty. The first sum on the right hand side of (5. 35) is disposed of
by the trivial Lemma 13 ; and the second, and likewise the corresponding
sum in (5. 36), by Lemma 15, which is still. available. We require only
that Q > Y on every n-t,, and this will certainly be so if /? < a.
(iii) J1 is disposed of by (6.121) or (6.122). Here we use the
latter inequality and observe that it is true a forliori for arcs 9X?.
Thus each of the terms J1, J,, Ja, q5 is of the form required. This
could naturally have been proved iore simply, had we been working from
the beginning in terms of the dissection we are using now.
8. 2. It remains to prove that

(a. 21)

It is here only that we use Hvpothesis

*r X.
If 18< a, as we have supposed, and x is on an 1st,, then either (i) it
is on an m, or (ii) it is on an 9X, for wlaich Q > Y. In case (i)
(8.22) j f) < BEna*+& < CTP-~
-__L_
a7) With %R,, m, for YR,, m,.

502
 Some problems of ‘Partitio Numerorum’ (VI),

on fixing E; and in case (ii)

on fixing E once more.

This completes the proof of Theorem 6. Theorems 6 and 7 require
no further proof, the first following from S as 2 followed from 1, and 7
being merely a restatement.
8.3. In proving Theorem 8, we revert to the method of P. N. I.
We have

(8.31) r,(n) = g-i1Jz >qpix

f”
s

say. We prove first that

I
(8.33) ISj
Taking S,II first, we have

]X,i <AxJ/fjsjdOI < C~‘e-2E”a-c’~~jflaL/d~/

m2 m2
by (8.22) and (8.23). Arguing as in 5 8.2, we find
1S, j < CEn8a-1-cC+E
< Cnsa-l-c.

This disposes of S2. That S, is of the form required is obvious;

it is in fact less than C (as in Lemma 13). As regards Sz, we have

if”-qq <AsIf- VI”+A,If- d ivy

r= cEnsax+s + cEnaxSE 1 y I-ab-1)

503
 3ti G. I-I. Hardy and J. E:. Littlewood.

(since every !I& is part of an !!I&). The first term here contributes less than

when E and cc are chosen appropriately. Alld the second term contributes
less than

This completes the proof of (8. ~3).

3.4. The proof of Theorem 8 is thus reduced to a proof that

We write

7 being, as in P. N. 1, the arc of r complementary to 6 (or !%!, I*

On 7, 101 >F$-, and so
2

since ea > 2 and a is a c.

Finally

S6 z=z (g fis=-l 2 (~-~)3e,(-np)=n”.-‘(~(n)-e,(n))

p,qsna

(the Y of Gz (n) being now v). Hence, by Lemma 3 (3.313), (8.42),

and (8.49,
s 6 C nga-1 G (n) --i- 0 (@a-l--c) + 0 (n8a-l-cl(85-9 = es (n) + 0 @pa-l-c);

which cordpletes the proof.

504
 Some problems of ‘Partitio Numerorum’ (VI). 37
t

8.5. Theorem 9 is an obvious corollary of Theorem 8. We conclude

the memoir by showing that when, as is usually the case‘j8), r(k) < k + 1,
Theorem 9 may be deduced in an elementary manner from Theorem 6.
We know that
$r,,,(m) - Bn.

Hence, if Hypothesis K is true, there are more than B,d-& numbers

less than n and representable by k k- th powerP). If n, is such a number
fess than n, and n is not representable by 2 k + 1 powers, then none of
the numbers n - n, is representable by k -+ 1. That this should happen
for an infinity of values of n would plainly contradict Theorem 6.

See P. N. 4, 184, for an analysis of the exceptional cases.

In fact, in the notation of 8 7, l;lS= 1 (0 = k).

( Eingegangen am 18. Juni 1924.)

CORRECTIONS
In the last displayed formula but two on p. 32, read vv,l-z’(n) on the right.
p. 10. Jn the second line of the proof of Lemma 2, read s = 4 in place of k = 4.

505
 518 G. H. HARDY aad J. E. LITTLEWOOD [Feb. 10

SOME PROBLEMS OF “PARl’ITIO NUMERORTTM” (VIII) :

THE NUMBER I’(k) IN WARING’S PROBLEM

By G. EL HARDY and J. E. LITTLEWOOD.

[Received 28 April, 1927, -Read 10 February, 1927.3

1. Introduction.

1 , 1. In this memoir we continue our researches concerning the

number F(k), some of whose properties were developed in the fourth and
sixth memoirs of the series t l

In 5 2 we give a formal prdof of what we have stated before: concern-

ing the arithmetical interpretation of F(k), viz. that F(k) is (except when
k = 4) the- least humber s such that every arithmetical progression con-
tains an infinity of numbers which are sums of at most s positive k-th
powers.
In 5 5 3-5 we are concerned with upper bounds for F( FE). Since each
of k+l and I’(k) is a lower bound for G(k), it is of interest to determine
completely all cases in which I’(k) > k+ 1. Our earlier researches left
little doubt which these cases are, but an exact proof that F(k) < k+l

t G. Ix. Hardy and J. E. Littlewood, “ Some problems of ‘ Partitio Numerorum ‘: (IV) :

The singular series in Waring’s Problem and the value of the number G(74 “, Math. Zeit-
shrift, 12 (1922), 161488 : and (VI) : “ Further researches in Waring’s Problem “, itlid.,
23 (19251, I-37. We refer to theac memoirs as P,‘N, 4, P.N. 6, Some errors in P.N. 4 are
corrected in P.N. 6, but the main results of P.N. 6 will not be required here.
We assume that the reader is acquainted with the ideas and notation of P.N. 4, but, for
the sake of clearness, we summarise the principal results, so far as they are relevant to our
present purpose, in ,lj$ 1.2-l. 3.
We may add that I‘ P.N. 7 “, which is still unpublished, contains an application of our
a&hods to the problem of the order of magnitude of the difference between successive primes.
We prove (subject to our generalized form of the Riemann hypothesis) that

1928, 4 (with J. E. Littlewood) Proceedings of the Londorn Mathe-

506 madica2 Society, (2) 28, 518-42.
1927.] THE NUMBER r(k) IX WARING’S PROBLEM. 519

in all others was lacking ; and this we now supply, the key to the proof
being contained in our present Theorem 3 (Q 4).
In the cases which were formerly the main obstacle, however, we
now find that we can prove I’(F). < k as easily as l’(k) < k+l. We
are naturally led to attempt to determine all cases in which I’(k) > k,
and this, we find, involves the solution of a further problem, to which
the key is <Theorem 2 (5 3). This theorem, like Theorem 3, is difficult ;
but the trouble involved in proving it is not wasted, partly because the
theorem has a certain intrinsic interest, and partly because the lemmas
on which it depends are important in the actual calculation of r(k) for
particular values of k.
In $ 5 we prove Theorem 4, which shows that l?(k) < k except in five
standard cases, in each of which r(k) is determined explicitly. To these
five cases we add a sixth, in which we can prove that r(k) = k.
Finally, $ 6 is concerned with questions of numerical calculation.

1 . 2. It is convenient to insert here a brief summary of the notations

and conclusions of P.N. 4, so far as these are directly relevant. It is
assumed there that k > 2, and we shall suppose this here also, except
where the contrary is explicitly indicated.
The singular series is

S = S(k, s, n) = 5 A,!
q=l

A 1 = 1, (* > I),
where A, = A#, s, n>

S,,, being a generalized Gaussian sum and p running through values less
than Q and prime to 4. Here A,,+ = A&g if (4, q’) = 1, so that

where m is a prime and

X= = 1+&-)-A++... = EAh,.
A
WW series and product- are absolutely convergent for s > 42 (except for
k- 2). Given k and 25, we define 8 and + by

=elk§P @= 0+1 (m > a), q5= e-t2 b= a,

t In PX, 4 necessities of printing compelled US to write - ;t~ T.

$ PX 4, 176 (Theorem 3).
g z[ymeans c4 x is a divisor of y ” ; x”ly means “ xv is the highest power of x which
dividesy ’ ‘.

507
 52.0 G. H. HARDY and J. E. LITTLEWOOD [Feb. 10,

and we sometimes write p for a? We have + < k, except t when

k = 4, zr = 2.
We denote by
NI I+, v) = N(m”, v, s)

the number of primitive solutions of

i.e. the number of solutions for which 0 \( xp < sx (r < s) and not every
xr is divisible by m, If (&)@I 1~1then8

where B = 0 if ,B= 0 and B > 0 if p > 0. Thus in all cases X= > 0.

We denote by y& = &(k) the least number ‘yn such that

fors>mandalln,andby ym = r,(k) the least number m such fhat

xw >, h, = h(k, s, m) > 0

for s 2 ‘yy1and all n. Then & = yV except when k = 4, m = 2 (when

y; = 15 and ya = 16)l). Our number r(k) is defined by

r(k) = Max y-.

w
In order that X= > & > 0 for all n, i.e. that s 2 ym, it is sufficient
that xw > h, > C for all suftjciently large TZ. The necessary and
sufficient condition for this is that

for all n71, in which case N(&, n) > 0 for h > $t f.

1 I 3. Much of our analysis depends on a particular division into

classes of the residues fo modulus p- We suppose that m > 2, SQ that
__c__AI--f_____ _--_- -----_. ----?.---____lt

+ P.N. 4,178.
$ Not wq7L
§ P.N. 4, 166, Theorem 2. We do not ueed here the rtcfurtl vslus of B when.6 > 0.
11 P.N. 4, 178 (Lemma 19).
V P.N. 4, 178 (Lemma 19).
1 f This is & particular consequence of P.N, 4, Lemma 5,

508
1927.1 THE XMBER 1"(k) IN WARING'S PROBLEM. I 521

(f) = O+l. We write-t-

k-.- d k* = &ko,
where t? = (k-*, a-1) = (a-@k, a-l);

ana m-1
a--- E ’
~0 that (ko, d) = 11. Finally we write

d-1 d-1
c = c, = c,(k) = -=-I e+1= d +1,

and we divide the residues into c classes Co, Cl, . . , Cc-1 as follows& l

If G is a primitUive root of p, and % is prime to P, we have

7+,G G%h+e (mod p),
where
m. has one of the vrtlues 0, 1, . . , d-l, and e one of the dues 0, 1, . . . .
l

qO-l. The d values of no which have a common e we class together in

the class
c: (8 = 0, 1, l -.* +0-u,

and we denote a typical member of the class by a:, These & classes
contain all residues prime to p.
Next, if 1~~ is a residue for which ~~1 mi, where 0 < i < +, we can
write
32;s m’Ni E a’(j!“iqi+’ (mod p)t

ryci runs from 0 ta d-l, and e from 0 to &-1. We thus define +

new classes
cg (e = 0, 1, . . . . &-1)
of members at. Finally, the residue 0 is the sole member of SI class
C+0* The total number of classes is c, and we may denote them, in the

t We did not UWI the Bymbol k* in P.N. 4.

$ To fix ideas, we give the examplas :
k = 34, w = 3, 8 = 0, + = I, p = 3, k" = 34, E = 2, d = I, ko= 17;
k = 34, v = 17, 8=1, q=2, p=289, k" =2, E= 2, d = 8, k,, = 1 ;
k = 34, w =103, 8 = 0, tp = 1, p ==103, k* = 34: E = 34, d = 3, k" = 1;
34 = 30. 2.17 =17'1 2 .I =103O. 34 *I.
fi A misprint, Cc for Cc-l, runs through # 5.4-5.6 of P.N.4.

509
 522 G. H. HARDY a,nd J. F,. hTTLEWOOD [Feb. 10,

order in which they have been defined, by Co, Cl, . . ., CE--, and typical
members by Q, ‘al, .. ., ‘C-1,

The class Co, which is particularly important, consists of the residues

of k-th powers prime to m. In particular, 1 is an a~. Any two Q’S are
incongruent (mod p), and therefore incongruent (mod m). The product
of an au and an a, is an ar, and aoa,, where CI)-is given, may be identified
with any Q~by choice of %t.
One additional remark will be useful, Throughout P.N. 4 we sup-
posed that k > 2. But our analysis of the classes does not depend on
this assumption, and is equally valid when k = 1 or k = 2. Thus when
k = 1 we have 8 = 0, # = 1, 8 -1, d = a-1, c = 2. There are two
classes, Co consisting of all residues save 0, and C, consisting of 0 alone.
It is, in fact, obvious that all residues save 0 are representable by a single
first power to m, and 0 by two.

I . 4. We shall require, besides the ideas summarized in 5 5 1.2-1.3,

a number of propositions which we state here as lemmas. These lemmas
are either contained substantially in P.N. 4, or are easy developments
of propositions proved there.

LEMMA 1. If d > 1 uje have

Iho I_ 0 (mod p),

the summation ext8nding over every residue of Co.

since G+o is an Q different from 1 and therefore: incongruent to 1.

(mod m).

-b For all this see P,N. 4, 180-181. The point of the incongruence of the &s iti not made
explicitly. To establish it we observe that if G is a primitive root (mod p) it is a primitive
root (mod w). Hence from
GmO*oE @“‘+O tmod w)
we deduce successively
(mo-m;)qo ii 0 (mod m-l),
mo-m; s 0 (mod d),

510
1927.1 THE NUMBER r(k) IN WAR.ING’SPROBLEM. 528

LEITMA 3. Every residue (mod p), except 0, is representable (mod p)

as the sum of at w2ost c-l U& rind every residuJe by c. Tbis is true
for k>2.
When k > 2 this is contained in P.N. 4, Lemma 21, and the result
is valid for k = 2 (see the end of 5 1 . 3).
LEMMA 4. If 1 c ad p(d) is t72.e number of classes repre-

sentable by at t
‘YPLUS then
(1 l 41) &+I) > mill {p(c’>+l, c1.t
If (I . 41) is fake for cl, then

a’+ 1) = P(C’) < c.

Suppose that C* is a typical class of tha &?> classes, and that C, is a
c* . If all is any member of &, ar--+- 1 must belong to a C*, since no new
classes are representable by c’+l aok.
Similarly (cL~+ I)+ 1 = 0[,+2, aT+3, .,., and so all residues, belong to
some C*, which contradicts ,u(c’) < c.

LEMMA 5. If dl ) d md d, > I, then the residue 0 is representable

(mod p) by dl QO’S-
This is proved in P.N. 4, in the cowse of the proof of Lemma 21.
LEMMA 6. If m > 2 and k = k/ko, then
yww < 3 (k < a),
yw(k) = ywo (k > 2).
(i) Suppose, first, that k = dc > 2. Then $j C, d, and I/Q are the
same for k and k, and the classes Co for k and k, are identical since
each consists of the residues
G’GJlo (7~2~ = 0, 1, . . . . d-l);
which proves the second resnl tI, (The proof is valid for =rll k.)

t Compare P,N. 4, 183, where we prove that

v(c’+l) &nin (y(c’)+l, c-11,
+) beingthe number of representable cltbssesexclusive of Ccm2(the class consisfing of 0 only).

511
 (ii) Next suppose that k < 2. Then plainly 8 = 0, + = 1, and we
have to prove that any residue (mod XT) is representable by at most three
k-th powers. We have pointed out already (in 0 1 . 4) that, when
k = 1, two are sufficient. When k = 2 we must have 8 = 0, + = .I,
8 = 2, and c = 3, so that our conclusion follows from Lemma 3.

0i m = 2, 8 = 0, then ya = 2 ;
(ii) m = 2, 0 > 0, then ‘ye = a’+‘;
(iii) ‘cir > 2, E = a-l, then ym = zP1 ;

(iv) ‘GJ> 2, E = *(m-l), then yw = $(m’+‘- l), except in the

case a=3, 8--O (when y-=2).
See P.N, 4, Lemma 22. The exceptional case in (iv) is there omitted
by inadvertence. It aIrises from the failure of 2 < c- 1, implicitlJr
assumed, and true in general.

LEMMA 8. If
(i) k = 2@, 0 > 1, then r(k) = 2e+2;
(ii> k = 2@3, 0 > 1, then F(k) = 284g:
(iii) k = d(a-1) a> 2, 8 > 0, fhen r(k) = a%+1;
(iv) k = &&(a-l), m > 2, 0 > 0, then I’(k) = #(m’+l-l).
See P.N. 4, Theorems 6, 7, 9, 10.

LEMMA 9. If 6 = (a- 1, k) and

then I&,wI \( @-1)l/~*

See P.N.4, Lemma 13. If 6 = I the sum vanishes.

LEMMA HI?. If x,(p) is a chamcter (moda), K’ a K for dkh sf is

did&i& wit71 the principal charncter xl, and

-- _--- ------. ----_II___------- --P.--,+--.-+.----_-___L

t ThexKwill not beoonfused

with the xw of QI, whichdonot occuragain.

512
 525

These propositions also are included in the statement or proof of

P.N. 4, I-mnma Ii-L The nnmber of the K’ is E--I (EN. IV, Lemma 7).

2. A Tit hm2etica.l ~sograssions.

2 . 1. Xn this section we prove
THEOREM 1. TIM n.ecesswy and su#%cient condition that every aritb
ww!ical progression should corztnin an i@nity of numbers wh’t’c71~
a,re sums
of s posl’t.ive k-t h pouxrs is
s > r(k),
except tidier2 k = 4, in ,which case it is 5 2 15.

The materials for the proof aIre contained in P.N. 4. We suppose,

first, that k # 4.
(i) Suppose that ts2 r(k). Since 5 2 y=, the congruence

is soluble to any modulus d (for h 2 $ and therefore for all A), and
therefore to any modulus Q. If x1, x2, .* , x$ is a solution, then
l

for every 972, so that the progression NQ-+n contains an infinity of

numbers of the form required.
(ii) Suppose thalt every arithmetical progression contains numbers
representable by s k-th powers (and that k # 4). Then for any ZF, n there
are numbers &IZ+~Z representable by s AAh powers, and the congruence

is soluble. For each of the values n = 1, 2, .. . , ,9 the solution is

necessarily a primitive one, since otherwise n would be divisible by
d, and so by a++‘. A fortiori there is a primitive solution of the
congruence

for each of these values of 12,and since now n ranges through all residues

513
 526 G. H. HARDY and J. E. LI’M’LEWUUD [Feb. 10,

of the modulus we have lV(p, n) >r 0, and so s 2 & = y-. Since this
ie true fur all w, we have s 2 r(k).
The argument in (i) fails for k = 4 because yW and yW differ’ when
w = 2. We have, in fact, y2 = 16, y; = 15, while

for every other m. It is y& which is relevant in the argument (i) here,
and every progression is permissible when s 2 X5. On the other hand,
consider the progression 16m+ 15, No number of this is representable
by fewer than 15 fourth powers, and s 2 15 is necessary.

3. A theorem cone the case 6 = 0.

3 . 1. Our main theorem in this section is
THEOREM 2. If k>:2, t9=0, d>I, then

The principal difficulty lies in the representation of the residue 0

(mod a), aid the essential lemma is
LEMMA 11. If m > 2, 8 = 0, then the residue 0 hcts a representation
by at most s IQ’S wheneaer

If =l, a==-1, c= 2, and 0 is representable by 2 u/s. If

G= 2, “d = $(a-l), c = 3, and it is representable by 3. These cases (in
which the main inequality is nugatory) are therefore trivial, and we may
suppose that 6 > 2.
Let N+l be the number of solutions of

x = i xf E 0 (mod a)
r=l
(0 \( xT < a),

One solution is given by xLtr_C0 for all T, and so what we have to prove is
that N > 0. Now Ze,(pX) has the value a-l if X E 0 and the value
-1 if X +O. Hen&
w-x
ZS ;,Tu = c 2 e,tpm
P p xr=o

= tN+l)(a-1)-(d-N-1) = (N+l)p-ma,
(3.11)

514
 TEE NUMBER r(k) IN WARTMO'S PROEUJSM. c
627
1997 .]

NOW 8 = (a-l, k) = q
and so, by Lemma 9,

= d cE yk (m- l)+ c (-1).

xk g Yk

Given II: + 0, there are just 6 y’s for which .yk _c zk. Hence

(3.13) zpp,,I” = { l-+(a-l)}(a-l)- {m~-1-E(m-l)}

= (e-l)a(a-1).
From ( 3.12) and (3 . 13) we deduce
(3.14) 2p*,,p< ‘((v-l)~~}~--2q &/a = (e-l)‘-fCJqm-l)*
It then follows from (3 . 11) that N > 0 whenever

d -73 > (e-1)s-1 &(a-l),

and a fortiori when &-1 >, (e- l)s-l &, the inequality of the lemma.

3 . 2. There is a companion result for residues other than 0. This is

not required in the proof of any of our main theorems, but it is essential
in the calculation of special values of F(k), and has a certain independent
interest.
LEMMA 12. Suppose that t3= 0 and n * 0 (mod m)m Then n 72~s Q

represetltation by at most s m’s whenever

s> I, m > (e- l)W- ‘1.
As in Lemma 11, we may suppose G> 2. V\/‘e have to prove that
N ‘> 0, A7 being now the number of solutions of,
X-n (moda), O<g,<a.
Now
(8.21) ~s;,,Gk-~P) = ZI =riilew {p(X-7t))
P p rt=O
= c Ze,{p(X-n)}+ c r:e,{ptX-41
x=+&p x$&n P

z N(a-1)-(&-N) = Na-d.

515
 528 G. I'E. HARDY and 3. E. LITTLEWOOD [Feb. 10,

But, by Lemma 10,

SP,W = ~vx,.(p).
Hence
lb-d = CC* TX(p) d-v> = ~*T~xlp) e,(-np>,
P P
where T is the product of s T’S, X(p) the product of s x’s (and m a
character), and the numb,er of terms in 2 * is (a- 1)“. The inner sum
is - 1 if X (p) is the principal character, since n qk 0, and it ia the product
of B x and a T otherwise; so that, by Lemma 10, its absolute value does
not exceed l/a. Hence

IN a-d 1 < (e- 1)” ds+l),

and N> 0 if d> (e-1)” d8+l), the inequality of the lemma.

Since (~-1)~ > s(s-2), we have as a corollary

LEMMA 13. The result of Lemma 11 is true also for residues other
than 0.

We shall also require the trivial

LEMMA 14. If x is integral and not less than 4, then

(X- 1)2(X--f)lW) < x(2x - 1).

3. 3. Proof of Theorem 2. Since 23 2 d+l> 2, we have + = 1,

C= e+l. We begin by showing that we can make certain simplifying
assumptions.
First, we may suppose 8 > 2. For k = &, k = k/b = c Hence,
if 6 < 2, we have, by Lemma 6,

Next, we may suppose ko = 1, so that k = 6. For suppose the theorem

proved in this case. Then, since k = 6>2, we have, by Lemma 6,

Next, it follows from Lemma 3 that every residue (mod a), except
perhaps 0, is representable by at most c-l = g = k a’s, and we have
only to show that 0 is so representable.
Finally, it follows from Lemma 5 that 0 is 80 representable when-
ever d is even, and, in particular, whenever 8 is odd ; since we may then
take d, = 2. We mayI therefore suppose that B is even and d is odd.

516
1997 *] THE NUMBER r(kj IN WARING’S PROBLEM. 529

To sum up, if the theorem i .s false for some k and lt~f, then a k and =
(not necessarily the same) exist for wh ich
(A) m > 2, 8 = 0, k, = 1, k = E, c = e+l ;
(B) E is even, E > 4, d is odd ;
(C) th.e r&&e 0 is not yepresmtable (mod Icr> by fewer thaft -i-l a~‘#=

From these assumptions we deduce successively (D), (E), (F), (G)

below. Since (G) contradicts (D), the theorem is then proved.

3 . 41. (D) ‘GJ< 42+1), d < G-1 t l

Take s = g in Lemma Il. The result contradicts (C) unless

m < (e- l)a(r-nw~* Since e4, by (B) * 7 Lemma 14 giveP
JZ < &k-l) ; and then
d Z (a-1)/e < 2e-1.

3. 42. (E) There is an cxo,SUY a& such tkat $+I is &O an a~.
First, the number of classes representable by two or fewer -Q’S is
two at must; for otherwise we should have, by Lemma 4,
p(c-1) > min {p(2)+c-3, c} = c,
and all residues would be representable by c-l = 6 q’s, contrary to (C).
Next, if no a: exists, no aO+l is an a~, and, in particular, l+l = 2
is not an m, 80 that every *c~+l belongs to the same class as 2. It follows
that ho+1 = 2ah for every a0 and some a& Incongruent m’s correspond
to incongruent ~6’s and U; runs through the whole of Co with a~. Hence,
by Lemma. 1,
Z(q,+l)E2Eao, d~~u,~O;
and this is impossible because dJ m- 1.

3. 43. (F) ar gk 1 (mod a).

If qf E 1, 2 is an q, bv (E) ; and so 1, 2, 22, 23, .*. are all Q’S, Now
any number up to 2f+1 -i can be represented in the binary scale by at
most e non-zero digits ; and, after (D),
m < e(2e-11) < 2’+‘-2
(since G 2 4). Hence

snd zip G 0 is representable by at most g Q’ s, contrary to (C),

f To avoid any misunderktanding, we repeat (D)-(G) are not trw theorems ) but

deductions from hyptheres which prove untenable

517
 530 G. H. HAEbDY und J.n E. fZlITTLEWOOD [Feb. 10,

3.43. (G) d>%-I.

Tske a = ar in Lemma 2. Then a and l+a are Q’S, 80 that a”(1 +a)
is an m for every rrt. Since d is odd, Lemma 2 gives
0 I_ l+a-f=-a”+...+ad-~ = (l+a)+(l+a) a2+...+(1+a)ad-‘+ad--I.
This is a representation of 0 by &(&3)+2 = &(d+l) Q&. Hence,
by (CL
*(d+l) > E, d > 2E-1.
Finally, (G) contradicts (D) Hence (A), (B),
l (C) are not
simultaneouslv nossible : and this proves the theorem.

4. A theorem concerning the case B> 0.

4.1. THEOREM 3. If 8>0, d>2, then

ym < k*
The main difficulties here arise with residues (mod p) other4han 0.
We denote by N[m], where m > 0, any number that is a sum of
m kmth powers [positive or zero). The symbol is not a one-valued function
0f m; m (together with k) determines a certain class of integers, and the
symbol denotes an undetermined member of this class. An equation in
N’s implies that any number that is of the form on the left of the equa-
tion is also of the form on the right.

4.2. LEMMA 15. N[m]N[n] = N[mn].

For N[m]N[n] = i=l

21zf j=l
iz yj = i,j
c (S#j)“i
LEMMA 16. For any n
n E N[c] (mod a).
This is trivial if n s 0 (mod a), since 0 is the sum of f 0%. If
qt + 0 (mod ;IJ), let k’ = a, ‘and let c’, ah denote the number c and a
typical aa for the index k’. Then, by Lemma 3, n can be expressed as a
sum of at most c’-1 = g $8. Now an Q taken (mod a) is plainly an
ai ; moreover, since the &s are incongruent (mod a) ( 5 1 .3), and since
the number d is the same for k and 8 (for the prime ml, it follows that
the set of 4 ip1 identical with the set of a0 considered to mod pr. Hence
n is representnble (mod a) by not more than 8 c&s.

4.3. We come rzow to the lemma containing the crucisl step in the
proof l

518
 LEMMA 17 t . If d > 2 then there is 5 u, not. a multiple of a, such
that
urn E N[a--11 (mod 7~~).

Since d 2 3, there exist three k-th powers, P, & ,@ prime to m and

incongruent (mod m ) We define two positive integers a, b by
l

a E a’ (mod d) (0 < a < d),

b E pk (mod ma) (0 < b < aa).

Since neither a nor b is congruent to 0 or to 1 (mod a), there exist
integers p, v such that

(Y-l)is+2 < b < VIZ-1 (0 < v < 4.

Then Pm = a-j-m C a”+m (mod a2),

where l<m\<a-2.

Thus pm is representable (mod a2) as & sum of l+m < m-1 k-th powers,
and similarly for yt3. The result of the lemma therefore holds (with
U =p or u = V) unless p = v = ‘15, Now in this case, since a and b
cannot both be m2 -1, one at least, say 5, must satisfy I

so that d = a+~, where 2 < m < m-2. Then, h denoting any

positive integer,

-m= -hm+(hm-m) I_ hd+(hm-zr) (mod zz2).

In this we take 3t = h,,, the least integer not less than a/~. Then -m
is expressible (mod a2) as a sum of

=’ h,-+-(l~,m-a)
s??k

k-fh powem, and the result of the lelnma (with 24= -1) will certainly
hold if it is true that S, < a-l for every WI in 2 ‘< m < a-2. Now
--- .- --. _.. - . ^_ --- -. --- ------ -
t In our original version the form of this lemma, was slightly different. The present
form, which leads rather more simply to Theorem 3, was suggested to us by Mr. Ingham.
We ace indebted to him also for the proof given here of the l&ma ; this represents a great
gain in simplicity over the original. We have finally to thank Mr. Ingham also for detecting
& number of minur errors in the paper as a whole,

519
 532 G. H. HARDY and J. F,. LITTLEWOOD [Feb. 10,

?72(L- 1) < m-1. Hence

If now *(a-l) < ~72\( m-2, then sm < a--2+2 = m. If on the other
hand 2 \( 77~< 4(23-l), then sm \( +(a-I)+&(=-1) = a-l. In
either case sm \( m- 1, snd the proof of the lemma is completed.
~LEMMA 18. For any integer n
32m E N [~(a- I)] (mod ~2).
We choose II so that uv _I VL(mod n), u being the 2c of Lemma 17.
Then

nt;r~v.umG N[ E1l 2L’z3 (modm2) (Lenlma 16)

s N[e]N[a-I] I, N[&-I)] (mod a’)

(Lemmas 17, 15)

4 . 4. P7taof of Thkorem 3. For any ‘YLwe have

92 = N[~]+vz,a (Lemma 16)
= N[e]+N[+-l)]+%a’ (Lemma 18)
= N[~]+N[4’~~-1)]+~3N[e(a-l)]-+-~71,~~ (Lemma 18)
- .******m*

= N[e]+N[e(a-l)]+aN[e(a-l)]+...+aB-’N[e(a-l)]+n,+,ab+~
G ~[e+e(~--l)+~(~~-a>+...+e(a~-‘t38-~)] (mod p)
s N [~a~] (mod p) l

If 72 + 0 (mod p) the em8 k-th powers by which 12 is thus represented

(mod p) cannot all be divisible by m, since + < k. Hence any TZ qk 0
is representable (mod p) as a sum of at most E& a& The same thing
is true of the residue 0, by Lemma 5 (since 1 < d < &e). Hence

5. Determination of all CUSPSin which F(k) > k

5 , 1. We use r, like m, to denote an odd prime ; there is no danger
of confusion with the. common notationt . We distinguish six special

t “t 8 are giV8ll by k ; w is a prime which may assume any value when k is given, and B
is determ ,ined by k and W.

520
1927.1 THE NEWER I’(k) IN WARING’S PROBLEMS 533

classes of vallues of k, viz. :

I. k = 2’ (9 > 1).
II. k = 2%3 (9 > 1).
III. k = 2(x-I) (4 > 0).
IV. k = ~Tyr-1) (9 > 0).
v. k=7r -1, 1: not belonging to my preceding r:laas.
VI. k = *<T- l), k not belonginq I to any preceding class.
It is easily verified that these classes are mutually exclusive.
THEOREM 4. The values ofI I’(k) u;hen k belongs to a special class
are given by the formulae :

I. r(k) = yz = 23+t = 4k > k+l.

II. r(k)=Y~=P+“=$k>k+l.
III, I’(k) = ya = ccr3+l > k+l.
SV. r(k) = yr = *(r”‘“- 1) > k+ 1, except in tJw particular
case k z 3 z J- . 3 (3-l), when I?(k) = k+l.
7PT. r(k) = y7F = r = k+l.
VI. r(k) = yn = i&--l) = k.
W3zm k does not belong to a special class, then

The values of I’(k) in cases I-IV are given in Lemma 8, We suppose

then that k&es not belong t.0 any of these classes.
Let m be any prime. We shall prove that ypy < k, except when k
belongs to class V and m = ST. There are -four cases to consider :
0i 23 =2,0=0. Herey,=y2=2<k,byLemma7.
(ii) zis = 2, 0 > 0. Here G = 1, and ko is odd and greater than 3,
since otherwise k would belong to class I, class II, or class III. Hence,
by Lemma 7,
YW = y2 = Pt2 = ak/k, < k.
(iii) m> 2, 8 = 0. If &>l, then

by Lemma tj* If kO = 1, d> 1, then y,< k by Theorem 2. If b = 1,

d’= 1, then 6 = m -1, k = a-1; and this is case V, with r = m,

521
 534 G. IE. HARDY rind J. ‘E. LITTLEWOOD [Feb. 10,

(iv) ‘~if> 2, 8 > 0. If k0 > 1, then, by Lemma 3,

&+1-l
yTu\(c= -“+1 < 2ae,+l\( d&,+1 = k-j-1,

and so y- < k’. If & = I, then d > 2, since d = 1 or 2 makes k belong

to class III or class IV. Hence rmk k, by Theorem 3.
It remains only to observe that. ylr= n = k+l in case V, and
yr = &(n-1) = 3c in case VI, by Lemma 7. With this the proof is
completed.
6. The numerical calculation of r(k).
6 , 1. Table I gives the values of r(k) for k < 36, the special class, if
any, to which k belongs, and (under the heading a) the least w (in
case more than one exists) for which ym = r(k).
The value of r(k) may be found for any k by a finite process of cal-
culation. A routine method may be set out as follows.
We observe first that r(k) > 3. Since r&k’) 2 r(k) and
r(2”) = 2>+” > 8 when 9 2 2, 1‘t is necessary to prove this only when
k is an odd prime r. In this case there are ‘IT- 1 n-th powers (mod G),
at most (~-l)~+fi-l < G numbers are representable by one or two of
them, and so y,(r) 2 3.
We cannot prove that r(k) 2 4, though no case of l?(k) = 3 is known.
Lemmas 11 and . X3 show that y&. 3 provided that
w > (k 0 1)2(3--1)t@--2)
= (km I)+
The number y2 is given for any k by Lemma 7, and, in particular,
we know that for an even k > 3 we have y2 2 8, r(k)+ 2 8. If now no
(odd) zz exists for which yla> Max (~2, 3), then I’(k) = Max (yz, ‘3).
If, however, such m’s exist, they must satisfy 23r< (k-l)‘, they can
be found by a finite search, and r(k) is equal to the greatest of the
yp corresponding.
A prime ‘t5 giving in this way F(k) = T,> Max (yz, 3) must further
satisfy either (i) m 1k, or (ii) m G 1 (mod 8j, where 8 1k and 6 is not
unity or a positive power of 3. For if such a m satisfies neither (i) nor
(ii) we must have 8 = 0 and G= 2”, where a 2 0, in which case

In c&w (ii) m belogs to the progression &+l ; it is fairly certain in

practice that it will be the Zeust prime in its progression.
An examination of the prime divisors of k and of the least primes in
&he progressions ms+l generally suggests fairly strongly some numerical

,t So that the coaditiun 8 = 0 for the applicability of Lemmas 11 sad 18 is satisfied.

522
1927.-J THE XUMBER P(k) IN WARING’S PROBLEM. 535

value s for P(k). We have then to prove, first, that ym = s for the
s&c&d L~J, and then that yla < s both for every prime divisor of k and
for every m, in each of the progressions,
l
for which
‘G3 < (E- p-qKs-2~~

If k is odd, however, we need consider only, in the progressions, primes

satisfring
Y the more favourable inequality
25 \< (p- l)W-11,
For in this case the aok are equal and. opposite in pairs, and 0 is repre-
sentable by two of them, so that we may use Lemma 12 instead of
Lemma 13.
It is easy to compute r(k) with practicd certainty for values of k
considerably larger than 36 ; but the necessity of examining primes of the
order k4 makes a complete proof impracticable for those k dose F(k)
happens to be small.
6,2. An example : k = 34. We illustrate the process by proving
that r(34) = 10. Here 72 = 8, by Lemma 7. It is improbable. that
r(k) is so small as this, and fhe most likely values for the critical m
are 17 and 103, the smallest prime in the progression 17m+l, the only
progression that has to be considered. Since 103 < 172 = 289, we begin
by examining m = 103.
We denote bv Y, a typical number, other than 0, representable
(mod 103) by just’ T Q’S ; v1 is an ‘e. Since d = 3, 0 is representable by
three I&. Calculation gives
vl 1, 46, 56.
vq 2, 9, 47, 57, 92, 102.
vs 3, 10, 35, 45, 48, 55, 58, 65, 93.
v4 4, 8, 11, 18, 36, 49, 59, 66, 81, 91, 94, 101.
u6 5, 12, 19, 24, 34, 37, 44, 50, 54, 6Q, 64, 67, 74, 82, 90, 95.
p6 6, 7, 13, 17, 20, 25, 27, 38, 51, 61, ,68, 70, 75, 80, 83 96 100.
v7 14, 21, 23, 26, 28, 33, 39, 43, 52, 53, 62, 63, 69, 71, 73, 76, 84, 97.
v8 15, 16, 22, 29, 40, 72, 77, 79, 85, 89, 98, 99.
vs 30, 32, 41, 42, 78, 86.
YlfJ 31, 87, 88,
These numbers exhaust the residues : hence ~103= 10.
It remains to verify that y= < 10 for the remaining primes 34m+l,
and that y17 < 10. Taking m = 34m.+l, and s = 10 in Lemmas 11
and 12, we have to verify that any residue (mod a), other than 0, ig

523
 representable by fen *Q’S when
(6.21) m \( (34- 1)2*10’(10-1)< 2380,
and that 0 is similarly representable in the further range
(6 . 22) 2381 \( m < (34- 1)2(10-‘)‘(zo-2) < 2614.
The values of m satisfying (6 . 22) are 2381, 2551. For these d is 70
and 75, bofh divisible by dl = 5. Hence in these cases, byI Lemma 5,
0 is representable by five *m’s.
The a’s of the range (6 . 21) are 1.37, 239, 307, 409, 443, 613, 647,
919, 953, 1021, 1123, 1259, 1327, 1361, 1429, 1531, 1667, 1871, 1973,
2143, 2347.
We have fo show that any residue to any of these moduli is repre-
sentable by at most feri 34-th powers, and calculation shows that this
is possible with a good deal to spare. With the higher moduli it is not
necessary to calculate more than about ten (~0’s.
Finally, to prove that y17 < 10, we have fo show that all residues
(mod 289) a re similarly representable ; and here again there is some-
thing to spafe.
6 . 3. It is natural to expect that r(k) tends to infinity with k. But
so far are we from being able to prove this that (as we stated in $6 . I.)
we cannot prove even
lim l?(k) > 4 ;
and these problems seem to be extremely difficult. Thus to pruve
I’(k) > 4 for all, or all large, k, or, what is the same thing, to prove
P(r) 2 4 for all, or all large, T, it is necessary to prove, for each r, that
either yW (x) 2 4 or y= (TV)2 4 for some prime m = 2mr+l. The first
dternative is improbable for a given JT (since the number of Q’S is
dI = r-1). The second would be established if we could show that, for
any large n, a prime m = 2-mn+ 1 exists with d = am <d- I+, which
is roughly the same as 2m < & Such a theorem, if true, must be very
deep, and it is not easy fo find other ways of attacking fhe problem,
6 . 4. This being so, it is worth while fo carry our numerical com-
putations beyond the limits of Table 1, contenting ourselves wifh lower
bounds for F(k) when the exact calculation becomes impracticable. In
what follows we suppose that a> 2, and we use the following con-
siderations to assign a lower bound for yW (where we cannot appeal to
my more precise result).

t For if d + dz + d3 < w there must exist some residue representsbls by one,

three an’s,

524
 We use N, Cd) to denote either tlhe number of distinct residues (mod p),
other than 0, which are representable by s or fewer Q’S, or allay upper
bound of this number. Here d has its usual meaning. If N,(d) were
the exact number of residues, the notation would be inappropriate,
since it would depend on k otherwise than through d ; but if will appear
t.h@, by taking an appropriate upper bound, we can make it depend 0x1
d and s only. Similarly we use u,(d) to denote an upper bound for the
number of residues representable by cc Q’S and not by less+. We may
plainly take

Let P,(s) be the number of partitions of s into just r of 1, 2, 3, . . ,

repetitions being allowed and order counting. Then

(6 .42) P,(s) = ( FE; ) (r < s), Py(a)= 0 (r > s):0

Let
(6 .48) (Jr(s)= i PT(t) = fi PM = (;>.
t=1 t=!l

Finally, when d > 1, let D,(d) be aln upper bound for the number of
combinations of d Q’S just T at ,a fime, order not counting, repetitions
not being allowed, and any combination which contains a set of ~10’s
whose sum is congruent to 0 (mod p) being rejectedg . Evidently
D,(cE) = 0 when T > k ; alnd this is true also for r = d, since 2% e 0, by
Lemma 1. And
.(6 .44) D,(d) < ( d, (1’ < d)-
r
A little consideration shows that the number of residues which art’
representable by just s Q’S, and in which just r different q’s appear, is
at most P,(s) D,(dj. Hence we may take

y,(d) < d$ -c-(s) a(d),

T=l

It is evident that, if N,(d) < a@- 1, then yw 2 s+l. Hence (6 . 45)

may be used to find a lower bound for y-. If d is even., however, we

t Thus if - 1 is an %, erg+ I+ (-1) r8preSWtS a0 by three an’s, but this representation is

to be rejected.
z Pr (8) is the coefficient of 2’ in (z + z9 + x3 + +..)*.
8 Without this provision the number of combinations would of course be

525
 538 G. H. HARDY nnd J. E. TATTTLEWOOD [Feb. 10,

can do better than (6 . 45). For, if Cc= 2dl, the aa’s are of the type
+
- a,, + a27-0-y + adl,

and, in estimating D (d), we can reject any combination in which a, and

-a, both occur. This leads to the inequalities
d(d-2)...(d-2Tm+2) (o < ‘r < d) D (d) =
(6 946? Dr(4 < 0 p’t \ 1) r 0 (7 > d,).
l

d(d-2)...(d-22r+z)
(6 l 47) Tl n

We use (6.45) when d = 3, and (6 , 46) when d is even and greater

than 3. It so happens that we do not find it necessary to consider odd
values of d greater than 3.
The values of N,(d), given by (6 . 45) and (6 . 46), are set auf in
Table 2. If one of them is less than p, then a/T 2 s+l.
6 . 5. Table 3 gives (1) the values of I’(k) when k belongs to a special
class, (2) lower bounds for l?(k) in other cases, for 36 < k < 200. The
special values are starred ; in a/l other cases we know, of course, that
r(k) ,(,k.
The lower bounds are found in four ways.
(i) In the first place
l?(k) 3, Max r(6).
6(k
When the lower bound is found in this way, we write k >r 6 (e.g. 45 > 9)
in the k-column, 8 being the relevant divisor of k.
(ii) If m is the least prime 3k+l or dk+l = 2&k+l, we can obtain
B lower bound for yar, and so for I’(k), by the use of Table 2 and the
argument of Q6 . 4. In this case the relevant value of d is entered in
the second column, It so happens that this d is, in fact, the least d,
odd or even, for which dk+l is prime, except in the single case k = 62,
when d = 5 and d = 6 both give primes, and d = 6 gives the better lower
bound, In a larger table, of course, it might be necessary to consider
odd d greater than 3.
(iii) In one case, k = 197, when d = 18, m = 3547, we have, in
order to show that I’(197) > 4, gone beyond the) principles of 5 6 .4 and
resorted to calculationt . The tables then show that, *for k < 200.
r(k) > 3, and that, except for 2, 3, 7, and 19, r(k) > 4.
...--.x---c -.--.P-.---.-IIA- --~---.--~cL_I-~+
t For which we are indebted to Mr. F. (3. Maunsell. The Q’S occur in equal and
opposite pairs, half of them being
1, 291, 447, 1162, 1163, 1177, 1468, 1548, 155%
Thr residue 7 requirea five of these.

526
1927
l] TBE NuMBER r(k) IN MiARINGk PROBLEM. 539
(iv) In two cases, k = 38 and k = 62, the best result is derived from
the fact that y2 = 8 (Lemma 7). Here we have written 38 = 2 .19 and
62 = 2 31 in the k-column.
l

When the actual lower bound is given by (1) or (iv), we have still
entered in the second column the appropriate d, or “ 2 b”, where b is
a lower bound of it, so that it can be verified that (ii) does not give a
better result Further refinements of the principles employed in
(ii) appear to be practically ineffective.

6 6, We have explored the possibilities

l I’(k) = 3, I’(k) = 4 in the
range 2 < k < 3000; 0 ur results are that r(k) > 3 in all cases ; and
r(k) > 4, except for k = 2, 3, 7, 19, for which it is 4, and possibly (but
very improbably) k = 1163, 1637, 1861, 1997, 2053.
We need only consider odd k, since y2 2 8 for even k. In the second
it is easy to see that we can confine ourselves to proving that only these
exceptions occur among prime k < 3000. Fur this result, together with
the inequality r(kk’) > r(k), shows that r(k) > 3 for odd composite
k < 3000 (indeed, k <. 2002), and I’ (k) > 4, unless k is divisible by two
of 3, 7, 19. Now all products of two of these numbers, except 1g2 = 361,
occur in Table 3 with a r(k) > 4. 1II the case k = 361 we consider
y19(361), when ‘~34 = 1g3, CE= 18, snd the argument of 5 6 . 4 shows that
ylg > 4, It follows that we need consider only @ze values of k.
Considering now only primes below 3000, we calculate for each k the
least d (necessarily even) for which dk+l is prime. We find
d’< 12 for 200 < k i 227 ;
d < 16 for 227 < k i 540, except for k = 227 (d = 24) and k = 457
(d = 30) ;
d < 22 for 540 < k i 1000 ;
d < 26 for 1000 < k < lN2, except for k = 1163 (d = 32) ;
d < 30 for 1302 i k < 2000, except for k = 1637 (d = 38), k = 1863
(d=40), and k-1997 (d=44);
d < 34 for 2000 i k < 2500, except for k = 2053 (d = 46) ;
d ‘< 36 for 2500 k k < 3000.
The sufficient condition for r(k) > 3 is
N,(d) < a-1 = dk,
which is
(6 n61) d’+3d+8 < 6k;

527
 540 G. H. HARDY und J. IS. LITIXEWOOD [Feb. 10,

end that for r(k) > 4 is

(6.62) d3+4da+20d+32 < 24k.
The inequality (6 . 61) is satisfied in all cases with a good deal fo
spare, and the limit 3000 could be extended considerably without exces-
sive labour, The inequality (6 , 62) is satisfied by the upper bounds of
d set out for the various ranges of k above. Thus I’(k) > 4, except in
the exceptional cases, mwhich require detailed calculation. Mr. Maunsell
has carried this out for k = 227 and k = 4*57t , but beyond ‘this point the
work becomes very heavy.

TABLE 1,

k 3 4 5 6 7 8 9 10 11 12 13 14 35 16 17 18 19
PPP
r'(k) 4 16 5 9 4 32 13 12 I1 16 6 14 15 64 6 27 4
Clam IV I VI III - I TV IV VI II - VI VI I - XII -

w 3 2 11 3 3 2 3 5 23 2 53 29 31 2 103 3 229<
V
k 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
-_f_
I'(k) 25 24 23 23 32 10 26 40 29 29 3l$ 5 128 33 IO 35 36
class / III w v VI 11 - VI IV v VI v - 1 VI - VI v
---
m 5 7 23 47 2 101 53 3 29 39 31 311 2 67 103 71 73

TABLE 2 : Na (d).
I

4 5 6 7 8 9 10 11 12 13 14 I5 16 17 18 19 20 21
-. ---1---- - _-__
3 30 45 G2 84 108 135 165 198 234 273 315 360408 459 513 570 630 693
4 40 W 82 112 144 180 220 264 3X2 364 420 480 544 612 684 760 840
6 120 al0 336 504 740 990 13202052
8 320 680 1288
10 6801682
12 1288

$ For k = 337, w = 3449, the au’6 are

I, 78, 409, 543, 603, 621, 635, 1238, 1474, 1475, 1518, 2007,
and the mm numbers with negative sign. The residue 21 require8 5.
For k = 457, w = 13711, the Q’S are
1, 702, 792, 1060, 1394, 3186, 2425, 3127, 3148, 3149, 3442, 3726, 510'7, 6167, 6543,
and the same numb&with negative sign. The residue 52 requires 5.

‘; There is a misprint, 30 for 31, in the value of r (30) given in the table of RN. 4,

528
 THE NUMRER l‘(kj IN WARING’S PROBLEM.

37
k d r

39
i k

"$1 = $Y(3--l} -
cl

121
38 =,2.39 38 #82=83-l 1 53
#33 = 79-1 39 “83 = ;(lm-lf 2 8s
#40=41-l 41 84 > 42 23 249
l 41 =+(83-l} 41 85 >17 212 3G
*42=7(7-l) 49 w6=g173-1) 2 0G
+43= :,(87-l) 2 43 8% > 29 23 ), 29
a44 =4(,89-l) 2 44 *88=89-l I SC3
45 > 9 4 313 "S9 = $(179-'1) 2 89
"46=47-l 1 47 %I0 = ;(lsl-l, 2 90
47 G 36 31 > 13 G 28
+48=2~.3 - 64 92 > 23 23 223
49 4 310 93 4 314
"50 = g(loi -1) 2 50 94 3 214
*51=;j103-11) 2 51 +95=g(191-1) 2 35
*52=53-l 1 53 ‘t’36= 3.2;) 128
%3 = 8 (107 - 1) 2 53 97 4 314
"54 = 3"(3-1) - 81 "98 = ~(197-1) 2 98
%5 = ~ll(ll-1) --.- Go W=&(l9r,-1) 2 99
#)36= g(l334j 2 56 #loo = 5"(5-1) 125
57 4 & 11 101 a0
%8= ;(117-I) 1 59 *102=103-1 103
*59 = 4 (119 - 1) 2 59 103 28
+60=61-l 1 61 "104 =;(20'3-1) 104
61 G 27 #103= ?J211-I, 105
62=2*31 ii 28 +I06 = 107 -1 107
*63=+(187-l) 2 63 107 28
"64 = 26 25G 108 > 54 281
'65- - l(131-1) 2 65 109 25
G6> II. 3 311 "110 = 11(11-l) 121
67 4 212 "111 =:(223-l) 2 111
468 =&(137-l) 2 68 *11~=1~3-1 1 113
+69=;(139-11) 2 G9 =113 =;(2!d7-I) 2 113
+70=&(141-I) 1 71 x114 =g(azo-l,I 2 II4
71 8 25 115 > 23 24 323
#72=73-l I 73 "116=;(233-1) 2 IX
73 4 &la 117 > 9 26 313
+74 = ;(l49-1) 2 74 I18 > 39‘ 23 2 59
*73 = ;(151-1) 2 75 *119=&(239-l) 2 119
76 3 312 +120=~(241-1) 2 120
77 > 11 6 311 121 1 1x 26 311
#78 = 91S(13-1) - 84 322 3 316
79 4 2 13 223 > 41 34 241
*80=~(161-1) 2 80 124 > 4 3 216

529
 THE NTSMBF,R I’(k) IN WARTNG’S PROBI,EM.

Cl r

#125 = *@ix-l, 2 163

al26 = 127-1. 1 164 > 41
127 4 *165-&(331-l)
*12a = 2: al66 = 167-l
129 > 43 34 167
"180= 131-J 1 #168 =$(337-l)
*131= $(263-l] 2 169
138 > 33 3 170 > 10
133 12 171 > 9
#134 =:(269-l) 2 *I72 = 173- 1
*135 =i(271--1) 2 #I73 = ; (34% - 1)
*136 = g7(17-1) -- “174 = ;(349-1)
1.87 6 175 > 35
WI8 = 139- 1 “176 = 177-I
139 177 > 59
#la0 =&(281-l) “178 = 179-l
+141=+(283-l) a179 =$(359-l) 2
142 “180 = Ml- 1 1
143 1 11 181 6
144 > 72 1 182.
145 > 29 “183 = &(367- 1)
*I46 =+(293-l) 184 > 4t;
*147 = 47y7-1) - 185 > 37
*Ii48 = 149 - I I #186 T;=3(373-l)
149 8 187 > Ii
*I50 = 151-l 1 188 > 4
151 6 ":89=$(379-l)
132 > 8 33 #I90 = 191-l
%53 = g (307 - 1) 2 *191-383-I
154 3 JL192 = 3.2" -
#155 = g(311-1) 2 193 4
#15G = 13(13--l) %I4 = &(389--l) 2
157 6 195 3 39 24
+15s +317-q 1 2 196 > fL8 23
*159 = #{319-l) 2 197 18
160 > 32 23 WI98 = 199-I 1
161 6 199 4
"162=163-I 1 *2w =&(401--l) 2

530
(c) Goldbach’s Problem
 INTRODUCTION TO PAPERS ON GOLDBACH'S
PROBLEM
The main papers in this section are P.N. III (1922, 3) and EN. V (1924, 6). The
other three papers are historical reviews and a preliminary announcement.
In P.N. III and P.N. V Hardy and Littlewood deal with Goldbach’s Problem on
the assumption that the following hypothesis is true.
Hypothesis R. There exists & real number 0 < 3/4 such that all zeros of all L-series
L(s, x> formed with Dirichlet characters lie in the half-plane o < 0.
Hypothesis RI is weaker than the original Riemann hypothesis, but is asserted for
a larger class of functions. It remains unproved today. Throughout P.K III and
PX. V Hypothesis R is taken for granted.
Let N;j(n) be the number of representations of the odd positive integer n as a sum
of three primes. Let
c3 = n (1+(w-V3),
W

where w runs through all odd primes. The authors prove that

where p runs through all odd prime divisors of n. In particular it follows that
N,(n) > 0 for large n.
The proof runs on lines similar to those of the preceding sections, though new
difficulties cropped up and had to be overcome. The generating function is the third
power
f(x) = 2 (log WP,
W

where w runs through all odd primes.

The introduction of the factor logw is a familiar device to make the use of the
prime number theorem for arithmetical progression rii easier. The Farey dissection
works as before; there are no minor arcs. If
X=e -Y +z~iPlq
?

it follows from Mellin’s formula that, on a Farey arc around p/q,

1
x= = e2~iPdq - Y-T(s)w-8 as,
27Ti s
2--ico

and apart from negligible terms, f(x> can be expressed as a linear combinafion of
integrals 2+iw
1
Y-T(s) --L’b~J x) &
27ri s Lb x) 3
2--im

where x runs through all characters (primitive or not) mod q. The integrand is regular

533
 INTROIXJCTIXN TO PAPERS ON GOLDBACH’S PROBLEM

for CT> U, except for a pole at 8 = 1 when x is the principal character, and is mero-
morphic for all s.
Hence Cauchy’s theorem can be applied so as to shift the path of integration to
the line 0 = -& The residue at s = 1 gives the principal terms in the approxima-
tion for f(x); the contributions resulting from the zeros of L(s, x> give simple poles
in the strip 0 < 0 < 0. A careful estimate of their contribution to the value of the
integral gives an approximation forf(x) on the Farey arc. Then the older techniques
can be applied; the singular series, though troublesome, does not present serious
difficulties. And the asymptotic formula is proved.
In the rest of the paper, the authors apply their method heuristically to a large
number of problems which are exceedingly difficult. Of their large number of con-
jectures only one has so far been proved, namely by Linnik, who showed in his book
The diqxrsion method for binury additive problems (Leningrad, 1961) that every large
integer is a sum of two squares and a prime. The same result had been proved earlier
by C. Hooley (Acta Math. 97 (l&R’), 189-HO) under the assumption of the generalized
Riemann hypothesis.
In P.N. V (1924, 6)-f the authors assume that Hypothesis R holds with 0 = +,
and deduce that there are at most O(nB+E) even positive integers not exceeding n
which are not the sum of 2 odd primes. The paper stands to P.N. III in the same
relation as P.N. VI to the preceding papers on Waring’s problem.
Since the time of Hardy and Littlewood much progress has been made in Gold-
bath’s problem.
L. Schnirelmann (Iswestija Donskowo Polytechn. Inst. (Novotscherkask) 14 (1930),
3-28) proved the existence of a constant y such that every integer > I is the sum
of at most y primes. I. M. Vinogradov (Rec. Math. Moscou, (2) 2 (1937), 179-95)
proved that every large odd number is the sum of three odd primes, that is, he proved
the Hardy-Littlewood theorem without the use of Hypothesis R. His proof is based
on a clever introduction of an exponential sum, essentially transferring the sieve
method to exponential level. U. V. Linnik (Rec. Math. [Math. Xbornik], N.S. 19 (61)
(1946), 3-8) proved the result again. His proof reverts to the original Hardy-Little
wood pattern, but instead of using Hypothesis R he obtains an estimate for the total
number of zeros, in a particular region, of all the L-functions to a given modulus.
For expositions of modern work on Goldbach’s problem, the reader is referred to:
T. Estermann, Introduction to modern prime number theory (Cambridge Tract No. 41),
Cambridge, 1952.
L. K. Hua, Additive Primxahltheorie, Leipzig, 195%
K. Prachar, PrimxahEverteiZung, Berlin-Gsttingen-Heidelberg, 1957.
I. M. Vinogradov, The method of trigonometrical sums in the theory of numbers (trans.
K. F. Roth and A. Davenport), London, 1954,
H. H.
t An abstract appeared in Proc. London Math. Sot. (2) 22 (1924), xi.

534
 Note on Mews Shah and Wilson’s puper entitled: c On acn
em~*rical formula connected with Qoldbnch’s Theorem ‘. By G. He
HARDY, M.A., Trinity College, and J. E. LITTLEWOOD, M.A.,
Trinity College.
[.&&wed 22 January 1919: read 3 February 1919,]
1. The formulae discussed by Messrs Shah and Wilson were
obtained in the courseof a seriesof researcheswhich have occupied
us at variOus times during the last two years, A full account of
our tnethod will appear in due course elsewhere*; but it seems
worth while to give here some indication of the genesis of these
particular formulae, and others of the same character. We have
added a few words about various questions which are suggested by
Shah and Wilson’s discussion.
The genesisuf theformulae.
2. Let
f(x) = ai (n) xn = ZA (n) e-y = P(y)
and fK (x) = R (y) = QK (n) A (n) e-Y
where A (nj is equal to log23 when n is a prime p, or a power of 33,
and to zero otherwise, and 3~~(n) is one of Dirichlet’s ‘characters to
modulus q ‘$ Also let
X = xpq

where p is positive, less than q, and prime to g ; and suppose that

x tends to unity by positive values.
It is known that
&K (Y) A (Y) = 0 (l.L),
1
unless xK is the ’ principal ’ character x1, in which case

It follows that
(2*1)
and
(2 2)
l

* Au outline of one of its most important qplimtions is contained in & paper

entitled i A new solution of Waring’s Problem ‘, which will be published shortly ia
the Qzlar~rl~ ~aarnal of Muthenzatics.
t See Landau, Handbuclb, pp. 391 et seq.

1919, 1 (with J. E. Littlewood) Proceedkg8 of the Cambridge Philo-

80@4 Society, 1% 24654. 535
 246 iIf?+ Hardy and Mr Littlewood, Note on

Now
(23) f(x) ernMiq,:= 2 @@ai@ 2 A(n) x5
= IZA(n)xn
j=l s=4
Ifj is prime to q, we have*
(W e A (fw = .&) u=lY XK(j)f, (x),
a=
-i
where Xx is the character conjugate to xx, and +(q) is the number
of numbers less than and prime to q. It follows from (2.1) and
(2.2) that

If on the other hand i is not prime to y, the formula (2#4) is

untrue, as its right-hand ‘side is zero. But in this case A (0~) = O
unless TZis a power of q,- so that
(2.6) -

From (%3), (M), and (2m6) it follows that

(2.7)
where

the summation extending over all values of j less than and prime
to q. The sum which appears in (2.71) hk been evaluated by
Jensen and Ramanujant, and its value is P(Q), the well-known
arithmetical function”of q which is equal to zeio‘;nless Q is a product
p1p2 ‘-’ p, of different primes, and then equal to (- 1)~. Thus

(2’8) f(x)rV p(Q) l

4$(&l -xr*

3. The sum
Gw w (n) = 2 A (m) A (m’),
m+m’=It

* Landau, IA., p+ 421.

t J.L. W.V. Jensen, ‘Et nyt Udtryk for den talteoretiekeFunktion zp (n) = M(m)‘,
Saertry k q f Bere tniqg om den 3 Skandinuviske Ma temutiker- Ko&s, Krist iania,
1915 ; S. Ramsnujan, ‘On certain trigonometrical sums and their applicationti in the
theory of numbers ‘, Trans. Camb, Phil. SW, vol. 22, 1918, pp. 259-276.
$ If /& (g) is zero, this formula is fo be interpreted as meaning

536
 I247
Messrs Shah and Wilson’s paper

which appears on the left-hand side of Shah and Wilson’s equation

(Q, is the coefficient of x9&in the expansion of {f(x)]? And

If(x)}2NI~~(~2~~~~~~ne-~p~iq,

when x --Pi’q along a radius vector. Our general method ac-

cordingly suggests to us to take
fi (n) = n x $3 2pPd$
{ 1
where the summation extends over q = 1, 2, 3, . . . and all values
of p lessthan and prime to q, as an approximation to o (IZ). Using
Ramanujan’s notation, this sum may be written ,,

(3’2) fi(n)=nX
$; 2cq(n).
I 1
The series (3*2) can be summed in finite terms, We have

(3-3) cq(n) = Z&(i$ ?

the summation extending Over all common divisors 6 of q and rP;

and it is easilv verified. either bv meansof this formula or by means
of the definition of c,(k) as a tAgonometrica1 sum, that

whenever p and q’ are prime to one another. We may therefore

write

where the product extends over all primes P, and

x ,=l+A,+A,,+Amz+...=l+Aw,
since A, contains the factor p (q) and A+, A+, I . . are accordingly
zero!
If ‘Ris not divisible by W, we have c, (n) = p (P) = - 1 and
1 1
A !w=-m=-((Pi
while if 72is divisible by w we have
c,(n)=~(w)+w~(l)=a--1,
1
A &--’
Hence
n(n)=nrr(l +L.l)nql-&

* Ramanujan, L.C.,p- 260.

537
 248 Mr Hardy and MqmLittlewood, Note OII

where II’ applies to primes which divide 72 and II” to primes

which do not.
It is evident that fi (12) is zero if n is odd. On the other hand,
if G+Z
is even, we have

where m now runs through all odd primes and p through odd
prime divisors of n.
L

The formula w (4 - fi (4
is formula (2) of Shah and Wilson’s paper*.

The incorrectness uf Sylvester’s formula.

4. It is easy to prove that if any formula of the type
W) 0 (7%)
- cl-l (12)
be true, thelz C wuustbe unitiy. In other words, our formula is the
only formula of this type-which can possibly be correct. This
may be shown as follows.
Let

where n runs through all even values; and let s - 1 = t. The ~&es
is absolutely convergent if s > 2, t > 1. Replacing a (n) by its
expression in terms of the prime divisors of 32,and splitting up
f(s) into factors in the ordinary manner, we obtain
21-tR 2l+A~ (t)
f( s> j-J 1-2-t ’
(
say, where A is the same constant as in Shah and Wilson’s paper,
and P runs through all odd primes.
Lte

and supposethat t-1. Then

+ When n (n) = 0, the formula iti to be interpreted EBBmeaning w (n) WI (n).

538
 249

and so
(4*3) f(s)44~(t)-2(1-2-“)~(t)+ 4 =sJ-2*
-
This is a consequence of our hypothesis : the corresponding
consequenceof the hypothesis (4.1) would be
c
(4.31) f( SN-
> s-2’
On the other hand, it is easy to prove* that
(W w (1) + 0 (2) + ,,. + 0 (7~)- $4;
and from this to deduce that
$(s)=$f=.$?y&
when s+% This equation is inconsistent with (49) and (4*31),
unless C = 1.
It follows that Sylvester’s suggested formula is definitely
erroneous,
It is more difficult to make a definite statement about the
formula given by Brun. The formula to which his argument
naturally leads is Shah and Wilson’s formula (12); and this
formnls, like Sylvester’s, is erroneous. But in fact Brun never
enunciates this- formula explicitly. What he does is rather to
advance reasons for supposing that soqneformula of the type (49)
is true, and to determine C on the ground of empirical evidence?.
The result to which he is led is cquivalcnt to that obtained by
taking C = 1’5985/1 l320:3 = 1*21.07 3, The reason for so substantial
a dis&Dancy is in ef%ct that ex+ned in the last section of
Shah and Wilson’s paper.

Fwther resd ts.

5. The method of 5 2 leads to a whole series of results con-
cerning the number of decompositions of 12,into 3,4, or any number
of primes. The results suggested by it are as follows. Suppose
1
* Since ZA (n) 2” - Ix
1
as x-4, we have Izw (n) x”= (2x (n) xy w (1 -x)2;

and the desired result follows from Theorem 8 of & paper published by us in 1912
(& Tauberian theorerm concerning power wries and Dirichlet’s series whose co&Gents
8re positive ‘, Proc. London Math. SOL, ser. 2, vol. 13, pp. 174-192). This, though
the shortest, is by no means the simplest proof,
The formula (4-4) is substantially equivalent to Landau’s formula (10) in Shah
snd Wilaon’a paper.
j- Evidence connected not with Goldbach’s theorem itself but with a closely
related problem concerning pairs of primes differing by 2. See 5 7.
$ 145985 is km% constant, while 1*3203 is 2A,

539
 250 Mr Hardy and Mr Littlewood, Note on

that vr (n) is the number of expressions of n as the sum of Y primes

Then if r is odd we have
(5*11) UT(n) = 0 (C’)
if D is even, and
2B (P - I)‘- (P - 1)
(5+12) vr (a> - @qj w-l n
i ---($y+I--
if n, is odd, p being an odd prime divisor of n, and

where P runs through all odd primes. On the other hand, if r is

even, we have
(5%) Vp(n) = 0 (P)
if ~2is odd, and

where
(5#23)

if ‘fi is even. The last formula reduces to (1) of Shah and Wilson’s
paper when r = 2.
We have not been able to find a rigorous proof, independent
of all unproved hypotheses, of any of these formulae. But we are
able to connect them in a most interesting manner with the famous
’ Riemann hypothesis ’ concerning the zeros of Riemann’s function
&). The Riemann hypothesis may be kated as follows : r(s) has
no zeros whosereal part is greater than 8. If this be so, it follows
easily that all the zeros of c(s), other than the trivial zeros s = - 2,
s=- 4, lie on the line c = R(s) = 4, It is natural to extend
l .*,

this hypothesis asfollows: no oneof the functions defined, when u > 1,

by the series
L (s)= c 9,
possesses zeros whosereal pavt is greater than 4. We may call this
the extejj,dedRiemanrbhypothesis. This being so,what we can prove
is this, that gy the extended Rientann hypothesis is true, then the
formulae (5*11)-(5.23) are true *for all values of r greater than 4.
The reasons for supposing the extended, hypothesis true are
of the same nature as those for supposing the hypothesis itself
true, It should be observed, however, that it is necessary,before
we generalise the hypothesis, to modify the form in which it is
usually stated ; for it is not proved (as it is for c(s) itself’) that
L(s) can have no real zero between & and 1.

540
 Messrs Xkah and Wilson’s puper 251

6. A modification of our method enables us to attack a closely

related problem, that of the existence of pairs of primes differing
by a constant even number k.
We have
CA (n) A (n + k) rm+k = k/r If(reie) I* ehkie de,

where f(x) is the same function as in $1, and r is positive and less
than unity. We divide the range of integration into a number of
small arcs, cbrrelated in an appropriate manner with a certain
number of the points e2+‘Q,and approximate to If(reie) I2 on each
arc by means of the formula (2%). The r&nlt thus suggested is
that

where A has the same meaning as in 5 2 and p is an odd prime

divisor of k. From this it would follow that

w> C A(v)A(v+k)-2AnII
u< ?A
and that, if N&z) is the number of prime pairs less than n, whose
difference is lc, then
(6-2)

This formula is of exactly the same form as (l), except that p is

now a factor of k and not of n. In particular we should have
2An
(6’3) N2 @) - (logn)2 ’
and
4Av-t
N6 b) - jig Q l

We should therefore conclude that there are about two pairs of

primes differing by 6 to every pair differing by 2. This conclusion
is easily verified. In fact the numbers of pairs differing by 2, below
the limits *
100, 500, 1000, 2000, 3000, 4000, 5000,
are
9, 24, 35, 61, 81, 103, 125;
while the nulnbers of pairs differing by 6 are
16, 47, 73, 125, 168, 201, 241.

* To be precise, the numbers of pairs (p, p’) such that p’=p + 2 and p’ does not
exceed the limit in question.

541
 252 Mr Hardy I mad Mr Littlewood, Abte on

The numbers of pairs differing by 4, which should be roughly the

same as those of’ pairs differing by 2, are
9, 26, 41, 63, 86, 107, 121.

7. Brun, in his note already referred to, recognises the corre-

spondence between the problem of $$ 2-4 and that of the prime-
pairs dif’fering by 2, and realises the identity of the constants in-
volved in the formulae ; but does not allude to the more general
problem of prime-pairs differing by i?. He does not determine the
fundamental constant A, attempting only to approximate to it
empirically by means of a count of prime-pairs differing by 2 and
less than 100000, made by Glaisher in 1878% The value of the
constant thus obtained is, as was pointed out in § 4, seriously in
error. The truth is that when we pass from (6*1), which, when
Ic = 2, takes the form
C R(v)R(v+2)^,2An,
VC12

to (6m3), the formula which present.s itself most naturally is not

(69) but
(W

This formula is of course, in the long run, cquivalerltl tc, (6’3).

But
?z 2! 3! t
s l + logn + (logn)" + I.* ;
and the second factor on the right-hand side is, for 12= 100000, far
from negligible. Thus (6*3) may be expected, for such values of
n, to give results considerably too small.
If we take the lower limit of integration in (7.1) to be 2, we
find that the value of the right-hand side for 72 = 100000 is, to the
nearest integer, 1249, whereas the actual value of LX&~) is, accord-
ing to Glaisher, 1224:. The ratio is 1.02, and the agreement seems
to be as good as can reasonably be expected.
The calculation of prime-pairs has been carried further by
Mrs Streatfeild, whose results are exhibited in the following table:

* J. W. IL Glaisher, r An enumeration of prime-pairs ‘, MWX~I- 0f imthernatics,

vol, 8, 1878, pp, 28-33, The number of pairs below 100000 is 1225.
? The series is naturally divergent, and must be closed, after a finite number of
terms, with an error term of lower order than the last term retained.
+-I- Glaisher reckons 1 aB a prime and (1,3) a6 a prime-pair, making 1225 in all.

542
 Mews Shah and Wilson’s *paper
L 253

2A * dx
--- Ratio
h w)
--- I 2 (log sj2
--

1224 1249 x*020

2159 2180 1 *UlO

2992 3035 1 G14

3801 38463 1’012

4562 4625 1’014

5328 5381 1’010

8. In a later paper* Brun gives a more general formula relating

to prime-pairs (p, p’) snch that p = ap’ + 2. This formula also
involves an undetermined constant k. It is worth pointing out
that our method is equally applicable to this and to still more
general problems. Suppose, in the first place, that u (n) is the
number of expressions of TZin the form
n = ap + +‘,
where p and p’ are prim&. We may suwose without loss of
generality that a and b have no common factor.
The results suggested bv our method are CMfollows. If n. has
any factor in common with i and b, then

,_.---
zf(n)=o
1(lo;4I; 2

and this is true even when n is prime to both LTand b, unless oyae
of in, ct, 6 is even:. But if ‘n,a and b are coprime, and one of them
even, then

where A is the constant of 5 2, and the product is now extended

over all odd primes which divide ‘yzor a or b.

* ( Sur les nombres premiers de la forme q + b ‘, ArcGu for Mathmmztik, ~1.

24, 1917, no. 14,
f We might naturally include powers of primes.
r These results are trivial. If 72 and a have a common factor, it divides bp’,
and is therefore necessarily p’, which can thus assume but a finite number of values,
If n, a, b are all odd, either p or p’ mustnecessarily be 2.

543
 254 Mr Hardy mad Mr Littlewood, Note etc.

Similarly, suppose N(n) to be the number of pairs of solutions

of the equation
ap’ - bp = k
such that p’< n. It is supposed that a and b have no common
factor. Then

unless k is prime to both a and b, and one of the three is even.

If these conditions are satisfied

where p is now an odd prime factor of k, a, or b.

544
 Goldbach’s Theorem.
By G. H. Hardy’).

The famous problem of which I propose to speak tonight

is probably as difficult as any of the unsolved problems of
mathematics, and my lecture cannot be entirely easy. I must
be content if I can give you some rough notion of the nature
and of the history of the problem, and of the general ideas
which have guided the mathematicians of the past and of the
present ,in their efforts to find a solution,
Every even number is the sum of two primes:
this is ‘Goldbach’s Theorem’. I may perhaps begin by the
trivial observation that, if ‘prime’ is to mean what it means in
modern mathematics, the theorem is obviously false. It fails
for 2, which is a prime, but not the sum of two, We do not
nowadays call I a prime, for, if we do, the factorisation of
a number into primes is not unique. We must therefore insert
the words ‘greater than z’ in the enunciation of the
theorem.
The theorem is stated in Goldbach’s correspondense with
Euler in the year 1742, It would seem that he had been
anticipated by Descartes 2). It was conjectured independently
by Waring a little later, and it is, I believe, in Waring’s
Meditationes algebraicae (1770) that the conjecture
appears first in print. Each of these authors appears to have
observed that, if every even number (greater than 2) is the
sum of two primes, then e very number (greater than 5) is

‘) A lecture to the Mathematical Society of Copenhagen on 6. October 1921.

‘) In these matters of history I am content to follow Prof. L. E. Dz’ckson’s
History of the theory of numbers (Washington, xgrg, vol. x.pp.421
et se 4.). Dickson however attributes to Descartes the assertion that
‘every even number is a sum of I, 2, or 3 primes’, and here the word
‘even’ should surely be deleted.

1922, 1 Matemtisk Tidsakrift B, 1922, l-16. 545

 2 6. H. HARDY:

the sum of three. You will see later why this apparently
trivial remark should be interesting to me.
The numerical evidence for the truth of the theorem is
overwhelming. It has been verified up to 1000 by Cantor
(iQ$--I8gs), to 2000 by Aubry (1896~-rgo3), to 10,000 by
Haussner (1896) ; and further numerical data, concerning special
numbers or numbers of specified forms, have been accumulated
by Ripert ( Igo3), by Cunningham (rgo6), and by Shah and
Wilson (1919). But most of the modern computations have
been directed towards a more ambitious end, that of deter-
mining or verifying some asymptotic formula for the number
of decompositions into primes of a given even number 12.
I denote by Y (1~) the number of ways in which n = ZN
can be expressed as a sum of two primes, or the number of
solutions of the equation

72= 2N=p-+pt* (I)

According to Goldbach’s Theorem, v (72)> o if a > Z; and a
very hasty survey of the evidence is enough to make two
things clear. In the first place, a great deal more is true
than is asserted by the theorem. Not merely is v(n) positive,
but it is large when 72 is large, Secondly, while v (?t) tends
to infinity with n, it does not do so in any very regular
manner. The magnitude of v (n> does not depend merely on
the magnitude of n, but also on’ its arithmetic form.
It is important to observe that these conclusions are in
complete agreement with the a priori judgement of common
sense, We naturally argue thus. If wz < 72, and 7~ is large, the
chance that nz is prime is approximately I : log 72. If then
we write IZ in every possible way in the form 72 = m + m’,
the chance that b o t h m and m’ are prime is approximately
I : (log n)e. We should therefore expect the order of magnitude
of v(n) to be
(2)

On the other hand, the arithmetical form of 72 is plainly

also relevant, In the first place, it is obviously relevant
whether n is odd or even: if n is odd there is no representation,
unless 12=p + 2, and then only one. It is not quite so
obviously important to consider whether it is a multiple of 3*

546
 GOLDBACH’S THEOREM. 3

We have however to exclude all cases in which either m

or HZ’ is a multiple of 3. If 3 jlzl), the two sets of cases thus
excluded are the same, while if 3 + n the effect of the double
exclusion is cumulative. W should therefore. expect divisibility
by 3 to increase v (72); and a similar argument applies to any
other prime, so that Y (72) should be largest when n is com-
posed of a large number of different small prime factors. All
these rough expectations prove to be in complete accordance
with the facts.
Sylvester (1871) was the first mathematician to suggest an
asymptotic formula for Y (n). Sylvester’s rule is stated in
words, and when translated into symbols is as follows:

(3)

where n (7~1 is the number of primes not exceeding n, and the

product extends over all primes p for which 3 (- p 5- 72 and
p t It- We know, though Sylvester did not, that

(4)

This is the famous ‘Primzahlsatz’, and we can use it to simplify

(3). We also require a theorem of Mertens (1874)~ to the
effect that
(5)

where C is Euler’s constant. This theorem, which lies much

less deep than (4), shows that

where

Following Landau, I write 3 [ n’ for ‘N is divisible by $, and ‘3 + TZ’ to

express the contrary.
 4 G, H. HARDY:

the product extending now over all odd primes. We thus

obtain the formule

where p is an odd prime divisor of 72, and this we may

reasonably call ‘Sylvester’s formula’,
Sylvester’s formula is certainly wrong. It is correct,
apparently, in its most obviously interesting parts, in its crude
order of magnitude, and in the irregularly oscillating factor
which depends upon the arithmetic structure of 72. It is the
constant factor 4Re -X that cannot be correct. It is interesting
to pause for a moment to consider how such. a negative result
can be established.
There is no mystery at all about the a v e rage value of
v(z). It is quite easy to prove 1) that

(8)

If now we assume any asymptotic formula for v(n), we can

test it by its compatibility with (8). We may assume a fur-
mula of Sylvester’s type, but with an unspecified constant
factor B; and we find, as the test of compatibility, that B--A.
Thus the only possible form,ula of this type is

-?-&L~ (E).
y @’ = (log 7212
(9)
PD

Sylvester’s formula must be wrong, if not in principle, then

by a constant multiplier ~e--~ = I 123 -a), The formula (9),
n l l

on the other hand, seems almost certainly correct.

Sylvester’s contribution to the problem passed unnoticed
for nearly 50 years. I turn now to the writers, Merlin, Brun,
and Stackel, who have attacked it recently, and particularly
to Brun. The work of Mr Littlewood and myself belongs to
a different circle of ideas, and I shall speak of it later,

‘) The actual theorem is due to Landau (1900).

“1 There is considerably more difficulty in testing the discrepancy by com-
parison with the facts. Shah and Wilsov have given a very clear and
interesting discussion of this point.

548
 GOLDBACH’S THEOREM. 5

The writings of these three mathematicians have a striking

common characteristic: they use elementary methods only.
You may ask me what an ‘elementary’ method is, and I must
explain precisely what I understand by this expression. I do
not mean an easy or a trivial method; an elementary method
may be quite desperately ingenious and subtle. I am using
the word in a definite and technical sense, and in this 1 am
only following the common usage of arithmeticians. I mean,
by an elementary method, a method which makes no use of
the notion of an analytic function. And the question that I
wish to put to you is this: is it reasonable, in the present
state of mathematical knowledge, to ho pe to o b tai n an
elementary proof of Goldbach’s theorem?
If I reply to this question in the negative, as I must and
shall, if I say that I am compelled to regard all such efforts as
foredoomed to failure, I trust that you will not misunderstand
me. I cannot believe that the methods of Merlin and Brun are
sufficiently powerful or sufficiently profound to lead to a solution
of the problem. Rut I am very far from meaning that I regard
their work as devoid of interest and value. There is much in
Brun’s work in particular that seems to me very beautiful,
and some of his theorems ought, I think, to find their way
into every book on the theory of numbers.
We have however to take account both of the history and
the logical structure of our subject. Let us turn back then
for a moment to its central theorem, the Trimzahlsatz’ or
‘pfime number theorem’ expressed by the equation (4). It
seems plain that this must be at any rate an easier theorem
than Goldbach’s theorem. No elementary proof is known, and
one may ask whether it is reasonable to expect one. Now
we know that the theorem is roughly equivalent to a theorem
about an analytic function, the theorem that Riemann’s Zeta-
function 1) has no zeros on a certain line 2) A proof of such
a theorem, not fundamentally dependent upon the ideas of
the theory bf ftinctions, seems to me extraordinarily unlikely.
It is rash to assert that a mathematical theorem ca n n o t be
proved in a particular way; but one thing seems quite clear.
We have certain views about the logic of the theory; we think

549
 6 G. H. HARDY:

that some theorems, as we say, ‘lie deep’, and others nearer’

to the surface. If anyone produces an elementary proof of
the prime number theorem, he will show that these views are
wrong, that the subject does not hang together in the way
we have supposed, and that it is time for the books to be
cast aside and fur the theory to be rewritten.
You are probably familiar with the general idea of the
‘sieve’ or ‘crible’ of Eratosthenes. We write down all the
integers

up to x. We erase (or underline), first I ; then every even

number after 2; then multiples of 3, except 3 itself; and so
on, repeating the process I_-
for every p. The process comes to
an end when p exceeds IX, for then only primes are left.
Suppose now that p,, p,,are the first -Y primes,
’ l l J pr

that x is large and Y fixed, and that we use the sieve for
these primes only. The number of numbers left is approxi-
mately

it is easy to see that the error in this enumeration is at most

2r + Y., If 71, (x) is the number of numbers, not exceeding X,
and prime or not divisible by any of Y prime% then

“r(x)NZli(l
--;)m 1
(9)

We have supposed so far that Y is fixed. The result would

be much mare int,eresting if we could suppose that r is a
function of x;. Let us assume provisionally that this is legi-
timate, and take Y to be the largest prime not exceeding f>.
We obtain

by Mertens’ formula (5) This formula is false (so that our

assumption was illegitimate), and it is significant that it is

550
 GOLDBACH'S THEOREM. 7

wrong in just the same way as Sylvester’s formula (7). our

failure may help to deepen our scepticism as to the results to
be anticipated from the use of the principle of the sieve.
The ‘sieve’ used by Merlin and Brun for Goldbach’s theorem
(le crible de Merlin) is 0f.a slightly more complex kind.
We form the table

I, 2, 31 41 5, 6, 7, ml ., 72-I (m>,)
n- L n--2, n-3, n-4, n- 5, n-6, n-7, l . . ,
I (Id). J
(

* I
1

Were 72 is an even number, and the table, read vertically, shows

all possible decompositions 18= PZ+ m’ of 72 into positive
integers. We now perform the process of Eratosthenes on
both rows of the table, starting from the right of the lower
row, and operating with the first Y primes. We consider a
decomposition FE+ m’ to be erased when either of its consti-
tuents is erased.
There are two cases to be considered for each prime p.
If p 172, the erasures in the second row fall immediately below
the corresponding erasures in the first, and the number of
decompositions erased is approximateiy 7~ :p. If p + 72, the
erasures never correspond, and their number is approximate-
ly 2n :p, If then Ye(72) is the number of decompositions into
numbers ~lil,m’ which are prime or not divisible by any of
the first r primes, we have

vp (72) - 72 n pj *n (dj. (12)

P/nlp&, - 2+b P(P,-

This formula is correct so long as Y is fixed.

Let us assume once more that this formula is correct when
p, is about J& , We obtain

where the value z of fi is now excluded. But, if 3 ~2

- p S
- iz,
we have

by Mertens’ formula (5 , We thus obtain

551
 G, H. HARDY:

(I 3)

and this is the formula to which Brun’s argument naturally

leads’). The formula, like Sylvester’s, is wrong, and by a
factor 4e-lc = I 263 l It will be observed that
l l l . Sylvester’s
formula (7) is the geometric mean between (9) and (I 3).
I have explained that I do not believe that a proof of
Gdldbach’s theorem is likely to be found by methods such as
these. Hut it is certainly possible to pr,ove s om e t hi n g in
this sort of way, and what Brun has proved seems to me
very interesting indeed. He has proved, for example that
every large even number 72 can be expressed in the
form tit + m’, where wz and WJ’ are numbers composed
of at most g odd prime factors, and that the number
of such decompositions is of order w (log7z)a at least.
His method of proof is elementary enough, but a little com-
plicated, and the idea which underlies it can probably be
explained most clearly by reference to a simpler problem,
The number of numbers not exceeding X, and prime or
not divisible by any of the primes p,, p,,. 1,p,, is (if p, < X)
l

where [x] is the largest integer in X. This can be shown at

once by the method of Eratosthenes, and it is an imediate
deduction that

where ,-1 is a constant* No very interesting consequences can

be drawn from this; but Brun has shown that, if we are con-
tent to allow the first term on the right of (I 5) to be multiplied
by a constant less than T, we can materially reduce the order
of the second, considering as a function of Y’. More precisely,
he proves that

I) The formulae (13) is not enhnciated explicitly by Brun. His procedure

was rather to develop his argument until it leads to a formuIa o f th i s
t y pe, and to attempt to determine the constant on other grounds. The
determination of the only possible constant by averaging was effected in-
dependently by WckeI, and by Mr. Littlewood and myself.

552
 GOLDBACH’S THEOREM, 9

(I 6)
Suppose now that p, is about x6, Then the first term is (by
Mertens’ theorem) of order x : log x, while the second is of
or&r A, Thus N is at least of order z : log z, If a number
does not exceed z, and all of its prime factors are greater
than &, the number of its prime factors is 5 at most. We
are led to the theorem that the number of numbers less
than z, and with five prime factors at most, is at
least of the order x:logx.
The theorem is trivial, for Tschebyschef proved (and by
purely elementary methods) that the number of p rim es less
than x is of this order of magnitude. I have chosen this tri-
vial theorem, however, merely as a simple illustration, and
Brun has proved much more, In the first place, he has ex-
tended the argument to the numbers of an arbitrary arithme-
tical progression mk + Zl). This theorem also is of course not
new, for we know, from the work of de la Vallke Poussin,
that any such progression contains the prescribed amount of
primes. The method, on the other hand, is in this case most
interesting, for there is no elementary proof of any of the
theorems concerning the primes of an arithmetical progression.
What is still more important is that Rrun has been able to
treat the ‘crible de Merlin’ in the same kind of way, and to
deduce the very striking and beautiful theorem that I enunciated
a moment ago; a theorem, I may add, which Mr, Littlewood
and I are unable to prove analytically, with the much more
powerful machinery at uur command.
It seems clear however that, if we are to attack the main
problem with any prospect of success, the methods of the
theory of functions are indispensable; and I shall now attempt
to explain the method by which Mr. Littlewood and I have
attacked it. The general lines of the argument are obvious
and inevitable enough; the difficulties lie in carrying it through
to its conclusion.
We write
9( x 1 = ;T;B + x3 +

1I 7)
x5 + l l ’ z cxp,

1 The uni nteresting case in which K and 2 have a common factor being
1
naturally excluded.

553
 G, H. HARDY:

SO that ~~(72) is the number of representations of 72 as a sum

of y primes I). We have

(I 9)

by Cauchy’s theorem, the path of integration C being the

circle 1x 1=R, where o<R<l. We take

R = 1 - -!- . (20)
72

We do not base our analysis, however, on the actual formulae

that I have written, It is more convenient to use the formulae

where the summation applies to all sets p,, p,, l ; p, whose l l

sum is n. Goldbach’s theorem asserts that v,(7t) (or, what is

the same thing, that N2 (~2))is positive for 72 = 4, 6, . Our m l l

object is the more comprehensive one of finding an asymptotic

formula for v&); and it is easy to show that, if we can find
such a formula for .N,(fz), we can deduce one for v,(n) by
division by (log 72)’ 2).
Our fundamental idea is the same as that which has guided
us in our work on Waring’s problem. The unit circle is a
barrier of singularities for f(x); we may say, roughly, thatf(x)
becomes large when x approaches the circle. There are how-
ever certain special points of the circle in whose neighbour-
hood f(x) is largest, and whose contributions to the integral
(23) are of dominating importance. These points are the ‘ra-
tional points A& k’, for which

L V&Z) has no connection with the vr (n) of (12).

2, If o < b < I, we have (I - b) log PZ < logp < log n for neaily all the
primes in question,

554
 GOLDBACH’S TXEOREM.

It is obvious, for example, that the point O, I, for which x = I,

must be most important of all ; for f(x) is a series with posi-
tive coefficients, and increases most rapidly when z is positive.
In general, one may expect the importance of the point I!Z,k
to diminish as k increases; and in fact

( 2 5)

where p(k) and cp(k) are the well known arithmetical functions
of Mtibius and Euler I), if x = Xxlr,k and X - + I by positive
values.
Suppose that a formula of the type of (25) has been proved
to hold in the neighbourhood of every point I& k, The natural
procedure is then as follows, We divide the circle C into a
large number of small arcs &, each associated with a parti-
cular point k, k. We substitute for f(x), in the part of the
integral (23) which is taken along &, k, the approximation
derived from (25); we evaluate the resulting integrals ; and we
thus obtain an approximation for N,(z) in the form of an in-
finite series. This series, which we call the singular series,
can happily be summed in finite form.
The analysis to which this process leads is intricate and
difficult. It will be best that I should begin by stating what
we can prove. In the first place we have, I for reasons which
will appear later, to assume the truth of an unproved hypo-
thesis. The celebrated hypothesis of Riemann may be stated
asfollows: the real part of a zero of the Zeta.function
does not exceed 3, Our hypothesis is a natural extension
of this. The theory of the distribution of primes depends, in
general, on the theory of c(s). Their distribution in an arith-
metical progression mR + Z, where l is prime to k, depends
upon a number of associated functions denoted generically by
L(s). Thus when k = 4 there are two such functions, defined
respectively by
L,(s) = rs + 3-s+ y+ . . l , -I;, (s) =- I-S - 3-“-f- 5+ . . l . (26)

‘) p(k) = (- ~)m it k is the product of nt different primes, p(k) = o other-

wise; q(k) is the number of numbers less than and prime to K.

555
 12 G. H. HARDY:

The first is (I -2--9<(s), but the second is a new transcendant.

There are cp@) such functions associated with a given k.
The most natural generalisation of Riemann’s hypothesis is
HYPOTHESIS R. The real part or a zero ofL(s) does
not exceed ;2.
We do not actually need quite the full force of this hypo-
thesis, but I shall be content to state it in its simplest and
most striking form.
Our main theorem is then a follows: If hypothesis R
is true, then every largel) odd number is the sum of
three odd primes. The number Y&Z) of representa-
tions is given asymptotically by the formula

where

The product in (27) extends over the odd prime

divisors of B, and that in (28) over all odd primes.
The complete proof of this theorem, and of similar theorems
concerni ng representations of numbers by four or any larger
number of primes, will appear shortly in the A c t a Ma the-
m at ic a, At the moment I shall attempt to explain to you only
(I) how the final formula arises,
(2) why it should be necessary to assume hypothesis R,
(3) why our method succeeds for three or more primes, but
fails for two.
I shall not take these questions in the order in which I
have stated them: I begin with the second. You will remember
that it is necessary to approximate to the function f(x) in the
neighbourhood of the point k, R. Taking first the simplest
case, suppose that h = 0, k = I, xh, k = I.
We write x = P-Y, and use the well known formula

e--Y L & y-” r (s)ds (2 9)

J
of Mellin. From this we deduce

‘) That is, every such number from a certain point onwards.

556
 GOLDBACH’S THEOREM.

which is substantially (though not exactly)

W) (3 I >
‘50S
The integrals are taken along a line parallel to the imaginary
axis and passing to the right of the point s = I.
The subject of integration has poles when s = I, when s
is zero or a negative integer, and at all the complex zeros of
C(S), which we denote generally by p- If we assume Riemann’s
hypothesis, every p has the real part +, It is natural to sup-
pose that, if we move the path of integration further and
further to the left, and apply Cauchy’s theorem, we shali ex-
press f(x) in the form of an infinite series, in which the most
important terms will be the terms corresponding to the residues
for I and p. If we denote these terms by
I AP
-1 (3 2 )
*--
3’ + 2
!
3P
then the term cortesponding to p is of order y- i when y is
small, and we may hope that the order of the series will be
much the same. In this case we shall prove that f(z) NJ+,
which is the simplest case of (25), and tltat the order of mag-
nitude of the error is not notably greater than that of J!- g.
It Is evident that our estimate of the error will depend es-
sentially on our assumption.
Passing now to the general case, I denote by A(B) the
arithmetical function of 72 which is equal to log n when n is
prime and otherwise to zero. We have now

(33)

and we obtain

(34)

557
 14 G. H. HARDY:

Thus f(x) is expressed as the sum of a finite number of func-

tions, each of which plainly depends on the distribution
of primes in an arithmetical progression ntk j-Z. We
are thus led to make an assumption, analogous to the hypo-
thesis of Riemann, about all the functions L(S) on which this
distribution depends; in other words, we are led to hypothesis
R. Unless we make some such assumption our difficulties, al-
ready considerable, will be terribly increased.
We therefore assume hypothesis X. We now write X-e-’
and express f(x), by means of the formulae (33) and (34), in
the form of an integral
--I* Y-$ r(s) Z(s)ds, (35)
2m s
where Z(s) is a linear combination of the logarithmic derivatives
L’(S) : L(S) of the L-functions associated with modulus K. The
dominant term of f(x), in the neighbourhood of z~,~., is the
residue of the integrand for s = I, and a simple calctilation
shows that this is the function which appears on the right
hand side of (25).
We now return to the integral (23, and consider the se-
parate contributions of the arcs &,k, If we substitute for f(x),
on El*,k, the approximation given by (~5)~ and transform the
integral by the substitutions

(37)

The path’ of integration is now a segment of a straight line,

parallel to the imaginary axis in the plane of Y, and passing
to the right of the origin.
We can substitute for this segment the complete straight
line of which it is part, the error involved in this approxima-
tion proving to be unimportant. The integral can then be
evaluated in finite terms, and (37) assumes the form

(38)

We have fimlly to sum with respect to h and k, and we ob-

tain the formula

558
 72’--’
ilv, (72) - _ ” s
( Y- I)! ’
where

and h, in (4r), is less than and prime to k.

The series S, the ‘singular series’, is the series on which
the solution of all these problems depends. It can, as I said
before, be summed in finite termsl). If Y and n are of oppo-
site parity, then S = O. If they are of the same parity, then

s= m Lep--
zc, -- I)‘+~_--_
(- qyp- - I)
(p- t)’ - (- I)’
-- ,
i
(4 2 )

where
I- 0’ . \
c r- - I (431
II( (P - I)’ )
Tf Y- 3, we are led to (27), and if Y _I 2 to (g), the powers
of log 12 which appear in these formulae introduced being
in
the passage from iv to V; and there formulae are similar
for
every value of r.
There is unfortunately a vital difference between the case
r = 2, which corresponds to Goldbach’s theorem, and all the
rest. We have to fill in the skeleton which I have presented
to you, and to transform it into an accurate proof; and in
doing this we find ourselves compelled to suppose that r> 2.
It only remains that I should explain to you shortly the reason
for this regrettable limitation 2).
Our work depends upon a system of approximate formulae
for -f(x), each validbnear a particcllar point of the unit circle. Tt
will be sufficient fur my purpose to consider the point z -- I.
If x = e--y, and y is positive, our formula for@) is of the form

‘> The summation is a simple but entertaining exercise in elementary alge-

bra, which I must be content to take for granted.
‘) The explanation which follows must be taken merely as a first approxi-
mation to the truth.

559
 16 G, Ha HARDY: GOLDBACH’S THEOREM.

where E may be any positive number. Thus (f(x))’ consist of

two parts, a term y-r, which is exactly known, and an error
term whose order is, so far as our analysis shows, not less
than O(y-4 ‘); and each df these terms gives rise to a corre-
sponding term in the final formula. The contribution of the
dominant term can be found precisely, and is of order ~t+-l.
The contribution of the error term can only be estimated in a
cruder manner, and the best that we can say about it is that
it is of order 0(7&). This error must, if our approximation is
to succeed, be smaller than the dominant term; and this requires
that P- I >i-r or r>Z. In fact I have done rather more
than justice to our method, It would appear, from what I have
stated, that we stand on the very margin of success. But.
there are further complications in the analysis, and we fail bJ
a power 72).
I implied, when I was discussing the methods of Merlin
and Rrun, that it was hardly reasonable to suppose that they
could possibly succeed. You may naturally ask me what I
think of the prospects of our own, and the question is one to
which I find it rather difficult to reply. You must make me,
for the moment, a present of hypothesis R, I presume that
the hypothesis of Riemann will some day be proved. Hypo-
thesis i? will, I am sure, be proved within a week from then,
and the proofs will be substantially the same. There is nothing
whatever to suggest that, in these respects, one L-function
behaves unlike another.
Apart from this I would reply that, the hypothesis once
proved or granted, I see no particular reason why our method
should not succeed. It seems to me adequate for the problem;
the ideas which underlie it are not too easy and lie sufficiently
deep. It fails in detail, and not in principle, even as it is; the
failure is not a failure of the method, but of the analytical
powers of Mr. Littlewood and myself. The method seems to
embody the essential features of the problem, and leads,
naturally and inevitably, to what is plainly the real result. I
believe that, when the problem is solved, it will be solved in
some such way as this.

560
SOMEPROBLEMSOF 'PARTITIONUMERORUM';III: ON THE
EXPRESSION
OFA NUMBERAS A SUM OF PRIMES.
BY

G, H. HARDY and J. IL LITTLEWOOD.

New College, Trinity (2011ege,
OXFOI~D. CAISII~RIDGE.

I. Introductim.
I. I. asserted by GOLDBACH,
It was in a letter to EULER dated 7 June,
1742, that every even number 2m is the sum of iuw odd primes, and this propmi-
tion has generally been described as ‘Goldbach’s Theorem’, There is no reasonable
doubt that the theorem is correct, and that the number of representations is
large when ‘yn is large; but all attempts to obtain a proof have been completely
unsuccessful. Indeed it has never been shown that every nunlber (or every
large number, any number, that is to say, from a certain point onwards) is tile
sum of IO primes, or of I ooo ooo; and t*he problem was quite recently classified
as a>mong those ‘beim gegenwktigen Stande der Wissenscbaft unangreifbar?
In this memoir we attack the problem with the aid of our new transcen-
dental method in ‘additiver ZablentheorX8 We do not solve it: we do not
1 E. LANDAU, ‘Geliiste und ungelijate Probleme aus der Theorie der Primzahlverteilung und
der Riemannachen Zetafunktion’, Proceedi~~gs af fhe $fth I~te~~&tional Congwss of Mathenzaticians,
Cambridge, 1912, vol. I, pp. 93- x08 (p. 105). This address was reprinted in the Jahresbwicht
der Lkutschen Xxth.- T%-einigwag, vol. 2 I (rgn), pp. 208-228,~
* We give here a complete list of memoirs concerned with the various applications of
this method.
G. EL HARDY.
I. ‘Asymptotic formulae in combinatory analysis’, Conaptes rendus dt4 qua tri&l~e
Cong& des ma thematiciens Scandinavcs h Stockltoltig, 19 16, pp. 45-53,
2. ‘On the expression of a number as the sum 0.f any number of squares, and in
particular of five or seven’, Boceediwgs of the hWiona1 Academy of Sciences, vol. 4 (r918),
pp* x89-193.

1922, 3 (with J. E. Littlewood) Acta Mathematics, 44, l-70. 561

 2 G. H. Hardy and J. E. Littlewood.

even prove that any number is the sum of I oooooo primes. In order to prove
anything, we have to assume the trutb of an unproved hypothesis, and, even
on this hypothesis, we are unable to prove Goldbach’s Theorem itself. We show,
however, that the problem is nof ‘unangreifbar’, and bring it info contact with
the recognized met!hods of the Analytic Theory of Numbers.

3. ‘Some famous problems of the Theory of Numbers, and in particular Waring’s

Problem’ (Oxford, Clarendon Press, Igzo, pp. r-34).
4 ‘On the representation of a number as the sum of any number of squares, and
in particular of five’, T9-ansactions of the American MutJzematical Society, voI. 2 I (1920), pp.
255-284,
5 ‘Note on Ramanujan’s trigonometrical 8um cq (n)‘, Proceedings of the Cambridge
Philosophical Society, vol. 20 (Igzr), pp. 263-271,
G. H. HARDY and J. E. .LITTTZWOOD.
I, ‘A new aolution of Waring’s Problem’, Quarterly Jouwal of pwe and applied
mathematics, ~01, 48 (I919), pp. 272-293.
2. ‘Pu’ote on Mestirs. Shah and Wilson’s paper entitled: On an empirical formula
connected with Goldbach’s Theorem’, Proceedings of the Cambridge Philosoplticnl Society,
vol. 19 (19rs>, pp* 245-254*
3. ‘Some problems of ‘Partitio numerorum’; I: A new solution of Waring’s Pro-
blem’, h’nchrichten zton dw K GeseEZschaft de9- Wissenschaften zu Giiftingen (Igzo), pp. 33-54,
4 ‘Some problems of ‘Partitio numerorum’; II: Proof that any large number is the
sum of at most 21 biquadrates’, ikkfhemntz’sche Zeitschrift, vol. g ( ~zI), pp x4-27,

G. II, HARDY and S, RAMANJJAN,

1. ‘Une formule nsymptotique pour le nombre des partitions de pa’, Cotiaptes y*endus
de Z’Acaddmie des Sciences, 2 Jan. 1917.
2, ‘Asymptotic formulae in combinatory analysis’, Proceedings of the Londou Mathew
atical Society, ser. 2, vol. 17 (IgIg), pp. 75-115.
3. ‘On the coefficients in the expansions of certain modular functions’, Proceediqlgs
of t7ae Royal Society of London (A), vol. 95 (rgr8), pp. 344-155.

I. ‘Zur Hardy-Litt~lewood’sc~~en Losung des Waringschen Problems’, X~r7z~i(*lzte9~

von der II. Gesellschaft der Wissensc7iafiepz zu Giittingen (~gz I), pp. 88-p.

L. J. MORDHLT,.
I. ‘On the repreuontationa of numbers as the sum of an odd number of squares’,
Transactions of the lCambg-idge Philusophicul Society, vol. 23 (1919)~ pp. 361-372.
A, OSTROWSKI.
1. ‘Bemerkungen zur Hardy-Littlewood’schen Liisun g des Waringschen Problems’,
Mathematische Zeitschrift, vol. g (IgzI), pp. ~8-34~

S. RAMANJJAN.
1. ‘On certain trigono metrical auma and their applications theory of
hers’, Triwucztions of the Cam bridge Philosophical Society, vol. 22 (19 qg--76.

N. M. SHAH and B. M. WILBON.

1. ‘On an empirical formula connected with Goldbach’a Theorem’, P9*oceedings of
the Camb?*idge Philosophical society, vol. 19 (IgIg), pp. 238-244,

562
 Psrtitio numerorum. III: On the expressionof a number as 8 sum of prirms. 3

Our main result mav be stated as follows: ;f a certain hypothesis (a natural

generalisation of Riemank hypothesis concerning the zeros of his Zeta-function)
is true, then every large odd number n is the sum of three odd primea; and the
number of representations is given asymptotically by

-
(I. II) N3(n)NC3~--~~~((~~‘)(“‘“),
0 n --3p+3
IJ
where +Jruns through all odd prime divisors of n, md

(I. IZ) c3 = 11 I + @-x~j ’

( 1

the product extending over all odd primes 23’.

Hypothesis R.

I. 2. We proceed to explain more closely the nature of ollr hypothesis.

Suppose that q is a positive integer, and that

is the number of numbers less than q and prime to q. We denote by

x(n)==x&) (k=I,z,...:h)

one of the h Dirichlet’s ‘characters’ to modulus p I: xr is the ‘principal’ character.

By x we denote the complex number conjugate to x: X is a character.
By .L(q x) we denote the function defined for G > I by

L(s) = L(a + it) = us, x>

Unless the contrary is stated the modulus is p, We write

L(s) = L(s, X) ’
Ry
e=p+iy
------.---_ _--__.._
-. _ _._.. - -_-- . _------ -- _----. -- -. .- --- ---- -.------
’ Our notation, so far as the theory of of Landau’s
L-functions ie concerned, is that
Hadhdt der Lehve van de?* Vwteilung des- l+imzahlen, vol. I, book 2, pp. 391 et seq., except that
we use q for his k, k for his X, and a’ for a typical prime instead of r). ,4s regards the ‘Farey
dissection’, we adhere to the notation of our papers 3 and 4*
We do not profess to give a colnplete summarv of the relevant parts of the theory of
the L-functions; but our references to Landau shouldwbe safficient to enable a reader to find
for himself everything that is wanted.

563
 4 G. H. Hardy and J. E. Littlewood.

we denote a typical zero of L(s), those for which y = o, p-<

-= o being excluded.
We call these the non-irivlral zeros. We write N(T) for the number of e’ s of
L(s) fot which o --< y-<
-=:T.
The natural extension of Riemann’s hypot.hesis is
HYPOTHESIS R*.
Every e has 28 real pari less than or equal lo I?
2
Wo shall not have to use the full force of this hypothesis. What we shall
in fact assume is
HYPOTHESIS R. There is a number 0 < 3 such that
4
p&o
for every e 0f every L(s) .
The assumption of this hypothesis is fundamental in all our work; all the
of the memoir, SO /ur as they are novel, depend upon it “; and we shall not
repeat it, in stating the conditions of our theorems.
We suppose that 0 has its smallest possible value. In any case 0 S- l

-- 2

For,
.- if e is a complex zero of L(s), 4 is one of L(a). Hence I - @is one of
-W -s), and so, by the functional equation 3, one of L(s).

Further notatiw and terminology.

1. 3. We use the following notation throughout the memoir.

A is a positive absolute constant wherever it occurs, but not the same
constant at different occurrences. B is a positive constant depending on the
single parameter r. O’s refer to the limit process n-3 ao, the constants which
they involve being of the type B, and O’S are uniform in all parameters except r,
a’ is a prime. p (which will only occur in connection with n) is an odd
prime divisor of n. p is an integer. If Q=I, p= o; otherwise

o<p<q3 (p,q)=r*
(m, n) is, tile greatest common factor of ‘yn and n, By m n we mean that rh is
divisible by m; by m -, n the contrary.
/r(n), fl (n) have the meanings customary in the Theory of Numbers. Thus
A(n) is lug a if n = ~m and zero otherwise: /l(n) is (- I)~ if n is a product of
. . -- ---- - -----._ ----
1 The hppothesia must be stated in this way because
(a) it has not been proved that no I&) has real zeros between f and I,
(h) the L-functions a@ociated with inqjrimitive (uneigentlich) characters have zeros on the line Q= 0.
? Naturally mmy of the results stated incidentally do not depend upon the hypothesis.
3 Landau, p, 489. All references to ‘Landau’ are to his Had12tc~, unless the contrary is stated,

564
 Partitio numerorum. III: On the expression of a number as a sum of primes. 5

k: different prime factors, and zero otherwise, The fundamental function with
which we are concerned is

b 31) f(x) = 2 log ‘GIXT

a’
To simplify our formulae we write

Also

If ;ck is primitive,

(1. 33)

This sum has the sbsolute value 2 V(r.

The Farey d i~~ecfht.

I. 4. We denote by 1’ the circle

- --1
(1. 41) I X I =e-H-e 12.

We divide 1’ into arcs & which we call Farey arcs, in the following manner.
We form the Farey’s series uf order

(I* 42) N=[l/n,], ’

the first and last terms being 0 and -?m We suppose that ‘& is a term of t*he
I I
series, and $ and ‘4 the adjacent terms to the left and right, and denote by
Q cl
& (q > I) the intervals

by jo,l and jl,l the intervals (0, &--) and (I---&-~ 1) l These intervals just
--. -- ---- --. -- - ___-- -----. .-
1 %&d = 0 if (n&,4) > I.
’ Landau, p. 49~~ I’

565
 6 G. H. Hardy and J. E, Littlewood.

fill up the interval (0, I), and the length of each of the parts into which jpB is

divided by 2 is less than f-. and not less thall z-m If now the intervals &
P qN zpN
0
are considered as intervals of variation of Fzj where 0 = arg x, and the two
extreme intervals joined into one, weobtain the desired dissection of r into arcs&*?
When we are studying the arc &, we write
-2pxi

(1, 43) X = e q X = e&)X= e,(p)e-‘,

(1. 44) Y =$?+iu.

The whole of our work turns on the behaviour of f(x) as Iz I- 1, v - 0, and

we shall suppose throughout that o < 17--
_<_5.
2 When =1;varies on &, X varies
on a congruent arc &,q, and
zpz
0 =- argx--
i Q 1

varies (in the inverse direction) over an interval - +9tp,420 ~I&. Plainly 8,, q

and O&,q are less than -2fl and not less than --% so that
qN ClN
A
&q = Max (t&q9 &,J < - l

In all cases Y-8 = (q + id)-” has its principal value

exp (-8 log (q + io)),

wherein (since r is positive)

-- 1x<310g(.)il+iO)<f7c.
2 2

By N,(n) we denote the number of representations of n by a sum of r primes,

attention being paid to order, and repetitions of the same prime being allowed,
so t*hat

r The distinction between major and minor arc8, fundamental ilm our work on Waring’s
Problem, does not arise here.

566
 Partitio nuMerorum. III: On the expression of a number as a sum of primes. 7

1
(1. 47) 2 Yf-0n xn = (f(X))?
n- 2

Finally S, is the sing&r series

T
(1. 4%) I,*=
Q c,(-- ?a).

3. Prdiminary lemmas.

2. I. Lemma I. If p=%(Y)>0 then

(2. II) f(x) = fJX) + f&L

where

(2. 12) fJx) = &d(?i)P-2 log a(x*H3+ xa3+ *- -),

k, nl > I ti

(2* 14)

(2. 16) 24
IC /cI z--
---’
h

567
 8 . G. El. Hardy and J. E. Littlewuod.

We have

= 2 e* (pi) 2 d4(lq + j) e-(Zq+j) y

1--<s’<
---_q,(q,j)-1 z-0
2+p3

=r,eq(pj)&qlq -I-j)& / I'-*r(s)(Ep+pdS,

I
i 2 -'&
2 '

2 +,;a,
= I Y-T(s) Z(s) as,
2 '2% 2 cl
2--icD

where

Since (q;j) = I, we have 1

and so

where

Since Xk(j) = o if (q, j) > I, the condition (q, j) = I may be omitted or retainecl
at our discretion,
Thus”

568
 Partitio numerorum, III: OH the expression of & number as 8 sum of primes. 9

Again, if k > I we have1

GC
j-1 m-1

If jjk is a primitive character,

Vp
IQk I _ . ..-•
h

If x is imprimitive, it belongs to & = $3 where d> I l Then Yk

c ‘m) has the
period &, and

The inner sum is zero. Hence Cl, = o, and the proof of the lemma is completGP
2, 2. Lemma 2. We have

(2. 21) If,(x)1 < A(log (q+1))&

We have

But

< A log (q + I) log q 2 Ix I”‘< A(Iog (q + I))~ &j2r

r-l r=l

< A(log (q + I))A log ; <A (log (q + I,,+*

___-_- --- .-~- ,._- -._----_
l L:\ndau, p. 485. Th.le result iis stated there o~Iy for R yriluitive character, but tile proof
is valid also for an imprimitive character when (p, q) = I l

a Landau, pp. 4% 489, 4~

s See the additions1 note at the end.

569
 10 G. H. Hardy and J. E. Littlewood.

Also

and so

From these two results the lemma follows.

2. 3. Lemma 3. We have

(2. 31)

where

the ~‘8, b’s, b’s and t’s are constants depending upon Q and X, Q is o ov 1,

(2. 32) 6I E L
and
(2. 33) ozb<A log (q+x).

AH these results are classical except the last2

The precise definition of b is rather complicated and does not concern US.
We need only observe that b does not exceed the number of different primes
that divide 41$ and so satisfies (2. 33).
2. 41. Lemma 4. If o<y&-2 then

where

@k

1 Landau, pp. 509, 510,519.

1 Land2ku, p, 511(footnote).

570
 Partitio numer0ruxna HI: On the expressionof a, uumber as a sunz of pri’mes. ‘i 1

(2. 414) 6 = arc tan rl m

PI
We have, from (2. 13) and (2. 14),
2+im

(2. 415) f(2x 1 = 2nz i' Y-~I‘(s)Z(s)da

I

2---&l

say. But 1

where

[f(a)>, denoting general1y the residue of J(s) for s = O.

Now z

--- = 1% n? c Ev---log flu +z c ---

L’(s) EVlog a;,
L(s) +z
& v - 1tip-"v a'l-" - EY
V
v-1

8+a -I
I L(I-8)
pe ---,
-- 2 4 2 1 1 -L( I-8)

where & is the divisor of q to which 3~belongs, i: is the number of primes which
divide Q but not Q, a,, aa, . . . are the primes in question, and Ed is a root of

unity. Hence, if 0 = - I-9 we have

4 --
-A
application
1 This of Cauchy’s Theorem may be justified on the lines of the classical
proof of the ‘explicit forlllulae’ for #(z) and X(X): see Landau, pp, 333-368. In this case the
proof is much easier, since Y D-8r(s) tends to zero, when 1 t 13 a, like an expon.ential ?lrl.
Compare ppm 134- 135 of our memoir ‘Contributions to the theory of the Riemann Zeta-function
and the theory of the distribution of primes’, Acta Mathemuticu, vol. 41 (1917)~ pp. rxg-Ig6.
a Landau, p. 517~
 Again, if s = - I+it, Y=y+iO, we have
4

1 Y-8 1 = 1 Y I4 ex%p(” arc tan t) 3

1 Y-“I‘(s)I<AI YI:(ltl+z)-:,,p(- ~~:-3rc tan!.-y)l+

I
1
<A,YIi It I--2 e---d 1t 1’Y
10gtltl + 2)
and SO

(2. 418)

2. 42. T;TTenow consider R. Since

--A..--+ r =o (s=o),
z( S-,0 eI
we have

.F A,(b + b)- (b- b) (A, + A, log Y) + C,(O) + &(\I) log I’,

where each of the C’s has one of two absolute constant values, according to the
value of c1, Since

w<-b<~,
-- --- o.<.b<A log(q+I)> Ilog PI<AIog ‘<Al;-:,
rIi1

572
 Psrtitio numerorum. III : On the expression of n number as a sum of primes. 13

From (2. 419, (2. 416), (2. 4r8), (2. 421) and (z 15) we deduce

11bI+ I;- “+I$r:),

l

(2. 423)

Combining (2. 422) and (2. 423) with (2. II) and (2. ZI), we obta,in the result of
Lemma 4,
2. 5. Lemma 5. If q > I and xrc is a primitive (and therefore non-principal I)
dhmter, then

( 2. jr)

2ohere

(2.VI) 1C(I)I= xq- “IL(o)1 (Q= I),

(2. 524 pd(I)~=zq-qL’(o)( (e=o).
Further

(2. 53) I - Cl_<__-

9i (e) =7
<- Cl,

rind

This lemma is merely a collection of results which will be used in the proof
of Lemmas 6 and 7* They are of very unequal depth. The formula (2. 51) is
classical.2 The two next are immediate deductions from the functional equation
for L(s)? The inequalities (z. 53)
c follow from the functional equation and tlhe
- -_-- -~ -._- ~ ------ -. -. _---._ _-.
1 Landau, p, 480.
i! Landau, p, 507.
’ Landau, pp. 496, 4g7*

573
 absence (for primitive x) of factors I - E,,G;S from L. Finally (2. 54) is due
to GRONWALL.~
2. 61. ibuna 6. I/M(T)
Z’R the number of zeros q of L (s) for which

o<T-qyJ<T+1,
---
-- -- -_--
-_
then
(2. 611) M(T)<A(log (q+I))A log(T+z).

The e’s of an imprimitive L(s) are those of a certain primitive L(s) corres-
ponding to modulus &, where & (q, together with the zeros (other than s = o)
of certain functions

where
I &yI = L a,lq*
__--.-----1--1_4- __---L.---.--. - ----_ ---------Pp..-- -
l T. H. GRONWALL, ‘Sur les &ries de Dirichlet correspondant A des caract&res complexes’,
Rendico~~f~ de1 C%mZo Matemat& di Palemno, vol. 35 (lgq), pp. 145-159. Gronwall proves that

for every complex x, and states that the same is true for meal x if hypothesis B (or a n~.~uh
less stringent hypothesis) is satisfied. LAXTDAU (‘nber die Xlassenzahl imagingr-quadratischer
Zahlkijrper’, Giittinger Nuchrichten, x918, pp. 28j- 295 (p. 286, f. n. 2)) has, however, observed
that, in the case of a real x, Gronwatl’s argument leads only to the slightiy less precise
ineqnaIi ty

Landau also gives a proof (due to HIWKE) that

for the special character 3

( 12 > associated with the f undamell t,al d iscri tninnnt - q.
The first results in this direction are due to Landau himself (‘uber das Xchtverschwin-
den der Dirichletschen Reihen, welche koqlexen Charakteren entsprechen’, Math. AWK&Z,
vol. 70 (x911), pp. 69-78). Landau there proves that

for cotnplex x.
It is easily proved (see 11, 75 of Landau’s lab quoted nmnloir) that

1w> I < A(loga>‘,

so that any of these results gives us more than all that we require.

574
 Partitio numerorum. III: On the expression of a number as a sum of primes. 15

The number of /dv’s is less than A log (q + I), and each Ev has a set of zeros,
on tT= O, at equal distances

~-22% 2%

log ai > log(q) l)’

The contribution of these zeros to M(T) is therefore less-than A (log (q + I))‘;

and we need consider only a primitive (and therefore, if g > I, non-principal) L(s).
We observe :
(a) that CI is t-he same for L(s) and L(s);
(b) that L( s ) and z(s) are conjugate for real s, so that the b corresponding to
E(s) is 6, the conjugate of the b of L(s);
(c) that the typical e of z(s) may he taken to he either c or (in virtue of t.he
functional equat.ion) I -e, so that

is real.
Bearing these remarks in mind, suppose first tlhat 0 = I. We have then,
from (2. 51) and (2. ~zI),

since

(1-f) (I-&) = r.
Thus

(2. Grz) (2%(b)+XI< A log (q-t r).

On the ot#her hand, if a = o, we have, from (2. 51) and (z, ~zz),

and (2. 61,) follows as before.

2. 62. Again, by (2. 31)

L’(I)
(z+ 621) LII)=bfb

575
 16 G. 11. Hhdy and J. E. Littlewood.

for every non-principal character (whether primitive or not). In particular, when

x is primitive, WC have, by (2. 62d, (2. jq), and (2. 33)9

(2. 622) A(log(q+1))A.

Combining (2. GIZ) and (2. 622) we see that

S < A (log (q + I))A

ItIl(b A (log (q+ I.))4

2. ~3~ If now q > J , and 3~is primitive (so tha,t b = o) , and s = z + i T, we
have, by (2. XI), (2. x3), and (2. 624):

<A+Alog (q+G+A (log(q+I))A+A log (ITl+z)

< A (log (q + d)A log !ITl+ z),

Every term on the left hand side is greater than A, and t’he number of terms
is not less than M( 1’). Hence we obtain the result of the lemma. We have
excluded the case q=
I, when the result is of course classicaL
2. 71. Lemmu 7. We have

(2. 71’) Jlw4 (1%tq+ I))4

Suppose first that x is non-principal. Then, by (2. 621) and (2. 54),

576
 Pslrtitio numerorum. HI: On the expression of a number as a sum of primes. 17

We write
(2. 7x3) z=z,+z,*

where Z, is extended over the zeros for which I - @~!R(q) $0 and Z, Over

fhose for which R(q) = O, Now Z, = X’, where X’ is the S corresponding to a

primitive Z(s) for modulus &, where Q j q. Hence, by (2. 623),

(2a 714) 1 < A (log (Qt I))A < A (log (q + xp*

12 I

Again, the Q’S of Z, are the zeros (other than s = o) of

II (I-$+
v v

the &Q being divisors of g and zV an m-th root of unity, where m=rp(&) <ql;
1.
so that the number of &s is less than A log q and

where either ctiV= o or

Let us denote by gv a zero (other than s = o) of I - ev~F8, by & a ey for which

<
le vI --
--I, and by erry a ev for which (gv I> I. Then

Any ey is of the form

27ci(m + tow) ’
ev = logflv -’

where n2 is an integer. Hence the number of, zeros & is less than A log aV or
than A log (q+ I); and the absolute value of the corresponding term in our sum
is less than
A < A log 6
I mvI <Aq log (q-t 1);
(2. 716)
lel
l For (Landau, p. 482) ey = X(P;), where X is a character to modulus Q.

577
18 G. H. Hardy and J. El, Littlewood.

so that

(2* 717) 2 <4 (log(a+ w

I I
Q’V

RISO

(2. 718)

< A (log aJe 3 $ < A (log cq + I))?

m-1

Rem (2. 715), (2. 717) and (2. 718) we deduce

and from (2. 71x), (2. 714) and (2. 7x9) the result of the lemma.
2. 72. We have assumed that x is not a principal character: For the
principal character (mod. q) we have1

Since a=o, E=I, we have

= b-x
_-~.- -
8

Ibl<A log(q+I)+IZ(~-~--~+~~l.
I

This correeponds to (2. 712), and from this point the proof proceeds as before.
 Partitio numerorum. 311: On the expression of a number as a sum of primes. 19

2. %I. Lemma 8. If o < q(-- 25 then

(2e 811) 111

(Ix) + i
fI X=) ‘k-H- CI,GIc + P,
k-1
where
(2. 812) Gk= 2 r(e) Y--,
Qk

(2. 813) (PI<Al/~(log(p+l))A(q+?-:+,Y(:d-:),

(2. 814) d = arc tan 3 4

PI

This is an immediate corollary of Lemmas 4 and 7.

2. 02. Lemma g. If o <r_<;f then

2

(2. 821) f(x>=W@,

where

(2. 822)

(2, 823)

(2. 824)

We have

I?

where Z, extends over $s for which 1~11

-- r j those for which Iyl< Pa

h 2 1 we have

11YkPexp (y arc tan 5)

- (+ arc tan !$j

579
20 G. H. Hardy and J. E, Littlewood.

(since 1YI < A and, by hypothesis R, p(o>. The number M(T) of q’s for which
1yl lies between T and T + I (T _>_O) is less than A (log (q + I))~ log (T + z), by
(2. 611). Hence

III1 Y1@-‘i e-sly1 --< A (log (q + r))A 2 (n + I,“-; log (n + 2) ewdn

n -I 0

< A (log tq + 1,)~8-@-’ log (; + 2) 3

(2. 826) &ll’(e) Y-1 < A (log (q * I)yy yl-%I-@-~ log ; + 2 l

( 1

2. 83. Again, once more by (2. ~II), 2 has at most A (log (q + x))A terms.
2
We write

(2. 831) z2= z,, + 222’I

z 221
applying to zeros for which I - 0.2 pT_ 0, and 2
212 to those flor wFic)r p = o.

1Y-@I= f Y I-Pexp ‘y arc tan R) < A I Y 1-P;

t

and in 2, f, IT (e) I< A. Hence

7

I ~<~IYi-~~211~(e)l<~IYl-~~~~~<A(log(q+I))~(YJ-~.
(2. 832) p, *
I t
Again, in Z, 2, IYI<A and
I
-I<Aqlog(q+I),
Id
by (2. 716); so that

(2+ 833)

Mz *I2 la, < 4 (log (a + IHA*

From (2. 829, (2. 826), (2. 831), (2. 832), and (2. 833), we obtain
 Partitio wmerorum, III: On the expression of a number as a sum of primes, 21

say; and from (2. 8x), (z. 812), (2. 8r3), (2. 821), (z 822) and (2. 834) we deduce

that is to say (2. 823).

2. g. Lemma IO. We have

(20 91) h. = cp(q) > Aq (log q)-4

We have in fact 1

rp(q) > (I -a> e-G--3P (q > qo(4))

1% 1% Q

for every positive d, C being Euler’~ constank

3. Proof of the main theorems.

Approximation to u,(n) by the singulccr series.

3. IX. Theorem A. If r is an integer, r --> 3, and

so that
(30 114

then
nr-1 r--I+(d) nr-1
(3- 113) Yr (n) = ---- s, + 0 ( n 4 (log nJB ) m (G+,
(r -I)! l

-_I_

l Landau, p. 217,

581
 22 G. H. Hardy aud J. E, Littlewood.

where

It is to be understood, here and in all that follows, that O’s refer to the
limit-process n - 00 , and that their constants are functions of r alone.
If 7~2, we have

the path of integration being the circle Ix\= e--EI, where H = $ SO that

Using the Fmey dissection of order IV = [l&J, we have

say. Now

I~-cp’I~I~I(If”-ll+If’-2~I+***+l~1’-11)
< B(I@fy + prpr-‘ll)*
Also fX-~I=e~“<A. Hence

582
 Psrtitio numerorum. III: On the expression of a number as a sum of primes. 23

3. 12. We have r/=H=A and q-<

G I/i, and so, by (2. 8~23)~
?a

where 6 = arc fan -.7 We must now distinguish two cases. If [Oi<----- y, we have
loI

and

Tf on the other hand 17< 10 1$8p,p, we have

Thus (3, 123) holds in &her case. Also 0) --- 1: and

2

(3* *24) IQ,I<Ane+:(logn)A

3. 13. Now, remembering that r-==

> 3, we have

a,
l

l1
< Bh-(r-l} (.f + ,4)2)- 2 r-1) d,g

J
0

< Bh-(r-1) nr--2;

583
 24 G, II. Hardy and J. E. Littlewood.

by (3. 124) and (2. 91).

3. 14. Again, if arg x = q, we have

A
<--log
I- Ix I (-I--
I- I xI1<Anlogn.
Similarly

Hence

r-1+(&“)
<Bn 4 (log n)C

584
 Partitio namerorum. III: On the expression of a number as a sum of primes. 25

From (3. x6), (3, r17), (3# Irg), (3. 131) and (3e 141) We d&ce

(38 142) Y&t) = pJq(-np)lp,* + o(,_, +(@-J(log n,s),

where Zp,* is defined by (3. 11s).

3a 5 In hh* we write X = e- y, dX e- e-53 Y, so that Y varies on the
straight line from q+ iO,,, to q- S&, Then, by (Z 822) and (3. II@,

?I+ $,q
Now
l
t

M
q+p
I

(3* 154 Z Y-‘@dY + 0 ‘lq+ iO~-rdtj

(J
rl --ia
Ypq J

+I+ iOp,p 4l

?p-l
=27&---+o (jq+ iuI-‘do),
(r - I)!
OP

0, = Min (Op,q, &q) =>--I-

Pa
-2qN
l

From (3, qr), (3, x52) and (3. 153), we deduce

585
 26 G, H. Hardy and J. E. LiMewood.

Since 7+3
-- and O>zv zr<r--I--I=<=~--I~ ; and from (3. 142),
=2 2 4
(3. 154), and (3. 155) we obtain
n’--1 r
El (4)
(3* 156) Y,(n) = -~ e,(- np) + O(n~-l+(@-~+h~ n,B)
(r- I)!
I4 !? (P(4)
21 1

3. 16. In order to complete the proof of Theorem A, we have merely to

show that the finite series in (3. 156) may be replaced by the infinite series &. Now

and ~r<r---I+ Hence this error may be absorbed in the second term

of (3. 156), and the proof of the theorem is completed.

Summation of the singular series.

3. 21. Lemma x If

u?hew n is a positive integer and the summation extend,c over all posilive v&es of p
less than and prime to q, p = o being included when q = I, bzlt not otherwise, then

(39 24 cq c-n> = c,(n);

(3a 213) %Pt ( n ) = cq (n) Cq’(4

if (q, q’) = I; and

(34 214)

where d is a common divisor of q and 12,

The terms in p and g - p are conjugate. Hence c,(n) is real. As c,(n)
and c&-n) are conjugate we obtain (3. 212)?
-~--_-_______- -- -- -
1 The argutwmt fails if y = I or q = 2; but c,(n) = c,(-ok) = I, c2(n) = cs( - 91.)= - I.

586
 Yartitio numerwum. TIT: On the expression of a number as a sum of primes, 27

Again

cq(4 C@(n) = Zexp (

P#

where
P = pq’ + $q*

Whon p assumes a set of rp(q) values, positive, prime to Q, and incongruent to

modulus q, and p’ a similar set of values for modulus q’, . thera P assumes a set
of cp(q) rp(a’) = rp (4Qr) values, plainly all positive, prime to QQ’ and incongruent t#o
modulus q$. Hence we obtain (3. 213).
Finally, it is plain that

which is zero unless q 1n and then equal to Q. Hence, if we write

and therefore

by tbo well-known inversion formula of M6bius.l This is (3, 214)~~

3, 22. hmma 12. Suppose that r -> 2 and

Then
(3* 222) S r -*
-0
1 Landau, p. 577*
a The formula (3* 2x4) is proved ~~'RAMANJJAK ('Qn certain trigonometrical sums and t,heir
appIications in the theory of numbers, + Tmns. Camb. Phd. Sot., vol, 22 (IgIg), pp. 259-276 (p. 260)).
It had already been given for n = z by LAX'DAU (Ihdbuch (Igoy), p* 572: Landau refers to it as
a known result), and in the general case by JEXSEN (‘Et ngt Udtryk for den talteoretiske Funk-
tion z p(n) = M(n)‘, Den 3. Sk~difiaviske Ma fematikey-Kongms, K&iatzia 1914 Kristiania (191 j),
p, 145). Ramanujan makes a large nulnber of very beautiful applications of the su~ns in ques-
tion, and they may well be associated with his nalne.
 28 G. H. Hardy and J. E. Littlewood.

ij n and r are of opposite parity. But if TZ and r are of like parity then

-g+ (- I,‘#-I)
(39 2231 4sIr =zcJ
-I)‘-(-#
P

where p is an odd prime divisor of n and

(3* 224)

Let

cq(- n) = A,.

Then

if (qt q’) = I; and therefore (on the same hypothesis)

sines A&, 963, l ’ l vanish in virtue of the factor pt(q) .

3m234 If +n, we have

p(a) = - x, rp(ti) = W--r, ta(n)-!l(ti)=--r,

(3. 231) A I3
= WI
(- TP l

w - 9’

If on the other hand ti/ n, we have

(- 1)’
(3- 232) A a=
(a - i)r-l’

l Since 1c&2) I(2 6, where 6 1n, we have C&I) = O(I) when n is fixed and q-m. Also
by Lemma TO, ~$4) > Aq (log qjeA, Hence the series and products concerned are absolutely
convergent.

588
 Yartitio rwnerorum. III: On the expression of a number as a sum of primes. 29

Hence

(3a 233)

If 7i is even and r is odd, the first factor vanishes in virtue of the factor
for which ti = 2; if n is odd and r even, the second factor vanishes similarly.
Thus S, = o whenever n and r are of opposite parity.
Jf 12 and r are of like parity, the factor corresponding to G = z js in any
case z; and
(- r)r (p - I)'+(- 1P(p - I)
Sr C 2fi I---- -9
( (a- 1)” IN @ (P - I)’ - (- ar 1
a-3

as stated in the lemma.

Proof of the final formulae.

3. 3. Theorem B. Sttppose that r _>_

-- 3. Then, if n and r are of unlike parity,

(39 31) 3%1n) = O(nT-l),

But if n and r a.re of like purily then

9+(n) co ~
zcy nT-1 (P - I)'+(-I)'(P -~ - I)#
(5 34 - 1y-(-
(r-1)! n(P (P I)' 1'

where p is an odd prime divisor of n and

CID
(3* 33) cr s
(- I)’
n( I - ------~--- .
(a’- IY 1
e-3

This follows immediately from Theorem A and Lemma 12.~

3. 4. Lemma 13. If r => 3 and ~YLand r ure of like purity, then

v,(n) > Bnr-l,

for nl n,(r>+
- _-~----.
l3esults equivalent to these are stated in equations (5, II)-(5. zz) of our note 2, but
iricorrectly, a factor
(log n>-

being omitted in each, owing to a momentary confusion between V&Z) and A$(?&). The ~~(12)

of z is the N,(n) of this memoir.

589
 30 G. H. Hardy and J. E. Littlewood.

This lemma is required for the proof of Theorem C, If r is even

If r is odd

(P- rp-p++ -- (P --I>'-p OQ ti

n( (P - I)‘+ I: i
>
n(
-
(P - 1)’ 1
>
n(
a-3
I------
(a-1)3 ) =
A
’

In either case the conclusion follows from 13. 32).

3. 5. Theorem C. I/ r) 3 and n und r are of like parity, then

We observe first that

VVrite now
1
(3. 5d 9%= Y, t 3’r1r, N r = N’, + N”, ,
where Y$ and N’, include all terms of the summations for which

Then plainly

(3m 513) v’,(n) L- (I - 6>’ (log njr Nfr(n).

But q.(n) > Bnr- 1 for nzn,(r), by Lemtia 13; and so

for every positive 8.

590
 Partitio numerorum. 1X1: On the expression of a number as a sum of primes. ,31

am (3. 54, (3m54, (39 513), and (3. c514) we deduce

(I -dy (log ~z)~(N~--N”,)<z(~- u”,<(log

- 7a)rNrr

( I - 6)’ (log 7-q N, <

- Y,-+ oh,) -< (log njr Nr,

As 6 is arbibary, t,his proves (3. 51).

3. 6. Theorem D. Ewry large odd number n is ihe sum 01 three odd p imes,
The asymptotic formula for the number of representations M,(n) is

n” (P --I)@--)
(3+ 61) N3b) cv cqofl TT( pe-- -3P-I-3 1’

where p is a prime divisor of n and

This is an almost immediate corollary of Theorems B and C These theo-

rems give the corresponding formula for N,(n). If not all the primes are odd,
two must be z and ‘YL -4 a prime. The number of lsuch representat-ions is one ,
at most.

Theorem Every large even number n is the sum of four odd primes (of
E.
which one may be assignta.1 The asymptotic formula fur the lotal number of repe-
sentations is
n3 --HP”--3P-t 3)
(3m 63) N,(n) cv% --
3 4(log n>”

where p is an odd prime divisor o/ n and

m
I
(34 64) C zzz
4
I --- l

n( (a- I)4
1

a-3

This is a corollary of the same two theorems. We hnve only ta observe

that the number of representations by four primes which are not all odd is
plainly O(az). There are evidently similar theorems for any greater value of r,

591
 32 G. H. Hardy and J. E. Littlewood.

4. Remarks on 9Goldbach9s Theoremf.

Our method fails when r = z It does not fail in principle, for it

4* I. l

leads to a definite result which appears to be correct; but we cannot overcome

the difficulties of the proof, even if we assume tbat @= 4. The best upper
2
1
bound that we can determine for the error is too large by (roughly) a power n4.
The formula to which our method leads is contained in the following
Conjecture A. Every large elleghnumber is the sum of two odd primes. T?de
asymptotic formula for the number of representatives is

(49 11) N,(n) CKJ2C

where p is an odd prime divisor of n, and

(4. 12)

We add a few words as to the history of tJhis formula, and the empirical
evidence for its truIW
The first definite formulation of a result of this character appears to be
due to SYLVESTER", who, in a short abstract published in the Proceedings of
London Mathematical Society in 1871, suggested that

where

Since

Jp.I..I.)=n
a<G(I-(+2) a<&
n (+4I ti<G(I-$
-- .--- --- -~
1 As regards the earlier history of ‘Goldbach’s Theorem’, flee L. E. DICKSON,E&tory of
the Theory of Nambem, vol. I (Washington IgIg), pp. 42x-425.
a S. J. SYLVESTER, ‘On the partition of an even number into two primes’, Proc. Lotdon
Muth. Sm., ser. I, vd. 4 (187 I), pp* 4-6 (Math. Pupem, vol. 2, pp. Tog- 7 11). See also ‘On the
Goldbach-Euler Theorem regarding prime numbers’, Nature, vol. 5s (1896~7), pp. 196-197, 269
(Math. Papers, vol. 4, pp. 734-737)4
We owe our knowledge of Sylvester’s notes on the subject to Mr. B. M. WITAON OE Trinity
College, Cambridge. See, in connection with all that follows, Shah and Wilson, I, and Hardy
and Littlewood, 2.

592
 Partitio numerorum. 111: On the expression of a number as a sum of primes. 33

and 1

where C is Euler’s constant, (4. 13) is equivalent to

and contradicts (4. II), the two formulae differing by a factor z e-C= I: , 123 , . l

We prove in 4. 2 that (4+ XI) is the only formula of the kind that can possibly
be correct, so that Sylvester’s formula is erroneous. But Sylvester was the first
to identify the factor

to which the irregularities of N,(n) are due. There is no sufficient evidence to

show how he was led to his result.
A quite different formula was suggested by ST;~CKEL~ in 1896, viz.,

N,(n) cv-- --II(&)*

n
ilog n)e

This formula does not introduce the factor (4. 16), and does not give anything
like so good an approximation to the facts; it was in any case shown to be
incorrect, by LANDAU 3 in rgoo.
In IgrS there appeared an uncompleted essay on Goldbach’s Theorem by
MERLTN." MERLIN does not give a complete asymptotic formula, but recognises
(like Sylvester
. before him) the importance of the factor (4. 16).
About the same time the problem was attacked by BRUNT. The formula
to which Brun’s argument naturally leads is

l Landau, p. 218.
a P. ST~CKET~, ‘ober Goldbach’s empirisches Theorem : Cede grade Zahl kann als Summe
van zwei Primzahlen dargestellt werden’, Giittingw Nachrichten, x896, pp. zgz-zgg.
* E. LAXDAU, ‘ober die zahlentheoretische Funktion co(rt) und ihre Beziehung zum Gold-
bachschen Satz’, Gtittingw Ndwichten, fgoo, pp. 177-186.
4 3. MERLIN, ‘Un travail sur lee nombres premiers’, Bulletin des sciences mn thdm aiiques,
vol. 39 (xgq), pp. 121-136.
’ V. ‘uber das Goldbachsche Gesetz und die Anzahl de,r Prirnzahlpaare’,
BRUN, Archiv fez
Muthematik (Christiania), vol. 34, part 2 (x915), no. 8, pp. I -x5. The formula (4. 18)is not actually
formulated by Brun: see the discussion by Shah and Wilson, I, and Hardy and Littlewood, 2.
See also a second paper by the same author, ‘Sur les non&es premiers de la forme ap+ b’,
ibid., part. 4 (IgIrj, no. 14, pp. r-g; and the postscript to this memoir.

593
 (40 17) N,(n)cuzHnn
P
where

This is easily shown to be equivalent to

(4- 18) N,(n) cv 8 e--2Y c,

and differs from (4. II) by a factor 4e-2C= I .263 . . . The argument of 4. z
will show that this formula, like Sylvester’s, is incorrect.
Finally, in 1916 STXCKEL 1 returned to the subject in a series of memoirs
published in the Sitxungsberichte der Heidelberger Akademie der Wissen~chajten,
which we have until very recently been unable to cons&. Some further remarks
concerning these memoirs will be found in our final postscript.
4. 2. We proceed to justGfy our assertion that the formulae (4. 15) ancl
(4* 18) cannot be correct.
Theorem F. Suppose it to be true fhaP

N,(n) m A ---E---n (2)P

(4. 21) (lug 7Q2 p p - 2

if
n= 2a paplal . . . (a > 0, a, a’, l l l > 01,

and
(4m4

if n is Q&L Then

(49 23) A=zC,=

l P. STACKEL, ‘Die D~rstellung der geracien Zahlen als Summen van zwei Yrimzahlen’, 8
August 1916; ‘Die Lfickenzahlen r-ter Stufe und die Darstellnng der geraden Zahlen als Sum-
men und Differenzen ungerader Primzahlen’, I. Teil 27 Dezember 1917, II, Teil 19 Januar 1918,
111.Teil rg Juli rgr8.
’ Throughout 4, z A is the same contitant,

594
 Partitio numer0rum. III: On the expression of a number as a sum of primes. 35

Write

(49 24) iii(n)= AnTI[ (E) (n even), a(n) =o (7~ odd).

P
Then, by (4. 21) and Theorem C, now valid in virtue of (4m ZI),

it being understood that, when n is odd; this formula means

Further let

these series being absolutely convergent if w(s) > 2, D?(u) > 1. Then

-I)@'- I). .
2 -au
l
(P
C --au I-tZct’U
AZ P P
+ ’ ’ (p - 2) (p’ - 2) rn . I
00

say. Suppose now that U- I, and let

= a-- I)2 - z-. I

N (a -p--1 I c,
Hence

_ 595
 36 G. H. Hardy and 3. IL Litt,lewood.

On the other hand, when AZ- I,

and so1

It is an elementary deduction z that

lvhen s-z; and hence” that (under the hypotheses (4. 21) and (4. 22))

(4m29) J( s1 N-I.s-2

Comparing (4. 27) and (4. 29), we obtain the result of thf theorem.
4. 3, The fact that both Sylvester’s ancl Brun’s formulae contain an
erroneous constant factor, and that tOhisfactor is in each case a simple function
of the number s-c, is not so remarkable as it may seem.
In the first place we observe that any formulu in the theory of primes,
dedwed from considerations of probability, is likely to be erroneous in just this
Wag. Consider, for example, the problem ‘what is the chance that a large number
n should be prime?’ We know tha.t the answer is that, the c!lance is approxim-

ately -I-.
log n
Now the chance that n should not he divisible by any prime less t#han a
fixed number x is asymptotically equivalent to

n (1-s);
a’<X

l We here use Theorem 8 of our paper ‘Tauberitin theorems concerning power series and
Dirichlet’s series whose coefficients are positive’, Pnx London Math* Sot., ser. 2, vol. 13, pp,
174~-rgz* This is the quickest proof, but by no nleans the most eIementary. The formula
(4. 28) is equivalent to the formula

used by Landau in his note quoted on p. 33,

2 For general theorems including those used here as very special cases, see EC. KNOPP,
‘Divergenzcharactere gewisser Dirichlet’scher Reihen’, Actn Maihemnticn, vol. 34, 1909, pp. IQ-
q (e. g, Satz 111, p. 176)~

596
 Yartitio numerorum. 111: On the expression of a number as a sum of primes. 37

and it would be natural to infer1 that the chance required is asymptotically

equivalent to

But B

and our inference is incorrect, to the extent of a factor w-c.

It is true that Brun’s argument is not stated in terms of probabilities 3,
but it involves a heuristic passsge to the limit of exactly t#he same character
as that in the argument we have just quoted. Brun finds first (by ,zn ingenious
use of the ‘sieve of Eratosthenes’) an asymptotic formula for the number of
representations of n as the sum of two numbers, neither divisible by any fixed
number of primes. This formula is correct and the proof valid. So is the first
stage in the argument above; it rests on an enumeration of cases, and all refe-
rence to ‘probability’ * is easily eliminated. It is in the passage to the limit
that error is introduced, and the nature of the error is the same in one case
as in the other.
4. 4. SHAH a,nd WILSON have tested Conjecture A extensively by comparison
with the empirical data collected by CANTOR, AUBRY, HAUSSNER, and RXPEBT.
We reprint their table of resu1t.s; but some preliminary remarks are required.
In the first place it is essential, in a numerical test, to work with a formula
N&2), such as (4. II), and not with one for p,(n), such as (4. ~5)~ In our
analysis, on the other band, it is v,(n) which presents itself first, and the formula
for N,(n) is secondary. In order to derive the asymptotic formula for &(n),
we write
v2(n) = 2 log G log a’ cv (log 7.~)~ N&) l

The factor (log YZ)~is certainly in error to an order log n, and it is more natural5
to replace v&2) by
((lq %I2- 2 log n + l ) N, (n).
 38 G. EL Hardy &ad J. E. Littlewood.

For the asymptotic formula, naturally, it is indifferent which substitution

we adopt. But, for purposes of verification within the limits of calculation, it is
by no means indifferent, for the term in log n is by no means of negligible
importance; and it will be found that is makes a vital difference in the plausibility
of the results, Bearing these considerations in mind, Shah and Wilson worked,
not with the formula (4+ II), but with the modified formula

Failure to make allowances of this kind has been responsible for a good
deal of misapprehension in the past. Thus (as is pointed out by Shah and
Wilson l) Sylvester’s erroneous formula gives, for values of n within the limits
of Table T,’ decidedly belter results than those obtained from the ww~~dified
formula (4. II).
There is another point of less importance. The function which presents
it-self most naturally in our analysis is not

f(x) = 2 log 23xa

bUt

The corresponding numerical functions are not v,(n) and N,(n), but

(so that Q2(n) is the number of decompositions of n into two primes OT two powers
of primes). Here again, N,(n) and Q,(n) are asymptotically equivalent; the diffe-
rence between them is indeed of lower ordtlr than errors which we are neglecting
in any case; but there is sometJhing to be said for taking the latt,er as the basis
for comparison, when (as is inevitable) the values of n+ are not very large.
In the table the decompositions into primes, and powers of primes, are
reckoned separately; but it is the total which is compared with I. The value
of the constant 2 C, is I .3203. It will be seen that the correspondence between
the calculated and actual values is excellent.

l 2. c., p. 242.

598
Partitio numerorum. III: OH the extression of a number as a sum of urimes. 39

Table I.
-
12

30 = 243.5 6+4 =IO 22 0 * 45

32 = za 4+7 =I1 8 I .38
34 = 2. I7 7+6 =x3 9 I a44
36 = z2 l 3z 8+8 =I6 17 0 * 94
--
210 = 2-3-547 42 + 0 = 42 49 0 .85
214 = 2.107 I7 + 0 = I7 16 I.07
216 = 9. 33 28 + 0 = 28 32 o. 88
256 = z8 16 + 3 = 19 17 I l IO
--
2,048 = 2” 50 + 17 = 67 I .06
63
2,250 = 2 l 32, 53 I74 + 26 = 200 179 I. II

2,304 = 9.3’ x34 + 8 = x42 136 I .04

2,306 = 2. I 153 67 + 20 = 87 69 I .26
2,320 = 2.3.5.7.11 228 + 16 = 244 244 I . 00
.--~
3,888 = z4. 3’ 186, + 24 = 210 197 I .06

3,898 = 2 l 1949 gg + 6 = 105 99 I .06

3,990 = 2.3+7.rg 328 + 20 = 348 342 I l 02

4,096 = 2’” 104 + 5 = 109 I02 I .06

--
4,996 = 2’: l ‘249 q24 + 16 = 140 x*9 I m I8
4,998 = 2 ‘ 3 , 7” ,417 228 + 20 = 308 305 x .OI

5,000 = z3. s4 rjo + 26= 176 I57 I I2

-.~
8,190 = 2 4 32. 5.7 ,r3 578 + 26 = 604 597 I .OI
8,192 = 2’” 150 -+ 32 = r82 171 I .06
8,194 = 2 , 17 ,241 192 + IO = 202 219 0.92
- --
10,008 = 2?, 3? 139 388 + 30 = 418 396 x .06
10,010=2.$7.1x.13 384 -I- 36 = 420 384 I IO9
10,014 = 2,3. 1669 408 + 8 = 4x6 396 I A5
- --- --
30,030 =2.3.~.7.11.13 1,800 + 54 = 1854 1795 I .03
36,960=2:‘.3.5.7.~ 1,956 + 38 = 1994, 1937 I x3
39,270 = 2m3.~.7*11.17 2,152 + 36 = 2188 2213 0 .99
41,580 = 22 .3”.5.7.rr 2,~40 + 44 = 2184 2125 I .03

50,026 = 2.25013 702+8 ~710 692 I .03

50,144 = 25 I 1567 60: -t 32 = 706 694 I l 02

--

170,166 = 2 - 3.79 ’ 359 3,734 + 46 = 3780 3762 I ,oo

170,170 = 2.5.7,II.I3,X7 3,784 --I- 8 = 3792 3841 0 s 90

170,172 = 22, 3”. 29, 163 3,732 + 48 = 3780 3866 0.98
--

599
 40 G.+ 13. Hardy and 3. E. Littlewood.

0 ther probl

5 I. This last section is frankly conjectural, and is not to be judged by

the same standards as $5 r-3.
The problems to which we have applied our method may be divided roughly
into three classes. The typical problem of the first class is Waring’s Problem.
Our solution of this problem is not yet its conclusive as we should desire, and
we have not, exhausted the possibilities of our method, even when allowance is
made for still unpublished work; WC;cannot at present, prove, for example, that
every large number is the sum of 7 cubes or 16 biquadrntes. But our proofs,
so far as t-hey go, are complete.
The typical problem of the second claPs is that, considered in $5 r-3. The
arguments by which we prove our results are rigorms, but the results depend
upon the unproved hypothesis R.
There is a third class of problems, of which Goldbach’s Problflm is typical.
Here we are unable (with or without Hypothesis R) to offer anything approaching
to a rigorous proof. What our method yields is a formula, and one which seems
to stand the test of comparison with the facts. In this concluding section we
propose to state a number of further formulae of the same kind, Our apology
for doing so must be (I) that no similar formulae have been suggested
I before,
and that the process by which they are deduced has at any rate a8certain
algebraical interest, and (2) that it seems to us very desirable that, (in default
of proof) the formulae should be checked, and that we hope that some of the
many mathematicians interested in the computative side of the theory of numbers
may find them worthy of their attention.

Conjugate problems: the problem of prime-pairs.

5. 2. The problems to which our method is applicable group themselves in

pairs in an interesting manner which will be most easily understood hy an example.
In Goldbach’s Problem we have to study the integral

where

600
 Partitio numerorum, III: On the expression of a number as a sum of primes. 41

The formal transfol’mations of this integral to which we are led may be stated
shortly as follows. We divide up the range of integration into a large num-
ber of pieces by mea’ns of the Farey arcs &,q, + varying over the interval

(-----
2 p:‘G

q - p’q’
0
4
\ 2FX
-~ -t OP4l when x varies over &,,*. We then replace f(x) by the

appropriate approximation

,11(q) I P(59
(P(4)log --.- ‘p(4)Y&f
----=-
MP) I
( X 1 n

4)-CP-E by u, and the integral

Q

(5 22)

bY.
a3
,l-iw
(59 23) n ep (- n p) ----? dw = z 7r;nep (- np).
s (I - zwy
-Xi

We are thus led to the singular series S,.

Now suppose that, instead of the integral (5. zr), we consider the integral

23-c

(5 24) J(R) = -& f (R&) f (Re-“w) ekiwd&

27c .i‘
0

where now II: is a fixed positive integer. Instead of (5. 22), we have now

ekiu
-----
e,(k
(;-iu) (; + iujdZl

601
4.2 G. H. Hardy and J. E. Littlewood.

We are thus led to suppose that

(5 25) J(R) CA2 P

---(d 2e, (kp)
2 a cf(P) 1
-- 1
when B-e Ta, ?&--+ob,
The series here (which we call for the moment S’,) is the singular series 8,
with -- k in the place of n. On the other hand

if both G and G + k are prime, and ati = o otherwise. Hence we obtain

--
Here R-e It, but the result is easily extended to the case of continuous ap-
proach to the limit, I, and we deduce’

And from this it would be an easy deduction that the number of prime pairs
differing by II:, and less than a large number n, is asymptotically equivalent to

We are thus led to the following

Conjecture B. There are infinitely mnny prime pairs

for every even k. If P,(n) is the number o# @-s less than n., then

where C, is the constant of 5 4 und p is an odd prime divisor of k.

-- - --~~
1 We appenl again here ta the Tauberian theorem referred to at the end of 4* 2 (f. II. I).
This time, of course, there ia no question of an alternative argument.
2 Kate that S’, = o if k it3 odd, RB it should be.
 Part itio numerorum, III: On the expression of a. number as a sulzl of primes. 43

It will be observed that the analysis connected with Conjectures A and R,

which deal respect-ively with the equations
n-a+& d=a+k,
is substadially the same. It is pairs of problems connected in this manner that
we call conjugate problems I

Numerical verifications.
5. 31~ For k- 2, 4, 6 we obtain

(5 313)

Thus there should be approximately equal numbers of prime-pairs differing by z artd

by 4, but abo& twice as rr~any differing by 6 The actual numbers of pairs,
below the limits
IOO, 500, 1000, 2000, 3000, 4000, 5000
are

The correspondence is as accrrrate as could be desired.

5, 32, The first formula (5. 311) has been verified much more systematic-
ally. A little ca(d,ion has to be exercised in undertaking such a verificafion.
The formula (s* 26) is equivalent;, when k: = 2, to

and, \&en we pass from this formula to one for the number of prime-pairs, the
formula which arises most naturally is not (5. 311) but2
--_-p-p
1 This formula follow from (5, 321) in exactly the salne way that
x(x) cv Li x
follows fronl

603
 44 CL H. Hardy and J. E. Littlewood.
n1
dX
(5 322) P,(n) cv 2 c, --
(log s)L;

indeed it is not unreasonable to expect this approximation to be a really good

one, and much better than the formulae of 4, 4. The formula (5, 322) is nat-
urally equivalent to (5. 311). But

n
- dX.-- =-- r+2cp+ _.._. 31--+.., ,l
(log x)" (b;?a)y ( Iog 9% (log ny 1
s

and the second factor on the right hand side is (for such values of n as we
have to consider) far from negligible. It is for this reason that Brun, when he
attempted to deduce a value of the constant in (5. 311) from the statistical
results, was led to a value seriously in error.
We therefore take the formula (5* 322) as our basis for comparison, choosing_.
the lower limit to be 2. For our statistics as to the actual number of prime-
pairs we are indebted to (u) a count up to ~oo,ooo made by GLAISRER in 1878 z
and (b) a much xnore extensive count made for us recently by Mrs. G. A,
STREATRMLD. The results obtained by Mrs. Streaffeild are as follows.
--_--__._,- - .--_ ----_..-_ ---___. -- -.----_
I
n
’ dx
Ratio
2c2 (log7
‘2i 1

100000 1224 1246,3 I ,018

200000 2159 1 2179.5 1 l 009

300000 ( 2992 / 30354 ! I,Olj

400GOO 1 3801 3846 l 1 I .012

I I
500000 4562 4625.6 I .oq

600300 5328 5381 ’ 5 I .OIO

/“ooooo GO58 6118.7 I 010

800000 6763 6840 . 2 I .OII

~00000 7469 7548 ’6 I .or1

T000000 8164 8245 .6 1.010

--- _-.- ....-.- ---.-- ..._
_--.--_ ,---- _
1 The series is of course divergent,, and must be regarded as closed after a finite Ilumber
of terms, with an error term of lower or&x than the last term retained.
* J. ‘w, L. thAISHEEt, ‘An enu tnwntion of priule-pairs’, Messenger of Mtrthematics, vol. 8
(1878), pp. 28-33, Glaisher counts (I, 3) as a pair, SO that his figure exceeds Our8 by I.

604
 Partitio numerorum. III: On the expression of a number as a sum of primes.

5. 33, Similar reasoning leads us to the following more general resu

Conjecture C. If a, b are fixed positive integers and (a, b) = I, and N(n
the number of represenkdions of n in the fork
n=m+ba’,
then

N(n)

unless @,a)=~, (n, b)=1, and one und only one of n, a, b is wed But if
these conditions are satisfied then

2c
N(n) cv -J
ab

where C, is the constant of 6 4, and the product extends over all odd primes p which
divide n; a, or 6.
Conjecture D. If (a, b) = I md P(n) is ihe number of puirs of solutions of

such that a’ < n, then

unless (h7, a> = I, (k, b) = I, and just one of E, a, b is even. But if these conditions
are satisfied then

where p is un odd prime factor of E, a, or b.

It should
be clear that the theorems proved in $5 r-3 must be capable of
a similar generalisation. Thus we might investigate the number of representa-
tions of n in the form
n =aa+bd+cd’;

and here proof would be possible, though only with the assumption of hypo-
thesis R. We have not performed the actual calculations.

l This is trivial. If M and a have a common factor, it divides bti’, send must therefore
be a’, which is thus restricted to a finite nulnber of values. If ut, a, b are all odd, G or P”
must be 2.

605
 46 G. II. Hardy and J. E, Litt!ewod.

Primes of the forms rn” + I, urna + bm +- c.

5. 41~ Of the four problems mentioned by Landau in his Cambridge address,

two were Goldbach’s problem and the problem of the prime-pairs. The third
was that of the existence of un infinity of primes of the form m2 + I. i
Our method is applicable to this problem also. We have now to consider
the integral.

where f(x) is the same function as before and

3(x) = ix”‘2.
m-1

The approximation for 3(X) =3(Rewiv) on & is

where

-
and sP>Pis the conjugate of SPlp: and we find, as an approximation for J(R),

--I
1\n
We replace the ntegral here by

~ -.._ _------ _~_..~ --

l The fourth was that of the existence of a, prime between T?’ and (n + I)~ for every ?a> o.
The problem of prilnes awl2 + bm +c ~nust not be confused with the much sin~pler (though
still difficult) problelu of primes included in the definite quadratic form ax2 + zlxv + c$ in Tao
indspenden t varitcbies, This problein, of course, was solved in the classical researches of LIELA
VAT~T,&EPOUSSIN. Our method naturally leads to de la Vall&e Poussin’s results, and the formal veri-
fication of them in this manner is not without interest, Here, however, our method is plainly
not the right one, md could lead at best to a proof encumbered with an unnecessary hypothesis
and far more difficult than the accepted proof.

606
 Partitio numerorum. III: On the expression of a number as a sum of primes. 47

and we are led to the formula

where S is the eingular series

s- @bp,,r,(--1)).
pt *v-P(9)

Repeating the arguments of 5 5. 2, we conclude that the wmber P(n) of primes

of the form m2 + I and less than n is given asymptotically by

-
ha
(5 413) P(n) cv ----FL
log n

5. 42. The singular series (5. 412) msy be summed by the method of 5 3. 2.
Writing
S=zAq= I + A, -I- A, + #“,

there is no difficulty in proving that APqr = A,A,~ if (Q, q’) = x. Hence we

write 1

B=lIIIxa,
where
k =I+A,+&+q.*=I+A,.

If a’=2, A,=o, x~=I. If ti>~,~

and

---
_--- -__--P

l Even this is a formal process, for (5* 412) is not absolutely convergentU.
LJ See DTRICHLET-DEDEKIND, Vodesunpn iibev ZnhEe~~fheorie, ed, 4 (1894)~ pp. 293 et seq.

607
 48 G. H; Hardy and J. E. Littlewood.

Thus finally we are led to

Conjecture E. There are infinitely many primes of the form rn” -t I. The
nu.mber P(n) of such primes less than n is given asymptotically by
-
P(n)
v
co C -n-3
log ?a
where

c
W

-- I -I
( m’)I
=
II
a---I
e-3

5. 43. Generalising the analysis of 8s 5. 41, 5, 42, we arrive at

Conjecture F, Suppose thai a, b: c are inlhgers and a is positive; that (a, E, c) = I.;
that a +- b and c are not both even; and that D = b2- qac is not a square. Then
there are infinitely m.any primes of th.e form am2 + bm -+ c, The nzJmber P(n) of
such primes less than n is given asymptotically by

where p is a wmmun odd prime divisur of a and b, E is I if a + b is odd awl 2

if a -+ b is even, and

It is instructive here to observe the genesis of the exceptional cases. If

(a, b, c) = d > I, there can obviously be at most one prime of the form required.
In this case xa vanishes for every a for which m 1d, Jf a + b and c are both
even, am2 + bm + c is always even: in this case x2 vanishes. If D = kB, then

4a(anz2 + bm t- C)= (ZU~ + b)2- kB,

and
qua’= (zam + b)‘- k2

involves zam + bf k I4a, which can be satisfied by at most, a finite number of

values of ‘ytz. In this case no factor 3~~ vanishes, but t,he product (5. 4321)
diverges to zero,
5, 44. The conjugate problem is that of the expression of a number n
in the form
(5 441) n= am2+bm+H.

608
 Partitio numerorum. III: On tile expression of a n~mbcr z1s a sum of primes. 49

Here we are led to

Conjecture G, Suppose thatb are integers, and a > o, and let N(n)
a and
be the number of representations of n in the form ame + bm + a. Then if n, a, b
have a commun fuctor, or if n and a -+ b are both even, 01’if b2 + qan is a square, then

(5 442)
v--
N(n)=’ tlogn
J!-1’
In cl11other cases

where p is u cunamon odd prime divisor of a and b, and E is I if a + b is o&l and

z if a -+ 6 is even.
The following are particularly interesting special cases of this proposition.
Conjecture H. Every large number n is either a square or the sllrn of a prime
a12d a square. The number N(n) of representations is given asymptotically by

(5 444)

There does not seem to be anything in mathematical literature corresponding

to this conjecture: probably because the idea that every number is a square,
or the sum of a prime and a square, is refuted (even if I is counted as a prime)
by such immediate examples as 34 and 58. But t#he problem of the representa-
tion of an odd number in the form ‘t3 + zm2 has received some attention; ancl
it has been verified that the only odd numbers less than gooo, and not of the
form desired, are 5 777 and 5 9932
Conjecture 1. Every large odd nqlmber n is the slim of a prime and the dwble
01 a squure. The number iV (n) of representations is g&n usymFtoticaEEy by

1 By STERY and his pupil in 1856. See Dickson’s History (referred t,o on p. 32) p. 424.
The tables constructed by Stern were preserved in the library of H,urwitz, and have been very
kindly placed at our disposal by Mr. G. Pblya. These manuscripts also contain a table of
deconlpositions of primes q = 4 m + 3 into sums q =p + z p’, where JI and p’ are primes of the
form 4 m + I, extending as far as g = zog83. The conjecture that such a decomposition is always
possible (I being counted as a prime) was made by Lagrange in 1775 (see Dickson, 2. c., pa 424)

609
 50 G. II. Hardy and J. E. Littlewood.

5 45 We may equally work out the number of representations of 7~ as

the sum of a prime and any number of squares. Thus, for example, we find

Conjecture J. The numbers of representations of n in the foms

a,re given asymptotically by the formu.lue

(5 450
where:

(5, 454 N(n) L’vTi

2
Cdn” n 1(Pps
- d”(p +-- I)
- p’ + 1 1’

Here p is an odd prime divisola of n, and represeniations which differ only in the
sign or order uf the numbers m, , m,, . . . are counted as distinct.
The last pair of formulae ehould be capable of rigorous proof.

Prublems with cubes.

5. 5. The corresponding problems with cubes have, so far a$ we are aware,
never deen formulated. The problem which suggests itself first is that of the
existence of an infinity of primes of the form ?n3+ z or, more generally, ms + k,
where k is any number other tflan a (posit-ive or negative) cube.
Here again our method may be used, but the algebraical transformations,
depending, as obviously they must, on the theory of cubic residuacity, are
naturally a little more complex. As there is in any case no question of proof,
we content ourselves with stating a few of the results which are suggested.
Conjecture K. If k: is any fixed wmber other than a (positive or negative)
cube, then there are in finitely many primes of the form mg+ k. The number P(n)
of such pGmes less than n is given asymptotically by
1

610
 Part.itio nulmerorum. III: On the expression of a number as 3 sum of priuws. 51

where

I
and (-k), is equal to x or to - 3 according as -k is or is not a cubic residue of a’.

Conjecture L. Every large number n is eitlher a cube or the sum uf a prime

and a (positive) cube. The number N (n> of representations is given asymptotic&y by

the range of v&es of a beinq defined us in K.

Conjecture M. If k is’any fixed number other thun zero, there are infinitdy
many primes of the form ES+ ms -+ k, where 1 and m are buth positive. The number
p(n) of such primes less than n, every prime being counted m=&iply according to
its number of representations, is given asymptotically by

where p and 23’ are odd primes of the form 3r + 1, pj k, a, k, and

A-Z
A n’ =-.--.
a(*- I)

;f - k is a cubic residue of G,

in the conirary cuse. The pus&e sign is to be chosen if

- E 2
-3cci
C es = eP
( co13

crl= a + be being that complex prime factor 01 ti fur which a I_- - I:, b G o (mod. 3) ;
the negative in the contrary event. And A and B are defined by

A-za--b,3B=b,qa=A8+z7BL.

611
52 G. H. Hardy and J. E. Littlewood,

In purticulur, when k = I, the number of primes l3 + m3 + I is given asym.p-

totically by

primes susceptibleof multiple representdon being count& mdiiply.

Conjecture N. There are infinite1 6y many primes of the form k3 -I- I3 t m3,
where k, 1, m are all positive. The number P(n) uj such primes less than n, prime.s
susceptible uf multiple representdun being cuu~ted mdtz’ply, is given asymptotically by

where ti is a prime of the form 3m + r , and A has the meaning explained under MI

Triplets and other seque rices of primes.

5. 61. It is plain that the numbers a, ‘C;I t 2, a+ 4 cannot all be prime,

for at least one of the three is divisible by 3. But it is possible t#hat
G, G + 2, G + 6 or ‘E;I, G + 4, a+ G should all be prime. It is natural to enquire
whether our method is applicable in principle to the investigation of the
distribution of triplets and longer sequences.
The general *case raises very* interesting questions as to the density of the
distribution of primes, and it will be convenient to begin by discussing them.
We write

so that Q(X) = e([~]) is the greatest number of primes that occurs indefinitely
often in a sequence n-+-r, n+z, l , 72+ [x) of [x] consecutive
l integers. Tile
existence of an infinity of primes shows that e(x)>1-- for X> I, and nothing
more than this is known; but of course Conjecture B involves e(x)) - z for x23.
It is plain that the determination of a lower bound for e(x) is a problem of
exceptional depth.
The problem of tin upper bound has gceatep possibilities. We proceed to
prove, by a simple extension of an argument due to Legendrel,
 Partitio numerorum. III: On the expressionof zbnurrlber a8 a sum of primes. 53

Theorem G. If E> o then

Q(x)< (1fE)e-“log%g; c/
(x> x0 =X&F)),

where C is Euler’s con&ant. More generdly, if N (z, n) is the number of the integers
72+1, n-+-z, l . ., n + [x] that uw not divisible by any prime less than or equal to
log x, then

a(x) = rim N(x, n> < (I + e) C-C---L--.- (x> X,,(E)).

78---F)
00 log log x

It is well-known that the number of the integers r, 3-, . , ., [yl, not divisible
by any one of the primes p,, p2, . . , p,, is l

where the i-th summation is taken over all combinations of the Y primes i at,
a time. Since the number of terms in the total summation is zv, this is

We now take pl, p2, . om,pV to be the first P primes, write n +x alld TL
successively for TJ, subtract,, and take the upper limit of the difference as n- 00,
We obtain

as y-m. 1 If we take y = log x, and p, to be the greatest prime not less than y,
we have

Y<p,(logx,
- P=o
(&g,) 3

o(x) < (I +E)e-C--x- (x > x0 (d) J

log log x
the desired result.

. 613
 54 G. 1-i. Hardy and J. E. Littlewood.

An examination of the primes less than ZOO suggests forcibly that

e(x) 54x) (x>- - 2).

But although the methods we are about to explairl lend to striking conjec-
tural lower bounds, they throw no light OII the problem of an upper bound.
We have not succeeded in proving, even with our additional hypothesis, more
than the tielementary@ Theorem G. We pass on therefore to our main topic.
5. 62. We consider now the problem of the occurrence of groups of primes
of the form
72, M-a,, n ta,, . . . . n-ta,,

where a;,a,, . . ..a. are distinct positive integers. We write for brevitv L

Then, if (h, k)= I, we have

If fP =2+(f), r-1, o - 0, and O is sufficiently small in comparison wit.h

Q
I:- r, then
n
f( re--iv ) w *-qe--iO
/(a)
w ?

n --. ,11(q)
Ad
- Y(9)
Let us assume for the moment that

if +==pl+O, r-1, and 0 ia sufficiently small. Then our method leads us to write
Q’

614
 Partitio numerorum. III: On the expression of a wmber as a sum of primes, 55

2x
-I '
m--1 ,,i(,,Y) f(&P)e""i'Pdrp
f
2x .i (
0

on replacing the integral by one extended from - 7~ to YL Thus (5. 621)

suggests that

(5 624

where gm is determined by the recurrence formula

From tollis recurrence formula we obtain without difficulty

where ql. runs through all positive integral values, 13.).through all positive values
less than and prime to qj-; and Q is the number such that

p
QC9, Pi -j- P3
q2 -++ Pm
-;P (P, Q>= I l

If we sum with respect to the p’s, we obtain a result which we shall write in
the form

We shall see presently that the multiple series (5. 6251) is absolutely con-
vergent.
c
l?or greater precision of statement we now introduce a definite hvpothcsis.

615
 Hypothesis X. If woo,
I_ and r--+1, then

where S, is given by (5. 625) and (5. 6251).

Our deductions from this hypothesis will be made rigorously, and we sl~all
describe the results as Thcorem,s 2L I, X 2, a . +
,=j. G3, From (5. 626) it follotlJs, by the argument of 4, 2; that

as x-m, where tlw Eejt-hand side denotes the number of groups of m -+ I primes
31, n-+-n,,. . 0, Wl-U~~ of which all the members are less thau =t:.
We proceed to evaluate S,,. In t.he first place we observe that A(ql, p2,. .,, q,,>
is zero if any 4 has a square factor. Next we have

provided (q,, Q/J = I for all values of r and s. For, if we write

so that ilr = 4,.q’,., and suppose that pr and p’, run through complete set,s of
residues prime to 8,. (or q’l.) and incongruent to modllluS pl, (or q’).), tJhen c,. runs
through a similar set of residues for mudulus iI,.* Also @, Q’) = I and so
(PQ’ + P(C), QQ’) = I, Hence, since

y p,---j--E
I) P’ P&l+ P’Q 7
&c-Q Q’ I
Q&

the Q associated with 2: is Q&L Since ~(qq’) = y(q)y(q’) if (q, p’) = 1, (5 632)
* l . l

follows a6 once.
Assuming then the absolute convergence, more conveniently established
later, of the series and tile pruduet, we have

616
 Partitio mm1ci~orui1~. III: On tile expression of a number as a sum of primes. 57

and where 2 is extended Over all A’ s in which r of the nz places are fiiled by
1’
G’S and the remaining nz -r by 1’s.
Our next step is ta evaluate the A’s correspor ding to a prime a. Writing

place, sa,y the first, is filled

Ql =a, qy== I(?*> I), pf-= o(r> I ,$-a,

and so

When r > I places, say the first rs are filled by G’S, we have similarly

where the $5 run through the numbers r , 2, . , 0, a- I:, and & is determined by

Clearly

Q- jo~o (mod. a)), s=a( p -I-!=o (mod. aj) .

Hence
 58 G. II. Hardy and J. E. Littlewood.

Now

is evickntly a fwction of a, u,, . . +, ur which is unaltered by a permutation of

a,, ’ ’ a,. We denote it (dropping the reference to ti) by B,&,
l 1 a,, . . ., clJ,
the suffix nz being used to recall that a,, a,, . , ,, ar, or rather the n’s that
replace them in the general case, are selected from a,, a,, . . , a,. l

Then

Here we are supposing r > 2. We shall adopt t,he convention B,,Ja,) = o.

5. 64. We now digress forea moment to establish the absolute convergence
of our product and multiple series. We have

(5 641) m(k) = a-- I (~1 k), m(k) = -- I (a-;-k)*

Hence, when rl~ is large, every CG occurring in (5. 635), (5. 636), or (5. 637)
is equal to - I 2 It follows that

and so, since A (ql, q2, . . .) is the product of A’s each invulving only a single
prime a, that the multiple series and the product in (5. 633) are absolutely”
convergent.
5. 65. Returning now to X(G), we have, for r G > I,

the result being true for r= I in virtue of (5* 635) and our convention as to
B, (a,) l Hence

l It is here that we use the condition a, j=n,.

618
 Partitio numerorum. III: On the expression of a number as a sum of primes. 59

= Ym- -C?&,
say, where

the summation being taken over all combinations (without reference t.o order)
of U1, 9. ., a, taken r at a time.
Now

(5.653) L+1-(1-x) L= 1 -x-(~-x)g+x~~~+xc~~a,))(~+xcti(am++-l+x)

P-1
?n

=x(1 - x) + x2( I+ ca(G72+1)) n (I + xcf&)> l

r-1

Also

Here 2’ denotes a sum taken over the combinations of a,, a=, . . . , ant, F--- I: at
a time; and the equation holds even for r = nz + I if we interpret Cm++1 as zero,.
Hence, by (5. 637)s

C mfl,r = CwI,+

and therefore

r-2

619
 60 G. II. Hardy md J. E: Littlewood.

Using (5. 65r), (5. 653), and (6. 654), and observing that x(1-x) =--&x2, we
obtain

(5 655) X,,tr(N - (I--x> X?n(~) = X’(I + C&&7$+1)) jj (I + xw(ar)) -

r-1

To this recurrence formula we add the value of X&Z) for m = I, via.

5, 66. We can now deduce an exceedingly simple formula for X,,(G), viz.

where

(5. 662)

is the nwnber of distinct

residues of o, a,, a, j m. ., a, (mod, a)+
This is readily proved by induction. Let us denote the right1 hand side of
(j. 661) by XL; and let us consider first the case m, = I a
If a,zo (mod. a) we have ~=a:, c&,) =GC--r; if a,+ we have ‘I/---Z,
cd(aJ ==--I. In either case X1 = X’, l

Mow suppose the result true for m,, and consider. XPILtl. There are three

On the other hand I + ~(a,,+l) = a, ~(a,--~)~ +l) = c&Q; the right hand side
of (5* 655) vanishes; and so

X mt1= (1 -- x) x, = (I - x) X’, = X’,+1 l

---a
(ii )
%+1 -- -b
1’17 for some r&m. Here again ‘I’~+~ = vm. On the one
hand we have, as before, Xlrn+l = (I - x)X’,. On the other

I: + W(~,,+I) = 0, x t xcd(a,,--a,+I) = I - I w(o) = 0;

a -1

the right baud side of (5, 665) vanishe6, and Xm+l= X’,+l AXISbefore.

628
 Partitio numerorum. III : On the expressiw of a number as a sum of primes. 61

(iii) anz++, u.~++,(r~m). Here vllLtf= u,+ I = Y + I. Also all the c’s
concerned are equal to -I. Hence

X m+l -(I -X)Xm=--~x2(r-x)m=x(I-x)m+1,

or, since X, =X’,,

X m+l = (1 -xx). (I -Xl”(I + (,I) - 1)x) t x( I - xp+l

= ( I- x)m++ + ux) = X1,+1.

This completes the proof.

We now restate our conclusions in a more svmmetrical form.
Theorem X I.~ Let b,? b,, . ., b, be m distinct&gers,
l and P(x; b,, b,,d . . , b,,) l

the number of gruups n + b, ,, -+ b,, . . . . . r, n + b,,, between I and x anct con,sisting

wholly of primes. Then

(59 663) P(x) cv G(b,, b,, . . ., b,) L&,(a)

wh.en x- 00, where

Y = u(m; b,, b,, . . , bm) is the number of distinct

l residues of b,, b,, . . , , b,, to mo-
dulus a, and

Further

(59 665) G(b,, b,, ’ + ‘J bm) = CmWb,, b,, . . .j &J

where

(5 666)

(50 667)

and A is the product of the differences of the b’s,

l To avoid my possible misunderstanding, we repeat that these theorerm are consequences

of Hypothesis X.

621
62 G, II. Hardy and J. E. Littlewood.

The change from o, a,, . . , am to b,, b,, . , ‘, & is obtained

l by writing
n-b, for n and m for m+ I, The expression of G as the product of the con-
stant C, (depending only on ml and the finite expression H follows immediatelv c
from the fact that Y =m if Gjd, ti>m.

5* 67* There are plainly many directions in which it would be possible to

extend these investigations. We may ask, for example, whether there are
indefinitely recurring pairs, triplets, or longer sequences of primes subject to
further restrictions, such as that of belonging to specified quadratic forms. We
have considered one problem of this character only, which is interesting in that
it combines those contemplated in Conjectures B and IL Is there an infinity
of pairs of primes of the forms VP + I, mB + 3 ? The result suggested is as follows.
Conjecture P. There are infinitely many prime pair& of the form ms + I, m2 + 3.
The number of such peek less than n is given asymptotically by

neither, one, or both of - I and - 3 are qwhztic

where ly is o, 2, or 4 according ~118
residws of a.

Numericul verifications.
5. 68. A number of our conjectures have been tested numerically by Mrs.
STREATFEILD, Dr. A. E. WPSTERN, and Mr. 0. WESTEBN. We &ate here a few
of their results, reserving a fuller discussion of them for publication elsewhere.
The first of these numerical tests apply to conjectures E and P. In applying
such tests we work (for rewona similar to those explained in 4 I 4 and 5 .32) not
with the actual formnlae stated in the mmciations of those conjectures, but
with the asymptotically equivalent formulae

n
3 dx
(5* 682) Q (N m -2 C CW”CLi,VG.
s G(log xy 4
The number of primes less than ~UOOOOO and of t,he prime form ms+ I is
301~ The number given by (se 681) is 302.6. The ratio is x . 005, and the agree-
ment is all that could be desired.
The number of prime-pairs rtl* + I and fn8 + 3, both of whose members are
less than g~oooo~, is 57. The value given by (5* 682) is 48.9. The ratio is
 Partitio numeroram. 111: On the expressjon of a number as a sum of primes. 63

l858. The numbers concerned are naturally rather small, but the result is perhaps
a little disappointing.
A more systematic test has been applied to the formulae for trip1et.s and
quadruplets of primes, the particular groups considered being

ti, a’+~, a+6; a, a-+4, ti+6;

ti, a-+-z, a't6, a+B; fl, a+4, a’+6, a’+m.

Tbe t.wo kinds of triplet-s should occur with the Pame frequency. On the other
hand there should be twice as many qtiadruplets of the second type as of the
first. For o, 2, 6, 8 have 4 distinct residues to modulus 5 and o, 4, 6, IO but
3, while for all other moduli the number of residues is the same; ati the rakio
5-3: 5-4 is 2. The actual results are shown in the following table.

Triplets.
-- ~-----.~- I -- ~-. -.- -
X F’&;0,2,6!/ ~C&j&) / Ratio p,kQ,4*6) Ea tio

IO5 260 270. 78 I *of+1 249 I ,087

2.105 1 417 44om71 I f ’ 057 425 I I ’ 037
3 l lo5 566 589 * 89 I .042 588 1 .m3
4. TO” 718 727*43 4 I-013 748 0 ’ 972
5. IO” 833 857. IO I * 029 88x 0 ’ 973
6. IO” 950 986.92 1 l 033 X008 0 * 973
7* JO” 1073 IIOO. 16 I. 025 1x33 0.971
8. Id 1195 1zq rn64 I ,017 1231 0.988
9. LO” 1295 1327.97 I ,025 1331 Q ’ 998
IOU 1398 1437 ’ 59 I .028 1443 o-996

Quadruplets.
--~--- _-~~~- -----_
Lt: P, (2; o, 2,6,8) Ibti0 IPdx; 0,4,6, x0)/ 27 Cd L;i, (LX) Rdh

f ,063 1 80 So. 82 J .OIO

TO” 38 40 ’ 41
2 1 IO” 52 61. 18 I* 177 I24 T22. 35 0 - 987
3.d j 70 78.62 I . 123 160 157 24 l 0 * 983
4, IO5 87 94 l 28
I ti 084 I94 188.55 0 l 972
5. IO” 103 108.75 1 ,056 219 217.50 0 - 993
6.10’ 117 122 l 36 I a045 239 244.7’ I .024
7, IO” 133 I35 ’ 29 I l OT7 263 270 l 59 r ,029
8.10~ I41 ‘47 l 69 f 1 047 285 295 ’ 39 1.036
9 l IO” I56 159.64 I ,023 299 3x9 29 l
I : 068
.
084
IO6 x66 I71 22 l

- / -- I.031 316 342 l 42 I l

623
 64 G. El. Hardy awl J. E. Littlewood.

Here C, and C, are the constants

of Theorem X I:. The results are on the w\~ole
very satisfactory, t,hough there is a curious deficiency of quadruplets of the
second type between Sooooo and ~oooooo.

5. 691. We return to the problems connected with the function

Q(X) = lim (A+ + x)-x(n)). We shall adhere to the notation of Theorem X I,
n -m
and shall suppose in addition that x is integral and that o < b, < b, < aem< @le.
It follows from Theorem X I t-hat, if H(b,, b,, , , , , b,)+ o, groups: tb,, n + b,, w. , l

n + b, consisting wholly of primes continually recur, and we shall say, when

this happens, that b,, b2, . . . , b, is a possibZe m-group of b’s, We say also that
the n, -+-b,, . . . , n -+ b, is an nt-group of primes, If, in a possible group, rrz = g(x),
where x = b,,- b, + I, we shall call the group, either of primes or of ti’s, a maxi-
mum group. A set of x consecutive positive integers we call ar x-sequence;
and we say that the group n + b,, . , ,, dn+ b,,, is coniained i92 the ( bl?b - b, + I)-
sequence b, T- c < bm, and that its Zen@ is b,,, - b, + I l

Theorem X 2. If b,, b,, . S.) b,, have a missing residue (mod a) for each
G <-
- m, theyA these b’s form a possible group.
This is an immediate consequence of Theorem X I, since Y ~a- I for G> ‘172.
Theorem X 3. Let M (x, n) be, the number of distinct ir~tegers
I Ci , c,, . . , C~U, l

in the interval n < c <n -+-x, which are not divisible by any prime less thaw 01’
equal to

e(x) -e(x) -t &

[ 1-t 1 f
eL(x) = Max M (x, n).
04
Then

eA4 = e(x)-
Let e(x, p) be the number obtained in place of g,(x) when the Q(X) that
occurs in the definition of el(x) is replaced by pt. Clearly we have

(5* 694 e(x, ,” -- 1) ke(x,

- 111)
Z> g(x)
and

624
 Partitio numerorum. III: On the expression of a number as a sum of primes. 65

Further,

(5 $13)
implies
e(x, 4 = e(+

For let d,, d,, . c ., & be an increasing set of positive integers with the properties
(characteristic of a set of z = e(x, /l) such integers) that (a) there is an ~2 such
that n+d,, l n + d, are not divisible by any prime less
l ‘, or equal to \l, than

and (b) d,-d,+&x. -_.- Then n+d,, ,.,, n + d, form a ‘possible’ group of b’s,
since they lack the residue o for every prime less than or equal to Z. Hence
e(x) -> z = p(x, p), and so, by (5. e(x> = e(x, p>*
Ggd

Next we observe that a(x, jl) = e(x) for !L = x, since t,he inequality z< 1’1is
clearly satisfied in this case. Let now /lo be the least !I, greater than or equal
to e(x), for which &, /lo) = e(x)* Then @)_Fu~~x. We have then

and so

by (5, 6913). Thus

,I%iQh
- PO- 1)<e(x,po)
._I +[--I +I =&) + x +I X
PO [1PO

-_
-<e(x)+ [ $-) I + I =&c).
Hence
e(x) = eb PJ 2. e(x, e(x>j = &) l

But it is evident that e,(x) )e(x), and therefore cl(x) = e(x).

It follows from the theorem that, in a maximum group of primes of length
x, the remaining numbers of the x-sequence are all divisible by primes less than or
equal Jto $(x). We shall see presently that (on hypothesis X) c(x)<e(x) + log x
for large values of x.
5. 692. We consider now the problem of a lower bound for e(x). Let p8
denote the s-th prime.
Theorem X 4. Let r = r(n) be defined, fur every value of n, by

Then. pr+l pPM, . . . p pt+n is ‘a possible n-group of b’s.

625
 66 G, I-I. Hardy and J. E. Littlewood.

For the primes less than or equal to n are pi, p,, . . , JIM and the b’s lack l

the residue o for each of them.

From Theorem X 4 we deduce at once
Theorem X 5+ If x = pr+n - pr+l + I, pr <n < p,+l, then

As 8 numerical exampk, let n = 76501. we have 177525

= 76493, p7526= 76507.
Hence
r = 7525, n + r = 84026, p7E+T= 1076503

X = 1076503 - 76507 + 1 = 999997 l

Thus
e(Ioooooo) >- 7650x.

We may compare this with the numbers of primes in the first, second, and third
millions, viz.
78498 170433 1 67885 l

Theorem X 5 provides a lower limit for Q(X) when x has a certain form:
we proceed to consider the case when x is unrestricted.
Theorem X 6. We have

for sufficiently barge values of x.

When rye is large
m log log nz
P =m(logm+
W& loglogm)-m+O
( log 77% - 1 l

Let

Then we have, by straightforward calculations,

Take n = p,. Then

626
 Partitio numerorum. III: On the expression of a number as a sum of primes. 67

prrt+r= y ( I- --
1,; y+qQi:;y)2)
x = p?z+r - Pr+l + I < pnfr - pr

=jy (I-----2 +o(‘qgy)2)<y-;&=z,

1%Y
when y is large. Thus

e(z)‘e(s)~n=p,=-y---- + 0 Y (log 1% Y>’

-
log Y (4 Y>” ( - (1% YP 1

Since y is arbitrary, so is x, and the theorem is proved.

5. 693. We conclude our discussion of e(x) with an account of one or two
par titular cases. For a given x it is, of course, theoretically possible to deter-
mine the maximum number of integers in an x-sequence that are not divisible
by any prime less than x. On hypothesis X, this number is a(x). Thus
L. AUBRY l has shown that 30 consecutive odd integers cannot contain more than
15 primes (or more than 15 numbers not divisible by 2, 3, 5, or 7). Thus
e(5g) 515. On the other hand if we take, in Theorem X 5, n = 15, r = 6, we
see that the 15 primes from 17 to 73 give a possible group of Vs. Hence, on
hypothesis X,

e(59) Ze(57j = e(73 - 17 + 1) E

-> 15 ;

and so e(sg)- 15~ The value of rr(5g) is 17.

Similarly a 35-sequence cannot contain more than IO numbers not divisible
bv 2, 3, or 5, but the 10 primes from 13 to 47, and therefore the numbers o, 4,
6; IO, 16, 18, 24, 28, 30, 34, form a possible IO-group of b’s, so that e(35) = IO =
=z(35)-1. A. striking example of a maximum prime group n + b,, . . . , n t b,,,
corresponding to this group of b’s, is provided by 7a= 113143.
The best example of a close approach by e(x) to sz(x) occurs when x = g7*
Consider the 24 primes from 17 to 113. They are a possibls group of b’s if they
have a missing residue for each prime less than 24. We have only to test 17,
19, 23, and we find that 17 lacks the residue 8, Ig lacks I and II, and 23 lacks
3, 12, 16, and zz. Hence on hypothesis X, e(g7) -t. 24. On the other hand it

l I,. EmDICKSON,
2. c., vol. I, p. 355.

627
 68 G. H. Hardy and J. E, Littkood,

may be shown that a 97-sequence cannot

contain 25 numbers not divisible by
2, 39 jt 7, IL or 13. Let us denote the range n<- x -< n + 96 by R,. There
is one and only one value of n, not greater than z.3.5.7= 210, for which
R,, contains 25 numbers not divisible by 2, 3, 5, or 7, viz. nf = 101. If then
ni 2.3.5.7 . II, a.nd R, contains 25 numbers not divisible by z, 3, 5, 7, or II, n
must be one of the numbers IOI + ZIO ry1 (m = I), x, . ., IO); and on examination
l

it proves that we may exclude all cases but ry1= IO. Repeating the argument
we see that, if 76-z.- x.5.7. II. 13, and R, contains 25 numbers not divisible by
2, 3t j, 7, 11, Or 13, then n must. be one of the numbers 12= 2201: + 2310~9
(m-0,x, l *,12). All these turn out to be impossible and, since any R, may
be reduced (mod. 2.3 l . 13)~ it follows
. that no R, can contain more than 24
numbers not divisible by a prime less fhan or equal to 13. A fortiori it follows
that e(97)z 24, and so (on hypothesis X) e(g7) = 24. Since z(g7) = 25, the dif-
ference e - 7~ is here unity. Beyond x = 97 it would seem that Q(X) falls further
below n(x), at least within any range in which calculation is pract*icable.

Conclusiun.

5. 7. We trust that it will not be supposed t#hat we attach any exaggerated

importance to the speculations which we have set out, in this last section. We
have not forgotten that, in pure mathematics, and in the Theory of Numbers in
particular, ‘it is only proof that counts’. It is quite possible, in the light of
the history of the subject, that the whole of our speculations may be ill founded.
Such evidence as there is points, for what it is worth, in the opposite direct.ion,
In any case it may be useful that, finding ourselves in possession of an apparently
fruitful method, we should develop some of its consequences to the full, even
where accurate investigation is beyond our powers.

Postscript,

(I). Prof. Landau has calleda our attention to the following passage in the
Habilitutionsschrijt of PILTZ (‘Uber die Hgufigkeit der Primzahlen in arithmetischen
Progressionen und iiber verwandte Gesetze’, Jena, 1884)) pp. 46-47: -
‘Ferner tviederholen sich gewisse Gruppierungen der Primzahlen mit gewisser
Regelmtissigkeit, so ist z. B. die durchscbnittliche Htiufigkeit der Gruppen van

628
 Partitio numerorum. III: On the expression of a number as a sum of primes. 69

je 2 Primzahlen, die in gegebenem Abstand aufeinanderfolgen, fur die ungef&hre

GrSsse x der Primzahlen, proportional 5, wobei allcrdings dieser Ausdruck je

v I
nach dem gegebenen Abstand mit verschiedenen const.anten Faktoren behaftet

ist, die H&ufigkeit einer Gruppe van 3 Prim&den proportional -$ und so

(1 >
fort. rn . Die n&here Ausfiihrung
l l dieser und andrer Gesetzc . . , werde ich ein
andres Ma1 folgen lassen.’
All of this is of course in perfect agreement with the results suggested in
our concluding section.
(2). We must add a few words concerning the memoirs of Stackel referred
to on p. 34. These have only become accessible to us during the print-
ing of the present memoir, and it is not possible for us even now to give any
satisfactory summary of their contents; but Stgckel considers the problem of
‘prime-groups’ in much detail, and it is clear that he has anticipated some at
any rate of the speculations of 5 .6. The method of Stgckel, like that of Brun,
rests on the use of the sieve of Eratosthenes, followed by a heuristic passage
to the limit; but StZckel’s problem is much more general, and he has gone much
further than Brun in the determination of the constants in the asymptotic for-
mulae. It seems to be the principal advantage of our transcendental method,
considered merely as a machine! for the production of heuristic formulae, that
these constants are determined naturally in the course of t-he analysis.
(3). We should also refer to a later memoir of Brun (‘Le crible d’Eratos-
th&rle et le th6ori:me <de Goldbach’, Videnskapsselska~ets Skrijter, Mat.-naturv.
Klasse, Kristiania, 1920, No. 3). Brun proves, by elementary methods, (I) that
every large even number is the sum of two numbers, each composed of at most
9 prime factors, (2) that the number of prime-pairs a, a + 2, less than X, cannot
exceed a constant multiple of x (log x)-~.
Brun’s work enables us to make a substantial improvement in the elemen-
tary theorem G. Using the inequalities proved on pp. 32-34 of his memoir, we
can show that

e(x)< log
&- x’
(4). Prof. Landau has pointed out fo us an error on p. 9. It is not neces-
sarily true that Ck= o when ok is imprimitive: our argument is only valid when
& is divisible by every prime factor of 4.
The inequality (2. 16) is however correct. Suppose first that q = & (A > o).
Our argument then holds unless Q= I; in this case xk is the principal character and

629
 70 G. W. Hardy and Jm E. Littlewood.’

This inequality is then eady generalised to all values of Q. If p-q1 g,, where
(919 92) = I, then every 3~ (mod. q) is the product of a ;G~(mod. q,) and a x, (mod.
p& and it is easily proved that

=<q&=vy.
The conclusion now follows by induction.

CORRECTIONS
p. 5,jootmte 2. Read: p. 492.
p. 64. On the third line of § 5.691, for bm read bm.
1~. 67. The statement on the last hne but one, that the primes from 17 to 113 lack the r&due
8 (mod 17), is incorrect since 59 - 8 (mod 17).

COMMENTS
Lemma 11 and footnote on p. 27. There is a simple closed expression for c&), namely

which escaped the notice of both Ramanujan and Hardy, and was discovered by Holder in 1936.
See Hardy and Wright (4th ed.), p. 238.
Recent researches of Tur6n (in course of publication) have enabled him to prove some of
Hardy and Littlewood’s results under consideraMy weaker hypotheses.

630
 Prof. G. H. HARDY ami Mr. J. E. LITTLEWOOD.

The series in question is

Here qr runs through all positive integral vahles, and JIM through all such
valuea less than and prime to qr, and Q is the denominator of

g +E+...+F = $,
IrL
expressed in its lowest berms. The arithmetical function x(q) is de-
filied by
= p(q)
x(q) s
9w

where w assumes all prime V&M, and Y is -the number of distinct resi-
dues of the group of numbers 0, ctl, a2, .*., a,)&to modulus ‘is. It is plain
that v = ~t+l from a certain point onwards.
The series is of very great interest, for it is the series on lvhich the
asymptotic distribution of groups of primes

X922, 8 (with J, E. Littlewood) proceeding8 of the LW &fath-

maatiml Lsdety, (2) 20, xxx. 631
 SOME PROBLEMS OF “PARTIT NUMERORUM” (V) : A FURTHER
CONTRIBUTION TO THE STUDY OF GOLDBACH’S PROBLEM

By. G. H. HARDY and 3, E. LITTLEWOOD

1
[Received November 20Gh, 1922,-Read December 14th, 1922.J

1 1, This paper is a sequel to the third of the series.* We proved

l

there that if tt certain hypothesis, which we called Hypothesis R*, and

which is a natural generalisation of the hypothesis oE Riemann concerning.
the zeros of i(s), is true, then ervery large odd mmber is the mm of thee
p 7i772es ; and we determined an asymptkc formula for the number of such
representations. But our analysis broke down when the number of primes
considered is less than three, and we were unable to make any rigorous.
contribution to the study of Goldbach’s Problem itself.
In the present paper we prove (still, naturally, on the assumption of’
Hypothesis R’) that almost alt eveIt 1z2mbel*s of t,ulopl%~~es,that
af*e ~~~~1~s
is to say that the number of numbers less than 32, for l&ich “ Goldbach’s
Theorem ” is false, is o(u) when 7~ is large. We prove, in fact, consider-
ably more, but this is the essential result.
1 . 2. We reed the termidogy of P. N. 3, in so far as it is relevant,
here. t Our fundamental hypothesis 1 is

L(s) = CXCrn)
m” ’

* G. H. Hardy and J. E. Littlewood, ‘I Some problems of ‘ Partitio Numerorum ’ (III) :

On the expression of a number as a sum of primes ’ ‘, Acta M&hentatica, Vol. 44 (1922),
pp. l--70. We refer to this memoir as P. N. 3, The analysis which it contains has been
considerably simplified by Landau Fib Zur Additive Primzahltheorie “, Rend. di Palemzo
Vol. 46 (1922), pp. 349-3561,
+ We have modified the notation of P. N. 3, since we prefer now to denote a typical prime,
as is usual, by p. What were there p, q are now 3t, k ; what was there h is now q (k) ; and
what was there Q is now 4.
1 In P. N. 3 we did not actually use Hypothesis R #? but a slightly less drastic hypothesis
(Hypothesis R) which proved sufficient for our purpose.

1924, 6 (with J. E. Littlewood) Proceeding8 of the Ltmdon M&he-

632 matical Society, (2) 22, 46-56.
47 G. H. HARDYand IT. E. LITTLEWOOD [Dec. 14,

We write p for & prime,

f(x) = Elogpx~ (‘x1 <I),

(f (x 11a= G v&J(nt) x7’&,

v(?G)= u&z) = c logy logp’,

p+pc=m

F(x) = 2X(m) x” = EmS(m) xHc,

where S(m) = ck(-p$,

c],,(- nt> = 2 e-2h*ailk = C ek(- hm).

h h

Here ‘jib= I, 2, 3, +.., k = 1, 2, 3, ,,,, and 1~ is positive, less than k,

and prime to k, except when k I_ 1, when /A = 0, c~(-H~~)= 1 and A, = 1.
S(m) is the “ singular series “, and its sum is given by*

S(m) = 0 (m odd),

St m> =2cII ;=; (m even),

(4
where P is an odd prime divisor of ~72, and

c =- 1
q---q* >
P23 (
We use the machinery of the CcFarey dissection “, explained in our
memoirs on Waring’s Problem. The circle I? to which it is applied is
defined by
I X I = e-H = e-w,
and the “ Fprey arcs ” sre denoted by &,I; or (. The dissection is of order

37 = [&I,
and there is, in this problem, no distinction of “ major ” atid ” minor”
arcs,
When we are studying the
, arc &, k, we write

X = sk(h)ewY, Y = -A- +,i8, - 01,\( 0 < 80,

72

* For the formal summation of the singular series see P. N, 3, pp. 26-29.

633
1922.1 SOllE PROBLEMS OF LLPARTITIO NUBZERORUM" (v). 48

where o0 and 0; lie between 2r/kN and T/MT. The function which affords
an approximation to f(z) on &, 7,is

We write further

where

so that
F = F(x) = Cmx” C Afi(m) =
m k

Finally, we write
S(722) = s, (N + s, b0 ,

F = 27?2S(772)xn= = ~772&(772) X”‘+- ~772s&2) X”‘,

where the suffix 1 limits ?Cto the range k < U, and the suffix 2 to the
range k > v.* Then
F 1= C F,,k, F,= C Fhka
k<v,h ’ k’>v, h ’

2. Prelimilzary lemmas.
2 I 1, In all that follows A denotes an absolute constant, and O’s and
o’s are uniform in all parameters that occur, except E.

LEMMA 1. -0?2 &, ks

* =f+ =fw+h,k = 0 (grj(log ,l)d).

See P. N, 8, pa 23, equation (3. 124), and take @ = +.
LEMMA %--If x = 1x1@ = c-~+~, so ihat x lies on r, then,

$If(X) I” dw = O(72 1Og 72).

See P. N. 3, p. 24.

* Y is au integer as yet undetermhed. We ultimately take Y = [a$

49 G. H. HARDY and J. E. LITTLEWOOD [Dec. 14,

LEMMA3.- We have

From P. N. 3, p. 23 (bottom), with r = 3, w-e have

Hence

C 1+j%M,An C 7 ’ < An c - 1 < A72 c -

(log WA
< An(log n) Al
J% kh k $dk) ic k
( $(k) 1

LEMMAJ.-We have

c ~ p-+212 a0 = O(nC+y
5
fur every pmitiae E.

= A&+A I&,
say. But

by Lemmas 1 and 2 ; and

I& < AnG(log 72)A C 1 y5 1a de = 0 (,,e(log +) = o(?zs+E),

t
by Lemmas 1 and 3 ; whence the conch&n.

2.2. LEMMA 5. - If k > 1, Ip( > O,* ctnd k = klk,, where

(k,, nz) = 1, then
IAk(m>l < AykhC)A.
I 2

* So tuat 7~ has no repeitted factor (is puaddjrei). Otherwise Ak (m) = 0,

635
 1922.1 SOME PROBLEMS OF 'CCPARTfTIO NUMEROKUM " (v). 50

But, if k = I@,, where the p’s are primes, then*

I cd-y) 1= II (ps-1) < II ps = k,,
P,b P,I fn
whence the result of the lemma.

) S,(m) 1 < A (log WC)~d (772) v,

whence d(m) is the number of divisors of m.

By. Lemma 5,

the summation extending over those values of k for which k > U,

]dWI > 0.
Suppose that q, r2, -* ., rt are the prime divisors of m, and that
s = Ilki is a typical qedrntfrei divisor of ~72 ; and consider those tel*ms
of C for which

6 is fixed and k, varies. The contribution of these terms is less than

c (log SkaJA < A(log mJA c (1~~skdA
2
s 3 k
%> v Sk 2 7c,>v/6 2

< (loa dA
A(log m)A -0
V

2 (log MA A (1% dA c 1,
Hence k c A (log ,m>
1 2 v SI?)L
which proves the lemma.

2.3. LEMMA L-If x = lx ] ei”, us in Lemma 2, thea

5027F
1F,(x) 12 do = 0 n3(log >@AV)
(
*

* See P. N, 3, pa 28 (bottom).

636
51 G. H. HARDY and J. E. LITTLEWOOD [Dee. 14,.

For Fa(x) = %nS&n)x”, and SO

2rr
(2.31) 1F,(s) 12 dctl= AC (nzi3&1~))~ 1x la,lz
0
- o (log 4”’
x 17z2(log ?I&)” ( d(7lt))2 15 p),
- ( v2
by Lemma 6. But
c (d(??Z)) a = 0
5%<a

c 112(log ?r&>”(do)t))
IIt < 12

A
and so since /x/(1--
( 72 )

(2. 32) lzd (log 7?z)A (d(??zqa 1x 2H1= 0 ( n3 (log 92)” ) l

The lemma follows from (2. 31) and (2, 32).

LEMMA &--We have

This is trivial, since

Fh, k- +;,H = (;$)2 (s(ze,c-lr,)+)

xed-h) = O(I).
I--xe&-h,j”-
(

* In fact the sum is 0 (TZ (log +). See S. Ramanujan, “ Some form&e in the aaalytic-
theory of numbers ’ ‘, Messenger of Mitthematics, Vol. 45 (1916), pp* 81-84.

637
 1922,J SOME PROBLEXS OF "PARTIT NUMERORUM" (v) l 52

de 8 (log k)*
(?&-“+ea)o < An kP g

Hence c

2.4. LEMMA 10.-&ppose that x lies on &, k ami t?mt & f 1~ is a Fwey
UTC&$&vent from &. Then
A72 (log K)A
P%KI< (IzK--HP’

where

and

Suppose first that the arcs &, &,K are not adjacent in the dissection.
Then

h H
and SO I0 I>Al --- Z A IhK--HI 9
k K kK
< A (log K)A kaKa &z(log K)- -A
IF if, K Ka (hK-kkHJa < (hK-kHj2’
since ka < P < 12. The argument fails when the arcs are adjacent. In
this case 1hK-kH = 1. As x ie outside the arc &,Kt IwI> A/KN;
ana so

)FH,BI<A~K’Nl<Ala(logIi’)“= $2

The msult of the lemma is therefore true in any case.

2 5. LEMMA
l 11.-&d
G = z lFH,KI,

whlers the sign E implies a summation over thse pairs uf values (H, K),
I.
* So that G is a function of ?z rend k, as well $6 of x and Y.

638
58 G. H, HARDY arta J. E. LITTLEWOOD [Dec. 14,

&t&t from (ha, k),$w which K < v.* Th]erz

c1 G2d0 = 0 (n; (lug ~t)~ $1.

J th, k
Since G is a sum of less than v2 terms, we have

by Lemma 10. Hence

< A+ (log 12)~v2 C C 1

H, x It, k k(hK-kHj4’
lvhere now the inner summation is defined by

k \( N, Vb k) # 4, K),
and the outer summation by K \( V, But
1 1 1
c
h, Ii k(hK-kHJ4
x1 c
= r4 k k k /7?- kH\ 4

A X4
<-+ k -j;<A+Alog?~;
7c

I
and so c c
H,K h,k k(hK--kH)’ < A” ““”
c Ga d0 < Ad (lug 92)* v2. A V’ log ti < An (log n)A v4,
th, k

the result of the lemma.

3.1. THEOREM A ,-If Hypothesis EL* is true, then

5 (v(m)-Q(m))” = o(Tb~+E)
1
for ewypositive E.

639
 1922.1 SoME PROBLEMS OF “PARTITIO NUMERORUM " (v). 54

It is sufficient* to prove thatl

C ( Y(~IL)-R(w~))~ 1x 12m= O(&+@),

when ] x.1 = tiAH, or that

Now 2r,fa-Fladw= C If-Fl’d@

0 6

and

by Lemma 4. It is therefore sufficient to prove that

(3 11)
l C j jb2-F12 de = O(n-j.
Jt

3.2. Since F = Fl +FQ, we have

(3. 21)
and
C
t
]y!+ Fj2d0< AC
J!t~+2-Fl~2d8+A~
t I&\%;

(3 . 22) C IFz[‘dO < An3(log n) A va,

(log dA
k
by Lemma 7. Next

according as k < Y or k > V, the range of summation indicated by C being

defined as in Lemma 11. If k < Y, then, we have

Iq2-Flja< A I~a-F~~,1;/2+AI~F~,h.)2<A(~a-Fh,k12+AG2;

640
55 G. H, HARDY and J. IS. LITTLEW~~D [Dec. 14,

(3.23) Iq”-Ii;l,k lade

< A +A “‘bf dA
+ Ad (log PZ)~v*,

by Lemmas 8, 9, and 11. From (3 . Zl), (8 .22), and (8 I as), we conclude

that

C f 1qa-Fla d9 < Ad(log 7a)Aq+Ang(logn)Av4a

Taking Y = [GJ, we obtain (3. II), and complete the proof of the theorem.

3.2, THEOREMB. -If Hy@hesis R* is true, then the number of eves

numbers less tha7t 92,for which Goldbrcch’s Theorem is false, is O(d+‘) for
every positive E.

We have Q(m) = 0, if m is odd, and

if PJZis even ; and so

( v(m)--n(m))2 = (61(m))’ > Am’,

for every even number nz for which Goldbach’s Theorem is false. The
number of such numbers between $, and PZ (inclusive of the limits) is
therefore less than
B (&a$+‘,
where B(e) ia a function of e only ; and the number of sllch numbers less
fhan ?zis less

B (e)+ (1+(9)“+‘+($)1+‘+*,*) < .B(e)Iza+c*

4. ConclUSiu~2.
4.1. It may be observed that the main conclusion of P. N. 3, viz. that
evwy Eal*ge odd number is (on Hypothesia R’) the such of thq=eeprimes, ia

641
1922.1 SOME PROBLENS OF “ PARTfTfO NUME*RORUM " (v). 56

an immediate corollary of Theorem B. For, if 72 is a large odd number

which is not the sum of three primes, then none of the numbers
92-p (2 < p < 4
can be the sum of two, which obviously contradicts Theorem B.
The method which we have followed here has other important applica-
tions. It enables us, for example, to prove that almost all numbers are
8ums of Jive cubes,
sixtem biquadmtes,
two squaws ad a cube (or any odd power),
tm squares a9td a pri9ne?

-the last theorem being, naturally, subject to Hypothesis R*. We shall

prove the first two oE these assertions in the next memoir of the series, irl
which we return to Waring’s Problem,
(d) Inaugural Lecture

(Oxford 1920)
SOME FAMOUS PROBLEMS
of the

THEORY OF NUMBERS
and in particular

Waring’s Problem

UniverJity of Oxford

C. II HARDY, M.A., F.R.S.

Fellowof mw cozzege

OXFORD
AT THE CLARENDON PRESS
1920
SOME FAMOUSPROBLEMS OF THE
THEORY OFNUMBERS.
IT is expected that a professor who delivers an inaugural
lecture should choose a subject of wider interest than those
which he expounds to his ordinary classes. This custom is
entirely reasonable ; but it leaves a pure mathematician
faced by a very awkward dilemma. There are subjects in
which only what is trivial is easily and generally compre-
hensible. Pure mathematics, I am afraid, is one of them ;
indeed it is more: it is perhaps the one subject in the
world of which it is true, not only that it is genuinely
difficult to understand, not only that no one is ashamed of
inability to understand it, but even that most men are
more ready to exaggerate than to dissemble their lack of
understanding.
There is one method of meeting such a situation which
is sometimes adopted with considerable success. The
lecturer may set out to justify his existence by enlarging
upon the overwhelming importance, both to his University
and to the community in general, of the particular studies on
which he is engaged. He may point out how ridiculously
inadequate is the recognition at present afforded to them ;
how urgent it is in the national interest that they should be
largely and immediately re-endowed ; and how immensely
all of us would benefit were we to entrust him and his
colleagues with a predominant voice in all questions, of
educational administration. I have observed friends of my
own, promo&d to chairs of various subjects in various

647
 6 SOME FAMOUS PROBLEMS OF

demarcated region, but it has suffered under tile miss& of

philosophers for generations, and it is ours by right; we
propose to accept the mandate for it, and to offer it the
opportunity of &f-determination under the mathematical
flag. Such at any rate is the thesis which I hope it may
before long be my privilege to defend,
It seemed to me, however, when I considered the matter
further, that there are two fatal objections to maVthematical
philosophy as a subject for an inaugural address. In the
first place the subject is one which requires a certain
amount of application and preliminary study. It is not
that it is a subject, now that the foundations have been
laid, of any extraordinary difficulty or obscurity; nor that
it demands any wide knowledge of ordinary mathematics,
l3ut there are certain things that it does demand ; a little
thought and patience, a little respect for mathematics, and
a little of the mathematical habit of mind which comes
fully only after long years spent in the company of mathe-
matical ideas. Something, in short, may be learnt in a
term, but hardly in a casual hour.
In the second place, I think tha& a professor should
choose, for his inaugural lecture, a subject, if such a subject
exists, to which he has made himself some contri bution of
substance and about which he has something new to say.
And about mathematical philosophy I have nothing new to
say; I can only repeat what has been said by t#he men,
Cantor and Frege in Germany, Peano in Italy, Russell and
Whitehead in England, who have originatred the subject
and moulded it now into something like a definite form.
It would be an insult to my new University to offer it,
a wat?ered synopsis of some one else’s work. I have there-
fore finally decided, after much hesitation, to take a sub-
Iject which is quite frankly mathematical, and to give a
summary account of the results of some researches which,

650
 THE THEORY OF NUMBERS 7

whether or no they c0ntai.n anything of any interest or.

imputince, have at any rate the merit that they represenb
the best that I can do.
My own favourite subject has certain redeeming advan-
hges. It is a subject, in the first place, in which a large
proportion of the most remarkable results are by no means
beyond popular comprehension. There is nothing in the
lea& popular about its methods ; as to its votaries it is the
most beautiful, so by common consent it is the mod diffi-
cult of all branches of a difficult science ; but many of the
actual results are such as can be stated in a simple and
striking form. The subject has also a considerable histori-
cal connexion with this particular chair. I do not wish ta
exaggerate this connexion. It must be admitted that the
contributions of English mathematicians to the Theory of
Numbers have been, in the aggregate, comparatively slight.
Fermat was not an Englishman, nor Euler, nor Gauss, nor
Dirichlet, nor Riemann ; and it is not Oxford or Cambridge,
but Gattingen, that is the centre, of arithmetical research
to-day. Still, there hag been an English connexion, and it
has been fur the most part a connexion with Oxford and
with the Savilian chair.
The connexion of Oxford with the theory of numbers is
in the main a nineteenth-century connexion, and centres
naturally in the names of Sylvester and Henry Smith.
There ia a more ancient, if indirect, connexion which I
ought not altogether to forget, The theory of numbers,
more than any other branch of pure mathematics, has
begun by being an empirical science. Its most famous
theorems have all been conjectured, sometimes a hundred
years or more before they have been proved; and they
have been suggested by the evidence of b mass of compu-
tation. Even now there is a considerable part to be
played by the computer, n and a man who has to undertake

651
 8 SOME FAMOUS PROBLEMS OF

laborious arithmetical computations is hardly likely to

forget what he owes to Briggs. However, this is ancient
history, and it is with Sylvester and Smith that I am
concerned to-day, and more particularly with Smith,
Henry Smith was very many things, but above all things
a most brilliant arithmetician. Three-quarters of the first
volume of his memoirs is occupied with the theory of
numbers, and Dr. .Glaisher, his mathematical biographer,
has observed very justly that, even when he is primaril?
concerned with other matters, the most striking feature of
his work is the strongly arithmetical spirit which pervades
the whoie, His most remarkable contributions to the
theory are contained in his memoirs on the arithmetical
theory of forms, and in particular in the famous memoir on
the representation of numbers by mnns of five squares,
crowned by the Paris Academy and published only after
his death. This memoir is peculiarly interesting fo me, for
the problem is precisely one of those of which I propose to
speak to-day ; and I: mzey perhaps add one comment on the
surprising history set out in Dr. (Zaisher’s i&rod&ion.
The name of Minkowski is familiar to-day to many, even
in Oxford, who have certainly never read a line of Smith,
It is curious to contemplate at a distance the storm of
indignation which convulsed the mathematical circles of
England when Smith, bracketed after his death with the
then unknown German mathematician, received, a greater
honour than any that had been paid to him in life.
The particular problems with which I am concerned
belong to what is called the ‘ additive ’ side of higher
arithmetic. The general problem may be stated as follows.
Suppose that ‘M,is any positive integer, and

aj3 “2, a,, .*a

positive integers of some special kind, squares, for example,

652
 THE THEORY OF NUMBERS 9

or cubew, OF perfect kth powers, or primes. We consider all

possible expressions of’ m in the form
n = cw,+q- a.* +a,,
where B may be fixed or unrestricted, the Cwfsmay or may
not be necessarily distinct, and order may or may not be
relevant, according to the particular problem on which we
are engagd We denote by
r (N)
the number of representations which satisfy the conditions
of the problem. Then what ctm ‘we say about r (9%)1 Can
we find an exact formula for r@), or an approximate formula
valid for large values of rt ? In particular, is r (in) a&5uays
p&be? Is it always possible, that is to say, to find at
least one representation of SQof the type required? Or, if
this is not so, is it at any rate always possible when YLis
sufficiently large ?
I can illustrate the nature of the general problem most
simply by considering for a moment an entirely trivial
case. Let us suppose that there are three different a’s only,
viz, the numbers I, 2, 3 ; that repetitions of the same 61 are
permissible ; that the order of the cw’s is irrelevant ; and
that 8, the number df the a’s, is unrestricted. Then it is
easy to see that Y(%), the number of representations, ia the
number of solutions of the equation
n =x+Zy+3z
in positive integers, including zez0.
There are various ways of solving this extremely simple
problem. The most interesting for our present purpose is
*that which rests on aq analytical foundation, and usm the
idea of the generating fulzctim

653
 10 SOME FAMOUS PROBLEMS OF

in which the coefficients are the values of the arithmetical

functitin I. It follows immediately from the definition
of r(?z) that

and, in order to determine the coefikients in the expansion,

nothing more than a little elementary algebra is required.
We find, by the ordinary theory of partial fractions, that

f( 1 1 1 17 1
x= --
6(1-x)’ $- 4(1--~)~ + 72(1--s) + 8(1+x)
1 1
+
9 (l-ox) + 9(1-h)

lvhere o and o2 denote e usual the two complex cube roots

of unity. Expanding the fractions, and picking out the
coefficient of tlln, we obtain .

+) =---(9%+ 3J2 7 + (-v+ 2c*g2n=

12 72 8 ii 3’

It is easily verified that the sum of the la& three terms

can never be as great as & so that r(m) isr the integer
nearest to
(m
-. + q2
12

The problem is, zw I said, quite trivial, but it is interest-

ing none the less. A great deal of work has been done
on problems similar in kind, though naturally far more
complex and difficult in detail, by Cayley and Sylvester,
for example, in the last century, and by G&her, and
above all by Macmahon, in this. And even thirs problem,

654
 THE THEORY OF NUMBERS 11

simple as it is, has auf&i& content to bring out clearly

certain principles of cardinal importance.
In particular, the solution of the problem shows quite
clearly that, if we are to attack these cadditive ’ problems
by analytic methods, it is in the theory of integral power
aeries

that the necessary machinery must be found. It is this

characteristic which distinguishes this theory sharply from
the other great side of the analytic theory of numbers, the
t multiplicative ’ theory, in which the fundamental idea ie
that of the resolution of a number into primes. In the
latter theory the right weapon is generally not a power
series, but what is called a Dirichlet’s series, a series of the
fyl?e
c a,w8.
It is easy to see this by considering a simple example.
One of the most interesting functions of the multiplicative
theory is d(n), the number of divisors of n. The amociated
pow& series

is easily transformed into the series

called Lam bert’s series. The function is an interesting

one, but somewhat unmanageable, and certainly+not one of
the fundamental f m&ions of analysis. The corresponding
Dirichlet’s series is far more fundamental; it is in fact

the square of the fakous Zeta function of Riemann.

655
 12 SOME FAMOUS PROBLEMS OF

The underlying reason for this distinction is fairly

obvious. It is natural to mdtipZy primes and unnatural
to add them. Now
WP x ncs = (mn)-g,

su that, in the theory of Dirichlet’s series, the terms

combine natural ly with oqe another in a cmu .Itiplicative ’
manner. But
xm XX% = xm+y

so that the multiplication of two terms of a power series

involves an additive operation on their ranks. It is
thus that the Dirichlet’s series rather than the power
series proves to be the proper weapon in the theory of
primes.
It would be difficult for anybody to be more profoundly
interested in anything than I am in the theory of primes ;
but it is not of this t*heory that I propose to speak to-day,
and we must return to our general additive problem. As
noon as we try to specialize the problem in some more
interesting manner, two problems stand out as calling for
resesrch. Each of them, naturally, is only the representa-
tive of a class.
The first of these problems is the problem of pwtitims.
Let us suppose now that the cx’s are cony positive integers,
and that (as in the trivial problem) repetitions are allowed,
order Is irrelevant, and s is unrestricted. The problem is
then that of expressing ‘IL in any manner as a sum of
integral parts, or of solving the equation
n =x-t2y+32!+4u+5v+.;.,

and r(m) or, as it is now more naturally written, p (n), is

the number of warestrieted partitbm of in. Thus
5 = 1+X+1+1+1 = 1+1+1+2
=1+2+2=1+1+3=2+3=1+4=5,

656
 THE THEORY OF NUMBERS 13

so that p(5) = 7. The generating function in this case

wae found by Euler, and is

I do not wish ‘to discuss this problem in any detail

now, but the form of the generating function calls for
une or two general remarks. In the trivial problem the
generating function was rc&onaZ, with a finite number of
poles all situated upon the unit circle. Here also we are
led to a power series, or infinite product, convergent inside
the unit circle ; but there the resemblance ends, This
function will be recognized by any one familiar with the
theory of elliptic functions ; it is an elliptic modular
function ; and, like all such functions: it has the unit circle
as a continuous line of. singularities and does not exist at
all outside, It is easy to imagine the immensely increased
difficulties of any analytic solution of the problem.
It was conjectured by a very brilliant Hungarian
mathematician, Mr. G. Pdlya, five or six years ago, that
cony function represented by a power series whose
coefficients are integers, and which is convergent inside
the unit circle, must behave, in this respect, like one or
other of the two generating functions which we have
considered. Either such a function is a rational function,
that is to say, completely elementary; or else the unit
circle is a line of essential singularities. I believe that a
proof of this theorem has now been found by Mr, F, Carlson
of Upsala, and is to be published shortly in the Muthe-
mcttische Zeitschrift. It is difficult for me to give reasoned
praise to a memoir which I have not seen, but I can
recommend the theorem to your attention with confidence
as one of the most beautiful of recent years.
The problem of partitions is one to which, in collaboration
with the Indian mathematician, Mr. S. Ramanujan, I have

657
 14 SOME FAMOUS PROBLEMS OF

myself devoted a great deal of work. The principal result

of our work has been the discovery of an approximate
formula for @Q in which the leading term is
27r
- &-&
1 -46
de
--

and which enables w to approximate to p(m) with an

accuracy which is almost uncanny, We are able, fur
example, by using 8 terms of our formula, to calculate
p(200), a number >f 13 figures, with an error of 404.
I have set out the details of the calculation in Table I,

TABLE T
242w
3,972,998,993,185-896
36,282*978
- 87*5%
+ 5m147
+ I*424
+ 04071
+ O*OQO
+ 0*043
3,972,999,029,388+004

The value of p(200) was subsequently verified by Major

MacBIahon, by a direct computation which occupied over
a month,
The formulae connected with this problem are very
elaborate, and except on the purely numerical side, where
the results of the theory are compared with those of com-
putation, it is not very well suited for a ha&y expo-
sition ; and I therefore pass on -at once to the principal
object of my lecture, the very famous problem known,
after a Cambridge professor of the eighteenth century, as
Wiwing’s Probhm.

658
 THE THEORY OF NUMBERS laY

We suppose now that every 61 is a perfect &th power

,k, E being fixed in each case of the problem which we
consider; rfl2 may be of either sign if h is even, but must
be positive if i2 is odd. In either case \-Te allow nz to be
zero. Repetitions are permitted, as in our previous problems;
but it is more convenient now to take acco.unt of the order
of the br’s ; and s, which was formerly unrestricted, is now
fixed in each case of the problem, like h. The problem is
therefore that of determining the number of representations
of a number m as the sum of s positive E-th powers. Thus
Henry Smith’s problem, the problem of five squares, is the
particular case of Waring’s problem in which X;: is 2 and
s is 5. The problem has a long history, which centres
round this simplest case of squares ; a history which began,
I suppose, with the right-angled triangles of Pythagoras,
and has been continued by a long succession of mathemati-
cians, including Fermat, Euler, Lagrange, and Jacobi, down
to the present day. I will begin by a summary of what
is known in the simplest case, where the solution is
practically complete.
A number 32 is t,lie sum of two squttres if and only if
it is of the form
9a I jlpp,

where P is a product of primes, all different and all of the

form 4kfl. In particular, a prime number of the form
4 1; + I can be expressed as the sum of two squares, and
substantially in only one way. Thus 5 = I2 + P, and there
is no other solution except the solutions (+ 1)2+ (+ 2)“,
(+ q2 + ( $- q2, which are not essentially different, although
it is convenient to count them as distinct. The number of
numbers less than x, and expressible as the sum of two
squares, is approximately
CX
-3
&g X

659
 16 SOME FAMOUS PROBLEMS OF

where C is a certain constant. The last result was proved

by Lndau in 1908; all the rest belong to the clmsical
theory*
A number is the Burn of three squares unless it is of
the form

when it is not so expressible. Evq number may by

expressed by four squares, and u f’tiori by five or more,
It is this last theorem of Lagrange that I would ask you
particularly to bear in mind.
If s, the number of squares, is even and less than 10,
the number of representations may be expressed in a very
simple form by means of the divisors of 32. Thus the
number of representations by 4 squares, when ,iz is odd, ia
8 times the sum of the divisors of n; when rz~is even, it
is 24 times the sum of the odd divisors; and there are
similar results for 2 sqtiares, or 6, or 8.
When 8 is 3, 5, or 7, the number of representations can
also be found in a simple form, though .one of a very
different character. Suppose, for example, that s is 3. The
problem is in this case essentially the same as that of
determining the number of classes of binary arithmetical
forms of determinant --‘r’l ; and the solution depends on
certain finite sums of the form

extended over quadratic residues p or non-residues y of ‘YL.

When S, whether even or odd, is greater than 8, the
solution * is decidedly more difficult, and it is only very
recently that & uniform method of solution, for which I
must refer you to some recent memoirs of Mr, L. J. Mordell
and myself, has been discovered. For the moment I wish
to concentrate your attention on two points : the first, that
an expre8tiu~n by 4 squmea is dways ptible, while om3

660
 THE THEORY OF NUMBERS 17

by 3 is lnot ; and the second, that the existence of numbers

sot expressible by 3 squares is revealed at once by the quite
trivial observation that no number so expressible can be
congruent to 7 to modulus 8.
It is plain, when we proceed to the general case, that
any number 32 can be expressed as & sum of Lth powers ;
we have only to take, for example, the sum of ‘YLones. And,
when 83%is given, there is a /miniirnztm, number of Xc-th
powers in terms of which m can be expressed ; thus
23 = 2*z3 + 7*13

is the sum of 9 cubes ahd of no smaller number. But it is

not at all plain (and this is the point) that this minimum
number cannot tend to infinity with ITL. It does not when
b = 2 ; for then it cannot exceed 4.. And Waring’s Problem
(in the restricted sense in which the name has commonly
been used) is the problem of proving that the minimum
number is similarly bounded in the general case. It is not
an easy problem ; its difficulty may be judged from the fact
that it took 127 years to solve.
We may state the problem more formally as follows.
Let 16 be given. Then there may or may not exist a
number nz, the same for all values of n, and such that
3%can always be expressedSas the sum of wz k-th powers or
less. If any number ‘yrt possesses this pruperty, all larger
numbers plainly possess it too ; and among these numbers
we may select the Zeust, This least number, which will
plainly depend on k, we call g(E) ; thus g(k) is, by defini-
tion, the least number, if such a number exists, for which
it is true that

We have seen already that g(Z) exists and has the vaIue 4.
In the third edition uf his Meditationes Algebmicae,
published in Cambridge in 17 82, Waring asserted that

661
 18 SOME FAMOUS PROBLEMS OF

every number is the sum of not more than 4 squares, not

more than 9 cubes, not more than 19 fourth powers, et sic
deimeps. A little more precision would perhaps have
been desirable ; but it k~s generdly been held, and I do
not question that it is true, that what Waring is assepting
is precisely the existence of s(k). He implies, moreover, .
that y(2) = 4 and y(3) = 9 ; and both of these assertions
are correct, though in the first he had been anticipated ‘by
Lagrange. Whether ~(4) is or is not equal to * 19 is not
known to-day.
Waring advanced no argument of any kind in support
of his assertion? and it is in the highest degree unlikely
that he was in possession of any sort of proof. I have no
desire to detract from the reputation of a man who was
a very good mathematician if not a great one, and who
held a very honourable position in a University which not
even Oxford has persuaded me entirely to forget. Eut
there is a tendency to exaggerate the profundity implied
by the enunciation of a theorem of this particular kind.
We have seen this even in the case of Fernlat, a mathe-
matician of a class to which Waring had not the slightest
pretensions to belong, whose notorious assertion concerning
‘ Mersenne’s numbers ’ has been exploded, after the lapse of
over 2 50 years, by the calculations of the American computer
Mr. Powers. No very labol-ious computations would be
necessary to lead Waring to a highly plausible speculation,
which is all I take his contribution to the theory to be ;
and in the Theory of Numbers it is singularly easy to
speculate, though often terribly difficult to prove ; and it
is only proof that counts.
The next advance towards the solution of the problem
was made by Liouville, who established the existence of
g (4). Liouville’s proof, which was first published in 185 9,
is quite simple and, as the simplest example of awl important

662
 THE THEORY OF NUMBERS 19

type of argument, is worth reproducing here. It may be

verified immediately tbhat

+~dc+z)4+(x-z)4+(t+y)4 ++yy
+tx+ty +(z- t)*+(y+z)4+(y-X)4;
and since, by Lagrange’s theorem, any number X is the
sum of 4 squares, it follows that any number of the form
6x2 is the sum of 12 biyuadrates. Hence any number of
the form 6 (X2 + Y2 + Z3 + T2) or, what is the same thing,
any number of the form 6 ~yyt,is the sum of 48 biquadrates.
But amy number ss is of the form 6m+r, where r is
O, I, 2, 3, 4, or 5. And therefore #IL is, at worst, the sum
of 53 biquadrates. That is to say, Iq (4) exists, and does
not exceed 53. Subsequent investigators, renning upon this
argument, have been able to reduce this number to 37; the
final proof that g (4) =< 37, the most that is known at present,
was given by Wieferich in 1909. The number
79 = 4a24 + 1 t&l4
needs 19 biquadrates, aml no number is known which needs
more. There is therefore still a wide margin of uncertainty
as to the actual value of g( 4).
The case of cubes is a little more difficult, and the
existence of g (3) was not established until 1895, when
Maillet proved that g(3j s 17. The proof then given by
Maillet rests upon the identity
6x(x2+y’+x”+t2)
= (x+y)3+(x-y)3+( x+~)3+(X-z)3+(X+t)3+(X--t)3,

and the known results concerning the expressiog of a

number by 3 squares. It has not the striking simplicity
of Iiiouville’s proof ; but it has enabled successive investi-
gators to reduce the number of cubes, until finally Wieferich,

663
 20 SOME FAMOUS PROBLEMS OF

in 1909, proved that g (3) s 9. As 23 and 239 require

9 &es, the value of g(3) is in fact exactly 9. It i8 only
for E = 2 and k = 3 that the actual value of g(k) has
been determined. But similar existence proofs were found,
and upper bounds for g(k) determined, by various writers,
in the cases hi = 5, 6,7, 8, and 10.
Before leaving the problem of the cubes I must call your
attention to another very beautiful theorem of a slightly
different character. The numbers 23 and 239 r’equire
9 cubes, and it appears, from the results of a survey of
all numbers up to ~U,OUO, that no other number requires so
many. It is true that this has nut actually been proved ;
but it has been proved (and this is of course the essential
point) that the number of numbers which require as many
cubes- as 9 is $w&.
This singularly beautiful theorem, vrhich is due to my
friend Professor Landau of Gijttingen, and is to me as
fascinating as anything in the theory, also dates from
19O9, a year which stands out for many reasons in the
history of the problem. It is of exceptional interest not
only in itself but also on account of the method by which
it was proved, which utilizes some of the deepest results in
the modern theory of the asymptotic distribution uf primes,
and made it, until very recently, the only theorem of its
kind erected upon a genuinely transcendental foundation.
To me it has a personal interest also, as being the only
theorem of the kind which, up to the present, defeats
the new analytic method by which Mr. Littlewood and
I have recently studied the problem.
Landau’s theorem suggests the introduction of another
function of E, which I will call G(k;), of the same general
character as g(k), but I think probably more fundamental.
This number G(k) is defined as being the least number for
which it is true that

664
 THE THEORY OF NUMBERS 21

’ eve?y member FROM A CERTAIN POINT ONWARDS iS the

aurn of G(k) bth powers or less.’
It is Obvious that the exi&nce of g(k) involves that of
G(k), and that Q(k) s g(k). 1Yhen k = 2, both numbers
are 4 ; but G(3) s 8, by Landau’s theorem, while g(3) = 9;
and doubtless G(k) < g(k) in general. It is important also
to observe that, conversely, the existence of G(k) involves
that of g(k). For? if G(k) exists, all numbers beyond- a
certain limit y are sums of G(k) k-th powers or less. But
all numbers less than y are sums of y ones or less, and there-
fore g(k) certainly cannot exceed the greater of G(k) and y.
I said that G(k) seemed to me the more fundamental of
these numbers, and it is easy to see why. Let us assume
(as is no doubt true) that the only numbers which require
9 cubes for their expression are 23 and 239. This is a very
curious fact which should be interesting to any genuine
arithmetician ; for it ought to be true of an arithmetician’
that, ZN has been said of Mr. Ramanujan, and in his case
at any rate with abdolute truth, that Cevery positive integer
is one of his personal friends’. But it would be absurd to
pretend that it is one of the profounder truths of higher
arithmetic : it is nothing more than an entertaining arith-
metical fluke. It is Landau’s 8 and not Wieferieh’s 9 that
is the profoundly interesting number.
The real value of G(3) is still unknown, It cannot be
less than 4 ; for every number is congruent to U, or 1, or
- 1 to modulus 3, and it is an elementary deduction that
every cube is congruent to O, or 1, or - 1 to modulus 9.
Frum this it follows that the sum of three cubes cannot be
of the form 9n2 + 4 or 9 m + 5 : fur such numbers at least
4 cubes are necessary, so that G(3) 2 4. But whether
G(3) is 4, 5, 6, 7, or 8 is one of the deepest mysteries
of arithmetic.
It is worth while to glance at the evidence of computation.

665
 22 SOME FAMOUS PROBLEMS OF

Dam, at the instance of Jacobi,. tabulated the minimum

number of cubes for values of /n from 1 to 12,000, and Dau-
blensky Son Sterneck has extended the table to 40,000,
Some of the results are shown in Table II.

TABLE IX
12‘3 4 5 6 7 8 9
l- 1000 10 41 122 242 293 202 73 15 2
lUOO- 2000 2 27 113 283 358 194 23 - -
9000-10000 1 17 121 377 401 83 - - -
19000-20000 1 12 100 400 426 61 - - -
29000-30000 1 11 105 448 388 47 - - - .
39000-40000 1 13 117 457 384 28 - - -

In each row I have shown a typical thousand numbers,

classified according to the minimum number of cubes by
which they Can be expressed. There are 15 numbers only
for which 8 are needed, fhe largest being 454. There are
X21 for which 7 are needed, the two largest being 58 18 and
8042 ; the distribution of these 121 numbers in the first
9 thousands is
73, 23, 7, 6, 7, 4, 0, 0, 1.
If empirical evidence means anything, it seems clear that
G(3) 2 6. I am sure that Professor Townsend and Professor
Ijndemann have made countless generalizations on evidence
far less substantial.
It is also clear that, throughout von Sterneck’s tables,
there is a fairly steady,
v though latterly very slow, clecresse
in the proportion of numbers for which even 6 cubes are
required ; but that the table is not sufficiently extensive to
give any very decisive indication as to whether these num-
bers disappear or not. 1t seemed to me this was a case in
which further evidence would be worth having. To calculate
a systematic table on the scale required would be a *work
of years. It is possible, however, to obtain some indication

666
 THE THEORY OF NUMBERS 23

of tile probable truth, without any superhuman patience,

by exploring a selected stratum of much larger numbers.
Dr. Ruckle of Giittingen recently undertook this task at
my request, and I am glad to be able to tell you his
results. He found, for the 2,000 numbers immediately
belotr I ,OOO,OOO, the following distribution.

1 2 3 4 5 6 I
998UUU-999000 0 f 98 640 262 1 0
999uou--1000000 I 1 94 614 289 1 0

You will observe that the e-cube numbers have all but
disappeared. , and that there is a quite marked turnover
from 5 to 4. Conjecture in such a matter is extremely
rash, but 1 am on the whole disposed to predict with some
confidence that G (3) s 5. If I were asked to choose between
5 and 4, all I could say would ,be this. That G (3) should
be 9 would harmonise admirably, so far as we can see
at present, with the general trend of Mr. Littlewood’s and
my researches. But it is plain that, if the S-cube numbers
too do ultimately disappear, it can only be among numbers
the writing of which would tax the resources of the decimal
notation ; and at present we cannot prove even that G( 3) ,< 7,
though here success seems not impossible.
With the fourth powers or biquadrates we have been
very much more successful. I have explained that g(4)
lies between 19 and 37. As regards G (a), we have here no
numerical evidence on the same scale as for cubes. Any
fourth puwer is congruent to 0 or f to modulus 16, and
from this it follows that no number congruent to 15 to
modulus 16 can be the sum of less than 15 fourth powers.
Thus G(4) > 15; and Kempner, by a slight elaboration of
this simple-argument, has proved that G(4) 2 16. No
better upper bound wa8 known before than the 37 of
Wieferich, but here Mr. Littlewood and I have been able

667
 24 SOME FAMOUS PROBLEMS OF

to make a very substantial improvement, fist to 33 and

finally to 2 1. Thus G(4) lies between 16 and 21, and
the margin is comparatively small.
I turn now to the general case, In the years up to
1909, the existence proof was effected, and upper bounds
for g(k) determined, for the values of k from z to 8 in-
clusive and for Jz = IO. These upper bounds are shown
in the first row of Table III ; that for 10, which is not
included, is somewhere in the neighbourhood of 140,000.

TABLE III.
2 3 4 5 6 7 8
Y(k) e 4 9 37 58 478 3806 31353
&(#)q+2"-2= 4 9 19 37 73 143 279
NW2 4 PI 37 58 478 3806 31353
G(k)s(k-2)2k-1$5= (5) (9) 21 63 133 335 773
G(k)zk+1,4k 4 4 l6 6 7 8 32

In the second I have shown the beat known lower

row
bounds, which are given by the simple general formula
which stands to the left, in which [(Uk] denotes the
integral part of (s)kg It is easily verified, in fact, that
the number
([(gq-1) @+ 9-1,
which is less than Sk, requires the number of k-th powers
stated? It will be observed that the first three numbers
are those which occur in Waring’s enunciation.
Waring’s problem, as I have defined it-the prublem,
that is to say, of tiding a general existence proof
for g(k), and a frtiori for G(k)--wan ultimately solved
by Hilbert, once more in IW% I wish that I had time
to give a proper account of his justly famous memoir,
which raised the whole discussion at once on to an

1 This observation wm made by Bretschneider in 1853.

668
 THE THEORY OF NUMBERS 25

altogether higher level. As it is, I must confine myself

to one or two extremely inadequate remarks. The
proof falls into two parts. The first part is what I may
call semi-transcendental. It is not fully transcendental in
the sense in which, for exa#mple, the classical proofs in the
theory of the distribution of primes are transcendental, for
it does not make use of the machinery of the theory of
analytic functions of a complex variable ; but it uses the
methods of the integral calculus, and is therefore not
fully elementary. Hilbert set out with what would appear
at first sight to be the singularly ill.- adapted weapon of
a volume integral in space of 25 dimensions, a number
which he wzcs afterwards able to reduce to 5. The formula
which he ultimately used is
(Xl2 + xz” + x,2 + x2 + xj2p

-c (x,t,+x,f,+x,t,+x,t,+xgt,)2kdtl . ..dt.,

where C is 21certain constant, viz.

and the integration is effected over the interior of, the

hypersphere
t,2+t22+t,2+t,2+t,2 = 1.

Starting from this formula he was able, by an exceed-

ingly ingenious process based upon the definition of a
definite integral as the limit uf a finite sum, to prove
the existence in the general case of algebraical identities
analogous to that used by Liouville and his followers
when I;: is 4. It should be observed that Hilbert’s
proof is essentially an &&ewe proof; his method is
not effective for the actual determination of these identities

669
 26 SOME FAMOUS PROBLEMS OE

even in the simplest cases. The identities which are known

for special values of k have been obtained by commolz
algebra, and are, after the first few values of E, excessively
complicated. The simplest known identity for LY= IO, for
instance, is

22680 (xl2 + q2 + x,” + x42)5

0) (48)
= 92 (x1+x~*q+xpp+
- - ~(2x,+x,+s3y0

(12) (4)
+ 1q (x2+x,)“‘+ 92 (2xJ0,

where t11e in brackets show the number of

under the signs of summation. However, the identities
exist ; and it should be clear to you, after our discussion of
the case 1~= 4, that they enable us :tt once to obtain a
proof in succession for k = 2, 4, 8, 16: . .. and generally
whenever II: is a power of 2. This co&udes the first and
most characteristic part of HiTbert’s argument. The second
part, in which the conclusion is extended to every value of
k, is purely algebraical.
Hilbert’s work has been reconsidered and simplified by
a number of writers, most notably by Dr. Stridsberg of
Stockholm, and the ultimate result of their work has been
to eliminate the transcendental elements from the proof
entirely. The proof, as they have left it, is fully elemen-
tary ; it does not involve any reference to integrals, or to
any kind of limiting process, but depends simply on an
ingenious system of equations derived by the processes of
finite algebra. It remains a pure existence proof, and
throws no light on the value of g(k).
It would hardly be possible for me to exaggerate the
admiration which I feel for the solution of this historic
problem of which r have been compelled to give 80 bald

670
 THE THEORY OF NUMBERS 27

and summary a clescription. Within the limits which it

has set for itself, it is absolutely and triumphantly $uccess-
ful, and it stands with the work of Hadamard and de la
ValMe-Poussin, in the theory of primes, as one of the land-
marks in the modern history of the theory of numbers.
But there is an enurmous amount which remains to be
done, and it would seem that, if we are to interpret
Waring’s problem in the widest possible sense, if we are to
get into real con&t with the actual values of our numbers
g(k) and G(k), still more if we are to attack all the obvious
problems connected with the number of representatidas,
then essentially diKerent and inherently more powerful
methods are required. There is one armoury only in which
such more powerful weapons can be found, that of the
modern theory of functions. In short we must learn how
to apply Cauchy’s Theorem to the problem, and that is
what Mr, Littlewood and I have set out to do.
The first step is fairly obvious. The formulae are
slightly simpler when X: is ewe~iz,. The number of represen-
tations of ‘1%as the sum of s lath powers, which we may
denote in general by
%,$ (4
is* then the coefficient of xl2 in the generating function

where

This formula involves certain conventions as to the order

and sign of the numbers which occur in the representations
which me to be reckoned as distinct ; but the complications
so introduced are trivial and I need nut dwell on them.
The series is convergent when 1x 1~ 1, and, by Cauchy’e
 28 SOME FAMOUS PROBLEMS OF

Theorem, we have
1 (f(W (jx
rk,&) = Gi - p+l ?
s

the path of integration being a circle whose centre is at

the origin and whose radius is less than unity.
All this is simple enough; but the further study of the
integral is very intricate and difficult, and I cannot attempt
to do more than to give a rough idea of the obstacles that
have to be surmounted, -Let us contrast the integral for a
moment with that which would stand in its place in the
‘ trivial ’ problem to which I referred early in my lecture.
There the subject of integration would be a rationd func-
tion, with a finite number of poles all situated on the unit
circle. We could deform the contour into one which lies
wholly at a considerable distance from the origin and in
which, owing to the f’rcctor znfl in the denominator, every
element is very small when AUis large. We should have,
of course, to introduce corrections corresponding to the
residues at the poles ; and it is just these corrections which
MTould give the dominant terms of an approximate formula
bv means of which our coefficients could be studied. In
tie present case we have no such simple recourse ; for
every point of the unit circle is a singularity of an exceed-
ingly complicated kind, and the circle as a whole is a
barrier across which it is impossible to deform the contour.
It is of course for this reason that no successful application
of the method has been made before.
Our fundamental idea for overcoming. the difkulty is
as follows. Among the continuous mass of singularities
which covers up the circle, it is possible to pi& out a class
which to a certain extent dominates the rest. These special
singularities are those associated with the rational points
of the circle, that is to say, the points
x = e2rnipJ,

672
 THE THEORY OF NUMBERS 29

where p/q is a rational fraction in its lowest terms. This

class of points is indeed an i?zJilzite class ; but the infinity
is, in Cantor’s phrase, only an erlzurrnerable infinity; and
the points can therefore be arranged in a aimply i&Ae
series, on the model of the series

In the neighbourhood of these points the behaviour of the

function is, sufficiently complex indeed, but simpler than
elsewhere. The function hrts, to put the matter in a rough
and popular way, a general tendency ti become large in
the neighbourhood of the ‘unit circle, but this tendency is
must pronounced near these particular points. They are
nut only the sip@& but also the heaviest singularities;
their weight is greatest when’ the denominatur q is smallest,
decreases as q increases, and (as a physicist would say)
becomes infinitely small when g is infinitely large. There
is, therefore, at any rate, the hope that we may be able to
isulate the contributions of each of these selected points,
and obtain, by adding them together, a series which may
give a genuine approximation to our coefficient.
I owe to Professor Harald Bohr of Copenhagen a
picturesque illustration which may help to elucidate the
general nature of our argument. Imagine the unit circle
as a thin circular rail, to which are attached an infinite
number of small lights bf varying intensity, each illumi-
nating a certain angle immediately in frunt of it. The
brightest light is at x = I, corresponding to 13 = 0, Q = 1 ;
the next brightest at x = - 1, corresponding to p = 1,
4 = 2 ; the next at x = Pi/3 and eaaii3, and so on, We
have to arrange the inner circle, the circle of integration,
in the position of maximum illumination. If it is too far
away the light will nut reach it; if too near, the arcs
which fall within the angles of illbminatioti will be too

613
 30 SOME FAMOUS PROBLEMS OF

nan~w, and the light will not cover it completely. Is it

possible to place it where it will receive a satisfactorily
uniform illumination ?
The answer is that this is onZy possible when 1~ is 2.
Our functions are then elliptic functions ; the lights are the
formulae of the theory of linear transformation ; and we
can find a position of the inner circle in which it falls
entirely under their rays. We are thus led to a solution of
the problem of the squares which is in all essential respects
complete. But when k exceeds z the result is less satisfac-
tory. The angle of the lights is then too narrow; the
beams which they emit, instead of spreading out, with
reasonable regularity, are shaped like torpedoes or cigars ;
however we move our circle a part remains in darkness.
It would seem that this difficulty, which held up our
researches for something like two years, is the really
characteristic difficulty of the general problem. It cannot
be solved until we have found some other source of light.
It was only after the most prolonged and painful efforts
that we were able to discover such another source, It is
possible not only to hang lights upon the rail, but also, to
a certain extent, to cause the rail itself to glow. The
illumination which can be induced in this manner is irri-
tatingly faint, and it is Xor this reason that our results are
not yet all that we desire ; but it is enough to make the
dark places dimly visible and to enable us to prove a great
deal more than has been proved before.
The actual results which we obtain are these. We find
that there is a certain series, which we call the &g&r
series, which is plainly the key to the solution. This
aerie is

674
 THE THEORY OF NUMBERS 31

where

-a sum which reduces, when /C= 2, to one of what are

known as c Gauss’s sums ’ -and the summation extends,
Srst to all values of p less than and prime to q, and
secondly to all positive integral values of q. The genesis
of the series is this. associate with the rational point
auxiliary series

which (cc) is as simple and natural’ as we can make it, and

(b) behaves perfectly regularly at all points of the unit circle
except at fhe one point with which we are particularly con-
cerned. We then add together all these auxiliary functions,
and endeavour to approximate to the coefficient of our original
series by summing the auxiliary coefficients over all values
of p and q. The process is, at bottom, one of ’ decom-
position into simple elements ‘, applied in an unusual way.
Our final formula for the number of representations is

the second term denoting an error less than a constant

multiple of rtb, and c being a number which is less than
s
i- - 1 at any rate for sufficiently large values of s. The
iitA
second term is then of lower iorder than the first. Further,
the first term is real, and it may be shown, if s surpasses
a certain limit, to be positive. If both these conditions
are saCsfied, and 32 is sufficiently large, then y7;, 8 (3~)

675
 32 SOME FAMOUS PROBLEMS OF

cannot be zero, and representations of 121by s kth powers

certainly exist. The way is thus open to a proof of the
existence of G(k); if G(k) exists, m also does g @c)~and
Waring’s problem is solved.
The structure of the dominant term in our general
formula ia best realized by considering some special cases.
In Table IV I have written out the leading terms of S,

TABLE IV.

k- 2.

2 2 4
+- cos g n7r + GO9 +.u + **,a
5&S i ( P-= )I

lc = 4, s = 33.

k = 4, s = 21.

first k=2 and s is arbi .trary, and for 7

cubes and for 33 and 21 biquadratea. There are certain
characteristic8 common to all these series. The terms
diminish rapidly ; in eIM=h case only a very few are uf real
importance : and they are oscillatory, with a period which
increases as the amplitude of the oscillations decreases.
The series for the cubes is easily shown to be positive;
but we cannot; deduce that 7*3, 7 (n) is positive, and draw
consequences as to the representation of numbers by 7 cubes,

676
 THE THEORY OF NUMBERS 33

because in this case we cannot dispose satisfactorily of the

error term 0 frt”) in the general formula. In the two
citges relating ti fourth powers which I have chosen, the
discussion of the series itself is rather more delicate, for
there is in each of them one term which can be negative
and greater than I. But the discussion can be brought to
a satisfactory conclusion, and, as in this case we are able
to prove that the error term is really of lower order, we
obtain what we desire. Ewery large mmber is the mm of
21 fbwth powers or less; G( 4) 2 21, Further, we have
obtained a genuine asymptotic formula for the number
of representations, which can be used for the study of the
representations of numbers of particular forms. We can
show, for example, that a large number of the form
I6 32+ 11) can be expressed by 2 1 biquadrates in about 200
times more ways than one of the form 16~~ + 2.
If the method of which I have tried to give some general
idea +Is compared with those which have previously been
applied to the problem, it will be found that it has three
very great advantages, In the first place it is inherently
very much more powerful. It brings us for the first time
into relation with the series on which the solution in the
last resort depends, and tells us, approximately but truly,
what the number of representations really is. Secondly, it
gives us numerical results which, as soon as k exceeds 3, are
far in advance of any known before, These numbers
are those in the fourth row of Table III.1 It will be seen
that these numbers conform to a simple law, and that is
the third advantage of the method, that it is not a mere

f The thick type indicates a new result. The (5) and (9) in round
brackets are inferior to results already known, Our method is easily
adapted to deal with the case k = 2 completely ; but it will not at
present yield Landau’s 8, which is therefore tinclosed in square
brackets.

677
 34 THEORY OF NUMBERS

existence proof, but gives us a definite upper bound for G(k)

for all values uf k, viz.
G(k) s (k- 2) Zkml+ 5.
In the la& row of the table I have shown all that is known
about G(Ic) on the other side. In all cases G(k) 2 lc + 1,
while if 7Cis a power of 2 we can say more, namely that
G(k) >= 41~. A compa;rison between this row of figures and
that above it is enough to show the room which remains
for further research. It is beyond question that our
numbers are still very much too large ; and there is no
sort of finality about our researches, for which the best
that we claim is that they embody a method which opens
the door for more.
I will conclude by ode w.ord as to the application of
our method to another and a still more diEcult problem.
It was asserted by Goldbach in 1742 that every ew~3~NT&~-
&r is the SUV~ of two odd Times. Goldbach’s assertion
-remains unproved ; it has not even been proved that every
number 3’~is the sum of 10 primes, or of 100, or of any number
independent of #IL Our method is applicable in principle to
this problem also. We cannot solve the problem, but we
can open the first serious atttik upon it, and bring it into
relation with the established prime number theory. The
most which we can accomplish at present is a~ follows.
We have to assume the truth of the notorious Riemann
hypothesis concerning the zeros of the Zeta-function, and
indeed in a geaeralized and extended form. If we do this
we can prove, not Goldbach’s Theorem indeed, but the next
best theorem of the kind, viz. that eve3qyodd nzc&e~, at any
rate from a certain point onwards, is the ~U/IYLof three ocld
primes. It is an imperfect and provisional result, but it is
the of the proble1n.

678
 CORRECTION
23. 31. In the formula for r&)I on p. 31, read r(s/k) in the denominator.

COMMENT
A French translation, by A, Sallin, with notes by Hardy on subsequent developments, appeared
in I93 1 as Trois ~obl&nes c&&es de Za thkorie des no&es (les Prews Universitairw de France).

679
 ARRANGEMENT OF THE VOLUMES
VOLUME I

I, 1 Diophantine approximation

I. 2 Additive number theory

(a) Combinatory analysis and sums of squares
(6) Waring’s Problem
(c) Goldbach’s Problem
(d) Inaugural Lecture (Oxford, N2Q)

VOLUME II

II. 1 Multiplicative number theory (including the zeta-function)

II. 2 Other number theory

II. 3 Inequalities
VQEUME III
III. 1 Trigonometric series
(a > Convergence of a Fourier series or its conjugate
(b) Summability of a Fourier series or its conjugate
(c> The Young-Hausdorff inequalities
(4 Special trigonometric series
(e> Other papers on trigonometric series
III. 2 Mean 1 values of power series

VOLUME IV
IV. I Special functions
(a) Zeroes and asymptotic behaviour of particular integral functions
(b) Taylor series and singularities
(c) Orders of infinity
(d) Miscellaneou s
IV. 2 Theory of functions

VOLUME V
V. Integral calculus

681
 VOLUME VI
VI . Theory of series
VOLUME VII
VII. 1 Integral equations and integral transforms

VII. 2 Miscellaneous papers

VII. 3 Questions from the EdwationaZ Times

VII. 4 Obituary notices by G. H. Hardy

VII. 5 List of other writings

 LIST OF PAPERS BY G. H. HARDY

Abbreviations
N.I.C. Notes on some points in the intepal calculus
D.A. Some problems of Diophantine approximation
P.N. Some problems of ‘ Partitio Numerorum’
N.S. Notes on the theorv v of series

1899
1. Question 13848, EducationaL Times, 70, 43.
2. Question 13917, Educational Times, 70, 78-79.
3. Question 14124, Educational Times, 71, 100-101, I VII. 3
4. Question 14005, Educationat Times, 71, 111-112.

1900
1. On a class of definite integrals containing hyperbolic functions, Messenger of M&,e-
matics, 29, 25-42. V
2. Question 14243, Educational Times, 72, W-81.
3, Question 14271, Educational Times, 73, 36-37. I VII. 3
4. Question 14179, Educational Times, 73, 53-54.
5. Question 14317, Educational Times, 73, 61-63.

1901
1. On differentiation and integration under the integral sign, Quarterly Journal of Mathe-
matics, 32, 66-140. (Corrected in 1915, 2.) V
2. General theorems in contour integration, with some applications, QuarterZy Journal of
Mathematics, 32, 369-384. V
3. N.I.C. I: On the formula for integration by parts, Messenger of Mathematics, 30, 18%
187. V
4. N.I.C. II: Two general convergence theorems, Messenger of lkfuthematics, 30, 187-190. V
5, Question 14496, Educational Times, 74, 37-38.
6. Question 14447, Educationa Times, 74, 98-100.
7. Question 14467, Educational Times, 74, 111-l 12. I VII. 3
8. Question 14028, Educational Times, 74, 122-123.
9. Question 14369, Educational Times, 75, 135-136.

1902
1. The elementary theory of Cauchy’s principal values, Proceedings of the London Mathe-
matical Society, (1) 34, 16-40. V
2. The theory of Cauchy’s principal values (Second Paper : the use of principal values in
some of the double limit problems of the integral calculus), Proceedirngs of the London
Mathematical Society, (1) 34, M-91. V

683
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1902 (cont.)
3, On the FrullEtnian integral

Quctrterly Journal of MatheLatics, 33, 113-144. V

4. N.I.C. III : On the logarithmic criteria for the absolute convergence of an integral whose
upper limit is 00, Messenger of Mathematics, 31, 1-6. V
5. N.I.C. IV: On the integral rsin x$(x) dx, Messenger of Mathematics, 31, 6-S. V
6. A new proof of Kummer’s series for log l?(a), JJessenger of Afathematics, 31, 31-33. IV. 1 (d)
7. N.I.C. V: On absolutely convergent integrals of functions which are infinitely often
infinite, Messenger of Mathematics, 31, 73-76. V
s, N.I.C. VI: Absolute convergence of infmite multiple integrals, Messenger Mathe-
matics, 31, 125-128. V
9. N.I,C. VII: On differentiation under the integral sign, Messenger of Mathematics, 31,
132-134. V
10. NLC. VLIZ: Absolutely convergent integrals of irregular types, Messenger of Muthe-
matics, 31, 177-183. V
11. On the zeroes of the integral function
x-sinx = 5 (-)n-l *!,
1 l

Messenger aj Mathematics, 3 I, 16 l-l 65. IV. 1 (a)

12. Questions 1423, 2316, 3941, 4794, Educutionul Times, (2) 1, 25,
13. Question 14851, Educational Times, (2) 1, 58-59. VII. 3
14. Question 14055, Educational Times, (2) 2, 41-42,

1903
1. The theory of Cauchy’s principal values (Third Paper: differentiation and integration
of principal values), Proceedings of the London MathematicuZ Society, (1) 35, 81-107. V
2, On the continuity and discontinuity of definite integrals which contain a continuous
parameter, Quarterly Journal of Muthematics, 34, 28-53. V
3. Note on the limiting values of the elliptic modular functions, QuurterZy Journal of Mathe-
matics, 34, 76-86. IV. 1 (d)

4. N.1.C IX: On the integral r{A-@i~?x))#(x)dx, Messenger of Muthemutics, 32, l-3, V

5. On the zeroes of certain integral functions, Messenger of Mathemtics, 32, 36-45. IV. 1 (a)
co
log(ux2+2bx+c)2
6. On the integral dx, Messenger of Mathematics, 32, 45-50, V
s ax2+2px+y

7. N.1.C. X: On conditionally convergent infinite multiple integrals, Messenger of Mathe-

matics, 32, 92-97. V
8. NLC, XI : Some conditionally convergent infmite double integrals, Messenger of Mathe-
matics, 32, 159-165. V
9. %I,C. XII: On the operation which is the inverse of double integration, Messenger of
Mathematics, 32, 187-192. V
10. Question 14988, Educational Times, (2) 3, 94-95.
11. Question 14989, Educational Times, (2) 4, 69-70. VII. 3
12. Question 15019, Educational Times, (2) 4, 75.
13. Question 15265, Education& Times, (2) 4, 109-110. I

684
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1904
1. On the convergence of certain multiple series, Proceedings of the London Mathematical
Society, (2) 1, 124-128, VI
2. A general theorem concerning absolutely convergent series, Proceedings of the London
MathematicaZ Society, (2) 1. 285-90. VII. 2
3. On differentiation and integration of divergent series, Transactions oj the Cambridge
PhiLosophicaL Society, 19, 297-32 1. VI
4. Researches in the theory of divergent series and divergent integrals, Quarterly Journal
of Mathematics, 35, 22-66, VI
5. A theorem concerning the infinite cardinal numbers, Quarterly Journal ~j Mathematics,
35, 87-94. VII. 2

6. Note on the function ye ~(~2-f2)dt 1 Quarterly Journal of Mathematics, 35, 193-207, VII* 1

7. The asymptotic solution of certain transcendental equations, Quwterly Journal of

Mathematics, 35, 261-282. IV. 1 (a)
8. N.I.C. XIII : On differentiation under the integral sign (continued), Messenger of Mathe-
matics, 33, 62-67. V
9. The cardinal number of a closed set of points, Messenger of Mathematics, 33, 67-69. VII. 2
10, N.1.C XIV : Integrals whose d&continuities are everywhere dense, Messenger of Mathe-
matics, 33, 80-M. V
11. Note on divergent Fourier series, Messenger of Mathematics, 33, 137-144. III. 1 (b)
X2. On the zeroes of two classes of Taylor series, British Association Report, 441-443. IV. I (a)
13. Question 15300, Educational Times, (2) 5, 61,
14. Additional note on Question 15282, Educational Times, (2) 5, 113-114.
VII. 3
15. Question 15361, Educational Times, (2) 5, 118.
16. Question 15125, EducationaL Times, (2) 6, 31.

1905
1
1 l On the rootsof the equation - = c, Proceedings of the London Mathematical Society,
qx+ 1)
(2) 2, l-7, IV. I (a)
2. (With T. J. I’A. Bromwich.) Some extensions to multiple series of Abel’s theorem on the
continuity of power series, Proceedings of the London Mathematicd Society, (2) 2, 161-
189. VI
3. Note in addition to a former paper on conditionally convergent multiple series. Pro-
ceedings of the London Mathematical #ociety, (2) 2, 190-191. VI
4. On the zeroes of certain classes of integral Taylor series, Part I: On the integral function
co xm)
~c= pw Proceedings of the London Mathematicul Society, (2) 2, 332-39. IV. 1 (a)

5. On the zeroes of certain classes of integral Taylor series, Part II: On the integral func-

Xn.
tion - and other similar functions, Proceedings of the London Mathematical
c (n+a)%!
nIz=()‘

Society, (2) 2, 401-431. IV. 1 (a)

6. A method for determining the behaviour of certain classes of power series near a singular
point on the circle of convergence, Proceedings of the London MathematicaE Society,
(2) 3, 381-389. IV. l(6)
7. On a class of analytic functions, Proceedings of the London Muthematical Society, (2) 3,
441-460. xv, 1 (b)

685
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1905 (cont.)
8. On certain series of discontinuous fwlctions connected with the modular functions,
Quarterly Journal of Mathemcttics, 36, 93-123. IV. 1 (d)
9. Note on an integral function, Messenger of Mathemutics, 34, 1-2. IV, 1 (a)
10. N.I.C. XV: On upper and lower integration, Messenger of Mathematics, 34, 3-6. V
11. N.I.C. XVI: A class of conditionally convergent infinite multiple integrals, Messenger
of Muthematics, 34, 6-W. V
12. A generalization of Frullani’s integral, Messenger of Mathematics, 34, 1 l-18, and note,
p. 102. V
13. On the zeroes of a class of integral functions, Messenger of Mathematics, 34, 97-301. IV. I (a;)
14. On certain conditionally convergent multiple series connected with the elliptic functions,
Messenger of Mathematics, 34, 146-153. VI
15, Question 15686, Educational Times, (‘2) 8, 74. VII. 3
16. The expression of the double zeta function and double gamma function iti terms of
elliptic functions, Tranwctions of the Cambridge Philosophical Society, 20, l-35. IV. 1 (d)

1906
1. On Kummer’s for log r(a), Quarterly Journal of Mathematics,
series 37, 49-53. IV. 1 (d)
2. On double Fourier series, and especially those which represent the double zeta-function
with real and incommensurable parameters, Quarterly Journal of Mathematics, 37,
53-79. IV. 1 (d)
3. On the function p,(x), QuurterZy JournuZ gf Mathematics, 37, 146-172 (correction at end
of 1906, 5). lV* 1 (a)
4. On certain double integrals, Quurterly Journal of Muthematicsy, 37, 360-9. V

5. On the integral function @ Quarterly Journul of Muthematics,

37, 369-378. IV, 1 (a)

6. N.X.C. XVII: On the integration of series, Messenger of Mathematics, 35, 126-130. V
7. A formula for the prime factors of any number, Messenger of Mathematics, 35, 145-146. II. 1
8. N.I.C. XVIIl: On some discontinuous integrals, Meesenger of Muthemutics, 35, 158-
166. V
9. Some notes on certain theorems in higher trigonometry, Muthematic& Guzetk, 3, 284-
288. I’*

1907
1. The continuum and the second number class, Proceedings of the London Mat~ematicul
Society, (2) 4, N-17. VII. 2
2. Some theorems connected with Abel’s theorem on the continuity of power series, Pro-
ceedings of the London Mathematical Society, (2) 4, 247-265. VI
3. On the singularities of functions defined by Taylor’s series (Remarks in addition to
a former paper), Proceedings of the London Mathematical Society, (2) 5, 197-205. IV. l(6)
4, The singular points of certain classes of functions of several variables, Proceedings of
the, London Mathematical Society, (2) 5, 342-360. IV. 1 (b)
5. On certain oscillating series, Quarterly Journal of Mathematics, 38, 269-288. VI
6. Some theorems concerning infinite series, Muthematische Annulen, 64, 77-94. VI
7. N.I.C. XIX: On Abel’s lemma and the second theorem of the mean, Messenger of Muthe-
matica, 36, 10-13. ‘V
8. Higher trigonometry, Mathematical Guxette, 4, 13-14. VII. 2
9. A curious imaginary curve, Mathematical Gazette, 4, 14. VII. 2
10. The line at infinity, etc., Mathematical Gazette, 4, M-15. VII* 2

686
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1908
1. Generalization of a theorem in the theory of divergent series, Proceedings of the London
Muthemutical Society, (2) 6, 255-264. VI
2. The multiplication of conditionally convergent series, Proceedings of the London Mathe-
matical Society, (2) 6, 410-423. VI
3. Further researches in the theory of divergent series and integrals, Transactions of the
Cambridge Philosophical Society, 21, l-48. VI
4. (With T. J. I’A. Bromwich.) The definition of an infinite integral as the limit of a kite
or infinite series, Quarterly Journal of Mathematics, 39, 222-240, V
5. Some multiple integrals, Quurterl~ Journal of Muthematics, 39, 357-375. V
6. N,I,C, XX: On double Frullanian integrals, AZessenger of Mathematics, 37, 96403. V
7. N.I.C. XXI
: On a conditionally convergent multiple integral, Messenger of Muthe-
matics, 37, 127-130. V
8. KLC. XXII: On double Frullanian integrals (cont.), Messenger of Mathematics, 37,
154-161, V
9, Question 16257, Educ&mul Times, (2) 13, 79-W. VII, 3
10. Mendelian proportions in a mixed population, Science (American Association for the
Advancement of Science), new series 28, 4940. VI3. 2

1909
1, A note on the continuity or discontinuity of a function defined by an infinite product,
Proceedings of the London MathematicaZ Society, (2) 7, 40-48. VI
2. The theory of Cauchy’s principal values (Fourth Paper: the integration of principal
values-continued-with applications to the inversion of definite integrals), PTO-
ceedings of the London Mathematical Society, (2) 7, 181-208. V
3. On an integral equation, Proceedings of the London Mathematical Society, (2) 7, 445-
472. WI. I
I
4, K1.C. XXIII: On certain oscillating cases of Dirichlet’s integral, Messenger of Mathe-
matics, 38, l-8. V
5. On certain definite integrals whose values can be expressed in terms of Bessel’s func-
tions, Messenger of Mathematics, 38, 129-132. V
4. N.I,C. XXIV: Oscillating cases of Dirichlet’s integral (cont.), Messenger of Muthe-
mutics, 38, 17zi-185 (correction at end of 1911, 3). V

7. The integral Kxdx, Mathematical Gazette, 5, 98-103. V

s X
0
1910
1. The application to Dirichlet’s series of Borel’s exponential method, of summation,
Proceedings of the London Mathematical Society, (2) 8, 277-294. VI
2. The ordinal relations of the terms of a convergent. sequence, Proceedings of the London
Mathematical Society, (2) 8, 295-300. VII. 2
3. Theorems relating to the summability and convergence of slowly oscillating series,
Proceedings of the London Mathematical Society, (2) 8, 301-320. VI
4. The maximum modulus of an integral function, Quarterly Journal of Mathematics, 41,
1-9. IV. 2
5. On certain detiite integrals considered by Airy and by Stokes, Quarterly JournaZ of
Muthematics, 41, 226-240. IV. 1 (d)
6, N.1.C. XXV: Absolutely convergent integrals of irregular types (cont.), Messenger of
Mathematics, 39, 28-32. V
Xnp
7, The zeroes of the integral function -, and of Some similar functions, Messenger of
n2!
Mathematics, 39, 88-96. IV. 1 (a)

687
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1910 (cont.)
8. N.I.C. XXVI: On a case of term-by-term integration of an infmite series, Messenger of
Mathematics, 39, 136-139. V
9, To fkd an approximation to the large positive root of the equation e@ = 101oxlOelO’OxlO,
Muthematical Gazette, 5, 333-334. VII. 2

1911
1. Theorems connected with Maclaurin’s test for the convergence of series, Proceedings of
the London Muthemuticsl Society, (Z), 9, 126-144. VI
2. (With S. Chapman.) A general view of the theory of summable series, Quarterly Journal
of Mathematics, 42, 181-215. VI
3. N.I.C. XXVII: Oscillatjing cases of Dirichlet’s integral (cont.), Messenger of Mathe-
matics, 40, 44-53. V
4. A class of definite integrals, Messenger of 17Mathemutics, 40, 53-54. V
5, N.I.C. XXVIII: A conditionally convergent double integral, Messenger of Mathematics,
40, 62-69 l V
6. N.I.C. XXIX: Two convergence theorems, Messenger of Mathematics, 40, 87-91. V
7. N.I.C. XXX: A theorem concerning summable integrals, Messenger of Mathematics, 40,
108-112. VI
8, N,I.C. XXXT: The uniform convergence of Borel’s integral, Messenger of Mathematics,
40, 161-165. VI
9. Foprier’s double integral and the theory of divergent integrals, Transactions of the Cam-
bridge PhilosophicaE Society, 2 1, 427-45 1. VIII:. I

1912
1 Properties
l of logarithmico-exponential functions, Proceedings of the London Mathe-
matical Society, (2) 10, M-90. IV. 1 (c)
2. On the multiplication of Dirichlet’s series, Proceedings of the London Mathematical
Society, (2) 10, 396-405. VI
3, Some results concerning the behaviour at infinity of a real and continuous solution of
an algebraic differential equation of the first order, Proceedings of the London Mathe-
m,aticaZ Society, (2) 10, 451-468. VII. 2
4. (With J. I% L,) S ome problems of Diophantine approximation, 5th International Congress
of Mathemuticians, Cambridge, 1, 223-229. I. 1
5, Generalizations of a limit theorem of Mr. Mercer, Quarterly Journal of Mathematics, 43,
143-150. VI
6. Note on Dr. Vacca’s series for y, Quurterly Journal of Mathematics, 43, 215-216. IV. 1 (d)
7. Note on a theorem of Ces$ro, Messenger of Mathemat~ics, 41, 17-22. VI
8. N.I.C, XxX11: On double series and double integrals, Messenger of Mathematics, 41,
4&48. V
9, N.I.C. XxX111: Some cases of the inversion of the order of integration, Messenger of
Afathemtics, 41, 102-109. VII. 1

1913
1. (With J. E. L.) Th e relations between Borel’s and Ces&ro’s methods of summation,
Proceedings of the London Mathematical Society, (2) 1 I, L-16. VI
2. (With 3, E. L.) Contributions to the arithmetic theory of series, Proceedings of the Lon-
don Mathematical Society, (2) 11, 41 l-478. VI
3. An extension of a theorem on oscillating series, Proceedings of the London Mathematical
Society, (2) 12, 174-180. VI

688
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1913 (cont.)
4, On the summability of Fourier’s series, Proceedings of the London Mathematical Society,
12, 365-372. III. 1 (b)
5. Oscillating Dirichlet’s integrals : an essay in t.he ‘Infinit&rcalciil’ of Paul Du Bois-
R,eymond, Quarterly Journal of Mathematics, 44, l-40. IV. 1 (c)
6, A theorem concerning Taylor’s series, Quarterly Journal of Mathematics, 44, 147-60, IV. 2
7. Oscillating Dirichlet’s integrals (Second Paper), QuurterZy Journal of Mathematics, 44,
242-263. IV, 1 (c)
8. N.X.C. XXXIV: Absolutely convergent integrals of irregular types (cont.), Messenger
of Mathematics, 42, 13-18. V
9. N.I.C. XXXV: On an integral equation, Messenger of Mathematics, 42, 89-93. VII. 1
10. (With J. E. L.) Tauberian theorems concerning series of positive terms, Messenger of
Afathematics, 42, 191-192. VI
11. (U7ith J. E. L.) S ur la s6rie de Fourier d’une fonction & carr6 sommable, Comptes Rendus,
156, 1307-1309. III. 1 (b)

1914
1. Sur les zeros de la fonction c(s) de Riemann, Comptes Rendus, 158, 1012-1014, 11. 1
2. (With J. E, L.) D.A. I: The fractional part of &, Acta Mathemutica, 37, 155-191. I. 1
3. (With J, E. L.) D.A. II: The trigonometrical series associated with the elliptic 6-
functions, Acta Mathewzatica, 37, 193-239. r. 1
4, (With J. E. L.) Tauberian theorems concerning power series and Dirichlet’s series whose
coefficients are positive, Proceedings of the London Mathematical Society, (2) 13,
174-191. VI
5. Note on Lambert’s series, Proceedings of the London Mathematical Society, (2) 13, 192-
198. VI
6. Note in addition to a paper on Taylor’s series, Quarterly Journul of Mathematics, 45,
77-84. IV. 2
7. A function of two variables, Quarterly Journal of Mathematics, 45, 85-113. IV. l(6)
8. N.T.C. XXXVI: On the asymptotic values of certain integrals, Messenger of Mathe-
matics, 43, 9-13. V
9. N.I.C. XXXVII: On the region of convergence of Borel’s integral, Messenger of Mathe-
matics, 43, 22-24. VI
10. N.I.C. XXXVIII: On the definition of an analytic function by means of a definite
integral, Messenger of Mathematics, 43, 29-33, IV, 2
11. (With J. E. L.) Some theorems concerning Dirichlet’s series, Afessenger of Mathematics,
43, 134-147. VI

1915
1. (With J, E. L.) New proofs of the prime-number theorem and similar theorems, Quarter&
Journal of Mathematics, 46, 2 15-2 19. 11. 1
2. Correction of an error, Quarterly Jownal of Mathematics, 46, 261-262. 1
3. On the expression of a number as the sum of two squares, Quarterly Journal of Mathe-
matics, 46, 263-283. II, 2
4. The mean value of the modulus of an analytic function, Proceedings of the London
Mathematical Society, (2) 14, 269-277. III. 2
5. Proof of a formula of Mr. Ramanujan, Messenger of MathenWics, 44, 18-21. v
6. N.I.C. XXXIX : Further examples of conditionally convergent infinite double integrals,
Messenger of Mathematics, 44, 57-63. 1’
7. N.I.C. XL: Some cases of term-by-term integration of an infinite series, Messenger of
.- -- Mathematics, 44, 145-149. v

689
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1915 (cont.)
8. N.I.C. XLI: On the convergence of certain integrals and series, Messenger of Mathe-
matics, 44, 163-166. V
9. Sur le problbme des diviseurs de Dirichlet, Comptes Rendw, 160, 617-619. 11. 2
10. Prime numbers, British Association Report, 350-354. II. I
11 Example
l to illustrate a point in the theory of Dirichlet’s series, Z”he T&oEu Muthe-
m&ml Journal, 8, 59-66. VI
12. The definition of a complex number, Mathematical Gazette, 8, 48-49. VII. 2

1916
1. The application of Abel’s method of summation to Dirichlet’s series, Quarterly JOUWMZZ
of Mathermtics, 47, 176-192. VI
2. Weierstrass’s non-differentiable function, Tranmctio~ of the Amemkun Mathematical
Society, 17, 301-325. IV. 1 (d)
3, (With J. E. L.) D.A.: A remarkable trigonometrical series, Proceeding8 of the Nation&
Academy of Sciences, 2, 583-586. I. 1
4. On Dirichlet’s divisor problem, Proceedings of the London MathematicaZ Society, (2) 15,
l-25. II, 2
5. The second theorem of consistency for summable series, Proceedings of the &ondon
Muthemtical Society, (2) 15, 72-88. VI
6. The average order of the arithmetical functions P(x) and A(x), Proceeding8 of the London
Mathematical Society, (2) 15, 192-213. II. 2
7. Sur la sommation des s&ies de Dirichlet, Comptea Rem&m, 162, 463-465. VI
8. (With 5. E. L.) Theorems concerning the summability of series by Borel’s exponential
method, Rendiconti del Circolo mtemtico di Palermo, 41, 36-53. w
9. (With J. E. L.) D.A.: The series C e(h) and the distribution of the points (A,&, Pro-
ceedings of the Nutionul Amdemy of Sciences, 3, 84-88. I. 1
10. Asymptotic formulae in combinatory analysis, Quat&me congrb des Muth&uticiena
Scandinaves, 45-53. I. 2 (at)
00
11. Further remarks on the integral f& dx, Muthemctticul Gazette, 8, 301-303. V
s X
0

1917
1. (With S. Ramanujan) Une formule asymptotique pour le nombre des partitions de sz,
Comptes Rendw, 164, 35-38. I. 2 (a)
2, On a theorem of Mr G. Pblya, Proceedings of the Cambridge Philosophical Society, 19,
60-63. IV. 2
3. On the convergence of certain multiple series, Proceedings of the Cambridge PhiZosophicuZ
Society, 19, W-95. VI
4. (With S. Ramanujan) Asymptotic formulae for the distribution of integers of various
types, Proceedings of the London Mathematical Society, (2) 16, 112-132. I. 2(a)
5. N.I.C. XLII: On W eierstrass’s singular integral, and on a theorem of Lerch, Memenger
of Mathematics, 46, 43-48. VII. 1
6. N.I.C. XLIII : On the asymptotic value of a definite integral, and the coefficient in a
power series, Messenger of Mathematics, 46, 70-73. V
7. N.I.C. XLIV: On certain multiple integrals and series which occur in the analytic
theory of numbers, Memenger of Mathematics, 46, 104-107. V
8. N.I.C. XLV: On a point in the theory of Fourier series, Mesaewer of Muthemtics, 46,
146-149. III. 1 (a)
9, N.I.C. XLVI: On Stieltjes’ ‘probl&me des moments’, Messenger of Muthematics, 46, 17!%
182. VII. 1

690
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1917 (cmt*)
10. (With J, E, L.) Sur la convergence dea series de Fourier et des s&k de Taylor, Comptes
Rendus, 165, 1047-1049. n-r. 1 (a)
Il. Mr. S. Ramanujan’s mathematical work in England, Journal of the Indim Mathematical
Society, 9, 30-45. VII. 2

1918
1, (With J. E. L.) Contributions to the theory of the Riemann zeta-function and the theory
of the distribution of primes, Acta Mathemutica, 41, 119-196. IX. 1
2. (With S. Ramanujan) On the coefficients in the expansions of certain modular func-
tions, Proceedings of the Royal Society, (A) 95, 144-155. I. 2(a)
3. Sir George Stokes and the concept of uniform convergence, Proceeding8 of the Cambridge
Philosophical Society, 19, 148-156. VII. 2
4. (With J. E. L.) On the Fourier series of a bounded function, Proceedings of the London
Muthemmtical Society, (2) 17, xiii-xv. III. 1 (b)
5. (With S. Ramanujan) Asymptotic formulae in combinatory analysis, Proceeding8 of the
London MaMekaticul Society, (2) 17, 75-115. I. 2 (a)
6. N.I.C. XLVII: On Stieltjes’ ‘probl&me des moments’ (cont.), Messenger of Muthematica,
47, 81-88. VII. 1
7.N.I.C. XLVIII: On some properties of integrals of fractional order, Messenger of M&e-
matics, 47, 145-150. V
8, N.I.C. XLIX: On Mellin’s inversion formula, Messenger of Mathematics, 47, 178-184. VII. 1
9. Note on an expression of Lambert’s series as a definite htegral, Messenger of Mathe-
matics, 47, 190-192. Iv. 1 (d)
10. On the representation of a number as the sum of any number of squares, and in par-
ticular of five or seven, Proceedings of the Nution& Academy of Sciences, 4,189-193. I. 2 (a)

1919
1, (With J. E. L.) Note on Messrs. Shah and Wilson’s paper entitled: ‘On an empirical
formula connected with Goldbach’s theorem’, Proceedings of the Cambridge Philo-
sophical Society, 19, 245-254. I. 2 (c)
2, N.I.C. L. On the integral of Stieltjes and the formula for integration by parts, Messenger
of Mathematics, 48, W--100. V
3. N.I.C. LI : On Hilbert’s double-series theorem, and some connected theorems concerning
the convergence of infinite series and integrals, Memenger of Mathemutica, 48, 107-
112. XI. 3
4. A problem of Diophantine approximation, Joumul of the Indian Mathematical Society,
11, 162-166. T_ 1

1920
1. (With S. Ramanujan) The normal number of prime factors of a number n, Quarterly
Jownul of Mathematics, 48, 76-92. II. I
2. (With J. E. L.) A new solution of Waring’s problem, Quarterly JoumuZ of Mathematics,
48, 272-293. I. 2 (b)
3. Note on a theorem of Hilbert, Mathemtische Zeitschrijt, 6, 314-317. rx. 3
4. On two theorems of F. Carlson and S. Wigert, Acta Mathematics, 42, 327-339. IV. 2
5. (With J. E. L.) P.N. I: A new solution of Waring’s problem, Qiittinger Nachrichten
(1920), 33-54. I. 2 (b)
6, Additional note on two problems in the analytic theory of numbers, Proceeding8 of the
London Mathematical Society, (2) 18, 201-204. II. 2
-7. -(with J. E. L.) Abel’s theorem and its converse, Proceedinga of the London Mathe-
matical: Society, (2) 18, 205-235. VI

691
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

’ 1920 (cont.)
8. N.I.C. LII: On some definite integrals considered by Mellin, AZeassnger of Muthematics,
49, 85-91. VII. 1
9, N.1.C LPII: On certain criteria for the convergence of the l?ourier series of a continuous
function, Messenger of Mathematics, 49, 149-155. IIf. 1 (a)
10. On the representation of a number as the sum of any number of squares, and in particular
of five, Transactions of the American Mathematical Society, 21, 255-284. I. 2 (a)
If. Some famous problenzs of the theory of numbers and in particular Waring’s problem,
Inaugural lecture, Oxford, 1920. I. 2 (d)

1921
1. (With J. E. I,.) P.N. II: Proof that every large number is the sum of at most 21 bi-
quadrates, Mathematische Zeitschrift, 9, 14-27. I. 2(b)
2. (With J. E. L.) The zeros of Riemann’s zeta-function on the critical line, Muthematiache
Zeitschrift, 10, 283-317. II, 1
3. Note on Ramanujan’s trigonometrical function c&n), and certain series of arithmetical
functions, Proceedings of the Cambridge Philosophical Society, 20, 263-27 1. II. 2
4, A theorem concerning summable series, Proceedings of the Cambridge Philosophi&
Society, 20, 304-307. VI
5. A convergence theorem, Proceedings of the London Mathematical Society, (2) 19, vi-vii. 11. 3
6. (With J. E. IL.) On a Tauberian theorem for Lambert’s series, and some fundamental
I theorems in the analytic theory of numbers, Proceedings of the London Muthematicul
Society, (2) 19, 21-29. II. 1
7. N.I.C. LIV : Further notes on Mellin’s inversion formulae, Mesenger of Muthematics, 50,
16&171. VII. 1

1922
1. Goldbach’s Theorem, Matematisk Tidsskrift B, 1-16. I. 2 (c)
2. A new proof of the functional equation for the zeta-function, Matem&& Tidssbift B,
7 1-73. II. 1
3. (With J. E. L.) P.N. III: On the expression of a number as a sum of primes, Acta
Muthematica, 44, l-70. I. 2 (c)
4, (With J. E. L.) P.N. IV: The singular series in Waring’s problem and the value of the
number G(k), Mathematische Zeitschrift, 12, 161-188. I* 2 (b)
5; (With J. E. L.) D.A. : A further note on the trigonometrical series associated with the
elliptic theta-functions, Proceedings of the Cambridge PhibsophicaZ Society, 21, 1-5. I, 1
6. (With J. E. L.) D.A.: The lattice-points of a right-angled triangle, Proceedings of the
London Mathematical Society, (2) 20, 15-36, I. 1
7. (With T. Carleman) Fourier’s series and analytic functions, Proceedings of the Royal
Society, (A), 101, 124-133. IV. 2
8. (With J. E. L.) Summation of a certain multiple series, Proceedings of the London Mathe-
matical Society, (2) 20, xxx. I. 2 (c)
9. (With J. E, L.) D.A.: The lattice-points of a right-angled triangle, Hamburg Abhand-
lungen, 1, 212-249. I. 1
10. N.LC. LV: On the integration of Fourier series, Messenger of Mathemcttics, 51, 186-
192. III. 1 (e)
11. The theory of numbers, British Association Report, 90, 16-24. VII. 2

1923
1. (With J. E. I;.) On Limlelaf’s hypothesis concerning the Riemann zeta-function, Pro-
ceedings of the Royal Society, (A) 103, 403-412. II. 1

692
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1923 (cont.)-
2. A chapter frc)m Ramanujan’s notebook, Proceedings of the Cambridge PhilosophicaL
Society, 2 1, 492-503. IV. 1 (d)
3. (With 3. E. L.) D.A.: The analytic character of the sum of a Dirichlet’s series considered
by Hecke, Hamburg Abhandlungen 3, 57-68. I, 1
4. (With J. E. L.) D.A. : The analytic properties of certain Dirichlet’s series associated
with the distribution of numbers to modulus unity, Transactions of the Cambridge
Philosophical Society, 22, 519-533. I. 1
5. (With J. E. I;.) The approximate functional equation in the theory of the zeta-function
with applications to the divisor-problems of Dirichlet and Piltz, Proceedings of the
London Mathematical Society, (2) 21, 39-74. XI. 1
6, N.I.C. LVI: On Fourier’s series and Fourier’s integral, Messenger of Mathematics, 52,
49-53. III, 1 (e)

1924
1. (With J. E. L.) Solution of the Ces&ro summability problem for power-series and
Fourier series, Mathewmtische Zeitschrift, 19, 67-96. III. P (b)
2. Some formulae of Ramanujan, Proceedings of the London Mathematical Society, (2) 22,
xii-xiii. IV. 1 (d)
3. (With J. E. L.) Note on a theorem concerning Fourier series, Proceedings of the London
Mathematical Society, (2) 22, xviii-xix. III. I (b)
4. (With J. E. L.) The equivalence of certain integral means, Proceedings of the London
Mathewmtical Society, (2) 22, xl-xliii, VI
5. (With J, E. L.) The allied series of a Fourier series, Proceedkgs of the London Mathe-
matical Society, (2) 22, xliii-xlv. III. 1 (b)
6. (With J. E. L.) P.N. V: A further contribution to the study of Goldbach’s problem,
Proceedings of the London Mathematical Society (2) 22, 46-56, I. 2 (c)
7. (With J. E. L.) Abel’s theorem and its converse II, Proceedings of the London Mathe-
matical Society, (2) 22, 254-269. VI
8. N,I.C. LVII: On Fourier transforms, Messenger of Mathemcttics, 53, 135442. VII. 1
9. (With E. Landau) The lattice points of a circle, Proceedings of the Royal Society (A),
105, 244-258. II. 2

1925
1. (With 3. E. L.) P.N. VI: Further researches in Waring’s problem, M&hemab&he
Zeitschrift, 23, l-37. I. 2 (b)
2. The lattice points of a circle, Proceedings of the Royal Society (A), 107, 623-635. II. 2
3. What is geometry ? Mathematical Gazette, 12, 309-316. VII. 2
4, (With Jm E. L.) D.A. : An additional note on the trigonometrical series associated with
the elliptic theta-functions, Acta Mathemtica, 47, 189498. I. 1
5. (With J. E. L.) A theorem concerning series of positive terms, with applications to the
theory of functions, Meddelelser Kabenhavn, 7, Nr. 4, IV. 2
6. Note on a theorem of Hilbert concerning series of positive terms, Proceedings of the
London Mathematical Society, (2) 23, xlv-xlvi. II. 3
7. Some formulae in the theory of Bessel functions, Proceedings of the London Mathe-
m&ical Society, (2) 23, lxi-lxiii, IV. 1 (d)
8. (With E. C. Titchmarsh) Solutions of some integral equations considered by Bateman,
Kapteyn, Littlewood and Milne, Proceedings of the London Mathematical Society, (2)
23, 1-26, and Correction ibid. 24, xxxi-xxxiii. VII. 1
9. N.I.C. LVIII: On Hilbert transforms, Messenger of Mathematics, 54, 20-27. VII, 1
10. N.I.C. LIX: On Hilbert transforms (cont.), iMessenger of Mathematics, 54, 81-88. VII. 1
11. N,I.C, LX: An inequality between integrals, Messenger of Mathematics, 54, 150-156, 11, 3

693
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1926
1, A definite intsgral which occurs in physical optics, Proceedings of the London Mathe-
mticaE Society, (2) 24, xxx-xxxi. IV. 1 (d)
2. (With J. E. L.) Some prop&k of fractional integrals, Proceedings of the London Muthe-
matical Society, (2) 24, xxxvii-xii. III. 2
3, Note on the inversion of a repeated integral, Proceedings of the London &hthemuticd
Society, (2) 24, 1-li. V
4, (With J. E. L.) The allied series of a Fourier series, Proceedings of the London Muthe-
m&xl Society, (2) 24, 211-246. III. 1 (b)
5. (With J. E. L.) A further note m the converse of Abel’s theorem, Proceedings of the
London Mathematical Society, (2) 25, 219-236. VI
6. (With J. E. L. and G. Polya) The maximum of & certain bilinear form, Proceedings of the
London Mathematical Society, (2) 25, 265-282. II. 3
7. (With J, E, L.) Some new properties of Fourier constants, Mathematische An&en, 97,
159-209. III. 1 (c)
8, (With 5. E. L.) N.S. I: Two theorems concerning Fourier series, Journul of the Lon-
don Mathematical Society, 1, 19-25. III. 1 (a)
9. A theorem concerning harmonic functions, Journal of the London Mat7Lematical Society,
1, 130-131. Iv. 2
10. (With J. E. L.) N.S. II: The Fourier series of a positive function, Journal of the London
i’bfathematical Society, 1, 134-138. III. 1 (b)
11. (With S. ]Bochner) Notes on two theorems of Norbert Wiener, Journal of the London
Mathematical Society, I, 24&244. VII. 1
12. N.1.C. LXI: On the term by term integration of a series of Bessel functions, Messenger
of Mathematics, 55, 140-144. IV. 1 (d)
13. The case agakt the Mathematical Tripos, Mathematical Gazette, 13, 61-7 1. VII. 2

1927
1, Note on Ramanujan’s arithmetical function T(n), Proceedings of the Cambridge Philo-
sophical Society, 23, 675-680. II. 2
2. (With J. E. L.) N.S. III: On the summability of the Fourier series of a nearly con-
tinuous function, Proceedings of the Cambridge Philosophical Society, 23, 981-684. III. 1 (b)
3, (With J. E. L.) N.S. IV: On the strong summability of Fourier series, Proceeding8 of
the London Mathemcttical Society, (2) 26, 273-286. III. 1 (b)
4. (With J. E. L.) N.S. V: On Parseval’s theorem, Proceedings of the London MathematicaZ
Society, (2) 26, 287-294. III. 1 (e)
5. (With A. E. Ingham and G. Pblya) Theorems concerning mean values of analytic
functions, Proceedings of the RoyaL Society (A), 113, 542-569. IV. 2
6, (With J. E. L.) Elementary theorems concerning power series with positive coefficients
and moment con&ants of positive functions, Journal fiir Matkemutik 157, 141-
158. II. 3
7, N.I.C. LXII: A singular integral, Messenger of Muthemutics, 56, l&16. VII. 1
8, N.I.C. LXIII: Some further applications of Mellin’s inversion formula, Messenger of
Mathematics, 56, 186-192. VII. 1
9. Note on a theorem of Mertens, Journal of the London MathemuticaZ Society, ZY 70-72. II, 1
10, Note on the multiplication of series, Journal of the London Mathematical Society, 2,
169-171. VI
11, (With J. E. L.) N.S. VI: Two inequalities, Journal of the London Mathematical Society,
2, 196-201. II, 3

694
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1928
1, (With J. E. L.) A theorem in the theory of summable divergent series, Proceeding8 of
the London Mathematical Society, (2) 27, 327-348. VI
2. (With A. E. Ingham and G. Pdlya) Notes on moduli and mean values, Proceedings of
the London Mathematical Society, (2) 27, 401-409, IV. 2
3. (With J. E. L.) N.S. VII: On Young’s convergence criterion for Fourier series, Pro-
ceedings of the London Mathematical Society, (2) 28, 301-311. III. 1 (b)
4, (With J. E. L.) P.N. VIII: The number r(k) in Waring’s problem, Proceedings of the
London Mathematical Society, (2) 28, 518-542. I. 2 (b)
5. (With 3. E. L.) Some properties of fractional integrals I, Mathematische Zeitschrifi, 27,
565-606. III. 2
6. (With J. E. L.) A convergence criterion for Fourier series, Mathemutische Zeitschwft, 28,
612-634. III. 1 (a)
7. N.I.C. LXIV : Further inequalities between integrals, Messenger of Mathematics, 57,
12-16. II. 3
8. N.I.C. LXV: A discontinuous integral, Messenger of Mathematics, 57, 113-120. IV. 1 (d)
9. A theorem concerning trigonometrical series, Journal of the London NathematicaZ Society,
3, 12-13. III. 1 (d)
10. (With J. E. L.) N.S. VIII: Am inequality, Journal of the London Mathematical Society,
3, 105-l 10. II. 3 -
11. Remarks on three recent notes in the Journal, Joumul of the London Mathematical
Society, 3, 166-169. II. 3
12. A formula of Ramanujan, JournaZ of the London Mathemtical Society, 3, 238-240. XV. 1 (d)
13. (With J. E. L.) N.S. IX: On the absolute convergence of Fourier series, Journal of the
London Mathematical Society, 3, 25&253. III, 1 (a)
14. (With J. E. L.) N,S. X: Some more inequalities, Journal of the London Muthemutical
Society, 3, 294-299. rr. 3

1929
1. (With 3. E. L.) The approximate functional equations for c(8) and c3(8), Proceeding8 of
the London Mathematical Society, (2) 29, 81-97. II. 1
2, Prolegomena to a chapter on inequalities (Presidential Address), Jour& of the London
Mathematical Society, 4, 61-78, and addenda, ibid. 5, 80. II. 3
3. Remarks in addition to Dr. Widder’s note on inequalities, Journal of the London Mathe-
muticul Society, 4, 199-202. PI. 3
4. (With J. E. L.) A point in the theory of conjugate functions, Journal of the London
Mathematical Society, 4, 242-245. III. 1 (e)
5. (With E. C. Titchmarsh) Solution of an integral equation, Journ of the London Mathe-
m&d Society, 4, 300-304. VII. 1
6. An introduction to the theory of numbers, Bulletin of the American Mathematical Society,
35, 778-818. VII. 2
7. N.I.C. LXVX : The arithmetic mean of a Fourier constant, Messenger of Mathemutica, 58,
50-52. III. 1 (e)
8. N.I.C. LXVII: On the repeated integral which occurs in the theory of conjugate
functions, Messenger of Mathematics, 58, 53-58. VII. I
9. N.X.C. LXVIII: The limit of an integral mean value, Messenger of Mathematics, 58,
l&120. II. 3
10. ‘N.I.C. LXIX: 0 n asymptotic values of Fourier constants, i’kfe88enger of Mathematics,
58, 130-135. III. 1 (a)
11. (With J. E. L. and G. P6lya) Some simple inequalities Batisfied by convex functions,
Messenger of Mathematics, 58, 145-152. II. 3
12. Mathematical proof, Mind, 38, l-26. VII. 2

695
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1930
1. (With J. E. L.) A maximal theorem with function-theoretic applications, Actu Mathe-
matica, 54, U-116. II. 3
2. (With E. C. Titchmarsh) Self-reciprocal functions, Quarterly Journccl of Mathematics,
1, 196-231. VII. 1
3. (With J. E. L.) D.A.: A series of cosecants, Bulletin of the Calcutta Muthemuticul
Society, 20, 251-266. I. 1
4. (With 5. E. L.) N.S. XI: On Tauberian theorems, Proceedings of the London Muthe-
maticul Society, (2) 30, 23-37. VI
5. (With E. C. Titchmarsh) Additional note on certain integral equations, Proceedings of
the London Mathematical Society, (2) 30, 95406. VII. 1
6. (With J. E. L.) N.S. XII: On certain inequalities connected with the calculus of varia-
tions, Journul of the London Mathewtutical Society, 5, 34-39. II. 3

1931
1. Some theorems concerning trigonometrical series of a special type, Proceedings of the
London Mathematical Society, (2) 32, 441-448. III. 1 (d)
2. (With J. E. I;,) Some properties of conjugate functions, Journal fiir Muthematik, 167,
405-423. III. 2
3. The summability of a Fourier series by logarithmic means, Quarterly Journal of Muthe-
matics, 2, 107-112. III. 1 (b)
4. (With J. E. I;.) N.S. XIII: Some new properties of Fourier Constants, Journal of the
London Mathematical Society, 6, 3-9. IIL. 1 (c)
5. (With J. E. L.) N.S. XIV: An additional note on the summability of Fourier series,
Journal of the London Mathematical Society, 6, 9-12. III. 1 (b)
6. (With E. C. Titchmarsh) A note on Parseval’s theorem for Fourier transforms, Journal
of the London Mathematical Society, 6, 44-48. VII. f
7. (With J. E. I;.) N.S. XV: On the series conjugate to the Fourier series of a bounded
function, Journal of the London Mathematical Society, 6, 278-281, III. 1 (b)
8. (With J. E. L.) N.S. XVI: Two Tauberian theorems, Journal of the London Mathe-
matical Society, 6, 28 1-286. VI

1932
1. (With E. C. Titchmarsh) Formulae connecting different classes of self-reciprocal func-
tions, Proceedings ofthe London @athematicaZ Society, (2) 33, 225-232. VII. 1
2. On Hilbert transforms, QuurterZy Journal: of Muthematics, 3, 102~HZ. VII, 1
3. (With J. E. L.) Some integral inequalities connected with the calculus of variations,
Quurterly Journal Mathematics,
of 3, 241-252, II. 3
4. (With J. E. L.) S ome properties of fractional integrals II, Muthemutische Zeitschrift,
34, 403-439. III, 2
5. (With J. E. L.) S ome new cases of Parseval’s theorem, Muthematische Zeitschrift, 34,
620-633. III. I (c)
6. (With J. E. L.) An additional note on Parseval’s theorem, Mathemutische Zeitschrift,
34, 634-636. III* 1 (c)
7. (With E. C. Titchmarsh) An integral equation, Proceedings of the Cambridge PhilosophicaL
Society, 28, 165473. VII. 1
8. Summation of a series of polynomials of Laguerre, Journal of the London Muthematicul
Society, 7, 138-139, and addendum, ibid. 192. XII. 1 (e)
9. (With J. Ii]. L.) N.S. XVII: Some new convergence criteria for Fourier series, Journal
of the London Muthematicul: Societv. 7. 252-256. III. 1 la1

696
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1933
1. (IVith E. C. Titchmarsh) A class of Fourier kernels, Proceedings of the London Muthe-
mtical Society, (2) 35, 116-155. VII, 1
2. (With J. E. L.) Some more integral inequalities, The T6hoku Mathem.&cal Journal,
37, 151-159. II. 3
3. The constads of certain inequalities, Journal of the London Mathematical Society, 8,
114-119. II. 3
4. A theorem concerning Fourier transforms, JoumaZ of
the London Mathematical Society,
8, 227-231. VII. 1

1934
1. (With 3. E. L.) Theorems concerning Ces&ro means of power series, Proceedings of the
London Mathematical Society, (2) 36, 5 16-53 1. III, 2
2. (With J, E. L.) Bilinear forms bounded in apa’ce [p, q], Qua~terZy JournaZ of Mathe-
matics, 5, 241-254. 11. 3
3. (With J. E. I;.) Some new convergence criteria for Fourier series, AnnaZi Piaa, (2) 3,
43-62. III. 1 (a)
4. (With E. M. Wright) Leudesdorf’s extension of Wolstenholme’s theorem, Journal of the
London Mathematical Society, 9, 38-41, and corrigendum, ibid. 240, II. 2
J*E; On the summability of series by Borel’s and Mittag-Leffler’s methods, Journal of the
London Mathematical Society, 9, 153457. VI

1935
1. Remarks on some points in the theory of divergent series, Annuls of Mathematics, (2)
36, 167-M. VI
2. The resultant of two Fourier kernels, Proceedings of the Cambridge Philosophical Society,
31, l-6. VII. 1
3. [With E. Landau and J. E. L.) Some inequalities satisfied by the integrals or derivatives
of real or analytic functions, Mathematische Zeitschrijt, 39, 677-695. II. 3
4. (With J. E. I;.) An inequality, Mathematische Zeitschrift, 40, l-40. II. 3
5. (With J. E. I;.) The strong summability of Fourier series, Fundamenta Mathematicae,
25, 162-189. III. 1 (b)
6. (With J. E. L.) N.S. X.VIII: On the convergence of Fourier series, Proceedings of the
Cambridge Philosophical Society, 3 1, 3 17-323. III. 1 (c)
7. (With 3. E. L.) N.S. XIX: A problem concerning majorants of Fourier series, Quartedy
Journal of Mathemcttics, 6, 304-315. III. 1 (c)
8. Second note on a theorem of Mertens, Journal of the London Mathematical Society, 10,
91-94. II. 1
9. Some identities satisfied by i-nfinite series, Journal of the London Mathematicul Society,
10, 217-220. IV. 1 (d)

1936
11 (With J, E, L.) N.S. XX: On Lambert series, Proceedings of the Londorz Mathew&cd
Society, (2) 41, 257-270. VI
2. (With J. E. L.) S ome more theorems concerning Fourier series and Fourier power series,
Duke Mathematical Journal, 2, 354-382. III 1 (b)
3. (With E. C. Titchmarsh) New solution of an integral equation, Proceedings of the London
Mathematical Society, (2) 41, l-15. VII. 1
4. A note on two inequalities, Journal of the London Mathematical Society, 11, 167-170. 11. 3

697
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1937
1, Ramanujan and the theory of Fourier transforms, Quarterly Journul of Mathemtica, 8,
245-254. VII. 1
2, The Indian mathematician Ramanujan, American Mathematical Monthly, 44, 137-
155. VII. 2
3, (With J. E. L,) N.S. XXI: Generalizations of a theorem of Paley, Qumterly Journal of
Mathematics, 8, 161-171. III. 2
4, On a theorem of Paley and Wiener, Proceedings of the Cambridge Philosophicub Society,
33, 1-5. VII. 1
5, (With N. Levinson) Inequalities satisfied by a certain defmite integral, Bulletin of the
American Mathematical Society, 43, 709-7 16. II, 3
6. A formula of Ramanujan in the theory of primes, Journal of the London Mathematical
Society, 12, 94-98. II. 1
7, Another formula of Ramanujan, Journal of the London &fatIiemticaZ Society, 12, 314-
318. IV. 1 (d)

1938
1. A further note on Ramanujan’s arithmetical function T(n), Proceedings of the Cambridge
Philo8ophicuZ Society, 34, 309-3 15. II. 2

1939
1, A note on a differential equation, Proceeding8 of the Cambridge Philomphical Sot iety, 35,
652-653. VII. 2
2, Notes on special systems of orthogonal functions (I) : The boundedness of the generalized
Laguerre system, Journal of the London Mathematical Society, 14, 34-36. IV. 1 (d)
3. Notes on special systems of orthogonal functions (II): On functions orthogonal with
respect to their own zeros, Journal of the London Mathematical Society, 14,37-44. III. 1 (e)

1940
1, Notes on special sysbms of orthogonal functions (III) : A sysfem of orthogona1 poly-
nomials, Proceedings of the Cambridge PIiilosophimZ Society, 36, 1-8. A? 1 (d)

1941
1. (With J. E. I;.) Theorems concerning mean values of analytic or harmonic functions,
Quarterly Journal of Mathemutica, 12, 221-256. III. 2
2. Note on a divergent series, Proceeding8 of the Cambridge Philosophical Society, 37, l-8, VI
3. Notes on special systems of orthogonal functions (IV) : The orthogonal functions of
Whittaker’s cardinal series, Proceedings of the Cambridge Philosophical Society, 37,
331-348. III. 1 (d)
4. A double integral, Jourml of the London Mathematical Society, 16, 89-94, VII. 1

1942
1. Note on Lebesme’s constants in the theory of Fourier series, Joulmcal of the London
Mathmm&-cuZ Society, 17, 4-13, III. 1 (a)

1943
1, An inequality for Hausdorff means, Jou~l of the Lo&m Mathematical Society, 18,
46-50. II. 3

698
 COMPLETE LIST OF HARDY’S MATHEMATICAL PAPERS

1943 (cont.)
2. (With W. W. Rogosimki) Notes on Fourier series (I): On sine series with positive co-
efficients, Journal of the London Mathematical Society, 18, 50-57. III. 1 (d)
3. (With W. W, Rogosinski) Notes on Fourier series (II): On the Gibbs phenomenon,
Journal of the London Mathemtical Society, 18, 83-87. III. 1 (a)
4. (With J. E. L.) KS. XXII: On the Tauberian theorem for Bore1 summability, Jourruzl
of the London Mathemuticctl Society, 18, 194-200. VI

1944
1, (With J. E. L.) N.S. XXIII: On the partial sums of Fourier series, proceedinga oj the
Cambridge Philosophica Society, 40, 103-107. III. 1 (c)
2. Note on the multiplication of series by Cauchy’s rule, Proceedings of the Cambridge
Philosophical Society, 40, 251-252. VI

1945
1. (With W. W. Rogosinski) Notes on Fourier series (III): Asymptotic formulae for the
sums of certain trigonometrical series, Quarterly Journal of Mathematics, 16, 49-
58. III. 1 (d)
2, (With N. Aronszajn) Properties of a class of double integrals, Annals of Mathematics,
(2) 46, 220-241, and corrigendum in 47, 166. V
3, Riemann’s form of Taylor’s series, JournaE of the London Mathematical Society, 20,
48-57. VI
4. A mathematical theorem about golf, Mathemtical Gazette, 29, 226-227. VII. 2

1946
1, (With J. E. L.) N.S. XXIV: A curious power-series, Proceedings of the Cambridge
Philosophicat Society, 42, G-90. I. 1
2, (With W, W. Rogosinski) Theorems concerning functions subharmonic in a strip,
Proceedings of the Royal Society (A), 185, l-14. Iv. 2

1947
1. (With W. W. Rogosinski) Notes on Fourier series (IV): Summability (R,), pro-
ceeding8 of the Cambridge Philosophical Society, 43, 10-25. III. 1 (b)
2, A double integral, Joumaul of the London Mathematical Society, 22, 242-247. VII. 1

1948
1, (With J. E. L.) A new proof of a theorem on rearrangements, Journal of the London
Mathematical Society, 23, 163-168. II. 3

1949
1, (With W. W. Rogosinski) Notes on Fourier series (V): Summability (R,), Proceedings
of the CamKdge Philosophical Sot III. 1 @I
\ ,H’) .
/
., ? 4’