J. w. S.

Cassels (known to his friends by the Gaelic

form "Ian" of his first name) was born of mixed
English-Scottish parentage on 11 July 1922 in the
picturesque cathedral city of Durham. With a first
degree from Edinburgh, he commenced research
in Cambridge in 1946 under L. J. Mordell, who had
just succeeded G. H. Hardy in the Sadleirian Chair
of Pure Mathematics. He obtained his doctorate
and was elected a Fellow of Trinity College in 1949.
After a year in Manchester, he returned to Cam-
bridge and in 1967 became Sadleirian Professor.
He was Head of the Department of Pure Mathe-
matics and Mathematical Statistics from 1969
until he retired in 1984.

Cassels has contributed to several areas of number

theory and written a number;()f other expository
books:
• An introduction to diophantine approximations
• Rational quadratic forms
• Economics for mathematicians
• Local fields
• Lectures on elliptic curves
• Prolegomena to a middlebrow arithmetic of
cur.J!es of genus 2 (with E. V. Flynn).
Classics in Mathematics
J.W.5. Cassels An Introduction to the Geometry of Numbers
Springer
Berlin
Heidelberg
New York
Barcelona
Budapest
Hong Kong
London
Milan
Paris
Santa Clara
Singapore
Tokyo
 J.W.s. Cassels

An Introduction
to the Geometry
of Numbers

Reprint of the 1971 Edition

Springer
J.W.S. Cassels
University of Cambridge
Department of Pure Mathematics
and Mathematical Statistics
16, Mill Lane
CB2 ISB Cambridge
United Kingdom

Originally published as Vol. 99 of the

Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen

Mathematics Subject Classification (1991): lOExx

ClP data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme

Cassels, John W.S.:

An introduction to the geometry of numbers I J.W.S. Cassels - Reprint of the 1971 ed. - Berlin; Heidelberg;
New York; Barcelona; Budapest; Hong Kong; London; Milan; Paris; Santa Clara; Singapore; Tokyo:
Springer. 1997
(Classics in mathematics)
ISBN-13: 978-3-540-61788-4 e-ISBN-13: 978-3-642-62035-5
DOl: 10.1007/978-3-642-62035-5

This work is subject to copyright. All rights are reserved. whether the whole or part of the material is
concerned. specifically the rights of translation. reprinting. reuse of illustrations. recitation. broadcasting.
reproduction on microfilm or in any other way. and storage in data banks. Duplication of this publication or
parts thereof is permitted only under the provisions of the German Copyright Law of September 9.1965. in
its current version. and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.

© Springer-Verlag Berlin Heidelberg 1997

The use of general descriptive names. registered names. trademarks etc. in this publication does not imply,
even in the absence of a specific statement. that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.

SPIN 10554506 41/3143-543210- Printed on acid-free paper

 J. W. S. Cassels

An Introduction to the
Geometry of Numbers

Second Printing, Corrected

Springer-Verlag Berlin· Heidelberg· New York 1971

 Prof. Dr. J. W. S. Cassels
Professor of Mathematics. University of Cambridge, G. B.

Geschiiftsfuhrende Herausgeber:

Prof. Dr. B. Eckmann

Eidgeniissische Technische Hochschule Zurich

Prof. Dr. B. L. van der Waerden

Mathematisches Institut der Universitat Zurich

AMS Subject Classifications (1970): 10 E xx

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting, reproduction
by photocopying machine or similar means, and storage in data banks.
Under § 54 of the German Copyright Law where copies arc made for other than private use, a fee is payable
to the publisher, the amount of the fee to be determined by agreement with the publisher.

@ by Springer-Verlall Berlin· Heidelberg 1959, 1971. Library of Congress Catalog Card Number
75-154801.
 Preface
Of making many bookes there is no
end, and much studie is a weari-
nesse of the flesh.
Ecclesiastes XII, 12.

When I first took an interest in the Geometry of Numbers, I was

struck by the absence of any book which gave the essential skeleton
of the subject as it was known to the experienced workers in the subject.
Since then the subject has developed, as will be clear from the dates
of the papers cited in the bibliography, but the need for a book remains_
This is an attempt to fill the gap. It aspires to acquaint the reader with
the main lines of development, so that he may with ease and pleasure
follow up the things which interest him in the periodical literature.
I have attempted to make the account as self-contained as possible.
References are usually given to the more recent papers dealing with
a particular topic, or to those with a good bibliography. They are given
only to enable the reader to amplify the account in the text and are
not intended to give a historical picture. To give anything like a reason-
able account of the history of the subject would have involved much
additional research.
lowe a particular debt of gratitude to Professor L. J. MORDELL,
who first introduced me to the Geometry of Numbers.
The proof-sheets have been read by Professors K. MAHLER, L.J.
MORDELL and C. A. ROGERS. It is a pleasure to acknowledge their
valuable help and advice both in detecting errors and obscurities and
in suggesting improvements. Dr. V. ENNOLA has drawn my attention to
several slips which survived info the second proofs.
I should also like to take the opportunity to thank Professor F. K.
SCHMIDT and the Springer-Verlag for accepting this book for their
celebrated yellow series and the Springer-Verlag for its readiness to
meet my typographical whims.
Cambridge, June, 1959 J. W. S. CASSELS
 Contents Page
Notation VIII
Prologue
Chapter I. Lattices. 9
1. Introduction 9
2. Bases and sublattices 9
3. Lattices under linear transformation 19
4. Forms and lattices. 20
5. The polar lattice. 23
Chapter II. Reduction . 26
1. Introduction . . 26
2. The basic process 27
3. Definite quadratic forms 30
4. Indefinite quadratic forms 35
5. Binary cubic forms 51
6. Other forms. . . . . . . 60
Chapter III. Theorems of BLlCHFELDT and MINKOWSKI 64
1. Introduction . . . . . . . . . . . . . 64
2. BLlCHFELDT'S and MINKOWSKI'S theorems 68
3. Generalisations to non-negative functions. 73
4. Characterisation of lattices 78
5. Lattice constants . . . . . . . . . . . 80
6. A method of MORDELL . . . . . . . . . 84
7. Representation of integers by quadratic forms. 98
Chapter IV. Distance functions 103
1. Introduction . . . . . . 103
2. General distance-functions 105
3. Con vex sets. . . . . . . loR
4. Distance functions and lattices 119
Chapter V. MAHLER'S compactness theorem 121
1. Introduction . . . . . 121
2. Linear transformations . 122
3. Convergence of lattices. 126
4. Compactness for lattices 134
5. Critical lattices . . 141
6. Bounded star-bodies 145
7. Reducibility 152
8. Convex bodies. . . 155
9. Spheres . . . . . 163
10. Applications to diophantine approximation 165
Chapter VI. The theorem of MINKOWSKI-HLAWKA 175
1. Introduction . . . . . . 175
2. Sublattices of prime index . . . . . . . 178
 Contents VII
Page
3. The Minkowski-Hlawka theorem. 181
4. SCHMIDT'S theorems . . 184
5. A conjecture of ROGERS . 187
6. Unbounded star-bodies. . 189
Chapter VII. The quotient space. 194
1. Introduction . . . 194
2. General properties. . . . 194
3. The sum theorem . . . . 198
Chapter VIII. Successive minima 201
1. Introduction . . . . . . 201
2. Spheres . . . . . . . . 205
3. General distance-functions 207
4. Convex sets. . . . 213
5. Polar convex bodies 219
Chapter IX. Packings 223
1. Introduction . . . 223
2. Sets with V(.9') = 2" ,1(.9') 228
3. VORONOI'S results . . 231
4. Preparatory lemmas. 235
5. FEJES T6TH'S theorem 240
6. Cylinders. . . . . . 245
7. Packing of spheres. . 246
8. The product of n linear forms . 250
Chapter X. Automorphs 256
1. Introduction . . . . 256
2. Special forms . . . . 266
3. A method of MORDELL 268
4. Existence of automorphs 279
5. Isolation theorems. . . 286
6. Applications of isolation 295
7. An infinity of solutions. 298
8. Local methods 301
Chapter XI. Inhomogeneous problems 303
1. Introduction . . . . . . . . 303
2. Convex sets. . . . . . . . . 309
3. Transference theorems for convex sets 313
4. The product of n linear forms 322
Appendix. 332
References 334
Index . . . . . . . . . . . . . . • . . . . . 343
 Notation
An effort has been made to distinguish different types of mathemati-
cal object by the use of different alphabets. It is not necessary to
describe the scheme in full since an acquaintance with it is not pre-
supposed. However the following conventions are made throughout the
book without explicit mention.
Bold Latin letters (large and small) always denote vectors. The
dimensions is n, unless the contrary is explicitly stated: and the letter n
is not used otherwise, except in one or two places where there can be
no fear of ambiguity. The co-ordinates of a vector are denoted by the
corresponding italic letter with a suffix 1, 2, ... ,n. If the bold letter
denoting the vector already has a suffix, then that is put after the
co-ordinate suffix. Thus:
a = (llt, ... , a,,)
b, = (bI " ... , b",)
x: = (X~., ... , X~.).
The origin is always denoted by o. The length of re is
lrel = (xl + ... + x!)i.
Sanserif Greek capitals, in particular A, M, N, r, denote lattices.
The notation d (A), L1 (9'), V( 9') for respectively the determinant
of the lattice A and for the lattice-constant and volume of a set 9'
will be standard, once the corresponding concepts have been introduced.
Chapters are divided into sections with titles. These sections are
subdivided, for convenience, into subsections, which are indicated by a
decimal notation. The numbering of displayed formulae starts afresh
in each subsection. The prologue is just subdivided into sections without
titles, and it was convenient to number the displayed formulae con-
secutively throughout.
 Prologue
P1. We owe to MINKOWSKI the fertile observation that certain
results which can be made almost intuitive by the consideration of
figures in n-dimensional euclidean space have far-reaching consequences
in diverse branches of number theory. For example, he simplified the
theory of units in algebraic number fields and both simplified and
extended the theory of the approximation of irrational numbers by
rational ones (Diophantine Approximation). This new branch of
number theory, which MINKOWSKI christened "The Geometry of Num-
bers", has developed into an independent branch of number-theory
which, indeed, has many applications elsewhere but which is well worth
studying for its own sake.
In this prologue we first discuss some of the concepts and results
which will playa leading role. The arguments we shall use are some-
times rather different from those in the main body of the text: since
here we wish to make the geometrical situation intuitive in simple cases
without necessarily giving complete proofs, while later we may need to
sacrifice picturesqueness for precision. The proofs in the text are inde-
pendent of this prologue, which may be omitted if desired.
P2. A fundamental and typical problem in the geometry of numbers
is as follows:
Let l(x1 , ..• , x,,) be a real-valued function of the real variables
Xl' ... , x". How small can I/(ul , ... ,un)1 be made by suitable choice

of the integers u1 , ... , U" ? It may well be that one has trivially
1(0, ... ,0) =0, for example when l(x1 , ••• , xn) is a homogeneous form;
and then one excludes the set of values U 1 = U 2 = ... = Un = 0. (The
"homogeneous problem".)
In general one requires estimates which are valid not merely for
individual functions I but for whole classes of functions. Thus a typical
result is that if
(1 )
is a positive definite quadratic form, then there are integers u1 , u 2 not
°
both such that
(2)
where

Cassels, Geometry of Numbers

2 Prologue

is the discriminant of the form. It is trivial that if the result is true

then it is the best possible of its kind, since

u~ + UI U 2 + u~ ~ 1
for all pairs of integers uI , u2 not both zero; and here D = !.
Of course the positive definite binary quadratic forms are a par-
ticularly simple case. The result above was known well before the birth
of the Geometry of Numbers; and indeed we shall give a proof sub-
stantially independent of the Geometry of Numbers in Chapter II, § 3.
But positive definite binary quadratic forms display a number of argu-
ments in a particularly simple way so we shall continue to use them as
examples.
P3. The result just stated could be represented graphically. An
inequality of the type
t(X I , x 2) ~ k,
where t(xl , x2) is given by (1) and k is some positive number, represents
the region fA bounded by an ellipse in the (Xl' x 2)-plane. Thus our
result above states that fA contains a point (ul , u 2 ), other than the
origin, with integer coordinates provided that k~ (4D/W.
A result of this kind but not so precise follows at once from a
fundamental theorem of MINKOWSKI. The 2-dimensional case of this
states that a region fA always contains a point (ul , u 2) with integral
co-ordinates other than the origin provided that it satisfies the following
three conditions.
(i) fA is symmetric about the origin, that is if (Xl' X 2) is in fA then so
is (- Xl' - x 2 ).
(ii) fA is convex, that is if (Xl' x 2) and (YI' Y2) are two points of fA
then the whole line segment

joining them is also in fA.

(iii) fA has area greater than 4.
Any ellipse t(x l , x2)~k satisfies (i) and (ii). Since its area is
kn kn
2)'
(a ll a n -al2 -Df'
it also satisfies (iii), provided that kn> 4Df. We thus have a result
similar to (2), except that the constant (t), is replaced by any number
greater than 4/n.
p 4. It is useful to consider briefly the basic ideas behind the proof
of MINKOWSKI'S theorem, since in the formal proofs in Chapter 3 they
 Prologue 3

may be obscured by the need to obtain powerful theorems which are

as widely applicable as possible. Instead of the region JI, MINKOWSKI
works with the region .'7= i91' of points (ix v ixz), where (Xl' X z) is
in 3£. Thus.'7 is symmetric about the origin and convex: its area is i
that of 91' and so is greater than 1. More generally, MINKOWSKI considers
the set of bodies .'7 (u l , u z) similar and similarly situated to .'7 but
with centres at the points (u l , u 2) with integer co-ordinates.
We note first that if .'7 and .'7(u I , U 2} overlap then I (u l , u 2) is in 91'.
For let a point of overlap be (~l' ~z)' Since (~l' ~2) is in .'7(ul , u 2)

GGGO
the point (~I-UI' ~2-U2) must
be in .'7. Hence, by the symmetry
of .'7, the point (UI-~I' U2-~2) IS

GGGO
in .'7. Finally, the mid-point of
(UI-~I' U2-~2) and (~I' ~2) is m
Y) because of convexity, that is

G GGG
(iu l , iu 2 ) is in .'7, and (ul , u 2 )
is in [Jl, as required. I t is clear
that .'7(ul , u 2) overlaps .'7(u~, u~)
when and only when .'7 overlaps

GGGG
Y'(u l - u~, u 2 - u~).
To prove MINKOWSKI'S theorein,
it is thus enough to show that when
the .'7 (ul , u 2) do not overlap then Fig. 1
the area of each is at most 1. A
little reflection convinces one that this must be so. A formal proof
is given in Chapter 3. Another argument, which is perhaps more intuitive
is as follows, where we suppose that .'7 is entirely contained in a square
hl:;:;:X, Ix 2 1:;:;:X.
Let U be a large integer. There are (2U +1)2 regions .'7(ul , u 2 ) whose
centres (ul , u 2) satisfy

These .'7 (uI , u 2) are all entirely contained in the square

IxII:;:;: U +X, Ix 2 1:;:;: U +X
of area
4(U +X)2.
Since the .'7(uI , u 2) are supposed not to overlap, we have
(2U + 1)2 V:;:;: 4(U +X)2,
I The converse statement is trivially true. If (u l • u 2 ) is in JI then (lUI' tu 2)
is in both [/' and [/'(u l • u 2 ).
1*
4 Prologue

where V is the area of !/'; and so of each !/' (u I , u 2 ). On letting U tend

to infinity we have V ~ 1, as required.
P 5. A change in the co-ordinate system in our example of a definite
binary quadratic form t(x l , x2 ) leads to another point of view. We may
represent t (Xl' X 2) as the sum of the squares of two linear forms:

0)
where
Xl = IX Xl +Px2 , X 2 =rXI + bX2 (4)
and IX, p, }', b are constants, e.g. by putting

IX = aL, P= a11;aI2 ,
r = 0, b = all~ D~ .
Conversely if IX, p, r, b are any
real numbers with IXb-Pr=t=o and
{ Xl' X 2 are given by (4), then
a-ft,}'-d)
xi+ X~=all x~+ 2a l2 Xl X 2 + a22 x~,
with

(5)

is a positive definite quadratic form

with
Fig. 2
D = all a22 - ai 2 = (IX b - Pr)2. (6)
We now consider Xl' X 2 as a system of rectangular cartesian co-
ordinates. The points Xl' X 2 corresponding to integers Xl' x 2 in (4) are
then said to form a (2-dimensional) lattice I\. In vector notation /\ is
the set of points
(7)

where uI , u 2 run through all integer values.

We must now examine the properties of lattices more closely. Since
we consider /\ merely as a set of points, it can be expressed in terms
of more than one basis. For example
(IX - P, r - b), (- P, - b)
is another basis for I\. A fixed basis (IX, P), (r, b) for /\ determines
a subdivision of the plane by two families of equidistant parallel
lines, the first family consisting of those points (Xl' X 2) which can be
 Prologue 5

expressed in the form (7) with tt z integral and ttl only real, while for
the lines of the second family the roles of UI and tt z are interchanged.
In this way the plane is subdivided into parallelograms whose vertices
are just the points of I\. Of course the subdivision into parallelograms
depends on the choice of basis, but we show that the area of each
parallelogram, namely
jocb-Prj,
is independent of the particular basis. We can do this by showing that
the number N(X) of points of A in a large square

satisfies
.V(X) 1
---*--- (X -* (0).
4X2 ICIa-Prl
Indeed a consideration along the lines of the proof of MINKOWSKI'S
convex body theorem sketched above shows that the number of points
of A in !2 (X) is roughly equal to the number of parallelograms contained
in !2 (X), which again is roughly equal to the area of !2 (X) divided by
the area joc b - Pr j of an individual parallelogram. The strictly positive
number
d(A) = jocb - Prj (8)

is called the determinant of A. As we have seen, it is independent of

the choice of basis.

P 6. In terms of the new concepts we see that the statement that

there is always an integer solution of t(xl , X2)~ (4D/3)i is equivalent
to the statement that every lattice A has a point, other than the origin, in

(9)

On grounds of homogeneity this is again equivalent to the statement

that the open circular disc
£1): Xi+X~<1 (10)

contains a point of every lattice A with d (A) < (l)i, and the fact that
there are forms such that equality is necessary in (2) is equivalent to
the existence of a lattice Ac with determinant d (AJ = (1)1 having no
point in £1). So our problem about all definite binary quadratic forms
is equivalent to one about the single region £1) and all lattices. Similarly
consideration of the lattices with points in
6 Prologue

gives us information about the minima of indefinite binary quadratic

forms:
ip.f
"h"t mtegers
II (UI , Ua)i :
not both 0
and so on.
These considerations prompt the following definitions. A lattice A
is said to be admissible for a region (point-set) ffi in the (Xl' Xa)-plane
if it contains no point of ffi other than perhaps the origin, if that is a
point of f!4. We may say then that A is ffi-admissible. The lower bound
L1 (ffi) of d (A) over all ffi-admissible lattices is the lattice-constant of ffi:
if there are no ffi-admissible lattices we put L1 (R) = 00. Then any
lattice A with d (A) < L1 (R) certainly contains a point of ffi other than
the origin. An ffi-admissible lattice A with d (A) = L1 (ffi) is called critical
(for ffi): of course critical lattices need not exist in general.
The importance of critical lattices was already recognized by
MINKOWSKI. If Ac is critical for ffi and A is obtained from Ac by a
slight distortion (i.e. by making small changes in a pair of base-points)
then either A has a point in ffi other than the origin or d (A) ~ d (Acl
(or both).
As an example, let us again consider the open circular disc

Suppose that Ac is a critical lattice for~. We outline a proof that a

critical lattice, if it exists, must have three pairs of points ± (AI' A 2 ).
±(BI , Ba), ±(CI , Ca) on the boundary X~+X~=1 of~. For if At
had no points on X~ +X~ = 1, we could obtain an .!'}-admissible lattice
with smaller determinant from Ac by shrinking it about the origin, that
is by considering the lattice A =tAc of points (tXI' tXa), where
(Xl' X 2 ) E A and O<t<1 is fixed. Then d(A) =t2 d(A,J <d(Acl, and
clearly A would be also .!'}-admissible if t is near enough to 1. Hence Ac
contains a pair of points on X~ + X~ = 1, which, after a suitable rotation
of the co-ordinate system, we may suppose to be ±(1, 0). If there were
no further points of Ac on X~ + X~ = 1 then we could obtain a £il-
admissible lattice A of smaller determinant by shrinking Ac perpendicular
to the Xcaxis, that is by taking A to be the lattice of (Xl' tX2 ),
(Xl' X 2 ) E A" where t is near enough to 1. Finally, if Ac had only two
pairs of points ±(1,O), ±(BI , B 2 ) on the boundary, then it is not
difficult to see that it could be slightly distorted so that (1, 0) remains
fixed but (BI' B 2) moves along X~ +X~ = 1 nearer to the Xl-axis, cf.
Fig. 3.
This can be verified to decrease the determinant of the lattice
[indeed (1,0) and (BI' B,.) can be shown to be a basis for Ac], and for
 Prologue 7

small distortions the distorted lattice A will still be 2}-admissible.

Hence a critical lattice At (if it exists) must have three pairs of points
+
on X~ X~ = 1 : and it is easy to verify that the only lattice with three
pairs of points on xi + X~ = 1, one of them being ± (1, 0), is the lattice
N with basis
(1,0), (t, VI)·
This has the vertices of
a regular hexagon
±(1,O),
± (t, VI),
( I,P)

±(-t, VI)
on X~ +X~ = 1, but no
points in X~ +X~< 1.
We have thus shown

"<
thaL1(D) =d(N) =(-W
provided that g has
a critical lattice. MIN- Fig. 3
KOWSKI showed that
critical lattices exist for a fairly wide set of regions fJi by, roughly speak-
ing, showing that any fJi-admissible lattice A can be gradually shrunk
and distorted until it becomes critical. In the text we give a more
general proof of the existence of critical lattices using concepts due to
MAHLER which turn out to have much wider significance.

P 7. Another general type of problem is the typical "inhomogeneous

problem": Let t (Xl' ... , XII) be some real-valued function of the real
variables Xl' ... , X n • It is required to find a constant k with the following
property: If $1' ... , $n are any real numbers there are integers UI , ... , Un
such that

Questions of this sort turn up naturally, for example in the theory

of algebraic numbers. Again there is a simple geometric picture. For
simplicity let n = 2. Let fJi be the set of points (Xl' X 2) in the 2-dimen-
sional euclidean plane with
I/(xl , x 2)i ~ k.
Denote by 9l (u I , U 2), where U I , U 2 are any integers, the region similar
to fJi but with the displacement U I , u 2 ; that is 9l (ul , u 2) is the set of
points Xl' x 2 such that
8 Prologue

Then the inhomogeneous problem is clearly to choose k so that the

regions ~(14l' 142) cover the whole plane. In general one will wish to
choose k, and so fJI, as small as possible so that it still has this covering
property. Here we have a contrast with the treatment of the homo-
geneous problem in § 4, where the objective was to make the regions
[there denoted by 9' (14, v)] as large as possible but so that they did not
overlap.
In this book we shall mainly be concerned at first with the homo-
geneous problem. Only when we have a fairly complete theory of the
homogeneous problem will we discuss in Chapter XI the inhomogeneous
problem and its relation to the homogeneous one.
 Chapter I

Lattices
1.1. Introduction. In this chapter we introduce the most important
concept in the geometry of numbers, that of a lattice, and develop some
of its basic properties. The contents of this chapter, except § 2.4 and
§ 5, are fundamental for almost everything that follows.
In this book we shall be concerned only with lattices over the ring
of rational integers. A certain amount of work has been done on
lattices over complex quadratic fields, see e.g. MULLENDER (1945 a) and
K. ROGERS (1955 a). Many of the concepts should carryover practically
unaltered. Again, work on approximation to complex numbers by
integers of a complex quadratic field [e.g. MULLENDER (1945a), CASSELS,
LEDERMANN and MAHLER (1951 a), POITOU (1953 a)] and on the minima
of hermitian forms when the variables are integers in a quadratic field
[e.g. OPPENHEIM (1932a, 1936a, 1953f) and K. ROGERS (1956a)] may
be regarded as a generalization of the geometry of numbers to lattices
over complex quadratic fields. We shall not have occasion to mention
lattices over complex quadratic fields again in this book; we mention
them here only for completeness. For lattices over general algebraic
number fields see ROGERS and SWINNERTON-DYER (1958a).
1.2. Bases and sublattices. Let aI' ... , a" be linearly independent
real vectors in n-dimensional real euclidean space, so that the only set of
numbers t1, ... , tn for which tl ~ +... +
tn an = 0 is tl = t2 = ... = tn = O.
The set of all points
(1 )

with integral u 1 , .•• , Un is called the lattice with basis aI' ••• , a". We
note that, since aI' ... ,an are linearly independent, the expression of
any vector ~ in the shape (1) with real u1 , ••• , Un is unique. Hence if ~
is in A and (1) is any expression for ~ with real U 1 , ... , U II , then U 1 , ... , U II
are integers. We shall make use of these remarks frequently, often
without explicit reference.
The basis is not uniquely determined by the lattice. For let a~ be
the points
L;
a; = v;;a, (1~i, i~n), (2)

where Vii are any integers with

det(v;j) = ± 1. (3)
10 Lattices

Then
aj = L: wjiaj (4)
i
with integral Wi;' It follows easily that the set of points (1) is precisely
the set of points

where U~, ... , u~ run through all integers; that is ai' ... , all and a~, ... , a~
are bases of the same lattice. We show now that every basis of a a;
lattice 1\ may be obtained from a given basis a j in this yvay. For since
a; belongs to the lattice with basis ai' ... , an there are integers vij such
that (2) holds: and since a j belongs to the lattice with basis a~, ... , a~
there are integers Wii such that (4) holds. On substituting (2) in (4)
and making use of the linear independence of the a j , we have

L:
WiiVil =
{1 0
if i = I
otherwise.
Hence
det(w ij ) det (ViI) = 1
and so each of the integers det (wii) and det (ViI) must be ± 1; that is
0) holds as required.
We denote lattices by capital sanserif Greek letters, and in particular
by 1\, M, N, r.
If ai' ... , a" and a~, ... , a~ are bases of the same lattice, so that
they are related by (2) and (3), then we have
det(a~, ... , a~) = det(vii) det(a1 , ... , an) = ± det(a1 ,···, all),
where, for example, det (a\, ... , an) denotes the determinant of the
n xn array whose i-th row is the vector a i . Hence

is independent of the particular choice of basis for I\. Because of the

linear independence of ai' "', a" we have
d(/\) > o.
We call d (1\) the determinant of I\.
An example of a lattice is the set 1\0 of all vectors with integral
coordinates. A basis for 1\0 is clearly the set of vectors
i -1 zeros n - j zeros)
(
ei = ~,1,~ (1;2;; i;2;; tt);
and so
 Bases and sublattices 11

We note that the vectors of a lattice /\ form a group under addition:

if aE/\ then -a(/\; and if a, bE/\ then a±bEI\. We shall see later
(Chapter III, § 4) that a lattice is the most general group of vectors in
n-dimensional space which contains n linearly independent vectors and
which satisfies the further property that there is some sphere about
the origin 0 which contains no other vector of the group except o.
1.2.2. Let a 1 , ... , a" be vectors of a lattice M with basis b 1 , ... , b.-.
so that
ai = L vijb j (1 )
i
with integers Vij' The integer
1= Idet(v ..)1 = !det(a1 ..... a,,)1 _
'1 Idet(b1 ..... bn)1 -
Idet(a 1 • ....
d(M)
alln
is called the index of the vectors a1 , ... , an in M. From the last ex-
pression it is independent of the particular choice of basis for M. By
definition, I~ 0; and 1=0 only if a1 • ... , an are linearly dependent.
If every point of the lattice /\ is also a point of the lattice M then
we say that 1\ is a sublattice of M. Let a 1 , ... , a" and b 1 , ... , b n be
bases of /\ and M respectively. Then there are integers Vij such that (1)
holds, since ai~ M. The index of a 1 , ... , an in M, namely

(2)

is called the index of /\ in M. From the last expression the index depends
only on /\ and M, not on the choice of bases. Since a 1 , •.• , a" are
linearly independent, we have D> O. On solving (1) for the b i and
using (2), we have
Db i = L wija j ,
i
where the Wij are integers. Hence
DM(I\(M, (3)
where D M is the lattice of Db, bEM.
It is often convenient to choose particular bases for 1\ and M so
that (1) takes a particularly simple shape.

I
THEOREM I. Let /\ be a sublattice 0/ M.
A. To every base bl , ... , b" 0/ M there can be found a base aI' ... , an
all\ 0/ the shape
al=vllb l
a2 = V 21 b 1 + V 22 b 2
(4)

an = V" 1 b 1 + ... +V n" b",

where the Vi; are integers and Vii=!=O /oralli.
12 Lattices

B. Conversely, to every basis aI' ... , a" 0/ " there exists a basis
b1 , ••• , b" 0/ M such that (4) holds.

Proof of A. For each i (1~i~n) there certainly exist points a; in

" of the shape

where vii' ... , Vj; are integers and v;i*O, since, as we have seen,
Db/_A We choose for a j such an element of " for which the positive
integer Ivi;1 is as small as possible (but not 0), and will show that
a1 , ..• , a" are in fact a basis for A Since a 1 , .•• , a" are in /\, by con-
struction, so is every vector
(5)
where WI' . _', w" are integers. Suppose, if possible, that c is a vector
of" not of the shape (5). Since c is in M, it certainly can be expressed
in terms of bl , . _., b", and so can be written in the shape
c= tl bl + ... + tk b k ,
°
where 1 ~ k ~ n, tk=1= and t1, ... , tk are integers. If there are several
such c, then we choose one for which the integer k is as small as pos-
sible. Now, since Vu=1= 0, we may choose an integer s such that
(6)
The vector

is in " since c and a k are; but it is not of the shape (5) since c is not.
°
Hence t k- SVkk=1= by the assumption that k was chosen as small as
possible. But then (6) contradicts the assumption that the non-zero
integer Vkk was chosen as small as possible. The contradiction shows
that there are no c in " which cannot be put in the form (5), and so
proves part A of the theorem.
Proof of B. Let aI' ... , an be some fixed basis of A Since D M
is a sublattice of " by (3), where D is the index of " in M, there exists
by Part A a basis Dbl , ... ,Db" of DM of the type

Dbl=wUal \
D b2 = W 21 a l + w22 a2 (7)

Db .. = w.. lal + ... + w.... a n ,.

with integral wii and W ij =1=O (1~i~n). On solving (7) for aI' ... ,an
in succession we obtain a series of equations of the type (4) but where
 Bases and sublattices 13

at first we know only that the Vi; are rational. But clearly b 1, ... , b"
are a basis for M and so the Vi; are in fact integers, since the a i are
in M, and since the representation of any vector a in the shape
(t1' ... , tn' real numbers)

is unique by the i.ndependence of b 1 , ... , b".

From this theorem we have a number of simple but useful corollaries.
COROLLARY 1. In theorem I we may suppose further that

(8)
and that
Q~ Vi; < Vi; in case A, (9)
o ~ t'i; < Vii in case B. (10)
Proof of A. To obtain (8) it is necessary only to replace a i or b;
by -ai' -bi respectively if originally Vii<O. To obtain (9) we replace
the a i by
a; = ti1 a1 + ... + t;,i--1ai-1 + ai'
where the ti ; are integers to be determined. For any choice of the ti ;
the a;are a basis for I\. We have

where

and, for j < i, we have

v~.=t-.v,,+t.
11 I, 11 1,1'+lV'+l
1,1 .+ ... +tI,S-
.. IV·l-,1
1 '+V""

For each i we may now choose t i . i - 1, t i ,i-2' ... , til in that order so that

as was required.
Proof of B. Similar.
COR OLLAR Y 2. Let aI' "', am be linearly independent vectors of a
lattice M. Then there is a basis b 1 , ... , b" of M such that
a 1 =v11 b 1

a2=v21b1+V22b2

wt'th integers Vi; such that

vii>O O~Vi;<Vii (1~i<i~m). (11)

 Lattices

We can choose vectors a m -- l , ... , an in M such that aI' ... , an are

linearly independent. Corollary 2 follows now on applying Corollary 1
to the lattke 1\ with basis aI' ... , an'
COROLLARY 3. Let aI' ... , am (m < n) be linearly independent vectors
01 a lattice M. A necessary and sul/icient condition lor the existence 01
vectors a m +l , ... , a,. in M such that aI' ... , an is a basis is the lollowing:
every vector CE M which is 01 the shape

(12)

with real ul , ... , U m necessarily has u l , ... , U m integral.

If ~, ... , am is part of a basis aI' ... , an the condition is clearly
satisfied. Conversely if aI' ... , am satisfy the condition, let b l , ... , b n
be the basis of M given by Corollary 2 and let Vij be the corresponding
integers. Then c = b l , " ' , b m are of the shape (12) and indeed the
coefficient of ai in the expression for b; is v;;!. Hence Vii= 1 and so
v,;=O for i=4=j, that is ai=b i (1~i~m) and we may put ai=b i
(m+1~i~n).
In some contexts we shall need the following more specialized
corollary which follows at once from Corollary 3.
COROLLARY 4. Let bl , ... , b n be a basis lor a lattice M and let

A necessary and sufficient condition that

bl , ... , b rn - l , C
be part 01 some basis
b1 , ••. , bm - 1 , C, C rn + l , ... , CIt

01 M is that urn' urn+l' "', un have no common lactors =4= ± 1-

Proof. Clear.
The following characterisation of the index of a sub lattice 1\ of a
lattice M is sometimes useful. We say that two vectors c, d of Mare
in the same class with respect to 1\ if C - d is in I\. Clearly this is a
subdivision into classes: if C - d and d - e are in 1\, then C - e is in 1\.
LEMMA 1. The index 01 the sublattice 1\ 01 M is the number 01 classes
in M with respect to I\.
For let ai' b j be bases for 1\ and M respectively in the shape (4)
given by Theorem 1. Then clearly the index D of 1\ in M is
 Bases and sublattices 15

But now every C( M is in the same class as precisely one of the vectors

as is readily verified (d. proof of Theorem I, Corollary 1).

1.2.3. There is a useful transformation of the criterion of Theorem I,
Corollary 3, for deciding whether or not a set of vectors ai' ... , am
(m<n) of a lattice A can be extended to a basis for A
LEMMA 2. Let b1 , ... , b n be a basis for a lattice A and let

(1~i~m) (1 )

be vectors of A A necessary and sufficient condition that aI' ... ,am be

extendable to a basis ai' ... , an of A is that the m X m determinants formed
by taking m columns of the array
(1~i~1I1, 1~i~n) (2)
shall not have a common factor.
The condition is certainly necessary. For let a m +!, ... , an form a
basis with ai' ... , am, so that

(m + 1~ i~n) (3)
for some integers Vii' Since ai (1 ~ i ~ n) and b, (1 ~ i ~ n) are bases
of the same lattice, we have
det (Vi;) = ± 1. (4)
We may expand the determinant (4) by the first 111 and last (n - m)
rows [Laplace-development] and obtain

L V. w, = det (Vii) , (5)

l~'~R

where the v,: are the (:) determinants formed from columns of (2)
and W, is the (n - m) X (n - m) determinant formed from the remaining
(n-m) columns and the (n-m) rows,

Vii (m < ~.< /'"<)

= n, 1::;:;, 1= n ,

taken with an appropriate sign. Since the W, are integers, it follows

from (4) and (5) that the V, have no common factor.
The condition is also sufficient. For let c be a vector of A of the
shape
c = U 1 a 1 + .. "+ U," am (6)
16 Lattices

for real numbers u1 , ••• , um • On inserting (1) in (6) we have

2: u.v;j = integer = Ij (say) (7)
l~i~",

since b 1 , ••• , b" is a basis for A. We may solve (7) for the u j ' and
indeed in a multitude of ways. For example let ii j be the cofactor of
VI; in the expansion of the determinant

(1~i~tn, 1~i~lII).
Then
2: vjl;=l{u1 ,
l~j~m

so Vi U 1 is an integer. Similarly
v, u; = integer (1~i~1n),

where V, is any 11t Xm determinant formed from (2). Since, by hypothesis,

the V, are integers without common divisor, the Uj must be integers.
Hence by Theorem I, Corollary 3 it is possible to extend aI' ... , a", to
a basis aI' ... , a' II •
1.2.4 1 • We shall now apply Lemma 2 to obtain a result of DAVEN-
PORT (1955 a) about the way in which a basis for a lattice may be chosen.
This will be used only in Chapter V, § 10 and then only to prove a
result on Diophantine Approximation rather aside from the main theme
of the book.
THEOREM II. Let A be an n-dimensional lattice, let
ci (1~i~n-1)

be (n -1) arbitrary real vectors and let e> 0 be an arbitrarily small real
number. Then for all real numbers N greater than a number No depending
only on A, e and the c;, there exists a basis ai' ... , a" of A such that
(1~i~1t-1). (1)
Here
(2)
denotes the usual euclidean distance.
To prove Theorem II we shall need a result about the distribution
of integers prime to a given integer. We prove this first, and then
Theorem II.
LEMMA 3. For each 15>0 there is a number k(b) with the following
property: Every interval of length k (b) q6, where q is a positive integer,
contains an integer prime to q.
1 § 2.4 may weli be omitted at a first reading
 Bases and sublattices 17

Let
q= II p«f, (3)
l;;;'i~,J 1

where the Pi are distinct primes and the (Xi> 0 are integers. An integer
is prime to q if and only if it is not divisible by any of P1' ... , Pl'
Consider some interval
V<U~V+U (4)
of length U, where U, V are fixed integers. For j1 <j2 < ... <jS' where
s~J, let
M(j1' ···,js)
be the number of integers u in the interval (4) which are divisible by
Pi., Pi,' ... , Pi. (and perhaps also by other primes from Pl' ...• Pf)' We
show next that
(5)
s>o

gives the number of integers u in (4) prime to q. where U is the number

of integers u in (4). For let the integer u be divisible by precisely r
primes Pi' where r~1: say by P1 •...• P,. but not by P,+l ... ·.P,.
Then u is one of the integers counted in MU1' ...• js) if and only if
s ~ rand ;1' .... js is one of the (:) combinations of s out of the numbers
1. 2•...• r. Also u contributes 1 to U regarded as giving the number
of integers in (4). Hence the total contribution of u to (5) is

1 - (;) + (;) ... = (1 -1)' =0.

If, however, u is prime to q, then it contributes 1 to U but does not

contribute to the MU1 •... , ;s); so W is the number of integers in (4)
prime to q. as asserted. But

IM(;l, .... ;s)- u

Pi,'" Pi,
\<1,

since MU1' ...• js) is the number of integers

u = Pi,'" Pj,U',
where u' is an integer and
V <u's u+ V .
Pj,'" Pj, - Pi,'" Pi,
Since (5) contains 2' summands, we have

W> U{1 + L (~.~) .. } -2'=U II(1-~) -21 ~ 2-'U - 2'.

s>o Pit P], i P, .
;,<"'<i,
Cassels, Geometry of Numbers 2
18 Lattices

Hence there is an integer prime to q in the required interval provided

that

If 0 is the arbitrarily small number given in the lemma, we now have

Uo~) ~ II (46) ~ II (-;) = k(o) (say).

q j Pi P;;i,41/O P
where the second product is taken now over all primes less than 41/°.
This proves the lemma.
We shall use the lemma in the following apparently more general
shape.
COROLLARY. Let q, 0, k (15) be as in the lemma and let s, t be integers
0/ which t is prime to q. Then an interval 0/ length greater than k (0) qO
contains an integer u such that t u + s is prime to q.
For since t is prime to q we may write
s = SIt + s2q
for integers S1 and S2. Then
tu+s=t(u+sl ) +S2q·
Since t is prime to q we need only choose u so that u + S1 is prime to q;
and this is possible by the lemma.
We now revert to the proof of Theorem II. Let b l , ... , b" be any
basis for A, and let the given vectors c i be
(1~i~n-1) (6)

for real numbers Yw We shall choose a basis

ai = L vijbj (1~i~n) (7)
i
for A such that
(8)
where N> 1 is the given positive number, 0> 0 is arbitrarily small, and
the constant implied by the 0 symbol may depend on n, 15 and the Yii.
We shall choose the Vii so that for each I <n the two integers
(1 ~ i ~ I, 1 ~ j ~ I)
and

are non-zero and without common factor.

Suppose, first, that 1=1. We take for vn one of the integers nearest
to NYn which is not o. Next we choose for V12 the integer nearest to
 Lattices under linear transformation 19

N Y12 which is not 0 and prime to vll . For j> 2 we choose for VI; the
integer nearest to N Yl;' Then (8) holds for i = 1 and j =} 2 trivially
and for i = 1, j = 2 by Lemma 3, and since clearly V ll = 0 (N). The
integers Rl =Vll and 51 =V12 have the required properties.
Now suppose that 1>1, and that the Vi; with i<1 have already
*
been constructed. For j 1, 1 + 1 we take for V 1 j just the nearest
integer to Ny 1.;' On expanding R I and 51 by their last rows, we now
have
R I = ±vIIRI_l +A,
51 = ± VI,!+! 5 1 - 1 + vIlE + C,
where A, E, C are integers which have already been determined. Since
R I - 1 is prime to 51-1, we may choose the integer VII so that RI is not 0
and prime to 5 1 -- 1 , We choose for V II the integer nearest to Ny II for
which this is true, so that, by the corollary to Lemma 3,
vII -NYII=O(5Ll) = o(N(l-I) 0) ,
since 5 1 - 1 =0 (NI-l), being a sum of products of 1 -1 numbers Vi; each
of order N. Having determined VII we now take for VI,I+l the integer
nearest to Ny I,1 +! such that 51 is not 0 and prime to R I, so that
similarly
VI,!+! - NY!,I+! =O(5~) =O(N IO ).
This completes one stage of the induction. We have thus shown the
existence of integers Vi; satisfying (8).
From (7) and (8) we have
lai-Neil = o (N(n-l) 0) (1~i~1t-1).

The truth of the statement of the theorem now follows on taking <5 = eln.
1.3. Lattices under linear transformation. It is convenient here to
consider briefly the effect of a non-singular affine transformation
x-+X =(lX of n-dimensional space into itself. Let the transformation
X =(lX be given by
Xi = L (XiiXj (1~i~n), (1 )
l:;;;i:;;;n
where
X=(X1 ,···,X,,), x=(x1 ,···,xn)
are corresponding points in the transformation and (Xii are real numbers
such that
det((l) =det((Xii) =}O.
Let /\ be a lattice and denote by a./\ the set of points (lX, xEI\.
••. , b" is a basis for /\, then the general point b = U 1 b 1 + ... + uti b n
If b 1 ,
2*
20 Lattices

(u,., ... , U" integers) of A has the transform

o.b = o.(u1 b l + ... + u"b,,) = ulo.bl + ... + u"o.b".
Hence o.A is a lattice with basis o.bl , ... ,o.b", and

d (a. A) = Idet (a. b l , ... , a. b,,)i = Idet (a.) II det (bl , ... , bn)l = Idet (0.)1 d (A) .

We note two particular cases. First, if t =1= is a real number, then °

the set of tb, bEA is a lattice of determinant Itlnd(A) which we shall
denote by tl\. Secondly every lattice M can be put in the shape M =o.Ao,
where a. is of the type (1) and Ao is the lattice of integer vectors. For
if flt, ... ,a" is any basis for A, we may define IXi; by
a; = (lXii' ... , IX,,;) •

1.4. Forms and lattices. We consider first quadratic forms. Let

"
I(x) = ~ Iii Xi x; (1 )
i,i=l
where
(2)
be a non-singular quadratic form of signature 1 (r, n - r); that is, there
exist independent real linear forms

Xi = ~ diixi (3)
l;:>j;:;; ..

such that identically

I(x) = IJ?(X) , (4)
where
(5)
and
IJ?(X) = X~ + ... + X~ - X~+l - ... - X: (6)
(for r =0, n there are no positive or negative squares respectively).
We have clearly
(7)

Conversely, if di ; is any set of real numbers with det(d ji) =1=0, then (3).
(4) and (6) determine a quadratic form (1) of signature (r, n - r) and
(7) holds. We shall be concerned a great deal with the values which
I (x) takes when Xl' "', XII are integers. By (3), these are precisely the
1 Many writers define the signature of a form to be the number of positi"e
squares less the number of negative squares in (6). i.e. 2r - 11. But it is more
convenient to give explicitly the number of each kind of square than to do the
arithmetic every time.
 Forms and lattices 21

values which q; (X) takes when X runs through the vectors of the
lattice A with basis

Then, by (7), we have

(8)
In this way statements about different quadratic forms of signature
(r. n - r) at integral values are equivalent to statements about the
single form q;(X) and different lattices. For later reference we formulate
a typical result as a Lemma.
LEMMA 4. The following four statements about a number x are eqtti-
valent, where
q;(X) = X~ + ... + X~ - X~+l - ... - X!.

( i) In every lattice A there is a vector A +0 with

(ii) In every lattice A of determinant 1 there is a vector A +0 with

1q;(A)1 ~x.
(iii) In every lattice A of determi1tant d(A)~y'-n/2 there is a vector
.4+0 in
Iq;(A)1 ~ 1-
(iv) For every quadratic form L fii Xi Xi of signature (r, n-r) there
+
is an integer vector a 0 such that
I/(a)1 ~xldet(/;iWI".
That (i), (ii) and (iii) are equivalent follows from homogeneity, since
q;(tX) =t2q;(X) and since the set tA of all tX (XEA) is a lattice tA of
determinant Iti"
d (A); and we may choose t so that t" d (A) = 1. That
(iii) and (iv) are equivalent follows at once from the earlier discussion
and, in particular, from (8).
The foregoing argument is quite general. For example the behaviour
for integer values of the variables of any form I (~) of degree n which
can be expressed as the product of 1t real linear forms:

I(~) = IT (dil Xl + ... + di"x,,)

1;:;;;;:;;"
is equivalent to the behaviour of the function

q;(X) = Xl ... X"

at the points of an appropriate lattice A A single function q;(X) cor-
responds to the set of all functions I (~) that can be deduced from it
22 Lattices

by a real non-singular affine transformation

Xi = L: dijXj (d ij real, det (d ij ) =!= 0).
1.4.2. Of course the form rp(x) and the lattice A do not determine
the function I(x) uniquely, since I(x) depends on the choice of a
particular basis for A; and we shall discuss this ambiguity here. The
transformation
Xi=L:dijXj
j
of § 4.1 is just of the type
X=ax
discussed in § 3. Identifying these transformations we see that
A =aAo,
where /\0 is the lattice of all integer vectors; the particular basis
d1 , .•• ,dn
of A corresponding to the basis
i-I n-i )
(
e;= ~,1,~ (1~i~n)

of /\0' Hence any other basis d~, ... , d~ of /\ is of the shape

dj = aej,
where ej is some other basis for /\0' Let I' be the form corresponding
to the basis dj as I does to di . Then clearly there is the identical relation
I'({£') = I'(x~, ... , x~) = rp(x~ d~ + ... + x~d~) = I(x~ e~ + ... + x~e~).
But now since ej is a basis for /\0 we have
e; = (VI;' ... , vni) ,
where the Vi; are integers and
det (v ij ) = ± 1: (1 )
so that identically
j'(x') = I(x) (2)
if
Xi = L: Vii xi·
i
Conversely, if the Vii are integers such that (1), (2), (3), hold then j'
and I correspond to the same lattice /\. Two forms in this relationship
are said to be equivalent; they take the same set of values as the
variables run through all integral values, since, by '1) and (3), integral
x' correspond to integral x and vice versa.
 The polar lattice 23

It is sometimes useful to distinguish between det (v ij ) = + 1 (proper

equivalence) and det(vij) =-1 (improper equivalence) in (1). We shall
not do this, however, since it does not correspond to anything intrinsic
in the corresponding lattices.
1.4.3. The forms /(x) and <p(X) do not in general determine the
lattice uniquely, since for example a quadratic form /(x) of signature
(r,s) with r+s=n may be expressed in the shape
X~ + ... + X; - X;+l - ... - X;+s
in many different ways. Let a l , ... , a" and b 1 , ... , b" be bases of
lattices /\ and M respectively and suppose that
(1 )

for all integral 11 =(111' ... , Uti)' Since <p(X) is a form, (1) is an identity
in the variables U l , ... , u". Let w be the uniquely determined homo-
geneous transformation such that

Then

for all u, and so

<p(X) = <p(wX) (2)
for all X, by (1) and since every vector is of the shape X = L uj u j
for some real numbers u j • If the homogeneous transformation w
satisfies (2) we call it an automorph of <po We have just shown that if
(1) holds there is an automorph w of <p such that waj=b j . The con-
verse is, of course trivial that if w is an automorph of <p and wU j = b j
then (1) holds.
We shall study the automorphs of forms intensively in Chapter X.

1.5. The polar lattice l • We denote the scalar product of two n-

dimensional vectors x, y by

xy = Xl Yl + ... + x" y". (1 )

Let b l , •.. , b" be a basis of a lattice I\. Since the b j are linearly
independent, there exist vectors bt such that

b* b. = {1 if i = j (2)
I' 0 otherwise.

1 This section will not be referred to until Chapter \"III and will not be of
importance until Chapter X and XI.
24 Lattices

The lattice A* with basis bj is called the polar (or dual or reciprocal)
lattice of A. and bj is the polar basis to b;. The polar lattice A* of A
is independent of the choice of the particular basis. as we now show.
LEMMA 5. The polar lattice A* 01 A consists 01 all vectors a* such
that a*a is an integer lor all a in A. Then A is conversely the polar
lattice 01 A*. Further.
d(A) d(A*) = 1.
Suppose. first. that
a-£...
* - " ui b*j •
are in A* and A respectively. so that the ui • Vi are integers. Then
a*a = L ujvi
is an integer. Now let c be any vector such that ca is an integer for
all a in A. In particular
(1~i~n)

is an integer. Put a* = L ujbj. Then

(c - a*) b j =0 (1~i~n);

and so c =a* since the bi are linearly independent. This proves the
first sentence of the theorem. The second sentence follows immediately
from the first and also from (2). Finally. (2) implies that
det(bt •...• b!) det(b1 . . . . . b ll ) = 1.
and so d (A*) d (A) = 1. This concludes the proof of the lemma.
1.5.2. When Y=F0 is fixed. the points z such that yz=O lie in a
hyperplane through o.
LEMMA 6. A necessary and sulficient condition that there be 11 - 1
linearly independent points Ot •...• a,,-1 in A with ya;=O (1 ~i~n -1)
is that y = ta* tor some real t and some a* in A*.
Suppose first that yaj=o (1~i~n-1). Then by Theorem I
Corollary 2 there is a basis b j (1 ~i~n) for A such that
(Vi; =F 0)
for integers Vii' Hence ybj=O (1 ~i~n -1). Let yb .. =t. Then
clearly y =tb: where bj (1 ~i~n) is the polar basis to b i . This proves
half the lemma.
Suppose now that y = ta*. where a* EA·. If a* = 0 there is nothing
to prove. Otherwise. a* =mbt. where m is an integer and bt is primitive l .
1 That is, not of the shape u c·, c· EA· for an integer u > t.
 The polar lattice 25

Then bi can be extend to a basis bj for 1\.*. Let b j be the polar basis.
Then
(2~j~ n).

This concludes the proof of the lemma.

Let I\.(a*) be the set of a in I\. such that a*a = O. Clearly if a 1 and
a 2 are in I\.(a*), so is U1llt+u2a2 for any integers u 1 , U2. By Lemma 6
if a* (" 1\.* there are n -1 linearly independent points of I\.(a*), and so
in a sense I\.(a*) is an (n -1 )-dimensionallattice. The following corollary
makes those remarks more precise.
COROLLARY. Let b*=(bi, ... , b!) be a primitive point 01 1\.* and
suppose that b! =1= O. Then the set 01 (n -i)-dimensional vectors a' =
(a 1 , ... , an~l) such that lor some an the vector a = (a1 , •.• , an) is in I\. and
satisfies b*a =0 is an (n -i)-dimensional lattice M 01 determinant
d(M) =\b!\d(I\.).
We note that M is the projection on x" = 0 of the set I\.(b*) just
defined. Since b:
=1= 0, if an exists it is uniquely determined by a1 , •.. ,
an~l' and the condition b*a=O.
We may suppose that b*=b:, where bi, ... , b:
is a basis for 1\.*
and b j is the polar basis. After what was said before the enunciation
of the corollary, it is clear that the (n -i)-dimensional vectors formed b;
by taking the first n -1 coordinates of b i are a basis for M. We now
show that
(1 )

If in the determinant in the left the n-th coordinate Xn is replaced by

b: x for x = b1 , ... , b n , the value of the determinant is multiplied by
b!n=b!. Since b!bj=O for 1~j~n-1 and b!bn=1, the equation (1)
follows at once. In particular \ b!\ d(l\.) =d(M), as required.
1.5.3. Finally we must note the effect of homogeneous linear trans-
formations on the relationship between polar pairs of lattices. Let
X=1:X (1 )
be a non-singular homogeneous linear transformation given by

Xi = L: Tiixi
i
where
det(1:) =det(Tii) =1=0. (2)
If Y is any vector, we have

YX = L: ¥;X; = L: ¥;T;jXi'
i i~ i
26 Reduction

Hence
YX=yx, (3)
where
(1~j~n). (4)

Since det ('t) =f= 0, by hypothesis, the equations (4) define Y as a function
of y. We write
Y = 't*y,
where 't* is called the transformation polar to 'to

LEMMA 7. Let't be a non-singular homogeneous linear transformation,

1\ a lattice, and 't 1\ the lattice of points 'tX, x E I\. Then the polar lattice
of 't 1\ is 't* 1\ *, where 't* and 1\* are respectively polar to 't and I\.
This follows at once from Lemma 5 and equation (3) above, where
X ='tX, Y ='t*y.

Chapter II

Reduction
11.1. Introduction. In investigating the values taken by an algebraic
form I (x) for integer values of the variables it is often useful to sub-
stitute for I a form equivalent to it (in the sense of Chapter I, § 4) which
bears a special relation to the problem under consideration. This process
is independent of the geometrical notions introduced by MINKOWSKI
and depends only on the properties of bases of lattices developed in
Chapter I. Indeed only the lattice 1\0 of integer vectors comes into
consideration.
It is convenient to collect together in one chapter the various
applications of reduction. The later parts of the chapter involve some
moderately heavy computation. The beginner might well omit all after
the enunciation of the results in § 4.2. Indeed the next few chapters
are practically independent of Chapter II, which might well have been
deferred until later.
In § 2 we discuss the general method. In the rest of the chapter
we shall be mainly occupied in investigating
M(f) = inf [f(u) [
u~o
" integral

where f (x) is a form of a special type. Definite and indefinite quadratic

forms are treated in §§ 3.4 respectively and binary cubic forms in § 5.
 The basic process 27

The methods of this chapter have been successfully applied to

related problems: for example, when j (x) is indefinite, to the estimation of
inf j (u)
over integer vectors u =l= 0 for which j (ll) is positive [either in the strict
sense j (u) > 0 or the weak sense j (u) ~ 0: two distinct problems in
general] but we shall'do this only for binary forms.
A table listing the known results about quadratic forms is given in
an appendix. We shall be considering quadratic forms later from other
points of view.
DAVENPORT and ROGERS (1950a) have shown that in many cases
not merely one but infinitely many integer points u exist such that
j (u) satisfies the inequalities stated. This requires deeper methods than
those used here and will be discussed in Chapter X.
It should be remarked that there is a classical theory of reduction
for indefinite binary quadratic forms which we do not discuss here.
Although it comes into the general scope of reduction as defined here,
that is the choosing of bases with special properties, it is best under-
stood after the discussion of Chapter III. It is closely related to con-
tinued fraction theory. See Chapter X, § 8.
11.2. The basic process. We first discuss the standard procedure
for positive definite forms j (x); that is for forms such that j (x) > 0
for all real vectors x =l= o.
We note first that if j(x) is positive definite of degree r, say, then
there is a constant u> 0 such that
(1)
for all real x, where we have written
Ixl = (xi + ... + x!)~.
For on the surface of the sphere Ixl =1 the continuous function j(x)
must attain its lower bound u, so %>0; and then (1) follows by homo-
geneity. In particular, there are only a finite number of integral
vectors· u such that j (u) is less than any given number.
We now choose a basis for the lattice /\0 of integral vectors with
respect to the positive definite form j (x) as follows. Let e~ =l= 0 be one
of those integral vectors u for which j (u) is as small as possible. By the
argument of the last paragraph such u exist, and there are only a finite
number of them. If e~ were of the shape e~=ka, a~/\o, where k>1
is an integer we should have

• i.e. \'ectors whose co-ordinates are rational integers.

28 Reduction

contrary to the definition of e~. Hence by Corollary 3 to Theorem I of

Chapter I. we may extend e~ to a basis e~. b 2 • •••• b" of the lattice /\0
of integer vectors. We now choose e; (2;;:;'j;;:;' n) in succession. Suppose
that e~ . .... el~-l have already been chosen and are extensible to a base
e~ ..... ef- l • b i' .••• b n of /\0. Then e; is one of the finite number of
vectors with the property that e~ . ... ; ej is extensible to a base of /\0
and for which / (e;) is as small as possible. Such e; exist but are finite
in number. by argument used for e~. In this way we obtain a base
e~ • ...• e~: and for any given / (x) there are only a finite number of such
bases.
If the function / (x) is such that we may indeed choose
j-l n- j )
( .--"--, .--"--,

e:=e.=
, , .0
'
•...• 0.1.0 •...• 0 (1;;:;'j~n)

for the above basis, then / (x) is said to be reduced (in the sense of
MINKOWSKI). The above proof shows that every positive definite form
is equivalent (in the sense introduced in Chapter I. § 4) to at least one
and to at most a finite number of reduced forms.
We may make the definition of a reduced form more explicit. By
Corollary 4 to Theorem I of Chapter I (or by Lemma 2 of Chapter I).
a necessary and sufficient condition that ell ...• e j _ l and the integral
vector u = (u l , ... , u,,) be extensible to a basis for /\0 is that
g.c.d. (U;, ... , un) = 1- (2)
Hence the fOIT)1 I (x) is reduced if and only if
/(UI.···, un) ~ /(eJ
for all j and for all integers U l •...• Un satisfying (2).
11.2.2. When the form /(x) is not definite, then there is no generally
valid procedure to replace the reduction procedure for definite forms.

*
If we know (or may assume) that /(u) does not assume arbitrarily
small values for integral u 0 then it is possible to salvage something
of the reduction procedure. Let e> 0 be chosen arbitrarily small. By
hypothesis,
Ml = inf I/(u)l > o.
",,"0

*
integral

Hence we may find an integral e~ 0 such that

Without loss of generality e~ is not of the form ka where a is integral

and k> 1 is a rational integer. If e~, ... , ef- 1 have already been found.
 The basic process 29

write
M; = inf I/(u)1
where the infimum is over all integral vectors u such that ('~. 000. ('i-I. U
is extensible to a basis for Aoo Then

Mj~Ml>O.

and so we may choose ei so that e~. 000. ej is extensible to a basis and

II (eil! ~ Mj/(1 - e) 0

Let I'(x) be the equivalent form for which

I(e;) =I'(ei ) 0

Then we have
II'(u l • 0 ••• un)! ~ (1 - e) II' (ei )I
for all sets of integers U l • •••• Un such that goc. d. (u i • ...• un) = 1. But.
of course. there is no reason to suppose that there are only a finite
number of I' with this property and equivalent to a given f.
An alternative procedure which is sometimes possible is to find some
other form g(x). related to our given I. which is definite and to reduce
g(x)o We shall do this for binary cubic forms in § 6. This method goes
back to HERMITE. who applied it to indefinite quadratic forms as follows.
Let I (x) be an indefinite quadratic form of signature (r. n - r). so
that. as before.
I(x) =X~ + ... +X~ - X~+l-'" - X!. (1)
where the Xi are linear forms in Xl' .•.• XII' Then
g (x) = X~ + ... + X~ + X~+l + ... + X: (2)
is a definite quadratic form with the same determinant. except. perhaps.
for sign. The forms Xl' .... Xn are not uniquely determined by I(x)
but we say that I(x) is reduced (in the sense of HERMITE) if the form
g (x) is reduced in the sense of MINKOWSKI for some choice of Xl' .... X".
Clearly I (x) is always equivalent to a reduced form. since we may choose
any representation (1) and then apply the transformation which reduces
g(x). Reduction more or less of this kind was first introduced by
HERMITE. and has been further discussed. amongst others. by SIEGEL
(1940a). as a tool for investigating the arithmetical properties of quad-
ratic forms. In general a form I (x) is equivalent to infinitely many
HERMITE-reduced forms. but SIEGEL shows that it is equivalent to only
finitely many if the coefficients of I (x) are all rational.
We note here that the relationship between (1) and (2) allows
estimates for the minimum of a definite form to be extended to an
30 Reduction

indefinite one, since clearly 1/(:r) I~ g (:r) for all real vectors:r. But in
general, the information so obtained is quite weak.
11.3. Definite quadratic forms. We shall be considering definite
quadratic forms from many different points of view in the course of
this book. Here we see what can be done by reduction methods alone.
The study of reduction is of great importance in the arithmetical theory
of quadratic forms, see WEYL (1940a) or VAN DER WAERDEN (1956a),
who give references to earlier literature. Here we are concerned only
with the minima of forms.
Let
I (Xl' X 2) = III X~ + 2/12 Xl X 2 + 122 X~
be a positive definite quadratic form. We wish to prove that there are
integers (u I , u 2) =l= (0, 0) such that
I (u l , u 2) ~ (4D/3)~,
where
D = III 122 -/~2'
By taking an equivalent form, if need be, we may suppose that
M(f) = inf I(U I ,U2) =/ll'
U" u%
integers not both 0
We have
I(x l , X 2) = III (Xl + ~
III
X2)2 + ~ X~.
III
Put U2 = 1 and choose for UI an integer such that

l
UI + 11~_1
III
~ ~.
2
(1 )
Then, on the one hand,

and, on the other,

D
I(u l , 1) ~ 4 /11 + 7;;'
1
(2)
Hence

that is
I~l ~ 4D/3,
as required. That ~ here cannot be replaced by < is shown by the
form
IO(Xj, x 2 ) = xi + X 1 X 2 + x~
for which D =! and I (u 1 , u 2 ) ~ 1 for integers (u 1 , u 2 ) =l= (0,0). It is not
difficult to show by examining when equality can occur in the above
 Definite quadratic forms 31

argument that ~ can be replaced by < unless I is equivalent to a

multiple of 10' We do not go into details, since we shall prove this
later more simply.
11.3.2. As HERMITE noted, this argument can be extended to prove
the following theorem.
THEORDJ 1. A non-singular quadratic lorm

I(x) = L: Iii X; Xi
represents a value I (u) with
I/(u)1 ~ (!)(n-l)/2I D I1/n, (1 )
where u =1= 0 is integral and
D = det (Iii) .
By the remarks at the end of § 2.2 we may suppose, without loss
of generality, that I (x) is positive definite. We may now suppose, as
before, that
111~ I(u)
for all integral u =1= o. Then

I (x) = III (Xl + ~ X2 + ... + h X»)2 + g (X2' ... , x,.) ,

III III
where g(X2' ... , xn) is a definite quadratic form of determinant Dllll'
Since we may suppose the result already proved for forms in n - 1
variables, there are integers u 2 , .•. , Un not all 0 such that

g(u 2 , .. ·,u n ) <

= (~)~(n-2)( D_)'l/(n-l)
, .
J ./11
Choose the integer 1(1 so that

lUI + 1~
hi
U 2 + ... + h u" I ~ ~- .
hi 2
Then

and so
I11 :s;:
-- 3
(~)~ (n-l) Dl/n.

This proves the assertion. Unfortunately, the constant (!)~(n-l) is the

best possible only for n = 2. We shall show below that it is not the
best possible for n = 3, and since the above proof is by induction it
cannot be best possible for n ~ 3. It is possible to modify the argument
to give the best possible result for n = 3 [for a neat version of this see
;\10RDELL (1948 a) J, but we shall not do this. Instead we give a more
elegant, if more artificial, treatment depending on a more detailed
32 Reduction

examination of reduced forms which goes back essentially at least

as far as GAUSS.
11.3.3. We start with the consideration of a positive definite binary
form which is reduced in the sense of MINKOWSKI:
l(x 1 ,x2) =/11X~+2/12xlx2+/22xi.
After the substitution X1-*X 1 ' x 2 -* - X2 if need be, we may suppose
without loss of generality that
112-;;;;' o. (1 )
By the definition of reduction,
(2)
and
1(-1,1)-;;;;'/(0,1),
that is
(3)
By (1), (2), (3) we have
4D - 3/11/22 = 111122 - 4/~2-;;;;' t~l - 4/~2-;;;;' 0;
and so
Rl~/11/22~tD.
The sign of equality is required only when 111 =/22=2112; i.e. when
I(JJ) = III (x~ + X 1 X 2 + x~).
Before going on to ternary forms, we note that any form satisfying
(1), (2), (3) is reduced. This is a special case of the general theorem
that MINKowsKI-reduced forms can be characterised by a finite set of
inequalities, but here it is easy to verify directly.
Let U1 , U 2 be integers neither of which is O. If lu l l-;;;;.lu 2 1 we have
t (UI , u2) = lUll {t11 lUll ± 2/121 u21} + t 2 2 U~
-;;;;.1 u11{Ill Iu11- 2/121 u11} + 122u~
= UWll - 2/12) + 122u~
-;;;;'/11- 2/12+/22=/(-1,1);
and if 0< IUll ~ Iu 2 1 the same inequality follows on reversing the roles
of U 1 and U 2. Since 111-2112+/22-;;;;'/22' by (3), we have shown that
I (JJ) is reduced.
In particular, if t is any number -;;;;.! then the form
It = x~ + Xl X 2 + (t + i) x~
is reduced. Since
MUt} = 1(1,0) = 1
 Definite quadratic forms 33

and the determinant D 'of It is t, we see that

M(f)/Dl

may take any value t-i~ (t)i. This is in striking contrast with the
behaviour of indefinite quadratic forms (see § 4).
For later convenience we collect what has been proved so far and
express it as a theorem.

THEOREM II. A positive delinite binary quadratic lorm

l11x~ + 2/12 Xl X + I22X~

2

is reduced il and only il

\2/12\ ~ III ~ 122'
The three smallest values taken by I(u) lor a reduced lorm and integral
u=!=o are 111,122 and In - 2\/12\ + 122' where
11l~/22~/11-2\/12\ +/22'
For a reduced lorm
lulu ~ 4D/3,
where
D = In 122 - tt2'
The ratio e = InlDi may take any value in the interval

11.3.4. We now consider ternary quadratic forms. As we shall

later be considering definite quadratic forms in a wider context
(Chapter V, § 9, see also Chapter IX, § 3.3) we content ourselves with
the following.
THEOREM III. A. Let

be a positive delinite ternary quadratic lorm. Then there is an integral vector

u =!= 0 such that
I(u) ~ (2D)l,
where
D = D(f) = det(/'i)'
B. 11 I (x) is reduced, then

Cassels, Geometry of Numbers 3

34 Reduction

C. The signs 01 equality are required when and only when I(a:) is
equivalent to a multiple 01
lo(a:) =x~ +x~ +x; +X2X3 +XaX1 +X1X2.
We note again that we get as good an estimate for 111/22/33 as we
do for fl.l' This will be put in a wider setting in Chapter VIII, § 2.
Since lo(u) is an integer we have 10(u)~1. Since D(/o) =i, this
shows that the equality signs are required for 10' Part A of the theorem
follows from the rest. Hence we need only prove Part B and that
equality in B occurs only for multiples of 10'
Following GAUSS (1831a) we distinguish two cases. Suppose first
that

Then after a substitution

X j -+ ± x;
we may suppose without loss of generality that
112~O, 123~o, 131~o.
Write
(I;i = Ii;)' (1)
Then
{}i; ~ 0

for all i =t= j since I is reduced. For example

1(1, -1, 0) ~ 1(1, 0, 0)
gives {}2l ~ o. We have identically
2D -/1l/22/33 = {}32{}21 {}13 + L {/1l/23{}23 + 123 {}13{}21} , (2)
where the sum is over cyclic permutations of 1,2,3; as is readily
verified on expressing both sides in terms of the I;; alone 1 • Since all
the terms on the right-hand side of (2) are non-negative, we have
(3 )
as required.
The other case is when 112/23/;'1~o, and then we may suppose that
112~O, 123~O, 131~o.
We write now
"Pi; = Ii; + 21i;
and
Wi = 1(1,1,1) -I.. ·
This is an application of LITTLEWOOD'S Principle: all identities are trivial
1
(once they have been written down by someone else).
 Indefinite quadratic forms 35

Then "Pij~ °and Wi~ 0, since I is reduced. Then identically

6D - 3/11/22/33 = "P23"P3 1"P12 + } (4)
+ 2"P32 "PI 3"P2l + L {Ill (-/23) ("P23 + 2wl) + (-/23) "P13 "P21}'
Again all the terms on the right-hand side are non-negative, so (3) holds.
We leave to the reader an examination of when equality can occur.
A rather tedious investigation of cases shows that it can occur only
when
111 = 122 = 133
and either 2/23=2/31=2/12= ±1, or one of 2/23,2/31,2112 vanishes and
the remaining two are equal to ± 1. But all these forms are equivalent
to 111/0 (x), as is readily verified. For example,
x~ + x~ + x~ + Xl X2 + X2 Xa = 10 (Xl' X 2 + Xa , - Xa) •

GAUSS lists several other identities which could be used instead of

those here.
11.4. Indefinite quadratic forms. These will also be considered
again and again throughout the book from different points of view.
A table listing known results is given in Appendix A. We do not here
carry the reduction argument as far as it will go, but only far enough
to illustrate the different nature of the results from those obtained in
the definite case.
We shall continue to use the notation
M(I) = inf II (u)1 '
"'*'o
integral

where 1(x) is a form in any number of variables, and write

D = D (I) = det (Iii)
for a quadratic form Lliixixj=/(x).
There are two characteristic differences between the behaviour of
M(I) for definite and indefinite forms. For definite binary forms we
saw that M(I)/ID(I)I~ could take any value (! in an interval

°< (! ~ (t)~,
where (t)i was the maximum possible value. It is not difficult to verify
that definite quadratic forms in any number of variables behave simi-
larly, d. Chapter V, Lemma 6. The first difference in the behaviour
of indefinite quadratic forms is rather trivial: it is quite possible that
M(I) =0, and this may occur either because there is an integral U::f.:O
such that I(u) =0, or because there are integral U::f.:O such that I(u)
3*
36 Reduction

is arbitrarily small but not o. The second difference is deeper: the values
of M(/)/I D (1)\1 do not fill the complete interval up to the maximum
possible value.
The position for indefinite binary quadratic forms has been the
most investigated. Here a very great deal is known about the possible
values of M(I)/I D (/)1 i. The greatest value is (f)l, given by the multiples
of X~+XIX2-X~. Otherwise M(t)~(~_)lID(t)ll. A well-known theorem
of MARKOFF ("the MARKOFF chain") states that there are only de-
numerably many possible values of M(/)/ID(f) 11 greater than t. There
are certainly intervals to the left of i which contain no values of
M(I)/I D (f) I~· The author has given a proof of the Markoff chain theorem
in his Cambridge Tract [CASSELS (1956a)], to which the reader is referred
for references for the various statements made in this paragraph. Here
we shall be content with finding the two largest possible values of
M(/)/1 D(/)I!·
There is a similar state of affairs for ternary quadratics but much
less is known. The most complete information is due to VENKOV (1945 a)
who has found the eleven largest values of M(f)/1 D (/)11, but they do
not seem to follow any general pattern, except that they are all given
by forms with integral coefficients. There are two unsolved problems
about indefinite ternaries which appear completely intractable. It is
not known whether there are forms t with M(f) > 0 which are not
multiples of integral forms; and it is not known whether the set of
values of M(f)/ID(f)ll has any limit point other than o. These two
problems are closely related [CASSELS and SWINNERTON-DYER (19SS·a);
see also Chapter X, Theorem XII].
This phenomenon of "successive minima" (not to be confused with
the "successive minima" of a lattice with respect to a point set which
is discussed in Chapter VIII) occurs very widely with indefinite forms.
It takes a great many different shapes and a general theory hardly
exists *. It is not possible to predict when it occurs: for example it does
not occur in the problems discussed in § 4.5 or § 5.
It is not difficult to see how "successive minima" can occur. An
inequality of the type I/(u)1 ~ 1, where 1(;£) is an indefinite form and u
is an integer vector, is really a pair of alternatives
either I(u) ~ 1
or l(u)~-1.

Each of these inequalities may be regarded as a linear inequality in

the coefficients of I. If we consider a large number of different u then
• MAHLER has shown that the minima form a closed set. In fact this follows
at once from his compactness theorem of Chapter V.
 Indefinite quadratic forms 37

the various pairs of alternatives are a priori independent. It may turn

out. on combining the various alternatives. that some combinations of
alternatives are altogether impossible while other combinations of alter-
natives define a form I uniquely. An example may make this clearer.
Suppose that we are interested in binary quadratic forms for which
M(f)=1 and 1{1.O)=1.
Such a form has the shape

I{x) = } (i)
xr + ex Xl X 2 + .B xL
where the coefficients ex
and .B are to be investi-
gated. The only such form
which satisfies the ine-
qualities
I{O.1) ;;;:;-1.
1{1.1) ;S+1 .
Fig. 4
1{2.-1) ;S +1.
is xi + Xl x 2 - x~. as the reader will easily verify. Hence any other form
with I (i. 0) = 1 and M(f) = 1 must satisfy at least one of the inequalities
I{O.1);S+1. 1(1.1);;;:;-1. 1(2. -1);;;:;-1. The form X~+XIX2-X~ is
thus in a strong sense isolated from all other forms (1) with M(f) = 1.
For example if ex and .B are plotted as cartesian coordinates for the
form I. a condition I/{u l • u2)1;S 1 excludes a strip of the plane between
two parallel lines. The three conditions

I/{O.1)I;S 1. 1/(1.1)1;S 1. 1/(2. -1)I;S 1

exclude three strips. What is left consists of the point (1. - 1) and a
number of infinite regions which are separated from the point by one
of the strips (see Fig. 4).
In the actual proofs. this general principle tends to be obscured. If I
is an indefinite form and M(f) = 1 there is not necessarily an integral
vector u with II (u)1 = 1. though there are integral vectors with
1;;;:; II (u) 1< 1 + e for any given e> O. and further devices must be used
to deal with this. The difficulty is that if t> 1. then the form t I (x) = I' (x)
satisfies the same choice of inequalities "I (u);S 1 or I (u);;;:; - 1" as the
original/(x). Here t might be arbitrarily close to 1. that is. the coef-
ficients of I'(x) might be arbitrarily close those of I(x). Hence to pin
down I (x) uniquely we must some-how make use of the normalization
M(f) = 1. We do this by first finding the determinant of the form in
38 Reduction

question and then using this as part of our information. The actual
proofs will make the details clearer.
We shall later deal with isolation of this type from a more sophisti-
cated point of view (Chapter X). The treatment there will also help to
show why the additional devices just mentioned are effective.
11.4.2. The problem of the minimum of indefinite binary quadratics
has already been discussed in § 4.1. All we shall actually prove here
is the following.
THEORElI! IV. Let
I (x) = III x~ + 2/12 Xl X2 + 122 X~ (1 )
be an indelinite lorm and
D =D(I) =/1l/22-/~2'
Then
(2)

except when I is equivalent to a multiple alone 01 the two lorms

lo(x) = x~ + X I X2 - xi, (3)

11 (x) = x~ - 2x~ (4)
lor which M (I) =1 and ID I = i, 2 respectively.
That 10 and 11 are exceptional is clear, since they both represent
only non-zero integers for integral U=F o. The constant 100 in (2)
221
cannot, in fact, be improved since the next form of the MARKOFF chain is
12 = 5 xr + 11 Xl X 2 - 5x~
which has D (/2) = - 221/4 and can be shown to have M(l2) = 5.
We now prove Theorem IV. If M(I) =0 there is nothing to prove.
Otherwise, we may suppose, without loss of generality, that
M(I) = 1,
by considering tl instead of I, where t is a suitable number. By the
general argument of § 2.2, there is a form g (x) = g. (x) equivalent to
I (x) for which

where B is any given positive number in the range 0 < B < 1. Put
± g(1, 0) = (1 -11)-1,
where
 Indefinite quadratic forms 39

Since the equivalent forms I and g have the same determinant D, we

may write g(x) in the shape

± g(x) = (xl +t:XX2)2 -IDI (1-1]) xt (5)

1-'Y)

where ex = ex. is a real number, which may be supposed to satisfy

(6)
on replacing Xl by ±xI +vx 2 with a suitable integer l'. Since M(f) = 1,
we have either
(7)
or
(8)

for each pair of integers U 1 , U 2 not both 0. Of course as e changes there

is no reason to suppose that for fixed u the same alternative (7) or (8)
always holds.
We consider various suitable pairs of integers U 1 , U 2 and must con-
sider various cases according as (7) or (8) holds for the integers in
question. Since we wish to single out the forms (3) and (4), we natu-
rally choose values of u such that lo(u) = ±1 or Il(U) = ±1.
In the first place, (7) cannot hold with (u 1 , u 2) = (0, 1) since by (6)
it would imply ID I< 0, at least when 1] is small enough. Hence on
putting (u 1 , u 2 ) = (0,1) in (8), we have
(1-1])2IDI ~ (1-1]) +ex2 (9)
for all e less than some eo> 0.
We now consider the two possibilities when (u 1 , u 2 ) = (1,1). Suppose,
first, that there are arbitrarily small values of e such that (7) holds.
For these e we have, suppressing the suffix e, that
(1-1])2IDI ~ - (1-1]) + (1 +ex)2. (10)
On eliminating IDI between (9) and (10) we have
20c ~ 1 - 21]
and so
(11)
by (6). On substituting this in (9) and (10), it follows that IDI can
differ from i at most by terms of the order of 1]. But now ID I is in-
dependent of 1] and either 1] = 0 or 1] > 0 can be made arbitrarily
small. Hence IDI =i. We now revert to one particular g(x) =g.(x)
for which (10) is true, where now we have the additional information
40 Reduction

I DI =!. On substituting D = i. IX~ l-1] in (9), we have

1]2-21]~0.

Since 1]<2, this implies 1] =0. Hence IX =l and

± g{XI' x 2) = (Xl + lX2)2 - ix~ = lo{x1' x 2)·
Otherwise (1O) cannot hold, when 8 is small enough; and so for all 8
less than some 8 1 > 0 we have (8) with u = (1, 1), that is
(12)
We now consider the possibilities for u={-3, 2). Note that
11(-3,2) =1, where 11 is given by (4). If there are arbitrarily small
values of 8 such that (7) holds with u = (- 3,2), then for these 8

(13)
On eliminating IDI between (12) and (13) we have 4IX~1], so 0~4IX~1]
by (6). On substituting in (12) and (13) and using the fact that 1] =0
or 1] can be made arbitrarily small and positive, we find that ID I = 2.
Finally on putting I D I = 2, IX ~ 0 in (12) we get 1] = 0, so IX = 0 and
± g(x) = x~ - 2X~ = 11 (x) .
Otherwise for all 8 less than some 82> 0 we must have (8) with u =
(- 3,2), that is
4(1-1])2IDI ~ (1-1]) + (-3 + 2OC)2. (14)
But now the right-hand sides of (12) and (14) increase and decrease
respectively in O~IX~l. If IX~ 1~ we use (14) and if IX;;;; l~ we use (12).
In either case we obtain I D I ~ 2.21 + 0 (1]), so I D I;;;; 2.21 since I D I is
independent of 1].
It is at first sight remarkable in these proofs that the inequalities
obtained show that 1] = o. As already mentioned, this is tied up with
the phenomenon of "isolation" which we shall discuss more fully later.
11.4.3. We consider now the "one-sided" problem for ·indefinite
binary quadratic forms. In contrast with § 4.2 there is here no set of
successive minima. Theorem V A, which we now enunciate, is a special
case of Theorem IX of Chapter XI and is due to MAHLER.
THEOREM V. A. Let
I (x) = 111 x~ + 2/12 Xl X2 + 122 x~
be an indelinite quadratic lorm and

D = 111/22 - 1~2.
 Indefinite quadratic forms 41

Then there is an integral vector u =+= 0 such that

0< feu) ~ 21 Dil. (1)
The sign of equality is required when and only when f is equivalent to a
multiple of

B. For any e> 0 there are infinitely many forms, not equivalent to
multiples of each other, such that
M+ (I) = I(u)inf> 0 f(u) > (2 - e) IDli. (2)
u integral

We first prove A. That fo = X 1X 2 is exceptional is obvious, so we

need only prove (1) and that equality can occur only when stated. As
in § 4.2, we may suppose that
M+U) = 1,
where M + (f) is defined by (2). Hence, as in § 4.2, there is a form

g(z) = (Xl + CU.)2 - (1 -1]) IDI X~

1-"1
equivalent to f, where
O~oc~t

and 1] ~ 0 can be made arbitrarily small·. Suppose, first, that

g ( - 1, 1) ~ 1. Then
(1 -1])21 DI ~ (1 - OC)2 - (1 -1]) ~ 1],
which is impossible if'fJ is small enough, since IDI is independent of 1].
Hence g(-1, 1)~O, that is
IDlz
-
(1 -ex)· Z ~
(1 -"1)2 - 4 '

the sign of equality being required only when oc = t, 1] = 0; that is when

g(z) = (Xl + tX2)2 - !x~ = Xl (Xl + X2) = fo(X1, Xl + x2).
It remains to prove B. It will be shown in § 4.4 that the forms
f,,(x) =k(X~+X1X2) -x~
have

when k is a positive integer. Since

ID(f,,)! = lk 2 + k,
• More precisely, we should work with a family of forms g,(3:) as in § II 4.2.
Having once carried out this type of proof in full rigour. in the rest of this chapter
we shall be more informal.
42 Reduction

the ratio
111+ (/k)!1 D (/k)l~
may be arbitrarily close to 2.
Another simple proof of B would be by means of continued fractions.
11.4.4. As an interpolation between the problems of § 4.2 and 4.3
one may consider the forms / (x) such that there is no integral point
u =FO in
-a</(u)<b,
where a and b are given positive numbers.
For some values of a and b one may deduce the least possible value
of D (f) from the results of § 4.2. For example I if

a = 1, b=~
10
we certainly have
111(/) ;:s 1 ,
and so by Theorem IV either
ID(/)I ;:s 2
or / is equivalent to
t (xi + Xl X 2 - x~)

for some t. In the second case it is clearly enough that t?:.~. The
- 10
corresponding determinant is (~)2. ~ < 2. Hence we have an isolated
10 4
first minimum. Note that the form with the least IDI does not take
any values in the neighbourhood of - a.
For any given values of a and b the techniques of §§ 4.2, 4.3 some-
times apply. For example, the minimum determinant when a = 5, b = 3
is ID I = 24 given by 3xi - 8 x~; this being isolated. The verification of
this statement is left to the reader. Here we shall prove only the follow-
ing theorem due essentially to SEGRE (1945a).
THEOREM VI. Let
(1)
have determinant
(2)

Suppose that there is no integral u =F 0 such that

-a</(u)<b, (3)
where a>O, b>O. Then
ID I ;:s a b + 1max (a 2 , b2 ) • (4)

1 This remark was made to the author by Professor C. A. ROGERS.

 Indefinite quadratic forms 43

lib> a, the sign 01 equality is required when and only when

k =b/a (5)
is an integer and
!(z) = a!k(z) , (6)
where
Ik(z) =k(x~ +X 1 X 2) -x~. (7)

For k = 1, Theorem VI is contained in Theorem IV. When k is not

an integer, an explicit improvement of (4) can be given. When k is an
integer, there is isolation and much more is in fact known rSAWYER
(1953a), TORNHEIM (1955a)]. When b5:a the cases of equality may,
of course, be deduced from the theorem by interchanging a and b.
We may suppose without loss of generality that
a=1, b=k,

where at first k is not necessarily an integer. Let

c = M+ (/) =l(u»O
inf I (u) ,
so that
c;;;' k.

As in § 4.2 there is a form g(z) equivalent to I(z) of the shape

g(z) = _c_ (Xl + CXX2)2 _lEl (1 - T}) X~,

1-1) C
where
05:cx5:!
and T};?;. 0 may be chosen arbitrarily small.
Clearly g{O,1)<c, so g(o, 1)5:-1. Hence g(1, -1)<c, and so
g (1, -1) ~ - 1, that is

--+
ID-I 21 -1) c c2 (1-cx)2.
(1-1))2
Hence
ID I ;;;. c + i c2 ;;;' k + i k 2
with equality only when
T}=O, cx=!, c=k,
so

It remains to see whether Ik(Z) has any integral solutions U=F0 of

(8)
44 Reduction

Since Ik (1, 0) > 0 but Ik (1, x 2) -+ - 00 as x 2-+ + 00, there must be some
integer v ~ 0 such that
Ik(1,v}~O>lk(1,v +1}.
If (8) were insoluble, we should have

Ik(1,v)~k, Ik(1,V+1)~-1:
that is
v(k - v) ~ 0, (v + 2) (k - v) ~ O.
This is possible only when v =k, i.e. when k is an integer.
It remains only to show that when k is an integer there is no integral
U=f=O such that -1<lk(u)<k. Since the roots {} of Ik({}' 1)=0 are
irrational, it is impossible that Ik(U) =0. Hence we must deduce a
contradiction from
(9)
If there are several solutions of (9) we choose one for which the integer
I"11 is as small as possible. Clearly
"1 =f= O.
We require the identities
I. (z) = Ik {(k + 1) Xl - x2 , - k Xl + X 2}
= Ik{Xl + X2 , kX1 + (k + 1) x 2}
= (k + 2) Xl {(k + 1) Xl - X2 } - {(k + 1) Xl - X2 }2 - X~
= (k + 2) Xl (Xl + X2) - (Xl + X 2)2 - x~.
Since I. (u) > 0, the last of these identities shows that
"1 {(k + 1)"1 - "2} > 0
"1 ("1 + "2) > O.
On writing - U for U if necessary, we thus have

"1>0, (k+1)"1>"2>-"1' (10)

From the first two identities and the minimal properly of 1"11, we have
1"1 + "21 ~ "1'
I(k + 1) "1 - "21 ~ "1 :
and so, by (10),

But then
I.(u) = k"~ + U2(ku1 - u 2 ) ~ ku~ ~ k.
Hence our assumption (9) was false.
 Indefinite quadratic forms 45

By considering 1(1, v) for all integers v, the estimate (4) may be

improved when k is not an integer. Since Ik is the only form I satisfying
1(1,0) =k and
I(O,1)~-1, 1(1,k+1)~-1,

1(-1,1)~-1, 1(1,k)~k,

the form I" gives an isolated first minimum when k is an integer. The
proof of these statements is left to the reader (d. papers quoted at the
beginning of § 4.4).
11.4.5. We now consider indefinite ternary forms. As already noted
(§ 4.1) there is a set of successive minima, the first eleven having been
found by VENKOV (1945 a). There is a derivation of the first four
minima due to OPPENHEIM in DICKSON (1930a) and a neat proof of the
first minimum only by DAVENPORT (1947a). Here we shall prove only
the following result.
THEOREM VII. Let
I (a:) = L Iii Xi Xi (1)
be an indelinite ternary quadratic lorm with determinant
D(f) =det(fii) *0. (2)
Then
M(f) = inf I/(u)! ~ I!Dli, (3)
"*'o
integral

except when I is equivalent to a multiple 01

10 = X~ + Xl X 2 - X~ - X 2 Xa + X~. (4)
Further,
(5)
We first prove (5). Since lo(u) is an integer when u=f=o is integral,
it is enough to show that lo(u) =1= 0. Now
4/o(u) = (2Ut + U 2)2 + (2ua - U 2)2 - 6u~.

Hence it is enough to show that there are no integral solutions of

v~ +v~ = 6v~
other than Vl =V2 =Va=0. We may suppose that VI' V2 , Va have no
common factor. Then clearly Vl and Va must be divisible by 3. Then
v~ +v~ must be divisible by 9, so v2 is divisible by 3; a contradiction.
That the consta:tlt ! in (3) cannot be further improved is shown by
Il(a:) = x~ + X l X 2 - x~ - 2x~.
46 Reduction

The reader should have no difficulty in modifying the proof to show

that this is the only case when there is equality in (3) and that it is
isolated.
We may suppose as before that
M(f) = 1 (6)
and, by taking - 1for 1if necessary, that
D<O. (7)
We have to show that 1 is equivalent to 10 or D~ -~. It is convenient
to enunciate steps of the proof as propositions.
PROPOSITION 1. Either

M+ (I) = inf I(u) = 1 (8)

l(u»O
or
D~ -f. (9)

If (6) is true but (8) is false, there must be integral u such that
I(u) = - (1-1]tl , where 1] ~O may be chosen arbitrarily small. Hence
1(x) is equivalent to a form g (x) of the shape
(1 -1]) g(x) = - (Xl + rx. X2 + (J X3)Z + h(X2' x3),
where rx., (J are real numbers and the form
h(x) = hZ2 x~ + 2h23 X2Xa + h33 xi
must be positive definite. The determinant of h (x) is
h22h33 - h~3 = - (1 _1])3 D = (1 -1])31 DI.
After a transformation on the variables x 2 , x3 , we may suppose that
h(x) is reduced; and so
( 10)
by Theorem II.
We now consider the indefinite binary form
G(XI' x 2) = (1 -1]) g(XI' x 2 , 0) =- (Xl + rx.x z)2 + h22X~,
of determinant - h22 • Clearly
M(G) ~ (1 -1]) M(g) = 1 -1].
Hence, by Theorem IV, either
h ~ 221 (1 _ )2 ( t 1)
22- 100 1]
 Indefinite quadratic forms 47

or G(Xl' x 2) is equivalent to one of t(Xi+XIX2-X~) or t(xl-2xi) for

some number t with 1tl ~ (1-17). If the second alternative holds, we
must have t = -1, since G( 1, 0) = -1. Then there arc integral u1 , u 2
such that G(U)=+1, i.e. g(U 1,U2, 0) =(1-17t1, so

M+(f) = M+ (g) = 1
since 17~ 0 may be chosen arbitrarily small. Otherwise the first alter-
native, namely (11), holds, and so, by (10),
IDI2(1_17).]_.(221)2> 7
- 4 100 2
This proves the proposition.
We may now suppose that
M+(f) = 1. (12)
As before, there is a form g equivalent to t such that
(1 -17) g(x) = (Xl + (X X2 + (J XS)2 + h(X2' xs),
where 17 ~ 0 may be chosen arbitrarily small, and the form

h(X2' xs) = h22 x~ + 2h23 X2Xs + has x~ (13)

is now indefinite and has determinant
h22h33 - h~3 = (1 - 'Y))3 D < O. (14)
PROPOSITION 2. It U2 , u3 are integers not both 0, then either
h(u 2, u3) ~! - 'Y), (15)
or
(16)
or
- £- 17 ~ h(U 2 , u3) ~ - £ + 'Y) . (17)
Further, it (17) holds there is an integer v such that
Iv + 1- ((XU2 + (Ju3)! ~!'Y). (18)
We must first show that there are no integral solutions u4=O of

- 2 + 17 < h (u 2, u3) < - £- 'Y),

- ! + 17 < h(u2, u3) ~ - 1 + 'Y),
-1 +'Y)<h(u2,Us) <!-17'
We may clearly choose the integer UI so that respectively
1 ~ 1u 1 + (X u 2 + {J u3 1 ~ !,
i- ~ Iu 1 + (X u 2 + {J tt31 ~ 1 ,
48 Reduction

and
0;:;:;; lUI +«Ull+PUaI ;:;:;;i·
Then in each case we have
(1- t l) Ig(u)1 < 1 - fJ.
contrary to hypothesis.
Suppose that (17) holds. There is an integer t and a real number
l' such that by choice of sign

IXU 2 +PUa=t±1'. O;:;:;;1';:;:;;t·

We may clearly choose integers u~. u~' so that
Iu~ + IX U l + Pual = 1 - l'
Iu~' +lXu 2 +pual =1 +1'.
Then

and so
h(ua.ua) + (1- T)2 =g(U~,U2.Ua);:;:;; -1 +fJ, (19)
h(u2• us) + (1 + T)2 = g(u~'. U2• us) ~ 1 - fJ. (20)
By subtracting (19) from (20) we have
i(1-fJ);:;:;;1';:;:;;i·
This is equivalent to (18) and so proves the proposition.
COROLLARY. If (17) holds, then U2 and ua cannot have a common
factor except ± 1.
For if U2=VU~, Ua=vu~, where v> 1, none of (15), (16) or (17) would
be satisfied by h(u~, u~).
PROPOSITION 3. Either
IDI~i (21)
or, after an equivalence transformation, we may suppose that

-i-fJ;:;:;;h(1,O);:;:;;-i+fJ. (22)
h(1,-1)~!-fJ, (23)
-! - fJ;:;:;; h(1,1);:;:;; -! +fJ. (24)
h(2.-1);:;:;;-2+fJ. (25)
IIX - il;:;:;; ifJ. (26)

1.BI~fJ (27)
provided that fJ is less than some absolute constant fJo> O.
 Indefinite quadratic forms 49

Suppose, first, that there are no solutions of

- 2 + rJ < h(u2, u3) < 1 - rJ·

Then, by SEGRE'S Theorem VI, we must have
Ih22 hSS - h=31 ;;;; H2 - rJ)2 + (2 - rJ) (1 - rJ) = 1(2 - rJ) (1 - rJ)·
Hence, by (14),
IDI -~ 5(2-1]) ~~.
4 (1 - 1])2 - 2

Otherwise by Proposition 2 there is a solution of Ih(u 2 , Us) +11 ~'Y) and

by Proposition 2, Corollary we may suppose, after a suitable trans-
formation on x2 , X 3 , that
- 1 -'Y) ~ h(1, 0) = h22 ~ -i +'Y)' (28)
After a further substitution of the type X2~ ± x2 +v X s , where v is an
integer, we may suppose further that
(29)
We now consider h(u 2 , us) for various choices of u 2 , us. If h(O, 1)
~ - 1+rJ;
that is h33~ - 1 +'Y), we should have

h22hS3 - h~3 > 0,

contrary to the assumption that h is an indefinite form. Hence

h33> - !+rJ, and so, by Proposition 2,

h33 = h(O, 1);;;; 1 -rJ·

But now, by (29),
h(1, -1);;;; h22 + h33> - 1 +rJ,
and so, by Proposition 2 again,

h22 - 2h23 + haa = h(1, -1) ~! - 'Y). (30)

Hence
h(1, 1) = h(1, -1) + 4h23> - 2 +rJ
by (29).
We now consider the two remaining possibilities for h (1, 1) allowed by
Proposition 2. Suppose, first, that

so

Then, by (14),
(1-rJ)3I D I =h~3-h22h33;;;; -hU h3S'
Cassels, Geometry of Numbers 4
50 Reduction

so
IDI -~ (1 -t1]LJ_ > ~
(1 _7])2 2 = 2 '

which is all we require. We may therefore suppose that

-! - fJ ~ h(1, 1) = h22 + 2h23 + ha3 ~ -! +fJ·
We now invoke the part of Proposition 2 referring to oc and fJ with
(u 2 , u 3 ) =(1,0) and (1, 1). Hence there are integers v' and v" such that

Iv' + l- oc I~ IfJ,
Iv" + l- (oc + fJli ~ IfJ·
After a substitution of Xl +v' X 2 + (v" - Vi) Xa for Xl we may suppose
indeed that
1 .!-ocl~J-l1
2 - S·I'

II- {oc + fJ)1 ~ IfJ·

Then
IfJl~fJ·
We now consider
h(2, -1) = h(1, 1) + 3h22 - 6h2a ~ h(1, 1) ~ -! +fJ.
We cannot have h{2, -1)~-!-fJ, since then by Proposition 2 the
fractional part of 20c - fJ would be about i-, while we know that 20c - fJ
is 1 +O(fJ). Hence
4h22 - 4h23 + h33 = h(2, -1) ~ - 2 +fJ.
This completes the proof of the assertions of Proposition 3.
We now conclude the proof of the theorem. The inequalities (22)
to (25) of Proposition 3 are linear inequalities in h22' h23 , h3a · Put
h22 =-!+).fJ, (31)
h23 =- ! +,ufJ, (32)
h3a =1+vfJ· (33)
Then (22) to (25) become
1).\ ~ 1, {H)
). - 2,u + v ~ - 1, (35)
I). +2,u +vl ~1, (36)
4). - 4,u + v ~ 1- (37)
Hence
2v = (). - 2,u + ,,) + l). + 2,u + v) - 2). ~ - 4,
3v = 4). - 4fl + v + 2{). + 2,u + v) - 6), ~ 9,
so
Ivl ~ 3·
 Binary cubic forms 51

Hence
by (36). Hence and by (14),
(1 .- 1]) 31 D I = h~ 3 - h22 hsa = I + (- A- ,u + i V) 1] + 0 (1]2) • (38)
= I+O(1])
But D is independent of 1], so*
IDI =f.
Suppose, if possible, that 1] * O. On putting ID I = I in (38) we have

- A - ,u + iv = - ! + 0 (1]) .
For small enough 1] this contradicts (34), (35) and (36), since they give
- A - ,u + iv = - t A + i (A - 2,u + v) + i (A + 2,u +v)
~-t-i-i=-f·
Hence 1] =0, so by (13), (26), (27), (31), (32), (33), we have
g(x) = (Xl + IX2)2 - ix~ - x2 xa + x~ = 10 (x) .
Since g(x) is equivalent to I(x), this concludes the proof of Theorem VII.
1I.S. Binary cubic forms. We must first consider briefly the
algebra associated with a binary cubic form
(1 )
Such a form may always be split up into linear factors with real or
complex coefficients:
I(X I ,X2 ) =[J({}jXI +"PjX2 ), (2)
l:oi:o3
With the form is associated the discriminant

It is easily verified that

D (/) = 18 abe d + b ell -
i 4a e3 - 4d b3 - 27 a2d2 (4)
(see § 5.2). From (3) it follows that D(/) =0 if and only if I(xl , x 2)
has a repeated linear factor. Forms I with D(/) =0 are called singular.
The discriminant D (I) is an invariant of the cubic, in the sense
that if
(5)
• More precisely. we should have worked with a family of forms g.(:lJ) as in
§ II 4.2. each form with its own TJ = TJ. and O~TJ < E. Then A. p.. v depend on E,
but (38) is true for all sufficiently small E.
4*
S2 Reduction

identically for some numbers «, fJ, y, <5, then

D (I) = (<< <5 - fJy)8 D (I), (6)
as follows at once from (3) and the fact that
I'(xt , x2) = II (1'); xt + 1f'; XI), (7)
;
where
(8)

In particular, D(f') =D(/) if 1and I' are equivalent, since then (5) holds
for some integers (1., fJ, y, <5 with (1. <5 - fJ Y= ± 1-
If a, b, c, d are real, then either all the ratios 1f';I1'); are real or two
of them are conjugate complex and the third is real, since roots ~ of
an equation I(t 1) =0 with real coefficients occur in complex conjugate
pairs. This subdivides the real non-singular binary cubic forms into
two essentially distinct types. We show now that two forms in the
same type may be transformed into each other by a transformation of
the type (7) with real «, fJ, y, <5. It is enough to show that all forms I
of a given type may be transformed into an I' which is fixed for the type.
We may suppose without loss of generality that either

1');,1f'; are all real

or

in our two respective cases, where the bar denotes the complex conjugate.
Clearly these two cases are characterised by D> 0 and D < 0 respectively.
There exist numbers At, A2' Aa not all 0 such that

Al1f'1 + Aa1f'2 + Aa1f'3 = 0 }

(10)
At 1')1 + AI1')a + Aa1')3 = O.
If, say, Aa=O, we should have
{}t1f'z - {}a1f't = 0,
and so D(/) =0 by (3), contrary to the hypothesis that I is non-singular.
Hence AtAlAa=!=O and we may suppose, without loss of generality, by
multiplying Al' Aa, Aa by a common factor, that
(11)
We now distinguish the two cases according as (9t ) or (92) holds. If
(91 ) holds, we may suppose that AI' A2 , Aa are real and put
(f = 1, 2).
 Binary cubic forms 53

Then

and so, by (11),

I(Xl ,X2 ) = Xl X 2 (Xl +X2 )· (12)

If (92 ) holds, we may suppose that A2 =Al , As=Aa and put

eXl + el XI = Al (1?lXl + 'PI X 2) }

(13)
elXl +e X 2=A2(t?2 Xl + 'PI XI) ,
where e is a complex cube root of 1. Then, by (10),
Xl + XI = As (1?a xl + 'Paxa) (14)
and
I(xl , XI) = X~ +~. (15)
The coefficients IX, p, ')I, d in
Xl = IX Xl +Pxa, XI = ')I Xl + "XI
are real, since the two equations (13) here are complex conjugates one
of the other.
In the sense of § 4 of Chapter I the values taken by non-singular
binary cubic forms are the values taken by the function

or

at the points of a lattice. The reader will have no difficulty in verifying

that there is a corresponding result for singular cubic forms, with
tp(X) =X~X2'
tp(X) =xt
according as only two or all three of the linear forms 1?jXl +'Pj Xa are
multiples of each other.
It was first shown by MORDELL (1943 b) that if I is a real cubic form,
then there is an integer vector u =+= 0 such that

(16)

according as D> 0, D < 0 or D = 0, where e is an arbitrarily small positive

number. The third case, when I(z) is singular, may be dealt with
54 Reduction

trivially by MINKOWSKI'S theorem of the next chapter, so we do not

discuss it here. That the coefficients 49, 23 are best possible in their
respective cases is shown by the binary cubic forms

(17)
and
(18)

These have discriminants 49 and 23 respectively. Since they do not

represent 0 and represent integer values for integer vectors u, the:;:;; in
(16) cannot be replaced by <. It will be shown that < may be taken
in (16) for all forms not equivalent to (17) and (18).
The results (16) were not first obtained by reduction arguments.
DAVENPORT (1945a, b) has however given simple proofs by such
arguments.
This treatment consists in defining a binary cubic form as being
reduced if a certain definite quadratic form associated with it is reduced:
it is necessary to choose different quadratic forms according as D> 0
or D < O. DAVENPORT then shows for a reduced form that either (16)
is true with strict inequality for one of a prescribed set of u, or I (x)
is one of the forms (17), (18). We give the proof for D>O in full but
only sketch that for D < 0 since we shall later be using the case D < 0
to illustrate another technique.
It was shown by DAVENPORT (1941 b) that neither the 49 nor the
23 is isolated. We do not give the proof, which depends essentially on
the fact that although a cubic form I{x) is always indefinite the area
of the region
I/(x)1 < 1
is finite, and the forms (17), (18) take the values ± 1 only a finite number
of times: in contrast to the situation with indefinite quadratic forms.
11.5.2. In order to enunciate DAVENPORT'S result we must first
introduce a quadratic form associated with a cubic form

I (Xl' X 2) = a X~ + b X; X 2 + C Xl X~ + d X~ (1 )
= II ({)jX 1 + "p;x 2), (2)
1:::>;:::>3

(3)
(4)
where
 Binary cubic forms 55

On evaluating the partial differentials by (2), a brief calculation shows

that
(6)
the sum being taken over all cyclic permutations of 1,2,3.
We now show that the hessian is a covariant of the form I(xl , x2 );
that is if oc, {3, y, b are real numbers with
ocb-{3y=±1, (7)

then the hessian of the form j'(xl , x 2 ) defined by

IS
h' (Xl' X 2) = h (oc Xl + {3 X 2 , YXl + b X 2) .
Indeed this follows at once from (6) and the expressions (7), (8) of § 5.1,
on noting that
{);tp~ - {)~ tpi = (oc b - {3 y) ({)itpk - {)k tpi) = ± ({)itpk - -Ok tpi) ,
on using (7).
From either (5) or (6) we see that the determinant of h(xl , X 2 ) IS

(8)
In particular, h (Xl' X 2) is definite when and only when D> 0, i.e. when
1is a product of three real linear forms [when the {)i' tpi are real the form
(6) is clearly positive definite, but the converse is not so clear without
using (8) J.
When the {)i' tpi are real, the form 1 was said by HERMITE to be
reduced when the definite quadratic form h is reduced in the sense of
MINKOWSKI I .

Every form with real {)i' tpi is equivalent to a reduced form. For
the transformation which reduces the h (x) in MINKOWSKI'S sense also
reduces 1(x) ; since h (x) is a covariant of 1(x), as we have seen. Further,
this reduction can be carried out in only a finite number of ways since
we saw that a definite quadratic form can be reduced by only a finite
number of transformations.
II.5.3. We may now enunciate and prove DAVENPORT'S theorem:
THEOREM VIII. Let 1(x) be a binary cubic lorm with discriminant
D>O which is reduced in the sense 01 HERMITE (§ 5.2). Then

min{lf(1,o)l, 1/(0,1)1, 1/(1,1)1, If(1,-1J1}~(~t. (1)

1 He could not put it this way, of course!
S6 Reduction

The sign 01 equality is needed only when

±/(xl • ±x.) =x~ +X~X2- 2xlx~-x1 (2)

OT
(3 )
the ±signs being independent.
DAVENPORT actually proved that if I(x l • XI) is reduced. then at least
one of the five products
1/(1.0)/(0.1)1. 1/(1.0)/(1.1)1. 1/(0.1)1(1.1)1.
1/(1.0)/(1.-1)1. 1/(0.1)/(1.-1)1
is ~ (D/49)i. with equality only for the forms (2) and (3). as before.
We shall follow CHALK (1949) and prove another generalisation. Let

h (Xl' XI) = A x~ + B Xl XI + C x~
be the hessian of 1(z). so that
O~B~A~C. A>o (4)
CHALK'S result is that

min{l/(1.o)l. 1/(0.1)1. 1/(1.1)1. 1/(1. -1)I}~ ( .-d .

A '6

with equality only for the forms (2) and (3). Since 4AC-BI~3Aa.
and 4AC-BI=3DU) by (8) of § 5.2. this will be a stronger result
than Theorem VIII.
We may suppose by homogeneity that

A =7. (5)

We must then deduce a contradiction from

1/(1.0)1~1. 1/(0.1)1~1. 1/(1.1)1~1. 1/(1.-1)1~1.

except for the forms (2) and (3). On writing

1(z) = a x~ + b x~ XI + CXl X: + d X:
these inequalities are
lal~1, Idl~1. (6)

Ia + b + C + dl ~ 1. Ia - b + c - dl ~ 1. (7)
By taking -I for 1 we may suppose that
a~1. (8)
 Binary cubic forms 57

We shall require the identities

A=b2-3ac, B=bc-9aa, C=c 2 -3ba, (9)
from which follow
Bc-Cb=3Aa, Bb-Ac=3Ca. (10)
From (4), (5) and (9) we have
0;;;; bc - 9aa~ 7. (11 )

Suppose, if possible, that a>O. Then (11) gives

bcG9aaG9. (12)
If bGc>O there is a contradiction with (101) and if cGb>O there is
a contradiction with (102), so
b<O, C<O.
Then we should have
A=b2 -3 a c
=l b I2 +il ac l +llacl
G H£a 2 b2 c2 )t
by the inequality of the arithmetic and geometric means; and so, by
(12),

in contradiction with the normalization A = 7.

Hence we may suppose that
d<O,
and so, by (11),
b c ;;;; 7 - 9a 1al ;;;; - 2. (13)
If b<O<c we have a contradiction with (102), so

c<O<b,
and (13) becomes
(14)
Further, (5) becomes
7=A = b2 + 3a 1c1G b2 + 31 cI. (15)

On substituting (14) in (15), we have

7 G b2 + 31 c1~ b2 + 6/b,
and so
(16)
58 Reduction

Similarly we have
7~~+3Icl,
c
and so
1~ - c~ 2. (17)
Clearly a sign of equality can hold in (16) or (17) only if
a=-d=1, bc=-2. (18)
From (14), (16) and (17) we now have
9aldl~7+lbcl~11
and so
...--11
a~­
- 9 ' I dl~~
- 9 .
But now
11 11
a-b+c-d~--1-1+-<1
- 9 9'
and so
a-b+c-d~-1. (19)
We now consider the two possibilities for 1(1, 1). If
a+b+c+d~-1, (20)
then on adding (19) and (20) we have
a-lcl~-1, so Icl~1+a~2.
Comparison with (17) shows that IcI = 2, and, since there is equality
in (17), we must have (18); that is
a=-d=1, b=1, c=-2.
Similarly, if
a+b+c+d~+1,
then
b+d~+1, so b~2;
and we have
a = - d = 1, b = 2, c= - 1.
This concludes the proof of the theorem.
11.5.4. When the binary cubic form 1 has discriminant D(/)<O the
hessian form is indefinite, and so a reduction of the hessian does not
single out a finite number of reduced forms from amongst the forms
equivalent to I. However, if D<O then only one of the linear factors
of 1 is real, and 1 may be put in the shape
l(x I , x 2) = (-&aXl + 'f'aX2)(P x~ + QX I X 2 + R x~), (1)
 Binary cubic forms 59

where the form P x~ + QXl X 2 + R x~ is positive definite, since it is the

product of two conjugate forms with complex coefficients. DAVENPORT
following earlier workers calls such a form reduced if the quadratic
form
P x~ + QXl X 2 + R x~
is MINKowsKI-reduced, that is
IQI~P;::;;;R, (2)
and, further,
(3)
The last condition may be achieved by changing the sign of X2 if need
be, which does not affect (2). DAVENPORT (194Sb) proves
THEOREM IX. I II (x) is binary cubic lorm with discriminant D (f) < 0,
then there are integers u =j= 0 such that

I/(u)I;::;;;I~ll.
II, lurther, I (x) is reduced, then

min[I/(1,O)I, 1/(0,1)1, 1/(1, -1)1, 1/(1, - 2)IJ;::;;; I~ It,

with equality only when

1(Xl' X 2) = a (X~ + X~ X 2 + 2 Xl X~ + x~) .
We only sketch the proof and refer to the original memoire for the
details. We later give another proof of the first paragraph of the
theorem (Chapter III, Theorem VII).
We have to show that D (f) ~ - 23 when

1/(1,0)1;;;; 1, 1/(0,1)1 ~ 1,
1/(1, -1)1 ;;;;1, 1/(1'-f)I~1,
i.e. when
PI1fal;;;;1, RIV'al;;;;1, (~)
l1fa - V'al (P - Q + R) ~ 1, (4~
l1fa - 2V'al (P - 2 Q + 4R) ;;;; 1, (4a)

since P-Q+R, P-2Q+4R are positive by the positive definiteness

of the quadratic form. For fixed 1fa and V'a, the inequalities (2) and (4)
restrict the point P, Q, R in 3-dimensional euclidean space to lie in a
certain infinite region !/' bounded by planes. DAVENPORT shows,
further, that
- D(f) = {PV'~ - Q1faV'a + RV'~}2(4P R - Q2),
60 Reduction

and that ID(f)il is a convex function of (P, Q,R) for fixed {}a,"Pa'
Hence the maximum of DU) is attained at the vertices of g, where
three of the plane faces meet [since it is easily seen that ID I~ 00 as
max (I PI, IQI, IR I) ~ 00]. The proof then follows from a rather tricky
estimation of DU) at the vertices of g.
II.6. Other forms. We briefly survey here results on the reduction
of forms other than those already discussed.
II.6.2. For binary forms of degree n~ 4 there is more than one
invariant. For example, a binary quartic form 1(Xl' X 2 ) which is the
product of two pairs of complex conjugate linear forms may be reduced
to the shape

where
XI =OCXI +{JX 2 , X 2 =yxl +bx2 ,
for some real (I., {J, y, band ft =ft (/) is a real number lying in

Iftl <i·
Two forms with different ft cannot be transformed into each other by
a homogeneous linear transformation of the variables. Further, ft (/) is
an absolute invariant in the sense that ft (t I) = ft (I), where t is any num-
ber. Of course we still also have the discriminant

where
1(Xl' X2) = II (1J;Xl + "P; x 2)·
i
The problem for definite binary quartics was solved independently
by DAVIS (1951 a) and CERNY (1952a) in the sense that they found the
best possible function y (ft) of ft such that every form 1 with invariant ft
has
inf 1(u) ~ y (ft) {D U) }A.
"*,o
integral

DAVIS (1951 a) also gives some results for indefinite binary quartic
and full references to earlier work. It is no longer true, as it was for
quadratic and cubic forms, that forms 1with D (/) = 0 assume arbitrarily
small values. This case was completely elucidated by DAVENPORT
(1950a).
The methods of these authors combines reduction techniques with
other tools drawn from the geometry of numbers.
There does not seem to be any systematic work on binary forms of
degree greater than 4.
 Other forms 61

II. 6. 3. The only other types of forms I (Xl' .... xm) of degree n with
m> 2. n> 2 for which the best estimate of
M(I) = inf If(u)1
"*0
integral

is known appear to be the ternary cubic forms with real coefficients

which are expressible as the product of three real linear forms:
I (Xl' X2• Xa) = IT ({Jjl Xl + {J,2 Xa + (Ji3 Xa).
1;:;;;;:;;3
where either all the {Jih are real (first type) or {JSl' {Jaa. {Jaa are real and
{Ju=;ou (1 ~ k~ 3). There is an invariant

This is the only invariant in each type. since there are obvious real
transformations taking I into

and
Xl(X~ +X=).
respectively. The two types are distinguished by D>O and D<O
respectively. The following two results are known:
THEOREM X. Let I(xl • X2 • xa) be a lactorisable ternary cubic lorm with
D (f) > O. Then there exist integers u 4= 0 such that
DO
I I (u)1 < 9.1'
except when I is equivalent to a multiple 01 one 01 the lorms
149 = x~ + x~+ x~- x~ x 2 + 5x~ x3 -2 Xl X~ + 6 Xl X:-2 x2 x:- X: X3 - Xl X. X3•
lSI = X~+ X~+ ~+6x~xa-3 Xl X~+9Xl X:-3 X2 X:-3 Xl X2 Xa.
lor which M(f) = 1 and D (I) = 49. 81 respectively.
THEOREM XI. Let I (:I)) be a lactorisable ternary cubic lorm with
D (f) < O. Then there exist integers u 4= 0 such that

II (u)l ~ I~ Ii .
The sign 01 equality is needed when and only when 1(:1)) is equivalent to
a multiple 01 the lorm
123 = x~ + x~ + xg + 2x~xa - XIX~ + XIX: - xax: - 3X1 X2 X3 •
We note that 140.181 and 123 are all of the shape
Norm (Xl + tpXa +tpx3).
62 Reduction

where 1, tp, VI are a basis for the integers of a cubic field. We shall
discuss later the reasons why this might have been expected (Chapter X).
For 149,/81 and 123 we have VI =tp2, and tpsatisfies the respective equations:

tp3 + tp2 - 2tp - 1 = 0,

tp3-3tp-1 =0,
and
tp3-tp-1=0.

[By Norm is meant the product of the three forms obtained from the
given one by inserting the three pairs of conjugate values for tp and VI.]
The first equation here corresponds in an obvious way tQ the form in
Theorem VIII. The third equation here corresponds to the binary form

which is equivalent to that in Theorem IX on making the substitution

X1-+X1 , X 2 -+-X1 - X 2 ·
For D>O Theorem X gives the first two successive minima and
shows that the second minimum is isolated. The first minimum in
Theorem XI is not isolated; but there is a weaker sense in which it is
isolated [DAVENPORT and ROGERS (1950a, especially Theorem 14): see
also Chapter Xl Theorem X was obtained by DAVENPORT (1943 a). He
had already obtained the first minimum [DAVENPORT (1938 a) and a
simpler proof in DAVENPORT (1941 a)]. A slightly weaker form of
Theorem XI in which ID/23 Ii + e with arbitrarily small e> 0 appears
instead of ID/23 I! was given by DAVENPORT (1943 a); the full form is in
DAVENPORT and ROGERS (1950a). CHALK and ROGERS (1951 a) showed
that every factorisable ternary cubic form with D> 0 is either equi-
valent to a multiple of 1 or to a form g (;.c) with

Ig(1, 0, 0) g(O, 1, 0) g(O, 0, 1)1 ~

Di)3
(7.1
This is analogue of the results about the products of the diagonal terms
of definite quadratic forms obtained in § 3.
We do not prove Theorems X and XI here, since in Chapter X,
following MORDELL, we deduce Theorems X, XI from the corresponding
results for binary cubics (in which, as the reader will have noticed, the
integers 49 and 23 also occur). It is however worth sketching the
reduction which DAVENPORT use to prove Theorem X:
Let /(z) be a factor is able ternary cubic with D > 0, where we may suppose,
without loss of generality for our purpose that M(f) = 1. Hence t is equivalent
to a form g such that
g(1, 0,0) = (1 -7}t1 ,
 Other forms 63

where 7] ~ 0 is arbitrarily small. Hence we may write

3
(t - 7]) g(:z:) = II (Xl + r1.i Xa + PjX3),
;=1
We consider also the quadratic form

h (:z:) = ~ (Xl + r1.i Xa + Pi xa)2.

i
From the inequality of the arithmetic and geometric means h (u) ~ 3 (1 - 7])i for
u*'
all integers 0, and it is easy to verify that in fact h (u) ~ 3, with equality only
when U= (1, 0, 0). Hence h(:z:) may be reduced in the sense of MINKOWSKI by
a transformation of the type
Xl ->- Xl + Vn X 2 + V13 Xa
X2 ->- V a2 X 2 + V2a Xa

Xa ->- V32 X 2 + V33 Xa

where the vii are integers and V Z2 V33 - v2av32= ± 1. Since h(:z:) has determinant
(1 - 7])8 D(f) and is reduced, we have bounds for the coefficients. The proof now
continues by an intricate and delicate chain of computations using these bounds
and the fact that Ig(u)1 ~t for all integers u*,o.
DAVENPORT'S treatment of Theorem XI starts off with a similar reduction but
the completion of the proof requires different ideas and the detailed consideration
of an intractable 2-dimensional figure.
11.6.4. The corresponding problem for the product of n> 3 homo-
geneous forms in n variables has been much worked on. Estimates
but no precise results are known, and these estimates were obtained
by other methods. We shall consider the case of large 1t in Chapter IX,
§ 8. The best estimates for n = 4, 5 in print appear to be those of
ZILINSKAS (1941 a) and GODWIN (1950a) respectively; but GODWIN refers
to a better estimate for n = 4, presumably the Vienna dissertation of
G. BOHM (1942) also mentioned in KELLER'S encyclopedia article
[KELLER (1954a)] but unavailable to me.
There is however a striking result of CHALK on the product of. the
values taken by n linear forms when these values are positive. He
shows that if L I , ••. , Ln are n linear forms in n variables ;E = (Xl' ... , xn)
with determinant LI =1= 0, then there exist integers u =1= 0 such that

Li(U) > ° (1 ~i~ n), (1 )

II Li(u)~ILlI· (2)
;

That the implied constant 1 on the right-hand side of (2) is the best
possible is shown by the simple example Li = xi' CHALK'S theorem is
indeed more general than the form given here since it refers to the
product of inhomogeneous linear forms. Consequently we do not prove
it here, but later in Chapter XI, § 4.
64 Theorems of BLICHFELDT and MINKOWSKI

Chapter III

Theorems of BLICHFELDT and MINKOWSKI

111.1. Introduction. The whole of the geometry of numbers may

be said to have sprung from MINKOWSKI'S convex body theorem. In
its crudest sense this says that if a point set f/ in n-dimensional euclidean
space is symmetric about the origin (i.e. contains -:E when it contains :E)
and convex [i.e. contains the whole line-segment

when it contains :E and y] and has volume V> 2", then it contains an
integral point u other than the origin. In this way we have a link
between the "geometrical" properties of a set - convexity, symmetry
and volume - and an "arithmetical" property, namely the existence
of an integral point in f/. Another form of the same theorem, which
is more general only in appearance, states that if A is a lattice of
determinant d (A) and f/ is convex and symmetric about the origin,
as before, then f/ contains a point of A other than the origin, provided
that the volume Vof f/ is greater than 2"d (A). In § 2 we shall prove
MINKOWSKI'S theorem and some refinements. We shall not follow
MINKOWSKI'S own proof but deduce his theorem from one of BLICH-
FELDT, which has important applications of its own and which is
intuitively practically obvious: if a point set at has volume strictly
greater than d (A) then it contains two distinct points :El and :E2 whose
difference :El -:E 2 belongs to A
The theorems of BLICHFELDT and MINKOWSKI may be regarded as
statements about the characteristic functions of a set f/, that is the
function X(:E) which is 1 if :E E f/ but otherwise O. There are generalisa-
tions of the theorems of BLICHFELDT and MINKOWSKI to non-negative
functions ""(:E) due to SIEGEL and RADO. These we present in § 3.
We do not in fact use these theorems later.
In § 4 we use MINKOWSKI'S theorem to obtain a characterisation of
a lattice which is independent of the notion of a basis: a lattice is any
set of points A in n-dimensional space which (i) contains n linearly
independent vectors, (ii) is a group under addition, i.e. if :E and y
are in A so are :E ± y, and (iii) has only the origin in some sphere
x~ + ... + X~<1J2, where 1J> O.
In § 5 we introduce the notion of the lattice constant LI (f/) of a
set f/. This is a number with the property that every lattice A with
d(A)<LI (f/) has a point other than 0 in f/, while there are lattices
whose determinant d (A) is arbitrarily near to LI (f/) with no other
point than 0 in Y. In § 6 we discuss at length a method due to MORDELL
 Introduction 65

which uses MINKOWSKI'S convex body theorem to evaluate or estimate

A (9') for sets which mayor may not be convex. The idea is, roughly
speaking, to show that if a lattice 1\ of given determinant d (1\) = Ao
has no points except 0 in 9', then at least 1\ must have paints in various
sets abutting on 9'. Since these points belong to 1\, so do linear com-
binations of them. These combinations must be either 0 or lie outside 9'.
In this way more and more information about these points of 1\ near
9' is obtained, until there is a contradiction; the contradiction showing
that every lattice 1\ with determinant d(/\) =Ao has a point in 9'.
This method is particularly effective in 2 dimensions, since the relation-
ship of the various points to each other then springs to the eye. Con-
sequently in § 6.2 we give a series of simple lemmas about 2-dimensional
lattices which are non-the-Iess useful tools. MORDELL'S method is
applied, amongst other things, to finding A (9') when 9' is the region
IX~ +X~I <1. (1)
This is equivalent to finding the lower bound of the values taken by
a binary cubic form with negative discriminant. This question was
discussed but not answered in Chapter II. The proof given here is ~
conflation of several given by MORDELL. It uses essentially the algebraic
background. We remark in passing that MORDELL (1946a) has shown
that the result obtained generalizes to all regions which look sufficiently
like (1). Similarly, BAMBAH (1951a) has proved a result to show that
all sets which look sufficiently like
(2)
do, in fact behave like (2). The set (2) corresponds to binary cubic
forms with positive discriminant in the same way as {1} does to those
with negative discriminant. For example BAMBAH'S result applies to
regions 9' with hexagonal symmetry and six asymptotes at angles :rt/3,
the set of points between two asymptotes which do not belong to 9'
being convex. Compare Chapter X, § 3.3.
Finally, in § 7 we use MINKOWSKI'S theorem to obtain some results
about the representations of numbers by quadratic forms; for example
that every prime p = 4m + 1 can be expressed as the sum of the squares
+
of two integers; p = u~ u~. This is all rather aside from the main
theme of the book but the proofs are so elementary and so striking
that they deserve to be better known.
III.l.2. It is convenient to introduce here some important defini-
tions and notions.
The length of a vector :ll = (Xl' ... , x n ), namely
(x~ + ... +x!)l
Cassels, Geometry of Numbers 5
66 Theorems of BLlCHFELDT and MINKOWSKI

will, as usual, be denoted by

It satisfies the "triangle inequality"

for all vectors z and y. The length o~ a vector is not an invariant under
all unimodular transformations, unlike most of the concepts we work
with, but we shall be concerned only with the topology induced by the
metric Izl and not the metric itself. Let
(1~i,i~n) (1)
be a real transformation of determinant
(2)
Clearly
lyl2 = ~ (~>XijX;t;£ n3A2 'Lx: = n3A21 z 12,
where
A = max IOCi;l.
Since det(oci;)=t=O, we may solve (1) for the xi and obtain, say,

X,='LPiiYi'
;
(3)
Then similarly
IzI2~n3B2IyI2,
where
B = max IPiil.
Hence there exist constants cl • c2 independent of z and y such that 1

(4)

We shall often make use of the following consequences without

explicit reference.
LEMMA 1. Let A be a lattice in n-dimensional space. Then there exist
constants 'YJl. 'YJ2 depending only on A with the following properties
(i) If uEA, vEA and lu-vl<'YJl' then u and v are identical:
(ii) The number N(R) of points of A in a sphere Izl <R is at most
1/2 (R" + 1).

Both of these statements are trivially true for the lattice Ao of

points with integer coordinates. But now (d. § 3 of Chapter I) if A
1 This is a particular case of a result to be proved later (Chapter IV, Lemma 2
Corollary).
 Introduction 67

is any lattice with basis

bj=(fJli"",P"i) (1~j~n),

then the points of A are just the points (3) with Y EAo. The truth of
(i), (ii) in general now follows at once from (4) and the truth of (i), (ii)
for Ao.
111.1.3. We say that a sequence of vectors J!, (r = 1,2, ... ) con-
verges to the vector J!' as limit if
lim IJ!, - J!'I = 0
in the usual sense. Clearly a necessary and sufficient condition for this
is that the co-ordinates of J!, should converge to the corresponding
co-ordinates of J!', since clearly
maxlxil ~IJ!I ~nimaxlxil
for any vector J! = (Xl' ... , X,,). An immediate consequence of Lemma 1
ci) is that a sequence of vectors u, of a lattice A can converge only if u,
is the same for all sufficiently large r, say
u,=u'
A set [I' of points is said to be compact if every sequence of points
J!,t [I' contains a subsequence Y. =J!" (rl <r2 <···) which converges to
a limit in [1':
lim y;
5-+00
= Y' E f/ .
A classical theorem of WEIERSTRASS states that a set [l'in n-dimensional
euclidean space is compact if and only if it is both bounded (i.e. con-
tained in a sphere IJ! I< R for some sufficiently large R) and closed
(i.e. if J!, E f/ (1 ~ r < (0) and J!' = lim J!, exists, then J!' E f/).
For the sake of completeness we give a proof of WEIERSTRASS'S theorem.
Suppose first that .9' is a compact set. If.9' were unbounded, we could find a
sequence of points iX,E.9' such that JiX,J- 00, and then it clearly cannot contain
a convergent subsequence. Hence a compact set .9' is bounded. If.9' were not
closed, we could find a sequence of points iX,E.9' such that lim iX, = iX' is not in .9'.
Clearly every subsequence of the original sequence tends to IX'. Hence a compact
set .9' is closed. Now let .9' be a set which is both bounded and closed. We shall
show that .9' is compact. Let iX, (1 ~ I' < 00) be a sequence of points of .9'. We
may suppose that originally all the iX, are contained in a n-dimensional cube ~o
of side 2R for some R. This cube may be dissected into 2" cubes of side R by
taking planes through the centre of 'to parallel to the faces. For definiteness we
take the cubes of side 'to to be closed, that is to include their boundary points.
At least one of the cubes of side R must contain iX, for infinitely many 1'. Let'tl
be one of these. On repeating the original process with 'tl instead of 'to we obtain
t
a cube 't2 of side R contained in 'tl which contains iX, for infinitely many r.
And so on. In this way we obtain a sequence of cubes 'ts (0 ~ S < 00)' of side
5·
68 Theorems of BLICHFELDT and MINKOWSKI

21 -. R. such that ilS+1 is contained in il.. Each ils contains ;E, for infinitely
many". The cubes '1/. define a point ;E' which is contained in all of them. We may
now find a subsequence ;E" tending to ;E' as follows: ;E" is any point of the original
sequence in ilo: if "1' ...• ". have already been fixed with

then ".+1 is anyone of the infinitely many indices" >". such that ;E, is in ils '
Finally. since
;E' = lim ;E, •
s~oo '

the point ;E' is in Y, since Y is assumed closed.

There is a form of WEIERSTRASS' Theorem which is apparently
more general. Let
~", (1~k~m, 1~r<oo)

be a sequence of sets A, of m points ~", in a compact set [/. Then there

is a increasing sequence r l <r2 <··· of integers such that all the limits

lim~.h
s-+oo '
exist and are in [/. For if

~", = (Xl"""" X"",),

the sets A, may be represent by points X, with coordinates x;kr
(1 ~i~n, 1 ~k~m) in nm-dimensional space. Clearly the set [/m of
points X = (xi") with
(xu, ...• XII") E: [/ (1 ~ k ~ n)

is bounded and closed if [/ is. Hence the points X, have a convergent

subsequence X", Then the rs clearly do what is required.
[Alternatively one could make use of the so-called diagonal process. First
pick out a subsequence

of the A, such that Yl. is convergent. Then pick out a subsequence C,= (zu •...• Zmt)
of the 8. such that z.,
is convergent. The sequence Zu is also convergent. being a
subsequence of the convergent sequence YU' And so on. After m repetitions of
the process one obtains the required subsequence.]
111.1.4. By volume we shall mean in this book LEBESGUE measure
unless the contrary is stated. We shall however have no need of any
of the more recondite properties of measure; the sets we shall be mainly
concerned with have a volume by any definition. for example the
interiors of cubes or ellipsoids.
111.2. BLICHFELDT'S and MINKOWSKI's theorems. We use the no-
tation and results of Chapter I. To BUCHFELDT is due the realization
 BLICHFELDT'S and MINKOWSKI'S theorems 69

that the following almost intuitive result forms a basis for a great
portion of the geometry of numbers [BLICHFELDT (1914a)).
THEOREM I. Let m be a positive integer, A a laUice with determinant
d(A), and 9' a point-set 01 volume V(9'), possibly V(9') = 00. Suppose
that either
V(9') > md(A), (1 )

V(9') = md (A) (2)

and 9' is compact. Then there exist m + 1 distinct points ~l' ... , ~"'+l
01 9' such that the dillerences ~i -~; are all in A
Let b l , ... , b" be any basis of A and .let f1J be the generalized paral-
lelopiped of points
y1b1 + ... +y"b"
Then f1J has volume
(3)
Every point ~ in space may be put in the shape
~=u+1', uEA, 1'Ef1J,

and this expression is unique, since the points of A are just the
y1b 1+ .. · + y"b" , where Yl' ... , y" are integers.
This parallelopiped f1J will play an important part later (Chapter VII),
where it will be called a fundamental paral1elopiped for A
For each 11, E A let al(u) be the set of points v such that
1'Ef1J. v +UE9'.
Clearly the corresponding volumes V{al(u)} satisfy
~ V {al (un = V(9'). (4)
II

Suppose now that the first alternative holds, namely V(9'»md(A),

so that (4) implies
LV {al (u)} > m d (A) = m V(f1J).
II

Since the al(u) are all contained in f1J, there must be at least one
point 1'0 Ef1J which belongs to at least m + 1 of the al(u), say
1'oE: al(u;) (1 ~ i;;;; m + 1),
where the 11,; are distinct. Then the points
~;=1'o+u;
70 Theorems of BLICHFELDT and MINKOWSKI

are in !/ by the definition of 9t(u), and

Zi-Zj=Ui- j {~o u -r (i=j=j).

This proves the theorem for the first alternative.
Suppose now that the second alternative holds. Let s, (1 ;:;:;;r< 00)
be a sequence of positive numbers and

lim s, = O.

For each r, the set (1 +s,)!/ of points (1+s,)z, ZE!/ clearly has
volume

Hence, by what we have already proved, there exist points

Zj,<=(1+B,).9' (1;:;:;;j;:;:;;m+1)
such that

u.(i,i) (say) = Zi, - zi' {~/\o (i =j= i). } (5)

By extracting suitable subsequences of the original sequences, and then

calling them B" Zj, again to avoid introducing new notation, we may
suppose, without loss of generality, that

all exist. Since!/ is now assumed to be compact, the z; are in .9'.

Then, by (5),
Z: - z;
= ,_00
lim u,(i,j).

But now the u,(i, i) are in I\. Hence (d. § 1.3) u,(i, i) is independent
of r from some stage onwards:
U,(i,i) = u'(i,i)
Hence
z·, -z·, = '(..) {E /\
U t }
• 1 ,7 =l= 0 (i =l= j), '
as required.
For later reference (Chapter VII) we note that in the proof for the
first alternative we have implicitly proved the following:
COROLLARY. Let.9' be any set 01 points and let 9;. be the set 01 points
v of the fundamental parallelopiped which can be put in the shape
V=Z-U, zE.9', UE/\.
 BLICHFELDT'S and MINKOWSKI'S theorems 71

Then
V(~) ~ V(9").
If no difference Xl - x 2 between distinct points of 9" belongs to A then
V(~) = V(9").
The first paragraph is clear. The second follows since then no two
9l(u) overlap.
111.2.2. From Theorem I we deduce almost at once the following
theorem which is due, at least! for m = 1 to MINKOWSKI ("MINKOWSKI'S
convex body theorem").
THEOREM II. Let 9" be a point set of volume V(9") (possibly infinite)
which is symmetric 2 about the origin and convex 2 • Let m be an integer
and let A be a lattice 0/ determinant d(A). Suppose that either

V(9") > m 2"d(A),

or
V(9") = m 2"d(A)
and 9" is compact. Then 9" contains at least m pairs of points ± u i
(1 ~i~m) which are distinct from each other and from o.
Again we note that the possibility of infinite volume is not excluded.
Theorem I applies to the set -l9" of points -l X, X E 9" which has
volume rn V(9"). Hence there exist m + 1 distinct points

(1~i~m+1),
such that
1
z:l:; -
1
z:l:i
{EA
=+= 0 (i =+= i)
}.
We introduce an ordering of the real vectors and write

if the first non-zero component of :l:1-X2 is positive. We may suppose

without loss of generality that

Put

Then clearly

1 The general case is apparently due to VAN DER CORPUT (1936a).

2 For the definition of these terms see § 1.1.
72 Theorems of BLICHFELDT and MINKOWSKI

are all distinct. But - X m + I E f/ since X m + IE f/ and f/ is symmetric.

Hence
U; = ix; + !(-xm+!) E f/
by the convexity of f/. This proves the theorem.
For later use we note the
COROLLARY. Let f/ be symmetric about the ongm and convex.
A necessary and sufficient condition that f/ contain a point of /\ other
than 0 is that there exist two distinct points ! Xl' ! X 2E! f/ whose differ-
ence !XI -!X2 is in f/.
If f/ contains the point a E /\ then! f/ contains the two points !a
and - ! a whose difference is a; which proves part of the corollary.
Conversely, as in the proof of the theorem, if ! Xl' ! x2 are given, then
!xI - !x 2 is in f/.
Theorem II is the best possible of its kind for any m. For example
the convex symmetric set
(2~j~n),

has volume m 2" but contains only m - 1 pairs of points of the lattice /\"
of integral points other than 0 namely
± (u, 0, ... ,0) (1~u~m-1).

We shall return in Chapter IX to the general problem of finding

convex symmetric sets of volume 2" d (/\) which do not contain any
lattice points other than the origin.

111.2.3. Important examples of a convex symmetric point set are

those sets f/ defined by a set of inequalities of the type
(1~l~L),

where the ali are real or complex numbers. Such a set is clearly sym-
metric. It is also convex, since if x, yare in f/ and

(O~).~1),
then clearly

I~ al;z;1 ~ ).I~ al;x;1

1
+ (1 - ).) I~ al;y;1 ~ max {I~ ali
1 1
x;I' IL ali y;l} .
1

For sets f/ of this kind one can relax the condition of compactness
in Theorem II somewhat. We enunciate the theorem for the most
important case when the ali are all real. It will be observed that the
argument might be used for a wide class of convex sets Y.
 Generalisations to non-negative functions 73

THEOREM III. Let A be an n-dimensional lattice of determinant d (A)

and let aij (1~i,i~n) be real numbers. Suppose that l Cj>O (1~i~n)
are numbers such that
CI • •• C" ~ Idet (aij)i d (A) . (1 )
Then there is a point U EA other than 0 satisfying
ILaljUjl~CI
IL a,iujl < ci (2 ~ i ~ n). } (2)

Suppose, first, that

det (aij) =f: O.
Then (d. Chapter I, § 3) the points X = (Xl' ... , X,,) defined by
Xi = L aijxj XE A
j

form a lattice M of determinant

d (M) = Idet(aij) I d (A) . (3)
The inequalities (2) become
IXII ~ CI
IXil <ci (2 ~ i~ n). } (4)

These define a set !/ in the space of X of volume 2" ci ... c". Hence
if there is strict inequality in (1) the theorem follows from the first
alternative in Theorem II. Let now 8 be any number in
0<8<1.
Even if there is equality in (i), there is certainly a point X.E Mother
than 0, with co-ordinates (X lt , •..• X".), such that
IX Ie I ~ CI + 8 < CI + 1
IXi.I<Ci (2~i~n).
But now there are only a finite number of possibilities for X., by
Lemma 1 (ii). Since 8 is arbitrarily small, one of those possibilities
must therefore satisfy (4). This proves the theorem unless det(aij) =0.
But then it is readily verified that (2) defines a region of infinite volume.
and so Theorem II certainly applies.
111.3. Generalisations to non-negative functions 2• The results of
§ 2 may to some extent be generalised to non-negative functions ",(x)
1 Cj>0 follows from (1) except when det(aij) = o. But we do not exclude this.
2 The results of § 3 will not be used later.
74 Theorems of BLICHFELDT and MINKOWSKI

of a vector variable~. We suppose that tp(~) is integrable and write

V(tp) = f tp(~) d~, (1 )
- 00 <x,<oo
where
d~ = dx1 ... dx".
This notation is justified, since if tp is the characteristic function of a
sct Y, that is.
tp (~) = {1 if ~ E Y } (2)
o otherwise,
then V(tp) is just the volume V(Y) of Y.
We now have the following simple analogue of BLICHFELDT'S
Theorem I:
THEOREM IV. Let tp (~) be a non-negative integrable junction and let
A be a lattice oj determinant d (A). Then there is certainly a point Vo
such that
d(A) ~ tp(vo + u) ~ V(tp).
uEII

Before proving Theorem IV we note that it certainly implies the first

alternative form of Theorem 1. For if tp is the characteristic function
of a set Yand V(tp) = V(Y»md(A) for some integer m, then (3) gives
~tp(vo + u) > m,
u
and so
L tp(vo + u) ~ m + 1 ,
u

since now tp (~) is given by (2). But this means that there are m + 1
distinct vectors u i such that v o+ UiE Y, and this is just the conclusion
of Theorem 1.
The proof of Theorem IV follows that of Theorem 1. Let b 1 , .•. , b"
be a base of A, and ~, as before, the set of
y1b1+···+y"b" (O~Yi<1);
so that every ~ is uniquely of the shape
~=v+U, VE~, uEA.
Then
V(tp) =f tp(~) d~
=~ ftp(u+v)dv
uEII vE~

=
VE~
J {~tp(u
uEII
+v)} dv.

Since ~ has volume V(~) =d(A), the theorem now follows at once.
 Generalisations to non-negative functions 75

11.3.2. SIEGEL (1935il) has given a stronger form of Theorem IV which has,
however, remained rather sterile of applications. For notational simplicity we
enunciate it only for the lattice 110 of integral vectors. The function

cp(v) = ~!p(v + U) (1 )
"E A,

is periodic by definition. Its Fourier coefficients c (p) = C (PI' "', Pn), where pE 110 ,
are given by
c (p) = J cp(t,) e- 21Ii (pv) dv, (2)
9
where (p v) denotes the scalar product

PI VI + ... + Pn vn ·
On substituting (1) in (2), we have
c(p) = J !pix) e- 21Ii (px) dx, (3)
- 00 < XI <00
(1;;> i;;> u)

since pu is an integer when pEllo' uEllo . In particular,

Jcp(v)dv=c(o) = V(!p). (4)

9

But now, by a fundamental theorem in the theory of Fourier series,

J cp2(V) dv = ~ [c(p)i2. (5)
9 pEA,

Since cp(u) :2: 0 for all v, there must be some Vo such that

J cp2 (v) dv ~ cp(vo) J cp (v) dv = cp (vo) V(!p). (6)

9 9
On substituting the definition of cp (vo) and the values (3), (4), (5) in (6) we have

L IJI (vo + te) = cp(vo) ~ V(!p) + {V(!p) }-1 L 1f !p (x) e- 2ni (px) dx 12. (7)
"EA, PEA.
P400
This is SIEGEL'S inequality.
When a general lattice II is substituted for 110 on the left-hand side of (7) then
11* must be read for 110 on the right-hand side, where 11* is the polar lattice of II
defined in Chapter I, § 5.
111.3.3. We now give RADO'S generalisation of MINKOWSKI'S convex
body theorem II. [RADO (1946a), see also CASSELS (1947a).] RADO
considered very generally a homogeneous linear mapping A of n-dimen-
sional vector space into itself given by
(1 )
when X=Ax. We write det(A)=det(Aij).
THEOREM V. Let 1p(x) be a non-negative junction oj the vector x in
n-dimensional space which vanishes outside a bounded set, and suppose that

1p (Ax - AY) ~ min {1p(X) , 1p (Y)} (2)

76 Theorems of BLICHFELDT and MINKOWSKI

for all real vectors a: and y. Then

0) +~ '\1 (u);;::: Idet(A)1 V( )

1p ( 2 LJ 1p - d(lI) 1p (3)
"Ell

for any lattice A, where

V(1p) = J 1p(a:) da:.
-00<"1<00
(1 ~i:iO ,,)

Before proving Theorem V we note that it does in fact imply the

first alternative part of Theorem II. Let 1p(a:) be the characteristic
function of a convex symmetric set [1', so that V(1p) = V([I'). For A we
merely take Aa: =ta:, so that det (A) = (t)". The condition (2) is certainly
satisfied, since the right-hand side of (2) is 0 unless both a: and yare
in [1'; and then
Aa: - AY = ta: + t( - y)
is also in [I' by the convexity and symmetry. On the other hand the
left-hand side of (3) is P+ 1, where p is the number of distinct pairs
±uEA in [I' other than o. Hence if V(1p»m2"d(A), we have 1 +p> m,
that is p~ m; which is the conclusion of Theorem II.
To prove Theorem V we need an elementary combinatorial lemma.
LEMMA 2. Given any sequence of distinct vectors

we can construct another sequence

{W} :WO,W1 , ... ,w" ...

satisfying the following three conditions:
(i)wo=o,
(ii) W,=f= ±ws t"f r=f=s,
(iii) every w, is the dillerence between two of the first r + 1 elements
of {z}, say
(4)
We introduce an ordering of real vectors and write

if the first non-zero coordinate of a:1 - a:2 is positive. If a:1 =f= a: 2 then
either a: 1 > a: 2 or a: 2 > a: 1 • We construct w o, ... , w" ... in turn, so that
w,>o (r> 0).
The vector Wo is given. Suppose that W o , ... , W,_1 have already been
constructed, where r;;:;;1. There is a unique permutation Zkl(O~j~r)
 Generalisations to non-negative functions 77

of the vectors Zj (O~i~r) so that

The r vectors
Zkj - Zk. (i = 1, 2, ... , r)
are distinct from each other and from o. Hence we may choose as w,
one of them which is distinct also from the r -1 vectors WI' ... , W,-I'
Since wi> 0 (1 ;;;'i~r) we cannot have w,= -wi' Hence the w,
do what is required.
Theorem V, will be an almost immediate consequence of the following
Lemma.
LEMMA 3. Suppose that (2) holds and that det (A) =!= 0, so that a trans-
formation A-I reciprocal to A exists. Then
l: tp (A -1 U + A-I t) ;;;, tp (0) + l l: tp (u) (5)
UE/\ UE/\
*0
for every real vector t.
For fixed t let z, be the sequence of vectors Z of 1\ such that
tp(A-1 Z+A-1 t»0 arranged so that
tp (A -1 z, + A-I t) ;;:;; tp (A -1 Zs + A-I t) (r;;;' s). (6)
Let w, be the corresponding sequence defined by Lemma 2. We apply
(2) with
x = x, = A-1 Zlr + A-It
Y = Y, = A-1 Z mr + A-It,
where lr and mr are defined by (4). Then
min {tp (X r), tp (Yr)} ~ tp (A -1 Zr + A-I t) ( 7)
by (6), and since lr;;;'r, mr~r. But now, by (4) again,
A(Xr - Y,) = wr
and so, by (2) and (7)
tp(Wr );;:;; tp (A-l zr + A-It).
Similarly, on interchanging x, and Yr' we obtain
tp (- w r) ;;:;; tp (A-1 zr + A-I t) .
Hence, since tp;;:;;O, we have
l:tp (u) ;;:;;tp(wol + l: {tp(W r) +tp(-Wr }}
UE/\ r>O

;;:;; tp (A -1 Zo + A-I t) + 2 l: tp (A Zr + A-I t)

-1 (8)
r> 0

=- tp (A -1 Zo + A-I t) + 2 l: tp (A U + A-I t) ,
-1
u~/\
78 Theorems of BLICHFELDT and MINKOWSKI

since every vector uEA with 1JI(A-1 U+A-1 t) >0 occurs as a z,. But
now (2) with y = x implies that 1JI (0) ~ 1JI (x) for any x, and in particular

(9)
The truth of the lemma follows now at once from (8) and (9).
Finally Theorem V follows from (5) on integrating with respect to t
over a fundamental parallelopiped f!P of A defined as in § 2.1. The
left-hand side becomes

j L~l(A-IU + A-I t)} dt =::ooj;<~(A -1 t) dt = Idet(A)1 V(1JI).

(l"5,i"5, n)

The right-hand side of (5) is independent of t and so, on integrating

with respect to t, is merely multiplied by V(f!P) = d (A). This proves the
theorem.
RADO (1946a) discusses the homogeneous linear transformations A
for which there is a function 1JI (x) which is not identically 0 satisfying
(2). It turns out that A must satisfy pretty stringent conditions, and
that taking multiplication by {- for A is in a sense on the borderline of
what is possible.
111.4. Characterisation of lattices. We are now in a position to
give a characterisation of lattices in which the notion of a basis does
not appear.
THEOREM VI. A necessary and sufficient condition that a set of points
1\ in n-dimensional euclidean space be a lattice is that it should have the
following three properties:
(i) If a and b are in 1\ then a ±b
is in I\.
(ii) A contains n linearly independent points ai' "', all'
(iii) There exists a constant 'fj> 0 such that 0 is the only point of A
in the sphere
Ixl <17,
where, as usual,
Ixl = (x~ + ... + x~)~.
By the definition and Lemma 1 every lattice satisfies (i), (ii), (iii).
It remains to show that any set A satisfying (i), (ii) and (iii) is a lattice.
We note first that it follows by induction from (i) that if c 1 , ••• , c m
are any points of A and u 1 , ••• , Urn are integers, then

Secondly, we show that if

c j =(c 1 j,···,c ni )
 Characterisation of lattices 79

are n + 1 points of /\, then there are integers ui (1 ~ i ~ n + 1) not all 0

such that
LUiCj=O.

For by Theorem II there certainly exist points (u 1, ... , un+!) of the*0

(n + i)-dimensional lattice /\0 of integral vectors in the convex sym-
+
metric (n i)-dimensional set [I' of infinite volume defined by the n
inequalities
I
L Ci1Uil <'YJ/n (1~i~n).
ISjSn+1
Put

so that trivially

Then d = 0 by property (iii), as was required.

Now let Ml be the lattice with the basis ai' ... , a.. given by (ii).
Then Ml is a subset of I\. If /\ coincides with Ml there is nothing to
prove. If not, there is some vector h in /\ but not in MI. But now,
on applying the result of the previous paragraph to the n 1 vectors +
ai' ... , a .. and h, there must be .integers U 1 , ••• , un and v not all 0 such
that
(1 )

* *
Here v 0, since a 1 , ••• , a.. are linearly independent. Further,
v ± 1 since h is not in Ml by hypothesis. We may suppose that b is
chosen so that Ivi in (1) is as small as possible. Let p be a prime divisor
of v and write

Then
PhI = u 1 a 1 + ... + un an,
where not all of u 1 , ••• , un are divisible by p since hI is not in Ml (because
v was chosen minimal). Without loss of generality, p does not divide U I ,
and so
lp -mu 1 =1

for some integers I and m. Put now

a{ = tal - m b 1 }
(2)
a; = aj (2 ~ i ~ n),
so that conversely
a 1 = pa~ + mU2a~ + ... + mu.. a~ }
aj = a; (2~ j~ n) .
Let M2 be the lattice with basis at.
80 Theorems of BLICHFELDT and MINKOWSKI

Then Ml has index p in M2 , so in particular

d(M2) = P-Id(M 1) ~ id(M 1)· (4)
But now, by (2), a basis of M2 is in A and so M2 is entirely contained
in A. We may now repeat the argument. If M2 does not coincide with
A there is a third lattice M3 which is in A and contains M2 as a sublattice.
And so on. Now, by (4),
d (M,) ~ id(M'-I) ... ~ (!),-l d (M I).
If
d (M,) < (rJln)",
where 'fJ is defined in (iii) of the enunciation of the Theorem, then, by
Theorem II, M, would contain a point d =1= 0 with
(1~j~n)

contrary to hypothesis. Hence the chain of lattices M1 , ... , M" ... must
have a last, MR; and MR then coincides with A.
111.5. Lattice constants. We must now introduce a number of new
definitions relating to lattices and points sets. The new concepts will
be subjected to a searching analysis in Chapters IV and V; here we just
prove enough to show their use and to enable applications of MIN-
KOWSKI'S theorem to be made.
Let g> be any point set. If a lattice A has no points in !7 other
than 0 (if 0 is in !7), then we say that A is admissible for !7 or
!7-admissible. We call the infimum (greatest lower bound) of d (A) for
all A-admissible lattices the lattice constant of !7 and write
L1 (!7) = inf d (A) (A is !7-admissible).
If there are no !7-admissible lattices then we say that !7 is of infinite
type, and write L1 (!7) = 00; otherwise !7 is of finite type and
O~L1(!7)<oo. An !7-admissible lattice A with d(A)=L1(!7) is said
to be critical. Critical lattices playa very prominent role in Chapter V.
Of course in general there is no reason why a general set !7 shbuld have
critical lattices at all.
Our definitions do not quite correspond with those of MAHLER
(1946d, e). He is usually concerned with closed sets !7 and says that
A is !7-admissible if no interior point of !7 except 0 belongs to A, that
is if A is admissible in our sense for the set of interior points of !7.
Our usage is a compromise between MAHLER'S and that proposed by
ROGERS (1952a).

111.5.2. The definition of L1 (!7) may be stood on its head: L1 (!7) is the
greatest number L1 such that every lattice A with d (A) < L1 has a point
 Lattice constants 81

other than 0 in Y. The discussion of § 4 of Chapter I shows that many

of the results of Chapter II may be interpreted as giving the value
of L1 (Y) for certain regions Y. Take for example the statement that
if /(;X:)=/l1x~+2/12XlX2+/22X~ is a definite quadratic form and
D=/11/22-/~2' then there are integers u=(U1,U2)=FO such that
/(u) ~(4D/3)!, withequalityonlyforformsequivalentto/l l (x~ + X l X 2 + x~)
lTheorem II of Chapter II). This is equivalent to the statement that
the 2-dimensional set P): X~ + X~ < 1 (1)

has lattice constant L1 (P)) = (Wand that the critical lattices are pre-
cisely those with a base b1=(b l l , b21 ), b 2=(b12 , b22 ) such that
(b 11 Xl + b12 X2)2 + (b 21 Xl + b22 X2)2 = X~ + Xl X2 + x~ (2)
identically. The reader will have no difficulty in making the translation
for himself (d. Lemma 4 of Chapter n. We can also make a geometrical
interpretation of (2l. Put
b11 = cos{), b21 = sin{},
b12 = cos1p, b22 = sin 1p.
Then (2) is true provided that
cos {} cos 1p + sin {} sin 1p = ! ,
that is provided that
-{} - 1p = ± n/3 .
Hence the critical lattice has as basis two points at angular distance
n/3 on +
X~ + X~ = 1. A further point on X~ X~ = 1 is b l - b 2 , as is
clear from (2). It is readily verified that the six points ± bl , ± b 2 ,
± (b l - b 2 ) are the vertices of a regular hexagon inscribed in X~ + X~ = 1.
111.5.3. In this and in the next section we shall use MINKOWSKI'S
convex body Theorem II to evaluate or estimate L1 (Y) for various
sets Y. Theorem II is directly applicable when Y is symmetric and
convex, since it asserts that then
L1 (Y) ~ 2- 11 V(Y). (1 )
This applies for example to the circular discP): X~+X~<1 and gives
L1 (P)) ~ n/4 = 0.785 .... , which may be compared with the exact value
(!)t=0.866 ... obtained above.
Even if our region Y is not convex or symmetric, we may obtain
estimates for L1 (Y) below if a convex symmetric body .:T is inscribable
in it. Clearly L1 (Y) ~ L1 (.:T) if .:T is a subset of Y, since every Y -admis-
sible lattice is automatically .:T-admissible. Hence

Cassels, Geometry of Numbers 6

82 Theorems of BLlCHFELDT and MINKOWSKI

Consider for example the region

This contains the convex symmetric region

ff: IXII + ... + IX.. I < n
by the inequality of the arithmetic and geometric means. Now ff is
convex and symmetric, since it is defined by homogeneous linear in-
equalities, and its volume is

Hence
Ll (Y) ~ n"/n!.

We shall later obtain a rather better estimate than this (Chapter IX,
§ 8). We note the translation into the theory of forms: Let
Lj(~) =L CjiX i
I;:;;.;:;; ..

be real linear forms in the n variables ~ = (XI' ... , x,,) with det (cij)
Then there exists an integral u*o such that
*o.
117 L j (u) I~ :~ Idet (cii )!·
MINKOWSKI'S convex body theorem also permits the evaluation of
Ll (Y) for sets Y which are not symmetric in o. We reproduce here,
with his kind permission, Professor MAHLER'S elegant treatment of the
simplex, hitherto unpublished l . Let Y be an open simplex in n-di-
mensional space containing o. If the faces of Yare given by the equa-
tions
(O~i~n),

where the L j (~) are linear forms, then Y is the set of points satisfying

(O~i~n).

There is one non-trivial relation between the linear forms, say

L I1.j Lj(~) =0
O;:;;i~"

identically in ~, where the I1.j are real numbers, and without loss of
generality
11.0> o.
1 It is given, however, in his mimeographed lecture course, Boulder (Colorado),
U.S.A., 1950, together with other interesting results about non-symmetric sets.
 Lattice constants 83

If, say, Ct.l ~ 0, then Y would contain the infinite ray of points x satisfying

(j=l=O,1);
which is impossible, since Y is a simplex. Hence

(O~j~n).

We may suppose without loss of generality that

Ct. o = 1 = min Ct.j' (2)

1
and then
Lo(x) = - L Ct.jLj(x) , (3)
l~j~n

where
(4)
We show that
(5)
where V('if) is the volume of the parallelopiped
(1~j~n).

In the first place, if 1\ is a lattice with d(I\)<2~nv('if), then there is

a point a=l=o of 1\ in 'if. By taking -a instead of a if necessary, we may
suppose that

and then
(O~j~n) ;

so a is in Y. Hence ,1(Y) 2; rnv('if). On the other hand, we shall show

that the lattice M of points a such that

(1~j~n)

is admissible for Y.
If a is in Y, we must have Uj~O (1~j~n), and minuj~-1 if
a=l=o. But then, by (4), we should have

Lo(a) = - L Ct.j Uj 2; 1;
and so a is not in Y. Hence 0 is the only point of M in Y. Since
d(M)=rnv('if), this completes the proof of (5). We note that
2~n V('~) = Idol ~l, where do is the determinant of the n forms L 1 , .•• , L".
By (3) and (4), do is the least in absolute value of the determinants of
selections of n out of the n + 1 forms L o , ... , Ln.
Estimates of ,1 (Y) for non-convex sets Y may be obtained from
Theorem I instead of Theorem II. Let fJ£ be any set such that all the
6*
84 Theorems of BLICHFELDT and MINKOWSKI

differences
~l - ~2' ~l EfJl, ~2EfJl (6)
lie in !/. Then
LI (!/) ~ V(fJl),
since by Theorem I if d(A)<V(fJl) there exist two points ~I'~2EfJl
such that ~1-~2EA; and by hypothesis ~1-~2E!/. Of course if :Y
is a convex symmetric set inscribed in !/ we could take fJl = i:Y: but
then we get just the same estimate LI (!/) ~ z-" V(:Y) as by the use of
Theorem II. However MORDELL and MULLENDER found suitable sets
fJl in the case they were treating such that V(fJl) was greater than
2-" V(:Y) for any convex symmetric inscribed :Y. The increases are
usually comparatively small and obtained at the expense of some
complication. We refer the reader to MULLENDER (1948a) and the
literature quoted there for further information.
In Chapter VI are obtained upper estimates for LI (!/) in terms of
V(!/) which are valid for all sets (Minkowski-Hlawka Theorem and
related topics).
111.6. A method of MORDELL. In this section we develop a method
of MORDELL for finding LI (!/) precisely for point sets !/ which mayor
may not be convex. The method applies primarily to star bodies.
This class of sets is defined by the properties that the origin is an inner
point and any radius vector meets the boundary either not at all or in
precisely one point: in other words, if ~ is any vector other than 0, then
either tZE g for all t~ 0 or there exists a to such that tz is an inner
point of !/, a boundary point of !/ or not in !/ according as t<to,
t = to or t> to. We now have the I rather trivial
LEMMA 4. Let!/ be a star body and suppose that a constant Llo exists
with the following two properties.
(i) every lattice A with d(A) =Llo has a point other than 0 in or on
the boundary of !/.
(ii) there exist lattices A, with d (A,) = Llo having no points other
than 0 in the interior of !/.
Then LI (!/) =Llo. If further, !/ is open l , then the critical lattices are
lust the A,.
For suppose, if possible that M is an !/-admissible lattice with
d(M)<Llo. Let 1'>1 be defined by y"d(M) =Llo. Then the lattice I'M
of points y~, ~E M has clearly no points in or on the boundary of !/,
contrary to (i). Hence LI (!/) 6L10' On the other hand (1 +e)A, has
no points in !/ for any e> 0, where A, is one of the lattices given in (ii).
1 i.e. does not contain any of its boundary points. MINKOWSKI and following
him MAHLER define a star body to be closed. We depart from their nomenclature.
 A method of MORDELL 85

Hence LI (9") ;£ (1 + t: t LI 0' so .1 (9") = LI o' The truth of the last sentence
of the lemma is now obvious.
When the description of star-bodies by distance-functions is intro-
duced in the next chapter, Lemma 4 will fall into place as part of a
wider theory.
MORDELL'S method of finding .1 (9") for a given star-body 9" may
now be described. First one must make an intelligent guess Llo at LI (9"):
in particular so that (ii) of Lemma 4 is true. If Llo has been correctly
chosen, then it may be possible to verify (i) and to find all the Ac in
(ii) by the following general procedure, of which the details naturally
vary widely from case to case. We suppose for simplicity that 9" is
open. Let M be any 9"-admissible lattice with d (M) = Llo. Then if
~ (1 ;£ j;£ r) is any collection of closed convex symmetric sets each of
volume
(1;£ j;£ r),
there must be points P;=FO of M in ~ for 1 ;£j;£r. Since M is 9"-
admissible, the P; must lie in fJt;, the set of points of ~ which are not
in 9". We may now use the hypothesis that the Pi are in a lattice M
of determinant Llo to obtain further points of M. Since these cannot
lie in 9", this gives further information about the Pi' In the end it
may be possible to show that M is one of a set of lattices A" all of which
have points on the boundary of 9". Lemma 4 shows that LI (9") =.1 0 ,
Of course the power of the method depends on a suitable choice of the ~.
MORDELL'S method is at its best in dealing with 2-dimensional
regions, since for these it is easier to grasp the geometry of the figure.
Before giving some concrete examples we must therefore study the
geometry of a 2-dimensional lattice more closely.
111.6.2. Throughout § 6.2 we denote by A a 2-dimensional lattice.
We regard vectors as coordinates of points on a 2-dimensional euclidean
plane, and use the normal geometric language to discuss their relations.
By distance we mean the usual euclidean distance. For later reference
we formulate our conclusions as lemmas.
We say that a point u of a (not necessarily 2-dimensional) lattice
is primitive if it is not of the shape u =ku1 , where U1EA and k>l is
an integer.
LEMMA 5. Let U be a primitive point of the 2-dimensional lattice A
Then the points of A lie on lines IT, (r =0, ± 1, ... ) which are parallel
to 0 U and at a perpendicular distance
rd(A)/lul
from it!. Each line IT, contains infinitely many points of A and these
are spaced at a distance Iu I.
1 As before lui = (u~ + u:)l, that is the distance from 0 to u.
86 Theorems of BLICHFELDT and MINKOWSKI

This is just a re-statement in geometrical language of what is known

already. Since u is primitive, there is a point v which with u forms a
basis for A (Chapter I, Theorem I, Corollary 3). Hence

det(u,v) = ±d(A),

that is the perpendicular distance from v on the line through °and u

is d(A)Jlul. But now A is just the set of points

rv +su (r, s integers).

Clearly the points with r fixed but s varying lie on a line IT, with the
required properties.
LEMMA 6. Let u, v be points of the 2-dimensional lattice A such that
0, u, v are not collinear. Then a necessary and sufficient condition that
u, v be a basis for A is that the closed 1 triangle ouv should contain no
points of A other than the vertices.
The condition is clearly necessary, by Lemma 5, so we must prove
it sufficient. If there are no points of A in the triangle ouv other than
the vertices, then the same must be true of the triangles with vertices

-u,o, v-u (1)

and
-v, u-v, 0, (2)

since, for example if JJ is a point of A in (1), then JJ +u is a point of A

in triangle ouv. Similarly there can be no points of A in the images
of our first three triangles in the origin, since - JJ is in A if JJ is. Hence
there is no point of Ain the hexagon .?Ie with vertices ±u, ±v,
°
± (v - u) except and the vertices. By Theorem II
d (A) ;;;:; ! V(.?Ie) =! Idet (u, v)\ .
But
Idet (u, v)\ = I d (A),
where the integer I is the index of the points u, v in A (Chapter I, § 2.2) ;
and so I = 1, as required.
The analogue of Lemma 6 does not hold in space of dimension > 2.
LEMMA 7. Let !2 be an open parallelogram with °as centre of area
4d (A), which contains no other point of A than 0. Then A has a basis
consisting of the mid-point of one of the sides of !2 and a point on one of
the other pair of parallel sides.
1 i.e. the sides are counted as belonging to the triangle.
 A method of MORDELL 87

After a suitable transformation of coordinates, we may suppose

that fl is the parallelogram

fl: IX d<1, IX21<1

and that d (1\) = 1. By Theorem III there is certainly a point of 1\
other than 0 in IX d;;:;; 1, IX 21 < 1, and so 1\ must contain a point
u = (1, u 2) Iu 2 1 < 1.
Similarly, 1\ must contain a point
v=(vI,1)
But now, since d (1\) = 1, the index of (u, v) in 1\ is
1= Idet(u,v)1 = 1- U 2 V I •
But I is an integer and IU 2 vII < 1. Hence 1=1 and either u 2 = 0 or
VI=O.
LEMMA 8. Let 1\ be a lattice 0/ determinant d (1\) which has two points
other than 0 in the closed parallelogram with vertices 0, a, b, a + band
volume (area) d(I\). Then either
(i) the two points are collinear with 0,
or
(ii) one 01 the points is a and the other is on the line-segment b, a + b,
or
(iii) one 0/ the points is b and the other is on a, a +b.
For the points p, q, say, are of the type

where
0;;:;; Jli ;;:;; 1 , 0;;:;; "i;;:;; 1 (j = 1,2).
The index I of p, q in 1\ is

I = IJll "2 - Jl2 "11·

Hence Jll"2- Jl2"'1 = 0 or ± 1; which gives the three alternatives quoted.
111.6.3. We first illustrate MORDELL'S method with an example
where the amount of subsidiary argument required is a minimum.
Let f be the cross-shaped 2-dimensional region defined by

min {lXII, Ix 2 \} < 1, max {lXII, Ix 2 1} <!.

We shall show that
LJ(f) =2
88 Theorems of BLICHFELDT and MINKOWSKI

and that the only critical lattices of :% are those with the following
bases:
Al basis (1.1) and (1. -1)
AI basis (t. -t) and (-t.f)
As basis (t. t) and (t. t)·
It is readily verified that these lattices are :%-admissible and have
determinant 2. Hence by Lemma 4. it is enough to show that any
Xz :%-admissible lattice A with
r----- -----, d(A) =2 must be one of AI'
I h 0 :
J---I------i.' A2 • As·
/
',0
/
/
0' From now on we suppose
i-- that
I
I
I
I
d(A) = 2: A is :%-admissible.
0
I 0
XI The convex symmetric oc-
I C
I tagon
I
IL __
,,
~: IXII <to
,, /
/
/

IX21 <to
IXII + IXII < i
I I
IL _____ _ _ _ _ _ JI

has area
~ > 2I d(A).
Fig. 5

and so contains a point a =F 0 of A The only points of ~ not in :%

are the four triangles with IXII ~ 1. IXII ~ 1 (see Fig. 5). Hence. by
symmetry. we may suppose that A contains a point a = (a l • a2 ) with
a: (1)
By Theorem III there is a point b =F 0 of A in
IXII < 1 IX21 ~ 2.
On taking - b instead of b if necessary and using the fact that b is
not in :%. we may assume that. the coordinates of b satisfy
b: (2)

Similarly there is a point c of A satisfying

c:
Now we show that a. b is a basis for A We have
 A method of MORDELL 89

and
det(a,b)<1·2+1·1 =t,
so
det(a, b) = 2 or 4,
since det (a, b) is an integral multiple of d (1\). Suppose first, if possible,
that det (a, b) = 4, so that the index of a, b in 1\ is 2. For any integer
k> 1 the points k-1a, k-1b clearly lie in .YI" and so are not in the .YI"-
admissible lattice 1\: that is a and b are primitive points of A We
show now that l (b - a) is in A Since a is primitive there is a basis
a,d where, say, det(a, d) =d(/\) =2. Than b=ua+vd for some
integers u, v; and indeed v = 2 since det (a, b) = 4 = 2 det (a, d). Then
u is odd since b is primitive, so {(b - a) is in 1\ as asserted. But
} (b - a) is clearly in .:it, so we have a contradiction. Hence we can only
have
det(a, b) =2 =d(I\). (4)
This gives the estimate
(5)
since otherwise we should have the contradiction

Similarly
det(a, c) = - 2 = - d(/\) (6)
and
(7)
Since a, b is a basis for 1\ we have
c=sa+rb
for some integers r, s. On substituting this in (6) and using (4) we
obtain r = - 1 and so
b +c =sa (8)
i.e.
(9)
But
-~ < b1 + C1 < 3, 1~ a1 < I
by (1), (2), (3); so there are only the two possibilities
s =1 or s = 2.
First case s = 1. From (1), (2), (3) and (9) we have
b1 < 0, c2 < 0. (10)
From (4), (5), (6) we have
det(c, b) = 2
90 Theorems of BLICHFELDT and MINKOWSKI

that is
c1 b2 - c2 bI = 2.
But Cl~.i, b2~1 by (2) and (3); and O>bl~ -l, O>C2~ -l by (5),
(7) and (10). Hence (8) can hold only if
C1 = b2 = I, C2 = b1 = -l,
which gives the lattice 1\2'
Second case s =.2. By (1), (2), (3) and (9) we now have
bl~O, C2~O. (11)
We now consider the lattice-point
(db d2 ) =d= (b - a) = l' (b - c) .
By (2), (3) and (11) we have
o~ 2d1 = bI - Cl~ - 2,
O~2d2=b2-C2~2.

Since d cannot be in .:f(" we must have d1 = -1, d2 = + 1; that is

C1= b2 = 2, b1= c2= O. This gives a1= a2= 1. Hence 1\ = 1\1 .
In the proof we have made use of the symmetry of the figure. Since
1\1 remains unchanged under transformation of .:f(" into itself, but 1\2
and As may be interchanged, we have shown that 1\ is one of 1\1,1\2,1\3'
as required.
111.6.4. As a second example of MORDELL'S method we take the disc
~: x~ +x~<1,
which we have already discussed by other means (§ 5.2). We take
,10= {!)I in Lemma 4. The lattices Ac certainly exist; since they can
be taken to be the lattices with a basis consisting of two of the vertices
of an inscribed regular hexagon. We shall show that if d (1\) = WI,
then 1\ has a point other than 0 in ~ except when 1\ is a I\c.
There are certainly points of 1\ in the circle
x~+x:<2,
since this has area 2n> 22> 22d (1\). Since 1\ is ~-admissible, the point
must lie in 1 ~ x~ + x~ < 2. After a suitable rotation of the coordinate
system we may thus suppose without loss of generality that there is
a point p = (PI' P2) in 1\ with
P2=-(i)l, l~Pl<l·
But now, by Theorem III there is a point q = (ql' q2) other than 0 in
the half-open parallelogram
~: IXI + riX21 ~ 1, IX21 < VI
 A method of MORDELL 91

of area 2113= 4d (1\) (see Fig. 6). The only portion of fl not contained
in !!) is the curvilinear triangle ~ cut off by the arc of the circle between
a = (1,0) and b =(l, - VI)
and the image of ~ in the origin. We may
suppose without loss of generality that q is in ~.
Cleariy both p and q are primitive, since, if either were of the shape
ku with u"A and integer k>1, then u would be in!!). Further p=j=:q,
since P2= - VI but \ q2\ < I T·
We now apply Lemma 8. From what

"\--------
\
\
\
\
\
\
\
\ (10)=0

Fig. 6

has just been proved, p, q, 0 cannot be collinear. Hence either p =b

and q lies on the line-segment between a (inclusive) and a + b (exclusive)
or q = a and p lies on the line-segment between b (inclusive) and a + b
(exclusive); and by symmetry we may suppose the second holds. Then
p - q = p - a lies between b - a (inclusive) and b (exclusive). The only
one of these points not in £f} is b - a. Hence also p = b. Hence A
is the lattice generated by a and b. Since we made an arbitrary rotation
of the coordinate system this completes the proof of the result stated.
111.6.5. As a final application of MORDELL'S method we prove a
result about binary cubic forms which fills a gap left in Chapter II, § 5.
We use the same notation.
THEOREM VII. 11 l(x 1 , x 2 ) be a binary cubic lorm 01 determinant
D<o, there are integers (u 1 , u 2)=j=:o such that

\/(u)I~I~ll.
The sign 01 equality is needed when and only when I is equivalent to a
multiple 01
92 Theorems of BLICHFELDT and MINKOWSKI

This is the most important part of Theorem XI of Chapter II, which

was left unproved. As already remarked, the form here is transformed
into the form there by the substitution XI--+X I , X 2 --+ - (x I + x 2 ). We
already noted that the exceptional form does require the sign of
equality since it has D = - 23 and represents only integers other than O.
We must first express Theorem VII in a geometrical form. We saw
in Chapter II that any two binary cubics with negative discriminant
can be transformed into one another. It is convenient to take X1 + ~
with discriminant - 27 as standard. Then Theorem VII is equivalent to
THEOREM VII A. Let 1\ be a lattice with
23 )1
d (1\) = (27. = Llo (say) (1 )

in the two-dimensional space 01 vectors X = (XI' X 2). Then 1\ contains a

point other than 0 in
f/: IX~+X~I <1, (2)

except when 1\ has a basis a = (aI' a2L b = (bl , b2) such that identically

(3)
In stating the equivalence of Theorem VII and Theorem VII A we
have tacitly applied Lemma 4 to the star body f/. From now on we
shall be concerned only with Theorem VII A. We use capital letters
to denote points and coordinates, except that 0 is still the origin.
Further, 1\ is a lattice with d(/\) given by (1) which has no point other
than 0 in the set f/' defined by (2). The set f/' is shown in Fig. 7.
First, since Llo < 1, there is certainly a point P =F 0 of 1\ in the square
(4)

Since P does not lie in f/', either P or - P must lie in the first quadrant
and we may suppose without loss of generality that.
O~1l<1, O~Pz<1. (5)
From Fig. 7 (or from elementary algebra) we must have

ll+Pz~1. (6)

Suppose, if possible, that there were two such points, P and P'. Then
their difference P"=P-P' satisfies (4). Hence on interchanging P
and pi if need be, we may suppose that P" is in the first quadrant:
of course it may coincide with P or P'. Hence
p=p,+p".
 A method of MORDELL 93

But now, in the obvious notation, we have P;+P;~1, P;'+P;'~1.

since neither P' nor pIt is in f/. Hence we should have

in contradiction to (5). To sum up what we have proved so far: there

is precisely one pair of points ±PE/\ other than 0 in the square

o2L+P

o
2fl-P
B

\
\

Fig. 7

!Xl! < 1, !X 2 ! < 1. We denote from now on by P the point of /\ which

satisfies (5) and (6).
We now examine more closely the lattices which satisfy (3). We
must make use of the algebra developed in Chapter II, § 5. Let AI, E I ,
A 2 , B2 be any numbers such that identically

On equating the hessians of both sides, we obtain

 Theorems of BLICHFELDT and MINKOWSKI

The linear factors ·of both sides must coincide, and so, after inter-
changing AI, BI and A 2, B2 if need be we have

On comparing the coefficients on both sides of (7), we have

A~+A~=1
(9- V69)A~ + (9+ V69)A~ = o.
This determines At Ag uniquely, and so AI' A 2, B I , B2 since they are
all real.
Hence there are only two lattices of the type specified in the theorem,
namely those with base
A = (AI' A 2), B = (BI' B 2),
and

respectively.
The approximate values are
Al : 1.014, A2 : - 0·347
BI-:- - 0.017, B 2 : 1.0005.
All we shall in fact use are the inequalities

(8)

The signs of A 2 , BI are easy to establish and, since

A~ + A~ = B~ + Bg = 1,
by (7), the rest follows.
Comparison of discriminants on both sides of (7) gives
27(AIBa - Aa B l)6 = 23,
and so
AIBa-AaBI= ±Llo,
where in fact the + sign holds, but we do not use this information.
Let X =~~ be the transformation of the plane X = (Xl' X 2 ) into
the plane z = (XI' xa) given by
XI=AIXI-BIXa, X2=A2XI-BaX2'
 A method of MORDELL 95

Then the region -r-l!l' of points -r-l X, X E!I' is given by

IX~-XIX~-x~1 <1.
Further, -r- l A is a lattice of determinant
(9)
(d. Chapter I, § 3).
The region -r-l!l' is shown in Fig. 8. The line Xl = 1 touches 10 (x) =
°
xi - Xl X~ - X~ = 1 at X 2 = and meets it again at x 2 =-1. The line
X2

-- --

o
X,

--
-:1-(lJI- --

x 2 =1 meets xi-xlx~-x~=-1 at xl=O, ±1. Since no line meets a

cubic curve in more than three points, it follows readily that the whole
of the unit square

°
lies in -r-l!l', except for a small region fJi in Xl < 0, X 2 < and the image
- ~ of fJi in the origin.
Suppose first that (1, 0)E-r-11\. Since d (-r- l A) =1, there are points
of -r- l A on the line x 2 = 1 spaced unit distance apart, by Lemma 5.
Since none of these can lie in -r-l!l' the only possibility is that -r-1A = Ao ,
the lattice of points with integral co-ordinates. But then A = -r Ao ,
which is one of the exceptional lattices permitted by the theorem.
Similarly, if (0, 1) E-r-1A, then A = -r Ao. Hence from now on we may
assume that
(10)
By Lemma 7, either there is a point q=l=o of A in the square !l,
or (1, 0) EA, or (0, 1) EI\. But the second and third alternatives have
96 Theorems of BLICHFELDT and MINKOWSKI

already been disposed of, and so, since q cannot lie in ",-1.9', we may
suppose that q is in Bt. Further, q must be primitive, since if q =kqI,
with integral k>1 and q I E",-lA, then qi would lie in IXII<j, IX21<t
and so certainly 1 in ",-1.9', contrary to the hypothesis that ",-1 A is
",-l.9'-admissible. Hence q is unique by Lemma 8.
We require another point of ",-IA. The tangent to lo(~) =1 at
(0, -1) is

This meets fo(~) = 1 again at (...2...., -

25
28). Hence all of the parallelogram
25.

~': Ix1 1<1, IX1+3x21~3

lies in ",-1.9' except for the points ± (0, 1) and a region or in X 2 < 0,
Xl> 0 and its image in the origin. But now, by Theorem III, there is
a point P of ",-1 A in ~' and, since ",-1 A is ",-I .9'-admissible, we may sup-
pose by (10) that P is in .r. The point p is primitive since if p =kpi
with integral k>1, then PI lies in IX11 ~t, IX21~-i, so PIE ",-1.9'.
An application of Lemma 8 shows that P is unique.
We note that the point 2q-p clearly lies in X1<O, x2 >-1.
Since the point q is the only point of ",-1 A in Bt, it follows that 2 q - P
must lie in the region to(~)~-1.
The next stage is to show that P and q form a basis for A. We have
(11)
and
(12)
Hence
det ( ) _ p _ p {< 0
p, q - 1 q2 qi 2 > _ 1.1 - l.t> - 3.
Since det (p, q) is a multiple of det (",-1 A) = 1, the only possibilities are
det(p, q) = -1
or
det(p,q) =- 2.
In the first case, p, q are a basis. Suppose, if possible, that det (p, q)
= - 2. Since p is primitive, there is a basis p, '1', where det (p, '1')
= ±d(A) = ±1. Write q=up+v'l'
where u and v are integers. Then
det (p, q) = v det (p, '1') ,
1 For then I/o(%lI :5lxlI3+ Ix1 11x212+ IX213<t< 1. We shall not explicitly give
such trivial estimations later.
 A method of MORDELL 97

so V = ±2. Now u must be odd, since q is primitive, and so

t = Hp - q) E-r-1 /\..
But then, by (11) and (12),
o < tl < 1, - i < t2 < 0;
a trivial estimation shows that I/o(t)I<1, and so t would be in -r-I 9',
contrary to hypothesis.
To sum up: there is a basis PE.'T, qEfJl of -r-1 1\. The point 2q- P
lies in 10 (x) ~ - 1. There are no other points of .-1/\. in f/ or fJl.
We must now translate our facts about -r-1 /\. into facts about I\.
We write

The region -r fJl is bounded by the curve

X~+X~=1 ,

the transform of 10 (x) = 1, and the line-segment joining the points

-r(O,-1)=B, -r(-1,-1)=-A+B,
and so is roughly as shown in Fig. 7. The point

Q (say) =-rq
lies in -r fJl.
+
Similarly -rf/ is bounded by Xi X~ = 1 and the tangents at
-r(O, -1)=B and at -r(1,O) =A. We now show that -rf/lies in
(13)
Indeed, since B 2> 1, the tangent to X~ + X; = 1 at B has negative
gradient and so meets Xi +X~ = 1 again at a point in (13). Since-rff
lies below this tangent, its points satisfy X 2 < 1. Similarly, since AI> 1,
the points of -rff satisfy X 1 <1. They clearly satisfy X 1 >O, X2>O.
But now we saw earlier that there is only one point, P, of /\. in (13).
Since -rp is in -rf/ we must have
-rp =P.
To sum up the results of our translation: there is precisely one point
Q E /\. in -r fJl. This point Q together with the unique point P of /\. in
(13) form a basis for I\. The point 2 Q - P lies in Xi X~~ -1- +
Let £' be the mirror image of -r fJl in XI + X 2 = O. By symmetry
there is precisely one point L, say, of /\. in £': this point together with
- P forms a basis for /\., and the point 2L P lies in Xi X~~ 1- + +
Cassels, Geometry of Numbers 7
98 Theorems of BLICHFELDT and MINKOWSKI

But now every point of the triangle oLQ is in one of the regions
Y, T 31 and.Yf'. By hypothesis there is no point of 1\ in Y, and we
proved that Q, L are the only points of 1\ in T 31,.Yf' respectively.
Hence Q, L forms a basis of 1\ by Lemma 6.
We have three bases P, Q: Q, Land L, -P for 1\ and must study
their relations. Now
det (P, Q) = det (Q, L) = det (L, - P) = d (1\),
since the determinants are ± d (1\) and are clearly positive. Write
P=uQ+vL.
Then
det (P, Q) = v det (L, Q),
det (P, L) = u det (P, Q) .
Hence
P=Q-L.
We have now reached a contradiction, since
2Q -P=2L+P,
and this point has been shown to lie both in X~ + X~;S; 1 and in
X~ + X~;;;; -1. The contradiction shows that there are no Y-admissible
lattices with d (1\) = Llo except those mentioned in the enunciation of
the theorem.
We have shown rather more. Let the line joining Band - A + B
(which forms part of the boundary of T3I) meet Xl +X2 =O in the
point (- c, c). Then it is clear that our argument shows that there is
a point of every lattice 1\ with d (1\) ;;;; Llo in the bounded region
IX~ + X~ I < 1 , max {i XII, IX21} ;;;; c,
except when 1\ is one of the two critical lattices. That is, IX~ + X~ I< 1
is boundedly reducible and indeed fully reducible in the sense of
Chapter V, § 7.
111.7. Representation of integers by quadratic forms 1. In this
section we digress to present a number of results in the arithmetic
theory of quadratic forms which can be proved very simply by the
methods of the geometry of numbers. The principle tool is the following
lemma.
LEMMA 9. Let n, m, kl' ... , k m be positive integers and ajj (1;;;; i;;;; m,
The set /\ 0/ points u with integral co-ordinates
1 :;;'j:;;' n) be integers.

1 This section is not used later.

 Representation of integers by quadratic forms 99

satisfying the congruences 1.

L aijUj == 0
l:,>j:'>n
(k i )

is a lattice with the determinant

That A is a lattice follows, for example, from Theorem VI. Two

points u and v of the lattice Ao of all integral vectors are in the same
class with respect to A if and only if
L aijuj == L aijv j (k i ) (1~i~m).
i i
Hence the index I of A in Ao , that is the number of classes, is at most
II ki'so
d (A) = I d (Ao) ~ II k i
i
(compare Lemma 1 of Chapter I).
III. 7.2. As a first example we show that every prime number
p == 1 (4) is the sum of the squares of two integers. For then, as is well
known, there is an integer i such that
i2 + 1 == 0 IP).
The set of integers (u 1 , u 2 ) such that
u 2 == iU 1 (P) (1 )
is, by Lemma 9, a lattice A of determinant dIA)~p. Hence by MIN-
KOWSKI'S convex body Theorem II there is certainly a point of A in
the disc
~: x~ +x~< 2P
of area V(!Zl) =2ilP>2 2 dIA). Hence there are integers U 1 , u 2 not both 0
satisfying (1) and

But (1) implies

u~ + u~ - f4~ (1 + i2) == 0 (P) •
and so u~ + u~ = p, as required. The method is readily extended to show
that a positive integer is the sum of two squares provided that it is
not divisible by a prime p = -1(4).
111.7.3. As a second example, we shall show that every positive
integer m is of the shape
m = u~ + u~ + u~ + U:
1 By a"", b (k) we mean that a - b is divisible by k.
7*
100 Theorems of BLICHFELDT and MINKOWSKI

with integral u I , u 2 , u 1 , u 4 • We may suppose without loss of generality

that m is not divisible by a square other than 1, so

m=PI"'Pg ,
with distinct primes PI' "', pg • We now show that to every prime P
there exist integers ap ' bp such that

a! + b! + 1 == 0 (P) . (1)
Indeed when Pis odd the numbers
(0;;:;; a < iP)' (2)
and
- 1 - b2 (3)
are each a set of i (P + 1) integers which are incongruent modulo p.
Since there are only P classes modulo P, there must be some integer c
which is congruent to an element of each set (2) and (3), that is
a~ == c == - 1- b!, so a~ + b~ + 1 == O. If P= 2, then a2 = 1, b2 = 0 will do.
We now consider (d. DAVENPORT 1947b) the lattice of integral
U = (u 1 , ••• , u 4 ) which satisfy the 2g congruences

U1 == ap u 3 + bp u 4 (P) } (4)
u 2 == bp u 3 - ap u 4 (P)
for P= PI' ... , pg • By Lemma 9, these form a lattice 1\ of determinant
d(/\) ~ pi ... p~ = m 2•
Hence there is a lattice-point other than 0 in the set
x~ + x~ + x~ + x: < 2 m
of volume
in2(2m)2> 24m2~ 24 d(I\).
If u is this point, then
o< u~ + u~ + u~ + u: < 2 m
and, by (1) and (4),
u~ + u~ + u~ + ui == (a! + b! + 1) u~ + (a~ + b~ + 1) u~ == 0 (P)
for P= PI' ... , Pg; that is m divides u~ +... +u~.This proves the result.
111.7.4. A famous theorem of LEGENDRE states that a ternary
quadratic form t (Xl' X 2 , x3 ) with rational coefficients represents 0 if
obviously necessary congruence conditions are satisfied. Following
DAVENPORT and MARSHALL HALL (1948a) and MORDELL (1951 a) we
verify this in a particular case, to which indeed the general case may
be reduced by simple arguments.
 Representation of integers by quadratic forms 101

Let
1(:£) = al X~ + a2X~ + a3xt
where aI' a2, a3 are square-free integers no two of which have a common
factor, so al a2 a3 is square-free. We show that there exist integers u*o
such that I (u) = 0 provided that the following two .conditions are
satisfied
(i) there are integers AI' A 2, A3 such that
al + A~a2 == 0 (a3), a2 + A~a3 == 0 (a l ) , a3 + A~al == 0 (a 2)
and
(ii) there are integers VI' V 2 , V3 not all even such that
al VI2 + a2 V22 + aaV32 == 0 (22+A) ,

where A. = 1 or 0 according as ~a2a3 is even or odd.

Let
laI a2 a3 1=2 ApI,,·Pg
where PI' "., Pg are distinct odd primes and A. = 1 or O. We shall take
for 1\ the integral vectors U=(U I , U2 , u3) satisfying the following con-
gruence conditions.
(I) Let P be one of PI' "., pg • By symmetry we may suppose that
a3 == 0 (P). We impose the condition

u2 == A 3 uI (Pi·
Then
al u~ + a2u~ + a3u~ == al u~ + a2u~ == (a l + a2A~) u~ = 0 (P) .
(II",) Suppose A. =0, so aI' a2 , aa are all odd. Now
v2 == 0 or 1 (22)
for any integer v. In condition (ii) precisely one of VI' V 2 , Va must be
{'ven, say V 3 • Then
o == al v~ + a2 v~ + aa v= == al + a2 (22) .
We impose the two congruences

(2), }
(2) .
Then

(IIp) Suppose A. = 1, so one of ~, a2 , aa is even, say aa. Then

al VI2+ a 2 V22 'IS even, so VI' V 2 are both even or both odd. If VI' V 2 were
even then
102 Theorems of BLlCHFELDT and MINKOWSKI

would be divisible by 22, so V3 would be even. Hence VI' v2 in (ii) must

be odd, and

since v2 = 1 (23) if v is odd. We impose the two conditions

Then it is readily verified that

al u~ + a2u~ + aau~ = O. (23)

In any case the lattice 1\ of integers u has determinant

d (1\) ~ 2H2 PI'" Pg = 41 al a2aal,
and the congruence conditions imply that

al u~ + a2u~ + aau~ == 0 (mod 2A+2PI'" Pg = 41 al a2 aa!)'

But now, by MINKOWSKI'S convex body theorem, there is a lattice
°
point not in the ellipsoid
fI: Ia] I x~ + Ia2 1 x~ + Iasl x~ < 41 lIt a asl
2
of volume
V(fI) = ~ . 25 1a1 a2 tlal > 23 d(A).
3

If u*o is the lattice point in fI, we must have alu~+a2u:+tlaug=O,

since it is divisible by 4~a2as; and

We conclude with a couple of remarks. An obviously necessary

+
condition for solubility of lIt u~ a2 u~ t as u= = 0 is that lIt, as, as should
not all have the same sign. We did not use this at all. Hence this
condition must be derivable from the others. The reader migbt verify
that this can be done by means of the law of quadratic reciprocity.
In the second place we have not merely shown the existence of a
solution, but we have shown that there is one which satisfies

Iall u~ + Iasl u~ + Ias Iu= < 41 al as as I·

The right-hand side here may be improved to 2tllItasasi by the use
of the precise Theorem III of Chapter II instead of Theorem II, as the
reader can easily verify.
 Introduction 103

Chapter IV

Distance-Functions
IV.t. Introduction. In this chapter we introduce a number of
concepts which are useful tools in all that follows.
IV.t.2. A distance-function F(;lJ) of variable vector;lJ is any function
which is
(i) non-negative, i.e. F(;lJ) ~ 0,
(ii) continuous,
and
(iii) has the homogeneity-property that
F(t;lJ) = tF(;lJ) (t ~ 0, real).
The set [/ defined by
[/: F(;lJ) <1 (1 )
turns out to be a star-body in the sense already introduced in the last
chapter: that is, the origin 0 is an inner point of [/ and a radius vector
t;lJo (0 ~ t < (0)
either lies entirely in [/ [which happens when F(;lJo) =0] or there is a
°
real number to = {F(;lJo)} -1> such that t;lJo is an interior point of, a
boundary point of or outside of [/ according as t<to, t=to or t>to.
In § 2 we examine this relationship and show that conversely every
star-body [/ determines a distance function F(;lJ) such that the set (1)
is the set of interior points of [/. Since many, though not all, of the
point-sets of interest in the geometry of numbers are star-bodies, the
concept of distance-function plays an important rOle.
Most of the problems considered in Chapter II relate to star-bodies;
and then it is easy to write down the corresponding distance functions.
For example if 1(;lJ) is a positive definite or semi-definite 1 form, the set
1(;lJ) <1
corresponds to the distance-function
F(;lJ) = {I (;lJ) }1t"
where r is the degree of 1(;lJ). Again, if 1(;lJ) is an indefinite form of
degree r andk>O is a number, then the set
- 1 < 1(;lJ) <k
1 By semi-definite we mean that f (;E) Z 0 for all a: but f (;E) = 0 for some ;E * o.
104 Distance-functions

corresponds to the' distance function

F(3:) _ {k--
1/'1/(3:)1 1/' if 1(3:) ~ 0,
- . 1/(3:)1 11' if 1(3:) ~ O.
The reader will readily verify that both the functions just defined
are in fact distance-functions. One advantage of introducing distance-
functions is that some of the ideas of Chapter II can be carried over
to all star-bodies. A simple example of a 2-dimensional set which is
not a star-body is
0<X1 X 2 < 'I.
Clearly star-bodies!/' which are symmetric, i.e. have the property
that - 3:E!/' when 3:E!/' correspond to distance-functions which are
symmetric in the sense that
F( - 3:) = F(3:) .
K. MAHLER (1950a) and C. A. ROGERS (1952a) have investigated a
wider class of sets which ROGERS calls star-sets and which include the
closed star-bodies as a sub-class. A star-set is a closed set such that
t3: E!/' whenever 0 ~ t ~ 1 and 3: E!/'. They are important in connection
with certain problems ("bounded reducibility" d. Chapter V, § 7) and
we shall mention them again; but we refer the reader to the original
memoirs for the details.
IV.l.3. Convex sets:f( are important as MINKOWSKI'S convex-body
theorem shows. It turns out that the convex sets which have the origin
o as an interior point are precisely the star-bodies whose distance-
function satisfies the inequality
F(3: + y) ~ F(3:) + F(y) .
This we prove in § 3. We call such distance-functions convex.
In § 4 we show that an n-dimensional convex set :f( has a tac-
(hyper) plane 1 at every point a on the boundary of :f(; that is a
(hyper) plane

which passes through a and is such that :f( lies entirely on one side
of or in n; say
(all 3: E:f() .
Clearly if there is a tangent plane to:f( at a, then it is the only tac-
plane. But tac-planes exist even when tangent planes do not, and they do
1 We use the words tac-plane and plane for tac-hyperplane and hyperplane.
'When n = 2 the corresponding thing is called a tac-line. The term supportplane
(German: Stiitzebene) is sometimes used.
 General distance-functions 105

not need to be unique: for example when a is a corner of the square

IXII < 1, Ix2 < 1.
1

In discussing tac-planes it is convenient to introduce the polar

body of a convex body; a notion which we shall in any case require
in Chapters VIII and XI. Any plane n not passing through the origin
can be put in the form
n: y1x1+"'+Ynxn=1,
and so may be represented as a point y = (YI' ... , YII) in n-dimensional
space. It turns out that the points y corresponding to planes n which
do not meet! f themselves form a convex set f*, say, the polar of f.
Further, the relationship between f and f* is reciprocal in the sense
that f may be obtained 2 from f* in the same way as f* was obtained
from f.
An example of a pair of polar bodies are the generalized cube

~: maxlx;1 ~ 1
and the generalized octahedron

£1): 2:IY;I~1.
It is easy to see that a plane 2: Y;x;= 1 for fixed y can contain a point
of the interior of ~ only if y is not in £1); and vice versa. We discuss
polar sets in § 4.
There is a rich theory of convex sets but we do not prove more than
is relevant to the geometry of numbers. For the rest the reader is
referred to the report of BONNESEN and FENCHEL (1934a) or EGGLE-
STON'S tract (1958a).
IV.2. General distance-functions. We set up now the relationship
between distance functions and star-bodies sketched in § 1.2.
THEOREM I. A. II F(x) is any distance lunction then the set
Y: F(x) <1
is an open star-body. The boundary 01 Y is the set 01 points x with
F(x) = 1 and points with F(x) > 1 are exterior to Y (that is, have a neigh-
bourhood which does not meet Y).
B. Conversely any star-body .r determines a unique distance-Iunction
F(x). II Y is the set 01 interior points 01 .r then Y is related to F(x)
in the way described in A.

1 We say that two point-sets meet if they have a point or points in common.
2 Strictly speaking the set .%'"** obtained from .%'"* coincides with.%'" except
possibly on the boundary. The distance-functions of.%'" and.%'"** are thus the same.
106 Distance-functions

We note first that two distinct star-bodies 9";. and 5; determine the
same distance-function F(x) if they have the same set of interior points,
but a distance-function defines precisely one open star-body, namely
F(x) < 1 and one closed star-body, namely F(x) ~ 1. Distinct distance-
functions 1;, ~ always determine distinct star-bodies. For then 1; (xo) =1=
~ (xo) for some x o' say Fl (xo) < F; (xo); and then there is a t such that

1; (txo) < 1 < ~ (tx o);

so txo is in one star-body but not the other.
The proof of Part A of the theorem is nearly trivial. If F(xo) < 1,
then, by the continuity of F(x), there is a neighbourhood

Ix - xol < r;
of Xo which lies in Y; so Xo is an interior point of Y. Here we have used
the standard notation
Ixl = (x~ + ... + x!)~.
Similarly, if F(xo) > 1, then there is a neighbourhood of Xo which does
not meet Y. Finally, if F(xo) = 1, then every neighbourhood of Xo
contains points txo both with t>1 and t<1, for which F(tx o) > 1,
F(txo) < 1 respectively: and so Xo is a boundary point of Y.
It remains to prove B. If.'T is any star-body, we define a function
F(x) as follows:
(cx.) F(x) =0 if tXE.'T for all t>o. In particular F(o) =0.
(p) If tx is not in .r for all t> 0 then, by the definition of a star-
body, there is a to=to(x»O such that tx is interior to or exterior to Y
according as t < to or t> to; and tox is on the boundary of ,i/. vVe put

Clearly, if F(x) is a distance function, then it is related. to the set Y of

interior points of .r in the way described. Further, it follows trivially
from the construction of F(x) that it satisfies two of the defining pro-
perties of a distance function, namely F(x) ~ 0 and F(tx) = tF(x) for
all t> o. It remains only to show that F(x) is continuous.
We show first that F(x) is continuous at o. By the definition of a
star-body, the origin 0 is an inner point of .'T, so there is an r; > 0 such
that the sphere

is contained in.'T. Hence, if xo=l= 0, the vector

t' x o, if = r;/I xol
is certainly in .r, so
F(xo) ~ r;-ll Xo I. (1)
 General distance-functions 107

Since 'YJ is independent of x o, this proves the continuity of F(x) at the

OrIgIn.
We now prove continuity at a point xo=l=o. Let e>.O be arbitrarily
small. The point
(2)
is an interior point of 3 by the definition of F; and so there is a
neighbourhood
Ix - xII <'YJl'
which lies in 3, that is, (3) implies
F(x) ~ 1. (4)
Write
.r = {F(xo) + etly, 'YJl = {F(xo) + et 1'YJ2'
Then (3) is equivalent to
(5)
and, by the homogeneity property of F(x), which we have already
proved, the inequality (4) is equivalent to
F(y) ~ F(xo) + e. (6)
We have thus found a neighbourhood (5) of Xo in which (6) holds.
It remains to find a neighbourhood of points y in which
F(y) ~ F(xo) - e. (7)
If F(xo) ~ e, then (7) is true for all y, since F(y) ~ O. Otherwise one
considers the point
x 2 = {F(xo) - etlxo'
This is an exterior point of 3 and then the argument goes as before.
This completes the proof of the theorem.
There is the trivial corollary of which we leave the proof to the
reader.
COROLLARY. Let ~ (x) and ~(x) be distance functions. The star-body
~ (x) < 1 is a subset of the star-body ~ (x) < 1 if and only if

(8)
lor all x.
We record for later reference two results, the first of which we have
already proved.
LEMMA 1. For every distance-function F(x) there is a constant C such
that
F(x) ~ Clxl
lor all x.
108 Distance-functions

LEMMA 2. A necessary and sulficient condition that the star-body

F(or) < 1 be bounded is that F(or) =F 0 if or =F o. There is then a constant
c> 0 such that
(9)
for all or.
We proved Lemma 1 above with C =Tj-1, at least when or=FO; and
it is trivial when or =0. If there is a oro=FO with F(oro) =0 then toro
lies in F(or) < 1 for all t> 0, so F(or) < 1 cannot be bounded; which
proves half of Lemma 2. Suppose conversely that F(or) =F 0 if or =F o.
The function F(or) is continuous on the surface of the sphere 1orl = 1 ;
and so attains its minimum, say, at oro. Then F(oro)> 0, by hypothesis.
Put F(oro) = c. Then F(oro) ~ c if lor 1 = 1; and so (9) holds by homo-
geneity. This completes the proof of Lemma 2. We note that if (9)
holds, then F(or) < 1 is entirely in the sphere 1orl < c-1.
The following trivial corollary rids Lemma 1 of its dependence on
the particular distance-function lor I.
COROLLARY. Let ~ (or), ~ (or) be distance-functions and let ~ (or) < 1
be a bounded set (i.e. ~ (or) = 0 only for or = 0). Then there is a constant C
such that

for all or.

If, further, ~(or)<1 is bounded, then there is a c>O such that
C 1\ (or) ~ r; (or) ~ c1\ (or) .
The second part of the corollary may be picturesquely summed up
10 the slogan "for qualitative purposes there is only one bounded
star-body".
IV.3. Convex sets. A set.Yt' is convex if
tor+(1-t)y (0<t<1) (1 )

is in.Yt' whenever or and yare in.Yt'. It is said to be strictly convex

if the points (1) are all interior points of .Yt'.
We show first that if or1 , ... , or, are any points of.Yt' and

then
(2)
This is true for r = 2 by the definition of convexity, and it is true for
r>2 by induction, since we may suppose that t1=F1, and then
t1or1+ ... + t,or, = t1or1+ (1 - t1) y,
 Convex sets 109

where
y=--:r2
t2 + ... + --:r,EoA,
t, .go
1-tl 1-tl

since it is of the shape (2) with r -1 summands.

Almost immediate consequences are
LEMMA 3. A convex set f in n-dimensional space either lies entirely
in a hyperplane
1T: PI Xl + ... + p" X" = k
or it has interior points.
LEMMA 4. A convex setf with a volume V(f) such that 0< V(f) < 00
is bounded.
For if f does not lie in a hyperplane it contains n + 1 points

which do not lie in a hyperplane. The points L ti:ri with ti ~ 0, L ti = 1

are just the points of the simplex with vertices :r1 , ••• , :r"H. The
whole simplex must be contained in f , and since a simplex has interior
points this proves Lemma 3.
In Lemma 4, we note that f cannot lie in a hyperplane if V(f) > 0;
so we may suppose without loss of generality after a change of origin
that 0 is an interior point of f. There is then a number 11 > Osuch
that all the vectors
;-1 "-i
~~

11 ei = (0, ... ,0,11,0, ... ,0) (1~j~n)

are in f . If a = (aI' ... , a,,) be any other point of f , we shall show that
max Iail ~ 11-,,+1 (n!) V(.~).
1::;;1~"

If, say, a1=FO, then the whole of the simplex with vertices 0, a, rJe2'
... ,11e" is contained in f and has volume
(n!)-l· 11"-ll ~I·
Since this can be at most V(f), the result follows.
Finally we prove
THEOREM II. A convex body f of which 0 is an interior point is a
star-body. The corresponding distance function F(:r) satisfies the inequality
F(:r + y) ~ F(:r) + F(y) (3)
for all :r and y.
Conversely if F(:r) is a distance function for which (3) holds, then the
star-body
Ft:r) < 1 (4)
is convex.
110 Distance-functions

The converse is trivial. If F(x)<1, F(y)<1 and O<t<l, then the

inequality (3) applied to tx and (1 - t) Y gives
F{tx+(1-t)y}:£F(tx) +F{(1-t)y}
+
= tF(x) (1 - t)F(y)
<t+(1-t)
=1.
It remains, then, only to verify the direct assertion of Theorem II.
We define a function F(x) as follows:
F(x) = inf t-I, (5)
where the infimum is taken over all t such that
t>O, tXEX". (6)
Since 0 is an interior point of X", there certainly do exist t satisfying (6).
It follows at once from the definition that F(x) ~ 0, F(o) = 0; and that
F(sx) = sF(x) for all s ~ 0. Thus F(x) will be a distance-function if
we can prove continuity. We first prove the functional inequality (3)
and then deduce continuity from (3).
Let x,y be any two vectors and s, t any two positive numbers such that
SXEX", tYEX". (7)
Then
rsx + (1 - r) tYEX"
if O<r< 1. We choose r so that this point is multiple of x +y, i.e.
rs=(1-r)t; r=t/(s+t).
Then
~(X+Y)EX";
s +t
so

Hence
F(x + y) :£ F(x) + F(y)
since F(x), F(y) are the infima of S-I, t- 1 over s, t respectively which
satisfy (7).
The function F(x) is continuous at 0 by the same argument as was
used for distance functions. Since 0 is an interior point there is a neigh-
bourhood

of 0 contained in X", and so

 Convex sets 111

The continuity at a general point :ro is now immediate. We have

F(:ro + y) + F(y) ,
~ F(:ro)
and
F(:ro) ~ F(:ro + y) + F( - y) .
Hence
IF(:ro + y) - F(:ro) I ~ max F( ± y) ~ 'Y}-11 yl
±
<e
for any given e>O, provided that Iyl <'Y}e.
Finally we must verify that the set
F(:r) <1
is in fact the set of interior points of f. A point :r with F(:r) < 1 is
certainly in f since, by the definition of F(:r), there is a t> 1 such that
t:rEf; and so

is in f by convexity. Since F is continuous, the set F(:r) < 1 is open;

and so all its points are inner points of f . Conversely, if :r is an inner
point of f , there is a t> 1 such that t:rEf, and so F(:r) < 1 by the
definition of F(:r). From the definition of F, no point :r with F(:r) > 1
can belong to f. Points with F(:r) = 1 mayor may not belong to f
but, since F(:r) is a distance-function, they must be boundary-points
off.
For later reference we enunciate formally a result we have just
proved:
COROLLARY. Let F(:r) be a non-negative junction of the vector :r which
satisfies the two conditions
F(tx) = tF(x) if t> 0,
F(:r + y) ~ F(:r) + F(y) ,
and which is continuous at o. Then F(:r) is continuous for all :r; and so
is a distance-function.
IV.3.2. The next lemma is an essential preliminary to the treat-
ment of polar bodies and tac-planes.
LEMMA 51. Let ~,Jt;; be a closed convex sets having no point in
common. Then there is a hyperplane
1T: PI Xl + ... + Pn Xn = k
which separates ~ and Jt;: that is all the points of ~ are on the opposite
side of 1T from those of Jt;.
1 Proof given is valid only if at least one of .;t;, .)(2 is closed (as otherwise there
need be no minimum distance).
112 Distance-functions

Consider the distance l:rl - :r2 1 when :r1 ,:r2 run through the points
of ~, %2 respectively. Since Jt'i and Jt; are closed, this distance
attains its infinum at some points :rjEJtf (j = 1, 2); and :r~ =F:r~ since
Jt'i and ·""2 have no points in common. We show that the hyperplane 1T
which bisects perpendicularly the line-segment :r~:r~ will do what was
required. After a suitable rotation of the co-ordinate system and a
possible change of origin we may suppose that
:r~ = (- 'f}, 0, ... ,0), :r~ = ('f}, 0, ... ,0)

for some 'YJ> O. The plane 1T is then

1T: x1 =0.
Suppose, if possible, that there is a point z in Jt'i with ZI ~ O. By con-
vexity, the point
z, = (1 - t):r~ + t z (0<t<1)
is in ~. The distance 1z,- :r~ 1 is given by
1z, - :r~12 = (2'f} - t'f} - tz1)2 + L (tz;)2
2",;:;; ..
= 4'f}2 - 4('f} +ZI)'f}t + o(t2) < 4'f}2,
if t is small enough and strictly positive. This contradicts the definition
of :r~ and :r~. The contradiction shows that z cannot in fact exist, and
so proves the lemma.
COROLLARY. II;;r is a convex closed set and a a point not in ;;r,
there is a hyperplane separating % and a.
For we may put Jt'i =% and take %2 to be the set consisting of a
alone.

IV.3.3. In introducing the polar set of a given convex set %, we

confine attention to the case when % is bounded and can be described
by a distance function; that is 0 is an inner point and 0< V(%) < 00
by Lemmas 3 and 4 and Theorem II. If the reader is interested he will
have no difficulty in extending the results to the other cases using
Lemma 2.
We write
pa=pl~+ .. ·+p.. a..
for the scalar product of two vectors p and a.
THEOREM III. Let F(:r) be the distance-/unctzon associated with a
bounded convex set. For all vectors y let
F*( y ) -sup
_ a:y
F( ). (1 )
:I: ,*,0 a:
 Convex sets 113

Then F*(y) is the distance-function associated with a bounded convex set.

The relationship is reciprocal in the sense that
_ OlJY
F(al) - sup P( ) . (2)
II'*'O Y
The functions F and F*, or the convex sets associated with them,
are said to be polar to each other.
We must first show that F* is well-defined. Since the body F(al) < 1
is bounded, we have F(al) =F 0 if al =F 0 by Lemma 2, and indeed there is a
constant c>O such that F(al)~Clall. Since alY~lalIlYI, it follows that
F*(y) ~ c-1Iyl· (3)
Immediate consequences of the definition are that
F*(ty) = tF*(y) if t> 0, (4)
and
F*(y) >0 if y=Fo. (5)
Now if Yl' Y2 are any vectors, we have

F*(y
1
+ Y2) = sup
..,
:Jl(Yl + Y.)
F(:Jl)
::;; sup :JlYl
-.., F(:Jl)

= F*(Yl)
+ sup :JlY.
.., F(:Jl)

+ F*(Y2)'
I (6)

But now (3). (4), (5) and (6) show that F*(y) is the distance-function
of a convex set, by Theorem II and its Corollary. This convex set is
bounded because of (5) and Lemma 2.
It remains only to prove (2); and here we need the convexity of
F(al), which we have not yet seriously used. If al =0, then (2) is trivial,
so let alo=FO be fixed. From (1) we have
F(al) F*(y) ~ a:y (7)
for all al and y: and so certainly

F(a:o) ~ s~p F~o(:) • (8)

Let e> O. Then by Lemma 5 Corollary there is a hyperplane TT separating

alo from the set of a: such that
(9)
Since TT does not pass through the origin, it may be written in the shape
TT: a:yo = 1. (10)
Then F(a:) ~ (1- e)F(a:o) for all points a: on TT, since TT does not meet
(9); hence
(11)
Cassels, Geometry of Numbers 8
114 Distance-functions

since one need clearly only consider the x with xy = 1 in (1), by homo-
geneity, if y*o. Further,
xoYo> 1, (12)
since xi! is on the other side of 1T from the origin, which is a point of (9).
From (11) and (12) we have

sup~oJL;;::: a)oYo > (1 - B) F(x ) (13)

y F*(y) - F*(yo) 0 •

The required result (2) now follows from (8) and (13), since B is arbi-
trarily small.
This concludes the proof of the theorem. The reader will be able
to verify readily that the sets F(x) < 1 and F*(y) < 1 are related to
each other in the way described in § 1.3.
We have at once the
COROLLARY 1
F(x) F*(y) ~ xy
lor all x, y For any Yo* °there is an x o*°such that
F(xo) F*(yo) = xoYo; (14)
and vice versa.
We have already noted the first inequality, which is an immediate
consequence of the definition. By symmetry it is enough to show the
existence of x o, given Yo. The set fJI of points x with F(x) = 1 is
bounded; and it is closed since F(x) is continuous. Hence the continuous
function xYo attains its upper bound, say at xo. But we have already
seen that the upper bound is F*(yo), so (14) must hold.
We also shall need later
COROLLARY 2. Let ~, Jt; be convex sets with non-zero volume having
the origin as inner point and with respective polars ~* and :Yt;*. If .fl
contains .)("2 then :Yt;* contains ~*.
Let the corresponding distance functions be 1\ (x), F; (x), Fl*(X) ,
F2*(X). Then F;(x)~1\(x) by Theorem I Corollary. The definition (1) of
the polar distance-function then gives immediately F2*(Y) ~Fl*(Y) for ally.
The following corollary links polar distance-functions with the polar
lattices and transformations introduced in Chapter I, § V.
COROLLARY 3. Let F(x), F*(y) be a pair of mutually polar convex
distance-functions. Let 't be a homogeneous linear transformation and 't*
its polar transformation. Then F('tx) and F*('t*Y) are mutually polar.
For by the definition of 't* we have 'tx't*Y =xy for all x, y. The
truth of the corollary now follows from (1) and (2).
 Convex sets 115

IV.3.4. A hyperplane 1T through a point Xo on the boundary of a

convex set :fl is said to be a tac-plane to :fl at Xo if no interior point
of:fl is in 1T. The following Theorem IV is an almost immediate con-
sequence of the results of § 3.3. We shall need Theorem IV in the next
chapter, but § 3.3 only in Chapter VIII.
THEOREM IV. Let:fl be any convex body with volume V(:fl) such
that 0< V(:fl) < 00. Then at every point Xo on the boundary ol:fl there
is at least one tac-plane. There are precisely two tac-planes to :fl parallel
to any given hyperplane 1T.
We may suppose that 0 is an interior point of <X". Let F(x) be the
corresponding distance function. Then F(xo) = 1. By Corollary 1 to
Theorem III there is a Yo=F 0 with
(1 )
The plane
(2)
thus passes through xo' By the Corollary 1 to Theorem III we have
x Yo ~ F{x) F* (Yo) ,
so F(x) ~ 1 for all points of 1T'. Hence 1T' contains no interior point
of :fl, so is a tac-plane.
Any plane (2) for fixed Yo is a tac-plane at some point xo' For by
Corollary 1 to Theorem III there is an Xo such that (1) holds.
Hence if Yo is any vector, the two planes
xYo = F*(yo)
and
xYo = - F*( - Yo) (4)
are both tac-planes. It is clear that they are the only tac-planes parallel
to x Yo = O. The origin lies between the hyperplanes (3) and (4), and
hence so does the whole of the interior of :fl.
IV.3.S. In Chapter IX we shall need the following result.
LEMMA 6. Let.x;. and ~ be open convex sets in n-dimensional space
with
0< V(X'j) < 00 (j =1, 2).
Suppose that .x;. and ~ have no points in common but that a is a boundary
point 01 both .x;. and~. Then there is a hyperplane through a which
does not meet either .x;. or :fl2 (and so is a tac-plane to both .x;. and ~).
The proof follows that of Theorem III. We may suppose without
loss of generality that 0 is an inner point of .x;.. Let b be an inner
8*
116 Distance-functions

point of f 2• Then Jt;. and ~ may be described by distance-functions:

Jt;.: 1\ (x) < 1 ,
f 2: ~(x-b)<1.

For ·~·<t< 1 let f/ (i = 1,2) be given by

fi: 1\ (x) ~ t,
Jt1: F2 (x - b) ~ t,
so that :xjt is a closed subset of.Ytf. By Lemma 5 there is a plane rr
separating fi and fi. Since rr does not pass through the point
o Efi, it has an equation

Since t> i, the set fi contains a neighbourhood Ix I~ 1] of the origin,

where 1] > O. Since no points of this neighbourhood lie on rr, we have
(1~i~n). (1 )
Since b is on the opposite side of rr from 0, we have
L Pi/hi> 1. (2)
l~i~"

By (1) and WEIERSTRASS'S compactness theorem, there exist pi which

are the limits of Pit as t tends to 1 through a sequence of values
t1 <t2 <···<tm <··· which is the same for each i. By (2) not all the pi
are O. The plane TI' defined by
L: P; Xi = 1
i
clearly has all the properties required.
IV.3.6. The results of the rest of this § 3 will not be required until
Chapter VIII, but it is convenient to give them here. They show that
any two symmetric convex sets Jt;. and ~ with finite non-zero volumes
behave similarly.
For more precise results, generalisations to convex sets which are
not symmetric, and references to the literature, see for example BAM BAH
(1955 a), and for an interesting application see MAHLER (1955 a, b).
A closed "generalized parallelopiped" in n-dimensional space with
o as centre is the set of all points
(1)

where Xl' ... , x,. are fixed linearly independent vectors and tl , "', tn
run through all real numbers in
(2)
 Convex sets 117

A closed "generalised octahedron" with 0 as centre is similarly the set

of all vectors (1), where tl , ... ,tn run through all numbers in

0)
We first prove the following refinement! of a result of MAHLER
(1939 b).
THEOREM V. Let f be any closed symmetric convex set with volume
V(f) such that 0< V(f) < 00. Then there exist points ±X1 , ... , ±XnEf
such thatf is contained in the parallelopiped'C with faces ±TIJ(1 ~ J ~ n),
where TIJ is the hyperplane through the points xJ±Xj (j=t-j). Further,
the generalized octahedron £& with vertices ± Xi (1 ~ j ~ n) is contained
info
The last sentence is in any case trivial by convexity. We take for
Xl' ... , Xn points of fsuch that the volume of £& is a maximum. Such
a choice is possible since f is closed and bounded. If f were not
contained in 'C, there would be a point y on the opposite side of the origin
from one of the faces ± TIJ' say on the opposite side of TI". Then the
generalized octahedron with vertices ±XI , ... , ±X,,_l' ± Y would have
greater volume than £&, contrary to construction.
COROLLARY 1.
V(f) ~ V('C) ~ n! V(f) ,
V(f) ~ V(£&) ~ (n !tl V(f).
For the left-hand inequalities are trivial, and the right-hand ones
follow from them and V('C)~n! V(£&).
COROLLARY 2. Let f , !l' be any two closed symmetric convex sets of
finite non-zero volume. Then there is a homogeneous linear transforma-
tion ~ of the variables such that
n-l~ ff ( f (n~ ff
and
(n!tl V('t'") ~ V(~ ff) ~ (n!) V(f).
Let Xl' ... , Xn be the points of the theorem for f and let Yl' ... , Yn
be the corresponding points for ff. We determine ~ by the equations
(1~j~n).

Then the 'C, £& of the theorem are the same for f and ~ ff. The stated
results are now trivial, since n-l'C(£&.
1 Suggested by Professor C. A. ROGERS, who disclaims originality. The same
method proves a corresponding result for non-symmetric bodies in which the
inscribed and circumscribed bodies are both simplexes (d. MAHLER 1950a). There
are also results about inscribed and circumscribed ellipsoids (J OH:-I 1948 a).
118 Distance-functions

IV.3.7. As an application of the methods and results of § 3.6 we

prove the following result about the volumes and lattice constants
(Chapter III, § 5.1) of polar convex sets. We again denote the lattice
constant of a set !7 by J(!7).
THEOREM VI. Let f and f* be bounded symmetrical convex sets
which are mutually polar. Then

~ ~ V(f) V(f*) ~ 4n
(n!)2 - -
and
(/)2~ J(f) J(f*) ~ 1,
where (/) is the lattice constant 0/ the octahedron L Ixi I~ 1.
The first pair of inequality is MAHLER'S (1939a, b) and the proof
of the second pair is practically identical. When n =2 MAHLER (1948a)
has determined the best possible inequalities namely
i~J(f) J(f*) ~1,
equality on the left-hand side being necessary when f is a square and
on the right when f is a circle. For related inequalities and references
to later work see BAM BAH (1954c and 1955a).
We now prove the theorem for the lattice constants. The proof for
the volumes is similar. Let -r be any homogeneous linear transformation
and -r* its polar transformation, so
det (-r) det (-r*) = 1. (1 )

The bodies -rf, -r*f* are mutually polar by Theorem III, Corollary 3.
Since

it follows from (1) that

J{-rf) LI{-r*f*) = J(f) LI(f*).
Hence neither the hypotheses nor the conclusion of the theorem are
affected if f is subjected to a homogeneous linear transformation and
f* to the polar transformation.
Suppose first that f = fa is a parallelopiped. After the application
of a suitable homogeneous linear transformation we may suppose
without loss of generality that fo is the unit cube

IXil~1 (1~f~n).
We saw already in § 1.3 that fo* is the generalized octahedron
 Distance-functions and lattices 119

Hence
A (fo) A (JtO*) = A (JtO*) = f/> (2)
by the definition of f/>.
Now consider a general f , which we may suppose without loss of
generality to be closed. Let ~ and !l) be the parallelopiped and octa-
hedron given by Theorem V so that
(3)
The polar of the parallelopiped ~ is an octahedron ~* which is inscribed
in f* by Theorem III, Corollary 2. Similarly the polar of the octa-
hedron !l) is a parallelopiped !l)* and
~*)f*)~*. (4)
We now show that
(5)
where f/> is given in the enunciation. By Theorem V we have
(6)

But every octahedron may be transformed into any other by a homo-

geneous linear transformation, and so the ratio A(~)IV(!l)) is the same
for all octahedra~. In particular, taking !l) to be L IXi 1< 1, we have
LJ(!'J) _ n! r:p
V(!'J) - 2"
Similarly,

and (5) follows from (6).

But now from (3) and (4) we have
L1(f) L1(f*) ~ L1{!l)) L1(~*) ~ f/> L1{~) L1(~*) = f/>2,
on applying (2) with JtO=~. Similarly
L1{f) L1(f*) ~ L1(~) L1(!l)*) ~ f/>-1L1{!l)) L1(!l)*) = 1,
on applying (2) with JtO=~*.

IV.4. Distance-functions and lattices. In the further study of the

relationship between star-bodies (and in particular convex bodies) and
lattices it is convenient to work with distance-functions rather than the
star-bodies themselves. We write
F(I\) = inf F(a) , (1)
aEII
a,*,o
120 Distance-functions

for any distance function F and lattice A. In the language of § 5.1 of

Chapter III the lattice A is admissible for the star-body F(z) < k if
k~F(A) but not if k>F(A).
We have
F(tA) = It IF(A) (2)
*'
for any real number t 0, where tA is the set of ta, aE A. If t> this
follows from the property F(tz) = tF(z) of distance-functions, and if
°
°
t< from the further observation that A contains - a if it contains a.
I n particular,
{F(t A)}" _ {F(A)}"
d(tA) - d(A) , (3)
where n is the dimension of the space. We sum up the properties of
F(A) in the following theorem, which links our present point of view
with that of § 5.1 of Chapter III.
THEOREM VII. For any distance function F write
~(F) = sup {F(A)}" (4)
II d(A)

over all lattices A. Then ~ (F) < 00. Further,

~(F) = {L1(9'n-1 , (5)
where .1(9') is the lattice constant of the star-body
9': F(z) < 1.
If .1(9') = 00, then (5) is to be interpreted as ~(F) =0.

If ~(F) *,0, then the supremum in (4) may be confined to lattices

A with F(A) > 0, and then, by homogeneity, to those lattices with
F(A) = 1. Such lattices are admissible for 9' by definition, and so they
have d(A)~L1(A). This shows that
(6)
On the other hand, if A is 9'-admissible then F(A) ~ 1, and since there
are 9'-admissible lattices A with d(A) arbitrarily close to .1(9'), by the
definition of .1(9'), we must have
~(F) ~ sup {d(An-l~ {L1(9'n-1 • (7)
II is Y·admissible

Then (5) follows from (6) and (7).

If ~(F) =0, then F(A) =0 for every A. From (7) this can hold only
if there are no 9'-admissible lattices, i.e. if .1(9') = 00. Conversely if
.1(9') = 00 and A is any lattice, the lattice tA is not 9'-admissible, for
 Introduction 121

any t>O:
tF(A) = F(tA) < 1.
Hence F(A) =0 on letting t-+ 00.
We note also the rather trivial
LEMMA 7. Suppose that the distance-function F(;£) vanishes only for
;£ =0. Then every lattice A contains a point a=t=o such that F(A) =F(a).
In particular, F(A) > O.
For by Lemma 2 there is then a number c> 0 such that
F(;£) ~ cl;£l.
Hence
F(;£) ~ F(A) +1 (8)
implies that
1;£1 ~ c-1 {F(A) + 1}. (9)
But now by Lemma 1 of Chapter III there are only a finite number
of points of A for which 1;£ 1is less than a given bound, and so there
are only a finite number of points ;£ of A satisfying (9). If we take
a =t= 0 to be one of those points for which F(a) is least, then a enjoys
the properties required.

Chapter V

MAHLER'S compactness theorem

V.I. Introduction. So far we have been concerned with one lattice
at a time. In this chapter we are concerned with properties of sets of
lattices. We first must define what is meant by two lattices A and M
being near to each other; and this is done by means of homogeneous
linear transformations. A homogeneous linear transformation X ="t';£
of n-dimensional euclidean space into itself is said to be near to identity
transformation if the coefficients Tij in
(1~i~n)

are near those of the identity transformation, that is if

(1~i~n)
and
1Tij 1 <
(1 = '< n, 1 <
Z= = J'<
= n, Z' -I-
T J')

are all small. The lattice M is thought of as near to A if it is of the

shape "t' A where "t' is near the identity transformation, and where "t' A
denotes the set of "t'a, aE A. Roughly speaking, M is near to A if it can
122 MAHLER'S compactness theorem

be obtained from A by a small deformation of the underlying space.

Convergence of a sequence of lattices AT to a lattice A' may then be
defined in the obvious way.
MINKOWSKI (1904a and 1907a) already used the idea of the con-
tinuous variation of lattices to show that a bounded convex set
.9: F{x) < 1, (i)

where F(x) is the corresponding distance function, always has a critical

lattice A, in the sense of § 5.1 of Chapter III; that is
F(A,) = inf F{a) ~ 1 , (2)
"EA.
,*,0
and d (A,) is a minimum:
d(A,) =.1(.9) = inf d{A).
F(A)~l

A critical lattice A, has the property that if it is slightly distorted to

a lattice A with d (A) < d (A,) then F(A) < 1 ; that is A has a point other
than 0 in.9. From this, MINKOWSKI obtained important properties of
critical lattices and so gave an explicit process for finding .1(.9) for
convex bodies .9, at least in 3-dimensional space. This was generalized
and put on a more satisfactory basis by MAHLER (1946a), who gives
general conditions under which a sequence of lattices A, should con-
tain a convergent subsequence. In this way he showed that any star
body F(x) < 1 has a critical lattice if only there are any lattices A with
F(A»O. In an important sequence of papers, MAHLER (1946a, b, c, d, e)
extended much of MINKOWSKI'S work on critical lattice to general
star bodies and made other applications of his compactness criteria.
He has also [MAHLER (1949b)] considered the critical lattices of sets
which are not star bodies, but we do not go into this here.
In this chapter we first consider the properties of homogeneous
linear transformations which are needed for the treatment of convergence.
Then we prove MAHLER'S general criterion for a sequence of lattices A,
to contain a convergent subsequence. After that, we study the pro-
perties of critical lattices of sets .9 taking in turn general star bodies,
bounded star bodies, convex sets and spheres. As the sets become more
specialized, there is more and more precise information about the critical
lattices. Finally in § 10 we give an application to a problem in the
theory of Dophantine approximation.
V.2. Linear transformations. Convergence for lattices will be
defined in terms of homogeneous linear transformations, already intro-
duced in Chapter I, § 3. We operate in n-dimensional space with some
fixed euclidean coordinate system. If "Cij is a set of n 2 real numbers,
 Linear transformations 123

we denote by 'tX the transformation of our space into itself given by

the equations
(1;;:;; i;;:;; n),

where X ='tX. We write det('t) =det(Tii). If det('t) =0, the trans-

formation 't is singular: otherwise it is non-singular and possesses an
inverse, which we denote by 't-1 . By a +'t, where a and 't are trans-
formations, we mean the transformation

(a +'t)x =ax +'tx;

and by a't we mean the transformation
(a't) X =a('tx).
If a, 't correspond to the matrices of coefficients (Iii' and Tii' then the
coefficients of a +'t and a't are clearly

and
(1 )

respectively. We denote the identical transformation

by l.
We require a measure of the size of the coefficients of the matrix
of a transformation 'to We write

11't11 = n max ITiil.

Clearly
11- 'tIl = 11't11, }
(2)
Iia +'tll;;:;; Iiall + 11't11·
Further,
Ila'tll;;:;; Ilallll'tll (3)
since the coefficients of a't are given by (i). Further, if X ='tX we
have trivially
, IXil ;;:;; 11't11 max
max , Ix;l· (4)

From this it follows crudely that

(5)
sInce
, Ix,1 ;;:;; Ixl ;;:;; n~ max
max , IXil
for all x.
124 MAHLER'S compactness theorem

We shall also need to use the fact that if ~ is near to the identical
transformation l, then ~-l exists and is also near to l. This statement
is made more precise in the following lemma.
LEMMA 1. Let ~ = l + a be a homogeneous linear transformation with

lioll < 1. (6)

Then ~ is nonsingular and
p= l - ~-l (7)
satisfies
II p II =< 1
Iiall .
-ilail (8)

We note first that if p exists, the inequality (8) follows at once

from (2) and (3). We have
p= ~-lo = a - po;
so

as required.
It remains to show that ~ is nonsingular; and for this it is con-
venient to use another characterization of II~II. Put
~(x)=n-lLlxil, ~(x)=maxlxil· (9)
i 1

Then ~ (x) and ~ (x) are convex symmetric distance-functions vanishing

only at 0, and
(10)
for all x. Then clearly
II~II = sup!2i~1 (11)
.1',.,0 .F;.(or)
for all homogeneous transformations~. For any x we have, by (10),
(11), that
~ (x) = ~ (~x - ox) ~ ~ (~x)+ ~ (ox)
~ ~ (~x) + ~(ox) ~ ~ (~x) + 11 all ~ (x),
the last line by (11) with a for~. In particular, since 11011 <.1 by hypo-
thesis, we have ~ (~x) = 0 only when x = 0: that is ~x = 0 only when
x = 0: so ~ is nonsingular. This concludes the proof.
Our choice of 1I't11 to represent the "size" of 't is somewhat arbitrary. If F
is the distance-function of a symmetric convex bounded body, an alternative would
be to use
(12)

The reader will have no difficulty in verifying that (2), (3) and Lemma 1 continue
to hold when IIIIF is substituted for 1111. Since we have used lorl to denote the
 Linear transformations 125

size of the vector:l:, it might have been more tidy to use lI'tIlF., where 1<;,(:1:) = 1:1:1,
to measure the size of't. We have chosen 1I't1l because of its simpler expression
in terms of the Y'j. The choice of 1111 instead of some IIIIF is, for all essential
purposes, irrelevant, since it follows from Lemma 2, Corollary of Chapter IV that

where c; and c~ are numbers depending on the particular function F, but not on 'to
We shall also need later two lemmas relating to distance functions
and linear transformations.
LEMMA 2. Let F(z) be a distance function such that F(z) =0 only for
;l! = 0, and let 't be a linear transformation. Then there is a number C1
depending only on F and 't such that
F('t z) ~ C1 F(;l!)
for all z.
For

is clearly a distance function. The result now follows at once from

Lemma 2 Corollary of Chapter IV. If 't is non-singular we may apply
Lemma 2 with 't-1 instead of 't and obtain the
COROLLARY. If't is non-singular there is a constant Cz such that

F(;l!) ~ czF('tz).
LEMMA 3. Let F(z) be a distance function such that F(;l!) =0 only for
;l!=0. Then to every 8 in 0<8<1 there is an 1]=1](8»0, depending
only on F and 8, such that
1- £::;:
-
F('t:l:l ::;: 1
F(:I:) -
+£ (13)

for all homogeneous linear transformations 't such that 1

(14)
and all z.
By Lemma 2 of Chapter IV there exists a number c> 0 such that
F(z) ~ clzl (15)
for all z. Since F(z) is continuous in the sphere 1;l!1 ~ 2, there exists
a number 1]1 in 0<1]1 < 1 such that
1F(;l!I) - F(zz)1 < c 8,
whenever

1 As before, L denotes the identical transformation.

126 MAHLER'S compactness theorem

In particular, this is true when 1~21 = 1; and so, by homogeneity,

(16)
whenever
(17)
But now, by (5) and (14),

I't'~-~I =1('t'-l)~I~n!II't'-llll~1 <17I1~1;

provided that n~1J <1JI; which we may suppose.

But then from (15) and (16) with ~I ='t'~, ~2=~' we have
IF('t'~) - F(~)I < tF(~),

which is equivalent to (13).

V.3. Convergence of lattices. If /\ is a lattice and 't' a non-singular

homogeneous transformation, we saw already in Chapter I, § 3 that the
set of 't'a, aE /\ is a lattice 't' /\ with determinant
d('t'/\) = Idet('t')1 d(/\). (1 )

If M is any other lattice, it may be put in the shape

M ='t'/\
for some non-singular homogeneous transformation 't', and indeed in
infinitely many ways. For if aI' ... , an, b l , ... , b n are bases for /\ and M
respectively, there is a uniquely defined homogeneous linear trans-
formation 't' such that
(1~i~n);
and then
M ='t'/\.
We say that a sequence of lattices /\, (1~r<oo) tends to the lattice
I\' if there exist homogeneous linear transformations 't', such that
1\ = 't',/\' (2)
and
(3)

where l is the identity transformation and 11't'11 is as defined in § 2.

We write then
/\, -,>-/\'.

From (1) and (3) we have immediately

d (/\,) -'>- d (1\') .
 Convergence of lattices 127

If « is any non-singular homogeneous linear transformation it is also

immediate that
«/\, -+ «I\' .
Indeed
«/\, = «'t, «-I {«I\'}
and

so

by (3) of § 2.
LEMMA 4. A necessary and sul/icient condition that the se.quence 01
lattices /\, (1;:;;; r < 00) tends to I\' is that there exist bases
b~, ... , b~,
and
b~, ... , b~
01/\" I\' respectively, such that
bj -+ bi (4)
The last limit is meant, of course, in the sense of the ordinary con-
vergence of vectors: Ibi - bi 1-+ O.
The proof of Lemma 4 is almost trivial. Suppose first that /\,-+1\'
and let 't, be the transformation satisfying (2) and (3). Choose any
basis bi for I\' and put
bj='t,bj (1;:;;;j;:;;;n; 1;:;;;r<00). (5)
Then by (5) of § 2 and (3), we have
Ibi-bil =1('t,-L)bil;:;;;nill't,-Llllbil-+o (r-+oo).
Suppose conversely that the bases are given satisfying (4). We may
define't, uniquely by (5). Then clearly 11't,-LII-+o.
The following criterion is rather less trivial.
THEOREM 1. A necessary and sul/icient condit£on that /\,-+1\' is that
the lollowing two conditions be both satislied:
(i) il a' E1\', there are points a' E/\, lor r = 1, 2, ... such that
a' -+a' (r-+ 00). (6)
(ii) il c is not in 1\', there is a number 'Y}>O and an integer ro>O,
both depending on c, such that
la'-cl>'Y} (7)
lor all a'E/\, with r;;:;;ro.
128 MAHLER'S compactness theorem

It is quite straightforward that (i) and (ii) are satisfied when I\,--+/\'.
In (i) we have only to put
a' = 't,a' ,

where the 't, are the transformations such that

1\, = T, /\', 11't, - LII --+ o.
Then, as before,
II a' - a' II ~ n~ II T, - LIII a' I --+ 0 (r --+ 00) .

To prove (ii), we note that there certainly is an 'iJl> 0 such that

la' -ci >'iJ1 (8)
for all a' E/\'. Put
'iJ =1'iJ1' (9)

Suppose, if possible, that there is a point a' E1\, such that

la'-cl~1'}· (10)
Then
(11)
By the definition of T, we have
a' = 't,a' (12)
for some a' EI\. Then
a' - a' = p,a', (13)
where

Now
II p,II--+ 0 (r --+ 00)
by Lemma 1 and since IIT,- LII--+O. Hence by (5) of § 2 and (11), we
have
Ia' - a' I ~ n& II p, III a' I ~ nt II p, II {I c I + 1'}} < 1'}

for all r greater than some roo From this and (9) and (10) we have
la' - cl ~ 21'} =1'}1'
This is in contradiction to (8). Hence statement (ii) of the theorem is true.
We must now show that if (i) and (ii) of the theorem are true then
I\,--+/\'. We require a lemma of some independent interest.
LEMMA 5. Let C1 , ... , c n be linearly independent points 0/ a lattice 1\
but not a basis. Then 1\ contains a point
d = {}l c1 + ... + {}ncn,
 Convergence of lattices 129

where fA •...• fJ" are numbers such that

i ~ max
I
IfJil ~ i· (14)

We first prove Lemma 5. Since c1 • .... c n is not a basis. there cer-

tainly exist points
a = (Xl C 1 + ... + (X" en
in /\ for which (Xl . . . . . (Xn are not integers. We may suppose without
loss of generality that
(1~j~n).

Let t be the least non-negative integer such that

zt max I(Xi I ~ i·
I
Then

and
d = ia
will do what is required. A slight refinement of the argument. which
is left to the reader. shows that the i in (14) may be replaced by! but
by no larger number.
We now revert to the proof of Theorem I. Suppose that /\, and I\'
satisfy (i) and (ii). Let b~ • .... b~ be any basis for 1\'. By (i) there exist
sequences of points
bj_bj (1~j~n. bjE/\,). (15)
We show that bj (1 ~j~n) is actually a basis for /\, except. possibly.
for a finite number of r. For if bL .... b~ is not a basis for /\'. let
(16)
be a point of /\ with
(17)

which is given by Lemma 5. Since the fJi , are bounded. they contain
a convergent subsequence by a classical theorem of WEIERSTRASS (d.
§ 1.2 of Chapter III). say
(18)
where
r 1 <r2 < .. · <r,< .. ·
is an increasing sequence of integers. Then
d' (say) = LfJ;b; = limd".
i '-+00
Cassels, Geometry of Numbers 9
130 MAHLER'S compactness theorem

by (15), (16) and (18). Hence d'eA' by (ii) of the enunciation of the
theorem. This is a contradiction since
J.::;;
4-.
max If)'1 ::;;.l
I - 2'
1

by (14) and (15), and since bj (1 ~i~n) was defined to be a basis for A'.
The contradiction shows that bi is a basis for A, except for a finite
number of r. If the bi are changed for these exceptional r so that bi
(1 ~ i ~ r) is a basis for A, for all r this does not affect the limits (15).
Hence the criterion is certainly sufficient by Lemma 4.
V.3.2. In Chapter X we shall need the notion of a neighbourhood
of a lattice, and we shall mention it again in passing briefly in § 9 of
this chapter.
A set £ of lattices A is said to be a neighbourhood of the lattice M
if it contains all lattices
A =-rM (1 )
with

for some TJ> 0 depending on the particular neighbourhood. The neigh-

bourhood £ may contain other lattices A than those given by (1) and
(2); but there is some TJ> 0 such that it contains all these. If II is any
non-singular homogeneous transformation we show that the set II £ of
lattices rJ.A, AE £ is a neighbourhood of II M. Indeed II £ contains all
lattices
N = cr(IIM)
with

since then
N =IIA
where

and then

as in § 3.1.
Clearly the sequence A, (1 ~r<oo) of lattices tends to M if and
only if every neighbourhood of M contains all but a finite number of
the A,.
Although we nowhere use it, we note that it is in fact possible to
introduce explicitly a metric into the space of all lattices. Let A and M
be two lattices and let
 Convergence of lattices 131

where the infima are over all non-singular a and 't such that
/\=aM M='t/\.
Put
O(M,/\) = 0(/\, M) = max {log (1 +,u),log(1 +v)}.
Then we have the triangle inequality
0(/\, N) ~ 0(/\, M) + O(M, N);
since if
/\ = (l + PI) M, M= (l + P2) N;
then
/\ = (l + Pa) N,
where
II Pall = II PI + P2 + PI P211 ~ II PIli + II P211 + II PIli II P211 ;
and so
log (1 + II P31 \) ~ log (1 + II PI I\) + log (1 + II P211) .
The neighbourhood defined above is the one associated with this
metric, since if
/\=aM
with
lI a -lll< 1J<1;
then

where

and so
0(/\, M) <_11_.
1-11
V.3.3. The continuity of the distance-function F(z) of the vector z
is reflected as a semi-continuity of the function
F(/\) = inf F(a) (1 )
aEII
*0
of the lattice /\ considered in § 4 of Chapter IV. For certain later
applications it is useful to allow the distance-function F and the lattice /\
to vary simultaneously.
THEOREM II. Let /\, (1~r<oo) be a sequence of lattices tending to
the lattice 1\'. Let .F;(z) (1~r<oo) be a sequence of distance functions
which converge uniformly to the distance-function F' (z) in the unit sphere
Izl<1. Then
F'(I\') ~ lim sup F, (/\,). (2)
'--"00
9·
132 MAHLER'S compactness theorem

The proof is very simple. Since F,. (tx) = tF,. (x) for t> 0, the con-
vergence of F,.(x) to F'(x) is uniform in any bounded set of points;
in particular, since the distance-function F'(x) is continuous by defini-
tion, if a, is any sequence of points converging to a point a', we have
lim F,.(a,) =F'(a').
' ..... 00

But by Theorem I, every point a' =F 0 of N is the limit of points a,=F 0

of /\,. Hence
F'(a') = lim F,.(a,) ~ lim sup F,.(/\,) ,
1-+00 1-----+-00

since F,. (a,) ~ F,. (/\,). The result (2) now follows from the definition (1).
The sign of equality need not hold in (2) even when F,. =F' for alI r,
but we defer giving an example until § 10.5. However, much more
than Theorem II is true if F'(x) = 0 only for x = 0, i.e. if the set F' (x) < 1
is bounded (Lemma 2 of Chapter IV).
COROLLARY. Suppose that the hypotheses 01 Theorem I I hold and that
the only point x such that F'(x) = 0 is x = o. Then
lim }~(/\,)
' ..... 00

exists and is equal to F'(N).

The proof is similar to that of Lemma). By Lemma 2 of Chapter IV,
there is a c> 0 such that
(3)
for alI x. Let e> 0 be arbitrarily small. By the uniformity of the con-
vergence of F,.(x), there is an ro such that
1F,.(x) -F'(x)! <ce (4)

for all r ~ r0 and all x with Ix I = 1. Hence for all x whatsoever and
r~ro, we have
1F,.(x) -F'(x)1 <celxl ~eF'(x);
so
F,(x)
1 - e < F'(:E) < 1 + e. (5)

Now let /\,="t,/\', where "t, are homogeneous linear transformations

such that
(r~ 00)
in the language of § 2. Then
1- < F'(-r,:E) < 1+ (6)
e F'(:E) e
 Convergence of lattices 133

for all r greater than some rl , by Lemma 3. Hence by (6) and (5) with
"C'r~for ~ we have
(1 - e)2< F,(~,a:) < (1 + e)2
F'(a:)

for all r> inax (ro' r1). But now A, is just the set of "C',~ with ~EN,
and SOl
(1 - e)2~ F,(I\,) ~ (1 + e)2.
- F'(i\') -

Since e is arbitrarily small, this proves the corollary.

V.3.4. An almost immediate consequence of Theorem II, Corollary
is the following result, which shows that no bounded star body can
have successive minima in the sense of Chapter II, § 4.
LEMMA 6. Let F(~) be an n-dimensional distance function which
vanishes only when ~ = 0 and let 1] be any number for which

0<'7 < d (F) = sup

A
{F(I\)}"
d(l\) .
(1)

Then there exists a lattice M'I such that

{F(M'I)}" = 1] d (M'I) .
After Theorem VI we shall be able to replace the second < in (1)
by ~.
Suppose that 1] satisfies (1). Then there exists a lattice N, such that
{F(N)}" > 1] d (N) . (2)
Let b 1 , ''', b" be any basis for N; and for 0 < e < 1 let N. be the lattice
with basis

Then
d(N.) = ed(N)
and

Hence
(e ~ 0).

But now, by Theorem II Corollary, F(N.) is a continuous function of e.

Hence by (2) and (3) we may put M'I= N. for an appropriate value of e.
V.3.5. For the sake of completeness we enunciate the following
lemma, which interprets the uniformity of the convergence of F, (~) to
1 Note that F'(i\') '*' 0 by Lemma 7 of Chapter IV.
134 MAHLER'S compactness theorem

F'(~) in terms of the corresponding star bodies

.9;: F,(~) <1 (1)
and
f/': F'(~) < 1. (2)
Since we do not use the lemma we do not give the proof, but the reader
should have no difficulty in constructing one along the lines of the
proof of Theorem I of Chapter IV.
LEMMA 7. A necessary and sufficient condition that the sequence of
distance functions F,(~) tend to the distance-function F'(~) uniformly in
I~ I~ 1 is that the bodies .9; and f/' defined by (1) and (2) have the following
properties:
(i) If c is an (inner) point of f/' , then there exists an 'fJ> 0 and an
integer ro (depending on c) such that all points ~ of the neighbourhood
I~-cl <'fJ belong to .9; for all r greater than roo
(ii) If c is an exterior point to f/' (i.e. F' (~) > 1) then there is an
'fJ> 0 and r0 such that no point ~ of the neighbourhood I~ - c I< 'fJ belongs
to .9; for any r>ro'
VA. Compactness for lattices. In this section we are concerned
with conditions under which an infinite sequence /\, of lattices should
contain a subsequence Mt =/\" which converges to a lattice M/, not
necessarily belonging to the sequence.
The simplest such condition is when every lattice of the sequence
has a basis every point (If which lies in some fixed sphere
I~I;;;;R (1)
and d (/\,) is bounded below by a positive constant, say
d (/\,) ~ " >0 (all r) . (2)
Since all the lattices have bases in (1) we may by WEIERSTRASS' com-
pactness theorem, find a subsequence of lattices Mt = /\" with bases
bL ... , b~ in (1) such that all the limits
lim
t-+oo
b~
1
= b~1
exist. By (2) we have
Idet(b~, ... ,b~)1 = lim Idet(bL ... ,b~)1 = lim d(M,)~,,>o:
t-+oo t-+'Xl

and so b~, ... , b~ are linearly independent. Hence there exists a lattice
M' with basis b~, ... , b~ and, by Lemma 4,
(t~oo).
 Compactness for lattices 135

A slight extension of this idea gives the following theorem which

however turns out not to be very useful. We give it partly for historical
interest and partly because the lemma on which it depends will be used
later.
THEOREM III. Let A, (1 ~r<oo) be an infinite sequence 0/ n-dimen-
sional lattices enjoying the following two properties:
( i) there exists an R such that every A, has n linearly independent
points in the sphere

(ii) there exists a x> 0 such that

d(A,) ~ x
for all r.
Then A, contains a subsequence of lattices Me for which
M'= lim Me
1->00
exists.
The proof of Theorem III depends on the following lemma due to
MAHLER (1938a) and rediscovered by WEYL (1942a).
LEMMA 8. Let F(x) be any symmetric convex distance function and
al an be n linearly independent points of a lattice A. Then there
, ... ,

exists a basis bl , ... b n of A for which

I

F{bi ) ~ max [F(aj ) !{F(al ) I + ... +F(ai)}J.

Before proving Lemma 8 we show that Theorem III follows from it
by applying it to the convex function F(x) = Ixl and to the n linearly
independent points a~, ... , a~ of A, given by (i) of the theorem. Then
Lemma 8 shows that A, has a basis bJ:, ... , b~ with
Ibtl ~ max [Iail, Hlarj + ... + lail}] ~ nRj2.
We have thus reduced Theorem III to the trivial case discussed at the

I
beginning but with nRj2 instead of R.
It remains to prove Lemma 8. By Theorem I of Chapter I there is
a basis Cli •.. , c n of A such that

al=vllcl •
a2 = Vn c1 + V 22 C 2 '
(3)
a" = Vn1 c l + ... + v"ncn'
where the Vii are integers and Vii=F o. We shall take bi of the shape
(4)
136 MAHLER'S compactness theorem

where the tj i are numbers to be determined. Clearly 1)1' ... , b" is a

basis for 1\ for any set of numbers tj; such that bjE I\.
We distinguish two cases for each i. If vii= ± 1, we put b j = ± a j •
This certainly has the shape (4) and
F(bj) = F(aj).
Otherwise Iviii;;;:; 2. On solving (3) for the c j we have
c·1 = v~la·
11 1 +k . 1a1- 1
1.1- + ... + k·
]1-»
a_ (5)
where kj; are certain real numbers. Choose tji in (4) to be integers such
that

(6)
where

and
(i <i).
Then by the convexity symmetry and homogeneity of F(x) we have

F(b j) ~ F(liiaj) + ... + F(lj1 a1)

= IIii IF(aj) + ... + Ilj1IF(~)
l (7)
~ HF(aj) + ... + F(~)}.
This concludes the proof of the lemma.
When F(x) is the usual euclidean distance, an argument due to
REMAK (1938a) gives a sharper result. See also VAN DER WAERDEN
(1956a).
When
(8)
Lemma 8 gives
F(b j) ~ max (1, i)F(aj ).

V.4.2. We owe to MAHLER (1946d, e) a criterion for the existence

of a convergent subsequence of lattices in a sequence of lattices, which
is much more fertile of applications than Theorem III, and which may
be said to have completely transformed the subject. MAHLER proved
his criterion by using the theory of successive minima l of a sphere to
show that it is equivalent to that of Theorem III. We shall give
1 Not to be confused with the "successive minima" discussed in Chapter II
which are quite different.
 Compactness for lattices 137

MAHLER'S argument 1 when we discuss successive minima in Chapter VIII,

but here we give a direct treatment due to CHABAUTY (1950a), who
shows that it generalizes significantly to a more general situation (sub-
groups of locally compact topological groups). MAHLER'S criterion is
expressed in
THEOREM IV. Let /\, be any infinite sequence of lattices satisfying the
following two conditions
(i) d(/\,)~K for all lattices /\" where K is independent of r.
(ii) I/\,I~){>o for all r where){ is independent of r and, as usual.
1/\ 1= inf Ia I.
"Ell

Then /\, contains a subsequence M, = /\" which converges to a limit M'.

We prove 2 Theorem IV by induction. The result of the j-th stage
(1 ~ j ~ n) will be the following statement:
I$i: There exist j linearly independent points a 1 , ••• , ai and a sub-
sequence Nt = Nt (1 ~ t < (0) of /\, which satisfies the following conditions:
1$;: Each point a i (1 ~i~j) is the limit of points
alE Nt = Nt (1)
1$; I: Suppose that t1 < t2 < ... is any increasing sequence of integers
and there exist points Ct,E Nt, such that
lim ct =Y1a1 + ... +y·a·
s~oo ' 1 1
(2)
with real Y1' ... , Yi' Then Y1' ... , Yi must be integers.
Before continuing the proof we note that the statement 1$" implies
that the lattices Me = N: converge to the lattice M' with basis Ut, ... , an;
the parts I$~ and I$~' of corresponding respectively to (i) and (ii) of the
Theorem 1. Hence it suffices to prove I$n.
We do not give a separate proof of 1$1 since that is a simple version
of the deduction of 1$i+1 from I$j. For the rest of this section we shall
assume therefore that I$i holds for some j in 1 ~j < n and will deduce
1$i+1' The sequence N;+l will be a subsequence of the sequence Nt = Ni,
and the points Ut, ... , a j will be the same in I$i and I$j +1'
A non-singular homogeneous linear transformation of the variables
does not affect either statement I$i or the hypotheses of the theorem,
though it will in general replace the K and ){ in (i) and (ii) by different
numbers. Hence we may suppose without loss of generality that
i-I n-i

Ui = e = (0.:-::0, 1,~)
j (1~i~j).

1 The reader may prefer, instead of studying the proof here, to turn to §§ 1, 2
of Chapter VIII, which are independent of the intervening matter.
2 I now prefer the proof given by CHABAUTY (1950a) to the version given here.
138 MAHLER'S compactness theorem

Define the number 1p by

(4)
where K is the number occurring in hypothesis (i) of the theorem. By
Theorem III of Chapter III, each lattice Nt contains a point x =t= 0 with I

Ix;! ~f (1~i~i) }
(5)
IXil ~ 1p U+1~i~ n).
Let ct be one of the finite number of points of Nt other than 0 in (5)
for which
max Ix·11
j+l~i~n
(6)

is a mInImum. Since the c t are in the bounded set (5), they contain a
convergent subsequence, say
(7)
where

Write
(8)
so that clearly
IAil ~ f (1~i~j) }
(9)
IAil ~ 1p(i + 1 ~ i ~ n).
Suppose first, if possible, that AiH = ... = An = 0, so that ai+I is linearly
dependent on a l , ••• , a i . We are assuming statement @3i to be already
established. Hence by (7) we could apply @3;' with Yi=A, (1 ~i~i)
and it would follow from (9) that Al = ... =Ai=O, and so
lim ct, = o.
S--+ co

This contradicts hypothesis (ii) of the theorem, since ct'E Nt, = /\, for
some rand ct,=t= o. Hence the vectors aI' ... , ai +1 are linearly inde-
pendent. We put r. = Nt" and will show that the statement @:ii+I
now holds for N~+! = r..
The statement @:i;+l is trivially true. So far as ai-II is concerned,
@:if+I follows from (7); and so far as the remaining ai (1 ~ i ~i) are con-
cerned, @:if+! follows from @:ij since r.is a subsequence of Nt.
It remains to prove @:i;~l' Suppose, if possible, that there is an
increasing sequence of integers
(10)
and vectors
dS",!O[
>,.
1 The only property of !- we use is t < ! < 1.
 Compactness for lattices 139

such that
lim d'm = d (say) }
t-+oo (11 )
0lal + ... + OJ+1 a j+l'
°
=

where 1 , ... , OJ+1 are not all integers. By ®f+1' which we have already
proved, we may add integer multiples of aI' ... , a j +1 to the right-hand
side of (11), after appropriate modification of the sequence d S'". Hence
we may suppose in the first place that
IOJ +1 I ~ i (12)
and in the second place, by (3). that
Idil~i<! (1~i~j), (13)
where, as usual, d=(d l , ... , dOl)' From (8) and (12) we have

. max Id;I=IOj+1l.max
J+l~.~n J+l~.~n
IA;I)
(14)
~i. max IA;I
J+lS.~n

(15)
We now show that this in contradiction with the definition of the vectors
ct as the vectors x of Nt in (5) other than 0 for which (6) is as small
as possible. Since ct· --+ a j +1' we have
lim max ICit I = max IA;I, (16)
r---+oo j+l~i~n ' j+l~i~n
where

By (13) and (15) the vector d Sm certainly lies in the region defined by
(5) when m is large enough. Further, dSmE NT, where T =tsm' But
now, by (14) and (16), the function (6) is certainly greater for c T than
it is for d Sm when m is large enough, which contradicts the choice of CT.
°
The contradiction shows that if (11) holds then 1 " " , OJ +1 are all
integers; that is the statement ®;~l holds.
This ends the deduction of ®j -1 1 from ®j' and so concludes the proof
of Theorem IV.
We note a form which is often useful in applications and which does
not depend on the use of the special distance-function Ix I.
COROLLARY. Let F(x) be any distance function and let A, be any
intinite sequence at lattices satisfying the two conditions
(i) d(A,)~K tor all r, where K is independent at r.
(ii) F(A,)~x>o for all r, where x is independent at r and, as usual,
F(A) = inf F(a).
aE/\

Then A, contains a convergent subsequence.

140 MAHLER'S compactness theorem

For by Lemma 1 of Chapter IV there is a C> 0 such that F(3J) ~ C 13J I'
and so
IA,I ~ C-IF(A,) ~ C-I" > o.
V.4.3. An almost immediate consequence (d. MAHLER 1949a) of
Theorem IV is
THEOREM V. Let !I' be any open set. Let~,~, ... , ~, ... be a
sequence at open s1-tbsets at !I' such that
(i) ~ is contained in 9; it r < t,
(ii) the origin is an inner point at .9i,
(iii) every point 3J ot !l'is in ~ tor some r.
Then
LI(!I') = lim LI(~).
,-+00

We recall that. LI(!I') is the lower bound of d(A) over !I'-admissible

lattices A, i.e. A having only 0 in!l'. Clearly

LI(~) ~ LI(!I') (1 )
for all r. Suppose that
lim inf LI(~) < LI(!I'). (2)
'-+00

Then there is an increasing sequence of integers r l <r2 < .. · and

lattices A" such that
lim d(A,,) < LI(!I');
1-+00

and A" is ~,-admissible. By (ii) and Theorem IV we may extract a

convergent sequence of lattices from the A", so that without loss of
generality
lim A" = N; d(N) < LI(!I').
1-+00

Hence there is a point p =l= 0 of N in!l'. By (iii) then p is in !l'R for

some R. By (i) and since ~ is open by hypothesis, there is a neigh-
bourhood
13J-pl<17 (3)
every point of which is in ~ for all r~R.
But now
p = lim p', p'EA,
'-+00

by Theorem I. Hence 3J =p' satisfies (3) for all r greater than some ri .
For r> max (R, r1 ) this means that p, is in !I' contrary to our assumption.
The contradiction arose from the assumption that (2) is true. Hence
the theorem is true by (1).
 Critical lattices 141

When the //' and .~ are star-bodies, Theorem V follows fairly

immediatelyl from Theorem II but we shall in fact apply Theorem V
when !/ is not a star-body in Chapter VIII. The proof of Theorem V
gives also the following corollary which is a trivial consequence of the
theorem when the //' and .~ are star-bodies, but which is valid when
they are not.
COROLLARY. Suppose that lor 1 ~ r< 00 there is an fI',-admissible
lattice Ar with
d(A,) = .1 1 ,
lor some number .1 1 ' Then there is an f/-admissible lattice A with d (A) =.1 1 ,
V.S. Critical lattices. Let F(~) be a distance-function. It may
well be that F(A) = 0 for every lattice A, in which case we say, following
MAHLER, that the distance function and its associated star-body are of
"infinite type". An example of a distance function of infinite type in
two dimensions is
F(~) = JX~X2J!.
Any lattice A of determinant d (A)
with
= d contains a point a = (a 1 , a2 ) * 0

where e> 0 is arbitrarily small, by MINKOWSKI'S convex body Theo-

rem II of Chapter III. Then
F(a) ~ JedJ!
is arbitrarily small, so F(A) = o. It is not always possible to decide
whether a distance function is of finite type or not, for example, this
is not known in the case of the 5-dimensional distance functions

and
F(~) = Jx~ + x~ + x~ - x: - x:J~.
The problem whether these functions are of infinite type or not is equi-
valent to the problem whether all indefinite quadratic forms in 5 vari-
ables represent arbitrarily small values (including 0) or not for integer
values of the variables (d. § 3 of Chapter I). A classical theorem of
MEYER says that if the coefficients of the form are rational then it
represents O. Recently DAVENPORT and more recently B. J. BIRCH have
developed an attack on this problem but it appears to work only for
1 When the !I' and ~ are star-bodies, say, with distance-functions F(a:) and
Fr (a:) , the hypotheses of Theorem V imply that, for each x, Fr (a:) tends monotonely
to F(a:). Since Fr(a:) and F(a:) are continuous, this convergence must be uniform;
and so Theorem II applies.
142 MAHLER'S compactness theorem

indefinite forms in more variables than 5 [see DAVENPORT (1956a) and

later work of DAVENPORT and BIRCH]' The results of Chapters VI and X
sometimes permit one to decide whether a given distance function F(x) is
of finite type or not but beyond that very little is known. For another
unsolved problem of this type with important implications see CASSELS
and SWINNERTON-DYER (1950a).
Most of the investigations in the geometry of numbers are concerned
with distance-functions F of finite type, i.e. not of infinite type. Then

r5(F) = sup {F(!\W (1)

/I d(!\)

s strictly positive. Then by Theorem VI of Chapter IV,

0< r5(F) < 00 (2)

and
r5(F)L1(Y) =1, (3)
where L1(Y) is the lattice constant of the set
Y: F(x)<1. (4)

We recollect that a critical lattice for Y is a lattice 1\ which is

Y-admissible and which has determinant d (1\) = L1(Y) (Chapter III, § 5).
A general theorem of MAHLER states that a set Y of the type (4) always
has critical lattices if it has admissible lattices.
THEOREM VI. Let the distance-function F(x) be of finite type. Then
there exist lattices 1\ such that
F(I\) =1, d(l\) = {r5(F)}-l = L1(Y) ,

where r5 (F) is defined in (1) and L1 (Y) is the lattice constant of the region
defined by (4).
The proof of Theorem VI is now quite simple. By the definition of
L1(Y), there exists a sequence of lattices A, such that
F(A,) ~ 1, d (A,) ~ L1(Y). (5)

We may now apply Theorem IV Corollary 1, its conditions (i) and (ii)
being satisfied by (5). Hence there exists a convergent subsequence,
and so, after a change of notation, we may suppose that
A,~N

for some lattice N. By (5) we have

d (N) = lim d (A,) = L1 (Y).
,->00
 Critical lattices 143

By (5) and Theorem II. we have

F(/\,) ~ lim sup F(I\,) ~ 1 .
,-)000

If F(/\,) > 1 there would exist a real number ·0< 1 such that

in contradiction to the definition of £"1(9') as a lower bound. Hence

F(/\,) = 1. This concludes the proof of the theorem.
In evaluating £"1(9') for star-bodies 9' we may therefore confine
attention to critical lattices.
There is an alternative formulation of Theorem VI which does not
need to distinguish between the two cases 0 (F) = 0 and 0 (F) > 0:
COROLLARY. For every distance-function F(x) in n-dimensional space
there is a lattice M such that
d(M) = 1
and
{F(MW = 0 (F) = s~p {~((~r •
For if 0 (F) = O. any lattice M with d (M) = 1 will do. Otherwise
M =(}/\' will do. where I\' is a critical lattice and {} is chosen so that
d(M) =1.
V.S.2. It would be natural to assume that every critical lattice 1\
for a star-body
9': F(x) < 1
should contain a point a with F(a) = 1. but in fact this is not the case
even in 2 dimensions. Here we construct a counter-example using the
phenomenon of successive minima discussed in § 4 of Chapter II. Write
(1 )
Theorem IV of Chapter I when translated into our present language
implies that
{Fa (1\)}2~ d (1\)/8~ (2)

except when 1\ is a lattice 1\, with basis

a 1 = (all. an) • a2 = (a 12 • au)
such that

identically in u 1 • U 2 for some number k; in which case

(5)
144 MAHLER'S compactness theorem

In particular
(6)
Now consider the distance-function

(7)
so that
401
Po (:Il) ~ 1\ (:Il) ~ 400 Po (:Il) . (8)

r
From (8) and (2) or (5) we have

{1\ (A)}2~ (:~~ d (A)/8' (9)

jf A is not a Ac; and
(10)
respectively. Since
8-& (401)2 < 5-&
400 '

a critical lattice for 1\ (:Il) < 1 is necessarily a Ac.

We show now that equality holds in (10). After a possible inter-
change of Xl and X 2 we may suppose that
UI all + U 2 au = all (u l + w u 2)
UI au + U 2 au = au (ul + 11' u 2) ,
where
2w=1+5', 211'=1-5', k=all a2l ,
on factorising the right-hand side of (4). Here
wtp = -1.
Since

we have

for every positive integer t and certain integers u~), u~). Hence
11 (say) = (all w', a21 vi) E Ac·
But now, since wtp = -1, we have

F.I (.,1)
N -
- Ia11 aU Ij [1 + 100{i auiIauwI aui
+ au w 1}2
1~ IaII aU Ii (t ~ 00)
=kl.
Hence
 Bounded star-bodies 145

This with (10) gives

But now if a+o is a point of Ac ' we have trivially

{1\ (a)}2 > {F.,(a)}2 ~ d (Ac)/5~.

In particular, if k=1, so that d(Ac)=5t=LI(~), where ~ is the
region 1\ (z) < 1, there are no points a of Ac on the boundary 1\ (z) = 1
of Ac •
By an ingenious argument, again using the phenomenon of successive
minima, ROGERS (1947c) has constructed a distance-function F(z) such
that the critical lattice of the unbounded star-body F(z) < 1 has only
one pair of points ±a with F(±a) =1. All other points b+o of A
satisfy F(b) ~ t for some explicitly given t> 1. This is in striking
contrast with the results we shall prove in § 6 about bounded star-bodies.
V.6. Bounded star-bodies. For bounded star-bodies a great deal is
known about critical lattices. [See in particular MAHLER (d, e) and for
an extremely detailed treatment of the 2-dimensional case MAHLER
(a, b, c).] In contrast to the negative result of § 5.2 we now have
THEOREM VII. Every critical lattice A 01 a bounded star body !/ has
n linearly independent points on the boundary 01 !/.
For suppose not. Then there exists a basis b1 , ..• , b n of A such that
any point
(U1 , ••• , un' integers) (1 )

of A on the boundary of !/ has u" = O. Since!/ is bounded, there exists

a number Y such that if a point

with real Yl' ... , Yn is in or on the boundary of [/, then certainly

(1~i~n).

Now let e be a number in, say,

and let A. be the lattice with basis

b1, ... ,bn - 1 and (1-e)b n.

Consider a point
P. = u1b1 + ... + un-1bn- 1 + u,,(1 - e) b ll (2)
Cassels, Geometry of Numbers 10
146 MAHLER'S compactness theorem

of A., where ttl' ••• , Un are integers. If Un = O. then Pc is in A; and so

is either on the boundary of Y' or outside !7. If

m~x Ittl I > 2 Y,

1::;;1="
then certainly P. is outside !7. We need therefore consider only the
points with
max Iu;1 ~ 2 Y, tt" =t= O. (3)
But now for these Uj the corresponding point p given by (1) is an
exterior point of !7, since u.. =l= 0; that is some whole neighbourhood
of p lies outside!7. Hence Pc cannot be in !7 for all e smaller than
some eo, which may depend in the first place on u l , ... , U ... But there
are only a finite number of U l •••.• U" to consider. by (3). and hence
A. is !7-admissible if e is small enough. But now
d(AE) = (1 - e) d(A) = (1 - e) L1(!7).
since A was assumed to be critical. But this contradicts the definition
of .1 (!7) as the lower bound of the determinants of admissible lattices.
It is only exceptionally that there can be as few as n pairs of linearly
independent points ±a; (1~i~n) of a critical lattice on the boundary
of!7. Rather surprisingly. it is possible, however, at least when n = 2.
for a star-body to have a continuous infinity of critical lattices each
with only n pairs of points on the boundary, see OLLERENSHAW (1945 a).
COROLLARY. Suppose that ± a; (1 ~ i ~ n) are the only points 0/ A
on the boundary 0/!7. Then there exists an eo such that all points
(4)
with
(5)

are either in or on the boundary 0/ !7.

For "t, ... , a.. are linearly independent by the theorem; and so
there exists a basis b l , ... , b.. such that
(1~i~n) (6)
with integers Vii and vi.=t=O. Let A~ be the lattice with basis

where
(7)
and 'fIl • ••• , 'fI .. - l are small real numbers. As in the proof of the theo-
rem, if max 1'fI;1 is small enough. the only points of A~ which can lie
in or on the boundary of !7 are ±al ..... ±a,,-l (which are unchanged
 I30unded star-bodies 147

by the substitution of b~ for b,,) and ±a:, where

a~ (say) =v"lb 1 + ... +l'",,,-lb"_l +v""b~. (8)
But
d(Aq) =ldet(bl, ... ,b"l,b~)1 =ldet(b1 , ... ,b,,)1 =d(A) =.1(9').
Hence either a:is in .~, when there is nothing more to prove, or A~ is
critical, and then a: is on the boundary of 9' by the theorem. Since
every vector of the shape (4) can be put in the shape (8), where max l17i I
is small if max I8il is small, this proves the corollary.
V.6.2. For the continued study of the points of a critical lattice
on the boundary of a bounded star-body, we need an estimate of
det (aI' ... ,a,,) in terms of

where "t, ... , a" are any n-dimensional vectors. For our present purposes
any estimate, however crude, would suffice, but, since we shall later
need a more precise estimate, we prove it here.
LEMMA 9 (HADAMARD). Let aI' ... , an be n-dimensional vectors. Then

Idet (aI' ... , an) I ~ Iall ... Ian I·

We note that the simple example
;-1 n-;
a;=e;=(~,1,~)
shows that ~ cannot in general be improved to <. The inequality is
the n-dimensional analogue of the fact that the volume of a parallelo-
piped is at most the product of the length of the sides.
If the determinant is 0 there is nothing to prove. Hence we may
suppose that a 1 , ... , a" are linearly independent. We construct a
sequence of vectors c; (1 ~i~n) such that
C,C; =0 (i 4= i) (1)
(scalar product of two vectors), and
a. = til C1 + ... + ti,,-l Ci - 1 + c. (2)
for some real numbers t,;. Indeed if c1 = a 1 and the ci are defined
recursively by
C, = at - ~ (a i c;) IC; 1- 2Ci '
;<i
it is readily verified that the c, have the required properties. By (1)
and (2) we have
Ia.12 = aiai = t~ll cl l2 + ... + t~.i-ll Ci _ 1 12 + Icil2~ IC,12, (3)
and
(4)
10·
148 MAHLER'S compactness theorem

On the other hand. on regarding the c1 • •••• c .. in det(c1 ••••• c.. } first as
rows and then as columns and multiplying the two determinants to-
gether. we have 1
{det (c1 • •••• C ..)}2 = det {cjci } = II Ic,12. (5)
by (1). The required inequality now follows from (3). (4) and (5).
V.6.3. We may now show that. in principle. the evaluation of LI(.9')
for a bounded n-dimensional star-body .9' may be reduced to a finite
set of ordinary minimal problems. Except for convex bodies, for which
see § 7. this is hardly in practice a fruitful approach. though it might
well be adaptable to machine computation.
We may suppose without loss of generality that .9' is defined by
.9': F(a:} < 1, {1}
where F(a:} is a distance-function. By Lemmas 1 and 2 of Chapter IV.
there are numbers c>O and C such that
(2)
In particular. a lattice A admissible for .9' has no points in the sphere
1a:1 < C-l.
and so has
(3)
by MINKOWSKI'S convex body Theorem II of Chapter III, where is v..
the volume of the unit sphere Ia: I< 1.
Now let A be a critical lattice, so that there are (at least) n linearly
independent points fit, ... , a.. of A on the boundary F(a:} = 1 of .9'.
Then by (2) we have
(1~i~n), (4)
and so by HADAMARD'S Lemma 9 we have
Idet(a1 ,···, a..}! ~ c-". (5)
Hence in the language of Chapter I the index I of ai' ... , a .. in A is
1= Idet(fIt . .... a..H ~ (~)"v.:-l =I (6)
d(A) - c " o·

Hence by the corollaries to Theorem I of Chapter I. there is a basis

b1 , .... b" of A such that
(7)
1 Alternatively one may observe that. by
determinants.
(1).1 ~
•
x, 1 1.: x:1 c,l' and compare
C, 1 =
 Bounded star-bodies 149

where the vii are integers,

O:;::;;Vij<Vji (i < i), (8)
and
O<IIv ji =I:;::;;10 · (9)
There are thus only a finite set of possibility for the integers Vij. For
each set of integers Vij' the points a i on the boundary determine the
bi , by (6). The a i are to be chosen so as to make
d (A) = Idet("t, .... an)1
V11·"V nn

a minimum, subject to no points of A being in Y and, in particular,

subject to (3). Then.d (Y) is clearly the minimum of d (A) over the A
so obtained and over all of the finite number of choices for the Vij.
We now verify that if A is a lattice constructed with n points
al' ... , an on the boundary and satisfying (3), (5), (6), (7), (8), (9),
and if d were any point of A in Y, then d has the shape
d = u 1 b1 + ... + un b n ,
where bounds can be given for the integers Uj. Indeed then Id I:;::;; c-1 ;
and so for each integer i we have
Idet(Ut, ... ,ai - 1 ,d,ai +1' ... ,an)1 ~ c- n
by (4) and HADAMARD'S Lemma 9. Hence, if ()) is true, the index of
al, ... ,ai-l,d,aj+1, ... ,an in A is at most 10 fori=1,2, ... ,n: and
it is easily verified that this gives bounds for the ui . It is thus, in
principle, a finite problem to find .d(Y).
The lattice constants of a great many 2-dimensional bounded star-
bodies have been evaluated. There is a partial list in KELLER (1954a)
to which may be added among others the bodies discussed by OLLE-
RENSHAW (1945a, b, 1953g). The treatment of bounded non-convex
body in more than 2 dimensions by such methods seems inevitably
laborious. Perhaps the only cases worked out are those of N. MULLINEUX
(1951 a).
V.6.4. In the evaluation of .d(Y) for a given star set Y it is usually
best to combine the techniques just introduced with those discussed
in Chapter III. We consider an instructive example due to N. MULLINEUX
which we shall have occasion to discuss further in § 7.
LEMMA 10. Let k be an arbitrary positive number and put
D = (k2 + 4k)~,
and
g= l (k + 2 + D) ,
150 MAHLER'S compactness theorem

so that
g-l = j (k +2 - D) .

Let f/ be the 2-dimensional star-body defined by

- 1< Xl X 2 < k, !Xl + X2 ! < D.

Then
L1(f/) = D.
The only critical lattices have bases of one of the two following kinds
(i) the point (1, -1) and any point on Xl x2=D, +
( ii) the points
p = (_g-lt,gt-l ), q = (-t,t- l )

where t is any number in the range

1<t<g.
We must first verify that the lattices defined above are f/-admissible.
This is certainly true for (i). We now verify it for (ii). It is readily
+
verified that the line Xl x 2 = D meets Xl X 2 = - 1 in the points

Hence the points p and q above do lie on the portion of the boundary
of f/ given by Xl X 2= -1. The point
r =p - q ={j(-k + D) t, j(k + D)t- = (r I} I , r 2)
lies on

Further,

since 1 <t<g and

j (- k + D) t + j (k + D) t-l
equals D both for t = 1 and for t = g. Hence a lattice of type (ii) has six
points ±p, ±q, ±r on the boundary of f/. There can be no further
point of the lattice in f/, since it is easy to verify that every point
of f/ except ±r lies either strictly between the (infinite) line A through
p and r and its image - A in 0; or strictly between the line 1.1 through
q, r and its image - 1.1 in 0; for example the line A meets Xl X 2 = - 1
and X l x2 =k respectively apart from p, r in the points
(gt, _g-lt-1), g(k +D)t, }(-k +D)t- 1};

and for both of these !Xl +x2 1> D.

 Bounded star-bodies 1 S1

For later use we note that the whole of the line-segment joining
p, r must lie in Y except the end points, since a line can meet a hyper-
bola X I X 2 = -1 or X l x 2 =k in at most two points. Hence the whole
of the closed parallelogram with vertices at 0, p, rand - q must lie
in ,,/ except for p, rand - q.
We are now in a position to prove Lemma 10. Let M be a critical
lattice. Suppose, if possible, that there is no point of M on the portion

of the boundary of Y. Then the set of pointsl

M.: {(1-S)X I +eX2, eXI +(1-s)x 2},
for small enough e, will also be Y-admissible since
{(1 - S) Xl + S x 2} -:- {S Xl + (1 - S) Xz} = Xl + X2
and
{(1 - s) Xl + S x 2}{S Xl + (1 - s) x 2} = Xl X2 + (s - S2) (Xl - X2)2 ;;:;; Xl x 2·
Since
d(M.) = (1 -- 2s) d(M) < J(Y),
this contradicts the hypothesis that M is critical. Hence there is a
point q = (ql' q2) on the boundary Xl X2= - 1 of Y; and, by symmetry,
we may suppose that
-ql;;:;; 1;;:;;q2> o.
Suppose first that q =1= ( - 1, 1). Then
(ql' q2) = (- t, t-l)
for some tin 1 <t<g. Let us identify this q, with the q of the lattice
1\ introduced earlier, and let p, r have the meanings introduced then.
Since 1\ is admissible and M is critical, we have
d (M) ~ d (1\) .
The line A of points x with
det (x, q) = d (1\)
passes through p and r, so the line
det (x, q) = d (M) (1 )
must either coincide with it or lie between it and the line through 0
and q. But now q is a primitive point of M, since r-1q Y for any
1 This argument becomes more transparent on introducing temporarily the
co-ordinates Yl = } (Xl + X 2J, )'2 = ! (Xl - X t )·
152 MAHLER'S compactness theorem

integer r> 1; and so there are points of M on the line (i) and at a distance
Iq I apart. Hence there must be a point of M other than 0 and - q
in the closed parallelogram with vertices at 0, - q, p and". But we
have already seen that the only points of this parallelogram which are
not in f/ are the vertices p, " and - q. Hence either p or " is in M;
and in both cases then M coincides with A.
There remains the possibility that q = (- 1, 1). If the definition of
p and" is extended in the obvious way to t = 1, the situation remains
the same, except that now the whole line-segment joining p and" is
part of the boundary Xl +x2 =D of f/. Hence we may deduce only
that M has a basis consisting of (- 1, 1) and some point on Xl + x 2 = D.
For this type of proof compare OLLERENSHAW (1945b).
For later use we note that we have also proved the
COROLLARY 1. The only critical lattices lor
- 1< Xl X 2 < k, IXl + x 1-;;;, D
2

are those 01 type (ii), where now t is allowed also to take the value 1.
For the other lattices of type (i) have a point on -1<XI X2 <D,
IXI +x 2 1 =D. Here our usage differs from that of MAHLER (1946a),
since he calls a lattice admissible for a set f/ if it has no points other
than 0 in the interior of f/. Thus MAHLER calls the lattice of type (i)
admissible (and so critical) for the set of the corollary.
Lemma 10 may be regarded as a more precise version of Theorem IV
of Chapter II. To make the connection more clear we prove
COROLLARY 2. Ilk is an integer, the critical lattices 01 type (ii) are
admissible lor
-1<XIXa<k.

For the general point of a lattice of type (ii) is

;I: = ulP + ua",
where uI , Us are integers. Then
xIXS = (UIPI + U2'1) (UI P2 + u 2 rS) = - u~ + kUI U2 + ku~.
We showed in § 4.4 of Chapter II that -us +kUIUS+ku~ does not
take any values strictly between - 1 and +
k when k is a positive
integer and UI , U 2 are integers not both o.
V.7. Reducibility. It may happen that if ~ is a star-body, there is
some star-body ~ which is properly contained in ~ but which has
the same lattice constant: L1(~) =L1(~). We say then that ~ is
reducible. If no such f/2 exists, then ~ is said to be irreducible. Criteria
 Reducibility 153

for the reducibility of a bounded star-body have been given by MAHLER

(1946a) and ROGERS (1947a). Later, ROGERS (1952a) gave a most
ingenious example of a reducible star-body which does not contain an
irreducible star-body of the same lattice constant: but he was able to show
that if a rather wider class of point sets, which he calls "star sets",
is considered, then every bounded reducible star set contains an irreducible
star set. Convex 2-dimensional sets were considered in great detail by
MAHLER (1947a). Mrs. OLLERENSHAW (1953 b) has shown that the
n-dimensional unit cube is irreducible for all n and that the unit sphere
is irreducible at least for n ~ 5. She shows further that a 3-dimensional
cylinder is irreducible if its 2-dimensional base is irreducible.
We refer the reader to the papers quoted for the general theory.
The following lemma shows in a simple case the sort of ideas involved
in the proof that a star-body is irreducible.
LEMMA 11. The star-body
!l) : x~ + x~ < 1
is irreducible;
For suppose Y is a star-body strictly contained in !l). Then there
is a point p on the boundary of ~ which is not on the boundary of Y.
But now (§ 6.4 of Chapter III) there is a critical lattice /\ of !l) having
points at ±p. The only other points of /\ on the boundary of !l) are
the points ±q, ±r which, together with ±p, are at the vertices of a
regular hexagon. Since Y (~, the lattice /\ must be admissible for Y.
But now the only points of /\ on the boundary of Y can be ±q and ±r.
These points clearly do not satisfy the criterion of Theorem VII, Corollary.
Hence /\ is not critical for Y, that is
L1(Y) <d(/\) =L1(!l)).
Since Y is any star-body contained in !l), this proves the lemma.
A similar proof shows that MULLINEUX'S star-body ,fj' defined in
Lemma 10 is irreducible. Again, if p is a point on the boundary of Y
then, apart from a finite number of exceptional p, there is a critical
lattice for Y which has only three pairs of points ±p, ± q and ± r
on the boundary of Y; and the points ± q, ± r cannot be the only
points on the boundary of a critical lattice of any set :T contained in Y.
The finite number of exceptional points p for which such a lattice does
not exist cannot affect the argument, since if :T is properly contained
in Y there are infinitely many boundary points of Y which are not
boundary points of :T.
V.7.2. If Y is an unbounded star set but there is a bounded star
set :T contained in Y such that L1(:T) =L1(Y), then Y is said to be
154 MAHLER'S compactness theorem

boundedly reducible. Corollary 2 of Lemma 10 shows that the 2-dimen-

sional star-body
(1 )

is boundedly reducible when k is a positive integer, since L1(9;.) =

L1(Y,,), where y" is MULLINEUX'S set
Y,,: -1<X1 X2<k, !X1 +X 2!«k 2 +4k)1. (2)
On the other hand, 9;. is not boundedly reducible for every k.
Thus we saw in § 4.4 of Chapter II that the critical lattices M for y',
are admissible for !Xl X 2 ! < it, and so have no points on Xl X 2 = - t
But then, precisely as in the proof of Lemma 10, M cannot be critical
for a bounded set .'T contained in 9Jb' since the lattice M. of points

would be admissible for .'T for sufficiently small e.

The proof of Theorem VII of Chapter III shows that the 2-dimen-
sional star-body

is boundedly reducible, since the proof used only a bounded portion

of the set. MAHLER (1946a) has developed criteria for sets of certain
types to be boundedly reducible if their critical lattices are known.
Bounded reducibility is further discussed by DAVENPORT and ROGERS
(1950a). DAVENPORT and ROGERS introduce the concept of full redu-
cibility. If.'T is a set contained in the set Y' and L1 (.'T) = L1 (Y') then
clearly every lattice critical for Y' is also critical for .'T, but in general .'T
might have more critical lattices. For example when k is a positive
integer the sets defined in (1) and (2) have the same lattice constant,
but the critical lattices of y" of the type (i) of the enunciation of
Lemma 10 will in general have points in 9;.. On the other hand, the set
.'T;: -1<X1 X2<k, !Xl+X2!;£(k2+4k)~

has no more critical lattices then 9;. by Lemma 10 Corollary 1. If an

unbounded set Y' contains a bounded set .'Twith the same lattice constant
and no more critical lattices then Y' is said by DAVENPORT and ROGERS
(1950a) to be fully reducible l . They, following MAHLER, use the concept to
show that lattices of certain types have infinitely many points in certain
regions. We shall be discussing this from a rather different point of
view later in Chapter X. We do not discuss bounded and full reducibility
1 Their definition is not quite the same as ours since they use MAHLER'S defini-
tion of an admissible lattice. But it is not difficult to see that it is equivalent
to ours.
 Convex bodies 155

further but refer the reader to the papers quoted. The following example
illustrates the connection with the existence of infinitely many lattice
points in sets.
LEMMA 12. Let k be a positive integer and 1\ a lattice with
d (1\) ~ (k 2 + 4k)~.
Then there are infinitely many points of 1\ in

~: - 1~ Xl X 2 ~ k. (3)
There are infinitely many points of 1\ in

~: - 1< Xl X 2 < k, (4)

except when 1\ is critical for ~.
If 1\ contains a point (0, x 2 ) with X 2 =F 0, it contains all the points
(0, H 2) (r = 1,2,3, ... ) and so the lemma is trivially true. Otherwise it
°
suffices to show that for every e> there is a point (Xl' x2 ) of 1\ in .9';.
for which IxII;;;;; e; and that this point is in ~ unless 1\ is critical for 9;..
Let t be any positive number. Then the lattice I\t of points
(5)
has the same determinant as I\. Hence by Lemma 10, Corollary 1
there is a point of I\t in
(6)
and indeed in ~ unless I\t is critical for 9;.. But now the region (6)
is bounded, so all the points of (6) satisfy
IXII ;;;;;y
for some number y which depends only on k. Hence, by (5), the original
lattice 1\ contains a point (Xl> X 2 ) =j=o such that
-1~XIX2;;;;;k, IXII~yt-l.

Further, A is critical for !/ if and only if At is. Since yrl is arbitrarily

small when t is a arbitrarily large, this proves the result.
V.S. Convex bodies. For convex bodies stronger results than Theo-
rem VII hold about the lattice points of a critical lattice on the boundary.
The following theorem of SWINNERTON-DYER (1953 a) generalised an old
result of KORKINE and ZOLOTAREFF for spheres.
THEOREM VIII. Let;X" be a bounded open symmetric convex set in
n dimensions and let A be a critical lattice for f . Then A has at least
in(n+1) pairs of points ±a on the boundary of f .
156 MAHLER'S compactness theorem

We reproduce SWINNERTON-DYER'S elegant proof. Let b1 , ••• , b .. be

a basis for A and let A' be a lattice with the basis bj (1 ~i~n), where
(1)

and the aii and 17 are real numbers to be determined later. Let
±P1' ... , ±PN be the only points of A on the boundary of:£ and let
±p~, ... , ±P;' be the points of A' which correspond to them in an
obvious way. Let 1T1, ... , 1TN be tac-planes to:£ at PI' ... ,PN (Theo-
rem IV of Chapter IV). If there is more than one tac-plane, we choose
one arbitrarily. We then impose on A' the condition that PI
lies in 1TJ
for 1 ~ J ~ N. By (1), and since PJ lies on 1TJ, this imposes a condition
of the type
~ a··tW =0 (1~J~N), (2)
L. .. " "
1;:>;;:>
1;:>;;:> ..

where the numbers tlf> depend only on the point PJ and the choice of
tac-plane 1TJ. We also impose the conditions

(i=l=i)· (3)
The total number of linear conditions (2) and (3) imposed on the n Z
numbers ai i is !n(n-1) +N. Hence if N<!n(n+1), there exists a
set of real numbers aii not all 0 satisfying (2) and (3). We select any
one such solution and keep it fixed in what follows.
Since the points PJ lie on tac-planes to the open set:£, they do not
lie in:£. When 1171 is small enough, there are no further points of A'
in :£ other than 0, by the argument of § 6.1. Hence A' is admissible
for:£. Since A is critical, we must then have
d(A') =ldet(b~, ... ,b~)I~ldet(b1, ... ,b..)I =d(A) =.1(:£);

)
that is
1 + all17 ~217

1~ det (
. a21 17 .1:- ~2~17 :::~
a.. 1 17 anZ17 1 + a.. n17
= 1 + A 1 17 + A2172 + ... + An17n (say).
Since this must be true for all sufficiently small values of 1171. it follows
that

and
Az= - L aii aii + L ajiaii ~ o.
i<; i<;
 Convex bodies 157

Hence on using the symmetry conditions (3) we have

o ~ 2A2 - A~ =- ~ a~j.
1;;;.;;;"
1:Si.j;;;"

Hence ajj=O for all i andj; which is a contradiction. The contradiction

t
arises from the assumption that there are fewer than n (n + 1) pairs
of points of A on the boundary of f. Hence the theorem is established.
V.8.2. For bounded symmetric convex star sets the considerations
of § 6.3 about the maximum number of points of a critical lattice on
the boundary and about their index may be made much more precise,
as was shown already by MINKOWSKI. His results apply indeed not
merely to critical but to all admissible lattices. We recollect that a
body f is strictly convex if every point tp+(1-t)q (O<t<1) is an
interior point of f whenever p and q are distinct points in or on the
boundary of f.
THEOREM IX. Let A be an admissible lattice lor the convex symmetric
open set f. Then there are at most !(3"-1) pairs 01 points ±a 01 A
on the boundary olf. I l f is strictly convex, the number 01 pairs is at
most 2"-1.
The proofs are very simple. Suppose first that f is strictly convex.
Let b1 , ••. , b" be any basis for A and let
a =u1b 1 + ... +u"b"
be a point of A on the boundary of f. Then not all of Ut, ... , u" are
even, since otherwise !a would belong to A; and !a is certainly an
inner point of f. Let now
a' =U~b1 + ... +u~b",
if possible, be another point of A on the boundary of f such thatl
uj == uj (mod 2) (1~j~n).

Then !(a+a')EA. By the strict convexity, !(a+a') is an inner point

of f and so must be 0, that is a' = - a. Hence the total number of
boundary points is at most the number of residue classes for (U1' ... , u,,)
modulo 2 excluding (0, ... ,0), that is 2"-1, as required.
When K is not strictly convex one must work with congruences
modulo 3; the details are left to the reader.
THEOREM X. Let f be a convex symmetric open n-dimensional set
and A an admissible lattice lor f. I I ~, ... , a" are points 01 A on the
1 The notation means that Uj - ui is divisible by 2.·
158 MAHLER'S compactness theorem

boundary 01 .% then their index I satislies

I~n!. (1)

There is inequality in (1) il.% is strictly convex.

If a l an are linearly dependent. then their index is 0 and there
•...•

is nothing to prove. Otherwise. every point c of A may be put in the

shape
(2)

where VI • •..• Vn are rational numbers. The sets of numbers 'v such that
(2) is in A clearly form a lattice M of determinant

d(M) = _~dJ!'l~ __ = I-I.

Idet(a1 , .,.,a.. )1

Hence. by MINKOWSKI'S convex body Theorem II of Chapter III. there

is point v =1= 0 of M such that

(3 )
Let F be the distance function associated with.%. so that

F(a j ) =1 (1~j~n).

For the cEA given by (2) and (3) we thus have by the convexity and
symmetry of.%. that

(4)

But F(c) ~ 1 since A is admissible for .% and so I ~ n! as required.

If 1= n! and.% is strictly convex we should have F(c) < 1 unless both
M is a critical lattice for IvII + ... + IVn I< 1 and every point of M on
the boundary has n - 1 of the co-ordinates VI • •.•• vn equal to O. But
these two requirements are incompatible by SWINNERTON-DYER'S
Theorem VIII.
The l estimate for I in Theorem X can usually be much improved
and more information obtained about the relationship of a l •...• an to
a basis for the lattice. Thus for n = 3 we have
COROLLARY. 11.% is strictly convex and n = 3. then I = 1 or 2. II
I =2. then t(~+a2+a3)EA
For I~ 5. If 1= 5. then there are integers u l • U 2 • U 3 not all divisible
by 5 such that

1 We do not use the rest of § 3,2 later but do refer to it at the end of § 8.5.
 Convex bodies 159

We may suppose that 5 does not divide U l and, by taking 2c instead

of c if necessary, that
U l == ± 1 (mod 5) .

Hence by adding appropriate integer multiples of aI' a2, a3 to c we

may suppose, without loss of generality, that
ul =±1, lu 2 1;£2, lual;£2.
But then by the strict convexity we should have
F(c) < iF(al) + i-F(a2) + i-F(aa) = 1;
a contradiction. Hence I =f= 5. Similarly I =f= 3.
Suppose now I = 4. Then there exists a base bl , b 2 , ba for A such
that
al = VII bl ,
a 2 =v2l b l +V 22 b 2 ,

aa = Val bl + V32 b2 + V33 b a,

where
(i < i),
and

Then vn =1, since otherwise tUtEA andF{tal ) <F(al ) =1. If v22 =l=1,
then either ta 2 or t(a l +a 2) is in A; and again we have a contradiction.
Hence
VU =V 22 =1; so vaa=4.
If V3l were even, we should have either i aa or t(a 2 +aa) in A; so Val
is odd. Similarly, V32 is odd. Hence there is a point
c = t(u1 Ut + u 2 a 2 + aa) E A,
where U 1 , U 2 are odd. By adding integer multiples of Ut and a2 to e,
we may suppose that Ut = ± 1, u2 = ± 1. But then
F(c) < i{F(a 1) +F(a2 ) +F(aa)} =! < 1.
Hence I =l= 4.
Finally, when 1=2 it follows, just as for 1=4, that the only pos-
sibility is vn=v22=1, v21 =O and vas=2. Further, the argument that
Val' Va2 are both odd continues to hold. Hence t(Ut+a2 +aa)EA.
V.S.3. When:f{' is a bounded symmetrical strictly convex 2-dimen-
sional set, the lower bound 3 for the number of pairs of points ±a of
a critical lattice on the boundary given by Theorem VIII coincides with
the upper bound give by Theorem IX. We have indeed
160 MAHLER'S compactness theorem

THEOREM XI. A. Let f be an open convex symmetrical2-dimensional

convex body. Then a critical lattice 1\ olf has six points ±P. ±q, ±r
on the boundary 01 f such that
(1 )
and any two 01 P. q. r is a basis lor I\.
B. Further, il ±p, ±q, ±r are any poi1lts on the boundary olf
st"h that (1) holds, then the lattice M with basis P. q is admissible lor f.
There are no lurther points 01 M on the boundary. except when f is a
parallelogram and two 01 p. q. r are mid-points 01 its sides.
The first part of Theorem XI is an almost immediate consequence
of the last three theorems. By Theorem VIII there are three pairs of
points ±P. ± q, ±r on the boundary of f. By Theorem IX, the index
of p, q is 1 or 2. Since ip. iq are (inner) points of f, they cannot
belong to I\. Hence, if the index is 2, the point i(p +q) is in I\. It is
also in f or on the boundary of f. the latter only if f is not strictly
convex. If the index is 2, we may thus take i(p+q) =q' instead of q.
The index of p and q' is 1. Hence without loss of generality the index
of p and q is 1. Hence r =up +vq for some integers u and v, where
Iul ~2, Ivi ~ 2, since the indexes of P. r and of q, r are at most 2. Not
both u and v can be even, since otherwise ir would be in I\. If, say,
u=±2 is even, then v=±1 is odd, and r'=i(r+vq)=iup+vq
is in or on the boundary of f. It must be on the boundary since 1\ is
admissible. Hence by taking r' instead of r we may suppose, with~ut
loss of generality, that Iu I = Iv I = 1. By changing the signs of p and q.
where necessary, we may suppose that u =V = -1, that is, that (1)
holds. This proves A.
It remains to prove B. Suppose, if possible, that the point
c=up+vq
= (v - u) q + (- u) r
= (u - v)p + (-v)r
is in or on the boundary off for some integers u, v. If, say, lui> Ivi +1,
then the point
p = u-1c - vu-1q

would be an inner point of f, because we should have Iu-11 +Ivu-11 < 1.

Hence from the three expressions for c we deduce that
Ilul-Ivil ~ 1,
Ilu-vl-lull~1,
Ilu-vl-lvl\~1.
 Convex bodies 161

It is easy to see that the only integral solutions of those inequalities

giving primitive lattice points distinct from ±p, ±q, ±" are

± (u, v) = (2,1), (1,~) or (1, -1).

Hence after permuting p, q," cyclically if need be, we may suppose

that c =p - q is in or on the boundary of f. Since now

p=!c-!".. q=-!c-!",
the only possibility is that c is a boundary point.
We now show that f contains the whole parallelogram ~ of points

;x!=Ap+,uq
with
max{1 AI,I,uI} < 1.
Indeed
;x! = ec + U",
where
lei +Iul =!IA-,ul +!IA+,u1 =max{IAI,I,uI}·
But now the area V(~) of ~ is
V(~) = 41 det(p, q)1 = 4d (M).

On the other hand, by MINKOWSKI'S convex body theorem, we have

V(f) ~ 4d (M) .

Since f includes ~, and since f is open, the only possibility is that

f coincides with~. This concludes the proof of the theorem.
Theorem XI gives one a ready criterion for finding the lattice constapt
of 2-dimensional convex star-bodies. It is easy to see that if p is a
given point on the boundary of f , then there is precisely one hexagon
of boundary points ±p, ±q, ±" for which (1) holds. The lattice
constant of f is then the lower bound of det (p, q) for these hexagons.
V.S.4. As an application af Theorem XI we prove
LEMMA 13. Let!/' be a convex symmetric open hexagon. Then
LI(!/') =i V(!/,). (1)
The only critical lattice M is that which has points at the mid-points 01
all the sides 01 !/'.
By MINKOWSKI'S convex body theorem,
LI(!/') ~ i V(9'). (2)
Cassels, Geometry of Numbers 11
162 MAHLER'S compactness theorem

Let the vertices of ,,/ taken in counter-clockwise order be

a, - b,c, -a, b, -c.

Then the lattice M of the lemma has basis ita-b) and i(b-c). It
clearly contains also i(c-a). Hence, by Theorem XI, M is 9'-admis-
sible. We now show that
d(M) = i: V(9').

On dissecting 9' into triangles with a vertex at 0, we have

- V(9') :::: det (a, b) + det (b, c) + det (c, a) = 4 det (u, v)
on putting b = a + 2 tt, C = a + 2 v. This proves (3). Then (1) follows
from (2) and (3) since M is 9'-admissible.
Now let A be any critical lattice for 9'. Then d (A) = i: V(,'/). If
A did not have a point on a particular side of 9' there would be a sym-
metric convex set larger than 9' which contained no point of 9' except
0; which would contradict MINKOWSKI'S convex body theorem. Hence,
by Theorem XI, A has precisely 6 points ±p, ± q, ±1' on the boundary
of 9'; one on each side. If, say, the points ±P are not the mid-points
of their sides, then by rotating slightly the sides about ±p, leaving
the other pairs of sides fixed, it would be possible to find a convex
symmetric set or of volume V(Y» V(9') containing no points of A
except 0; again contradicting MINKOWSKI'S convex body theorem.
Hence ±p, ± q, ±1' are the mid-points of their sides, and A = M.
It would, of course, be possible directly to compute the determinants
of all lattices having points p, q, l' with p +q +1' =0 on the boundary
of 9' and to show that M gives a minimum.
V.8.5. MINKOWSKI (1904a) has extended the argument of Theorem XI to
3 dimensions and proved the following.
THEOREM XII. To lind the lal/ice constaflt LI(f) 01 an open symmetrical convu
set f in 3 dimensions it sulfices to consider the minimum 0/ the determinants 0/
lattices generated by three points aI' a 2 , a 3 on the boundary 0/ 1\ alld satisfying one
0/ the followinlf three condition.s:
(A) the points a 1- a 2, a 2- aa' aa- a 1are on the boundary 0/ f and - a 1+ ",+a 3,
a 1-a 2+a3 , "l+aa-aa are outside f.
(B) the points a 1+a2, a 2+03, a 3+o1 al·e on the boundary 01 f and a 1+a,+"8
is ott/side f.
(C) the points ~ + a 2, "2+"s' "s+ a 1 and a1+ a.+ a 3 are on the boundary 0/ f.
We refer the reader to the original paper for the proof. Alternatively the
reader may construct a proof by combining the ideas of the proof of Theorem XI
with those at the end of § 8.2. The corresponding result in 4-dimensional space,
which is fairly complicated, has been found by K. H. WOLFF (1954a), who states
that some of the auxiliary results are due to E. BRUNNGRABER (1944a).
 Spheres 163

MINKOWSKI (1904 a) used Theorem XII to find the lattice constant of the
octahedron

namely 19/108. The lattice constants of further convex 3-dimensional bodies have
been determined by CHALK (1950a) and WHITWORTH (1948a and 1951a). In all
cases a considerable amount of rather tedious detail is necessary.

V.9. Spheres. We now consider more particularly the n-dimensional

spheres
P},.: 1:1:12 = x~ + ... + < 1. x: (1 )

We denote the lattice constant of P}" by

r.. = L1(P},,). (2)
The value of r.. is known for 1 ~n~ 8, see Appendix A. We here find
again Fa, which we already found in another context in Chapter II,
Theorem III. From this the value of ~ will follow almost at once by
a general theorem of MORDELL in Chapter X.
We must first prove a result for spheres which is more precise than
the mere application of Theorem X.
THEOREM XIII. Let /\ be a lattice admissible lor P},,: 1:1:1 2<1; and
let ai' . '" an be points all\ on the boundary 0/ P}n. Then the index I
0/ a l , ... , an satisfies
I ~ {d(/\)}-l ~ {L1(~,,)}-1 = l. r..-(3)
For la;1 =1 (1~j~n), and so, by HADAMARD'S Lemma 9, we have

Idet (ai' .,,' an) I ~ I~ I ". Ian I = 1.

Since
I = Idet(ai' ... , an) I
d(/\.) ,

the first half of (3) follows. The second half of (3) is a trivial conse-
quence of the definition of r...
COROLLARY. 1/ n = 3 the index is 0 or 1.
For P}" has volume 4:r:/3, and so
Fa~ 71:16> t,
by MINKOWSKI'S convex body Theorem II of Chapter III.
THEOREM XIV.
Fa=ri.
A critical lattice lor P}s has a basis m l , m 2 , ma such that

Iulml + u 2 m 2 + uamal2 = u~ + tt~ + u~ + u 2 u a + uaul + UI U 2

identically in UI , U 2 and Us.
11*
164 MAHLER'S compactness theorem

Let A be a critical lattice for !iJa. By Theorem VIII there are at

least In(n+1} =6 pairs of points ±m of A on the boundary of !iJa
and by Theorem VII there is a linearly independent set of 3, say ~,
m 2, rna. By Theorem XIII, ~,m2' ma is a basis for A If
m = Ul~ + u 2 m 2 + uama
is another point of A on the boundary of !iJa, the only possible value
for the ui are 0, ± 1 by Theorem XIII. There can be at most one such
pair ±m with UlU2Ua=FO. For if, say,
m = ulm 1 + u 2 m 2 + uams, U 1 U2U a=f= 0,
m' = u~ml + u~m2 + u~ma, u~ u~ u~ =F °,
the index IU2U~-Uau~1 of~, m, m' is even, so must be 0. Similarly

som'= ±m. Hence there must beat least one point ~ml +u 2m 2+uarna
with U l U 2 Ua=0 on the boundary of !iJa other than ±~, ±m 2 , ±rna.
We may suppose without loss of generality that it is
m,=ml -m2·
Then neither m l +m 2 nor ~+m2±rna can occur as boundary points,
since they would give index 2 with rna and m,. Hence at least two
of the remaining possibilities
~±ma, m 2 ±ma, ~-m2±ma

must occur. Since ~-m2+ma and ~-m2-ma cannot both occur,

we may suppose without loss of generality that

occurs. Then m 2+rna and ~-m2-rna do not occur, since they

give index 2 with m 2 and m6: and ~ + rna cannot occur, since it gives
index 2 with m, and m 6 • Hence the only possibilities for ±m. are
ma - 'In l or m l - m 2 + rna.
In the second of these cases take m6 instead of rna. Then without loss
of generality

Write
I(ul , U 2 , u3} = IUl~ + u 2 m 2 + uam al2,
where ~,U2' ua are variables, so I(u} is a quadratic form. Then
1(1,0,0) = 1(0,1, O} = 1(0, 0,1)
= 1(1, 0, -1) = 1(0,1, -1} = 1(1, -1,0) = 1.
 Applications to diophantine approximation 165

Hence
IM=~+~+~+~~+~~+~~
with determinant D (f) = 1. and so
{det(~. m 2 • m a)}2=1.
as required.
V.9.2. Let f/ be a star-body and 1\ an f/-admissible lattice. We say that
1\ is extreme for f/ if there is a neighbourhood 2 of I\, in the sense of § 3.2, in which
every f/-admissible lattice M satisfies

d(M) ;;;; d(I\).

Clearly a critical lattice is extreme; but an extreme lattice need not be critical.
Some of the results proved already extend to extreme lattices, notably SWINNERTON-
DYER'S Theorem VIII.
The extreme lattices of n-dimensional spheres have been exhaustively studied.
For example there are six distinct types of extreme lattice for the 6-dimensional
sphere as was shown by BARNES (1957b). There is a general theorem of VORONOI
(1907 a) which helps to characterise the extreme lattices of an n-dimensional sphere
(they are "perfect" and "eutactic"). BARNES (1957a) has given an extremely
elegant proof of VORONOI'S characterisation. Unfortunately we cannot discuss
these points further here. so we refer the reader to the two papers by BARNES
where there are further references to the copious literature.

V.10. Applications to diophantine approximation 1. The theory of

Diophantine approximation deals with the approximation of rational or
irrational numbers by rational numbers with special properties. The
geometry of numbers has many applications to Diophantine approxima-
tion. The author's recent Cambridge Tract [CASSELS (1957a)] deals
with Diophantine approximation and we do not intend to repeat what
was done there. We give however a theorem of DAVENPORT generalizing
work of FURTWANGLER which is an interesting application of MAHLER'S
compactness techniques.
First. we note an obvious consequence of MINKOWSKI'S linear forms
Theorem III of Chapter III. Let (A •...• {},. be real numbers and Q an
integer. By Theorem III of Chapter III there exist n + 1 integers
"0 ..... ",.. not all O. such that

I"0{}1 - "II < Q-l/,. (1~i~n). (i)

l"ol~Q; (2)

since "o{}j-"I (1~i~n) together with "0 form n+1 linear forms in
"0 ..... ",. with determinant 1. Were "0=0. we should have 1"11 < Q-l/".
so "1=0 (1~i~n). Hence "0=1=0. and on replacing "0. " " " I f by
1 Not used later in book.
166 MAHLER'S compactness theorem

- U o, ... , - U .. if need be, we may suppose that

(2')
Further, (1) may be written

(1')

which shows that the u;luo are good rational approximations to the {};,
all with the same denominator uo.
We may look at (1) and (2') from another point of view. On elimi-
nating Q we have

There are in fact infinitely many solutions uo>O, u1 , ••• , u .. of (3). If

all of {}1' ... , {}.. are rational, this is trivial since then there exist integers
vo>O, VI' ••• , V .. such that

(1~j~n),
and then we may put
(O~j~ n),

where r is any positive integer: and then the left-hand side of (3) is 0.
Otherwise we may suppose that {}1 is irrational. Suppose that R integral
solutions u}'l (O~j~n, 1~r~R) have already been found with u~»O.
Since {}l is irrational, we may choose Q so large that

(1~ r~ R).

For this value of Q the solution of (1) and (2') gives a solution of (3)
which is clearly not identical with any of the earlier one~.
V.I0.2. For different purposes one may be interested in different
properties of the approximations u;/uo to the 0;. For example, instead of

max IuO{}i - ujl

1~1~"
we may wish to make
(1) .
or
(2)

small. Or again one may be interested in "asymmetric" inequalities,

of the type
- k0 U o
-1/.. ..- 2::! k1 U o
2::! U o{}i - ui --- -1/.. (1~j~n), (3)
 Applications to diophantine approximation 167

where ko and kl are positive numbers. All these different problems

may be brought into one general shape. Let (j)(xl , ... , x .. ) be a distance-
function of n variables. How small can 1
U o (j)" (u o1}1 - Ul , ... , U o1)" - Un)

be made for infinitely many sets of integers uo> 0 and U1 , ••• , Un? We
write
D ((j) : 1}1, ... , 1}n) = lim inf U o (j)" (u o1}1 - u1 , ... , U o1)n - un) (4)
110-+ 00
tlO! "II ''OJ UrI integers
and
D((j)) = sup D((j) :1}1, ... ,1},.l; (5)
lih··"lJ,.

so that D ((j)) is the number we wish to estimate.

The non-negative function F(x o, ... , x,,) of n 1 real variables +
defined by
+1 _{ XO(j)"(xl,· .. ,X,.) if xo~O}
pn (x o, ... , X,.) - (6)
- Xo (j)" (- Xl' .•. , - X,.) if Xo ~ 0
is a distance-function when (j) is a distance function of n variables:
since it clearly has the three defining properties that it is non-negative,
continuous and satisfies
F(txo, .. ·,txn ) =tF(xO, .. ·,xnl
when t> o. By definition, F is symmetric:
F(- Xo, ... , - xn) =F(xo, ... , x,,). (7)
It satisfies the identity
F(t" x o, t-l Xl' ... , t- l X,,) = F(xo, ... , x,,) (8)
for any t>o, since
(j)(t- l Xl' ... , t- l Xn ) = t-l (j)(X l , ... , X .. ).
As in § 4 of Chapter IV we write
pH (J\)
c5 (F) = s~p d (J\) ,

where the supremum is over all (1J + i)-dimensional lattices, so that

c5(F) = {Lf(9')}-l,
where 9' is the (n + i)-dimensional star-body
9': F(xo, ... ,xn )<i.
DAVENPORT'S result may now be put in the following shape.
1 By tP" is meant the n-th power of tP.
168 MAHLER'S compactness theorem

THEOREM XV. Let tJ> and F be related as above. Then

D(tJ» ~ ~(F) (9)
always. If tJ>(:I:) =0 only for :I: =0. then
D (tJ» = ~ (F) . (10)

The first part of Theorem XV is due essentially to MAHLER and is

related to the theory of automorphic bodies which we shall study in
Chapter X. When D(tJ» =0. there is nothing to prove. Otherwise.
let c be any positive number such that
c < D(tJ». (11)

Then. by the definition of D(tJ». there are real numbers {}l' .... {},. and
an integer 00 such that
(12)

whenever U o• ...• u,. are integers and

uo~ 00· (13)
In particular. {}l ..... {},. are not all rational; and so there exists a
number ,,>0 such that
(14)

for all integers U o• ...• u,. with

O<Uo~ 00.
Clearly
,,~t<1. (15)
Let Ml be the n + i-dimensional lattice of points
(xo... ·• x .. ) = (u o• UO{}l - U1 • .... uo{),. - u"). (16)

where uo, "', u,. run through all integers. Clearly

(17)
The function
= max [F(xo ..... x"). - - m?-x Ix;1
,1/("+1) ]
~(xo ... ·.x,.) ( 18)
" 1:>,::0"
is clearly an (n + i)-dimensional distance-function and
~ (-:I:) = 1\ (:1:) (19)
by (7). We show now that
(20)
 Applications to diophantine approximation 169

Consider a point (16) of M1 , where, by (19), we may suppose that uo~O.

If uo= 0 but not all of u1 , ... , un are 0, then the second term of the
outer maximum in (18) is

x Iu·I I -~--~c
1 1
-c -
/(n+1)
x
ma l:;;;i;:;on
c /(n+1)
,,-
1/(n+l)
,

by (15). If O<uo~ Uo, then the second term of the outer maximum
in (18) is still ~ c 1/(n+l), by (14). If U o ~ [fo, the first term of the outer
maximum in (18) is ~cl/(n+1) by (12). Hence in any case,
F;. (x) ~ c1/(n+l)

for all XE Ml except o. This completes the proof of (20).

For positive integers r = 1,2, ... write more generally

F,.(Xo,···, Xn) = max [ F(x o,· .. , Xn), - - max IXii .

cl/(n+l) 1 (21)
r" 1;:;0,;:;0,
Then
F(x) ~ F,. (x) ~ 1\ (x) (21 ')
and
lim F,. (x)
,--+ 00
= F(x) (22)

uniformly in any bounded set of points x. We have the identity

(23)
by (8).
Let M, be the lattice

Clearly
(24)
and
p,n+l (M,) = 1\n +l (Ml ) ~ c, (25)
by (17), (20) and (23). Consequently, by (21'), we have the weaker
assertion
F;.n+1 (M,) ~ c> 0 (1~r<oo). (26)
By (24), (26) and Theorem IV Corollary, there exists a convergent
subsequence of the M" say
M,,-+ N.
By (24) we have
deN) =1. (27)
Since (22) holds uniformly in any bounded set, we have
p+l (N) ~ lim sup P;;+ 1(M,,) ~ C, (28)
'--+00
170 MAHLER'S compactness theorem

by (25) and Theorem II. Hence

F,,+l (1\) > F n +1 (N) >
o(F) = s~p d (1\) = diN) =C.

Since Cwas any positive number smaller than D (C/» , this proves
o(F)~D(c/», the first part of Theorem XV.
The second part of Theorem XV requires quite different techniques
and uses the basis constructed in Theorem II of Chapter 1. By the
Corollary to Theorem VI, there is a lattice 1\ with
d(/\) = 1 (29)
and
p+l (1\) = 0 (F) . (30)
We denote the (n + i)-dimensional vector (xo, ... , x,,) in which xi = 1
but the remaining co-ordinates are 0 by
i n-i
ei=(~,1,~) (O~j~ n).
By Theorem II of Chapter I, with e =} and n + 1 for n, there exists,
for all sufficiently large numbers N, a basis ao' aI' ... , a" of 1\ such that
lai-Neil <NA (1~j~n). (31)
Then
aj =N ~ tj.e. (1~j~n), (32)
0::;;.::;;"
where
It.·11 -11 -~ N-! (1~j~n) (33)
and
Itjil ~ N-i (1~j~n, o ~ i ~ n, i =l= ;) . (34)
Since ao, aI' ... , a" are linearly independent, there are feal numbers
.1.0, AI, ... , An such that
eo = .1.0 a o + Al a1 + ... + An an ,
where we may suppose that
Ao~ 0,

on taking -ao for a o if necessary. Since d(/\) =1, we have now

.1.0 = .1.0 Idet (a o,"" an) I
= Idet (eo, aI' ... , an) I
= N"{1 +O(N-!)},
where the constant implied by the 0 depends only on n. We may
thus write
(35)
 Applications to diophantine approximation 171

where 01 , ••• ,0.. are certain real numbers, and

(36)

Let (l be any number such that

<5' < <5 (F).

We wish to show that

for the 01 , ••• ,0.. we have just constructed; provided that N is greater
than some No which may depend on <5' and the function tP. After the
first part of Theorem XV, this will complete the proof of the theorem.
If <5 (F) = 0 there is nothing to prove. Otherwise we may suppose without
loss of generality that
o< <5' < <5 (F) . (37')
To prove (37) we may clearly confine attention to integers uo, "', u .. ,
if any, for which
uo>O, uotP"(Y1, .. ·,y..)~<5(F), (38)
where we have put
(1~i~n). (39)

So far we have not used the fact that tP(:r) = 0 only for :r = o.
By Lemma 2 of Chapter IV, this implies that

or some c>O. Hence, by (38), we have

U o max IYil .. ~c- .. <5(F). (40)
l~I:O"

We now consider the point

of A. By (35) and (39) this is of the shape

Y = p,uoeo + L Yj"j;
1:oi:;;"

and so, by (32), has co-ordinates (Yo, ... , Y.. ), where

Yo=p,uo+N L YjtjO' (41)

l~i:o"

Y;= (1~i~n). (42)

172 MAHLER'S compactness theorem

Let e be an arbitrarily small positive number to be determined later.

By (33). (34) and Lemma 3. the inequality

f/J (f ti I Yi •...• f t i" Yi) ~ (1 + e) f/J(YI' ...• y,,)

holds for all real numbers YI •...• y" whatsoever. provided that N is
greater than a number depending only on the number e and the func-
tion rJ>. Hence. by (42).
rJ>{\i •...• V,,) ~ (1 + e) N rJ>(YI' ...• y,,). (43)
By (40) and (41). we have
(44)
for some l:{, which will depend. of course. on N. But now YEA and
pH (A) = d (F). by hypothesis. Hence

d(F) ~ YorJ>"(\i • ...• V,,). (45)

by the definition (6) of F. From (36). (37'). (43) and (44). we have
UOrJ>"{YI' ... , y,,) ~ (N",utl(1 + e)-.. -ld{F) > d' (all uo~ Uo),
provided that first e is chosen small enough. then N is chosen large
enough, and finally Uo is chosen large enough. This concludes the proof
of (37). and so of the theorem.
V.tO.3. The condition that rJ>(z) =0 only for z =0 is necessary for
the second part of Theorem XV. The case when n = 2 and
rJ>2(XI • x2) = IXl X2 1
represents a fascinating problem of LITTLEWOOD. It is not in fact
known whether there exist numbers {h and {}2 such that

lim inf uol UO{}1 - ulll U O{}2 - u 2 1 > O•

....... 00

where uo• u l • u 2 are integers. The corresponding function F(xo. Xl' XI)
is given by
F3(Xo• Xl' X 2) = IXOXIXsl:

and for this we have DAVENPORT'S result that

d(F) = 1/7.
which we shall prove in Chapter X. But it follows from work of CASSELS
and SWINNERTON-DvER (1955a) and from DAVENPORT'S results about
the successive minima of F, that at least
 Applications to diophantine approximation 173

There is a companion result to Theorem XV, also due to DAVENPORT,

which relates to the approximation of a single linear form to O. Here
one is concerned with
D'(l/J:{}l' ... , (},,) = lim inf IUo + Ul (}l + ... + u"{},,Il/J"(ul , ... , U,,),
max 1
... 1•.••• 1....1-0-00
..0+ ...".+···+ .... ".. 0;:0
"'J .'0' "" integers
where the condition UO+Ul{}l+'" +u,,{},,~O may clearly be omitted
if l/J is symmetric. Then Theorem XV remains valid if D (l/J) is replaced
by
D'(l/J) = sup D'(l/J:{}l'"'' (},,);
8 1 ,···,iJ"
and the proof is substantially similar.
V.I0.4. Note that we have not shown the existence in the second
part of Theorem XV of {}l' "', {}" such that
lim infuol/J"(uo{}l - Ul , ... , uo{}" - u,,) = «5 (F) :
"o~OO

and indeed in general such {}l' ... , {}" do not exist!. When n = 1, however,
a (}l does exist, as is easy to show. Here, of course, the only possibility
for the distance function l/J(Xl) of one variable is
l/J(xl) = { k Xl if Xt~ 0
-tXt if Xt~ 0,
where k and t are positive constants. As in the proof of the second
part of Theorem XV, we consider a lattice 1\ with
d (1\) = 1 , p2 (1\) = «5 (F) .
Let
a= (a o • at). b = (b o , bt )
be a basis for 1\, where without loss of generality
bt > 0 aobt - atbo = d(/\) = 1. (1 )
Put
{} = {}t = iltlbt· (2)
After Theorem XV it is enough to show that
lim inf Uo l/J(uo{}
"0-+ 00
+u l) ~ «5 (F).

As in the proof of Theorem XV, it is enough to consider value of U o

and ~, such that
(3)
where c is a constant such that l/J( Xl) ;?; cIXII for all Xl'
1 For example when n=2 and If>2(Xl.X2)=X~+x~. as one may show by
"isolation" techniques. Cf. Chapter X.
174 MAHLER'S compactness theorem

We consider now the point

y = U oa + UI b = (u oao + ul bo, U oal + U I bl) = (Yo' ~)
of A. By (1) and (2), we have
IP(~} = bl IP(u o{} +u l ). (4)
But now, by (3), we have

lim Yo = lim (a o + bo~) = ao - bo{} = bI I , (5)

"0 -to co 140 140

by (1) and (2). But

Yo IP(~) ~ c5 (F) ;
and so

by (4) and (5).

In particular, Theorem IV of Chapter II shows that
lim infuol uo{}
"0-+ 00
+ ull ~ 5- 1
for all {}: and there exist numbers {} for which the sign of equality is
required. Indeed the "successive minima" of Theorem IV of Chapter II
correspond to a sequence of successive minima here. The original proofs
of this used continued fractions, but there is a proof due to C. A. ROGERS
which uses the isolation techniques which will be discussed in Chapter X
and which is given in the author's Tract (CASSELS 1957a).
V.tO.5. The proof of Theorem XV gives a simple case when in-
equality necessarily occurs in Theorem II, that is, when we have a
convergent sequence of lattices,
M,-7M'
and a distance function F such that
F(M') > lim supF(M,}.
'--+00

Let F be the distance-function and M, the lattices occurring in the first

half of the proof. Then
F(M,) =0
for all r, since M, has points with xo=O. On the other hand, we con-
structed a convergent subsequence M" of the M, such that
M,,-7 N,
where
p+l (N) ~ D(IP: {}1"'" {}n).
The right-hand side here may well be strictly positive, as § 10.4 shows.
 Introduction 175

Chapter VI

The theorem of MINKOWSKI-HLAWKA

VI.l. Introduction. Hitherto we have been primarily concerned to
estimate the lattice constant L1(9') of a set 9' from below, that is to
find numbers L10 such that every lattice A with d (A) < L10 certainly has
points other than 0 in !/. In this chapter we are concerned with estimates
for L1(9') from above; that is we wish to find numbers L11 such that
there are certainly lattices A with d (A) = L11 which have no points other
than the origin in /1', i.e. are 9'-admissible.
HLAWKA (1944a) showed that if 9' is any bounded n-dimensional
set with a volume (content) V in the sense of JORDAN 1 and if L1 1> V,
then there is a lattice A with d (A) = L11 which is admissible for 9'. He
showed, further, that if 9' is a bounded symmetric star-body, then it
is enough that
(1 )
where
(2)
thereby confirming a conjecture of MINKOWSKI. These results were put
in a wider setting by SIEGEL (1945 a). Denote by N,9'(A) =N(A) the
number of points of A other than 0 in a set 9'; and by P,9'(A) = P(A)
the number of primitive 2 points of A in 9'. SIEGEL 3 gave a very natural
way to define averages over the set of all lattices A with a fixed deter-
minant d (A) = L1 1. If 'IjJ (A) is any function of a lattice A, let us denote
this average by
9Jl {'IjJ (A)}. (3)
II
SIEGEL showed that
9Jl {N,9' (A)} = V(9')/L1 1 , (4)
II
and
9Jl{P,9'(A)} = V(9')g(n) L1 1, (5)
II

where 9' is any bounded set, not necessarily a star-body and not
necessarily convex, which possesses a volume V(9') in JORDAN'S sense.
1 This is rather more restrictive than the sense of LEBESGUE, but if the volume
is defined in the sense of JORDAN it is also defined in that of LEBESGUE and equal
to it. Let X (xl be the characteristic function of .9', that is X (xl = 1 if xE 5 and
X (xl = 0 otherwise. Then .9' has a volume in the sense of JORDAN if X (xl is integrable
in the sense of RIEMANN, and the volume is equal to the integral of X (xl over all
space.
Z That is points aE A which are not of the form a = Rb, where bE A and
k> 1 is an integer.
3 For a particularly simple exposition of SIEGEL'S averaging process, see
MACBEATH and ROGERS (1958al.
t76 The theorem of MINKOWSKI-HLAWKA

HLAWKA'S theorems follow at once from (4) and (5). If .1 1 > V(9') , then,
from the definition of the average, there must certainly by (4) be at
least one lattice, say M, such that N9'(M) ~ rol(N9'(I\)) < 1. Since
A
N9'(M) is an integer, we must have N9'(M) =0, so M is 9'-admissible.
Similarly, if 9' is a symmetric star-body and Ll1> V(9')/2C(n), then
there must be some lattice N for which P9' (N) < 2. Since 9' is sym-
metric, points of N, other than the origin, occur in pairs, ±a, so
P9'(N) =0. Hence 9' contains no primitive points of N and, being a
star-body, can contain no points of N at all other than o.
The constant C(n) occurs in (5), roughly speaking, because the
probability that a point of a lattice 1\ chosen at random should be
primitive is {C(n)}-l. More precisely, the ratio of the number of primitive
points of 1\ to the total number of points of 1\ in a large sphere Iall < R
tends to {C(n)}-l as R~oo.
When 9' is convex, improvements of the Minkowski-Hlawka theo-
rem were obtained fairly soon after the original proof [see e.g. MAHLER
(1947b), DAVENPORT and ROGERS (1947a) and LEKKERKERKER (1957a)].
However, even so, the smallest value of

Q(9') = V(9') (6)

.1(9')
is not known even for 2-dimensional symmetric convex sets: though
the same conjecture was made independently by REINHARDT (1934a)
and MAHLER (1947c) that it is attained when 9' is a certain "smoothed
octagon", that is an octagon in which the corners are replaced by certain
hyperbolic arcs.
Mrs. OLLERENSHAW (1953a) has given an example of a 2-dimensional
non-convex symmetric star-body 9' for which Q(9') is smaller than for
the REINHARDT-MAHLER convex octagon and constructed from it a set
which is not a star-body for which

Q = 1.3173 ....
It is not known whether this is the smallest possible value for a 2-dimen-
sional set.
For a long time no improvement was obtained on the Minkowski-
Hlawka theorem for general sets or for star-bodies. However, almost
simultaneously, improvements were made by ROGERS (1955 a, 1955 band
1956a) and SCHMIDT (1956a and 1956b). ROGERS'S work depends on
elaborate estimates of the average

(7)
 Introduction 177

for positive integers k, where we have used the same notation as in (4).
In a later paper ROGERS (1958a), using ideas of SCHMIDT combined
with his own, shows that there is an absolute constant C such that
V(.9') 1 4
Q(9') = - - ~-nlog- - 210gn - C (8)
Ll(9') - 2 3

for all symmetric sets 1, provided that the dimension n is greater than
some absolute constant no. We shall not discuss ROGERS'S work further
but refer the reader to the original memoires. SCHMIDT, on the other
hand, uses an elegant device which is more effective than ROGERS'S
method for small dimensions but much less effective when the dimension
is large. We shall discuss it more in detail in § 4.
The work just described can be generalized in several directions.
In the first place, instead of operating with the number Ny (A) defined
above, one may consider more generally
2:.1(a), (9)
DEli
*0
where 1(~) is some function defined at all points of space and which
may be subjected to certain conditions (e.g. that it be non-negative or
Riemann-integrable). If f(~) is the characteristic function of 9', then
the sum (9) is just Ny(A). Again, one may confine the sum in (9) to
primitive points of A, when there is an analogue of Py(A). In fact
most of the work so far described has dealt with generalisations of this
kind. Again, it was shown by MACBEATH and ROGERS (1955a) that the
Minkowski-Hlawka theorem extends to more general sets of points than
lattices. It is enough for A to be any set of points such that the ratio
of the number of points of the set A in the sphere I~I < R to the volume
of the sphere should tend to a finite non-zero limit d as R -+ 00. Indeed
(4) continues to hold with a modified definition of the mean !In and with
Lll =d-1•
Finally, we observe that MAHLER'S Theorem V Corollary of Chap-
ter III often permits the results of this chapter to be extended to un-
bounded sets 9' on taking .9, to be the set of points of 9' in the sphere
1~I<r.
VI.l.2. In this book we shall not consider any of these generaliza-
tions in detail. In § 3 we shall prove the Minkowski-Hlawka Theorem
in its original formulation, that is, the existence of a lattice A admissible
for a symmetric star-body 9' with finite volume V(9') and with deter-
minant arbitrarily near to V(9'). We shall use an averaging argument,
but the type of average will be chosen to facilitate the proof, not for
1 Professor ROGERS tells me that Dr. SCHMIDT has obtained an improvement
of (8) which is in course of publication in Acta Mathematica.
Cassels, Geometry of Numbers 12
178 The theorem of MINKOWSKI-HLAWKA

any deeper reason l . Then in § 4 we shall give an improvement of the

Minkowski-Hlawka theorem using SCHMIDT'S ideas but not carrying
the detail quite so far as he does.
The arguments of §§ 3. 4 depend on a thorough investigation of the
properties of sublattices of prime index in a lattice and this is carried
out in § 2. These investigations further enable one to prove the result
conjectured by ROGERS that if Y is a symmetric star-body and
md (A) < L1 (Y) for some integer m and some lattice A. then Y contains
at least m pairs of points ±a E A other than o. This we do in § 5.
In § 6 we give an entirely different generalization of the Minkowski-
Hlawka Theorem which applies only in 2 dimensions. We show namely
that certain sets Y of infinite volume (= area) are of finite type, that
is. possess admissible lattices. The proof depends on a generalization
of a theorem of MARSHALL HALL (1947a) due to the author (CASSELS
1956a).
We do not use the contents of this chapter later in the book.
VI. 2. Sublattices of prime index. An important tool in the work
of both ROGERS and SCHMIDT is the existence of sublattices of a given
lattice with certain special properties. We shall use the definition and
properties of an index introduced in Chapter I.
LEMMA 1. Let p be a prime number and A an n-dimensional lattice.
Let a 1 , •••• a R be any points 0/ A which are not 0/ the shape pa, a E A
and let k1 • ... , k R be real numbers. Then there is a lattice M 0/ inde.x p
in A such that
(1)

Let b1 , ... , b n be a basis for A. Let c1 , .... Cn be integers and

O~Ci<P (1~j~n), (2)
(c 1 .... ,cn ) =4= (0 .... ,0). (3)

Let M (c1 , •• '. cn) be the lattice of points U 1 b1 + ... + Un b n, where

u 1 , ... , Un are integers, such that
U 1 Cl + ... + UnC n == 0 (P).
Clearly M(c 1 , ... , cn ) is of index p. There are P" -1 such lattices and
we now show that a point a r belongs to precisely pn-1 - 1 of them.

1 Other averaging processes have been used. For a particularly brief proof of
Theorem II using one of them. see CASSELS (1953 a). It has been shown by ROGERS
(1955a) that many of the averaging processes that can be used to prove the
Minkowski-Hlawka Theorem are essentially equivalent to SIEGEL·S.
 Sublattices of prime index 179

We have

where V,l"'" v'n are integers not all divisible by p, by hypothesis.

Without loss of generality, V rl is not divisible by p. The congruence
(4)
then determines C1 uniquely if C2 , •.• , c.. are given subject to (2). In
particular, (4) gives C1 = 0 if already c2 =··· = ; = 0; contrary to (3).
But cz, "', Cn may be given any other of the p ,. 1_1 possible sets of
values s';~ject to (2). Hence the average of th: left-hand side of (1)
over all lattices M = M (c1 , •.. , cn) is given by the right-hand side, and
so (1) must be true for at least one of them.
We have at once the
COROLLARY 1. Let p be a prime number and let aI' ... , at> be p points
0/ A none 0/ which is 0/ the shaP.e Pb, bE A Then th Ie is a lattice M
0/ index p in A which contains none 0/ aI' ... , at>.
For we may put k, = 1 for 1-;;'r-;;'p. For the lattice M of the theorem
we have
" 1 s: p.. - l _ 1 P< 1-
L....J - pn_1
orE M

The number p of points in the corollary cannot be replaced by p + 1.

It is easy to see that if aI' a 2 are any two points of A, then at least
+
one of the p 1 points
~, a2 +ra1 (O-;;'r-;;,p-1)
is in each sublattice of index p.
More generally we have the following corollary, due to SCHMIDT in
essence.
COROLLARY 2. Suppose that the number R 0/ points a, satisfies
R< pm+l_1
P-1
jor some integer m. Then there is a lattice M 0/ index p in A such that
pm-l_ 1
" k, -;;,
L....J pm-1 " k,.
L....J
(5)
orE M l;;;,;;;R

(i.e. n in (1) may be replaced by m).

If the dimension n of the space is -;;, m the result follows at once since
pm-l_ 1 > p.. -l_ 1
pm-1 = pn_ 1 if m~ n.
12*
180 The theorem of MINKOWSKI-HLAWKA

When n>m we use induction on the dimension n. We say that two

vectors a and a' of /\, neither of the shape pb, b (/\, are proportional
mod p if there is an integer u and a vector c of /\ such that
a =ua' +pc. (6)
Clearly u is prime to p. The relationship is a symmetric one between
a and a', since there is an integer v such that uv == 1 (P); and then
va =a' +pc'
or some c' E I\. Further, if a proportional both to a' and a", then a'
s proportional to a". We thus have a subdivision into classes or "rays".
The number of rays is clearly
P"-1
P-1
Since we are now supposing that n>m, at least one of these rays
must contain no members of the set a, (1 ~r~R). If c is in this ray,
it is of the shape c = w b where b is primitive and w is an integer priJl.le
to p. Hence the primitive point b is in the ray, and we may suppose
that b = b1 , where b1 , ... , b" is a basis for I\. Then every point a, is
of the shape
a,=v,lb1 +··· +t1,,,b,,,
where by the construction of b1 , at least one of V,2' ... , v,,, is not divisible
by p. Hence if we make a, correspond to the vector
0, = (V,2' .•• , v,,,)
in the (n -i)-dimensional lattice /\0 of points with integer coordinates,
then ii, is not of the shape p b, b E /\0' Since we are assuming that the
corollary has already been proved for smaller values of n, there exist
integers c2 , ••• , c" such that

The lattice M of points

with
C2U2 +"'+C"U,,=O (P)
then does what is required.
VI.2.2. A refinement of the argument gives a rather more special
result than Lemma 1 in which now the k, must be non-negative.
LEMMA 2. Let p be a prime-number and /\ an n-dimensional lattice.
Let a o, ... , a R be any R + 1 points all\ which are not 0/ the shape Pb,
 The Minkowski-Hlawka theorem t8t

bE" and let kl' ... , kR be non-negative real numbers. Then there is a
lattice M 01 index p in " such that
ao~M
and
(1)

We may choose a basis bl , ••• , btl for" such that

where Vo is some integer, which is not divisible by p by hypothesis.

For integers c; (2~j~n) with
O~c;<p (2~i~n), (2)
denote by N (c a, ... , cp) the lattice of points
ulbl + ... + u"b,.,
where the integers u1 , •.• , u" satisfy
Ul +CaU 2 +····+c,.u,.=O (p).
Clearly ao~ N (c a, •.. , cp).
For 1 ~r~R, let
a, = V,l bl + ... + v," b".
By hypothesis, not all of the integers V,l' ••• , v," are divisible by p.
If all of V,I' •.• , v," are divisible by p, then V,l is not divisible by p:
and so a, does not belong to any N (c a, •••• c,,). If, say, V,I is not divisible
by p, the condition
v,l+cav,a+···+c"v,,,=O (P)
is satisfied for precisely one value of CI if Ca • •••• C" are fixed; that is
a, belongs to precisely p"-ll of the p,,-l lattices N (c a, ••• , c,,). Hence
if M runs through all the p,.-l lattices N (c a, ... , c,,) the average value
of the left-hand side of (1) is
P-l};' k"
where };' denotes that the r for which v,a • ... , v," are all divisible by p
must be omitted. Since k,~o for all r, by hypothesis, this shows that
at least one of the lattices M =N(ca, ... , c,,) satisfies (1).
VI.3. The Minkowski-Hlawka Theorem. Following ROGERS (1942 b
and 1951 b) we now prove the following theorem of HLAWKA.
THEOREM 1. Let 1(a:) be a Riemann-integrable lunction 01 the variables
a: = (Xl' ... , X,,) which vanishes outside a bounded set. Let L1 l >O and 8> 0
182 The theorem of MINKOWSKI-HLAWKA

be given. Thm there is a lattice M of determinant .11 such that

.11 L f(a) < Jf{~) d~ + 6, (1)

oEM
,*0
where

We may suppose that I(x) vanishes outside the cube

(1~i~n). (2)

Let p be a prime number and let "I> 0 be determined by the equation

pTJtI = .11. (3)
We may choose p so large that
PTJ>S. (4)
Let A be the lattice of points

where u1 , ••• , uti are integers, so

d(A) =TJtI. (6)
Now
TJtI L f(a) <JI(x)dx +!s (7)
oEII
0,*0

if "I is small enough, by the definition of Riemann integration; and

so (7) is true when P is large enough, by (3).
A point a of A other than 0 for which f{a) =FO lies in (2); and
so cannot be of the shape pb, bE A by (4). Hence we may apply
Lemma 1 where ~, ... , a R are all the points a of A other than 0 at
which f(a) =FO and
k, = I (a,) .
Then M has determinant
d{M) = Pd(A) = pTJ" = .1 1, (8)
and
L f(a) ~ P":~~ 1 L f(a). (9)
oEM p oEII
,*0 0'*0

Finally, (1) follows from (3), (7) and (9), when p is chosen large enough.
As in § 1 we have the
COROLLARY. Let !/' be a set with Jordan-volume V(!/,) and let
.1 1 > V(!/,). Then there is a lattice M with d(M) =.11 which is admissible
for !/'.
 The Minkowski-Hlawka theorem 183

For let I (x) be the characteristic function of Y, and choose e so

that ,11> V(Y) +e. The number of points of M other than 0 in Y is
then
L I (a) < ,11 1 {V(.9') e} < 1,
oEM
+
0*0

by (1). Since the number is an integer, it must be 0.

VI.3.2. The result corresponding to Theorem I in which only primi-
tive points are summed over is:
THEOREM II. Let I (x), ,11 and e be as in the enunciation 01 Theorem I.
Then there exists a lattice M 01 determinant d (M) = ,11 such that

C(n) ,11 L* I (a) < f I (x) dx + e,

oE M

where the star (*) indicates that only primitive points are to be summed
over.
We only indicate briefly the modification required to the proof of
Theorem 1. In any case Theorem II is embraced in the generalization
of Theorem I to point sets /\ other than lattices due to MACBEATH and
ROGERS (1955 a), which was discussed in § 1. The exposition still follows
ROGERS (1947b and 1951 b).
In the first place, it is trivial that a point of M in the cube (2) of
§ 3.1 is a primitive point of M if and only if it is primitive as a point
of I\. Hence it is enough to show that

limrl"L*/(a) = {C(n)}-l fl(x)dx. (1 )

'1->0 oEA
Now
00

L I(a) = L L* I(ra).
aEA ,=10EA
0*0

Hence by MOBIUS' inversion formula [e.g. HARDY and WRIGHT (1938a)

Chapter XVI], we have

L*/(a) =L,u(r) LI(ra).

oEA , oEA
0*0
Hence
'f}" 'L* I (a) = 'L Ilr~) <1(rfj),
aEA ,;';1

where, for any ;>0, we have put

<1(;) =;" L 1(;11).
" integral
"*0
184 The theorem of MINKOWSKI-HLAWKA

But now a (E) is bounded for all E. and

lima(E)
e-+o
= f I(z) dz.
The result now follows on letting p-? 00, so "l ~ 0, since

L IJ,.~) = {C(n)}-l.
'~l

As in § 1 we have the
COROLLARY (The "Minkowski-Hlawka Theorem"). Let 9' be a bounded
symmetric star-body with volume V(9') and let 2C(n) Ll 1 > V(9'). Then
there is a lattice M with d(M) =Lll which is admissible lor 9'.
VI.4. SCHMIDT's theorems. We are now in a position to illustrate
SCHMIDT'S method of improving the corollaries to the last two theorems.
We first give a simple example
LEMMA 3. Let 9' be a symmetric star-body in n-dimensions with
Jordan-volume V(9') and let Lll be any number such that

3C(n) Lll> (1 + 21 -,,) V(9').

Then there is a 9'-admissible lattice M 01 determinant Ll1 •
Let g (:1:) be the characteristic function of 9', and let
I(z) = g(z) + 2g(2z),
so that
3 if zEi9'
{
1(:1:) = 1 if zE9', zEEt9'
o otherwise
and

Choose e so small that

By Theorem II with Ll1/2 for Lll and this e, there is a lattice A with
determinant

such that
L I (a) <6.
"E II. primitive

Since I(-z) =/(:1:), by the symmetry of 9', there is thus no primitive

point of A for which 1(0) =3, and so no point of A at all in i9' except o.
Further, there are at most two pairs of primitive points say ±a1 , ±a2
 SCHMIDT'S theorems 185

of 1\ in Y. By Lemma 1 Corollary 1, there is a lattice M of index 2

which contains neither a1 nor a2 • Since ai' a2 are not in l Y, the points
2~, 2a 2 of M are not in Y. Hence M is Y-admissible. Since

d(M) = 2d(l\) = L1 1 ,
the lattice M does what is required.
VI.4.2. When n = 2, the result of Lemma 3 is no stronger than
Theorem II Corollary.
By further elaboration, SCHMIDT (1956a) improved Lemma 3 some-
what but for values of n at all large Lemma 3 is weaker than the follow-
ing Theorem III which applies to all Jordan-measurable bounded sets
not merely symmetric star-bodies. To obtain results about symmetric
sets, Theorem III should not be applied to Y directly but, say, to the
"half-set" g;, of points

Then
V(g;) = l V(Y) ,
and a lattice Mis g;-admissible if and only if it is Y-admissible. There
is thus an additional factor 2 for symmetric sets.
THEOREM I II. Let Y be any hounded n-dimensional]ordan-measurable
set 01 volume V(Y\ and let Lfl be any number such that
(1 )
Then there is a lattice M 01 determinant Lfl having no points, except pos-
sibly 0, in Y.
Let g (z) be the characteristic function of S and put
j(z) = g(z) + 2g(2Z) + 3g(3z) + 6g(6z). (2)
Then
JI(z) dz = (1 + 2·2-" + 3 . r" + 6·6-") Jg(z) dz
= (1 + 21 -") (1 + 31 -") V(Y).
By Theorem I there is thus a lattice 1\ of determinant

d (1\) = Lfl/6, (3)

such that
L: 1(a) < 12. (4)
oE/\
0,*,O

We shall construct a lattice M of index 6 in 1\ with the required properties.

We classify the points a of 1\ in Y, other than 0, into four types
:t1 , :t2 , :ta and :te:
186 The theorem of MINKOWSKI-HLAWKA

(i) a is in Xl if it is not of either the shape a = 2b or a = 3 b with

bEl\.
(ii) a is in X2 if it is of the shape a = 2b but not of the shape a = 3 b,
with bEl\.
(iii) a is in Xa if it is of the shape a = 3 b but not of the shape a = 2b,
bEl\.
(iv) a is in Xa if it is of the shape a =6b, bEl\.
Let Nl , N2 , N3 , !Va be the numbers of lattice points in the correspond-
ing classes. Then by (2) and (4) we have
Nl + 3 N2 + 4N3 + 12!Va < 12, (5)
since, for example, the coutribution to (4) of aEXa is
1+2+3+6=12.
In particular, by (5),
!Va=o.
Suppose, first, that N3> o. We apply Lemma 2 with P= 2, taking
ao to be one of the N3 points in X3 and aI' ... , a R to be the remaining
points in Xa (if any) together with any points in Xl' The numbers k,
of the lemma are taken as 1 if a,EX l and 4 if a,E1:3 • Then, by Lemma 2,
there is a lattice r of index 2 which contains N;, N; points of Xl' X3
respectively, where
(6)

(the - 4 being the contribution of ao, which is definitely lost). All the
points of X2 are, of course, in r. By (5) and (6) we have
2 N; + 3 Nz + 8N; + 4 ~ 11.
Hence N; =0 and Nt' +Nz~i, so Nt' +Nz~3. But now 'by Lemma 1,
Corollary 1 there is a sublattice M of r of index 3 which contains
+
none of these N..' N 2 points. Then M does what is required.
We may thus suppose now that
Na=O.
We now apply Lemma 1, Corollary 2 with P= 3 to the points a, with
k,=1 if a,EXl and k,=3 if a,EX 2 • Since there are at most
11 < (33 -1)/(3 -1)

points a" we may take m = 2, so, in the notation of the corollary,

pm-l_ 1 1
P"'-1 -T'
 A conjecture of Rogers 187

Hence there is a sublattice r of index 3 which contains N{ , N2' points

of ~, 't2 respectively, where

N{ + 3N; ~ : (Nl + 3N2) ~ ~ < 3.

Hence N; = 0 and N1' ~ 2. By Lemma 1, Corollary 1, there is a sub-
lattice M of r of index 2 which contains non~ of these N1' points. This
lattice M does what is required.
Thus in every case we have constructed a lattice M of index 6 in 1\
which is admissible for g. Since
d(M) = 6d(l\) = ,11'
the lattice M has all the required properties.
As SCHMIDT remarks, Theorem IIi can be improved somewhat at
the expense of further elaboration; but for large n is weaker than
ROGERS' results which we referred to in § 1 and which we cannot prove
here. In particular the factor (1 + 21 -") (1 + 31 -") on the left of (1) may
be replaced by something smaller if g is a star-body, since then a point
in r-1g is automatically in rIg if t~r.
VI.5. A conjecture of Rogers. We digress now from the general
theme of the chapter to prove a result which was conjectured by ROGERS
(1951 a), who compares it with the generalization of Theorem II of
Chapter III from m = 1 to m> 1. It was proved by ROGERS when the
number m occurring in it is a prime and by SCHMIDT (1955 a) for all
except a finite number 1 of m. It has been proved generally in a rather
wider context by the author (CASSELS 1958a). We do not use it later.
THEOREM IV. Let g by a symmetric star-body and let 1\ be a lattice
with
md(/\) < ,1 (g), (1 )

where m~ 1 is an integer. Then g contains at least m pairs ±a of points

of 1\ other than o.
Theorem IV is an immediate consequence of the following theorem
in which the reference to star-bodies disappears.
THEOREM V. Let aI' ... , a R be primitive points 01 a lattice 1\ and let

1r (2)
be positive integers. Then there is a lattice of index at most
(3)
1 For all m;;;; 107 and all sufficiently large m, according to the review in Mathe-
matical Reviews!
188 The theorem of MINKOWSKI-HLAWKA

which contains none 01 the points

±i,a, (4)
We show first that Theorem V implies Theorem IV. Suppose that
1\ in Theorem IV contains fewer than m pairs of points of f/. Since f/
is a star-body. the points of 1\ in f/ can be put in the shape (4). where
the number of pairs is
J<m.
Hence by Theorem V there is a lattice M of index ;;;; m in 1\ which con-
tains none of these points. i.e. M is f/-admissible. Since

d (M) ;;;; m d (1\) < LJ( f/).

by (1). this is a contradiction to the definition of LI(f/).
The proof of Theorem V depends on the following lemma. which
gives the existence of primes with certain properties. It is due to
SYLVESTER (1892a) and was rediscovered by SCHUR (1929a) who
gave a rather simpler proof. The proof is in any case rather involved.
so we do not give it here but refer the reader to the original papers.
LEMMA 4 (SYLVESTER). Let X. Y be integers and
1;;;;X;;;;Y.

Then there is a prime number p>X which divides one 01 the numbers

Y+1 ..... Y+X.

We now prove Theorem V. Suppose first that R = 1. Since fit is
primitive. it may be taken as part of a basis for 1\:

where n is the dimension. Clearly the lattice M of points

~ b1 + ... + Un b,I'

where ~ •...• u,. are integers and

1).(] +
does all that is required.
We now consider the case whenR>1 and use induction onJ. Without
loss of generality
fl = 1;:i;,&R
max f,· (5)
 Unbounded star-bodies 189

Let p be the prime given by SYLVESTER'S Lemma 4 with

x = min VI ';2 + ... +iR) ,

Y = max &1' ; 2+ ... +;R) .
Then
p>X~j, (6)

Since p divides one of the numbers Y + 1, ... , Y + X, we have

[:]+[~]<[X~Y],
that is
(7)

where for any real number x we denote by [x] in this proof the integer
such that [x] ~ x < [x] + 1. By Lemma 2, there is a lattice r of index p
which does not contain ~ and such that

that is
" .: : ; [ii + ...P + iR ].
L.." 1,- (8)
o,Er

By (6), if a point i,a, in (4) with r> 1 is in r, then a, is in r. Since

~ is not in r, the only points (4) with r = 1 in r are the

± i~ (p~) (9)

But now, by the hypothesis of the induction argument, there is a lattice

M of index at most

1 + [; 1+ o~/'~ 1 + [;] + [i2+'; + iR]

in r which contains none of the points (4) at all. The index of M in A
is p times the index of M in r; and so, by (7), is

~p{1 + [;] + [i2+·;+iR]}~J<J+1.

This concludes the proof of Theorem V.
VI.6. Unbounded star-bodies. The results of §§), 4 extend to un-
bounded star-bodies. For example we have
THEOREM VI. Let f/ be a bounded or unbounded symmetric star-body.
Then
,1(f/) ~ {2C(n)}-1 V(f/). (1)
190 The theorem of MINKOWSKI-HLAWKA

When !/ is bounded this is just Theorem II, Corollary. When !/

is unbounded it follows from Theorem II, Corollary together with Theo-
rem V, Corollary of Chapter V.
In the same way any of the other estimates of §§ 3, 4 may be extended
to unbounded star-bodies !/, or indeed, to any open sets of finite volume
of which the origin is an inner point.
VI. 6.2. There certainly exist star-bodies !/ of finite type [i.e. with
L1(!/) < 00] and infinite volume. A 2-dimensional example is

(1 )

for which L1(!/1) = St, as we saw in Chapter II. More generally, in

n-dimensions the body

is of finite type but infinite volume, since admissible lattices are given
by the norm-forms of totally real algebraic fields of degree n (see
Chapter X). In general, in more than 2 dimensions it is very difficult
to decide whether a given star-body is of finite type or not. Two 3-
dimensional examples are discussed in CASSELS and SWINNERTON-DYER
(19SSa), for which a decision on this point would have interesting reper-
cussions. In 2 dimensions however there do exist general criteria which
we shall now discuss.
VI.6.3. From now on we put 1
n =2.
In an obvious sense, the body ~ defined in (1) of § 6.2 has two pairs
of asymptotic arms, the asymptotes being the Xl and X 2 axis. It is
possible to inscribe in !/ arbitrarily narrow parallelograms with one
pair of sides parallel to an asymptote and area 1, for example

In a sense ~ is a limiting case, since if it is possible to inscribe in a

star-body !/ parallelograms with centre the origin and arbitrarily large
volume (area), then !/ is of infinite type by MINKOWSKI'S convex body
Theorem II of Chapter III. Roughly speaking, any star-body with a
pair of arms wider than those of ~ is of infinite type. We now show
that a 2-dimensional star-body may have any finite number of arms
like those of ~ and still remain of finite type.
1 It is customary to call 2-dimensional star-bodies "star domains" but we do
not follow this usage. Similarly we may sometimes continue to speak of volume
where area is more usual.
 Unbounded star-bodies 191

THEOREM VII. Let

lor some s> 0, and let

(1~j~J)

be any linite number of indefinite quadratic forms. Suppose that the

distance-function F(XI' x2) satisfies
(2)

for all (Xl' X2)' Then the star-body

[/: F(XI' x 2 ) <1 (3)

is of finite type.
The exponent 2 in (2) is dictated by reasons of homogeneity.
We shall deduce Theorem VII from the following generalization of
a theorem of MARSHALL HALL (1947a) which is due to the author
(CASSELS 1956a).

THEOREM VIII. Let PI' "', PK be any real numbers. Then there exists
a real number oc such that

Iu! I(oc +- Pk) u +- vi> 8(K ~ 1)2 (1~k~K) (4)

for all integers u =1= 0 and v.

We first deduce Theorem VII from Theorem VIII, and then prove
Theorem VIII in § 6.4. After a suitable rotation of the co-ordinate
system, we may suppose without loss of generality that

f;(1,O) =1=0 (1~j~]);

and so
f; (Xl , xa) = Aj(XI +- {);X 2 ) (Xl +- CP; x2) (1~j~J) (5)
for real numbers;', {);, CPt such that

A; =1= 0, 1J; =1= cp;.

But now
If; (Xl' x2)1 ~ P; min {lx2 (Xl +- {); XI) I ' IXI (Xl +- CP; xl)I}, (6)
where
'"; = II A; II {); - CPt I > 0;
since if, for example

then we have
192 The theorem of MINKOWSKI-HLAWKA

We apply Theorem VIII where the Pl' ... ,PK are the {}i' fIJi in some
order, so K = 2J. Let IX be the number given by Theorem VIII, so that
IU{(IX + {}i) u + v}1 ~ '1 > 0 } (7)
IU{(IX + fIJi) U + v}1 ~ 7J > 0
or integers u =1= 0, v, where
7J = {8(2] +1)2}-I.
Let /\ be the lattice of points
(Xl' X 2) = R(IXU + V, u), (8)

where u, v run through all integer values and R is a positive number

yet to be chosen. If u=I=O we have, by (6) and (7)
Ili{R (IXU + v). Ru}1 ~ft;R27J. (9)
If however u=O but v=l=O then. by (5),
1/;(Rv, 0)1 ~ IA;I R2. (10)
Similarly
(11)

for all (Xl' X2)E/\ other than 0, on distinguishing the two cases u=l=O
and u =0, v=l=O in (8). We may choose R so large that the right-hand
sides of (9), (10) and (11) are all not less than 1. Then for all (Xl' X2)E/\
except 0, we have, by (2),
F(x}, x 2) ~ min 1/;(x}, x 2)! ~ 1;
O;fi,;;fi,J

that is /\ is ..9'-admissible. This concludes the proof of Theorem VII.

VI. 6.4. We now prove Theorem VIII which was enunciated in
§ 6.3. Write
(1)

We shall construct a sequence of open intervals J_I , J o, .1;., ... which

enjoy the following three properties:
(i)m Jm+l is contained in J m.
(ii)m J m is of length ,,-2m-2.
(iii)m the inequality
Ui(1X +PI<) u + vi> t,,-4 (1~k~K) (2)
holds for all numbers IX in J m and for all integers v and u with
O<u~"m.
 Unbounded star-bodies 193

If we can construct the J". we shall have proved Theorem VIII,

since there is a number a. contained in all the intervals J". and then
(2) holds with this a. for all integers u>O and v.
We may take J_ 1 to be the interval O<a.< 1, since there are no
integers u in (3) with m = -1. We thus assume that J". has already
been constructed and construct J".H. By (ii)"., the open interval J".
is the set of a. satisfying
a.' < a.< a." (4)
for some numbers a.' and a." for which
(5)
For each k (1 "!i:k"!i:K), there is at most one fraction Vk/U k in its lowest
terms such that
- (:: +Pk)EJm , 0< uk"!i: ""'+1, (6)
since two fractions v/u with 0<u;~:~,,"'+1 differ by at least "-2,,.-2.
By (iii)"., we have
Uk> "m (1"!i: k~ K). (7)
Let f§ be the set of a. such that
(8)
and
(9)
for all k in 1"!i: k"!i: K for which a Vk!U k of the type (6) exists. Then ~
consists of at most K + 1 intervals. Their total length is
a." - a.' - ,,-2 ".-4 - L ,,-4 u;2
k
~ ,,-2m-2 _ (K + 1) "-2,,.-4
= (K + 1) "-2,,,-4,
by (1), (5) and (7). We may therefore find in ~ an open interval J"'H
of length exactly "-2,,.-4. Then J"'H satisfies (i)". and (ii)".H' by
construction. It remains only to verify (iii)"'H. We may clearly suppose
that U and v are coprime and that
,,"'< U ~ "",+1 (10)
by (i)". and (iii)",. If V/U =Vk/Uk is a fraction of the type (6), then
UI(a.+Pk)U+vl>i,,-4 (11)

for all a.EJm-H' by (9). Otherwise - (: + Pk) is not in J"" and so

I: + (a. I
+Pk) > ~ "-2,,,-4
for all a.EJ"'H' by (8); then (11) follows, by (10). Thus J"'+l has all
the required properties.
Cassels, Geometry of Numbers 13
194 The quotient space

Chapter VII

The quotient space

VII. I. Introduction. Before resuming the general study of the geo-
metry of numbers, it is convenient to introduce here the concept of the
quotient space of an n-dimensional space by a lattice. This concept
plays an important role in the discussion of inhomogeneous problems in
Chapter XI: but we shall also need it in Chapter VIII as it gives the
most natural interpretation of MINKOWSKI'S theorem about the succes-
sive minima of a convex body with respect to a lattice.
In § 2 we give the definition and most important properties of a
quotient space. In § 3 we prove a result which will be basic for one
topic in Chapter XI.
VII.2. General properties. Let A be a lattice in n-dimensional eu-
clidean space. Two points Yl, Y2 of the space are said to be congruent
modulo A, written
(1)
if the difference Yl- Y2 is in A This relationship is clearly symmetrical
in Yl and Y2. If
Yl == Y2 (A) , Y2 == Ya (A) ,
then
Yl == Ya (A).
The points Y may therefore be divided into classes t) so that two points
y and y' are congruent if and only if they are in the same class. A class t)
consists of all the points Yo+a, where Yo is some fixed member of t)
and a runs through all points of A
If
Y' == Y (A) , z' == z (A) ,
then clearly
Y' + z' == Y + z (A) .
Hence there is no ambiguity in deIining the sum t) + 3 of two classes
as the class to which Y +z belongs when Y, z are any members of t), 3
respectively.
Similarly, if t is an integer, the definition of tt) as the class to which
ty belongs when Y is in t) is unambiguous. On the other hand, if t is
not an integer, it is not, in general, true that ty'=ty when y'=y.
Hence tt) for real numbers t other than integers must be left undefined.
So far, of course, we have only followed the standard procedure for
finding the quotient group of an abelian group (namely the additive
group of all vectors) by a subgroup (namely the additive group of vectors
in A). We shall say that the classes t) are points of the quotient space
9l/A, where [}l will denote the original n-dimensional euclidean space.
 General properties 195

VII.2.2. Let F(x) be any distance function defined in 9t and put!

F(t)) =infF(y) (1)
yEI)

for t)EBi/A This is the function which will be important in inhomo-

geneous problems (Chapter XI). Note that
F(o) =0, (2)
where 0 is the class to which 0 belongs. For reference we enunciate the
principal properties of F("£), "£EBi/A, in the following lemma.
LEMMA 1. Let F(x) be a distance lunction and let F("£) be delined, as
above, lor "£EPi/A Then
(i) F(t"£) ~ tF(!) lor integers t ~ 0.
(ii) II F(x) is convex, then so is FI"£), in the sense that

F("£ + t)) ~ F("£) + F(t))

lor all ~, t).
° °
(iii) I I F(x) = only lor x = 0, then F(t) = only lor ~ = o. Further,
lor each t)EBi/A there is a YEt) such that F(t)) =F(y).
(iv) II 1\ (x), ~(x) are two distance lunction and 1\(x)~cF;{x) lor
some number c and all xEPi, then 1\ (~) ~ c~ (~) lor all FPi/A
Here (iv) is an immediate consequence of the definition (1). By the
definition of a distance function, we have F(tx) =tF(x) for all real t> 0.
Hence, if t>o is-an integer, we have
F(t!) =infF(y)~infF(tx) =tinfF(x) =tF(~).
YEI~ zE~ zE~

This establishes (i). The proof of (ii) is similar and may be left to the
reader.
It remains to prove (iii). Let t)EBijA and let yoEt), so that the general
element of t) is yo+a, aEA By Lemma 2 of Chapter IV, there is a
constant c>o such that F(x)~clxl for all x; and so
F(Yo+a) ~ cIYo+al;S cllal-IYoil·
In particular, if F(a+Yo)~F(yo), we have

lal ~ IYol + c-1F(yo)' (3)

There are only a finite number of aEA in (3). Hence there exists an
aoEA such thatF(yo+ao) = inf F(Yo+a). By definition,F(t)) =F(yo+ao)'
aEA
Further, F(t)) =0 only if F(Yo+a o) =0, that is t) =0.
1 There should be no confusion with the usage of Chapter IV, since there the
arguments were lattices; and here they are classes with respect to a lattice.
13·
196 The quotient space

VII.2.3. Let~, (1~r<oo) be a sequence of elements of BIll\. We

say that the sequence tends to ~' EBI/A if

lim I~, - ~'I = 0, (1)

'-+00

where, in conformity with the notation of § 2.2, we have written

1~ 1= inf 1xl·
orE~
(2)

LEMMA 2. A necessary and sufficient condition that ~,~~' is that

there exist elements y,E tJr and y' EtJ' such that

y,-+y'. (3)

Suppose, first, that the y" y' exist such that (3) holds. Then

I~, - ~'I ~ Iy, - y'l;

so (1) holds, that is ~r~~'.
Suppose, now, that (1) holds. By Lemma 1 (iv) there exist Z,E~,-~'
such that
1z,l = 1 ~, - ~'I-

Let y' be any element of ~' and put y,=1/' +zr. Then the Yr clearly
have all the properties required.
VII.2.4. Let
(1 )

be any basis for I\. Then every point x of space can be put uniquely
in the shape
(2)

for some real numbers ~l' ... , ; n; and x EA if and only if ~l' ••• , ~n are
integers. Hence to every vector x there is a unique aEA such that

(3)
where
(4)

In other words, every ~t.Bl/A has precisely one representative YE~ in

the half-open parallelopiped fjJ defined by (3) and (4). We say that
this parallelopiped is a fundamental parallelopiped for I\. Different
bases b j in general give rise to different fundamental parallelopipeds.
An immediate consequence of Lemma 2 and the existence of a
fundamental parallelopiped is
 General properties 197

LEMMA 3. The quotient space fJll/\ is compact. That is, any sequence
I), (1 ~r<oo) 0/ elements 0/ fJlj/\ contains a convergent subsequence:

t)" ~ t)'. (5)

The fundamental parallelopiped & is not compact, since although
it is bounded it is not closed. Let 1# be its closure, that is the set of
points (3) with O~1Jj~ 1 (1 ~i~n). Let y, be the representative of I),
in &. By WEIERSTRASS'S compactness theorem (§ 1.3 of Chapter III),
there is a convergent subsequence

y,,~y',

where y'E9. Then (5) holds by Lemma 2, where y'Et)'.

VII.2.S. We are now in a position to introduce a measure into the
quotient space fJlll\. Let 5 be any set of elements of fJlll\. We call a
set .9 of elements of fJI a set of representatives for 5 if (i) for each
tES there is precisely one ZE t which belongs to .9 and (ii) each zE.9
belongs to an tE S. We say that S is measurable if at least one set .9
of representatives is measurable.
Let ~ be the set of elements ZE& of the shape

z=y+u, yE!i, uEA,

where .9 is any measurable set of representatives of 5 and & is a funda-
mental parallelopiped. By Theorem I Corollary of Chapter III, the set
~ is measurable, and
V(~) = V(.9).

In particular, if .9, .9' are any two measurable sets of representatives

of S, we have V(.9) = V(.9'). This common value will be denoted by

m(S)

and will be called the measure of S.

Clearly the measure of the whole of the quotient space is the volume
of the fundamental parallelopiped &, that is d(/\).
Let't be any homogeneous mapping of n-dimensional space fJI onto
itself. In a natural way, it gives a mapping of fJlj/\ into fJlj't/\, which
we may also denote by 'to If m' is the measure defined in fJlj't/\ in the
way that m is defined in fJll/\, then clearly

m'('tS) = Idet('t)1 m(S)

for any set S in fJljl\.
198 The quotient space

VII.3. The sum theorem l • If C and 0 are two sets of points in

+
the quotient space 1;1//\ we denote by C 0 the set of all points

c + b, where CE C, bE O.
This section is devoted to proving
THEOREM I. Let C and 0 be non-empty sets in 1;1//\ with measures
m(C) and m(O) respectively.
(i) II m(C) +m(O»d(/\), then C +0 is the whole space 1;I/A.
(ii) II m(C) +m{O)~d(A), then m{C + 0) ~ m(C) + m(O).
This theorem is due to MACBEATH (1953a). It was discovered inde-
pendently by KNESER (1955 a), who first recognized its importance for
the geometry of numbers. Theorem I is, in fact, now only part of a
much wider theory, for which see KNESER (1956a) and the literature
cited there. It falls into the same circle of ideas as the so-called "rJ. +P
hypothesis" about the densities of sequences of integers which was first
proved by MANN. As all this is rather aside from the main theme of
the book we do not discuss it further. It is convenient to prove Theo-
rem I here but the application to the geometry of numbers will not be
made until Chapter XI.
Part (i) of Theorem I is easy. Suppose that there is a point !" of
1;1//\ which does not belong to C + O. Then none of the points

!"-C, CEC (1 )
can belong to D. We may denote the set (1) by !"- C. Clearly

m(!"-C)=m(C). (2)

But 0 and !" - C have no points in common, so

m{!" - C) + m(O) ~ m(1;I/A) = d(A).

Then m(C) +m(O)~d(/\), by (2) and (3). This proves (i).
In what follows we denote, as is conventional, by C f"\ 0 and C v 0
the sets of points which belong to both C and 0 and to either C or 0
(or both) respectively. We note for further reference the identity

m(Cf"\O) + m(CvO) = m(C) + m(O); (4)

which becomes clear on noting that points of C f"\ 0 occur in two sets
on each side of (4), but points of C v 0 other than those of C f"\ 0 occur
1 The results of § 3 will not be needed until Chapter XI.
 The sum theorem 199

in precisely one set on each side. Further, we show that

C + D) (C("\D) + (CvD) (5)

( ) means "contains"). For let

Suppose b belongs to C: then we may regard a as belonging to D since

it belongs to both C and D. Hence a + b = b +aE C +D. Similarly, if b
belongs to D we regard a as belonging to C.
It follows from (4) and (5) that, if the conclusions of Theorem I are
true when C("\ D, C v D are read for C, D respectively, then the con-
clusions are also true for C and D themselves. This is one of the principal
ingredients of the proof. The other is provided by
LEMMA 4. There is some !EfJl/A such that

d(A) mf(C +!) ("\D} = m(C) m(D).

Before proving Lemma 4 we complete the proof of Theorem I with
its use. Let C, D be two sets with

m(C) =yd(A), m(D) = CJd(A)

and

If y = 0, the conclusions of the theorem certainly hold, since C is non-

empty, by hypothesis, and if CE C the set C+ D, which is contained in
C +D, has measure m(D) = m(C) + m(D). We may thus suppose without
loss of generality that
(6)

Now let! be given by Lemma 4, and put

C1=(C+!)("\D, D1={(C+;r;)vD}-!.
Write
m(C1) =y1d(A). m(Dl) = CJ1d(A) ,
so that

and
Yl = yCJ
by (4) applied to C +! and D and by Lemma 4 respectively. Further,
200 The quotient space

by (5) applied to C +~ and D. We may now repeat the process on

C1 , D1 • In this way we get a sequence of sets C" D, with measures
y,d (/\), b,d (/\) respectively, such that

C+D)C,+D" (7)
and
y, + b, = l' + b, (8)
1', = 1',-1 b,-I' (9)
But now, by the argument used when l' =0, it is certainly true that
m (C, +- D,) ~ m (D,) = b, d (/\). (10)
It follows from (6), (8) with 1 - 1 for 1 and (9), that

y,~ 1',-1 (1 - 1"-1);

and so

Hence
1', --+ ° (1--+00). (11)

b,--+y +b (1 --+ (0), (12)

by (8). But
m(C + D) ~ b,d(/\) , (13)
by (7) and (10). In letting 1 --+ 00 in (13) and using (12) we have
m(C + D) ~ (1' + b) d(/\) = m(C) + m(D)
as required.
It remains only to prove Lemma 3. We note, first, that
(14)
varies continuously with~. This is clearly true with the "well-behaved"
sets C and 0 to which we will wish to apply Theorem I, but it is in fact
true for all measurable C and 0, see for example A. WElL (1951 a).
In the second place, in an appropriate sense, to be explained more fully
below, the average of (14) as ~ runs through fJt//\ is m(C) m(D)/d(/\).
Perhaps the simplest way is to observe that we may introduce integration
in fJt//\ in the obvious way. Let tp(~) be a function defined in fJt//\ and
let 1(a:) be the function in fJt such that

when a: belongs to the class~. Then we write

f tp(~) d~ = f 1(a:) da:,
fit/A !1'
where f!jJ is a fundamental parallelopiped. Exactly as in § 2.5, one may
show that this definition is independent of the choice of fundamental
 Introduction 201

parallelopiped fP. Let tp(~), X(~) be the characteristic functions of C,D

respectively; so that
m {( C + ~) "D} = Jtp (t) + ~) X(t)) d t).
at/A
Then

But
Jtp(t) +~) X(t)) d~ = X(t)) m(C).
dl/A

Hence, on interchanging the order of integration in (15), we obtain

J m{(C +~)" D}d~ = m(C) JX(t)) dt) = m(C) m(D).

dl/A dl/A

Since 9l/A has measure

m(91/A) = J 1 d~ = d(A),
dl/A

the truth of Lemma 4 now follows from the continuity of m{( C +~) "D}
and the connectedness of 9l/A.

Chapter VIII

Successive minima
VII!.t. Introduction. For some purposes one requires to know not
merely that a lattice A has a point in a set .9', but that it has a number
of linearly independent points in .9'.
Let F(a;) be an n-dimensional distance function and A a lattice. If
for some integer k in 1 ~ k ~ n and some number A the star-body

(1 )

contains k linearly independent points

(2)

of A, then so does 1-'.9' for any I-'>A, since the points (2) are also in
1-'.9'. We define the k-th successive minimum Ak =Ak(F, A) of the dis-
tance function F with respect to the lattice l A to be the lower bound
of the numbers A such that A.9' contains k linearly independent lattice
points. Clearly

1 Or of the lattice with respect to the distance functio~.

202 Successive minima

The numbers AI' ... ' A" defined above certainly exist, since if
aI' ... , a" are any n linearly independent points of A, then, trivially,

Ak~ A,,~ max F(a;).

1~1;$"

In the notation of § 4 of Chapter IV we have

A1 = F(I\) = inf F(a). (4)
aE/\
,*0
Hence, by the definition of
d (F) = sup po (A) (5)
/\ d(A) ,
we have
A~~ d(F) d(I\). (6)
The remarkable inequality
AI ... A,,~ 21(,,-1) d(F) d (1\) (7)
was discovered independently by ROGERS (1949a) and CHABAUTY
(1949a); and CHABAUTY (1949a) and MAHLER (1949a) independently
produced examples to show that if ~ is any number < 21 (,,-1) then there
are distance-functions F and lattices 1\ such that
(8)
We shall give the elegant proof of (7) in § 3 and give the construction
of the counter-example to show that it cannot be improved in the case
n = 2. The difficulties in extending the counter-example to n dimensions
are purely algebraic. It can be shown easily by means of an example
that

can be arbitrarily small, so there is no lower bound analogous to the

upper bound (7) [but see (13) below for symmetric convex F].
The inequality (7) holds with a suitable definition of the terms not
merely to star-bodies F(x) < 1 but to all point sets [/' whatsoever.
There have been several different definitions of the successive minima
of an arbitrary set [/'. We do not discuss these further, but refer the
reader to the papers quoted for the extensive literature.
It was shown already by MINKOWSKI (1896a, § 51) that, when F(x)
is the euclidean distance lxi, the inequality (7) may be replaced by
AI . .. An ~ d (F) d (1\) . (9)
We give his proof in § 2. More generally, it has been conjectured that
(9) holds for all symmetric convex distance functions. In § 4 we shall
show for these F that
(10)
 Introduction 203

which is equivalent to (9) when n =2. The inequality (10) was ap-
parently discovered by CHALK and ROGERS (1949a) and CHABAUTY
(1949a) independently. It has been shown by WOODS (1956a and
1958b) that (9) continues to hold for n = 3 when F is symmetric and
convex and for n = 2 when F is convex but not symmetric: the proof
is distinctly intricate and we do not discuss it here. For general n
and symmetric convex F, RANKIN (1953a) indicates that the constant
2*(,,-1) can be replaced by a rather smaller one.
For symmetric convex functions F and any n, there is a result going
back to MINKOWSKI (1907a) which may be regarded as a substitute
for the unproved conjecture that (9) holds. In our notation, MIN-
KOWSKI'S convex body Theorem II of Chapter III states that

(11)

where VF is the volume of F(:e) < 1; and so Ai VF is the volume of the

body F(:e) < AI' which, by hypothesis, contains· no point of A except o.
MINKOWSKI'S theorem is that in fact

(12)

The proof of (12) remains difficult. Simpler proofs than the original
have been given by DAVENPORT (1939c) and WEYL (1942a). We follow
WEYL in § 4, since the ideas introduced will be needed in Chapter XI.
For symmetric convex F there is also an inequality
2"
AI' .. A" Vj...~ In. d (A) , (13)

the almost trivial proof of which is also given in § 4. From (12) and
(13) it follows that the product Al ... An is determined by VF and d (A),
except for a factor which is bounded in terms of n.
In general, it is hopeless to expect more information about successive
minima than can be deduced from the formulae for the product AI'" A".
For example, let AI' "', A" be any numbers such that

Al~A2~"'~A,,; A1 •• ·A,,=1.
Then the lattice A of points
(Ut, ... , Un' integers)
has d (A) = 1 and has successive minima AI' ... , An with respect to the
distance function
max 1x·11 '
F(:e) = 1;:;;,;:;;"
as is easily verified.
204 Successive minima

VIII. 1.2. For later purposes we shall often need the following two
simple lemmas.
LEMMA 1. Let AI' ... , An be the successive minima 01 a lattice A with
respect to a distance lunction F associated with a bounded star-body
F(:ll) < 1. Then there exist n linearly independent points ~, ... , an EA
such that
(1~i~n).

II aEA and F(a) <Ai' then a is linearly dependent on ~, ... , ai-I.

For by the definition of An there are n linearly independent points
of A in
F(:ll) < An + 1. (1)
By Lemma 2 of Chapter IV, the set (1) is bounded and so contains only
a finite number of lattice points. Only these points need be considered
in the definition of the Ai. The truth of the lemma is now obvious.
LEMMA 2. Let AI' ... ' An be the successive minima 01 the distance
lunction F with respect to the lattice A. Then there is a basis
bl , ... , bn
01 A such that, lor each i = 1,2, ... , n, the inequality
F(x) <Ai
implies that
:ll = U l b1 + ... + u j - l b j - 1
lor integers ut, ... , Uj-l·
When F(x) = 0 only for :ll = 0, this is a trivial consequence of Lem-
ma 1, since we may choose b l , ... , b n so that a i for each i is dependent
only on bl , •.• , bi • by Theorem I of Chapter 1.
Otherwise a slightly more refined argument is needed. In general,
the Ai will not be all unequal, but there are numbers
#1<fl2<···· <fl.,
for some s in 1 ~s~n,such that
All = flt if
where
o =ko<~< ... <k.=n.
By the definition of successive minima, there is no point of A with
F(a) <#1 except, possibly 1, o. Since

fl2> All"
For a general distance function F(z) there is, of course, no reason why Al
I
should not be o. Indeed, if F(z) = IXl ••• X" Il/n, we have Al = ... = A" = 0 for the
lattice 11.0 of points with integer coordinates.
 Spheres 205

the are kl linearly independent points

(2)
of A in F(~) <Ps, and, since

every other point of A in F(~) <Ps is linearly dependent on them.

Similarly, we may find ks linearly independent points of A in F(~) <Pa
such that every other point of A in F(~) <Pa is linearly dependent on
them. Since Ps<Pa we may suppose that kl of these k2 points are
~, ... , ak, already determined. We may thus denote by

the maximal linearly independent set of points of A in F(~) <Pa without

disturbing the notation (2). And so on. In this way we obtain kS - 1 < n
points

of A such that
(t ~ s).
By Theorem I of Chapter I there is a basis bl , " ' , b .. of A such that,
for each i = 1, ... , k s - l , the vector aj is linearly dependent on bl , ... , b j
only. This basis clearly has all the properties required.
VIII.2. Spheres. We first prove the results for spheres. since they
are simplest and the treatment forms the model for what follows.
THEOREM 1. Let
(1)
and let AI' ... , A.. be the successive minima 01 a lattice A with respect to
Eo. Then
d(A) ~ AI'" A.. ~ d(Eo) d(A). (2)

The left-hand side of (2) was substantially proved in Theorem XIII

of Chapter V. We have on the one hand

1det(~ •...• a ..)1 =I d(A) ~ d(A),

where I is the index of ~, ... , a.. in A. and, on the other hand,

Idet(~, ...• a..)1 ~ 1~1···la.. 1

by HADAMARD'S Lemma 9 of Chapter V. If now the aj are the linearly
independent vectors of A with F(a;) = Ai given by Lemma 1, the required
inequality follows at once.
206 Successive minima

It remains to prove the second part of (2). As in the proof of Lemma 9

of Chapter V, there is a set of mutually orthogonaF vectors C1 , " ' , c n
such that
bi = tjl Cl til ci + ... +
for some real numbers ti; (n~j), where bi is the basis given by Lemma 2.
By incorporating a factor in c; we may suppose, without loss of generality,
that
(1~i~n).
Then

and so
IL ui bj12 = L (L Ui tj
• J~'
Y (3)
We now show that
L .1.;-2 (L Ujtjit~ 1 (4)
i i~i

for all sets of integers tt =1= o. For let Ul , ... , Un be integers, and suppose
that
(j> J). (5)
Then ulbl+ .. ·+unb n is not dependent on bl, ... ,b j - l ; and so
ILUjbiI2~A}. (5')
Further, (5) implies that all the summands in (3) and (4) with i?- ]
are O. Hence, and since Ai~Aj if j~J, the left-hand side of (4) is
L A;-2 (L uitiiy~ L AJ 2(L ujtii)2 = AJ 21L ui bil2~ 1,
i;;:'j i~i i~j i~i i

by (3) and (5'). Hence if I\' is the lattice with basis

(1~j~n),
we have
IL u bil
j 2 ~1
for every point L ui bj =1= 0 of 1\'; that is

Po (1\') = II\' I ~ 1. (6)

On the other hand,
d (1\') = All ... .1.;;-1 d (1\). (7)
But now
IA'I"
- IMln
- < sup - - = t5 (Eo) , (8)
d(A') = M d(M)

1 We say that two vectors a, b are orthogonal if their scalar product ab

vanishes.
 General distance-functions 207

by the definition of d (Po). The right-hand side of (2) follows now from
(6). (7) and (8). This concludes the proof of Theorem I.
VIII.2.2. As was remarked in Chapter V. the theory of successive
minima shows that the hypotheses of Theorem III and IV of Chapter V
are equivalent. This we do now.
LEMMA 3. The lollowing two statements A and B about a set 2 01
n-dimensional lattices 1\ are equivalent. where x. K. ,10' ,11 are supposed
to depend on 2 but not on I\.
(A) there exist ,11 < 00 and x> 0 such that d (1\) ~ ,11' and 11\1 ~ x > 0
lor all I\E 2.
(B) there exist ,10> 0 and K < 00 such that d (1\) ~ ,10> 0 and the sphere
Ix I~ K contains n linearly independent points 01 1\. lor all 1\ E2.
If AI •...• An are the successive minima of Po (x) =Ixl with respect
to 1\. then clearly (A) and (B) are equivalent to

(A) d(I\)~L11' Al~X>O.

and

respectively. We now use the inequality

d (1\) ~ AI' .. An ~ d (Po) d (1\) (1 )
of Theorem I. Suppose first that (A) holds. Then
d(l\) ~ {d(Po)}-IAl ... An~ {d (Fo)}-I xn = ,10 (say).
and

These are the two conditions (B).

Suppose now that (B) holds. Then
AI~ (AnAn-l ... A2t l d(l\) ~ K- n+lL1 o=y. (say).
and
d (1\) ~ )'1'" An ~ Kn = ,11 (say).
These are the two conditions (B).
VIII.3. General distance-functions. We first prove a lemma which
will be required later. Just in this section we denote by {x} the fractional
part of x. that is. the number such that
O~{x}<1. x-{x} = integer.
LEMMA 4. Let 'f}1 •...• 'f}n be any real numbers. Then there tS a
number 'f} such that
(1 )
208 Successive minima

For any number ~ we have clearly

{~} + {_ n ={O if {n.=O} ~ 1-

1 otherWIse
Hence

Thus there is at least one k such that (1) holds with rJ =rJk'
We shall require only the more specialized
COROLLARY. Let fll' ... , fl .. be any numbers such that

°< fll~ fl2~ ... ~ fln' (2)

°
Then there exists a real number fl> and positive integers m1, ... , m"
such that
(i) mj +1!mj is an integer (1 ~j< n),
(1~j~n),
and

We shall in fact take all the mj to be powers of 2, say

m·=il
1 (1~j~n). (3)
Let
fli = 2'11 (1~i~n) (4)
for real numbers rJi; and let rJ be the number given by Lemma 4. By
subtracting an appropriate integer from rJ we may suppose, by (2) and
(4), that
rJ ~ rJl~ rJ2~ .. , ~ rJ,,'
If now fl = 2~ and the integers Ii are defined by
rJi - rJ = li + {rJj - rJ},
then the numbers mj defined by (3) clearly satisfy (i) and (ii). Further,
by the lemma,

which is just (iii).

VIII.3.2. We are now in a position to prove
THEOREM II. Let F(~) be a distance-function and A.1 , ••• , A.n its succes
sive minima with respect to a lattice A. Then
(1 )
 General distance-functions 209

We denote by b 1 , ••• , b1l the basis for /\ given by Lemma 2. Let

fl and the integers mj be given by Lemma 4, Corollary when flj = Aj
and let I\' be the lattice with basis

Then
d (I\') = II (fl mj) -1 d (/\) . (2)

We now show that

F(/\,) ~ 1. (3)
Any point a of I\' other than 0 may be put in the shape

a = u 1 b~ + ... + U J b J' UJ =F 0,
where u1 , ••. , U J are integers. Then

where
(1~j<]),

are integers, since u j and mJjm j are integers. By Lemma 2, since vJ=F 0,
we have
F(flmJa) ~ AJ'
Hence
F(a)~~ ~1.
JlmJ
This proves (3).
Finally,
F" (I\') < (4)
d (I\') = b (F) ,

by the definition of b(F). The required inequality (1) now follows from
(2), (3), (4) and the inequality

II(/~j) ~2l(n-ll
1

of Lemma 4, Corollary.
A rather more detailed argument shows that the sign of equality
in (1) cannot hold if F(x) < 1 is a bounded star-body. Then it is possible
to ensure that there are not n linearly independent points a of I\' with
F(a) = 1, so I\' cannot be critical, and there is inequality in (4). See
ROGERS (1949a).
Cassels. Geometry of Numbers 14
210 Successive minima

VII!.3.3. We now show that the constant 2~(n I) in Theorem II

cannot be improved. For reasons of algebra we treat only the case
n = 2.
For general n see MAHLER (1949a) or CHABAUTY (1949a).
We first consider a point set which is not a star-body. Denote by
C(j' the set of points
C(j': (±t,O)

and by C(j" the set of points

where

Finally, let Y be the set of points which belong neither to C(j' nor
to C(j". Clearly Y is open, and if any point x is in Y, then rx is in Y
for O~ Irl ~ 1: so Y has some of the attributes of a star-body. We shall
later modify Y slightly to obtain a set .c;; which actually is a star-
body.
There certainly exist Y-admissible lattices I\, i.e. lattices having
only the origin 0 in Y. For example the lattice 1\2 of points

where U1 , U2 are integers, is Y -admissible, since if U2 =f= ° the point

(2u1 , u 2) is in C(j" and if u 2 =O, but U1=f=O, then (2u\, u 2 ) is in C(j'. We
shall next show that
(1 )
that is that every Y-admissible lattice 1\ has determinant d (1\) ~ 2.
Let 1\ be any Y-admissible lattice. By MINKOWSKI'S ·convex body
Theorem II of Chapter III, there is certainly a point x other than 0
of 1\ in

This point is not in Y, so must be in C(j' or C(j" and hence has the shape

We may suppose without loss of generality that b\ is primitive. There

is then a vector
b 2 = (b\2' b22 ) ( 1\,
which, with b\, forms a basis. Hence
 General distance-functions 211

Since b 2 is in the 9'-admissible lattice A, it must be in re' or re", so

b12lb 22 = rational.

Similarly b i +b 2 is in re' or re", so (bu +bl2 )/b22 is rational; and hence

bn lb 22 = rational.

There thus exists a real number ~ > 0 and integers B u , Bu , Bn such

that

Without loss of generality; B u , Bl2 and B22 have no common divisor

except ±1.
Let v be the product of the primes which divide Bn but not Bl2 •
Put
B~2 = V Bll + B l2 .

We wish to show that B~2 is prime to Bn: and must distinguish two
cases for the prime divisors p of Bw If P does not divide Bl2 , then it
divides v. If p divides Bl2 then it does not divide B u , since B u , B l2 , Bn
have no non-trivial common divisor; and p does not divide v. In both
cases p does not divide B~2' Hence, on replacing b 2 by b 2 +vb1 , we may
suppose that Bl2 and B22 have no common non-trivial divisor.
Now b 2 is in the 9'-admissible lattice 1\, so is in re' or re". Hence
I~I ~1,
since Bn and Bn have no common factor. Similarly b i is in re' or f(j",
and so

Hence
either I Bnl = 1, I~I ~ 21 ,
or IBnl ~ 2, I~I ~ 1.
In either case,
d(/\) = IBllBn~21 ~ IBn~21 ~ 2.
This concludes the proof of (1).
We denote, as usual, by fl9' the set of points
fl9': fl~, zE9';
and by Ao the lattice of points ("I' "2) with integer "I' "2' Clearly if
fl;;;;;' 2- 1 there are no points of 1\0 except
0 in ,It 9'; if r l <fl ;;;;;'1, there
are only the further points (± 1, 0) of 1\0 in fl9'; while if fl> 1, the
points (± 1, 0) and (0, ± 1) are in fl9'. If 9' were a star-body F(z) < 1,
14*
212 Successive minima

these statements would imply that the successive minima of 1\0 were
Al = 2-', A2 = 1. Hence
).IA2 = 2' (L1(9'))-ld(l\o);
which is the case of equality in Theorem II if (.1 (9'))-1 is written for
b(F), the two being equal for star-bodies.
It remains now to modify 9' so as to obtain a bounded star-body,
in such a way that its successive minima with respect to 1\0 remain
2- 1 and 1, and so that its lattice constant is arbitrarily close to 2. We
do this by replacing the lines in rr' and rr" by narrow wedges.

Fig. 9. The shaded portion is ", (tI)

Let £>0 be arbitrarily small. For any vector Y=(Yl' Y2)'*0. let
1r, (y) be the set of points z for which
ir,(y) : XIY1+X2Y2-£-llxIY2-X2YI!;;;;Y~+Yi. (2)

Then ir, (y) is an infinite wedge having a vertex at Y, see Fig. 9. Its
precise shape is not important. The two sides of the wedge make the
small angle ±arc tan £ with the outward radius vector from 0 to y .
Now let rr; be the set of points in ir, (2 6, 0) and 1r, (- 26, 0) and

U2'*
let rr;' be the set of points in ir, (u 1 , u 2 ) for any pair of integers with
0. Finally let y. be the set of points in

which do not lie either in rr; or in rr:'. Clearly y. is a star-body, since

there are only a finite number of the wedges composing rr; and ~:'
which have points in common with the disc (3). Indeed, by (2) and
(3), the distance-function F.(z) as!'ociated with y. may be written down
explicitly.
 Convex sets 213

Since .9, is contained in f/, and since the points (±2i , 0), (0, ±1)
are evidently still boundary points of .9" at least when 8 is small enough,
it follows that the two minima A1 and A2 of Ao with respect to F.(~)
are r1 and 1 respectively.
Further,
.1(.9,) ~ .1(9') = 2.
Indeed
lim .1(.9,) = .1(f/) = 2
....... 0+
by Theorem V of Chapter V. Hence there exist e such that
21 t5 (F.) d (Ao) = 2* (.1 (.9,) )-1
is arbitrarily close to A1 A2 • This shows that for n = 2 the constant
26(,,-1) = 21 in Theorem II cannot be improved.

VIII.4. Convex sets. We shall often have occasion to refer to the

results of § 3.1-3.4 of Chapter IV and in particular to the properties
of tac-planes.
We first need a general lemma about convex functions.
LAMMA 5. Let F(~) be a symmetric convex distance function associated
with a bounded convex body F(~) < 1. Let c =F 0 and let Tr be the plane
through the origin parallel to a tac-plane at c to F(:I:) <F(c). Then
F(y + p.sc) !?: p.F(y + sc) (1 )
for aU y in Tr, all real s, and all It in
0<p.<1-
If s = 0 there is nothing to prove. Otherwise we may suppose, by
homogeneity, that
s = 1,
since s-ly is in Tr if Y is. Then
F(y + c) ~ F(c) , (2)
by the definition of a tac-plane. Then, by convexity,
F(y+c) ~F(y+p.c) +F{(1-p.)c} }
(3)
=F(y + p.c) + (1 - p.)F(c).
The required inequality (1) with s = 1 now follows from (2) and (3).
We may now prove
THEOREM III. Let F(~) be a symmetric convex distance-function
associated with a bounded body F(:I:) < 1 and let Al , •.. , A" be the successive
214 Successive minima

minima 0/ a lattice 1\ with respect to F. Then there is a lattice I\' with

determinant
d (1\') = ( AI~:l ) d (1\) (4)

and successive minima Ai (1 ~j~n), where

(5)
Let b 1 , " ' , b n be the basis for 1\ given by Lemma 2. Let ± c be
the points on the boundary of F(x) < 1 at which the tac-plane is parallel
to the plane IT through b 1 , .•• , b n - 1 (Theorem IV of Chapter IV). Then
every point in space can be uniquely put in the shape
x=y+sc, YElT. (6)
We put
fl = An - 1/A n ,
and define I\' to be the lattice of all points
Y+flsc, y+sc,:/\. (7)
Then (4) dearly holds. If s =1= 0 in (7), the point Y + s c is not linearly
dependent on b 1 , ... , b n - 1 ; and so
F(y + sc) ~ An'
Hence
(s =1= 0) (8)
by Lemma 5. On the other hand, the points of I\' with s = 0 are just
the points of 1\ which are linearly dependent on b 1 , ... , b n - 1 . Hence
(8) implies (5).
COROLLARY 1.
).~-1 An ~ b (F) d (1\).
For in the proof of the Theorem put

fl = AI/An
instead of A,.-l/A,.. Then (8) becomes
F(y +flsc) ~flJ.,. = Al (s =1= 0); (8')
so
F(a');;;; Al
for all a' EI\' except o. That is,
F(I\') ;;;; AI'
Further,
d (1\') = fl d (1\) = t-t.) d (1\) . (4')
 Convex sets 215

But
P (1\') ~ 1J (F) d (1\')

by the definition of 1J (F); and then the corollary follows from (4')
and (8').
COROLLARY 2.
1 1
AI'" A,,~ 22(n-l)-,.- 1J(F) d(/\).
We only sketch the proof. By varying fl in the proof of the theorem,
we may obtain a lattice I\' with successive minima where Ai,
(1~j<n),
and
d (1\') ~ A~:1 d (/\) .
Then
Al ... A" < A; ... A~
d(i\) = d(R)'
Hence it is enough to prove the corollary when An - 1 =An • But it is
easy to see that if two of the numbers 'fJj in Lemma 4 are equal, then
the right-hand side of (1) of § 3.1 may be replaced by ~ (n -1) - : .
When this improvement is inserted in the proof of Theorem II, it gives
the corollary.
VIII.4.2. Before treating MINKowSKI'S estimates for the product of
the successive minima of a bounded symmetric convex body in terms
of the volume we must first prove a result, which we shall also use
later, relating to convex bodies and the quotient space !Jill\. We
shall use the concepts and notation of Chapter VII. As was done
there, we denote the points of fJl by small bold letters and those of
!JIj/\ by small gothic letters.
THEOREM IV. Let F(;r) be a convex symmetric distance-function
associated with a bounded convex set
Y: F(;r) <1 (1 )
of volume
Vp= V(Y). (2)
Let /\ be a lattice with successive minima AI, ... , An with respect to F.
For real t>O denote by Sit) the set of t.)E!JII/\ which have at least one
representative y in t Y (i.e. F(y) < i). Then ihe measure m {S (i)} of S(i)
satisfies ihe inequality
= tn VF if i ~ VI 1
m{S(i)} 1 ~(-Pl) .. ·aAJ)tn-JVF if !AJ~i~-~AJ+1J
~ (-}A}) ... (iAn) Vp if t ~ iAn.
216 Successive minima

We first examine how the hypotheses and conclusion are affected

by a homogeneous linear transformation T. Let N=TA, F'{~) =F(T-l~).
The successive minima of N with respect to F' are the same as those
of A with respect to F. Clearly
Vp, = Idet(T)I VF ,
and by the remarks at the end of § 2.5 of Chapter VII we have
m' {T S(t)} = Idet (T)I m {S (t)},
where T S(t) is the image of S(t) in the natural mapping of 9t/A onto
9tlTA; and m' is the measure in 9tlTA But TS{t) =S'(t) is the set in
9tlT A defined in respect of F' and N as S(t) was defined in terms of
F and A. Hence a homogeneous linear transformation multiplies both
sides of (3) by the same factor Idet (T)I.
We may therefore suppose without loss of generality that the basis
b l , ... , b,. for A given by Lemma 2 is just
;-1 ,.-;
b;=e;=(~,1,~); (4)
and that A = Ao is the lattice of points with integer coordinates.
We now obtain a formula for m {S (t)} valid when
t;£ jAJ+1' (5)
and J = 1, 2, ... , n -1. Let
~l = (xu, ... , X,.l)' ~2 = (xu, ... , x1I2)
be two points of F(OI:) <t;£jAJ+1; and suppose that
(6)
Then
F(~l - 01:2) ;£ F{:r.l ) + F(~2) < AJ+1'
Since Zt -0I:2 EAo, we have now
(i > J), (7)
by (4). Further,
(8)
where At is the J-dimensionallattice of points with integral co-ordinates.
Clearly (7) and (8) together imply (6). Denote by 9tJ the J-dimensional
euclidean space and by mJ the measure in 91J/At. For given (n - J)-
dimensional vector z = (zJ+1' ... , ZII)' denote by SJet, z) the set of points
of 91J/At which contain representatives {Xl' ... , XJ} E 91J such that
F(x1 , ••• , XJ, zJ+1' ... , Zn) < t. (9)
Then we assert that (5) implies
m{S(t)} = J m,{SJ(t,z)}dz (10)
 Convex sets 217

In the first place, 5J (t, z) certainly has a J-dimensional measure, since

F(~) is continuous by its definition as a distance function. Then, if z
runs through all (n - f)-dimensional space and y = (YI, ... , YJ) runs for
each z through a complete set of representatives for 5J (t, z), it follows
from the equivalence of (6) to (7) and (8), that
~ = (YI' .. ·' YJ' ZJ+I' .. ·' zn)
runs through a complete set of representatives for 5 (t). We may, for
example, normalize the y by taking always O~ Yi< 1 (1 ~i~J). This
proves (10).
The next stage is to show that if s is any number ~ 1, so
O<t~st, (11)
then
mJ{5 J (st,sz)};;;; mJ{5 J (t,z)} (12)
for any (n - f)-dimensional vector z. This is certainly true if the right-
hand side of (12) is O. Otherwise, there is some J-dimensional vector
yo= (YIO' ... , YJo) such that
F(yo,z) < t
where, in an obvious notation, (Yo, z) = (YIO' ... , YJo, ZJ+I' ... , zn): and
similarly later. Let y be any J-dimensional vector with
F(y, z) < t.
Then by the convexity and homogeneity of F(~), we have
F{y + (s -1) Yo, sz} =F{(y,z) + (s -1) (Yo,z)}
~ F(y, z) + (5 -1) F(yo, z)
< t + (s -1) t
=st.
Hence, if Y runs through a complete set of representatives for 5J (t, z),
then y + (5 - 1) Yo runs through representatives of distinct elements I of
5J (st, sz), when Yo is kept fixed. This proves (12).
Suppose, now, that
0< t ~ st ~ V'J+!' (13)
Then, by (10) and (12) we have
m{5(st)} = J mJ {5J {st, z)} dz 1
= sn-J J mJ {5 J (st, sz)}dz
(14)
~sn-J J mJ {5 J(t,z)}dz
= sn-J m {5 (t)},
1 Of course not every element of SJ (s t, sz) necessarily has a representative
of the type y + (s - 1)yo' What is important, is that distinct y mod At give
distinct y + (s - l)yo modAt.
218 Successive minima

where in the second line we have replaced z by sz and in the third

line we have used (12).
When
(15)
we have the simple equation
m{S(t)} = V(tY') = tnVp., (16)
where t Y' is the set F(;r) < t. Indeed, if ;rI and ;r2 are any two points
of tY' with ;rI==;r2 (Ao), we have
F(;rI - ;r2) ~ F(;rI) + F(;r2) < 2t ~ AI;
and so ;rI = ;r2'
We may now prove (3) . For t ~ .pI' the truth of (3) follows from
t
(16). Suppose that (3) is already proved for t ~ AI' where 1 ~ J ~ n - 1.
Its truth in the range t AI ~ t ~ t Al +1 then follows from (13) and (14)
t t
with t = AI' Finally, the truth of (2) for t~ An is trivial, since Y' (tI )
includes Y'(t 2) if tI~t2: and hence m {S(t)} increases with t.
VIII.4.3. Theorem IV provides the kernel of the proof of the follow-
ing theorem of MINKOWSKI.
THEOREM V. Let F(;r) be a symmetric convex distance-function as-
sociated with the bounded set F(;r) < 1 of volume VF . Let AI, ... , An be the
successive minima of a lattice A with respect to F. Then

(1)

In Theorem IV the measure m {S (t)} for any t can be at most the

measure of the whole space f!.ljA, namely d (A). On applying this remark
when t=jAn to the inequality (3) of § 4.2 we get the right-hand side
of (1) at once.
Now let aI' ... , an be the linearly independent points of A with

F(a j ) =Aj

given by Lemma 1. By the homogeneity and convexity of F(;r), all

points
(2)
such that
(3)

lie in F(;r) < 1. Hence Vp. ~ V' where V'is the volume of the set of (2)
subject to (3). But clearly

= -I det(a I ,· .. , a,.) 1=
I 2 n 2n I
V -d(A), (4)
n! n!
 Polar convex bodies 219

where I is the index of a 1 • •••• a" in /\. This proves the left-hand side
of (1). since I~ 1.
COROLLARY. The index I of a1 • .••• a" is at most n!.
This follows from (4) and the right-hand side of (1). (Compare the
proof of Theorem X of Chapter V.)
VIII.5. Polar convex bodies. Let /\* and F* be the respective polars
of the lattice /\ (Chapter I. § 5) and the symmetric convex distance-
function F (Chapter IV. § 3). MAHLER (1939b) has shown that the
successive minima of /\* with respect to F* are determined by the suc-
cessive minima of /\ with respect to F apart from factors which have
bounds depending only on the dimension n. Thus relationship will be
exploited in Chapter XI in the discussion of inhomogeneous problems
and is of importance in other contexts. The theorem is. of course.
closely related to Theorem VI of Chapter IV dealing with the lattice
constants of mutually polar convex bodies.
THEOREM VI. Let AI' ...• An be the successive minima of a lattice /\
with respect to the symmetric convex distance-function F and let At I A!
••••

be the successive minima of the polar lattice /\ * with respect to the distance-
function F* polar to F. Then
(1~j~n). (1 )
We attack first the left-hand inequality. By Lemma 1 there exist
linearly independent vectors ai' af of /\ and /\* respectively such that
F(ai ) = Ai' F*(aj) = Ai. (2)
By Theorem III of Chapter IV we have
F(~) F*(~*) ~ ~~*

(scalar product) for any two vectors ~ and ~*. On applying this
~=±a,. ~*=±a' for any pair of indices i,j we have

since F(~) and F*(~) are symmetric. But a,af is an integer by Lemma 5
of Chapter 1. and so
either A, Ai ~ 1 or a,af = O. (4)

Let I be a fixed index. The vectors ~ such that ~at = 0 (1 ~ i ~ I)

form an (n - I)-dimensional subspace. Hence by the linear independence
+
of the a i there is some a j with j ~ n 1 - I which does not lie in this
subspace; that is
220 Successive minima

for some i, i with

i~I, i~n+1-I.

Then ).:~).1, ).j~)."+I-I' and so, by (4),

Since this is true for any I, this gives the left-hand inequality of (1).
We now prove the right-hand inequality in the enunciation. Let
a j (1 ~i~n) be as above. Then (d. Chapter I, § 5) there are n primitive
vectors bf of A* such that

(i =1= i) . (5)

Since the a j are linearly independent, the n equations a j ;):* = 0 are

satisfied only by ;):* = 0: and so

(1~i~n). (6)
Hence the b! are linearly independent.
By Theorem III, Corollary 1 of Chapter IV, there are vectors ;):;
such that
F(;):;) F*(bf) = ;):ib;*. (7)

Without loss of generality

(1~i~n). (8)

The next stage is to show that for fixed J the determinant D J formed
from;):J and the a j (i=l=]) has absolute value at least d(A). For fixed J,
there is a basis ct, ... , c! for A* with

(9)

Let c j (1 ~i~n) be the polar basis, so that, by (5) and (9),

aj = L VjiC; (i =1= J) (10)

1:0;:0"-1

for some integers Vj;. Further,

;):J = ± c" + L t, c i
l~i:o,,-1

for some real numbers ti , by (8) and (9). Hence

DJ = Idet (ai' ... , a.1-1';):J' a HI' ... , a,,)1 = Idet (Vji};,~.fll det (cl , ... , c,,)I.
,*"
 Polar convex bodies 221

The first factor here is a non-zero integer since the a i are linearly in-
dependent; the second factor is just d (/\). Thus

(11)
as required.
The points

with
ItJI F(xJ) +i*,J
L Itil F(a < 1j)

lie all in the set F(x) < 1 of volume Vp. This set of points has volume

Hence, and by (11),

(12)

and finally
(13)

by (7), (8). The inequality (13) holds for each integer] and for the
independent vectors by of /\*.
Now AI;:;;; A2;:;;;"';:;;; An and so, for each integer], there are the n+ 1 - ]
linearly independent vectors b*=br (J~j~n) of /\* such that F(b*)
~n!Ajl. By the definition of A:+ 1 -J it follows that

A:+1-J;:;;; n!Ajl.
This is the required inequality and so concludes the proof of the theorem.
The applications of the theorem are usually only qualitative so the
magnitude of the factor n! on the right-hand side is usually irrelevant.
MAHLER (1939b) showed that the weaker inequality

AJA!+l-J;:;;; (n!)2
can be deduced very simply from the left-hand inequalities, Theorem V
and Theorem VI of Chapter IV. We have

~A1· .. An~2"d(/\),
VpoAt ... A!;:;;; 2"d(/\*) ,
222 Successive minima

and so
VF VF* n AjA:+1_;
;
~ 22K d (A) d (A*).

Now
d(A)d(A*) =1
by Lemma 5 of Chapter I, and
22 "
VF~·~ (n!)1

by Theorem VI of Chapter IV. Further,

JI
;
A;A:+1_;~AJA:t-l-J

for any particular J by the left-hand inequality of Theorem VI. Hence

AJ A: _J ~ (n !)2, as required.
VIII.5.2. In Chapter XI we shall also need the following result of
which the proof is similar to that of Theorem VI.
THEOREM VII. Let F(~) and F*(~) be polar symmetric convex distance
functions. Let b 1 , ••• , b" be any basis of a lattice A and bt, ... , b: the
polar basis of the polar lattice A *. Then
2"d(A) F*(bj) ~ n! VF F(b;) n (1)
;*J
for each integer J = 1, 2, ... , n.
For the deduction of (12) from (5) and (6) in § 5.1 did not depend
on the fact that the a i gave the successive minima for F. Hence (12)
of § 5.1 remains true if b j is read for aj' where ~j is to be given by (7)
and (8) of § 5.1. On substituting (7) and (8) into (12) of § 5.1 the required
result follows.
COROLLARY [M. RIESZ (1936a), K. MAHLER (1939a, b)). tet AI' ... ' A"
be the successive minima of F with respect to A Then the basis b i may be
chosen so that
F(b 1) = Al
2F(b;) ~j Ai (2~j~n) }
(2)

and
(3)
The existence of a basis b j satisfying (2) follows at once from Lemma 8
of Chapter V on defining a j there to be the linearly independent points
with F(a i ) =A;. But now on multiplying (1) by F(b J) and using Theo-
rem V, we have
2"d(A)F(b J)F*(bj) ~ n! ~ II F(b j ) ~ (-i)"-I(n!)2VF
l~j:;;"
n Aj~2(n!)2d(A).
l;:;;j;:;; ..
 Introduction 223

Chapter IX

Packings
IX.1. Introduction. If Y is any n-dimensional set and y a point,
we denote by Y + y the set of points
(1 )
By a packing of Y in some other set :T we shall mean a collection of
sets
(2)
each of which is contained in :T, and no two of which have points in
common. If:T is the whole space 9t we speak simply of a packing
of Y. If the Yr in (2) run through the points of a lattice A then we
say that the packing is a lattice packing. In this chapter we examine the
consequences of these ideas for the geometry of numbers. This chapter may
be regarded as a sequel of Chapter III but we shall also require some of
the general properties of convex bodies discussed in Chapter IV. We
shall find that the general theory of packings is relevant even to strictly
lattice-theoretic problems.
There is an admirable account of the theory of packing in FEJES
T 6TH (1953 a) and a conspectus of the more important results in
BAMBAH and ROGERS (1952a).
IX.l.2. The three following theorems show the relevance of packings
to the theory of Chapter III. We give the simple proofs here
THEOREM I. A necessary and suflicient condition that the lattice A
give a packing 01 the set Y is that no diflerence ~1 - ~2 01 two distinct
points 01 Y belong to A
Suppose, first, that ~1-~2=aEA Then the sets Y=Y+o and
Y +a both contain the point ~1 =~2+a, and so overlap. Conversely,
suppose that the sets Y +a1 and Y +a2 have the point y in common
where "t, a 2 are in A Then the two points y-a1=~1' y-a2=~2 are
in Y, and their difference a 2 -"t is in A
BLICHFELDT'S Theorem I of Chapter III shows that
V(Y) ~ d(A)
whenever A packs Y. The following theorem shows when the sign of
equality can occur. To avoid irrelevant topological considerations we
confine attention to rather special sets Y.
THEOREM II. Let Y be a bounded open star-body and A a lattice with
V(Y) = d (A). (1 )
224 Packings

(A) 1/ A packs !/', then every point in space either belongs to precisely
one set!/' +a, aEA and is not a boundary point 0/ any other!/' +a,
or is a boundary point 0/ at least two such sets !/' + a.
(B) 1/ every point 0/ space either belongs to or is a boundary point 0/
at least one set!/' +a, then A packs !/'.
By hypothesis, there is an R such that !/' is contained in

We now prove (A). Suppose, first, that A packs!/' and that there
is some point y which is not in or on the boundary of any!/' +a. aEA
We may choose e in the range 0<£<1, so small, that the sphere 9;.
of points a: with
9;.: Ia:-yl<e (2)

is completely outside the finite number of bodies!/' + a with aE A and

Ia - y 1< R + 1. By the definition of R, the set!/' + a certainly contains
no points a: of 9;. if Ia - yl ~ R + 1. We may suppose, further, that e
is so small that the only point of A in Ia: I < 2 £ is o. Let

!/" = !/' v 9;.

be the set of points belonging to either!/' or 9;.. Clearly, if a:1 and a: 2
are distinct points of !/" the difference a:1 -a: 2 cannot belong to A
Hence
V(!/");;;;; d(A)

by BLlCHFELDT'S Theorem I of Chapter III. But then V(!/,) < V(!/") ,

which contradicts the hypothesis. Suppose now that the hypotheses
of (A) are fulfilled and that there is a point y which is on the boundary
of precisely one !/' +a, aEA. Suppose, without loss of generality,
that y is on the boundary of !/'. As before, there is an £>0 such that
9;. defined in (2) contains no point or boundary point of any !/' + a
with aEA, a=Fo. But then the point (1 +7J)y, for sufficiently small
7J> 0, is in 9;. and is not a point or boundary point of!/,. On taking
(1 +7J)y instead of y, we thus have the case first considered. No point
can belong to more than one !/'+a, aEA by the definition of a packing.
If y were a point of !/'+ a and a boundary point of !/'+ b, where
a, bEA, then there would be points in the neighbourhood of y in both
!/' + a and!/' + b, since!/' is open. This completes the proof of (A).
We now prove (B). If!/' is not packed, then, by Theorem I, there
are points a:1 and a:2 such that
 Introduction 225

Since!/' is open, by hypothesis, there is an e>O such that both spheres

~: 1;1: - ;1:11 < e,

~: 1;1: - ;1:21 < e
are contained in!/'. We may suppose that e is so small that ~ and ~
have no points in common. Let!/" be the set of points which belong
to !/' but not to ~. Clearly every point in space is either an inner
point or a boundary point of !/" + a for some aE A, since every point
+
of !/' is either in !/" or in !/" ao' Let !/" be the closure of !/,'. Since
!/' is a star-body and ~ is a subset of !/', we have

V(!/,') = V(!/,') < V(!/,) = d (A).

This is a contradiction with the Corollary to Theorem I of Chapter III

since we are supposing that every point, and so every point of the
fundamental parallelogram, is of the form z + a where ZE !/" and aE A.
This completes the proof of Theorem II.
THEOREM III. A necessary and suflicient condition that the convex
symmetric set!/' admit the lattice A is that A give a lattice packing of t!/'o
This follows at once from Theorem I and Theorem II, Corollary of
Chapter III.
We shall consider only packings of convex sets!/' in what follows,
and we shall suppose that !/' is symmetric, whenever this gives any
simplification of proofs or results.
IX.1.3. MINKOWSKI'S convex body Theorem II of Chapter III states
that if // is an n-dimensional symmetric convex body of volume
V(!/,) >2"d(I\), then the lattice 1\ cannot be !/'-admissible. In § 2 w~
discuss when a lattice I\c can be admissible for a convex symmetric
body of volume 2"d(Ac). Of course then by MINKOWSKI'S convex body
theorem we have
L1(//) = r"V(!/,) , (1 )

and the lattice Ac is critical.

Even when!/' is the cube Ix;1 < 1 (1 ;;;;;l;;;;;n), the critical lattices were
not completely known until HAJ6s (1942) confirmed on old conjecture
of MINKOWSKI. We quote the result here, but shall not prove it since
it depends on considerations of group-algebra remote from the other
topics in the book.
THEOREM IV. A necessary and sufficient condition that a lattice A
be critical tor Ix;I<1 (1;;;;;l;;;;;n) is that, after a suitable permutation ot
Cassels, Geometry of Numbers 15
226 Packings

the axes 0/ co-ordinates, it has a basis 0/ the shape

b 1 = (1,0, 0)
b 2 =(bI2 ,1,0, 0)
b 2 = (b1a , b2a , 1,0, ... ,0)

b" = (bIn' ... , bn-1,n, 1).

The reader will readily verify that a lattice of the stated kind has
determinant 1 and no points other than 0 in Ixii < 1 (1 ~j ~ n). For
the proof of the converse the reader is referred to the original paper of
HAJOS (1942) and to REDEl (1955 a) where there are references to the
considerable amount of later literature. We proved HAJ6s' Theorem for
n = 2 incidentally as Lemma 7 of Chapter III.
MlNKOWSKl (1896a) showed that any convex symmetric set Y with
LI(Y) = 2 -n V(Y) must have very special properties, for example that
it must be a polyhedron bounded by at most 2" - 1 pairs of hyperplane
faces. We prove this in § 2.
IX.l.4. VORONO! (1908a) suggested a simple way of finding open
convex symmetric sets Y such that
V(Y) = d (A)
and which are packed by a given lattice I\. If g(x) is any positive
definite quadratic form, the set of points such that
g(x) infg(x + a)
< aE/\
has this property. The condition
g (x) < g(x + a) ,
for any given a, is linear in the coefficients XI' ... , Xn; so Y is convex.
Y is clearly symmetric. It is not difficult to verify that Y is, in fact,
bounded; and that then the infimum in (1) may be replaced by a
minimum over a finite number of a depending on A and the function
g(x), but not otherwise on the individual x. Not every open convex
symmetric Y with V(Y) =2"d(A) for which A is admissible may be
obtained in this way, but VORONO! was able to show that all, in a
sense, sufficiently general such Y could be. Unfortunately the excluded
cases include some of great interest, such as those covered by HAJOS'S
Theorem IV.
We do not discuss the case of general dimension n in this book but
deal in detail with n = 2 in § 3. As a byproduct we obtain a result
about the inhomogeneous problem for definite binary quadratics.
 Introduction 227

IX.I.5. Let:ft be any open 2-dimensional set and !/ the 3-dimen-

sional set of points
tt': (X I ,X 2 ,Xa) (X I ,X2 )E:ft !xa!<1; (1)
that is, a generalized cylinder of height 2 and with cross-section :ft.
Then a plane section
X3 = constant

of a lattice packing of tt' gives, in an obvious way, a packing of :ft,

but not necessarily a lattice packing. The idea of using non-lattice
pac kings in this context is apparently MAHLER'S (1946g). In this way
we are led to consider non-lattice packings of 2-dimensional sets. This
we do in § 5, after some preparatory lemmas in § 4. It turns out, as
was proved independently by ROGERS (1951a) and FEJES TOTH (1950a)
[see also FEJES TOTH (1953a)] that, in a sense which will be made
precise, no packing of convex symmetric open sets is closer than the
closest lattice packing. It appears unlikely that this result extends to
higher dimensions. For a discussion of this point see FEJES T6TH
(1953 a).
In § 6 we use the packing results to show that
Ll (tt') = Ll (:ft), (2)
when :ft is convex and symmetric and tt' is defined in (1). This result
was originally proved independently by CHALK and ROGERS (1948a)
and YEH (1948a). An example was given by ROGERS (1949b) which
shows that (2) need not hold when :ft is a symmetric non-convex
2-dimensional star-body, and DAVENPORT and ROGERS (1950b) gave an
example to show that then the ratio Ll(tt')jLl(:ft) may be arbitrary small.
VARNAVIDES (1948a) has shown that (2) continues to hold in one inter-
esting non-convex case. It is trivial that Ll(tt')~Ll(:ft) for any:ft, since
if 1\ is a 2-dimensional admissible lattice for tt', the 3-dimensionallattice
of points
(Xl' X 2 , X 3 ) (Xl' X 2 ) E1\, X3 = integer

is clearly admissible for tt' and has the same determinant as I\.
There is an interesting unsolved problem in this connection. Let
~ and Jt; be convex symmetric bodies in n l and n 2 dimensions respec-
tively and let tt' be the (nl + n 2)-dimensional "topological product" of
~ and Jt;; that is the set of points

;r = (y,z), YE~, ZEJt;.

The argument above shows that
Ll (tt') ~ Ll(~) Ll(~). (3 )
15·
228 Packings

Can it ever happen that there is strict inequality here? The cylinder
is, of course, the case ~ = 2, n z= 1. WOODS (1958a) has shown that there
is equality in (3) for n1=3. n z=1 when ~ is a 3-dimensional sphere.
IX.l.6. In §§ 7, 8 are given applications by BLICHFELDT techniques
based on packing considerations, or at least BLICHFELDT'S Theorem I
of Chapter III, to the estimation of the lattice constants of the sets
x~+ ... +x:<1
and
IX1 ... X,.1 <1
respectively. The relationship of BLICHFELDT'S results to later work
will be discussed there.
IX.2. Sets with V(9') = 2" ,1(9'). We prove here the following result
of MINKOWSKI (1896a).
THEOREM V. Let 9' be an open symmetric n-dimensional convex set
which admits a lattice A with d(A) =2-"V(9'). Then 9' is defined by
m ~ 2" -1 inequalities 1 01 the shape
1~/;iXil < 1. (1 )
1

For each I (1 ~ I ~ m) the planes

'Lflixi= ±1
i
give an (n -i}-dimensional pair of faces of [/, and each such face
contains a point of /\ as an inner point (i.e. for each I there are lattice
points satisfying (2), and (1) for i=t=I).
By Lemma 4 of Chapter IV the set 9' is bounded since 0< V(9')<oo.
By Theorems II and III, every point either belongs to precisely
one set
Y(a): 'i9'+a, aEA,
in which case it is not a boundary point of any Y(b), bEA or it is a
boundary point of at least two Y(a). Hence every boundary point of
Y(o) =19' is also a boundary point of some Y(a), a=t=o: and, by the
boundedness of !7, only a finite number of a can occur in this way.
We note now that, for fixed a, the set of points which are on the
boundary of both Y(o) and Y(a) is convex. For if ~,y are two such
points, the point
t~ +
(1 - t)y (0 < t < 1) (3)
1 In fact there are at most 3" - 3 faces [GROEMER, MZ 79 (1962) 364 - 37 5], and
both .9'and its faces are centrally symmetric. Estimate 3"-1 is easy (HLAWKA,
1949a). Both GROEMER and HLAWKA give generalizations.
 Sets with V (9') = 2" L1 (9') 229

is certainly either a boundary point of .r(a) or belongs to .r(a) by

convexity, and similarly for .r(o). Hence (3) is a boundary point of
both .r(o) and .r(a) by Theorem II.
In particular, if z is common to the boundaries of .r(o) and .r(a)
then so is l a - z by the symmetry of Y. Hence so is also
!a =!z + !(a - z)
a common boundary point.
Denote by
(i~ k~ K) (4)
the points c of 1\ such that the boundary of .r(c) has n linearly inde-
pendent points in common 2 with that of .r(o), and denote by
(i~l~L) (5)
the remaining points b of 1\ such that the boundaries of .r(b) and
.r(o) have points in common. From what has just been shown, the
points common to the boundaries of .r(o) and .r(bz) lie in a linear
subspace of dimension at most n - 2 (not, of course, necessarily, passing
through the origin. In fact, it cannot pass through the origin).
We show now that every boundary point z of .r(o) is also a boundary
point of a .r(ck ). The set of boundary points x of .r(o) in any neigh-
bourhood
(6)
of z is (n -i)-dimensional, and so cannot be exhausted by the at most
(n - 2)-dimensional sets of boundary points in common with the .r(bz).
Hence there must be points in (6) which are common boundary points
of .r(o) and a .r(ck ). Thus z itself is a boundary point with a .r(ck )
as required, since there are only a finite number of C k •
[More precisely, let 9' be F(:r) < 1, where F(:r) is a distance-function. We may
suppose, without loss of generality, that z = (1, 0, ... , 0). If z is common to the
boundary of 9""(0) and .r(bz), the common boundary points of 9""(0)' and .~(bz)
satisfy at least two distinct equations

T1(X1-1) + L,T;X;= 0,
;<?;2
and so at least one equation
L,S;X;= o.
;<?;2

There is an equation of this type for each 1 for which z is on the boundary of .r(b/).
If xi' ... , X" are chosen so as not to satisfy any of these conditions, and arbitrarily

1 -z is on the boundary of 9""(0), by symmetry, and then a-z is on the

boundary of .r(a).
2 That is, the common boundary of .r(o) and .r(e) is a convex (n - i)-dimen-
sional set with centre te,
by what has been already proved.
230 Packings

small. then the point

t{F(I. X 2 • •.•• X II )}-1 (1. x 2 • •••• xn)

is arbitrarily close to % and not on the boundary of any Y(b l ).]

Now we consider the boundary common to .9""(0) and .9""(ck). We
saw already that lCk is one point of the common boundary. Let
(1;£;j;£;n-1) (7)
be 1~ linearly independent points on the common boundary. (They exist
by the definition of the cd Then the points
lCk - Ykj
are also on the common boundary, by symmetry; and hence, by con-
vexi t y 1, so are all poin ts
(8)
with
(8')
Let TTk be the (hyper)plane through lCk and the lCk±Yk;' Clearly any
plane other than TTk through lCk contains points of .9""(0); and so TTk
must be the only tac-plane to .9""(0) at Jck. The equation of TTk may he
written in the shape
(9)

since TTk cannot pass through the inner point 0 of .9""(0). The correspond-
ing tac-plane - TTk through - lCk is obtained by changing the sign of
the fk; in (9). Hence every point of the open set .9""(0) satisfies the
inequalities
IIfk,x;1 <i· I
(10)

Further, every point y, which does not belong to .9""(0) is of the

shape Y =tyo, where t~ 1 and Yo is a boundary point. We saw already
that every boundary point of .9""(0) is also a boundary point to some
.9""(±ck) and so satisfies
± Ifk;x; = l
for this k. Hence Yo. and a fortiori y; cannot satisfy (10). Thus .9""(0)
is precisely the set of ;£ which satisfy (10). Since .9"=2.9""(0), the
corresponding equations for .9" are (1).
1 The point (8) is
to (tck) + ~ It;1 (lCd-Ykj).
1:;;;:;;,,-1
where the ± prefixed to Yki is the sign of tl • and
to+lt1 1+"'+I/''-11 = 1.
 VORONOi's results 231

Some of the inequalities (10) may be identical, since it is quite

possible that the pairs of tac-planes ±TT" may be the same for distinct k.
We may suppose that (10) for 1~k~m gives a complete set of distinct
inequalities, where m~K. We saw that there is a unique tac-plane at
c", and so, since the planes ± TTl (1 ~ l ~ m, l =F k) are certainly tac-
planes and are distinct from TT k , they cannot pass through c". Hence
x =c" satisfies
(1~l~m, l=Fk), (11)
and
(12)

To complete the proof of the theorem, it remains to show that

m~ 2"-1. As in the proof of Theorem IX of Chapter V, it is enough
to show that the points -i(e,,-e,) are not in A for 1 ~k<r~m. But
from what has just been proved, the point ·He" - e,) certainly satisfies
I 1<
~ II i
1
1, for 1 ~ l ~ m, there being strict ineqmility for l = k, r because
then (11) holds for x=·~e,., te" respectively. Hence -~-(ek-e,)E9', so
cannot be in A, since A is 9'-admissible by hypothesis.
IX.2.2. When n = 2, it is possible to specify completely the convex
symmetric sets 9' with ..1(9') =!V(9').
THEOREM VI. A necessary and sulficient condition that the latt1"ce A
be admissible lor the convex open symmetric 2-dimensional set 9' with

V(9') = 4d(A)
is that either
(i) 9' is a parallelogram and 1\ is generated by a mid-point 01 one
side and a point on one 01 the other pair 01 sides or
(ii) 9' is a hexagon and 1\ is the lattice generated by the mid-points
01 any two non-opposite sides. Then A contains the mid-points 01 all the
sides.
That 9' is a parallelogram or hexagon follows from Theorem V,
since 2"- 1 = 3 for n = 2. The lattices A are critical by MINKOWSKI'S
convex body theorem. The critical lattices of parallelograms and
hexagons have already been determined in Lemma 7 of Chapter III and
Lemma 13 of Chapter V respectively.
IX.3. VORONOi's results. We already saw in § 1 that if g(x) is a
positive definite quadratic form and A a lattice, then the set of points
such that
~(x) < inf g(x + a)
"Ell
*0
232 Packings

form a convex symmetric body f/' of volume 2n d (1\). We shall show

that when n = 2 every symmetric convex hexagon j(> and its unique
critical lattice can be related in this way by a suitable quadratic form
g(~). On the other hand. if f/' is a parallelogram. then 1\ must be the
particular critical lattice generated by the mid-points of the sides.
These results are clearly invariant under homogeneous linear trans-
formation so we may suppose without loss of generality that 1\ = 1\0
is the lattice of points with integral co-ordinates and that

g (Xl. X 2) = a X~ + 2 h Xl X 2 + b X~
is reduced. in the sense that
(1)

If U I • U a are integers not both O. the condi tion

(2)
is

Since (-u1 • -u2 ) occurs as well as (ul • U 2). we thus have the infinitely
many conditions
(3)
where
(4)
In particular.
21X1 1<a )
21X21 <b (5)
2IXI+Xal<a+2h+b=c (say).
where
O<a;£b;£c;£a+b. (1')
The set j(> defined by (5) is a proper hexagon unless h = O. when it
degenerates into a parallelogram. The area V(j(» of j(> is reasily com-
puted from (4) and (5) to be
V(j(» = 4 = 4d (Ao) .
But f/' is a subset of j(> and V(f/') =4. by Theorem II. Hence f/' =j(>.
since both are open. This implies that the infinitely many inequalities
(3) all follow from (5). which the reader may verify directly with little
trouble.
Further. every non-degenerate convex symmetric hexagon j(> with
its critical lattice may be generated in this way. as we now show. The
 VORONOl's results 233

hexagon is given by three inequalities

(1=1,2,3). (6)
where

and
'iJ! =lliXl +l2i x 2
is the scalar product. The three 2-dimensional vectors are linearly
dependent and, by multiplying them by suitable factors, we may suppose
'i
without loss of generality that

'1 +'2 + fa = 0,
and, on re-indexing, that
kl -;;;' k 2 -;;;' ka · (7)
On taking Xi = 'iJ! (i = 1,2), the inequalities (6) become

(8)

Further,

since the hexagon.Tf is not degeneratic, by hypothesis. We may identify

(8) and (5) by putting
2 kl = a, 2 k2 = b, 2 ka = c = a + 2 h + b,

though of course the Xl' x 2 in (4) are not necessarily to be identified

with the Xl' x 2 in (6). Let xi, x~ be defined in terms of Xl' X 2 by the
analogue

of (4). On comparing with the earlier part of this section, we see that
the unique critical lattice of .Tf must be given by integral values of
xi, x~. We may thus suppose, without loss of generality, that (xi, x~)
was in fact the original co-ordinate system (Xl, X2), and then we have
the situation already discussed.
IX.3.2. From the results of § 3.1 we deduce the so-called "hexagon-
lemma" of DIRICHLET 1 which illustrates the connection between homo-
geneous and inhomogeneous problems that will be discussed in more
detail in Chapter XI.
I For an alternative derivation of the lemma and a partial generalization to
n dimensions, see MORDELL (1956a).
234 Packings

THEOREM VII. Let

g (Xl' X 2) = a xi + 2 h Xl X 2 + b x~ (1 )
be a quadratic lorm, reduced in the sense that
0~-2h~a~b. (2)

Then to every real point Xo = (Xl 0' x 20 ) there is a point u = (ul , u 2 ) with
integer co-ordinates, such that
4(ab-h2)g(xo+u)~abc, c=a+2h+b. (3)
The sign 01 equality is required when and only when
2(ab - h2) (xo + v) = ± {bra + h), - a(b + h)}, (4)
where v has integral co-ordinates.
For by the results of § 3.1 and by Theorem II there is certainly a
point xo+u with integral u in the closed hexagon

where
(5)
But the positive definite quadratic form g (x) can reach its maximum
in £ only at the vertices l of £. It is now readily verified that the
vertices are of the shape (4) and that the value of g (x) at all the vertices
is given by the right-hand side of (3). The calculations are facilitated
by the identity
g(X2' - Xl) = (ab - h2 ) g(x),
where Xl' X 2 are given by (5).
Finally, the ~ in (3) cannot be replaced by < if Xo is any vertex
of £, since
g(x) = inf g(x + u)
"EA,

for the points x of £. This last remark also shows that it was sufficient
to compute g(x) at anyone vertex a'l (say) since, from the nature of a
critical lattice, all the other vertices are of the shape ±xI + w, where W
has integral co-ordinates.
IX.3.3. Theorem VII gives yet another proof of the result that a
definite ternary quadratic form I(x) represents an number a~ (2D)! for
integral values of the variahles not all 0, where D is the determinant
of I(x) (§ 3.4 of Chapter II). We may suppose, without loss of generality,
1 Perhaps the easiest way to see this is to make a homogeneous linear trans-
formation Y='tz so that g(z) = iyi 2 , when it is obvious.
 Preparatory lemmas 235

that I(~) is reduced in MINKOWSKI'S sense (d. Chapter II, § 2.1). We have
I(~) =ax~ + bx= + c x: + 2hx1 X 2 + 2g xlxa + 21 x 2 xa }
= a (Xl + (xXa)2 + 2h(Xl + (Xxa}lX2 + PXa) + b(X2 +PXa)2 + y x: (1)
for some (x, p, y. We may suppose that h~ O. Then
(2)
and
(3)
for all integers u1 , u 2 , by the condition of the reduction. But now, by
Theorem VII, we may choose u 1 , u 2 so that
I (u1 , u2 , 1)::;;
-
ab(a + 2h+ b)
4 (a b _ h2)
+ y. (4)
Hence from (1), (3), (4) we have
4D = 4(ab - h2)y ~ 4b(ab - hI) - ab(a + 2h + b)

Now
12h + }al ~ -i a ,
by (2); and so
4D ~ 3ab 2 - a2 b ~ 2ab 2 ~ 2aa,
by a further application of (2). This is the required result. Further,
using the knowledge of the cases of equality in Theorem VII, it is easily
verified that 2D=a3 can occur only for forms equivalent to multiples
of the critical form
Xl2 + X2 + Xa -
2 2 Xl Xa - Xa Xa - Xa Xl'
IX.4. Preparatory lemmas. In §§ 5, 6 we shall need three lemmas,
each of independent interest, which it is convenient to prove first. We
use the word polygon to mean indifferently a 2-dimensional set bounded
by a finite number of line-segments or the boundary of such a set.
Which is meant will be clear from the context. We shall say that a
convex polygon is circumscribed to a convex set :Yt" if it contains :Yt"
and if every side of the polygon is a tac-line l of :Yt". The first lemma is
an analogue of Theorem XI of Chapter V due to REINHARDT (1934a),
and found independently by MAHLER (1947c).
LEMMA 1. Let:Yt" be a convex symmetric open 2-dimensional set. Then

LI(Y') = i inf V(Jf'), (1 )

where Jf' runs through all symmetric circumscr£bed hexagons at,d V(Jf')
is the area 01 Jf'.
1 We speak of a tac-line in 2-dimensions instead of a tac-plane.
236 Packings

Let yt be any circumscribed hexagon and I\(yt) the critical lattice

of yt; so that
d {I\(yt)} = t V(yt)
by Lemma 13 of Chapter V. But I\(yt) is certainly admissible for f,
and so the left-hand side of (1) is at most equal to the right.
When f is a parallelogram, the lemma is trivial, so we suppose f
is not a parallelogram. Let M be a critical lattice for f so that, by
Theorem XI of Chapter V, it has precisely 6 points ±p, ±q, ±r on
the boundary of f, where p, q is a basis and

p+q+r=o.
Let yto be the hexagon formed by tac-lines at ±p, ±q, ±r to f,
taking the corresponding tac-line - TI at p to the tac-Iine TI taken at p,
if that is not unique, etc. Then yto is a symmetric hexagon circum-
scribed to f. The lattice M is admissible for yto by Theorem XI of
Chapter V, and so
L1(f) = d(M) ~ L1(yto) = t V(ytoL
by Lemma 13 of Chapter V. This concludes the proof of Lemma 1.
IX.4.2. The following lemma due to DOWKER (1944a) relates the
areas of circumscribed polygons to a convex set f, which need not be
symmetric. We sketch the proof, for which see also FEJES TOTH (1953 a).
LEMMA 2. Suppose that there exists a circumscribed (n +
1)-gon 91',,+1
and a circumscribed (n -1)-gon &'''-1 to a convex set f. Then there exists
a circumscribed m-gon with m ~ n and area

If ~, a 2 , tla are three points on the boundary of f then in this

proof we mean by

that~, a2 , tla occur in that order on traversing the boundary of fin,

say, a counter-clockwise direction.
°
Let the sides of &',,-1 be the lines 1, ... , a,,-1' By definition, these
are tac-lines to f. Let a j (1~j~n-1) be a point on the boundary
of f at which OJ is a tac-line. If OJ is a tac-line at several points, then
we choose a j once for all. We may suppose without loss of generality
that

Similarly let ~jand b j be defined with respect to &'"H' where 1 ~j~ n+1.
 Preparatory lemmas 237

We distinguish two cases. Suppose, first, that three of the b; occur

between two of the ai' say,
a 1 -< b1 -< b 2 -< ba -< a 2 ,
where the symbol between "t and b 1 means that possibly a 1 = b1 , but
otherwise "t -< b1-< b 2 • Let &'~ have sides aI' ~2' a 2, "', a" and &'~' have
sides~l' ~3' ... , ~"+1' Then
V(&',,+1) + V(&'''_l) ~ V(&,~) + V(&,~'), (1)
as is clear from Fig. 10. Indeed the difference between the two sides
of (1) is the sum of the areas of the two 4-gons whose sides are formed
by aI' ~a, ~1' ~2 and aI' a2'
~a, ~2 respectively. From (1)
P
-;,.=:-- ...,---~-----T----------""fq
,;>,,<-...

we have
az '...,, \
\ 1
1
, \ 1
min{V(&,~), V(&,~')}
IJ~
"
'u \\ 1
1
~.... \ I
"J.v
:;;; HV(&',,-l) + V(&'''+1)}' 'f-.,
, ........ I
1
1

which proves the lemma in ~\ . . . . . . ,j

~w 1
this case. \\
, 1
1

The polygons &'~, &'~' \1

'P1'
may have fewer than n sides, 1
1
since some sides of &'''+1 may ~'
coincide with those of &'''-1' 1

But this possibility is covered

by the enunciation of the Fig. to. From the figure,
lemma. We shall not repeat V(ao'R_') - V(ao'~) = V(pqr) ,
this remark which will apply V(9',;') - V(ao'R+t) = V(uvw)
and clearly
at a later stage in this proof V(pqr):<:V(uvw).
and also to the proof of The point labelled 0, should be labelled 0,
Lemma 3.
If the first case does not happen, then, since there are two more
b's than a's we have, on re-indexing if necessary, that
"t.:S b1 -< b 2 < a 2 < a'_ l -< b.< b'+1 -< a.
for some s. Let &'~, &'~' have sides

aI' ~2' ... , ~., a., ... , a,,-l

and

respectively. Then again

V(&'''+1) + V(&'''-l} ~ V(&,~) + V(&,~'),
the difference being the sum of the areas of the 4-gons a 1 a 2 ~1 ~2 and
a'-la'~'~'+1' see Fig.H.
238 Packings

COROLLARY 1. Let U(n) denote the infimum 0/ the areas 0/ circum-

scribed m-gons with m'5, n. Then
U(n) '5, U(n -1) (2)
and
2 U(n) '5, U(n - 1) + U(n + 1). (3)
The first inequality is a trivial consequence of the definition, the
second follows at once from Lemma 2.
It is convenient to extend the definition of U(n) to non-integral
value of the argument. For
t~3 put

U(t) = (i-I) U(n) + l U(n+1),

if
t=n+l, O'5,l'5,1.

COROLLARY 2. Let /11' ... , /1R

be numbers such that

/1,~0 (1'5,r'5,R), L,u,=1.

Then
U(L/1,t r) '5, L/1, U(tr) , (4)
, ,
Fig. 11. The sum of the areas of the shaded regions is where t, (1 ~ r ~ R) are any num-
V(9'n+ ,) + V(~_,) - V(9'':) - V(9'~') b'h
ers w1f t,~ 3.
The inequality follows at once from Corollary 1 if R = 2 and then
follows easily for general R by induction.
By a similar argument to that used for Lemma 2 DOWKER (i944a)
proved
LEMMA 3. Suppose that f is symmetric as well as being convex. Let
.9'2n be an 2n-gon ci,cumscribed to f. Then there is a symmetric 2m-gon
with m'5, n, also circumscribed to f ot area at most V (.9'2 n).
Let the sides

of .9'2n be tac-lines at

where

Let
(5)
where the bar denotes the image in the origin. Then, by symmetry,
the ~i are the sides of the circumscribed polygon .9'2n' which is the
 Preparatory lemmas 239

image of &2n in the origin. By the convexity and symmetry we have

ai+1 < u j < u j +1 < aj

for every j.
If .9'2n is not already symmetric, we may suppose without loss of
generality that Un=F b n and, by changing the orientation of the indexing
if need be, that

Then

by (5). There is thus a greatest j in w~j<2n such that

b<a<b.<a
J= J J= J'
and for this j clearly
Uj < b j < b j +1.:s: u j +1'
It is not excluded that b i +2 also lies between u j and UjI.1' Without
loss of generality j = n; and then

by (5).
Let &;n, .9';~ have sides

a1, "', an, /3"+1' ... , /32n and /31' ... , /3n, a,,+1' ... , a 2n ,

so .9';n and &;~ are symmetric, by (5). Precisely as in the second case
of the proof of Lemma 2 we have

V(.9';,,) + V(.9';~) ;£ V(.9'2") + V(.9'2") = 2 V(.9'2");

and so either .9';" or .9';~ satisfies the requirements of the lemma.
COROLLARY. For convex symmetric f ,
,1(f) =! V(6) ,
where V(6) is the infimum of the areas of circumscribed m-gons with m;£ 6.
This follows at once from Lemma 1 and Lemma 3.
IX.4.3. We shall also need EULER'S formula for convex polyhedra
in a slightly unusual form (d. FEJES TOTH 1953a). Let v" (1 ~n;£N)
be points in the plane (vertices). Let As (1 ~ s;£ S) be curves joining
one vertex to another vertex or, possibly coming back to the same
vertex (the edges). The reader may think of the As as line-segments or
composed of a finite number of line-segments. We suppose that no
point of As except its ends is a Vn and that no two As cross. Finally we
suppose that it is possible to get from anyone vertex to any other along
240 Packings

the As' Then the whole plane is dissected by the As into a number rp
of connected pieces (the "faces") one of which contains all points
outside a large circle 1a:1
=R. Then EULER'S formula is
LEMMA 4.
rp+N=S+2.
This may be readily verified by induction on S.
IX.S. FEJES TOTH's theorem. In this section we prove a result
due to FEJES TOTH (1950a), see also FEJES TOTH (1953a). He proves
something more general and also gives interesting related results but
we give here only what is needed to treat the lattice constants of
cylinders.
THEOREM VIII. Let.Yf be a convex open polygon with at most 6 sides.
Let .:f be any convex open set and suppose that the sets
~ = .:f + a:, (1 ~ r ~ R)
are packed it~ .?It', i.c. the ~ are subsets 01 .?IF and no two have points in
common. Then
R U(6) ~ V(.Yf) ,
where U(6) is the lower bound 01 the areas 01 m-gons circumscribed to .:f
with m~6.
The notation U(6) is in conformity with that of Lemma 2, Corollary.
FEJES TOTH'S own version of his proof is very compact, and we have
found it desirable to expand it.
IX.S.2. The stages in the proof of Theorem VIII are enunciated for
convenience as propositions.
PROPOSITION 11. Let.Yf be a convex open 2-dimensional polygon and
let.Yt", (1 ~r~R) be open convex sets packed in.Yf. Then there are open
convex polygons fl, (1 ~ r~ R) such that fl, contains .Yt", and
( i) the fl, are packed in .Yf,
(ii) il 0' is a side 01 fl, then either,
(iiI) 0' is part 01 the boundary ol.Yf,
or
(ii2) there is a subsegment 0" 01 0' containing more than a single point
which is part 01 the boundary 01 a fls' (s=4=r), and
(iii) il 0' is a side 01 .Yf then some subsegment 0" 01 .Yf consisting 01
more than a single point is part 01 the boundary 01 some fl,.
Note that the .Yt", are not required to be similar to each other. We
shall give two proofs of proposition 1. The first is by transfinite induc-
1 Mr. H. L. DAVIES has pointed out that this Proposition is false as it stands by
giving a counter example. The proof of Theorem VIII can, however, be salvaged.
 FEJES T6TH'S theorem 241

tion (ZORN'S Lemma). It involves the minimum of geometric argument,

but is non-constructive. The second, which will only be sketched, gives
a process for constructing the !!2, in a finite number of steps.
If {f'} and {f"} arc two packings of R open convex sets in ,Yf,
we write
{f'} < {fll}

if f;' contains f; for 1 ~r~R, not necessarily strictly. We denote the

set of all such packings by n
and verify three statements about the
symbol <.
(I) If {f'} < {fll} and {f"} < {f'} then {f'} = {fll}, in the
sense that the sets f,' and f," are identical for 1 ~ r~ R. This is
trivial.
(II) If {f'} < {f"} and {fll} < {f'"}, then {f'} < {f"'}. This
is again trivial.
n
(III) Suppose that ii is any subset of the set of packings such that
if {f'} and {fll} are in n
then either {f'} < {fll} or {~"} < {f'}.
Condition (III) states that then there is some packing {f} in n (not
necessarily in ii), such that {f'} < {%} for all {f'} in ii.
To prove (III) we take for:%, the union of :r, for all {f} in ll.
We must verify that {f} is a packing of convex open sets, and do this
for the properties in turn:
First, :%, is open. For if Zo is a point of .i", then it is a point of f,'
for some packing {f'} of ii. Since f ; is open, a neighbourhood of Zo
is in f;, and hence also in :%" as required.
-
Secondly, f, is convex. For let z}, z. be any points of 1;" say, -
~l Ef;, Z2E f;', where {f'}, {f"} are packings of ii. By the hypotheses
of (III) we may suppose, by interchanging Zl and Z2 if necessary, that
{f'} < {f"}. Then ZlEf;(f,". Since z2Ef,", the whole segment

(O~t~1),

is in f;'; and so in :%" as required.

- -
Thirdly, :r, and ~ have no points in common if r=Fs. For suppose
zoE.i"" zoE"f.. Then zoEf;, zoEf;' for some packings {f'}, {fll} inll,
where again without loss of generality {f'} < {f"}. Then ZoEf; (f;',
so Zo is common to f;' and ~", contrary to the hypothesis that {fll}
is a packing. This concludes the verification of (I), (II) and (III).
We say that a packing {fl'} is maximal if

{fl'} < {f'}

Cassels, Geometry 01 Numbers 16
242 Packings

implies {~/l} ={~'}. By ZORN'S Lemma, since (I), (II), (III) are satis-
fied, to any packing {~} there is at least one maximal packing {~/l}
such that
{~} -< {~/l}.
But it is easy to see that in a maximal packing {~/l} the sets~:' must
be polygons fl, which satisfy the conditions (i), (ii) and (iii) of Proposi-
tion 1. Since this will be clear from the constructive proof which we
give later, we do not give the detailed argument here. This concludes
the first proof of Proposition 1.
We now sketch a second, constructive, proof of Proposition 1. The
fundamental process is this. If ~ is any open convex bounded set and

Bountlory of JI!

Fig. 12. r,' consists of Jr, together with the shaded region

P is any point not in~, then the open convex cover of ~ and p is the
least convex set which contains ~ and has p as a boundary point:
that is, ~ is the set of
tp+(1-t)q, qE~,

If P is on the boundary of ~, then the open convex cover of q and ~

is just~. Otherwise the convex cover has as boundary the two tac-
lines from p to ~ together with a portion of the boundary of ~.
If now ~, ... , ~R are the sets of Proposition 1, we form the'poly-
gons fl, by successively taking the convex covers of the sets .;r, and
suitably chosen points. Let a be any point on the boundary of ~
and a a tac-line at a. Consider points q on a along one direction, say,
to the right of a (see Fig. 12). If q2 is to the right of ql' then the open
convex cover of q2 and ~ contains that of ql and ~. For some q
to the right of a on a it is possible that the open convex cover of ~
and q overlaps some other body Jt; of the original packing. Since the
.;r, are open, there is then a p fartherst to the right along a such that
 FEJES T6TH's theorem 243

the open convex cover of p and Jt';. contains no points of any f, (r=l= 1).
It is possible that p = a. We then get a new packing {f'} on replacing
Jt';. by the portion of open convex cover of Jt';. and p which is in 1 Jf.
If the open convex cover of p and Jt';. does not meet any f, (r +-1)
for all q to the right of a, then f{ is to be the set of points in :K which
are in the convex cover of Jt';. and any point q to the right of a on ex.
Similarly one may consider points to the left of a alongex~
We may repeat the process on the new sets {f'} and will indicate
how after a finite number of steps. it must come to an end with polygons
fl, have the properties (i), (ii), (iii) of Proposition 1. We denote the
sets at the i-th stage by {fi}, so {fi- 1} < {fi}. Suppose first that
there is a pair of indices r, s such that f/ and f,1 have a boundary
point a in common. Then f,/+1, f,1+1 are obtained from f/, f,1 by
taking ex to be a common tac-line (Chapter IV, Lemma 6) to f/, f,1
at a and by applying the above process both to f/ and f,1 and both
to right and left along ex. Once this has been done for a pair of indices
r, s at the i-th stage we do not do it again for the same pair of indices at
a later stage. If there is no pair r, s of indices for which :f/, f,1 have
a common boundary point and which have not already been treated,
then there may be a body f,/ with a boundary point a on the boundary
of Jf. If so, we take ex to be the side of Jf on which a lies (both sides
in tum if a is a vertex of Jf) and apply the process. Again, once this
has been done for f,/ and a side of Jf we do not do it again for the
same r and the same side of Jf. Neither of the first two steps may be
allowable. Suppose that one of the f,/ is not a polygon. Then a is
taken to be any point on the boundary of f,! which is not in a line-
segment forming part of the boundary of f,s nor on the boundary of
.Jt1 (s =l= r). Finally, if all the :7("/ are polygons and the first two stages
are impossible, then a is taken to be any vertex of a f,! at which at
least one of the two sides is not also a tac-line to some f,1 (s=l=r).
It is clear that the steps outlined above will come to an end. And
the final set of .Yt[is clearly a set of polygons fl, having the properties
(i), (ii). (iii) of the enunciation.
IX.S.3. The next stage is an ,application of EULER'S formula (Lem-
ma 4) to the configuration of Proposition 1-
PROPOSITION 2. Let fl, 01 Proposition 1 have q, sides (1 ~r~R).
Then 2

In the application of EULER'S formula, the faces will be the polygons

fl, together with flo, the set of points not in or on the boundary of Jf.
1 The reader is reminded that .1f' is the set in which the :r, are packed.
2 The proof assumes facitly that every vertex of .1f' is a vertex of a ~,.
16*
244 Packings

Lemma 4 is not immediately applicable, since not every point is in or

on the boundary of a ~,. The set of points which do not enjoy this
property is clearly open and so consists of a finite number ~, ... , 9'L
of connected open sets. By (ii) and (iii) of Proposition 1, anyone of
these sets, say, ~ cannot contain the whole of a side a of a ~,. We
now apply Lemma 4 where the "vertices" are of the following kinds
(at) the sets ~ (1 ~l~L),
(fJ) points not on the boundary of an ~ but on the boundary of
at least three ~, (O~r~R),
(r) vertices of fe.
The "edges", for the purpose of Lemma 4, are the segments of the
sides of the ~, joining the "vertices". Then every side of ~, gives rise
q;
to at least 1 but possibly more "edges". Let be the number of "~dges"
surrounding ~" so
(1)
Since every "edge" belongs to precisely two ~, (O~r~R), the number
of "edges" is
S =1 L q;. (2)
O;;;;,;;;;R

Let fe have precisely h sides, so

h~6. (3)
Every vertex of type (at) or (P) above belongs to at least three ~,
(O~r~R) and there are at most h vertices of type (r). Vertices of
type (r) are on the boundary of ~o and at least one ~, (r=l=O). Hence
the total number of "vertices" N satisfies
(4)

Finally, the number of faces rp is

rp=R+1. (5)
From (i), (3), (4) and EULER'S
IJi+N=S+2
(Lemma 4), we get
L q;~6R-6+2h.
O;:iO,;:iOR

But clearly q~~qo=h, by (i), and so, by (i), (3),

L q,~ L q;~6R.
l;:iO,;;;;R l:i!,:iOR

This concludes the proof of Proposition 2.

 Cylinders 245

IX.S.4. The proof of Theorem VIII is now comparatively rapid.

Let U(t) and fl" q, have the meanings they had in Lemma 2, Corol-
laries 1, 2 and Propositions 1, 2. Clearly
V(fl,) ~ U(q,)
and so
V(Jt") ~ L V(fl,) ~ L U(q,),
since the fl, are packed in Jt".
Hence by Corollaries 1, 2 to Lemma 2 and by Proposition 2 we have
R-l V(Jt") ~ L R-l U(q,) ~ U {R-l L q,} ~ U(6).
l:;;,;:>R l:;;,:;;R

This is just the assertion of Theorem VIII, and so concludes the

proof.
IX.6. Cylinders. We now make the application of Theorem VIII
to the lattice constants of cylinders adumbrated in § 1.5.
THEOREM IX. Let.:f be a convex symmetric 2-dimensional star-body
and ~ the set 01 points
f(/: (Xl' X2 , Xa) (Xl' X 2) E.:f, IXal < 1.
Then
L1(~) = .1 (.:f) .
We may suppose without loss of generality that .:f, and so ~, is
open since the presence or absence of boundary points does not affect
the value of the lattice constants L1(~), L1(.:f). It was shown already
that

whether or not .)(' is convex, so it remains only to show that

d (1\) ~ .1 (.:f) (1 )
for any ~-admissible lattice J\.
We prove (1) by computing in two ways the number N =N(X) of
points of 1\ in a large cube
IXil <X (1~i~3)·
In the first place we have the trivial estimate
(2)
as X-+oo.
By Theorem III, since 1\ is ~-admissible, it gives a packing of l~.
Let C be the set of N cylinders
246 Packings

where
aEA, maxlail <X. (4)
I

These cylinders are all contained in the cube

m<:Lxlxil <X +R, (5)
I

iftre is contained in 1:1:1 <R. We consider only the packing of the

cylinders C in (5).
For Iyl <X +R, let L(y) be the number of cylinders of C which
meet the plane X3 = y, that is the number of aE A satisfying (4) for which

la3-YI<~-·
These L(y) cylinders give rise to a packing in the square
1xii < X +R (j = 1, 2)
of L(y) sets similar and similarly situated to tf. Hence
L(y) U'(6) < 4(X + R)2 (6)
by Theorem VIII, where U'(6) is infimum of the areas of circumscribed
m-gons to t f with m ~ 6.
But clearly
X+R
J L(y)dy =N
-X-R
from the definition of L(y). Hence
U'(6) N < 8(X + R)3, (7)
by (6).
Since Rand U'(6) are independent of X, the comparisGn of (1) and
(7) as X ~ 00 gives
d (A) ~ U'(6).
But
U'(6) = 4L1(tf) = L1(f)
by Lemma 3, Corollary. This completes the proof of (1), and so of the
theorem.
IX.7. Packing of spheres. The unit sphere
~,,: 1:1:1 <1
in n dimensions has volume
v.. = V(~,,) (1)
 Packing of spheres 247

where r (1 + ;) is the usual gamma function. In this section we estimate

the lattice constant
1',. = ,1(2),,); (2)
and are primarily interested in the behaviour of r" when n is large.
In the literature it is customary to use y" defined as the lower bound
of the constants y~ such that every positive definite quadratic form
L I,; x, x; in n variables represents a number ~y~ Idet(/i;W'" ("HER-
MITE'S Constant"). By the arguments of Chapter I, § 3 we have

(3)
We shall need to know the asymptotic behaviour of the volume v".
From STIRLING'S formula l we have
lim n
"-+00
V:'" = 2ne, (4)
where

From MINKOWSKI'S convex body Theorem and the Minkowski-

Hlawka Theorem we have
(5)
where C(n) is RIEMANN'S function. These inequalities lead by (3) and
(4) to
lim sup ny;I~ 2n e (6)
,,-+'X)

and
lim inf ny;l~ ine. (7)
"-+00

Of course the factor 2C (n) in (5) has no effect in (6) and might as well
have been replaced by 1. Indeed none of the improvements of the
Minkowski-Hlawka Theorem discussed in Chapter VI affect the constant
on the right-hand side of (6). On the other hand, BLICHFELDT (1929a)
has improved (7) to
lim inf ny;I~ ne, (8)
which appears to be the best asymptotic form to date 2. The argument
is a purely packing one and makes no use of the fact that only lattice
packings are relevant to (8). BLICHFELDT'S results have been improved
by RANKIN (1947a), and yet further, by a more perspicuous argument,
1 See any reputable text book on analysis, for example WHITTAKER and
WATSON (1902a) Chapter XII.
2 The improvement in (8) announced by CHABAUTY (1952a) is not correct,
see the review by RANKIN in Maths. Reviews 14, 541.
248 Packings

by ROGERS (1958c). Their methods yield considerable improvements

for small values of n, but do not improve the constant in (8).
BUCHFELDT'S methods may be applied to other sets than spheres,
see RANKIN (1949a, b, c) and 1955a and the literature cited there.
There is a detailed discussion of the non-lattice packings of 3-dimen-
sional spheres in FEJES T6TH (1953a), see also S. MELMORE (1947a).
I have been helped by my recollection of a seminar talk by Professor
RANKIN in Cambridge in the late 1940s on BUCHFELDT'S method.
IX.7.2. We observe first that BUCHFELDT'S Theorem I of Chapter III
may be generalized to pac kings and indeed takes a quite simple shape.
Let /7 be any bounded n-dimensional set and suppose that the sets
(1)
are packed in some set Y. Then trivially
V(Y) ~ R V(/7). (2)

Suppose now that there is some function rp (~) of the vector variable
Z such that
(i) rp(z) =0 if Izl ~ e for some e
and
for all z,

whenever (1) is a packing of /7.

Let Y(e) be the set of points at a distance ~ e from Y, including
the points of Y itself. Then, in the first place,
f 'P(z) dz ~ V{Y(e)} (dz = dx1 ••• dx,.) (3)
9"(C/)

by (ii) and, on the other hand, by (i)

f 'P(z) dz = L f rp(z - zr) dz = R f rp(z) dz = R V(rp) (say), (4)

9"(C/) r

since all points with rp(z-zr) =1=0 lie in Y(e). The comparison of (3)
and (4) gives
(5)
Of course the characteristic function of /7, which is 1 on /7 and 0
elsewhere, has the properties (i) and (ii). With this as the function rp,
the inequality (5) is rather weaker than (2). because we have replaced
V(Y) by V{Y(e)}: though of course this can be avoided by a refinement
of the argument. BLlCHFELDT observed that there are sometimes
 Packing of spheres 249

functions rp which give a better estimate than the characteristic

function.
For example, if [/' =P)" is the sphere of unit radius, the necessary
and sufficient condition that open spheres of radius 1 and centres :Ill' :Il2
shall not overlap is I:Ill - :Il21;;;:; 2. The following lemma may be regarded
heuristically as showing that, in a packing of spheres, a point may be
on the boundary of two spheres but cannot be too near the boundaries
of more than two spheres simultaneously.
LEMMA 4. Put
rp(:Il) = max{o, 1 - 1:Il12}. t (6)
Suppose that :Il, (1 ~ r ~ R) are any points such that
(7)
Then
L rp (:Il - :Il,) ~ 1 (8)
l::i!,::i!R
lor all points :Il.
We may clearly suppose without loss of generality that
0< rp(:Il- :Il,) = 1 - tl:ll- :Il,12
for 1 ~r~R.
If Yl> ... , YR and yare any real numbers, we have
R L (y - y,)2 =L (y,- Ys)2+ (Ry - Ly,)2;;;:; L (y,- Ys)2.
, ,<s , ,<s

Hence, on applying this to the individual co-ordinates, since 1:Il1 2 =

x~ + ... + x~, we have

R ~ 1:Il-:Il,12~~I:Il,-:IlsI2~2R(R-1),
l::i!,::i!R ,<s
by (2). But this is just the same as (8).
From this we have almost immediately
THEOREM X. Let:ll, (1 ~r~R) be points in the n-dimensional sphere
(9)
and let

Then
R ~ 2-,,/2 (1 + ;) (X + 26)". (10)

If rp(:Il) is as in Lemma 4, we have

V(rp) = Jrp(:Il) d:ll = J(1- ~ 1:Il1 2) d:ll = (1 + ;rlv,., 2"/2

Izl'<2
250 Packings

where V" is the volume of the unit sphere. The result now follows from
(5), since .9'"(e) is now the sphere

Izl <X +2~,

which has volume (X + 2i)"v".
COROLLARY 1. The lattice constant r.. and volume v" of the unit sphere
Izl < 1 satisfy
1',.:?; r"j2 (1 + ;r1v".
If I\. is admissible for Iz 1< 1, then the points z, of 21\. satisfy the
conditions of the theorem. The number of points of 21\. in

.9'": Izl <X

is
{d (21\.)}-1 V(.9'") + 0 (X"-l) = r"{d (I\.)}-l X" v,. + 0 (X.. - 1).
On comparing with the theorem and letting X _ 00, we obtain the
required inequality.
COROLLARY 2.
lim inf ny.;l:?; n e,
where y: =r..-2 •

This follows from Corollary 1 and (4) of § 7.1.

IX.S. The product of n linear forms. Denote by .;v,. the n-dimen-
sional set

and let

The set .;v,. plays an important part in algebraic number theory (see
Chapter X), but the only precise values of v" known are
v: = 51, v~ =7
given by Chapter II, Theorem IV and Chapter X, Theorem V respec-
tively. Here we shall be concerned with estimates for v" when n is large
rather as in § 7. For information about what is known for n =4 or 5
see Chapter II, § 6.4.
In Chapter III, § 5.3 we already gave MINKOWSKI'S estimate

J(.#.:) ~ n"
, , - n!

which by STIRLING'S Formula, gives

lim infv,,:?; e = 2·71828 ....
,,--+00
 The product of n linear forms 251

BLICHFELDT has given an elegant proof that

lim inf v,. ~ (2n)1 e1 = S ·30653 ... : (1 )
,.-400

and this we obtain in this section.

The estimate (1) is not the best yet found. ROGERS (1950a) has
shown indeed that
lim inf v,. ~ 4n- 1 el = S . 70626 ....
,.-400

His intricate and laborious proof may be regarded as an elaboration of

BLICHFELDT'S.
Since.#;. has infinite volume, there is no estimate of L1 (.#;.) above
from the Minkowski-Hlawka Theorem. Indeed, work of SCHOLZ (1938a)
on the discriminants of algebraic number fields gives some reason to
suspect that lim sup v,. = 00.
,.-400

In § 8.2 and 8.3 we give two lemmas and then in § 8.4 BLICHFELDT'S
proof of (1).
IX.8.2. The following Lemma of SCHUR (1918a) also occurs in the
theory of the "transfinite diameter" in analysis.
LEMMA 6. Let ~l' ..• ,~,. be real numbers. Then

11. (~. - ~;)2~ -D", (L• ~f)l"'(m-l),

.</
(1 )
where
-Dm={m(m-1)}-lm(m-l).12.22 ..... m"'. (2)
The continuous function II
(~'_~i)2 of the m variables~. attains its
k;
maximum, -D say, on L~~=1, say at ~.=rJ; (1~i~m). Then, by
homogeneity,
(L ~~)-l"'(m-l) II. (~. - ~;)2~-D (3)
• .</
for all ~i' with equality when (~i) = (rJ.). The derivative of the logarithm
of the left-hand side of (3) with respect to each variable must vanish
at the maximum (~j) = (rJi); and so

L:-- 1 - = m(m-1)1]i (1~i~m). (4)

;,*,' 1]i - 1]; 2

since L rJ~ = 1. Let

l(rJ) = II ("I - "I.) is)
be a polynomial in the variable "I. Then (4) is
/"(1]i) m(m -1)1],
2/'(1],) = -- 2 -
(6)
252 Packings

The polynomial

is of degree at most m-1, since the coefficients of r/" vanishes. By

(5) and (6) we have h(17i) =0 (1 ~i~m); and so h(17) vanishes identically:
1"(17) - m(m - 1) 17 /'(17) +m 2 (m - 1) 1(17) = O. (7)

The difterential equation (7) determines 1(17) completely in terms of, say,
1(0) and /,(0). Hence we may determine the symmetric functions ~ 17~
and II (1]i-1]i)2 in terms of 1(0) and /,(0). Since ~17~ =1 and the coef-
ficient of 17 m in I (17) is 1, this determines I (17) completely, and so also
II (1]i _1];)2 = {}. (8)

It is simpler, however, to use a more indirect approach which will now

be described.
The resultant of two polynomials, say,

with highest coefficient 1 is defined to be

R (f{J, 1p) = II
i,;
(rt.i - P;)

= II
i
1p(rt.,)

= (-1)/J n f{J(/J;)
;
= (_1)IJ R(1p, f{J).
If
w(17) = II
l~k;:oK
(17 - Yk)

is a third polynomial with highest coefficient 1, and if

identically for some number ,l and polynomial X (17), then

R(f{J,w) =,ll R(f{J,1p). (10)
by (92),
In particular, if 1(17) is defined by (5), we have

{) ~;i!lJ;:o.~17i -17;)2 = (- 1)lm(m-~;;lJ!,(17i)) (11)

= (_1),m(m-l)mm R(f, 11).
 The product of n linear forms 253

where

has highest coefficient 1. More generally, put

I.('YJ) = (m,:-!k)! pl('YJ);

so that I. ('YJ) has highest coefficient 1. Then on differentiating (7) k

times one readily obtains
(12)
Hence
- (m - k - 1) }",-k-l I I
R(/k,/Hl) = R(lHl' Ik) = { m(m _ 1) R( HI' H2)' (13)

on using the rules of operation (9.) and (10). But I",('YJ) =1 and
1",-I('YJ) ='YJ+y for some number I' (in fact 1'=0); so
(14)

by (9 2). The required value (2) for {} follows now from (11), (13) and (14).
IX.8.l. We also require an estimate of the number {}", occurring in
the last lemma.
LEMMA 7. II
G", = 1 . 22 ••••• m"',
then
lim sup {m- 2 10g G", -llog m};:;;; - 1.
"'-+00
Put
g(x) =xlogx (x>o).
Then
g'(x) = log x + 1
increases with x; and so
g(x + t) + g(x - t) ~ 2g(x) (1)

for any t, since if t> 0 we have

g(x + t) - g(x) = tg'(~I)
g(x) - g(x - t) = tg'(~2)'
where ~2< X<~I' so g'(~2) <g'(~l)·
In particular,
IH i
J g(x) dx = J {g(l + t) + g(l- t)}dt~g(l), (2)
1-6 0
254 Packings

for any integer l. Thus

m+'
10gG,,. = L g(l) ~ J g(x) dx =! (m + !)210g(m +!) - 1(m + !)2 + y,
2;>;t;>;m ~

where y is independent of m. The required estimate now follows at once.

COROLLARY. II .om ={m(m -1)}-i m(m-1 l Gm is the number delined in
Lemma 6, then
lim sup {m- 2 10g{}m + ! log m} ~ - 1.
m->oo

This is immediate. It is not difficult to see that "lim sup" may be

replaced by "lim'~, but we do not need this.
IX.S.4. We can now prove BLICHFELDT'S Theorem on the product
of linear forms discussed in § 8.1.
THEOREM XI. Let v: be the lattice constant 01 the set

Then
lim inf v,. ~ (2n)1 et . (1 )

Let A be a lattice which is admissible for .;v, and let m be an integer

whose value will be settled later.
Consider the sphere
fi}: l:.el < e,
where e is chosen so that
V(fi}) = md(A);
that is
e" v.. = m d (A) , (2)

where v.. is the volume of l:.el < 1. By BLICHFELDT'S Theorem I of

Chapter III, there are m points :.e1 , ••• ,:.em in fi} whose differences
:.e.. -:.ei all lie in A. Put
(1~i~m).
and write
S,,=Lx~ . (1~k~n).
l;>;";>;m
Then
L s" =l:;;i:;;m
1:;;";>;,.
L 1:.e. 12~ me2 ;
and so

by the inequality of the arithmetic and geometric means.

 The product of n linear forms 255

Now let

Then, on the one hand

(4)

by Lemma 6, where {)". is the number defined there. On the other hand,

where

The points ~i-~i belong to A, which is %-admissible; and so

(i=f:i).
Hence
II ~~1. (5)
l;;;k;;;"

On eliminating ~, Sk from (3), (4) and (5) we get

1=
<.0."
"V".
(!!'_ e2)1"".(".-1). (6)
n

Hence, on eliminating e between (2) and (6),

{d(AW'" ~XIX2X3' (7)

where
Xl = m-i{);l,,,.(,,.-l),
X2=n 1 V,t''',
and

Now Xl is independent of nand

lim inf XI~ el , (8)
"'-+00

by Lemma 7 Corollary. Further, Xa is independent of m, and

lim Xa = (2:rte)1 (9)

"-+00
by (4) of § 7.1. Finally,
limXa = 1 (10)
if, say, m=n-+oo.
Since v" is the infimum of {d(AW'" over %-admissible lattices, and
since the product of the right-hand sides of (8), (9) and (10) is the right-
hand side of (1), this proves the theorem, by (7).
256 Automorphs

Chapter X

Automorphs
x.t. Introduction. A homogeneous linear transformation w is said
to be an automorph of a point set .9 if .9 is just the set of points w:I:,
:l:E.9. The automorphs of a set .9 evidently form a group. Many of
the point sets of interest in the geometry of numbers, or which occur
naturally in problems arising in other branches of number-theory, have
a rich group of automorphs which is reflected in the set of .9-admissible
lattices. Already in the work in which he introduced the notion of
limit of a sequence of lattices, MAHLER (1946d, e) laid the foundations
for future work and indicated some fundamental theorems. Since then
much has been done but some challenging and natural questions remain
unanswered.
MAHLER (1946d, e) considers star-bodies with groups of automor-
phisms having special properties which he calls automorphic star-bodies.
In this account we prefer in each case to state the properties of the
group of automorphs which are required to hold.
We shall say that a homogeneous linear transformation w is an
automorph of a lattice A if wA =A, that is if A is precisely the set of
wa, aEA This is really a special case of the definition at the beginning
of the chapter since A is a point set. Since

d(wA) = jdet(w)i d(A) ,

we must have
det(w) = ± 1.
We say that w is an automorph of a function 1(:1:) of the vector :I: if
I(w:l:) =/(:1:),

for all:l:. In particular, w is an automorph of the distance-function

F(:I:) if and only if it is an automorph of the star-body

.9: F(:I:) < 1,

since .9 and F(:I:) determine each other uniquely. Clearly

F(w A) = F(A)
for a lattice A if w is an automorph of the distance-function F(:I:), since
F(A) = inf F(a) ,
oEII
*,0
by defimtion.
 Introduction 257

If Y is any point set and 't a non-singular homogeneous linear

transformation, then
,1('tY) = \det('t)\ ,1(Y).
since a lattice 1\ is admissible for Y if and only if 't 1\ is admissible
for 'tY.
In this chapter we shall make great use of the properties of homo-
geneous linear transformation expounded in Chapter V, § 2. In par-
ticular we write
<p=p+a, ~ =pa
if
<px=px+ax, ~x=p(ax),

respectively for all x.

X.1.2. We first give three theorems which are already in MAHLER
(1946a, b) but not all as formulated here. We give also the proofs:
their brevity shows the power of MAHLER'S techniques, particularly in
the striking Theorem III.
THEOREM I. Let F(x) be a distance-function with an automorph w
such that
det (w) =1= ± 1-
Then F(I\) =0 for all lattices I\.
By taking w-1 instead of w if need be, we may suppose that

\det(w)\ < 1-
If there is a lattice 1\ with F(I\) =1= 0, then there is a critical lattice M
for F(x) < 1, by Theorem VI of Chapter V. But then

F(wM) =F(M) = 1,
and
d(wM) = \det(w)1 diM) <.d(M),
in contradiction to the definition of a critical lattice.
For example, Theorem I shows that
\x~ x2 \ <1
is of infinite type since it has the automorphs xC"~ix, x2--+4x of de-
terminant 2. This was our example of a star-body of infinite type in
§ 5 of Chapter V.
THEOREM II. Let F(x) be a distance-function. Suppose that every
point Xo with F(xo) = 1 is ot the shape
(1 )
Cassels, Grometry of Nurnbf"rs 17
258 Automorphs

where w is an automorph 0/ F, and Co is in a compact set «1'. Then lor

every lattice 1\ with F(I\) = 1 there exists a lattice M with
F(M) = 1, d(/\) = d(M)
having a point in f'{/.
Since F(x) is continuous, the set f'{/' of points cEf'{/ with F(c) = 1 is
compact if '6' is compact. Since Co as defined in (i) has F(co) =F(xo) = 1,
we may suppose without loss of generality that
F{c) =1 (CEf'{/) . (2)
Since F(I\) = 1, there is a sequence of points a,E 1\, not necessarily
distinct, such that
F(a,) ~ 1 : F(a,) -r 1 (r -r 00).
Then b,= {F(a,)}-la r satisfies F(b,) = 1; and so
b, = w,c,
for some automorph w, of F and some c,E«1'. Since f'{/ is compact, we
may suppose, after extracting a subsequence and re-indexing, that
C, -r c' Ef'{/ (r-roo).
Let
1\ = Wr 1\,.
Then, since Idet (wr)i = 1 by Theorem I, we have
F(I\,l = F(I\) = 1 , d (I\r) = d (1\)
and
F(ar)CrEI\,.
By Theorem IV, Corollary of Chapter V, the sequence I\r contains a
convergent subsequence, and so, without loss of generality,
I\,-r M
for some lattice M. Then
d (M) = '-+00
lim d (1\,) = d (1\)
and
F(M) ~ lim supF(I\,)
'-+00
= F(I\) = 1
by Theorem II of Chapter V. Further, M contains
e' = lim F(a,) c"
'-+00
so
F(M) ~ F(c') = 1, (4)

by (2). From (3) and (4) we have F(M) = 1. This concludes the proof.
 Introduction 259

COROLLARY. There is a critical lattice for F(x) < 1 having a point c

in ~ with F(c) = 1-
For if 1\ is critical so is M. This corollary is in contrast with the
example given in § 5.2 of Chapter V of a star-body no critical lattice
of which has points on the boundary. Note that the corollary does
not affirm that every critical lattice of F(x) < 1 has points on F(x) = 1 ;
the author [CASSELS (1948a)] has given a rather artificiaP counter-
example of a body F(x) < 1 satisfying the hypotheses of Theorem II
and having critical lattices with no point on F(x) = 1.
As an example of Theorem II consider the body %: IXl X 2 Xal < 1
with its distance-function IXIX2Xal!. Here ~ may be taken to be the
single point c=(1, 1, 1); since every point XO=(XIO' x 20 ' xao) with
IXloX2oXaol=1 is of the shape
XO=WC,
where w is the automorph

of %. Hence there are critical lattices for %with a point at (1,1,1).

If one is concerned only with the evaluation of .1(%) and not with the
enumeration of all the critical lattices, it is enough to consider critical
lattices with a point at (1,1,1).
THEOREM III. Let the point-set f7 be a subset of the star-body ,,/
with .1(9') < 00. Suppose that for every r there is an automorph w, of ,,/
such that w r f7 contains every point of 9' which is in Ix 1< r. Then
.1(f7) = .1(9').
Clearly
.1(f7) ~ .1(9').
By Theorem I we have det (w r ) = ± 1, and so
,1(f7) = .1(w,f7) ~ .1(~),
where ~ is the set of points of 9' in Ixl < r. But
lim .1(~) = ,1(9')
,-->00

by Theorem V of Chapter V, so .1 (f7) = L1 (9'), as asserted.

Clearly one may formulate theorems similar to but more general
than Theorem III by making use of the full force of Theorems II and V
of Chapter V. The argument used in the proof of Theorem III was
already used in the proof of Theorem XV of Chapter V.
1As Professor ROGERS remarks, it is quite likely that the 3-dimensional body
IXII max (x:' xi) < 1 furnishes a natural example.
17'
260 Automorphs

As an example of Theorem III one may take for f/, 9" respectively
the sets

and
9": IXlxsxal < 1, Ixsl < e, IXal < e,
where e is any fixed positive number. Then the automorphism w, may
be taken to be
XI=r-2e2xI' X 2 =re-I x s , Xa=re-Ixa,
where X = W,:I:. In this example one may deduce that a lattice A
with d(A) <LI(f/) has infinitely many points in f/.. For A must have
a point in 9" for every e>O. If A has no point a=!=o with a2 =aa=0,
this implies that A has infinitely many points in f/; and on the other
hand, if a = (al , 0, 0) is in A, then all the points ma (m = 1, 2, ... ) are
in A, so there are still infinitely many points of A in f/. Indeed the
argument shows that for any e> 0 there are infinitely many points
of A in :r. This sort of argument was already used for Lemma 12 of
Chapter V about the existence of infinitely many points in -1 <Xl x 2<k.
There we could prove rather more since this set was shown to be bound-
edly reducible. In § 7 we shall make a systematic study of when there
are infinitely many points of a lattice in a star-body following DAVEN-
PORT and ROGERS (1950a).
X.l.3. The point sets with a large group of automorphisms with
which we shall be concerned will be mainly constructed simply from an
algebraic form ffJ (:1:). For example ffJ (:I:) may be Xl XI' Xl X2Xa' Xl (X: + X:)
or X~ + X~ - xt and the set f/ may be defined by
IffJ (:1:)1 < 1 (1 )
or
o ~ ffJ(:I:) < 1 {2}
or
0< ffJ(:I:) < 1 (3)
or
- k < cp(:I:) < 1, (4)
where k and 1 are positive numbers. Of course (2) and (3) are not star-
bodies. Apart from sets especially constructed from sets of the type
(1)-{4) to act as counter-examples, other sets with large groups of
automorphisms have proved intractable. For example the lattice
constant of
IXII max(xt x~) < 1
is not known, though it would be of some interest in the theory of
simultaneous approximation and the problem has had considerable
 Introduction 261

attention [see DAVENPORT (1952a) and CASSELS (1955a) and the refer-
ences given there J.
We shall make continual use in this chapter of the results of Chap-
ter I, § 4 about the relationship of lattices to forms.
A particular kind of lattice plays a special rl'>le in connection with
sets of the type (1)-(4), where IP(:I:) is an algebraic form. It is useful
to introduce some new terminology. If IP(a) is an integer for all aEA
we say that IP is integral on A. If, further, IP (a) = 0 for aE A only when
a =0, we will say that IP is non-null on A (the trivial zero at 0 being
disregarded). Finally, if there is some number t=l=O such that tIP is
integral on A we say that IP is proportional to integral on A. Then IP
is integral on IWI'" A, where m is the degree of IP'
In many, if not all, cases where the form IP has infinitely many
automorphs and the critical lattices Ac for one of the sets (1)-(4) are
known, it turns out that IP is proportional to integral on Ac' Indeed
in some cases IP is proportional to integral on every known admissible
lattice, and it is suspected, but not proved, that no other admissible
lattices exist. In other cases, there certainly do exist admissible lattices
on which IP is not proportional to integral, but the critical lattices are
not amongst them.
Before discussing the general· properties of a lattice A on which a
form l IP is proportional to integral and illustrating it with concrete
examples, it is convenient to prove a simple lemma.
LEMMA 1. Let l' > 0 and m> 0 be integers and let
y(u l , ... , u,)

be arbitrarily given numbers lor integers ull in

(1~e~r). (5)

Then there is a uniquely delined polynomial I (u) 01 degree m in the variables

ul, ... , u, such that

I(u) =y(u) (6)

lor all integers u = (Ut, ... , u,) in (5).

This is certainly true when T = 1. For T> 1 we use induction on T.
We may write
I (u) = L u~ gp (~ , ... , U,_l), (7)
0;:;;;1';:;;;'"

where the gl' are polynomials to be determined. For any fixed values
of Ul , "'J U,-lJ the equations (6) determine uniquely the values that
must be taken by gl' (u1J . . . J U,_l) in (5); and then there are uniquely
1 We recollect that the word "form" implies homogeneity.
262 Automorphs

determined polynomials taking these values, since we assume that the

lemma has already been proved with r -1 for r. Alternatively one
could observe that the determinant of the (m + 1)' equations for the
(m + 1)' coefficients in I(u) have determinant
II (v_u)2m=!=O.
O;i;u<v;i;m

COROLLARY. II the y (ul , ... , uri are rational, so are the coellicients in f.
This follows at once from the proof.
Now let rp be a form which is integral on the lattice /\ with basis
bl , ... , b n • Put
I(u) = I(u l , ... , un) = rp (f
U i bi )· (8)

By Lemma 1, Corollary, the coefficients in the form I(u) are rational.

Conversely if the coefficients in I (u) are rational, then rp is proportional
to integral on I\.
We shall now describe in some detail what happens in some special
cases which have been extensively investigated.
Suppose, for example, that

so that I (u) in (8) is any indefinite ternary quadratic form of signature

(2,1) (d. § 4 of Chapter I). No-one has yet been able to construct a
ternary quadratic form which can be shown not to take arbitrarily
small values for integral u, apart from the multiples of forms with
integral coefficients. OPPENHEIM (1953 b, c) has shown I that an in-
definite quadratic form which takes arbitrarily small values of one sign
also takes arbitrarily small values of the other. Such a form then takes
values in every interval, since
I(ru) = r2/(u)
and I(u) may be taken arbitrarily small of either sign.
The situation is much the same when
rp(x) = Xl X 2 Xa'

Then the function I (u) given by (8) is the product of three real linear
forms:
(9)

1 He also shows that if an indefinite quadratic form is not a multiple of a

form with integral coefficients and takes the value 0 then it also takes arbitrarily
small non-zero values for integer values of the variables if the number of variables
is greater than 5.
 Introduction 263

and conversely every product of three linear forms (9) with

gives rise in this way to a lattice A. A classical theorem which we

shall prove in § 4 states that if the coefficients in I (u) are rational and
I(u) may be expressed as the product of three real linear forms and if,
further, I(u) =j=-O for integral u*,o, then
I (u) = t n (fJil UI + /3i2 U 2 + /3i a u a) ,
l&i&a

where /311' /312' /313 are numbers in a totally real cubic field srI and /3i k
is the conjugate of /31k in the conjugate field sri.
On the other hand, there are certainly lattices 1\ which are admis-
sible for

and on which Xl x 2 is not proportional to integral. This follows at once

from the theory of continued fractions: alternatively it is not difficult
to modify the proof of Theorem VIn of Chapter VI.
A rather more interesting case is
(10)
Since

where i2 = - 1, there is a connection with the cubic fields that are not
totally real, similar to that of Xl X 2 Xa with totally real fields: it is clas-
sical, and will be proved in § 4.4 that if Xl (X~ + xi) is proportional to
integral and non-zero on 1\, then 1\ arises from a cubic field. But there
certainly are other admissible lattices for
(11)

Let T be any transformation X

Xl
X2=
= tIl Xl
=TX

t22 X2+ t 2a Xa
I
of the special type

Xa= tS2X2+taaXa,
where

Then there are clearly constants C, c depending on T, such that

00 > C~ I q; (-r:r) I ~ c> 0

Iq;(:r) I -
264 Automorphs

for all z. Hence if A is admissible for Icp (z) I< 1, so is t ~ A for some
number t; and in general cp (z) will not be proportional to integral on
~ A if it is on A.
This does not exhaust the admissible lattices for Xl (X: + X:) < 1. One way to
show this is to use the arithmetic-geometric mean inequality in the shape

Hence any lattice admissible for IXI x2 xal < t is also admissible for IXI (x= + x:lI < I;
for example 2-! M has this property if Xl X 2 xa is integral and non-null on M (i.e.
when M arises from a totally real cubic field); and it is easy to see that Xl (X: + x~)
cannot be proportional to integral on M. [In fact the xl-eo-ordinates of M for
a lattice on which Xl X, xa or Xl (X: + x~) is non-null and proportional to integral
determine the relevant cubic field completely and it cannot be both totally real
and not totally reaL] More generally, one can construct admissible lattices by the
methods of Chapter VI, Theorem VIII, compare CASSELS (1955 b) for a closely
related problem.
It is an interesting problem to decide for any given form cp (z) if
there exist admissible lattices for a set Icp (z) 1< 1 on which cp (z) is not
proportional to integral. CASSELS and SWINNERTON-DYER (1955 a) have
considered the special cases cp (z) = Xl X 2 Xa and x~ + x~ - x~, but they only
transform the problems into another one. For another line of attack,
see ROGERS (1953b). It is reasonable to think that essentially new ideas
will be required even to cope with Xl X 2 Xa or x~ + x~ - x~.
X.l.4. An important part in the theory is played by so-called isola-
tion theorems. Their importance was first apparently recognised by
DAVENPORT and ROGERS (1950a) though there are foreshadowings in
MAHLER (1946e) and indeed in REMAK (1925a). A new type of isolation
theorem is proved and exploited in CASSELS and SWINNERTON-DYER
(1955a).
The phenomenon of isolation takes various forms all of which state,
roughly speaking, that lattices in the neighbourhood of a given lattice
M, with certain exceptions, are much worse behaved than M itself.
Thus one result we shall prove is that if Xl X 2 Xa is integral and non-null
on a lattice M, then to every e> 0 there is a neighbourhood ~ of M
in the sense of § 3.2 of Chapter V, depending on e, such that

inf IXI x 2 x3 1 < e

:rEA
,*,0

for all A E ~ except the A of the shape t M, for a number t. This is a

particularly sweeping result .. Perhaps more typical is the isolation
theorem for xdx~ + x:). This states that if

inf Ixdx~ + x~)1 = 1,

:rEM
:r,*,o
 Introduction 265

and if Xl (X~ + x~) is proportional to integral on M, then there exists

an rJo> 0 and a neighbourhood Q of M, such that
inf Ixdxi
;rEA
+ x~)1 < 1 - rJo
*0
for all /\E ~ except those of the type 't'M, where 't' is of the special type
with T12=T13=T21=Tal=O already discussed in § 1.3. Note that for
Xl x 2 xa , the number e could be chosen at will, whereas for Xl (xi + x~)
both rJo and Q are fixed by the lattice M.
All isolation theorems have the same general type of proof. In the
first place, it is shown that if the form rp (x) is, say, integral or integral
and non-null on a lattice M, then rp (x) and M have a group Q M of
automorphs w in common; that is
rp(wx) = rp(x), wM = M.
For the special forms X I X 2 ' X I X 2 Xa and xl(xi+x~) these automorphs
are given by the theory of units in algebraic number fields, and for
x~ + xi - x~ by the theory of indefinite ternary quadratic forms; but
we shall, in fact, find it easy to handle the group Q M without these
theories and using only MAHLER'S compactness theorem I. A lattice /\
near M, in the sense of MAHLER, is one of the shape
/\ ='t'M,
where 't' is near the identity transformation. Suppose that there is an
aoE M such that rp (a o) takes some interesting value (I.. Then
rp(wao) = rp(ao) = (I., WEQM'

Then /\ contains the point 't'wao' Although I't'ao-aol is small when 't'
is near the identity, it does not follow that I't'wao- waol is uniformly
small for all w, since in general w may be chosen so that wao is arbi-
trarily large. By suitable choice of w in Q M one may then show the
existence of a point 't'wa o in /\ ='t' M having the properties desired in
the problem in question, unless the transformation 't' satisfies certain
conditions. Sometimes one must start not with one point ao, but with
several, aI' ... , aT' so as to eliminate 't' of different kinds. This general
attack will be clearer from the examples in § 5. Isolation theorems may
be used to discuss the existence of infinitely many lattice points in
regions, as will be shown, following DAVENPORT and ROGERS (1950a), in § 7.
X.1.S. Before going on to the main subject matter of the chapter
we shall discuss in § 2 certain special forms and their groups of auto-
morphs. In § 3 we shall then discuss a method of MORDELL which shows
lOne of MINKOWSKI'S first applications of the geometry of numbers was in
fact to the theory of units in algebraic fields.
266 Automorphs

how a bound for the lattice-constant of an n-dimensional body may

be obtained from a bound for that of a related (n -i)-dimensional body.
When the original n-dimensional body is of a special type having many
automorphs, MORDELL showed the argument can be carried a stage fur-
ther. In particular it gives the lattice constants of the 3-dimensional
sets !Xl X 2 X 3 ! < 1 and !Xl (x~ + x~)! < 1. In § 8 we discuss briefly the
relevance of continued fractions to forms and bodies with automorphs
and the possibility of generalisation.
X.2. Special forms. We discuss first the automorphs of the form

<p(a:) =t!l"X;} {J!~s(X~+k+X~+S+k)}: n =r+2s, (1)

where both the possibilities r = 0 and s = 0 are permitted. We may

write
(2)
where
Z, =xj
(3)

and i2 = - 1. If the XI are all real, then z; is real for 1 ~ j ~ rand z,H
and z,+ s+ k are conjugate complex numbers for 1 ~ k ~ s; and conversely,
if the ZI (1 ~ 1~ n) are of this shape then the XI are real. Let now w
be a real automorph of <p(a:). In the obvious way it gives rise to an
w
automorph of 1f' (z). Let Z = wz. Then
(4)

identically in Zl' ... , Z,,' where the ZI are linear forms in Zl' ... , z". The
only possibility is that Z L =).1 ZI where L = L (I) is a permutation of
1, ... , nand ).1' ... , )." are real or complex numbers. For our purposes,
it is enough to consider the automorphs
(1~l~n), (5)
where
(6)

by (4). But the transformation w transforms the real point a: into the
real point X = wa:. Hence Zl' ... , Z, are real and Z,H' Z'+s+k are
conjugate complex, and so
).i = real (1~j~r) }
(7)
).,+ k' ).,+s+ k conjugate complex (1 ~ k ~ s).
 Special forms 267

Conversely, if the numbers Al satisfy (6) and (7), then (5) defines a real
automorph CI) of tp(~).
We shall also need the transformation CI)* polar to CI), that is the
transformation such that identically

when
x = CI) ~, y = CI)* y.
Now
L XIYI = LZIWI,
I I
where ZI is given by (3) and

(1~i ~ r)

2W,+k = Y,+k - i Y'+S+k} (1 ~ k ~ s).

2W,+s+k = Y,H + i Y,+s+k

Hence if CI)* induces the transformation (;,* in the w-co-ordinates, we

must have

where W = w*
w. In particular, the transformation CI)* is also an auto-
morph of tp (~).
X.2.2. We shall require also to know something of the automorphs
of the form
tp (~) = x~ + ... + x~ - ~+l - .•. - x:, (1 )

where possibly' = n, so there are no negative terms. For completeness

we prove the well-known
LEMMA 2. II tp(~) is delined by (1) and ~o is any point, then there
is an automorph CI) 01 tp (~) such that, lor some number t,
CI) ~o = (t,o, ... ,0)
or
CI) ~o = (0, ... , 0, t)
or
CI) ~o = (t, 0, ... , 0, t)

according as tp (~o) > 0, tp (~o) < 0 or tp (~o) = O.

This is certainly true for n = 2, since then there are the well-known
automorphs X =CI)~ given by

Xl = Xl cos{} + Xa sin {}, Xa = - xlsin{} + xacos{} (2)

268 Automorphs

for any real {} when r = n = 2, and by

Xl + X 2 = k(XI + X 2), Xl - X 2 = k-I(XI - X 2) (3)
when r = 1, n =2 and k ma.y take any values except k = 0.
Next, the lemma is true when r = n. For we may suppose it proved
for n - 1. There is then an automorph WI acting only on the first n - 1
co-ordinates such that
Xl = WI Xo = (u, 0, ... ,0, x"o)
for some u. Then an automorph W2 acting only on the first and last
co-ordinates makes
W 2 Xl = (t, 0, ... ,0)
for some t. Then W = W 2 WI does what is required.
Finally, the lemma is true in general. For we may find in succession
automorphs WI' W 2 , Wa such that for some numbers u, v we have
Xl = WI Xo = (u, 0, ... ,0, Xr+],Q, ... , X"o) ,
X2 = W 2XI = (u, 0, ... ,0,0, ... ,0, v);
and then
X3 = W 3X 2 = (t, 0, ... ,0) or (0, ... ,0, t) or (t, 0, ... ,0, t) .
COROLLARY. 1/ W is the automorph constructed above, then the polar
w* is also an automorph.
It is readily verified that the polars of the special transformation (2)
and (3) are automorphs of cp(x). The required result now follows by
induction.
[It is in fact true that if w is any automorph of IP (z) then its polar is also an
automorph. This is most easily proved using matrix theory. Let w for the nonce
denote the matrix whose elements are the coefficients in the transformation w
and let E be the matrix with 1 in the first I' places on the diagonal, - 1 on the
remaining diagonal places, and 0 elsewhere. The fact that w is an automorph is
expressed by
W'EW = E, (4)
where the dash (') denotes the transposed. On taking the reciprocal of (4) we
obtain
(5)
But the polar w· of w is clearly w· = W'-I; and so w· is an automorph of IP by
(5). since E-1 = E.)
X.3. A method of Mordell. In this section we discuss a method of
MORDELL for estimating the lattice constant of an n-dimensional set
by reducing the problem to an (n -i)-dimensional one.
Let [I' be any n-dimensional set, not for the moment necessarily
endowed with any automorphs and A a lattice. In Chapter I, § 5 it
 A method of Mordell 269

was shown that if b is any point of the polar lattice /\*, then there
are n - 1 linearly independent points of /I. on the plane
lTb: ;rb=Q

(scalar product). The plane lTb cuts Y in an (n -i)-dimensional set

Y b . In an obvious sense, there is an (n -i)-dimensional lattice /\b
consisting of the points of /\ in lTb' Hence, if we can show that there is a
point other than 0 of the (n -1 )-dimensionallattice /\b in Y b , then there

°
is certainly a point other than 0 of /\ in Y. If bn=f=O, for example, one
could project Y b on to the hyperplane Xn = and use Lemma 6 Corollary
of Chapter I. For this procedure to be effective, the vector bE /\* must
be chosen so as to give a good (n -i)-dimensional problem in lTb; and
so in general we have replaced one n-dimensional problem by another,
rather vaguer, one for the polar lattice, together with an (n -1)-
dimensional problem.
In this shape the technique has been applied by MULLENDER (1950a)
and DAVENPORT (1952a) to the enigmatic 3-dimensional starbody

IXII max (xi, x~) < 1.

Making use of the known (d. § 3.3) lattice constant of the set

IXII (xi + x~) < 1 ,

they select a point b of /\* for which bdb~+ b~) is small and then treat
the 2-dimensional problem in lTb'
MORDELL (1942a, 1943a, 1944b) observed that it is sometimes
possible to make the n-dimensional problem for the polar lattice the
same as the original problem; and then the n-dimensional problem is
reduced entirely to one or more (n -i)-dimensional problems without
the need to solve an n-dimensional auxiliary problem. The sets Y for
which this procedure is feasible are those with a large group of auto-
morphs, so it is appropriate to discuss them in this chapter. From one
point of view it may be regarded as based on a generalization to non-
convex bodies of the results in Chapter VIII about polar convex bodies.
X.3.2. We first consider quadratic forms, for which OPPENHEIM
(1946a) has given a neat treatment following MORDELL (1944b).
THEOREM IV. Let r"s=LI(24) be the lattice constant 01 the (r+s)-
dimensional star-body
24,s Ix~ + ... + x~ - X~+l - ... - x~+sl <1 (1 )
lor r;;;;O, s;;;;O. Then
(2)
where the lirst or second term is omitted il r = °or s = °respectively.
270 Automorphs

Write
'P(x) = 'P" s(X) = X~ + ... + X~ - X~+1 - ••• - x~+s, (3)
and
(4)

for any lattice A. Then, by homogeneity,

r..-2 _ s~p {l1P1d (A)}'+s
',S - (A) 2
(5)

over all lattices A, with the natural convention that if I'P I (A) = for °
all A, then 4, S= 00; as most probably happens when r > 0, S > 0,
r+s~5 (see appendix A).
We show first that
{I'PI (A)}'+S-l ~ C- 2 {1'P1 (A*)} d 2 (A) , (6)
where A* is the polar lattice of A and

C= min (4-l,s, 4,S-1)' (7)

It is enough to show that
{I'PI (A)}'+S-l ~ C- 2 1'P(b)i d2 (A) , (8)

where b is any primitive point of A*. After Lemma 2 we may suppose

that b is one of the points
b l = (t, 0, ... ,0), b 2 = (0, ... ,0, t), b 3 = (t, 0, ... ,0, t), (9)

where b l , b 2 , b3 occur only if r>O, S>O and both r>O, S>O, respectively.
Consider first b = b l , where
'P (b l ) = t2 • (10)

By the results of § 5 of Chapter 1 there is a basis aI' ... , an for A such

that
bl a l =1, blai=O (2~j~n):
so that al=(t-l,a~) and ai=(O,aj) for j=l=1, where aj is an (n-1)-
dimensional vector. Hence the points of A in Xl =0 form an (n -1)-
dimensional lattice M in the space with co-ordinates X 2 , ... , Xn with
basis aj (2~j~n). Further,

d(A) = Idet(al, ... ,a.. )i = It-llldet(a~, ... ,a~)1 = Itl-1d(M). (11)

But now by (5) with r -1, S for r, s we have

 A method of Mordell 271

by (10) and (11). This proves (8) in the case b = b l since the left-hand
side of (12) is not less than {Irp I(I\)},+S-l. The proof of (8) in the second
case, when b = b 2 in (9), is similar except that the roles of rand s are
interchanged.
It remains to consider the case

b = b a = (t, 0, ... , 0, t) ,

which occurs only when r>O, s>O, so

rp(b) =0.

There then exist a basis ai' ... , an of 1\ such that

ba; = ° (2~i~n).

Introduce new co-ordinates xj by

,
Xn = Xl - X"' xj=Xj (i=l=l,n),
so that
rp (:r) = Xl, X""+2
X2 + ... + X,' 2
- 'X'+l
2
- ... -
'2
X,+s-l,
and the points a2 , ••• , an lie on X~ = 0. The points of 1\ on X~ = form °
an (n -1)-dimensionallattice M, and rp(:r) with x~ =0 depends only on
the n-2 variables X 2 , ••• , xn - l • Hence Irp(:r) I takes arbitrarily small
values on M; for example, by the degenerate case of MINKOWSKI'S
convex body Theorem, there are points of M other than 0 with

X~=o, Ixjl<e (2~i~n-l),

where e> 0 is arbitrarily small, since this set has infinite (n - 1)-
dimensional volume. Hence (8) holds also when b = ba ; and so generally.
This concludes the"proof of (6).
We may also apply (6) to the lattice 1\* with its determinant
d (1\ *) = d- l (I\) and its polar lattice 1\:

{Irp I (1\*) }'+S-l ~ C- 2 {Irp I (I\)} d- 2(1\) . (6')

On eliminating Irpl (1\*) between (6) and (6')", we obtain

{Irp I (1\) }('+S) (,+s-2) ~ C- 2(,+s) {d (1\)}2 (,+s-2).
This implies the required result (2) on using (5) and (7).
In general there is no reason to expect there to be equality in (2),
but this sometimes happens, as in the following
COROLLARY.
272 Automorphs

By Theorems III and VII of Chapter II, we have

Fa 0= 2- 1, 1;,1 = r;,2 = mi.
Hence the theorem shows that ~,o and 1;,2 have at least the values
specified. The forms
!{(x2 + xa + X4)2 + (X2 + Xl )2 + (xa + Xl )2 + xU
and
X~ - x~ - x§ - x~ + X 4 Xl + X4 X 2 + x 4 xa + 2x1 xa + 2Xl X 2

have signature (4,0), (2,2), and determinants i,! respectively and do

not represent members less than 1 in absolute value for integer value
of the variables not all 0, as is easily verified. This proves the corollary
on making use of the relationship between forms and lattices of Chapter I,
§ 4 (especially Lemma 4).
Again, as MORDELL observed, Theorem IV gives Fa,o once 1;,0 is
known. Again, the method of proof of Theorem IV gives the lattice
constant [i, see OPPENHEIM (1953 b)] of
o < x~ + x~ - x~ - x~ < 1,
once that el, see DAVENPORT (1949a)] of
O<x~+x~-x~<1
is known. These sets are not star-bodies. It is necessary to choose the
+
point b of A* so that b~ b~ - b~ - b~ is numerically small and negative.
It is possible to use MORDELL'S method to obtain information about the
critical lattices wheR there is equality in (2). We do not do this here
since we shall do something similar for products of linear forms in § 3.3.
X.3.3. Before applying MORDELL'S method to ternary cubics we
must translate Theorem VIII of Chapter II out of the language of forms
into that of lattices.
LEMMA 3. The lattice-constant of the 2-dimensional set
.9'": Itp(x) I <1, (1 )
where
tp(x) =XlXZ(Xl +X2), (2)
is LJ(.9'") = 71. There are precisely two critical lattices, Ml and M2. These
lattices have only 0 in common.
Let {}l' {}2' {}s be the roots of
(}3+{}2-2{}-1=0 (3)
in some order. Then the lattice M({}l'{}2,{}a) with basis aI' a! defined by
71a l = ({}2 - {}a, {}a - (}l)' 71a2 = {{}l ({)! - (}a). {}2 ({}a - (}l)} (4)
 A method of Mordell 273

is one of the two critical lattices. I f-&~ , -&;, -&; is a permutation of-&I' {}2' -&3'
then M(-&1' -&2' -&3) = M(-&~, -&;, -&;) if and only if the permutation is an
even one.
The geometrical purport of the lemma becomes clearer if new co-ordinates
YI' Y2 are introduced by the equations
1 V]
xI=YI' x2=-2YI+2Y2'
so

In YJ' Y2 co-ordinates, the region .'T has three asymptotes at an angle of 2n!3
and is carried into itself by either a rotation through 2n/3 round the origin or by a
reflection in an asymptote. Thc two critical lattices given by the lemma are then
each invariant under a rotation through 2n/3 and each is carried into the other
by a reflection in an asymptote. The reader may find it instructive to draw a
figure of the critical lattices each with 6 pairs of points on the boundary. For a
treatmcnt of sets .'T' which have similar symmetry and convexity properties to
.'T by the geometrical methods of Chapter III see BAMBAH (1951 a).
In what follows we do not introduce YI' Y2 as above but we do maintain the
essential cyclic symmetry between XI' x 2 and - XI - x 2 •
We note that the roots of (3) are
Ll
0'1 = 2cos~
2n Ll
0'2 = 2cos~
4n e"3 = 2cos~
6n
(5)
7 ' 7 ' 7 '

so that -&1' -&2' -&3 are a permutation of 8 1, 8 2, 8 3, We have the trivial

identities
82=8~-2, 83=8~-2, 8 1 =8;-2, 81=1-82-8~ etc. (6)
The value of J(.'1} follows at once from Theorem VIII of Chapter II,
so it remains only to verify the statement about the critical lattices.
By Theorem VIII of Chapter II, if M is critical there is certainly a basis
u 1 , u 2 of M such that
1p (u 1 u 1 +
U 2 u 2 ) = - to (u 1 , u 2 ) , (7)
where
(8)
for one may interchange the two elements of the base given by Theo-
rem VIII of Chapter II or take - Uk for Uk (k = 1, 2). Let

Uk = (au, a2 k) (k = 1, 2), (9)

and define numbers a3k by

(k = 1, 2). (1O)
Then (7) becomes
II
1~j::;:;3
(a j1 u 1 +a j2 u2) = IT
l::;:;j::;:;3
(u 1 +8 j u 2 )· (11)
Cassels, Geometry of Numbers 18
274 Automorphs

Hence
(12)
where {fl' D2 • Da is some permutation of 8 1 , 8 2 , 8 a. From (10) and (12)
we have
lajl = D;+l - O;H }
(j=1.2.3). (13)
l a;2 = D; (Of+l - Of H)
where ° 4 =°1 • OS=02 and l is some number. By (11) we have
IT afl = 1.
f
and so in fact
la = (Dl - ( 2 ) (0 2 - Oa) (Oa - ( 1) (14)
=±7. (15)
where the value ± 7 may either be checked directly from (5) or from
the fact that the square of the right-hand side of (14) is the discriminant

°°
of the cubic f0 (u l • u 2 ) by definition (§ 5.1 of Chapter II). We note that
1 , 2 , Oa determine a l and a 2 absolutely uniquely. by (14).

But now we have the identity

fo(w+v.v) =fo(-v.w).
Hence if the point aa of M(°1 , °
2• Oa) is defined by

we have

and so a 2 • aa must correspond to a permutation O~. O~. O~ of 1 , 2 • Oa; °°

and it cannot be the identical permutation by the last sentence of the
previous paragraph. Hence the cyclic change of bases of M (°1 , 2 , Oa): °
(a l • a 2) --+ (a 2 • aa) --+ (aa. al ) --+ (~. a 2 )--+

° °°
must correspond to a cyclic permutation of 1 , 2 , Oa. Hence there are
°
at most two distinct lattices M (°1 , 2 • Oa), for the permutations 1 • D2 • Da
of 8 1 , 8 2 , 8 a.
lt remains to show that M (°1 , '19 2 , Oa) is distinct from M (81 .82 , 8 a)
if D1 • O2 • Oa is an odd permutation of 8 1 , 8 2 , 8 a. We may suppose now.
without loss of generality. that

°=8 °=8
1 2, 2 1, Oa=8a·
From (4). (6). (13) and (15). a point b of M (81 .82 , 8 a) has
7ibj = P(8 j ) (j = 1. 2. 3). (16)
 A method of Mordell 275

where P(t) is a polynomial in the variable t with (rational) integer

coefficients. We may suppose, by (3), that P(t) is of degree ~2: and
then it is completely determined by anyone of bl , b2 , bs . If b is also in
M(e2 ,el ,ea), then it is also of the shape
7l b1 = Q(e2 ). 7!b 2 = Q(e1 ). 71bs = Q(eaL
for some polynomial Q (t) of degree ~ 2 with integer coefficients. But
now p(ea) =Q(ea), and so the polynomials P(t) and Q(t) are identical.
Hence
p(e z) = p(el ): (17)
and so
p(ea) = p(e z) = p(el ) , (18)
since p(e j ) {J' = 1,2,3) are conjugates l . Finally,

by (16) and (18), and so

b1 = bl = 0:
That is, 0 the only point common to M (e1 , e z, ( 3) and M(e 2 , e1 , ( 3 ),
as required.
X.3.4. We now apply MORDELL'S method to prove results for
XlXIXa and Xl(X~+X:). These are equivalent to weaker forms of Theo-
rems X and XI of Chapter 2, where the relevant literature is cited.
We shall later prove something rather stronger by the use of isolation,
but will not prove the full force of Theorem X of Chapter 2 in this book.
The methods extend to products of n real or complex forms in n dimen-
sions in a way which will be obvious, but do not then give the exact
lattice constants [MORDELL (1941a) and (1943a)].
THEOREM V. A. The lattice constant 01 the 3-dimensional set
.Ai: Ix1x1xal <1 (1)
is L1(.Ai) = 7. Denote by Nl the lattice with basis
b 1 =(1,1,1), b z ={lh,f)2,{}a). ba=({}~,{}:,U:), (2)
where {}l' {}2' {}a are the roots 01

in some order. All the critical lattices A 01 .A'i which have a point a lor
which
(4)
1 Alternatively, (17) means that p(ef - 2) = p(e1 ); and so the polynomial
e.
+ t 2 _ 2t - 1. One may now put t = and obtain
P(t 2 _ 2) - P(t) is divisible by t 3
p(e.) = p(ea).
1S*
276 Automorphs

are 01 the shape

A=wN l • (5)
where w is an automorph 01 ~.
B. The lattice constant 01
A;: IXII (x: + x~) < 1 (6)
is LI(A;) =1(23)t. Denote by N2 the lattice with basis

(1.1.1). {DI• ~ (D2+Da). 2ti (D2-Da)}.)

;i (D~ - D:)} ,
(7)
{O~. ~ (D: +D~).
where i 2 = -1 and {)l is the real, and D2 , Da are the complex roots 0/

D3-D2+1 =0. (8)

Every critical lattice A lor A; which possesses a point a with
1~(a:+a:)1 =1 (9)
is 01 the shape
A =wN a, (10)
where w is an automorph 01 A;.
We first prove Theorem V. A. The lattice NI given by the theorem
is certainly ~-admissible, since a point a of NI has co-ordinates
(J' = 1, 2, 3), (11)
where ul , ua' Us are integers. Then ~aaaa is a rational integer by its
symmetry in DI , Da, Da. If alaaaa=O, then one of the a; is 0, say
uI+uaDI+UsD~=O; and this is impossible unless uI=uz=ua=O, since
DI does not satisfy any equation of degree less than 3. Further,
d(NI) = Idet(bl.ba.ba)i =1(DI-{)a)(Da-{)a)({)a-{)I)l =7, (12)
as was verified already in the proof of Lemma 3. The lattices obtained
by different permutations of DI , Oz, {)a in (2) all differ from each other
by an automorph of ~, namely a peI'fhutation of the co-ordinate axes.
Write

and, as before,
I<pI (A) = aEA
inf l<p(a)i.
~o

We show first that, for any lattice A.

(13)
 A method of Mordell 277

where A* is the polar lattice of A. The proof follows closely the pattern
of the proof of Theorem IV. It is enough to show that
(14)
where b is any primitive point of A*.
Suppose, first, that !p (b) = 0. Then, after applying a suitable auto-
morph of .Ai furnished by § 2.1, we may suppose without loss of
generality 1 that
b=(1,O,O) (15)
or
b=(1,1,O). (16)
In the first case, (15), the plane
bz = 0,
which must contain two linearly independent elements of A is iust the
plane x1 =0, and all points on it satisfy !p(:':) =0. Hence (14) certainly
holds in this case. In the second case, (16), there are two linearly inde-
pendent points of A on
(17)
For these points
(18)

and the 2-dimensional set IX~X31 <8 is of infinite type for any 8>0.
Hence there are certainly points aEA other than 0 with 1!p(a)I<8.
This proves (14) in the case b is given by (16).
There remains the case when !p(b) =1=0 and so, after the application
of a suitable automorph, we may suppose that
b = (t, t, t) , t> 0, (19)
and so
(20)

We have supposed that b is primitive, and so, by Lemma 6, Corollary

of Chapter 1, the 2-dimensional set of points (Xl' X 2) such that
(Xl' X 2 , - Xl - X 2) E A
is a lattice M of determinant
d(M) =td(A).
But now
inf Iflta2(flt +a2)1 ;£{7-1tl(M)}I= rlt1dl(A) ,
(II....)EM
,*,0

1 For we may suppose that b1

according as b.= 0 or b.* o.
* O. ba= O. One gets the shape (15) or (16)
278 Automorphs

by Lemma 3, the exponent I being correct for reasons of homogeneity.

A fortiori
Irpl (/\) ;;;:; 7-~ tt dt (/\).
This proves (14) when b is given by (19) and (20); and so completes the
proof of (13) and (14).
On interchanging /\ and 1\* in (13) and using d(/\*) =d-l (/\), we
have
{I rp I (/\*)}2;;;:; 7-1{I rp I (/\)} d- a (/\) . (13')
On eliminating Irp I(1\*) from (13) and (13') we obtain
Irpl (/\);;;:; ;-1 d (/\), (21)
so LI(~);;;:; 7, since NI is the set of x with I rp (x)1 < 1; and then LI(~) = 7
since we have already exhibited an admissible lattice NI , with d (N I ) = 7.
It remains to consider the critical lattices /\c with a point on the
boundary, and we may suppose, after the use of a suitable automorph,
that
(1,1,1)E/\c d(/\c) =7. (22)
Clearly then the 2-dimensional lattices considered above will turn out
to be critical for the relevant 2-dimensional sets, and it is necessary
only to check that this can happen only when /\c = NI for a suitable
choice of {)l' {)2' {)a, where NI is defined in Theorem V. A.
We note first that
Irpl (/\:) = ;-2 (23)
by (13) and (13'). Hence the lattice M; of points
(Xl' X 2) with (Xl' X 2 , - Xl - X 2) E/\c*, (24)
which has determinant

must be one of the two critical lattices for

IXl X 2 (Xl + X 2) I < 7-2 (25)
given by Lemma 3. But we have already seen that NI for any choice
of {)l' {)2' {)3 is critical, and so the lattice
M~ = M~ ({)l' {)2' {)a),
defined by putting Nl = Nl ({)I , {)2' {)3) for /\c in (24), is also critical.
Clearly, by the proof of Lemma 3, both critical lattices of (25) occur as
M~ ({)l' {)2' ()3) for suitable choice of {)l' {)2. f}3' Hence we may suppose
without loss of generality, that
M; = M~;
 Existence of automorphs 279

that is, the polar lattices /\: and N: are identical at least on the plane
Xl + X2 + Xa = o.
Let now b = (bl , b2 , ba) be any point of /\: (and so of N:) with
(26)

Then the lattices /\~, N~ consisting of the points of /\t and of Nl re-
spectively in the plane
(27)
must both be critical, in the obvious sense, for the 2-dimensional section
of IXl X2 Xai < 1 by the hyperplane (27). By Lemma 3, there are only
two critical lattices and these have only the origin in common. Hence
/\~ and /\~ must be identical, since (1. 1, 1) belong to both lattices, by
(27). Thus /\t and /\1 coincide on any hyperplane (27) such that the
point b satisfies (26).
But now N:has a basis b:. bt,
b: (say) such that b b: =b:,
satisfies (26), for we have only to choose a suitable basis b: for the b:,
section of Nt by Xl + X 2 + Xa = 0 and extend it to a basis for N:. Let
bl , b 2 , ba be the polar basis for Nl . Then. on putting b = b:, b: in
(25) in turn, we see that /\t contains all points a of Nl such that either
b: a = 0 or b: a = 0;
that is all points of Nl of the shape either
U2 b2 + Ua ba or VI b l + Va b a,

where ul • U2• VI. V3 are any integers. Hence /\t must contain each point
U1 b 1 + u 2 b 2 + uaba = (u 2 b 2 + u 3 b 3 ) + (ul b l + Oba)
of N1 • Since d(Nl } =d(/\t}; we then have /\t=N l , as required.
This completes the proof of Theorem V. A. That of Theorem V. B
is similar except that Theorems VII and VII A of Chapter III are used
instead of Lemma 3. The details may be left to the reader.
X.4. Existence of automorphs. In this section we prove the exist-
ence of common automorphs of a lattice /\ and a form ffJ (~) which is
integral and non-null on I\, and make deductions about the possible
such /\ in a special case.
We shall require a quantitative form of MAHLER'S compactness
criterion, Theorem IV of Chapter 5.
LEMMA 4. There is a number
(1)
280 Automorphs

depending only on the integer n>O and the numbers ,11>0, ">0, 8>0
with the following property: amongst any No lattices 1\; (1 ;£i;£N) in
n-dimensional space such that
d (1\;) ~ ,11 (2)
and
11\;1 ~ ", (3)
there is at least one pair, say 1\1' 1\2' such that
1\2 = ~1\1 (4)
and the linear transformation ~ satisfies
(5)
where l is the iaentity transformation.
We recollect that
11\1 = aEA
inf lal, (6)
*0
and that the symbol 110'11 for a linear transformation X = 0':£ with
X;=L(I;"X" is 110'11 =nmax l(lf"l.
It would be possible to modify the proof of Theorem IV given in
Chapter V but it is simpler to follow the alternative proof sketched in
§ 2.2 of Chapter VIII. We suppose we have No lattices I\f, where No
will be determined later. By Lemma 3 of Chapter VIII there is a
,10> 0 and a K depending only on ,11 and ", such that any I\i satisfying
(2) and (3) has
d(I\;) ~ Llo >0 (7)
and has n linearly independent points in the sphere
1:£1 ~K.
By Lemma 8 of Chapter V, there is then a basis

of I\i with
(8)
Let '1/>0 be arbitrarily small, to be chosen later. Then, by (8), if No
is greater than an N1 depending only on n, '1/, Llo, K, that is on n, '1/,
Ll 1, ", there are two A; say 1\1 and 1\2' such that
Ibn - bi2 1<'1/ (1;£ i~ n). (9)
Since the b o are linearly independent, we have

b i2 - b n = L" 0';; bi1

;=1
 Existence of automorphs 281

for some numbers a;i' But now on solving for the aii from (7), (8)
and (9), we have
\aij\~aOrJ (1~i~n, 1~i~n),
where a o is a number depending only on ,1oK and n; a crude estimate
being
ao = n! ,10 1 (n K)"-1
obtained by estimating the elements of the matrix reciprocal to the
matrix with columns b n (1 ~i~n). Hence
\Ia\l<e
+
if rJ chosen to satisfy n a orJ < e. Hence 't' = l a has 't' Al = A2 and
I\'t'-ll\<e. Since A1 ='t'-IA 2 we have also 1\'t'-I- l l\<e, because (9) is
symmetric in AI' A2. This concludes the proof.
X.4.2. We shall also require the following rather trivial lemma which
says, roughly, that a form rp (z) cannot be integral on too many essen-
tially distinct lattices.
LEMMA 5. Let rp(z) be a lorm integral on a lattice A. Then there is
an rJ> 0 depending only on rp (z) and A with the lollowing property: II
rp (z) is integral on 't' A and

then 't' is an automorph 01 rp(z).

Let rp (z) be of degree m and let b1 , ••• , b" be a basis for A. If 't'
satisfies (1) with sufficiently smallrJ, we have
Irp ('t' f Uj bj) - rp (f I
Uj b j ) < 1 (2)

for all integers Uj such that

(3)
Then (2) implies
(4)

for the integers (3), since both sides of (4) are integers. By Lemma 1,
it follows that (4) holds for all real numbers Uj. Since every z is of
the shape LUjb j with real Uj' we have rp('t'z) =rp(z) for all z, as required.
COROLLARY. Suppose, Iurther, that rp(z) is non-null on A and that

d(A) ~,11
lor some ,11' Then rJ may be chosen depending only on rp and ,1., but not
otherwise on A.
For then
\rp(a)\ ~1 (aEA, a=Fo);
282 Automorphs

and so
IAI~c>o
for some c depending only on cp. Hence, as in the proof of Lemma 4,
there is a basis b1 , ... , b" of A with
(1;£j;£n)
for a K depending only on .d 1 and c, i.e. on .d 1 and cp. Hence all the
u,
points 1.: b, subject to (3) lie in a sphere
(5)
Then (2) holds for small enough 1J depending only on cp and K, since
cp(x) is uniformly continuous in (5). Hence the corollary follows.
XA.3. We are now in a position to prove the main theorem on the
existence of automorphs.
THEOREM VI. Let the form cp (x) be integral and non-null on the
lattice A and let a be any automorph 0/ cp (x). Suppose e> 0 is given
arbitrarily small. Then there is an automorph 't 0/ cp (x) with

I/'t-lll<e, (1)
such that
(2)
is an automorph 01 A for certain integers u, v with
o~u<v. (3)
It is not excluded, of course, that w may be the identical trans-
formation.
We have
Icpl (A) = inf\ cp(a)1 ~ 1, (4)
aE/\
,*0
by hypothesis, and so
(5)
for all integers u. Hence
laUA\ ~ c> 0 (6)
for all u and some constant c> O. Further,
d (aft A) = d (A) (7)

for all u since det (a) = ± 1 by Theorem I. By (6) and (7) we may apply
Lemma 3 to the aU A (1;£ u;£ N), if N is some large enough number,
to obtain two lattices a" I\. and a" I\. such that
aU A= 'tav A (u < v) (8)
 Existence of automorphs 283

and
(9)
We may suppose, by choosing a smaller number instead of the ori-
ginal e if necessary, that e<1], where 1] is the number in Lemma 5,
Corollary with L11=d(I\). We may then apply Lemma 5, Corollary with
a V 1\ instead of 1\ and deduce from (8), (9) that't is an automorph for
<p(x). Hence w defined in (2) has all the properties required.
Theorem VI becomes false if the condition that <p(x) be non-null
on 1\ is omitted, as is shown by the 2-dimensional example where 1\ = 1\0
is the lattice of integral vectors, <p (x) = Xl X 2 , and a is the automorph
x1 -?2x1 , X 2 -?!X 2 . But in more dimensions it is sometimes possible to
use the idea behind Theorem VI to construct automorphs of 1\ even
when <p (x) may be null on 1\, for example, by restricting attention to
automorphs leaving fixed an element or elements of 1\ or of the polar
lattice 1\*.
X.4.4. Theorem VI takes a particularly simple shape when

<p (x) = { II
l~i~'
Xi}{ n
I ;i;k;i;s
(X~+k + X~+S+k)}' (1)

+
where n = r 2s, which is substantially equivalent to, but rather
stronger than, DIRICHLET'S theorem on the existence of units in an
algebraic number field. We write as usual
(1~j~r)

Z,+k = X'+k + i X,+s+k } (2)

z,+s+.~ = X,+k - t X,+s+k (1 " k " 'i. )
It is convenient to work with the zi rather than the xi' so we shall speak
of the zi as the appropriate complex co-ordinates. We shall also say
for brevity that a set of numbers Ai (1 ~j~n) is compatible with <p(x) if
Ai = real (1 ~ j ~ n)
A,+k' A,+s+ .• conjugate complex (1 ~ k ~ s).
THEOREM VII. Let <p(x) be given by (1), and let Aj (1 ~j~n) be
numbers compatible with <p (x) such that
17 Ai = 1.
l~i~n

Suppose that <p (x) is integral on 1\ and that to> 0 is given arbitrarily
small. Then there are numbers Wi compatible 'with <p (x) and an integer
m> 0 such that
(1~j~n), (3)
284 Automorphs

and such that the translormation (a) given in the appropriate complex co-
ordinates by
(1~i~n)
is an automorph 01 A
The automorphs of rp (a:) were discussed in § 2.1. From what is said
there it is clear that if Z ='U is an automorph of rp given in the ap-
propriate complex co-ordinates and if
II,; -lll < n, (4)
where n is the dimension, then T must be of the shape
(1~i~n); (5)
that is, there can be no permutation of the forms on the right-hand side:
indeed, if '; is written as Zi= L T,IIZ,<> the inequality (4) implies
II

(1~i~n),

and the only automorphs of this kind are (5). If Z = AZ is given in

complex co-ordinates by
Z; = liZ; (1 ~i~ n),
it follows now that A and '; commute. Hence applying Theorem VI
with 0" =A we have

where m = v - u. Then (a) does what is required.

We shall later require to know slightly more about the automorphs
(a) of lattices on which rp (a:) given by (1) is integral; and it is convenient

to prove it here.
LEMMA 6. Let rp (a:} given by (1) be integral on A and let the automorph
Z =(a)Z 01 A be given in the appropriate complex co-ordinates by
(1~i~n).

Then the Wi are algebraic units, that is they satisly an equation 01 the type
I(w;) = 0,
where
(6)
lor some m and c1 , ••• , cm - 1 are rational integers.
Let b l •.•.• b" be a basis for 1\. so that

(a) bi = L mi" b"

1:0;11:0; ..
(7)
 Existence of automorphs 285

for some integers mj k' Since the w bi are a basis, we have

det(mik) = ± 1.
Let
(1~k~n) (8)

in the appropriate complex co-ordinates and let B be the matrix of

which the rows are given by (8). Then (7) takes the shape
Bw =mB, (9)

where m is the matrix with elements mik and w is the diagonal matrix
with elements WI' .•• , Wn on the diagonal. Hence
w =B-1 mB,
and Wt, ... , Wn all satisfy the equation I (wi) = 0, where
I(t) =det(tl-m),
which is of the form (6).
The two following corollaries are immediate
COROLLARY 1. WI'"'' Wn satisly the same equation 01 type (6) with
m=n.
COROLLARY 2. II Wj is rational, then Wj= ±1.
Although we do not need it later it is interesting to note that Theorem VII
and Lemma 6 rapidly gives a complete characterisation of the lattices A on which
q:>(:I:) is proportional to integral and non-null, at least when r>O. We only sketch
the proof, for details see BACHMANN (1923a) Kap.12.
LEMMA 7. All the lattices A on which q:>(:I:) is proportional to integral may be
obtained in the following way. Let srI' ... , sr"
be a set of conjugate algebraic fields
of degree n over the field of rational numbers, wI/ere srI' ... , sr,
are real and sr,+k,
sr,+s+k are conjugate complex (1 ~k~s). Let Yll' ... , YIn be linearly itldependent
elements of srIover the rationals and iet Ylk (t ~l~n) be the conjugate of Ylk in sri'
Let M be the lattice with basis
(t:::;;k:::;;n)
in the appropriate complex co-ordinates. Then a necessary and sufficietlt condition
tllat q:> (:1:) be proportional to integral and non-null on a lattice A is that A be of the
shape

wllere t is real, "" is an automorph 0/ q:>(:I:), and M is of the type fust described.
When r > 0, the proof is shorter than the enunciation. By applying Theorem VI I
with

we deduce the existence of an automorph (a) of q:> (:1:) and A with

(10)
286 Automorphs

Since WI' ...• w" all satisfy the same equation of degree 11. they must all by (10)
be precisely of degree n and so conjugates. Let b l •...• b n be a basis for /\ and
use the notation (7). (8). Then it follows from (9) that
(Pil' ...• {Jj II)
is an eigenvector belonging to Wj of the matrix tn. But clearly m has a set of con-
jugate eigenvectors
(Yil' ...• Yin)
in the fields sr j generated by Wj; and if these are identified with those of the enuncia-
tion it is easy to see that the lattice M has the required properties.
When y= O. the position is more difficult since it may be impossible to achieve
that the Wj are all of degree n. though it is possible to make them all of degree
tn. Let a = (IX\ •...• IX,,) and b = (Pl' ...• {In) be two linearly independent vectors
of /\ in the appropriate complex co-ordinate system. Then 91 (ua + v b) is a poly-
nomial in the variables u and v with coefficients proportional to integers. and it
vanishes for integers u and v only when u = v = O. Hence lXI/PI is an algebraic
number. Similarly. if c= (YI' ...• Yn)E/\ is linearly independent from a and b.
then the ratios IXI/YI' PI/YI are of degree n as is also (PIX I + qPI)/:'l for any integers
p and q. It is not then difficult to deduce that lXI/PI is in a field of degree n depending
only on /\ and not on the choice of a and b; and the rl)st follows with some little
trouble. We do not go into details as we do not use the result.
X.S. Isolation theorems. As was stated in § 1 there is a wide
variety of isolation theorems, and it hardly seems worth while to for-
mulate theorems of great generality. We shall instead consider only
three concrete cases.
We shall need the following simple Lemma which is really a simple
case of KRONECKER'S Theorem and belongs of right in Chapter XI.
LEMMA 8. Let a., p, 1', 6 be real numbers with a.6 - PI' =t= o. Suppose
thata.IPisirrational. Then to every number e>O there is an rJ =rJ (a.,p, 1',6, e)
with the following property:
For any numbers A, I-' there are integers m, n such that
Ima.+np-AI<e, Imy+n6-I-'I~rJ·
By MINKOWSKI'S linear forms Theorem there are integers (m, n) =t= (0,0)
such that Ima.+nPI is arbitrarily small; and ma.+np=t=O since a.IP is
irrational. Hence there are integers (ml' nl ) and (m2' n 2) such that
o<lmla.+nJPI<e, O<lm 2 a.+n 2 PI<e,
and

Put
(j = 1, 2),
so that
(j = 1, 2),
Let e, f1 be the solution of
eXl + f1 X z = A,
 Isolation theorems 287

and choose integers a, b such that

la-el~t, Ib-O'I~t·
Then
laXl + bXz - AI = i(a - e) Xl + (b - 0') X2 1 ~t (IXll + IXzl) < £,
and
la~+bYz-,tl~t(I~1 +IYzJ) ='fJ (say).
The lemma now follows on putting
m=aml+bm z , n=anl +bn2 •
X.S.2. Perhaps the simplest isolation theorem is that for X l X 2 and
is due to C. A. ROGERS [unpublished, but see CASSELS (1957 a), Chapter II
where an application to the "Markoff chain", due to ROGERS, is given].
THEOREM VIII. Let Xl x 2 be integral and non-null on the 2-dimensional
lattice A and let there be a, bE A such that
al a 2 = - ex < 0 < bl b2 = p. (1)
Then there are numbers 'fJo>O, 'fJl>O with the following
properties:
Let 't' be a linear transformation and suppose that
11't'- LII<'fJl (2)
and
1'12 =l= 0, (3)
where the transformation X ='t':V is given by
Xl = TllXl + 1'12 X 2 , X 2 = T2l Xl + T 22 X 2 •
Then there is a point c =l= 0 ol't'A such that
- ex (1 - 'fJo) < cl Cz< P(1 - 'fJo) , Icll < 1.
We may suppose without loss of generality that
al>O, bl>O
and so
a2< 0, b2 >0.
By Theorem VII, there is an automorph X =W:V of A of the shape
Xl =WlXl , X 2 =W Z X 2 ,
where
0< % < 1 < W 2 , WI Wz= 1.
Then A contains all the points
am=(w;m~,w;'az)' b".=(w;mbl,w;'b z), (4)
where m is any integer, positive negative or O.
288 Automorphs

We must now distinguish two cases according ro the sign of Tl2 .

Suppose, first, that
Tl2>O. (5)
Let the integer m be determined by

as is possible, since Tn a,. > 0> Tl2 a 2 • Then

(6')
Hence
£0;' =O(T1h, (7)
where the constant implied by the 0 symbol may depend on aI' a2, and
where we assume TJl in (2) chosen so that, say, ITn -11 < t. Put
c='t'a,,,, (8)

where a", is given by (4). Then, in the first place, it follows from (6)
and (7) that

so
Icll <1
if TJl is chosen small enough. Secondly, it follows from (6) or (6') that
(9)
But now, by (7),
£02'" C2 = Tn al £02 2 '" + Tn a2 = Tn a2 + 0 (Tn). (10)

Put TJO=tW22. Then since a2<O<al , we have from (7), (9) and (10),
that

provided that Tl2' T2l are small enough and that Tn, Tn are near
enough to 1, which may be achieved by taking TJl small enough in (2).
This concludes the proof when T12 >O. The proof when T12 <O is
completely similar, except that b is used instead of a.
COROLLARY. Under the hypotheses 01 the theorem except (3), there is
an TJ2 such that il

and 't' is not an automorph 0/ Xl .%2, then 't' A contains a point c with
al a2(1 - TJo) < c) c2 < bl b2 (1 - TJo) •
 Isolation theorems 289

For if or is not an automorph, then either 'f12=F0 or 'f21=F0. If 'f12=FO,

then the theorem applies; and if 'fa I =F 0 then the theorem can be applied
with the roles of Xl and x 2 interchanged.
Note that Theorem VIn works with the values of XIX Z at two
distinct points of A. This rather restricts its field of application. The
other isolation theorems which we shall discuss require at most knowledge
of the value of the function at only one lattice point.
X.S.3. Before discussing the isolation results for
rp (or) = Xl Xa Xa
we require a simple lemma.
LEMMA 9. Let Xl Xa xa be integral and non-null on A. To every 8> 0
there is an 'Y/ > 0, depending on A, with the following property:
To any numbers e>O, 0'>0 and index k =1,2 or 3 there is an auto-
morph X =wor of A:
Xi =w,xi (1;£;i;£;3). (1 )
with
wi>O (1;£; i;£; 3), WIWzWa = 1 (2)
and
1 - 8 < eCOl_ < 1 + 8, 'Y/-l< COk <'Y/. (3)
co. G

For by Theorem VII there are certainly automorphs .s and ~ of A

defined by
Xi =8j Xj, Xj = V'j Xi (1;£; i;£; 3)
respectively, with
8 1 >1, 0<8z<1, 0<83 <1, 8 1 8 a8 3 =1,

Put
(4)
so
(5)
and
PI> 0, pz< 0, Pa< 0,
ql< 0, qz> 0, qa< O.
Hence
PI qa - Paql = P2Qa - PaQ2 =F O. (6)
We now show that
(7)
is irrational. If not, there would be an automorph A =.s"~o with
integers (u, v) =F (0,0) for which, in an obvious notation, At = )'2' But
Cassels, Geometry of Numbers 19
290 Automorphs

then, Al would be rational, since AI' A2 , Aa satisfy a cubic equation with

integer coefficients by Lemma 6, Corollary 1. Hence Al = A2 = Aa= 1, by
Lemma 6, Corollary 2; that is
u Pj + v qj = 0 (1 ~ i ~ 3) ,
which contradicts (6). By (6) we may now apply Lemma 8, with

A= log e, (X = P2 - PI' {J = q2 - qI'

p = log 0', I' = Pk'
and
min \log (1
±
± £)j, Ipg1}

or £, 'fJ respectively. Then

w =,sm$",
where m and n are given by Lemma 8, clearly has all the properties
required.
It is now a simple matter to prove
THEOREM IX. Let XI X 2 Xa be integral and non-null on 1\ and let £1>0
be arbitrarily small. There exists an 'fJ1>0, depending on £1 and 1\, such
that il
1I'1'-tll<'fJ1 (8)
and t'1' is not an automorph 01 Xl X 2 Xs lor any number t, then the lattice
'1' A contains a point c =1= 0 lor which
Ic1 c2 cs l<£I' (9)
Let '1' be given by Xi = 2: TijXj , when X ='1':11. If '1' is not an auto-
j
morph, there is a Tij=F 0 (i =Fi). We shall suppose that
(9')

this being one of twelve possible cases i • Now 1\ certainly does contain
some point a with
~>0>a2'

We shall pick one such point and keep it fixed in all that follows, so
that numbers depending only on a and 1\ will be said to depend only
on 1\, etc.
By Lemma 9 with an £ > 0 to be chosen later and

0' = 1, k =3, (10)

1 For the maximum in (9') may correspond to anyone of the six pairs (i, i)
with i*i; and the maximal Tij may be either positive or negative.
 Isolation theorems 291

there is an automorph W of A with

1_e<_a1'fU w1 <1+e, r(l< W3< 1], (11)

a2 T 12 W 2
where
(12)
is independent of the Tij' Since Tn is assumed near to 1, say ITn- 11 < t,
it follows from (11) and WI wawa = 1 that
1]'-1 Tia< WI < 1]' Tta, 1]'-1 Ti-i < Wa < 1]' Tj}, (13)
where
(14)
is independent of the Tij-

We put

Then by (9'), (11) and (13), we have

WIIICII = wIll al TuWl + aaT12w2 + aaT13Wal
~ wIl{1 al TuWl + aaT12W21 + 1aa T13 Wa I}< "'Ie + ~1 Tia.
where
"'1 = "'1 (A) , ~l = ~1 (A, e) .

It is important that "'1 is independent of 8. Hence

Willcil < 2"'18, (15)
provided that Tn is smaller than a number depending on 8. Similarly,
but more simply, by (9'), (11) and (13),
wi l l c21 <wilwlIT21all +IT22a21 +wi lwaI T13 aal
< I T22 aal + ~aTt2'
where
~2 = ~a(A, E);
and so
(16)
provided that 1'12 is small enough and 1'22 is near enough to 1. Similarly
(17)
if Ta3-1 and Tn are small enough. From (15), (16) and (17) we have
ICl Cacal < 81 "'1 aa aa I8.
Since e is arbitrarily small, we may put 8 1 =81 "'la2aal 8, where 81 is the
number in the enunciation.
This completes the proof.
19·
292 Automorphs

Note that we have used the full force neither. of Lemma 9 nor of
the inequalities (13).
The proofs of the following two corollaries maybe left to the reader.
COROLLARY1. Theorem IX remains valid i/lcl c2 c3 1<8l is replaced
by 0<1 c1 c2 cal <81'
COROLLARY 2. To every 82 >0 there is an '1/2>0 depending only on
A, 82 such that, i/
II-r - LII < '1/2
and one 0/ 1'12,1'13,1'21,1'23 is not 0, then there is a cE-rA with

o < ICl C2 Cal < 82' ICli < 1, Ie2 1 < 1.

Corollary 1 is proved in CASSELS and SWINNERTON-DYER (1955a).
A somewhat weaker form of Corollary 2 is in DAVENPORT and ROGERS
(1950a).
X.S.4. We now discuss
!p(:r) = Xl (X; + x~).
As in § 4.4 it is convenient to introduce the appropriate complex co-
ordinates
(i2 = -1).

A transformation Z =-rz corresponds to a real transformation for the

real variables :r if and only if it is of the shape

Zi = ~ Ti"z", (i)
where "
1'12 = Tn. Tn = Tal. Tn = Tn T23 = T32' Tn = Taa (2)
and the bar (-) denotes the complex conjugate.
THEOREM X. Let
!p(:r) = Xt (x; +~)
be proportional to integral and non-null on A and let
A = 1!p1 (A) = oEA
inf I!p(a)!. (4)
,*0

Then there are numbers '1/1>0, '1/2>0 with the following properties:
Suppose that -r is a homogeneous trans/ormation in the appropriate
complex co-ordinates given by (1) and (2) such that
1I-r-LII<'1/l' (5)
 Isolation theorems 293

Then
(i) II T12=TlS=FO, there lS a C=(Yl,Y2'YS)=F0, in complex co-
ordinates, in TA such that
(6)
(ii) II TSI=T21=FO, there is a C=(YI'Y2,YS)=FO in TA such that
(7)
By Theorem VII there is an automorph Z =wz in complex co-
ordinates of the shape

Define numbers T>O and X by

WI = P, W2 = T-Ie(x) , Ws = T-Ie( - X), (8)
where
e(x) = e2n'x.
If X were rational, say X =f4/V, the transformation WV would have two
equal eigenvalues w~, w~, which would thus be rational and so 1,

°
contrary to hypothesis (d. proof of Lemma 9). Hence X is irrational.
Thus by Lemma 8 with e = 1, there is a number 'Y/s> with the follow-
°
ing property: To every pair of numbers e> and 'P there are integers
u and v such that
lux+V-'PI<l (9)
and
-1 Tau
'Y/s < - <'Y/s· (10)
e
We now prove (i). Since IP (:1:) is proportional to integral on A,
there is an aEA of the shape
a=(tX),tX 2 ,0Cs). tX2=Ce(O), IXs=Ce(-O), (11)
where
tX1>o, C>O, A=tX1 C2
and A is defined by (4). Put
T12 = -ae('P). T13= -ae(-'P), (12)
where a> 0. Then a is small when liT - LII is small. We now choose
integers u and v to satisfy
lux+v-('P+O)i<l (13)
[d. (9)] and (10) with
(14)
294 Automorphs

so that
-2 2aC Tau
'YJa < Tn ell
<1- (15)

Since Tn is near 1, there are two constants 1'/, 'YJ", depending only on 1\
(and a), such that
0<1'/ a-I < T" <1'/, a-i. (16)
We shall show that the point
C ='t"w-ua = (YI'Y2,Ya) (17)

satisfies the conditions of Theorem X in case (i). In the first place,

T 2N IYII = lOCI Tn - T h aC{e(1J +tp - uX) + e(-1J- tp +UX)}lj

= lOCI Tn - 2 ThaC cos 2n(1J +tp - ux)i (18)
~ OCI Tn (1 - i 'YJa 2) < OCI (1 - ! 'YJa 2)
by (11), (13), (15), provided that 1I't" - LII is small enough. Further,
T-uiYal = T- u1Y21 ~ ITnl OCI T-au+CI Tnl +CI Tnl <C(1 + e) (19)
for any given e> 0, provided that 1I't" - LII, and so also a, is small enough.
From (18) and (19) we then have
IYIY2Yai < OCIC2 (1 -1 'YJa 2) (1 + e)2 < OCIC2 (1 - t 'YJa 2) = A(1 - t 'YJa 2) ,
if e was chosen suitably. Since (16) and (18) clearly imply IYll < 1 if
1I't" - LII is small enough, this completes the proof of (i) of the theorem
WI
·th 'YJ2=S'YJ3.
1-2

The proof of the second part is similar on considering 't"wUa with

suitable positive integer u. The details may be left to the reader.
For a later application we note the
COROLLARY 1. The numbers 'YJl an.d 'YJ2 may be chosen so that the con-
clusion 01 the theorem holds unilormly lor all lattices 1\ =A M, where M
is some lixed lattice on which g; (01)) is proportional to integral and non-
null and A runs through all automorphs 01 g;(0I)).
It is clearly enough to consider the case when Z =AZ is of the type
Z;=A;z;. Then w is an automorph of 1\ if it is of M. Hence the only
non-uniformity is possibly introduced by the point a. But clearly there
is a number R depending only on w, and so only on M, such that
Iw"al <R for some k. If w"a is taken for a, there is then complete
uniformity in the estimates.
COROLLARY 2. When 't" is any automorph 01 g; (01)) with
 Applications of isolation 295

then

We may suppose that A=1 and that a=r=(1,1,1). For any

integer u positive or negative we have
Icp('tw"e) I = ITllil T22 e (U X) + T2S e( - u X)12,
where X is given by (8). By Lemma 8, we may choose u so that
IT 22 e(ux) +T23 e (-UX)i
is arbitrarily near to II id -I i23ll, and the corollary follows.
Note that

dJ~~) = Iillllli2212 -li2s 121~ Iill I{I T221 -I T231}2,

with equality only when i 22 =O or T 23 =0, i.e. when 't is an automorph
of cp (or).
X.6. Applications of isolation. Following DAVENPORT and ROGERS
(1950a) we first use isolation to strengthen Theorem V. For XI(X~+X~)
it gives the best result to date, but for Xl X 2 Xs more is known, see Theo-
rem X of Chapter II, which is not proved in this book.
THEOREM XI. A. There is an- 111 > 0 such that every lattice A admis-
sible for

and with

is of the shape
A=twNI'
where t~ 1, w is an automorph 01 .Ai, and NI is delined in Theorem V.
B. There is an 1J2> 0 such that every lattice A admissible lor
.A'; : IXl (x~ + x~) I < 1
and with

is 01 the shape
A ='tw N2 ,
where N2 is delined in Theorem VB, w is an automorph 01 .A'; and 't
is a transformation Xi= l: TikXk with T12=T)3=T21=T31=O.
k
We first prove A by reductio ad absurdum. Suppose, if possible,
that 1JI does not exist. Then there exists an infinite sequence of admis-
sible lattices M, (1;;;;r<oo), none of the shape twN 1 , and such that
d(M,) -+ 7.
296 Automorphs

Now

by Theorem V, and since M, is ~ admissible; and so there is a sequence

of points

such that

On replacing M, by w, M" with a suitable automorph w, of ~, we may

suppose that
a, = (5,,5,,5,), 5,--+ 1
By MAHLER'S compactness principle, there is a convergent subsequence
of the M, which we may also call M" say
M,--+ M. (1 )
Then d (M) = 7 and M is ~-admissible, so is critical. Further, (1,1,1) EM
and so, by Theorem V, we have
M=SN1
where..s is an automorph of~. In particular, X 1 X 2 X a is integral on M.
But now
M, ='f,M
for transformations of, such that
1I'f, - 111--+ 0
Since M, is Ai-admissible, the transformation of, must be of the shape
'f,=t,~, for some number t, and some automorph ~, of Ai, by Theo-
rem IX, provided r is sufficiently large. T!lis contradicts the definition
of the M,. The contradiction proves Theorem XI A.
The proof of Theorem XI B is similar but using Theorem X instead
of Theorem IX. The details may be left to the reader. The only point
to notice is that if 'f and ware as enunciated in the theorem, then
'fW =w''f' for some w', of' with similar properties to wand 'f respectively.
COROLLARY TO THEOREM XI. B. To every s> 0 there is an"l ="la (s) > 0
such that every admissible lattice A for oN; with

d(A) < t(23)i(1 + "la) (2)

is of the shape A ='fW N2 , where of, ware as in the theorem and

1I'f-lll<s.
 Applications of isolation 297

We take 'YJa<'YJ2 for the 'YJa of the theorem, so that A ='twN 2 • We

may suppose that Tll>O and then, incorporating an appropriate auto-
morph in (0, that

Then
(4;)

where we use the appropriate complex co-ordinates for 't as in § 5.4.

But now
(5)
by Theorem X, Corollary 2 since A is .At;-admissible; and so, in particular,
ITuls-1 Tnl s < 1 +
{lTul-ITul}l 'YJa·
Hence if 'YJs is small, either ITnl/I Tnl or ITul/I Tnl is small; and we
may suppose the latter on incorporating in w, if necessary, the trans-
formation which interchanges X 2 and X 3 • We may further incorporate
in w a transformation of the type
XC+ X1' x 2 -e(x)xs , xs-e(-X)xa,
where e(x) =ebil. and X is chosen to make Tn real and positive. Then
from (4) and (5) we see that T 22 -1 and Tn are small if 'YJs is small.
Since Taa=Tu and T2a =T32' this proves the corollary by (3), and since
the remaining terms T;k are O.
X.6.2. The following interesting result about X1 X 2 X a has no analogue
for Xl (X: + x:), since it depends on the fact that 8 in Theorem IX may
be chosen arbitrarily. There is, however, a corresponding result for
~+x=-x:, see CASSELS and SWINNERTON-DvER (19SSa).
THEOREM XII. Suppose that lor some number D there are inlinitely
many lattices M, (1 ~ l' < (0), admissible lor
.Ai: IX1 x2 xal < 1,
with d(M,)~D; and such that no two, M', M", say, are 01 the shape
M"=twM', where t is a number and w an automorph 01.Ai. Then
there is a laUice A admissible lor .Ai with d (A) ~ D on which Xl X 2 xa is not
proportional to integral.
For the lattices M, have a convergent subsequence, say, without
loss of generality
M,-A (1'_00).
If X1 X 2 X 3 were proportional to integral on A, then by Theorem IX and
since M, is .Ai-admissible, we should have for all sufficiently large r
M, =t,w,A
298 Automorphs

for some numbers t, and some automorphs 00, of .Ai. This clearly
contradicts the hypotheses of the theorem.
As stated in § 1, it is unknown whether such a D or such a A exists.
X.7. An infinity of solutions. We now prove some results of DA-
VENPORT and ROGERS (1950a) about the existence of infinitely many
points of a lattice in certain point-sets with groups of automorphisms.
They prove more than we do here; the reader is referred to their inter-
esting memoire for the details.
The following trivial lemma gives almost all we need for the first
type of result.
LEMMA 10. Let n be some group of homogeneous linear transforma-
tions 00. Suppose that for every :1:*0 and every number r tnere is an ooEn
such that
100:1:1> r.
Then for every pair of numbers c, C with

O<c<C<oo (1 )
and every number r there is a finite set of elements 001, ... , 00", of n such
that
m~x 100;:1:1 >r (2)
1;:;i,S;".
for aU:I: in

This is a simple application of the HEINE-BoREL covering theorem.

The infinitely many open sets .r,(oo) of points z such that loozl >r
cover the compact set (3). Hence a finite covering may be selected from
the .r, (00).

*
THEOREM XIII. Let the boundedly reducible 1 star-body !/ have a group
n of automorphisms 00 such that to every z 0 and every r there is an
OOEn such that IOO:l:I>r. Then to.every integer k>O there is a bounded
set .9;. contained in !/ such that every lattice A with d (A) < LI (!/) has at
least k points in .9;. other than o.
That .9;, exists is equivalent to the statement that !/ is boundedly
reducible. We suppose .9;. has been found and deduce the existence of
.9;.+1' We may suppose without loss of generality that .9;. is the set
of points of !/ in some sphere
1:1:1 ~ C = Ck •
Further, there is a positive number Ck < C such that the entire sphere
(4)
1 For definition, see Chapter V, § 7.2.
 An infinity of solutions 299

is contained in!/'o We denote by g;,+l the set of points of !/' in

Ixl ~ CHI' (5)

where
CHI> max {C k , (k + 1) ck }
is so large that (5) contains all the sets wi l g;,(1 ~j~m), where the wi
are given by Lemma 10 with c = Ck and r = C = Ck. We must verify
that g;,+l has the required properties.
By hypothesis, if d (A) < A (!/') there are k points of A in g;, other
than o. If one of them, say a, is in Ix I < Ck' then all the points
la (1~l~k+1)

are in Ixl~CHI and in!/', so in !/'HI' as required. Otherwise, there

is a point b of A in g;,+l for which
c=ck~lbl~C=Ck·
Hence there is an automorph wi of the set ~, ... , w'" stich that
Iwjb I> C. Hence b ~ wi l g;,. But now, since Wj is an automorph,
we have
Idet wil =1,
and so
d(wiA) = d(A) < A(!/,).
Hence by the defining property of g;, there are k points of wi A in g;"
that is there are k points of A in W,-:-l g;,. These together with b give
k + 1 points of A in !/'HI' as required.
COROLLARY. When!/' is fully reducible I, the conclusions of Theo-
rem XIII continue to hold when d(A) =..1 (!/'), provided that A is not a
critical lattice of !/'.
For the existence of .9i is equivalent to the statement that !/' is
fully reducible, and the induction now goes as before.
When the star-body!/' is not boundedly reducible only slightly less
than Theorem XIII is true.
THEOREM XIV. Let!/' be a star-body and Al any number in
0< Al < A(!/,).
Then to every integer k there is a bounded star-body g;, (depending also
on AI) such that every lattice with d (A) ~ Al has at least k points other
than 0 in g;,.
We may suppose that !/' is open. Suppose, if possible, that for
every integer r there is a lattice A, with d (A,) ~ Al which contains no
1 For definition. see Chapter V. § 7.2.
300 Automorphs

point other than 0 of !l' in l;r I~ r. Then MAHLER'S compactness

theorem applies, and there is a lattice I\' which is the limit of a conver-
gent subsequence of 1\.,. Since d (I\') ~ Lll and I\' is !l' -admissible, this
contradicts the definition of Ll(!l'). The contradiction shows that ~
exists. The induction from Y,. to y"+l then goes exactly as for Theo-
rem XIII.
X.7.2. Where they apply, isolation theorems may give stronger
results than those § 7.1", as the following example shows.
THEOREM XV. Put
1P(;r) = Xl H + x~). (1)
There is a number 1)0> 0 such that every lattice I\. has one 01 the lollowing
two properties.
(i) there is a number t such that the set 01 xl-co-ordinates 01 tl\. is
identical with the set 01 xcco-ordinates 01 the critical lattice N2 01 lIP (z)1 < 1
occurring in the enunciation 01 Theorem VB, or (ii) lor every e> 0 there
is a point a =f= 0 01 I\. such that

IIP(a)1 ~ ---;-
(23)
(1 -1)0) d(l\.) , I~I < e. (2)

We will choose 1)0 later in the course of the proof. Suppose that (ii)
is false for some particular I\. and e. For integers r = 1, 2, ... , let 1\., be
the set of points (r2 Xl' r-l X 2 , r-l Xa), (Xl> X 2 , Xa) E I\. Then there IS a
convergent subsequence
(3 )
and M is admissible for
IXl (x~ + x~)1 <~(1 -1)0) d(I\.).
(23)

Hence by Theorem XI B, Corollary for any given eo we may choose

1)0 =1)0 (eo)
so small that
(4)
where 'r, ware as in Theorem XI Band t is some number. We take
for eo the number 1)1 which occurs in the enunciation of Theorem X
and its Corollary when M = N2 • By (3) and (4) we now have
(5)
for some Ok such that
IIOk - III < eo,
for all sufficiently large k. Clearly M, does not contain any points
c = (711,712,713) with 17111 < 1 and 17'17127131 < ~ (1-1)0) d (I\.) if r is suf-
(23)
 Local methods 301

ficiently large. Hence, by Theorem XI and its Corollary, if 1]0 is small

enough, there is a (/=(/k such that (/12=(/13=0 in the obvious notation;
indeed this happens for all sufficiently large k. But then, by (5) this
implies that (i) holds. This concludes the proof of the Theorem.
+
There is a similar result where Iall < e in (2) is replaced by a~ a: < e,
d. DAVENPORT and ROGERS (1950a).
X.S. Local methods. For many questions concerning indefinite
quadratic forms the appropriate tool is the theory of continued fractions.
We only mention the topic briefly here since the application to specific
problems not infrequently involves detailed calculation. Continued
fractions appear very naturally from the point of view of the geometry
of numbers. We sketch the connection here and refer the reader to
the author's Cambridge Tract [CASSELS (1957a)], where they are intro-
duced in a similar spirit 1 in a slightly different context, for a fuller treat-
ment and references. There a knowledge of the geometry of numbers
could not be assumed. For another account of the relationship of
continued fractions to quadratic forms see, for example, DICKSON (1929a).
Characteristic applications of local methods are MARKOFF'S original
treatment of his chain theorem (MARKOFF 1879a), [there is an account
in DICKSON (1930a); compJ.re Chapter II, §4)], the paper of BLANEY
(1957a) that will be dis(U;sed in Chapter XI, § 4, and the paper of
BARNES (1951 a). But applications are almost everywhere dense in the
literature.
Let us suppose for convenience that the 2-dimensional lattice 1\ has
no point except 0 on either axis. Then no two distinct points of 1\ have
the same xl-co-ordinate or the same x2-co-ordinate. There certainly
exist points ~o= (x10 , x 20) of /\ such that 0 is the only point of /\ in

IX11 < Ixlol, IX21 < Ix20 1·

Let ±~l = ± (X11' x21)=f=o be the points in IXll <Ixiol for which IX21
is least. Then there is no point except 0 in

IXll < IXIOI, IX21 < IX21I, (1 ')

and a fortiori in
IXll < IX11I, IX21 < Ix21 l·
We may then repeat the process with a; instead of ~o to obtain a se-
quence of points a;, ~2' • .•• Similarly we may start with ~o and inter-
change the roles of Xl and x 2 to obtain a sequence of points ~-1' ~-2' ....
There is thus a sequence of
(-oo<i<oo)
1 Which goes back to FELIX KLEIN (189Sa and 1896a).
302 Automorphs

such that there is no point of /\ except 0 in

IXII < lXIii, IX21 < IX2,i+ll· (1 )

Clearly a necessary and sufficient condition that a point YE/\ should
occur as ±zi for some i is that there should be no point of /\ except 0
in IxII < IYII, I x21< I Y21· Hence the sequence of pairs ± zi is completeiy
determined by I\, although the particular pair chosen to be ±zo is,
of course, arbitrary. If W is any automorph of Xl x 2 then the sequence
of pairs for w/\ is either ±wzi' if w does not interchange the axes of
co-ordinates, or ±wz_i (i.e. in the reverse order) if it does.
Since there is no point of /\ in (1) except 0, there is no point of /\
in the closed triangle with vertices 0, zi' zi +l except the vertices;
and so zi' zi+l is a basis of /\ for each i, by Lemma 6 of Chapter III.
We must now introduce an asymmetry between the XI - and x2-axes to
study the relationship between the various bases zi' zi+l' We choose
zi to be that point of the pair ±zi for which
(all i) . (2)
Then
(3)
since otherwise zi+l-zi would lie in (1). Since both Zi-l, Zj and
zi' zi +1 are bases, we must have

(4)
for some integer ai' Since
x2,i+l> x 2,i > X 2, i-I'
we must have
ai > O.
Then we must have the - sign in (4), since

IXI,i+l1 < IXlil < IXI,i-II,

and (3) holds for every i. Hence there is a sequence of integers aj> 0
such that
Zj+l - Zi-l = ajzi'
It may be shown that if two lattices have the same sequence of integers
ai then they are identical up to a transformation of the type

Further, to every sequence of positive integers aj there is a lattice.

Hence it is natural in 2-dimensional lattice problems about Xl X 2 to
consider not the lattice /\ itself simply, but the sequence ai' It turns
 Introduction 303

out that the behaviour of any particular basis, say :£J' :£1+1' of A is
influenced very strongly by the value of aj for j near to ] but only
very weakly by a;. for i remote from]. In many problems it is possible
to study the behaviour of only a few a j at a time. Hence the name
"local methods".
It would be interesting if local methods could be successfully extended
to problems in more than 2 dimensions, for example to problems relating
to Xl max (x:, x:), Xl (x~ + x~), x~ + x~ - x~ or Xl X 2 %3 • The difficulty is
not to find the analogues of the xi but to devise techniques to cope
with their interrelations. Continued fractions have however been
generalized to 2-dimensional lattices over a complex quadratic field,
i. e. substantially to certain special 4-dimensional lattices, see POITOU
(1953a) and the references there given.

Chapter XI

Inhomogeneous problems
XLt. Introduction. As previously, we say that points:£1 and:£2 are
congruent modulo A, written
:£1 = :£2 (A) ,
where A is a lattice, to mean that x 1 - :£2E A The set of points :£ con-
gruent to a given point :£0 modulo A is called a grid1 @: A will be called
the lattice of the grid and we shall call
d(@) = d(A)
the determinant of the grid. The characteristic inhomogeneous problem
of the geometry of numbers is to find conditions under which a grid
has a point in a given set 9'.
There is a wide variety of different problems. Thus one may be
concerned with all grids of given determinant d (@) or one may have
information about the lattice A Many of the fundamental techniques
for inhomogeneous problems are natural extension of those for lattices
[compactness theorems and so on; for bodies with automorphs see
SWINNERTON-DYER (1954a)J. For some specialized problems some
extremely powerful and delicate techniques have been developed which
would take too much space to discuss properly. Hence this last chapter
will have more the character ofa report and less that of a detailed
exposition.
1 Other terms are inhomogeneous lattice or non-homogeneous lattice.
304 Inhomogeneous problems

XI.l.2. The following simple result due to MACBEATH (1951 a) helps

to fix ideas.
THEOREM I. Let the set Y' have finite volume V(Y') and let e>O be
given arbitrarily small. Then there are grids @ with d (@) = e having no
point in Y'.
We may choose R so large that the portion of Y' in l:rl ~R has
volume < ie. Let A be the lattice with basis

b 1 = (4 R. o. o..... 0)
b2 = (0. 'Y}. O. O•...• 0)
b 3 = (0. o. 'Y}, O•...• 0) (1 )

b,. = (0.0.0, ...• 0, 'Y}).

where
(2)

Every point :r1 of space is congruent modulo " to precisely one point
of the parallelopiped
f!J: {Ylbl+ ... +Ynb,,}
(-j ~ Yj <j).

The volume of f!J is V(f!J) =d(") =e. by (2). If a point :r'=Lyibj

of f!J is congruent modulo " to a point ir1 in Iir11 ~ R, then clearly
IY~I ~i· Hence the set of points of f!J with this property has measure
at most j e. But now the set of points ir2 of Y' with Iir21 > R has volume
at most ie by construction; and hence so has the set of points ir" of
f!J which are congruent to at least such one point (compare the proof
of Theorem I of Chapter III). Thus the set of points of f!J congruent
to a point of [I' has measure at most ie + ie< e= V(f!J). There is
thus a point iroEf!J which is not congruent to any point of [1'. The
grid @ of all points congruent to iro modulo A clearly has all the proper-
ties required.
Xl.l.3. We shall mainly be concerned with star-bodies [I' defined
by a distance-function.
[1': F(ir) < 1. (1 )

For any lattice" and any point :ro we write l

(2)

1 So m (~o) = F(~o) in the notation of Chapter VII § 2.2, where ~o is the element
of the quotient space to which ~o belongs.
 Introduction 305

and
It (/\) = sup m (:ro,/\)' (3)
or,
Clearly
,u (t /\) = Itill (/\) (4)
for any t =F O.
The infimum in (2) need not be attained, though it clearly is attained
when the set F(~) < 1 is bounded. The function m (~o) need not be
continuous, but it is semi-continuous:

lim sup m (~) ~ m (~o) . (5)

.e-+,r,

Indeed given any 8>0 there is a point aE/\ such that

and then
F(~ + a) < m(:ro) + 8
for all ~ in a neighbourhood of ~o, by the continuity of F(;J!); so
m (~) < m (~o) + 8 in this neighbourhood. Again, when F(~) < 1 is
bounded, the function m (~) is readily seen to be continuous. The reader
will be able to supply the proofs of the positive statements just made
on the lines of the proof of the semi-continuity of the function F(/\)
in Chapter V, § 3.3. Examples to show that the infimum in (2) need
not be attained and that m (~) need not be continuous are provided in
2 dimensions for certain lattices /\ when F(~) = IXl X21l. This case has
implications in the theory of algebraic numbers and has been extensively
investigated both because of this and because of its intrinsic interest;
see BARNES and SWINNERTON-DYER (1952a, band 1954a) and BARNES
(1954a), where there are extensive references to earlier work. There is
some work on similar lines for IX1 X2X3 1l (n=3), but it has not been
carried so far, see DAVENPORT (1947c), CLARKE (1951 a) and SAMET
(1954a, b).
From the definition (2) it follows that m (~) may be regarded as
defined on the quotient space al//\ (compare Chapter VII). Since this
is compact, it follows from (5) that the supremum in (3) is always
attained; that is, there is an ~l such that

Of course the infimum in (2) need not then be attained for ~l=~O'
With unbounded sets F(~) < 1 there may be again a phenomenon of
successive minima; that is, it may happen that

sup m (:ro) < ,u (/\) .

m(.r.l *I'(A)
Cassels, Geometry of N urn bers 20
306 Inhomogeneous problems

Indeed some rather elaborate patterns of successive minima have been

found, see the papers of BARNES and SWINNERTON-DYER just quoted.
The quotient
{fl (A))"
(6)
d(A)

is unchanged on replacing A by tA, by (4). We shall write

b(F) = inf {fI(A)}" (7)
A d(A) ,

where possibly b(F) = O. If the set F(:r) < 1 has finite volume ~o', we
now show that
(8)
Let A be some lattice and e> 0 be arbitrarily small. There is a point
:rl congruent to any given point :ro and satisfying
F(:rl) < fl (A) + e. (9)
Hence the set (9) must have volume at least d (A). Since the volume
of the set of points :rl satisfying (9) is
{fleA) + e}" VF ,
the required result (8) follows.
We shall show in § 3 that if the body F(:r) < 1 is bounded, the
infimum in (7) is attained; that is there is a lattice M such that
{p (M)}" = b (F) d (M) .
We shall treat the estimation of b (F) for convex distance-functions
F in § 2 where the relevant literature will also be discussed.
When VF = 00 it is, of course, still possible that b (F) > O. In par-
ticular, DAVENPORT (1951 a) showed this to be the case for the 2-dimen-
sional distance-function
(10)
His estimate,

was improved to
b(F) ~ 451'2

by the author [CASSELS (1952a)], with a probably simpler proof. This

has recently been improved by ENNOLA (1958a) to
b(F)~(16+6~)-I= 30.!, ... ,
by a modification of DAVENPORT'S original method. On the other hand,
Miss PITMAN (1958a) has shown that
b(F)~~2
 Introduction 307

More recently I, she has obtained an even smaller upper bound for
b(F).
The problem of determining b (F) for F given by (10) is closely related
to the problem of determining the real quadratic numberfields with a
Euclidean algorithm. DAVENPORT extended his work to number-fields
of two other types corresponding to
P=XI(X~+X~) and p= (x~+x~)(x~+x:).
These results were proved by the author [CASSELS (1952a)] much more
simply and with a better estimate of b(F).
HLAWKA (1954c) has generalized these results to any distance-
function F(;x:) in n variables which may be put in the shape

{F(;x:)}" = {F, (Xl' "., x,)}, {F,,_, (X'+l' "., x,,)}"-',

where F" F" _, are r- and (n - r)-dimensional distance-functions such
that the star-bodies F, (;x:) < 1 and F..-, (;x:) < 1 are bounded. We do not
prove these results here. A closely related problem is treated in the
author's tract [CASSELS (1957a) Chapter V, § 6], where there are further
references.
In general it appears to be a difficult problem to decide whether
b (F) = O. Thus it does not appear to be known whether this happens for 2
F(;x:) = Ix~ + x~ - xW n=3
or
n =3.
XI. 1.4. It follows at once from MACBEATH'S Theorem I that
~(F) = sup {p(AW
A d(A)

IS 00 whenever VF<oo. In § 4 we shall be concerned with ~(F) for

F = Ixl ... x"I I !".
It was conjectured by MINKOWSKI that ~(F) =2-", but this has been
proved only for n =2,3,4. We shall give references and a further
discussion in § 4. We shall also give a result of CHALK about the set
XIX2."X,,~1 Xj>O (1~i~n)

(not a star-body!) and quote other work about sets defined in term
of Xl'" X".
1 I am grateful to Miss PITMAN for allowing me to refer to this unpublished work.
now published. Acta Arithmetica 6 (1960). 37-46.
2 The first case has been settled by E. S. BARNES [J. Austral. Math. Soc.
2 (1961/62) 9-10]. who shows that b(F) =0.
20·
308 Inhomogeneous problems

The value of ~ (F) for

F(3!) = Ix~ + x~ - xW (n = 3)
has been found by DAVENPORT (1948a) who showed it to be isolated
and the investigation of the successive minima was carried further by
BARNES (1956a). More recently BIRCH (1958a) has found ~(F) for
F(3!) = Ix~ + ... + x~ - x~+l - ... - x~,li (n = 2r),
for all r~ 2. Estimates for
F(3!) = Ix~ + ... + x~ - X~+l - ... - x!ll
with r>O, n-r>O have been given by BLANEY (1948a) and improved
by ROGERS (1952a) and Miss FOSTER (1956a). All the work just
described is of course equivalent to finding the best possible constant
rj"s such that
sup inf I/(u+uo)l~rj"sIDII/"
U o real U integral

for all indefinite quadratic forms I of signature (r, s) with r +s =n and

with determinant D. We shall not discuss this work further in this
book but refer the reader to the original memoires.
XI.1.S. For some functions F(3!) there are inequalities, valid for
all 1\, connecting
p (1\) = sup inf F(3!)
"'. ",=",.(A)
and
F(I\) = inf F(x)
"'EA

or, more generally connecting fl (1\) and the successive minima of F(3!}
with respect to I\. When F(3!) is convex, there are further relations
with the corresponding quantities for the polar distance-function F*(3!}
and the polar lattice 1\*. These relations go under the general name of
transference theorems l (Obertragungssatze). Thus DIRICHLET'S hexagon
Theorem VII of Chapter IX may be regarded as a very precise trans-
ference theorem for Ix~ + xW.
We shall discuss transference theorems
for convex functions F(3!) in § 3. Much interesting work has been done
on transference theorems for the non-convex F(3!) defined by
{F(3!)}" = II (X~+k + x~+s+k),
Ixl ... x,1 l;i;k;i;s
where n=r+2s, but here we can only refer the reader to the paper
of DAVENPORT and SWINNERTON-DYER (1955a), where references are
given to earlier work. There is a striking related result in SWINNERTON-
DYER (1954a).
1 Presumably because information is transferred from one problem to another.
 Convex sets 309

There is a further type of result which may most appropriately be

mentioned here since they are transference theorems of a sort. BARNES
(1950a) showed that if

and if A has the basis "t, a2 then

2f1(A) ~ max {F(a 1) , F(a 2 ) , minF(al ± a2 )}.

±
Other results of this general kind are known, see BAMBAH and K. ROGERS
(1955 a) and the references given there. In particular, K. ROGERS (1953 a)
showed that BARNES' result is true for all distance-functions F(~) such
that F(~) < 1 has the same general appearance as I Xl x2 1< 1. The proofs
are all elementary and tend to involve a tedious splitting of cases. We
do not discuss them further in this book.
XI.2. Convex sets. In 2 dimensions the problem of finding 11 (F) in
the notation of (7) of § 1.3 for convex functions F is completely solved
by the following result [BAM BAH and ROGERS (1952a)].
THEOREM II. Let Y' be a closed 2-dimensional convex set and .11 some
number. A necessary and sufficient condition that there exist a lattice A
with d (A) = .11 such that every point is congruent modulo A to a point
of .9' is that there exist a convex hexagon 1 .?If inscribed in .9', which is
symmetrical about some point and has an area V(.?If) =.1 1.
Note that Y' is not required to be symmetrical about any point.
Suppose, first, that .?If exists. We may take the centre of .?If as
origin o. Let A be a critical lattice for 2.?1f. Then d (A) = V(.?If) = .1 1,
by Lemma 13 of Chapter V. Hence by Theorems II, III of Chapter IX
applied to 2 Jt', and since Jt'is closed, every point is congruent modulo A
to a point of .?If, and so of Y'.
Suppose now that there exists a A such that every point is congruent
modulo A to some point of Y'. If Y' is unbounded, there is clearly
nothing to prove, so we may suppose without loss of generality that Y'
is bounded. We shall construct the hexagon.?lf in stages. Suppose, first,
that there is ana=l=oEA such that Y'andY' +ahaveinnerpointsincom-
mono By taking 2s a with suitable integer s ~ 0 instead of a, we may sup-
pose without loss of generality that Y' + 2a and Y' have no inner points
in common. Then there exist points c and d on the boundary both of Y'
and Y' + a such that the portion of the boundary of Y' between c
and d (taken in an anti-clockwise direction, say) lies in Y' +a and the
portion of the boundary of Y' + a between d and c lies in Y'. Then
c - a and d - a are common to the boundaries of Y' and Y' - a. Let
1 A parallelogram being allowed as a degenerate hexagon.
310 Inhomogeneous problems

~ be the portion of f/' lying between the line joining C and d and the
Jine joining C - a and d - a. and taken closed; i.e. including the points
of f/' on those lines. Then clearly ~ is convex and every point of the
plane is congruent modulo A to a point of ~. After a finite number
of steps (since f/' is bounded) we obtain a closed convex set .r ( f/'
such that every point is congruent modulo A to a point of .r but no
two sets.r and.r +a. aEA have inner points in common. Then every
boundary point of .r is also a boundary point of .r + a for some a ==1= 0
in A. Since.r and .r +a are convex. this common boundary is either
a point or a line-segment. Since.r is bounded. only a finite number
of a come into consideration. and so .r is a convex polygon. We must
now show that it is symmetric about some point. Let the vertices of
.r be cl ..... c.... where the line segment cici+l is the common boundary
of.r and.r +ai • aiEl\. Then the line-segment (ci-a j ) (Cj+l- aj) is the
common boundary of .r and .r - a;. Hence m is even. m = 2l. and

Hence
t(C; + Ci+/) = t (C;+l + ci+l+ /)

for each i. so e =t(cj+c;+I) is independent of i. Clearly.r is sym-

metric about e.
We may suppose without loss of generality that e =0. Then A
gives a lattice packing of .r (or. more precisely. of the interior of .r)
and every point is congruent to some point of .r modulo I\. Hence.r
is a hexagon by Theorems II and VI of Chapter IX. This concludes the
proof of Theorem II.
Using known results about hexagons inscribed in convex sets. BAMBAH
and ROGERS (1952a) deduce in our notation (§ 1.3) that

1;£ ~bW);£ i

for a convex 2-dimensional distance-function F inequality and the

stronger inequality

f F is symmetric. The equalities on the right-hand side are attained

when F(z) < 1 is a triangle and a circle respectively. The left-hand
inequality. which is valid whether F is convex or not. was obtained
in § 1.3.
There is a theory of lattice coverings and non-lattice coverings which
is closely analogous to the theory of packings discussed in Chapter IX.
 Convex sets 311

For details in 2 dimensions see FEJES T6TH (1950a and 1953a) and
BAM BAH and ROGERS (1952a).
Not much is known about b(F) in more than 2 dimensions. When
F(x) < 1 is the unit 3-dimensional sphere, the precise value has been
found by BAM BAH (1954 b), and other proofs have been given by BARNES
(1956b) and FEW (1956a); but all proofs are fairly complicated. The
4-dimensional sphere has been considered by BAMBAH (1954a), who
obtains an estimate for b (F) and gives a conjecture for the correct
value. Estimates for b(F) above and below and also for the corres-
ponding number for non lattice coverings have been obtained for
n-dimensional spheres, see BAMBAH and DAVENPORT (1952a), DAVEN-
PORT (1952b) and WATSON (1956a) for the lattice case, and ERDOS
and ROGERS (1953 a) and ROGERS (1957 a) for the non-lattice case, the
last treating general convex sets. Very recently ROGERS (1959a) has
obtained much stronger results by more powerful methods.
XI.2.2.1 ROGERS (1950b) has given an elegant proof of the following
result relating b (F) to the function
c5(F) = sup {F(A)}"
A d(A)
introduced in § 4 of Chapter IV.
THEOREM III.
b (F) ~ r" 3,,-1 c5 (F)
for all symmetric convex n-dimensional distance-functions which vanish
only at the origin.
ROGERS (1950b) also proved a similar result for non-lattice packings
and coverings, and indeed with the smaller constant 2-1 instead of
2-"3,,-1. Before proving Theorem III we note the following
COROLLARY.
VFb(F) ~ 3"-1,
where VF is the volume of F(x) < 1.
For VF c5 (F) ~ 2" by MINKOWSKI'S convex body theorem.
ROGERS proves Theorem III by considering a critical lattice M for F,
that is
F(M) = 1, d(M) = {c5(F)}-1. (1 )

We use the notation of § 1.3; in particular

m (xo) = inf F(x) .
or 5 or.(M)

1 When n is at all large, the results of this section are superseded by ROGERS
(1959 a ).
 I
312 Inhomogeneous problems

As was shown in § 1.3, there is then a point :r1 such that

m (:r1) = sup
al,
'In (:ro)

= fl(M) (2)
= fl (say).
Then
m (3 :r1) -;£ fl '
and so, since F(:r) < 1 is bounded, there is an aE M such that
F(3:r1 - a) = m (3:r1) ~ fl.
Then
(3)
and so 1a is not in M.
Let A be the lattice of points

b + ; a, bEM, r = integer,
so
d(A) = td(M).
Hence
{F(A)}" -;£ 15 (F) d (A) = 115 (F) d (M)
by the definition of 15 (F) ; that is, there exists a point b + ...':-a+o of A
such that 3
{F(b + ; a)}"-;£; 6(F)d(M). (4)

We may suppose without loss of generality that r=O or ±1. If '=0,

we have b+o, and so

I
F(b) ~ F(M) = 1,
and (1) and (4) are in contradiction. Hence r= ±1, and

F(b ± 1a) = F {b ±:r1 =t= (:rl -la)}

~ F(b ± :r1 ) - F(:rl -la)
(S)
~,u-1,u
= ifl,
by (2) and (3).
On substituting (5) in (4) we obtain

L-;;;;
d(M)
r"3,,-115(F)
.
(6)

Since the left-hand side of (6) is at most b (F), by the definition of b (F)
as an infimum (§ 1.3), the theorem follows.
 Transference theorems for convex sets 313

XL3. Transference theorems for convex sets. In this section we

consider for a symmetric convex n-dimensional distance function F which
vanishes only at 0 the relationships between the function
fl = fl (1\) = sup inf F(x) (1)
=
or, or or, (J\)

discussed in § 1 and the successive minima AI' ... , All of F with respect
to 1\ which were discussed in Chapter VIII.
We first prove the inequality
(2)
Let b1 , ••• , b" be any basis for I\. Then by the definition of fl and the
fact that F(x) < 1 is bounded, there are vectors cjEI\ such that
F(t b j - c j ) ~ fl.

Hence the vectors d j = b j - 2cj all satisfy

F(d j ) ~ 2fl·
Since the d j are linearly independent, as is easily seen 1 by considering
congruences modulo 2, the left-hand side of (2) follows.
We now prove the right-hand side of (2). There are linearly in-
dependent vectors aj of 1\ such that
F(u,) = Aj •
Every vector Xo is thus of the shape
xO=~lal + ... +~lIall
for some real numbers ~1"'" ~". Put

where
IUj-~jl ~t,
and U 1 , ... , u" are integers. Then, clearly,
F(xo - a) = F {L (~j - ui ) aj}
I

~ L I~i -
j
ujl F(a j )

~t LF(a j )
=t LA j •

1 For suppose that ~ rj d j = 0, where the 1'; are integers which, without loss
j
of generality, may be supposed to have no common factor. Then ~ I'j b j = 2~ I'j ('j'
j
Since the b j are a basis, all the I'j must be even. A contradiction!
314 Inhomogeneous problems

This proves the right-hand side of (2).

On making use of the inequalities

of Theorem V of Chapter VIII. we may deduce estimates for p above

and below in terms of
Al = inf F(a} = F(A}.
"*0
Ell

From the left-hand sides of (2) and (3). and since

1.1~ A2~ ... ~ I.... (4)
we have
(5)

On the other hand. the maximum of Al + ... + I." for given Al and pro-
duct At ... ).,. is clearly attained when Al = 1.2= ... = 1... - 1. Hence. by (2)
and (3).
(6)

Both the inequalities (5) and (6) may be improved. The problem
of obtaining an estimate above for p in terms of Al is an old one which
has been attacked by many methods. The latest result due to KNESER
(1955a) and BIRCH (1956a) will be proved as Theorem V. The inequality
(5) has attracted much less attention. We sketch a proof of an improve-
ment due to BIRCH (1956b). as Theorem IV. BIRCH actually proves
something slightly stronger than Theorem IV and gives examples to
show that it cannot be much further improved.
THEOREM IV.

for convex symmetric n-dimensional distance-functions.

BIRCH'S proof is very simple. We may suppose after a suitable
homogeneous linear transformation that A = Ao is the lattice of points
with integer co-ordinates. and that
F(O • ...• O. 1} = AI.
Let ~ be the (n -i}-dimensional projection of the set
fl!/: F(~) ~ p

on to the hyperplane x" = o. Then every point with x.. = 0 is congruent

modulo Ao to a point of ~. so ~ has (n -1}-dimensional volume
 Transference theorems for convex sets 315

v.. -1 (ff) ~ 1. Further, fl!/' contains the points

± (O, ... ,0, fl!AI).
Some elementary geometry 1 now shows that the volume of fl!/' must
be at least
V{u!/') ;;:::: ~. £ . v: 1 (ff);;:::: 2p.
r - n Al ,,- - n~ •

Since V{fl!/') =fl"VF , and since we have assumed that A=Ao, so

d (A) = 1, the truth of the theorem follows.
THEOREM V. Let
Q= 2"~(1\) =q+" (7)
Al VF
where q is an integer and 0;;;;,,<1. Then
fl ;;;; tAl (q + ,,11") . (8)
Further
(9)
provided that Q~ n.
Note that q +,,1/"~ Q and q~ 1 by (3). The inequality (8) is
KNESER'S (1955 a) 2 and (9) is BIRCH'S, though the remark that (9) holds
already for Q~n is KNESER'S [see BIRCH (1956a)]. BIRCH proves similar
results involving other minima A2 , ••• , A"-I.
Before proceeding to the proof we note that (9) cannot be further
improved 3. Let
F(~) = max {lXII, ... , Ix"I},
and let A be the lattice of points
(~, .•• , Un_I, Qu,,),
where Q is any number ~ 1 and u1 , ••• , u" run through all integers.
Clearly
Al=1 VF =2";
and so Q is in fact the number given by (7). Further, fl =tQ, as is seen
by considering
~o = (0, ... , 0, t Q) .
1 The details are given in the author's tract [CASSELS (1957a)] page 84 Lemma 1.
The easiest way is to replace f/': F(z) < 1 by a body of the same volume symmetric
in x,,=O. on replacing for each (Xl' ...• x n - I ) the segment of xn such that
(Xl' ...• X,,) Ef/' by the one of equal length symmetric in x,,= 0 (STEINER sym-
metrization). The result is trivial for the symmetrized set.
2 Professor KNESER tells me that he can show that < can be substituted
for ~ in (8) except when Q is an integer.
3 BAMBAH (1958a) shows that (8) and (9) may sometimes be improved if i5(F)
is known.
316 Inhomogeneous problems

It was long conjectured that (9) was valid for all Q, but the following
example, due to KNESER and BIRCH (see BIRCH 1956a), shows that in
fact the weaker inequality (8) cannot be improved for 1 ~ Q< 2. Let

and let A be the lattice of points

where O~ e< 1 is fixed and u l , ... , u" runs through all integers (note
the change of sign in the last co-ordinate). Then

d(A)=1+e", Al =···=A,,=1, p,=t(1+e},

as is readily verified. No case appears to be known when (9) is false
and Q~2.
Now to the proof of Theorem V. We work in the quotient space
Bt/A and use the notation of Chapter VII and of Theorem IV of Chap-
ter VIII. In particular, we denote by S(t) the set of points t) of Bt/A
which have representatives Y in Bl such that F(y) < t. By Theorem IV
of Chapter VIII the measure m{S(t}} satisfies

= t"lj- if t ~ tAl. (10)

{
m{S(t)} -:?t(.!A)"-lV,
- II 1 F
if .!A~t~.!A.
II 1 - - II 1\ (11)

We shall also need the inequality

for any 1x~0, t2~0. This follows at once from the "Sum Theorem",
Theorem I of Chapter VII. Indeed, S (Ix +t 2) contains the sum S (Ix) + S(t a),
where addition of sets is as defined in § 3 of Chapter VII, since
F(Yl+YZ)<Ix+ts if F(Yl) <Ix and F(Ya)<ta.
We also remark that p, is the lower bound of the numbers t such
that m{S(t)}=d(A). Clearly m{S(t)}=d(A) if every point of Bl is
congruent modulo A to a point ~ with F(~)<t. Conversely, suppose
that m{S(to)}=d(A). Let e>O be arbitrarily small. Then m {S(e)}>O
by (10). and so every point of Bl/A belongs to S(to)+S(e) (S(to+e) by
the first part of the "Sum Theorem" I of Chapter VII.
We now prove (8) very simply. By (10) we have

m{S(tAl)} = (tA.S'VF = Q-1d(A)

and
 Transference theorems for convex sets 317

Hence, by repeated use of (12), we have

m [5 a;'l (q +"lin)}] ~ q m {5 (t;'l)} + m {5 (t ,,11" ;'1)}

= (q + X) Q-ld(A)
=d(A),
as required.
To prove (9) we need (11) as well as (10), where now
Q~n.

We must distinguish two cases. Suppose first that

QAl~ nAn'
Then 2/-l;S Al Q by (2), which proves (9) in this case. Otherwise, by (11),

m {5(~ .~~)} ~ ~ (iAit V F = d(A)jn,

by the definition (7) of Q. Hence, by repeated use of (12), we have

m {5 (tQ AI)} ~ d (A),

which completes the proof of (9).

XI.3.2. We are now in a position to prove the result enunciated
in § 1. 3 that when the star-body F(x) < 1 is bounded, then b(F) is an
attained minimum, that is, in the notation of § 1. 3, there exists a lattice
M such that
{,u (MW = b (F) = inf {,u (t\W
d(M) /\ d(t\) •

We must use the transference theorem of § 3.1 to ensure that we

may apply MAHLER'S compactness criterion. Write

so that
F(x) ~ cFa(x) , .c> 0

for some c and all x, since F(x) < 1 is bounded. Hence clearly
/-l(O)(A) ;S c-l/-l (A) ,
where the superfix (0) indicates that the quantity is relative to Fa. In
particular, if /-l (A) is bounded above for some set 2 of lattices A, then
so is ",(0) (A); and hence AiO) is bounded below a strictly positive number
by Theorem IV [or by the weaker inequality (5) of § 3.1]; that is
inf lal
IAI = aE/\
,*,0
is bounded below.
318 Inhomogeneous problems

Now we select a sequence of lattices /\T (1 ~ r < 00) not necessarily

distinct such that
d(/\T) = 1
and
{,u (/\T)}n --* b (F) .
From what was proved in the last paragraph, 1/\,1 is bounded below
by a positive number. Hence by MAHLER'S compactness criterion there
is a convergent subsequence; without loss of generality
/\, --* M.
Then M clearly has the properties required.
XI.3.3. Let /\ and /\* be polar lattices in the sense of Chapter I,
§ 5. It was there shown that a necessary and sufficient condition that
a point x belong to /\ is that the scalar product xa* be an integer for
all a*E/\*. We develop now what may be regarded as a quantitative
generalization of this statement. For a real number; we denote by
11;11 the difference between; and the nearest integer either above or
below taken positively, that is

II; II = m=O, ±1,

inf
±2, ...
I; - mi·
There will be no possibility of confusion with the notation II 't II where 't
is a homogeneous linear transformation.
THEOREM VI. Let F(x) be a symmetric convex n-dimensional distance
function corresponding to a bounded set F(x) < 1 and let F*(x) be the polar
distance-function. Let /\ and /\* be polar lattices. For any point xo write
(1 )
and
K(x) = sup II a* xoll (2)
° a"EN F*(a*) ,
~o

where a* xo denotes the scalar product. Then

0)
The precise values of the constants in (3) are immaterial: what
matters is that the ratio K(xo)/m (xo) lies between constants. Theorem VI
goes back in essence to KHINTCHINE (1948a). KRONECKER'S Theorem
follows from it in a few lines [compare Chapter V, § 8 of the author's
tract (CASSELS 1957a), where a less general form of Theorem VI is
given].
 Transference theorems for convex sets 319

We first prove the right-hand side of (3). Let

:1)1 == :1)0 (A) . (4)
Then :1)1 a* differs from :l)oa* by the integer (:1)1 - :1)0) a* and so

II:I)oa* II = 11:1)1 a* II ~ 1:1)1 a* I· (5)

But now, by the definition of a polar function (Theorem III of Chap-
ter IV), and since F(:I)) is symmetric, we have
(6)
Hence
(7)
and so
!1:1)0 a* II ~ m (:1)0) F*(a*) , (8)
on taking the infimum of the right-hand side of (7) over all:l)1 == :1)0 (A).
This is just the right-hand side of (3).
To prove the left-hand side of (3) we need the dual bases b i and b;*
of Theorem VII, Corollary of Chapter VIII, for which
F(bi)F*(b;)~(t)"-I(n!)2 (1~i~n). (9)
Let :1)0 be any point, so that
:1)0 = ~1 b1 + ... + ~"b"
for some real numbers ~i' Then, by (2),
II~ill = lib! :1)011 ~ K(:I)o) F*(bf) (10)
for 1 ~i ~ n. Choose integers ui so that
IUi-~il =1I~ill. (11)
and let

so

Then by (9), (10) and (11),

m (:1)0) ~ F(:I)1)
~ Lj I~i - uil F(b j)

= L II~iIlF(bj)
j

~ K(:I)o) L F*(bf) F(b j )

j

= 2"~1 (n !)2 K(:I)o).

which is the left-hand side of (3).
320 Inhomogeneous problems

XI.3.4. In this section we prove a rather specialized transference

theorem which we shall need in § 4. The proof uses the so-called tech-
nique of the additional variable which has often been used with success l .
For example, the best result in the direction of Theorem V until the
work of KNESER was proved by HLAWKA (1952a) using this technique.
[It is reproduced in the author's tract (CASSELS 1957a) in a special case.]
LEMMA 1. Let Fa (~) = I~ I, where ~ = (Xl' X2 , x 3) is a 3-dimensional
vector. Let AI' A2' A3 be the successive minima 0/ a lattice A with respect
to Fa and let
It = sup inf
z, z=z.(A)
Fa (~) .
Then
(1 )
and
(2)

We first prove (2). There are linearly independent points a j of A

I
such that Ia,l = Ai' Let c l , c2 , c 3 be a set of mutually orthogonal
vectors such that

a1 = cl
a2 = V2l c1 + c 2 0)
a 3 = V 31 c1 + V32 C 2 + c3

for real numbers vii' Then

(1~i~3)· (4)

But now, if ~o is any point, it is possible to choose integers "3' "2,"1

successively in that order, so that

~1 = ~o +"1 a1 + "2 a2 + "3"a = ~1 C1+ ~2C2 + ~3 c3 ,

where the numbers ~i satisfy

(1~i~3)·
Hence
I~~I = ~~ Ic112 + ... +~: Ic312 ~ i (A~ + ~ + A:)
by (4). This establishes (2).
We now construct a 4-dimensionallattice M as follows. There is a
point ~o such that
fl = fl (A) = inf 1~ I· (5)
z=z.(A)

1 Apparently first used by MORDELL (1937a).

 Transference theorems for convex sets 321

Let the number e be defined by

e2+,tt2 = At (6)
so
e ~ lA.a. (7)
by (2). Then M is the set of all 4-dimensional points
X = (a:. eu). (8)
in an obvious notation. where u runs through all integers and the vector
a: satisfies the congruence
(9)
Clearl}
d(M) =ed(A).
If XEM and u*,0 we have
IXI2 = 1a:12 + e2u2 ~ A=
by (5) and (6) or by (7) according as u = ± 1 or lui> 1. The values
taken by IX I with u = 0 and X EM are precisely those taken by Ia: I
with a:EA Hence the four successive minima of the function IXI with
respect to M are AI. A2 • A.a. A4 • where AI. A2 • Aa are the minima of 1a:1
with respect to A. as already defined. and
A4 ~ A.a.
(Indeed A4=A.a. but we do not need that.)
By Theorem I of Chapter VIII and Theorem IV. Corollary of
Chapter X. we have
AI A2A :;£ AI A2Aa A4j
;;£ 14:01 d (M)
(10)
= 2d(M)
= 2e d (A).
r..
where 0 is the lattice-constant of the 4-dimensional sphere IX 1< 1. On
eliminating e between (6) and (10). we obtain the required inequality (i).
We shall actually need Lemma 1 in the following shape:
COROLLARY. To every point a:o there is a point a; =a:o (A) such that
1 ""'l
~_12:::;;:~A2{~}i
- 4 3 Al A2 A3 •
(11)
In the first place.
3e + e-3 = e + e + e + e-3 ~ 4 (12)
for every number e> 0 by the inequality of the arithmetic and geometric
mean. Hence it follows from (1) that 1'2 is at most equal to the
Cassels, Geometry of Numbers 21
322 Inhomogeneous problems

righthand side of (11). on using (12) with

e= {
d(l\)
).1).2).3
}i •

But now. since Izl < 1 is bounded. there is certainly an Z1 such that
int Izl ~ p.;
IZll = z_z,
and the corollary follows.
XI.4. Product of n linear forms. Let
~ (z) = Ixl •.. x"I I ,,,. (1 )
As in § 1.4 we put
f-ll (A) = sup int ~ (z) • l) = sup (PI (I\))" (2)
z, z ..z,(A) 1 A d(l\) •

There is a famous conjecture of MINKOWSKI that

l)l = 2-".

That ~ ~ 2-" follows at once by considering the case when A =Ao

in (2) is the lattice of points with integer co-ordinates andzo= (t ..... t).
Clearly then ~(zo) ~t for all z =
zo(I\,). and d(Ao) =1-
It is well known that
{p.I(A)}" ~ 2-"d(A)
if A is a sublattice of the integer lattice 1\,. The proof is simple. The
lattice A has a basis
b; = (bl ; •.•.• bi; • O•...• 0) •

where the bii are integers and bi;=I=O. bi;==O for i>i. For any real
numbers (XlO •.•.• x"o) we can thus choose integers u l • •..• u". in order.
so that
Iu;b jj + ... +u"b; .. +xiol ~jlbiil·
For zl=Utbl +··· +u"b,,+zo. we then have

as required.
The conjecture (3) has been proved only for n = 2. 3. 4. A great
many proofs of (3) for n = 2 for have been given; we shall present one in
§ 4.2 due to SAWYER. This has the advantage that it gives naturally a
result for the .. asymmetric" distance function I
1
1\ ,(z) = { k IXI x2 1i if Xl x2 ~ O} .
, llx1 X 2 1 if X1X2~0
10f course Fk,/(a:) < 1 is symmetric about 0; but it is not symmetric in the
four quadrants.
 Product of n linear forms 323

where k and l are positive numbers. These arise quite naturally even in
originally symmetric problems; indeed the result we shall prove was
first obtained by DAVENPORT (1948a) as a tool in his work on the "sym-
metric" problem for indefinite ternary quadratic forms. Further results
about F;. I have been obtained, notably by BLANEY (1950a), BARNES
and SWI~NERTON-DYER (19S4a) and, as an adjunct to anotherinvestiga-
tion, by BARNES (1956a). We refer the reader to these papers for further
details.
When n = 3, MINKOWSKI'S conjecture (3) was proved by REMAK
(1923 a, b) and a simplified proof was given by DAVENPORT (1939a).
We give DAVENPORT'S proof in § 4.3, having already paved the way in
§ 3.4. A proof for n = 3 using different ideas has been given by BIRCH
and SWINNERTON-DYER (1956a).
When n =4 a proof of (3) has been given by DYSON (1948a) following
the same general line as REMAK'S proof. It is an extremely powerful
piece of work and requires tools from topology as well as from number-
theory proper.
For n>4 only estimates for '1)1 are known. It was shown by TSCHE-
BOTAREW (1934a) that

and this was improved by MORDELL (1940a) and by DAVENPORT (1946a)

to

where 'YJ .. s a number <1 such that 'YJ..---*(2e-1J-1 as n---*oo. Recently

WOODS (1958c) has shown that TSCHEBOTAREW'S result may be im-
proved simply by using BLICHFELDT'S theorem instead of MINKOWSKI'S
convex body theorem. MORDELL (1959a) remarks that this improvement
can be combined with the earlier techniques. In particular, DAVENPORT'S
'YJn can be replaced by a number which is asymptotically t'YJ .. for large 'YJ.
We give TSCHEBOTAREW'S result with its impressively simple proof in
§ 4.4.
Some further results of a general nature are known about this
problem. BIRCH and SWINNERTON-DYER (1956a) have shown that

for all lattices 1\ in a certain neighbourhood of the integer lattice 1\0'

and give some other facts relating to the general conjecture. The author
(CASSELS 1952b) has shown that for any 8>0 and every n there are
infinitely many lattices 1\ such that

21*
324 Inhomogeneous problems

and such that no two lattices A, A' of the set are of the type A' = t CJ) A,
where t is real and CJ) an automorph of 1\(~); so if MINKOWSKI'S con-
jecture is true then the first minimum is certainly not isolated. ROGERS
(1954c) has investigated the least number ,u~ (A) such that for every
e> 0 and every ~o there are infinitely many solutions of
1\ (~) < ,u~ (1\) + e
and obtained general conditions for 1\ under which ,u~ (A) =,udl\).
CHALK (1947a, b) has obtained the complete answer for what may
be regarded as an extreme asymmetric version of MINKOWSKI'S problem.
He shows namely that for any lattice 1\ and any point ~o there is an
= =
~1 (XlI' .•. , X" 1) ~o (1\) such that

(1~i~n), (4)
X11 ••• X"l ~ d (1\). (5)
That ~ in (5) cannot always be replaced by < is shown by the simple
example when 1\ =1\0 is the lattice of integer vectors and ~o=o. The
case n =2 was obtained by DAVENPORT and HEILBRONN (1947a). When
n=2, BLANEY (1957a) has given an interesting strengthened form:
namely that for every ~o there is an ~1 = (xu, x2l ) = ~o (1\) such that
(i = 1,2)
and
i (126* - 11) d(/\) ~ XII X21 ~ d (1\),
where the ~ on the left cannot be replaced by < for a certain lattice
A The proof is a classic example of the local methods discussed in
general terms in § 8 of Chapter X. COLE (1952a) has shown that to
=
every ~o there is an ~1 ~o such that
(1~i~n-1)
and
X11 ••• x"_1,1Ix"11 ~ id (1\).
CHALK (1947b) discusses when for given ~o there are infinitely many
~1 = ~o (1\) satisfying (4) and (5). The principle behind the proof of
CHALK'S theorem is similar to TSCHEBOTAREFF'S, and we prove it in
§ 4.4. The idea has been put in a much more general form by MACBEATH
(1952a) and C. A. ROGERS (1954b), but we do not go into that here.
XI.4.2. The proof of MINKOWSKI'S conjecture in 2-dimensions may
be made to depend on the following lemma due to DELAUNAY (1947a).
He used it as a tool to investigate ,Ill (1\) (in the notation of § 4.1) for
individual 2-dimensionallattices 1\; and the so-called "algorithm of the
 Product of n linear forms 325

divided cell" has been exploited further by BARNES and SWINNERTON-

DYER (1954a), and BARNES (1954a, 1956c). It was remarked by
DELAUNAY (1947a) that the lemma does not generalize to 3 or more
dimensions; and the same counter-example in 3 dimensions was given
by BIRCH (1Q57a) in ignorance of DELAUNAY'S example.
LEMMA 2. Let /\ be a 2-dimensional lattice and let Xo be a point not
congruent modulo /\ to a point on either co-ordinate axis. Then there are
4 points :11., x 2 , x 3 , x" each congruent to Xo modulo I\, where x; is in the
j-th quadrant, so that
(1 )

and x 2 - Xl' X3 - Xl is a basis lor I\.

The four points Xl' X 2 , X 3 , X, forms a "divided cell" -of the grid @
of points X == Xo (/\). Simpler proofs of Lemma 2 have been given by
BAMBAH (1955b) and REDEl (1959a). We followREDEI.
The proof depends on the following two propositions.
PROPOSITION 1. Let YI' Y2' Y3' y, be lour points 0/ @ such that the
quadrilateral YIY2Y3Y' is convex and contains no other point 0/ @ in its
interior or boundary. Then Yl Y2 Y3 y, is a parallelogram and Y2- Yl'
Y3- YI is a basis lor I\.
This follows almost at once from Chapter III, Lemma 6.
PROPOSITION 2. Let TI be a tine containing points 0/ @ in 3 quadrants.
Let Yl be a point 0/ @ in the remaining quadrant. Suppose that there are
points Y2' Y3 0/ @ on TI such that Yl' Y2' Y3 are the only points 0/ the closed
triangle YI Y2 Y3 in @. Then Lemma 2 1:S true.
For the line TI' through YI and parallel to TI also contains points
of @ in three quadrants. It is then easy to pick out a divided cell with
a pair of opposite -sides on TI and TI'.
We now revert to the proof of Lemma 2. We can find 4 points
Zl' Z2.' Z3' Z" with z; in the j-th quadrant, such that the (not necesarily
convex) closed quadrilateral Zl Z2 Z3 z, contains as few points of @ as
possible. The following three cases are all that can occur.
(i) The quadrilateral Zl Z2 Z3 Z, is convex. It is then a split parallelo-
gram by Proposition 1.
(ii) Three of the points Zl' Z2' Z3' z, are collinear. If, say, Z2' za, z,
are on a line TI, then Lemma 2 follows from Proposition 2 applied to
Zl and TI.
(iii) One point, say Zl' is an inner point of the convex cover of the
remaining three. By the minimal defining property of Zl' Z2' za, z"
any point of @ in the closed quadrilateral Zl Z2 za z, other than Z2' za, z,
must be in the first quadrant; and such points exist since Zl is one.
We may thus choose a point t of @ in the first quadrant and in the
326 Inhomogeneous problems

triangle Z2 Zs z" such that the only points ofin the closed triangle
@
Z2 Zs t are the vertices. Lemma 2 now follows from Proposition 2 on
putting
YI = zS, Y2 = Z2' Ys= t.
Since we have now disposed of all three cases, this concludes the
proof of Lemma 2.
COROLLARY. II zi = (xli' X 2i ) then

1IIxlix2il ~z-8{d(A)}'.
i
For the area of the divided cell is d(A). It is also the sum of the
areas of the four triangles $f with vertices 0, zi' iri +1 (1 ~ i ~ 4; irs = irl ).
But now the area of $f is

Hlxli X2,i+11 + IXI ,i+1 X2il},

and so
2d(A) = L IXli x2,i+11 + L IX2i XI ,i+1l·
i i
The required inequality now follows on applying the inequality of the
arithmetic and geometric means to the 8 terms on the right-hand side.
We can now prove DAVENPORT'S generalization of MINKOWSKI'S
conjecture for n = 2.
THEOREM VII. Let (J, (J be positive numbers and
16eO"~1.

Then to every 2-dimensional point Zo and every lattice A there is a point

=
ir' iro (A) such that
- (J d (A) ~ x~ x~ ~ a d (A). (2)

The case e=a=i is, of course, MINKOWSKI'S conjecture l for n=2.

When Zo is congruent to a point on an axis modulo A, there is nothing
to prove. Otherwise we show that one of the four points iri (1~i~4)
given by Lemma 2 will do. If not, we should have

IXl l X21 1> ad (A), IX13X231 > ad(A),

IXl2 xul > (J d (A), IX14 X ul > (Jd(A);

which is in contradiction with Lemma 2, Corollary.

The reader should not find it difficult to verify that when (J =a = 1
the only case when the equality signs are needed in (2) is when
A =twAoandiro= tw ('Lt) (A), wheret > 0, wisanautomorphofxl x 2 and

1 Proved by MINKOWSKI in this case.

 Product of n linear forms 327

Ao is the lattice of integers. DAVENPORT (1948 a) showed that the equality

signs may be needed when e=FO'. On the other hand it follows from
CHALK'S Theorem of § 4.4 that something stronger is certainly true if
e>1 or 0'>1; and BLANEY (1950a) has given stronger results which
cover the cases when e or 0' is near 1.
XI.4.3. We now give the REMAK-DAVENPORT proof of MINKOWSKI'S
conjecture in 3 dimensions, which depends on the following
LEMMA 3. Let A be any 3-dimensional lattice. Then there exist
numbers Pj>O, (1 ~i~3) such that there are no points 0/ A other than 0
in the ellipsoid
(1)

but there are three linearly independent points 0/ A on the boundary 0/ tf.
We call the ellipsoid tf free if 0 is the only point of A in it. We shall
assume that a free ellipsoid cannot have three linearly independent
points of A on the boundary, for some particular lattice A, and will
ultimately deduce a contradiction.
We note first that
P1P2P3~(~r{d(A)}-2>0 (2)

for any free ellipsoid, by MINKOWSKI'S convex body theorem: the

constant in (2) is not important; all that is important is that it is positive.
Secondly, if ±fIt, ±oa, ±Os are three linearly dependent pairs of
points of A on the boundary of a free ellipsoid, we must have

for some choice of the three ± signs, since the ±OJ lie on a plane through
the origin and so are points of a2-dimensionallattice on the boundary
of an ellipse which contains no point of the lattice (Theorem XI of
Chapter V).
Thirdly, under our hypothesis, if there are two pairs of points
± fit and ± 02 of A on the boundary of a free ellipsoid, they cannot
both lie in the same co-ordinate plane, say, x1 =0. For then we should
have
p2a~1 + P3a~1 = 1, 01 = (0, an, as1)
p2a~2 + Paa~2 = 1, 02 = (0, au, as2)·
If P1 is decreased but Pa, Pa kept constant, the points 01' Os remain on
the boundary and the volume of the ellipsoid increases. Ultimately
there must come a third point on the boundary for some value of P1'
°
since it is impossible to decrease P1 to without a point of A entering
328 Inhomogeneous problems

the ellipsoid. by (2). Hence for some PI the ellipsoid is free but there
are points a,.. az. as on the boundary. where as is not on x1 =0. This
contradicts the hypothesis whose absurdity we wished to prove.
Fourthly we show (on our hypothesis) that if there is a free
ellipsoid (1) with the points ±a1 • ±azEA on the boundary. then there
is one with ±a,.. ±az and ± (a,. +a2) on the boundary. For put
as=a,.+a2 • a4 =a,.-aZ ' and write
(1~i~4).
Then
(f=1.2) }
(f=3.4).
There are numbers ql. qz. q3 not all 0 such that
(i = 1.2). (4)
and after a change of sign. if need be. we may suppose without loss of
generality thatl
qla~4 + qza~4 + q3a~4 ~ O. (5)
We now consider the ellipsoids
(PI + t ql) ~ + (PI + t qz) ~ + (P3 + t q3) x~ = 1
where
t~o.

Since at least one of ql. qz. q3 is negative by (4). as t increases from 0

the inequality (2) with Pi+tqi for Pi must fail for some t; so there
must be some value of t at which for the first time a lattice point enters
the ellipse 8. This cannot be a4 • by (5). and so must be ±as= ± (a,. +az)
by the second remark; which concludes the proof of the fourth remark.
We now prove the lemma. It is clear that we can obtain free ellipsoids
with two pairs of points ±a,.. ±azEA on the boundary by varying the
parameters Pi appropriately. By the fourth remark. there is then a free
ellipse with a,.. al' a,.+az on the boundary. Then by the fourth remark
applied to a,. and a,. +az there is a free ellipse with a,.. a,. +az and
2a1 + a z on the boundary. By induction. there is an ellipsoid

pin) ~ + p~n) x: + p~n) x~ < 1

with a,.. n a,. +a z• (n + 1) a,. +az on the boundary. In particular.
pin) (nau + ~z)z+ p~n)(na21 + a22)2+p~n)(naal + aaz)z= 1. (6)
1 It is readily verified that there cannot be equality in (5). since the deter-
minant of the three forms in q1' qat qa in (4) and (5) does not vanish. But we do
not need this.
 Product of n linear forms 329

We distinguish three cases. Suppose, first, that an =1= 0, a 21 =l=O, a31 =1=O.
Then, by (6),
(j=1,2,3) (n-+oo),
in contradiction to (2). Suppose now that precisely one of an, a 21 , aal
vanishes, say, an=U, a21 =1=O, a31 =1=O. Then by the third remark above
we have a12 =1= 0, and so
(j = 2, 3),
again in contradiction to (2). Finally, suppose that two of an, a 21 , aal
vanish, say, a n =a 21 =O=l=aai' Then a l2 =1= 0=1= a22 , and so
P1(n) <= a:- 2
12
(j = 1,2), p~n) -+ 0,

again in contradiction with (2). Since we have reached a contradiction

in every case, we have proved the absurdity of our initial hypothesis
and so the lemma is true.
MINKOWSKI'S conjecture for n = 3 now follows in a few lines from
Lemma 3 and Lemma 1 Corollary.
THEOREM VIII. Let A be any 3-dimensionallattice and Xo any point.
Then there is an Xl = (xn, x 21 , xal ) == Xo (A) such that

(7)
Let PI' P2' Pa be the numbers given by Lemma 3, so that A has no
point in Pl x~ + P2 x~ + Pa x~ < 1, but three linearly independent points
on the boundary. Hence the three successive minima of A with respect
to the distance-function
F(x) = (PI xi + P2 x~ + Pa x~)~ (8)
are
(9)
We may now apply Lemma 1 Corollary to the lattice M of points
(Pi Xl' PR x 2 , pg Xa), (Xl' x 2 , Xa) E A,
with determinant
d (M) = (PI P2 Pa)! d (A)
and with successive minima with respect to Ixl given by (9). Hence to
any Xo there is a congruent Xl such that

p. xl. +P,r.. +p,r.. '" : W~U' ) (10)

= -4 (PIP2Pa)i{d(A)}i.
The required inequality (7) now follows at once from (10) and from the
inequality of the arithmetic and geometric means.
330 Inhomogeneous problems

The reader should have no difficulty in showing that the sign of

equality in (7) is required only when A =twAo for some number t=j=0,
and some automorph w of "1"2"3, where Ao is the lattice of points with
integral co-ordinates; and then only for ~o==two(-Lt,t) (A). Note
that to have equality in (7) one must have equality in both applications
of the inequality of the arithmetic and geometric means; that going
from Lemma 1 to Lemma 1 Corollary and that going from (10) to (7).
XI.4.4. We now prove 1 the theorems of TSCHEBOTAREW and CHALK.
Since CHALK'S theorem is slightly simpler, we prove that first.
THEOREM IX. Let A be an n-dimensional lattice and ~o a point.
Then there is an ~1 = ("11' ... , "nl) == ~o (A) such that
";1>0 (1~i~n), (1)
"l1 ... "nl~d(A). (2)
There certainly is a point ~2 = ("12' ... , "n 2) = ~o (A) for which
";2>0 (1~i~n). (3)
If II "i2~d(A), then we may put ~1=~2' Otherwise we have
II ";2> d(A), (4)
and so, by MINKOWSKI'S convex body theorem, there is a point a=!=o
of A such that
(1~i~n). (5)
By considering 2'a instead of a with a suitably chosen integer r~O,
we may suppose, further, that
la,1 ~tl"'21 (6)
for at least one integer J. Then the two points
~2 ± a = ~± = ("i', .. " ,,;=)

are both congruent to ~o and lie in the quadrant ";>0 (1~i~n).

Further,
II
;
xt II X7
';
II X,
1 =II (Ix;l-a,I) :5:~
; ,. ; "
x~ -4'

since by (5) and (6) every factor on the right-hand side is ~ 1, and one
at least is ~ 1. Hence choosing for ~3 that one of ~± for which II x;
is least, we have
"i3>0 (1~i~n); II
;
"i3 ~ (1)6 II "j2"
i
1 Following MACBEATH (1952a), but in our special cases the argument can be
simplified.
 Product of n linear forms 331

If IT xi3~d(A). then we may put :1:1 =:1:3 , Otherwise we repeat the

process with :1:3 instead of :1:2 and obtain an :1:, with
(1~i~n); IT xi' ~ (l)l IT xi3 ~! II X,2'
i i i
And so on. Clearly an :1:1 is reached in a bounded number of steps,
with a bound that can be given explicitly in terms of II xi 2' This
concludes the proof. i
A similar idea gives TSCHEBOTAREW'S
THEOREM X. Let A be any n-dimensional lattice, e an arbitrarily
small number and :1:0 a point. Then there is a point :1:1 = (X11' "', x n1 ) =:1:0
(A), such that
IX11 ••. xn1 1 ~ (rn/2 + e) d(I\). (7)
Let t be the number such that
(2- n/2 + e) tn = 1 , (8)
so
(9)
If IT IXiol~(rn/2+e) d(I\), there is nothing to prove, so we may
i
suppose that
VIXiOI > (rn/2+e)d(l\) }
(10)
= t-nd(I\).
By MINKOWSKI'S convex body theorem, there is a point a=!=o in 1\ for
which
(1~i~n). (11)
As in the proof of Theorem IX, we may suppose, on taking 2'a with
suitable integer r~O, that
(12)
for some J. Put
:I:± =:1:0 ±a,
so that
If1 xi If, Xi
n. x.10 =II (1 _ x_
a~' )
,10
•.
(13)
1
But
332 Appendix

by (12). Hence, on taking for z,. that one of z± for which II Ixi I is least,
we have

where
52 = max {I 1 - t21, 11- it21}< 1.
As in the proof of IX we reach an ZI satisfying (7) after a finite
number of steps, the number of steps being bounded by a number
depending only on n, II IXiOI and E. This concludes the proof.
i

Appendix
In this appendix we list the lattice constants of some sets connected
with quadratic forms and give further references and some additional
comments. We write
~"s(z) = x~ + ... + x~ - x~+1 - ... - x~+s,
and denote by r"s the lattice constant of the set

I~"sl < 1
in n-dimensional space, where
n=r+5.
Results about definite forms are usually given in terms of 1'.. where
1': = r,;:o·.The first 8 values are known:
I'~ = 1, I'~ = t, I'~ = 2,
I'~ = 8, 1': = ~:-, I'~ = 64,
The value of 1'1 is trivial; the values of 1'2,1'3'1', have been found
in this book (Chapter II, Theorems II,. III and Chapter X, Theorem IV,
Corollary). For references and a list of the corresponding critical forms
see CHAUNDY (1946a), who gives proofs that 1':=28, l'~g=2J.°/3; but
CHAUNDY'S proofs contain a lacuna. Presumably his line of argument
would lead to incorrect results by n = 12; see COXETER and TODD (1953 a)
for a special form in 12 variables.
For indefinite forms we have
n2 _5
.lI,I-4"
1;~1 = I;.~2 = !
I~~2 =!
To3,1 =.lI,3
2 nll
= 4"7
 Appendix 333

due to HURWITZ, MARKOFF, OPPENHEIM and OPPENHEIM respectively,

the proofs being reproduced in DICKSON (1930a). We have proved all
except the last line in the book (Chapter II, Theorems IV, VII and
Chapter X, Theorem IV, Corollary). All are isolated. The successive
minima of IfIJI, I I< 1 are the MARKOFF Chain (see Chapter 2, § 4). The
first 11 minima for IflJ2,II < 1 and the first 7 minima for IflJ2, 21 < 1 have
been given by VENKOV (1945a) and OPPENHEIM (1934a) respectively.
It is conjectured that IfIJ" sl < 1 is of infinite type when r>O, s>O,
r +s~ 5, see DAVENPORT (1956a)l.
Let B" s be the lattice constant of

0< fIJ"s < 1.

Then
B1,1
2 I
=4;

B 2,1
2 1
=4;,

The value of BI,I is given by Theorem V of Chapter II. The results in

the second row are due to DAVENPORT (1949a); both are isolated and
something is known about further minima, see OPPENHEIM (1953a).
The results in the third row are due to OPPENHEIM (1953 b) and again
something is known about successive minima. In all cases the critical
lattice has points a=F0 at which fIJ"s(a) =0.
Let A"s be the lattice constant of

O~fIJ"s<1.
Then

A22,2 = 81 Aa2 ,1> Ata~-~-.

64-' = .2I
a2'

The value of Al I follows at once from Theorem VI of Chapter II. The

rest are due to BARNES (1955a) and BARNES and OPPENHEIM (1955a).
If a quadratic form in n~ 3 variables takes arbitrarily small non-zero
values of one sign then it also takes arbitrarily small values of the
other sign. If a quadratic form represents 0, has two of its coefficients
in an irrational ratio and has n~ 5 variables, then it takes arbitrarily
small values of both signs (OPPENHEIM 1953c, d).

1 For later work on this problem, mainly due to DAVENPORT and BIRCH, see
RIDOUT (1958a).
 References
BACHMANN, P. (1923 a): Die Arithmetik der quadratischen Formen II. Leipzig and
Berlin.
BAMBAH, R. P. (1951a): On the geometry of numbers of non-convex star-regions
with hexagonal symmetry. Phil. Trans. Roy. Soc. Lond. 243, 431-462.
(1954a): Lattice coverings with four-dimensional spheres. Proc. Cambridge
Phil. Soc. 50, 203-208 (1954).
(1954 b) : On lattice coverings by spheres. Proc. Nat. Inst. Sci. India 20, 25 - 52.
(1954c): On polar reciprocal convex domains (and addendum). Proc. Nat.
Inst. Sci. India 20,119-120,324-325.
(1955a): Polar Reciprocal Convex bodies. Proc. Cambridge Phil. Soc. 51.
377-378.
(1955b): Divided cells. Res. Bull. Panjab Univ. 81, 173-174.
(1958a): Some transference theorems in the geometry of numbers. Mh. Math.
62, 243-249.
'-. and H. DAVENPORT (1952a): The covering of n-dimensional space by spheres.
J. Lond. Math. Soc. 27, 224-229·
-, and C. A. ROGERS (1952a): Covering the plane with convex sets. J. Lond.
Math. Soc. 27, 304-314.
-, and K. ROGERS (1955a): An inhomogeneous minimum for non-convex star
regions with hexagonal symmetry. Canad. J. Math. 7, 337-346.
BARNES, E. S. (1950a): Non-homogeneous binary quadratic forms. Quart. J. Math.
Oxford (2) 1, 199-210.
(1951 a): The minimum of the product of two values of a quadratic form 1.
Proc. Lond. Math. Soc. (3) 1, 257-283.
(1954 a): The inhomogeneous minima of binary quadratic forms IV. Acta
Math. 92, 235-264.
(1955a): The non-negative values of quadratic forms. Proc. Lond. Math. Soc.
(3) 5,185-196.
(1956a): The inhomogeneous minimum of a ternary quadratic form. Acta
Math. 96, 67-97.
(1956b): The coverings of space ey spheres. Canad. J. Math. 8, 293-304.
(1956c): On linear inhomogeneous Diophantine approximation. J. Lond. Math.
Soc. 31, 73-79.
(1957a): On a theorem of Vorono!. Proc. Cambridge Phil. Soc. 53, 537-539.
(1957b): The complete enumeration of extreme senary forms. Phil. Trans.
Roy. Soc. Lond. 249, 461 - 506.
-, and A. OPPENHEIM (1955a): The non-negative values of a ternary quadratic
form. J. Lond. Math. Soc. 30, 429-439.
-, and H. P. F. SWINNERTON-DvER (1952a, b): The inhomogeneous minima of
binary quadratic forms I, II. Acta Math. 87, 259-323; 88, 279-316.
- - (1954 a): The inhomogeneous minima of binary quadratic forms III. Acta
Math. 92, 199-234.
BIRCH, B. J. (1956a): A transference theorem of the geometry of numbers. J.
Lond. Math. Soc. 31, 248-251.
 References 335

BIRCH, B. J. (1956b): Another transference Theorem of the geometry of numbers.

Proc. Cambridge Phil. Soc. 53, 269-272.
(1957a): A grid with no split parallelopiped. Proc. Cambridge Phil. Soc. 53,
536.
(1958a): The inhomogeneous minima of quadratic forms of signature O. Acta
Arithmetica 4, 85-98.
-, and H. P. F. SWINNERTON-DYER (1956a): On the inhomogeneous minimum of
the product of n linear forms. Mathematika 3, 25 - 39.
BLANEY, H. (1948a): Indefinite quadratic forms in n variables. J. Lond. Math.
Soc. 23, 153-160.
(1950a): Some asymmetric inequalities. Proc. Cambridge Phil. Soc. 46,
359-376.
(1957a): On the Davenport-Heilbronn Theorem. Mh. Math. 61, 1-36.
BLICHFELDT, H. F. (1914a): A new principle in the geometry of numbers with
some applications. Trans. Amer. Math. Soc. 15, 227-235.
(1929a): The minimum value of quadratic forms and the closest packing of
spheres. Math. Ann. 101, 605-608.
(1939a): Note on the minimum value of the discriminant of an algebraic field.
Mh. Math. Phys. 48, 531 - 533.
BONNEsEN, T., U. W. FENCHEL (1934a): Theorie der Konvexen Karper. Ergebnisse
der Math. usw. 3 (1). Berlin.
BRUNNGRABER, E. (1944 a): trber Punktgitter. Diss. Wien.
CASSELS, J. W. S. (1947 a): On a theorem of Rado in the geometry of numbers.
J. Lond. Math. Soc. 22, 196-200.
(1948a): On two problems of Mahler. Proc. Kon. Ned. Akad. Wet. 51,854-857
(= Indag Math. 10, 282-285).
(1952a): The inhomogeneous minimum of binary quadratic, ternary cubic and
quaternary quartic forms. Proc. Cambridge Phil. Soc. 48, 72- 86, 519- 520.
(1952 b) : The product of n inhomogeneous linear forms in n variables. J. Lond.
Math. Soc. 27, 485-492.
(1953a): A short proof of the Minkowski-Hlawka Theorem. Proc. Cambridge
Phil. Soc. 49,165-166.
(1955 a, b): Simultaneous diophantine approximation I, II. J. Lond. Math.
Soc. 30, 119 - 121 and Proc. Lond. Math. Soc. (3), S, 435 - 448.
(1956a): On a result of Marshall Hall. Mathematika 3, 109-110.
(1957 a) : An introduction to diophantine approximation. Cambridge: Cambridge
University Press.
(1958a): On subgroups of infinite abelian groups. J. Lond. Math. Soc. 33,
281-284.
W.LEDERMANN and KMAHLER (1951a): Farey section in k(i) and k(e).
Phil. Trans. Roy. Soc. Lond. A 243, 585-628.
-, and H. P. F. SWINNERTON-DYER (1955a): On the product of three homogeneous
linear forms and indefinite ternary quadratic forms. Phil. Trans. Roy. Soc.
Lond. 248, 73-96.
CERNY, K (1952a): On the minima of binary biquadratic forms I (in Russian).
Czechos. Math. J. 2, 1 - 56.
CHABAUTY, C. (1949a): Sur les minima arithmcHiques des formes. Ann. Sci. Ec.
Norm. Sup., Paris (3) 66, 367-394.
(1950a): Limite d'ensembles et geometrie des nombres. Bull. Soc. Math.
France 78,143-151.
(1952 a): Empilement de spheres egales dans R". etc. C. R. Acad. Sci., Paris
235, 529- 532 (but see Math. Reviews 14, 541).
336 References

CHALK, ]. H. H. (1947a, b): On the positive values of linear forms I, II. Quart.
]. Math. Oxford 18, 215-227; 19, 67-80 (1948).
- (1949a): Reduced binary cubic forms. ]. Lond. Math. Soc. 24, 280- 284.
- (1950a): On the frustrum of a sphere. Ann. of Math. (2) 52, 199-216.
-, and C: A. ROGERS (1948a): The critical determinant of a convex cylinder.
]. Lond. Math. Soc. 23,178-187.
- (1949a): The successive minima of a convex cylinder. ]. Lond. Math. Soc.
24, 284-291-
- (1951): On the product of three homogeneous linear forms. Proc. Cambridge
Phil. Soc. 47, 251-259.
CHAUNDY, T. W. (1946a): The arithmetic minima of positive quadratic forms.
Quart. ]. Math; Oxford 17 (67),166-192.
CLARKE, L. E. (1951a): On the product of three non-homogeneous forms. Proc.
Cambridge Phil. Soc. 47, 260-265.
- (1958 a): The critical lattices of a star-shaped octagon. Acta Math. 99,
1-32.
COHN, H. (1953 a, b): Stable lattices I, II. Canad.]. Math. 5, 261 - 270; 6,265 - 273
(1954).
COLE, A. ]. (1952a): On the product of n linear forms. Quart.]. Math. Oxford (2)
3, 56-62.
CORPUT, ]. G. VAN DER (1936a): Verallgemeinerung einer Mordellschen Beweis-
methode in der Geometrie der Zahlen. Acta Arithmetica 2, 145- 146.
COXETER, H. S. M., and]. A. TODD (1953 a): An extreme duodenary form. Canad.
]. Math. 5, 384- 392.
DAVENPORT, H. (1938a): On the product of three homogeneous linear forms I.
Proc. Lond. Math. Soc. (2) 44, 412-431-
(1939a): A simple proof of Remak's Theorem on the product of 3 linear forms.
]. Lond. Math. Soc. 14, 47-51-
(1939 b): On the product of three homogeneous linear forms III. Proc. Lond.
Math. Soc. 45, 98-125.
(1939c): MinkowSki's inequality for the minima associated with a convex body.
Quart.]. Math. Oxford 10,119-121-
(1941 a): Note on the product of three homogeneous linear forms. ]. Lond.
Math. Soc. 16, 98-101-
(1941 b): On a conjecture of Mordell concerning binary cubic forms. Proc.
Cambridge Phil. Soc. 37, 325-330.
(1943 a): On the product of three homogeneous linear forms. IV. Proc. Cam-
bridge Phil. Soc. 39, 1-21-
(1945 a, b): The reduction of a binary cubic form I, II. ]. Lond. Math. Soc.
20,14-22,139-157.
(1946a): On a theorem of Tschebotareff. ]. Lond. Math. Soc. 21, 28- 34 and
Corrigendum 24, 316 (1949).
(1947a): On a theorem of Markoff. ]. Lond. Math. Soc. 22, 96-99.
(1947b): The geometry of numbers. Math. Gazette 31,206-207.
(1947c): On the product of three non-homogeneous linear forms. Proc. Cam-
bridge Phil. Soc. 43, 137- 152.
(1948a): Non-homogeneous ternary quadratic forms. Acta Math. 80, 65 - 95.
(1949a): On indefinite ternary quadratic forms. Proc. Lond. Math. Soc. (2)
51,145-160.
(1950a): Note on a binary quartic form. Quart. J. Math. Oxford (2) 1, 253-261.
(1951 a): Indefinite quadratic forms and Euclid's algorithm in real quadratic
fields. Proc. Lond. Math. Soc. (2) 53, 65-82.
 References 337

DAVENPORT, H. (1952a): Simultaneous diophantine approximation. Proc. Lond.

Math Soc. (3) 2, 406-416.
(1952 b): The covering of space by spheres. Rend. Circ. Mat. Palermo (2) 1, 92-107.
(1955a): On a theorem of Furtwangler. J. Lond. Math. Soc. 30, 186-195.
(1956a): Indefinite quadratic forms in many variables. Mathematica 3, 81-101.
-, and M. HALL (1948a): On the equation ax2+b y 2+ cz2=0. Quart.]. Math.
Oxford 19, 189-192.
-, and H. HEILBRONN (1947a): Asymmetric inequalities for non-homogeneous
linear forms. ]. Lond. Math. Soc. 22, 53-61.
-, and C. A. ROGERS (1947a): Hlawka's Theorem in the geometry of numbers.
Duke Math. ]. 14, 367-375.
- (1950a): Diophantine inequalities with an infinity of solutions. Phil. Trans.
Roy. Soc. Lond. 242, 311- 344.
- (1950b): On the critical determinants of cylinders. Quart.]. Math. Oxford
(2) 1, 215-218.
-, and H. P. F. SWINNERTON-DYER (1955a): Products of inhomogeneous linear
forms. Proc. Lond. Math. Soc. (3) 5, 474-499.
DAVIS, C. S. (1951 a): The minimum of a binary quartic form. Acta Math. 84,
263-298.
DEL AUN AY, B. N. (1947 a): An algorithm for divided cells [Russian]. Izv. Akad.
Nauk SSSR. (Ser. Mat.) 11, 505- 538.
DICKSON, L. E. (1929a): Introduction to the theory of numbers. Chicago, Ill.:
Chicago University Press.
- (1930a): Studies in the Theory of Numbers. Chicago, Ill.: Chicago University
Press.
DOWKER, C. H. (1944 a): On minimum circumscribed polygons. Bull. Amer. Math.
Soc. 50, 120-122.
DYSON, F.]. (1948a): On the product of four non-homogeneous forms. Ann. of
Math. (2) 49, 82-109.
EGGLESTON, H. G. (1958a): Convexity. Cambridge: Cambridge Univ. Press.
ENNoLA, V. (1958a): On the first inhomogeneous minimum of indefinite binary
quadratic forms and Euclid's algorithm in real quadratic fields. Ann. Univ.
Turkuensis (Turun Yliopiston ]ulkaisuja) A 1 28, 1- 58.
ERDOS, P., and C. A. ROGERS (1953a): The covering of n-dimensional space by
spheres. ]. Lond. Math. Soc. 28, 287 - 293.
FEJES T6TH, L. (l('50a): Some packing and covering theorems. Acta Univ. Szeged,
Acta Sci. Math 12/A, 62-67.
- (1953a): Lagerungen in der Ebene, auf der Kugel und im Raum. Berlin:
Springer.
FENCHEL, W. (1937 a): Verallgemeinerung einiger Satze aus der Geometrie der
Zahlen. Acta Arithmetica 2, 230- 241.
FEW, L. (1956a): Covering space by spheres. Mathematika 3,136-139.
FOSTER, D. M. E. (1956a): Indefinite quadratic polynomials in n variables. Mathe~
matika 3,111-116.
GAUSS, C. F. (1831a): Besprechung des Buchs von L. A. Seeber: Untersuchungen
liber die Eigenschaften der positiven ternaren quadratischen Formen usw.
Gottingsche Gelehrte Anzeigen 1831, ]uli 9. Reprinted in Werke (1876).
Vol. II, 188-196.
GODWIN, H.]. (1950a): On the product of five homogeneous linear forms. ].
Lond. Math. Soc. 25, 331- 339.
HAJ6s, G. (1942a): Dber einfache und mehrfache Bedeckung des 71-dimensionalen
Raumes mit einem Wlirfelgitter. Math. Z. 47, 427-467.
Cassels, Geometry of Numbers 22
JJ8 References

HALL, M. (1947a): On the sum and product of continued fractions. Ann. Math.
(2) 48, 966- 993.
HARDY, G. H., and E. M. WRIGHT (1938a): An introduction to the theory of
numbers. Oxford.
HLAWKA, E. (1944a): Zur Geometrie der Zahlen. Math. Z. 49, 285-312.
(1949a): Ausfiillung und Oberdeckung konvexer Karper durch konvexe Karper.
Mh. Math. Phys. 53, 81-131.
(1952a): Zur Theorie des Figurengitters. Math. Ann. 125, 183-207.
(1954a): Grundbegriffe der Geometrie der Zahlen. ]ber. DMV 57,37-55.
(1954b): Zur Theorie der Oberdeckung durch konvexe Karper. Mh. Math.
Phys. 58, 287-291.
(1954c): Inhomogene Minima von Sternkarpern. Mh. Math. Phys. 58, 292- 305.
] OHN, F. (1948 a): Extremum problems with inequalities as subsidiary conditions.
Studies and essays presented to R. COURANT, p. 187-204. New York.
KELLER, O. H. (1954a): Geometrie der Zahlen. Enzyklopadie der Math. Wissen-
schaften, Bd. 12 , H. II, Teil iii.
KHINTCHINE, A. Y A (1948a): A quantitative formulation of Kronecker's Theory of
approximation [in Russian). Izv. Akad. Nauk SSSR. (Ser. Mat.) 12,113-122.
KLEIN, F. (1895 a): Ober eine geometrische Auffassung der gewahnlichen Ketten-
bruchentwicklung. Nachr. Ges. Wiss. Gattingen 1895, 357-359.
- (1896a): Ausgewahlte Kapitel der Zahlentheorie. (Ausgearbeitet von A. SOM-
MERFELD und PH. FURTWANGLER.) Leipzig.
KNESER, M. (1955a): Ein Satz tiber abelsche Gruppen mit Anwendungen auf die
Geometrie der Zahlen. Math. Z. 61, 429-434.
- (1956a): Summenmengen in lokalkompakten abelschen Gruppen. Math. Z. 66,
88-110.
LEKKERKERKER, C. G. (1956a): On the Minkowski-Hlawka theorem. Proc. Kon.
Ned. Acad. Wet. Amsterdam A 59 (= Indag. Math. 18),426-434.
- (1957 a): On the volume of compound convex bodies. Proc. Kon. Ned. Acad.
Wet. A 60 (= Indag. Math. 19), 284-289.
MACBEATH, A. M. (1951 a): The finite volume theorem for non-homogeneous lattices.
Proc. Cambridge Phil. Soc. 47, 627-628.
(1952a): A theorem on non-homogeneous lattices. Ann. of Math. (2) 56,
269-293.
(1953a): On measure of sum sets. II. The sum-theorem for the torus. Proc.
Cambridge Phil. Soc. 49, 40-43.
-, and C. A. ROGERS (1955a and 1958b): A modified form of Siegel's mean-value
theorem I, II. Proc. Cambridge Phil. Soc. 51,565-576 and 54,322-325 (1958).
- - (1958 a): Siegel's mean value theorem in the geometry of numbers, Proc.
Cambridge Phil. Soc. 54, 139-151.
MAHLER, K. (1938a): On Minkowski's theory of reduction of positive definite
quadratic forms. Quart.]. Math. Oxford 9, 259-263.
(1939a): Ein Obertragungsprinzip fiir lineare Ungleichungen. Casopis pro pest.
mat. a fys. 68, 85-92.
(1939 b): Ein Obertragungsprinzip fiir konvexe Karper. Casopis pro pest. mat.
a fys. 68, 93-102.
(1946a, b, c): Lattice points in two-dimensional star domains I, II, III. Proc.
Lond. Math. Soc. (2) 49,128-157,158-167,168-183.
(1946d, e): On lattice points in n-dimensional star-bodies. I. Existence theorems.
Proc. Roy. Soc. Lond. A 187, 151-187. - II. Reducibility theorems. Proc.
Kon. Ned. Akad. Wet. 49, 331- 343, 444-454, 524- 532, 622-631.
(1946f): The theorem of Minkowski-Hlawka. Duke Math. J. 13, 611-621.
 References 339

MAHLER, K. (1946g): On lattice points in a cylinder. Quart. ]. Math. Oxford

17, 16-18.
(1947 a): On irreducible convex domains. Proc. Kon. Ned. Akad. Wet. 50,
98-107 (= Indag. Math. 9,3-12).
(1947 b): On the area and the densest packing of convex domains. Proc. Kon.
Ned. Akad. Wet. 50, 108-118 (= Indag. Math. 9, 14-24).
(1947c): On the minimum determinant and the circumscribed hexagons of a
convex domain. Proc. Kon. Ned. Akad. Wet. 50, 692-703 (= Indag. Math. 9,
326-337)·
(1948a): On lattice points in polar reciprocal convex domains. Proc. Kon.
Ned. Wet. 51, 482-485 (= Indag. Math. 10, 176-179).
(1949a): On the minimum determinant of a special point set. Proc. Kon. Ned.
Akad. Wet. 52, 633-642 (= Indag. Math. 11,195-204).
(1949b): On the critical lattices of arbitrary point sets. Canad. ]. Math. 1. 78-87·
(1950a): The geometry of numbers (Duplicated lectures, Boulder, Colorado,
U.S.A.).
(1955a, b): On compound convex bodies I, II. Proc. Lond. Math. Soc. (3) 5,
358-384.
MARKOFF, A. (1879a): Sur les formes quadratiques bin aires indefinies. Math. Ann.
15, 381-409.
MELMORE, S. (1947a): Densest packing of equal spheres. Nature, Lond. 159,817.
MINKOWSKI, H. (1896a): Geometrie der Zahlen. Leipzig and Berlin.
(1904a): Dichteste gitterfiirmige Lagerung kongruenter Karper. Nachr. K.
Ges. Wiss. Gattingen 1904, 311- 355. (Reprinted in Gesammelte Abhandlun-
gen II, 3-42).
(1907a): Diophantische Approximationen. Leipzig.
MORDELL, L. ]. (1931 a): Indefinite quadratic forms in n variables. Proc. Roy. Soc.
A 131, 99-108.
(1937 a): A theorem of Khintchine on linear diophantine approximation.
]. Lond. Math. Soc. 12, 166-167.
(1940a): Tschebotareff's theorem on the product of non-homogeneous linear
forms. Vjschr. naturforsch. Ges. Zurich 85, Beiblatt (Festschrift Rudolf Fueter)
47- 50.
(1941 a) : The product of homogeneous linear forms. J. Lond. Math. Soc. 16, 4-12.
(1942a): The product of three homogeneous linear ternary forms. J. Lond.
Math. Soc. 17. 107-115.
(1943a): The product of n homogeneous forms. Mat. Sbornik (Rec. Math.) 12
(54), 273-276.
(1943 b): On numbers represented by binary cubic forms. Proc. Lond. Math.
Soc. (2) 48, 198- 228.
(1944a): Lattice points in the region Ix3 + y 3 1;;;; 1. J. Lond.Math. Soc. 19, 92-99.
(1944 b): Observation on the minimum of a positive definite quadratic form in
eight variables. J. Lond. Math. Soc. 19, 3-6.
(1946a): Further contributions to the geometry of numbers for non-convex
regions. Trans. Amer. Math. Soc. 59,189-215.
(1948a): The minimum of a definite ternary quadratic form. ]. Lond. Math.
Soc. 23, 175- 178.
(1951a): On the equation ax2 +b y 2_ cz2=0. Mh. Math. 55, 323-327.
(1952 a): The minima of some non-homogeneous functions of two variables.
Duke Math. ].19,519-527.
(1956a): The minimum of an inhomogeneous quadratic polynomial in n
variables. Math. Z. 63, 525-528.
340 References

MORDELL. L. J. (1959a): Tschebotareff's theorem on the product of non-homo-

geneous linear forms II. (To appear.)
MULLENDER. P. (1945a): Toepassing van de meetkunde der getallen op ongelijkhe-
den in K(I) en K(i lim).Diss. Amsterdam.
(1948a): Lattice points in non-convex regions. I. II. III. Proc. Kon. Ned.
Acad. Wet. 51,874-884.1251-1261; 52.50-60 (1949) [= Indag. Math. 10,
302-312. 395-4lJ5; 11. 18-28 (1949)].
(1950a): Simultaneous approximation. Ann. of Math. 52, 417-426.
+
MULLINEUX. N. (1951 a): Lattice points in the star-body Ixf x~ - xii;::;; 1.1 xal ;::;;Vi
Proc. Lond. Math. Soc. (2) 54, 1-41.
OLLERENSHAW. K. (1945a): The minima of a pair of indefinite harmonic binary
quadratic forms. Proc. Cambridge Phil. Soc. 41. 77-96.
(1945 b): The critical lattices of a circular quadrilateral formed by arcs of
three circles. Quart. J. Math. 17. 223 - 239.
(1953a): An irreducible non-convex region. Proc. Cambridge Phil. Soc. 49.
194-200.
(1953b): Irreducible convex bodies. Oxford Quart. J. Math. (2) 4, 293-302.
OPPENHEIM. A. (1932a): The lower bounds of indefinite Hermitian forms. Quart.
J. Math. Oxford 3,10-14.
(1934a): Minima of quaternary quadratic forms of signature O. Proc. Lond.
Math. Soc. (2) 37. 63-81.
(1936a): The lower bounds of Hermitian quadratic forms in any quadratic
field. Proc. Lond. Math. Soc. (2) 40, 541- 555.
(1946a): Remark on the minimum Iilf quadratic forms. J. Lond. Math. Soc.
21,251-252.
(1953 a): One-sided inequalities for quadratic forms. (I) Ternary forms. Proc.
Lond. Math. Soc. (3) 3, 328-337.
(1953 b): One-sided inequalities for quadratic forms. (II) Quaternary forms.
Proc. Lond. Math. Soc. (3) 3, 417-429.
(1953C, d): Value of quadratic forms I, II. Quart. J. Math. Oxford (2) 4,
54-59,60-66.
(1953e): Value of quadratic forms III. Mh. Math. 57. 97-101.
(1953f): One-sided inequalities for hermitian quadratic forms. Mh. Math. 57,
1-5.
PITMAN. J. (1958 a): The inhomogeneous minima of a sequence of symmetric MAR-
KOV forms. Acta Arithmetica 5, 81-116.
POITou. G. (1953a): Sur l'approximation des nombres complexes par les nombres
des corps imaginaires quadratiques etc. Ann. Sci. tc. Norm. Sup. Paris (3)
70, 199-265.
PRASAD, A. V. (1949a): A non-homogeneous inequality for integers in a special
cubic field. Froc. Kon. Ned. Akad. Wet. Arnst. 52. 240-250. 338-350
(= Indag. Math. 11.55-65.112-124).
RADO, R. (1946a): A theorem on the geometry of numbers. J. Lond. Math. Soc.
21. 34-47.
RANKIN. R. A. (1947a): On the closest packing of spheres in n dimensions. Ann.
of Math. 48. 1062-1081.
(1949a. b. c): On sums of powers of linear forms I. II, III. I. Ann. of Math.
50, 691 - 698. - II. Ann. of Math. 50, 699-704. - III. Proc. Kon. Ned.
Akad. Wet. 51, 846-853 (1948) (= Indag. Math. 10, 274-281).
(1953a): The anomaly of convex bodies. Proc. Cambridge Phil. Soc. 49. 54- 58.
(1955a): The closest packing of spherical caps in n dimensions. Proc. Glasgow
:Vlath. Assoc. 2, 139-144.
 References 341

RtDEI, L. (1955a): Neuer Beweis des Haj6sschen Satzes tiber die endlichen abel-
schen Gruppen. Acta Math. Hungarica 6, 27-40.
- (1959a): Neuer Beweis eines Satzes von DELaNE tiber ebene Punktgitter.
J. Land. Math. Soc. 34, 205-207.
REINHARDT, K. (1934a): Ober die dichteste gitterfOrmige Lagerung kongruenter
Bereiche in der Ebene usw. Abh. Math. Sem. Hansische Univ. 10, 216-230.
REMAK, R. (1923a, b): Verallgemeinerung eines Minkowskischen Satzes I, II.
Math. Z. 17,1-34; 18, 173-200 (1924).
(1925 a): Ober die geometrische Darstellung der indefiniten binaren quadrati-
schen Formen. Jber. DMV 33, 228-245.
(1938a): Ober die Minkowskische Reduktion. Comp. Muth. 5, 368- 391,
RIDOUT, D. (1958a): Indefinite quadratic forms. Mathematika 5,122-124.
RIESZ, M. (1936 a): Modules teciproques. Comptes Rendus, Congr. Intern. des
Math., Oslo 2, 36-37.
ROGERS, C. A. (1947a): A note on irreducible star-bodies. Proc. Kon. Ned. Akad.
Wet. 50, 868-872 (= Indag. Math. 9, 379-383).
(1947b): Existence theorems in the geometry of numbers. Ann. of Math. 48,
994-1002.
(1947c): A note on a problem of Mahler. Proc. Roy. Soc. Lond. A 191,503- 517.
(1949a): The product of the minima and the determinant of a set. Proc. Kon.
Ned. Akad. Wet. 52, 256-263 (= Indag. Math. 11, 71-78).
(1949b): On the critical determinant of a certain non-convex cylinder. Quart.
J. Math. Oxford 20, 45-47.
(1950a): The product of n real homogeneous linear forms. Acta Math. 82,
185-208.
(1950b): A note on coverings and packings. J. Land. Math. Soc. 25, 327-331.
(1951 a): The closest packing of convex two-dimensional domains. Acta Math.
86, 309- 321-
(1951 b): The number of lattice points in a star-body. J. Land. Math. Soc.
26,307-310.
(1952a): The reduction of star-sets. Phil. Trans. Roy. Soc. Land. A 245,59-93.
(1952 b): Indefinite quadratic forms in n variables. J. Land. Math. Soc. 27,
314-319·
(1953a): Almost periodic critical lattices. Arch. der Math. 4, 267-274.
(1954a): The Minkowski-Hlawka Theorem. Mathematika 1, 111-124.
(1954b): A note on the theorem of Macbeath. J. Land. Math. Soc. 29,133-143.
(1954c): The product of n non-homogeneous linear forms. Proc. Land. Math.
Soc. (3) 4, 50 - 83.
(1955a): Mean values over the space of lattices. Acta Math. 94, 249-287.
(1955b): The moments of the number of points of a lattice in a bounded set.
Phil. Trans. Roy. Soc. Land. A 248,225-251-
(1956a): The number of lattice points in a set. Proc. Land. Math. Soc. (3)
6, 305- 320.
(19 57 a): A note on coverings. Mathematika 4, 1 - 6.
(1958a): Lattice coverings of space: the Minkowski-Hlawka theorem. Proc.
Land. Math. Soc. (3) 8, 447-465.
(1958b): Lattice coverings of space with convex bodies. J. Land. Math. Soc.
33,208-212.
(1958c): The packing of equal spheres. Proc. Lond. Math. Soc. (3) 8, 609-620.
(1959a): Lattice coverings of space. Mathematika 6, 33-39.
ROGERS, K. (1953 a): The minima of some inhomogeneous functions of two variables.
J. Land. Math. Soc. 28, 394-402.
342 References

ROGERS, K. (1955a): On the generators of an ideal, with an application to the

geometry of numbers in unitary space U 2 • Amer. J. Math. 77, 621-627.
- (1956a): Indefinite binary hermitian forms. Proc. Lond. Math. Soc. (3) 6,
205-223·
-, and H. P. F. SWINNERTON-DYER (1958a): The geometry of numbers over
algebraic number fields. Trans. Amer. Math. Soc. 88, 227-242.
SAMET, P. A. (1954a, b): The product of linear non-homogeneous linear forms I, II.
Proc. Cambridge Phil. Soc. 50,372-379,380-390.
SAWYER, D. B. (1948a): The product of two non-homogeneous linear forms. J.
Lond. Math. Soc. 23, 250-251,
(1950a): A note on the product of two non-homogeneous linear forms. J. Lond.
Math. Soc. 25, 239-240.
(1953 a): The minima of indefinite binary quadratic forms. J. Lond. Math.
Soc. 28, 387 - 394.
SCHMIDT, W. (1955 a): tJ'ber hohere kritische Determinanten von Sternkorpern.
Mh. Math. 59, 274- 304.
(1956a): Eine neue Abschatzung der kritischen Determinanten von Stern-
korpern. Mh. Math. 60, 1 - 10.
(1956b): Eine Verschariung des Satzes von Minkowski-Hlawka. Mh. Math.
60,110-113.
(1958 a): The measure of the set of admissible lattices. Proc. Amer. Math.
Soc. 9, 390-403.
SCHOLZ, A. (1938a): Minimaldiskriminanten algebraischer Zahlkorper. Crelle 179,
16-21,
SCHUR, I. (1918a): tJ'ber die Verteilung der Wurzeln bei gewissen algebraischen
Gleichungen mit ganzzahligen Koeffizienten. Math. Z. 1, 377-402.
- (1929a): Einige Satze tiber Primzahlen mit Anwendungen auf Irreduzibilitats-
fragen I, II. Sitzgsber. preuO. Akad. Wiss. 1929, 125-136,370-391.
SEGRE, B. (1945a): Lattice Points in infinite domains and asymmetric diophantine
approximations. Duke Math. J. 12, 337-365.
SIEGEL, C. L. (1935 a): tJ'ber Gitterpunkte in konvexen Korpern und ein damit
zusammenhangendes Extremalproblem. Acta Math. 65, 309- 323.
(1940a): Einheiten quadratischer Formen. Abh. Hans. Univ. 13, 209-239.
(1945a): A mean value theorem in the geometry of numbers. Ann. of Math.
46,340-347.
SWINNERTON-DYER, H. P. F. (1953a): Extremal lattices of convex bodies. Proc.
Cambridge Phil. Soc. 49, 161-162.
(1954a): Inhomogeneous lattices. Proc. Cambridge Phil. Soc. 50, 20-24.
(1954b): The inhomogeneous minima of complex cubic norm forms. Proc.
Cambridge Phil. Soc. 50, 209-219.
SYLVESTER, J. J. (1892a): On arithmetical series. Messenger of Maths. 21,1-19,
87-120 (= Collected works 4,687-731).
TORNHEIM, L. (1955a): Asymmetric minima of quadratic forms and asymmetric
diophantine approximation. Duke Math. J. 22, 287-294.
TSCHEBOTAREW, N. (1934a): Beweis des Minkowskischen Satzes tiber lineare in-
homogene Formen [Russian]. Ucen. Zapiski Kazansk. Gos. Univ. 94, 3-16.
There is a German version in Vjschr. naturforsch. Ges. Ztirich 85 (1940), Bei-
blatt (Festschrift Rudolf Fueter) 27 - 30.
VARNAVIDES, P. (1948a): On lattice points in a hyperbolic cylinder. J. Lond.
Math. Soc. 23,195-199.
VENKOV, B. A. (1945a): On the extremal problem of Markov for indefinite ternary
quadratic forms [Russian]. Izv. Akad. Nauk SSSR. (Ser. Mat.) 9, 429-494.
 Index 343

VORONOI, G. (1907a): Sur quelques propriet6s des formes quadratiques positives

parfaites. Crelle 133, 97 -178.
- (1908a): Recherches sur les paralleloedres primitifs I, II. Crelle 134,198-287;
136,67-181 (1909).
WAERDEN, B. L. VAN DER (1956a): Die Reduktionstheorie der positiven quadrati-
schen Formen. Acta Math., Stockh. 96, 265-309.
WATSON, G. L. (1956a): The covering of space by spheres. Rend. Circ. Mat.
Palermo (2) 5, 1 - 8.
WElL, A. (1951 a): L'integration dans les groupes topologiques et ses applications
(Actualites Sci. Ind. 869-1145). Paris: Hermann.
WEYL, H. (1940a): The theory of reduction for arithmetical equivalence. Trans.
Amer. Math. Soc. 48,126-165; 51, 203-231 (1942).
- (1942a): On geometry of numbers. Proc. Lond. Math. Soc. (2) 47, 268-289.
WHITWORTH, J. V. (1948a): On the densest packing of sections of a cube. Ann.
Mat. pura appl. (4) 27,29-37.
- (1951 a): The critical lattices of the double cone. Proc. Lond. Math. Soc. (2)
53, 422-443-
WHITTAKER, E. T., and G. N. WATSON (1902a): A course of Modern Analysis.
Cambridge: Cambridge University Press.
WOLFF, K. H. (1954a): tl"ber kritische Gitter im vierdimensionalen Raum. Mh.
Math. 58, 38- 56.
WOODS, A. C. (1956a): The anomaly of convex bodies. Proc. Cambridge Phil.
Soc. 52, 406-423.
(1958a): The critical determinant of a spherical cylinder. J. Lond. Math. Soc.
33,357-368.
(1958b): On two-dimensional convex bodies. Pacific J. Maths. 8, 635-640.
(1958c): On a theorem of TSCHEBOTAREFF. Duke Math. J. Z5, 631-638.
YEH, Y. (1948a): Lattice points in a cylinder over a convex domain. J. Lond.
Math. Soc. 23,188-195.
~ILINKAS, G. (1941 a): On the product of four homogeneous linear forms. J. Lond.
Math. Soc. 16, 27-37.

Index
Commonly used symbols are listed first, followed by words and phrases in
alphabetical order
Izl viii LI (.9') 80
1I't1i 123 det('t) 123
11,11 318 e(x) = exp(2niX) 293
== (a ==b(k» 99£.n. F(A) 119
(Yl == YI(A» 194, 303 F(~) 195
{x} 207 L 124
[~ 164, 247 M(f) 26
Fm ..269 M+(j) 41
d(A) 10 m(S) 197
D(f) (for quadratic forms) 35 m(zo) 304
- (for cubic forms) 51 ptA) 305
g',$ 269 at, at/A 194
g .. 163, 246 V(.9') viii
b (F), il (F) 307 V(",) 74
d(F) 120
344 Index

admissible 6, 80 LITTLEWOOD'S problem 172

- (in sense of MAHLER) 152 local methods 301
affine transformation 19
MARKOFF chain 36
automorphs, automorphic star bodies
meet (of two point sets) 105f.n.
256
metric (in space of lattices) 130
basis (of lattice) 9 MINKOWSKI'S convex body theorem 71
BLICHFELDT'S theorem 69 - linear forms theorem 73
boundedly reducible 154
non-null (function non-null on a lattice)
class: see congruence class 261 .
compact 67 non-singular: see singular
compatible 283
octahedron (generalised) 105, 117
congruent 194, 303
orthogonal 206f. n.
congruence class 194
continued fractions 301 packing, lattice packing 223
con vergence (of lattices) 126 parallelopiped (generalised) 116
convex (point set) 2, 64 polar basis 23
- (distance function) 104 - convex body 105, 113
critical 6, 80, 141, 142 - distance function 113, 114
- (in sense of MAHLER) 152 - lattice 23
cube (generalised) 105 - transformation 26, 114
cylinder (generalised) 227 primitive (lattice point) 24, 85
proper equivalence 23
determinant viii, 5, 123
proportional to integral 261
distance function 103
divided cell 325 quotient space 194
equivalent (forms) 22, 23 red ucible 153, 154
extreme 165 reduction in sense of MINKOWSKI 28
finite type 80, 141 semi-definite 103 f. n.
fully reducible 154 signature (of quadratic form) 20
fundamental parallelopiped 69, 196 simplex (lattice constant of) 82
grid 303 singular cubic forms 51
- transformations 123
hexagon lemma (of DIRICHLET) 233
star body 84
hessian 54
- set 104, 153
homogeneous problem 1
sublattice 9
improper equivalence 23 successive mllllma (for homogeneous
infinite type 80, 141 minimum of quadratic forms) 36
infinitely many lattice points in a set (of distance function with respect to
155,298 lattice) 201
inhomogeneous problem 7 - (inhomogeneous problems) 305
invariant 51 support-plane: see tac-plane
isolation 38, 286 SYLVESTER'S lemma 188
symmetric (point set) 2, 64
Jordan-volume 175
tac-plane, tac-line 104, 115
lattice viii, 9
transference theorems 308, 313
- (inhomogeneous) 303£.n. transformation: see affine transforma-
lattice-constant viii, 64, 80
tion
linear transformation: see affine trans-
triangle inequality 66, 104
formation
length (of vector) viii, 66 vectors viii, 4
LITTLEWOOD'S principle 34 volume 68, 175
 Springer
and the
environment
At Springer we firmly believe that an
international science publisher has a
special obligation to the environment,
and our corporate policies consistently
reflect this conviction.
We also expect our business partners -
paper mills, printers, packaging
manufacturers, etc. - to commit
themselves to using materials and
production processes that do not harm
the environment. The paper in this
book is made from low- or no-chlorine
pulp and is acid free, in conformance
with international standards for paper
permanency.

Springer