Minkowski geometry is a non-Euclidean geometry in a finite number of dimensions

that is different from elliptic and hyperbolic geometry (and from the Minkowskian
geometry of space-time). Here the linear structure is the same as the Euclidean
one but distance is not "uniform" in all directions. Instead of the usual sphere
in Euclidean space, the unit ball is a general symmetric convex set. Therefore,
although the parallel axiom is valid, Pythagoras' theorem is not.
This book begins by presenting the topological properties of Minkowski spaces,
including the existence and essential uniqueness of Haar measure, followed by the
fundamental metric properties - the group of isometries, the existence of certain
bases and the existence of the L6wner ellipsoid. This is followed by character-
izations of Euclidean space among normed spaces and a full treatment of two-
dimensional spaces. The three central chapters present the theory of area and
volume in normed spaces. The author describes the fascinating geometric inter-
play among the isoperimetrix (the convex body which solves the isoperimetric
problem), the unit ball and their duals, and the ways in which various roles of
the ball in Euclidean space are divided among them. The next chapter deals with
trigonometry in Minkowski spaces and the last one takes a brief look at a number
of numerical parameters associated with a normed space, including J. J. Schaffer's
ideas on the intrinsic geometry of the unit sphere. Each chapter ends with a section
of historical notes, and the book ends with a list of 50 unsolved problems.
Minkowski Geometry will appeal to students and researchers interested in ge-
ometry, convexity theory and functional analysis.
ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS

EDITED BY G.-C. ROTA

Editorial Board
R. Doran, M. Ismail, T.-Y. Lam, E. Lutwak, R. Spigler

Volume 63

Minkowski Geometry
 ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICATIONS
4 W. Miller, Jr. Symmetry and separation of variables
6 H. Minc Permanents
11 W. B. Jones and W. J. Thron Continuedfractions
12 N. F. G. Martin and J. W. England Mathematical theory of entropy
18 H. O. Fattorini The Cauchy problem
19 G. G. Lorentz, K. Jetter, and S. D. Riemenschneider BirkhofJinterpolation
21 W. T. Tutte Graph theory
22 J. R. Bastida Field extensions and Galois theory
23 J. R. Cannon The one-dimensional heat equation
25 A. Salomaa Computation and automata
26 N. White (ed.) Theory ofmatroids
27 N. H. Bingham, C. M. Goldie, and J. L. Teugels Regular variation
28 P. P. Petrushev and V. A. Popov Rational approximation of real functions
29 N. White (ed.) Combinatorial geometries
30 M. Pohst and H. Zassenhaus Algorithmic algebraic number theory
31 J. Aczel and J. Dhombres Functional equations containing several variables
32 M. Kuczma, B. Chozewski, and R. Ger Iterative functional equations
33 R. V. Ambartzumian Factorization calculus and geometric probability
34 G. Gripenberg, S.-O. Londen, and O. Staffans Volterra integral and functional
equations
35 G. Gasper and M. Rahman Basic hypergeometric series
36 E. Torgersen Comparison of statistical experiments
37 A. Neumaier Interval methods for systems of equations
38 N. Komeichuk Exact constants in approximation theory
39 R. A. Brualdi and H. J. Ryser Combinatorial matrix theory
40 N. White (ed.) Matroid applications
41 S. Sakai Operator algebras in dynamical systems
42 W. Hodges Model theory
43 H. Stahl and V. Totik General orthogonal polynomials
44 R. Schneider Convex bodies
45 G. Da Prato and J. Zabczyk Stochastic equations in infinite dimensions
46 A. Bjomer, M. Las Vergnas, B. Sturmfels, N. White, and G. Ziegler Oriented
matroids
47 E. A. Edgar and L. Sucheston Stopping times and directed processes
48 C. Sims Computation with finitely presented groups
49 T. Palmer Banach algebras and the general theory of*-algebras
50 F. Borceux Handbook of categorical algebra I
51 F. Borceux Handbook of categorical algebra II
52 F. Borceux Handbook of categorical algebra III
54 A. Katok and B. Hassleblatt Introduction to the modern theory of dynamical systems
55 V. N. Sachkov Combinatorial methods in discrete mathematics
56 V. N. Sachkov Probabilistic methods in discrete mathematics
57 P. M. Cohn Skew fields
58 Richard J. Gardner Geometric tomography
59 George A. Baker, Jr., and Peter Graves-Morris Pade approximants
60 Jan Krajicek Bounded arithmetic, propositional logic, and complexity theory
61 H. Groemer Geometric applications of Fourier series and spherical harmonics
62 H. O. Fattorini Infinite dimensional optimization and control theory
ENCYCLOPEDIA OF MATHEMATICS AND ITS APPLICA TION S

Minkowski Geometry

A. C. THOMPSON
Dalhousie University

CAMBRIDGE
UNIVERSITY PRESS
 Published by the Press Syndicate of the University of Cambridge
The Pitt Building, Trumpington Street, Cambridge CB2 I RP
40 West 20th Street, New York, NY 10011-4211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia

© Cambridge University Press 1996

First published 1996

Library of Congress Cataloging-in-Publication Data

Thompson, Anthony c., 1937-
Minkowski geometry / AC. Thompson.
p. cm. - (Encyclopedia of mathematics and its applications:
v.63)
Includes bibliographical references.
ISBN 0-521-40472-X (hc)
1. Minkowski geometry. I. Title. II. Series.
QA685.T48 1996
516.3'74 - dc20 95-46491
CIP

A catalog record for this book is available from the British Library.

ISBN 0-521-40472-X hardback

Transferred to digital printing 2004

 CONTENTS
±ÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐ

Preface Page ix
Acknowledgements xv

o The algebraic properties of linear spaces and convex sets 1

0.1 Linear spaces 1
0.2 Convex sets 6
0.3 Notes 9

1 Norms and norm topologies 13

1.1 Norm topologies 14
1.2 The unique linear topology on Rd 27
1.3 The Hahn-Banach theorem 32
1.4 The existence and uniqueness of Haar measure 36
1.5 Notes 42

2 Convex bodies 45
2.1 Separation and support theorems 46
2.2 Support functions and polar reciprocals 48
2.3 Volumes and mixed volumes 53
2.4 Various derived metrics 60
2.5 Approximation of convex sets and the Blaschke selection
theorem 64
2.6 Notes 71

3 Comparisons and contrasts with Euclidean space 75

3.1 The Mazur-Ulam theorem 76
3.2 Normality in Minkowski space 77
3.3 The Uiwner ellipsoid 80
3.4 Characterizations of Euclidean space 85
3.5 Notes 94

v
vi Contents

4 Two-dimensional Minkowski spaces 99

4.1 Inscribed regular hexagons and other constructions 100
4.2 Sets of constant width and equichordal sets 106
4.3 Lengths of curves, perimeter of the unit ball 111
4.4 The isoperimetric problem in a Minkowski plane 118
4.5 Isoperimetric inequalities 123
4.6 Transversality 125
4.7 Radon curves 127
4.8 Notes 129

5 The concept of area and content 135

5.1 Requirements and examples 137
5.2 The role ofthe function aB 141
5.3 The properties and the normalization of I 145
5.4 The isoperimetrices that arise from Examples 5.1.4 150
5.5 Further properties of I 171
5.6 Notes 182
6 Special properties of the Holmes-Thompson definition 187
6.1 The convexity of the area function a 187
6.2 Properties of the mapping I 195
6.3 Cauchy's formula for surface areas 201
6.4 Integral geometry in Minkowski spaces 205
6.5 Bounds for the surface area of B 212
6.6 Miscellaneous properties 215
6.7 Notes 222
7 Special properties of the Busemann definition 229
7.1 The convexity of the area function a 229
7.2 Properties of the mapping I 233
7.3 Area and Hausdorff measures 237
7.4 Bounds for the surface area of B 242
7.5 Notes 245
8 Trigonometry 251
8.1 The functions cm and sm 251
8.2 The function a 258
8.3 Trigonometric formulas 260
8.4 Differentiation of the trigonometric functions 264
8.5 Notes 271
9 Various numerical parameters 275
9.1 Projection constants 276
9.2 Macphail's constant 283
 Contents vii

9.3 The inner metric 286

9.4 The girth, perimeter, inner radius and inner diameter of X 288
9.5 Five examples in Rd 293
9.6 Relationships with the Banach-Mazur distance and extreme
values 300
9.7 Notes 304
10 Fifty problems 307

References 313
Notation index 331
Author index 335
Subject index 339
 PREFACE

In choosing a title for this volume I faced two problems. Firstly, should it be
"Minkowski" or "Minkowskian Geometry"? Secondly, how could I avoid accu-
sations of false advertising from students of relativity who expect Minkowski(an)
geometry to deal with one time-like and several space-like dimensions?
In an attempt to resolve the first problem I made two very long lists, including
the following items:
Abelian group Banach space
Boolean algebra Blaschke sum, product
Euclidean geometry Fourier series
Gaussian integer Galois theory
Hamiltonian circuit Hausdorff measure, topology
Hermitian matrix Hilbert space
Newtonian mechanics Lebesgue integral
Riemannian metric von Neumann algebra

I then convinced myself that I detected two slight trends: the second list tended to
predominate in more recent times, and this tendency was less pronounced in the
applied or physical areas. My proposed solution to both problems is to suggest
the use of Minkowski geometry for the present topic, i.e. the theory of finite di-
mensional normed linear spaces, which (as Dunford and Schwartz [130], p. 372,
say) "is primarily due to Minkowski", and to use Minkowskian geometry for that
other creation of Minkowski 's, the theory of linear spaces with an indefinite inner
product. To the linguistic purists, I apologize for the juxtaposition of the terms
Euclidean geometry and Minkowski geometry in several places.
Space to Euclid and Newton was uniform and "isotropic" - the same in all di-
rections. Such a notion flies in the face of daily experience, where the connotation
of "up" and "down" is different from that of "east" and "west". There are preferred

ix
x Preface

directions. Another good example is the preferred directions that cause crystals to
grow as polyhedra and not spherically like soap-bubbles. This book is about that
kind of geometry - a geometry in which, to someone peering in from outside, it
appears that the unit for measuring length is different in different directions and
hence unit "circles" and "spheres" are not the familiar round objects from Eu-
clidean geometry but are some other convex shape (called the unit ball). Beings
living inside the geometry would have a harder time detecting the preferred direc-
tions because yard-sticks, compasses and other measuring devices would, as they
were rotated, adjust (in the view of someone outside) to the change. Nevertheless,
internal measurements can be made to verify that this geometry is not Euclidean.
One is the circumference of a unit "circle" (or the ratio of the circumference to the
radius of any "circle"). This measurement yields a number somewhere between 6
and 8 but not usually the familiar 2rr. Moreover, this number varies as the plane
of the circle varies in space, indicating clearly that space is not uniform.
The most fundamental change from Euclidean geometry is the lack of a satis-
factory idea of "perpendicular". There are shortest distances from a point to a line
or plane. Unfortunately, this minimum may be attained by many points and the
perpendicular relationship it yields is not symmetric. If "perpendicular" and "right
angle" have gone then that staple of high school geometry, Pythagoras' theorem,
has also gone. All these topics are examined in this book. However, the core of
the book is Chapters 5, 6 and 7, which deal with the notion of surface area and the
solution of the "isoperimetric problem". This problem asks which surface, enclos-
ing a fixed volume, has the smallest surface area. The solution, the isoperimetrix,
is a fundamental object in the geometry, and it is the interplay between this object
and the unit ball which gives the geometry its special interest and flavour.
The geometry is "affine" in the sense that the properties examined are inde-
pendent of the choice of basis vectors. Thus, although we are dealing with finite
dimensional vector spaces over lR which can always be regarded as lR d , we try to
avoid coordinates whenever possible. Another view of this aspect is to say that the
properties examined are invariant under invertible linear transformations. There-
fore, as far as Minkowski geometry is concerned, all d-dimensional ellipsoids are
equivalent as unit balls. They correspond to Euclidean geometry because, with a
suitable choice of coordinates, an ellipsoid is the familiar unit sphere. Likewise all
d-dimensional parallelotopes are equivalent to the unit cube as unit ball.
The subject is a 20th century one. It originated in the late 19th century with the
work of Brunn in convexity theory and the work of the Italian school (see Monna's
book [391]) on abstract linear spaces. These strands were brought together by
Minkowski. In part, Minkowski geometry shares its development with the theory
of infinite dimensional normed spaces (Banach spaces) and, in part, it has much in
common with the theory of convexity. The subject received large impetus from the
work of Herbert Busemann in the middle of the century. He showed how significant
the solution to the isoperimetric problem is for the geometry of the space and how
 Preface xi

the roles of the unit ball in Euclidean space are shared between the unit ball and the
convex body that is the (normalized) solution to the isoperimetric problem. The
connections with convexity theory have become stronger in recent years with the
development of the "duality" between projection bodies and intersection bodies
which Erwin Lutwak has vigorously pursued. Since Dvoretzky's theorem [133]
appeared, the connections with Banach space theory have largely revolved around
the "local theory" of Banach spaces.
The emphasis throughout the book is on geometry in a classical sense. There
is, for example, frequent use of figures to illustrate an example or part of a proof.
One of Busemann's aims was to develop enough of the theory of Minkowski
spaces in a synthetic way to enable the theory of Finsler spaces (which corre-
spond to Minkowski spaces in the same way that Riemannian spaces correspond
to Euclidean ones) to proceed with less emphasis on coordinates and analytic
expressions. I hope that the present volume helps to further Busemann's goal.
The book is aimed at readers who wish to be introduced to the geometry of
Minkowski spaces. The general level of background knowledge required is not
more than that of a senior undergraduate or beginning graduate student. For most
of the book what is needed are: (a) a familiarity with the ideas of linear algebra, in
particular the concepts of vector space, subspace, convex set, linear transformation,
linear functional and dual space; and (b) a familiarity with the ideas of metric space
topology, in particular the concepts of continuity, norm, open, closed and compact
sets.
The level of mathematical sophistication and mathematical background needed
is not, however, constant. There are three areas where I have used more than the
minimum just listed. The first is the development of Haar measure in Chapter 1.
This concept is of such fundamental significance to the subsequent development
that I thought a rather full treatment should be included. In ]Rd, Haar measure is
usually viewed as a multiple of Lebesgue measure and because Lebesgue measure
is generally developed with the tools of Euclidean geometry in hand, I felt it would
be a good idea to make clear that those tools are not necessary for its development.
Nevertheless, I have limited the discussion to the commutative case appropriate for
]Rd. The second is the Brunn-Minkowski theory of convex bodies, which is used
heavily in Chapters 5-7. The particular topics needed are Minkowski's theory of
mixed volumes and a number of deep inequalities. This theory is introduced in
Chapter 2, where I have tried to present clear statements of the results needed but
with either no proofs or only brief outlines of proofs. The third area is the use of
multilinear algebra in parts of Chapters 5 and 6. On this topic I have included a
rather sketchy outline of the background needed. Readers who find some of the
sections on these topics difficult are advised to skip them on a first reading. It is
possible to read later sections without having absorbed all that has gone before.
A glance at the Table of Contents will give the reader a good idea of what is
discussed. Unfortunately, many topics have been left out. The main criteria for
xii Preface

deciding what to include were:

(i) the view that this is a geometry book,

(ii) the knowledge and competence of the author,
(iii) the limitations of space and
(iv) the availability of other treatments.

Some readers may be disappointed to find very little about the "local theory of
Banach spaces", butthere are treatments of this subject in, among other places, the
books by Pisier [430] and by Tomczak-Jaegermann [513]. I would have liked very
much to include some sections on differential geometry at the end of Chapter 8
but, for lack of space, have not done so. For the same reason the planned chapter on
bodies of constant width in Minkowski spaces was reduced to a section of Chapter
4. The second criterion led to the exclusion of anything about Finsler geometry
and anything about minimal surfaces with boundary. I have tried to compensate
for some of these omissions with the brief notes at the end of each chapter and the
reference list at the end of the book, but I am conscious of having an inadequate
knowledge of the literature. To those readers who find some of their favourite
topics either not covered or not sufficiently covered and citations missing from the
reference list, I apologize.
A few comments about notation. When an equation is used to define the quantity
on the left-hand side, the symbol := is used. I have tried to be systematic about
usage and letter styles because I believe that such systematization is a great help to
the reader. For example, most small Greek letters are either real numbers or real-
valued functions. Uppercase Roman letters serve a variety of purposes: some are
arbitrary sets, some (K, B) are convex sets, X and Yare linear spaces, Land Mare
subspaces and T is a linear operator. There is some duplication; e.g. P is sometimes
a polytope, sometimes a parallelotope or parallelogram and sometimes (in Chapter
9) a projection. Most lowercase Roman letters are vectors (points). Exceptions are
d (see below) and i, ... , n, which are integral indices. At times accepted usage
overcomes these conventions. For example, I have used A. for Lebesgue measure
in various dimensions including d -dimensional measure in a d -dimensional space.
But I have used V for mixed volumes, which makes Minkowski's inequality for
mixed volumes appear a little unusual. As just indicated, I have used d rather than n
for the dimension of a typical Minkowski space. This frees n for its traditional role
as a typical integer. However, since the dimension d appears in a number of integral
formulas, the reader should be on guard not to confuse it with the symbol for the
differential. The numbers d and d - I also appear frequently as superscripts to
denote the particular measure to be considered (volume or area); the reader should
be sure not to confuse this usage with an exponent. I hope the Notation Index will
be helpful.
Definitions, propositions, lemmas, theorems and corollaries are all numbered
consecutively in each section. Thus Corollary 4.3.5 is the fifth item in §4.3. The
 Preface xiii

long §5.5 has been divided into subsections but this does not affect the numbering
of items. Equations and figures are numbered consecutively in each chapter. Thus
(6.24) denotes the 24th equation in Chapter 6.
A final comment on terminology. Chapters 6 and 7 discuss the consequences
of two particular definitions of Minkowski surface area. One is due to Busemann,
or at least he is responsible for the development of the concept and much of the
theory. The other, as far as I know, first appeared in [262J. Descriptive terminol-
ogy (the intersection definition and the projection definition) did not seem helpful.
The first draft, by various circumlocutions, avoided giving names to either defini-
tion. I accepted Erwin Lutwak's criticism of the clumsiness of this solution. With
reluctance I have adopted his suggestion.
 ACKNOWLEDGEMENTS

A book of this kind is not an independent production. Many people have con-
tributed and helped over the years. I have learned mathematics from many people
in many places, too numerous to name. My colleague Michael Edelstein gave much
encouragement and advice during the years that we were together at Dalhousie.
A former colleague, Raymond Holmes, was instrumental in initiating the work
that now forms Chapter 6. Erwin Lutwak suggested I write this volume and his
encouragement at the other end of a telephone line has been indispensable.
The first draft was written mostly while I was on sabbatical leave in Freiburg
im Breisgau. Thanks are due to Dalhousie University for that leave, which was
essential for the book. It is a great pleasure to acknowledge the warm hospitality
of the people in Freiburg who made our stay there in 1990-1991 and again in
the summer of 1992 very memorable: Dr. Rolf Schneider and Dr. O. Kegel of the
Mathematisches Institut, Albert-Ludwigs Universitat, not only for their kindness
in providing office space, access to the fine library and all the ancillary services
but also for their personal hospitality; Dr. John Wieacker for many conversations
on the subject matter of this book and on many topics of wider interest over good
coffee; and the many friends my wife and I made in Freiburg.
A large number of people helped with the collection of information for the
book by sending me copies of preprints and offprints. Particularly helpful in this
respect were Richard Gardner, Helmut Groemer, Heinrich Guggenheimer and
Horst Martini, who provided drafts ofthe manuscripts for [172], [206], [227] and
[49] respectively. Erwin Lutwak and Horst Martini also assisted greatly by pointing
out many of the works included in the reference list. Peter Gruber was particularly
helpful in providing comments and references for §3.4.
Several people read all or parts of the book at various stages. I would like to
thank all of them for their help and advice. They include Rolf Clack, who was
the only one I allowed to see the very first attempts. He spent many hours trying
to make sure that the book really was aimed at the beginning graduate student,

xv
xvi Acknowledgements

that it was relatively free of non sequiturs, sadistic twists of logic and half-hidden
pitfalls. Erwin Lutwak, Rolf Schneider and Dean Tsaltas patiently read later drafts
and made numerous suggestions for improvements as well as indicating errors. I
thank them for their time and effort. The first draft was carefully and ably typed in
I5fEX by Sabine Linsenbold in Freiburg and Maria Fe Elder in Halifax. Both had
to contend with poor handwriting as well as difficult mathematical symbols and
terminology. Ms. Linsenbold deserves special credit for handling all this so well
in a foreign language. Various figures were produced using both Mathematica and
MAPLE by Brian Ingalls and Jennifer Overington (both former honours students at
Dalhousie). The remaining figures were drawn by Hugh Thompson using AutoCAD.
Keith Fordham helped me to produce the indices. Lauren Cowles at Cambridge
University Press showed remarkable patience as various deadlines passed, and the
production staff has done an excellent job of producing the volume.
The later part of the work was done with some financial support from NSERC
(Canada) under research grant 4066. I am indebted to BCS Associates (Moscow,
Idaho) and Springer (Berlin) for permission to reproduce the quotation on p. 10.
 o
The algebraic properties of
linear spaces and convex sets

This short chapter gives the background from linear algebra and the basic facts
about convex sets that are needed throughout the book. The first section describes
the notation and the conventions that will be used for concepts from linear algebra.
The substantive items concern dual spaces, dual bases, the second dual and adjoint
linear maps.
A large variety of facts about convex sets are needed in the later chapters. The
definition of a convex set and some examples of such sets are already needed
in Chapter I, where norms are introduced. Therefore, the second section of this
chapter introduces the idea of convexity via a series of examples and definitions.
However, most of the important facts about convex sets (e.g. the separation and sup-
port theorems) are more easily discussed using such topological notions as closed,
bounded, interior and continuous linear mapping. Consequently, after dealing
with topological ideas in Chapter 1, we return to a fuller study of convex sets in
Chapter 2.

0.1 Linear spaces

Throughout the book X (and occasionally Y) will be used to denote a finite di-
mensional vector space over the field lR of real numbers. The dimension of X will
usually be d. Small Roman letters v, w, x, y, z will be used for vectors and small
Greek letters ex, {3, y for scalars.
Subspaces of X will be denoted by capital Roman letters such as L, M; other
capital Roman letters (e.g. B, C, K) will be used for convex sets, and T will indicate
a linear transformation between vector spaces. The identity map is denoted by 1.
A basis for X will often be denoted by (bl, b 2, ... , bd ) and the usual basis for lR d
will always be (el, e2,· .. , ed)'
The vector space of linear functionals on X, i.e. all linear mappings from X into
lR with the usual pointwise operations, is called the dual space of X and denoted
2 The algebraic properties of linear spaces and convex sets

by X*. Elements of X* will be indicated by the letters f, g (often with subscripts).

If (b l , b2, ... , bd) is a basis for X then a linear functional is entirely determined
by its values at each b i and, on the other hand, assigning arbitrary values to each bi
determines a linear functional on X. Those linear functionals b7 (i = 1, 2, ... , d)
defined by

(i,j=l, ... ,d) (0.1)

are evidently linearly independent and span X* because feb)~ = ¢) if and only
if f = ,,£1=1 ¢i b7· Thus (bi, bi, ... , b'd) is a basis for X * and is called the dual
basis to {b l , b2, ... , hd}. If a vector x has coordinates (~l, ~2, ... , ~d) relative
to {b l ,b2, ... ,bd } and if fin X* has coordinates (¢1,¢2, ... ,¢d) relative to
{bi,bi, ... ,b'd}then

(0.2)

When we use coordinates in this way, in order to preserve the distinction between
X and X* we shall always regard elements of X as column vectors and elements
of X* as row vectors. Then an element of X* acts on an element of X by the
usual process of matrix multiplication; e.g. the element f (1, 2, 3) in (lR3 )*
evaluated at

x=G)
in lR.3 is

f(x) = (1,2,3) G) = 10. (0.3)

However, to save space, a column vector will usually be written as the transpose
of a row vector, thus:

At this point it is usual to omit the transpose, identify row and column vectors
and think of Equation (0.3) as an inner product:

( (1,2,3), (3,2,1») = 10.

In general, if (b l , b2 , ••. , hd ) is a basis for X then two vectors x and y in X have

representations
 0.1 Linear spaces 3

relative to this basis. One can identify the vector y with the linear functional
d

Iy := (111,112, ... , I1d) = L 7

l1i b
i=1

and hence define the inner product

(as in Equation (0.2».

The relationship between y and Iy can be described by saying that X and X*
are identified by means of the linear map that sends each basis vector bi in X to the
corresponding dual basis vector b7 in X*. An inner product defined in this way from
a particular choice of basis for X is called a Euclidean structure or a Euclidean
geometry on X. Since the basis (b l , b 2 , ••• , b d ) is entirely arbitrary and different
bases give rise to different Euclidean structures, this process of constructing an
inner product is rather artificial.
In the same way that a choice of basis gives rise to an inner product on X, it also
gives rise to one on X*. It follows that if x and yare vectors in X which correspond
(as above) to linear functionals Ix and Iy in X* then (x, y) = (fx, Iy). Thus the
virtue of a Euclidean structure is that the mapping from X to X* defined above
not only is a linear isomorphism but also preserves inner products and, hence, is
an isometry relative to the metrics in both X and X* generated by the two inner
products. By contrast, in a more general Minkowski geometry the map bears little
or no relation to the metric structure and is, therefore, of much less use.
The null space or kernel of a linear functional I in X* is denoted by 11-, i.e.

11- := {x EX: I(x) = OJ.

If x is in 11- then we can say that I is perpendicular to x.

With a Euclidean structure the identification of X and X* allows the relationship
of perpendicularity to be construed as a relationship between elements of X. We
say that vectors y and x are orthogonal if

(y,x) =0.

The notion of orthogonality is central to much of elementary Euclidean geometry. A

great deal of the rest of this book can be viewed as finding partial substitutes for this
relationship. In summary, then, there is always a notion of perpendicularity between
certain elements of X and X*. The notion of orthogonality between elements of
X is special to the Euclidean case. This is the prime reason why it is important
in Minkowski geometry to preserve the distinction between elements of X and
elements of X*.
Despite the foregoing discussion, for the purposes of calculation it is frequently
convenient to adopt an auxiliary Euclidean structure. By this we mean an arbitrary
4 The algebraic properties of linear spaces and convex sets

choice of basis and the inner product derived from that basis by the corresponding
identification of X and X*.
If L I is a subspace of X then, since a basis for L I can be extended to a basis
for X, we can write X as the direct sum of LI and a complementary subspace L2:

In X*, two subspaces are defined by the equations

Lf := {f E X* : fey) = 0, Vy ELi}, i = 1,2.

Proposition 0.1.1 With the preceding notation, we have X* = LtEBLt andLj, L2

are isomorphic to Lt, Lt respectively.

Proof First, it is clear that Lt n Lt = {OJ. If f E X* and x E X then x =

XI + X2 for unique elements Xi in L i . Define /; on X by fi(X) := f(Xi). Then
fl E Lt, 12 E Lt and f = 12+ II sothatX* = LtEBLt. Eachfunctionalg E Lt
can be regarded as an element g' of Lj by setting g'(xI) := g(XI), i.e. g' = giLl.
Conversely, each functional g' E Lj can be extended to an element g of Lt by
letting g(xI + X2) = g'(xI). It is easy to show that this defines an isomorphism
between L j and Lt and, similarly, for L2 and Lt. •

We repeat and extend the definition of f 1- •

Definition 0.1.2 If f E X* then the subspace of X that is annihilated by f

is denoted by f1-, i.e. f1- := {x EX: f(x) = OJ. Similarly, if x E X then
x1- := {f E X* : f(x) = OJ. Furthermore, each translate of the kernel of a
non-zero linear functional, i.e. {x EX: f(x) = a}, is called a hyperplane in X.

Having defined X* and seen that it is another vector space, it is logical to

consider X** := (X*)*. Each vector x in X, however, defines an element J x in
X** by

(Jx)(f) := f(x). (0.4)

If J x = 0 then f (x) = 0 for every linear functional f in X* and hence x = O.

Thus J : X r-+ X** is an injection. Since, for finite dimensional spaces, X and X*
have the same dimension, it follows that so do X and X** and hence that J maps
X onto X**. Therefore, X and X** are isomorphic. Moreover, the isomorphism
J is natural in the sense that it is independent of basis. In fact, J can be defined
for arbitrary vector spaces and it is a surjection if and only if the space is finite
dimensional. We shall see later that J is also natural with respect to the norms
on X and X** in that it is always an isometry. There is no problem, therefore, in
always identifying X and X**. For this reason we shall sometimes use a pairing
 0.1 Linear spaces 5

notation (j, x) to denote the value of f at x in order to emphasize the symmetry

between f and x in this relationship. Using this notation we shall write (x, f) in
place of (lx, f).
If T is a linear transformation between vector spaces X and Y then there exists
a dual transformation T* between y* and X* defined by

(T* f, x) := (j, Tx) (0.5)

for all f in y* and x in X.

If, in this situation, T is an isomorphism and (b l , b2, ... , bd) is a basis for X
then (Tb l , Tb 2, ... , Tb d) is a basis for Y. The dual basis (br, bi, ... , b'd) for X*
is defined by Equations (0.1). Since T*T*-Ib; = b; we have

(T*T*-lb7, b j ) = 8ij

and hence, using (0.5),

Thus (T*-Ibi : i = 1,2, ... , d) is the basis in y* dual to (Tbl, Tb 2 , ... , Tbd) in
Y.
Equation (0.5) provides another reason for writing elements of X as column
vectors and elements of dual spaces as row vectors. If the spaces X and Y have
dimensions d and n respectively and if bases for X and Y have been chosen then,
relative to these bases, a vector x in X has a representation

and T has a representation as an n x d matrix M T. Then T x is represented by

If y* is given the basis dual to that for Y then f in y* is represented as

and hence, from (0.2), the right-hand side of (0.5) is

(0.6)

It follows that this same matrix product, with the multiplication associated differ-
ently, must also represent the left-hand side of (0.5). Thus

(0.7)

and the matrix for T* (relative to the two dual bases) is the same matrix MT but
written on the right of the row vectors.
6 The algebraic properties of linear spaces and convex sets

0.2 Convex sets

Schneider remarks in [475] that it is surprising how rich a theory is generated by
the simple definition of convexity. In this section we begin with the definition and
then give a list of examples of convex sets. This is followed by various related
definitions. The part of the theory of convexity that we shall need is postponed
until Chapter 2 after we have discussed norms and norm topologies. Nevertheless,
a few topological words (closed and interior) creep in at the end ofthis section. It
may be supposed that the finite dimensional space under consideration is equipped
with an inner product which generates a Euclidean metric and that these terms refer
to the topology derived from that metric. This will be clarified in the next chapter.

Definition 0.2.1 A non-empty subset K of a linear space X is said to be convex if

ax + (1 - a)y is in K whenever x and yare in K and 0 ::S a ::S 1.

The set {ax + (l - a)y : a E [0, I]) is called the line segment joining x and y
and is denoted by [x, y]. Thus K is convex if it contains the line segment joining
any two of its points.
The vector operations on X can be extended to the collection of convex sets as
follows:

K, + K 2 := {x: x =x, +X2, Xi E K i },

aK I := {x : x = ay, y E Kd.

It is easy to see that if K, and K 2 are convex, so are all linear combinations of K I
and K 2 •
It is also clear that the associativity and commutativity of the vector operations
imply the same for the extended operations. Furthermore, we have

and, if a ~ 0 and .B ~ 0,

(a + .B)K = aK + .BK.
The first of these is easy. For the second, if y = (a + .B)x E (a + .B)K then
y = ax + .Bx is also in aK + .BK. On the other hand, if y = aXI + .Bx2 with
a,.B ~ 0 then y = (a + .B)[a/(a + .B)XI + .B/(a + .B)X2] is in (a + .B)K because
a(a + .B)-IX, + .B(a + .B)-I X2 E [x" X2] S; K.
Because this distributive law applies only to non-negative scalars, the linear
combinations of convex sets we consider will usually have non-negative coeffi-
cients. Also, for many of the convex sets that we shall consider, multiplication by
negative scalars is irrelevant because they are symmetric in the following sense.
 0.2 Convex sets 7

Definition 0.2.2 A set K (which need not be convex) is said to be symmetric (with
respect to 0) if (-l)K = K.

The unqualified adjective" symmetric" will always mean "symmetric with respect
to 0".

Example 0.2.3 (convex sets)

(i) A single point {x} is convex.

(ii) Any subspace M of X is convex. Hence every translate of a subspace (i.e.
a set of the form M + {x}) is convex. Such a set is called afiat. The dimension of
a flat is the dimension of the subspace of which it is a translate. Each hyperplane
in X is a flat of dimension d - 1.
(iii) A line segment [x, y] is convex.
(iv) A sum of line segments is convex.
In particular, if {XI, X2, ... , xd is a linearly independent set in X then 2::7=1 [0, x;]
(or just 2::[0, Xj]) is convex and is called the parallelotope spanned by the vectors
Xj. This idea is especially important when the vectors form a basis.
More generally, if {Vj : i = 1, 2, ... , n} is an arbitrary set of vectors we let
[vd := [-2- l vj, 2- l vd (so that[vd is symmetric and of the same length as [0, v;])
and then let
n
Z[vd7=1 = Z[vJ, V2, ... , vn ] := ~)v;]. (0.8)
j=1

This convex set is called the zonotope spanned by the vectors Vj. Each zonotope is
symmetric. When {XI, X2, ... , xd is linearly independent, the zonotope 2::7=1 [x;]
is a translate of the parallelotope 2::7=1 [0, x;] by the vector -2- 1 2::7=1 Xj.
The class of zonotopes will play an important role in what follows.
(v) If {XI, X2, ... , Xk} does not lie in any (k - 2)-dimensional flat in X then
°
we call the set {y EX: y = 2::7=1 otjXj with otj 2: and 2::7=1 otj = I} the simplex
spanned by the vectors Xj and use the notation S[XI, X2, ... ,Xk] or S[Xj ]f=1 for
°
this set. In ~d, the special simplex S[O, el, e2, ... , ed] with vertices at and the
standard basis vectors will be denoted by Sd. It is straightforward algebra to show
that a simplex is convex.

The example of a simplex is a special case of a general phenomenon. An

arbitrary intersection of convex sets is again convex, for if both x and y are in
every member of the intersection so is the line segment joining them. Hence if A
is a subset of X then we may consider first the family of all convex sets containing
A (this is non-empty since it contains X itself) and then the intersection of this
family. This is the smallest convex set containing A and is called the convex hull
8 The algebraic properties of linear spaces and convex sets

of A and denoted by conv A. One can show that

conv A = {y EX: y = taixi; ai :::: 0, t a i = 1, Xi E A, n arbitrary}.

(0.9)
It follows that S[Xl, X2, ... , xn] = conv{xl, X2, ... , x n }.
It is an important theorem of Caratbeodory [90] that if dim X = d then the sums
in (0.9) can be restricted by considering only n = d + 1. In other words, conv A is
the union of all those simplices with vertices in A.
We now return to the list of examples of convex sets.

Example 0.2.4

(i) If {VI. V2, ... , Vn} is an arbitrary finite set of vectors in X then the convex
setconv{vl, V2, ... , Vn} is calledapolytope. In particular, a simplex and a zonotope
are both polytopes. Throughout the book a polytope is a convex set.
(ii) If f E X* and if a is a real number then the closed half-space {x EX:
f(x) :::: a} and the open half-space {x EX: f(x) > a} are both convex. We
shall use the notation H[f, a]+ to denote the closed half-space and H(f, a)+ for
the open half-space. If the functional f is obvious we may suppress reference to
it, and when a = 0 we may sometimes write H+ for the set {x EX: f(x) :::: O}.
It follows that every intersection of half-spaces is also convex. In Chapter 2
we shall show that the following converse of this is true: each closed convex set is
an intersection of closed half-spaces.

Definition 0.2.5 The dimension ofa convex set K is the dimension of the smallest
flat containing K.

An alternative approach is first to define the dimension of a simplex S[Xl, X2,

... , Xk] to be (k - 1); (recall that in Example 0.2.3(v) {Xl, X2, ... , Xk} does not lie
in any (k - 2)-dimensional flat so this definition is in agreement with Definition
0.2.5). Then the dimension of a convex set K is the dimension of a maximal
simplex contained in K.
Since a simplex has non-empty relative interior (i.e. interior points relative to
the smallest flat containing it) the same is true of an arbitrary convex set. It follows
that there is no real loss of generality in supposing that a single convex set has
non-empty interior since we can usually restrict discussion to the flat containing
the set.

Definition 0.2.6

(i) A convex subset E of a convex set K is said to be an extreme subset of K

if, whenever xE E and x = aXl + (1 - a)x2 with Xl, X2 E K and a E (0, 1), then
Xl,X2 E E.
 0.3 Notes 9

(ii) If P is a polytope then the extreme subsets of P are called the faces of
P. If P is of dimension d then the (d - I)-dimensional faces are called the facets
and the O-dimensional faces are called the vertices of P .

Example 0.2.7 If K is a cube in]R3, then the extreme subsets (faces) of K are (i)
K itself (of dimension 3); (ii) the facets of K (of dimension 2); (iii) the edges of
K (of dimension 1); and (iv) the vertices of K (of dimension 0).

Definition 0.2.8 A point x of K such that {x} is an extreme subset of K is called

an extreme point of K .

If P is a polytope, the extreme points of P coincide with the vertices of P . Every

closed and bounded convex set K has a relative abundance of extreme points, in
the sense that such a set is the convex hull of its extreme points. In other words,
each point of K can be represented as a point of a simplex whose vertices are
extreme points of K.

Definition 0.2.9 A closed convex set is said to be strictly convex if its only extreme
sets (other than itself) are extreme points (i.e. if there are no line segments - or
larger extreme sets - in the boundary).

0.3 Notes
Although the concepts from linear algebra are well known, their origins and his-
tory are not. A very readable account of the development of the idea of a linear
space is contained in Chapter 2 of Monna's book, Functional Analysis in Histor-
ical Perspective [391]. The chief names and dates he gives are Bolzano (1804),
Laguerre (1867), Grassmann (1844 and 1847) but especially the book Calcolo
geometrico . .. [404], written by Peano in 1888, which, in Chapter 9, gives the first
definition of a linear space.
An inner product on a real vector space can also be introduced axiomatically
as a bilinear functional which is symmetric and positive definite. It is not always
immediately obvious that, in the finite dimensional case, every such functional
arises from a choice of basis in the way described in §O.l. As an example, consider
the usual inner product

(p, q) := 10 1
p(x)q(x) dx

defined on the d-dimensional space of real polynomials of degree not exceeding

(d -1). However, given an inner product, the familiar Gram-Schmidt orthogonal-
ization process produces a basis with respect to which the inner product conforms
to the situation described in §O.l.
10 The algebraic properties of linear spaces and convex sets

For an infinite dimensional, complete, inner product space X the Riesz repre-
sentation theorem asserts that the dual space (now consisting of all the continuous
linear functionals on X) is isomorphic to the space itself. This theorem is due,
independently, to both Riesz [440] and Frt!chet [163, 164] in 1907. Their proofs
were for the space L2[0, 1]; for the proof for an abstract inner product space see
Stone [495], p. 62, or Riesz [443].
For a more detailed account of dual bases and the isomorphism between X
and X** see, e.g., the book by Halmos [240]. In particular, Proposition 0.1.1 can
be found in §20. A full discussion ofthe place finite dimensional spaces occupy
among all vector spaces, including the theorem that X is finite dimensional if and
only if it is isomorphic to its second dual, is contained in Kothe [301], §9.5.
The origins of the study of convexity are clearer. In 13 pages Gruber [213] gives
a masterly account of the history of the subject, and there is another good survey
of a similar length by Fenchel [150]. However, for a succinct summary one cannot
do better than quote the initial paragraph of the foreword to Bonnesen and Fenchel
[51]:

Convex figures have always played an important role in geometry. Brunn however was the
first to make a comprehensive investigation of those figures that are characterized by the
convexity property alone. In two works: "Uber Ovale und Eiftachen" [60] and "Uber Kurven
ohne Wendepunkte" [61] that appeared in 1887 and 1889 he proved, along with a variety
of results about convex regions and bodies, a theorem on the areas of the intersections of a
COnvex body with parallel planes. This theorem subsequently turned out to be fundamental.
The significance of this theorem was emphasized by Minkowski. In several works, in par-
ticular in "Volumen und Oberftache" [388] (1903) and "Zur Theorie der konvexen Korper"
[390] which was laid out on a large scale but remained unfinished, Minkowski introduced
basic concepts such as support fUnction, mixed volume, and so on. Here he created the
fonnal tools appropriate for problems about convex regions and bodies. Minkowski above
all opened the way to various applications, especially to the isoperimetric (isepiphane) and
other extremal problems. Furthennore, he discovered the close connection of these concepts
and theorems with the question of detennining convex surfaces by means of their Gaussian
curvature and proved profound theorems in this regard.

Following Minkowski were, among others, Caratbeodory [90], whose basic

theorem on convex hulls we mentioned, Radon [433, 434] and Helly [255]. The
latter two both proved theorems of a combinatorial type which are closely related
to that of Caratheodory. Radon proved that every set containing (d + 2) points in
]R.d can be expressed as the union of two disjoint sets whose convex hulls have a
common point. Helly's theorem states that if AI. A2, ... , Ak are convex sets in
]R.d and if every (d + 1) of these sets have a common point then the intersection
of all of them is non-empty. These three theorems and the connections between
them can be found in Schneider [479]. The article by Danzer, Griinbaum and
Klee [120] is a good survey of these theorems and subsequent generalizations.
This article begins with an interesting short biography of Helly which explains the
 0.3 Notes II

circumstances during and after the First World War that led to ReIly's theorem,
proved in 1913, being first published by Radon [433]. A very recent survey with
up-to-date material on these theorems is the article by Eckhoff [135].
The result that a closed, bounded convex set is the convex hull of its extreme
points is, for finite dimensional spaces, due to Minkowski [390]. The generalization
of this result to infinite dimensional topological vector spaces states that a compact
convex set is the closed convex hull of its extreme points and is due (for normed
spaces) to Krein and Milman [305]. This result can be regarded as the starting
point for the Choquet theory of convex sets, the aim of which is to represent points
of a convex set by means of integrals with respect to a measure located on the
"boundary" of the set (see, e.g., Phelps [427]).
 1
Norms and norm topologies

Minkowski geometry is the study of the interplay between two structures. The
first is a finite dimensional, real linear structure and the second is a metric struc-
ture derived from a norm. The preceding chapter was concerned with the linear
structure. However, before we look at the detailed metric properties derived from
a specific norm, we turn to the more general topological properties induced by a
norm. The essential fact is that there is only one Hausdorff linear topology with
which a finite dimensional space can be endowed. This implies that between any
two d-dimensional Minkowski spaces there is a linear homeomorphism or, more
briefly, an isomorphism. Any property of the one space which depends only on
the topology and the linear structure carries over to the other via the isomorphism.
Stated more succinctly, the isomorphic theory of Minkowski spaces is rather unin-
teresting; the dimension is the only isomorphic invariant and all Minkowski spaces
are completely classified up to isomorphism by this invariant. On the other hand,
as we shall see, the isometric classification is too complicated. There is a vast
array of non-isometric Minkowski spaces, even in two dimensions. Perhaps this
is why Minkowski spaces have received less attention than they deserve - on the
one hand, they are too simple and, on the other, too complicated.
The chapter has four sections and, for the most part, the ideas discussed are
the familiar ones presented in a course on metric space topology or at the begin-
ning of a course on functional analysis. The exception is the material in the final
section.
The first section gathers together the elementary facts about norms and the
metric space topologies they generate. Important relationships discussed are, first,
that between a norm and its unit ball; second, that between continuity and bound-
edness for linear maps; and third, that between a norm and the dual norm on the
dual space. The section ends with a variety of examples of norms on ~d.
The second section is devoted to showing that the topology on any d-dimen-
sional Minkowski space is the familiar Euclidean topology. This is the first of

13
14 Nonns and norm topologies

several uniqueness results that make finite dimensional spaces special. All these
results can be thought of as consequences of the local compactness of Minkowski
spaces.
Since the topology is the familiar Euclidean one, it follows that closed and
(Euclidean) bounded sets are compact (in the Minkowski norm topology). In fact,
more is true: any two norms on a d -dimensional space are equivalent. This implies
that not only the topologies but also the uniformities are the same. Thus, the
collection of bounded sets is the same for all norms and we can say that a set is
compact if and only if it is closed and bounded. Moreover, the collection of Cauchy
sequences is independent of the norm and all Minkowski spaces are complete.
Another consequence is that all linear maps between finite dimensional spaces are
continuous and, in particular, the algebraic dual space and the topological dual
space coincide.
The centrepiece of the third section is the Hahn-Banach theorem on extending
a continuous linear functional from a subspace to the whole space. The important
point here is not the existence of a continuous extension - that is straightforward
because all linear maps are continuous - but that there is one which preserves
the norm. This theorem has a variety of consequences dealing with supporting
and separating hyperplanes, which will come in Chapter 2. The most important
consequence for the third section is that every finite dimensional normed space is
isometrically isomorphic to its second dual space.
The final section contains less standard material. It deals with another unique-
ness theorem stemming from the local compactness. On lRd there is only one
Hausdorff linear topology, only one linear uniformity and, up to a scalar normal-
izing factor, only one translation-invariant measure - Haar measure. The essential
ideas for this result are from group theory. The space (lR d , +) is a commutative,
locally compact, topological group. The central chapters of the book (Chapters 5-
7) will depend fundamentally on this result. It is often convenient (especially when
making detailed calculations using Haar measure) to suppose that a Minkowski
space has an auxiliary Euclidean structure. We would like to emphasize, however,
that this is only a convenience and a Euclidean structure is not inherent in the de-
velopment. For this reason, at this point we avoid talking about Lebesgue measure
and give an account of the construction of Haar measure.

1.1 Norm topologies

Definition 1.1.1 A norm on a real linear space X is a mapping 11.11 from X into
lR+ that satisfies the following axioms:

(i) Ilxll:::: o with equality ifandonly ifx =0,

(ii) Ilaxll = lalllxll(a E lR),
(iii) Ilx + yll :s Ilxll + Ilyll·
 1.1 Norm topologies 15

The last condition is called the triangle inequality. From the norm we can define
a metric 8 (x, y) := Ilx - y II and then the triangle inequality for 8,

8(x, z) ::::: 8(x, y) + 8(y, z),

follows directly from (iii) because

IIx - zll = II (x - y) + (y - z)1I ::::: IIx - yll + lIy - zll·

Condition (i) implies that 8 (x, y) :::: 0 with equality if and only if x = y and the
symmetry of 8 follows from (ii) with a = -I. Thus 8 is a metric which generates
a topology on X called the norm topology. We shall denote this metric space by
(X, 11.11). A Minkowski space is a pair (X, 11.11) in which X is finite dimensional.

Definition 1.1.2 The unit ball B = B[O, 1] in (X, 11.11) is the set
B[O, 1] := {x EX: IIxll ::::: I}.

We shall use Bx if the dependence on X needs to be explicit. More generally,

B[x, p] := {y EX: IIx - yll ::::: pl. The unit sphere in (X, 11.11) is the boundary
of the unit ball, aB. Thus

aB := {x EX: IIxll = I}.

Definition 1.1.3 An isometry from the normed space (X, 1I.lIx) to (Y, 1I.lIy) is a
mapping T such that for all x I, X2 in X we have

Proposition 1.104 If (X, 11.11) and (Y, 11.11) are two Minkowski spaces of the same
dimension and ifT is a linear isometry of X onto Y then
(i) TBx=B y ,
(ii) T(aB x ) = aB y ,
(iii) ifF is an extreme subset of Bx then T (F) is an extreme subset of By.

Proof Since a linear map sends 0 in X to 0 in Y, (i) and (ii) are evident from the
preceding definitions.
Now suppose that F is an extreme subset of Bx and suppose y E T(F) with
y = aYI + (l - a)Y2 for some YI, Y2 E By. Then y = Tx for x E F and, by
(i), Yi = T Xi for some Xi E B x (i = 1, 2). Since T ~ I is linear these equations
imply x = aXI + (l - a)x2 and therefore, by Definition 0.2.6, Xi E F and hence
Yi E T (F) as required. •
16 Norms and norm topologies

Proposition 1.1.5 fl(X, 11.11) is a Minkowski space then

(i) each translation is an isometry 01 (X, 11.11);

(ii) the mappings 1 and -1 are isometries 01 (X, 11.11).

Prool
(i) For all x, y in X we have

8(x, y) = Ilx - yll = II (x + z) - (y + z)11 = 8(x + z, y + z).

(ii) That 1 is an isometry is clear and for -1 we have

8(-x, -y) = II - x + yll = II (-1)(x - y)1I = IIx - yll = 8(x, y). •

It is possible that the mappings listed in Proposition 1.1.5 (and compositions of

-1 with a translation) are the only isometries of (X, 11.11). For example, this is so
if]R2 is given the norm

{ 1~21
1~11+2-11~21 ifsgn~1=sgn~2 and 1~21::::21~tI,
IIxll = 1I(~1'~2)111:= ifsgn~1 =sgn~2 and 1~21 > 21~11,
(~f +~f)'/2 ifsgn~1 = -sgn~2.

The unit ball for this norm is shown in Figure 1.1. If XI := (1,0)1, X2 := 1)1 (!,
and X3 := (0, 1)1 then aB consists of the line segments [XI, X2], [X2, X3] and the
circular arc from X3 to -XI and the images of these under -1.

Figure 1.1
 1.1 Norm topologies 17

To prove that the only isometries are compositions of those listed in Proposition
1.1.5, we first need the Mazur-Ulam theorem (see 3.1.2 below), which states that
any isometry of X onto itself that leaves the origin fixed is necessarily linear. Hence
any isometry of a Minkowski space is a composition of a translation and a linear
isometry. Proposition 1.1.4 implies that a linear isometry T maps line segments in
aB to line segments of the same length. Since IIxl - x211 = v's/2 and IIx2 - x311 = 4
it follows that T Xi = ±Xi and T = ±1.

Proposition 1.1.6 The unit ball B in a Minkowski space (X, 11.11) is symmetric
and convex. Moreover, with respect to the norm topology, B is closed and 0 is an
interior point.

Proof The symmetry of B is evident from Definition 1.1.1 (ii) with a = -1. If
x, Y E B and 0 :::: a :::: 1 then

Ilax + (1 - a)yll :::: allxll + (1 - a)llyll :::: a + (l - a) = 1.

This follows from Definition 1.1.1 (i) and (ii) and the facts that IIx II :::: I, II y II :::: 1.
Thus ax + (1 - a)y E Band B is convex.
The triangle inequality implies that Ilix II - lIy III :::: IIx - y II and hence the norm
is continuous as a mapping from (X, 11.11) to R Since B is the inverse image of a
closed set, it is closed. Likewise {x : IIx II < I} is open and so 0 is an interior point
of B. •

Conversely, if B is a symmetric convex set in X with the property that each line
through 0 meets B in a closed, bounded segment of positive length then we may
define a functional 11.11 B on X by

IIxliB := inf{~ E lR.+ : x E ~B}.

Definition 1.1.7 Thefunctionalll.IIB is called the Minkowskifunctional of B.

Proposition 1.1.8 If B is a symmetric convex set in X such that each line through
o meets B in a non-trivial, closed, bounded segment then 11.11 B is a norm on X.

Proof From the definition, IIxllB ::::: O. If xt:-O then the ray joining 0 to x meets
B in a segment [0, ax] with 0 < a < +00. Hence IlxliB = a-I and so IIxllB is
a strictly positive real number. That IlthllB = flllxllB for fl ::::: 0 follows readily
from the definition and the symmetry of B implies that II - x II B = IIx II B.
If x, y are non-zero elements of X then x' := x/llxllB and y' := y/IIYIIB
are in B (because the line through 0 and x meets B in the closed segment