Mathematics: Form and Function

Saunders Mac Lane

Mathematics
Form and Function

With 116 Illustrations

Springer-Verlag
New York Berlin Heidelberg Tokyo
Saunders Mac Lane
Department of Mathematics
University of Chicago
Chicago, Illinois 60637
U.S.A.

AMS Classifications: 00-01, 00A05, 00A06, 03A05

Library of Congress Cataloging in Publication Data

Mac Lane, Saunders
Mathematics, form and function.
Bibliography: p.
Includes index.
1. Mathematics-1961- . I. Title.
QA39.2.M29 1986 510 85-22160

© 1986 by Springer-Verlag New York Inc.

Softcover reprint of the hardcover 1st edition 1986

All rights reserved. No part of this book may be translated or reproduced in any
form without written permission from Springer-Verlag, 175 Fifth Avenue, New
York, New York 10010, U.S.A.

Typeset by House of Equations Inc., Newton, New Jersey.

9 8 7 6 5 432 I

ISBN-13: 978-1-4612-9340-8 e-ISBN-13: 978-1-4612-4872-9

DOl 10.1007/978-1-4612-4872-9
Preface

This book records my efforts over the past four years to capture in words a
description of the form and function of Mathematics, as a background for
the Philosophy of Mathematics. My efforts have been encouraged by lec-
tures that I have given at Heidelberg under the auspices of the Alexander
von Humboldt Stiftung, at the University of Chicago, and at the University
of Minnesota, the latter under the auspices of the Institute for Mathematics
and Its Applications. Jean Benabou has carefully read the entire
manuscript and has offered incisive comments. George Glauberman, Car-
los Kenig, Christopher Mulvey, R. Narasimhan, and Dieter Puppe have
provided similar comments on chosen chapters. Fred Linton has pointed
out places requiring a more exact choice of wording. Many conversations
with George Mackey have given me important insights on the nature of
Mathematics. I have had similar help from Alfred Aeppli, John Gray, Jay
Goldman, Peter Johnstone, Bill Lawvere, and Roger Lyndon. Over the
years, I have profited from discussions of general issues with my colleagues
Felix Browder and Melvin Rothenberg. Ideas from Tammo Tom Dieck,
Albrecht Dold, Richard Lashof, and Ib Madsen have assisted in my study
of geometry. Jerry Bona and B. L. Foster have helped with my examina-
tion of mechanics. My observations about logic have been subject to con-
structive scrutiny by Gert Miiller, Marian Boykan Pour-El, Ted Slaman,
R. Voreadou, Volker Weispfennig, and Hugh Woodin. I have profited from
discussions of philosophical issues with J. L. Corcoran, Philip Kitcher, Leo-
nard Linsky, Penelope Maddy, W. V. Quine, Michael Resnik, and Howard
Stein. Some of my earlier views on various issues have been constructively
examined by Joel Fingerman, Marvin 1. Greenberg, Nicholas Goodman,
P. C. Kolaitis, J. R. Shoenfield, and David Stroh. I am grateful to all those
people-and to a number of others-even in the numerous cases where I
have not followed their advice.
Vl Preface

My wife, Dorothy Jones Mac Lane, enthusiastically supported the whole

project from its inception. Her encouragement was vital.
Springer-Verlag in general, and the editorial staff in particular, has pro-
vided steady support in the process oftuming my manuscript into a book.
Dune Acres, Indiana SAUNDERS MAC LANE
July 4, 1985
Contents

Introduction
CHAPTER I
Origins of Formal Structure 6
I. The Natural Numbers 7
2. Infinite Sets 10
3. Permutations 11
4. Time and Order 13
5. Space and Motion 16
6. Symmetry 19
7. Transformation Groups 21
8. Groups 22
9. Boolean Algebra 26
10. Calculus, Continuity, and Topology 29
1I. Human Activity and Ideas 34
12. Mathematical Activities 36
13. Axiomatic Structure 40

CHAPTER II
From Whole Numbers to Rational Numbers 42
I. Properties of Natural Numbers 42
2. The Peano Postulates 43
3. Natural Numbers Described by Recursion 47
4. Number Theory 48
5. Integers 50
6. Rational Numbers 51
7. Congruence 52
8. Cardinal Numbers 54
9. Ordinal Numbers 56
10. What Are Numbers? 58
Vlll Contents

CHAPTER III
Geometry 61
I. Spatial Activities 61
2. Proofs without Figures 63
3. The Parallel Axiom 67
4. Hyperbolic Geometry 70
5. Elliptic Geometry 73
6. Geometric Magnitude 75
7. Geometry by Motion 76
8. Orientation 82
9. Groups in Geometry 85
10. Geometry by Groups 87
II. Solid Geometry 89
12. Is Geometry a Science? 91

CHAPTER IV
Real Numbers 93
I. Measures of Magnitude 93
2. Magnitude as a Geometric Measure 94
3. Manipulations of Magnitudes 97
4. Comparison of Magnitudes 98
5. Axioms for the Reals 102
6. Arithmetic Construction of the Reals 105
7. Vector Geometry 107
8. Analytic Geometry 109
9. Trigonometry 110
10. Complex Numbers 114
II. Stereographic Projection and Infinity 116
12. Are Imaginary Numbers Real? 118
13. Abstract Algebra Revealed 119
14. The Quaternions-and Beyond 120
15. Summary 121

CHAPTER V
Functions, Transformations, and Groups 123
I. Types of Functions 123
2. Maps 125
3. What Is a Function? 126
4. Functions as Sets of Pairs 128
5. Transformation Groups 133
6. Groups 135
7. Galois Theory 138
8. Constructions of Groups 14?
9. Simple Groups 146
10. Summary: Ideas of Image and Composition 147
Contents IX

CHAPTER VI
Concepts of Calculus 150
I. Origins 150
2. Integration 152
3. Derivatives 154
4. The Fundamental Theorem of the Integral Calculus 155
5. Kepler's Laws and Newton's Laws 158
6. Differential Equations 161
7. Foundations of Calculus 162
8. Approximations and Taylor's Series 167
9. Partial Derivatives 168
10. Differential Forms 173
II. Calculus Becomes Analysis 178
12. Interconnections of the Concepts 183

CHAPTER VII
Linear Algebra 185
I. Sources of Linearity 185
2. Transformations versus Matrices 188
3. Eigenvalues 191
4. Dual Spaces 193
5. Inner Product Spaces 196
6. Orthogonal Matrices 198
7. Adjoints 200
8. The Principal Axis Theorem 202
9. Bilinearity and Tensor Products 204
10. Collapse by Quotients 208
II. Exterior Algebra and Differential Forms 210
12. Similarity and Sums 213
13. Summary 218

CHAPTER VIII
Forms of Space 219
I. Curvature 219
2. Gaussian Curvature for Surfaces 222
3. Arc Length and Intrinsic Geometry 226
4. Many-Valued Functions and Riemann Surfaces 228
5. Examples of Manifolds 233
6. Intrinsic Surfaces and Topological Spaces 236
7. Manifolds 239
8. Smooth Manifolds 244
9. Paths and Quantities 247
10. Riemann Metrics 251
II. Sheaves 252
12. What Is Geometry? 256
x Contents

CHAPTER IX
Mechanics 259
1. Kepler's Laws 259
2. Momentum, Work, and Energy 264
3. Lagrange's Equations 267
4. Velocities and Tangent Bundles 274
5. Mechanics in Mathematics 277
6. Hamilton's Principle 278
7. Hamilton's Equations 282
8. Tricks versus Ideas 287
9. The Principal Function 289
10. The Hamilton-Jacobi Equation 292
II. The Spinning Top 295
12. The Form of Mechanics 301
13. Quantum Mechanics 303

CHAPTER X
Complex Analysis and Topology 307
1. Functions of a Complex Variable 307
2. Pathological Functions 310
3. Complex Derivatives 312
4. Complex Integration 317
5. Paths in the Plane 322
6. The Cauchy Theorem 328
7. Uniform Convergence 333
8. Power Series 336
9. The Cauchy Integral Formula 338
10. Singularities 341
II. Riemann Surfaces 344
12. Germs and Sheaves 351
13. Analysis, Geometry, and Topology 356

CHAPTER XI
Sets, Logic, and Categories 358
I. The Hierarchy of Sets 359
2. Axiomatic Set Theory 362
3. The Propositional Calculus 368
4. First Order Language 370
5. The Predicate Calculus 373
6. Precision and Understanding 377
7. Gi.idel Incompleteness Theorems 379
8. Independence Results 383
9. Categories and Functions 386
10. Natural Transformations 390
II. Universals 392
12. Axioms on Functions 398
13. Intuitionistic Logic 402
Contents Xl

14. Independence by Means of Sheaves 404

IS. Foundation or Organization? 406

CHAPTER XII
The Mathematical Network 409
I. The Formal 410
2. Ideas 415
3. The Network 417
4. Subjects, Specialties, and Subdivisions 422
5. Problems 428
6. Understanding Mathematics 431
7. Generalization and Abstraction 434
8. Novelty 438
9. Is Mathematics True? 440
10. Platonism 447
II. Preferred Directions for Research 449
12. Summary 453

Bibliography 457

List of Symbols 461

Index 463
Introduction

This book is intended to describe the practical and conceptual origins of

Mathematics and the character of its development-not in historical terms,
but in intrinsic terms. Thus we ask: What is the function of Mathematics
and what is its form? In order to deal effectively with this question, we
must first observe what Mathematics is. Hence the book starts with a sur-
vey of the basic parts of Mathematics, so that the intended general ques-
tions can be answered against the background of a careful assembly of the
relevant evidence. In brief, a philosophy of Mathematics is not convinc-
ing unless it is founded on an examination of Mathematics itself.
Wittgenstein (and other philosophers) have failed in this regard.
The questions we endeavor to answer come in six groups, as follows.
First, what is the Origin of Mathematics? What are the external sources
which lead to arithmetic and algebraic calculations and thence to
mathematical theorems and theories? Or, are there internal sources, so
that some of these theories develop just from imagination and introspec-
tion? This is close to the traditional question: Is Mathematics discovered
or invented?
Second, what is the Organization of Mathematics? Clearly a subject so
large and diverse as Mathematics requires a quite extensive and sys-
tematic organization. Traditionally, Mathematics is often split into four
parts: Algebra, analysis, geometry, and applied Mathematics. This subdi-
vision is handy at first, say for the arrangement of undergraduate courses,
but it soon needs refinement. Thus number theory is to be included,
perhaps as a part of algebra, but often using analysis as a tool. Finite (or
"discrete") Mathematics is presently popular-but is it algebra, or logic, or
applied Mathematics? Algebra soon splits into group theory, field theory,
ring theory, and linear algebra (matrix theory). These split up again:
number theory can be elementary, analytic, or algebraic; research in
group theory is sharply divided between finite groups and infinite groups,
while ring theory is split into commutative and non-commutative ring
theory, with different uses and different theorems. Analysis can be real
2 Introduction

analysis, complex analysis, or functional analysis. In geometry, algebraic

geometry is based on projective geometry, differential geometry is close to
parts of analysis, and topology has branches labeled point-set topology,
geometric topology, differential topology, and algebraic topology. The
fourth part, "applied Mathematics", is even more varied, since it may
refer primarily to classical applied topics such as dynamics, fluid mechan-
ics, and elasticity, or primarily to more recent applied topics such as sys-
tems science, game theory, statistics, operations research, or cybernetics.
Finally, the active study of partial differential equations is in part applied
Mathematics (especially when numerical methods are involved), in part
analysis, and in part differential geometry (especially when invariant
methods using differential forms are involved). But this list of subdivi-
sions is incomplete, for example, it omits logic and foundations and their
applications in computer science.
In sum, these subdivisions of Mathematics are imprecise and necessarily
involve overlaps and ambiguity. The use of even finer subdivisions (as in
the sixty-odd fields used by Mathematical Reviews to organize current
research papers) still presents corresponding difficulties. Should we con-
clude that the real organization of Mathematics cannot be accomplished
simply by subdivision into special fields? Are there deeper methods of
organization? What is the proper order of the parts of Mathematics, and
which branches belong first? Are there even parts of Mathematics which
are unimportant or mistaken?
Since Mathematical ideas often arise in prescribed order, one may also
ask whether a foundation of Mathematics provides a good organization of
the subject.
Our survey will indicate that each part of Mathematics inevitably has
an aspect which is formal. Factual problems necessitate calculations, but
the calculations then proceed by prescriptions or by rule, rather than by
continued attention to the facts of the case-yet the result of a good for-
mal calculation does agree with the facts. Proofs in geometry flow by log-
ical argument from axioms, but the resulting theorems fit the world.
Therefore we must inquire as to the relation of the formal to the factual.
Thus we begin the first chapter by exhibiting a few of the basic formal
structures of Mathematics.
This leads to our third question: Are the formalisms of Mathematics
based on or derived from the facts; if not, how are they derived? Alterna-
tively, if Mathematics is a purely formal game, why do the formal conclu-
sions fit the facts?
The fourth question is this: How does Mathematics develop? Is it
motivated by quantitative questions which arise in science and engineer-
ing, is it driven by the hard problems which have arisen in the Mathemat-
ical tradition, or is it pushed by the desire to understand the tradition
better? For example, how much does number theory owe to the repeated
attempts to prove Fermat's last theorem? Is the solution of a famous
Introduction 3

problem the pinnacle of Mathematical accomplishment-or should there

be comparable credit to the more systematic work in the introduction of
new ideas by comparison, by generalization, and by abstraction? For that
matter, how does abstraction come about, and how do we know which
abstraction is appropriate?
These questions about the dynamics of the development of Mathemat-
ics touch on a further-and difficult-topic: How does one evaluate the
depth and importance of Mathematical research?
Careful methods and canons of proof developed first in geometry
(Chapter III). Subsequently the calculus worked well, but without careful
proof, using dubious notions of infinitely small quantities. This led to the
problem of finding a rigorous foundation for the calculus (Chapter VI).
These two cases present the fifth general question, that about rigor. Is
there an absolute standard of rigor? And what are the correct foundations
of Mathematics? Here there are at least six competing schools of thought,
as follows.
Logicism: Bertrand Russell asserted that Mathematics is a branch of
logic, and so can be founded by a development from a careful initial
statement of the principles of logic. Moreover, Whitehead and Russell
carried out such a development in their massive (but now neglected) book
Principia Mathematica.
Set Theory: It is remarkable that (almost) all Mathematical objects can
be constructed out of sets (and of course sets of sets). Hence arises the
view that Mathematics deals just with properties of sets and that these
properties can all be deduced from a suitable list of axioms for sets-either
the Zermelo-Fraenkel axioms, or these axioms with supplements, some
perhaps still to be discovered.
Platonism: This set-theoretic description of Mathematics is often cou-
pled with a strong belief that these sets objectively exist in some ideal
realm. Indeed some thinkers, such as Kurt Godel, may consider that we
have special means (not the usual five senses) for perceiving thiS ideal
realm. There are other versions of platonism for Mathematics, for exam-
ple one in which the ideal realms are comprised of numbers and spatial
forms (the "ideal triangle").
Formalism: The Hilbert School holds that Mathematics can be regarded
as a purely formal manipulation of symbols, as though in a game. This is
the manipulation done when we write rigorous proofs of Mathematical
theorems from axioms. This idea was part of the Hilbert program: To
show that some adequate system of axioms for Mathematics is consistent,
in the exact sense that proofs in the system could never lead to a contrac-
tion, such as the contradiction 0 = l. To this end, the proofs were to be
viewed as purely formal manipulations and were to be studied objectively
by strictly "finite" (and hence secure) methods. As yet, such a consistency
proof has not been achieved, and Godel's famous incompleteness theorem
(to be discussed in Chapter XI) makes it unlikely that it can be achieved.
4 In trod uction

Intuitionism: The Brouwer school holds that Mathematics is based on

some fundamental intuitions-such as that of the sequence of natural
numbers. It holds, moreover, that proofs of the existence of Mathematical
objects must proceed by exhibiting these objects. For this reason intui-
tionism objects to some of the classical principles of logic, more explicitly
the tertium non datur (either p or not p). There are a number of variants
of intuitionism, some emphasizing the importance of finding proofs which
are constructive.
Empiricists claim that Mathematics is a branch of empirical science, and
so should have a strictly empirical foundation, say as the science of space
and number.
In recent years, these (and other) standard views as to the nature and
foundation of Mathematics have not been very fruitful of new insights or
understanding. For this reason, we do not wish in this book to assume
anyone such position at the start. Instead, we intend to examine what is
actually present in the practice and in the formalism of Mathematics.
Only then, with the evidence before us, will we turn to the question of
what is and what ought to be a foundation of Mathematics.
Our last and most fundamental question concerns the Philosophy of
Mathematics. This is actually a whole bundle of questions. There are
ontological questions: What are the objects of Mathematics and where do
they exist (if indeed they do exist)? There are metaphysical problems:
What is the nature of Mathematical truth? This is a favorite question,
given that the philosophers' search for truth often will use the truths of
Mathematics as the prime example of "absolute" truth. There are
epistemological problems: How is it that we can have knowledge of
Mathematical truth or of Mathematical objects? Here the answers may
well depend on what is meant by such truth or by such an object.
There are also more immediate or more practical questions. If
Mathematics is just formal· or just logical deductions from axioms, how
can Mathematics be so unreasonably effective in science (E. Wigner)? Put
differently, why is Mathematics of such major use in understanding the
world?
The various schools on the foundations have correspondingly various
attempts to answer these questions, none of them generally convincing.
Often-especially in work by philosophers-they are anchored almost
exclusively in the most elementary parts of Mathematics-numbers and
continuity. Much more substantive material is at hand. This is why we
begin with a fresh view of the variety of Mathematics.
To this end, Chapter I starts with the traditional idea that Mathematics
is the science of numbers and space-but shows that this starting point can
lead directly to some basic formal notions (transformation group, con-
tinuity, and metric space) in defiance of the usual historical order. The
next chapter describes the natural numbers as a structure, with both sur-
face and deep aspects. The traditional foundations of geometry are sum-
Introduction 5

marized in the third chapter, with emphasis on the ubiquitous role of

groups of motions and on the remarkable observation that almost all the
basic geometrical ideas can be developed in just two dimensions. The
familiar (but extraordinary) fact that very many measures of magnitude
(in time, space or quantity) can all be consigned to one structure-that of
real numbers-is the subject of Chapter IV. The next chapter discusses
the origins of the idea of "function" and the troubles in defining it. This
leads through transformations to groups again and to the question: Why
do the very simple group axioms lead to such deep structural results? The
analysis of "effect proportional to cause" is the starting point of linear
algebra (Chapter VII), but its ramifications (such as the notion of an
eigenvalue) extend beyond algebra. The next chapter deals with some of
the aspects of higher geometry: What is a manifold? Some of these ideas
are closely tied to classical mechanics, which illustrates (Chapter IX) the
intricate connection between applied and pure Mathematics. Chapter X
in complex analysis returns to the study of functions-this time holo-
morphic functions; they are closely tied to the manifolds of Chapter VIII
and to the origins of topology. At the end, the book returns to questions
of foundations (Chapter XI) and then to the six philosophical questions
raised above. With this sample of the extensive substance of Mathematics
at hand, these questions take on a different and more illuminating form.
Our discussions of the scope of elementary Mathematics do assume
some acquaintance with Mathematics; however, we endeavor to motivate
and define explicitly all the Mathematical concepts which play a role in
our discussion. Each defined word is italicized. A reference to §VII.6 is
to the sixth section of chapter seven, while (VII.6.5) is to the fifth num-
bered equation of that section; references within a chapter omit the
chapter number.
Since our survey touches upon many parts of classical elementary
Mathematics, we assume that the reader has at hand some of his own
familiar texts for possible reference. We add only occasional supplemen-
tary references to the Bibliography at the end, in the form Bourbaki
[1940]. There are a number of references to Survey of Modern Algebra and
to Algebra, both books written in some combination by Birkhoff and Mac
Lane. Homology and Categories Work (short for "Categories for the
Working Mathematician" refer to books by Mac Lane alone. We do note
here a few other overviews of Mathematics. That magnificent multivol-
ume monster by Bourbaki (for example, [1940]) is a splendid formal
organization of many advanced topics, formulated in blissful disregard of
the origins and applications which are important to our present purpose.
On a more elementary level the 1977 essay by Garding covers, with
different emphasis, many of the topics on which we touch. Davis and
Hersh [1981] has a more popular scope.
CHAPTER

Origins of Formal Structure

Mathematics, at the beginning, is sometimes described as the science of

Number and Space-better, of Number, Time, Space, and Motion. The
need for such a science arises with the most primitive human activities.
These activities presently involve counting, timing, measuring, and mov-
ing, using numbers, intervals, distances, and shapes. Facts about these
operations and ideas are gradually assembled, calculations are made, until
finally there develops an extensive body of knowledge, based on a few
central ideas and providing formal rules for calculation. Eventually this
body of knowledge is organized by a formal system of concepts, axioms,
definitions, and proofs. Thus Euclid provided an axiomatization of
geometry, with careful demonstrations of the theorems from the axioms;
this axiomatization was perfected by Hilbert about 1900, as we will indi-
cate in Chapter III. Similarly the natural numbers arise from counting,
with notation which provides to every number the next one-its successor,
and with formal rules for calculating sums and products of numbers. It
then turns out that all these formal rules can be deduced from a short list
of axioms (Peano-Dedekind) on the successor function (Chapter II).
Finally, the measurements of time and space eventually are codified in the
axioms (Chapter IV) for the real numbers. In sum, these three chapters
II-IV present the standard formal axiomatization of the science of
number, space, and time.
This development of the formal from the factual is a long historical
process in which the leading concepts might very well have come in a
different order. Our concern is not the historical order, but the very possi-
bility of a development of form from fact. To illustrate this, we start
again from number, time, space, and motion and build up directly some
of the general concepts of modern Mathematics. Thus counting leads to
cardinal and ordinal numbers and to infinite sets and transformations.
The analysis of time leads to the notion of an ordered set and a complete
ordered set; these concepts fit also with geometrical measurement. The
study of motion (in space) and of the composition of two motions suggests
I. The Natural Numbers 7

the notion of a transformation group. Comparison of this notion of com-

position with the arithmetic operations of addition and multiplication
leads by further abstraction to the concept of a group. On the other
hand, motion involves continuity, and the formal analysis of continuity
gives rise to a simple axiomatic description of space as a metric space or,
more intrinsically, as a topological space. Thus this chapter introduces the
idea of the formal in terms of certain basic structures: Set, transformation,
group, order, and topology. With Bourbaki, we hold that Mathematics
deals with such "mother structures". Against the historical order, we hold
that they arise directly from the basic stuff of Mathematics.

1. The Natural Numbers

In order to list, label, count, enumerate, or compare it is convenient to use
the single system of natural numbers, written in our conventional decimal
notation as

0,1,2,3, ... ,9,10,11, .... (1)

The same natural numbers could be written in other notations-with the

base 2 instead of 10, or as Roman numerals, or simply as marks

I,II,III, .... (2)

These numbers are used to list in order the objects of some collection of
things, or simply to label these objects, or to count the collection, or to
(thereby) compare two collections. From these activities, several
Mathematical concepts arise together

set· number· label· list .

At this point the word "set" simply means a collection of things: A group-
ing or assemblage S of objects (say, of physical objects or of symbols)
such as the collection of two turtle doves, three french hens, four colley
birds, or five gold rings-or the two collections

S = {A,B,C}, T = {U,V,W} (3)

of three letters each, written with the conventional bracket notation for a
set or collection. At this stage, the word "collection" is appropriate,
because all that matters about a set (or collection) is that it is determined
by specifying its elements; one does not yet need more sophisticated
notions, such as sets whose elements are themselves sets, or sets of sets of
sets. or sets of subsets.
8 I. Origins of Formal Structure

In these terms, one can give semi-final descriptions of the (at first)
highly informal operations of listing, labeling, counting, and comparing.
To "list" a collection such as {A,B,C} means to attach in regular order a
numeral to each object in the collection; one usually begins with the
numeral I and proceeds in order, say, as {A 1,B2 ,C3}. Note that the
numerals will be adequate for this process in all cases only if there is
always a next numeral; this is one origin of the idea that every natural
number n has an immediate successor sen) = n + 1. To "label" means to
attach the same numerals to the objects of the· collection, but irrespective
of their order, as in {A 2 ,B3,C1 }. To "count" a collection means to deter-
mine how many numerals (or which numerals) are needed to label all the
objects in the collection. In this connection, note that the count, done
properly, always comes out to the same answer. In particular, the
numerals needed do not depend on the order in which the objects of the
collections are counted: Whether it is {A 1,B2,C3}, {B 1,A 2,C3} or
{C 1,B2,A 3 }, it always ends at the same 3. Comparing two collections,
such as {A,B,C} and {U, V, W} means matching each object of the first
collection with some object of the second, until both are exhausted, as in
{AIW, BIV, CIU}. Of course, it might happen that one collection is
exhausted before the other; the first is then "smaller" in the comparison.
The result of this comparison does not depend on the order in which
objects are matched: {A,B} in any order is smaller than {U, V, W}. There
are many pairs of collections to be compared, but it again turns out that it
is not necessary to compare each pair; it is enough to compare finite col-
lections with the standard initial segments of the positive natural
numbers:

{I,2,3}, {I,2,3,4,}, {I ,2,3,4,5}, etc.

In this context, one says that the collection {A,B,C} has the cardinal
number 3, in symbols

# {A,B,C} = 3. (4)

As noted, this means that there is a one-to-one correspondence f

f: 1 I--> A, 2 I--> B, 3 I--> C (5)

which matches the standard collection {I,2,3} to the collection {A,B,C}.

The collection {u, v, W} has the same cardinal number, by the correspon-
dence

g:l I--> U, 2 I--> V, 3 I--> W. (6)

The formal definition of this matching process states that a bijection b (a

one-to-one correspondence) from a collection S to a collection T is a rule
I. The Natural Numbers 9

b which assigns to each element sin S an element b(s) in T, in such a way

that every element t of T occurs for exactly one s. This means that the
inverse of b (b read backwards) is a bijection from T to S; thus the inverse
of the bijection f of (5) above is

B ~ 2, c ~ 3. (7)

"Composed" with the bijection g of (6) this gives a bijection, f -I fol-

lowed by g, directly from {A,B,C} to {U, V, W{ as

B ~ v, C~W (8)

Thus the elementary observation that the two collections

{A,B,C} and {U, v, W} have the same cardinal number,

#{A,B,C} = #{U,V,W},

suggests the more general process of "composing" bijections, one followed

by another. Indeed, these ideas about bijections can be used to provide a
formal definition of the (cardinal) natural numbers (§II.8).
But, whatever the natural numbers are (or however they may be
defined) their primary function is to serve in calculations of sums, prod-
ucts, or powers.
The sum of two numbers is the cardinal number one gets by combining
two sets with the two given numbers, provided these sets are disjOint-that
is, have no common elements. Thus if A, B, C, U, V above are all
different, the sum 3 + 2 = 5 is

3 +2 = #{A,B,C,U,v},

and similarly for other sums. The product 2· 3 can be described "geomet-
rically" as the cardinal number of a 2 X 3 square array

2. 3 -- # {(A,U)(B,U)(C,U)}
(A, V)(B, V)( C, V) .

Here the three columns are three disjoint sets, so the product can also be
described as an iterated sum

2·3 = 2 + 2 + 2.
Similarly, the exponential 23 can be described as an iterated product

it can also be described as the cardinal number of the set of all functions
from a 3-element set {I,2,3} to a 2-element set {O,l}.
IO I. Origins of Formal Structure

These three arithmetic operations were invented (or discovered?)

because they have all manner of practical uses in financial or scientific
calculations. But, to make such calculations we never bother to reduce
each operation to its original meaning, as this meaning has just been
described. Instead, for the usual decimal notation, one may use a com-

°
puter or employ the familiar rules: The addition and multiplication tables
for the digits from to 9, plus the rules for carry-over of tens. These rules
are "formal" in the basic sense of the word: They do not refer to the
meanings of the decimals or of the arithmetic operations (though they can
be rigorously deduced from these meanings). Instead they simply specify
what to do, and specify that correctly. Thus if one counts two disjoint col-
lections as having 5 and 17 members, respectively, and then adds the
decimals 5 and 17 according to the rules, the sum is always the count for
the combined collection-and similarly for the product. To be sure, items
can get lost from collections and calculators can make errors, but then
there are further rules to make checks, like the rule of "casting out 9's"
(replace each decimal by the sum of its digits, then add or multiply,
according to the case). For numbers written in bases other than tens,
there are corresponding rules for calculations and for checks (what does
one cast out?).
This example gives a clear indication of what we intend to mean by for-
mal: A list of rules or of axioms or of methods of proof which can be
applied without attention to the "meaning" but which give results which
do have the correct interpretation.

2. Infinite Sets
The collection of all the natural numbers,

N = {0,1,2,3, ... }, (1)

° °
starts with and has to each number a successor; hence it is infinite. His-
torically, one started with I and not 0, but we need as the cardinal
number of the empty set.
The infinite set N of all natural numbers includes many finite subsets

{0,1,2}, {1,3,5,7}, {2,4,16},

as well as infinite sets, such as the set P of all positive natural numbers

p = {1,2,3,4, ... },

the set E of all even positive numbers, and the set S of all positive multi-
ples of 6. These various infinite sets may be compared as follows:
3. Permutations II

p 1

IIII
{ 1, 2, 3, 4, ...

E { 2, 4, 6, 8, ... l (2)

S
1111
(6, 12, 18, 24, ... );
the result shows that there are just as many even positives as there are
positives all told; ben) = 2n defines a bijection b: P~E. Similarly
c(2m) = 6m is a bijection c: E~S. In the comparisons (2), c(b(n)) = 6n
gives a "composite" bijection c·b: P~S.
A set X is called denumerable when there is a bijection f: N ~X. Thus
the comparisons (2) indicate that P, E, and S are all denumerable; as a
matter of fact, any subset of N is either finite or denumerable.
Two sets X and Y have the same cardinal number when there is a bijec-
tion f: X ~ Y. This definition includes the finite cardinals 0,1, ... already
discussed in §l, and the cardinal number called ~o (aleph-naught) of N,
E, P, and all other denumerable sets. In this way, the elementary activity
of counting leads to infinite cardinal numbers-of which ~b is only the
first. We will later see that the set of all points on a line is infinite but not
denumerable.
One can also formally describe when a set is infinite: When its cardinal
number is not finite, or, equivalently, when it has a proper subset S for
which there is a bijection S ~ X.
Finitists hold that infinite sets (and geometrical infinities) are just con-
venient fictions, while only the finite is "real". This we must later con-
sider. For that matter, is a finite set real? On the fourth day of Christmas,
did my true love send me four colley birds or a set of four colley birds?
Where is the set?

3. Permutations
A finite set, counted in any order, leads to the same (finite) cardinal
number. The count is not changed by "permuting" the things counted.
But one may also count how many permutations there are. Thus the set
{l ,2,3} has six permutations

(123), (231), (312), (213), (321), (132).

12 I. Origins of Formal Structure

Such counts are useful in gambling or speculating. Choose three cards in

succession from an (ordered) deck of thirteen; what is the chance that
they come out in a direct or reverse order? It is the ratio of favorable
cases [(123) or (321)] to the total number 6 of cases (of permutations).
This is the root of probability, though in the end the definition of a pro-
bability must be more sophisticated than the simple ratio of favorable
cases to total cases.
A permutation can be viewed "dynamically" -say, as an operation mov-
ing the original order (123) to the order (312) by the bijection

1 f-> 3, 2 f-> I, 3 f-> 2.

This is usually written as a cycle (132), standing for f-> 3 f-> 2 f-> 1.
Any permutation of {l ,2,3} can be viewed as a bijection

b: {l,2,3} --:7 {1,2,3}

As a bijection, it has an inverse, and any two permutations of {1,2,3} have

a permutation as their composite.
Permutations also arise in algebra. Thus, given the polynomial

(1)

what permutations of the subscripts will leave the polynomial unchanged?

To begin with, one may interchange 1 and 2, or interchange 3 and 4, or
do both interchanges, or do neither. These we may list as the permuta-
tions

(12), (34), (12)(34), I', (2)

here (12)(34) is 1 f-> 2, 2 f-> I, 3 f-> 4, 4 f-> 3; it is the composite of the

two cycles (12) and (34). Also I (do nothing) is the "identity" bijection
1 f-> 1, 2 f-> 2, 3 f-> 3 4 f-> 4. But the polynomial (1) is also left
unchanged by the following four permutations which interchange the two
factors:

(13)(24), (14)(23), (1324), (1423). (3)

This completes the list. Of the 24 possible permutations of the set

{l,2,3,4,} exactly eight leave this polynomial unchanged; of these eight,
four leave the factors unchanged. One may also wonder at the sequence
24, 8, 4. One may also experiment with other polynomials. Thus the poly-
nomial
4. Time and Order 13

has more symmetries (12 permutations!) while the polynomial

(4)

allows only four permutations (the four group)

(12)(34), (13)(24), (14)(23), I. (5)

In this list, the composite of any two permutations still leaves the polyno-
mial (4) unchanged, so the composite is also in the list. Such a list of per-
mutations is called a permutation group. The combined list (2) and (3) is
also such a group.

4. Time and Order

The passage of time suggests the ideas "before" and "after"; when the
instant t of time comes before the instant t' we write t < t '. Moreover, if
in tum t' is before til, then it is apparent that t is also before til. This can
be stated formally in the transitive law

t < t' and t' < til imply t < til (1)

for the "binary relation" <. Moreover, for any two distinct instants of
time, one must come before. In different language, for all t and til exactly
one of

t < t' or t = t' or t' < t (2)

must hold. This statement is the law of trichotomy.

But the "before" and "after" of time is not the only example of these
two laws. There is a "discrete" example. For natural numbers, m < n
means that n comes after m in the list of numbers succeeding m; here
both laws (1) and (2) hold:

0<1<2<3< .. · (3)

The usual order of the positive and negative integers provides another
instance of these laws:

. .. -3 < -2 < -1 < 0 < 1 < 2 < 3 < (4)

as does the usual ordering of the rational numbers, suggested by the

display
Q: - '/5 < ... < 0 '" < '/5 ... < '/4
(5)
< '/3 < ... < 2/3 ... < I ...
14 I. Origins of Formal Structure

There are numerous other examples of these two formal laws. Hence it is
handy to have a name for this combined situation, as it might apply to
any set X (of instants of time or of integers or of rationals ... ).
A binary relation < on a set X specifies that x < y is true or false for
any two elements x,y in X; one might also say that the relation amounts
to specifying a set: The set of all those ordered pairs (x,y) with x < y. A
linearly ordered set is then a set X with a binary relation < for which the
laws (l) and (2) hold; in other words it is a set equipped with a transitive
and trichotomous relation <. One can then invent (or discover?) many
other examples of linearly ordered sets: Finite ones such as
1 < 2 < 3 < 4 or long infinite ones such as

o< 1< 2 < 3 < ... < w < w + 1 < w + 2 < ... , (6)

where w is the first thing beyond all the finite natural numbers. (This
linearly ordered set is actually the start of the infinite ordinal numbers.)
This definition is an easy first (of many) cases of a list of axioms
describing a common situation with many different examples. As in other
cases, the choice of axioms can vary. Thus, rather than using "before" and
"after", the passage of time can be described by the notion "not later
than", usually written t < t '. This alternative can be formalized for any
linearly ordered set X. Define x < y to mean x < y or x = y. This
binary relation on X is then

Transitive: x <; y and y <;z imply x<; z,

Reflexive: x <; x for all x,
Antisymmetric: x <; y and y <; x imply x = y.

Finally, corresponding to trichotomy, it has the property:

For all x and y in X, either x < y or y <; x.

Conversely, let any set X have a binary relation < with these four proper-
ties, and define x < y to mean that x < y but x oF y. Then X is indeed
a linearly ordered set and the originally given relation < is related to <
as before. In brief, the same notion of linear order can be defined in two
formally different ways, via < or via <. In general, the same situation
may often be defined in two or more formally different ways.
One also asks when two "models" of the axioms are "essentially" the
same-in the sense that the linearly ordered set of natural numbers has
the same "order type" as the ordered set of even positive natural
numbers:

2<4<6<8<10< ...
4. Time and Order 15

So for linearly ordered sets X and Y an order isomorphism f: X ..... Y is

defined to be a bijection of the set X on the set Y such that order is
preserved: For all XI and X2 in X,

implies [XI < [X2. (7)

When there is such an isomorphism [, X and Yare said to have the same
order type. (This is like the definition of "same cardinal number" except
that now one also keeps in mind the order of the elements being com-
pared.) One can then readily prove (say) that any linearly ordered set of 4
elements is order isomorphic to the standard such set: 1 < 2 < 3 < 4.
A general question is then at hand: Can one describe a particular
model of the axioms by giving enough additional axioms to determine the
model uniquely (i.e. uniquely up to an order isomorphism?) In the present
case, can one give properties of an ordered set X which imply the
existence of an order isomorphism X ..... N (or X ..... Q, the ordered set of
rationals, or X ..... R, the ordered set of reals?)
The answers are "yes". To get at the case of the reals R, one must for-
mulate the sense in which a real number (an instant of time) can be
approximated by rational numbers. For example, the real number 7T is
determined by the usual sequence of decimal approximations

3.14,3.141,3.1415,3.14159,3.141592, ....

Indeed, 7T is the "least upper bound" of this set of rational numbers. For-
mally, in a linearly ordered set X an element b is an upper bound for a
subset S of X if s ~ b for every s in S. Also, b is a least upper bound for
S if no b' with b' < b is an upper bound for S. This implies that if S
has a least upper bound, that least upper bound is unique. (This is the
sense in which 7T, for example, is determined uniquely by its decimal
expansion). Also, the set X is unbounded if there is in X no upper bound
and no lower bound. (For example, the ordered set N has a lower bound
0, hence is not unbounded).
The crucial property of the ordered set of real numbers is completeness:
Every non-empty subset S with an upper bound has a least upper bound.
The ;;tdditional fact that every real number can be approximated by
rationals can be made formal by stating that the set Q of rational
numbers is "dense" in R. Here a subset D of a linearly ordered set X is
said to be dense in X if, for all X < Y in X there is always a d in D
between X and y, so that X < d < y. It is then clear that the ordered set
R is complete, unbounded, and has a denumerable dense subset. Also
one can prove that any linearly ordered set X with these three properties
is order isomorphic to R (see Hausdorff). In the proof one uses a charac-
terization of the order type of Q: It is denumerable, unbounded, and
dense (as a subset of itself).
16 I. Origins of Formal Structure

This result does provide a description of the order of the real numbers.
In Chapter IV we will combine this with a description of their algebraic
properties. These properties also arise from experience with the passage
of time. Once intervals of time are measured by a clock (or an hourglass)
one can add one interval to another, and regard each instant of time t as
the end of an interval (from some starting time). This addition is then an
operation which produces to each pair t, t' of instants their sum, t + t',
with properties such as t + t' = t' + t and

(t + t ') + til = t + (t' + til)

-just like those for the addition of natural numbers. Again, different
examples lead to the same formal law.

5. Space and Motion

Space can be regarded as something extended or as a receptable for
objects or as a background for ideal "figures". These aspects are all
closely tied to the notion of motion through space, while motion provides
the notion of measuring distance in space. Space and motion crop up
together everywhere, from physics to physical exercise.
Idealization of the notion of space suggests that chunks of space are
made up of figures which are filled up with "points". A point is in space,
but without extent. In the extreme analysis, the space consists just of
points-but to make this work the points must have added structure, say
that described by giving the distance p(p,q) from the point p to the point
q. This distance is to be measured along straight lines and is a number-
at first, just some rational number. But some lines must be vertical (for
balance) and others horizontal. Thence comes the idea of perpendicular
lines (the word suggest the vertical, as in the perpendicular version of
gothic architecture). This leads to right triangles. These lead in turn to
the pythagorean theorem and the discovery that the hypotheses of an isos-
celes right triangle with both legs of length 1 is measured by v1-which
cannot be a rational number (because v1 = min in lowest terms would
give m 2 = 2n2, forcing m and then n to be even). Thus it is that space,
measured with distances, requires not rational numbers but real numbers.
Thus, given the real numbers, one is led to describe space-or a chunk
of space-as a collection of points p, q,... together with a non-negative
real number p(p,q) which is the measure of the distance from p to q. It is
the same as the distance from q to p:

p(p,q) = p(q,p) for all p,q; (1)

it is zero only when the points coincide:

5. Space and Motion 17

p(p,q) > 0; p(p,q) = 0 if and only if p = q. (2)

Moreover the intent is that this distance is the shortest from p to q. (The
straight line is the shortest distance between two points.) In particular, this
means that the distance from p to q is not lessened when it is measured
along two straight lines going through a third intermediate point r. This
amounts to the (Figure 1) triangle axiom: For all p, q and r in X,

p(p,q) < p(p,r) + p(r,q). (3)

Thus arises the concept of a metric space: A collection of points p, q

together with the real number distances p(p,q) which satisfy the axioms
(l), (2), and (3). The evident chunks of space-a square, a cube, a
cylinder, a blob, a dumbbell, each with the usual measure of distance-are
all metric spaces in the sense of this definition, as is the whole of our
("ordinary") three dimensional space. Non-Euclidean geometry (Chapter
III) provides natural examples of such spaces as do the curved spaces to
be considered in Chapter VIII; there are also bizarre examples-such as
"a space" with infinitely many different points, with distance 1 between
any two different points (try to fit that into the plane). Despite such
bizarre examples, many elementary properties of space can be formalized
and studied for a general metric space. In other words, given numbers, the
Mathematical study of space need not start with the conventional ideas of
Euclidean geometry, but instead with an axiom system-that of metric
space-which applies to many different examples of "space".
Motion can be described in any metric space-push the points around,
keeping fixed their distances apart. More formally, if F is a figure (a col-
lection of points) in a metric space X a motion of F will at each time t
take each point p of F to a new position (a new point) Mtp in X. This
passage must be "continuous" (an idea to which we will soon return) ..
Moreover, the motion must be rigid-the distance apart of any two points
must stay the same during the motion; in other words, for all times t and
all points p and q of F, the distance p must satisfy

(4)

We speak of such a motion (p,t) I--> Mtp as a parametrized motion of the

figure F, with t as the parameter.
It is perhaps easier to consider just a "completed" motion-the passage
from the initial position p to the final position MtlP at some chosen time

Figure I
18 I. Origins of Formal Structure

tl. This is called a rigid motion M; it assigns to every point p of the figure
concerned a new point Mp such that, for all p and q,

p(Mp,Mq) = p(p,q); (5)

put briefly, a rigid motion is a bijection of space which preserves distances

between points. For example, a rigid motion of an equilateral triangle
into itself could be a rotation (by 120°, 240°) or a reflection of the triangle
in one of the three altitudes or the identity motion (every point stays put).
There are thus six such motions (symmetries) of the triangle. For motions
of the plane as a whole, we will see in Chapter III the use of three typical
motions: A translation (every line stays parallel to its original position), a
rotation (one point is fixed) and a reflection (all the points on a line stay
fixed. These are not all: Moving a triangle ABC into a congruent triangle
A 'B'C' (Figure 2) may require a translation (A to A ') followed by a rota-
tion about A '; in other words, a composite motion.
From such examples arises the idea of the composition of two motions
M and N-first move by M and then move the result by N, to give the
composite motion C with

C(p) = N(Mp). (6)

We write C = N·M for the composite and observe at once that if M and
N are rigid motions, so is C. For parametrized motions the addition of
time intervals usually corresponds to composites, in that

(7)

The axioms for a metric space show that any rigid motion M keeps dis-
tinct points distinct. Indeed, p =1= q implies by axiom (2) that p(p,q) =1= 0
and hence by the definition (5) of a motion that p(Mp,Mq) =1= 0), hence
Mp =1= Mq by axiom (2) again.
In studying the symmetry of a figure F, we usually consider a motion M
of F "into" itself; that is, a motion M such that p in F moves to some
M(p) in F and such that every point q of F comes from some p in F, so
that q = M(p). By the above, the motion M is therefore a bijection of F
to F, and so has an inverse M- I which is also a rigid motion of F to F.
However, the reader might wish to construct an infinite figure F (say
one in the plane) and a rigid motion M of F into F which is not onto F.
B

~c
B'

A A'~C'
Figure 2
6. Symmetry 19

6. Symmetry
Symmetrical objects are all about us. There are many (man-made) sym-
metrical figures (Figure 1). Each of the figures has vertical symmetry,
horizontal symmetry, and rotational symmetry. The vertical symmetry V
can be construed as a reflection of the figure in its vertical axis, and simi-
larly for the horizontal axis, H. The rotational symmetry can likewise be
regarded as a 180· rotation R of the figure about its center. If we think of
the figure as a metric space X, each of these symmetries is a rigid motion
M of X onto itself, and these four motions are the only such. This sug-
gests a definition of a symmetry of a figure F: A rigid motion of F onto
itself. In particular the different figures of (1) have by this definition the
same symmetry (later called the four-group).
By this definition, the composite of two symmetries of F is again a sym-
metry. Thus vertical reflection followed by another vertical reflection is
the identity (which thus must count as a symmetry). Again, vertical
reflection followed by horizontal reflection is the 180· rotation. This one
may check by actual experiments with a rectangular card-or one may
label the vertices of the rectangle by numbers 1, 2, 3, 4 so that V amounts
to the permutation (12)(34), H is (14)(23), and the composite H· V (first
apply V then H) is

1 ~ 2 ~ 3, 2 ~ 1 ~ 4, 3 ~ 4 ~ I, 4 ~ 3 ~ 2; (1)

this is the permutation (13)(24) given by the 180· rotation. Thus the total
list of symmetries for the Figure I is

(12)(34), (14)(23), (13)(24), I. (2)

This is identical to the list (3.5) of permutations allowed by the polyno-

mial (Xl - X2)(X3 - X4) of (3.4). Thus the same symmetry turns up in
both geometric and algebraic circumstances. This suggests that the under-
lying symmetry here-in this case the "four group"-must itself be some-
thing "abstract"; neither geometric nor algebraic; or perhaps both. It
need not depend on numbers-the dumbell of Figure I has no convenient
corners to be numbered!
There are many different types of such symmetries. In three dimensions,
one has the symmetry of the regular tetrahedron, or of the cube, or of the
icosahedron, or of the octahedron. In the plane there are symmetrical
figures such as those of Figure 2. For the equilateral triangle there are six

~I 0-0
Figure 1
20 I. Origins of Formal Structure

symmetries, accounting for all six permutations of the three vertices-or

just as well, all six permutations of the three sides. For the square and
also for the decorated square there are eight symmetries all told-four
reflections (vertical, horizontal, and two reflections in the diagonals) and
four rotations (counting the identity as a rotation through 360°!) If one
labels the four vertices as in Figure 2, the eight symmetries turn out to be
exactly the eight symmetries (3.2) and (3.3) listed in §3 for the polynomial
(Xl + X2)(X3 + X4). This again indicates that algebra and geometry
have in common some underlying, more abstract, form.

2 4

Figure 2

The frieze of a Greek temple, such as that suggested by the scheme

(Figure 3) has "more" symmetry. One considers it as a "linear orna-
ment", extending to infinity in both directions: one may picture it more
schematically as in Figure 4, with nodes labeled by numbers. There are
then infinitely many symmetries: Vertical reflection (n to - n), translation
T to the right by two units and repeated such translation Tn, n times, as
well as the inverse translation T- (two units left) and its iterates T- n .
There is also a different rigid motion S-translate one unit right and reflect
in the horizontal axis. Then the composite S·S is just T, so that all the
symmetries of this figure are "generated" by V and the "slide reflection" S
and its inverse. If we erase the lower spikes in Figure 4 we get fewer sym-
metries (no S, but V and 1). The reader may try to find linear ornaments
with still different symmetries. (There are just seven sorts).
Three dimensional infinite symmetry comes in much greater variety.
There the origin is not just from architecture, since the classification of
three dimensional symmetries is the first step in the classification of cry-
stals by the "crystallographic groups" .

• !.!.!
Figure 3

I I I
-3 -2 -1 0 2 3

Figure 4
7. Transformation Groups 21

7. Transformation Groups
A permutation of a set, a symmetry of a figure, and a motion of
Euclidean space are all examples of "transformations". A transformation
T of a set X is a bijection T: X ..... X; that is, a one-to-one correspondence
x I--> Tx on the elements x of X. Thus each transformation T has an inverse
T- 1: X ..... X; any two transformations Sand T have a composite S· T -first
apply T and then S.
A transformation group G on a set X is a non-empty set G of transfor-
mations T on X which contains with each T its inverse and with any two
transformations S, T in G their composite. This implies that G always
contains the identity transformation I on X:

(1)

A transformation group on a finite set (and especially on the typical finite

set (l,2, ... , n}) is usually called a permutation group. The symmetric
group of degree n is the group of all n! permutations of {I , ... ,. n }.
Symmetry groups of figures or formulas are the leading examples of
transformation groups, and the source of the "abstract" concept. This is a
typical example of Mathematical experience leading to a formal
definition. But we are also led to explicate when two transformation
groups are "essentially" the same. To do this, one may examine a case
such as the representation in §6 of each symmetry of the square X by a
permutation of the vertices of that square. This takes place by labeling
the vertices by numbers, say by a function f: {I ,2,3,4} ..... X which puts each
number on the corresponding vertex. The labeled vertices are all different;
that is, fk = fm implies k = m; one says that the function f is injective
(an injection, or one-one into). With these labels, each motion T: X ..... X of
the square sends each vertex to a vertex, so determines a permutation
# T: Y ..... Y of the set Y of vertices. Thus # T does to k what T does to fk;
in other words,

(2)

for k = 1,2,3, or 4. This equation can be written in terms of composites

of functions as

(3)

or displayed in a diagram of the corresponding functions as

#r
y------
• y

f~ ~f (4)
x - - - - - -• x .
r
22 I. Origins of Formal Structure

This exhibits I as comparing the action of # T on the vertices Y with the

action of T on X. This diagram is called commutative because (3) holds:
Both paths from upper left to lower right have the same result. This
example (and many others like it) suggests a general formalization of the
idea of comparing a transformation group H on a set Y with G on X: A
map of (H, Y) to (G,X) is a function I: Y --X and a function #: G--H
such that (4) commutes for every transformation T in G. In case this I is
an injection (as in the case above), the equation (3) shows that giving I
(giving the labels of the vertices) completely determines # . If moreover I
is a bijection, it has an inverse I - I so that # can be described directly by

(5)

to find the permutation, label each vertex by I, look to see where the ver-
tex goes, and read off its label (by I-I).
This result does formalize the evident fact that the permutations of a
typical set {l,2,3,4} of 4 things represent also the permutations of any set
of four things. Generally, if sets Y and X have the same cardinal number,
by a bijection I: Y --X, then the correspondence # of (5) is a bijection
from the permutation group of X to that of Y. Note incidentally that #
goes in the direction opposite to I.
However, this notion of a map is a bit complicated. Moreover, it
doesn't directly handle all the desired comparisons. Thus in (6.1) the dum-
bell Yand the perimeter X of the rectangle clearly have the "same" sym-
metries, but there is no evident way to get a map f: Y --x to make such a
comparison. Indeed, there is no such f -because the dumbell Y has a
center point left fixed by all the motions and there is no such point on the
perimeter of the rectangle. The two transformation groups in this case
can at least be compared through some intermediary-mapping each (say)
into a common (containing) such rectangle.
To summarize: symmetry forces us to consider transformation groups,
and even forces thoughts as to more abstractions from this notion.

8. Groups
For any three transformations R, S, and T of a set X the iterated compos-
ite, by its definition, satisfies

«R,S)'Dx = R(S(Tx)) = (R,(S'D)x,

so composition of transformations is associative. Now, in a transforma-
tion group G, forget the fact that the elements T of G transform things,
and use only the properties of composition. It is then a group in the sense
of the following definition of an "abstract" group:
8. Groups 23

A group is a set G equipped with three rules, as follows:

(i) A rule assigning to any two elements s, t of G on element st, called
their product, such that the product is associative,

r(st) = (rs)t, (1)

for all r, s, t in G.
(ii) A rule determining an element e (the unit, often written as e = 1) of
G such that, for all t in G,

te = t. (2)

(iii) A rule assigning to each t in G an element t - 1 in G such that

tt- 1 = e. (3)

In every transformation group, composition has these properties, so every

transformation group is a group. Moreover (and vice versa) Cayley'S
theorem asserts that every group G arises in this way from a transforma-
tion group; just take the set X of points to be transformed to be the set G
itself, while each t in G is the transformation sending x in G to the prod-
uct tx in G. But transformations are not the only sources of groups. With
multiplication taken to be the product, the positive rational numbers or
the positive real numbers or the non-zero complex numbers constitute
groups. If addition is taken to be a "product", the real numbers (the
instants of time) form a group, as do the ordinary clock hours (12 = 0).
Other groups, as we will see, arise in number theory. Groups such as
these, where the product is commutative,

st = ts (4)

for all sand t are called abelian groups.

There are many consequences of the simple axioms (i), (ii), and (iii) for
a group. They include easy consequences such as the cancellation law
(st = s't implies s = s ') or the rules

te =t= et, (5)

which might as well (for the sake of symmetry) be used as axioms in place
of (2) and (3). A group G may have subgroups S (a subset which is itself a
group under the same multiplication (and inverse)). If G is finite, its cardi-
nal number is called its order. One proves that the order of a subgroup is
always a divisor of the order of the group; this serves to understand and
explain some of the observations made above about the orders 8 and 4 of
subgroups of the symmetric group of four things. There are all manner of
constructions of particular groups. Thus to each positive n the cyclic group
24 I. Origins of Formal Structure

Zn of order n consists of all rotational symmetries of a regular n-gon; it is

an abelian group. To any two groups G and H one may construct a group
GX H, their "product". It is abelian when G and H are. The "structure
theorem" for finite abelian groups G states that every such group is a
iterated product

(6)

of cyclic groups. Moreover, the orders m] , ... ,mk of these factors can be
chosen so that each is a multiple of the next (and their product is the
order of G). We will be concerned with the origins of this theorem in
number theory (the multiplicative group of residues prime to m, modulo
m), in topology (the homology group of a finite complex described in
terms of Betti numbers and torsion coefficients). We are also concerned
with the question of the proper generality of such a theorem (is it really a
theorem about finitely generated modules over a principal ideal ring
(Algebra, p. 384)?). We are concerned with the additional concepts which
such a theorem brings to attention-for example, the notion of direct
product of groups and its eventual conceptual generalization to products
of other types of objects (rings, spaces) and finally, to products of objects
in a category (Categories Work, p. 68 or Chapter XI below).
For non-abelian groups G there is no structure theorem as simple as (6).
For example, the symmetric group on 3 letters {l,2,3} has order 6 but it is
not cyclic nor is it a product of cyclic groups (though it does have sub-
groups of orders 2 and 3). For such non-abelian groups there are instead
much deeper structural results (Chapter V). One may then ask why the
very simple group axioms lead to such deep structure.
Return to the idea of comparing two groups. For the case of a bijection
f: Y -->X which gives a map of one transformation group (H, Y) to (G,X)
we used #: G-->H with

for all Tin G, as in (7.5). Then for any composite S· T in G one gets

Thus arises the definition: for any two groups G and H a homomorphism
# or b: G-->H is a function assigning to each sin G an element bs in H in
such a way that

best) = (bs)(bt). (7)

(It follows that be = e and b( t - ]) = (bt) - ]). If b is a bijection, It IS

called an isomorphism; hence the simple formulation of the answer to a
question from §7: Two figures have the same symmetry if their symmetry
groups are isomorphic.
8. Groups 25

But geometry is not the only source of the idea of isomorphism. The
familiar property of the logarithm (say to base 10),

log(xy) = log x + log Y

states that the logarithm is an isomorphism of the multiplicative group of
positive real numbers to the additive group of (positive, 0, and negative)
real numbers. Note also that log 1 = 0 and log (x -1) = -log x.
There are many other familiar examples of homomorphisms: absolute
values of numbers, determinants of matrices, or the way in which each
symmetry of a square produces a permutation of the diagonals of that
square.
This discussion has summarized, in an especially striking case, the way
in which some underlying informal idea arises and then is formalized by
generalization and abstraction. First the study of motion, of symmetries
and of permutations suggests the "idea" of composition. One formaliza-
tion of this idea is that of a transformation group (we will later meet other
formalizations of composition). The notion "transformation group" gen-
eralizes a variety of examples in a way which assists the understanding of
the common properties and the comparisons between examples. But it
turns out that some of the most useful properties involve not the composi-
tion of transformations, but just composition; moreover this behavior of
composition is like that of addition or multiplication of numbers. By an
effort of abstraction this leads to the more abstract notion "group" and to
the study of the extensive properties of such groups.
Our presentation implicitly argues that the notion of a group comes
early (and prominently) in the order of mathematical ideas. Historically,
this was not the case. Geometry was treated by axioms and not via groups
(as it might have been-see Chapter III). Classification of crystals by their
symmetry groups apparently was not developed till the 19th century. The
first groups explicitly recognized as such were permutation groups, and
the first explicit use of group notions came with Galois (1832), who used
homomorphisms (i.e., normal subgroups) to prove theorems about the
solution of algebraic equations (Galois theory: §V.7). For the rest of the
19th century groups (usually described with confused definitions) were
chiefly permutation groups. When Cayley defined an abstract group in
1852, nobody paid any heed. When he repeated his definition in 1871 he
found three different groups of order 6 (not two), because he failed to
recognize an isomorphism Z6 == Z2 X Z3. In 1905 Burnside published a
definitive monograph "Theory of Finite Groups". It dealt in fact with
abstract groups, but called them "groups of substitutions". That doubtless
expresses the intuitive base for group theory. It would seem that only the
20th century saw the full utility of the notion that a variety of examples
could be understood well by an axiomatic description of the concept.
The axioms for a group provide a pattern for other axioms of algebraic
structure. Instead of stating an axiom "there exists in G an element e such
26 I. Origins of Formal Structure

that always te = t", our axiom (ii) has specified that the element e is
"given". Indeed it can be "given" as a function e: {*} ~ G mapping the
one point set {*} into the element e of the set G. Such a function is a nul-
lary operation (on the set G). Thus the group axioms provide three opera-
tions

e: {*} ~G, (8)

a binary operation (multiplication), a nullary operation (unit), and a unary

operation (inverse). These operations are required to satisfy certain identi-
ties (1), (2), and (3) which can be regarded as identities between "compos-
ites" of the initial operations (8).
Much the same pattern applies to operations of addition and multipli-
cation (the axioms (§IV.3) for a ring or a commutative ring) and for the
axioms on the algebraic operations for lattices, vector spaces, and the like.
Groups have been variously generalized. There are, for example, gen-
eralizations made by deletion of axioms. Drop the unary operation of
inverse (and the axiom (iii) pertaining thereto) and one has the axioms for
a monoid. Drop also the axiom (ii) for the unit e to get the axioms for a
semi-group, and observe that there are various motivations for these dele-
tions; semi-groups arise in the operation of finite state machines (the
sequences of states form a semi-group) and in the composition of opera-
tions in functional analysis-but semi-groups do not have as rich a struc-
ture as do groups (How does one account for such varying richness of
structures?) We will repeatedly examine generalizations by deletion.
These and many other cases illustrate the general notion of an algebraic
structure: A set X with nullary, unary, binary, ternary ... operations satis-
fying as axioms a variety of identities between composite operations.
"Universal algebra" is concerned with the general properties of such
structure. There is also a "many-sorted" universal algebra for those struc-
tures involving more than one set. A first example (two sorts) is a transfor-
mation group: A set X together with a group G of transformations on X.
An even more decisive example is that of a ring R and a left module
(§VII.11) over that ring. More recently, many-sorted universal algebra has
proved useful in the computer science of data types.

9. Boolean Algebra
Another example of an algebra is provided by the operations such as the
intersection and the union of subsets Sand T of a given set X. If we write
XES for "x is an element of S" and ¢ ; > for "if and only if', these
operations are specified by giving the elements of the resulting subset of X
as follows:
9. Boolean Algebra 27

Intersection XES n T <=> XES and x E T, (I)

Union XES U T <=> XES or x E T, (2)

~ XES ~ T <=> if XES, then x E T, (3)

<=> x E T, or not (x E S).

They correspond exactly to the three propositional connectives "and",

"or", and "if then". They may also be pictured by Venn diagrams; if the
set X is taken to be all the points in a rectangle while Sand T are respec-
tively the points inside the ovals Sand T, then two of these operations
may be indicated by shaded areas as in Figure 1. There is also a unary
operation, the complement.., S of S:
x E .., S <=> not (x E S) (4)

These various operations n, u, ~, .., satisfy certain algebraic identities

which can all be deduced from a suitable list of axioms, the axioms for
Boolean Algebra. Thus the set P(X) of all subsets of X is a Boolean alge-
bra.
There also are operations on infinite families of sets. Thus if Si is a sub-
set of X for each i in some "index" set I, the (infinite) Union and intersec-
tion are defined by
x E USi <=> For some i in I, x E Si, (5)

x E n Si <=> For every i in I, x E Si (6)

These operations correspond to the logical quantifiers "There exists an i"

and "For all i", respectively. These connections with logic will be
explored further in Chapter XI.
Boolean algebra provides a Mathematical way of representing proper-
ties, in that each property H of elements of a set X determines a subset of
X; namely, the subset S consisting of all those elements which have the
property

S = {x I x E X and x has H}. (7)

snr s~r

Figure I. Boolean operations.

28 1. Origins of Formal Structure

This subset is sometimes called the extension of the property H, to

emphasize the notion that differently formulated properties may have the
same extension-and that Mathematics has to do with extensions rather
than with meanings. This in turn involves the "extensionality" axiom for
sets-that a set is completely determined just by specifying its elements.
This means that the equality of two subsets of X is described by the state-
ment

S = T ¢:;> (For all x in X, XES ¢:;> x E 1), (8)

while the inclusion of one subset S in another is described by

SeT ¢:;> (For all x in X, XES ~ x E 1); (9)

here the arrow ~ stands for "implies".

This inclusion relation is transitive, reflexive, and antisymmetric, as
these properties were defined in §4 above. In general, an ordered set W is
a set W (such as P(X» with a binary relation (such as SeT for
S, T E W) which is transitive, reflexive, and antisymmetric. An ordered
set is often said to be partially ordered (a poset) because it need not satisfy
the "trichotomy" property which holds for a linear order.
It is important to recognize that many orders are just partial orders and
not total orders (i.e., not linear). However, in many domains of the appli-
cation of Mathematics to social phenomena, there is a strong tendency to
order ideas, people, and institutions in a linear way-for example, accord-
ing to rank on some imagined numerical measure. The more relevant
notion of partial order seems little known and less used.
Diagrammatic presentation of an inclusion relation is suggestive. Thus
the various inclusions of the subsets of a three-element set can be pictured
by the rising lines in Figure 2, where the bottom symbol 0 denotes the
empty subset. The Boolean operations on subsets may be visualized in
this figure. For example, the union {I,2} of the subsets {I} and {2} is the
smallest subset which lies "above" both the subsets {I} and {2}; in this
way it is the least upper bound, as defined in §4, of {I} and {2}. Gen-
erally, the union S U T of two subsets Sand T of a set X has the proper-
ties

ScSuT, T C S U T, (10)

S c Rand T C R ~ S uTe R, (11)

which state that it is the least upper bound of Sand T in the partial order
given by inclusion. In an exactly dual way, the intersection S nTis the
greatest lower bound of the subsets Sand T. In other words, both these
Boolean operations can be described directly in terms of inclusion,
10. Calculus. Continuity, and Topology 29

Figure 2. Lattice of subsets.

without any use of membership. In Chapter XI we will see further exam-

ples of sets treated without the use of elements.
There are corresponding definitions for other inclusion relations. In
general (and in view of diagrams like that above) a poset is said to be a
lattice when it has a top element 1, a bottom element 0 and when each
pair of elements have a least upper bound (called their join) and a
greatest lower bound (called their meet). The lattice of subobjects of an
algebraic object is a way of describing some of the structure of that object.

10. Calculus, Continuity, and Topology

Many notions besides those of transformation groups arise from the
mathematical analysis of motion. The complex motions of the planets
and the varying velocities of falling bodies suggest the idea of "rate of
change": Velocity as rate of change of distance or acceleration as rate of
change of velocity. These ideas were codified in the notion of the deriva-
tive, subsequently formalized (Chapter VI) in the rigorous foundation of
the calculus, as based on the axioms for the real numbers. This uses the
definition of the derivative by means of limits and thus the consideration
of a class of "good" functions-those which are differ!!ntiable. As a first
example of this circle of ideas, we examine here another good class-the
functions which are continuous.
A rigid motion M: F ~F of a figure is continuous because (by rigidity)
the distance from Mp to Mq must equal that from p to q. For a function
f: R ~ R on the real numbers R continuity means considerably less: Just
that fx and fy will be close if the originals x and yare sufficiently close.
This formulation is still pretty vague; it should mean that one can make
fx and fy "as close as you please" by requiring x to be "suitably close"
to y. This is still vague. "As close as you please" should mean "within a
specified measure () (a positive real number) of closeness; "suitably close"
should mean that one can specify a measure of closeness (again a positive
real number £) which will do the job. All this (and we have telescoped a
30 1. Origins of Formal Structure

long and painful historical development) comes down to make the famil-
iar (but meticulous) £ - 8 definition of continuity: A function f: R ..... R
is continuous at a point a E R if
For all real £ > 0 there is a real 8 > 0 such that, for all x in R, (1)
If Ix - a I < 8, then If(x) - f(a) I < £. (2)

If this statement holds for all points a E R, the function f is called con-
tinuous; the class of all such continuous functions is called C.
Note that the statement involves both propositional connectives ("if ...
then") and the so called "bounded" quantifiers (For all real numbers,
there exists a real number). Thus it is that careful formulations lead to the
use of concepts of formal logic.
Topological and metric spaces arise from analysis of this definition of
continuity. The inequalities used in the definition arise from ideas of
approximation (approximations of the value b = f( a) to within the accu-
racy £) and so implicitly involve the open interval I( (b) =
{y I I y-b I < £} of center b and "radius" £. In the familiar represen-
tation of the function f by its graph (the set of points (x, f(x)) in the
plane), this open interval appears as an open horizontal strip of width 2£
around y = f( a) (Figure I). The definition is concerned with those
points x E R for which f( x) lands in this interval I = I( (b )-this set of
points is usually called the inverse image of I under the function f, in
symbols:

f-II = {x Ix E R andf(x) E I}.

Indeed, if Xo E f-II (that is, if f(xo) E I), then one can prove from the
definition of continuity that there is an open interval (on the x axis) of
center Xo wholly contained in f-II. This amounts to the

Theorem. The function f: R ..... R is continuous for all a E R if and only

if the inverse image f -II of every open interval of R is a union of open
intervals.

f(a) = b 2€

1-------+
--------+-----~-------+---------x

Figure I
10. Calculus, Continuity, and Topology 31

In other words, continuity can be described wholly in terms of open

intervals.
Continuity is also needed for functions f defined on only a part of the
real line, for functions of two or more variables, or for functions defined
on a surface, etc. But this new definition requires no new ideas. The
absolute value I x - a I which appears in the definition (l) and (2) is
just the distance p(x,a) on the line R from the point x to the point a.
Hence the same definition will work with any suitable notion of
distance-that is, for a metric space.

Definition. If X and Yare metric spaces, a function f: X ~ Y is continu-

ous at a point a E X if for all real t: > 0 there is a real 8 > 0 such that
for all x in X, if p(x,a) < 8 then p(f(x),f(a» < t:.

In particular, this defines continuity for functions of two real numbers

x,y. Just regard (x,y) as the (coordinates of) a point in the plane R X R
with the usual (pythagorean) metric PI = p,

p«x,y),(a,b» = [(x - a)2 + (y - bil l / 2 ,

Then in the definition of continuity the open interval consisting of all x
with I x - a I < 8 is replaced by the open disc with center (a,b) and
radius 8. This disc consists of all points (x,y) in the plane with

(3)

As in the theorem above, one can prove that a function f: R X R ~ R of

two real variables is continuous if and only if the inverse image of each
open interval in R is a union of open discs in R X R.
But the plane is also a metric space with a distance P2 described as the
"shortest path along the rectangular grid", so that

pz«x,y), (a,b» = Ix - a I + Iy - b I'

or with the distance given by

P3«X,y), (a,b» = Max( I x - a I, Iy - b I)·

For a function f(x,y) of two real variables, anyone of these distance for-
mulas may be used to define the continuity of f in the usual t: - 8 style.
These different distances all yield the same continuous functions, so that
continuity, viewed invariantly, depends not on distance but on something
else more intrinsic. What is it? The answer is well-known. Each distance
Ph pz, or P3 will determine a series of "open circular discs" -circular in
that metric-of the respective forms (interior of) circle, diamond, or square
(Figure 2). Within each disc we can readily draw a smaller disc of the
32 I. Origins of Formal Structure

Figure 2

other forms with the same center-so we get the same E - 8 continuous
function by just choosing different 8's to a given E.
What is the intrinsic formulation? In the alternative description of con-
tinuity stated in the theorem above there appeared unions of open inter-
vals and unions of discs. In any metric space X the open disc with center a
and radius 8 may be defined to be the set of all points x in X with
p(x,a) < 8. Now define an open set U in X to be any union (finite or
infinite) of open discs. Equivalently, a subset U is open in X if to each
point a E U there is a 8 > 0 such that p(x,a) < 8 implies x E U -
every point a in U is contained in an open disc centered at a and within
U. Now the continuity of f: X ...... Y is expressed in terms of open sets: f is
continuous if and only if the inverse image of any open set of Y is an
open set of X. This is the desired intrinsic formulation, independent of the
perhaps accidental choice of a metric. Specifically, the three different
metrics P = PI, P2, and P3 described above for the plane all yield the
same open sets, because any union of circular open discs is also a union
of open squares or of open diamonds, and conversely. In this way, the
notion of "open set" is more intrinsic than that of distance.
This suggests that a space can be defined directly in terms of its open
subsets. A topological space is a set X in which certain of the subsets U are
distinguished and called the open sets, and in which these open sets are
required to satisfy the following three axioms:
1. The intersection of two open sets is open;
2. The union of any collection of open sets is open;
3. X itself and the empty subset 0 c X are open.
A topology on a set X is then the specification of open subsets which
satisfy these axioms. Thus any metric X determines a topology, in which
the open sets are the unions of discs open in the metric. There are also
topologies not defined by way of a metric. For example, there is a topol-
ogy on the set N of natural numbers for which the open subsets are the
empty subset and all those subsets S with finite complement (in N). There
are many other examples of topologies.
The definition of continuity (the inverse image of any open set is open)
now applies to any function f: X ...... Y between topological spaces
10. Calculus, Continuity, and Topology 33

X and Y. The three axioms on open sets are enough to prove most of the
basic facts about continuous functions-for example, the fact that the
composite x I--> g(j(x)) of two continuous functions g and j is again con-
tinuous.
To describe continuity at a single point of a space, one may use the
notion of "neighborhood"" A neighborhood of a point a in a topological
space X is any open set of X which contains a. One then says that a func-
tion j: X --> Y between topological spaces in continuous at one point
a E X if to each neighborhood V of j(a) there is a neighborhood U of a
for which j(U) C V. This definition agrees with the previous notion of
continuity at a point for a metric and expresses the intuitive idea that
"nearby" points in U go into nearby points in V. Moreover, j is continu-
ous if and only if it is continuous in this sense at each point a E X.
Extensive experience has shown that this description of a "topology" in
terms of open sets and neighborhoods is extraordinarily effective in for-
mulating all sorts of Mathematical facts in a geometric form. The concept
of "topology" has been appropriately abstracted from the many examples
of "continuity".
The notion of a topological space was first presented by F. Hausdorff
in a famous (and beautiful) book Mengenlehre, His definition was formu-
lated differently, in terms of selected neighborhoods, and included an
added axiom (the Hausdorff separation axiom): Two distinct points have
disjoint neighborhoods, A topological space with this property is called a
Hausdorff space,
We have now seen a number of Mathematical concepts which are
described as sets-with-structure. Thus a linearly ordered set is a set
equipped with a binary relation < having certain specified properties. A
group is a set equipped with a binary, a unary, and a nullary operation,
which together satisfy certain identities. A Boolean algebras is similarly a
set with appropriate operations. A topological space X is a set-with-
structure, where in this case the "structure" consists of a specified collec-
tion of the subsets of X, namely the collection of all open sets. This kind
of structure is quite different in style from the algebraic structures. There
are also structures of a mixed kind. For example, there are cases of
motions (e.g., translations or rotations) which deal with a set of motions
which is both a group and a space. This leads to the notion of a topologi-
cal group, Such a group is a set G which is both a group and a topological
space and in which the group operations-both the product G X G -> G and
the inverse G->G-are continuous, It is this last condition which ties the
two structures together (to make the definition complete, one must know
how the topology on G induces, in a natural way, a topology on G X G).
As in this case, most composite axiomatic structures (combinations of two
kinds of structure on the same set) involve one or more axioms expressing
the formal connection between the two structures-here between the
group structure and the topology.
34 I. Origins of Formal Structure

Here is another example of a mixed structure: A linearly ordered group

G is a set which is a group and also has a linear order, with the added
axiom that a < b in G and I < e implies both ae < be and ea eb.<
This added axiom is the one which ties together the two structures of
order and of multiplication. There are many examples of such linearly
ordered groups-positive rational numbers or real numbers under multi-
plication, or integers with multiplication replaced by addition.
We will see that many Mathematical notions can be described as set-
with-structure.

11. Human Activity and Ideas

This chapter, starting from the study of number, space, time, and motion,
has led to the description of various formal notions-especially cardinal
number, permutation, linear order, group, continuity, and topology. Each
notion represents a type of formalization in Mathematics. The formaliza-
tion may take the guise of a rule (e.g., a multiplication table), a simple
definition (the same cardinal number), a more subtle definition (that of
continuity), a list of axioms describing the common properties of several
systems (linear order), a less evident such list (a group), or a list of axioms
deemed sufficient to describe exactly one object (the real numbers as an
ordered set). In some cases, like that of topological space, the axioms
serve to help understand the common features of a wide variety of situa-
tions.
These formal notions arise largely from premathematical concerns
which can best be described as "human cultural activities". For this rea-
son, our analysis of the genesis of Mathematics will note a number of such
activities. Often it is illuminating to say that the activity leads first to a
somewhat nebulous "idea", which is finally formalized, perhaps formal-
ized in several different ways. For example, the process of counting sug-
gests the idea of "next" -the next item to be counted or the next number
to be used in the count or the next thing in some ordered list. This general
idea "next" may then be formalized by a rule for adding one to each
decimal or by the axioms on the operation which provide to each natural
number its successor. The idea "next" appears in other forms: The next
(infinite) ordinal beyond a given set of ordinals or the next step (after
choice of alternative) in some computer program. Or the frequent obser-
vation of steady changes may suggest the (nebulous) idea of steady
change, formalized (say) by what we called a parametrized motion.
This type of source for Mathematical form, in the cases we have noted
so far, may be summarized in a table, where each activity suggests an idea
and its subsequent formalizations (Table I).
This tabulation is intended to be suggestive but not dogmatic. Each
"idea" is intended to have some intuitive content; it may serve as the car-
II. Human Activity and Ideas 35

Table 1.1

Activity Idea Formulation

Collecting Collection Set (of elements)
Counting Next Sucessor; order
Ordinal number
Comparing Enumeration Bijection
Cardinal number
Computing Combination (of nos) Rules for addition
Rules for multiplication
Abelian group
Rearranging Permutation Bijection
Permutation group
Timing Before and after Linear order
Observing Symmetry Transformation group
Building, shaping Figure; symmetry Collection of points
Measuring Distance; extent Metric space
Moving Change Rigid motion
Transformation group
Rate of change
Estimating Approximation Continuity
Limit
Nearby Topological space
Selecting Part Subset
Boolean algebra
Arguing Proof Logical connectives
Choosing Chance Probability (favorable /total)
Successive actions Followed by Composition
Transformation group

rier for the well known phenomenon of "Mathematical Intuition". The

same idea may arise from different activities, and may well be the back-
ground for several different formalizations. We have tried to use familiar
words to describe each idea but this does not represent any established
consensus or precise definition. On the other hand, each notion, as con-
ventionally formalized, has a rigorous definition (within some context).
The table is by no means complete; as the reader keeps it in mind he
may find new examples in subsequent chapters.
Even after the basic Mathematical notions have been developed out of
these activities and ideas, there continue to be inputs from outside
Mathematics. These inputs often take the form of Mathematical questions
arising in other sciences and requiring application of Mathematics. Thus
the primitive sort of study of motion noted above becomes later the sub-
36 I. Origins of Formal Structure

ject of dynamics (in physics) or that of celestial dynamics in astronomy.

The study of social changes in part becomes the study of marginal costs or
econometrics. In general, under the genesis of Mathematics we intend to
include all sorts of inputs from scientific and other cultural activities.
Some formal Mathematical notions have a more complex origin. Such
is the case for the notion of a "set". The idea of a collection is surely there
when we count, but on this level it is hardly a useful candidate for formal-
ization. Infinite collections also arise, perhaps at first in observations and
in Euclid's proof that there are infinitely many prime numbers-but then
one soon has other infinite collections. They are often subsets of (say) the
set of all natural numbers, but the notion of a subset is not really forced
on our attention until we try to describe the completeness of the ordered
set of reals (Every bounded subset has a least upper bound) or the princi-
ple of Mathematical induction (Every set of natural numbers containing
zero and the successor of each of its elements contains all the numbers).
Even here we might dispense with subsets: Completeness can be described
by convergent sequences and induction can be described by properties.
But Boolean algebra is unthinkable without subsets. The more sophisti-
cated notion of a set whose elements are themselves sets does arise later.
The set of integers modulo 6 will be described as the set whose elements
are the congruence classes such as {l,7,13,19, ... }, right now a topological
space is most clearly defined as a set with a specified set of its subsets
(namely, the open subsets). However, in both of these cases the use of sets
of sets can be avoided by using relations: the relation of congruence
module 6 (Gauss) or the relation stating that the subset U is a neighbor-
hood of the point p. The real motivation for the full use of set theory lies
much deeper, and will be explored in Chapter XI, where we will note the
curious fact that abstract set theory arose from the study of trigonometric
series!

12. Mathematical Activities

The genesis of the more complex mathematical structures tends to take
place within Mathematics itself. Here there are a variety of processes
which may generate new ideas and new notions. We list a few of these
processes in tentative form for further refinement after our more detailed
studies.
(a) Conundrums. Finding the solution of hard problems is one of the
driving forces of Mathematical development. Fermat asserted without
proof that the equations xn + yn = zn for n > 2 have no solutions in
integers. As we will see in Chapter XII, this apparently innocent diophan-
tine equation was one historical source of the whole development of alge-
braic number theory in the 19th century-and so was the principal origin
12. Mathematical Activities 37

of such algebraic notions as that of "ideal" -although the arithmetic

theory of quadratic forms also played a role.
The problem of solving polynomial equations by formulas involving
radicals was a historically important conundrum. For quadratic polynomi-
als the solution is easy, by the familiar "quadratic formula". Early alge-
braists found no such formulas for solutions of the general equation of
5th degree. Using permutations of the roots, it was eventually showed (by
Lagrange) that such a solution was impossible-but the first real insight
into the reasons for the impossibility came with Galois in 1832, (see
Chapter V); this was the point where the notion of a group first explicitly
arose.
Our presentation has in effect argued that the notion of a group could
have arisen otherwise-but in historical perspective the solution of
different Mathematical problems is a vital element in the progress of the
science (and is often viewed as the characteristic aspect of that science).
(b) Completion. The whole list of natural numbers arises by starting
with the first few 0, 1, 2, 3, ... , 9 and asking that there always be a suc-
cessor. Then subtraction, alas, is not always possible-until one creates all
the integers. To insure the possibility of division, one must then have all
the rational numbers, and so on to the real numbers and then to the com-
plex numbers. In many other cases, the need to complete a structure
under some partiaily defined operation brings out a new structure.
(c) Invariance. A non-trivial homogeneous equation

ax + by + cz = °
has infinitely many non-zero solutions, but all can be expressed as sums of
multiples of some two solutions-because, as we know, the set of all solu-
tions (x,y,z) is a plane through the origin in 3-space and any vector lying
in that plane is the sum of multiples of two suitable such. Again the solu-
tions of the homogeneous linear second order differential equation
d 2x Idt 2 = -k 2x all have the form

x = A cos kt + B sin kt ;

they are expressed here as linear combinations of two particular solutions

cos kt and sin kt. These two parallel situations serve to suggest the struc-
ture of a vector space (Chapter VII), the idea of a basis for such a space,
and the need to describe its properties independently of anyone choice of
basis.
(d) Common Structure (Analogy). This example exhibits also the motive
of finding a common structure (here, that of a vector space) underlying
different but similar phenomena (here, geometry, linear equations, and
linear differential equations). Another such instance is given by a descrip-
tion (§4) of linear order. The symmetry group as the commonality of two
38 I. Origins of Formal Structure

superficially different symmetrical figures is another example. Again, the

definition of continuity for a function on a metric space exhibits the com-
mon features of the definition of continuity for functions of one or several
real variables.
(e) Intrinsic Structure. Sometimes a (hidden) formal structure will seem
to explain the facts. Thus a permutation group on n letters always has
order dividing n!; the explanation is provided by the well known theorem
on the possible orders of a subgroup of finite order. The observation of §9
above that different metrics in the plane yield the same continuous func-
tions j(x,y) of two variables is explained by shifting attention from the
(various) metrics to the (common) topology which they define in the
plane. There will be many such cases in which the appropriate abstract
notion serves to provide an understanding of Mathematical phenomena.
Perhaps the progress of Mathematics depends on a counterpoint between
solving conundrums and searching for the concepts which provide better
understanding.
(f) Generalization. The process of generalization in Mathematics takes a
number of different forms. There are generalizations from many concrete
instances to a "general" law; thus 2 + 3 = 3 + 2 and 4 + 7 = 7 + 4
lead to the commutative law x + y = y + x for addition. Or some case,
already general, is extended to be more so: Finite cardinals suggest
infinite cardinals, but the finite laws do not all carry over. Similarly, finite
groups generalize to infinite groups-but theorems on the orders of sub-
groups drop out.
The generalization from analytic geometry in 2 and 3 dimensions to
that in n dimensions is straightforward (but it took a long time histori-
cally); the generalization was not really relevant or interesting until there
was considerations of "events" described by more than three numbers
(coordinates) and a recognition that a geometrical language helped to
visualize and describe such events. Axiomatic structure (like the structure
of a group) may be generalized by deletion of axioms. One of our aims is
to examine enough cases to obtain a more complete typology of generali-
zation in Mathematics.
(g) Abstraction. This process typically consists in getting (some of) the
same results under weaker or more "abstract" hypotheses. A standard
example of abstraction is the shift from groups of transformations to
"abstract" groups. In the concrete case, the elements of a transformation
group are actual transformations acting on a specified set, and the group
multiplication is the composition of transformations; this multiplication is
automatically associative. In the abstract case, the elements of a group
and their multiplication can be arbitrary (i.e., "abstract")-but the multi-
plication must satisfy the necessary group axioms, in particular the associ-
ative law. Under this process of abstraction, two transformatiori groups on
different sets may turn out to be "the same" (i.e., isomorphic). However,
no new groups turn up in this process, in view of the famous theorem of
12. Mathematical Activities 39

Cayley (§8) which asserts that every (abstract) group is isomorphic to a

group of transformations. Hence all those theorems about transformation
groups which do not involve the elements being transformed (e.g., which
are not theorems about transitive or primitive permutation groups) remain
valid for abstract groups. Much the same happens when the "algebra of
sets" is abstracted to Boolean algebra.
Other examples of abstraction can involve real generalization. Thus
Hilbert and others considered number rings-subsets of the complex
numbers which are closed under addition, subtraction, and multiplication.
The abstract notion of a ring (due to Emmy Noether) is that of a set of
elements with operations of addition, subtraction, and multiplication sub-
ject to suitable axioms (§IV.3)-and such an abstract ring need not be iso-
morphic to any ring of numbers.
In these cases, abstraction amounts to considering a structure by
neglecting the provenance of its elements, but keeping "all" the opera-
tions on these elements, with (one hopes) all the formal properties of
these operations formulated as axioms and their consequences.
(h) Axiomatization. This process typically asks: Given a long list of
"all" the theorems on a given topic, can one deduce them all from a suit-
able shorter list-a list which will then constitute the axioms for that topic.
The right choice of axioms can lead to great insight and better under-
standing. Thus Euclid produced congruence axioms for triangles sufficient
to prove all the known theorems about congruence of such figures. In
more recent times, one first observed that much of the geometry of three-
dimensional Euclidean space could be described in terms of vectors
v = (x,y,z) with vector addition

(x,y,z) + (x',y',z') = (x + x', Y + y', z + z')

and the multiplication (x,y,z) I-> (ax,ay,az) of a vector by a scalar.
Axiomatization then asks: what short list of properties of addition and
multiplication by a real scalar will give all these theorems? The answer is
the usual list (Chapter VII) of axioms for a vector space over the reals,
plus the assumption that there is a basis of three vectors. Many other
successful examples ofaxiomatization spring to mind. Processes of
abstraction, as recounted above, also will usually involve axiomatization.
The task ofaxiomatization can be hard, especially for such quasi-
mathematical subjects as mechanics, thermodynamics or (in economics)
utility theory.
(i) The Analysis of Proof One way to find axioms is to ask for a
minimum list of properties needed to carry out a given proof or proofs.
For example, the axioms for a commutative ring are essentially the
minimal list needed for the standard algebraic manipulations of addition,
subtraction, and multiplication. Again, new Mathematical ideas can
sometimes be disentangled by analyzing what makes a given proof tick. A
40 I. Origins of Formal Structure

striking example is the proof that a function f: 1-> R, continuous on a

closed interval I of the reals, is uniformly continuous there. A straightfor-
ward direct proof can be given, using the basic properties of the real
numbers. This proof, originally given by the German Mathematician
Heine, and further developed by the French Mathematician Emil Borel,
leads to the Heine-Borel theorem: If the closed interval I is the union of
an (infinite) collection of opens sets Vi, so that I = U Vi, then it is a
union of a finite number of these open sets,

for some finite list of indices it , ... ,ih. In current terminology, this pro-
perty states that I is a compact subset (of R) and so leads to the idea of
compactness for topological spaces.
At the end of our study of structure, we will return to a more detailed
examination of these processes, internal to Mathematics, for the genera-
tion of new notions. They playa role counterpuntal to the input of prob-
lems from the sciences outside Mathematics. Both are accompanied by
the continued search for deeper properties of the notions already at hand.

13. Axiomatic Structure

In the next three chapters we will indicate how number, space, and time
can be described by axioms; that is, by axioms for the natural numbers,
the Euclidean plane, and the real line which describe these structures
uniquely. In classical terminology, these axiom systems are categorical, in
the sense that any two "models" of the axioms, taken within an inclusive
set theory, are isomorphic-as in the case described in §4 for the reals (we
will also note another "first order" version of these axioms where there
can be non-isomorphic non-standard models). Thus these structures are
closely attached to the traditional view that (say) the axioms of Euclidean
geometry describe one specific object-physical space.
In this first chapter, we have deliberately followed a different order of
axiomatics, emphasizing those systems of axioms (linear order, group,
metric space) which have many essentially different models. This use of
axioms is historically more recent than the categorical axiomatization of
geometry. In particular, it allows for the view that the formal systems stu-
died in Mathematics come in a great variety and are intended primarily to
help organize and understand selected aspects of the "real world" without
being necessarily exact descriptions of a part of that unique world. For
example, our presentation allows that the first step in the formalization of
space could be the description of figures and chunks of space as models of
metric space and not as subsets of Euclidean space. This is by no means
the conventional view.
13. Axiomatic Structure 41

Nevertheless, this chapter has started from the conventional idea of

Mathematics as the science of number, time, space, and motion, to go
beyond these topics to related more general formal notions of cardinal
number, permutation, order, transformation, group, and topological space.
Mathematical experience, as suggested in our subsequent chapters, shows
that each of these notions plays a basic role in Mathematics. We have
deliberately put them first to let the reader judge their importance. This
does not mean that they need be prior to the classical description of
number and space, but simply that they appear in parallel to these classi-
cal notions.
The linear order of a book does not allow the actual presentation to be
in parallel.
CHAPTER "

From Whole Numbers to

Rational Numbers

I. Properties of Natural Numbers

Various human activities such as listing, counting, and comparing lead, as
we saw, to the natural numbers
N = {O,I,2,3,4, ... }
and to the operations of addition, multiplication, and exponentiation on
these numbers. These operations have a variety of general properties. For
example, addition for all natural numbers k, m, and n satisfies the equa-
tions

m + °= m, m + n = n + m, (1)

k + (m + n) = (k + m) + n. (2)

These rules can be proved from the definitions of the operations. For
example, the commutative law (1) holds because, when two disjoint finite
sets are combined, the cardinal number of the combined set does not
depend on which of the two sets is taken first. On the other hand, the
rules are formal in the sense that they can be used directly without atten-
tion to their "meaning". For example, the associative law (2) tells me that
if I add a long column of figures in three successive groups, subsequently
combined, the final result will be the same, irrespective of the order in
which the three are combined. A similar rule will work for more than
three groups. Moreover, these (long-established) rules are inviolate: If it
doesn't tum out as they specify, I know that I have made a mistake some-
where. This is the merit of a formal rule: Once firmly established, it can
be applied mechanically and is an infallible guide.
Multiplication has corresponding formal properties:

m·l = m, mn = nm, (3)

k(mn) = (km)n. (4)

2. The Peano Postulates 43

Together, addition and multiplication satisfy the distributive law

k(m + n} = km + kn. (5)

Again, this law can be used formally, without attention to its origin in the
definitions of addition and multiplication, as suggested in the following
display:

(6)
m n m n

There are many other properties of these operations. For example,

every square, on division by 4, leaves a remainder 0 or 1 (never 2 or 3). If
b > 1, then every natural number n can be written in terms of b as

n = ak b
k + ak-l bk-I +. .. + al b + aD (7)

for some natural k and with coefficients ai all satisfying 0 ai < b. In<
particular, if b = 10, this is the decimal expansion of n, and its properties
lead to the familiar formal rules for manipulating decimals.

2. The Peano Postulates

Each of these properties, and many more, of the addition and multiplica-
tion of natural numbers could be demonstrated directly from the
definitions of these operations on finite cardinal numbers. Such proofs
would be cumbersome. However, a remarkable fact emerges: Both addi-
tion and multiplication can be described just in terms of the number zero
and the single operation "add 1", and their properties can be derived
from a short list of axioms on the single operation. These axioms are the
Peano postulates. The idea is that the natural numbers can be listed, start-
ing with zero, so that to each number n there is always a "next" number,
its successor n + 1, and so that this process exhausts all the natural
numbers. Thus we can state formally:
The (natural) numbers N with zero and "successors" s form a collection
with the following five properties (the Peano postulates):
(i)0 is a number;
(ii)If n is a number, so is its successor sn;
(iii)0 is not a successor (i.e., sn is never O);
(iv) Two numbers n, mwith the same successor are equal (i.e., if
sn = sm, then n = m);
(v) Let P be a property of natural numbers. If 0 has P, and if sn has P
whenever n does, then P holds for all natural numbers.
44 II. From Whole Numbers to Rational Numbers

This is a typical description of a structure by axioms. There are certain

primitive (or undefined) terms: here the terms "number", "zero", and
"successor". The statements of the axioms use only these terms and the
standard logical connectives: "if ... then", "not", "and", "equality", "for
all", "there exists". Such a statement is called a formula (or a formal
statement) in the language of Peano arithmetic (for more detail, see
Chapter XI). In particular, a "property" of the number n, as used in pos-
tulate (v), should be one which is described by such a formula, involving
n.
The induction axiom (v) is vital; it expresses the intuitive idea that tak-
ing successors exhausts all the natural numbers. It is very useful practi-
cally, in proving all sorts of formulas involving general n (for example, the
formula for the sum I + 22 + 32 + ... + n 2 of the first n squares)
and for proving such results as the binomial theorem.
Sometimes the induction axiom is formulated in terms of sets rather
than properties, as follows:
(v') If S is a set of numbers containing 0 and if every n in S has its suc-
cessor in S, then S contains all (natural) numbers.
This axiom implicitly refers to "all" subsets of N, so it is sometimes called
a "second order" axiom, because the quantifier "all" is applied not just to
elements of N, but also to subsets. More specifically, this form of the
axiom means that we are considering the natural numbers in a context of
sets, and that proofs of theorems about natural numbers may use not just
the Peano axioms, but properties of sets, as these might be formulated in
axioms for set theory. In this respect, it is like the completeness property
of the real numbers (§I.4).
The set-theoretic induction axiom (v') does include the property-
theoretic version (v), because the usual axioms for set theory do specify
that every (formal) property of elements of a set N does determine a sub-
set of N. This transition from properties of numbers to sets of numbers is
a familiar one. The use of properties may be called "intensional",
because a property is described by a formula. Thus the properties "n is
odd" and "n leaves the remainder I on division by 2" are verbally
different, but describe the same set {l,3,5, ... }. On the other hand, as in
§1.9, the use of sets is extensional: As soon as two sets include the same
elements, they are equal. The "extent" of the set is all that matters.
However, the induction axiom (v) for properties is weaker than that for
subsets. Since a property, as explained, can be expressed in a finite list of
words in a fixed language, the number of properties of natural numbers is
denumerable. However, for the usual notions of sets, a "diagonal" argu-
ment (see Chapter XI) shows that the number of subsets of N is not denu-
merable but larger. This observation has consequences. One can formu-
late theorems about natural numbers which are true within set theory but
which cannot be proved from the Peano axioms with induction in the
 2. The Peano Postulates 45

form (v). Using sets, one can also construct a model of these axioms
which is "non-standard" in the sense that the numbers are not exhausted
by taking successors (Chapter XI). However, this is not so for the set-
theoretic version with axiom (v'). As we will soon see, one can prove that
these Peano postulates do determine the natural numbers up to an iso-
morphism.
From the Peano postulates one can define all the familiar arithmetic
operations by recursion-specifying in succession the results of the opera-
tion. Thus addition (explicitly, the operation "add k") is defined by the
two recursion equations

k + 0= k, k + sn = s (k + n); (1)

multiplication (multiply by k) is defined by the recursion

k·O = 0, k(sn) = k + kn; (2)

while exponentiation to the base k is given by the recursion

(3)

The other properties (commutative, distributive, etc.) of these operations

may then be proved by induction on n. Each of these pairs of formulas
defines a function f(n) = k + n, kn, or e
by first giving the value f(O)
and then the value f(sn) in terms of some known function g of f(n); in
(1) g is s, in (2) it is k+ -, and in (3) it is k· -. The principle involved
may be stated as a

Recursion Theorem. If X is a set, a an object of X and g: X ~ X a func-

tion, then there is a unique function f: N ~ X with

f(O) = a, f( sm) = g(fm ) , (4)

the latter for all natural numbers m.

A proof of this theorem uses axiom (v') and must depend upon the
set-theoretic definition of "function" as a table of values (see Chapter V),
but the idea behind any such proof is a straightforward use of induction,
as follows. For each n let n denote the finite set n = {0,l,2,3, ... , n} and
consider the following property P of n: There is a unique function
fn: n ~ X which satisfies (4) for m = 0,1,2, ... ,n-l. Clearly, P holds
for n = O. If n has the property P, we have the (unique) function fn
whose values may be listed as

o 2
foO /2 2
46 II. From Whole Numbers to Rational Numbers

The desired value of f at n + I can then be adjoined to this table; it is

g(jnn); from the (formal) definition of function it follows that this gives a
(unique) function fn+ I: n + I ~x. Thus n + I has the property P.
Hence by induction all numbers n have P. This gives a list of functions
fo,fl , ... ,fn,fn+h . .. which "fit together" and so (again by the
definition of a function) produce the desired function f on all of N.
The recursion theorem can be stated conveniently in diagrammatic
form, if we observe that an object a of the set X is the same thing as a
function a: I ~X from the one-point set I into X (the function whose only
value is a). Thus we are given the functions represented by solid arrows in
the diagram 0
• N • N
I I
I I
I I
II If
I
I
If
I
I
(5)
I I
a t g t
• X • X;

the recursion theorem then asserts that there is a unique function f

(shown by the two dotted arrows) such that the diagram "commutes"-
that is, such that fO = a (left-hand square commutes) and gI = f·s
(right-hand square commutes).
In this theorem, f is said to be defined by "primitive recursion" from g.
More elaborate versions of the theorem are true. For example, one might
assume instead of g a function h = h(x,n) of two variables (a function
h: X X N ~X) and replace the second equation of (4) by the condition
f(sm) = h(jm,m). Recursions with a "parameter" -such as the parameter
kin (1), (2), and (3)-are also valid.
By recursion one may also prove the

Uniqueness Theorem. The set-theoretic Peano postulates determine the col-

lection N of natural numbers uniquely, up to an isomorphism of the struc-
ture given by 0 and successor.

If N' is another set with a distinguished object 0' and a "successor"

function s ': N' ~N " a homomorphism of the O,s structure is a function
f: N~N' with
f(O) = 0', fsn = s 'fn, for all n. (6)

Such an f is an isomorphism of N to N' if it is a bijection (a one-to-one

map of N onto N '). By the recursion diagram (5), with X replaced by N "
there is such a homomorphism. Since N' with s' also satisfies the Peano
postulates, the same recursion diagram produces a unique homomorphism
g: N' ~N in the opposite direction, as in the bottom row of the commuta-
tive diagram below:
Now compare the composite function g-j: N ---N with the identity func-
tion I: N ---N. They both make the diagram
0
-N ~N
II II
II II
Ii
I"g·! IiII,g.! (8)
II II II
0 H tt
- N -N

commute. Hence, by the uniqueness assertion of (5), g-j = I. Similarly,

rg is the identity function. Thus f has g as a two-sided inverse under
composition, and so is a bijection.
This result is typical of the axiomatic description of sets with structure.
At best, such a description can determine the model only "up to isomor-
phism". As in this case, an isomorphism means a bijection from one
model to another which "preserves" all the primitive terms involved in the
axioms-as in (6) above. In this case, there are in fact many different but
isomorphic models. For instance, if 100 is viewed as the zero, then the
even natural numbers starting with 100 form a model for the Peano pos-
tulates when the assignment n I-> n + 2 is taken to be the successor func-
tion.

3. Natural Numbers Described by Recursion

The Peano postulates are not the only possible axiomatic description of
the natural numbers. Instead, one can take the recursion theorem as the
(sole) axiom. In detail, this axiom assumes that the natural numbers are a
set N with a distinguished object 0 and a function s: N --- N which together
satisfy (for all a f X and all g: X ---X) the recursion theorem, as pictured
in the diagram (2.5). This approach to the natural numbers was first made
explicit by Lawvere; it is described in some detail in (the first edition of)
Mac Lane-Birkhoff.
The logical equivalence of the two approaches is readily verified. Thus,
we have already seen that the Peano postulates imply the recursion axiom.
Conversely, one may prove that the recursion axiom implies all the Peano
postulates. The most interesting part of this demonstration is that for the
48 II. From Whole Numbers to Rational Numbers

axiom of mathematical induction, for a subset S of N, as summarized in

the following diagram:
0
• N • N
I I
I I
II Ih
I
Ih
I
t s t
0
.. s • s (I)

0
Ii Ii
.. N
" • N

Since S is a subset of N each of its objects x is in N, so the assignment

x f-> x is a function i: S --> N (the inclusion function, as displayed in the
lower part of (1». The induction assumptions on S state that 0 is in S,
giving the function 0: l-->S, and also that each n in S has its successor sn
in S, so that the assignment n f-> sn for n in S is a function s: S --> S as
shown. By the recursion axiom, there is a function h: N --> S with
hO = 0, sh = hs, as displayed in (1). The composite function
f = i·h: N -->N then satisfies the same recursion conditions
10 = 0, sf = fs as the identity function N -->N. Since our axiom asserts
that the conditions determine the function uniquely, f must be the iden-
tity. Thus, each number n is n = fn = i(hn), which states that n is the
element hn of S and hence that the elements in S include all the elements
n ofN.
This case illustrates a general point: The axioms needed to describe a
Mathematical structure (here to describe the structure of N, unique up to
isomorphism) are themselves by no means unique. The recursion
theorem of (2.5) is an especially convenient form of axiom; it states that
the diagram l-->N -->N is "universal" (that is, maps uniquely into every
other such diagram I-->X -->X).

4. Number Theory
Once the Peano postulates are at hand, they yield all manner of specific
results. Division is sometimes but not always possible, but if one tries to
divide m by n one obtains a quotient q and a remainder r, which may be
o but in any event less than n, as in the equation m = nq + r with
<
o r < n. This result is known as the division algorithm. Those natural
numbers which have no divisors (except, of course, for themselves and 1)
are the primes; they appear in a curious irregular order:

2,3,5,7, II, 13, 17, ....

4. Number Theory 49

Every number n can be factored into a product of primes (some of which

may be repeated). No matter how this factorization is obtained, the result-
ing prime factors are unique, except, of course, for their order. The proof
of this unique factorization theorem rests on the division algorithm. From
the prime factorization of two numbers one may read off their greatest
common divisor; however, this also could be found directly from the
numbers by the Euclidean algorithm, which is just an iteration of the divi-
sion algorithm.
The curiously irregular sequence of primes noted above is infinite, by a
proof which goes back to Euclid. One is soon led to try to estimate how
thick the primes are. If '7T(n) denotes the number of primes less than or
equal to n, the prime number theorem (proved with more sophisticated
means) will tell how fast '7T(n) grows as n approaches infinity. Again, if we
arrange all the numbers according to their remainder on division by 3, we
get the following three arithmetic sequences

036 9 12, ... ,

147 10 13, ... ,
2 5 8 11 14, ... .

Except for the prime 3 in the first sequence, all the primes must fall in the
last two sequences. It turns out that there are an infinite number of
primes in each of these two arithmetic sequences, and that they are, in a
sense, equally distributed between those two sequences. More generally,
Dirichlet's theorem asserts that any arithmetic sequence nd + r, for fixed
d and r and increasing n, will have an infinite number of primes, provided
only that d and r have no common factors except 1.
Every number can be written as a sum of at most four squares or of at
most nine cubes. These results have relatively elementary proofs; by
much deeper analysis for Waring's problem, similar results hold for higher
powers. By trial, one can verify that each small even number can be writ-
ten as a sum of two primes. Goldbach (in 1742) conjectured that this was
always true. To date, no one has proved this to be so. The best results to
date are VinogradofI's: Every sufficiently large odd number r is a sum of
three primes, and Chen's: Every sufficiently large even number is a sum
p + b, where p is a prime and b is either a prime or a product of two
primes.
Problems in Diophantine equations ask for solutions in integers and in
natural numbers. The equation x 2 + i = z2 has infinitely many (well-
known) solutions in non-zero integers x, y, and z, but the equation
x4 + l = z4 has none. Fermat stated, and no one has yet proved, that
xn + yn = zn for n > 2 has none. That the numbers of such solutions is
finite has just recently been proved (the Mordell conjecture). Pell's equa-
tion x 2 - Di = 1 has an infinite number of integer solutions, of
relevance to algebraic number theory.
50 II. From Whole Numbers to Rational Numbers

This is but a small sample of the wealth of questions arising for the
natural numbers. All these results are ultimate consequences of the struc-
ture specified with such simplicity by the five Peano postulates.

5. Integers
To keep accounts of gains and losses, subtraction is needed. Within the
natural numbers, subtraction is not always possible, but it becomes possi-
ble when the set N of all natural numbers is expanded to the set Z of
integers. One can formally define the integers (and arithmetic operations
upon them) in several ways. Perhaps the simplest is that of adjoining to
N a new copy of the positive numbers, each prefixed by -, as
-1,-2,-3,-4, .. '. Then addition of the old and new integers is
defined for natural numbers nand m in N in cases:

n + (- m) = n - m, if n >m,
-(m-n), if n < m,

(-n) + (-m) = -(n+m).

With this definition, subtraction is always possible in Z; moreover, similar

definitions describe the appropriate multiplication and order in Z.
Another approach to Z observes that subtraction amounts to solving for
x an equation n + x = m; the ordered pair (m,n) is then introduced for-
mally so as to denote "the" solution to this equation. The familiar rules
for manipulating differences m - n translate to give definitions of sum
and product of such pairs by the formulas

(m,n) + (m',n') = (m +m', n +n'), (m,n )(m',n') = (mm' +nn', mn' +m') .

But beware: the pairs (m,n) and (m+h,n+h) should count as the same,
hence one defines (m,n) = (r,s) if and only if m + s = n + r, and
verifies that this artificial equality satisfies the expected rules; in particu-
lar, that sums and products of equals are equal. The integers, defined to
be these pairs with this equality, do not literally contain the natural
numbers from which we started, but the meaning of subtraction suggests
that each n in N be identified with the pair (n,O); this identification
preserves addition and multiplication. Stated more formally, this says that
the function N ...... Z given by n ~ (n,O) carries sums to sums, products to
products, inequalities to inequalities, and distinct numbers to distinct
integers; it is thus a monomorphism of the structure described by
+, X, and <.
These two constructions of the integers give essentially the same result.
Specifically, the map n ~ (n,O), -m ~ (O,m) is an isomorphism (for
6. Rational Numbers 51

the structure +, x, and <) of the integers as first constructed to those

described by pairs and equality. In either construction, the integers Z can
be described as the minimal way, unique up to isomorphism, of embed-
ding N in a larger structure in which subtraction is always possible,
preserving all the algebraic properties of + and X. Here as elsewhere,
what matters is not an exact description of what an integer is, but a
description of the structure of all the integers, up to isomorphism.

6. Rational Numbers
To keep accounts, one often needs to divide numbers evenly into parts-
and this often cannot be accomplished with whole numbers. Fractions
provide the answer. They are introduced individually, as 1/2, 2/3, 1/5,
4/5, etc., and then manipulated in the evident way:

m m' mn' + m'n m m' mm'

n
+n'- nn' n n' nn' .
(1)

This practical process suggests a corresponding formalization. The ini-

tial observation is that division is not always possible within the set N + of
all positive natural numbers. Hence one is led to introduce the set Q + of
all ordered pairs (m,n) of positive natural numbers, defining addition and
multiplication by the evident translations

(m,n) + (m ',n ') = (mn' + nm', nn '), (m,n)(m ',n ') = (mm ',nn ') (2)

of the practical rules (1) and taking care to define the equality of (m,n)
and (r,s) by ms = nr. When m in N+ is identified with the pair (m, 1) this
again defines a minimal expansion of the set N + to a larger set in which
division is always possible and in which all the rules of arithmetic still
hold. As before, there is nothing unique about the formulation of this
construction. Instead, one might have used only those pairs (m,n) in which
m and n have no common factor (except I); in that case, one must modify
addition and multiplication in (1) to reduce each answer to lowest terms.
With this inconvenience, the "artificial" definition of equality of pairs is
avoided. Again, what matters is only the resulting structure up to isomor-
phism.
The system Q of all rationals may then be obtained from Q+, the posi-
tive rationals, by simply adjoining zero and negative rationals. Alterna-
tively, one may construct Q directly from Z by using all pairs (a,b) of
integers a,b in Z, with the same addition and multiplication as in (2)-and
the important proviso that the "denominator" b is never O.
As in previous cases, what matters is not the explicit definition of a
rational, but the resulting structure.
52 II. From Whole Numbers to Rational Numbers

7. Congruence
A typical clock runs up to the figure of 12 hours and then repeats, but one
can still do arithmetic on the limited list of hours: Seven hours after nine
o'clock is four o'clock. Similarly, in the decimal system there are only ten
digits 0, 1,2, ... , 9; the usual rules for addition and multiplication, ignor-
ing the carryover to the tens' place, work perfectly well for the manipula-
tion of these digits by themselves:

6 +7 = 3, 8 + 7 = 5, 8·3 = 4, 3·9 = 7.

These rules ignore all the multiples of 10; in a sexagesimal system there
are similar rules which ignore all the multiples of sixty. "Casting out
nines" is a rule for checking arithmetic calculations. This rule for checking
a multiplication says: Add up the digits of each factor, multiply the result-
ing sums, and check this against the digit sum of the original purported
answer. Thus 32 times 27 calculates to 864. To check, 32 becomes 5, 27
becomes 9, 5 times 9 is 45, with digit sum 9. This checks with the digit
sum in the purported answer, which is 8 + 6 + 4 = 18 with digit sum 9.
What happens here is that 32 is replaced by 5, casting out the difference
which is 27, or three nines. The reason it works is that factors differing by
a multiple of nine will have a product differing (at most) by a multiple of
9. In brief, arithmetic operations are valid "casting out 9's" or "modulo
9".
These examples each involve the use of a modulus: 12, 10, 60, or 9, as

°
the case may be. The general procedure is similar. For integers a, b, and
any natural number m oF as modulus, one writes a =: b (mod m), or
says that a is congruent to b for the modulus m, when the difference
a - b is a multiple of m. Then one readily proves the arithmetic rules: If
a =: band c =: d, both mod m, then

a + c =: b + d (mod m), ac =: bd (mod m). (1)

This congruence modulo m behaves like equality; also it is reflexive, sym-

metric and transitive. ("Transitive" is defined in § 1.5; a relation such as =:
is symmetric when a =: b implies b =: a for all a and b.)

the same remainder r, with °, ; ;

Two integers a and b are congruent modulo m if and only if they leave
r < m, upon division by m. As a result,
calculations modulo m amount to calculations with a finite list of objects
(to wit, with the remainders 0,1,2, ... ,m -I). All the rules for addition
and multiplication-commutative, associative, and distributive laws-still
hold for these finite calculations. Thus the remainders modulo m form an
(abelian) group under addition. Under multiplication, the non-zero
remainders modulo a prime p also form a group of p - I elements. This
is not the case for a composite modulus m, such as 2·3, because there
°
2·3 =: (mod 6) so that neither 2 nor 3 can have a multiplicative inverse
7. Congruence 53

modulo 6. To get a multiplicative group for such a composite modulus m,

on must use only those remainders r which have no factor in common
with the modulus m. The number of such remainders is denoted by cp (m),
while cp is called Euler's cp -function. For a prime p or for integers m,n
with greatest common divisor lone readily calculates that

cp(p) = p-I, cp(pk) = (p_l)pk~l, cp(mn) = cp(m)cp(n). (2)

These formulas provide for a computation of any cp (m) from the prime
decomposition of m. We cite them to emphasize that the formulation of
congruence arises both from practice (multiplying hours or digits) and
from number theory.
To say that calculations with congruences are calculations with the
remainders is a bit artificial. Thus modulo 5 one could replace the five
remainders 0, 1,2,3,4 by the remainders -2, -1,0, 1,2 or by -4, -3,
- 2, -1, 0. Here, as always, mathematicians strive for an invariant formu-
lation. Each remainder r stands for (and may be replaced by) the
"congruence class" emr of all integers a with a == r (mod m). To add the
class emr to the class ems one may then take any representative a in emr,
any b in ems, add a and b, and take the class of this sum a + b as the
sum emr + ems. One must then prove that the resulting sum of classes
doesn't depend on the representatives a and b chosen-but this fact is just
a restatement of the rule (1) for adding two congruences. With this fact
established, we see that the collection Zm of all these congruence classes
em forms a system with binary operations of addition and multiplication,
and that the function em from Z to Zm, as in

carries the addition and multiplication of integers to that of congruence

classes. (It is thus a first example of a homomorphism of + and X.) This
gives an "invariant" formulation of the calculation with remainders.
Thus we have (at least) three descriptions of the algebra of integers
modulo m: As the ordinary integers taken with a new equality,
congruence modulo m; as the algebra of remainders modulo m; or as the
algebra of congruence classes, modulo m. The last description is the more
invariant-and the more sophisticated, since it involves a set whose ele-
ments are sets (a collection of congruence classes). However, all three con-
structions yield isomorphic results, and the results are useful (and practi-
cally indispensable) for the statement of simple number theoretic facts.
For example, for any integer x, one has always

x 2 == ° or 1 (mod 4), x 2 == 0, 1, or 4 (mod 8).

Another problem is that of finding a common solution x for two (simul-

taneous) congruences:
54 II. From Whole Numbers to Rational Numbers

x == b (mod m), x == c (mod n). (3)

In this situation, the "Chinese remainder theorem" states that if m and n

have no common factors (except I), there always is a solution, unique
modulo the product mn.

8. Cardinal Numbers
The question arises: What, after all, is a natural number? One explana-
tion says that it is a cardinal number. For this purpose, as in § 1.2, define
two sets Sand S' to be equinumerous (or, cardinally equivalent), in sym-
bols S == S', when there is a bijection b: S ~S'; that is, a one-to-one
correspondence between Sand S'. This relation between sets is reflexive
(because the identity function is a bijection), symmetric (because the
inverse of a bijection is again such), and transitive (because the composite
of two bijections is again such).
Arithmetic operations work appropriately under this equivalence rela-
tion. To add two collections Sand T, first make sure that they are disjoint
(have no objects in common), then take as sum the collection S + T of
all objects in either S or T. This disjoint union is preserved by bijections:
if b: S ~S' and c: T ~ T' are bijections, they determine together a bijec-
tion S + T ~S' + T'. Therefore

S == S' and T == T' imply S+T==S'+T'. (I)

To get a product of S by T one needs to add T copies of S; that is, to

count the points in a rectangle [=:=J S units wide and T units high. Call
such a rectangular set S X T the product of the sets Sand T; it can be
described as the set of all the ordered pairs <s,t> where s E Sand
t E T. The use of bijections shows, much as in (l), that

S == S' and T == T' imply SXT==S'XT'. (2)

Exponentiation is repeated multiplication; thus S3 = S X S X S is

the set of all ordered triples <SJ,S2,S3> of elements Sj in S (with possible
repetitions). This is equivalently the collection of all functions i I-> Sj on a
three-element set {l,2,3} to the set S. Generally, the exponential set ST
denotes the set of all functions f: T ~S. Here one again has

S == S' and T == T' imply ST == S,T' (3)

Explicitly, a bijection b: S~S' yields by f ~b'f a bijection ST ~S'T,

while a bijection c: T ~ T' yields by f' ~ f ' .c a bijection S T' ~ S T, going
in the opposite direction.
8. Cardinal Numbers 55

On this basis, a finite cardinal number is just a finite set taken "modulo"
cardinal equivalence; that is, with cardinal equivalence taken as the
equality. In other words, a number is "represented" by a finite set, and
two sets count as the same if they are in bijection. Alternatively, if we
want not a "representative" of the number, but a single object, we may
define the cardinal number of S to be the set of all sets S' equinumerous
with S:
card S = {S' I S == S '} (4)

The arithmetic rules above for congruence then justify definitions of the
sum and the product of cardinal numbers by

(card S) + (card T) = card( S + T), (card S)( card T) = card( S X T) (5)

This definition of cardinal number is strongly analogous to the invariant

definition of a congruence class module m, as in §7. Historically,
congruence classes understood well before the more general definition of
cardinal number. There is a difference: The congruence class emr of r
modulo m consists of all integers a in a given set of integers with a == r
(mod m), but the cardinal of S, as defined in (4), consists of all sets S' in
the whole world with S' == S.
Some discomfort with this wide scope of things is in order. If one is
using here the "naive" idea that a set S is any collection whatever of
identifiable things (and that two sets are equal when they contain exactly
the same things), then there can be paradoxes. To formulate one paradox,
note that there can be sets (such as the set of all infinite sets) which are
members of themselves-and other sets not members of themselves. Thus
Bertrand Russell, noting that we can define a set by specifying its ele-
ments, considered the set R consisting of those sets which are not
members of themselves:

R = {S I not (S ES)}. (6)

In other words

S ER if and only if not (S ES). (7)

He then asked: is R a member of itself, or not? If so, then R E R, so by

(7), not (R E R). If not so, then not (R E R), so by (7) read backwards,
with S = R, one has R E R- a contradiction either way.
This Russell Paradox exhibits a difficulty with the naive notion of set,
when coupled with the idea that one can specify a set by collecting all the
objects with a given property P to form a set T:

T = {X I X has the property P}. (8)

56 II. From Whole Numbers to Rational Numbers

(Call this the naive comprehension axiom.) One (standard) way to avoid
this trouble is to apply this comprehension axiom only to construct subsets
of some already given set W-as, for example, subsets of the set Z. This
"bounded" comprehension axiom then reads: Given a set Wand a prop-
erty P of sets, one may form the set

T = {X I X E Wand X has the property P)}. (9)

Later (in Chapter XI) we will examine this axiom in the context of a
systematic axiomatization of "set" and "element of a set". For this pur-
pose, we also assign to each set U its power set: All subsets of U:

PU = {S I S c U}. (10)

For the present, to explicate cardinal numbers, start with some (infinite)
initial set Vo, called a type or a "universe"; for example, Vo might be the
set of natural numbers or the set of all real numbers. The "next" type VI
is to consist of all the subsets S of Vo, so can be described as VI = P Vo.
Now we can describe the cardinal number of such a set S as the set of all
equinumerous S' which are also subsets of Vo

card S = {S' IS' E VI and S == S'}. (II)

This cardinal number is then a set in the next higher type V2 = PVI •
This approach makes the cardinal number depend on a choice (or hierar-
chy) of types, but the definition (11), in contrast to (4), uses only the
bounded comprehension axiom so that Russell's paradox no longer arises.
The use of successive "universes" Vo, VI, V2 is a highly simplified version
of the "type theory" invented by Russell to avoid his paradox. (More care
is needed in details; for example to arrange that a type is closed under
products.)
Any such approach defines the finite cardinal numbers in terms of sets
and bijections and derives their arithmetic properties (addition, multipli-
cation, exponentiation) from properties of sets (and bijections). It serves
also to define the cardinal number of an infinite set, and arithmetic opera-
tions upon such infinite sets.
In this way the natural question: What is infinity? receives a numerical
answer. There are different sizes of infinity, and they can be measured by
suitable "cardinal" numbers which are subject to arithmetic operations.

9. Ordinal Numbers
The cardinal number of a finite set S does not depend on the order in
which the elements of S are presented or counted. However, for other
purposes, we often list the elements of a set in a given order, as we count
9. Ordinal Numbers 57

first, second, third, ... , and not just one, two, three, .... This leads to the
notion of an ordinal number. Such a number is attached to a suitable
linearly ordered set (P, <). If P I is a second such set, an order-
isomorphism, as in §IA, is a bijection I: P~P which preserves the order,
in the sense that PI < P2 in P implies IpI < Ip2 in pi; when there is
such an isomorphism, the ordered sets are said to be ordinally equivalent
(in symbols, P~PI). Then a finite ordinal number can be defined to be a
finite linearly ordered set P taken modulo ordinal equivalence-that is,
with ordinal equivalence as equality. Alternatively, as with congruence
classes of integers modulo m or as with cardinal numbers, we can define
finite ordinal numbers as equivalence classes in a type VI

ord P = {P I IP I EO VI and P ~ Pi} (1)

Arithmetic operations apply to ordered sets. Thus, if P and Q are dis-

joint ordered sets, we can order their (disjoint) union P + Q by taking
the elements of P, in order, before those of Q, in order. The product
P X Q of two linearly ordered sets has a lexicographic order defined by

If Q is finite, the exponential set pQ of all functions I: Q~ P has a similar

lexicographic order: Given functions I and g, find the first q in the given
order of Q for which I q "*
gq, and then specify that I < g if and only if
I q < gq. For each of these three operations we have the rules for
equivalence, such as:

P ~ P' and Q~ Q I imply P X Q ~ pi XQ ' . (3)

The product of two ordinal numbers is then defined by setting

(ord Q)(ord P) = ord(P X Q); (4)

here the reversal of the factors follows tradition. The result is independent
of the choices of P and Q within the classes ord P and ord Q because of
the rule (3). The definitions of sum and exponents for ordinals is analo-
gous. For finite ordinals, these operations agree with the corresponding
arithmetic operations on finite cardinals, because the cardinal number of a
finite ordered set does not depend on the order.
This is not so for infinite sets. If P is an infinite ordered set, the set
ord P defined in (3) is usually called the order type of P. For example, the
denumerably infinite set N+ of positive integers with its usual order has
an order type quite different from that of the set N- of negative integers
(N+ has a first element, but N- has none). The order type of N- is not
regarded as an infinite ordinal number.
58 II. From Whole Numbers to Rational Numbers

For ordinal numbers, the leading idea is that there is always a "next"
one-the first ordinal "beyond" a given set of ordinals. More generally,
this should mean that any set of ordinal numbers has a first element; this
leads to the definition: A (linearly) ordered set P is well-ordered when
every non-empty subset S c P has a first element in the given order. For
the natural numbers N, the axiom of induction is equivalent to the asser-
tion that every non-void subset T of N has a first element-that is, to the
requirement that N in its given order is well-ordered.
Generally, an ordinal number is defined to be the order type in the
sense (3) of some well-ordered set. This definition provides, for example,
for many infinite but denumerable ordinal numbers. The ordinal number
of the (well-ordered) set N is usually written as w; then the arithmetic
operations apply to w as well as to finite ordinals. For example,
w + w = w2 is the order type of two copies of N, one following the
other. On the other hand, 2w is the order type of w copies of 2, one fol-
lowing the other; hence 2w = wand 2w =1= w2! The first few infinite ordi-
nals come in order as

0,1,2, ... , w, w+ 1, w+2 ... , w2, w2+ 1, w2+2, . .. , ... , (5)

where each ... stands for the whole string of natural numbers from 3 on.
This display suggests that the ordinal numbers are themselves well-
ordered, where ord P < ord Q means that P is order isomorphic to an
"initial segment" of Q. Moreover, each ordinal number is the order type
of the ordered set of all preceding ordinals. This motivates an alternative
description of an ordinal number as the ordered set of all the preceding
ordinals:

°= 0,1 = {O}, 2 = {O,l}, ... , w = {0,1,2, ... }. (6)

In other words, each ordinal nun /er is a member of each succeeding

ordinal. This suggests a formal definition of an ordinal number as a set
whose elements are linearly ordered by the membership relation. We will
return to this idea in §XI.2 when we analyze more closely the notion of a
set.

10. What Are Numbers?

This chapter has proposed several alternative answers to this question, as
follows:
(a) The natural numbers are just a sequence of marks

I, I I, I I I, I I I I, I I I I I,···
lO. What Are Numbers? 59

for use in counting or labeling. Any other sequence of marks would

do as well, provided the sequence always provides a way to write
down the next (the successor) mark. This is the case with decimal
notation

0,1,2, ... , 10,11,12, ... , 100, ... , 200, ....

(b) Natural numbers are finite sets, with equinumerous sets regarded as
equal numbers, while successor means "adjoin one more element".
(c) Natural numbers are cardinal-equivalence classes of finite sets.
(d) Natural numbers are ordinal-equivalence classes of finite ordered sets,
while successor means "adjoin one new element, to come after all the
others".
(e) Natural numbers are finite sets of sets, linearly ordered by the
membership relation, as in (6).
In each of these cases, the natural numbers as described do satisfy the
Peano postulates. From this multiplicity of answers to our question, we
must conclude that there is no answer to the question: What is a natural
number? There are various alternative concrete descriptions, depending
on the sort of counting intended or on the prior assumptions. In each
case, the description provides "numbers" which do satisfy the Peano pos-
tulates. Hence we conclude that one does not define what a natural
number "is", by itself. Instead, one defines the system of all natural
numbers, with successor operation. Then N is any such system which
satisfies the Peano postulates. This means that there are many such
systems within set theory-but that they are all isomorphic, just as in the
case above of the decimal notations.
Note that the postulates themselves are by no means unique; for exam-
ple, the Peano postulates may be replaced by the recursion theorem as an
axiom. Here, as in other axiomatic descriptions of mathematical objects,
there are a variety of choices for lists of axioms. Number theory, like
other subjects in Mathematics, is not the study of a unique model nor yet
the examination of a unique axiomatic system-it is rather a study of the
form exemplified by the various models and specified in the axioms.
To summarize: The natural numbers start out from elementary opera-
tions of counting, listing, and comparing; they then develop into effective
tools for calculation. The rules for calculation are formal and can be
organized as the consequences of simple systems of postulates. The conse-
quences of these postulates include the remarkably varied and rich prop-
erties studied in number theory-properties by no means apparent in the
original processes of counting and listing. They are none the less implicit
in these elementary human activities. Number theory is inevitable.
But number theory is not self-contained. First, calculations with natural
numbers do not allow all subtractions or all divisions, and so require the
60 II. From Whole Numbers to Rational Numbers

constructions of other systems of numbers such as Z and Q. Secondly,

number theory suggests and requires calculations with congruences and
then with integers modulo m-Mathematical objects which are really not
simply numbers, but classes of numbers. Finally numbers also lead inex-
orably to the study of sets and functions. Thus bijections are required to
explain when two finite sets have the same cardinal number, binary rela-
tions are required to understand the ordinal numbers implicit in the list-
ing process, while subsets are required to formulate the induction axiom.
Then the explication of congruence modulo m and of cardinal and ordinal
numbers requires sets of sets, namely, sets of congruence classes (see also
§4).
Thus understanding simple and elementary Mathematical ideas leads
inevitably to general and abstract notions.

References. There are many sources for number theory; one good one is
Hardy and Wright [1983]. For a truth about natural numbers not
demonstrable from the Peano postulates, with induction on properties, see
the last article in Barwise: Handbook of Logic.
CHAPTER III

Geometry

This chapter is concerned with the origins of elementary plane geometry,

its close relation to groups of transformations, and the notions of non-
Euclidean geometry. Subsequently, n-dimensional geometry will appear
with linear algebra in Chapter VII, manifolds and spaces with curvature
in Chapter VIII, and topology, in its connections with complex analysis, in
Chapter X. These and other aspects of geometry pervade Mathematics.

I. Spatial Activities
Geometrical figures and facts arise from a wide array of activities and
observations. Some involve motion; thus in watching motion we see
objects falling in vertical lines, water waves expanding in circles from a
dropped stone, dun eland grasses moved by the wind to describe semicir-
cles on the sand, tree branches oscillating back and forth in the wind, long
straight lines of ocean waves approaching a beach, and the like. There are
also motions which we initiate and then watch-the thrown ball falling,
the log rolling downhill, the circle turned into a wheel for a cart, and so
on. Some activity involves construction. Thus a post or a column will stay
upright in balance if it is set perpendicular to the floor as a vertical; a
three-legged stool is more likely to be steady than one constructed with
four legs; rods joined to form a triangle stay rigid, but rods joined in a
square may wobble. Also, pieces of a board cut apart will fit together
again to make up the same size board. Mapping a labyrinth or painting a
scene calls for reproducing a variety of shapes, each on a smaller scale.
Fitting an object inside another highlights differences in shapes. To check
ahead of time that a fit will be possible may require a measurement of
length or of circumference. Estimating separation at a distance requires
lines of sight, angles, and triangles. These are but a few of the many
activities that go into the formation of geometrical iOeas.
62 III. Geometry

From these activities one may disengage various figures: Circles and
ovals; lines, horizontal, vertical, and perpendicular; triangles, equilateral
or otherwise; squares and rectangles; cubes, cylinders, as with logs or
rods, and spheres, as with oranges or balloons. One also discovers a
variety of facts about these figures: A triangle is rigid because three sides
(or two sides and the included angle) determine the triangle up to rigid
motion. A right triangle can come in various shapes, but in every case the
square constructed on the longest side is the sum of the squares on the
two shorter legs (a fact said to have been useful in re-establishing prop-
erty lines after the Nile had been in flood). Two different perpendiculars
to the same line would never meet. Two triangles with corresponding
angles equal have corresponding sides proportional-and so on.
After a considerable array of facts about geometrical figures are at
hand, some connections between these facts come to light. Once the area
of a triangle can be computed from the lengths of its sides, it is possible to
fit four congruent triangles together so as to prove the Pythagorean
theorem, by adding up the five areas in Figure 1. Or, the Pythagorean
theorem can be deduced from facts about ratios of sides in similar trian-
gles.
Explicitly (Figure 2), the perpendicular from the right angle divides the
hypotenuse e in Figure 2 into pieces hand k with h/a = ale and
k/b = ble. Clearing of fractions, he = a 2 and ke = b 2 , so that (h + k)e
= e 2 = a 2 + b2, q.e.d. Many more such logical relations between
geometrical facts gradually appear.

Figure 1. The Pythagorean theorem.

/
a
/ k
/
/
/
b

Figure 2
2. Proofs Without Figures 63

Nevertheless, it is truly remarkable that our experiences of spatial

extent and of motion through space can be organized so completely. By
taking certain figures-line, plane, segment, angle, and circle-as basic
ones, and by assuming certain simple facts as axioms, other geometric
figures can be constructed and other geometric facts deduced from the
assumed axioms. Some of the resulting facts are surprising; for example,
the three medians of a triangle meet in a point, as do the three altitudes.
The resulting deductive structure of Euclidean geometry is a model of
Mathematical method.

2. Proofs Without Figures

In elementary Euclidean geometry, proofs of various facts about figures
are developed from the axioms, with occasional implicit use of intuitively
evident properties of the figures concerned. For instance, to find the per-
pendicular bisector of a segment AB one constructs two circles with
centers A and B and both with radius AB; these circles obviously intersect
on opposite sides of the line AB, .... But what is meant by a "side" of AB
and why must the circles intersect? (See Figure l.) If all facts are to be
deduced from axioms, then these plausible observations must also be
proved from the axioms, rather than from inspection of the figure. Once
this austere requirement was recognized, it turned out to be possible to
add suitable new axioms to those of Euclid so that all arguments could be
free of any intuitive or pictorial content. This meant that geometry was
not really "about" the figures themselves, but was about certain
corresponding notions defined exclusively in terms of basic notions to be
taken as primitive or undefined. This austere program was systematically
carried out by David Hilbert in a book, Grundlagen der Geometrie (B. G.
Teubner, first edition 1899, twelfth edition 1977). His axioms were for
both plane and solid geometry; we will examine only those for the plane.
He presented the axioms in the five following groups:

Group I (Incidence Axioms). These axioms involve only three primitive

notions: "point", "line", and "the point P lies on the line I". They
require:
(I) Two distinct points P and Q lie on one and only one line. (Except for
the meticulous logical formulation, this is the familiar requirement
that "two points determine a line".)
(2) There are at least two points on every line.
(3) There are three points P, Q, R not all on a line.
With these axioms, one can define a triangle (the figure formed by three
non-collinear points). However, one cannot yet define "right angle".
64 III. Geometry

Figure I. Do the circles meet?

Group II (Axioms of Order). In Euclidean geometry a line does not have

a selected direction, so we cannot describe the order of points P, Q on a
line in terms of binary relations such as "P lies to the left of Q" or "P
comes before Q" or "P < Q". Instead, the description is given means of a
ternary relation "B lies between A and C', available when A, B, and C
are all known to lie on a line. The axioms then read:
(I) If B is between A and C, then B is between C and A.
(2) If A and B are distinct points on a line t, there exists on t a point C
between A and B and a point D such that B is between A and D.
(In elementary geometry, one often speaks of "prolonging the line seg-
ment AB beyond B. This axiom provides the existence of at least one
point D in this prolongation.)
(3) If A, B, and C are three distinct points on a line I, then exactly one of
these three is between the other two.
With these axioms, one can define the segment AB for A =1= B, to consist
of all the points C between A and B on the line determined by A and B.
One is now in a position to formulate such pictorially obvious facts as '~
line divides the plane into two parts". Specifically, this means that given a
line I, all points in the plane but not on I can be put in one of two non-
empty disjoint sets U and V (the two "sides" of t) such that any segment
AB joining a point A in U to a point B in V meets I in one point, while
any segment joining two points of U (or, two points of V) does not meet I.
To establish this "intuitively evident" separation property, it is necessary
to assume an additional axiom:
(4) Pasch's Axiom. If a line I meets the segment AB of a triangle ABC
and does not contain C, then I meets either AC or BC (Figure 2).
In pictures, this states that a line which crosses into the triangle over side
AB must again come out, either through AC or through CB (or through
C). From the austere, no-pictures, point of view, it was this axiom which
was most sharply missing in the Euclidean formulation. This axiom will
show that each triangle divides the plane into two parts, an "inside" and
2. Proofs Without Figures 65

A~------------~--------~B

Figure 2. Pasch's axiom.

an "outside". One can also prove that any closed polygon (which does not
cross itself) divides the plane into two parts, and inside and an outside-
but the proof is quite involved because the polygon may be convoluted,
with many reentrant angles.
With these axioms, one can also define that is meant by a ray or "half
line" starting at a point A. If B is a second point and I the line determined
by A and B, then the ray r from A containing B consists of all the points
between A and B, the point B, and all points D with B between A and D.
An angle L rs between two rays rand s can then be defined as the figure
formed by two (distinct) rays rand s from the same point. In particular,
this defines the (three) angles of a triangle ABC. One also defines straight
angle (formed by two rays on the same line).

Group III (Congruence). Next one introduces two new undefined terms,
"Segment AB is congruent to segment A'B"', in symbols AB == A'B', and
"Angle rs is congruent to angle r' s "', in symbols Lrs == Lr's '. The
corresponding axioms require:
(l) Given a segment AB and a ray r from A " there is a unique point B'
on r with AB == A 'B'.
In simpler language, this states that we can "layoff" the length AB on a
given line from A ' in a given direction.
(2) Congruence of segments is reflexive, symmetric, and transitive.
(3) AB == A 'B' and BC == B'C' imply AC == A 'C', provided B is
between A and C and B' between A ' and C'.
This amounts to describing the addition of segments.
(4) Given an angle Lrs and a ray r' from A' on a line I, there is a unique
ray s' from A ' on a given side of I so that Lrs == Lr 's '.
This axiom specifies that one can "layoff" the angle Lrs from a given ray
r' at A ' and on a given side of the line of r '.
(5) If two triangles ABC and A 'B'C' have AB == A 'B', BC == B'C' and
LB == LB', then also LA == LA' and LC == LC'.
66 III. Geometry

Given this much about the two triangles, one proves that AC == A 'C' and
hence that the triangles are congruent. This is the familiar first congruence
theorem (side-angle-side, or SAS) of Euclidean geometry. It is convention-
ally "proved" by moving one triangle till its parts coincide with the other.
Though we regard such "motion" as lying in the practical origins of
geometry, it use in a formal axiomatic proof is not acceptable; hence this
axiom. From the congruence axioms, one can define right angles.

Two lines are defined to be parallel when they do not meet (have no
point in common). One then requires the following famous axiom:

Group IV (The Parallel Axiom). Given a point A not on the line m, there
is at most one line through A parallel to m.

Group V (Continuity Axioms). Examples show that the axioms stated so

far do not yet insure that the Euclidean plane contains all the points that
should be there, and no more. To achieve this, two more axioms are
needed. To state the first, notice that a segment AB on a ray from A can
be laid off repeatedly to give n points B = B), B2 , ... ,Bn on the ray with
AB J == BiBi+1 and each Bi between A and B i+ l , for i = 1, ... , n-l
(Figure 3).
(1) Archimedian axiom. Given C on that ray from A, there is a natural
number n with C between A and Bn.
That is, multiples of ABI will eventually exceed the given segment AC; in
other words, C, as measured by the unit AB, cannot be infinitely far away.
(2) Completeness. The system of points and lines satisfying the given rela-
tions of incidence and congruence cannot be a part of a larger system
of points and lines with the same relations satisfying all the same
axioms.
These are the two axioms of continuity as formulated in Hilbert's book.
However, this pair of axioms can be replaced by a single axiom which
concerns the division of a line I into two rays. Each point 0 on I divides I
into two rays; if S is the set of all points A on one ray and T that on the
other (and 0 is put in neither ray), then 0 lies between any point A of S
and any point B of T. The axiom desired is essentially a converse to this:

A c

Figure 3. The Archimedean axiom.

3. The Parallel Axiom 67

Dedekind's Axiom. Let all the points on a line I be divided into two non-
empty disjoint subsets Sand T in such a way that no point of S is
between two points of T and no point of T is between two points of S.
Then there is a unique point 0 on I with the following property: For
points A, B not 0 on I, 0 is between A and B if and only if A lies in Sand
B in T, or A in T and B in S.
In other words, the point 0 divides the line into Sand T. We will see in
Chapter IV that this axiom is modeled on a similar axiom for the real
numbers.

From this list of axioms all of plane geometry can be developed in

austere and logical completeness, even though the brevity of our presenta-
tion does not make clear the austere beauty of the resulting development.

3. The Parallel Axiom

The parallel axiom (Group IV above) may be intuitively less convincing
than the other axioms, and so has been the source of extensive discus-
sions. The axiom as we have formulated it above requires only that there
be at most one parallel to a given line m through a given point not on m.
This formulation is chosen because we can and will prove (from the previ-
ous axioms) that there is at least one such parallel.
For the proof, consider a line k which meets two other lines I and m in
points A and B, respectively; in classical terminology, k is transversal to I
and m, as in the Figure 1. In this figure, angles (like a: and 13) between k
and I or m which have one ray along AB are called interior angles. For
this figure, we prove

Lemma 1. If the interior angles a: and 13 on the same side of the transversal
k to I and m have sum a: + 13 = 180·, then the lines I and m are parallel.

(Note incidentally that the statement, carefully formulated, makes use of

the "sides" of k.)

---'----'-----'--{J _ _ _m_ _ _ ~-~}

Figure I
68 III. Geometry

PROOF. If I and m are not parallel, then by the definition of parallel these
lines must meet in some point P. By symmetry, we may take P to be on
the same side of k as the angles a and f3. Then (and with covert refer-
ences to Figure 2, which cannot "really" exist) we layoff on m, on the ray
opposite BP, a segment BP I == AP. Since LP I BA + f3 is a straight angle,
so equal to 180·, it follows that the angle pi BA must equal a. Therefore
the triangles piBA and PAB have corresponding sides (P' B == PA ),
angles a, and sides BA == AB congruent, so by the basic congruence
axiom (side-angle-side, or SAS), they are congruent triangles. Hence angle
P'AB of the first triangle is congruent to angle PBA (or f3) of the second.
Since f3 + a = 180·, this in turn means that the angle P'AB + BAP =
180· is a straight angle, so that P'AP is a straight line. It must then be the
given line I, which is thus revealed to be a line meeting m in two different
points P' and P on opposite sides of k (note the essential use of "sides" of
a line, which can be defined by using Pasch's axiom). From this contradic-
tion, we deduce that I must have been parallel to m.

Theorem 2. Through a point A not on a line m there exists at least one line
I parallel to m.

PROOF. On the line m there is at least one point B. Join this point to A.
The segment AB makes with m two angles y and f3 which together form a
straight angle, so y + f3 = 180·. Layoff the angle y at A, with one side
along the segment AB and the other on a ray AP along the (new) line I';
then BAP, on the same side as f3, is equal to y (Figure 3). By the lemma,
I' must then be parallel to m, as desired.
Up to this point, we have not invoked the parallel axiom.

Theorem 3. When a line k is transversal to two parallel lines I and m, the

alternate interior angles formed are equal.

In Figure 1, with I parallel to m, this means that f3 equals LBAC.

PROOF. In Figure 1 we first show that the angles a and f3 on the same side
of k have sum 180·. If not, take y with f3 + y = 180·, I' as in Figure 3,

/
I AI

~rT--r----
./' ex
./"""~ j - - - - -

- .......... "

- -------------~
~ ~ ..--- ........ / >,
p'C ./' m(J
_ _--
IE
I

Figure 2
3. The Parallel Axiom 69

i'
------p

Figure 3

so that I' is parallel to m by Lemma l. But the parallel axiom asserts

there is only one parallel to m through A. Therefore I = I', so that a = y
and a + f3 = 180°. The requisite equality of alternate interior angles fol-
lows.
(In describing the proof, we have really used the figure. The reader
should convince himself that this prop is not necessary.)

Corollary. The sum of the three interior angles of a triangle is 180°.

PROOF. Through a vertex A of the triangle construct a line I parallel to the
base of the triangle. Then three angles at A on one side of I add up to
180°; by alternate interior angles, these equal the three angles of the trian-
gle (Figure 4).
This result is characteristic for Euclidean geometry; in fact, the parallel
axiom itself can be proved from the assumption that the sum of the angles
in a triangle is always 180°. From it one can also prove that the sum of
the interior angles of any quadrilateral is 360°. Similarly, the sum of the
interior angles of an n-sided polygon in the plane must be (n - 2) 180°-
though the proof takes some care where the polygon is not convex. But
the basic fact, illustrated by the various corollaries, is that about the sum
of the interior angles of a triangle.
There are other (equivalent) formulations of the parallel axiom. Thus
for Euclid the axiom stated that for a transversal k as in Figure 1, if the
sum a + {3 of the interior angles on one side of k is less than 180°, then
the lines I and m meet on that side of k.

A
--------- - - -I- - - - -

BL--------------------~C

Figure 4
70 III. Geometry

4. Hyperbolic Geometry
A set of axioms is said to be an independent set if no one of these axioms
can be deduced from the others. It is desirable and appropriate (though
not necessary) that the axioms for a basic structure, such as that of the
Euclidean plane, be independent. In particular, there is the question: Is
the parallel axiom independent, or can it be deduced from the others?
This question has had considerable historical importance. For example,
one might try to prove the parallel axiom by assuming the contrary (more
than one parallel to m through a point A) and deducing a contradiction.
There were several attempts to do this, most notably one in which Sac-
cheri in 1733 deduced a large number of consequences, some of them
perhaps bizarre-but none a contradiction. Nevertheless, he concluded
that Euclid's parallel postulate was "vindicated". Then in the 19th century
Bolyai, Lobachevsky, and Gauss took the opposite view, preparing to
develop systematically a non-Euclidean geometry (specifically, a hyper-
bolic geometry) on the assumption that there is more than one parallel,
and hence that the angle sum in a triangle is not 180°. When this is done
systematically, it turns out that the angle sum is always less than 180° and
that the difference between 180° and the sum is proportional to the area
of the triangle.
This striking development raised (at least) two questions: "Is the result-
ing geometry consistent?", and "Does it fit the real world?" To answer the
latter question, one must propose a specific "real world" interpretation of
the primitive concepts of the geometry-say, by taking a straight line to be
the path (in vacuum) of a ray of light, while an angle is the thing meas-
ured by a surveyor "turning off' with a transit the angle between two rays
of light. With this interpretation, it appears that Gauss (who was also
active as an astronomer) measured the angle sum for the triangle formed
by chosen "points" on the peaks of three convenient mountains in Ger-
many; the resulting angle sum was 180°, within the accuracy of the mea-
surements then made. While the result indicates that there is not a
flagrant deviation from Euclidean geometry on this interpretation, it does
not provide any clear decision between the reality of Euclidean and
hyperbolic geometry. It even suggests that there might never be such a
decision, in view of the inevitable margin of error in the measurements
made in any such interpretation. The terms involved in the interpretation
are also open to question; for example, in general relativity theory the
path of a light ray may not be "straight" in the intended sense. This ulti-
mately brings up another and more profitable thought: Any geometrical
axioms, Euclidean or non-Euclidean, offer a mathematical structure which
may be open to a variety of different interpretations to suit a variety of
geometrical (or even non-geometrical) circumstances.
There remains the question of the consistency of the assumptions of
hyperbolic geometry. By definition, these assumptions are consistent if
4. Hyperbolic Geometry 71

they lead to no contradiction-but it is clearly not enough to just observe

that they haven't yet led to a contradiction. One way to approach this
consistency would be to provide within Euclidean geometry an interpreta-
tion of hyperbolic geometry, say, by proposing certain Euclidean objects
to serve as the "pseudo-points", the "pseudo-lines", and the "pseudo-
distance" of a hyperbolic geometry, and proving from the Euclidean
axioms that these "pseudo" objects do satisfy the axioms for point, line,
and distance in hyperbolic geometry.
This can be done. Let C be a circle in the Euclidean plane. Let all
Euclidean points within C count as pseudo-points, while diameters of C
and the arcs of Euclidean circles which are orthogonal to C count as
pseudo-lines, and a pseudo-point is on a pseudo-line in the evident
(usual) sense of a point on a circle. It is then true that two distinct
pseudo-points A and B do determine a unique pseudo-line. For either
A and B lie on a diameter, or among the circles passing through A and B
there is (by a continuity argument) exactly one which is orthogonal to the
circle C. The axioms of incidence in Group I then hold, while with the
usual Euclidean relation of betweeness the axioms of Group II also hold.
The notion of pseudo-congruence can then be defined in a standard
way by first introducing a "pseudo-distance". Given pseudo-points A, B
on a pseudo-line as in Figure 1, let Sand T be the Euclidean points in
which the (orthogonal) circle meets the base circle C, and let AT, BT, etc.
denote the Euclidean lengths of corresponding arcs on this circle. Now
define a (pseudo) distance by

.
dIS(AB) AS /
= lo~ [AT BT
BS' 1 (1)

and observe that this means that as B approaches T along the pseudo-line,
with A fixed, the pseudo-distance AB will approach infinity. Then with
this "metric" (i.e. in this metric space), the length of the whole pseudo-
line will be infinite, as one might wish. All the congruence axioms for seg-
ments thus hold. The angle between two pseudo-lines at a point of inter-

Figure 1
72 III. Geometry

section is then taken to be the Euclidean angle between the Euclidean cir-
cles. One can then define congruent angles and layoff such angles as
required in the axiom. Moreover, drawing a triangle formed by three
pseudo-lines will readily show that the angle sum in such a (pseudo) trian-
gle will turn out to be less than 180·. All the axioms for hyperbolic
geometry hold in the interpretation. In particular, Figure 2 suggests that
through a pseudo-point A not on a pseudo-line m there will be many
pseudo-lines not meeting m, and hence parallel to m.
This Euclidean interpretation of hyperbolic geometry proves that hyper-
geometry are multiple. On the one hand, this geometry arises from mea-
surements: Do the angles in a triangle of light rays really add up to 180·,
This contradiction could then hold for the pseudo-points and the pseudo-
lines of this model, and so would apply to those points and circles of
Euclidean geometry-it would then be a contradiction in Euclidean
geometry. In this way, the interpretation provides a proof of relative con-
sistency. We will return later (Chapter XI) to the deeper question of abso-
lute consistency.
The intellectual origins of the structure of a hyperbolic non-Euclidean
geometry are multiple. On the one hand, this geometry arises from meas-
urements: Do the angles in a triangle of light rays really add up to 180·,
or to something else? On the other hand, this geometry also arises from a
formal study of the axioms for geometry. Is the parallel axiom necessary,
or can it be deduced from the other axioms? If not, what does this say
about geometries? The result is clearly the conclusion that there can be
different forms of geometrical theories. In the actual historical develop-
ment, this 19th century conclusion was a tremendous shock. The result
emphasizes the subtle nature of the primitive terms of a formal axiomatic
system: The straight line, assumed as a primitive term in the Hilbert
axiomatics, can indeed be variously interpreted-in particular, by objects,
such as our pseudo-lines, which are intuitively by no means straight.
The development of non-Euclidean geometry represents a major change
in the nature of Mathematics, from a science (of number and space) to a
study of form.

Figure 2
5. Elliptic Geometry 73

5. Elliptic Geometry
There is another non-Euclidean plane geometry, called elliptic geometry,
in which there are no parallel lines. The axioms needed to describe such a
structure must deviate from the parallel axiom and from other axioms of
Euclidean geometry, since the latter axioms by themselves (as in §3) sup-
pose the existence of some parallel lines. Instead of examining these
axioms, we will describe elliptic geometry by Euclidean models.
The proof of Lemma 3.1 already suggests what must happen in an
elliptic plane: Two perpendiculars at different points to a line k are likely
to meet on both sides of k (so something must break down in the
Euclidean description of the two "sides" of a line). More explicitly, if the
perpendiculars at two points A and B on k meet at some point P, then the
triangle APB has its base angles at A and B equal, hence is isosceles.
Then take another point C on k so that AB = BC (and B is between
A and C); then PB == PB, AB == BC, so the triangle PBA is congruent to
PBC, and then the corresponding sides PA and PC are equal (Figure 1).
Therefore PC is a third perpendicular to k, and all three of these perpen-
diculars to k (at points A, B, and C) meet at the single point P-and there
will be many more such perpendiculars from P to various points along k.
This configuration seems implausible in our usual plane, but it suggests
a different interpretation. Let a "point" be a Euclidean point on some
fixed sphere, while "line" is a great circle on that sphere. Then the equa-
tor is a "line", and so is the Greenwich meridian-and the various meridi-
ans from the north pole N do realize on the sphere the curious behavior
suggested in Figure 2. In this interpretation one may define congruence
in terms of distance (the length of arc along a great circle) and angle mea-
sure (the usual angle between two great circles). The result is a geometry
in which the appropriate axioms hold. In particular, it is clear from exam-
ples that the sum of the three interior angles in a triangle for this
geometry is always greater than l80 o-a result opposite to that in hyper-
bolic geometry. This model is sometimes called double elliptic plane
geometry-"double" because any two lines (two great circles) meet in two
points (diametrically opposite on the sphere).
p

Figure 1
74 III. Geometry

Figure 2. Elliptic geometry.

The axiom that two "lines" meet in a "point": can be rescued by con-
structing a single elliptic interpretation: Identify diametrically opposite
points in the previous model. Thus a "point" of this new model is a pair
of diametrically opposite points on the fixed sphere, a "line" is a great cir-
cle on the sphere, the "distance" from A to B is the shorter of the great-
circular arcs from A to B or its diametrical opposite, and so on. Put
differently, the geometry of great circles on the sphere is a non-Euclidean
geometry.
This is a special case of the geometry of geodesics on a curved surface
S. A curve y on S is said to be a geodesic if, given two nearby points A
and B on y, the length of y from A to B is less than the length of any
other curve from A to B on the surface. (It is necessary to speak here of
"nearby" points; on the sphere, a great circle from the north pole N is no
longer a shortest distance when one travels along it beyond the south
pole.) Hyperbolic plane geometry can also be represented by such a
model. Take the tractrix, a curve in the xy plane given by the equations in
a parameter 0,

x = sin 0, y = cos 0 + log(tan 0/2), (I)

as in Figure 3. Rotate this curve about the y axis to form a surface. The
y

L-----------------~x

Figure 3
6. Geometric Magnitude 75

geodesic curves on this surface form a model of hyperbolic geometry. The

surface itself turns out to have constant but negative curvature, in the
same sense that the sphere has the same (constant but positive) curvature
at all of its points. In this and related ways, plane geometry evolves to
geometry on a curved surface (see Chapter VIII).

6. Geometric Magnitude
Hilbert's (and Euclid's) axioms for Euclidean geometry have been
expressed in terms of congruence, and not in terms of distance, as it is
usually measured by a real number. In other words, the geometric
approach has been formulated independently of any use of real numbers
(but does employ natural numbers, to state the Archimedean axiom). In
fact, the geometric approach can be used to give a geometric description
of magnitudes (i.e., of real numbers). We now sketch briefly how this
might be done.
Fix on a line k and choose on it a point 0 (as origin) and another point
U, calling the segment OU the unit magnitude. Then by a magnitude one
means any segment OD from the origin on this line k; one of the two rays
from 0 on k is chosen to be positive, and the segments OD on that ray are
called positive magnitudes (Figure I). These magnitudes are ordered;
specifically, any negative magnitude is by definition less than any positive
magnitude; again, given two positive magnitudes OD / and OD, the first is
less than the second when D / lies between 0 and D on k. Integral magni-
tudes (OU2,OU3, •.. ) may be obtained by simply laying off the same seg-
ment OU repeatedly along the ray au.
A segment AB anywhere in the plane is then measured by some (posi-
tive) magnitude: Simply layoff AB as a segment OD along the positive ray
on k from 0, and use the corresponding magnitude OD as the measure.
These magnitudes can be added (in the evident way) and multplied, by
using similar triangles (see §IV.3). A magnitude OD is rational if there are
integers m and n with m(OD) = n(OU). There are irrational magnitudes,
such as the hypotenuse of an isosceles right triangle with legs of unit
length. Any irrational magnitude can be approximated by rational ones.
Such a geometric theory of magnitude was present in Euclid's
geometry. The theory there avoided the choice of any unit of measure-
ment, and so operated wholly with ratios. Thus what we have called the
magnitude OD would appear in Euclid as a ratio OD IOU. Euclid then used
proportions, which are simply equalities of ratios; our approach would

o u D

Figure I
76 III. Geometry

also need such proportions to compare magnitudes for different units. A

crucial point in the Euclidean approach was a definition of such a propor-
tion. This definition, due to Eudoxus, uses integral multiples as follows:
An equality OD/OU = AE/AV holds if and only if, for all whole numbers
m and n,

mOD> n OU implies mAE> n AV,

and

mOD < n OU implies mAE < n AV.

In effect, this describes the ratio OD/OU in terms of the set of those frac-
tions n /m with n /m < OD IOU; similar ideas will appear in our descrip-
tion in Chapter IV of real numbers in terms of rationals. The essential
observation is that the axioms for plane geometry suffice to give a
geometric theory of real numbers.
Angular magnitudes may also be compared and measured. Here an
angle, as in the axioms, is an angle rs between two rays rand s emanating
from a common origin O. Once a unit angle has been chosen, it can be
bisected repeatedly to yield binary fractions of this unit and then, by
approximations, real number measures of all angles. Once the perimeter
27T of the unit circle has been determined, one may take as unit an angle
of one radian-that is, the angle at the center of a circle of radius I which
is sub tended by a circular arc of length 1. This assigns to a right angle the
familiar measure 7T 12, and gives measures ranging from 0 to 7T for all the
angles between two rays rand s. Angle measures greater than 7T apply
only to angles from a ray r to a ray s in (say) the counterclockwise
direction-and this uses the notion of an orientation, to be discussed
below in §8.
For Greek Mathematics, as this discussion indicates, magnitudes were
geometric rather than arithmetic; western Mathematics, as we will see in
Chapter IV, has reversed this emphasis. Oswald Spengler, in his book The
Decline of the West, has argued that this means that there are two wholly
different "Mathematics", as parts of two different cultures. Our position is
rather that congruence and geometric ratios on the one hand and Dede-
kind cuts on the other are just two different careful formalizations of the
same underlying idea of magnitude-and that this point exemplifies the
unity of idea behind the inevitably varied form.

7. Geometry by Motion
Geometry need not be static. Intuitively, it is concerned with the ways in
which "objects" can be moved around in an ambient "space". From this
7. Geometry by Motion 77

point of view space is really there just as a receptacle for motion. For
example, the congruence theorems for triangles in plane geometry can be
viewed as descriptions of conditions when one triangle can be moved so
as to coincide with another. The motions involved include translations,
rotations, and reflection. All are familiar from early practical activities; all
are examples of "rigid motions".
A rigid motion of the plane is a bijection A 4 A', B 4 B' on the points
of the plane such that each segment AB is congruent to its image A 'B'; in
other words, it is a transformation preserving distance (§I.5). From this it
follows that every angle is congruent to its image, for every angle LABC is
part of a triangle ABC which is moved to an image triangle A 'B'C' with
corresponding sides congruent-and hence, by a congruence theorem, with
corresponding angles congruent. Since a straight line is the shortest distance
between any two of its points, it also follows that a rigid motion must take
the points on a line I into points on some line I'; in other words, it moves I
to I'. Moreover, it preserves betweenness on I: If C is between A and B on I,
then the image point C' is between A ' and B' on I'. A rigid motion also
takes each side of I into one of the sides of I'; indeed, D and E lie on the
same side of I when the segment CD does not meet I; consequently the
images D' and E' lie on the same side of the image line I'. All told, a rigid
motion of the plane is an "automorphism": An isomorphism A 4 A "
I 4/' of the whole structure of the plane.
From the definition it follows at once that the composite of two rigid
motions is again rigid, and also that the inverse of any rigid motion is
itself a rigid motion. Hence the rigid motions of the plane form a group.

Theorem 1. Any rigid motion which leaves each of three non-collinear points
fixed is the identity.

Indeed, if the rigid motion M leaves two distinct points 0 and A fixed, it
must leave the whole line I through 0 and A fixed. Since a third point not
on the line is fixed, the motion carries each side of the line I into itself.
Since any other point B on one side of the line is determined by the dis-
tances OB and AB in the triangle OAB, this point B must also be fixed.
The several familiar kinds of rigid motions can be described and
analysed directly from the axioms, although they are often described by
equations in cartesian coordinates.
Intuitively, a translation is a motion which moves every point in the
same "direction" by the same amount. Formally, a translation is a motion
such that, if A 4 A', B 4 B', then AA' is parallel to BB' and A' B' is
parallel to AB. Given A 4 A' and B, these conditions fix B' as the fourth
vertex of the parallelogram with sides AA ' and AB, as in Figure 1. It fol-
lows that, given points A and A " there is a unique translation T which
moves A to A '. If C is a third point in the plane which is similarly
translated to a point C', consideration of the congruent triangles
78 III. Geometry

c -----------
r
/
/
/
/
A~
Y' _ _ _ _ _ _ _ _ _ _ _ ~

Figure 1

A CB == A 'C' B' yields the congruence BC == B' C'. This last congruence
means that the translation T sending B to B' and C to C' does preserve
distance. Hence T is a rigid motion, according to our definition of such
motions.
The identity is a translation, the inverse of a translation is a translation,
and the composite of two translations is again a translation. Hence all the
translations of the plane form a group under composition, called the
translation group H of the plane. It is an abelian group. If each translation
T is represented by a vector (from a chosen point 0, the origin, to the
image TO = 0' of that point), the translation group H is just the group of
these vectors under the operation called vector addition. Once 0 is chosen,
the possible translations of the plane are determined by the points 0' of
the plane; that is, by the vectors 00' in the plane.
The intuitive idea of a rotation is direct: Take a circular disc, fix the
center 0 and spin the disc. This idea leads to a formal definition: A rota-
tion R of the plane is a rigid motion which leaves exactly one point fixed;
it is called a rotation "about" that point. For completeness, the identity
transformation is also counted as a rotation (about every point in the
plane). From this definition it follows that the inverse of a rotation is a
rotation-about the same point O. All the rotations about a point 0 do
form a group, but to show this one needs

Theorem 2. If the distinct segments OA and OA ' are congruent, with 0 =1= A,
there is exactly one rotation R about 0 taking A to A '.

To construct such an R, first assume that 0, A and A ' are not collinear.
We need to construct the image X' of each point X =1= 0, A. Since the
desired rotation must preserve angles, the angle X 'OA' must be chosen
equal to the given angle XOA; this is surely possible, since the congruence
axiom 111.4 of §2 specifies that a given angle between rays can be "laid
off' on a given side of another ray. It remains to choose which side of
OA '. We know in general that all the points on one side of OA must go
into points on one side of OA '; after inspecting Figure 2 we propose the
rule (which could be formulated without "looking" at a figure) that the
side of OA containing A' rotate to the side of OA' not containing A. In
more detail:
7. Geometry by Motion 79

y'

A'

x'
x
A

Figure 2

(i) If X and A' are on the same side of OA, put the image point X', with
OX == OX', on the side of OA ' opposite A;
(ii) If Y and A' are on opposite sides of OA, put the image point Y', with
oY == 0 Y', on the side of OA ' containing A.
This prescription covers all cases. From the figure, it then appears that
LXO Y == LX '0 Y'; from the axioms of congruence one can prove that this
is always the case, so that the rotation X I-> X', Y I-> Y' does carry each
angle at 0 into a congruent angle and each segment from 0 into a
congruent segment. From this it then follows by the congruence axioms
that this rotation R carries every segment XY into a congruent segment
X' Y'. Therefore R is a rigid motion leaving only 0 fixed, as desired.
It remains to prove that this transformation R is the only rotation about
o taking A to A '. But the only alternative would be to make the opposite
choices in rules (i) and (ii) above; in this case one sees that the ray OB
bisecting the angle AOA' (see Figure 2) must go into itself, so that B == B'
would be an added fixed point, contrary to the definition of a rotation.
(Incidentally, the opposite choice of the rules (i) and (ii) would construct a
different rigid motion; namely, the motion reflecting the whole plane in
the angle bisector OB.)
In the excluded case when A and A' are on the same line through 0,
with A =1= A " the image of each point X may be constructed by prolong-
ing XO over 0 to the point X' opposite X, with OX == OX'. This motion
X I-> X' is then a rotation about 0 by the angle 'fT, sometimes called a
"half-tum". In each case, the rotation R constructed is the only rotation
with A I-> A '. From this uniqueness it follows that the composite of two
rotations about 0, if not the identity, leaves only 0 fixed, hence is a rota-
tion. Therefore all these rotations about 0 form a group.
It is remarkable that the simple idea "Hold 0 fixed and swing A to A '
leads to such a subtle proof, using sides of lines. This subtlety reflects the
80 III. Geometry

difference between the Euclidean study of figures and the more

"dynamic" idea of moving all the points in the plane.
Note however that rotation R, as constructed, simply specifies where
each point eventually ends up (A to A', X to X') and not how the point
"moves" there. In other words, this rotation R is not a motion Rt
parametrized over time t, in the sense described in §I.5. Thus R is neither
clockwise nor counterclockwise from A to A '! On the contrary, we will use
the properties of rotations given in Theorem 2 to explain, in §8, the choice
of a "counterclockwise" sense.
Next we combine rotations and translations. Since a rigid motion which
leaves fixed three non-collinear points must be the identity (Theorem I),
two rigid motions which have the same effect on three non-collinear
points must be equal. This observation makes it possible to compute
several useful composites. Thus if R is a rotation about the point 0, while
T is a translation taking 0 to some point 0', then the composite

(I)

is the translation TJ taking 0 to the point RO'. This can be proved for-
mally from the axioms, or checked visually by showing that the compos-
ites RT and TJ R have the same effect on the points of the right angle
AOB in Figure 3. Similarly (Figure 3 again), if T translates 0 to 0' and A
to A " while R rotates OA to OA", then the composite

(2)

is the rotation about 0' taking 0 'A' to 0 '(TA "). Recall here that TRT- J is
called a conjugate of R, so this equation states that translation T conju-
gates a rotation about 0 into a rotation (through the same angle) about 0 '.
Similarly, equation (I) states that the conjugate of a translations by a rota-
tion is another translation.
These equations enable us to prove a familiar fact:

Theorem 3. Composites of rotations and translations of th' plane form a

group. If 0 is any chosen point, an element of the group can be written
uniquely as a product T· R with T a translation and R a rotation ahout O.
This group is called the group Eo of proper rigid motions.

B"
A"
RO'
B B' B B'
B'~A'"
\~
\
\\
\~---1--+-_ _ _-!:-:-_-I A'
A o A 0' A'

Figure 3
 7. Geometry by Motion 81

PROOF. Consider the collection of all composites of a finite number of

rotations and translations; it is a group, since the inverse of a translation
or a rotation is again such. By equation (2), any rotation R I about a point
0' may be replaced by translations and a rotation R about O. By equation
(l), any composite R· T may be replaced by a composite T I • R for some
translation T 1 • Hence all the composites can be rewritten in the desired
form T·R. This form is unique, for take any point A =1= O. Then T·R takes
the segment OA to a congruent segment O'A', so (Figure 4), T must be
the translation taking 0 to 0' and R the rotation taking OA to OA 1 where
Al == T-1(A').
This last argument also proves the following familiar fact:

Corollary. If OA and 0'A' are congruent segments, with 0 =1= A, there is a

unique proper rigid motion taking 0 to 0' and A to A '.

In other words, a proper rigid motion is determined by its effect on any

two distinct point 0 and A.
The reflection L in a line I is also a rigid motion. Under this reflection,
each point A is sent to the point A' with I a perpendicular bisector of
AA '. Thus (Figure 5) A' is on the opposite side of I (and at the same dis-
tance from I). Again, AB is congruent to A'B', so reflection is a rigid
motion by our definition. It has an evident physical meaning (reflect, as in
a mirror at I). It differs from translation and rotation in that it "turns the
plane over".

o,L
Figure 4

B
I
A
I I
I II
I
I I
I
I I
I A'
I
B'

Figure 5
82 III. Geometry

Theorem 4. Let I be any line in the plane. Any rigid motion of the plane is
either a proper rigid motion P or a composite P·L where L is the reflection
in the line I.
PROOF. The motion is determined by what happens to each of three non-
collinear points; that is, by what happens to a triangle. (Hence the impor-
tance of triangles in Euclidean geometry.) So consider the triangle OAB
with side OA along the given line I and its image 0'A 'B' under the motion
(Figure 6). Since OA == 0 'A', there is a proper motion (rotation R fol-
lowed by translation 1) which takes OA to 0'A '; one can then place the
image B' on the desired side of I, by using first a reflection L in I, if
necessary. Hence the motion has a unique representation as either T·R or
T· R· L, with R a rotation about O.
These theorems serve to show that groups (transformations and their
composites) are firmly embedded (though not explicit) in classical
Euclidean geometry.

I'
0'
B I
I

o
~A I
I" ................ " .......
..... ......: A'

Figure 6

8. Orientation
Hitherto "angle" has meant the angle between two rays rand s from a
point O. For trigonometry and elsewhere we also need to use the angle
from the ray r to the ray s. But this is ambiguous (Figure 1) unless we
specify that we mean the "counterclockwise" angle from r to s. But this is

Figure I. Clockwise and counterclockwise.

8. Orientation 83

meaningless for the plane as it has been axiomatized, because the counter-
clockwise sense, viewed from in front of the plane, will be clockwise when
viewed from behind the plane. Thus we must make a deliberate choice of
one of the two possible "senses" for angles. Such a choice is called on
orientation of the plane, and the choice of an orientation is an additional
structure on the Euclidean plane.
Practically speaking, someone (the first clockmaker?) chooses a direc-
tion of rotation for his clock; then one can carry this clock (or copies, such
as wrist watches) around so as to determine the same clockwise sense for
all other clocks. Here to "carry around" means a proper rigid motion; this
idea we can now formalize (in the plane).
To fix the ideas, consider, instead of clock-hands, ordered right angles
from AO to EO, with legs say of unit length, as in Figure 2. Here
LAOB and LA '0' B' have the same sense, while LA "0" B" has the oppo-
site sense. Formally, two such ordered right angles have the same sense if
and only if the first can be carried into the second by a proper rigid
motion (a rotation followed by a translation). Since there is exactly one
such proper motion carrying OA into the congruent segment 0'A " there
are exactly two possible senses, as given by the opposite perpendiculars
0' B' and 0' B'" in the middle of Figure 2. Put differently, the ordered
right angles with unit legs fall into two distinct classes (or "orbits") under
the action of the group of all proper rigid motions. Choosing one of these
classes is the choice of a "sense" or an orientation of the plane; for
instance, the chosen sense may be called the counterclockwise one. The
choice is usually made by picking some one right angle, from OA to OB, to
be counterclockwise.
Once the orientation is made by this choice of LAOB, one can introduce
the usual four "quadrants" I-IV for this right angle and then cartesian
coordinates, since we now have a "positive" direction OA or OB on each
axis of Figure 3. We can also consider and measure ordered angles (say,
the angle from the ray OA to the ray Ot in Figure 3) on the familiar radian
scale from 0 to 2'17, so that these angles are real numbers taken modulo 2'17.
We can also speak of the "left" side of the ray OA; it comprises all the

B'

0'

o
A'

B"

Figure 2
84 III. Geometry

II
-+.---'----- A

III IV

Figure 3

rays from 0 in quadrants I and II. We will write sLr for rays rand s from
oto say that s is on the left side of r.
A plane with orientation is really not the same object as one without.
The plane with an orientation has more structure-namely, the choice of
the orientation. At the same time, it has less symmetry; the automorphism
group of the oriented plane is the group of all proper rigid motions (i.e.,
no reflections), while that of the unoriented plane is the group of all rigid
motions, including the reflections. This is a first example of a striking
observation: A geometrical "thing" (the plane) can be formalized in
different ways to give different mathematical objects. In the oriented
plane, one has angles up to 2," and coordinates as well-but these are not
present in the unoriented plane described by the Euclidean axioms.
The added structure of the oriented plane has been described as a
"choice of an orientation". It can be formulated in other ways, so as to
resemble an "axiomatic" structure. For example, consider the relation sLr
for "s is on the left side of r", where rand s are rays from a common
point O. If -r denotes the ray opposite r, but along the same line through
0, this relation has the following properties:

(i) sLr if and only if ( - r)Ls,

(ii) sLr if and only if not (-s)Lr,
(iii) If a translation carries the rays rand s at a point 0 to rays r' and s'
at a point 0', then rLs implies r'Ls'.
These three statements can be regarded as the added axioms describing
an oriented plane. From them one can prove directly, without using rota-
tion, that a Euclidean plane has exactly two orientations. On this basis
one can then define a rotation about the point 0 to be a rigid motion leav-
ing 0 fixed and mapping rays r, s from 0 to rays r', s' in such a way that
rLs implies r'Ls'. One can then easily prove that there is exactly one such
rotation carrying a segment OA into a specified congruent segment OA' (as
in Theorem 2 of §7).
These related notions "rotation" and "orientation" are known to be
subtle-and are often left to the intuition. Our discussion, both in
9. Groups in Geometry 85

Theorem 2 of §7 and in the formulation of the relation L, does not follow

any standard source in the literature. It is intended to show how the very
natural "idea" of rotation requires a subtle formulation. Still, the idea is a
familiar one: In three dimensions, one cannot rotate a left-hand glove into
a right-hand one!

9. Groups in Geometry
Rigid motions not only carry points into points, but they also carry seg-
ments into (congruent) segments, angles into (congruent) angles, and so
on; one says that the group E of all rigid motions "acts" on the set of seg-
ments or on the set of angles, and that the "orbit" of an angle is the set of
all congruent angles. This example motivates a general definition. A group
G is said to act on a set X when for each g in G and each x in X there is
given an element gx in X (the result of g acting on x) in such a way that

Ix =x (1)

for all gl and g2 in G, for the identity I of G, and for all x in X. The orbit
of an element x in X under this action is the set Gx of all gx for g in G,
and the action is transitive when Gx = X for all x; that is, when to each
x and y in X there is at least one g with gx = y. In any case, the ele-
ments fin G which leave a given point x fixed form a subgroup Fx of the
group G, called the subgroup fiXing x or the isotropy group of x.
For example, when the group Eo of proper rigid motions acts on the
points of the plane, the isotropy group Fx of each point x is the group of
rotations about x. Any two such groups Fx and Fy are isomorphic. All the
proper motions carrying x to y can be written as products TR for R in Fx
and T any motion (say a translation) carrying x to y. All these motions
constitute a so-called "coset"

TFx = {all TR IRE Fx} (2)

of Fx. The whole group Eo consists of the various disjoint cosets

TFx, T' Fx corresponding to the various points Tx = y, T'x = z, ... of
the plane.
These observations hold in general. For any group G acting on a set
X, hx = kx for hand k in G if and only if k - Jh is in the isotropy group
Fx. Thus all the elements k in the coset

hFx = {all hf I f E Fx} (3)

map x to the same point on the orbit, so that the set of cosets corresponds
to the set of points in the orbit, as in the bijection
86 III. Geometry

r r r r
Fx hFx IFx mFx '" .

(4)
x hx Ix mx, ..

This is the (geometric) origin of the notion of a "coset". If points x and y

are in the same orbit, so that hx = y for some h, then Fy = hFxh -I; this
states that the isotropy subgroups Fx and Fy are conjugate subgroups of
G, isomorphic under the bijection g I-> hgh -I. This is the geometric ori-
gin of the notion of "conjugation".
The group E of all rigid motions acts on the set X of all segments AB.
Here the orbit of each segment is the set of all congruent segments, while
the subgroup fixing a segment AB consist of four motions, {I,L,R,L·R},
where L is the reflection in the line of AB and R is the half turn (rotation
by 180) about the midpoint of AB.
Similarly, the reader may examine the associated action of the group E
on the set of triangles, on the set of lines, on the set of rays, and on the set
of circles in the plane.
Consider the subgroups of E:

E =:) Eo =:) H =:) {l}, (5)

where Eo is the group of proper rigid motions and H the group of transla-
tions. The plane has just two orientations, call them i and t . The
whole group E acts on the set of these orientations, in that each rigid
motion M induce a permutation a(M) of the set {i , t }; thus when
M E Eo is a proper rigid motion, the orientation is unchanged, so a( M)
is the identity permutation, while a reflection L interchanges the two
orientations. For the composite M·M' of two motions the effect is

a(M·M') = a(M)a(M'); (6)

this equation states that a preserves composites, hence is a homomorphism

of groups. This homomorphism carries the subgroup Eo to the identity
permutation and its coset EoL (all products ML, for M in Eo) to the per-
mutation interchanging i and t '
A direction (e.g., the "direction" of a ray ro) in the plane may be
described formally as the orbit of the ray ro under the translation group.
Every rigid motion M carries directions to directions, so defines an action
of E on the set of directions. Let a(M) denote the permutation of the set
of directions induced by the rigid motion M; thus a is a homomorphism,
as in (6), while the set of all motions M with a(M) the identity is the sub-
group H of translations.
In the terminology of group theory, this subgroup H is a normal sub-
group of E because every conjugate MTM- 1 of a translation is a transla-
tion. We showed that every proper motion in Eo had the form T· R, with
10. Geometry by Groups 87

T a translation and R a rotation about a fixed point 0, so that Eo is the

union of the cosets of H

HR = {all TR IT E H}.

The cosets correspond to the rotations R, so that the cosets may be said to
form a group Eo IH isomorphic to the group of rotations about O.
These observations, and many like them, indicate the very close relation
between Euclidean geometry and group theory-so close that one might
say that groups were implicit (though never explicit) in traditional
geometry. For these reasons, it is clear that the basic ideas of group theory
belong early in the conceptual order of mathematical structures. Histori-
cally, groups did not appear until the 19th century, implicitly with Gauss
and others and explicitly with Galois. When they were fully recognized,
they were applied promptly to geometry, by Klein, Lie, and others. In
particular, hyperbolic and elliptic geometries also involve appropriate
groups of motions-as does solid geometry.

10. Geometry by Groups

Space is not just a static array of geometric figures; it is primarily a field
for motion. This can be stated more emphatically in the observation that
the axioms for the Euclidean plane can be formulated wholly in terms of
the group E of rigid motions.
We will sketch briefly how this can be done. In any group, an element h
of order 2 (an element h =!= I with h 2 = I) is called an involution. In the
group E of all rigid motions the reflection Lk in the line k and the rota-
tion (half-tum) RA by 180 about the point A are both involutions.
0

Theorem 1. A rigid motion M of the plane which is an involution is either a

reflection Lk in some line k or a half-turn RA about some point A.
PROOF. Since M2 = I but M is not the identity, there is a point B with
MB = C =!= B. Then MC = M2B = B, so the motion M interchanges the
two points Band C. Let I be the line which is the perpendicular bisector
of the segment BC, as in Figure 1. The points on I are the points equidis-
tant from Band C; hence the given involution M must carry points on I
to points on I (it thus leaves line I as a whole fixed). Moreover, M must
leave fixed the midpoint 0 of the segment BC; 0 lies on I. Take a point P
on I but different from O. Then M must send P to another point on I at
the same distance from Band C-thus either to P itself or to the point pi
on the opposite side of 0 with PO = OP'. The reflection L[ sent P to P;
the half-tum Ro sends P to P'. Since a rigid motion is determined by
what it does to three non-collinear points, this proves that the given invo-
lution M is either L[ or Ro.
88 III. Geometry

.
B

p o p'
.
C

Figure 1

Theorem 2. Every rigid motion is a composite of involutions.

PROOF. It will suffice to show that every rotation and every translation is a
composite of involutions. To this end, consider the composite of two
reflections L k • L[ in distinct lines k and I. If k and I are not parallel, they
meet in a point 0, say at an angle $ The composite motion Lk L[ leaves
0 0

(Figure 2) the point 0 fixed and preserves orientation, hence is a rotation

about o. Now L[ leaves I fixed, while Lk takes I to the line I' through 0
making with k the same angle $ as does I, but on the opposite side of k.
Therefore L k • L[ is the rotation about 0 through the angle 2$. Hence
every rotation is a composite of two reflections.
In the remaining case k and I are parallel, so have a common perpen-
dicular mo The composite Lk L[ then turns out to be translation along m
0

by twice the distance from k to I.

The objects of plane geometry can now be identified with elements in
the group E of rigid motions. Since each rotation RA by 180 is the com-
0

posite of two reflections, the points A may be identified with those involu-
tions which can be written as the composite of two involutions, while the
lines k are iderttified with those involutions (to wit, the L k ) which are not
such a composite. The point A lies on the line k if and if only if L k • RA =
RA ·Lk . If the motion M takes the point A to the point B, then the half-
tum RB is clearly MRA M- 1, so the motion M conjugates the involution
RA to the involution RBo One may similarly describe the effects of any
motion on a line, represented as an involution.
This start can be continued to give a complete list of axioms on the
group E sufficient to describe E as (generated by the involutions

Figure 2
II. Solid Geometry 89

representing) the points and lines of the Euclidean plane. Details are
given in Bachmann (1973) and in Guggenheimer (1967). The reader may
wish to formulate for himself some geometric facts in group-theoretic
form. For instance, AB = BC states that B is the midpoint of AC!

11. Solid Geometry

The geometry of "our" space is three dimensional, so is harder to visual-
ize on paper. As for the plane, experience reveals many facts about solid
figures. It again turns out that these facts are logically related to each
other and so can be derived from suitably chosen axioms. These axioms
are stated in terms of basic figures: points, lines, planes, spheres, and the
like. As formulated in Euclid or as presented in modern form in Hilbert
they involve little more than the axioms for plane geometry. The surpris-
ing fact is thus the observation that so much of geometry is already con-
tained in the geometry of the plane (but not yet in the geometry of the
line). For example, the Pythagorean theorem for plane right triangles
readily produces the distance formulas of analytic geometry, both in the
plane and in space. Again, angles necessary for plane geometry (and not
present in one line alone) are vital to trigonometry, periodic functions,
navigation by spherical trigonometry, and the group of the circle-but the
corresponding notion of "solid angle" is very little used. It was surely a
remarkable discovery (was it made by the Greeks?) that concentration on
the plane would yield so much of the ideas of geometry. This discovery
rested in turn on the prior formulation of the "idealized" notion of a
plane (or of aline), but this latter discovery is not so remarkable (it is
surely suggested by diagrams in the sand or on paper, by maps, etc).
The group of rigid motions of Euclidean solid geometry is a much
bigger and richer group than that of the plane-but it is still true that any
rigid motion of space is a composite of rotations, translations, and
reflections. For example, a rotation R is a rigid motion which leaves fixed
all the points on some line k; then k is called the axis of rotation. It is a
rotation through an angle (J, because each plane perpendicular to the axis
is rotated by R through the same angle (J (watch the orientations!). In this
sense a solid rotation is just a combination of plane rotations in parallel
planes.
All the rotations about a point 0 (i.e., all with axis through 0) form a
group. This is a consequence of one of the fascinating facts about such
solid rotations, the rule for the composition of two rotations about the
same point O. Take the sphere of radius 1 and center 0, and consider a
rotation R about the axis AO through an angle (J followed by a rotation S
about a different axis OB through an angle </>. To find the composite S·R,
draw the arc AB of a great circle joining A to B and then construct great
circular arcs AC and BC making angles (J12 and cp 12, respectively, with
90 III. Geometry

AB (Figure 1). This gives a "spherical triangle" ABC. One may show that
S·R leaves C fixed and hence is a rotation (through what angle?) about
CO. This curious fact about rotations is one of the results of solid
geometry much used in the study of rotations in mechanics. However,
even this apparently "solid" geometrical result has its predecessor for the
plane: There is a similar description of the composite of two rotations
(about different points) in the plane.
The orientation of 3-dimensional space also involves no essentially new
ideas. To describe it, consider three perpendicular rays x, y, z from a point
o (they are, in effect, coordinate axes). Then the ordered triple (x,y,z) is
defined to have the same orientation as a similar triple (x ',y ',z ') if and
only if there is a rotation leaving 0 fixed and carrying x to x', y to y', and
z to z '. This definition provides for exactly two orientations at 0 Gust as in
the plane). For, one can always rotate x to x' (about an axis perpendicu-
lar to the plane determined by x and x '). The new position of y can then
be rotated (about the axis x') to the position y' (Figure 2). The resulting
position of z must then be perpendicular to x' and y' and hence must
coincide either with z' or with its opposite-therefore there are just two
orientations at O. Note that the triple (x,y,z) has the same orientation as its
cyclic permutation (y,z,x)-just rotate x to y (about z) and then the new y
(about the old y) to z. In other words, the orientation of (x,y,z) is

Figure 1

y
y'

°v----x
z
z'
x'

Figure 2
12. Is Geometry a Science? 91

unchanged by a cyclic permutation (an even permutation). However, it is

changed by interchanging x and y to give (y,x,z), which is an odd permu-
tation. The orientation determined by the first three fingers, in order, of a
left hand is called the "left-handed" orientation.
This description of just two orientations extends readily to dimensions
higher than three. In general, an axiomatic treatment of n-dimensional
Euclidean geometry in terms of the appropriate "flat" subspaces (points,
lines, planes, hyperplanes ... ) is readily possible-and a bit dull.
In all these respects, three dimensions is much like two dimensions-but
there are important points of difference, such as the much greater variety
of figures in 3-space: Spheres, cylinders, pyramids, regular solids, irregular
solids, and the like. Decisive new notions, such as the Gaussian curvature
of a surface at a point, arise from the study of these figures, as we will see
in a later chapter. There we will also examine a wide variety of other
geometries.

12. Is Geometry a Science?

This chapter has summarized the elementary aspects of the remarkable
way in which primitive experiences of space and motion can be arranged
and refined in a precise deductive system organizing and relating
geometric facts and figures. At first blush, the resulting system of
Euclidean geometry can appear to be the science of space, just as mechan-
ics is the science of motion or biology is the science of life-forms. But if it
were to be the science of space, it should be the unique such; this is not
the case. We will not attempt here a precise description of what is meant
by a science. Geometry was "extracted" from experience, but it seems
clear that we do not test the theorems of geometry by experiments, as we
would the propositions of physics. Even if measurements were to show the
Pythagorean theorem false for some proposed right triangle, we would not
doubt the proofs of that theorem. We would be more likely to alter the
definitions used in those measurements of distance or of right angles, or
we would perhaps decide that those measurements of distance and of an
angle belonged to some different geometric theory. In the language of
Karl Popper, statements of a science should be falsifiable; those of
geometry are not.
We are more concerned with the positive aspects of the question: What,
then, is geometry? It is a sophisticated intellectual structure, rooted in
questions about the experience of motion, of construction, of shaping. It
leads to propositions and insights which form the necessary backdrop for
any science of motion or of engineering practices of construction. But at
the same time, the structure of geometry is not unique. The axioms can be
formulated as in the Euclidean tradition, as it was refined by Hilbert-or
they can be put in group-theoretic form. Moreover, non-Euclidean
92 III. Geometry

geometries of plane or of space are indeed possible and useful, and they
can be reduced to axiomatic form. Hence geometry is a variety of intellec-
tual structures, closely related to each other and to the original experi-
ences of space and motion.
There arise from this study other structures which are less geometric-
distance and angle as geometrical magnitudes, algebraic manipulations of
these magnitudes, and thus real numbers, developed geometrically. There
also arise structures which are not geometric at all-groups are implicitly
present in the transformations of geometry. Logic is (historically) first
fully exemplified in the deductive structure of Euclidean geometry. Con-
tinuity and topology are hidden there as well.
The development of geometry also turns up a number of general ideas.
The very formulation of axiomatic geometry requires ideas of line, plane,
angle, and triangle. Later developments tum up more subtle ideas, such as
that of orientation or of composition of motions. Geometry is indeed an
elaborate web of perceptions, deductions, figures, and ideas.
CHAPTER IV

Real Numbers

1. Measures of Magnitude
This chapter will explore the origins and development of the system of
real numbers. These numbers form the most central structure in all of
Mathematics. Like other structures it arises from more primitive human
activities-in this case, from the measurement and comparison of magni-
tudes of various kinds.
Comparison of magnitudes, in its simplest qualitative form, may just
assert that "A is bigger than B" without specifying by how much A is
bigger. The idea of such qllalitative comparisons leads to the formal
notion of a linear order, as already discussed in §I.4. Now we are con-
cerned with the associated quantitative question "A is how much bigger
than B?". Such questions arise in a number of different regards "A is how
much further away" or "how much heavier", or "how much longer", or
"how much later". Such comparisons are not limited to just two objects
A and B, but may well intercompare the sizes of three, four, or many
objects. This makes it effective not just to compare two objects, but to
locate all the relevant objects on some on scale of sizes. Once a unit of
size is chosen, the scale becomes a scale of numbers. It is a familiar but
nonetheless remarkable fact that one single scale of numbers will be
applicable to each of many types of quantitative comparisons: To dis-
tance, to weight, to length, to width, to temperature, to time, to height,
and so on. Once a unit is chosen, each of these magnitudes exemplifies
one and the same scale: That of real numbers, considered as a scale laid
out as the points of a line with chosen origin and unit point; and so
emphasizing the interpretation of the scale by distances:

I I
-2 -1 o 2 3

The ubiquity of this scale may account in part for the prominence of the
real numbers in Mathematics. This ubiquity can also be read as a state-
94 IV. Real Numbers

ment about the nature of the world: All sorts of physical magnitudes can
be reduced to one scale-a situation which has sometimes misled social
scientists to use fake magnitudes.
The importance of this scale depends also on the fact that the scale is
complete. It is not limited just to the whole number points and not even
just to the rational points. All the points at irrational distances from 0 are
to be included, as, for example, Y2. By the Pythagorean theorem, this
number, considered as a length, is the diagonal of a unit square; it cannot
be expressed as a rational distance as Y2 = min. (Proof: reduce to
lowest terms, then one of m or n is odd, but Cv1)2 = 2 = m 2 1n 2 gives
2n 2 = m 2, so m and then also n must be even.) There are many other
such irrationals which are algebraic (roots of a polynomial equation with
integer coefficients, as in x 2 - 2 = 0). In addition, there are numbers on
the scale which are transcendental (i.e., not algebraic). The first example is
77, the ratio of circumference to diameter in any circle, but the proof that
77 is transcendental is not easy. Another transcendental number is the base
e of natural logarithms.
There are many more transcendentals. Completeness (to be formulated
below) is the assurance that they are all there. They are not all in use at
anyone time-but their presence for potential use is what makes the scale
effective.

2. Magnitudes as a Geometric Measure

The basic scale of magnitude can be regarded as either an arithmetic or as
a geometric structure. The arithmetic approach starts just with the scale of
whole numbers (positive, zero, and negative), then adjoins the magnitudes
represented by the rational numbers, and finally those represented by irra-
tional (algebraic and transcendental) numbers. This strictly arithmetic
view of magnitude will be developed below, in §6.
The scale of magnitudes may also be regarded as primarily geometric.
Then the basic model of a magnitude is a line segment AB (or, if you will,
the length of that line segment). Then, as discussed in Chapter III on
geometry, two line segments AB and CD represent the same magnitude
when they are congruent, so that CD can be "moved" to exactly cover AB.
Thus a magnitude is not really a single segment AB, but all the segments
CD which are congruent to AB-in set-theoretic terms, the magnitude AB
is the (congruence) class of all segments CD congruent to AB. Also the
magnitude AB can be compared with any other magnitude RS by "laying
off" the segment AB along the ray of the segment RS, with A at the posi-
tion of R; the congruence axioms of plane geometry provide precisely for
such a comparison. It follows that the points on a single line represent all
the possible magnitudes of segments so that the line (with zero and unit
point chosen) is the desired scale of magnitudes.
2. Magnitudes as a Geometric Measure 95

Angles present another type of geometric magnitude. As in Chapter III,

the axioms of plane geometry provide for the congruence of angles and
the comparison of angular magnitudes. The reduction of these angular
magnitudes to linear magnitudes (i.e., to the real number scale) involves a
subtle development. First one measures the circumference of a circle, say
a circle S 1 of radius 1. This can be done by inscribing regular polygons in
the circle, starting say with an equilateral triangle and using successive
bisections (by ruler and compass) to get inscribed regular polygons of 6,
12, 24, ... sides, as suggested in Figure 1. The (rectilinear) perimeter of
each such polygon can then be calculated; as the number of sides
approaches infinity, these perimeters approach a definite limit, the real
number 2'TT.
Now the radian measure of an angle () at the center of the circle is to be
the length of the arc which () subtends on the circumference. Also, bisect-
ing the angle halves the arc length, so this gives the usual radian measures
'TT for a straight angle and 'TT /2 for a right angle. Successive bisections
(again by straightedge and compass) then provide angles with radian
measures 'TT /2 n . Any angle can be approximated, as closely as desired, by
multiples of such angles: Any angle () = LA CP at the center C of a circle
of radius I has measure to (Figure 2) described as follows:

measure(()) = to, o < to < 2'TT, to = length of arc AP. (l)

Any real number t can be written modulo 2'TT as t = to + 2'TTk for some
integer k, and so determines the angle () = ()t with measure ()t = to. The
function t ~ ()t then "wraps" the whole real line around the circle S 1,
sending t to the point P; i.e., to the angle ()t = LA CPo Moreover this
wrapping function is periodic:

(2)

Figure 1
96 IV. Real Numbers

~ _ _-.,7r/2

Figure 2

Alternatively, the methods of calculus could be used to determine circular

arc lengths and hence radian measure and the wrapping function 0/. In
either way, this function plays a central role in analytic trigonometry (§9).
The development of Euclidean geometry from its foundations in the
axioms also provides a treatment of the magnitudes of areas (of triangles,
rectangles, circles, etc.) and a comparison of these magnitudes with those
measured along the real line. This is carried out in Hilbert's Foundations
of Geometry. Thus one might say that the axiomatic foundations of
geometry must provide for the comparison and reduction to a single scale
of the various types of geometric magnitudes-including also volumes.
The same must be done in any completed system of non-Euclidean
geometry.
Measurements can also reduce other types of physical magnitudes to a
geometric form. Thus weights on a balance are compared with numbers
(the number of weights needed to achieve balance), while the weights
measured on other instruments give the magnitude of a weight by a
pointer position along some geometric scale. A thermometer does the
same for the magnitudes of temperature, while a clock hand reduces time
to angular magnitudes.
To summarize: Many comparisons or measurements of various magni-
tudes can be reduced to the single scale provided by the real numbers.
And in those cases where a single real number does not suffice to measure
a magnitude, it is often fitting to use several such numbers-as when the
size of a plane figure is given by its width and its height, or of a solid
figure by width, height, and depth. In other words, that scale of real
numbers has many different realizations.
3. Manipulations of Magnitudes 97

3. Manipulations of Magnitudes
When two objects are put together, side by side or end to end, the magni-
tude of this combination is just the sum of the two separate magnitudes.
This operation, called addition, arises for all sorts of magnitudes-for dis-
tance, weight, time, height, area, and the like. Geometrically, the addition
of two segments consists in laying off one segment after the other along
the line. This geometric operation corresponds exactly to the arithmetic
operation of adding numbers.
A second operation on magnitudes, that of multiplication, is suggested
both by the multiplication of numbers and by geometric formulas; thus
the area of a rectangle is obtained by multiplying its base by its height. A
complete geometric description of the multiplication of segments requires
more than one line in the plane. Thus to multiply two positive linear mag-
nitudes x and y one may represent them on two intersecting linear scales:
Segment OA and then AB on the first line with measures I and x, respec-
tively, and then OA' with measure y on a second ray from O. Then
drawing (Figure I) BB' parallel to AA ' constructs similar triangles OAA '
and OBB', while the proportionality theorem for similar triangles makes
OA 'lOA = A'B' lAB and hence shows that A 'B' represents the magni-
tude xy. This may be regarded as the geometric definition of multiplica-
tion of magnitudes.
Other types of magnitudes such as weight or time may more easily be
multiplied first by whole numbers-thus to multiply the weight of a given
item by three, take the combined weight of three such items. By division,
this extends first to multiplication by rational numbers and then by con-
tinuity to the multiplication of a weight by any number, rational or irra-
tional. It is again remarkable that one gets the same operation of multipli-
cation for all these types of magnitudes.
To summarize: The "practical" operations of addition and multiplica-
tion on various types of magnitudes lead to the algebraic operations of
sums and product for the real numbers on the linear scale. The various
rules for these manipulations of numbers were well known before they

O . . . . . : : : : : - - - - ; - - - - - A . L - - - - -X- - - -B
-:-------

Figure 1
98 IV. Real Numbers

were codified by axioms. We record the codification: The real numbers

form an abelian group under addition, under addition and multiplication
they form a commutative ring, moreover, one which is a field. Here a
commutative ring R is a set of elements (numbers) which is an abelian
group under an operation of addition, which has an associative and com-
mutative binary operation of multiplication with a unit (a number 1 such
that r·l = r for all r in R) and in which both distributative laws hold:

a( b + c) = ab + ac (b + c)a = ba + ca. (1)

(Since multiplication is commutative, either one of these two laws would

suffice; it is a curious observation, repeated in other axiom systems, that
here a single axiom, the distributive law, suffices to tie together the
separately axiomatized additive and multiplicative structures.) Finally, a
field is a commutative ring in which every equation xa = 1 with a =1= 0
has a solution x, necessarily unique. The remarkable fact is that all the
algebraic rules for the manipulation of sums and products of real numbers
follow from this simple list of axioms-a list far more perspicuous than the
axioms of plane geometry, reflecting in part the fact that the line (the
scale) is geometrically simpler than the plane, and that the axioms for the
line more strictly separate the algebraic structure from the order structure
(§4).
For the immediate purposes of this chapter, we could have formulated
these axioms just as properties of the real numbers. With the more gen-
eral terms (ring and field) defined above, we can note that there are
already at hand other examples of fields (the rational numbers Q and
Z/(p), the integers modulo a prime). There are many more examples of
commutative rings, including the system Z of integers and Zn, the integers
modulo any n. More generally, a ring is defined by dropping the commu-
tative law ab = ba in our definition of commutative ring, but retaining
both distributive laws (1). A division ring is a ring in which every equation
xa = 1 and ay = 1 for a =1= 0 has a solution x or y. These general
notions, involving a possibly non-commutative multiplication, are
recorded here for convenience. They do not belong here in the order of
ideas, which now should concern only the scale of reals and its remark-
able properties.

4. Comparison of Magnitudes
Many practical observations about magnitudes amount to the determina-
tion of which of two magnitudes is the greater. In geometry, the notion of
betweenness provides for such a comparison of the magnitudes of two
segments. On a line, the directed segment AB is less than the directed seg-
ment AC if and only if B lies between A and C. If the real numbers band
4. Comparison of Magnitudes 99

e represent these segments AB and A C, then this defines the relation

b < e (b less than e) for real numbers band e. The axioms appropriate to
this relation "less than" can be derived (though we will not carry this
derivation out) from the geometrical axioms for betweenness. They state
that the real numbers form an "ordered field". Here a field F is said to be
ordered if it is linearly ordered by a binary relation a < b with the follow-
ing two properties, for all field elements a, b, and e:

a<b implies that a + e < b + e, (1)

a < band 0 < e imply that ae < be. (2)

(Note that here too there is just one axiom connecting order and addition,
and just one axiom connecting order and multiplication.) Observe also
that the rational numbers (as well as the reals) form an ordered field in
this sense.
These axioms provide for all the usual manipulations of inequalities. In
particular, a real number e is positive if and only if e > 0, while the abso-
lute value I b I of a number b is b or -b according as b >
0 or b < o.
Then one proves rules such as

I ab I = I a I I b I ' la+bl < lal + Ibl·

These absolute values formulate the notion of magnitude without direc-
tion, and are convenient for expressing ideas of "nearness". Thus a is near
to b when I a - b I is small.
Inequalities and absolute values are the customary formal tools for
expressing ideas of approximation. These ideas arise in many ways,
perhaps most vividly in the observation that a real number, considered as
an infinite decimal, can be successively approximated by finite decimals;
that is, by rational numbers with denominators powers of 2 or of 10. To
formalize this idea, one may start by recalling the Archimedean axiom
from geometry (§III.2): that multiples of any segment OA with 0 =1= A
suffice to bound all larger segments-any segment OB is exceeded by some
multiple of the first segment OA. In arithmetic terms, this amounts to
observing that the order of the real numbers satisfies the following
"Archimedean" law:

Archimedean Law. If a and b are positive, then there is a natural number

n such that na > b:
In other words, no matter how small the positive number a may be,
every other positive number b will be exceeded by some multiple of a.
This amounts to asserting that there are no infinitesimal real numbers a-
though for other purposes, as we will see, it is convenient to think about
infinitesimals.
The Archimedean law also implies the important property that each
real number can be approximated-to any desired degree of accuracy-by
100 IV. Real Numbers

a rational numbers. This amounts to saying that there is a rational

between any two reals.

Theorem. IfO < b < c, there is a rational min with

b<mln<c. (3)

The idea of the proof is that c - b > 0, so must exceed the reciprocal
1In of some integer n. This can be formalized as follows. Since
c - b > 0, the Archimedean law provides a natural number n with
n(c-b) > 1; that is, with c > b + lin. By the same law and mathemati-
cal induction, there is also a smallest natural number m with m·l > nb.
This means that b < min. Since m is the smallest such natural number,
we must also have m - I <
nb; that is, min <
b + I In and hence
min < c. Thus min is between band c, just as required.
A similar argument will show that the interval from b to c contains a
rational with denominator a power of 2. This suggests again that every
such number, such as b, can be approximated by rationals with denomi-
nators powers of 2.
Now that the notions of "order" and "absolute value" are formalized,
one may also give a formal definition of the idea that a sequence con-
verges to a limit. To say that a sequence {an} = al ,a2, . .. of real
numbers an converges to a real number b as limit should mean that the
successive terms an get close to b-ultimately closer than any preassigned
measure. Here "ultimately" is to mean "after some index n", while the
preassigned measure is to be a positive (but "small") real number, usually
written £ > 0. The idea "ultimately closer than this £" is better expressed
in the opposite order, that given such an £ one can find an index n beyond
which an is indeed that close to b. In formal language, this becomes the

limit if and only if to each real £ > °

Definition. The sequence an of real numbers has the real number b as a
there exists a natural number k
with the property: if n > k, then I an - b I < £.

In this formal definition, the anthropometric ideas (given an £, we can

all £ > °
find an index n) have been replaced by the use of logical quantifiers (For
there exists a k ... ). The particular choice of k is not relevant,
for once k works for an £, so will any larger natural number k '. The infor-
mal idea that an gets close to b seems to have disappeared, but it is
covered in the phrase "for all positive £"-and hence in particular for a
small £. Indeed, once the statement is true for one £, it automatically holds
for any larger £2, and with the same k. For instance, by the Archimedean
law there is to each positive £2 a natural number m > °with 11m < £2.
Hence for convergence it is enough to require that to each natural
4. Comparison of Magnitudes 101

number m there exists k so that if n > k, then I an - b I < 11m. In this

way different formal statements deseribe the s.ame intuitive idea of conver-
gence.
This sophisticated definition of convergence to a limit is a remarkable
achievement, since it both codifies the intuitive intentions about conver-
gence and makes possible proofs of the various expected formal conse-
quences. For example, if a sequence {an} converges to limits band b',
then necessarily b = b' (because they differ by less than any t: > 0).
Also, if {an} converges to band {cn} to d, then {an + Cn} converges to
b + d and {anc n} to bd. Moreover, in a convergent sequence the succes-
sive terms must not only get close to the limit, but also "ultimately" close
to each other. This can be expressed formally by the theorem that any
convergent sequence {an} is Cauchy-where a sequence is said to be
Cauchy if to every real t: > 0 there is a natural number k such that
n > k and m > k imply I an - am I < t:.
The ideas of approximation represented in the definition of conver-
gence are essentially the same as those involved in the definition (§IA) of
continuity. Indeed, using the axiom of choice (!), one can prove that a
function f: R ..... R on the reals is continuous at a real number b if and
only if f maps every sequence an converging to b into a sequence f(a n)
converging to f(b). In brief, f continuous means exactly that f preserves
convergence.
Infinite sums also come up and must be explained in terms of conver-
gence. Thus every unending decimal fraction is really an infinite sum; for
example, 1/6 is the unending decimal

.166666 ... = 1/10 + 6/102 + 6/103 + 6/104 +

For m a positive integer, the binomial theorem gives

m(m-l) 2
(1 + x)m = 1 + mx + x + ... + xm,
1·2
but when m is a fraction this becomes an infinite series in powers of x

m(m-1) 2
(1 + xt = 1 + mx + 1·2
x +

+ m( m - 1) ... (m - n - 1) X
n
+ ...
1·2 ... n
The formula suggest an infinite number of additions. This infinite opera-
tion is not literally possible; instead one approximates by finite sums.
The study of convergent series is essentially equivalent to that of con-
vergent sequences. An infinite series is a formal infinite sum

c] + C2 + C2 + ...
102 IV. Real Numbers

of real numbers Ci, and is said to converge to the limit b if and only if the
sequence Sn = Cj + C2 + ... Cn of "partial sums" of the series con-
verges to b. One could, vice versa, define convergent sequences in terms of
convergent series, and the understanding of either notion is improved at
the hand of examples of series which do not converge-as for instance
with the harmonic series 1 + 1/2 + 1/3 + 1/4 + .... Infinite series also
occur essentially in complex analysis (§X.7), to expand known functions
and to define new ones.

Commenl. Here is an example of an ordered field in which the

Archimedean law fails. It is suggested by the manipulation of the power
series which occurs in analysis. Introduce a formal symbol I and consider
all possible infinite power series which occurs in analysis. Introduce a for-
mal symbol I and consider all possible infinite power series of the form

with real number coefficients and with only a finite number of negative
powers of I-but with no convergence required. Add two such "formal"
series by adding the corresponding coefficients, and multiply two such
~ries in a purely formal way; for example

(ao +aj/+a2/2+ ... )(bo +b j /+b 2/2+ ... )

By suitable calculations, one may verify that these elements form a field,
called the field of formal power series R( (I)). One may order this field by
specifying that s > 0 if and only if its first non-vanishing coefficient a- n
is positive. Then also s > 0 if and only if s - I is positive. This means, in
partic~lar, that 1 > 0 and also that.1 > 0; how~ver,. no. integral. multipl.e
of I wIll be as large as I-so that I IS a sort of mfimteslmal, whIle I - IS
"infinitely large" compared to 1. Moreover, the infinite series s of ele-
ments in this field is convergent in the sense of our definition of
convergence-using among the ( > 0 all the powers ( = In for n > o.

5. Axioms for the Reals

The practical understanding of real magnitudes and their uses in approxi-
mations leads eventually to the idea of characterizing the field R of real
numbers by a suitable list of axioms. It is not enough to require just that
R be an ordered field, for there are many such fields, including the field Q
of rational numbers and the formal power series field R«/)). The crucial
feature is a completeness axiom, stated in §I.4: Every non-empty set of
reals with an upper bound has a least upper bound. An ordered field with
5. Axioms for the Reals 103

this property is called a complete ordered field. The real numbers form
such a field. The formal power series do not.
This completeness axiom implies the Archimedean law. For suppose to
the contrary that there were positive reals a > 0 and b > 0 such that no
multiple na exceeded b. Then the set S of all multiples na, for n a natural
number, has an upper bound, namely b. By the completeness axiom, S
then has some least upper bound, call it b*; thus b* ;;:: na for all n. This
also means that b* ;;:: (n+ l)a for all n, and so that b*-a ;;:: na for all n.
Thus b* - a is less than b* and is also an upper bound for S, a contra-
diction to the choice of b* as a least upper bound for S.
The force of the completeness axiom is to insure that all the real
numbers that ought to be there are there. For example, the irrational Y2
must be there, as the least upper bound of the set 1, 1.4, 1.41, 1.414, ... of
rationals (those approximating Y2). Similarly 'TT must be there, as the least
upper bound of the set of decimals 3, 3.1, 3.14, 3.141, 3.1415, 3.14159, ....
Indeed, since there is a rational between any two reals, any real can be
expressed as the least upper bound of a set of rationals. The completeness
axiom is also used in more sophisticated ways, for example in the proof of
Rolle's theorem and of the mean value theorem of the calculus (Chapter
V).
There are other equivalent forms of the completeness axiom. Instead of
requiring that every bounded set of reals have a least upper bound, one
may require anyone of the following:

Dedekind Cut Axiom. If the set R of reals is the union of two disjoint
non-empty subsets L and V such that x EL and y E V imply x < y, then
there is a real number r such that x <
r for all x EL and r <
y for all
YEV.

Cauchy Condition. Every Cauchy sequence of real numbers has a limit.

Weierstrass Condition. If a series c) + c2 + ... + cn + ... of posi-

tive real numbers Cn > 0 is such that there is an upper bound b for all
partial sums Sn = c) + ... + Cn, then the series converges.

It is illuminating to establish the equivalence of the different complete-

ness axioms. The Dedekind cut axiom is perhaps the more geometric (and
has already appeared (§III.2) in the completeness axioms for the geometry
of the plane): ArIy cut of the real line into a lower part L and an upper
part V must be a cut at some real number r. Indeed, each real number
determines just two such cuts, one with L consisting of all x <r, the
other with L consisting of all x < r. In this way, real numbers can be
completely described by cuts.
The completeness axioms also serves to determine the real numbers
uniquely, up to an isomorphism of ordered fields. We sketch the proof.
104 IV. Real Numbers

°
multiplication. Then °
Suppose that R' is any complete ordered field, with unit element I' for
< I', for otherwise I' < 0, which would give

natural number n > °

< -I' and so I' = (-1')( -I') > 0, a contradiction. For each
the multiple n I' must then be positive, and all
these elements n I' are different elements in our field R '. Therefore
If> (n) = n I' defines an injective map If>: N ~ R '. Because R' is a field,
this If> can be extended to a map If>: Q ~ R ' by setting
If>(n/m) = If>(n)l!f>(m) whenever m =1= 0; in other words, R' contains a
copy If> (Q) of the ordinary rational numbers Q. Also, If> preserves the
order. By the Archimedean law, each element r' in R' must then be a
least upper bound of a set L' of these rationals, indeed it is the least
upper bound of the image If> (L) of the set

L = {nlm 1m =1= 0, If> (nlm) < r'}.

This set L has the special property that x < y and y E L imply x E L, so it
is the lower half of a "cut" in the rationals. Since it is bounded, it has an
(ordinary) real number r as least upper bound, and r in turn determines
L. The one-to-one correspondence r ~ r' then maps the ordinary real
numbers on the given complete ordered field R '. One can show that it
preserves sums, products and order, hence it is the desired order isomor-
phism R == R'.
Note that this argument makes essential use of sets, such as the sets L
of rationals. In this it is like the use of sets in the induction axioms to
prove that the Peano postulates uniquely determine the natural numbers.
We now have two different axiomatic descriptions of the reals: Here,
the realjield, described as a complete ordered field, in §I.4 the real contin-
uum, described as an unbounded ordered set with a denumerable dense
subset. Each description determines the set of reals uniquely, up to an iso-
morphism of the structure concerned. However the structures are drasti-
cally different. The real continuum has only the order structure, and there
are many automorphisms of this structure. The real field has both order
and algebraic structure, and its only automorphism (by a proof like that
just above) is the identity automorphism. These differing structures on the
same "thing" (here the reals) are much like the differing structures on the
plane (without and with orientations, as in §III.8). These differing struc-
tures furthermore reflect practical differences. Thus the ordered contin-
uum handles comparisons of many items where one knows only which of
two items is the larger, with no measure of "how much" larger.
In either structure, the real numbers form a "continuous" scale. Physi-
cists (and others) sometimes suggest that an "atomic" or "discrete" scale
would be more "real"; finitists propose finite scales of magnitude. These
proposals turn out to be hard to execute-the continuous scale of reals
works smoothly!
6. Arithmetic Construction of the Reals lOS

6. Arithmetic Construction of the Reals

The axiomatic approach assumes that the real numbers are already there,
as geometrical or other magnitudes, and describes them-uniquely up to
an isomorphism-by axioms. There is a wholly different and deliberately
arithmetic approach, in which one starts from the natural numbers and
constructs the reals as sets of natural numbers. Since we already have con-
structed the rationals, it will be sufficient (and more appropriate) to con-
struct the real numbers directly from the rationals and sets of rationals in
such a way that the resulting reals do satisfy the axioms.
This construction can be done in several ways, in parallel to the
different forms of the completeness axiom. The construction by Dedekind
cuts in the rationals is perhaps the most direct. Define a Dedekind cut in
the field Q of rational numbers to be a pair of disjoint non-empty subsets
(L, U) with union Q such that x ELand y E U imply x < y; to get
uniqueness, require that the "lower set" L have no maximal element. Any
such cut is determined by the set L alone, because it is a bounded non-
empty set L of rationals, with no maximal element, which does not con-
tain all rationals but has x ' E L whenever x' < x E L. Call such a set L a
real number. To add two such real numbers Land L', take the sum L" to
be the set of all sums x + x' of rationals x EL and x' EL'. Multiplica-
tion is more delicate, because the product of two negative rational
numbers is positive. If both the real numbers Land L' contain positive
rationals, take their product LL' to be the set of all products xx' of
rationals x EL and x' EL', with at least one of x,x' positive. The remain-
ing cases of multiplication can then be handled by replacing the set L by
its negative, defined as the set of all rational y with x + y < 0 for every
x E L. Finally, define the order by specifying that L =< L' if and only if
L c L'. A systematic proof (cf. Landau [19S1]) then shows that the real
numbers L so constructed do form a complete ordered field R. Moreover,
the rationals Q can be embedded in R by the monomorphism (of ordered
fields) which sends each rational x into the set L of all rationals y with
y < x. (This may recall the uniqueness proof of §S.)
There is an alternative construction of the reals as Cauchy sequences of
rationals. Since different such Cauchy sequences may have "the same"
limit, one must for this construction consider two Cauchy sequences {an}
and {b n } as equivalent when the sequence an - bn converges to O. Under
suitable operations of addition and multiplication, and with a suitable
linear order, these equivalence classes of Cauchy sequences form a com-
plete ordered field. By the uniqueness of such fields, this field is iso-
morphic to the field of reals constructed by Dedekind cuts. However, in
other foundations of Mathematics in terms of elementary topoi the
Dedekind and Cauchy sequence constructions may yield differing results
(see Johnstone [1977]).
106 IV. Real Numbers

It was primarily George Cantor who developed this construction of real

numbers by Cauchy sequences. Weierstrass has a quite similar construc-
tion related to his version of the completeness axiom. He considered
sequences c = {cn } of positive rational numbers Cn with bounded partial
sums; that is, such that there is a rational b with c, + ... + Cn b <
for every n. If c' is a second such sequence, then c <
c' is defined to
mean that for each n there is a natural number m with

c, + .,. + Cn < c', + ... + c'm.

If both c < c' and c' <
c, then c and c' are said to be equivalent. The
equivalence classes of such sequences then constitute the desired positive
real numbers, with operations of addition and multiplication defined by

(This latter is essentially the formal product of the two intended infinite
series.)
This construction really amounts to the use of the sequence
Sn = C, + .. , + Cn which is both monotone and bounded:

S, < S2 < ... < Sn < ... < b.

In particular, it must therefore be a Cauchy sequence. The choice of a
construction by series or by sequences then depends on one's
preference-and Weierstrass, in all of his extensive work, was inclined to
use series. However, one cannot say that a real number is a Dedekind cut
or is an (equivalence class of) Cauchy sequences; the cut or the sequence
is only a means to an end, the construction of the reals, defined only up to
an isomorphism. The one idea of a real number has different formaliza-
tions.
All of these constructions are arithmetic, in that they build up the real
numbers from the arithmetic of the rational numbers. They are, however,
not purely arithmetic, since each of the constructions also makes intrinsic
use of set theory-the set of all Dedekind cuts, each cut itself a set, or the
set of all equivalences classes (= sets) of Cauchy sequences. Hence the
appearance, in the constructions, that the reals are "purely" arithmetic is
deceptive. It would be more accurate to say that the scale of real magni-
tudes requires both sets and arithmetic, as well as geometric understand-
ing.
Our axioms for the reals are the standard ones, but there are other
models of the reals-as indeed there are multiple models for many
mathematical ideas. The "non-standard" reals, built by a sophisticated
7. Vector Geometry 107

construction, contain both the standard reals and actual infinitesimals;

they can be used to prove theorems (for the ordinary reals) (Lightstone
and Robinson [1975]).
Other approaches restrict attention to those real numbers which are
explicitly "constructible" in some specific sense, perhaps by means of
recursive functions of natural numbers. An efficient version of construc-
tive analysis is due to Bishop [1967]; the corresponding real numbers are
described in Myhill [1973]. Earlier, L.EJ. Brouwer developed intuition-
ism, with emphasis on real numbers given by choice sequences (Troelstra
[1977]; see also Dummett [1977]). In my view, such requirements of con-
structivity tend to emphasize arithmetic at the expense of geometry. Since
the real numbers should subsume both arithmetic and geometric aspects
of mensuration, these practical uses seem better formalized by the stan-
dard axioms for the field of real numbers.

7. Vector Geometry
The real numbers provide a one-dimensional geometry-a line with a
chosen origin and unit. The higher dimensional analog is vector geometry.
Thus in the Euclidean plane, choose an origin 0 and a segment OU
representing the unit of distance. Then each point P in the plane is
represented by the directed segment OP, called a vector v. Vectors OP and
OQ can be added by the usual parallelogram law: Form the parallelogram
with edges OP and OQ, and take the diagonal vector OR as the sum (Fig-
ure 1). Under this addition, the vectors form an abelian group; it is just
the group of all translations of the plane (§III.7). Using the chosen unit
segment OU, each real number r can be laid off along the line OU. The
geometric construction of multiplication by r (Figure 3.1) then defines an

°
operation of multiplying the vector v = OR by the real number (scalar) r
to give the vector rv along the line of OR. If r > it is in the direction of
v; if r < 0, it is in the opposite direction, in both cases with I r I times

---- ------
Q

o
------ -
u

Figure 1
108 IV. Real Numbers

the length. This operation of multiplication of a vector by a scalar has the

following formal properties

r(v+w)=rv+rw, lv=v, (1)

(r + s)v = rv + sv, (rs)v = r(sv). (2)

for all vectors v and wand all scalars r,s E R. In view of these properties
of addition and of scalar multiplication, we say that the vectors form a
vector space over the real numbers.
Many geometric facts can be established readily by the use of vector
algebra. For example, this algebra gives an easy proof of the theorem that
the three medians of a triangle meet in a point.
Vectors may also be introduced in 3-dimensional spaces (and in higher
dimensional spaces); moreover, exactly the same formulas (1) and (2)
apply, whatever the dimension. More generally, a vector space V over a
field F is an abelian group (under addition) with another operation
F X V --+ V, written (r, v) f-> rv for elements rEF, v E V, which satisfies
with addition the four laws (l) and (2) above. One can also describe for-
mally the "dimension" of such a space. In the plane, if Ul and U2 are two
non-collinear vectors, every vector v in the plane can be written as a
"linear combination" V= Xl Ul + X2U2 with unique choices for the
scalars Xl and X2. In three dimensions, we need three such vectors Ul, U2
and U3, along three "axes". More generally, vectors Ul , . . . ,Un in V are
said to form a basis for the vector space V if every vector v in V has a
unique expression

v= Xl UI + ... + XnUn ;

then the Xi are the coordinates of v relative to the basis Ui' A first theorem
of vector geometry (= linear algebra) states that any two (finite) bases for
a given vector space must have the same number of elements. This
number n is the dimension of the space.
Vectors also provide a convenient description for certain functions
transforming the plane into itself (and leaving the origin fixed):
Transformations such as expansions from the origin (v f-> rv) for a fixed
r, compressions on a line through the origin, and shears. All of these
transformations preserve the vector operations. More generally, a linear
transformation on V is a function t: V --+ V which preserves both addition
of vectors and multiplication of vectors by scalars. This means that the
equations

t( v + w) = tv + tw, t( rv) = r( tv ) (3)

hold for all vectors v and wand all scalars r. For example, a rotation
about the chosen origin 0 is a linear transformation (it maps the parallelo-
8. Analytic Geometry 109

gram of addition into a parallelogram). The same notion of linear

transformation applies to higher dimensional vector spaces, and is the
cornerstone of the study of linear algebra and matrix theory (Chapter
VII).
The general idea that some geometrical facts can be handled well by
appropriate algebraic operations has now appeared in the use of real
numbers for one-dimensional geometry and of vectors for higher dimen-
sions. The addition of vectors was possible only with the choice of an ori-
gin O. Without a choice of origin one still has some algebraic operations
such as the construction of the midpoint, written (l /2)P + (l /2 )Q, of
two points P and Q or of the point (2/3)P + (l/3)Q which is one third
of the way from P to Q. The general such operation is the formation of
the weighted average

(4)

of n + I points with weights real numbers Wi with sum

Wo + ... + Wn = 1. A transformation preserving such averages is
called an affine transformation; the resulting geometry (in any number of
dimensions) is affine geometry. It is possible to write a complete set of
axioms for the weighted average operations (4), but they are cumbersome
(Algebra, first edition). Even the simplest ternary operation
Po - PI + P2 of this type has not proven to be very useful. It seems
more efficient to reduce to vector algebra by choosing an origin. Addition
is a very convenient operation! Binary operations are much handier than
ternary ones.

8. Analytic Geometry
Another and earlier reduction of plane geometry to algebra is provided by
the familiar method of cartesian coordinates. Given an orientation, a
choice of origin and unit, and two perpendicular coordinate axes, each
point P is represented by its coordinates, a pair (x,y) of real numbers.
Each line in the plane may be described as the set of those points whose
coordinates satisfy a linear equation, while the distance between two
points is given by the familiar formula in coordinates, derived from the
Pythagorean theorem. In this way, all sort of geometric facts about the
plane are handled by algebraic machinery; in effect, the plane is reduced
to the cartesian product R X R = R2 of two copies of the real line R.
However, this reduction does depend on the choice of origin and of axes,
and one must betimes verify that truly geometric facts are independent of
the choice of coordinates.
Geometry and coordinates arise first in dimensions 2 and 3. The need
for higher dimensional geometry is motivated by phenomena which need
110 IV. Real Numbers

specification of more than three coordinates: Events in space-time need

four (position and time); in dynamics, the initial conditions for a particle
need six, three for position and three for velocity. The use of such a six-
dimensional phase space is a first example of the importance of mechanics
as a motivation for mathematical developments (Chapter IX). To be sure,
this example could be (but usually is not) described without explicit use of
coordinates, as the product of a three-dimensional position space and
another three-dimensional velocity space. A general n-dimensional
Euclidean space may be constructed as the n-fold product Rn of real lines,
with points the n-tuples of real numbers (Xl, ... ,xn ). This is a vector
space over R, with basis the n "unit vectors" (0, ... , 0,1,0, ... ,0). Such
use of coordinates allows for the efficient treatment of higher-dimensional
phenomena in dimensions where the corresponding geometric intuitions
are weak or non-existent. It serves to extend geometric ideas beyond the
ordinary three dimensions.
Much of mathematical physics deals with phenomena in three dimen-
sions and so often requires a formulation in triads of equations, one for
each coordinate in R3. When R3 is regarded as a vector space, many of
these equations can be written as single vector equations. This accounts
for the popularity of vector analysis in Physics. It includes the use of inner
products, to which we now tum.

9. Trigonometry
Trigonometry is essentially a procedure for turning angular measures into
linear measures. This appear directly in the definition of the two basic

°
trigonometric functions sin () and cos () of the angle (). In the oriented
plane take a point P on the unit circle with center at the origin so that
the segment op makes the given angle () with the X coordinate axis. Then
cos () and sin () are defined to be the X and y coordinates of P; since the
circle has radius 1, this immediately gives the identity

(1)

This defines sin () and cos () for angles () of all sizes (not just for an angle ()
in the first quadrant, as displayed in Figure 1).
When angles are measured (in radians) by numbers t we can also think
of the sine and the cosine as functions of the number t, so that

Sin t = sin( angle of t radians). (2)

Thus there are really two legally different functions: The Sine of a
number, here with capital S, and the sine of an angle, with lower case s.
This pedantic (but real!) difference is usually ignored. It implicitly
9. Trigonometry III

L-------~~~--------~-----x
(1,0)

Figure 1

involves the wrapping function (}t of §2 above. Recall that this function
sends each real t to the point P = (cos (}t, sin (}r) on the circle such that
the length of the counterclockwise circular arc from (1,0) to P is congruent
to t, modulo 2'1T. Then the definition (2) reads

Cos t = cos( (}t ) : (3)

it really amounts to a composite function t-'>(}t -,>sin«(}t). Since the wrap-

ping function has (}tHIT = (}t we get

Sin(t + 2'1T) = Sin(t), Cos( t + 2'1T) = Cos( t) . (4)

In other words, Sin and Cos are periodic functions, of period 2'1T, as in the
familiar graph of Figure 2. Now that this has been stated, we drop the S
in Sin and the function (}r; they would just get in the way of trigonometric
manipulations, but we emphasize that the wrapping function accounts for
radian measure and the periodicity (4) of the trigonometric functions.
Indeed, the whole study of periodic functions, as carried on in Fourier
analysis, concerns the expression of more or less arbitrary functions f(t)
of period 2'1T in terms of the periodic functions sin nt and cos nt for the
natural numbers n.
A rotation of the circle of Figure 1 about the origin through the angle ()
will carry the point (1,0) to our point P and the point (0,1) to a
corresponding point P'; thus this rotation has the effect

sin t
112 IV. Real Numbers

(1,0) f-> (cos O,sin 0), (0,1) f-> (-sin O,cos 0) (5)

Since (I,Q) and (0,1) form a basis for the 2-dimensional vector space, any
point Q with coordinates (x,y) can be written as the linear combination
(x,y) = x(l,O) + y(O,I) of the two basis vectors. Now rotation is, as
already noted, a linear transformation, so preserves linear combinations
and thus by (5) has the effect

(x,y) f-> (x cos 0 - y sin 0, x sin 0 + Y cos 0).

When one writes (x ',y') for the coordinates of the rotated point, this
yields the customary equations in coordinates

x' = x cos 0 - y sin 0, (6)

y' = x sin 0 + Y cos 0

for a rotation (x,y) f-> (x',y') of the plane about the origin. Also, if one
takes Q = (x,y) to be that point Q on the unit circle of Figure 1 for
which OQ makes the angle cp with the positive x-axis, then x = cos cp,
y = sin cp. Now the rotation by 0 clearly carries Q to a point QI where
OQ' makes the angle cp + 0 with the positive x-axis. The coordinates
(x',y') of Q' are then cos(cp + 0), sin(cp + 0) and the equations (6) for
rotation become

cos( cp + 0) = cos cp cos 0 - sin cp sin 0, (7)

sin(cp + 0) = cos cp sin 0 + sin cp cos o.

These "addition formulas" for the trigonometric functions are thus
another expression of the relation of angular to linear measure.
This relation also accounts for the use of trigonometric functions in cal-
culating the connections between the sides and the angles of a triangle, as
suggested by the Euclidean congruence theorems. In a triangle OAB with
sides a, band c, as illustrated in Figure 3, drop a perpendicular from A to
the opposite side at M. If 0 and cp are the angles of the triangle at 0 and

b c

e ¢
O~~-----M~---------------L--~B

Figure 3
9. Trigonometry II3

at B, the definition of the sine function shows that the length AM can be
expressed either as b sin () or as c sin €J>; this gives the law of sines:
blc = sin €J> Isin (). On the other hand, one may compute the length c of
the side AB from the Pythagorean theorem for the right triangle AMB to
be

(8)

This is the law of cosines. When u = OB and v = OA are regarded as vec-

tors of the respective lengths I OB I = a, I OA I = b, the expression
ab cos () which appears in (8) is called the inner product u· v of the two
vectors; it is a scalar (a real number). If A has coordinates (XI,YI ) and B
coordinates (X2,Y2), the lengths a, b, and c in (8) can be written in terms
of these coordinates by the Pythagorean theorem; the formula (8) then
becomes a formula for the inner product u·v:

(9)

This again is a linearization of angular phenomena.

With these ideas and formulas we have presented all of the essential
concepts of trigonometry in less than 6 pages. It is curious (and trouble-
some) that the standard elementary presentations of trigonometry have
been inflated to inordinate lengths-an inflation hardly justified by the
inevitable subtlety wrapped up in the comparison of linear and angular
magnitude.
In (9), the inner product is expressed in terms of a choice of coordi-
nates. However, as in other cases, the inner product can be described in a
more invariant way by suitable axioms. It has three characteristic proper-
ties: Linearity, symmetry, and positive definiteness, as expressed by the
three equations
(10)

u·v = V·U, (11)

u·u > 0, u·u = 0 only if u = 0, (12)

valid for all vectors u, v, VI, and V2 and all scalars tl and t2' All the pro-
perties of the inner product follow from these equations. Indeed, given
these equations-for a product U·V of vectors in a two dimensional vector
space over R-one can prove that there is a choice of basis and
corresponding coordinates for which the given inner product is expressed
by the formula (9). Such a basis is called a normal orthogonal basis (for
the given inner product).
We will extend these ideas to higher dimensions in §VII.5.
114 IV. Real Numbers

10. Complex Numbers

Each extension of the number system to a larger system is driven by the
need to solve questions which the smaller system cannot always answer.
Subtraction is not always possible in the natural numbers; hence construct
the system Z of integers. Division (by non-zero integers) is not always
possible in Z; hence construct the rationals Q where division is possible.
Cauchy sequences in Q need not converge to a limit in Q; hence construct

equations such as x 2 + 1 = °
and x 2 + x + 1 = °
the reals R to include such limits. Finally, there are plausible polynomial
with real
coefficients which do not have real solutions; hence construct the complex
numbers C to provide solutions for such equations.
Were there a solution i for x 2 + 1 = 0, there would also be combina-
tions a + bi for any real numbers a and b. They could be manipulated
by using the algebraic rules for the reals and the fact that i 2 = -1. This
would yield in particular the following rules for addition and multiplica-
tion:

(a + bi) + (e + di) = (a + e) + (b + d)i, (1)

(a + bi)·(e + di) = (ae - bd) + (ad + be)i. (2)

However, the symbol i imagined here cannot possibly be a real number,

because the square of any non-zero real is necessarily positive. Therefore
this i was originally (in the 18th century) thought to be an "imaginary"
number-not in any sense real. Early in the 19th century, these practical
rules were turned into a formal definition. One does not need the actual
symbol i; a complex number a + bi could be defined arithmetically to be
just an ordered pair (a,b) of real numbers. When the addition and multi-
plication of such pairs is defined, following (1) and (2), by the rules

(a,b) + (e,d) = (a+e,b+d), (a,b)(e,d) = (ae-bd,ad+be) (3)

it follows readily that these pairs do form a field, call it C. The given real
numbers are embedded monomorphically in C by the map a-->(a,O); in
particular I becomes (1,0). Moreover, by the multiplication rule (3) the
pair i = (0,1) has i 2 = (-1,0). Then any pair (a,b) can be rewritten as a
linear combination

(a,b) = (a,O) + (O,b) = a + bi.

In other words, this strictly arithmetic construction has produced things
which act exactly like the desired complex numbers, and i, defined as a
pair (0,1), need no longer be regarded as imaginary, although this word is
still used for i and its real multiples bi.
This arithmetic construction may also be viewed geometrically. The
complex numbers, rewritten as z = x + iy for real x and y, correspond
10. Complex Numbers 115

exactly to the points (x,y) in the cartesian plane (Figure I) with the real
numbers x + iO placed along the x axis, as usual, while the purely ima-
ginary numbers iy lie along the y axis. This provides a geometric con-
struction of the complex numbers with each complex number z a vector
OZ from the origin, while the addition of complex numbers is just addi-
tion of the corresponding vectors. The complex number z may also be
described with polar coordinates: The counterclockwise angle 0 from the
positive x-axis to the vector OZ and the length r >
0 of this vector; one
calls 0 the argument of z and r its absolute value, and writes

z = r( cos 0 +i sin 0), r = I z I, o= arg z. (4)

This polar form is especially useful for the multiplication of complex

numbers. Thus if w = s(cos cf> + i sin cf» is a second complex number,
the product zw, as defined by (3), can be rewritten, using the addition for-
mulas for sin and cosine, as

zw = rs(cos(O + cf» + i sin(O + cf». (5)

In other words

I zw I = I z I I wi, arg(zw) = arg z + arg w; (6)

thus multplication of complex numbers multiples their absolute values

and adds their arguments. This may be stated geometrically: Multiplica-
tion by the complex number w is a transformation z 1-+ zw of the z-plane
into itself which rotates the plane counterclockwise about the origin by
the angle arg wand at the same time stretches the plane away from the
origin uniformly by the factor I wi. This geometric description of com-
plex multiplication illustrates the tight connection between the algebra of
complex numbers under addition and multiplication and the geometry of
the plane under translation, rotation, and expansion. Thus i, once ima-
ginary, now means rotate by 90°! The subsequent development of calculus
for complex numbers, as we will see in Chapter X, illustrates many more
such relations between algebra and geometry.
The complex numbers thus are constructed by adjoining to the field of
real numbers one solution of the polynomial equation x 2 + 1 = O. It is a
y

z =x + iy

---¥~--~-----------------.x

Figure I
116 IV. Real Numbers

remarkable fact that this one adjunction suffices to provide solutions for
all such polynomial equations. Indeed, the fundamental theorem of alge-
bra first proved by C.F. Gauss asserts that any equation
anx n + an_1X n- 1 + ... + alX + ao = 0 with complex coefficients ai
and with an =F 0 has at least one complex number as root. The proof is
not easy. At least one form of the proof (presented in Birkhoff-Mac Lane,
Chapter V) uses heavily the geometric realization of the complex numbers
in the complex plane. Another proof will be indicated in §X.9.
This construction of the complex numbers does not depend on any spe-
cial virtues of the equation x 2 + 1 = 0; other polynomial equations can
serve as well. For example, the equation x 2 + x + 1 = 0 also has no
real root, because the minimum of x 2 + x + 1 for x real lies at
x = - 1/2 and is 3/4 > O. One might then introduce a new symbol w for
a non-real root of this equation and manipulate the combinations a + bw
for a and b real, using the rule w2 = - W - 1, derived from this equation,
in multiplying two such symbols. The resulting symbols again form a
field; however, this field is isomophic to C under the map sending w to
- 112 + (V3/2)i. Indeed w3 = _w2 - w = 1, so this symbol w is in
fact a (complex) cube root of 1. By the rule for products of complex
numbers, the argument of any cube root of I must be 0 or ±2'7T/3; we
have taken w = cos(2'7T13) + i sin(2'7T13).
There could be a different choice as

w' = cos( 4'71' /3) + i sine4'71' /3) = cos(2'7T /3) - i sin(2'7T /3).

The change from w to w I amounts to replacing i by - i. (Switching the

two roots of x 2 + I = 0.) For any complex numbers, this replacement
a + bi I-> a - bi carries sums to sums and products to products-
because the definition of sum and product depend only on the rule that
i 2 = -1, a rule equally valid for i and for -i. This replacement,
a + bi I-> a - bi, called complex conjugation, is then an automorphism
(an isomorphism with itself) of the field of complex numbers. It is a first
example of the automorphisms studied in the Galois theory of polynomial
equations (see §V.7).

11. Stereographic Projection and Infinity

Division by zero does not work; if attempted, it leads inevitably to para-
doxes. Nevertheless, it is suggestive to think that the reciprocal w = liz
of a complex number z should approach infinity as z approaches zero.
This does not mean that there is a number 00 in the field C of complex
numbers, but it does mean that there is a geometric model which adds a
"point" at "infinity" to the complex plane.
Place a sphere of radius I so that it is tangent to the complex plane at
its south pole, at the origin of the plane. Then project each point P of the
II. Stereographic Projection and Infinity 117

sphere onto the plane. This means that one prolongs the line NP from the
north pole N of the sphere to P till it meets the plane at a point P' (Fig-
ure 1). This provides a transformation P I-> P' which is a bijection from
the sphere (omitting the north pole) to the plane. It is called stereographic
projection. It carries the south pole of the sphere to the origin in the plane
and the equator to the circle of radius 2 with center at the origin. Each
line L in the plane is the projection of a circle passing through the north
pole of the sphere-it is the circle cut out from the sphere by the plane
passing through the line L and the pole N. Also, if two lines Land L' in
the plane meet at an angle (J, the corresponding circles on the sphere meet
at the same angle (J on the sphere. Thus stereographic projection is a con-
formal transformation, in the sense that it preserves angles-though it
manifestly does not preserve distances.
Under this stereographic projection all the points P on the sphere reap-
pear as points P' on the plane-except for the north pole N on the sphere.
But one can "extend" the plane by adding a point 00, called the point "at
infinity" and then specify that stereo graphic projection sends the north
pole N on the sphere to 00 on the extended plane; it is then a bijection
from the whole sphere to the whole (extended) plane. As a point P' on
the plane moves along a straight line L away from the origin, the
corresponding point P on the sphere moves toward the north pole N.
Hence the point 00 at infinity in the extended plane lies on all lines in the
plane, moreover, as a complex number z "approaches 00" in absolute
value, the corresponding point in the extended plane approaches this
point 00. This can be formalized by defining "neighborhoods" of 00 so as
to make the extended plane a topological space.
The virtue of the extended plane (or the corresponding Riemann
Sphere) is that functions such as w = liz are now defined for all z.
Specifically, for z = rete =1= 0, the corresponding w = liz is (lIr)e -te
(draw a picture), while for z = 0, W = 00 and for z = 00, W = 0. The

N~,
\ '"
\ '-
\ Q \-,,>O..,.--------~
/-------,--" \
\ Q'
\
\

Figure I. Stereographic projection.

ll8 IV. Real Numbers

resulting transformation w = liz of the extended plane to itself can be

shown to be conformal (with the inevitable definition of angles at the
point (0). The same holds for the more general fractional linear transfor-
mations

w = (az + b)/(cz + d), ad - bc ¥= 0,

where the point z = -die which makes the denominator zero is sent Lv
the point at 00 in the w-plane. These transformations form a group, with
many fascinating properties. But the extended complex plane is useful in
many other ways, for the effective geometric understanding of more gen-
eral functions w = f(z) of a complex variable z (Chapter X).

12. Are Imaginary Numbers Real?

As the name "imaginary" for i recalls, the introduction of the complex
numbers in Mathematics met with doubts and resistance, and it was
thought that numbers such as 4i were not really there, hence were "ima-
ginary". This doubt can be put as a matter of principle as follows: If
Mathematics is the science of number, space, time, and motion, and if
number covers just whole, rational and real numbers, then there is no
place in this science for these imaginary numbers. Therefore, the gradual
general acceptance (after 1800) of the complex numbers as reasonable and
useful does mark a departure from the "number, space, time, and
motion" concept of Mathematics. In other words, once the construction of
the complex numbers is accepted as "real", it becomes more natural to
accept many further "constructions"-of higher-dimensional spaces,
quatemions, groups, and the like. The complex numbers represent a
major first step in developing the present scope of Mathematics.
The acceptance of complex numbers depends only in small part on the
standard "explanation" (§ 10) that a complex number is a point in the
plane or an ordered pair of real numbers. The ordered pair approach does
represent the fact that a complex number has a real part and an ima-
ginary part, but the formal definition of ordered pair is not really relevant.
The points in the plane can be interpreted as complex numbers only after
a choice of origin and axes. It is not just that geometry justifies the com-
plex numbers; it is also that the algebra of complex numbers helps
explain geometry. Addition gives translation, multiplication rotation.
The complete acceptance of complex numbers came primarily in their
many uses in helping to understand other parts of Mathematics. Thus,
they produce two square roots for every non-zero real number; this is just
a small part of the benefit from the fundamental theorem of algebra. Par-
ticularly striking is the so called "irreducible" case of the cubic: After the
formula (with cube roots) for the general solution of a cubic equation was
13. Abstract Algebra Revealed 119

found, it developed that there can be real cubics with all real roots for
which the formula inevitably involves the use of complex numbers (For
details, see for example Birkhoff-Mac Lane, Chapter XV, Theorem 22.)
In other words, there are phenomena with real numbers which can be
properly explained only with complex numbers. In electricity and magne-
tism, complex numbers provide a convenient formalism. Finally, many
ordinary functions (eX, sin x) of a real variable are better understood
when they are extended to be functions (e Z, sin z) of a complex variable
z. As we will see in Chapter X, the use of complex numbers yield a much
deeper understanding of the nature of a "well behaved" function. In
short, the real "foundation" for complex numbers lies in their manifold
uses in better understanding of Mathematics. This is typical of the con-
ceptual expansion of Mathematics.
The complex numbers are just one of the many new aspects which
expanded the scope of Mathematics in the 19th century. Others include
the use of integers modulo m and congruence classes, non-Euclidean
geometry, n-dimensional geometry, decompositions of algebraic numbers
into "ideal" factors (§XII.3), non-commutative multiplication in groups
and quaternions, infinite cardinal and ordinal numbers, and the various
necessary uses of sets and of logic, especially in the foundation of calculus
(Chapter VI). These 19th century developments made obsolete the simple
view that Mathematics is the science of number, space, time, and motion.

13. Abstract Algebra Revealed

The Peano axioms describe the natural number uniquely, up to isomor-
phism. The axioms for the real numbers describe these numbers uniquely,
again up to an isomorphism. Such is not the case with the axioms for a
commutative ring or a field-there are non-isomorphic models. Our suc-
cessive constructions of the number systems

N,Z,Q,R,C
have aimed to preserve the appropriate algebraic properties. All of these
systems, save the natural numbers, are commutative rings and the final
three are fields. Thus properties deduced from the axioms for a field hold
in all three cases-as well as in other fields such as the finite fields Z/(p)
for prime p. The familiar formula for finding roots for a quadratic
equation applies to anyone of these fields, as do the less familiar solu-
tions of cubic and quartic equations. The same applies to the solution of
simultaneous linear equations and to determinant formulas for their
solution-although the determinant can also be given a geometric
interpretation as an area (for 2 X 2 determinants) or as a volume in the 3
X 3 case (Birkhoff-Mac Lane, §10.3). Matrices and linear algebra work
120 IV. Real Numbers

for every field. In this way the rules for the manipulation of ordinary
numbers lead to codified rules for the manipulation of all sorts of
numbers, and the various newer sorts of numbers were constructed so as
to obey these rules. But it turns out that addition and multiplication
apply not just to "numbers" but also to other mathematical objects: To
polynomials, to congruence classes of integers, to formal power series (§4)
and the like. Algebra replaces numbers by symbols; it is then possible that
these symbols do not stand for numbers or "quantities". Hence it
develops that the actual subject matter of algebra is not the manipulations
of numbers under addition and multiplication, but the manipulation of
any objects satisfying the rules (the axioms for a ring or for a field) for
such manipulation. Algebra is thus inevitably abstract, even though this
was not fully recognized until the 1920's when Emmy Noether and her
disciples saw clearly that many algebraic phenomena could be grasped
more effectively when stated and proved abstractly.

14. The Quaternions-and Beyond

The real numbers provide an algebraic formulation of one-dimensional
geometry in algebraic terms, by means of a field. The complex numbers
provide a corresponding field for geometry in dimension 2. Why not more
such fields for geometry in higher dimension? The answer is that there
cannot be such fields.
But there is something more-in four dimensions; that is, in R4. Pick a
basis 1, i, j, k of the vector space R4, so that any q in R4 has the form of
a "quaternion"
q = t + xi + yj + zk (I)

with real coefficients t, x, y, z. For this basis, introduce the multiplication

table
·2= j·2= k 2
l 1
=-, ij = -ji = k. (2)

These equations (and the associative law) imply that j k = - k j = i and

ki = -ik = j. Hence by distributivity we can find the product of two
"quaternions" q, each of the form (1), while the sum of two quaternions is
found by term-by-term addition. With these operations the quaternions
(1) do form a division ring (as this was defined in §3). Division is possible,
because for any quaternion q ¥= 0 the multiplicative inverse is
(t - xi - yj + zk)/(t2 + x 2 + y2 + z2)1/2.
The striking fact is that the multiplication (2) of the quaternions is not
commutative. There is no alternative multiplication table (2) which would
make it commutative; that is, there is no way to define a multiplication on
the vector space R4 extending the multiplication on R so that this multi-
15. Summary 121

plication, with vector addition, makes R4 a field; i.e. a system with com-
mutative multiplication. In a field, the equation x 2 + 1 = 0 can have at
most two solutions; in the division ring of quaternion there are many:
±i, ±j, ±k, for example. Moreover, the multiplication table (2) is not
just "pulled out of the air"; it can be used to describe rotations in three
space much as the complex numbers of absolute value 1 describe rotations
in the plane. Also the multiplication rule for two "pure" quaternions
(t = 0) reads

(xi + yj + zk)(x'i + y'j + z'k) = -(xx' + yy' + zz')

+ (yz' - y'z)i + (zx' - z'x)J + (xy' - x'y)k.

The first term on the right is the "inner product" of the vectors (x,y,z)
and (x',y',z'); the remaining terms form the "vector product". It is no
wonder that texts in Physics still write vectors with the three basis units
i, J, and k.
All this suggests that the quaternions are inevitable (though perhaps
shocking to the commutativity-minded). They are. Consider any division
ring D which contains the real numbers R as a subring; then D is a vector
space over R. A famous theorem then asserts: If D is finite-dimensional
(over R), then it must be isomorphic (as a ring, preserving R) to the
quaternions, or to the complex numbers, or to the reals themselves. There
are no such finite dimensional division rings beyond dimension 4, and in
dimension 4 the quaternions (possibly with a different choice of basis) is
the only possibility!

15. Summary
Measures of many different sorts of magnitudes are all reduced to and
codified by the real numbers. The simple formal axioms for the reals
involves both algebraic aspects (the field axioms on addition and multipli-
cation), ordinal aspects (the linear order), and aspects of continuity (the
completeness requirement). Of these, the completeness requirement is the
most geometric and the deepest-and has the most varied expression. For
the algebraic axioms it is remarkable that multiplication appears to have
no natural geometric representation on the line, but only one in the plane.
One dimensional geometry is evidently quite weak.
The arithmetic constructions of the reals from the rationals are varied,
straight-forward, and not of great weight. More significant is the use of
the reals to reduce problems of geometry to questions in algebra, by way
of vectors or of coordinates. The reduction of angular measure to linear
measure is the source of trigonometry and the starting point for the alge-
braic treatment of vector spaces with inner products. It is also vital to the
122 IV. Real Numbers

introduction of complex numbers. The real numbers (in dimension one)

and the complex numbers (dimension two) both satisfy the axioms for a
field-but in dimensions higher than 2 there is no corresponding field of
"numbers".
Originally, it was the manipulation of magnitudes which led to the alge-
braic operations of addition and multiplication-but the same operations,
with corresponding properties, hold for many systems which are not
"magnitudes", hence the development of algebra as abstract. This abstrac-
tion is a natural sequel to the recognition of the real existence of ima-
ginary numbers.

References. The systematic construction of number systems can be found

in many texts, for example in Gleason [1966]. There is also a famous
classical presentation by Landau [1951]. He was an expert on austere and
rigorous formulation. The historical aspects are well suggested in
Sondheimer-Rogerson [1981].
CHAPTER V

Functions, Transformations,
and Groups

Firm formal definitions come first in any systematic foundation of

Mathematics. Hitherto, however, we have not formally defined the notion
"function", although we have used the notion extensively and consider it
central to the organization of Mathematics. This delay was deliberate,
since it has allowed the assembly of part of that wide variety of examples
from which arises the general and abstract notion of function. Though
some may hold that "abstract notions are difficult to understand" we hold
with G. Kreisel that these notions "in fact, are usually introduced to make
concrete situations intelligible" (Math Reviews 37 (1969) # 1224).

1. Types of Functions
Functions probably first appear with the practical experience that the size
of some one magnitude depends on the size of another-the weight of a
block of ice depends on the size of the block, or the distance travelled
depends on the speed, or the area of a rectangle depends on its dimen-
sions, or the angle subtended by a circular arc depends on the length of
that arc. Thus practical problems, physical problems, geometric facts, and
algebraic manipulations all indicate that one "quantity" may depend
upon others. This leads to an idea of functional dependence, often
described in suggestive but imprecise ways. A modern version of such an
informal description would say that a function acts like a machine which,
given any number as input, will produce as output some other number,
"depending" on the input.
Algebraic operations provide many examples of functions of numbers.
For instance, the operations "add 2 to the given number" or "multiply the
given number by 3" or "take the square of the given number" are func-
tions which we indicate by the "assignment" notation (with a barred
arrow f--» to designate the destination of the given number x, y, or z:
124 V. Functions, Transformations, and Groups

x ~ 2 + x, y ~ 3y, (1)

One such function, followed by a second one, will give a composite func-
tion. In elementary parlance, "substituting" 2 + x for y and then 3y for z
gives the composite of the three functions (1) as the quadratic function

x ~ 36 + 36x + 9x 2 . (2)

If the input x is a real number, so is the output; the result is a function on

the reals to the reals, in symbols R ~ R. Or, one may restrict the input x
to be rational (or allow x to be a complex number), obtaining thereby
from the same formula different functions Q ~Q or C ~C, respectively.
Many algebraic operations produce functions which are not defined for
all real numbers. This is usually the case for a rational function (the quo-
tient of two polynomials). Thus the assignment

x ~ (3x 2 - 1)/(x 2 - 3x + 2)

produces from x a real number if and only if the denominator is not zero;
that is, if and only if x =1= - I and x =1= - 2. It is then a function to the set
R of reals from that set with -I and -2 deleted: R - {-1,-2} ~R.
The assignment x ~ + Vx (take the positive square root of x) acts only
on non-negative reals (or on non-negative rationals). (The two-valued
expression ± Vx does not count for us as a function.) A function may be
determined by one of several different operations, with the choice depend-
ing in the location of the input number. Thus the absolute value of a real
number may be described by the assignments

x ~ x if x > 0, x ~ -x if x < O.

There can be all manner of such combinations, often represented by

graphs. In some cases, the function constructed may not be readily visual-
ized by a graph, as for instance for the assignments x ~ 0 if x is rational
and x ~ I if x is irrational.
Geometric definitions also produce numerical functions such as the trig-
onometric functions 8 ~ sin 8 and 8 ~ cos 8 with their familiar periodic
graphs. These functions are not polynomial or rational, but can be
expressed or even defined by formulas, specifically by infinite power series
(see Chapter X) such as

sin 8 = 8 - ff /3! + OS /5! - ....

These series, or others technically more convenient, may then be used to

calculate the familiar tables of the trigonometric functions. There is also
an inverse function "arc sin", intended to send each number x into an
angle 8 with sin 8 = x. For real angles, only the values between
2. Maps 125

- 1 and + 1 are possible for x = sin {}, and there are then many choices
for the angle. To get a function, one may choose the angle always
between - 'TT 12 and 'TT 12 inclusive; this gives the function "arc sin" from
the interval consisting of all real x with - I < < x I to the interval of all
real {} with - 'TT 12< {} < 'TT 12. Recall that such inverse functions are
needed both in solving triangles and in performing certain integrations,
such as f (l-x 2)-I!2dx. Another basic function which is not algebraic is
the exponential eX and its inverse, the logarithm log x - defined only for
positive real x.
"Variable quantities" in physics and other sciences are essentially mag-
nitudes measured (according to established scientific rules) by real
numbers. Thus weight, length, volume, temperature, velocity, and the like
are such quantities. Various elementary laws of physics assert that one
such quantity is a function of another. Thus, for materials of fixed density
p, weight w is w = p V, where V is the volume. The distance transversed
by a falling body is a function of the time t of fall, according to the famil-
iar formula s = gt 2/2. For an ideal gas, pressure P, volume V and
temperature T are related by the law PV = kT, for a suitable constant k.
Thus each of these three quantities is a function of the other two. There
are many such relations between variable quantities, and the study of the
manner of their variation is at the origin of much of Mathematics.

2. Maps
Functions of points arise in geometry. The problem of representing the
globe or a part of the globe on a piece of paper is the problem of map-
ping a portion of the sphere S2 on the Euclidean plane R2, by some func-
tion S2 ..... R2. The path of a projectile in space is specified by giving its
coordinates x, y, and z as they depend upon time t by three functions
x( t), y( t), z( t); that is, by a single function from (part of) the time axis
into R3. These and many similar examples provide "maps" -that is,
functions-from one set X of objects, points, or numbers into some other
set Y. In other words, functions need not be numerical. The operation of
rotating a plane about a given point and through a given angle maps each
point of the plane into another point, hence is a function R2 --> R2,
expressed in coordinates via trigonometry as in the linear equations
(IV.9.5). More generally, any linear transformation T of the plane is a
function R2 --> R2 which takes linear combinations to linear combinations.
Hence, writing each vector of R2 as a linear combination
(x,y) = x(l ,0) + y( 0, I) of the standard basis vectors, the whole
transformation T is determined by the images T( 1,0) = (a,e) and
T(O,I) = (b,d) of the two basis vectors; indeed one may use these
numbers to write T as

T(x,y) = (ax + by, ex + dy). (1)

126 V. Functions, Transformations, and Groups

We recall the convenient shorthand in which such a linear transformation

T, expressed thus in terms of coordinates, is represented by a 2 X 2
matrix of real numbers,

[a
cd'
b). (2)

the composition of two such transformations then corresponds to the fa-

miliar operation of the multiplication of two 2 X 2 matrices. There are
corresponding rules for n X n matrices in the case of n-dimensional
space. In any event, a matrix (2) is essentially a shorthand record of a
function on vectors to vectors, represented in coordinates as in (1); a
matrix is a model example of a Mathematical object which is neither a
number nor a point.
There are many other sorts of non-numerical functions. Thus in
Chapter I we already met correspondences (§I.l), bijections (§I.l), rigid
motions (§1.5), symmetries (§I.6), permutations (§I.3), transformations
(§I.7), and unary operations (§I.8). For these we have already used the
modern notation f: X ..... Y for a function on the domain X to a codomain
Y, with x I-> fx E Y for each x in X. There are all sorts of such. On a
group G, the inverse is a function G ..... G, while group multiplication is a
function of two variables in G. In a metric space, the distance is a real-
valued function of pairs of points. In Boolean algebra, "intersection" and
"union" are functions of pairs of sets. In geometry, "length" is real-valued
function of curves. Historically, functions were perhaps first explicitly
recognized as functions R ..... R in the calculus, as with Newtons "ftuents"
(functions of time). Now the more general notion of an arbitrary function
X ..... Y, in parallel to the notion of a set, plays a powerful unifying role in
Mathematics.

3. What Is a Function?
The various intuitive ideas about functions and functional dependence are
helpful but vague. Here are some of them.

Formula. A function is a formula in a letter x. When x is replaced by a

number, the formula produces a number, the value of the function for the
given argument x.
This description is not very helpful for notions of functional depen-
dence of variable quantities in physics, but it does describe the elementary
Mathematical functions well: Polynomial, rational, algebraic, trig-
onometric, and exponential functions. This ideas of "formula" does need
to be specified. It should include algebraic or "analytic" formulas, perhaps
also the formulas given by infinite series, but it doesn't seem to encompass
functions defined by several different formulas in different portions of the
3. What Is a Function? 127

domain. The essential problem remains: What sorts of formulas are

envisaged? Are all functions given by formulas? "Formulas" depend on
the symbolism, but functions depend upon the facts.

Rule. The variable y is a function of the variable x when there is given a

rule which to each value of x produces the corresponding value of y.
This description, and its variants, has for generations puzzled the stu-
dents of calculus. It has a pleasant generality: Any "rule" will do. Also,
the use of rules clearly takes care of the case of functions defined by
different formulas in different portions of the domain; any such collection
of alternative formulas is clearly a rule. Nevertheless this is not a formal
definition, since it uses the undefined words "correspond" and "rule".
Even if one defines a rule as something expressed in a specified formal
language, there are troubles. (In the usual formal languages, the set of all
formal expressions is denumerable, while the set of possible functions on
R to R is not denumerable.)

Graph. A function is a curve in the (x,y) plane, such that each vertical line
x = a meets the curve in at most one point with coordinates (a,b). When
it does so meet, the number b is the value of the function at the argument
a. For other arguments a, the function is undefined.
This description emphasizes the geometric aspect. It is persuasive for
functions of real numbers which are smooth or at least continuous, but it
doesn't fit well with a function which jumps from 0 to 1 as the variable
changes from rational to irrational. It involves also the (undefined) notion
of a curve, and so makes arithmetic depend upon geometry.

Dependence. The variable quantity y is a function of the quantity x if and

only if a determination of the value of x also fixes the value y, so that y
depends on x.
This, a physicist's definition, is also not formal.

Table of Values. A function is determined by a table of values, which

opposite each entry for the first quantity x lists the corresponding numeri-
cal value for the second quantity y.
This is a hard-nosed, no nonsense definition; evidently inspired by
tables of trigonometric or logarithmic functions. Trouble is, the actual
tables are finite while most of the intended functions have infinitely many
different values. It is not clear what would be meant by an infinite table.

Syntax. A function f on the set X to the set Y is a symbol f such that

whenever the term x stands for an element of X, then the string of
symbols fx stands for an element of Y, the value of f at the argument x.
This doesn't really describe functions, but just the use of symbols for
functions. It is a mute protest against the confusion of standard notations,
in which f(x) ambiguously denotes a function of x and a value of that
128 V. Functions, Transformations, and Groups

function. (Thus, strictly speaking, "sin" is the trigonometric function,

while "sinO" denotes its value at the angle 0.)
Enough. The variety of these descriptions of "function" illustrate well
one of our theses as to the nature of Mathematics: Human activities and
facts about phenomena together indicate many examples of dependence
of one item upon another. This leads to useful but informal ideas about
dependence and functions-and poses the problem of producing a formal
definition, necessary to make unambiguous Mathematical statements
about functions.

4. Functions as Sets of Pairs

A formal definition of "function" must be stated in the context of some
axiomatic system. There are at present two such definitions. One of these
directly axiomatizes the notion of function in terms of the composition of
functions (categories, in Chapter XI). The other operates in terms of
axioms for sets, to be given in full in Chapter XI. These axioms assume
that everything (in Mathematics) is a set, and are formulated in terms of
one primitive notion, membership, written x E A for "x is a member of
(i.e., an element ot) the set A". The equality of sets is then defined by (cf.
(1.9.8)

A = B ~ for all x, (x E A ~ x E B), (1)

where the double headed double arrow ~ is short for "if and only
if'. An axiom (the axiom of extensionality) then requires that equals can
be substituted for equals:

A = B ~ for all C, (A E C ~ B E C); (2)

here the double arrow ~ is short for "implies". Another axiom asserts
that for any two sets a, b there is a set {a,b} such that

x E {a,b} ~ (x = aorx = b); (3)

in other words, {a,b} is the set whose only elements are a and b. In partic-
ular, {a,a} is the set whose only element is the set a; it is usually written
{a}. On this basis we can define the ordered pair of two sets a, b to be the
set

<a,b> = {{a}, {a,b}}. (4)

This is a set whose elements are sets of sets. One can then prove that

<a,b> = <a',b'> ~ (a = a' and b = b'), (5)

4. Functions as Sets of Pairs 129

which states that this set <a,b> has the property one would expect of an
ordered pair of any two things. Other axioms prove that, given two sets
X and Y, there exists a set

X X Y= {<x,y> I x E X, Y E Y} (6)

whose elements are exactly all the ordered pairs of elements x EX and
y E Y. This cartesian product of sets has already been used, for example
in the description of the coordinate plane as the cartesian product R X R.
We also use the formal definition of a subset SeX: A set such that
XES implies sEX.

Definition. A function f on the set X to the set Y is a set SeX X Y of

ordered pairs which to each x E X contains exactly one ordered pair
<x,y> with first component x. The second component of this pair is the
value of the function f at the argument x, written f(x). We call X the
domain and Y the codomain of the function f.

This provides a formal definition which, in plausible ways, does match

the intent of the various preformal descriptions of a function. Specifically,
it does provide y, depending on x. If one imagines a list of all the ordered
pairs <x],y] >, <X2,Y2 >, <X3,Y3 >, ... in the set S, then this list is
just the (usually infinite) table of values of the function. If one visualizes
the sets X and Y as "spaces" of some sort, then the cartesian product
X X Y is a "space" (in Figure I a cylinder) and the function S is a subset
of the product which meets each "vertical" subspace {x} X Y in exactly
one point. Thus S is a curve in this product space. In view of such exam-
ples, the set S of ordered pairs is often called the graph of the function-
though in our definition, the function is its graph.
From this formal definition (and the axioms of set theory) one can
derive all the formal properties of functions. For example the definition of
equality for sets provides an equality condition for functions. Two func-
tions f,g: X -+ Y with the same domain and the same codomain are equal

y s

C)x
Figure 1
130 V. Functions, Transformations, and Groups

(as sets of ordered pairs and thus as functions) if and only if, for all
x E X, f(x) = g(x) in Y.
Our examples of functions have involved many instances of functions
constructed by composition, such as (2x + 2)2 or cos 7(). There is a
corresponding formal definition. Given any two functions f and g, when
the codomain of f is the domain of g, as in
f g
X ~ Y ~ X,

one can define the composite function h = g. f to be the following set of

pairs

{<x,z> Ix EX, Z E Z, :3y E Ywith <x,y> E fand <y,z> E g}. (7)

Here j y is a (logical) quantifier meaning "there exists a y"; moreover we

have identified (as we should) f with "its" set S of ordered pairs. One
readily shows that this set (7) of ordered pairs is indeed a function
g·f: X ->Z and that its value, for each argument x E X, is

(g·f)(x) = g(j(x», x EX. (8)

Here the usual convention that f(x) means "apply the function f to the
argument x" has the consequence that the composite function g' f means
"first apply f, then apply g". The definition of the composite function g. f
is then equivalent to the statement that (8) holds, and that

domain(g·f) = domain f, codomain(g· f) = codomain g. (9)

This description of composite functions does match the examples such as

cos 7(}. It illustrates the fact that the formal definition of functions by
ordered pairs is needed only at the start, and soon is replaced by
equivalent (derived) statements like (8) and (9) in more conventional
language, not explicitly using ordered pairs. For example, the formula (8)
suffices to prove associativity, a central property of composition: Given
three compos able functions
f g h
X ~ Y ~ Z ~ W,

one necessarily has h·(g·j) = (h·g)·f.

In defining the composite g. f we have required that the domain of g be
exactly the codomain of f. In this view, a function f involves all of these
things: A domain, all the pairs <x,Jx>, and the codomain. The set of
pairs determines the domain, but does not determine the codomain.
Changing the codomain changes the function. Thus if R + is the set of
non-negative reals, the assignment x ~ x 2 defines a function R ->R + and
a different function R ->R with the same values but a larger codomain.
4. Functions as Sets of Pairs 131

This convention (that a function carries with it a specified codomain) is

usually not made in elementary Mathematics. However, it is useful in
keeping composites in order and it is essential for some concepts in topol-
ogy (see §XI.9 below). Also it is essential to make sense of statements such
as "this function is a bijection". For example, the set of all pairs
<a,a +2 > for a an integer determines a bijection Z ..... Z and also gives a
function Z ..... Q which is not a bijection.
Formally, this definition of function is fine-but it does appear to rest
upon a possibly quite artificial notion of an ordered pair. Also, why
should an ordered pair be construed to be a set? However, with this con-
struction the ordered pair (4) has exactly the formal property (5) one
might require of such a pair. The crucial observation is that this formal
definition does indeed provide for all the intended practical properties of
functions. In particular, two functions are equal if and only if they have
the same domain, the same codomain, and the same values.
The definition (7) of the composition of functions does include the com-
position of bijections, as this was defined in Chapter I at the hand of
several examples. For economy of presentation, we should have defined
"composition" just once. However, the repetition of the definition may
serve to make a philosophical point: The definition of composition for-
malizes an idea ("carry out two operations in succession") which has
arisen from a wide variety of activities. These activities include the succes-
sive comparison of sets for size (§I.2), the "operation" of composing two
permutations, the essentially "static" compositions of symmetries (§I.6)
and the more "dynamic" composition of motions (§I.5). Moreover, there
is the composition of arithmetic operations by substitution, as in (1.2).
Composites such as these, as presented by formulas (e.g. sin 80), are
perhaps the most elementary, since they can be handled by rules for
manipulating formulas in ways which require no formal definition of
"composite". On the other hand, the composition of motions is more
geometric and hence more "visible". The philosophical point is that all
these varied human activities fall-quite naturally and easily-under the
one general notion of "composition", and the formulation of this general
notion helps to organize and understand the various cases. This, we hold,
is typical of Mathematical abstraction.
Binary relations, such as linear order, inclusion, and congruence rela-
tions, provide another example of a set-theoretic formalization of a practi-
cal notion. Thus in ordinary language the phrases "brother of' or "larger
than" provide two-place relations which appear in context in phrases such
as "A is the brother of B" or "s is larger than t". There is also a notion of
composite relation; for example, "uncle of' means "brother of a parent
of'.
The composition of functions is a special case of the composition of
relations. A (binary) relation from the set X to the set Y is a subset
ReX X Y; the intention is that it consist of those ordered pairs
132 V. Functions, Transformations, and Groups

<x,y> of elements x E X and y E Y which stand in the intended rela-

tion to each other. One may often write yRx for <x,y> E R. If
S c Y X Z is a second relation from Y to Z, the composite relation S·R
is defined to consist of those ordered pairs <x,z> for which there exists
at least one element y E Y such that <x,y> E Rand <y,z> E S.
This composition is associative. One may develop a consequential algebra
of relations extending Boolean algebra. There are many Mathematical
examples of binary relations, such as the inclusion relation c for subsets
of a set and the linear order relation for real numbers. There are also
cogent philosophical arguments for the importance of relations. (One can-
not reduce all relations between objects x and y to conjunctions of
separate properties of x and properties of y.) Nevertheless, despite various
attempts, the properties of relations do not seem to playa central role in
Mathematics; instead, relations are often replaced by operations or func-
tions. For example, congruence is replaced by congruence classes.
There are a few useful general facts about functions on arbitrary sets.
They all have been suggested by specific examples. Thus a function
f: X ~ Y may have as "inverse" some function g from Y to X-just as
square root is inverse to square and log inverse to exponential. To define
this formally, first introduce for each set X the identity function Ix: X ~X
which sends each element x E X to itself. A more vivid description states
that the graph of the identity function is the diagonal subset ~ of X X X,
consisting of all the ordered pairs <x,x> for x EX. If two functions
f: X ~ Y and g: Y ~X have composite g·f = Ix the identity, then the
function g is said to be a left inverse of j, while j is called a right inverse
of g. Thus the definition of a logarithm, IOlogl()x = x, rewritten

expIO(logIO x) = x, x > 0,

states that log: {x I x > O} ~ R is a right inverse of exp (it is also a left
inverse). The operation "take the positive square root" is a function
{x I x > O} ~R. The equation x = (Yx"y means that "square" is its
left inverse-but not its right inverse. In general again, g' f = Ix and
f·g = ly together mean that g is a (two-sided) inverse of f. When
f: X ~ Y has such an inverse g, that inverse is unique; moreover the sets
X and Y have the same size (formally, have the same cardinal number).
When any f: X ~ Y is regarded as a comparison of the set X to the set
Y one naturally may ask when f hits all elements of Y (is "surjective") or
when f keeps distinct elements of X distinct (is "injective"). Formally, a
function f: X ~ Y is called an injection when Xl =1= X2 in X implies
jx] =1= jX2 in Y; intuitively, an injection maps the set X in one-to-one
fashion onto some subset of Y, called the image of f. Unless X is the
empty set </>, a function f with domain X is an injection if and only it has
a left inverse g; specifically, g takes each element of the image back to the
necessarily unique element of X from which it came, and the remaining
elements of Y (if any) back to any old element of the (non-empty!) set X.
5. Transformation Groups 133

A function f: X ...... Y is a surjection when to each y EO Y there is at least

one element x EO X with fx = y. The axiom of choice, commonly
assumed among the axioms of set theory, asserts that any surjection f with
a non-empty domain X has a right inverse g. Specifically, such a right
inverse "chooses" to each y EO Y an element x EO X with fx = y, and
the axiom of choice is there to provide for such a potentially infinite col-
lection of choices.
Finally, a function f: X ...... Y is a bijection if and only if it is both an
injection and a surjection; that is, if and only if there exists to each y EO Y
exactly one x EO X with fx = y. Thus a bijection establishes a one-to-
one correspondence between the elements of X and those of Y. A function
f is a bijection if and only if it has a two sided inverse.
One may note that these definitions are formal, but make very little use
of the ordered pairs used in the formal definition of a function.
A function f: X ...... Y which is not injective "collapses" some of the ele-
ments of X, and the collapse may be formalized by a binary relation E1
"has the same f-image as" on the domain X. Specifically, for x and y in
X, take xE1y to mean fx = fy; the relation EI may be called the "ker-
nel" or the "kernel pair" of f. It is reflexive, symmetric, and transitive. A
reflexive, symmetric, and transitive relation E on a set X is usually called
an equivalence relation or a congruence relation; the prime example is
congruence modulo m on the set Z of integers. In that case, congruence is
the kernel of the projection Z ...... Zm to the integers modulo m. By the same
construction, any equivalence relation E on a set X is the kernel of a sur-
jection p: X ...... Y: one simply takes p(x) = {x' I x' EO X and xEx '}, and
Y to be the set of all these equivalences classes. Each element x' in X
belongs to exactly one of these equivalence classes, while p turns the
equivalence relation E in X into equality in Y. This important construc-
tion has already been used to construct cardinal numbers, ordinal
numbers, the set of "directions" in the plane (§III.9) and the Cauchy real
numbers. It is a handy set-theoretic way of specifying new Mathematical
objects.
This section has indicated how the notion of a function, originally sug-
gested by formulas and functional dependence, can be formalized and
applied to sets quite different from sets of numbers, and so provide a set-
ting for the comparison and mapping of arbitrary sets.

5. Transformation Groups
The analysis of the symmetry of figures, of formulas, of ornaments or of
crystals leads inevitably, as in Chapter I, to the study of transformation
groups. Such groups also arise in geometry, as groups of rigid motions,
and can even provide a foundation for Euclidean geometry (§IIl.lO). In
this section we will note additional examples of transformation groups.
134 V. Functions, Transformations, and Groups

The symmetric group Sn on n letters consists of all the n! permutations

of the set {1,2, ... , n} of the first n positive integers; any other set of n
elements has the same symmetry group. Every permutation can be written
as a composite of transpositions, where a transposition interchanges two
integers and leaves all others unchanged. A permutation is even if and
only if it can be expressed as a product of an even number of transposi-
tions. Within Sn the even permutations form by themselves a transforma-
tion group An with n!!2 elements, called the alternating group, a subgroup
of Sn. For example, all the permutations of four letters XI,X2,X3,X4 which
do not change the sign of the product

constitute the alternating group A 4 • For an equilateral triangle, each sym-

metry is determined by the permutation induced on the vertices of the tri-
angle, so the symmetry group of the triangle is isomorphic to S3' For
analogous reasons, the symmetry group of an equilateral tetrahedron is
isomorphic to S4 while the alternating group A5 appears as the symmetry
group of the icosadedeon. In general, finite symmetry groups can always
be expressed as subgroups of some permutation group Sn.
Linear algebra is replete with transformation groups, beginning with the
group of translations of the line, which is isomorphic to the additive group
of real numbers. A linear transformation t: V ~ V of a vector space into
itself is said to be non-Singular if and only if it is a bijection. The set of all
such non-singular transformations on V is clearly a transformation group,
called the general linear group GL( V). When V is the space F n of n-
tuples of elements of a field F, such a non-singular transformation is
represented, as in §2 above, by an n X n matrix with entries in F and
determinant =1= O. The group of all these non-singular matrices is written
GLn(F); especially prominent are the groups GLn(R) and GLn(C), The
subgroup of those real matrices with determinant I is called SLn (R). In a
vector space over R with an inner product, the linear transformations
which preserve the inner product are called orthogonal, and constitute the
orthogonal group On = On (R). The matrices representing such an orthog-
onal transformation must have determinant ± I; those with determinant
+ I preserve orientation and constitute the special orthogonal group SOn.
There are corresponding constructions for complex numbers. In an affine
space A, the bijections t: A ~A which preserve weighted averages (that is,
satisfy t(~WiPi) = ~WJPi for weighted averages of points Pi, expressed as
in (IV.7.4», are called affine transformations. In this way each type of
geometry can be summarized by the transformation group of all those
bijections of the space which preserve the intended geometric structure.
This was the central observation of Felix Klein's Erlanger Program for
geometry (1872). Two figures count as "equivalent" in a geometry when
there is a transformation of the group carrying the first figure into the
6. Groups 135

second. Thus in the Euclidean plane two triangles are equivalent in this
sense if and only if they are congruent, in the oriented Euclidean plane
two triangles are equivalent if and only if they are congruent and have the
same orientation, while in the affine plane any two non-degenerate trian-
gles are equivalent.
This description of the Erlanger program fits all manner of geometries.
Thus topology involves for each topological space X the group of all those
bijections t: X ..... X which are continuous with a continuous inverse. More
generally, a homeomorphism t: X ..... Y of topological spaces is a continuous
bijection with a continuous inverse. Thus the surface of a sphere is not
homeomorphic (why?) to the torus (the surface of a doughnut), but is
homeomorphic to the surface of an ellipsoid or of a cube. However the
latter homeomorphism cannot be "smooth" (at the corners of the cube);
this suggests still another kind of geometry, that of eX! structure, to be
discussed in Chapter VIII.
All told, the initial and quite intuitive observations of symmetry, of
motion, and of transformations in geometry and of their composition
leads to the formal set-theoretic notion of a transformation group. Its
study involves algebra, geometry, and continuity.

6. Groups
The definition of a group G by axioms, as given in §I.8, arises by abstrac-
tion from the notion of a transformation group-one ignores the fact that
the elements are transformations of something but retains the composition
operation, the identity, and the inverse and requires the associative law
(valid automatically in case the group elements are indeed transforma-
tions). No further axioms are needed to formalize the properties of com-
position, because these axioms suffice to prove the Cayley theorem that
every (abstract) group G is isomorphic to a transformation group G'.
Specifically, each element g E G can be regarded as a transformation
g': x f-> gx of the elements x of the set G. By the associative law,
k(gx) = (hg )x, so the composition hg of group elements matches exactly
the composition h 'g' of the corresponding transformations; thus the bijec-
tion g f-> g' is an isomorphism of the abstract group G to the transforma-
tion group G'. It is called the (left) regular representation of G, because
each element g of G is "represented" by the operation "multiply on the
left by g".
Other groups arise not as groups of transformations but as groups of
numbers under multiplication (or addition; cf. §J.8). Most of these groups
are abelian, but algebraic considerations also yield examples of non-
abelian groups, such as GLn and the multiplicative group of non-zero
quaternions (§IV.14). From given groups one may construct new groups
as products or as semidirect products (§8). One may also construct groups
136 V. Functions, Transformations, and Groups

ad hoc by listing the elements and writing a suitable multiplication table.

Indeed, the richness of group theory may be ascribed in part to its multi-
ple origins.
The simplest (multiplicative) groups G are those "cyclic" groups gen-
erated by a single element a, and so consisting just of the identity and
powers a,a -1,a 2,a -2, ... of that element. The order of an element a is by
definition the least positive integer m with am = I, the identity element.
This order may be infinite, in which case the cyclic group is infinite-and
isomorphic to the additive group Z. The order m may be finite, in which
case the (cyclic) group is isomorphic to the group of all rotations of a reg-
ular m-gon into itself and to the additive group Zm of integers modulo m.
In any finite group, the order of each element is a divisor of the order
(number of elements) of the whole group. Much of the structure of a par-
ticular group may be revealed by examining the lattice of its subgroups
and their inclusion relations-where a subset S of a group G is a subgroup
if and only if it is non-empty and closed under composition and inverse.
The comparison of two (multiplicative) groups G and H is made by a
function or "morphism" preserving the group structure. Specifically, as in
§I.8, a homomorphism t: G-'>H of groups is a function t such that
t(glg2) = (tgl )(tg2) for all pairs of elements gbg2 E G. It follows also
that t(g -I) = (tg) -I for all g and that t(l) = 1. When first studied,
homomorphisms were taken to be always surjective (onto the codomain
H) but it is more convenient to allow the codomain to be larger than the
image. In particular, when S is a subgroup of G, the inclusion S c G can
be considered as a function from S to G which is a homomorphism. In
general, a homomorphism t is called a monomorphism when t is an injec-
tion, an epimorphism when t is a surjection, and an isomorphism when a
bijection. A homomorphism t: G-'>G of G to itself is an endomorphism,
and an automorphism when it is also a bijection. In fact, the set Aut G of
all automorphisms of G is itself a transformation group; it expresses the
symmetry of G. The mono-, epi-, endo-, and auto-terminology applies also
to "morphisms" of other types of algebraic systems such as rings, fields, or
monoids-but any homomorphism of fields is necessarily a monomor-
phism.
Homomorphisms arise naturally for transformation groups. For exam-
ple, the group Q of the square has 8 elements (four rotations, four
reflections). Each symmetry t E Q induces a permutation t' of the diago-
nals; the permutations of the (two) diagonals constitute the symmetric
group S2, and the correspondence t I--> t' is an epimorphism Q -'> S2.
Again each symmetry t E Q must either leave the square right side up or
tum it over; in other words, t determines a permutation t" of the two
faces of the square, and this correspondence t I--> t" is an epimorphism
Q-,>S2-different from the previous epimorphism! More elaborate exam-
ples can be constructed from the group of symmetries of a cube.
A homomorphism t: G-'>GLn(F) of a group into a general linear group
is called a representation of G. The systematic study of such representa-
6. Groups 137

tions reveals many deeper properties of groups. They include the

representation of G by permutations, since any permutation of the sym-
metric group Sn can be regarded as a permutation of the n unit vectors of
the vector space P, and so can be represented (isomorphically) as an
n X n matrix with entries zero except for one entry I in each row and
each column.
For each element k of a group G, the operation g f-+ kgk -I is called
conjugation by k. It is an automorphism k.: G~G, called an inner auto-
morphism. Moreover, the correspondence k~k. is itself a homomorphism
G~Aut G -trivial if G is abelian. For a transformation group G, conjuga-
tion by k has "geometric" meaning; for instance (§III.9), it carries the
subgroup fixing one point x isomorphically onto the subgroup fixing the
point k(x); cf. (III.9.3) and (111.9.4).
Any homomorphism maps subgroups to subgroups. A subgroup N of a
group G is called a normal subgroup if and only if it is mapped onto itself
by every inner automorphism k. of G; that is, if and only if n E Nand
kEG imply knk- I E N. For any homomorphism t: G~H, the image is
the subset (subgroup) of all elements t(g) for g E G and the kernel K is
the set of all those nEG with t(n) = I in H. One proves readily that the
kernel K is always a normal subgroup of G. Moreover every normal sub-
group N is the kernel of an epimorphism. This basic result we state in the
correct (but not customary) conceptual form, as in our book Algebra:

Theorem. For any normal subgroup N of G there is a group H and an epi-

morphism t: G ~ H with kernel exactly N. If s: G ~ L is any group
homomorphism with kernel containing N, there is a unique homomorphism
S': H~L with composite S'·t = s, as in the commutative diagram

G • H

~ is'
s I
I
t
(1)

One says that s I is induced by s. The group H may be constructed

explicitly, for example by taking as its elements t(g) the sets gN of all mul-
tiples gn of g by an element n E N. Such a set gN is called a coset, as in
§III.9; observe that g and gm for m in N give the same coset. To make the
set H of all cosets into a group, their product is defined by
(gl N)(g2 N) = (gl g2 )N. Then g f-+ gN is an epimorphism to the group
of cosets, and it follows that any s as stated does factor uniquely through
t, as in the commutative diagram (1). Because of this factorization, one
says that t is universal among the homomorphisms from G with kernel
containing N.
The explicit description of the elements of H as cosets is irrelevant,
since the group H is uniquely determined (up to an isomorphism) by its
universal property-for any other homomorphism s: G~H' with the same
138 V. Functions, Transformations, and Groups

universal property, the diagram (I) produces an isomorphism s': H == H'.

This group H, constructed by cosets or otherwise, is called the factor group
(quotient group) G/ N. A very simple example is the subgroup (n) of all
multiples of an integer n in the additive group Z; the factor group is the
additive group Z/(n) of congruence classes modulo n, because these
congruence classes are exactly the (additive) cosets of (n). Again, A4 is a
normal subgroup of the symmetric group S4, and S4/A4 == S2. In the
oriented plane, the group of translations is a normal subgroup of the
group of all proper rigid motions (cf. (III.7.1» and the corresponding quo-
tient group is isomorphic to the group of rotations about a point. Such
examples, plus others from Galois theory (§7), led to the general
construction-and manipulation-of factor groups. Contrary to common
custom, they are not best understood in terms of cosets.

7. Galois Theory
Symmetry and its formalization by transformation groups arises not only
in geometric situations, but also in purely algebraic cases. We have
already noted a first example, in the operation C ---C of conjugation
x + iy ~ x - iy for complex numbers. This operation interchanges the
two complex cube roots wand w2 of 1. Conjugation leaves only the real
numbers fixed, and can be viewed geometrically as a reflection of the
complex plane in the real axis. As a symmetry group, it consists of just
two transformations: conjugation and the identity. It arises from the equa-
tion x 2 + 1 = 0, since it interchanges the two roots of this equation.
Now consider a polynomial equation of degree n such as

an '1= O. (1)
We will assume that the coefficients an,an-l , ... ,ao lie in some subfield
F of the field C of complex numbers, and that the polynomial f is irredu-
cible over F-that is, cannot be factored in F as f( x) = g( x )h( x) into two
polynomials g and h of lower degree.
Galois theory is concerned with formulas for or properties of the roots
of f(x) = O. By the fundamental theory of algebra, there is at least one
complex root al and hence a corresponding factorization of f(x) as
f(x) = (x - al)g(x), where the second factor g(x) is a polynomial (of
degree n - I) with complex coefficients. This polynomial in its turn has a
complex root a2 so continuation of this process ultimately gives n roots
and a factorization
(2)

From the fact that f is irreducible, one may prove that these n roots
an are all different.
al , . . . ,
7. Galois Theory 139

For n = 2, the familiar process of "completing" the square produces

formulas involving a square root for the two roots 0'1 and 0'2. For polyno-
mials of degrees 3 and 4 there are similar but more complicated (and less
useful) such formulas, involving cube roots and fourth roots (Survey,
§V.5). The Galois theory will reveal why there can be no corresponding
formulas valid for the roots of all polynomials f of degree 5-or higher.
To begin with, shift attention from the n roots 0'1, . . . ,O'n to the set of all
complex numbers which can be obtained by rational operations (addition,
subtraction, multiplication, and division) from these roots and the ele-
ments of the given field F. All these numbers constitute a sub field

of the field of complex numbers. It is called a splitting field for the poly-
nomial f over the base field F, because it is a smallest field containing F
in which the polynomial f "splits" into linear factors. In fact (forgetting
the complex numbers) this property determines the field F( 0'1, . . . , O'n)
up to an isomorphism leaving the elements of F all fixed. In the case of
the polynomial x 2 + lover R, the splitting field is just the field C of
complex numbers.
Now consider the symmetries of the splitting field N relative to F. A
symmetry is by definition a transformation t: N -> N which is an automor-
phism of fields (i.e., a bijection which preserves sums and products) and
which leaves fixed all the elements of the base field F. All such sym-
metries constitute a transformation group G, the Galois group of the split-
ting field N-and of the polynomial f -over the base field F. Since each
symmetry t leaves all the coefficients of f fixed, it must carry any root 0'
of f(x) = 0 into another root of this polynomial. Hence such a symmetry
induces a permutation of the roots 0'1, . . . ,O'n of f; since N is generated
by these roots, the symmetry is determined by the permutation, so that G
is isomorphic to a subgroup of the symmetric group Sn. This means in
particular that the Galois group G is finite. It can be described either as a
group of automorphisms of N of as a group of permutations of the roots.
Linear algebra provides a more specific measure of the size of the
Galois group. If we neglect multiplication in the splitting field, then addi-
tion, plus multiplication by elements of F, still make N a vector space over
the field F-and the dimension of this vector space is equal to the order of
the Galois group. For algebraic reasons, this dimension is called the
degree [N:Fj of N over F. (This is a striking simple case of the use of
geometric dimensions in algebra.)
For examfle, the easy equation x 3 - 5 = 0 over the rationals has
three roots: Vs, w3 Vs, and w2 3 Vs, where the first root is the real one,
while w is a complex cube root of unity. The whole splitting field over Q
then can be built up in two steps from Q-first the real root, then the
others-as
140 V. Functions, Transformations, and Groups

It has degree 6 over Q and has a subfield QeVS) of degree 3 over Q, as

well as another sub field Q( w) of degree 2.
For any polynomial equation f(x) = 0 over F any solution of the
equation by "radicals" (roots of equations x P - a, say with p a prime) is
found to involve a process of building up the splitting field N by stages,
with intermediate fields K. Here enters the fundamental theorem of
Galois theory, which establishes a bijection between the intermediate
fields K, with N ::::> K ::::> F, and the subgroups S of the Galois group G.
Each intermediate field K determines a subgroup, call it K# :

K# = {all t E G I for all b in K, tb = b}.

Each subgroup S of G determines an intermediate field, call it Sb

Sb = {all bEN I for all s in S, sb = b}.

A subtle argument (Algebra or Survey) then shows that these two assign-
ments K I-> K# and S I-> Sb are mutually inverse functions and so pro-
vide the asserted bijection from intermediate fields to subgroups.
In this bijection, the whole field N is left fixed only by the identity auto-
morphism, so N# = 1 is the one element subgroup of the Galois group
G. We may then picture the correspondence briefly as follows

u n
S = K# = Gal( N:K)

u n
G = F# = Gal(N:F).

Note that a larger subgroup S will leave fewer elements b fixed, so

corresponds to a smaller subfield, hence the subgroups appear "upside
down" in the diagram. Also the subgroup K# = S consists of all the
automorphisms of N which leave K pointwise fixed; it is just the Galois
group of N over K, as indicated in the figure. Also one can prove that
S = K# is a normal subgroup of G if and only if K is itself a splitting
field over F; in this case the factor group G / S is exactly the Galois group
of Kover F. Factor groups here enter essentially.
The bijection between intermediate fields and subgroups has remark-
able consequences. First the Galois group G, as a permutation group, is
finite. Hence it has only a finite number of subgroups. Consequence:
There are only a finite number of intermediate fields! (For x 3 - 5 there
are four such.)
7. Galois Theory 141

Next, building up the splitting field N in stages amounts to the con-

sideration of the corresponding chains of subgroups

G = Go :J GI :J G2 :J ... :J Gm - I :J Gm =I (3)

of a group G. If each successive group Gk in this chain is a normal sub-

group of its predecessor Gk- I and if each quotient group Gk- I IGk is sim-
ple (§9), for k = I, ... , m, this chain is called a composition series for G:
it exhibits G in pieces Gk-I/Gk. Clearly any finite group G has such a
composition series: Just choose G1 to be a maximal proper normal sub-
group of G, and continue down. There can be many different such compo-
sition series, but the Jordan-Holder theorem (with a perspicuous proof,
Algebra, p. 430) asserts that any two such series for a finite group G have
the same length of m and, up to order and isomorphism, the same set of
factor groups

called the composition factors of G. A group is said to be solvable when

these factors are all cyclic groups.
This name comes from the striking theorem which states that a polyno-
mial f over F has roots given by formulas involving pth radicals (for
various prime numbers p) if and only if the Galois group of f over F is
solvable. The essential reason is this: When the field F already contains all
the pth roots of unity, the Galois group over F of an irreducible equation
x P - a is a cyclic group of order p. Since pth roots of unity also lead to
cyclic Galois groups, this means that fields built up by radicals must have
solvable Galois groups.
The most general equation of 5th degree will have as Galois group the
whole symmetric group Ss. A composition series for this group reads

Ss :J As :::J I,

where As is the alternating group of all even permutations; the quotient

group Ss lAs is cyclic of order 2, and the alternating subgroup As can be
shown, by direct calculation, to be a simple group. This is the group-
theoretic reason for the in solvability of the quintic equation by radicals.
In the hands of Evariste Galois, this Galois theory was the occasion for
the explicit discovery of the uses of groups, normal subgroups, and factor
groups. Note in particular that the use of factor groups is essential-just
the axiomatic description of symmetries by a "group" is not enough!
This short sketch has omitted many interesting aspects of the theory.
The splitting fields can be characterized directly, independent of anyone
polynomial which is split. The idea also works for splitting fields of
infinite degree, provided however that the Galois group is then considered
as a topological group, while its closed subgroups corresponding to inter-
mediate fields. It applies to finite fields, and provide an easy proof that for
142 V. Functions, Transformations, and Groups

each prime p and each natural numbers n there is up to isomorphism

exactly one finite field with pn elements-and this field is the splitting field
of the polynomial x P• - x over the field Z/(p) of the integers modulo the
prime p. However, the Galois theory does not apply without restrictions
to infinite extensions of such fields (called fields of characteristic p,
because the p-fold sum 1 + 1 + ... + 1 is zero in these fields).
The Galois theory is a decisive demonstration of the utility of group-
theoretic and abstract methods: Practical formulas for solving equations
are best understood by groups of symmetries, their subgroups and factor
groups. The Galois theory is an example of a piece of Mathematics which
is deep but not difficult. Originally, however, it was obscure; it was not
understood by the contemporaries of Galois, and some expositions pub-
lished as late as 1925 were wholly confusing. It became really perspicuous
only when the advent of modern abstract methods in algebra, in the
hands of Richard Dedekind, Emmy Noether, B. L. van der Warden, and
Emil Artin, yielded the present elegant and conceptual proofs of the basic
theorem of Galois theory (see for example, Artin [1949] or the treatments
in Survey [1977] or in Algebra, second edition; [1979]). A text by Kaplan-
sky [1972] provides a wealth of examples.

8. Constructions of Groups
The remarkable richness of group theory rests in part on the extensive
variety of specific groups, especially finite groups. Initially, Mathemati-
cians wanted to know all groups, and so were tempted to try to list all the
possible finite groups of a given order, of course counting isomorphic
groups as the same. Thus, there are two different groups of order 4, the
cyclic group 24 and the group of 4 symmetries of a rectangle (the so-
called four group). In a group of order 5, any non-identity element must
have order 5, so any such group is (isomorphic to) the cyclic group Z5.
There are two different (i.e., non-isomorphic) groups of order 6, the cyclic
group Z6 and the symmetric group S3, and only one group Z7 of order 7
because 7 is prime. The continuation of such a simplistic catalog, how-
ever, soon becomes cumbersome and is not very enlightening. Too much
depends on the arithmetic properties of the order of the group.
Instead, long experience with examples has developed more useful con-
structions.
If a prime p divides the order n of a group G, as n = pen' with n' rela-
tively prime to p, then G has at least one subgroup P of order pe, and any
two such subgroups are conjugate. These subgroups P are called the
Sylow subgroups of G; one shows that any subgroup S of G with order
some power of the prime p is necessarily contained in a Sylow subgroup P
(i.e., one of order pe). The investigation of these subgroups (Algebra,
8. Constructions of Groups 143

chap. XII) is a first step in the examination of the arithmetic influence of

the order of a group on its structure.
New groups may be constructed, in various ways, from given groups. If
A and G are given groups, their (direct) product A X G consists of the
ordered pairs (a,g) of elements a E A and g E G with termwise multipli-
cation, as in

(a,g)(b,h) = (ab,gh). (1)

The functions (a,g) f-+ a and (a,g) f-+ g provide two homomorphisms
'TTl: A X G-+A and 'TT2: A X G-+G, called projections. Given any
homomorphisms tl: H -+ A and t2: H -+ G from a group H to the factors
A and G, there is a unique group homomorphism s: H -+ G which pro-
duces t) and t2, by composition with the projections, as t) = 'TTl'S and
t2 = 'TTl'S. This is expressed by the (commutative) diagram of group
homomorphisms

(2)

This property of the projections characterizes the product A X G, up to

isomorphism; one says that 'TTl and 'TT2 are a "universal" pair of morphisms
to the pair (A,G) of groups.
The direcL product A X G of abelian groups is clearly abelian. For
abelian groups, the construction of direct products gives a satisfactory
description of finite abelian groups A. Every such group is (isomorphic to)
a direct product of a number of cyclic groups of prime power orders:
(3)

There can be various such representations of A, but the prime numbers

p) , ... ,Pk and their exponents e) , ... ,ek are the same in all; they are
"invariants" of the group A.
No such simple direct product decomposition applies to infinite abelian
groups. To be sure, the multiplicative group of positive rational numbers
is a sort of product of infinitely many infinite cyclic groups (one generated
by each prime number) but there is no such decomposition of the quo-
tient group QjZ of the additive group of rational numbers. Some ques-
tions about the structure of "large" abelian groups remain quite unsolved!
The direct product A X G contains isomorphic copies A X 1 and
1 X G of the given factors A and G; each is a normal subgroup of
A X G and each element (a, I) of the first factor commutes with every
element (l,g) of the second. This is not the case with the free product
144 V. Functions, Transformations, and Groups

A * G. Its elements are all possible words algi a2g1 ... angn which are for-
mal products of elements ai E A and gi E G, with each ai (except possi-
bly for al) and each gj (except possibly for gn) 1= 1. Two such words are
multiplied by juxtaposition and subsequent cancellation (when possible);
thus ag multiplied by g-Ibg l becomes agg-Ibg l = (ab)g, where ab is the
given product in factor A. With some care, one may prove that this multi-
plication is indeed associative, and that a f-+a·l and g f-+ 1· g are monomor-
phisms k l : A ~A * G and k 2 : G~A *G which enjoy the following "univer-
sal" property: To any morphisms tl: A ~H and t2: G~H into an arbi-
trary group H there is exactly one morphism s: A * G ~ H which yields tl
and t2 as the composites tl = s·k l and t2 = s·k2 ; that is, which makes
the following diagram

(4)

commutative. This diagram may be obtained from the direct product

diagram (2) by simply reversing all the arrows; one says that the diagrams
are dual to each other. This diagram (4) means that A * G is the "most
general" group generated by A and G.
In the complex plane the modular group provides a striking example.
This group consists of all the fractional linear transformations

Z f-+ (az + b)/(ez + d) (5)

of the plane (z a complex number) with integral coefficients and such that
the determinant ad - be = 1. This group (which plays a considerable
role in number theoretic investigations) can also be described as the mul-
tiplicative group of all 2 X 2 matrices of integers

which are unimodular (i.e., have ad - be = 1), modulo the subgroup

consisting of the two matrices ±I, where I is the identity matrix. (The
matrix - I yield the identity transformation (5).) The group contains a
subgroup of order 2, generated by the transformation z f-+ -liz and a
subgroup of order 3, generated by z f-+ (z - 1 )Iz. It is exactly the free
product of these two cyclic groups!
The free product F = Z* Z of two infinite cyclic groups is called the
free group on two generators (namely, the generators, call them a and b,
of the two cyclic factors Z). Given any two elements g and h of any group
8. Constructions of Groups 145

H, the universal property (4) implies that three is a unique homomor-

phism t: F-.H with tea) = g and t(b) = h. Therefore any group H
which can be generated by a finite number of elements can be represented
(usually in many ways) as a factor group of a free group Z* ... *Z on a
finite number of generators. A surprising theorem asserts that Any sub-
group of a free group is free. There is a purely algebraic proof, but the rea-
sons lie in geometric facts, depending on representation of a group as a
group of paths in a space.
One may systematically construct groups from subgroups and factor
groups. Thus given groups A and G, one may seek to construct all the
groups E with A as normal subgroup and G == E IA as the corresponding
factor group. Then E is called an extension of A by G. If k is the inclusion
of A in E and '1T the projection on G, the extension can be pictured as a
sequence of group homomorphism
k 7T

1 -7 A -7 E -7 G -7 1, (6)

where 1 designates the group with just one element. This sequence is said
to be exact because at each node the image of the incoming homomor-
phism is "exactly" the kernel of the outgoing homomorphism. Thus
exactness at A means that k is a monomorphism, and exactness at E
means that the (normal) subgroup kA is the kernel of the projection '1T.
The direct product A X G yields one such extension E, but there are
many others. For instance, if E is the symmetric group Sn and A the alter-
nating subgroup An, the group G in (6) is cyclic of order 2, but Sn is not
the direct product An X Z2.
In any such sequence, the requirement that A (or k(A» be normal in G
means that each element e E E induces an inner automorphism
a l->eae- 1 onA, call it cpe. Hence cp is a homomorphism E-.Aut A. If A is
abelian (and in some other cases), A will be in the kernel of this map, so
that cp "induces" a homomorphism 0: G-.Aut A. Conversely, given A, G,
and such a homomorphism 0: G-.Aut A, sending each g E G to a I->ga,
one may construct a group E on the set A X G by the multiplication

(a,g)(b,h) = (agb,gh). (7)

Then a -.(a, 1) and (a,g) -. g does yield an exact sequence (6). This group E
is called' the semidirect product of A and G, with "operators" O.
The dihedral group t:.n is such a semidirect product. For each natural
number n, the group consists of all the 2n symmetries of a regular n-gon
t:.n (thus, for n = 3, of an equilateral triangle). Call a transformation t of
t:.n even ( + 1) if it leaves the n-gon right side up and odd ( - 1) if it turns
the n-gon over; these labels constitute an epimorphism '1T: t:.n -,Z2 to the
cyclic group ( ± 1) of order 2; its kernel is the normal subgroup consisting
of all n possible rotations of the n-gon. If R is such a rotation though 2'1T/n
146 V. Functions, Transformations, and Groups

and D any odd transformation (a reflection), the equations

R n = I, D2 = 1, and DRD- i = R n - i show that f:. n is a semidirect prod-
uct. The group of orientation preserving rigid motions of the Euclidean
plane is similarly a semidirect product, with A the normal subgroup of all
translations and G the group of rotations about a point.
In the semidirect product (7), the function g f-+ (1 ,g) is a right inverse
(1 of the projection 'IT; one says that the exact sequence (6) is split by (1.

Many other extensions are not split. In particular, when A is abelian (and
written additively) all extensions of A by G with operators () can be
described by giving a multiplication on the set A X G by the formula

(a,g)(b,h) = (a + gb + j(g,h), gh)

where the function j: G X G~A satisfies an identity sufficient to insure
that this multiplication is associative (Homology, p. Ill). This identity, in
its tum, has repercussions in the study of division rings (where the multi-
plication must be associative) and in algebraic topology (where a great
deal of information is codified in exact sequences).
Other aspects of groups enter vitally in number theory, in harmonic
analysis, in quantum theory (the eight-fold way) and elsewhere. Mackey
[1978] is a good reference.

9. Simple Groups
Any group G has itself and 1 as normal subgroups; it is simple when it has
no other normal subgroups. The Jordan-Holder theorem (§7) for compo-
sition series indicates that any finite group may be constructed as an
iterated extension of simple groups, though the order in which the simple
factors may occur in the series can vary. For this and many other reasons,
it would be useful to know all the finite simple groups. They include the
cyclic groups of prime order, the alternating groups An for n ;) 5, and
others. From an examination of other examples, W. Burnside had, about
1900, conjectured that all finite simple groups except the cyclic ones had
even order. This conjecture was proved true, with a complicated argu-
ment, by Feit and Thompson in 1962. Their methods were novel and
powerful and were then further extended, till there is now a complete list-
ing of all possible finite simple groups. This list begins with certain famil-
iar infinite families: The cyclic groups of prime order, the alternating
groups An for n ;) 5 and 16 families suggested by geometric construc-
tions; for example, one may modify the definition of the special orthogo-
nal group by replacing the field of real numbers by a finite field, so as to
get a finite group. There are then 26 simple groups which do not fit in any
such systematic infinite family. They are called the sporadic groups. For
example the largest one, called the "monster", has order the following
product of prime powers:
10. Summary: Ideas of Image and Composition 147

This number is approximately 8·ro53 . This large group has been con-
structed as a group of the transformations preserving certain geometric
structure in a space of 196,883 dimensions! This dimension, in its turn, is
closely connected with coefficients in the expansion of certain modular
functions.
These remarkable results serve to illustrate the richness of group theory.

10. Summary: Ideas of Image and Composition

The practice of mapping one thing (a point, a number, or some other
object) into some image thing arises in many connections: In the successor
operation for the natural numbers, in addition and multiplication in arith-
metic, in algebraic manipulation, in trigonometric formulas, in infinite
series, and in all sorts of geometric transformations. These experiences
lead to the notion of forming the image of a thing and the images of all
the things in some collection. In its turn, this notion has various intuitively
attractive but inexact formulations, all dealing with the way in which one
thing or quantity may "depend" on another. This in turn finally leads to
the precise but meticulous concept of a function as a suitable set of
ordered pairs of things; from this formal definition the abstract properties
of functions flow easily.
Repetition or iteration of such functions leads to the notion of
composition-one map followed by another. This notion of composition
turns out to be extraordinarily fruitful. It has a variety of formal realiza-
tions, as the composition in permutation groups, in transformation groups,
in abstract groups, in monoids, in categories, and in semigroups. In each
of these cases the formal properties include the associative law, valid for
any composition of functions.
In this chapter we have followed this development from functions and
transformations through transformation groups (geometric and otherwise)
to abstract groups and some of their remarkably rich properties. One is
led to ask why group theory, starting with such a simple system of axioms,
turns out to be so deep. The explanation may depend on several features
of the axiomatics for groups, such as the following:
(a) Multiple origins (geometric transformations, algebraic operations)
(b) Decisive uses (Galois theory; geometry; number theory)
(c) Additional applications (Quantum Theory, crystal structure)
(d) Exact representation in terms of the origins: Every abstract group is
isomorphic to (at least one) transformation group
(e) Multiple models: Many, but not too many, finite groups.
Listing these features does not provide a real explanation of the depth of
group theory, unless we could show that other axiom systems with these
148 v. Functions, Transformations, and Groups

five features also had deep consequences. This we can do only tentatively
in a few cases. Thus monoids do have properties (a) and (d), but do not
have a wealth of additional applications-and have "too many" finite
models with "too little" structure. In set theory, properties of intersection,
union, and complement can be abstracted in the axioms for a Boolean
algebra. Here (d) holds, in view of the Stone representation for Boolean
algebras, (and (a) and (c) hold as well). However, the finite models of
Boolean algebra are dull: For each natural number n, there is just one
model with 2n elements-namely, the Boolean algebra of all subsets of an
n element set. There are in Mathematics many other axiomatic notions-
but few of them enjoy all five of the properties (a}-(e) above.
The material of this chapter may be summarized in the following net-
work, which indicates, only in part, how these ideas are all interrelated.
10. Summary: Ideas of Image and Composition 149

Functions - Transformations - Groups

Arithmetic Mechanics Geometry

Counting Formula
I
Quantity Graph
/
I
Next ~ DepeJdence~ Transformation

II
Successor Composition Permutation

\
I
Addition Translate
I
Multiplication

I
1~ 1--
R R FUNCTION--I X ~ Y - { Rotate
1

Winding fcn
I
Sine, Cos

I
1
Table -....;o:..;cr.=.de:..:cr.:..,e1d:..p:...:a::;ir - sets~ ~I
Rigid rotion

Log Image - Inverse Transformation GROUP

I
Power Series
',j"UO~UOO c S,""U,' 1
I Symmetry

C,t"rny - S_'""P~· Moooid _ABSTRACT GROUP \

/
",.L, _____

GALOIS THEORY ~ Factor Group~ Normal Subgroup

HOMOM~RP~~ \ 1ISMS
A:C"

Subgroup Primes

/' com~iOO S,ri" \ Sy;),,:';:::,

proiucts/ Finitnbelia~GrolPs _______ \

SJru~4 L
Semi-Direct Products Abelian Groups Infinite Groups Finite Groups

: :~:i~::~'
Homj'0gy Group Reprerntations ~ The Monster
SOOP'i

AnalY~ Geometry~ ~ Number Theory

CHAPTER VI

Concepts of Calculus

1. Origins

Many sorts of calculations press themselves upon us. Thus, given a piece
of surface, how does one calculate its area? Or, given a section of a curve,
how does one calculate its length or the direction of its tangent line at
some point? More generally, how does one calculate the rate at which this
or that variable quantity is changing with time? The striking discovery (by
Newton and Leibniz) that there were systematic methods to calculate all
these things, and many more like them, had a major influence on the
directions and structure of Mathematics. For a considerable period, more
practical calculations of such things tended to dominate conceptual under-
standing, in a way that emphasizes the observation that Mathematics
takes its origin in human activities.
The calculation of areas began in Euclidean geometry, with formulas
for the areas of triangles and squares. Next in line was the area of a cir-
cular disc. This could be determined by inscribing and circumscribing reg-
ular polygons in the disc; as the number of sides of these polygons
increased, the area inside the circle was "pinched" between the (calcul-
able) areas inside the larger and the smaller polygon-and this pinching
process produced the area inside the disc as a sort of limit. This method,
neatly adapted to the case of the circle, suggested to Archimedes and oth-
ers a search for extensions of the method to calculate other areas-that of
an elliptical disk, and that for more irregular figures. Similarly the meas-
urements of length began with the Pythagorean theorem used to deter-
mine the length of a slanted line and hence the perimeter of any polygon.
The circumference of a circle could then be found (§III.2) through succes-
sive approximation by inscribed polygons. These and other problems of
measurement (of volumes, weights, centers of gravity, and the like)
emphasized the ubiquitous role of approximation and may have indicated
the need for a systematic understanding of the way in which such succes-
I. Origins 151

sive approximations can converge to a desired limit-an understanding

then provided in the calculus by the general definition of the integral as
the limit of a sum.
Also, how does one construct tangents? Drawing the tangent line to a
circle at some point P is hardly problematical, since the line through P
and perpendicular to the radius there is evidently the correct tangent-just
touching the circle evenly on both sides of P. For an ellipse, such as a
simple construction will succeed only for tangents perpendicular to the
major or minor axis at one end of that axis. At other points, here and on
the hyperbola, the parabola or on other curves in the plane, more sophis-
ticated methods, some known to the Greeks, are needed to draw an exact
tangent; with time, it became plausible that the tangent line to a curve at
a point P might be well approximated by the line (the secant) determined
by two points P' and P" on the curve taken close (and then closer) to P.
The need to locate such tangent lines is suggested not just by the geometr-
ical aspects of plane or spatial curves, but also by mechanical situations
where moving things fly off "on a tangent"-from a bobsled or a roller
coaster, for example. Thus the determination of tangent lines again can
suggest their calculation by the limit of some systematic sequence of
approximations.
Observations of motion raise analogous questions of approximation.
Terrestrial and heavenly bodies move about, but with velocities which
often do not stay constant. There still should be some way of calculating a
velocity for such motion-an instantaneous velocity which might, on
reflection, be measured by using as approximations suitable average velo-
cities over shorter and shorter time intervals. Similar thoughts might then
arise for other measurable quantities which vary with time, but not at a
uniform rate. In this way attention might be directed to the calculation in
general of such instantaneous rates of change-again by some sort of
scheme of successively better approximations.
Such problems can also occur for "rates" which are not time rates but
rates relative to some other changing quantity. An elementary example
might be the rate at which the area of a circle changes as the radius alters.
Decisive examples arise in Mathematical Economics. There, in analysing
the costs of producing more, it is wiser not to consider just the average
cost for the production of one widget out of a total of n such; but to use
the marginal cost-meaning roughly the additional cost of producing the
last one of those n widgets or, more conceptually, the rate at which the
cost of production changes relative to the number n of widgets produced.
This example of the use of derivatives may have developed historically
much later than the original discovery of the calculus; however, it is in-
trinsically one of the appropriate origins for the idea of a systematic cal-
culation of relative rates of change for variable quantities.
This summary description provides a conceptual (though not a
specifically historical) account of the multiple origins of the calculus.
152 VI. Concepts of Calculus

2. Integration
It is remarkable that so many different processes of approximating total
measured quantities by adding together little bits of these quantities can
all be subsumed under one process, that of integration. Area, volume,
length, pressure, moment of inertia, weight, and the like all can be
managed by such sums. To be sure, the usual formal definition of the
Riemann Integral is usually presented as the calculation of an area-
specifically the area below a curve y = f(x), above the x axis, and
bounded left and right by the ordinates x = a and x = b. This area is
broken up into the usual thin vertical and rectangular strips of width dx
running from x = a to x = b, as in Figure 1. The area of such a strip is
altitude f( x) times base d, hence is written f( x )dx and the sum of them
all-and hence the total area desired-is the definite integral

J
b
f(x)dx. (1)
a

On one reading, the width dx of each strip is an infinitesimal increment in

x (hence the notation dx), there is an infinite number of strips, and so the
integral sign (an elongated S) represents an infinite sum of infinitesimal
quantities, giving the desired area. Another reading, devoid of any such
uncertain appeal to the infinitely small, would approximate the desired
area by a finite sum of the areas of rectangles of finite width. To this end,
subdivide the interval from x = a to x = b in n intervals with successive
endpoints a = Xo < XI < X2 < ... < Xn = b; call such a subdivi-
sion 11. In the ith interval from Xi-I to Xi the function f (if continuous)
will take on a maximum Max;j and a minimum Min;j (Figure 2), and
the presumptive true area Ai under this portion of the curve will lie
between the areas of the two rectangles which have width Xi - Xi-I and
height Min;j and Max;j, respectively, as expressed by the inequalities

(Min;j)(xi - Xi-I) < Ai < (Max;j)(xi - Xi-I).

Adding up these rectangles gives for our subdivision 11 a "lower sum"

L(j,I1) and an "upper sum" U(j,I1)

y
/'
-
--r---.... 1"--_
/

x
a b

Figure I. The Riemann integral.

2. Integration 153

Figure 2

n n
L(j,a) = ~ (MinJ)(xi - Xi-I) < U(j,a) = ~ (MaxJ)(xi - Xi-I); (2)
i=1 i=1

the desired area under the whole curve must be squeezed between these
two sums. To actually express this area, we must then take a limit of these
sums for successively finer subdivisions a, so let the size of a be measured
by I a I ' the maximum of all the interval lengths I Xi - Xi-I I . When f
is continuous (on the whole interval a =< X =< b) it is then the case (as we
will soon see) that for each measure ( > 0 of approximation there is a
~ > 0 such that, whenever I a I <~, then I U(j,a) - L(j,a) I < t:.
This implies that Land U have a common limit as I a I approaches 0,

lim L(j,a) = lim U(j,a) =

I <1 I .... 0 I <1 I .... 0
ia
b
f(x)dx; (3)

this limit is the definite integral (3) of the function f over the interval
a =< X =< b. By its construction, it represents what the area under the
curve over this interval should be; if you wish, it is the definition of this
area; for those applications such as pressures and volumes it can also
serve as a definition. Such definite integrals, so defined as limits of sums,
are still written in the classical notation (3) suggesting the infinite sum of
infinitesimals; in fact that intuitive view of the matter allows us to easily
set up the integral representing all sorts of other quantities: The weight of
a thin slab of known but variable density, the water pressure on a slab-
like portion of a dam, or the volume enclosed by a surface of revolution.
Much of the instruction in elementary integral calculus consists of repeti-
tive exercises practising the formulations of such integrals. When they
cannot be done by slices or slabs, they can be managed (with more tech-
nique but in a wholly similar spirit) by multiple integrals.
This formal definition of the Riemann integral of a continuous function
replaces the intuitive idea of successive approximations by the use of lim-
its and of the standard logical quantifiers (for all ( > 0 there exist a
~ > 0 such that ... ) needed for the exact description of these limits.
Various general properties of the integral also follow directly from this
154 VI. Concepts of Calculus

definition. Linearity is an example: the definite integral from a to b of the

sum of two continuous functions is the sum of the integrals of these func-
tions separately. Again, if a < b < c, the integral of a given continuous
function from a to c is the sum of the integrals from a to b and from b to
c. This property is really using the idea of "composing" a path (of
integration) from a to b with a path from b to c-an idea already present
in group theory, and one which will recur in more elaborate cases of line
integrals.
So much for the initial formulation of measurement by integration.
However since we do not really carry out infinite sums or limits of an
infinite number of successive finite sums, the real "calculus" of such
integrals rests on the connection with differentiation, to which we now
turn.

3. Derivatives
The derivative of a variable quantity y with respect to another such quan-
tity x on which it depends is to be the instantaneous rate of change of y
relative to the change in x. This description has intuitive appeal, espe-
cially in the special case when x is taken to be time. Instantaneous rates
are modeled on average rates: For a value of YI at time XI, changed to a
value Y2 at a different time X2, the average rate of change is the ratio
(Y2 - YI )/(X2 - XI) of change in y to change in x. The "instantaneous"
aspect might again be formulated by an infinitesimal change dx from X to
x + dx. Then if y depends on x by a function y = f(x), the instantane-
ous rate of change is the ratio [f(x + dx) - f(x)]ldx or, in an evident
extension of the dx notation, dy /dx, the quotient of two infinitesimals.
Such infinitesimals serve for quick calculation. For example, if f(x) = x 2 ,
then the instantaneous rate is

dy _ (x + dX)2 - x 2 _ 2xdx + dx 2 = 2xdx = 2x

(1)
dx dx dx dx '

since the square (dx)2 of an infinitesimal surely vanishes, at least in com-

parison with its first power. Other rates easily follow. Moreover, if y
depends on x and x in turn on time t, cancellation of the infinitesimal dx
glVes

dy dx _ dy
(2)
dx·Tt - dt'
and one has (but hasn't quite proved) the important (because necessarily
most useful) "chain rule" for the derivative of a composite function. Thus
calculus with infinitesimals is intuitive and an efficient means of calcula-
tion.
But what are these infinitesimals? The archimedean law for the real
numbers (§IV.4) says that a positive number, no matter how small, has
4. The Fundamental Theorem of the Integral Calculus 155

arbitrarily large multiples-in effect, that a positIve number cannot be

infinitesimal. With Bishop Berkeley, one may conclude that "infinitesimals
are the ghosts of departed quantities."
What remains are limits. For each value x = a and each actual finite
increment x = a + h with h =1= 0 one may form from a given function
y = j(x) the very same ratio [j(a + h) - j(a)]lh as before. The deriva-
tive, when it exists, is then defined to be the limit of this ratio as h =1= 0
approaches 0; with either standard notation dy Idx or j'(x), this means
that the derivative is

[ dy 1 = j'(a) = lim j(x + h) - j(a) . (3)

dx x=a h~O h
The necessary meticulous t: - 8 definition of the limit in this description
must again use the quantifiers (for all t: there exists a 8) to achieve preci-
sion. With this precision (as careful students of the calculus know) one
can then prove that the derivatives of x 2 and of xn are what they should
be and that the desired rule (2) for the derivative of a composite function
does hold for suitable differentiable functions-not by simple cancellation
of infinitesimals, but by limits taken after cancellation.
This discussion of alternative views of the calculus may serve to support
our thesis that Mathematics is not just formalism and is not just empiri-
cally convenient ideas, but consists in formalizable intuitive or empirical
ideas. Calculus fits this thesis as to the nature of Mathematics because it
starts from problems (such as the calculation of areas or of rates of
change) and from these problems develops suggestive ideas which ulti-
mately can be fully formalized. For calculus, the initial formalization
dealt chiefly with practical rules for finding and manipulating derivatives
and integrals. This is not yet a complete formalization, which was first
provided in the 19th century by the rigorous t: - 8 treatment by limits.
Another thesis asserts that the same intuitive idea can be variously for-
malized. For the calculus, we now know that this is the case. Indeed, the
earlier and originally wholly speculative use of infinitesimals can be
rigorously formalized in at least two different ways: By using Abraham
Robinson's non-standard model of the reals, as presented for example in
Keisler's text [1976] on calculus, or by using Lawvere's proposals to use
"elementary topoi" in which the real line R is presented not as a field but
as a ring in which there is a suitable infinitesimal neighborhood of 0 (see
Kock [1981]).

4. The Fundamental Theorem of the Integral

Calculus
The essential connection between differentiation and integration is pro-
vided by the theorem of the title of this section. This theorem goes back
to a simple intuitive idea: That the total change in a quantity ought to be
156 VI. Concepts of Calculus

exactly the sum of the successive small instantaneous changes-an idea

which in other forms arises early in geometry, with the rule that the whole
is the sum of its parts. Specifically, if the quantity in question is a function
F(t) of time t, the total change from time t = a to a later time t = b is
just the different F(b) - F(a). Now suppose this function F(t) has at each
time t a derivative pet) = f(t). The instantaneous change in an
infinitesimal interval dt of time is then the product of this interval dt by
the instantaneous rate of change F'(t), and so the sum of all the successive
instantaneous changes is the definite integral f~ f( t )dt. The suggested
theorem then reads

Theorem. If the function F(t) has a continuous derivative f(t) = F'(t) on

the interval a < <
t b, then

ia
b
f(t)dt = F(b) - F(a). (I)

The expression on the right is often written as [F(t)]%, while the function
F is called the indefinite integral ff( t )dt.
Before we discuss a rigorous version of the "infinitesimal" motivation of
this theorem, let us consider its utility. It provides a formula for calculat-
ing the (Riemann) integral f%, provided the function f( t) inside the
integral is known to be the derivativef(t) = F'(t) of some other function
F. If we have differentiated enough such functions F and have prepared a
"Table of Integrals" giving the results, we may then hope to find to the
given "inside" function f( t) a suitable "primitive" F( t). When this is the
case, the definite integral (representing an area, a volume, or some other
quantity to be determined) may be calculated as the difference
F(b) - F(a) of two values of this function F. Thus the fundamental
theorem conceptually ties the process of differentiation to that of integra-
tion, and provides means of (sometimes) calculating definite integrals.
But alas, our table of integrals may not contain any function F with the
desired derivative f. For instance, if we have so far differentiated only
polynomials, powers, square roots and the like, we will have no function
of t with derivative the square root (l - t 2 )1I2. Progress on the
corresponding definite integral f% (I - t2)1/2dt, with well chosen limits
a and b (i.e., with 0 < < <a b I) is then possible only when we use
trigonometric functions; set t = sin 0 for some angle 0, determine that
the derivative of sin 0 is cos 0, and then use the standard rules for
"change of variables" - dtldO = cos 0 or dt = cos OdO -to get
4. The Fundamental Theorem of the Integral Calculus 157

where the new limits 0' and f3 for the integral over 0 are chosen so that
sin 0'= a, sin f3 = b. Now, if we have noticed that the derivative of
sin 0 cos 0 is 2 cos20 - 1, we have a primitive, so we may again use the
fundamental theorem to calculate

f f3 cos 0dO =
a
2 (1/2)[0 + sin 0 cos Olr
In particular,

(2)

This is a small indication of the way in which integration forces the con-
sideration of derivatives of new classes of functions-here the trig-
onometric functions; later elliptic functions. A still simpler example is the
integral fdx/x, which leads to the function lo~ x. In the present case, the
entry of trigonometric functions is no accident-after all, the integral on
the left of (2) is just a representation of the area '1T / 4 of the quadrant of a
unit circle, and the trigonometric functions serve to provide rectangular
coordinates for points on the circle.
Now return to consider a possible proof of the fundamental theorem.
We can derive it from two more primitive facts; as follows.
Lemma A. If a continuous function F( t), defined for all t with a <; t <; b,
has derivative 0 for every such t, then F(t) is constant for all t with
a <; t <; b.
This is intuitively plausible: If the rate of change of F is zero every-
where then it doesn't change at all, hence is constant. Later we will return
to examine a rigorous proof of this lemma from the Law of the Mean
(§7).
Lemma B. If the function f(t) is bounded between two constants m and M,
so that m <; f(t) <; M for all t with a <; t <; b, then the definite integral
of f(t) satisfies the inequalities

(b - a)m <; ia
b
f(t)dt <; (b - a)M. (3)

If one considers the definite integral as a measure of the area under the
curve y = f(x) and between the ordinates x = a and x = b, these two
inequalities are evident, since (b - a)M is the area of a rectangle includ-
ing all that area under the curve, while (b - a)m is the area of a rectan-
gle wholly under the curve. An exact proof of (3) from the limit
definition of the integral in §2 is straightforward: In the sum (2.2), each
term is by hypothesis bounded as in .

m(xi - Xi-I) <; (Min;f)(xi - Xi-I) <; (Max;f)(xi -Xi-I)

<; M(x; - Xi-I)'
158 VI. Concepts of Calculus

The whole sum is then squeezed between m(b - a) and M(b - a), so the
same must be true for the limit of this sum; that is, for the definite
integral.
As for the fundamental theorem (1), consider the definite integral

fa
I
G(t) = J(t)dt

with a variable upper limit t. To get at its derivative, change t to t + h.

By Lemma B

hm < G(t + h) - G(t) = f I

l+h
J(t)dt < hM,
where m and M are now lower and upper bounds for J(t) on the short
interval from t to t + h. Since J is continuous at t, both bounds can be
made close to J(t) by taking h small. Therefore G '(t) = J(t). (Geometri-
cally, the rate of change of the area under the curve is the length of the
ordinate J(t).) This means that F(t) - G(t) has derivative zero. By
Lemma A, it is constant; since G(a) = 0, this constant must be F(a). Thus

G( b) = J
a
b J( t )dt = F( b) - F( a),

which is the fundamental theorem.

5. Kepler's Laws and Newton's Laws

The problems presented by mechanics have had a decisive influence on

the development of calculus. In mechanics, one needed to analyse the
ways in which bodies move-both heavenly bodies and terrestrial ones,
falling freely under gravity, sliding down inclined planes, or propelled by
other forces. Thus Isaac Newton developed the calculus in order to
account for Kepler's laws describing the motions of the planets. Kepler,
on the basis of extensive earlier empirical observations, asserted that each
planet moves around the sun in a plane containing the sun, that the orbit
of the planet is an ellipse in that plane, with the sun at one focus, and
that the variable speed of the planet in its orbit is such that the radius
vector from the planet to the sun would sweep out equal areas in equal
times.
In the calculus, one had to hand the conceptual description of velocity
and acceleration. For a body moving in a straight line, with position given
at any time t by its distance s from some origin on the line, the velocity v,
as we have already seen, is just the rate of change ds/dt of s with respect
to time; Newton wrote this rate (this derivative) as s, and called it the
fluxion of the fluent s. One also needs to consider acceleration, described
5. Kepler's Laws and Newton's Laws 159

as the rate of change of velocity with respect to time; this acceleration is a

second derivative a = d 2s/dt 2, also written as s.
For bodies moving in
three-dimensional space, referred to coordinates x, y, and z, the velocity is
a vector with components x, y, and i, while the acceleration is the vector
x, z.
y, In this way, the concept of the derivative enters in the description
of the kinematics of motion.
Newton's laws of motion describe the effect of external forces on the
motion of a body (or better, of a particle). When there are no such exter-
nal forces, the body continues to move in a straight line at a constant
velocity; in other words, its acceleration is zero. When there are external
forces, the acceleration is proportioned to the force. More exactly, the
acceleration vector is in the same direction as the force; specifically mass
times acceleration is equal to force. Thus for one-dimensional motion
along a straight line under a constant force there will be a constant
acceleration, call it g, and the law reads a = s = g. Here we are given
the (constant) derivative g; one function with this derivative is gt and by
Lemma A of §4 any other function with this derivative can differ from gt
by most a constant, call it vo, so that

oS = gt + Vo. (1)

This constant Vo represents the initial velocity (that at time t = 0). By a

similar argument, one finds

S = (1/2)gt 2 + vot + So (2)

where So is a constant, the coordinate (initial position) of the body at time

t = O. Thus Newton's laws and a simple integration easily suffice to
derive the familiar formula (2) for a body falling under (constant)
acceleration of gravity.
The corresponding derivation of Kepler's laws for the motion of planets
rests also on the assumption that the only force acting on the planet is the
force of gravitational attraction to the sun. This force is directed from the
planet to the sun, and its magnitude is given by the inverse square law
and so is proportional to the product of the masses of the sun and the
planet and the square reciprocal of the distance r from sun to planet. At
anyone time the sun, the planet and this velocity vector of the planet
relative to the sun determine a plane (Figure 1). The force of gravity is
directed toward the sun and in this plane. Hence the acceleration of the
planet perpendicular to that plane is zero-and remains zero. Thus
Newton's laws easily prove Kepler's first law, that the planet moves in a
plane. For coordinates x and y in that plane, with origin at the sun one
x
may now express the acceleration and ji in terms of the corresponding
polar coordinates rand 0 and the components of force as

(3)
160 VI. Concepts of Calculus

~~--~~--------------------x

Figure 1. Planetary orbit.

where A is a proportionality constant. A suitable integration of these

equations (Chapter IX) proves Kepler's other laws. This represents the
initial triumph of the calculus-a clear derivation of the empirically based
laws for the motion of the planets from more conceptual laws of motion.
This triumph is all the more remarkable because it shows that the same
laws of motion apply both to the planets and to motions here on earth.
We have already noted the derivation of the equation of motion for a
freely falling body. For a projectile much the same argument works. If a
body is projected from the origin in the (x,y)-plane, with the vertical axis
y positive upwards, and with initial velocities Vx and Vy along the two
axes, the accelerations of the body are given by Newton's laws as

y = -g, x= 0, (4)

where g is the (constant) acceleration of gravity. Integrating twice gives

Thus, substituting x for t in the first equation, the path of the body is the
locus of the equation

This is a parabola in the (x,y)-plane.

The case of a projectile illustrates in an especially simple way one of
the characteristics of formal mathematical calculations. One need not
watch the whole path of the projectile; rather the Mathematics predicts
this path from physical laws (here Newton's laws) and initial
observations-here the initial position and velocity of the body. The calcu-
lation then proceeds without further attention to the actual observation of
the body, but comes out with a (more or less accurate) description of the
successive positions of the body. The same formal character applies to
more careful or accurate calculations, such as those taking air resistance
into account.
6. Differential Equations 161

6. Differential Equations
This discussion of the motion of planets and of projectiles is a typical
instance of the general method of describing physical and other
phenomena by differential equations. A typical differential equation, now
of the first order and not "second order", will determine some unknown
function x of the variable t by giving the first derivative of x,

dx/dt = h(x,t), (1)

in terms of a known function h of the two quantities x and t. The under-

lying way of determining a solution is then the same basic notion: That
the total change in a variable quantity x ought to be the sum of the suc-
cessive instantaneous changes h(x,t)dt. The expression of this idea in the
fundamental theorem of the integral calculus can be rewritten (see (4.1»
as

F(b) = F(a) + i a
t
f(t)dt.

In words, the final values of the variable F (or x) is determined by the ini-
tial value F(a) and all the intermediate values f(t) = F'(t) of the deriva-
tive on the interval a <; t <; b. This underlies the use of "initial values"
in getting solutions of differential equations. In explicit terms, this
amounts to a search for some function x = g(t) defined for a suitable
range of t and such that this function with its derivatives satisfies the
given differential equation (1):

g '(t) = h(g(t),t) (2)

for all the intended t. To do this, we must again canvass functions g whose
known derivatives might satisfy (2); elementary treatments provide vari-
ous rules and tricks to do this, for example by changing the variables x or
t to other more manageable quantities. As with elementary integration,
the known functions may not suffice to get the solution. At this point one
might use numerical methods, or might try to invent a new function g to
do the job. Such invention might seem a chancy matter, but this need not
be so. Given the general definition of a function, it is possible to prove
"existence theorems" which specify conditions under which a solution g to
equation (2) must exist and (perhaps) conditions which make the solution
unique. One such existence theorem is the Picard theorem.
To formulate this theorem, we consider the initial condition that x = Xo
when t = to and assume for the equation (1) that the function h on the
right is continuous, say in the square D consisting of all (x,t) such that
I x - Xo I < a and It - to I < a. A solution of (1) with the given
initial conditions is then a function g(t), defined for t in some interval
162 VI. Concepts of Calculus

I t - Xo I < ~ about to, and with I get) - Xo I < a there, such that
g(to) = Xo and that (2) holds for all I t - to I <~. One version of the
Picard theorem then asserts that there is a ~ > 0 for which such a solu-
tion exists and is unique, provided that the given function h satisfies the
Lipschitz condition: There is a constant M > 0 such that, for all (X,tl)
and (X,t2) in the square D

(3)

Incidentally, this Lipschitz condition will certainly hold if the function h

has a continuous partial derivative with respect to t everywhere in the
square D.
This result will illustrate the care necessary to formulate exactly the
intuitive idea that a solution ought to exist and ought to be unique-at
least ought to if the differential equation formulates correctly a physical
situation known to have a physical solution.
Our illustrations from mechanics had to do with a second order
differential equation such as x = j(x,t). In this case the appropriate ini-
tial conditions amount to giving both Xo and its first derivative Xo at the
initial time t = to. This one may see by reducing the single second order
equation to a pair of first order equations; use the velocity v = x as
another variable; the equations are then

v = j(x,t), x = v

and there is a corresponding existence theorem. For a particle moving in

three dimensions the corresponding second order differential equations
involve three second derivatives x, ji, and ii, and reduce to first order
equations in the six variables x, y, z and x, j, z. The space of these six
variables-called a phase space-is a simple example of the need for higher
dimensional spaces in the understanding of mechanics (Chapter IX). For
k moving particles, one needs a phase space of 6k dimensions! Mechanics
motivates higher geometry.

7. Foundations of Calculus
The differential and integral calculus, as we have seen, starts with prob-
lems of calculating various quantities and the invention of uniform
methods of making these calculations. These methods rest initially on
vague but persuasive ideas of infinitesimals, infinite sums, and of rates of
change; their very success and rapid development forcibly raises the ques-
tion of formulating a rigorous foundation for these methods. This founda-
tion, developed in the 19th century, must start from a clear notion of a
function (the thing which might have a derivative). The foundation then
7. Foundations of Calculus 163

rests on the notion of a limit, the definition of derivative and of integral in

terms of limits, and the careful proof of a number of necessary facts about
these processes. In particular, this involves the proofs from the axioms for
R of the following sequence of intuitively plausible properties of any real

°<
valued function f, continuous on the unit interval [0,1], which consists of
allx, x < 1:

f: [0,1] ~ R, continuous. (1)

(i) The function f is bounded above and below; that is, there are real
numbers m and M such that, for all x in [0, I],

m < f(x) < M.

(ii) The function f takes on its maximum; that is, there is a point Xo in
[0,1] such that f(x) < f(xo) for all x E [0,1].
The maximum can be determined from property (i) and continuity as
the least upper bound of the set of all the real numbers f( x) for x E [0, I].
This is typical of the use of the least upper bound property of the reals.
(iii) The function f takes on its minimum in the interval [0,1].
(iv) The function f takes on all intermediate values. It is enough to
prove that if f(O) < ° and f(l) > 0, then there is an x E [0,1] with
f( x) = 0. That is, f takes on the value of 0.
(v) The function f is uniformly continuous in [0,1]; that is, for each
( > ° °
there is as> such that, for all Xl, x2 in [0,1],

requires that for all Xl and all ( > ° °

Compare this with ordinary continuity in the interval. That continuity
there is as> (perhaps depending
on Xl and () such that, for all X2 in [0,1], the implication (2) holds. Thus
for ordinary continuity one must produce for each Xl and ( a S which
works in (2) for that one value of Xl; for uniform continuity one must
produce for each ( a single S which works for all points Xl in the interval.
Thus uniform continuity is (2) with prefix the sequence of quantifiers

(V' ( here is short for "for all (> 0"), while ordinary continuity
throughout the interval is the same statement (2) preceded by the
sequence of quantifiers

In other words, interchanging a universal quantifier (V' Xl) and an

existential quantifier (:J S) makes a difference (examples below)-while
164 VI. Concepts of Calculus

interchanging two universals such as (\I {) and (\I XI) makes no

difference in meaning.
We continue with the properties of the continuous function f of (l).
(vi) (Rolle's Theorem). If f(O) = f(l) and f has a first derivative at
every point X with 0 < X < I, then there is a point ~ 0 < g < I, with
f'm = O. (In other words, the graph of f has a horizontal tangent some-
where.)
For ~ one can take a point where f attains its maximum as given by
property (ii), and then show from the definition of the derivative that f'
must be zero there. This argument fails if this g is 0 or I; one may then
choose a point ~ where f attains its minimum.
(vii) (Law of the Mean). If f has a first derivative at every point x with
o < x < I, then there is a point ~,O < g < I, where f'a) =
f( I) - f(O). Put differently, the graph of f has somewhere a tangent
parallel to the longest secant.
This law follows from Rolle's theorem applied to the function

g(x) = f(x) - x[f(l) - f(O)]

which does satisfy the hypothesis g(O) = g(l) of Rolle's theorem. The law
of the mean also proves Lemma A of §4.
All of these theorems apply at once to real-valued functions f continu-
ous on any finite closed interval, say the interval [a,b] of all x with
a < < x b. In that case, for example, the law of the mean states that

feb) - f(a) = f'W (b - a)

for some point g with a < g < b: The total change in the function is the
length of the interval times the rate of change at some intermediate point
g. This law is another way of expressing one of the basic ideas of calculus:
The first derivative gives a linear approximation for the function.
Uniform continuity, as formulated in (v), is powerful. For example, it
can be used in connection with our description in §2 of the definite
integral fb f( X)dx. Given { > 0, uniform continuity yields a 8 > 0 so
that I XI G_ X2 I < 8 implies I f(xI) - f(X2) I < {/(b - a). Thus if
we subdivide the interval of integration into pieces Xi - Xi-I of length
less than 8, the maximum and the minimum of f in any subinterval differ
by less than deb - a), and this in tum implies that the lower sum L(j,a)
and the upper sum U(j,a), defined as in §2, differ by less than
(d(b - a»~xi - Xi-I) = €. Hence they do approach the same limit, as
we asserted in (2.2) in defining this limit to be the definite integral.
The assertion in (v) of uniform continuity depends essentially upon the
hypothesis that f is continuous in the closed interval [0, I ]-the interval
including its endpoints. For example, the function g(x) = l/x is continu-
ous on the open interval (0, I), consisting of all real numbers x with
7. Foundations of Calculus 165

°><
f °
x < 1, but g is not uniformly continuous on this open interval-given
0, as x gets closer to even smaller ~ is required in (2). In fact the
uniform continuity property (v) is a consequence of the following property
of the topological space X = [0,1].

Lemma (Heine-Borel alias Borel-Lebesgue). If the interval [0,1] is con-

tained in the union of open intervals Vi (jor i in any index set I), then [0,1]
is contained in the union of a finite number of these Vi.
One says that [0,1] is covered by the Vi when it is contained in their
umon.
PROOF. Use least upper bounds! Let [0,1] be covered by some collection

°°
{Vi l i E I} of open intervals. Consider the set S consisting of all those
real numbers x E [0,1] such that the closed interval [O,x] from to x can
be covered by a finite number of the Vi. This set S is bounded by and 1,
hence has a least upper bound, call it ~. If ~ is less than 1, this ~ belongs to
one of the V, say to Vj-and Vj contains numbers e > f Since ~ is a
least upper bound of S, Vj also must contain some XES. Then the
interval [O,x] is covered by a finite number of the Vi, and this number,
together with Vj, cover the interval [O,e] stretching beyond ~, a contradic-
tion to the choice of ~ as the least upper bound of S. Therefore ~ = 1, so
the interval [0,1] is covered by a finite number of U's.
One should note that this is not a "constructive" argument. It does not
produce the finite cover, but argues by contradiction.

°
Heine-Borel suffices to prove uniform continuity (property (v) above).
For let f > be given. Since f is continuous at each point Xo there is for
each Xo an open interval VXo (with center xo) such that
I f(xl) - f(x2) I < { whenever Xl and X2 are in Vxo' Now (a small
trick) consider the intervals VXo c Vxo with the same center Xo and half
the radius. Since each Xo E VXo ' these smaller intervals cover [0,1]. By
Heine-Borel, a finite number of them will cover. Take ~ to be the
minimum radius of this finite number of intervals. Clearly ~ > O. Now
consider any two points Xl and X2 in [0,1] at distance apart less than ~.
Since the intervals cover, Xl is in some interval Vxo' By the trick, Xl and
X2 lie together in the larger interval VXo so I f(xl) - f(x2) I < f, as
desired.
With this result, we have completed the announced proof of the
existence of the definite integral. However, for double integrals we will
need Heine-Borel for a square. The proof of uniform continuity, how-
ever, involves nothing really new. To understand it, think of the interval
(or the square) as a metric space and hence a topological space and call
such a space compact when it satisfies Heine-Borel. In other words X is
compact if every covering of X by open sets Vi contains a finite sub cover-
ing. Then the proof we have just given for (v) really proves
166 VI. Concepts of Calculus

Theorem 1. Any continuous real valued Junction J: M -->R on a compact

metric space M is uniformly continuous.

There are many compact spaces. For example, any product of compact
spaces is compact; in particular the unit square or cube is compact. Any
closed subset C of a compact space X is compact; here a subset C c X is
said to be closed when its compliment X - C is open in X, while the open
subsets of C (which define its topology) are just the intersections U n C
with open subsets U of X. A subset of the line or the plane is compact if
and only if it is both closed and bounded. On the other hand, the open
interval of all real x with 0 < x < 1 is not compact; when it is covered
by expanding proper open subintervals, no finite number of them will
suffice to cover. Similarly the whole real line R is not compact, and it car-
ries functions such as x 2 which are continuous but not uniformly continu-
ous.
Compactness is related to convergence. One readily proves that in a
compact space X every infinite sequence of points Xn has an infinite
subsequence converging to some point of X; here the convergence of a
sequence of points is defined just as was the convergence of a sequence of
numbers (§IV.4). This indicates again why R is not compact, because the
sequence of natural numbers has no convergent subsequence. A metric
space X is compact if and only if every infinite sequence of points in X
has an infinite subsequence converging to a point in X -but this result
fails for general topological spaces with convergence defined using neigh-
borhoods. The recognition of the importance of compactness and of its
description by coverings is a major step in the understanding of topologi-
cal spaces. It developed only slowly-and was not really codified until
Bourbaki, in his influential 1940 volume on topology, insisted.
This concept of compactness is only one of the many issues arising from
the development and foundation of the calculus. Geometry and mechan-
ics had led to the intuitive ideas of rates of change, area, and summation.
The somewhat vague formulation of these ideas by means of
infinitesimals proved to be powerful tools in the 18th century, but various
difficulties developed. For example (Titchmarsh [1932], 1.75), the
sequence of functions In(x) = n 2x(l-xt converges to zero for all x in
the interval 0 < < x 1, but the definite integral of the In(x) from 0 to I
is n 2 /(n + 1)(n + 2), so the sequence of integrals does not converge to O!
(The original convergence is not "uniform".) This is typical of problems
arising with the interchange of two infinite processes (here convergence
and integration). Such difficulties eventually required a more sophisticated
formulation of the calculus, based on careful development of the concept
of a limit. This concept, in its turn, involves a careful use of logical
quantifiers and suggests the consideration of limits in R2, in R 3, and in
much more general topological spaces. The study of the definite integral
inevitably involves new notions of uniform continuity and then compact-
8. Approximations and Taylor's Series 167

ness, and the proofs must rest upon a careful axiomatization of the real
numbers-an arithmetization of one-dimensional geometry. The axioms
alone do not suffice-there is a whole list of subtle consequences. Also,
compactness in its turn must be disentangled from the more pictorial
notions "closed and bounded". There are many more developments,
especially of more general integrals, such as the Lebesgue integral. Thus,
all told, the initial intuitive ideas and problems, and their extensions, and
applications, lead to notions which in their turn demand more subtle con-
cepts internal to Mathematics.

8. Approximations and Taylor's Series

Ideas of successive approximation underly the calculus. Areas are
approximated from inside and outside the simple areas, and integrals are
approximated by finite sums. Tangent lines to curves are approximated by
secants, instantaneous velocities are approximated by average velocities,
and similarly for higher derivatives. For example, the first derivative of a
function y = j(x) allows an approximation of that function by a linear
function, as in

j(x) '" j(a) + (x - a)!'(a), (I)

where the linear function on the right agrees with j and with j' at
x = a; in other words, it is the linear function whose graph is the tangent
line to y = j(x) at x = a. Moreover, one can estimate the error in such a
linear approximation. If the first derivative j '(x) is continuous on the
closed interval from a to x, while the second derivative j"(x) exists there,
a simple iteration of the Law of the Mean (§7) shows that there is a real
number ~ between a and x such that

(X a)2
j(x) = j(a) + (x - a)j'(a) + j"(~) 2
(2)

In other words, if the second derivative j" is small in the interval, the
linear approximation (1) is good.
This formula (2) is a special case of Taylor's theorem. If a function
j(x) has a continuous (n - l)st derivative j(n-Il(x) on the closed inter-
val from a to x and if the nth derivative exists on the open interval from a
to x, then there is a real number ~ between a and x such that

where k! = 1·2 ... k. In other words, the first n - 1 derivatives of f pro-

vide coefficients for an approximation to j by a polynomial of degree
168 VI. Concepts of Calculus

n -1. The polynomial is uniquely determined as the one of that degree

which at x = a agrees with f and with the first n - I derivatives of f.
The error in this approximation is then measured by a value of the nth
derivative, as in the last term of (3); there are also other (integral) formu-
las for this "remainder term", and there are other types of polynomial
approximations.
For a function f(x) which has derivatives of all orders, this process sug-
gests the formation of an infinite power series

(4)

called the Taylor's series for f. The explicit meaning of such infinite sums
is then given by the convergence of the partial sums, as discussed in
§IV.4. Often the remainder formula given by Taylor's theorem (3) can be
used to show that the Taylor series converges for some (or even all) values
of x, or that it converges "uniformly" for some closed interval of values of
x. This leads to the familiar power series for eX, sin x, and cos x. They-
and their variants-are then at hand for the computation of tables of trig-
onometric functions. Indeed one can use the power series to define analyt-
ically the function sin x, thereby avoiding the winding function of
§IV.2-and obscuring the reasons for the periodicity of sin x. The
numerous other applications of such power series indicate that the initial
elementary idea of approximation, in particular linear approximations,
does indeed have extensive consequences. This does not justify the
overuse of merely linear approximations in the multiple regression
methods (least squares) so popular in econometrics.

9. Partial Derivatives
For elementary problems in algebra, it is always vital to know whether to
add or to multiply. For example, the total change due to two separate
causes clearly ought usually to be the sum of the separate changes. On
the other hand, one must multiply to get a composite rate of change: For
example, given the exchange rate from pounds to dollars and the rate
from dollars to francs, the product will be the exchange rate from pounds
to francs.
These two simple observations of the practical meaning of addition and
multiplication appear formally in the chain rule for differentiation in the
calculus. This we have already used: If z = g(y) and y = h(x) are two
functions with continuous derivatives, then in the relevant range
z = g(h(x)) is a function of x and has derivative

dz dz dy
z'(x) = g'(y)h'(x), or (1)
dx dy dx·
9. Partial Derivatives 169

The proof requires a little care with the limits entering in the definition of
the derivatives involved; the limits involve the formulas for the
corresponding finite increments ~x = x - Xl, for then

~z ~ ~y
~X ~y ~x·

Thus (1) expresses the underlying reason why one should multiply rates.
The corresponding chain rule is more striking for functions of several
variables, such as a quantity z given as a function z = f(x,y) for all
points (x,y) in some open set U of the cartesian (x,y)-plane. There is no
problem in finding what derivatives might mean here; if one holds y fixed,
the quantity z remains just a function of X; its derivative, when it exists, is
called the partial derivative with respect to x. Thus at a point (x,y) in U
this derivative, for h =1= 0, is

az -_ f' x (x,y )
-a - l' f(x
- 1m
+ h,y)h - f(x,y)
. (2)
X h~O

Holding X fixed, there is a similar partial derivative f 'y(x,y). This is the

more explicit notation, specifying both variables and indicating which is
"held fixed" and which is variable. The notation az / ay for the same par-
tial derivative is incomplete unless an indication is added to show which
other variable is at hand and is held fixed. (The observation is important,
say in thermodynamics, where several alternative pairs of independent
variables may be to hand.)
For z = f(x,y) these two partial derivatives give the rate of change of z
when the point (x,y) in the plane moves horizontally (in the X direction)
or vertically (in the y direction). As such, they do not give complete infor-
mation about all the possible rates of change of z. One would at least also
want the rate at which z changes when the point (x,y) moves off in
another direction-say in a direction making an angle 0 with the positive x
axis. A linear change in this direction is then given, in terms of a parame-
ter t, by

x(t) = x + (cos O)t, yet) = y + (sin O)t.

The intuitive principle that the total change is the sum of the separate
changes in x and in y then suggests that the derivative of z with respect to
the parameter t in the direction 0 should be the linear combination

-dz = [az]
- cos 0 + [az].
- sm 0, (3)
dt ax ay

usually called the directional derivative. When both the partial derivatives
of z are continuous, this formula holds; it is a special case of the following
170 VI. Concepts of Calculus

chain rule. If the functions x = g( t) and y = h( t) have continuous

derivatives, giving values x,y in the set U C R2 where the function
z = f(x,y) also has two continuous first partial derivatives, then
z = f(g(t),h(t» is a function of (suitable values of) t with the continuous
derivative

dz =~dx + az dy. (4)

dt ax dt ay dt

This clearly includes the motivating case (3). A similar formula applies
when z is given as a function of more than two variables or when these
variables x and y depend not on one but on several parameters.
This chain rule (4) has several different aspects.
First, think of dx = (dx/dt)dt as an infinitesimal change in x, caused
by the (equally) infinitesimal change dt in t. Then, multiplying (4) by dt
and cancelling gives
az
dz = - d x
az
+ -dy. (5)
ax ay
This expression is called the total differential of z; for given values of x
and y it gives the total change in z due to infinitesimal changes dx and dy;
we will soon give a less "infinitesimal" interpretation.
Starting from a point xo, yo with finite changes x - Xo and y - yo,
the formula (5) suggests a linear approximation z - Zo to the change in
z:

(z-zo)= [~]
~ 0
(x-xo)+ [~] (y-yo).
~ 0
(6)

This suggests (and correctly) that there is a Taylor's formula and also a
Taylor series, each valid for functions z of two variables x and y which are
sufficiently "smooth". Here and later we will use the term smooth for
functions with enough continuous derivatives, in cases where we do not
wish to specify in detail how many such derivatives are in fact needed.
In the chain rule (4), dz / dt can be regarded as an "inner product" of
two "vectors", as follows

dz = ~ dx + ~ dy = [~, ~]. [ dx, dY ]. (7)

dt ax dt ay dt ax ay dt dt

The first factor on the right is called the gradient of z = f( x,y); it is

defined at each point (xo,Yo) of the plane, and is written

('7/)0 = [af , af 1 . (8)

ax ay X=Xo, y= Yo
9. Partial Derivatives 171

This vector "points" in the direction of the maximum rate of increase of

the function f, and has that rate as its length. The function f determines
one such vector at each point of the plane. At each point, all such vectors
for all f form a two-dimensional vector space, called the cotangent space
attached to the plane at that point.
The second vector in the product (7) depends on the functions
x = get), y = h(t). They describe a continuous path passing through the
point xo, yo; such a path is also called a parametrized curve, consisting of
the points (g(t),h(t)) each labelled with the corresponding value of the
parameter t. Such a curve is the trajectory of a moving point. At time to,
where x = xo, and y = Yo, the velocity of this moving point is the
second factor of (7)

x 9l
[ ddt , ddt ]1==1 = (g'(to), h '(to)). (9)
0

It is called the tangent vector to the path at the point. All the tangent vec-
tors to trajectories through the point (XO,yo) form a two dimensional vec-
tor space, called the tangent space To to the plane at this point.
Now the chain rule, in the form (7), gives dz /dt as the product of the
gradient vector of z by the tangent vector to the path. If one considers
both vectors to lie in the same two-dimensional space, this product is just
the inner product as described in §IV.9. However, it is preferable to con-
sider the tangent and cotangent spaces as conceptually distinct. The
"product" in (7) is then a real-valued function of two vectors, one from
each space. This function is linear in each vector when the other is held
constant, so is said to be bilinear; in Chapter VII we will indicate why it
makes the cotangent space "dual" to the tangent space.
The cotangent space may be constructed formally as follows: Take all
smooth functions f(x,y), each defined in some neighborhood of (xo,Yo);
they form an (infinite dimensional) vector space under addition of values
and multiplication by real constants. Call two such smooth functions
z = f(x,y) and w = k(x,y) cotangent (or equivalent) at the point (XO,Yo)
when they have the same first partial derivatives there:

The equivalence classes of these functions under this relation then form
the desired two-dimensional space, called the cotangent space at the point
(XO,yo). This construction works not just for the plane, but for other
curved surfaces such as the sphere, when the coordinates (x,y) in the
plane are replaced by suitable coordinates, such as latitude and longitude
for the sphere.
172 VI. Concepts of Calculus

On the other hand, every smooth function f has a gradient V f as in

(8). In particular, the coordinates x and yare smooth functions, with gra-
dients Vx = (1,0) and Vy = (0,1). Hence every gradient can be
expressed at each point as a linear combination of these two gradients, in
the form

Except for notation, this is just the definition

df = [ ~l dx + [ ~l dy
of the total differential. Thus the differential, born as an infinitesimal, may
be defined to be the gradient V f -a vector in the cotangent space. This
is why the cotangent space differs from the tangent space.
The tangent vector at to to the path get), h(t) also determines the usual
tangent line to that path, with the parametric equations

x - Xo = g'(to)(t - to), y - yo = h '(to)(t - to), (11)

where Xo = g(to) and yo = h(to).

The chain rule (4) also has a three-dimensional interpretation. The
function z = !(x,y) represents a height z above (or below) the point (x,y)
in the plane, and so may be pictured by a smooth surface S at these
heights above some portion of the plane. The tangent plane 'TT to this sur-
face at a point p = (xo,yo,zo = !(XO,yo» is by definition the plane (if
there is one) containing all the tangent lines at p to all the smooth curves
on S passing through p. Such a smooth trajectory is given by
x = get), y = h(t) and z = !(g(t),h(t»; its tangent line at p is given by
the parametric equation (II) plus the corresponding equation for z:

Z - Zo = [ : lo (/ - /0) (11')

But these three equation (11) and (11') together satisfy the linear equation
(6) for the approximate change z - Zo in z. This linear equation (6)
represents a plane in 3-space; since it is satisfied by (11), it must be the
tangent plane to the surface S according to our definition.
In this way, the chain rule combines ideas from geometry (tangent
planes), from mechanics (velocity vectors), from calculus (linear approxi-
mation), and from algebra (dual spaces), with the results appropriately
added or multiplied. It gives meaning to the "total differential".
Some of these ideas are more vivid in pictures. Thus the gradients of a
function f(x,y) defined in the whole (x,y)-plane give a vector at each
10. Differential Forms 173

point in the plane-hence a vector field in the plane (Figure I). Alterna-
tively the loci where f(x,y) = constant give a family of curves in the
plane-the contour lines for f. Their use is suggested by topographic
maps, picturing in the plane the varying heights above sea-level (Figure
2). When f is smooth the gradient vectors, if non-zero, are orthogonal to
the contour lines; for topography, they represent the direction of fastest
ascent. For Mathematics, their exploitation has proved decisive in topol-
ogy and Morse theory.

Figure I. A vector field.

Figure 2. Contour lines.

10. Differential Forms

The one dimensional definite integral

fba f(x)dx

taken along the x-interval from a to b adds up successive increments dx in

x, each weighted by the value f(x) of a function f. In two dimensions the
analogous "line" integral

f(c.d)(p(x,y)dx + Q(x,y)dy)
(a.b)

is taken along some path in the plane and adds up infinitesimal incre-
ments both in x and in y, each weighted by a function P(x,y) or Q(x,y).
These ideas will be clarified if we begin by thinking first just about the
gadget which sits under the integral sign.
174 VI. Concepts of Calculus

A first order exterior differential form w in an open set U of the x,y

plane is an expression
w = P(x,y)dx + Q(x,y)dy (1)

where P and Q are smooth functions of a point (x,y) in U. Such a

differential form can also be described as a function which assigns to each
point (x,y) in U, in a smooth way, a cotangent vector Pdx + Qdy at that
point. For each smooth function z = f(x,y) the total differential
dz = f' x dx + f 'ydy is such a differential form w. However, there are
many differential forms which are not total differentials. This is because of
the important theorem which states that the order of partial differentiation
can be interchanged, at least for a function f(x,y) with continuous second
partial derivatives:

~-~ (2)
ayax ax ay
There is a standard proof, using the mean value theorem, of this intui-
tively plausible result. It means that a differential form w in (1) which is a
total differential must satisfy the condition

ap aQ . (3)
ay ax'
such a differential form is said to be closed.
A variety of examples, some to appear below, suggest the following for-
mal definition of a line integral h w. Here the "line" L means a smooth
curve in the region U of the (x,y )-plane where w is defined. The integral
is again a limit of a sum: Subdivide the curve L at points PO,PI , ... ,pn,
with Pi = (Xi,Yi), into n "short" pieces and form the sum
n
~ P(Xi,Yi)(Xi - Xi-I) + Q(Xi,Yi)(Yi - Yi-I), (4)
i=1

each summand is then the value of the differential form w at one of the
points Pi (or, alternatively, at some other point on the part of the curve
from Pi-I to Pi), with the differentials dx and dy replaced by the actual
increments Xi - xi-lor Yi - Yi-I in the coordinates. Under suitable
hypotheses, this sum approaches a limit as n ..... 00 and the length of each
piece of the curve approaches zero; this limit is the line integral f L w. The
idea here is just that of the ordinary integral; the identity in ideas can be
expressed formally. Represent the line L by smooth parametric equations
X = g( t), y = h( I); the line integral is then equal to an "ordinary"
definite integral between suitable limits in the parameter I,

Jw =
L
flo
I)
[P(g(t),h(t»g '(I) + Q(g(t),h(/»h '(/)]dl ;
10. Differential Forms 175

moreover the standard formulas for change of variable in an integral

show that the integral on the right (between suitable limits in t) is
independent of the choice of a smooth parametrization.
Examples are numerous. In physics, the work done by a (vector) force F
in a (vector) displacement D is defined to be the product of the force by
the displacement in the direction of the force; that is, to be the inner
product F· D of these two vectors. If the force in question varies with
position in the plane as F = (P(x,y),Q(x,y)), then the work done under
a small displacement ~x,~y is approximately the inner product
P(x,y)~x + Q(x,y)~y, so that the sum (4) will approximately measure
the total work done along the line L-and the line integral, in the limit,
measures the exact amount of work.
The flow of fluid in a plane provides another example of a line integral.
If the fluid is flowing everywhere with the same constant vector velocity
V, with x and y components M and N, then the amount of fluid crossing
a position vector (~x,~y) in unit time will be the length of that position
vector times the component of velocity V perpendicular to it; in other
words, it will be the inner product of V and the vector (-~y,~x) perpen-
dicular to the position vector: The same formula - M(x,y)~y + N(x,y)~x
applies where the fluid velocity V = (M(x,y),N(x,y» varies with the point
(x,y). Adding up all the contributions and taking the limit, as in the line
integral, shows that the integral f L(Ndx - Mdy) measures the total flow
of fluid across the line L in unit time. This example (and others) shows
how the line integral is another embodiment of the general idea of
measuring a total quantity by adding up all the little bits (intuitively, the
infinitesimal bits).
A double integral f f Af(x,y)dxdy of a smooth function f over a
bounded area A of the (x,y) plane is again a suitable limit of a double
sum ~ !(XiYj )(Xi+ 1 - Xi)(Yj + 1 - Yj)· There are multiple interpreta-
ij
tions. Thus when !(x,y) = I the double integral is just the area A, while
for a slab A with density !(x,y) varying from point to point, the double
integral is the weight of the slab. Here the fundamental theorem of the
integral calculus reappears in the form of the Gauss lemma ("the diver-
gence theorem"). Suppose that the area A is bounded by a smooth closed
curve C, called the boundary C = aA of A, and let w be a smooth
differential form defined throughout A and on C. Then

fy(x,y)d< + Q(x,yjdy] ~ n [~; - ;~ ]dXdY ; (5)

the line integral of w is reduced to a double integral (which in the case of

fluid flow can be interpreted as a total "divergence"). The outline of the
proof of (5) is straightforward: Suppose that P = 0, so that only Q is
present, and suppose also that each horizontal line meeting the curve C
176 VI. Concepts of Calculus

meets C in just two points-one, (XI,YI), at the left and another, (x"y,), at
the right, as in Figure 1. Then the double integral of aQ/ax can plausibly
be replaced by an iterated integral-first on x, then on y, and the integral
on x along a horizontal line (i.e., the sum along a horizontal strip) is by
the fundamental theorem just [Q(x"y,) - Q(X[,YI )]dy; the second
integration on y then produces the line integral of Qdy in two pieces-left
side and right side of A. As the arrows in the figure may indicate, it is
here important to integrate along the curve C in the appropriate direction,
so that the area A always lies to the left of C as traversed. (This requires
the geometric ideas of orientation, as discussed in Chapter III.) Our con-
cern here is not with the necessarily careful details of rigorous proof of (5)
(say, for more convoluted boundaries C) but with the more sweeping
observation that what appears in the Gauss lemma (6) is another (neces-
sarily formal and careful) realization of the idea that the total change in a
quantity (the total along the boundary curve) is exactly the sum of the
infinitesimal changes-just the idea which was already expressed (§4) in
the fundamental theorem.
The same idea reappears in higher dimensions. An integral over a two-
dimensional surface S in Euclidean 3-space has the general form of a
double integral
Sfs[L(x,y,z)dydz + M(x,y,z)dxdz + N(x,y,z)dxdy],
where L, M, and N are smooth functions of the coordinates x, y, and z
and the integrand is a second order differential form. Such an integral
may be defined as an appropriate limit of a sum or by reduction to a dou-
ble integral taken in the plane of two parameters sand t when the surface
S is given by parametric equations x = g(s,t), y = h(s,t), z = k(s,t). If
the surface S is the total boundary S = av of some volume V, Green's
theorem (also called Gauss' lemma) asserts that

Sfav[Ldydz + Mdxdz + Ndxdy] =

SSJ,rv [ -aL
ax
- -aM + -aN dxdydz.
ay az
1 (6)

(x/, y/) 1=========:::::;7 (x" y,)

Figure 1
10. Differential Forms 177

Similarly Stokes' theorem deals with the integral of a first order form
taken over a curve as bounding a piece S of surface:

i ~
[Pdx + Qdy + Rdz] = If.s [ -aQ
ax
- -ap] dxdy +
ay

[ - - daQ
y ]d z .
+ [ -aR ] z + -aR
ap d
--dx
ax az ay az (7)

In each of these versions (5), (6), and (7) of the fundamental theorem,
the differential form w which on the left is integrated over the "boundary"
determines another differential form which appears as the integrand on
the right; the latter is called the exterior derivative of w, and is written as
dw:

d[P(x,y)dx + Q(x,y)dy] = [aaxQ - ap ]dXdY ,

ay

d[Pdx + Qdy + Rdz] = [ -aQ - -ap] dxdy +

ax ay

[ -aR 1
x d z + [ -aR - -aQ dydz,
- - dap 1
ax az ay az

d[Ldydz + Mdxdz + Ndxdy] = [ -aL - -aM + -aN] dxdydz.

ax ay az

With this description of the exterior derivative of a form, all of these ver-
sions of the "fundamental theorem" of the calculus can be written in a
uniform way. Consider a bounded smooth portion V (volume, surface,
area) of some space with a smooth boundary av. Then

I avw = I vdw,
where the smooth differential form w (and thus it exterior derivative) must
be defined throughout the region V and on its boundary. Note that the
formulas above for these exterior derivatives can all be obtained from a
simple memnonic: For a function P or L, dP (or dL) is the usual total
differential; for a product, d(PQ) = (dP)Q + PdQ, for a variable x,
d(dx) = 0, and the differentials dx, dy of the variables are multiplied with

(dx + dyi = °
the understanding that every square is zero. Then (dx)(dx) = 0, while
and hence dxdy =
-dydx. From the rule (2) for inter-
changing the order of two successive smooth partial derivatives, it then
follows that a second exterior differential ddw is always zero.
178 VI. Concepts of Calculus

These various observations form the starting point of many important

developments in topology and geometry (see Chapter VIII). These tech-
niques are also heavily used in classical Mathematical Physics in the
language of vectors rather than that of forms.

11. Calculus Becomes Analysis

At about this point in the conceptual development the subject of
calculus-rules for the effective calculation of magnitudes by instantaneous
rates and infinite sums-is gradually transformed into analysis. Here
"real" analysis deals with functions on the reals R to R and examines
their deeper properties as to integration, limits, and derivatives. It has
many aspects not covered in this book save in the following brief notes.
The (Riemann) integral which we have discussed involves a variety of
unpleasant problems such as those of the termwise integration of a con-
vergent series of functions or the interchange of order of two successive
integrations. Many of these become smooth when the integral is suitably
generalized to the Lebesgue integral fLf(x )dx, available for a larger class
of functions f. One approach to the development of this integral is by step
functions. So consider bounded functions f on an interval of the real line
and sets S of points on that interval. Cover S by disjoint open intervals
and take the sum of the lengths of these intervals. The greatest lower
bound of all these sums is the "outer measure" of S. A set E is called
measurable if for each £ > 0 there is an open set U containing E and such
that the outer measure of the difference U - E is less than t:. This also
gives the measure (the "width") of E. A step function is one which is con-
stant on each of a number of disjoint measurable sets Ei and zero else-
where. The integral of a step function f is an evident sum of the terms
"width of each step times the value of f on that step"; when f(x) is
approximated from below by a step function sex) and from above by
another such t(x), and when the least upper bound of the lower approxi-
mations equals the greater lower bound of the upper ones, this common
value is the Lebesgue integral f L(f(X) )dx, and the function f is said to
be Lebesgue integrable. There are evident algebraic properties for the
measures of intersection and union of sets; indeed such measures are
essential to the conceptual axiomization of probability theory. There is an
especially simple definition of a set of measure zero and it is useful to
assert that two functions are equal almost everywhere when they are equal
everywhere except on a set of measure zero. Now identify two functions f
which are equal almost everywhere and consider the set L 2 of all
(complex-valued) functions f on the interval whose absolute value
squared is Lebesgue integrable. To each function f in L 2 we can assign a
norm Ilfll, defined by the integral over the interval

IIfll2 = S I f(x) I 2dx.

I \. Calculus Becomes Analysis 179

The "distance" between two functions f and g in L 2 is then the norm

Ilf - gil. This distance makes L 2 into a metric space, and it proves to be
very useful to apply geometric ideas about spaces to the analytically
described set L2. It is an especial virtue of the Lebesgue integral that this
metric space is complete-every Cauchl sequence in L 2 has a limit there.
Any two functions f and g in L also have an "inner product"
ff(x)g(x)dx, so L2 is an infinite dimensional inner product space
(§VII.5).
The Lebesgue integral, as distinguished from the Riemann integral,
shows that the same primitive idea of "adding up the little pieces" can
have more than one formal expression (in fact, more than just these two!).

°
For differential equations, a central idea is that a smooth function f(t)
can be determined completely by its initial value at t = and knowledge
of all of its first derivative. The same idea appears for smooth functions
f(x,y) of two or more variables: They can be determined by "boundary
conditions" and knowledge of the partial derivatives. Such P.D.E. (partial
differential equations) arise not just from the conceptual analogy of these
ideas but from many different cases in theoretical physics where exactly
this kind of data is at hand for quantities depending on several indepen-
dent variables.
For example, let a one-dimensional wave at time t = °
be represented
by a height y (above each point x) given by a smooth function y = f(x)
as at the left of Figure 1. If this wave keeps the same form as it moves to
the right at a constant velocity c (in units distance per time), then the
height at position x at time t is given by u = f(x - ct) (try it out!). This
function u satisfies the first order P.D.E. au / at = - c au / ax and (from
Rolle's theorem) any smooth function u satisfying this P.D.E. must have
the form u(x,t) = k(x - ct) for some smooth function k. Thus the solu-
tion k = k(x,t) of the P.D.E. is determined by its initial values k(x, 0).
Similarly a wave moving steadily at velocity c to the left is given by
v = g(x + ct). Both u and v satisfy the second order P.D.E.,

(1)

o I ""------/ r

1=0 1=2

Figure \. Motion of a one-dimensional wave.

180 VI. Concepts of Calculus

called the wave equation; the most general smooth solution of this equa-
tion have the form f(x - ct) + g(x + ct) for arbitrary smooth f and g.
In three space dimensions x, y, and z the wave equation reads

(2)

it applies to many sorts of quantities u varying in time and space.

Potential. Under Newton's law inverse square for gravitational attraction,

the potential energy of a unit mass at distance r from an attracting body
at the origin is the work done by gravity in bringing the unit mass up
"from infinity". Since the force of gravity is proportional to l/r2,
this potential (the integral over r) is proportional to l/r =
(x 2 + i + z2)-1/2. This function u = l/r of x, y, and z satisfies
(except at the origin) the P.D.E.

a2u a2u a2u

ax 2 + ai + az 2 = o. (3)

By addition, the potential due to several attracting bodies or due to a

smooth distribution of such bodies will also satisfy this P.D.E., called
Laplaces's equation for the potential u.
A smooth function u which satisfies the equation (3) in an open set Vof
R3 is said to be harmonic in V. For example, an electric charge distributed
over a closed surface S in space will define a potential u constant on S
and harmonic inside S. This and related physical situations suggests the
following Dirichlet problem: Let V be an open set in R3 bounded by a
suitable surface S, while f is a smooth real-valued function defined on S.
Extend f to a function u defined on S plus V and harmonic in V. Diri-
chlet intended to solve this problem by considering, among all functions u
extending f, one which minimized the triple integral

(4)

he could then show that this minimizing function is harmonic. The

integral (4) is non-negative, so has a greatest lower bound over all u, but
it is not clear that there is an actual function taking on these greatest
lower bound as the minimum. The clarification of this situation led to
major advances in analysis (Monna [1975]).
Harmonic functions (of two variables) also arise in complex analysis
(Chapter X).
Harmonic motion is governed by the ordinary differential equation
II. Calculus Becomes Analysis 181

(5)

for (one-dimensional) position x as a function of time t. The most general

solution of this equation has the form

x = A cos kt + B sin kt (6)

for constants A and B. This function x is periodic in t, of period 27T Ik.

Moreover, (6) asserts that the vector space (over the reals) of all the solu-
tions is a two dimensional vector space, with basis the two solutions cos kt
and sin kt. Thus harmonic motion reduces to trigonometry; these two
basic solutions can be regarded as the two projections on the axes of a
point moving at uniform velocity around a circle.
In determining the solutions of P.D.E. it is often convenient to first
solve the P.D.E. for specified simple boundary values, then suitably
adding up the results. To do this one must express the finally desired
boundary value as a sum (or a series) of simpler functions. For a function
f(x) with period 27T in x a typical such expression is the Fourier series

1 00
f(x) ~ lao + ~ak cos+kx + bk sin kx. (7)
k=]

The study of such series expansions had a major historical role in the con-
sideration of the question of Chapter V: "What is a function"?
Now suppose that the series (7) does converge to f(x) for all x and try
to determine the coefficients ak and bk by multiplying f and the series by
sin mx or cos mx and then integrating from 0 to 27T. One would like to
integrate the resulting series term-by-term. This is a typical example of the
problem of double limits (interchange convergence and integration). This
is a general question, with many ramifications in analysis. When it can be
justified in the present case, it leads to simple formulas for the coefficients
in (7):

1 1
am = -7T S f(x)cos mxdx,
0
2'IT
bm = -7T S f(x)sin mxdx.
0
2'IT
(8)

The study of the actual convergence of such series makes effective use of
the Lebesgue integral, and leads to many deep theorems. There is also a
uniqueness question: If two such series (7) converge to the same function
f(x), are their coefficients all the same? This question reduces at once to
the question: If such a series (7) converges to 0 for all x, 0 x < <
27T, are
the coefficients all zero? The answer is yes, and also yes if one assumes
convergence to 0 for all x except for a finite number of points. In fact, the
182 VI. Concepts of Calculus

question of what additional exceptions are possible is what first led

George Cantor to consider abstract sets: The sets S such that convergence
to 0 for all x ( S implies zero coefficients. It is striking that such specific
questions about convergence were the principle origin of the general con-
cepts of set theory! The relevance of this origin has been emphasized by
quite recent research by Solovay in set theory.
These Fourier series deal with functions f(x) periodic of period 2'7T;
such a function can alternatively be described as a function f(O) defined
for all points 0 on the circle (of radius 1); here the additivity of angles
enters, so the circle is a topological group both compact and abelian. Gen-
eral aspects of this subject of "harmonic analysis" (this term is derived
from harmonic motion) concern functions defined on other topological
groups-there are remarkable ramifications.
At a maximum or a minimum, a smooth function y = f(x) necessarily
has derivative zero; this fact is used in the elementary (and useful) deter-
mination of maximum and minimum points. There are many more
sophisticated minimum problems: Along which curve will a moving parti-
cle descend from A to B in the least possible time? Which simple closed
curve of given length will enclose the greatest area? In these cases one
tries to find not the values of x which make a given function y = f(x) a
minimum but the values of a given function f(x) (the form of that curve
of descent) which make a given quantity a minimum. Typically, that
quantity is usually measured by an integral whose integrand is some
expression F involving both x, values of the function y = f(x) at interest
and the values of its derivative-sayan integral

f b

a
F(y,y ',x )dx, y = f(x).

Such questions of choosing a minimizing function f( x) arise in the "least

action" principles of Mechanics (Chapter IX) as well as in the Dirichlet
problem (4) above. These minimum (or maximum) problems constitute
the Calculus of Variations-a subject with many conceptual questions,
sharp need for a careful ( - ~ foundation, a number of (recently
developed) uses in practical problems of optimization and finally many
striking conceptual ideas. For example, what happens to the elementary
observation that between any two minimum of a smooth function
y = f(x) there lies a maximum? (The answer leads to the so called
"Morse relations" on the numbers of minima, of saddle points, and of
maxima, say for functions of two variables).
In the Calculus of Variations one maximizes or minimizes an integral
over "all" smooth functions f(x); it turns out to be suggestive to think of
the collection of all of these functions as a "space". The study of such
spaces of functions is one of the reasons historically for the introduction
of the general idea of a metric space, as in the case of the space L 2.
12. Interconnections of the Concepts 183

12. Interconnections of the Concepts

Algebra (e.g., group theory) and geometry (e.g., vector spaces) could each
be described as the study of models of a simple formal axiomatic system.
The calculus is not easily so described; it is rather the development, from
intuitive methods of calculation, of a variety of formal concepts (limit,
derivatives, integrals, differential forms) all found within one overarching
axiomatics (that for the real numbers and sets of reals and sets of func-
tions on the reals). Nevertheless it has the same basic character as do alge-
bra and geometry: It begins with practical problems and intuitive ideas
useful for calculations which turn out to be strictly formalizable-and
which have been so formalized, both with f - 8 and with actual
infinitesimals. Moreover, the various ideas involved are tightly connected
with each other, as suggested in the following diagram.
184 VI. Concepts of Calculus

Function Real Numbers Infinitesimals Space

~ ~~ ~~ /
Changing Quantity Approximation Total Quantity

Tangents velrity
--;::::
/ ~
/ ""'-
Area Arc Length

M"W'~ r:/=APPro< '.Som, VO~~ I)"k

j,,,,,,,,,,", "'" \ Mi,iI, Som,
Mech~ ~ LImIt Inte(al-MechaniCs

/1 ~
\j
Rt Nm Add Up Alii Ch",,, / \ j,fi,iI,

Rr" ~FUNDAMENTAI TIlEOREM

Moo, IV.lo", ~ I,,, t''''1 Leb"."

Compactness Diff Equations Surface Integral

I
Infinite Series Existence Theorems Metric Space

I
Taylors Series Partial Derivatives Gauss-Green-Stokes

p.J~n ~rfaces
\ ?tiruEO=,
Fourie! Series - - - - - - - Rule /

~
Fluid Motion

/
Periodicity Potential--- Laplace Equation Exterior Derivatives
I
Trigonometry W'1 Diri'W,I, probl,m
I
Harmonic ~DIYI'sStinSbu/tl'oBnOsUndary ~Qond'ts, M h ,~TOPO~DgYfC t' I G t
uan urn ec ames I ,eren la eome ry

' - - - - - - - - - Topological Groups

CHAPTER VII

Linear Algebra

An algebraic approach to problems of plane geometry led us in §IV.7 to

introduce two-dimensional vector spaces over the field R of real numbers,
while in the calculus gradients and tangent lines lead to cotagent and
tangent vector spaces. Three dimensional vector spaces are standard in
physics, while the algebra of vectors is an effective way of handling
geometrical ideas in dimensions higher than 3. Analysis soon produces
infinite-dimensional spaces such as L 2 (§VI.lI). This chapter will summar-
ize the properties of such linear vector spaces over an arbitrary field, not
necessarily R or C.

1. Sources of Linearity
To say that an effect is "linear" means that the effect respects proportions
and that the effect of a sum is the sum of the separate effects. The formal
description of such linearity can be stated in the context of a "linear" vec-
tor space defined much as in §IV.7. Such a space, defined over a given
field F of "scalars", is a set V of "vectors" equipped with two operations:
addition (v, w) 1-+ V + w of two vectors v and w, and multiplication
(a, v) 1-+ av of a vector v by a scalar a E F. The axioms require that the
vectors form an abelian group under addition, and that multiplication by
scalars satisfies the identities
a( v + w) = av + aw, Iv = v, (1)

(a + b)v = av + bv, (ab)v = a(bv) (2)

for all vectors v and w in the space V and for all scalars a, b in F. These
two operations of addition and scalar multiple yield the more general
linear combinations, such as the combination
186 VII. Linear Algebra

of n vectors Vi E V with corresponding scalar coefficients ai E F. All the

linear combinations of n given vectors VI , ..• ,Vn form a subspace of V,
the subspace spanned by V. Linear independence of the VI , ••• , Vn means
that 'J.aivi = 0 only when all the coefficients ai are O. Moreover, the
Vi , ... , Vn form a basis of the finite dimensional space V when every vec-
tor in V can be expressed in just one way as a linear combination 'J.aivi.
The vectors of a basis are necessarily linearly independent. Any two bases
of the same space have the same number of elements, and this number n
is the dimension of the space, dim V. All these concepts depend on the use
of linear combinations.
Just as for the plane, the geometry of a vector space is expressed in
terms of its transformations-more exactly, its linear transformations-into
itself or into some second vector space W (over the same field F). Such a
transformation is by definition a function T from V to W which preserves
all linear combinations, in the sense that always

(3)

It is enough to require just the two simpler identities

T(av) = aT(v) (4)

stating that T preserves sums (is additive) and that T preserves scalar mul-
tiples (is homogeneous). If both S: W -+ U and T: V -+ Ware linear, so is
their composite S· T: V -+ U. A linear transformation T: V -+ V of V into
itself is a linear endomorphism. It is said to be non-singular when it has a
two-sided inverse T- I : V -+ V. This inverse (unique as always) is neces-
sarily linear.
This "transformation" language serves to codify the many cases of
Mathematical operations which are additive and homogeneous.
In calculus, the operations of differentiation and integration are linear.
Thus on an interval of the real axis, say the unit interval
I = {x I 0 < < x I}, consider the set Coo = C[oo of all those functions
f: I -+R which have continuous derivatives of all orders. This set Coo is a
vector space over R with addition and scalar multiple defined in "term-
wise" fashion by the equations

(f + g)(x) = f(x) + g(x), (aj)x = a(f(x» (5)

for all x in I. The operation of differentiation, say DF = df /dx, is then a

transformation D: Coo -+ Coo which is linear according to elementary rules
of the calculus ("The derivative of a sum is the sum of the derivatives").
The operation of integration, sending each f to f~f(x)dx, is also linear
("The integral of a sum"). This space Coo is clearly very large; it has no
I. Sources of Linearity 187

finite basis and so is infinite dimensional. It is just one of many examples

of a function space (a space whose elements are functions of some
specified type). It has turned out that it is very illuminating to use the
geometric language of vector spaces for this and many other such function
spaces.
For harmonic motion, we observed in §VI.ll that all solutions for the
differential equation d 2xldt 2 = -k 2x were uniquely linear combinations
of two solutions, sin kt and cos kt. This differential equation for harmonic
motion is a homogeneous linear differential equation with constant
coefficients (i.e., a linear combination, over R, of derivatives
(dnx Idt n , ... , dx Idt, x). For such an equation of order n, it is similarly
the case that all Coo solutions are linear combinations of n linearly
independent solutions.
In the calculus for functions of two variables x and y, we have seen in
§VI.9 that the chain rule suggests two different vector spaces attached to
each point in the (x,y)-plane: The tangent space, consisting of all the vec-
tors tangent to the various possible paths through the point, and the
cotangent space, consisting of all the differentials of smooth functions at
this point. The relation between these two spaces is one source of a notion
of "dual" vector spaces (§4), and is a background on which much higher-
dimensional geometry is developed. If we think of a path as the trajectory
of a moving particle, it becomes clear that the tangent space is suggested
by the development of mechanics, as we will see in more detail in Chapter
IX below.
For algebra, vector space dimensions provide a measure of size. Thus
when a polynomial f(x) with coefficients in F is irreducible over F (has no
proper polynomial factors over F), it is always possible (§ 10) to construct
a field K = F( a) generated by F and by one root a of f( x); this field is
unique up to an isomorphism of fields leaving all elements F fixed. This
field K can be viewed as a vector space over F (vectors are elements of K,
added as there, and multiplied by elements of F). As such, its vector space
dimension is precisely the degree of the given irreducible polynomial f(x)
(Special case: The complex numbers C arise from R by "adjoining" a root
of x 2 + 1, and form a two dimensional vector space over R). This use of
dimensions has already been noted in §V.7 for the Galois group, whose
order is the dimension of the splitting field. Similar dimensions occur
throughout the study of fields (for example, in the determination of all
finite fields) and in the study of quaternions and other rings containing a
field (often called linear algebras over that field).
The notion "vector space" thus organizes diverse examples. Given all
these examples it is remarkable that the axioms for a vector space (though
stated by Peano in 1888 for spaces over R) were not generally recognized
by Mathematicians until Weyl's 1913 publication of his book on relativity
(where he needed to use vector and affine geometry).
Our formulas (1) and (2) describe V as a left vector space over the field
F, in the sense that the scalar multiples, a,b,. .. appear to the left of the
188 VII. Linear Algebra

vectors. The same vector spaces can be described with scalar multiples on
the right. It will be convenient to use such right vector spaces in the next
sections.

2. Transformations versus Matrices

There are two styles of doing linear algebra: Geometrically (or "invari-
antly") with linear combinations of vectors, or explicitly, with coordinates.
These coordinates depend on a choice of basis. In detail, if the (right) vec-
tor space V has a finite basis of vectors U\ , ... , Un, then every vector v in
V has a unique expression as a linear combination

of these basis vectors Ui with scalar coefficients Xi (here written on the

right). These scalars are the coordinates X = (XI, ... , xn) of the vector v
relative to the basis Ui. One can say that the vector "is" the n-tuple X of
numbers Xi, so that V is the space F n of these n-tuples. One can then
express all properties of vectors in term of these numbers. Thus a linear
endomorphism T: V ~ V is completely determined by giving the images
n
TUj = ~Uiaij (1)
i=1

of the n basis vectors Uj ; hence T is determined by the array

A (2)

of coefficients-an n X n matrix. By linearity TCi.ujxj) = 'i.i,jUiaijXj so

the coordinates X'i of the transformed vector v' = Tv are given by the
equations
n
X'i = ~aijXj' i = l , ... ,n. (3)
j=\

This is the familiar way in which the square matrix X also represents the
transformation T in terms of coordinates (or "variables") Xi; in particular,
the j th basis vector U j is sent into the j th column of the matrix A.
The coordinate representation (3) and the basis representation (1) of T
are equivalent. however, the matrix A for T depends on T and on the
2. Transformations versus Matrices 189

choice of basis. This leads to a central question: When do two different

square matrices represent the same linear transformation (§ 12 below)?
The practical issue of solving simultaneous linear equations is another
source of linearity: Given x '\ , ... , x' m and the m equations (3)-this
time with i = 1, ... , m-to find all the x j' if any, to the given x 'i. These
m equations now represent a linear transformation T: V ~ W from the n-
dimensional space V of the (Xi, . . . , X n ) to the m-dimensional space Wof
the X'i. Familiar facts about the solutions now translate to observations
about this transformation and its rectangular matrix. Thus the image of T
is the subspace 1m T of W consisting of all the vectors Tv for v in V; that
is, the set of all resulting values (x '\ , ... , x' m); its dimension is the rank
of T (and of its matrix A). The null-space of T is the subspace Null T of V
consisting of all those vectors v of V with Tv = 0; its dimension is the
nullity of T (and of its matrix A). An elementary construction of bases
shows that these dimensions ("dim") satisfy

n = dim(domain T) = dim(lm T) + dim(Null T). (4)

This equation in dimensions subsumes the facts about solutions of n

simultaneous linear equations. For given x 'j, the equations (3) have a
solution if and only if the vector (x '\ , ... , x' m) lies in 1m T. In particu-
lar, if n = m and rank A = n (i.e., the rows of A are linearly indepen-
dent), then the solution always exists and is unique-and the solution is
given, via the usual determinants, by Cramer's rule. For n homogeneous
equations (i.e., when all X'i = 0) the number of linearly independent
solutions (the nullity) is n minus the rank. These results provide another
example of the translation of algebraic calculations into geometric con-
cepts.
Products of matrices arise from composites of the corresponding
transformations. If the square n X n matrix B represents a second linear
S: V ~ V by SUk = ~Ujbjb then calculation with (2) shows that the com-
posite transformation T·S (first apply S, then T) has
n n
TSUk = ~Ui ~aijbjk' (5)
i=\ j=\

So the matrix C of the composite transformation has entries

Cik = ~jaijbjk-the ith row of A times the kth column of B. This explains
the familiar row by column definition of the product C = A· B of two
square matrices. This matrix product is associative because the composi-
tion of transformation is associative, but this can also be verified directly
from the row by column definition. In this way the "idea" of composition
is realized for computations by the matrix formalism-especially handy for
numerical calculations, for example for "sparse" matrices (those with only
a few non-zero entries).
190 VII. Linear Algebra

Rectangular matrices have corresponding products. A linear transfor-

mation S from an I-dimensional space to an m-dimensional one plus a
choice of basis in each space yields an m X I matrix B (m rows, I
columns). When T is a transformation from this m-dimensional space to
one of dimension n, with an n X m matrix A, the matrix product AB
representing the composite T·S is defined row by column, as in (5). Again
a triple product of matrices, when defined, is associative; and there is a
suitable distributive law for the addition of matrices. For these calcula-
tions it is convenient to think of a vector X as a column-an n X 1
matrix, then the action (3) of a matrix A on a vector X can itself be writ-
ten as a matrix product X' = AX. (Here the matrix A appears on the left,
just as the corresponding transformation and as for functions generally;
this is why we used scalars on the right, to balance functions on the left.)
The columns of a matrix A can be changed to rows by defining the tran-
spose matrix A T to be the matrix with entries (A T)ij = A j i. Then
(ABl = BTAT.
The whole subject can be developed exclusively in terms of n-tuples X
(though this popular approach misses the geometric invariance, and so
loses some understanding). Then the typical finite dimensional vector
space F' consists of all such n-tuples (all n X I columns) with the evident
addition and multiplication by scalars, and a subspace is a collection of
n-tuples closed under these operations. An n X n matrix A is non-singular
if and only if its rows are linearly independent. It is equivalent to require
that the columns be linearly independent, or that there exists a matrix B
with product BA = J, the identity matrix. Then B is unique, hence is the
inverse A -I of A, and A - IA = J = AA - I. When A is non-singular, the
linear equations (3) have a unique solution X for given X'; to wit,
X = A -IX'. The (usual) determinant I A I of a matrix A satisfies
IAB I = IA I I B I ' so is a homomorphism (of multiplication) to F.
Under composition, all the non-singular matrices with entries in a given
field F clearly form a group, the general linear group GLn(F). Equivalently,
it can be described as the group of all non-singular linear endomorphisms
T: V -> V of an n-dimensional vector space over F. Geometrically, it con-
sists of all the non-singular affine transformations of space which leave a
point (the origin) fixed-it contains contractions, expansions, shears,
reflections, and rotations about the origin. Linearity is a common formal
property of these quite various transformations.
Matrix multiplication is not commutative, because the composition of
linear transformations is not necessarily commutative. In the 20th century,
matrix multiplication had triumphant (and unexpected) applications in
quantum mechanics. However, the ideas underlying matrix theory first
developed from the study of elasticity or at least from the use of many-
variable calculus to get linear approximations. Thus consider a small
smooth deformation R3 ->R3 of 3-space, keeping the origin fixed. Each
point with coordinates x, y, z is deformed to a point x', y', z'; these
3. Eigenvalues 191

latter quantities, as functions of x, y, and z, have Taylor expansions begin-

ning with linear terms
x' = allX + a12Y + a\3z +

Z' = a31x + a32Y + a33 z +

Properties of the resulting deformation then depend in the first approxi-
mation on the behavior of the array of coefficients aij, though it was a
long time before Cayley in England and other algebraists on the continent
took the step of detaching the coefficients A = (aij) from the environ-
ment. Again in this case, an abstract idea of matrix manipulation arises
from more concrete problems.

3. Eigenvalues
A matrix description of any mathematical object depends on a choice of a
basis, and so requires attention to the effect of a change in that basis-a
change which replaces old coordinates Xi by new coordinates Yi, where
Yi = "i.PijXj and the square matrix P with entries Pij is non-singular
(because one can change back). Two matrices A and B which represent
the same endomorphism of V relative to (possibly) different bases are said
to be similar. Thus in old (X) and new (Y) coordinates the same transfor-
mation reads
X' = AX, Y' = BY.
The change of coordinates is expressed by the matrix equations Y = PX
and Y' = PX', so an easy calculation with matrix products gives
B = PAP-I. In words, square matrices A and B are similar if and only if
there is a non-singular matrix P with B = PAP-I. This is the formal set-
ting; we will return later (in § 12 below) to the explicit question of deter-
mining when two given matrices are similar.
What is the simplest matrix similar to a given matrix A? In particular,
can A be similar to a diagonal matrix D-one with entries (say)
AI , . . . , An along the main diagonal,

D= (1)

o
192 VII. Linear Algebra

and zeros elsewhere? The corresponding linear transformation then simply

multiplies the ith basis vector by Ai. Such a scalar Ai is called an eigen-
value of A. Specifically, A is an eigenvalue of an endomorphism T if and
only if Tv = VA for some non-zero vector v; equivalently, in matrix
language, A is an eigenvalue of a matrix A if and only if AX = XA for
some X =1= O. An eigenvector X for A is then any vector (zero or not) with
AX = XA for a scalar A. The notion of an eigenvalue is manifestly invari-
ant under change of coordinates, and the above discussion proves the

Theorem 1. An n X n square matrix A is similar to a diagonal matrix if

and only if A has n linearly independent eigenvectors.

In terms of the identity matrix I (diagonal with 1's down the diagonal)
the equation AX = XA can be written as (A - AI)X = O. Hence a scalar
A is an eigenvalue if and only if this system of n homogeneous linear
equations in X has a solution X =1= 0; that is, if and only if A - AI is a
singular matrix. But a square matrix is singular if and only if its deter-
minant is 0, and the determinant I A - AI I of A - AI for A n X n is a
polynomial of degree n in A,

This is called the characteristic polynomial of A; we have proved the

Theorem 2. The eigenvalues of A are the roots of its characteristic polyno-

mial.

Therefore by the fundamental theorem of algebra, a real (or a complex)

matrix A does have eigenvalues.
The notions of eigenvector and eigenvalue dominate much of the use of
linear algebra, especially in analysis and physics. The initially simple ideas
of linearity have remarkable ramifications. As an elementary sample, con-
sider a system of simultaneous first order linear (and homogeneous)
differential equations

dXi n
~aijxj, i = l , ... ,n (3)
dt j=1

in n (real) variables Xi' The real matrix A with entries aij has n eigen-
values AI, ... , An, possibly not all different and real or possibly complex.
If A is similar to a diagonal matrix, the corresponding change of coordi-
nates to YI , ... , Yn then transforms the equations (3) to the form

dYi
dt = AiYi, i=l, ... ,n.
4. Dual Spaces 193

In this form, some solutions are evident: Yi = C/\i t , for i = 1, ... , n

and for arbitrary constants (constants of integration) Ci . If some eigen-
value Ai is real and positive, the solution Yi grows exponentially. However,
Ai may be complex. This is a typical case in which a "real" problem in
real numbers leads to formulas involving complex numbers-here requir-
ing the exponential function eZ for complex values of the exponent z.
The characteristic polynomial (2) of a matrix A has an unexpected
property. In (2), substitute the matrix A for the (scalar) variable A. the
resulting (matrix) expression is always zero. This is the Cayley-Hamilton
theorem.

4. Dual Spaces
Vector spaces come naturally in pairs, acting on each other. Thus the
matrix product of a row by a column is a scalar; by this product, the "row
vectors" may be regarded as linear functions of the "column vectors", and
vice-versa. In §VI.lO the directional derivative of a function z = f(x,y)
along a path turned out to be a "product" of the gradient of f by the
tangent vector to the path, so the gradients (cotangent vectors) are thereby
linear functions of the tangent vectors, and vice-versa. To develop this
idea, notice first that scalar-valued functions on any set X constitute a vec-
tor space. Specifically, if F is a field, the collection F X of all functions
f,g: X --> F is a vector space when the vector operations are defined by the
equations

(f + g)(x) = f(x) + g(x), ( af)( x) = a(f( x » (1)

for all x on X and all scalars a E F. One says of (1) that the operations
are defined pointwise. For an infinite set X, this vector space FX is infinite
dimensional, but when X is finite, say with n elements, FX is isomorphic to
the familiar vector space F' of n-tuples of elements of F.
If X = V is itself a vector space over F, it is natural to consider just
those functions f: V -->F which are linear. Under the pointwise opera-
tions (1) they constitute a vector space (a subspace of FV)

V* = hom(V,F) = {f I f: V ~ F linear} (2)

called the dual space of V. This space is conceptually different from V, but
when V is finite dimensional, it has the same "size" as V. Specifically, if V
has a basis of n elements u I , . . . , un, then the value of the function f on
any vector with coordinates Xi may be calculated as

(3)
194 VII. Linear Algebra

Thus the linear function I is completely determined by its n values

/ = lUi on the n given basis vectors of V. This suggests that these /
ought to be coordinates of I. Indeed, one can define n special vectors
Ul , ... , Un of the dual space V* by setting

i = j, (4)

= 0, i =1= j,

where the values ui(ui) = 8ii , called the Kronecker 8, are the entries of
the n X n identity matrix. Then I = ~iUi' so the Ui do form an n-
element basis of the dual space V*. This basis Ui , ... ,Un of V* is called
the dual basis to u 1 , . . • ,un. (The latter, we note, is written with indices
upstairs, all so arranged that sums ~/Ui' ~UiXi over an index i have one
index up, the other down.) In any event, we conclude for the dimensions

dim V finite implies dim V* = dim V. (5)

Despite this comparison, it is not in order to identify the space V with

its dual V*, because they are conceptually different and because the
difference appears explicitly for infinite dimensional spaces. For example,
take V to be the space of all finite sequences (XO,Xl , ... ,Xn,O,O, . .. ) of
elements of F-equivalently, all functions N ..... F with only a finite number
of non-zero values. Every vector in this space is a finite linear combination
}:xiui of the sequences u i, where u i is zero except for the scalar 1 in posi-
tion i. These vectors u i thus form a (denumerable) basis for V, so any
linear function I: V ..... F is determined by its values lUi = Yi on these
basis vectors. This means that the dual vector space V* can be represented
as the space F" of all infinite sequences (YO,Yl, ... ) of elements Yi ( F.
This space V* is much "bigger" than V; for example, if F is the field of
rational numbers, V is denumerable but V* is not. This illustrates why V*
is not V.
We have "constructed" V* from V, but the two spaces can better be
handled symmetrically. The process of evaluating a function I E V* at a
vector v ( V is a function

e: V* X V ~ F, e(j, v) = I(v) (6)

of two variables, I in V* and v in V. For fixed I, this function is linear in

V-because I was defined to be a linear map I: V ..... F. For fixed v, this
function is linear in I -because this linearity is required in the vector
space operations defined by (l) on the dual space V*. All told, the above
function e of two variables is bilinear-i.e., linear in each variable.
In general, for any two vector spaces Wand V over the same field F, a
function b: W X V ..... F of two variables is said to be bilinear if and only
if b(w, v) is linear in v for each fixed w E Wand linear in w for each fixed
4. Dual Spaces 195

v. This in turn means that, for each fixed wo, b(wo, -): V ->F is an ele-
ment of the dual space V*, so that there is a map

W -?> V*, Wo ~ b(wo, -), (7)

linear from W to V*. Similarly, b determines a linear map

V -?> W*, Vo ~ b( -,Yo). (8)

Theorem. Let Wand V be finite dimensional vector spaces over F If

b: W X V -> F is bilinear, and if for all Wo and Vo

b( wo, -) = 0 implies Wo = 0, b( -,vo) = 0 implies Vo = 0, (9)

then b determines isomorphisms W == V* and V == W* by (7) and (8).

(A bilinear b with property (9) is called a dual pairing; the theorem

then says that dually paired spaces are duals.)
PROOF. The hypothesis (9) insures that each map (7) and (8) has null
space zero; by using the basic fact that rank plus nullity equals dimension
of domain it follows that both linear maps are isomorphisms.
This abstract formulation recaptures the geometric relation between the
cotangent and the tangent spaces (§VI.9). At that point, the chain rule for
functions of two variables involved the product of the cotangent vector
(the gradient of a function) by a tangent vector (to a path). This product
is bilinear and moreover satisfies (9), so is a dual pairing. Therefore, by
this theorem, it makes the cotangent space the dual of the tangent space
(and vice versa). Thus calculus ties in to algebraic duality.
The matrix product of a row Y = (Yl , ... ,Yn) by a column
X = (Xl, ... ,Xn )-an n X 1 matrix-is a scalar ~YiXi; this is a dual
pairing; thus the one-row matrices are the duals of the one-column
matrices.
Neither picture of duality provides room for a double dual V** or for
higher duals. For finite dimensions there is no such need. The evaluation
v ~ e( -,v) of (6) turns each vector v into a linear function on the dual
space V*; that is, into an element of the double dual space V**. This map

v ~ e( -,v): V -?> V** (10)

is an isomorphism when V is finite dimensional. Put differently, the

evaluation e of (6) is a dual pairing of V and V*, so by the theorem pro-
duces an isomorphism V == (V*)*. This is called a natural isomorphism
because it is defined directly in (10), using only the given structure and
not any special choices, such as a choice of basis. For a finite dimensional
196 VII. Linear Algebra

V there is also an isomorphism V == V*, say that sending each basis vec-
tor u i of V to the corresponding dual basis vector Ui of V*-but this iso-
morphism depends on a choice of basis, so is "unnatural".
Matrices work backwards or forwards. Thus a rectangular m X n
matrix A gives on column vectors X a linear transformation X f-> AX of n
space to m-space. Equally well, on row vectors Y it yields a different
linear transformation Y ~ YA of m space to n space. Alternatively, A for
basis vectors comes on the right in (2.1), for coordinates, on the left in
(2.2). This may be explained by duality. If T: V ~ W is a linear transfor-
mation (say, the transformation on column vectors given by the matrix A),
then each vector I in the dual space W* is really a linear map f: W ~ F,
hence the composite IT: V ~ F is an element of the dual space V*, so that
I f-> IT is itself a linear mapping

T*: W* ---7 V* (11)

-with direction opposite that of T. Moreover (TS)* = S*T*; one says

that the formation of the dual applies to spaces V f-> V* and to their
maps, as T f-> T*, is a contravariant functor (of spaces and maps).
The construction of V* can be formulated more generally. The addition
(term by term) of two rows of the same length suggests the addition, entry
by entry, of two matrices of the same size. More conceptually, for given
vector spaces V and W, two linear transformations T, T': V ~ W have a
sum T + T': V ~ W by the (pointwise) formula

(T + T')v = Tv + T'v.

With pointwise multiplication of each T by scalars, this makes the collec-

tion of all the "homomorphisms" T: V ~ W into a vector space, called

hom( V, W) = {T I T: V ---7 W linear}. (12)

In case W is just the base field F, regarded as a vector space over itself,
this space hom( v,F) is just the dual space V*. If V and W have dimen-
sions m and n, a choice of basis in each replaces each T by a matrix and
so proves the vector space hom( V, W) isomorphic to the space of all
n X m matrices with entries in F. Its dimension is then nm, with a basis
(say) those matrices which have one entry 1 and all other entries zero.

5. Inner Product Spaces

For many geometric purposes, linear vector spaces are not enough. For
example, a linear transformation does carry straight lines to straight lines,
but it can distort angles, change circles into ellipses, and generally alter
distances. However, in discussing trigonometry and the rotations of the
5. Inner Product Spaces 197

plane, we already saw in §IV.9 the utility of the inner product of two
plane vectors in describing angles and distances. It turns out here (as in so
many other cases) that the two-dimensional case is typical: One can get
the appropriate "Euclidean" geometry in an n-dimensional space simply
by adjoining to the vector space structure the additional structure of an
inner product.
An inner product space E is thus defined to be a vector space E over the
field R of reals with an inner product-a real valued function U'v of two
vectors u and v-which (as in equations (10), (11), and (12) of §IV.9) is bi-
linear, symmetric, and positive definite. Such spaces are at hand. Thus for
each dimension n the space Rn of n-tuples (XI, ... ,xn) of real numbers
has the "standard" inner product

(XI, ... ,XIlHYI , ... ,Yn) = xIYI + ... + xnYn· (1)

There is a Hilbert space z2 consisting of all those infinite sequences of

(Xlo ... ) of reals for which the sum of the squares is convergent. Here "':ix?
and "':iy? convergent imply the convergence of the inner product

Similarly, analysis uses the space L 2 of equivalence classes of all those

real-valued functions f (say on the unit interval) which are Lebesgue
square integrable (§VI.ll). The inner product in this space is the Lebes-
gue integral

Sof(x)g(x)dx.
I
f . g = (3)

With any inner product, the length or the norm 1 u 1 of each vector can
be defined by the equation

1U 12 = U • u, 1u 1 > 0, (4)

while the angle 0 between two vectors u and v is determined by a variant

of the law of cosines as

u . v = I u I I v I cos 0; (5)

in particular, u is perpendicular (orthogonal) to v when u·v = 0). If this

angle 0 is to make sense, one must have the Schwarz inequality: For any
two vectors the norm of the inner product is

lu,vl:>;;; lui Ivl· (6)

This can be proved from the axioms of an inner product space. Moreover
one can define a distance between two vectors u and v-that is, the dis-
tance between their end points-as I u - v I . From the Schwarz inequal-
198 VII. Linear Algebra

ity it follows that this distance satisfies the standard triangle inequality.
Hence, with this distance, the inner product space is a metric space-as it
should be. In fact, in Rn with the standard inner product (l), the distance
between two n-tuples X and Y is just

we have returned to the starting point of geometry with the Pythorgorean

theorem-now in n dimensions.
For an arbitrary basis u I, ... , un of a finite dimensional inner product
space, the inner product of any two basis vectors ui·u j is a scalar gij, and
these scalars determine the inner product of any two vectors in terms of
their coordinates as

[~UiXi]' [~ujyj ] = ~ijXiyj; (7)

l,j

the right hand side is the coordinate expression of the inner product as a
"bilinear form" in the x's and the y's. However, for the standard inner
product the matrix G = (gi j ) of this form is just the n X n identity
matrix. This can be arranged in any finite dimensional inner product
space, by choosing c normal orthogonal basis-one in which the inner
product ui·u j is the Kronecker ~ij. This requirement means geometrically
that each basis vector has length I and any two are perpendicular (ortho-
gonal). Any given basis vi, ... , vn can be made orthogonal as follows:
First tip v2 to be perpendicular to v 1 (by adding to v2 an appropriate mul-
tiple of v I). Then tip v3 to be orthogonal to v 1 and v2, and so on. Finally
shrink each of the resulting vectors to a length 1. This is called the
Gram-Schmidt process. Similar uses of normal and orthogonal bases crop
up for infil'ite dimensional spaces in Fourier series, Hilbert spaces, and
elsewhere. The idea is transportable from geometry to analysis.
The inner product is automatically a dual pairing of the space with
itself. Hence each finite dimensional space with an inner product is iso-
morphic to its dual space-because each vector v is also a linear function
"inner product with v". Hence for such spaces one can drop the distinc-
tion between a space and its dual.

6. Orthogonal Matrices
In a space with an inner product one has all the concepts of Euclidean
geometry: One can define spheres and rigid motions. For such a space E,
the appropriate endomorph isms are the orthogonal transformations; those
functions T: E --'>E which are linear and which preserve the inner product,
in the sense that Tu· Tv = U· v for all pairs of vectors u, v. This means also
6. Orthogonal Matrices 199

that T preserves distances and angles, and carries orthonormal vectors to

such. Indeed, if T carries anyone normal orthogonal basis of E into a
normal orthogonal basis, then T is an orthogonal transformation-because
the inner product is computed by the standard formula (5.1) from the
coordinates relative to this basis.
A square matrix A is orthogonal if and only if the corresponding
transformation is orthogonal (relative to the standard inner product (5.1».
Since this transformation on column vectors X is X I--> AX, it carries the
successive basis vectors into successive columns of A. Hence the matrix A
is orthogonal if and only if its columns are normal and orthogonal-each
of length I and any two orthogonal. If A T is the transposed matrix, this is
exactly the statement that the matrix product ATA is the identity matrix I,
so that the transpose A T of A is its left inverse. But such a left inverse is
also a right inverse. Hence a square matrix A is orthogonal if and only if

(1)

its transpose is its inverse. Since the determinant of a product is the prod-
uct of the determinants, this equation implies that an orthogonal matrix
has determinant ± I And if the orthogonal matrix A has a real eigenvec-
tor X *" 0, the equation AX = X\ and preservation of length implies that
A = ± I-the only possible real eigenvalues of an orthogonal matrix.
In two dimensions, an orthogonal transformation is a rigid motion leav-
ing the origin fixed, so must be either a rotation or a rotation followed by
a reflection. Hence the only 2 X 2 orthogonal matrices are ±Al/ where

All = [cos (} -sin (} 1 (2)

sin {J cos {J

is the matrix of a rotation through the angle {J about the origin, just as in
equation (IV.9.6). The characteristic polynomial of this matrix is
A2 - 2A cos {J + I; hence the eigenvalues are A = cos {J ± i sin {J-
non-real complex numbers unless {J = 'TTk for some integer k. Thus,
except for the identity rotation and the half-turn, there are no (real)
eigenvectors-no vectors rotated into a multiple of themselves. The eigen-
values are complex numbers of absolute value 1.
This fact is general. The eigenvalues of an orthogonal matrix A (and
hence of an orthogonal transformation) are always complex numbers of
absolute value 1. To see this, we treat the real-number entries in A as
complex numbers, and let A act on the complex vector space enof
columns X of n complex numbers. In this space en,
the inner product
should be given by the formula

(Xl,X2, ... , Xn)-(Yl , ... ,Yn) = X1YI * + ... + xnYn * (3)

200 VII. Linear Algebra

where y* denotes the complex conjugate of y. This inner product is bi-

linear and positive definite, because yy* ~ O. However, it is not sym-
metric. Instead it is antisymmetric, in the sense that complex conjugation
gives

(v·u) = (u·v)*

for any two vectors u and v. A vector space over C with such an inner
product (bilinear, positive definite, and antisymmetric) is called a unitary
space. The theory of these "unitary" spaces over the complex numbers is
wholly analogous to the theory of inner product spaces over the reals. In
particular, a linear transformation is unitary if and only if it preserves the
inner product, and an n X n matrix A of complex numbers is unitary if
and only if its conjugate transpose is its inverse. It follows that the eigen-
values of a unitary matrix (a transformation) all have absolute value l. In
particular a real orthogonal matrix is necessarily unitary-and hence the
result cited above about its eigenvalues. This is again a case where prob-
lems about real numbers require complex numbers for a solution.
From the definition it follows that the n X n orthogonal matrices form
under multiplication a group, called the orthogonal group On. Equivalently,
it can be described as the group of all orthogonal endomorphisms of an
n-dimensional inner product space. This group includes reflections which
invert the orientation of space, where orientation can be described as in
§III.8. The matrices of such transformations have determinant ± I. The
proper orthogonal matrices (those with determinant + I) form a subgroup
SOn of On, called the special orthogonal group.
An orthogonal transformation may also be described as a rigid motion
which leaves the origin (the vector 0) fixed. Just as in the plane, the most
general rigid motion of an n-dimensional inner product space is an
orthogonal transformation followed by a translation.
To summarize, we see that the geometric ideas of space, transformation,
and rigid motion extend naturally beyond two and three dimensions (and,
for analysis, into infinite dimensions) and have an effective algebraic for-
malization by vector spaces, linear orthogonal transformations, and their
representations by matrices.

7. Adjoints
As we have seen, any linear transformation T: E ..... E has a dual
T*: E* ..... E*. But if E is an inner product space, E can be identified with
E* by regarding each vector v of E as the linear function v· -, "inner
product with v". Then T*v, by definition of the dual T*, is just the compo-
site function v· T. In other words,

(T*v)· u = v· Tu (1)
7. Adjoints 201

for all vectors U in E. Since a vector is determined by giving its inner

product with all other vectors u, this equation defines T*v and hence
defines the linear transformation T*. It is called the adjoint of T. The
definition (1) is often written with a bracket notation for inner product, as

<T*v,u> = <v,Tu>. (2)

Now suppose that T is given by a matrix A with entries aij' relative to

some normal orthogonal basis Ui' Then, as in (2.1), TUj = ~Uiaij.
According to the definition (1)

(T*Uk)'Uj = Uk·Tuj = uk·~uiaij = akj'

i

because Uk'Ui = l}ki. For the same reason,

j ,k = I, ... , n.

Therefore T*Uk = ~uiaki' In other words, T* is given by the transposed

matrix AT whose ij entry is aji. In short, adjoint transformations have
transposed matrices (relative to orthonormal bases). This is the intrinsic
meaning of a transposed matrix.
Adjoints probably appeared first in the study of homogeneous second
order linear differential equations. Such an equation may be written as
L(y) = 0 where y is a (smooth) function of x, ' denotes d/dx and L is the
differential operator

L(y) = (py')' + ry' + qy,

with p, q, and r suitable functions of x. Classical texts then simply decree
that the "adjoint" operator on a function z is (Frank-von Mises (1930], p.
363)

M(z) = (pz')' - (rz)' + qy.

More conceptually, regard Land M as linear operators on a suitable

space of functions in which the inner product of functions y and z is the
integral, over some x interval from a to b, of the product function yz.
Then if y and z vanish at the ends of the integral, integration by parts
shows that

f f
b b
L(y)z = yM(z).
a a

Thus L is indeed adjoint to M in exactly the sense (1). Corresponding

notions for partial differential operators involve Gauss-Green-Stokes.
These ideas turned out to be essential in the applications of Hilbert space
to quantum mechanics.
202 VII. Linear Algebra

We return to the case of a finite-dimensional inner product space E.

A linear endomorphism T: E ..... E is self-adjoint when T* = T; that is,
when its matrix A relative to any normal orthogonal basis is symmetric
(equal to its transpose). Thus, for an inner product space E:

Theorem. All the eigenvalues of a self-adjoint T: E ..... E are real.

The proof must use complex numbers (to prove a theorem about real
vector spaces!); it will be convenient to formulate it in terms of a (real)
symmetric matrix A representing T. By the fundamental theorem of alge-
bra, the characteristic polynomial I A - AI I of A has (possibly) complex
roots; they are the eigenvalues of A. Take one such eigenvalue,\. There
must then be a non-zero and (possibly)-complex column vector X which is
an eigenvector, so that AX = X'\. Let * denote the operation which
transposes each matrix and takes the complex conjugate of each entry;
then X*A * = ,\ *X*, where X* is a row and A * = A because A * is real
and symmetric. By the associative law for matrix multiplication (row by
matrix by column)

(X*A)X = X*(AX),

or, evaluating each by AX = X,\ and X*A = ,\ *X*,

'\*(X*X) = '\(X*X).

But X*, 0, and so X*X *'

O. Hence ,\ * = '\, so that the eigenvalue ,\ is
real, as asserted.
In this proof the matrix A originally gave an endomorphism X f-> AX of
the real vector space E, but we changed it tei give an endomorphism, writ-
ten also X f-> AX, bf a complex vector space E' of the same dimension.
This change of field was done in terms of a basis of E. It is desirable to
describe the change invariantly without using a basis. This we will soon
do, in §9.

8. The Principal Axis Theorem

In the analytic geometry of the plane, ellipses and hyperbolas can be
described as the loci of equations

x2
7±17=1
i (I)

(+ for ellipses, - for hyperbolas). What about the locus of a more gen-
eral quadratic equation

ax 2 + 2bxy + ci = I? (2)
8. The Principal Axis Theorem 203

This equation can also be written in matrix form as

(3)

with a 2 X 2 symmetric matrix, call it A, in the middle. Can the equation

be so simplified that the matrix A becomes diagonal, as in (1), just by
rotating the coordinate axes? This amounts to multiplying the column vec-
tor of coordinates by an orthogonal matrix P; this will evidently replace
the matrix A in (2) by the matrix pTAP. But P is orthogonal, so its trans-
pose pT is its inverse P -I, and thus A has been replaced by the similar
matrix (§2) P-IAP. This means that we now think of A not as the
coefficients in the quadratic form (2) but as the matrix of a (self-adjoint)
transformation. In a moment we will prove

Theorem 1. In an inner product space, each self-adjoint T: E -'> E has a

diagonal matrix relative to a suitable orthonormal basis of E.

For the equation (2) this theorem means that with new (rectangular)
coordinates x' and y' the equation reads

where A and p. are the (real) eigenvalues of the 2 X 2 symmetric matrix

A. Then if both are positive, the equation represents an ellipse (taken rela-
tive to its "principal" axes); if A > 0 and p. < 0, it represents a hyper-
bola, again relative to "principal" axes. If both A and p. are negative or
zero, there is no (real) locus, while if /l = 0 and A > 0 the locus is a line
x' = A-1/2. In each case, the theorem has reduced the curve to principal
axes by a rotation (possibly followed by a reflection) of the coordinate
axes.
To prove the theorem for a space E of any finite dimension n, take a
real eigenvalue A of the self-adjoint transformation T. The corresponding
eigenvector can be multiplied by a suitable scalar to have length 1. We
can then make this vector the first vector of an orthonormal basis of E.
The matrix A for the transformation T relative to this new basis must then
carry the eigenvector with coordinates (1,0, ... ,0) into (A,O, ... ,0), so
the latter vector is the first column of the matrix, and the matrix has the
form

A B

o
204 VII. Linear Algebra

where B is some (n - 1) X (n - 1) matrix. But A is symmetric, so the

first row here is really (,\,0, ... ,0) and the lower corner B is also sym-
metric. Induction applied to this B then gives the result desired.
This theorem is sometimes called the principal axis theorem because of
its application to the geometry of ellipses and hyperbolas, and also to the
corresponding result in three dimensions for ellipsoids and hyperboloids
of one or two sheets. The theorem can be stated simply as a theorem
about matrices, in two different ways:
For every real symmetric matrix A there is an orthogonal matrix P with
P-1AP diagonal, or
For every real symmetric matrix A there is an orthogonal matrix P with
pTAP diagonal.
In both cases, the diagonal entries are the eigenvalues.
For the first statement, A is considered as the matrix of a self adjoint
transformation T; for the second A is considered as the matrix of a qua-
dratic form XTAX = ~XiaijXj; it then amounts to the assertion that every
such quadratic form can be made diagonal by an orthogonal change in
the coordinates X. This is a striking example of the fact that a formal
result (here, one about matrices) has two different geometric
interpretations-one for transformations, and the other for quadratic
forms.
There is still another formulation of Theorem 1. For each eigenvalue ,\
of T: E ..... E consider the subspace EA consisting of all the corresponding
eigenvectors. This subspace has at least the dimension 1, and in fact its
dimension is the "multiplicity" of the eigenvalue '\; that is, its multiplicity
as a root of the characteristic polynomial. The theorem can then be read
to say that every vector v in E can be written uniquely as a sum ~AVA of
vectors VA in EA' This is said to make the space E the direct sum of its sub-
spaces EA' In the infinite dimensional case a more complicated statement
of this character will apply; it is called the spectral theorem.
Suitable versions of all these results apply also to spaces over the com-
plex numbers. They deal not with matrices which are symmetric but with
hermitian matrices-those equal to their conjugate transpose. This case is
specifically relevant in analysis.

9. Bilinearity and Tensor Products

For three vector spaces U, V, and Wover the same field F, a function

B: U X V --3> W (I)

is said to be bilinear when the values B(u, v) are linear in u E U for each
fixed vector v and also linear in v E V for each fixed u. For example, the
9. Bilinearity and Tensor Products 205

inner product (5.7) is bilinear. It turns out that all of the bilinear functions
B described under (1) can be represented by linear functions and a single
bilinear function ® from U X V to a new space, called the tensor product
of U and V. The properties of this space are very useful for a wide range
of geometric and algebraic problems. Its existence is asserted in the fol-
lowing theorem.

Theorem 1. Given two vector spaces U and V over a field F there is a third
vector space U ® V and a bilinear function

®:UXV-"7U®V (2)

with the following property: Whenever B is a bilinear function (1) on U and

V to any third space W over F, there is a unique linear transformation
T: U ® V ~ W such that

B(u,v) = T(u®v) (3)

holds for all vectors u E U and v E V

This property (3) may also be formulated in a diagram

UXV~fI
B
tv
I
3'T,
(4)
I
t
w ,
to express the idea that the new bilinear function ® has the property that to
any old bilinear B there is exactly one linear T (dotted arrow) which makes
T ® = B; that is, which makes this diagram "commute".

The force of this theorem is to reduce as much as possible about bi-

linear functions B to linear functions T -on the grounds that we know
more how to handle the linear ones. Since all the bilinear functions Bare
obtained-and uniquely-from the one bilinear ® it is appropriate to call
this one bilinear function ® a universal one; specifically, universal for the
given spaces U and V.
Here is one employment for the tensor product space. Just as we have
constructed the dual space V* = hom(V,F) and the space hom(V, W) of
linear transformations (or matrices), so we need a space

Bilin( U, V; W)

of all bilinear functions B: U X V ~ W. But by the theorem the bijection

B I-> T of the diagram (4) reduces this space to one previously con-
structed, by an isomorphism of vector spaces
206 VII. Linear Algebra

Bilin( U, V; W) == hom( U 0 v, W). (5)

This result is sometimes formulated by saying that bilinear functions are

represented by U 0 V; in some sense they are "objectified" by U 0 V.
There are similar representations for trilinear functions.
Before we indicate how to construct the desired universal bilinear func-
tion 0 of (4), let us observe that it will not matter how it is constructed,
provided only that it has the property stated under (4)-in other words,
the formal property matters, and not the explicit elements of the desired
space U 0 V. More exactly, suppose that we had two such universal bi-
linear functions 0 and 0 into (say) two apparently different spaces such
as U 0 Vand U 0 V. Then we have a diagram
u x v --®.::....--... u ® V---------.. . S'T
tI ' - - - - - - -.... /
I I
SI IT (6)
I I
I I

D
u x v ----"=--.... u"D v
with both 0 and 0 bilinear and universal. Now since 0 is universal, it is
universal for 0, and so there is a linear T as shown, with T 0 = O. Vice
versa, since 0 is universal, there is a linear S as shown, with SO = 0.
This means that the composite S·T: U 0 V ..... U 0 V has (S·1)0 = 0.
But the identity map I of U 0 V also has this property I 0 = 0, while
the definition of the universality of 0 said that there is only one such
linear map. Therefore S· T is the identity of U 0 V. By exactly the same
argument T·S is the identity (of the second space U 0 V). In other
words, S is a two-sided inverse for T, so T is an isomorphism. The two
spaces U 0 V and U 0 V are isomorphic-and by an isomorphism which
carries the universal bilinear 0 in the first space into the 0 for the second
space.
Note first that this is really not an argument just about the universal
bilinear function, but about a universal what-not. Another example has
already appeared. In §V.S, the free product A *G of two groups was
defined to have as elements certain words algi ... angn in elements of A
and of G; then there were two group homomorphisms k l : A ..... A *G+-G
and k2 which were "universal" among all pairs of homomorphisms from
A and G to a third group H, as in (V.S.4)

A --?> H <E- G.

The argument we have just given for the tensor product shows that the
free product A *G with the homomorphism kl and k z is determined
uniquely, up to an isomorphism respcting kl and k2, by this universal
property. In other words, what matters about the free product A *G is not
its specific construction by means of words, but its universal property.
9. Bilinearity and Tensor Products 207

Recall also that the (direct) product A X G of two groups with its projec-
tions A --A X G~G was also shown to have a universal property-see
(V.S.2). In other words, the product of groups (or of spaces) need not be
constructed from ordered pairs of elements. Any other construction will
do, provided only that it yields the universal property!
Now we can return to formulate our construction for the tensor product
® of spaces, as in (2). Suppose for instance that U and V are finite dimen-
sional, with bases say UI , . . . , Urn and VI , . . . , Vn . First take mn symbols
Uj ® Vj for i = I, ... , m and j = I, ... , n and manufacture a new
vector space U ® V == pnn over F with these symbols as basis. Its vectors
are then formal linear combinations "J:.ai/ui ® Vj). Now consider any bi-
linear B to any space W. By its bilinearity, the values of B are all deter-
mined once one knows its values B( Uj, Vj) on the basis vectors. Now if we
define T by

T( ~aij( Uj *X Uj» = ~aijB( Uj, Vj)

i,j

we get a linear map T: U ® V ~ W with the desired property that

T® = B; clearly this is the only such T with this property.
This construction used arbitrarily chosen bases, but by earlier remarks,
the same construction with any other bases would give an isomorphic
result. The construction also gives the dimension for the tensor product
(of finite dimensional spaces) as

dim(U ® V) = (dim U)(dim V). (7)

The construction also shows that the universal bilinear function ®

yields "product" vectors U ® v in the tensor product. These product vec-
tors certainly span U ® V, but there are in general many vectors in
U ® V which are not just products U ® v, but linear combinations of
such products.
For an infinite dimensional space U (or V) much the same proof will
give the existence of the tensor product U ® V-once one has proved
(using Zorn's lemma) that every vector space has a (possibly infinite)
basis-one such that each vector in the space is uniquely a linear combina-
tion of a finite number of basis vectors. A more invariant construction of
U ® V, done without any use of bases, will appear in the next section.
The tensor product allows us to "change the field" (of scalars). For
example C is a two-dimensional vector space over the reals R. For any
real vector space V the tensor product V ® C has scalar multiples (on the
right) by C. Thus V ~ V ® C changes the real vector space V to a com-
plex one (of the same dimension)-as needed in §7 above.
Starting from one vector space V and using also its dual space V* one
may now construct all sorts of tensor spaces, such as the space
V* ® V® V-of mixed tensors with one contravariant and two covariant
208 VII. Linear Algebra

indices. Once given a basis u 1 , •.. ,un of V one then has the dual basis
Ul , . . . ,Un of V* and therefore, in the above case, a basis for
V* ® V ® V consisting of n 3 vectors Ui ® u i ® Uk. One may then exhi-
bit a tensor by its coordinates relative to such a basis. One may readily
derive the (many indexed) formulas for changing coordinates from one
basis to another. They were once used to define what a tensor is, but this
we have avoided by the invariant definition of the tensor product itself.
For an inner product space V = E, there are also formulas for shifting
upstairs indices downstairs-on the basis of the canonical isomorphism
E == E*.
There are many other useful properties of tensor products. For exam-
ple, if T: U -+ U ' and S: V -+ V' are two linear transformations, then
< u, v> f-> Tu ® Sv is a bilinear function U ® V -+ U I ® V '; hence there
is by universality a corresponding linear function

T ® S: U ® V -+ U ' ® V' (8)

with (T ® S)(u ® v) = Tu ® Sv. In other words, the tensor product

operation applies not just to spaces, but also to the linear transformations
between them. It is said to be functorial (Chap. XII).
The triple tensor product U ® (V ® W) of three vector spaces pro-
duces a trilinear function (u, v, w) f-> U ® (v ® w) which is clearly univer-
sal among trilinear functions on U X V X W. Now, reassociating
parentheses, (u ® v) ® w is also a univeral trilinear function to the space
(U ® V) ® w. Since the universal property determines the space up to
an isomorphism, there is an isomorphism

U®(V® W) ==(U® V) ® w, u ® (v ® w) f-> (u ® v) ® w (9)

It is convenient to identify these two spaces by this isomorphism-then the

tensor product of vectors has the handy property of associativity.

10. Collapse by Quotients

In this chapter, our subject is linear maps. One needs means to construct
them; one such is the process which collapses a vector space to a quotient
space.
The study of congruence of integers modulo an integer m leads (as
described in §II.8) to the construction of the ring Zm of integers modulo
m, with the corresponding homomorphism (or projection) Z -+Zm from
the ring Z of integers. The effect of this homomorphism is to map all the
multiples of m to zero-in effect, to collapse the set of all multiples of m.
Similarly for groups: The factor group G/ N collapses the elements of any
given normal subgroup N.
10. Collapse by Quotients 209

A corresponding process applies to a vector space V. If one wishes to

ignore a particular vector Uo and its multiples, one may map V into a
newly constructed space in which exactly the subspace consisting of these
multiples is mapped to zero. For example, in cartesian 3-space the vertical
projection onto the x - y plane collapses exactly all the multiples of the
unit z-vector. The same process works more generally to collapse any sub-
space S of V. There is a new space V / S, called the quotient space, and a
linear transformation P: V ..... VIS with P(S) = 0 and the following
universal property: Any linear transformation T: V ..... W into a third vec-
tor space W for which T(S) = 0 can be written as a composite T = T' P
for a unique linear transformation T': ViS ..... W, as in the following com-

vt,
mutative diagram:
p

v~. P(S)<O,

T 1 (1)
I
t
W , T(S) = o.
This may be proved by taking P to map each vector v of V into the
"coset" (or hyperplane) S + v consisting of all vectors s + v for s in S;
these cosets themselves form a vector space under operations such as the
addition given by (S + v) + (S + w) = S + (v + w). Essentially,
these cosets are just like the congruence classes Cm(a) modulo m used to
construct the integers modulo m, because each coset S + v consists of all
vectors v' in V with v' - v in S (in effect, with v' == v (mod S)). How-
ever, once the universal property (1) is established, it no longer matters
how the elements of this quotient space V / S are described, because it is
this universal property which formalizes everything about the quotient
(and which proves that the quotient is unique up to an isomorphims of
vector spaces).
This result allows one to collapse any given collection of vectors
Sl , . . . ,Sk of a space V-simply take the subspace S to be the set of all
linear combinations alsl + ... + aksk of these vectors; that is, the
subspace spanned by the Si; clearly any transformation T which maps all Si
to zero must map to zero all of the subspace.
As a typical example of the use of this process, let us describe another
construction of the tensor product U ® V of two vector spaces U and V.
First take all pairs < u, v> of vectors u f U and v f V, and write each pair
as if it were a product U·V. Form the (very large) vector space L which has
all these pairs U·V as its basis; thus L consists of all finite linear combina-
tions of symbols Ui'Vi with coefficients in the field F. Now <u,v> f-> U·V is
a function F: U XV ..... L, but it is by no means bilinear. One can force it
to be bilinear simply by collapsing in L all the things which would be zero
if F were bilinear-thus collapsing all elements
210 VII. Linear Algebra

(2)

and similarly for linearity in the second factor v. Now F becomes a bi-
linear functions U X V -->LIS into the collapsed space. One may verify
that it is universal, so that this construction LIS is indeed a tensor prod-
uct.
This example illustrates a use of infinite dimensional spaces to construct
finite dimensional ones-and also shows that there are many different for-
mal ways to construct what is effectively the same tensor product.
Mathematics uses many other types of "collapse" for other structures.
In each case one must note first what can be collapsed; that is, what can
be mapped to zero by a homomorphism: For vector space, any subspace;
for groups any normal subgroup. For a ring R, define an ideal of R to be
a subset K c R such that k( and k2 in K imply k( + k2 in K, while k in
K and r in R imply that rk and kr are in K. In any homomorphism
f: R -->S of rings, the subset of R consisting of all k E R with fk = 0 is
necessarily an ideal. This ideal is called the kernel of the homomorphism.
Conversely, given any ideal K in a ring R, there is a ring R / K and an
homomorphism R --> R / K with kernel K which is universal among
homomorphisms from R with kernel containing K. This quotient ring may
be constructed by taking its elements to be cosets K + r of K. This con-
struction includes the case of integers modulo m, because the set of all
multiples of m is an ideal in the ring Z of integers. It also includes one of
the constructions of the complex numbers C from the field R of real
numbers. For this, first form the ring R of all polynomials in a symbol x
with real coefficients. In this ring, the set of all multiples of the polyno-
mial (x 2 + I) is an ideal K, and the quotient ring R / K is (isomorphic to)
the complex numbers-because the collapse of x 2 + I to 0 forces x (when
collapsed) to satisfy the equation x 2 = - I defining the basic complex
number i.
This process also makes Galois theory possible without any recourse to
complex numbers. Over any (abstract) field F the polynomials in x form a
ring R = FIx] in which the multiples of anyone polynomial f(x) form an
ideal (f). If f is irreducible, the quotient ring R/(f) is a field containing
F and an element ~ (the coset of x) with f(~) = O. In short, R/(f) is a
field F(~) generated by F and one root of f; the remaining roots can be
adjoined in the same way to give an (abstract) splitting field and its Galois
group.
This notion of ideal, useful for collapses, will turn out to have other
uses in number theory and arithmetic (§XII.3).

11. Exterior Algebra and Differential Forms

For differential forms and the relations (Gauss-Green-Stokes) between
surface integrals and volume integrals we needed to manipulate
differentials dx and dy according to the formal rules
II. Exterior Algebra and Differential Forms 211

(dx)(dy) = -(dy)(dx), (dX)2 = O. (1)

These are rules intended to apply to the vectors dx, dy, dz in the
cotangent space at a point, say, in Euclidean 3-space. They need not be
mysteious, because they are algebraic rules which can be achieved by suit-
able collapse of formal products taken over any vector space V. This col-
lapse leads to exterior algebra.
Starting with V, the successive tensor products yield a string of vector
spaces

F, v, V0V, V0V0V, ... (2)

Their elements are called (covariant) tensors; those in the n-fold tensor
product Tn being tensors of rank n while (for completeness) the scalars in
F count as tensors of rank O. Taken together, all these tensors form an
algebraic system with the following rules of operation: Any tensor can be
multiplied by a scalar; any two tensors of the same rank can be added;
any two tensors can be multiplied by using the product 0; for example,
the product of the tensors u and v 0 w is

u·(v 0w) = u 0v 0w. (3)

All these tensors, with these algebraic operations, form the so-called tensor
algebra T( V). This algebraic system isn't quite a vector space, because we
do not (and do not need to) add two tensors of different ranks; for the
same reason it is not quite a ring under sum and product. Technically, it is
called a graded algebra; i.e., both a graded vector space and a graded ring
(with grading by ranks); see Algebra XVI.4). The reader will find it
straightforward to write down the axioms for such a graded system, noting
that the convention (8.9) about triple tensor products ensures that the
multiplication is associative. However, the multiplication is not commuta-
tive.
Every tensor is a sum of products of vectors, so that T( V) is generated
by the vector space V; it is the "most general" graded algebra so gen-
erated, and so is called the free graded algebra over V. It thus has a
"universal" property, as follows: For any graded algebra A, each linear
transformation V -->A into the vector space of elements of grade I in A can
be extended uniquely to a morphism T( V) -->A of graded algebras. In
consequence, any such A generated by V can be obtained by suitably col-
lapsing the tensor algebra T( V).
For the graded algebra of differentials (1) we need a product with
u 0 v = - v 0 u and u 0 u = O. These identities do not hold in the
tensor algebra T(V). Hence we simply take the appropriate quotient of
T(V). This can be done by collapsing in T(V) all the elements t ® t (for
any tensor t of positive rank) and all the multiples (left and right) of any
such t 0 t by another tensor. The resulting quotient algebra E(V) is
212 VII. Linear Algebra

called the exterior algebra of V. We will write its elements as e, f, g and

the product of two such as e f. By its construction, this algebra satisfies
A

(4)

for all e of positive rank. This in its turn implies

Since eAe and fA fare 0, this gives the desired equation

(5)

again for e and f of positive rank.

To pin this algebra down, write Ek = Ek(V) for its elements of rank k,
so that Ek is a vector space quotient of the k-fold tensor product
Tdv) = V ® ... ® V. Since we have divided out only by "squares"
t ® t with t of positive rank, nothing in rank 0 or I has been collapsed, so
that E) (V) is just T) (V) = V. Now suppose that V is finite dimensional,
with some basis U) , . . . , Un. Then Tk ( V) has as a basis all k-fold tensor
products of these basis vectors, possibly with repetitions. In Ek( V) all pro-
ducts with a repeated Ui become 0 by (4), while any two different Ui can
be interchanged, by (5). Hence in Ek ( V) every element is a linear combi-
nation of the k-fold "exterior" products

(6)

of basis vectors with indices in ascending order. For k > n there can be
no such product, so Ek(V) is zero for k > O. For smaller k, it turns out
that the (k,n -k) possible products of (6) do form a basis. This one may
see more easily for n = 3, say. Here we have the list of basis elements
suggested:
Eo: I,
E) : U), U2, U3,
(7)
E2: U) AU2, U) AU3, U2 AU3 ,

E3 : U) AU2 AU3 .

All the products of these elements can be computed by the rules (4) and
(5); with these product rules one verifies that any
t = a) U\ + a2U2 + a3u3 has tAt = o. Hence the required collapse
already happened when all the elements in the list (7) are taken to be
linearly independent. Therefore they are indeed linearly independent in
E(V).
For U) = dx, U2 = dy, and U3 = dz and so V = R3 this exterior alge-
bra then does provide a formal setting for the calculations with
12. Similarity and Sums 213

differential forms in three variables (and similarly for higher dimensions!).

From this geometric origin of exterior algebra there also flows an unex-
pected byproduct. For any n X n matrix A with entries aij in a field think
of the i-th row R; as an element Ri = '2.aiUj of rank 1 in the exterior alge-
bra. Then the exterior product of all n rows computes out to

(8)

where I A I is the usual determinant of the matrix A (try it out for

n = 3). Indeed, the rules (4) and (5) simply encode the formal rules
about calculating determinants-when two columns are equal, the deter-
minant is 0 (rule (4»; when two columns are interchanged, the sign of the
determinant changes (rule (5». The consequene is that exterior algebra
can be used to provide a definition of determinants and a complete
development of their properties (Algebra, §XVJ.7). This can be done in
invariant style, defining the determinant not of a matrix A but of the
(more geometric) linear endomorphism represented by A. Here again the
tension between geometry and matrices reappears and is resolved.
Observe also that linear transformations T: V .... W will induce
corresponding homomorphism E(T): E( V) .... E( W) on the associated exte-
rior algebras; in brief, this contruction from V is functorial.

12. Similarity and Sums

Now we return to the central question of linear algebra: When do two
square matrices represent the same linear endomorphism T: V .... V, rela-
tive to possibly different bases? Or, given such a T, can one find a sim-
plest matrix representation? In case V = E is Euclidean and G is self-
adjoint, the principal axis theorem of §8 gives the answer: The matrix
may be chosen to be diagonal, with diagonal entries the (real) eigenvalues
of T. This answer will not extend to all other cases. For example, the
2 X 2 matrix

(1)

has zero as its only eigenvalue, but it is not similar to the matrix of all
zeros because it does not represent the zero endomorphism. Hence finding
a "simplest" matrix for a general T will require something much more
elaborate than just diagonal matrices.
One approach is to consider V not just as a vector space but as a
"richer" algebraic system (a "module") given by the additional presence
of T and its iterates. What this means is that any vector v of V can be
multiplied not just by scalars in the ground field F, but also by polynomi-
als in T with coefficients in F, as in
214 VII. Linear Algebra

We call this the "product" of v by the polynomial

ao + al x + ... + akxk. The set of all these polynomials in x forms a
commutative ring, the polynomial ring F[x]. the multiplication of the vec-
tor v by such polynomial scalars satisfies all the usual rules for such multi-
ples. Hence the definition (l) makes Va" module" over F[x 1in the sense
of the following definition.
A module M over a ring R is an abelian group under addition with an
operation (a, v) f-> av of multiplication for each element v E M and each
scalar a in R such that

a( v + w) = av + aw, I v = v. (3)

(a + b)v = av + bv, (ab)v = a(bv). (4)

Higher geometry and analysis present many other examples of such

"modules"-abelian groups with actions by various rings of "operators".
For instance, differential forms constitute a graded module with an opera-
tor d where d 2 = 0.
The module axioms (3) and (4) are identical with those for a vector
space over a field-except that now the scalars a in the ring R need not
have multiplicative inverses, and the multiplication of two scalars need
not be commutative. Much of the elementary theory of vector spaces car-
ries over unchanged to modules; in particular, the definition of a linear
map T: M ~N of one module into another. Given a submodule S of M (a
subset of M closed under all the operations) one can again construct the
quotient module M / S and a linear map P: M ~ M / S which sends S to
zero and is universal with this property. Given two modules M and N
over the same ring R one can form their direct sum M EEl N, consisting of
all ordered pairs of elements (v, w) with v E M, WEN, with the opera-
tions

This is a module over R; it has linear maps

M -E-- M EEl N --33- N

to the factors, given by v ~ <v, W> f-> w. They are "universal", so make
M EEl N a product of M and N in the sense used in §V.9 for groups. It also
has linear maps

M --33- M EElN ~ N

from the factors, given by v f-> (v,O) and W f-> (O,w). They are also
universal, so make it a "coproduct" -the construction which for groups
was called a free product in §V.9.
12. Similarity and Sums 215

For vector spaces over F, the direct sum of FEB· .. EB F of a number

(say n) of copies of the field F is simply the standard vector space of n-
tuples of scalars in F. Every finite dimensional vector space has this form,
once a basis is chosen. This last fails for modules: There can be many
modules over a ring R which are in no wise direct sums of copies of R.
For example, take the quotient of the polynomial ring F[x] by the ideal
(x 2) consisting of all multiples of x 2. This quotient F[x]/(x2) is still a
module over R = F[x] but it is much "smaller" than R. Every element
comes from one of the polynomials a + bx, for a and b in F. Multiplica-
tion of this element by x is a linear transformation T sending a to ax and
bx to 0 (x 2 = O!). In other words, the module F[x ]/(x 2) is just a two-
dimensional vector space over F with a linear endomorphism T acting on
the basis vectors by UI f--> U2 and U2 f--> O-and so with exactly the matrix
(1) noted above. In this way every module over the polynomial ring F[ x]
is just a vector space V over F together with a linear endomorphism
V --> V, given by multiplication by x.
Now in the polynomial ring F[x], just as in the ring Z of integers, there
is a division algorithm, from which one may prove that every ideal K in
F[x] is 0 or is an ideal K = (f(x» which consists of the multiples of a
single polynomial [(x), which might as well be taken to be monic (i.e.,
with leading coefficient I). The corresponding quotient module
F[x]/(f(x» is a vector space V over F with dimension the degree m of the
polynomial [(x). Take as basis of this vector space the successive powers
1, x, x 2 , ... , xm - I. Then the corresponding endomorphism T: V --> V,
"multiply by x" sends each basis vector into the next one, but the last
x m - I into

-am_Ix m-I

where the ai are the coefficients of [(x) = ao

+ '" + am_Ix m- 1 + xm. Therefore the matrix of T, relative to this
basis, has the appearance (for m = 4)

0 0 0 -ao
1 0 0 -al
0 0 -a2 (6)
0 0 -a3

It is called the companion matrix (of the polynomial [(x». These matrices
will be the building blocks for the analysis of similarity.
The analysis uses modules over special kinds of rings. An integral
domain is a commutative ring in which bc = 0 implies b = 0 or c = O. A
principal ideal domain is a domain D in which every ideal K is an ideal
K = (d) consisting of all the multiples (in D) of some one element d.
216 VII. Linear Algebra

Both FIx] and Z are principal ideal domains. For modules over such aD
one can prove (with considerable trouble; Algebra, Chap. X) the

Theorem. If a module M over a principal ideal domain D is spanned by a

finite number of its elements, then there is a number r and non-zero elements
d l , . . . ,dk of D, each a multiple of the next, such that M is the direct sum

The integer r and the ideals (d; ) are uniquely determined by M.

This result does include vector spaces (as modules over a field F), since
F is a principal ideal domain whose only ideals are (0) and F itself. How-
ever, the result (7) exhibits the complexity of modules relative to vector
spaces; for a finite dimensional vector space V over F one has only
V == F $ ... $ F-just the first r summands of (7), with r the dimen-
sion. For general D the additional modules D/(d;) here are called cyclic
modules because, like cyclic groups (§V.6) they are generated by one
element-the I of D.
What does this theorem mean when D is the polynomial ring F[x]? A
module M is then just a vector space V together with a linear endomor-
phism, "multiply by x". In case V is finite dimensional, there can be no
factor D = F[ x] in the decomposition (7), because each such factor D
would correspond to an infinite dimensional vector space F[x]. hence
there is only the string of cyclic modules D/(d;). Each of these
corresponds to a transformation with matrix a companion matrix (6).
Hence the

Corollary. For any linear endomorphism T: V ..... V of a finite dimensional

vector space V over a field F there is a basis of V in which the matrix of T,
displayed as

(8)

consists of companion matrices C; of polynomials d; E F[x] down the diag-

onal. Moreover, each d; is a multiple of d;+ I, i = 1, ... , k-1.

The number k and the polynomials d;, chosen monic, are uniquely
determined by the transformation T.
These monic polynomials are called the invariant factors of T. They
form a complete system of invariants under similarity: Two square
12. Similarity and Sums 217

matrices are similar if and only if they have the same invariant factors.
Correspondingly, the representation (8) is called a canonical form for a
(square) matrix, under similarity-every square matrix is similar to exactly
one such canonical form. Much more can be said (but not here) about
this "rational canonical form" and other canonical forms for matrices,
such as the Jordan canonical form.
What does the theorem mean when the principal ideal domain D is the
ring Z of integers? Consider any (additive) abelian group A. In A a
repeated sum (say three times, as a + a + a) amounts to the multiplica-
tion of a E A by an integer, here by 3, as 3a = a + a + a. This
"scalar" multiple enjoys the properties (3) and (4) used to define a
module. In other words, an abelian group A is just the same thing as a Z-
module.

Corollary. Any finitely generated abelian group A can be represented as a

direct sum

A=ZEB EB Z EB Z/(m]) EB ... EB Z/(mk) (9)

of a number r of infinite cyclic groups Z, plus cyclic groups of orders

m],m2 , ... ,mk where each mi is a multiple of mi+], i = 1, ... , k-\.
The numbers rand m] , ... ,mk are invariants of A.

This result, which describes all finitely generated abelian groups by

'invariants" rand mi, is especially useful in algebraic topology. There each
"homology group" A of a polyhedron is abelian and finitely generated, so
is determined by its "Betti numbers" r and its "torsion coefficients" mi'
In the canonical form (8), each invariant factor d; can be written as a
product of powers of irreducible polynomials, called the elementary
divisors of T; there is a corresponding decomposition of the companion
matrix Ci . For abelian groups, for example, each mi can be written as a
product p:1 ... p;'.of powers of primes Pi. Then just as Z6 == Z2 ® Z3,
there is an isomorphism

(10)

Combined with (9) this gives the decomposition of finite abelian groups
already announced in (V.8.4)-the direct product there is the same as the
direct sum here.
Our main observation now is that this result (9) for abelian groups and
that for companion matrices are essentially the same. Separate proofs
(often presented by mysterious manipulations of matrices) would come
down to the use of essentially the same devices, and these devices are
codified-and hence better understood-by using the concept of a module,
generalized from that of vector space.
218 VII. Linear Algebra

13. Summary
Linear algebra starts with geometrical pictures of vectors and with ele-
mentary ideas about "linear" operations-those which preserve sum and
proportion. The exact formulation of these ideas is possible only with the
notion of a linear transformation between vector spaces. The manipula-
tion of such transformations involves both conceptual formulations and
calculations with matrices and presents a central problem, that of similar-
ity: When do two matrices represent the same linear endomorphism?
Geometric and analytic considerations introduce various constructions on
spaces-dual spaces and inner product spaces. Linear functions lead to
bilinear functions and more general products, using the tensor product of
spaces. In the presence of an inner product, the similarity problem for
such symmetric matrices can be solved by the use of eigenvectors and
eigenvalues. In the general case, one of the canonical forms for matrices
under similarity requires that vector spaces over a field be generalized to
consider modules over a suitable ring. Further developments of linear
algebra, not summarized here, involve the same pattern-an interaction
between geometry and analysis, leading to successive formal generaliza-
tions helpful in formulating and understanding aspects of linearity and its
manifold uses, for linear approximations and analytic operations.

References. Survey of Modern Algebra was the first text in [1941] in

English to effectively combine the Anglo-Saxon matrix methods with the
conceptual view of vector spaces (Hermann Weyl). This connection with
Hilbert space (von Neumann) is emphasized in the classic text of Halmos
[1942]. The use of modules is emphasized in Algebra and in Hungerford
[1974]. Modern computers have made big matrices manageable, but
unfortunately many current texts of elementary linear algebra bury the
ideas under a morass of muddled matrix manipulations with no under-
standing of the concepts.
CHAPTER VIII

Forms of Space

Perceptions of space and of motions in space have led mathematicians to

describe a wide variety of formal geometrical structures. In this chapter
we will introduce a few of these structures, beginning with the description
of arc length and of various curvatures, and going on to topological
spaces, sheaves, manifolds, and the like. It will appear that the role of
intuitive ideas is very important in the analysis of such geometric
structures-and that it often is a long time before evident geometric intui-
tions are brought to a clear formal expression. These expressions provide
many different forms for the elusive idea of "space".

1. Curvature
In Euclidean and Non-Euclidean geometry, the phenomena of space were
analysed in terms of lines, triangles, angles, and congruence; in brief, such
a geometry is primarily linear. Many other geometrical phenomena, how-
ever, involve curved lines in the plane and twisted curves and curved sur-
faces in three-dimensional space. By approximating these curves by
straight lines, the methods of calculus come into play, leading to the sub-
ject of differential geometry, to which we now tum. The resulting elemen-
tary methods of analysing curvature lead inevitably to a study of the
intrinsic geometry of surfaces and to problems in the classification of sur-
faces and higher-dimensional manifolds.
A smooth path (that is, a parametrized curve) in the x,y plane is given,
as in §VI.9, by a pair of smooth functions

x = g(t), y = h(t) (1)

defined for some interval tl < <

t t2 of the real parameter t. The
tangent vector at any point p = (g(to ),h(to)) of the path is the vector

(g'(to), h '(to)) (2)

220 VIII. Forms of Space

in the tangent space Tp of the plane at that point. If the tangent vector is
never zero, the path is said to be regular. The parameter t can be changed,
say to u = k(t), provided the smooth function k has k '(t) =1= 0 throughout
the interval in t. Such a change alters the length of the tangent vector (2)
but not its direction-and gives the same set of points (x,y). This collection
of points is the curve traced out by the path (1). Thus a change of param-
eter gives the same curve, traversed at a (possibly) different speed and
with points labelled with (possibly) different parameter values.
One first wants the length of such a curve. To measure the length of this
curve from tl to t2 one inscribes a polygon in the curve, calculates the
length of the polygon by the Pythagorean theorem for each piece, and
takes suitable successive such polygons with shorter and shorter sides. In
the limit, this gives the length expressed as the (Riemann) integral

(3)

(More generally, a possibly not-sa-smooth curve is said to be rectifiable

when this limiting process leads to a definite result.) The length s is a
function of t; that is, the length from tl up to t is then the integral

(4)

then s determines t, so one may use this arc length s as a parameter in

place of t. The differential of this arc length s is then given by the qua-
dratic expression

(5)

it is the Pythagorean theorem in infinitesimal form. It can be regarded as

an inner product dx dx + dy dy in the tangent plane Tp. This inner prod-
uct structure on the tangent plane Tp applies to all the curves through p.
For curvature, the circle provides a typical example. The smaller the
radius r, the greater the curvature. Here one defines the curvature of the
circle to be Ie = llr. For a general curve the natural approach is then to
approximate the curve at each point p by a circle through that point.
Specifically, take p and two nearby points p' and pI! on the curve.
Through them there passes a (unique) circle C. Because the curve is
smooth, this circle C will approach a unique limiting position Cp ' except
when p is a point of inflection. The curvature of this limiting circle, called
the osculating circle, is by definition the curvature of the given curve at
the point p in question. From this description as a limit one can obtain an
analytic formula giving the curvature Ie at each point as a function of the
parameter value t at that point:
g '(t)h I!(t) - h '(t)gl!(t)
Ie = (g '(t)2 + h '(t)2)3/2
(6)
1. Curvature 221

Here the term in the denominator is non-zero because the parametrization

has been assumed to be regular, as defined above. For a curve given in
the familiar form of an equation y = f(x) in the coordinates x and y,
one may regard the x coordinate as the parameter; in that case the for-
mula (6) for the curvature K simplifies to

K = y"(x) (7)
[l + y '(x )2]3/2 .

However, the choice of a coordinate as the parameter has no intrinsic

geometric meaning, since the same geometric path may be referred to
different coordinates. The arc length s, measured from some starting
point, provides an intrinsic parameter, and the curvature K at each point is
more naturally expressed as a function K = K(S) of the arc length s up to
that point.
Now in view of the "intrinsic" geometric definition of arc length and of
curvature, a rotation or a translation of the curve will not alter its arc
length between two points or the curvature as a function of this length.
Hence these quantities are said to be Euclidean invariants of the curve, in
the sense that they are unaltered by the stated (Euclidean)
transformations. Much more is true: The quantities arc length and curva-
ture form a complete system of invariants in the following sense. If two
smooth curves have the same curvature function K = K(S) when the arc
lengths are measured from points p I and P'I on the curves, then there is a
rigid motion of the plane which carries the first curve exactly onto the
second in such a way that PI coincides with P'I.
For curves in space there is an additional invariant, the torsion, which
measures the rate at which the curve deviates from one lying wholly in
one plane. The definition again involves a suitable use of limits to formal-
ize the intuitive notion, while the curvature and the arc length of a space
curve are defined much as in the plane. It is then true that arc length, cur-
vature, and torsion together provide a complete set of Euclidean invari-
ants for a smooth space curve.
In space, a smooth path is given, in terms of a parameter u, by three
smooth functions of u

x = g(u), y = h(u), z = k(u).

It is often convenient to write the coordinates (x,y,z) as a single vector p,
in bold face type. Again, the differential of arc length at p is
ds 2 = dx 2 + d/ + dz 2. If the arc length s is taken as the parameter, the
tangent vector t at each point has length 1. This means that the inner
product f.t = 1 is constant. Differentiating this equation with respect to
the parameter s gives 2t·t' = 0, so that the vector t' representing the rate
of change of the tangent vector (with respect to arc length) is perpendicu-
lar to the tangent vector t. Now the unit normal vector of the space curve
222 VIII. Forms of Space

is defined to be the vector n in the direction of t " but with length 1. This
means that the derivative t' is proportional to n; indeed, t' = Kn because
the curvature K (of a plane curve or a space curve) represents the rate at
which the tangent is turning. For a space curve one may take a third unit
vector perpendicular to both the normal n and the tangent t and suitably
directed; it is called the binormal b. Thus at each point of the space curve
one has an orthonormal basis t, n, b (for the tangent space at that point).
The curvature and torsion together specify how these basis vectors (the
"frame") changes as one moves along the curve-as in the equation
t' = Kn and other such equations constituting the "Frenet formulas".
This method of "moving frames" has an immediate pictorial content
(think of the three first fingers of a hand stretched out to be orthogonal
and moving along the curve). This method applies not just to curves, but
to surfaces and beyond, and is a powerful formal tool in the use of cal-
culus to understand geometry.
In brief, the lines, circles and planes of elementary Euclidean geometry
can be used to approximate the curves (and later the surfaces) in space,
thereby making the ideas of the calculus apply in a geometric context.

2. Gaussian Curvature for Surfaces

Just as two dimensions (the plane) are decisive for the understanding of
Euclidean geometry, so two dimensions (a surface) are decisive for the
understanding of curvature. Various surfaces, variously curved, are at
hand in three-space: Spheres, ellipsoids, cylinders, paraboloids, hyper-
boloids, toruses and the like. Some of these surfaces may be described as
the locus of a single smooth equation, just as the sphere of radius r with
center at the origin is (by Pythagoras again!) the locus of the equation
x2 + i + z2 = r2. The coordinates x, y, z of these points P on the
sphere may also be represented as functions of suitable parameters. The
most familiar parameters are latitude () and longitude cf> (the names recall
the problems of navigation along the long direction in the Mediter-
ranean). They are described by the familiar Figure I or by the
corresponding parametric equations read off from this figure:

x = r cos () cos cf> ,

y = r cos () sin cf> , (1)

z = r sin ().

In the case of a curve, the choice of a parameter amounts to the choice of

a continuous map of an interval into the (set of points of) the curve. As
for the sphere with these parameters, we note that the longitude cf> is
2. Gaussian Curvature for Surfaces 223

.. x
~+-,----+---+-

Figure I. Spherical coordinates.

ambiguous at the north and south poles of the sphere; moreover It IS

periodic, with period 2'TT. Hence we had better say that the spherical coor-
dinates </> and (J in the equations (1) describe a continuous bijection of the
open square /2

-'TTi2 < (J < 'TTi2, (2)

in the (J-cp plane onto all of the sphere except the north pole, the south
pole and the "international date line" (the half great circle where cp = 'TT).
This map /2 ~S2 is called a chart on the sphere. Since the sphere is mani-
festly not the same figure as the rectangle, it is impossible to cover the
whole sphere with one such chart. Two charts will do-this one and a simi-
lar one where the "poles" are taken on the present equator. One may then
say that the sphere is obtained by pasting together two such charts (where
they overlap). Other surfaces can be similarly described by pasting
together several charts, where each chart is a map defined on an open
square like (2) (or more generally on an open set in the plane). For exam-
ple, the surface of a torus can be covered by three such charts, but no two
will cover the whole torus.
Often, a piece of a surface is given by a smooth equation z = f(x,y) in
rectangular coordinates x, y, and z; and this is just a chart with the x - y
coordinates as parameters.
The curvature of such a piece of surface should somehow measure how
rapidly the tangent planes to the surface change direction as one moves
from point to point. This change can be observed equally well by the "tip-
ping" of the normal vector (the vector orthogonal to the tangent plane).
Specifically, using the calculus and an equation z = f(x,y) as in §VII.9
one can determine at each point p of the surface S a plane tangent to S
and hence a vector 0 of length I orthogonal to the tangent plane; this vec-
tor is called a unit normal vector to S. If we choose these vectors smoothly
at nearby points of the chart we get a field of such vectors, one attached to
each point of this piece of the surface. (A reversal of sign, from 0 to - 0 ,
gives a second choice of such a local field.) Now the curvature of the Sur-
224 VIII. Forms of Space

face ought to be measured by the way in which these unit normal vectors
vary (tip about) from point to point on the surface. The idea is much the
same as that for a curve, where the curvature measures the rate at which
the tangent vector (or equally well the normal vector) turns. This intuitive
idea for measuring curvature for a surface can be formalized in several
ways.
First, one may try to reduce the question to one dealing directly with
the curvature of plane curves. To do this, consider all the planes contain-
ing the point p and the unit normal vector D at this point. They will all be
orthogonal to the tangent plane to the surface at p, and each will cut out
from the surface a plane curve which has some definite curvature K at p.
If the surface is a sphere, these "sectional" plane curves will all be circles
of the same curvature, but in general the sectional curvature K will vary as
the plane through D rotates about the axis D. This is the case, for example,
with the different curves cut out in this way at one vertex of an ellipsoid.
This example suggests that it is reasonable to concentrate attention on the
maximum curvature Kl and the minimum curvature K2 of these "sec-
tional" plane curves. These two are called the principal curvatures of the
surface at the point p and their directions (in the tangent plane at that
point) are called the principal directions at p. (In fact the principal curva-
tures can be calculated as eigenvalues of a suitable matrix, and the two
principal directions turn out to be eigenvectors orthogonal to each other.)
At any point on a convex surface (such as an ellipsoid) both the principal
curvatures will have the same sign (say, positive). However, at a saddle
point ! of a surface (for example at the origin for the surface
z = x - i) the principal curvatures have opposite signs, because one
of the sectional curves bends up, while the other bends down at p. If one
wants a single measure of the curvature of the surface at the point, one
may take a suitable combination of the two principal curvatures-for
example, the mean curvature (Kl + K2)!2. But notice that this mean cur-
vature might come out to be zero at a saddle point, although the surface
is surely "curved" there.
A different and deeper concept was developed by Gauss. On a little
piece A of the surface about the point p, take all the unit normal vectors D
at points of A and translate them so that they all start from the origin O.
The ends of these unit vectors then trace out a region B on the unit sphere
about the origin. Moreover, the more sharply the surface curves at p, the
larger the region B. Hence, assuming that areas can be measured, one
defines the Gaussian curvature of the surface at p to be a limit

K = Lim area B ( on the unit sphere) , (3)

area A (on the surface)
where the limit is taken as all the dimensions of A approach zero. (Later,
we will assign a sign to this limit.)
This intuitive description can be formalized in various ways. One such
method uses a process of "covariant differentiation", which amounts in
2. Gaussian Curvature for Surfaces 225

this case to differentiating along a vector v in the tangent plane at the

point p on the surface. Specifically, consider each of the three com-
ponents of the normal vector D, and its derivative (a real number) in the
direction v. These three derivatives are the components of a vector 'V vD,
called the derivative of the vector field D in the direction v. Since all the
vectors of the field D have unit length, the inner product n'D is I; there-
fore the derivative 'V vD will be perpendicular to D-and so will be a vector
in the tangent plane at the point p. In other words, the correspondence

is a function mapping the tangent plane Tp into itself; one then sees that
it is a linear transformation
L: Tp ---?- Tp.

With careful formulation (see for example Barrett O'Neill [1966]) it turns
out that this transformation is self adjoint (i.e., has a symmetric matrix
relative to any basis of Tp). This transformation is sometimes called the
"shape mapping" for the surface in question.
This shape mapping clearly contains information as to how rapidly the
unit normal is turning, for motions in any direction from the point p.
Hence it is not surprising that it can be shown (see e.g. O'Neill loco cit.)
that L contains all the information about the curvature of the surface at
the point p. Specifically, the determinant of the shape mapping (calcu-
lated as the determinant of anyone of the 2 X 2 matrices representing
that mapping) is the Gaussian curvature K of the surface, complete with
the desired sign. Like every self-adjoint transformation, this particular
transformation L can be brought to principal axes. It turns out that these
axes are exactly the principal directions of curvature, as defined above,
while the eigenvalues (the diagonal entries) are the two principal curva-
tures K) and K2. Since the principal axes of a symmetric matrix are ortho-
gonal, this shows that the principal directions of curvature are ortho-
gonal-as one may readily see in particular surfaces such as the ellipsoid.
Moreover the Gaussian curvature (the determinant of the matrix) is the
product K = K)K2 of the two principal curvatures (this defines the sign of
K). This indicates that the Gaussian curvature is positive when the surface
is convex or concave at the point in question (i.e., when the surface near
the point lies all on one side of the tangent plane there). On the other
hand, it also indicates that the Gaussian curvature is negative at a saddle
point, where the surface lies on both sides of the tangent plane.
The detailed demonstration of these results requires time and care and
can be done by various techniques, not necessarily formulated in terms of
the "shape transformation" L. However, the use of this transformation
does illustrate in striking form the close relation which obtains between
the algebraic properties of symmetric matrices and the geometric study of
curvature.
226 VIII. Forms of Space

3. Arc Length and Intrinsic Geometry

On a sphere, the length of an arc of any smooth curve can be computed
by an integral like (1.4), where the differentials dx, dy, and dz are com-
puted in a suitable chart directly from the expression for the curve in the
coordinates of latitude () and longitude cf>. The formal equations (2.1) for
the cartesian coordinates x, y, and z give (with r held constant)

dx = - r sin () cos cf> d() - r cos () sin cf> dcf>,

dy = -r sin () sin cf> d(} +r cos () cos cf> dcf>,

dz =r cos () d(}.

Hence the three-dimensional element ds 2 = dx 2 + di + dz 2 of dis-

tance, when restricted to the sphere, can be expressed formally as

(1)

One may also motivate this formula by drawing on the sphere a small
"spherical triangle" made up of arcs of circles with hypotenuse ds, hor-
izontal side r cos () dcf> and vertical side rd(}; then (1) is the Pythagorean
theorem for this infinitesimal "right triangle". What is more (and more
than we will prove here), this formal calculation gives the correct
integrand ds for the arc length of a spherical path written in latitude and
longitude coordinates for a parameter t as () = g(t), cf> = h(t). On the
practical side, such length calculations are needed in ocean navigation,
where the great circle joining two points provides the shortest possible
(smooth!) voyage between those two points. On the theoretical side, such
calculations of length form the starting point of much of Riemannian
geometry (§ 10 below).
There are similar formulas for the length of arc applicable to paths on
other surfaces in 3-space. One assumes that a piece of the surface is given
in terms of two parameters u and v by smooth functions

x = f(u,v), y = g(u,v), z = h(u,v). (2)

This is to be a chart for a piece of the surface, defined for u, v in some

open set of the u, v plane, so that (2) is to be a bijection of this open set
onto an open subset of the surface. The representation is to be regular, in
the sense that the matrix

aj ag ah
au au au (3)
af ~ ah
av av av
3. Arc Length and Intrinsic Geometry 227

has rank 2 everywhere (i.e., has its rows always linearly independent).
This means that at each point in the u, v plane the functions (2) maps the
tangent plane of the (u, v)-plane linearly to the tangent plane at the
corresponding point on the surface-and that (3) is the matrix of this
linear mapping. Then the same sort of formal calculations as those for the
sphere yields a formula for arc length (that is, for a ds 2) of the form

ds 2 = E du 2 + 2 F dudv + G dv 2 , (4)

where E, F, and G are smooth functions of the parameters u and v in the

open set at issue; these functions are expressible in terms of the given
functions f, g, and h of (2) and their derivatives. Moreover, this formal
expression can again be used to determine actual lengths of arc for
smooth paths on the surface. This "quadratic differential form" (4) is
called the first fundamental form for the surface. There is also a "second"
such form, which we will not discuss here.
Now there arise questions about the "intrinsic" geometry of the
surface-geometric properties of figures in the surface which can be deter-
mined entirely in terms of measurements in the surface which make no
reference to the ambient space-the rest of the space surrounding the sur-
face. These properties would include quantities which could be computed
just from the first fundamental form ds 2, because this form depends only
on measurements of arc length for paths in the surface. Not all measur-
able quantities for a surface are "intrinsic" in this sense. For example, the
two principal curvatures are not intrinsic. This can be seen in the example
of a circular cylinder-such a cylinder can be "rolled out" onto a plane
without changing the lengths of any arc on the cylinder-but clearly
changing one of the two principal curvatures of the cylinder. An amazing
theorem found by Gauss asserts, however, that the Gaussian curvature
(unlike the principal curvatures) is intrinsic-it depends only on the
coefficients in the first fundamental form (4).
The proof of this result is beautiful but subtle. Initially, one can express
the Gaussian curvature K in term of the coefficients of the two fundamen-
tal forms; a calculation then shows how to express it just in terms of
E, F, and G and their derivatives.
Another version of this argument uses small "triangles" ABC on the
surface. Each edge of the triangle is to be a geodesic-so that the length of
AB, for example, represents the shortest distance from A to B. From the
metric one can also measure each angle of the triangle, in radians, and
hence determine the excess E(ABC)-the sum of the three angles, less '!T.
One may then prove that the Gaussian curvature K at each point p is
given by the limit

E(ABC)
lim
A,B,C~p area(LiABC)
228 VIII. Forms of Space

with limit taken as the three vertices approach the point p. Thus, just as
on the sphere, positive curvature means angle sum in a triangle larger
than 17, and larger positive curvature corresponds to larger angle sum.
This result makes the intrinsic character of curvature visible. It also recalls
the results on angle-sums in non-Euclidean geometry (Chapter III).
The discovery of this result raised in explicit form the problem of study-
ing the geometry of the surface itself, intrinsically and without reference
to the ambient space. This leads to the problem of describing surfaces and
higher-dimensional varieties without using any ambient space. This is the
origin of the idea of an (intrinsically) "curved space", as used in relativity
theory. Thus it is that the Euclidean viewpoint for geometry is tran-
scended.

4. Many-Valued Functions and Riemann Surfaces

There is a second important origin for the idea of considering surfaces
"intrinsically". In solving quadratic equations and elsewhere one comes
up against "functions" such as u = ± Vx which are two-valued (or
many-valued). At first, mathematicians were inclined to treat such a for-
mulas as an ordinary "function" u = f(x), with an ambiguity indicated
by the "±", and indeed this is the usual practice in writing the standard
formula for the solution of a quadratic equation. But the two-valued func-
tion u = ± Vx can also be displayed by a graph in the (x,u) plane which
gives both the values of u for given x as the coordinates of points on the
parabola u 2 = x of Figure 1. In this way, one can avoid the hopelessly
ambiguous two-valued function u = ± Vx by observing that both
u and x are single valued functions, not of the number x, but of the point
p on the parabola. A similar analysis applies to more involved cases of
two-valued functions. Thus for the function u = ± V(x 2 -1)(x 2 -4) one
may say that both u and x are single-valued functions of a point p on the
quartic curve u 2 = (x 2 - I )(x 2 - 4), a curve which has the general form
shown in Figure 2.
u

----4-----------------~x

Figure 1
4. Many-Valued Functions and Riemann Surfaces 229

11

-----=+--4--+---+---,~--_x

Figure 2. Quartic curve.

In brief, these many-valued functions of a real variable can be made

into single-valued functions, not of a number but of a point on a curve. In
effect, such a curve displays (in a continuous fashion) all the values of the
function as the projections on the axes u and x, with nearby values
corresponding to nearby points.
This geometric process becomes more decisive and more illuminating
when it is used not for functions of a real variable but for functions of a
complex variable. Such a variable z = x + iy has two real dimensions
x and y, so instead of a curve one must use a surface. As an example,
consider the two values w = ± Vz for the square root function. This is to
be represented by the surface S whose points are all pairs (w,z) of com-
plex numbers with w 2 = z, where two points (w1,zd and (W2,Z2) count as
neighboring when both I W1 -W2 I and I Z1 -Z2 I are small. Thus the
surface is the geometrical "manifold" of all solutions of w 2 = z. The
correspondence (w,z) 4 z is a continuous projection p: S ->Z of this
manifold into the complex plane Z. Now the point (w,z) with w 2 = z on
S is really completely determined just by the number w, so the surface S
is "really" just the complex w-plane W, while the projection W = S -> Z is
continuous. Since each z =1= 0 has two complex square roots W1 and W2,
the projection W ->Z sends two points (W1,Z) and (W2,Z) to each point
z =1= 0 of Z. In this sense, W consists of two copies of Z.
To see this in more detail, we use polar coordinates rand 0 in the z-
plane, so that z = r(cos 0 + i sin 0) = re i8 . Since the multiplication of
complex numbers multiples their absolute values and adds their argu-
ments, the two square roots of z =1= 0 are

o~ 0 < 271"

when Vr is taken real and positive. The values W1 of the first square root
cover all the upper half of the w-plane, except for the negative real axis OR
(Figure 3), while the values W2 cover the lower half-plane, omitting the
positive axis OA '. Each of these half-planes is the image under Vz of the
whole of the z-plane, so we need two copies of the z-plane, as in Figure 4.
230 VIII. Forms of Space

B ___ ~
j __ ~v~A
o

Figure 3
To make this image continuous, each copy of the z-plane should be cut
apart by a slit along the positive z-axis. Now the whole surface S' is
obtained by pasting together the two halves of the w-plane from Figure
3-ray OB in the top half pasted to ray OB' in the bottom half and then
ray OA' in the bottom half pasted (point for point) to the ray OA in the
top half. Now exactly the same surface S can be described by pasting
together the two slit copies of the z-plane (Figure 4): The lower slit OB in
the top plane is pasted to the upper slit OB' in the bottom plane while the
upper slit OA in the top plane is pasted (point for point) to the lower slit
OA ' in the bottom plane. The result, seen end-on from the point at 00 on
the positive real axis, is illustrated in Figure 5 which is slightly inexact,
because the ray 000 represents two separate rays, meeting only at the ori-
gin O. All told, the surface S on which the function Yz becomes single-
values is thus represented by pasting together two copies of the slit z-
plane, so that each copy carries one of the two possible values (one of the
"branches") of Yz. Note that the location of the slit could be changed by
replacing the positive real axis by any other smooth curve running from 0
to 00. Different such choices of a slit give the same topological surface,
which is described intrinsically as the manifold of all pairs (w,z) with
w2 = z.
This first example illustrates a general method for turning a many-
valued "function" w of a complex variable z into a single-valued function
not of z but of a point on a surface S, a Riemann Surface. First decom-
pose the z-plane by slits so that w becomes single-valued when it is fol-
lowed continuously along paths constrained not to cross the slits. This
yields several single valued functions Wi which may be called the branches
of w. For each such branch wi' take one copy of the slit z-plane to carry
these values wi. Then paste these copies together along the slits
appropriately matched so as to get the desired surface S. This is a process
which visualizes the surface S, which is described intrinsically as the man-
ifold of all the pairs (w,z) involved.

e A o_ _ _ _ _ B'
---C.-

B A'
bottom

Figure 4
4. Many-Valued Functions and Riemann Surfaces 231

2nd plane
/
/
/
/
/
/

1st plane

Figure 5

For this function w = Yz it is also natural to say that z = 00 yields

w = 00, and so to use instead of the z-plane the Riemann sphere both for
z and for w (see §IV.lI). Then the surface is constructed by taking two
copies of the Riemann z-sphere, slitting each along a great circle (the real
axis) from 0 to the north pole, and then joining one edge of each slit to
the other edge of the slit on the other sphere, much as in Figure 5 above.
This process yields a two-dimensional closed surface on which both
wand z are single-valued, including values 00.
There is another way to visualize this surface. First distort each sphere
continuously so that the opened slit becomes a circular hole; the needed
connection of the edges of the slit can then be visualized by connecting
these two holes by a cylinder, as in Figure 6 below. A further distortion of
this figure shows that our Riemann surface is "really" just the surface of a
single sphere (that is, is homeomorphic to the sphere).
A more striking case is that o( the two-valued function
w = ± V(z2-1)(z2-4) which occurs in the integrand of a certain "ellip-
tic" integral which cannot be integrated in elementary terms. Again we
wish to construct a "Riemann surface" S whose points are all pairs (w,z)
with w 2 = (z2-1)(z2-4), so that both wand z will be single-valued
functions of a point (w,z) on this surface. Now for each z except z = ± I
or z = ±2 there are two corresponding values of w, so the projection
(w,z) f-> z makes this surface cover the z-plane twice. One may write w
as a product

w = ± Vz-=1 Vz+l Vz-2 Vz+2.

Figure 6
232 VIII. Forms of Space

We do not need explicit formulas for the two roots WI and W2, but we
note that for values of z very near z = I the values WI and W2 will inter-
connect much as did the values of the square root Yz-=l (just like Vz").
In other words, as the point z moves once around z = ± I in a small cir-
cle, the first root WI will change into the second root W2, and W2 into WI.
There will be a similar interchange near the three points
z = - I, Z = +2, and z = - 2, and our two copies of the z-plane
should be connected correspondingly. This can be done if we take two
copies of the z-plane and slit each copy along the real x-axis from + I to
+2 and from - I to - 2. These slits will insured that a circle or other
closed path not crossing the slits cannot go around just one of the points
± I, ±2, but must (Figure 7) go around two (or four) of them-and so
following such a closed path will not change one value Wi into the other.
Now to follow what happens to Wi across the slits we paste the two z-
planes together, so that the upper edge of a slit on the top z-plane
attaches to the lower edge on the bottom, and vice versa, much as in the
previous Figure 5. Also the two copies of each point ± I and ±2 are
identified. Now the two copies of the z-plane can carry the two values
WI and W2 so that they fit together across the slits. The result is a
Riemann surface such that both z and ware single-valued and continuous
functions of a point on the surface.
This description of the surface is not really intrinsic, since we could
have used differently placed slits, say slits along semicircles from - 2 to
+2 and - I to + 1. However, any such description provides a picture of
the intrinsic manifold ("algebraic curve") of all solutions of the equation
w2 = (z2 - 1)( z2 - 4). The previous Figure 2 displays just the "real"
points on this curve.
As before, one may make a more "geometric" picture of this surface by
starting with two slit Riemann spheres instead of two slit z-planes, then
distorting each of the slits into an open circular hole. The result is then
two spheres, each with two circular holes which are to be joined rim to
rim. This join can be done by using two cylindrical tubes, as in the Figure
8. A distortion of the resulting surface shows that it is "like" the surface of
a doughnut-a torus. For these and more elaborate Riemann surfaces one
wishes to formulate an intrinsic description.

Figure 7. Slits.
5. Examples of Manifolds 233

Figure 8. Distorted torus.

5. Examples of Manifolds
The sphere can be described as the "manifold" of all solutions (x,y,z) of
the equation x 2 + i + z2 = l. In many other cases the set of all solu-
tions of some problem or the collection of all "things" with some observ-
able property can be regarded as a geometric entity-curve, surface, or
solid-in which "nearby" solutions are pictured as "nearby" points.
One simple example is the manifold of all possible quadratic polynomi-
als ax 2 + bx + c with real coefficients. Since such a polynomial is fully
determined by the three real numbers a, b, and c, it can be represented
by the points in Euclidean 3-space with these coordinates a, b, c. Hence
this manifold of quadratic polynomials is just the space R3.
In mechanics one has occasion to consider a double pendulum in the
plane, with a second bob B hanging by a string from the end of a first bob
A, the latter hung from a fixed point P (Figure 1). To study the possible
motions, one needs first of all the "manifold" of all possible positions of
this double pendulum. Now a position is described completely by giving
two angles () and If>; to wit, the angle made by each bobstring with the
vertical. Here the possible values of the angles () and If> range from 0 to
2'1T radians, with 2'1T counting the same as zero (Figure 2). Hence the possi-

Figure I. Double pendulum.

234 VIII. Forms of Space

21T r----~p~__,

Q Q'

Figure 2

ble positions can be considered to be the points of a 2'1T X 2'1T square in

which each point P of the top edge is to be identified with the correspond-
ing point P' of the bottom edge, while the left-hand edge is similarly
identified point by point (Q with Q ') with the right-hand edge. If one
replaces the square by a flexible piece of paper, making the identification
by pasting, one may see that the top to bottom identification by itself pro-
duces a cylinder which on left-to-right identification becomes a torus. In
other words, the torus is the geometric manifold of possible positions of
the double pendulum; on this torus one may picture various periodic
motions of the pendulum. It is striking that the same torus arises both
from the pendulum problem and as a Riemann surface. This illustrate the
ambiguity of geometric form.
The Mobius band is another example. Take a long rectangle and then
identify points on the left-hand edge-but in upside down order-with
points on the right-hand edge, as in Figure 3. The resulting surface or
band has a "twist" and is one-sided. Moreover, it has an edge (a "boun-
dary") which is a single twisted circle.
A related example is the "projective plane"-described as the manifold
of all possible lines passing through the origin 0 in 3-space R 3, with the
obvious meaning of "nearby" lines. In brief, it is the manifold whose
"points" are the lines through o. Since each line through the origin meets
the unit sphere about the origin in two diametrically opposite points, this
manifold may also be described as the manifold whose points are pairs of
diametrically opposite points on a sphere. More simply we might use just

Figure 3. Mobius band.

 5. Examples of Manifolds 235

single points-say the points in the lower-half hemisphere, except that

then each pair of diametrically opposite points on the equator of the hem-
isphere must be identified. If we think of this as a circular disc with these
diametrical points identified, the picture is almost like that of the torus
above. However, if we try to carry out the identifications by pasting
together corresponding points from the equator of the hemisphere we
arrive at a figure like 4, with A to be pasted to A " B to B' and so on. One
cannot actually carry this out in space without making one edge PA Q'
pass through P' BQ' where it really should not-but one can imagine the
resulting manifold existing, though not in familiar space R3. This mani-
fold is clearly a surface-hence two-dimensional.
This manifold may also be pictured by projecting the lower hemisphere
from the center (of the hemisphere) onto the plane 'iT tangent to the hemi-
sphere at the south pole. The points below the equator then project onto
all the points of the plane 'iT, while the points on the equator are projected
off to "infinity". The manifold is thus pictured as an ordinary plane plus a
stock of points at infinity, and one may show that any two parallel lines in
the ordinary plane meet at exactly one of these points at infinity. This is
why this manifold is called the projective plane. There are many more
such manifolds, such as a projective 3-space or the manifold of planes
through the origin in R4. They arise intrinsically and not just as subsets of
Euclidean space.
The projective plane of Figure 4 is a "one-sided" surface-one may
move from the inside (say, at A ') to the outside (say, at A) without going
through the surface. In this respect, it is like the Mobius band. Indeed, if
we cut off the bottom (the disc below the circle TSVR of Figure 4) what is
left is just the Mobius band! To see this, project the rest of Figure 4 onto
the plane through TSVR. After some distortion this gives the ring on the
left of Figure 5, where Q'PQ is still to be pasted (reversed) on QP'Q'. To
do this, cut the ring apart vertically along RQ and Q' S, and turn the
right-hand half over. Now Q'PQ can be pasted to QP'Q-while RQS is
still to be pasted to RQ'S-again with a reversal. But this (as on the right
of Figure 5) is just a distorted Mobius band, with the arc RQS pasted in

Q'

R
,.-------......\
T s v

Figure 4. Projective plane.

236 VIII. Forms of Space

R R S
I Q

1'01'
I
I

T V T P p' V

Q'
I
I
I Q'
I
S S R

Figure 5. Mobius band.

reverse on the arc RQ'S-and a single free edge STRVS (a circle) as

boundary. In other words, the projective plane arises from the Mobius
band by pasting in a disc (the bottom disc of Figure 4) along the circular
edge of the Mobius band of Figure 3!

6. Intrinsic Surfaces and Topological Spaces

Examples have indicated the need to describe a two-dimensional surface
intrinsically, without thinking of the surface S as something embedded in
ordinary 3-space. Thus the idea of Gaussian curvature is intrinsic, while
the various Riemann surfaces and other manifolds all illustrate this
need-most dramatically for the projective plane, which simply cannot be
embedded in 3-space without some singularity. It turns out that the intrin-
sic description of a surface proceeds in several stages, each stage a formal-
ism which describes more of the structure of the surface.
First, a surface S is at least a topological space (Chapter I). This
requirement provides points and a sense in which two points on the sur-
face may be regarded as "nearby". A "distortion" of a surface is just a
homeomorphism of spaces.
Second, a surface S is two-dimensional. This is a "local" property of the
surface, which will require that each point has a neighborhood which
"looks like" a neighborhood of a point in the Euclidean plane. "Local"
properties are to be contrasted with "global" properties of the whole sur-
face.
Third, a surface S should normally be smooth; this means that one
should know which curves on S are smooth and which functions on the
surface are differentiable. In turn, this provides means to construct
tangent and cotangent planes to the surface.
Fourth, the surface may be equipped with a metric, which will make it
possible to measure the length of smooth curves on the surface as well as
the angles between such curves-thus providing for the basic operations of
6. Intrinsic Surfaces and Topological Spaces 237

mensuration on the surface. This metric structure on the surface is usu-

ally expressed by giving a suitable differential form ds 2, where ds is the
"element of arc", used in integration to get the length of arcs on the sur-
face.
To fill in the details, we turn first to the topology. A surface S is at least
a topological space. This means that certain subsets of S are specified as
"open", and that these open sets satisfy the axioms given in §I.lO. Then
we can define when a function I on the surface is continuous-for any
function Ion S with values in S, or in the reals R, or in any other topo-
logical space (as always, I continuous means that the inverse image of
each open set is open). For example, if Sand S I are two surfaces, a
homeomorphism (see §V.5) S --> S I is a function I which is a bijection of
the set S onto the set S I such that both I and its inverse I-I are continu-
ous. Thus the Riemann surface of z = Vw pictured as two spheres joined
by a tube is homeomorphic to the sphere; similarly the two-spheres pic-
ture (Figure 4.8) of the Riemann surface of z = V(w 2 - 1)( w 2 - 4) is
homeomorphic to the torus (but a torus is not homeomorphic to a sphere).
For a homeomorphism f: S -->S I it is not enough to require that I be con-
tinuous and a bijection. For example the map (j ~ e 2'TTiIJ is a bijection of
the "half open" interval [(j I 0 < (j < 27T] of the real axis onto the unit
circle in the complex plane-but this interval is clearly not homeomorphic
to this circle, because the inverse function 1- 1 is not continuous at 1. On
the other hand, the circle is homeomorphic to the ellipse or to the peri-
phery of the square.
The definition of a topological space is very general. Any surface with
fins and filaments attached or with holes punched out is a topological
space. Any subset Y of the plane R2 is a topological space, provide one
takes as the open sets U of Y the intersections V n Y with Y of arbitrary
open subsets V of R2. In fact, this procedure makes a subset Y of any
topological space X into a topological space (with the so-called induced or
"subspace" topology). This choice of a topology on Y makes the inclusion
Y c X of Y in X into a continuous function Y -->X; moreover, it provides
the fewest open sets in Y necessary to have such continuity. The indicated
generality in the definition of a topological space is useful not just for the
study of "pathological" spaces (of which there are many) but also to allow
a flexible way of handling many mathematical objects as spaces. One
example is the consideration of "spaces of functions". Thus the set of all
continuous real-valued functions I on the unit interval [0 < <x 1] of
the reals is a metric space with the distance p(j,g) between two functions
I and g taken to be the maximum value attained by the quantity
I/(x) - g(x) I on the interval (recall from §I.lO that every metric space
is also a topological space). Other ways of specifying when two functions
are close to each other can be similarl~ used to construct topological
spaces, as for example with the space L of Lebesgue square integrable
functions.
238 VIII. Forms of Space

Any set X whatever can be regarded as a topological space when all the
subsets of X are taken to be open; this gives the so-called discrete topol-
ogy on X. For an infinite set X there is a topology in which the empty
subset and all subsets V with a finite complement are taken to be open.
This clearly satisfies the axioms for open sets, but is far from the usual
geometric pictures of a space.
In view of this generality, it has been useful to consider various re-
stricted classes of topological spaces. As in §I.lO a space X is said to be
Hausdorff when it satisfies the following "separation" axiom:

Hausdorff Axiom. To each pair p =1= q of distinct points of X there exist

disjoint open sets V and V in X with p E V and q E V.
(In other words, different points p and q have disjoint open neighbor-
hoods V and V.) All the subspaces of R2 or of Rn are clearly Hausdorff
spaces, as is any metric space. There are, however, many non-Hausdorff
spaces; the simplest is the Sierpinski space which has exactly two points
p and q, with open sets the empty set 0 and the sets {p} and {p,q} (the
point p is open, but not the point q!).

For many purposes one may wish to put a space together out of smaller
pieces, just as a sphere may be described by two overlapping charts or a
circle by two or more overlapping intervals. This pictorial idea (scissors
and paste!) can be formalized in the notion of a "covering". An open cov-
ering of a space X indexed by a set I is a family Vi, for indices i E I, of
open sets Ui of X, such that X is the union of the Ui (i.e., such that every
point of X lies in Vi for some index i E I). The space X is thus com-
pletely determined by knowing all the subspaces Vi and how they over-
lap. For example, a subset V of X is open if and only if each intersection
V n Vi is open in the (induced) topology of the space Vi; for the same
reason a function f on X to some other space is continuous if and only if
the restriction of f to each of the open sets Vi is continuous. This result is
the reason why it is useful to consider coverings by open sets-and why
one describes surfaces by overlapping open charts.
Coverings are also used in the definition of compactness (§VI.7).
For many purposes, one studies not general topological spaces, but the
more special compact Hausdorff spaces-in particular surfaces which are
both compact and Hausdorff.
It is also appropriate to ask when a space is connected. The intuitive
idea is that a space is connected when it does not "fall apart" into two or
more pieces. This can be turned into a formal definition in at least two
ways. One definition reads: A space is connected when it is not the union
of two disjoint, non-empty open sets V and V. For example, the topologi-
cal space consisting of the whole real line minus the origin is not con-
nected, because it is the union of two (disjoint) half lines, each open-
because the origin has been omitted.
7. Manifolds 239

A second definition of connectivity uses paths, where a path in any

space is defined just as for paths in the plane: A path in X is a continuous
map f: I ~X of the unit interval I = {t I 0 ::;;; t ::;;; I} into the space X; it
is of course called a path from f(O) to f(1). Then a space X is said to be
path connected (or, sometimes, "arcwise connected") when there is at least
one path joining any two points of the space. One may readily show that
any path-connected space is connected in the previous sense. Indeed, if it
is not so connected, and "falls apart" into disjoint open sets U and V, join
a point of U by a path f to a point of V; then the open sets f- IU and
f-IVare disjoint open sets which disconnect the interval I-but the inter-
val is a connected space. However, there are connected spaces which are
not path-connected! Here again we have an example of two different for-
malizations of the same intuitive idea!

7. Manifolds
The second aspect of the "intrinsic" description of surfaces is the sense in
which any surface is "two-dimensional". This should mean that each
point of the surface has a neighborhood which looks like (i.e., is
homeomorphic to) a neighborhood of some point in the Euclidean plane.
We can express this condition in terms of the "charts" which have already
been used (in navigation and) in our description of the intrinsic geometry
of the sphere. Specifically, if S is any topological space a (2-dimensional)
chart on S is a homeomorphism

U c S, (1)

of an open subset U of S to an open set V in the Euclidean plane. The

open set U is called the domain of the chart. An atlas for S is a family of
charts </> i: Ui ~ Vi, indexed by the elements i of some set I and such that
S is the union of the domains of these charts (i.e., such that every point of
S belongs to at least one chart). A two-dimensional topological manifold
(a surface) is a topological space which has such an atlas; usually one also
requires that the surface be connected in the sense described above; this
will make it likely that there are charts which overlap. Here two charts </>
and </> I have an overlap when the intersection U nU' of their domains is
nonempty. In this event, since </> -I and </> I-I exist and are continuous,
the images </> (U nU') = Wand </> '(U nU') = W' of the overlap are
open sets Wand W' in the plane, and the relevant restrictions </>1 and </> 'I
of </> and </> I yield a diagram (see Figure I)

W' , (2)
240 VIII. Forms of Space

u u'

Figure I. Overlap of charts.

which is a homeomorphism

</>'1'</>1-1: W ~ W'; (3)

this is informally called the overlap (or patching) map for the two charts,
because it codes the way in which the two charts are to be pasted together
on the manifold. In other words, a surface is a topological space obtained
by pasting together open sets (more explicitly, open discs) cut from the
Euclidean plane.
This description of a surface S is not yet an invariant one, because one
and the same surface may have many different atlases. It is for this rea-
son that the definition reads "A surface is a topological space for which
there exists an atlas" (e.g., there can be many different atlases for the
sphere!). Alternatively one might take all the possible charts for S; they
form an atlas called the maximal atlas because it contains every other
atlas. Then a surface would be defined as a topological space which is
covered by the domains of all of its two-dimensional charts.
As an example, consider the projective plane, defined to be the mani-
fold of all lines through the origin of R3. As we have seen in §5, it can
also be described as a circular disc in which diametrically opposite pairs
of points on the circumference have been identified. It can be covered by
three charts, for example, by three sectors of the disc, each extended a bit
over its boundary so as to be an open set (see Figure 2). In this atlas, the
charts V and V' overlap right side up along their extensions on AO and
also along BC-but upside down there because CB is pasted upside down

c B

u u'
o
u"
c
A

Figure 2. Projective plane with three charts.

7. Manifolds 241

on Be. Thus U nU' falls into two connected pieces Wi and W2 . If we

choose the overlap map on the first piece so as to preserve the orientation,
then on the second piece it must reverse the orientation. We could have
made the intersection of the charts connected by taking more and smaller
charts as in the string of beads in Figure 3-but it still turns out that at
least one of the overlap maps W ~ W' must reverse the orientation of the
open sets Wand W' of the Euclidean plane-because CB goes upside
down onto Be. In other words, there is no way to pick "matching" orien-
tations on all the charts of an atlas for the projective plane. This surface is
therefore said to be non-orientable (and this property also means that the
projective plane turns out to be a "one sided" surface whenever it is
embedded in an (orientable) Euclidean space.
These observations about orientability can be formalized so as to apply
to any surface. First, the basic axioms of plane geometry (§III.8) provide a
choice of orientation for the Euclidean plane R2; this choice may be pic-
tured as a choice of one of the possible directions of rotation as the
"clockwise" one. Then any overlap map (3) between two charts sends an
open set W of R2 to another open set W' of R2. We can assume that W
and W' are both nonempty and connected. At anyone point of W, the
overlap map will either preserve the chosen orientation (i.e., preserve the
chosen sense of rotation) or reverse it. Since W is connected, orientation
preserved at one point of W will mean orientation preserved at all points
of W. Now if a surface has an atlas in which every nonempty overlap
map preserves the orientation, we say that the surface is orientable, and
that it is oriented by a choice of orientation in one chart (and, by con-
tinuation, in every chart). This formal definition matches the intuitive
notion. Indeed, the chosen orientation in R2 may be pictured by a small
circle with a sense of rotation (a "clock"). Each chart cf>: U ~ W transfers
(by cf> -I) this picture to the piece U of the manifold; when the overlap
maps can all be made invertible, these directed small circles can be moved
continuously along the overlap from the domain of one chart to the
next-and this cannot be done on the projective plane. However, the plane
itself, the sphere and the torus are all orientable in the sense of this
definition. This development shows how the elementary Euclidean notion

Figure 3
242 VIII. Forms of Space

of orientation (Chapter III) extends to a much more general geometric

context.
One may also readily prove that a surface is connected if and only if it
is path-connected.
With these definitions in hand, we may ask what surfaces there may be,
and in particular: What are all the compact connected orientable sur-
faces? Here two surfaces are to count as the same if they are
homeomorphic; thus the surfaces of the sphere and of an ellipsoid are the
same, but they are evidently different from the surface of the torus. The
torus, in its tum, is different from the pretzel with two holes (Figure 4)
and so on. This sequence "and so on" gives all the possibilities. Any com-
pact connected orientable surface is homeomorphic to the sphere or to a
pretzel with g holes, where the natural number g >
I is called the genus
of the surface. Two such surfaces of different genera cannot be
homeomorphic. The proof of this theorem (to be found in many elemen-
tary texts on algebraic topology) uses a description of surfaces in which
the edges of a closed polygon are suitably pasted together. For example,
the torus has been described above as a square in which the four edges in
cyclic order are pasted according to bab -Ia -I (Figure 5). Similarly an
octagon with the edges paired up in cyclic order in the pattern
aba -Ib -Icdc -Id- I will produce a pretzel with two holes (just think of
two tori, connected along the circle e of Figure 6). The proof of the gen-
eral theorem about surfaces amounts to showing that any such pasting of

Figure 4. Pretzel with two holes.

b b

a
a

Figure 5. Torus as pasted square.

7. Manifolds 243

Figure 6. Pastings for a pretzel.

polygons to form a connected orientable surface can be reduced to one of

the standard forms for a sphere or a pretzel. The result has evident mean-
ing for the Riemann surfaces we have already described; indeed, every
compact Riemann surface for a suitable function w = w(z) is such a sur-
face, and the genus is a significant invariant.
For non-orientable surfaces there is an exactly similar theorem. It can
be formulated in terms of the Mobius band, which, as we have seen, is a
non-orient able surface with a single circular arc as boundary. So take any
orientable surface, cut out a circular disk with a circular arc as boundary,
and along the boundary paste in a Mobius band-boundary to boundary.
The resulting surface is evidently non-orientable, and the general theorem
states that any compact connected non-orientable surface is
homeomorphic to one of those so constructed. For example, this construc-
tion on the sphere will produce the projective plane (Mobius band plus
disc!).
For noncompact surfaces the result cannot be so easy-for example one
can produce such a surface by removing a finite or infinite number of
points or slits from a compact surface.
For higher dimensions, our formal definition of a surface (a two-
dimensional topological manifold) carries over at once to a definition of
an n-dimensional topological manifold; the only change is that charts are
now homeomorphisms </>: U ~ VeRn into open subsets V of Rn. For
each chart </> the cartesian coordinates Xi of Rn, regarded as functions
Rn ~ R, yield by composition n functions Xi</>: U ~ R, i = 1, ... , n, which
are called the local coordinates for that chart. A point in the domain U of
the chart is completely determined by its local coordinates, while the over-
lap functions describe the possible changes in local coordinates. Some
more old-fashion treatments of manifolds operated with local coordinates
only-but the full notion of a manifold requires both its global description
as a topological space and its local description by charts.
For n > 3 the classification of all possible manifolds is a difficult prob-
lem. There are also infinite-dimensional analogs of manifolds (Lang
[1967]).
244 VIII. Forms of Space

8. Smooth Manifolds
The surface of a rectangular box (a parallelopiped) is a two-dimensional
manifold according to our definitions-it is a compact and connected
topological space which is, moreover, homeomorphic (say by radial pro-
jection) to the surface of a sphere. However, a curvature can hardly be
defined at the corner points of such a box, and at these points there is also
no tangent plane. In order to describe tangent planes, intrinsic curvature,
and the like we must restrict attention to smooth surfaces.
The idea, as in §VI.9, is that a function is smooth when it has enough
derivatives. To expand this, consider first a function f of two real vari-
ables XI and X2 which is defined in some open set W of the plane R2,
thus f is a function

(1)
Since W is open, each point (XI,X2) in W has an open neighborhood also
contained in W; this allows the usual definition of the partial derivatives
of f at that point. Then f is said to be smooth of class C:IJ when f has
partial derivatives of all orders at each point of its domain W, while f is
said to be of class Ck when it has continuous partial derivatives of all or-
ders up through order k.
Consider next a function I/;

(2)

between two open sets W' and W in the plane R2. Here each point p in
W' is determined by its coordinates, say X I and X2, while each point in
W is determined by its coordinates, say Ui and U2; they can be regarded
as functions UI,U2: W ..... R. In this way the function I/; of (2) is described
by two composite functions UI·I/; and U2 '1/;: W' ..... R

(3)

This is the expression of I/; in terms of (local) coordinates. The function I/;
is defined to be smooth if both these coordinate functions f I and f 2 are
smooth (have continuous partial derivatives of all orders).
This definition has three basic properties:

1° Composition. The composite of two smooth functions is smooth. In

particular, if 1/;: W' ..... W in (2) and f: W ..... R in (1) are both smooth, so is
their composite j-1/;: W' ..... R-because its partial derivatives are given by
the chain rule.

2° Restriction. A smooth function remains smooth under any restriction of

its domain. Specifically, if 1/;: W' ..... W in (2) is smooth, while V' is any
8. Smooth Manifolds 245

open subset of W', then the restriction of l{; to V'is a smooth function
I V': V' ---> W.
l{;

3° Patching. A function put together from smooth pieces is smooth. For

example, if f: W ---> R in (I) is smooth and if the open set W is covered by
a family of open sets Vi, for i is some index set I, then if all the restric-
tions I I Vi: Vi --->R are smooth, so is I-because each derivative of I at a
point of W can be calculated from the values of I in one of the sets Vi
which contains that point. In this case, one says that I is "patched
together" from the smooth pieces I I Vi'
A smooth surface is now described by requiring that there be an atlas
of charts with smooth overlaps. Indeed, recall from §7 that any two charts
of a surface S,
</>'
S :J V' -3> V' C R2

have on the intersection V n V' of their domains an "overlap" map

</> 'I</> I-I
R2 :J W' = <p (V n V') -3> W = <p '(V n V') c R2

given as a composite of (suitable restrictions of) <p I and <p -I. This overlap
map W' ---> W goes between subsets of R2, so, as in (2), we can determine
when it is smooth. Hence one can say that the charts <p and <p I have a
smooth overlap if the composite <P/I<PI- I is smooth. If we write UI and U2
for the local coordinates in V' and Xl and X2 for the local coordinates in
V, this amounts exactly to saying that the local coordinates UI,U2 in the
overlap are smooth functions (on <p (V n V')) of the local coordinates
XI,X2·
Now define a smooth surface S to be a (topological) surface S together
with an atlas A of charts <p i: Vi ---> Vi for i E I such that any two charts <p i
and <p j in the atlas A have a smooth overlap. As before, this atlas may
not be maximal, but we can construct the maximal atlas by the following
details. So consider any other chart cf>: V ---> V for S which is "smooth for
A" in the sense that <p has a smooth overlap with every chart <p i in the
atlas A. We want to consider the collection A * of all such charts <p. We
claim that if <p I: V' ---> V' is another such chart, smooth for A, then
<p and <p I have a smooth overlap. Indeed, for each chart <p i of the given
atlas A, there is a (possibly empty) intersection Vi = V n V' n Vi of
the domains and a diagram (Figure I). Then, for suitable restrictions of
the charts <p we have smooth overlaps

the first because cf> has smooth overlaps with all cf> i, and the second simi-
larly for <p'. By composition (principle 10 above) the map
246 VIII. Forms of Space

~.
¢J' 0 w'

't ¢J i
'V ¢Ji(V;l = ¢Ji(un U' n Ui )

Figure 1. Overlapping charts.

cp 'cp - I: cp (Vi) -> cp '( V;) is smooth. On the other hand the domains
Ui of
the given charts cover the whole surface S, so the intersections
cp (Ui nUn U ') cover the whole of W = cp (U n U '). Therefore, by
the patching principle, cp 'cp -I: W -> W' is smooth on the whole of W. In
other words, the atlas A * of all charts smooth for A has the property that
any two of its charts cp and cp , have a smooth overlap. Among all smooth
atlases for the smooth surface S, it is the maximal one-and every smooth
surface can be described in an invariant way by such a maximal atlas. The
description evidently depends on the three basic properties (composition,
restriction, and patching) for smooth maps.
The same ideas will describe smooth manifolds M of any dimension.
One also has the notion of a smooth map

(): M ~ M' (4)

between two such manifolds M and M ' (of the same or of different
dimensions). Indeed, for each point p of M choose charts for M and M'
with domains U and U' such that p E U and ()(p) E U'. Then if U has
dimension k and local coordinates XI , • . . ,Xk from the given chart, while
U' has local coordinates UI , . . . ,Un, the map () is expressed, much as in
(3), via the local coordinates by n real-valued functions

The map () is then defined to be smooth if these n functions are smooth;

this does not depend on the choice of the charts because the overlap func-
tions between any two charts are themselves smooth, and because the
composite of smooth functions is smooth.
In the simple case of a real-valued continuous function

f:M~R (6)

one needs only one chart for the reals R, considered as a one-dimensional
manifold. Thus the definition above states that such a function f is
smooth if on the domain U of each chart of M with local coordinates
9. Paths and Quantities 247

Xl , ... ,Xk the corresponding real-valued function I(Xl , ... ,Xk) is of

class Coo. For any open set V in M, this also describes what is meant by a
smooth function V ->R, since V itself is a smooth manifold (with charts
the intersections with V of all the charts of M). It is then straightforward
to prove that a map 8: M ->M' between smooth manifold M and M' is
smooth if and only if, for every smooth chart of M' with local coordinates
Uj, all the real-valued composites uj 8 are smooth in the sense described

under (6). By the patching property, it is enough to require this just for
the charts of some atlas covering M' -but the notion of smoothness is
intrinsic, since it is independent of any particular choice of charts or of
coordinates.
The plane, the sphere, the torus and the projective plane, with the respec-
tive atlases which we have described, are all smooth two-dimensional mani-
folds, with the expected smooth maps. There are similar smooth manifolds
in higher dimensions. For that matter, the circle, the open interval and the
whole real line are smooth one-dimensional manifolds.

9. Paths and Quantities

Next we study maps between manifolds. The calculus for a smooth sur-
face S involves pairs of smooth maps
h f
I ~ S ~ R. (1)

Here I is an interval, say the interval - 1 < t < I of R, with t as local

coordinate. For suitable coordinates Xl,X2 on S these two functions hand
I are expressed by formulas such as

we may suppose that h(O) = (0,0) is the origin (of local coordinates) on S.
The map h is a path through h(O), parametrized by t, while the map I
is a (real-valued) quantity (e.g., a physical quantity) defined everywhere
on the surface S. Such a quantity on S may be pictured by its level lines
(the loci I(Xl,X2) = constant} on S, as in Figure 1. At the origin t =
the path has a tangent vector with coordinates
°
dhl dh2]
[ (2)
dt'dt 1=0

while the quantity f has a gradient (or differential) given (as in §VI.9) by
coordinates

(3)
248 VIII. Forms of Space

Figure I. Level lines =countour lines.

this is its "cotangent" vector at the origin; it "points" in the direction of

the maximum rate of increase of 1 at the origin. Indeed the directional
derivative of 1 at the origin in the direction of the path h is defined to be
the derivative there of the composite function Ih. It can be expressed in
coordinates by the usual chain rule as

d(fh)
dt
I -
1=0 -
[XaXl
dh l
dt +
X
a ah2]1
X2 at 1=0·
(4)

This formula exhibits the cotangent vector (grad 1)0 of (3) as a linear
function of the tangent vector (h' J,h '2) of (2), and it indicates that two
quantities 1 and g have the same cotangent vector at this origin when
they have the same directional derivatives along all paths-and dually that
two paths through the origin have the same tangent vector there if and
only if they give equal directional derivatives there for all quantities.
This last remark suggests how one can define tangent vectors (or
cotangent vectors) without any use of coordinates. So consider all smooth
paths through the origin and all smooth quantities 1 there. Under addi-
tion and multiplication by reals a (real scalars a E R) these quantities do
form a (high-dimensional) vector space. Without using coordinates, each
smooth composite function Ih has a derivative

d(fh)
Do(fh) = ( j ( 11=0 (5)

at the origin; moreover, this derivative is linear as a function of the quan-

tity f. Hence the assignment h ~ Do(fh) makes each path h a real-
valued linear function on the vector space of all quantities. Now call two
paths hand k through 0 tangent at 0 (and write h ...... ok) when they deter-
mine in this way the same linear function; that is when Do(fh) = Do(fk)
for every quantity I. This relation h ...... ok is reflexive, symmetric, and tran-
sitive. Hence we can introduce for each path h the equivalence class

'Toh = {all smooth paths k I h ...... ok} (6)

9. Paths and Quantities 249

and call it the "tangent vector" defined by the path h at O. In much the
same way, one may call two quantities !,g: S ..... R "cotangent" at 0 (in
symbols, ! ~og) when Do(fh) = Do(gh) for all smooth paths through O.
This relation is again reflexive, symmetric, and transitive, so we may intro-
duce for each! the class

do! = {all gig smooth and ! ~og} (7)

and call it the gradient (or the differential or the cotangent vector) for !
at O.
NOW!I_Ogl and h-Og2 imply that

for each a E R. Hence the vector operations of addition and multiplica-

tion by real scalars carry over from the vector space of functions to the
classes do!, so these cotangent vectors, as constructed, do form a vector
space over R. It is a two-dimensional space, since the chain rule shows
that each do! is determined by two coordinates (~)o and (~)o rela-
aXI aX 2
tive to the basis doxb dox2. This completes the invariant description of
the cotangent space TO(S) to the surface S at the point O. Moreover, each
path h through 0 gives by the assignment

do! I-> Do(fh)

a linear function To(S) ..... R on the cotangent space. This function, deter-
mined by h, actually depends only on the equivalence class 'Toh of the
path h. Now the linear functions TO(S) ..... R are the elements of the dual
vector space TO(S)*. Hence each tangent vector 'Toh can be taken to be an
element of the dual space TO(S)*. The sum of two tangent vectors can
then be defined to be their sum as elements of this dual space-and the
sum agrees with that which would be calculated from the sum of their
components in a local coordinate system. Hence the dual space TO(S)*
provides an invariant (coordinate-free) description of the tangent space
To(S) to the surface S at the point O. In this way, both for surfaces and
for manifolds of higher dimension, the tangent and cotangent spaces arise
naturally and together from the smooth structure and from the smooth
paths and quantities defined by that structure. Many texts do not explain
this natural origin.
Each point of the surface S has its cotangent space, a vector space of
dimension 2. Now consider the set of all cotangent vectors at all points of
S; they form a geometric object called the cotangent bundle res. It is also
a smooth manifold; for a given chart of S, each cotangent vector d! is
determined by four real numbers: The local coordinates (Xl,X2) of the
point of cotangency, plus the values XaXI and XaX 2 of the partial deriva-
250 VIII. Forms of Space

tives at that point of some quantity f. These four coordinates transform

smoothly on the overlap between two charts on S. Hence the cotangent
bundle res for S is a four dimensional manifold. Moreover, the function
p: res --'>S which sends each cotangent vector dof to the point of
cotangency (dof f-> 0) is a smooth map, called the projection of the
cotangent bundle on its base S. A cross section w of this projection is a
smooth map w: S --'> res such that pw: S --'> S is the identity; in other
words, a cross section is a smooth map which sends each point (XI,X2) of
the surface S to a vector i X I>X 2 )f cotangent at that point. In local coordi-
nates Xl and X2 the differentials dXI and dX2 of the coordinates form a
basis of each cotangent space, hence each cross section w has (in local
coordinates) an expression of the form

(8)

where gl and g2 are smooth functions. Expressions such as (8), usually

called differential forms, can be described intrinsically as cross sections of
the cotangent bundle. In particular, each smooth function f: S --'> R on the
surface determines such a differential form; in local coordinates

(9)

This idea of a cotangent bundle also arises from mechanics (Chapter IX).
The tangent bundle T.S is similarly described; its points are all the
tangent vectors at all the points of the surface S; they form a four dimen-
sional smooth manifold with the evident coordinates, and with a smooth
projection T.S --'>S. A smooth cross section of this projection is called a
vector field on S. Such a field assigns to each point of S a tangent vector
at that point-and this in a smooth way. For example, if S is the plane
with coordinates x,y, a pair of simultaneous differential equations
dx
dt = f(x,y), d: = g(x,y)

with smooth f and g is in effect a vector field in the plane.

These few definitions are just the starting point for the study of the cal-
culus on manifolds. For example, integration of differential forms can be
defined, and the fundamental theorem of calculus has higher dimensional
analogs in the theorems of Gauss, Green, and Stokes. Higher derivatives
of smooth quantities also enter. For example, the k-jet of a quantity
f: S --'> R is the object determined by the collection of all the partial
derivatives of f up through order k, with respect to some set of local
coordinates.
The topological structure of a surface or a manifold is given in a very
simple way, by specifying the open subsets. A smooth structure requires a
10. Riemann Metrics 251

more elaborate description, in terms of an atlas. Familiar topological

manifolds, like Rn , also have a familiar (standard) smooth structure-but
just recently it has been discovered that R4 has a second smooth
structure-not equivalent to the standard one. (This cannot happen for
Rn, n =1= 4.) However, consider the 7-sphere S7; that is, the locus in R8 of

xf + xi ... + xl = 1.

It has a standard smooth structure-with charts given (say) by projection

onto coordinate hyperplanes of R8. However, it has also another,
essentially different, smooth structure.

10. Riemann Metrics

The final step in the intrinsic description of a surface S is the metric,
defined by specification of the lengths of curves on S. This will be done
by integrating a formula for the differential ds of arc length. For the
sphere (and other surfaces) we have seen that ds 2 can be given in terms of
suitable local coordinates u and v by a formula of the form

ds 2 = E(u, v)du 2 + 2F(u, v)dudv + G(u, v)dv 2,

where E, F, and G are smooth functions. More generally, one may express
such a Riemann metric in terms of local coordinates Xi as a "quadratic
differential form"
2
ds 2 = ~ gi,j(Xi,Xj )dXidxj, (1)
i.j=1

where the gij are smooth functions of the local coordinates and are sym-
metric (gi j = gji) and positive definite. The latter means, as usual, that
(al,a2) =1= (0,0) implies ~ gij(xi,Xj)aiaj > 0. A Riemann metric on a sur-
face means a metric (1) on each chart, such that the metrics agree (under
change of coordinates) on each overlap of charts.
To interpret this, consider a path h: 1--> S on the surface for which the
image h(/) lies in the domain of the chart involved. We have already seen
that each differential form

represents a linear function on each tangent vector h '(t), by the formula

I dh l 2 dh2
w(Toh) = g (XI,X2) dt +g (XI,X2) dt'
252 VIII. Forms of Space

In just the same way, the quadratic ds 2 of (1) is a quadratic function of

tangent vectors, as in

(2)

By the assumption about overlaps of charts, the value of Q( Toh) at any

value of t does not depend on the choice of a chart. Finally, the length of
the path h from h(tl ) to h(t2) is defined by the integral

(3)

The metric may also be used to describe corresponding formulas for

areas.
The meaning of ds 2 can be formulated in terms of vector space con-
cepts. At each point of the surface S the form ds 2 involves a positive
definite 2 X 2 matrixg ij = gij (XI,X2). Ifv and ware two tangent vectors
to S, expressed in the same local coordinates as v = (VI, V2) and
W = (WI,W2), then the formula (2) gives a bilinear, symmetric, and posi-
tive definite inner product
2
(V,W) = ~ gij ViWj . (4)
i.j = I

Thus specifying a Riemann metric on the surface S amounts exactly to

specifying an inner product in the tangent plane at each point of S, in
such a way that this specification is smooth (i.e., smooth as a function of
the local coordinates). This description (and the formula above) apply
equally well for smooth manifolds of higher dimensions. It is the starting
point of an extensive theory of Riemannian manifolds.
We have thus completed the intrinsic definition of surfaces. A surface is
a topological space covered by charts with smooth overlaps and with a
Riemann metric. On this basis there is a remarkable further development
of geometric properties of surfaces-and similarly of higher dimensional
manifolds. In particular, the Gaussian curvature of a surface may be
defined directly in terms of the metric. With this, the intrinsic definition of
a surface is complete.

11. Sheaves
Continuous functions on a topological space X can be described just in
terms of the open subsets of X, since a function f: X ~ Y is continuous
when the inverse image under f of each open subset of Y is open in X.
Smooth functions on a manifold, on the other hand, require more for
 II. Sheaves 253

their definition: A whole array of charts. There is a different approach

which describes directly just the collections of continuous or of smooth
functions (call them the "good" functions). For this approach one must
consider not just the good functions defined on all of X, but those defined
on each open subset U of X, together with the way in which these func-
tions restrict to smaller open subsets. Such collections of functions (or of
other objects) are called "sheaves". They play a major role in higher
geometry.
In the case of two topological spaces X and Y, the topology determines
for each open set U C X the set

qu, Y) = {f: I f: U ---3>- Y continuous} (1)

of all continuous functions on U to Y. Moreover, if V C U is a smaller

open set in X, each continuous f on U has a restriction f I V to V which
is continuous there; this gives an operation "restriction"

f ~flV, r~: q U, Y) ---3>- q V, Y).

For fixed Y, we now examine the way in which the set qu, Y) depends on
U.
A presheaf P on a topological space X is a function which assigns a set
P( U) to each open set U of X and a restriction map

r = r~: P( U) ---3>- P(V) (2)

to every inclusion V C U of open sets, but in such a way that rlJ is

always the identity and also so that restrictions compose. This means that
for each nested array
WCVCU (3)

of open sets in X one has the composition rule for restrictions,

rfi, = r~r~: P( U) ---3>- P( W). (4)

In general, think of P( U) as the set of all "good" functions on the open set
U, together with the operation r~ of restricting any good function to an
open subset Vof U. For example, for each fixed Y, the sets q U, Y) of (l)
constitute a presheaf on X because, given a nested array (3), the restric-
tion of a continuous function f from U to W is the same as the restriction
of f to V, further restricted to W. If X is a smooth manifold, the sets
S(U,R) of smooth functions f: U ..... R is also an example.
Cross-sections provide another example. If p: E ..... X is any continuous
map of topological spaces, a cross section of p over an open set U is a
function w: U ..... E such that the composite pw: U ..... X is the identity on U.
Each cross section on U has an evident restriction to each open subset
254 VIII. Forms of Space

v c V, and with this restriction the cross sections of p form a presheaf

X( V,p).
Still another example, for any topological space X, is the set
g( V) = {W IW c V is open in X} (5)
of all open subsets W of the given open set U. Here the restriction map is
to be the intersection
(6)
The next observation about continuous functions is that a continuous
function on a large open set can be "pieced together" from values on
smaller open sets when these given values match. Specifically, given open
sets VI, V 2 C X consider continuous function fl: VI -> Y and h: V 2 -> Y
which match on the overlap VI n V 2 (i.e., the restrictions of fl and h
to VI n V2 are the same). Then there is a unique continuous function
f: VI U V2 -> Y which restricts to fi on each Vi-as for example when
VI and V2 are overlapping intervals on the line, as in Figure 1.
The idea here is that good functions which match on all overlaps can be
pieced together. When this applies to a presheaf P, that pre sheaf is called
a sheaf.
Formally, a sheaf S on a topological space X is a presheaf with the fol-
lowing property: For every covering
V= UVi (7)
iEi

of an open set V of X by open sets Vi and for every list of elements

fi E S( Vi) which "match", in the sense that any two fi and fi have the
same restriction to S( Vi n Vi)' there is a unique f E S( V) with restric-
tion fi on each Vi. In other words, given (7) and fi E S (Vd with
U'f" "",u jf
TV'Ji = J.V i' (8)
for all i and j, there exists a unique f E S( V) with
T~f = fi, i E I. (9)
This property: Things that match as in (8) can be pieced together as in
(9), turns out to be a central characteristic of collections (sheaves) of
"good" functions, and of other geometric objects.

--~--~--~--~--------~x
J
o 2 3

Figure I. Pasting two curves.

 II. Sheaves 255

The examples of presheaves given above are all also examples of

sheaves. For instance, the matching condition clearly holds for the
presheaf C( V, Y) of continuous functions and for the presheaf of cross
sections of a map p: E ..... X. The matching conditions all hold for the
pre sheaf Q of open sets described under (5). For, given an open covering
(7) of an open set V and given open subsets Wi C Vi for each i E I
which satisfy the matching conditions

Wi n Vj = Wj n Vi
for all i and j, the union W = U Wi is the unique open subset of V with
W n Vi = W; for each i. Hence Q is a sheaf.
The matching-piecing condition applies also to smooth functions.
Thus for an open set V on a smooth manifold and a covering V = U Vi,
a function f: V ..... R such that its restrictions Ji: Vi ..... R are all smooth is
necessarily smooth itself-because the derivative of J at a point of V
depends only on the values of J in any specific neighborhood of the point.
Hence for any smooth manifold M the sets

Cfi(U) = {f I J: V ---?- R smooth} (10)

define a sheaf on M, sometimes called the "structure sheaf' on M.

The notion of a structure sheaf can also be used to give an alternative
definition of what is meant by a smooth structure on an n-dimensional
manifold. For this approach, one starts with certain model spaces-the
open sets V of Rn , each V equipped with its structure sheaf S v of smooth
real-valued functions on open subsets of V. An n-dimensional smooth
manifold is then defined to be a topological space X equipped with a
sheaf S and which has a covering by open sets Vi, for i in some I, such
that S restricted to anyone Vi is "isomorphic" to the structure sheaf S Vi
of some open set V; of Rn. We omit the details, which need a description
of isomorphisms and of more general morphisms from one "space X with
a sheaf S" into another such object. However, the general intent should
be clear: A manifold can be described as a topological space on which the
"good" functions are specified, by giving the sheaf of such functions.
There are corresponding sheaf definitions for other types of manifolds-
for example, for Ck manifolds and for complex analytic manifolds. The
same approach has recently been especially successful in describing "alge-
braic" manifolds. In other words, the notion of a sheaf has proved to be
an effective conceptual tool in describing a considerable variety of sorts of
spaces; the elementary notion of space thus acquires (and requires) a
sophisticated conceptual development.
One may also need algebraic structures on sheaves. For example, the
structure sheaf Coo of a manifold is actually a sheaf of rings: Each set
C fi( U) of smooth functions is a commutative ring under the evident addi-
256 VIII. Forms of Space

tion and multiplication of smooth functions, and the corresponding re-

striction maps r~ of (2) are homomorphism of rings. Similarly, sheaves of
abelian groups on a topological space may be used to describe the con-
nectivity properties of that space. Sheaves also enter in certain indepen-
dence proofs in set theory (Chapter XII).

12. What Is Geometry?

Within Mathematics, geometry plays a significant role, as we have seen;
however we have not yet seen just what it is that makes a given piece of
Mathematics "geometry". The starting point for geometry is the idea that
there should be a "science of space", but the subsequent development of
this science goes in many directions, some of them unexpected and some
not strictly "geometrical". Space, to start with, is fiat, three-dimensional,
and extended. Indeed it is this unlimited "extension" of space which
introduces infinity into geometry, formally by way of the Archimedean
axiom (though this axiom also has a strictly arithmetic formulation with
real numbers). Similarly geometry involves the consideration of rigid
motions, in part as a means of comparing differently situated but similar
figures and in part because the very idea of geometric extent arises from
the experience of motion. This use of rigid motions implicitly involves the
compositions of motions and hence groups of motions-even though this
idea of group was formalized only late in the historical development of
geometry.
On the other hand, geometry employs simple physical instruments, the
straight-edge and the compass, and then idealizes these instruments as
perfectly straight lines and perfectly round circles. Using these idealized
forms the raw empirical facts of geometry are organized as systematic
consequences of a system of postulates. Amazingly, most of these postu-
lates are formulated just for the geometry of the plane, even though such
an ideal plane is not initially present in the rough original experience of
space. Moreover, this Euclidean plane has not just one structure, but two:
One structure not oriented, the other oriented and hence equipped with a
smaller group of symmetries.
The Euclidean plane is an effective formal object for expressing many
geometric phenomena, and moreover provides a handy tool for managing
three-dimensional space-but geometry does not stop short at the three
dimensions of our ambient physical space. All manner of appearances
require geometrical formulations with more than three dimensions. Linear
phenomena in these higher dimensions can be managed readily, partly
with algebraic means via coordinates and partly by extending suitable
insights from 2 and 3 dimensions. The resulting study of n-dimensional
geometry then appears in a double light, both as linear algebra and as
vector geometry-and the tension resulting from this double classification
12. What Is Geometry? 257

reappears repeatedly, as for example in the alternative between matrices

and linear transformations. Such double billing means that any answer to
our question "What is geometry?" must inevitably mix geometry with
other sources of Mathematics. Similarly, elementary geometrical notions
of angle, mensuration, and trigonometry reappear in higher dimensions in
the guise of an inner product structure for Euclidean vector spaces. Here
again the principal axis theorem, alias the spectral theorem, formulates
both the geometric properties of the axes of conic sections and the eigen-
values appropriate to linear differential operators. Therefore geometry is
not just a subdivision or a subject within Mathematics, but a means of
turning visual images into formal tools for the understanding of other
Mathematical phenomena.
But geometry is not just one form of space. The possibility of different
geometries was first indicated by the development of non-Euclidean
geometry. It turned out that the Euclidean parallel axiom is just not a
consequence of the other axioms of Euclidean geometry; instead it
represents one possible choice among several: Through a point not on a
given line in a plane there may be no parallel lines, one parallel line, or
many parallels, and the axioms expressing these varying possibilities can
be supported by geometric models-in which the "lines" mentioned in the
axioms are not the expected Euclidean lines. What then matters is the
coherent development of angle sums and other metric (and group-
theoretic) properties from the axioms. At this point in the historical
development it becomes clear that geometry is not the science of a unique
form of "space", but the extended study of many space-like forms.
Within anyone of these spaces one then comes to consider all sorts of
figures-not just the lines and triangles of linear algebra. Curves, surfaces,
and curved solids appear both analytically as the loci of equations and
geometrically as configurations in space. Curvature and global form is of
their essence. Locally such figures can be specified when point-coordinates
are given by functions of suitable parameters, such as time, latitude, and
longitude-but the full nature of these figures requires a subtle global
description. At first, they may appear to be just figures embedded in some
standard Euclidean space, but this is not adequate. Their properties, in
particular their curvature properties, have an intrinsic existence, indepen-
dent of any Euclidean embedding. In fact there are many "manifolds",
such a manifold of lines, which are originally given directly as manifolds
and not in some Euclidean space. In this chapter we have been concerned
primarily with the intrinsic description of such manifolds. This description
has required quite a number of separate steps, involving separate kinds of
"space" and "manifolds". It turns out to be convenient to begin with the
very general notion of a topological space, where everything is described
by the notion of a neighborhood (via open sets). Such spaces provide a
satisfactory generality and a way of using geometrical language for ana-
lytic purposes, as with function spaces. For the much more specific notion
258 VIII. Forms of Space

of a manifold the essential "local" idea is that of a "chart" -an idea

derived perhaps from the use of maps in navigating the globe. At any
rate, the global manifold is pieced together from overlapping local charts,
each a bijective image of a chunk (open set) of Euclidean space (or of
some other model space). Specifying the overlap maps between adjacent
charts is then an essential feature; according as the overlap is continuous,
or continuous and smooth, or complex analytic one has a topological, or a
differentiable, or a complex-analytic manifold. In this case again a simple
geometric object-for example a torus-can be viewed as with different
structures-simply as a topological manifold (or space), or as a smooth
manifold, or as a complex one (e.g., a Riemann surface). These alterna-
tives, in cases more general than that of the torus, later involve a remark-
able variety of problems. (Can one topological manifold have different
smooth structures?)
For a manifold, the charts are primarily a means to an end-in this case
to the determination of those functions on the manifold which are con-
tinuous, or differentiable, or complex-analytic, as the case may be. If one
instead considers directly such sets of "good" functions and their behavior
under restriction of the domain one is led to the notion of a "sheaf'. This
concept, with its subsequent extensive use elsewhere in Mathematics, pro-
vides a striking example of the way in which pictorial ideas from
geometry turn into conceptual idealizations.
This was the last item of this chapter, but it is by no means the end of
geometry. A few (among many) other examples will appear later. In the
next chapter we will see how certain problems in mechanics naturally lead
to the study of tangent bundles of manifolds. Then complex analysis
(Chapter X) will make use both of topology (connectivity) and manifolds
(Riemann surfaces). Algebraic varieties (loci described in coordinates by
polynomial equations) play an important role, for example in number
theory (Chapter XII). Here again geometric intuition aids mightily in the
understanding of parts of Mathematics which initially are not at all
geometric.
Beyond these few examples lie many other parts of geometry. They are
enormously varied; they do not really provide a simple answer to our
original question "What is geometry?". It would seem that the original
and pictorial study of space and motion, pursued systematically and ima-
ginatively, leads to the study of many different kinds of Mathematical
forms. The resulting development is partly geometrical and partly concep-
tual, and uses geometric insights to provide both organization and under-
standing. Geometry is not so much a subdivision of Mathematics as one
source of Mathematical form.
CHAPTER IX

Mechanics

There is a remarkable interaction between theoretical constructions in Sci-

ence and conceptual notions in Mathematics: the same idea-or the same
idea, disguised, may arise both in Science and in Mathematics. This
interaction is especially visible in the relations between theoretical Physics
and pure Mathematics, where it first became apparent in mechanics at the
hands of Galileo and in the invention of the calculus by Newton. One
may say that the calculus (which dominated the development of
Mathematics for at least two centuries) was developed by Newton in order
to tackle problems of mechanics, especially those of celestial mechanics.
This is only one of the striking interactions we have in mind. Today there
is a amazing confluence of the gauge theories in Physics (for the Yang-
Mills equations) and the geometrical theory of connections on fiber bun-
dles. It is the aim of this chapter to sketch some of these developments.
Though we cover only a small number of items, they are intended as sam-
ples of many decisive interactions between Science and Mathematics.

1. Kepler's Laws
Fascination with the motions of the planets and the stars is endemic. For
the Greeks, a precise description of planetary motion was provided by
ptolomaic astronomy. Given that the circle was for Greek geometry the
dominant representation of repetitive or periodic motion, there was an ini-
tial inclination to think that the planets move in circular orbits with the
earth as center. However, such an orbit did not suffice to explain the
appearances-in particular the observation that at times the planets
seemed to move backwards, in a retrograde motion. To account for this,
Ptolemy provided epicycles (small circles superimposed on the original
circular orbits-and with enough epicycles most of the motions (perhaps
almost all) could be explained. Such an explanation required a consider-
able number of epicycles; it served until Copernicus introduced a
260 IX. Mechanics

heliocentric theory. It was then Kepler who used many careful measure-
ments of the positions of the planets (i.e., their orbits) relative to the sun.
After elaborate calculations he was able to propose his three laws about
these orbits of the planets. His laws read
(1) The planets describe orbits in a plane containing the sun, in such a
way that the areas swept out in equal times are equal. (This refers to
the area swept out by the radius vector from the sun to the planet.)
(2) Each planetary orbit is an ellipse, with the sun at one focus.
(3) The square of the period of each planet in its orbit is proportional to
the length of the major axis of the ellipse.
In this form, Kepler's laws are a summary (in geometrical terms) of
facts of observation. Newton, using the calculus for this purpose, was able
to deduce these laws from more basic principles of motion. In particular,
Newton's second law of motion states that the force on a particle is its
mass m times its acceleration. Here the force F and the acceleration a are
three dimensional vectors, written with bold face letters as is the custom
in theoretical physics. Thus Newton's second law takes the form of a vec-
tor equation

ma = F, (1)

where the acceleration a is measured relative to some "inertial" base.

Such a vector equation is independent of the choice of coordinates; given
a choice of coordinates, it is equivalent to three scalar equations in the
three (rectangular) components of sand F.
The calculus enter essentially because the acceleration of the particle is
defined to be the second time derivative of its position vector s (the vector
from the origin of the inertial base to the planet). Using Newton's nota-
tion of one dot for one time derivative, this definition can be written as
a = s. In particular, if the force F on the particle is zero, so is the
acceleration a. From this it follows that the vector velocity v = s is con-
stant. This is Newton's first law: A particle under the action of no forces
remains at rest or moves in a straight line with constant velocity. Newton
also formulated a universal law of gravitation: The force of gravitational
attraction between two point masses m] and m2 is directly proportional to
the product of these two masses, inversely proportional to the square of
the distance r between them, and directed along the line joining the
bodies. In other words the magnitude F of the force is given by the inverse
square law

(2)

where y is a constant (the gravitational constant) in suitable units. This is

called a "universal" law because it applies effectively both to the motions
I. Kepler's Laws 261

of planets and stars, and to the more local motions of bodies (such as pro-
jectiles) in the earth's gravitational field. Newton presented both kinds of
applications in his famous "Mathematical Principles of Natural Philoso-
phy" (1687), but the presentation there did not explicitly use the calculus.
The full development of the subject of "Newtonian Mechanics", with its
use for extended bodies, fluid mechanics, and the like, required more than
a century and was advanced by many other noted mathematicians, espe-
cially Leonard Euler (1707-1783).
Newton did deduce Kepler's laws from his; let us summarize the pro-
cess. Consider the motion of a planet of mass m relative to the sun. We
neglect the gravitational forces from the other planets, so that we have a
two-body problem, with sun and planet as the two bodies. At a chosen ini-
tial time, the vector force on the planet is directed toward the sun, and the
planet has some initial vector velocity. If this velocity happens to be
directed toward (or away from) the sun, all the subsequent motion (by
Newton's second law) must be along the line joining the planet to the sun.
Leaving aside this exceptional case, the initial velocity and the line from
planet to sun together determine a plane containing both the velocity and
acceleration vectors. Hence, by Newton's second law, all the subsequent
motion takes place in that plane. This is the derivation of that part of
Kepler's first law which asserts that the orbit is planar. The proof did not
use the inverse square law.
To get the shape of the orbit, we describe the plane of the orbit by rec-
tangular coordinates x and y relative to the sun 0 as origin. The gravita-
tional force F on the planet P is then proportional to Ilr 2 and directed
from P to the origin 0, where r is the distance OP, This suggests the intro-
duction of polar coordinates (r,(}), as in Figure 1. These polar coordinates
convert to rectangular coordinates (x,y) by the familiar equations

x = r cos (), y = r sin () (3)

which embody the definitions of sine and cosine. Therefore the force F is
a vector of magnitude F with components - F cos () and - F sin () along
the x and y axes. Thus Newton's second law, written in components,
becomes

mx = -F cos (), my = - F sin (); (4)

Figure 1. Coorinates for an orbit.

262 IX. Mechanics

here the double dots represent the second derivatives of the (variable)
coordinates with respect to time. These two differential equations can be
combined to give the equation

m(xy - yx) = °
which states that the time derivative of yx - xy is zero, so that
yx - xy = k (5)

is a constant k.
This formal deduction has substance. The quantity yx - xy on the left
of (5), when multiplied by the mass m, is called the angular momentum of

°
the planet P about the point 0. So the proof has shown that the angular
momentum about is constant when the force is directed toward (or away
from) 0. In polar coordinates, the equation (5) for angular momentum
becomes, by (3),

(6)

this disp!ays the relation of angular momentum about °

to- the angular
velocity () about 0. In rectangular coordinates, consider the position and
velocity vectors as three dimensional vectors with components (x,y,O) and

°
(x,y,O). Their vector product (§IV.14) is then (O,O,yx - xy); multiplied
by m, it is the momentum, considered as a vector through perpendicular
to the plane of the motion.
This is also related to the area A = A(t) swept out by the radius vector
OP in time t; Figure 2 indicates that the increment in area A(t) due to an
increment 6,() in the polar angle is approximately the area 'hr 26,() of a cir-
cular sector of radius r and angle d(): Hence, by .the usual limiting pro-
cedures, the rate of change of area is A = (1/2)r2(). Thus the equation (6)
contains that part of Kepler's first law which states that area is swept out
at a constant rate; again the conclusion does not depend on the inverse
square law!
To get the equation of the orbit, we must use the inverse square law in
the form F =: mp,lr2, where p, is some constant. By (6), this can be written
F = (mp,lk)(); then the second order differential equations (4) become

x= - (p, Ik)cos () (), y = - (p, I k )sin () ().

OL----L-------+----

Figure 2. Area swept out.

I. Kepler's Laws 263

One integration then yields

x= -(M/k)(sin () + A), j = (M/k)(cos () + B),

where the constants of integration A and B are to be determined by the
initial velocity at the initial state. When these two equations are entered
into the expression (5) for the constant k of angular momentum, we
obtain

k 2 = M(x cos () +y sin ()) + M(xB + yA).

Since r =x cos () +y sin (), this equation can be written as

(7)

By putting in the polar coordinate values (3) of x and y, this is the equa-
tion of the orbit written in polar coordinates. More directly, the right
hand side in (7) is proportional to the distance from the line whose equa-
tion is that right hand side set equal to O. It is easier to see this by choos-
ing the initial conditions at a point where () = 0 and x
is zero. Then
A = 0 and the equation (7) reads

(8)

Now k 2 /MB - x on the right is clearly the horizontal distance from the
planet P to the (vertical) line x = k 2 /MB. Then equation (8) states that
the planet moves so that its distance r from the sun at 0 is always a fixed
proportion B of its distance from a vertical line. But a conic section (Fig-
ure 3) can be defined, in terms of a point 0 as focus and a line D as direc-
trix, as the locus of all those points P with distances from 0 and D so that
PO = ePD, where the constant e is the eccentricity of the conic. The orbit
must then be an ellipse (e < 1), a parabola (e = 1), or a hyperbola
(e > 1); in the case of the planets we know by observation of their
recurrence that the orbit must be an ellipse (with the sun at one focus). As
one knows, an ellipse can also be defined in terms of two foci, as the locus
of points P such that the sum of the distances from P to the two foci is a
constant 2a; then 2a is the length of the "major axis" of the ellipse. A
straightforward analysis will then give Kepler's third law about the square
of the periods of different planets. But we have already established the
essential point: When velocity and acceleration are represented by deriva-

CJ)-+--I"-
Figure 3. An ellipse as a locus.
264 IX. Mechanics

tives, as defined in the Calculus, then Kepler's laws, summanzmg the

observations of the positions of the planets, can be deduced from
Newton's second law of motion and the inverse square law of gravitation.
Moreover, the deduction leads naturally to physical quantities such as
momentum and to Mathematical ideas-the choice of suitable coordinate
systems and the process of integrating differential equations, using initial
conditions to determine the inevitable constants of integration. The
further scope of these ideas is tremendous!

2. Momentum, Work, and Energy

Newton's third law of motion states that action equals reaction. This
means, say, that in a mechanical system consisting of just two particles the
force exerted by the first particle on the second will be equal and opposite
(equal in magnitude and opposite in direction) to the force exerted by the
second particle on the first. It is this law which gives rise to the principle
of the conservation of momentum of an isolated mechanical system.
The (linear) momentum of a mass particle is a vector defined as the
mass times the vector velocity. Thus, for a particle of mass m with coordi-
nates XI,X2, and X3 the momentum is the vector mv with coordinates
mXI, mX2, and m"i 3 . Since the acceleration is the time derivative of the
vector velocity v, Newton's second law (l.l) becomes the statement that
the vector force on a particle is the time rate of change of momentum; in
particular, when the force is zero, the momentum is conserved. Also, for
an isolated system of two particles, exerting equal and opposite forces on
each other, the total momentum of the system is constant. The same con-
clusion holds for systems of n particles.
Next comes the notion of work-specifically, the work done by a vector
force F in moving a particle along a path. In the simplest case, when a
constant force F moves a particle by a distance d along the direction of
that force, it is said to do work of magnitude the product Fd. More gen-
erally, the work done by a vector force F in a vector displacement w
depends only on the component of F along the direction w, so is the prod-
uct of that component by the length of w; in more invariant terms, it is
the inner product F·w of these vectors. (The inner product is often called
the scalar product, to distinguish it from the vector product which is also
used in mechanics.) This formula, work equals F·w, suggests that force
acts as a linear function of displacement, so that a variable force will
appear as the dual of a vector field; that is, as a differential form.
This leads to the following formulation. In an open set U in R 3, with
rectangular coordinates x I, x2, and x 3, a field of force F has three com-
ponents FJ, F2, and F3 depending on position, as Fi = Fi(X l,x 2,x\ and
so determines a (smooth) differential form

(1)
2. Momentum, Work, and Energy 265

on U. We want the work done in moving along a path, given as usual by a

smooth map u: I ..... U of a time interval, say I = {t I 0 t tI}' This < <
work will be approximately the sum of small bits of work along short
pieces of the path; hence, by the main idea of integration, the work done
by the force F along the path u is (defined to be) the line integral

f u
w= f u
(FIdx I + F2dx 2 + F3dx 3)
(2)
= f o(FIX ll.I
+ ·2
F2 x + .3)d
F3X t.

This definition of work as an integral does include the special case (of a
constant force), as introduced above. What we call a "differential form" in
(1) is in physics texts usually called the differential 8W of work, but the
idea is the same-it is the thing which is to be integrated (along paths) to
get the work done along these paths; the symbol 8W does not necessarily
mean that there is a function W of position with this differential.
Now use Newton's second law for a particle of mass m moving under
this force field F along this path u, given by smooth functions Xi(t). Then

i = 1,2,3.

The indefinite integral of this product is (1/2 )(m(x I )2), so that the work
done along the path is

f U
w = (l/2)m[(xI)2 + (X 2 )2 + (x3)21~1, (3)

where the notation at the right means the difference of the values of the
bracketed expression at tl and O. This leads to the definition of the kinetic
energy T of the particle as

(4)
and to the theorem that the work done along the path is equal to the
change (3) in the kinetic energy along that path. In brief, if v is the mag-
nitude of the velocity, then the kinetic energy is
T = (l /2)mv 2 .
Sometimes the differential form w = 8W
of work may be the
differential w = -dV of an actual function - V of position. This means
that the force F; in each direction Xi is the partial derivative

F,. = --
av i = 1,2,3. (5)
I ax i '
This implies that the work done along any smooth path from a point b to
a point c is just the difference V(c) - V(b)-and so is independent of the
266 IX. Mechanics

choice of path. This function V-unique up to a constant-is then called

the potential energy and the forces are said to be conservative. For exam-
ple, under the inverse square law for a particle moving under the gravita-
tional attraction of another particle at distance r there is a potential
energy of the form

v= -ymlr; (6)

indeed, the partial derivative with respect to the coordinate r is the

desired inverse square force -ymlr2. (The minus sign indicates that the
force is directed toward the origin.) The potential energy per unit mass-in
this case V = -y Ir-is often just called the potential. For several attract-
ing sources there can be similar potentials; the study of these potentials is
extensive, with profound connections to partial differential equations and
to probability theory.
When the forces are conservative, the work done along a path, as
expressed in the integral (3), is exactly the change in its potential energy
along that path. It follows that in such a system the sum of potential and
kinetic energy remains constant. This is a first form of the famous physi-
cal principle of the conservation of energy.
The harmonic oscillator is a simple example of a conservative system. A
particle moves along a line under a force of attraction directed toward the
origin and proportional to the distance x from the origin. Thus by
Newton's law, the motion is governed by the second order differential
equation

(7)

for some constant k. This force -kx may be derived from a potential
kx 2 /2 + C, for any constant C (potential energy is determined only "up
to" an additive constant). The equation (7) has the well known integral
A cos kt + B sin kt or

x = A cos(kt + cp), (8)

where A and B or A and cp are two constants of integration. In the form

(8) the constants have an immediate interpretation, because A is the larg-
est value of x (at kt = -cp), so is the amplitude of the motion, while the
angle cp measures the "phase" by which the motion is translated from the
simple cosine function A cos kt. The formula (8) also yields the velocity
as a function of time,

x= -kA sin(kt + cp).

The motion may be visualized by plotting x and x together in a plane,
called the phase plane. When the constant k is 1, the plot (Figure 1) is a
 3. Lagrange's Equations 267

------~------~~--~-----+x

Figure 1. Phase plane.

circle. A point Q moves around the circle at uniform speed; its projection
P on the x axis is the point executing the original harmonic motion. Such
"harmonic oscillators" are important because general periodic motions
(oscillations) can be built up (via Fourier series) out of linear combina-
tions of these "simple" harmonic oscillations. The reappearance of the
geometrically defined trigonometric functions cos kt and sin kt in this
connection with mechanics is a remarkable (and classical) instance of the
interrelation of mathematical constructs.
The idea of using the phase plane with coordinates both position and
velocity is relevant, because the initial conditions for the motion (position
and velocity) fix a point in this phase plane. This point (and thus these
conditions) determine the constants of integration A and <I> ; geometrically,
this point determines the circle above. Put differently, this point is a posi-
tion on the real axis and a velocity there, hence is a point on the tangent
bundle for the x axis. In this form, as we will soon see, the idea general-
izes.

3. Lagrange's Equations
The calculation of planetary orbits in § I started with the usual rectangular
coordinates x and y in the orbital plane of the planet, but soon switched
to polar coordinates in that plane. It would have been convenient to have
the equations of motion written directly in terms of polar coordinates.
This can be done, and done in a way to suggest the form of such equa-
tions in any other coordinate system.
The transformation from rectangular to polar coordinates rand () reads

x = r cos (), y = r sin (). (1)

From these we can express the rectangular components of velocity as

x= - r sin () () + r,
cos ()
(2)
j = r cos () () + sin () r ;
268 IX. Mechanics

hence the kinetic energy becomes

(3)

Differentiating the equations (2) again with respect to time gives the rec-
x
tangular components and ji of the acceleration vector a. The radial and
angular components are then (Figure I)

ar = x cos () + ji sin () = ;: - rif,

(4)
alJ = - x sin () + ji cos () = rji + 28;.
The force has two corresponding components, Fr along the radial direc-
tion and FIJ perpendicular to this direction, but for this "angular" direc-
tion it is more natural to use the torque TIJ about the origin, defined as FIJ
times the lever arm r from O. For these two components, Newton's second
law then reads
( ..
Fr=mr-r ()'2)
, TIJ = m(r2jj + 2rO;). (5)

The right hand side at first looks mysterious; for example, the term -r82
on the right in the first equation is often shifted to the le~t side, where it is
called the centrifugal force due to the angular velocity (). (Swing a horse
chestnut on a rope and you can feel the centrifugal force!) However, this
right hand side can be explained more systematically in terms of the par-
tial derivatives of the polar coor~inate kinetic energy function T of (3).
This T is a function of r, ;, and (); calculating its partial derivatives, one
finds

Fr = !£ [ aT 1 - aT,

:n
(6)
dt a; ar

T, = ~ [ (7)

We can even make these two equations look alike by subtracting aT / a()
on the right of (7)-a zero term because the kinetic energy T is in fact
independent of ().

------~o~-------------------x

Figure I. Angular acceleration.

3, Lagrange's Equations 269

As a result of these apparently mechanical computations, a miracle now

happens: Equations like these hold in any coordinate system. Specifically,
if q is anyone coordinate and T the kinetic energy, as it depends on q and
the q-component q of velocity, Newton's equation for this coordinate
becomes

d [
dt aq 1- aq
aT aT = Qq, (8)

where Qq (like the torque in (7)) is a "generalized" force in the q-

direction, and the partial derivatives with respect to q hold constant all the
other coordinates of the system. This is the Lagrange equation for the
coordinate q.
This Lagrange equation (8) does include the polar coordinate equations
(6) and (7) for the coordinates rand (J. The miracle is that this one form
of equation applies to any choice of coordinate system. This exemplifies
the observation that one Mathematical form has many realizations.
We will derive the equations (8) and the definition of generalized force
for the general case of N particles, where the position in space of each one
is given by the usual three cartesian coordinates; the position of the whole
system is then determined by k = 3N coordinates Xi, for
i = I, .. , , k = 3N; in effect we are considering the motion of a single
point in the space Rk. Then Newton's law makes

i= l, ... ,k, (9)

where each Fi is the appropriate force while mi is the appropriate mass

(thus ml = m2 = m3 is the mass of the first particle). The kinetic energy
T depends on the velocities according to the standard formula

T = (l/2)(mlxf + , .. + mnx;) , (10)

Then mi xi = a,T, so that Newton's law (9) can be written as

aXi

~[£l=F
dt axi l'
j l, ... ,k. (11)

Since T does not depend on the Xi, this is a special case of the Lagrange
equation (8)
Now replace the xi by new coordinates. These new coordinates are
traditionally denoted by q's, so that the xi are given by smooth functions
hi ,

x - hi( q,
i - n)
I ... ,q, j= 1, ... ,k, (12)
270 IX. Mechanics

in terms of the new coordinates. The partial derivatives ah i /aqi are tra-
ditionally written as axi /aqi, thereby avoiding the use of a letter h for the
function, and using the letter xi to denote both the quantity x J (the coor-
dinate) and the function hi of the q's. The components of velocity
dx i / dt = Xi along a path are then, by the chain rule,

axi axi
x
'J'
= --q
aql
·1
+ +--nq,
'n
j = l , ... ,k. (13)
aq

Consequently the kinetic energy T becomes a function of both the qi's and
the i/,s. In these coordinates we will show that Newton's second law takes
the form of the Lagrange equations

i = l, ... ,n, (14)

where the Qi are suitable "generalized" forces; for examples, torques.

Since the Lagrange equations hold for the case of rectangular coordi-
nates we can prove that they hold generally by showing that they remain
true under any change (12) from old coordinates xi (not necessarily rec-
tangular) to new coordinates ql , ... , qn. It is convenient to think of the
xi as coordinates of a point in one configuration space D, say the space
D = R3N, while the q's are coordinates in some possibly different
configuration space C of dimension n; the change of coordinates (12) is
then interpreted as a smooth map cp: C -> D of spaces; such a map is often
called a change of base (e.g., in algebraic geometry).
In order to transfer Lagrange's equations from C to D we need to
change both paths (trajectories) and functions (kinetic energy). The
diagram

u <P f
I ---7 C ---7 D ---7 R, (15)

with I an interval of time, suggests what happens. Each path u in the

space C becomes by composition with cp a path cp·u in D, while each func-
tion f or quantity defined on D becomes by composition with cp a function
fcp on C; one says that f is pulled back to C. Also the differential form w
which determines the work (and thus the forces) on D can also be pulled
back along the smooth map cp. First, each of the k coordinates xi on D
becomes by a composition a quantity xi cp on C, with differential there
given by the usual formula

j = l , ... ,k.
3, Lagrange's Equations 271

(Here axi / al are just the partial derivatives ah i / aqi of (13).) By these
equations, the differential form w = 'i.Fi dx i for work w on the original
manifold D becomes a differential form

cf>*w F dx 1, = ~ ~ F -axi
= ~, 1 - , dq',
, , 1 aq'
1 1 I

on the second configuration space C; collecting the coefficients of each

differential dqi, it has the expression
n
cf>*w = ~ Qidqi, (16)
i=1

The quantities Qi so obtained on C are by definition the generalized forces

in the direction { The essential point of this definition is that the work
done by given forces, obtained as the integral of the differential form w
over a path, is preserved; by using the definition of the line integral
(straightforward, as specified by (2.2» one shows, for each path u: I ..... C,
that

f </>·u
w = f u
cf>*w. (17)

The components xi of velocity are expressed as functions of the q's and

the q's by the equations (13), and by these equations the partial deriva-
.
tIves f h f . 'i( 1 n .1 . n) are
0 t ese unctIons x q " .. ,q ,q , ... , q

(18)

If I is a time interval, with parameter t, then along each path u: I ..... C all
the coordinates qi, l and xi become, by composition with u or with cf> ·u,
functions on I (i.e., functions of t). By the chain rule for differentiating a
composite,

Since ql = dql idl, the right hand side is the same as in the second equa-
tion of (18). hence

(19)

valid along any path u. This means that the partial derivatives here are
those of functions of q 1 , ••. ,qn, q 1 , ••• , qn; they then become functions
of I alone via u.
272 IX. Mechanics

With these preparations, we can now deduce the Lagrange equations

for the coordinates q by proving that: A path u: I ..... C satisfies these equa-
tions whenever the composite path </> ·u: I ..... D satisfies the Lagrange equa-
tions

d [ aT
dt ax}
I -
aT - F
ax} - }, j I, ... , k

in the x-coordinates on D. Indeed from the definition (16) of the general-

ized forces Qi one has

Qi = L F- ax} = L
J
. J aq' .
J

by the rule for the differentiation of a product. After substituting (19) in

the second term here, the second and third terms combine to give
aT / a{ As for the first term, the chain rule gives

aaq~ = L} I aa.~
x
aa~~q + aa~x aa~~q ;
here ax} / ai/ = 0, while ax} / ai/ is ax} / aqi by (18) so that the whole
becomes

i=l, ... ,n.

These are Lagrange's equations for the path u on C, as desired.

The Lagrange equations apply also to the case of motion under
constraints-for example a so called "holonomic" constraint, where a par-
ticle or particles is required (i.e. is constrained) to move only along some
curve or surface in Euclidean space. An example is a sphere rolling along
a smooth table or a pendulum where the pendulum's bob is suspended by
a weightless string from some fixed point 0. Then the motion of the bob is
just the motion of a point on a sphere, and so can be described in terms
of coordinates on the sphere-say the latitude and the longitude. In this
case the configuration space C is a manifold (the sphere) of dimension
n = 2, while the original configuration space D is R , Euclidean space of
dimension k = 3. The argument above applies directly to this case and
gives the Lagrange equations which describe the motion in terms of lati-
3. Lagrange's Equations 273

tude and longitude. Notice here that the dimension 2 of C is smaller than
that of D (hence k =I n in the formulas above) and that the configuration
space C is a manifold, included in the Euclidean space D by the inclusion
map cJ>: C -> D. Moreover the coordinates qi must be regarded as local
coordinates, valid in some chart-the longitude is not defined at the north
pole of a sphere!
In such a case there are forces Fj of "constraint" which hold the parti-
cle to the submanifold, as well as "external" forces Fj. Thus in the
Euclidean space the j th component of the total force is Fj = Fj + Fj.
But the forces of constraint act orthogonally to the submanifold C; as a
result the pullback to C of the differential form for work done by the con-
straints is cJ>*CJ.Fjdx j ) = o. More simply, "the forces of constraint do no
work". What this means is that the required generalized forces Qi on C
may be calculated by pulling back only the external forces ¥.
Lagrange's equations apply also to more general ("non-holonomic")
constraints, such as the case of a ball rolling on a rough table. At anyone
time, the ball is constrained to move on a three dimensional manifold,
with local coordinates the two rectangular coordinates of the point of con-
tact in the plane (the table) plus one angle (the angle of rotation of the
sphere about the diameter through this point of contact). However, by rol-
ling suitably forward and backwards, one may see that the ball can reach
any position in a four dimensional manifold (two coordinates in the plane
and two for rotation of the sphere). The Lagrange equation in such a case
may be derived much as in the argument above, replacing the holonomic
constraint cp: C -> D by a map cp: C X 1-> D which at each time tEl
gives the position cP t = cP ( - ,I) of the constraint to C at the time t. The
generalized forces Qi for the Lagrange equations of motion on C are then
obtained by pulling back the work w along CP*t; this is classically called a
"virtual" displacement.
Lagrange's equations have a simpler form when the forces are conser-
vative. There is then a potential V, which is a quantity defined on D such
that the forces Fj are given by the partial derivatives - 3V / 3x j . This
potential is then also a quantity defined on C (as the composite V·cJ». The
generalized forces Qi on C then are

.= ~ = _~ ~ =_
t } 3x3qi t
p. j 3x j 3V .
Q, 3x j 3qi 3qi .

In other words, they are also derivatives of the potential V in the new
coordinates. One may then introduce L = T - V, a quantity defined on
D and on C, and called the Lagrangian; the Lagrange equations for con-
servative forces then take the form, for coordinates qi on the configuration
space C,

i = l , ... ,n. (20)

274 IX. Mechanics

There are many practical applications of the Lagrange equations. In

the case of a simple pendulum, a particle of mass m is constrained to
move on a vertical circle of radius r under the force of gravity, with the
standard acceleration g. There is one generalized coordinate, the angle 0
(of the pendulum arm) to the vertical. The force is conservative, with
potential energy V = -mgr c~s O. The kinetic energy, transformed from
rectangular coordinates, is mr2(jl 12. Lagrange's equation in the coordinate
o then becomes
d .
dt (0) = - (g Ir)sin O.

An exact integration requires elliptic functions. If the angle 0 is small, sin

omay be approximated by 0, and the equation becomes that of a simple
harmonic oscillation.
The spherical pendulum consists of a particle of mass m constrained to
move on a sphere of radius r. The appropriate coordinates are longitude </>
and latitude 0 (measured from the south pole of the sphere). By the stan-
dard equations transforming rectangular to spherical coordinates one
obtains

V = - mgr cos o.
The two Lagrange equations in the coordinates 0 and </> then are

The second one integrates once to give ci> = h Isin 20, for h a constant of
integration. The first equation then takes the form

jj = h 2cos 0 _ K sin (J.

sin3(J r

If h = 0, this reduces to the case of the simple pendulum. The integration

in other cases is more complex.

4. Velocities and Tangent Bundles

The Lagrange equations,

d [ aT
dt ail
1 aT
aq'
i = l , ... ,n.
4. Velocities and Tangent Bundles 275

with Qi a function of position, are n second order differential equations in

the n coordinates qi, as one may exhibit by the expansion

(1)

In general, a solution will be determined by giving the 2n initial values of

the position and velocity coordinates qi and i/
Alternatively, we can think
of these 2n quantities as 2n independent variables. Then we replace
d 2qJ/dP by diJ' /dt and rewrite (1) as

n
a2T di/ a2T dqJ aT
- - . =Qi.
~ ail ai/ Tt + ail aqJ dt aq'
(2)
J=I

Since i/ is now a coordinate and not a derivative, we must add the equa-
tions

dqi _ .i
dt - q, i = l , ... ,n. (3)

We now have 2n first order differential equations in the 2n variables qi

and i/ for position and velocity (phase space), with coefficients functions
of these variables.
For rectangular coordinates, the equations (2) have the simpler form

In general, the matrix with coefficients a2T / ai/ aqJ is usually non-
singular; when this is the case, the equations (2) can be solved explicitly
for the derivatives dqJ /dt. If we label the m = 2n variables i ' ... ,ym,
the equations (2) and (3) then have the general form

d/ _
Tt- G( 1
kY,""y
m)
, k=l, ... ,m (4)

with each Gk a smooth function of the y's. A trajectory of the given

mechanical system is then a solution l = yk(t), k = 1, ... , m, of these
equations. Here the m quantities Gk can be considered as the m com-
ponents of a tangent vector at the point with coordinates yl , ... ,ym, so
that the Gk describe a vector field (see §VIII.9) on (part of) the m-
dimensional space Rm. A solution of the equations is then a smooth path
in Rm whose tangent vector at each point is the vector given by the field at
that point. More informally, a solution is a path which threads through
the given vector field, as suggested by Figure I for m = 2. In these terms,
one can readily visualize the corresponding "existence theorem", which
276 IX. Mechanics

Figure l. Paths through a vector field.

asserts that such a smooth system always has, for each given initial point,
a smooth solution defined for some sufficiently small interval of time-so
solutions exist, even though they may not be expressed in terms of
elementary functions.
The phase space for simple harmonic motion (§2) exhibited both posi-
tion and velocity. Similarly the quantities qi (position) and i/
(velocity)
are the coordinates for the phase space of the motion. A point in this
space is a point in the configuration space C plus a tangent vector at that
point. In other words, it is a point on the tangent bundle B. C to C-which
we write with the letter B, to avoid confusion with the use of the letter T
for kinetic energy. Thus the device of thinking of a point moving in such
a phase space is really equivalent to the geometric idea of using a tangent
bundle to a manifold.
Our proof that the Lagrange form of the equations is preserved under
change of base can be illuminated in these terms. Each smooth map
1> : C -'>D of manifolds carries tangent vectors (to paths) in C into tangent
vectors to D and hence induces a map B.1> : B. C -'> B. D on the tangent
bundles. On each tangent space, it is the linear map given in coordinates
by the equations (3.13). The determinant of this linear map, with entries
ax} / aqi, depending on the q's, is called the Jacobian of the transforma-
tion 1>.
In these terms, the diagram (3.15) can be expanded to represent all the
coordinates involved in the change of base (qi coordinates on C, x j on D):
qi
R .. B.C B.¢ • B. D
T
.R

(5)
qi 7T 7T xi

J - - - - -..... C - - - - - - " D - - - - - - . R.
U xi
5. Mechanics in Mathematics 277

Here the map 'IT is the projection of each tangent bundle (phase space)
onto its base, and the coordinates qi (or xi) on the base become, by com-
position with 'IT, coordinates qi.'IT on the bundle. This diagram also
displays the other functions involved. Thus the kinetic energy T, originally
given as a quantity on the initial phase space (i.e., a function T on B. D to
the reals) becomes by composition with B. cp a quantity on the new phase
space. By convention, the letter T usually stands for either (or both) of
these functions. The whole proof of §3 can then be reformulated in these
terms.
For iterated constraints one has successive smooth maps

of configuration spaces. In this case, successive images of the tangent vec-

tors of E are just the images under the composite map cp. '1'; in other
words, B. (cp '1') = (B. cp )(B. '1'). The fact that application of B. respects
composition in this way is sometimes formulated by saying that the
"tangent bundle is a functor B. on the category of manifolds"; see §XI.9.

5. Mechanics in Mathematics
We pause to emphasize the many and remarkable exchanges of ideas
from mechanics to Mathematics and back. To begin with, the very notion
of the calculus was found by Newton in order to formulate the mechanics
of planetary motion. The idea of rate of change (derivative) was needed
in order to get at the ideas of velocity and acceleration necessary for
mechanics. This in tum led to the consideration of (ordinary) differential
equations and their solutions (exact or approximate) subject to initial con-
ditions. Since these differential equations are typically of second order,
their initial conditions involved initial position and velocity and so led to
the use of first order differential equations on tangent bundles (phase
spaces) where the location of a point is determined by position and
velocity. For several particles, the configuration spaces often required
spaces of more than three dimensions. Both the configuration spaces and
the phase spaces were originally described just in terms of coordinates,
but the effective formulation of the differential equations of motion called
attention to the need for some invariance (say in the form used for the
Lagrange equations) under changes of coordinates. Once this is con-
sidered, one is really dealing with a conceptually described idea of a
smooth differentiable manifold. Although the explicit general definition
of such a manifold was not formulated till the 1930's, it was implicitly
present for at least a century before that in the minds of both geometers
and physicists.
278 IX. Mechanics

This is not all. The physicist's notion of the "differential of work" is the
Mathematician's "differential form". The total work in a process is
obtained by summing up differentials, and this amounts to the definition
of line integrals. Other such connections of ideas will appear later in this
chapter. They are discussed here as examples of the very many cases of
ideas connected from theoretical Physics to Mathematics. They cannot all
be listed in anyone book, but we can mention the recent exciting parallels
which have appeared between gauge theories (the Yang-Mills equations
in quantum field theory) and the Mathematical study of connections in
fiber bundles. The interaction between tensor analysis and relativity
theory is another such case (Weyl [1923]), as is the use of matrices and of
Hilbert spaces in quantum mechanics.
In such cases, a particular idea may arise first in Physics or first in
Mathematics or often with apparent independence in both. At issue is not
the question of which comes first, but the remarkable fact that both
come-that ideas from physical problems and from apparently pure
Mathematical speculation come together.

6. Hamilton's Principle
The form of the Lagrange equations in a configuration space C is
independent of the choice of coordinates in that space. This fact needs
explanation, and has one: Hamilton's principle asserts that the solutions
of Lagrange's equations are exactly those paths in C which "minimize"
the integral of the Lagrangian function L along the path. We consider a
conservative system, and a Lagrangian L which depends not only on posi-
tion and velocity, but also on time t. Such a "time dependent" Lagrangian
is thus a smooth function L: B. C X [~R, where [ is the time interval,
say all I with 0 < <I 1. Thus for each choice of coordinates ql , ... ,qn
on C, L appears as a f unction . L( q,
I...n ,q ' ) G'Iven two
,q I , ... , q. n,I.
fixed points a and b in the space C, we consider smooth paths from a to
b; that is, smooth functions u: [~C with u(O) = a and u( I) = b. Each
point of the path has a tangent vector at that point, so u( t) together with
these tangent vectors defines to a path UB: [~B. C in the tangent bundle;
we say that UB lifls u over 'iT because the projection 'iT: B. C ~ C will again
give u as the composite u = 'iT·UB. Also, sending each time t to itself
yieldsbyl I-> (uB.t,t)amapu#:[~B.C X [.
The integral J to be minimized is the integral of L along this path,

SoL(ql , ... ,qn,i/ ' ... ,il,t)dl = Jro(L·u# )dl;

I I
J(u) = (I)

the second version of this formula indicates that one is simply integrating
the composite function L·u# on [. We want to compare this integral J(u)
with the same integral along other smooth paths in C with the same end-
6. Hamilton's Principle 279

points (at the same starting and ending times t = 0 and t = 1), we aim
to find those paths u for which J(u) is a minimum or at least is "station-
ary" (see below).
Now to minimize (or to maximize) a smooth function f of a real vari-
able x one first finds where the derivative of f vanishes. Let us say that f
is stationary at some value Xo if df /dx = 0 for x = Xo. Thus a function
is stationary both at a maximum and at a minimum-and also at a hor-
izontal inflection of the curve y = f(x); to distinguish these cases one
needs more than just the vanishing of the first derivative. Now we wish to
find similar stationary positions not for a function like f(x) of a number
x, but for a function J(u) of a curve u; now u can vary over the set of all
curves from a to b. Such a problem is said to belong to the calculus of
variations (§ VI. 11 ). Here "variation" refers to the variation of the path;
this we interpret as an embedding of the given path u(t) in a one-
parameter family U(t,£) of paths, where £ is a real parameter, say £ E [.
Thus U is a smooth function U: [ X [---C, such that U(t,O) = u(t) is the
given path, while all the "varied" paths have the same endpoints a and b,
so that U(O,£) = a and U(l,£) = b for all £ in [. The situation may be
pictured as in Figure I (for a two dimensional space C). For each such
family U, one can again form the integral J( U); it will be a smooth func-
tion of £. We say that the original integral J(u) is stationary if dJ(U)/d£ is
zero for £ = 0 whenever u is embedded in a smooth family U of paths.

Hamilton's Principle. For each time-dependent smooth Lagrangian func-

tion L on a manifold C, a path u: I ---C from a point a to a point b
satisfies Lagrange's equations for L if and only if the corresponding
integral J(u) of the Lagrangian is stationary at u.
Since the integral J( u) depends only on the function L and the path u
and not on the coordinates, this theorem explains the fact that the
Lagrange equations have the same form, independent of the choice of
coordinates.
Note that this principle compares the actual path u (which does satisfy
the Lagrange equations) with other nearby paths, which do not in general
satisfy the Lagrange equations-but which do have the same starting and
ending points and times.

1
U
..
,~,
1

Figure 1
280 IX. Mechanics

The method of proof is adequately described by considering the case of

just one coordinate qI in configuration space (i.e., for n :::: I).
First assume that Lagrange's equations hold. We wish to show that J( u)
is stationary; that is, that the derivative dJjd£ :::: 0 for every family U. For
sufficiently smooth functions we may calculate this derivative of the
integral J by differentiation under the integral sign, as in

dJ ::::
d£
51 dL
0d£
dt:::: fl [
0
aL dq
aq d£
+ aL dq ] dt .
aq df.

Now along the lifted path in B. C one has

dq_d[dq ]
(f; -
dt (f; .

With this substitution, the first term in the integral above can be
integrated by parts according to the familiar formula for two functions v
and w oft;

f ov -d
l dw dv ('I
dt :::: - J, -d wdt
tOt
+ (VW)t=I-(VW)t=o. (2)

In the present case v:::: aLI aq and W :::: dqldf.; also all the varied paths
have the same endpoints, so that W :::: dq I d£ is zero at t :::: 0 and at
t :::: 1. Thus we get

(3)

Now set f. :::: 0 in L. The term in the large brackets is then exactly the
left-hand side of Lagrange's equation for the function L, so its vanishing
(for all t along the path) does imply that d.J j df. :::: 0 and hence that J is
stationary, as desired.
For the converse, assume that J is stationary for every family U, pick a
smooth function TJ: I ~R with TJ(O) :::: TJ(l) :::: 0 and use these functions to
construct a variation of the path as

q(t,£) :::: q(t) + £TJ(t).

For this special family U the formula (3) still holds, this time with
dqldf. :::: TJ, so

o :::: So
I -- -
d [aL]
dt
+ -aL
aq aq
I
TJ(t)dt. (4)
6. Hamilton's Principle 281

This holds for every such smooth function 1/. The following lemma then
shows that the term in the large brackets must vanish-in other words,
Lagrange's equation must hold.

Lemma. If M: I -->R is a smooth function of t for which

I
foMIJdt = 0 (5)

for every smooth 1/: I -->R with 1/(0) = 1/(1) = 0, then M is identically zero.
PROOF. Suppose not, so that M(t3) =1= O-say M(t3) > O-for some t3 in
the interval I. We can assume that t3 =1= 0 and t3 =1= 1. Since M is continu-
ous, one must then have M(t) > 0 in some small interval about t3. Now
choose b to be a bump function b: I -->R: smooth, zero outside this small
interval, positive inside the interval and equal to 1 at t3. Next choose the
variation to be 1/ = bM. Then

J.oM1/dt = f
I I

0
bM2dt > 0,

in contradiction to the hypotheses (5).

The methods used in this argument are not limited to mechanics; they
also apply when the Lagrangian is replaced by other smooth functions K.
In the plane with coordinates x and y, consider all smooth curves
y = y(x) from a point (XO,Yo) to a point (xl,YI). One wishes a curve for
which an integral

S XI
K(y,y I,X )dx (6)
Xo

is stationary, where y I = dy /dx. The proof above shows that it is neces-

sary and sufficient that the curve satisfy Euler's equation

~ aK _ aK = 0, (7)
dx ay' ay
which for K = L is Lagrange's equation. Such problems arise early in
many connections-for example, in the problem of finding the path of
quickest descent in a vertical plane (the brachistocron) from the point
(xo,Yo) to the point (xl,YI ). For this purpose, one really wishes not just a
curve where the integral (6) is stationary, but one where it is actually a
minimum (or a maximum) among suitable comparison curves. This study
gave rise to a remarkable array of rigorous methods in the Calculus of
Variations-including methods which would apply when the curves used
are allowed to have comers or when the minimizing curve is subjected to
various kinds of "side conditions", On the other hand, in mechanics the
idea of characterizing trajectories as those paths which minimize a suit-
282 IX, Mechanics

able integral has several different forms, with different choices of integral
and "side conditions". Several of these forms go by the name "Principle
of least action". For our form, the "action" can be defined as the integral
over the trajectory of the difference between kinetic and potential energy
(the Lagrangian L = T - V). This is in contract to Newton's second law,
which describes the same trajectory in terms of a local property (a second
order differential equation). For an eloquent description of the physics of
the principle of least action we refer to Feynmann, Leighton, and Sands,
vol. II, lecture 19.
To summarize: The explanation of the invariant form of Lagrange's
equations depends on minimizing a suitable integral over families of
curves, and is intimately tied to the development of the Calculus of
Variations-which in recent times has reappeared as the theory of optimal
control.

7. Hamilton's Equations
Another, more invariant form of the equations of motion is provided by
Hamilton's equations. They use as coordinates not positions and velocities
but positions and the corresponding momenta; they assume that the
forces are conservative and so are derived from a potential function V. In
place of the Lagrangian, they use the total energy H
H=T+V
(the hamiltonian), considered as a function of position and momentum.
Consider first the case of N particles, with position determined by
3N = k coordinates xi (or qi) and the trajectories given by Newton's laws
in the form (3.9)

i=l, ... ,k. (1)

Since the forces are conservative, Fi = - aVI axi - aHI ax i, while

the ith component of momentum is Pi = midxi Idt. These equations are
thus rewritten as

dPi
(2)
dt

On the other hand

dl
dt = Pi Imi can be found from the kinetic energy

dx',]2
2
2" ~ mi [dt
1
T = ~~~.
2 mi
7. Hamilton's Equations 283

as Pi1mi aT! api, so that, since V does not depend on the momenta,
dqi aH (3)
dt api
In other words, in the 2k dimensional momentum phase space, with
k = n and with cartesian coordinates ql, ... , qn (for position) and
PI , ... , pn (for momenta) the trajectories are the solutions of the 2n first
order differential equations

dPi dl aH i = l , ... ,n. (4)

dt dt a/'
These are called Hamilton's equations.
These equations hold in much greater generality, for any constrained
motion subject to conservative forces. To show this, we will deduce them
from the 2n first order Lagrange equations

d [ aL 1 aL - 0 i = 1,2, ... ,n, (5)

dt ail -al-'
. the n
m 2 coord'mates q I , ...
n , q ,'q I , ... , q. n . Wh en th erst
fi n coord'1-
nates q I , . . . , qn are fixed, they determine a point x in the configuration
space C, the second n coordinates qI , . . . ,qn are then those for a vector
in the tangent space BxC to C at x. We propose to simplify these equa-
tions by replacing these latter n coordinates by the momenta PI , ... ,pn,
defined as functions of the q's and q's by
aL
Pi = aqi' i = l , ... ,n. (6)

In the cartesian coordinate case this formula does give exactly the usual
components of momentum. Moreover, as in that case, we will assume that
the matrix with entries

a2L i,j = 1, ... , n.

al ai/'
is always non-singular, so that we can solve the n equations (6) for the i/
as functions

q. J -_ q. J(q I n )
, ... ,q ,PI, ... ,pn (7)

(we will also use the partial derivatives of these functions). Thus L, origi-
nally given as a function on the velocity phase space,

L = L(q,
I
... ,q n,q·1 , ... , q.n) (8)
284 IX. Mechanics

can then also be expressed as a function of the p's and q's on the momen-
tum phase space. (Mind your p's and q's!)
Now define the Hamiltonian as the quantity
n
H -- ~
~ Pjq.j - L, (9)
j=l

to be considered via (7) as a function of the p's and q's. Its partial deriva-
tives relative to the p's are then found by the usual chain rule for the
derivative of a composite as

aH .i ~ ai/
ap ·
I
= q + j=l
~ Pj apl'

in the partial derivatives on the right, the other variables to be held con-
stant are those indicated in the functional dependencies (7) and (8). But
because of the definition (6) of the pi, the two sums here cancel; with the
second set of Lagrange's equations (5) the result becomes just

i=l, ... ,n. (4a)

For the remaining coordinates q 1 , ••• , qn the partials of H, again with

dependencies (7) and (8), are

By the definition of the p;'s, the two sums again cancel; with the first of
Lagrange, it becomes

i=l, ... ,n. (4b)

We have thus deduced Hamilton's equations (4) from Lagrange's equa-

tions; we have presented this as a sort of trick which makes a transforma-
tion of variables so as to simplify the form of the Lagrange equations. The
trick seems to consist in looking at the Lagrange equations, putting Pi for
the terms aLI ail so that the Lagrange equations are automatically in the
solved form for dPi Idt-and then pulling H out of the air by the definition
(9). Indeed, many standard treatments derive the equations in just this
way, usually without bothering to explain (as we did in (7) and (8» which
other variables are to be held constant in which partial derivatives.
But what appears as a trick is in fact an idea-an idea which must have
been clear to Hamilton when he did it. But we claim that in general most
of the formal tricks appearing in Mathematics are really ideas in
7. Hamilton's Equations 285

disguise-ideas presented as manipulations because the manipulations can

be made explicit while the ideas are a bit nebulous. So let us get behind
this one trick.
Fix the coordinates q!, ... , qn; they determine a point x in the
configuration space C. The velocity is then the vector with components
q! , ... ,qn in the tangent space BxC to C at the point x; call this vector
space W, as in the diagram (10):
q' l e p
R ••---BxC=W------------W*=BxC 1. R

(lO)

R X.EC R

in this diagram each "quantity" is represented as a function (arrow) on

the appropriate space to the reals R. In particular, everything starts with
the Lagrangian L, which for our fixed point x is a quantity Lx on
BxC = W; this is a vector space on which the n quantities q! , ... ,qn are
the coordinates. Now the main terms of the Lagrange equations have the
n partial derivatives aLx / a{ They depend on the particular choice of
coordinates q! , ... , qn, but they are exactly the coefficients (for these
coordinates) of the total differential dL x of the function Lx :
n aLx .j
dLx = ~ - . j dq . (11 )
j=! aq

Now a differential is just a vector in the cotangent space W* = BXC, the

space dual to W-hence our picture also includes the cotangent space,
projected down by '1T to x E C. This should actually be said more carefully.
The differential dLx is taken at some point w of W-the point with coordi-
nates q! , ... ,qn; hence the differential is a (cotangent) vector in the
cotangent space to W at this point w. But the tangent space to Wat any
one point w may be identified with W-just translate each vector from the
origin in W to a vector starting at w. By the same device (a "canonical"
isomorphism) the cotangent space to W at this point may be identified
with the dual space W*. Now we have the idea behind the trick change of
variables: It is to transform each point w of the tangent (or velocity) space
into that point of the dual space W* given by the differential dL x of Lx at
286 IX. Mechanics

that point w. The usual dual coordinates PI , ... ,pn of this point in the
dual space are then, as in (11), the quantities

aLx
Pi = -'-i
aq
exactly as in (6). The transformation now appears in the figure (10) as the
arrow (): W -> W*; it is classically called the Legendre transformation for
the quantity L.
Now that we have the transformation () we would like it to have an
inverse ()- I-and it would be handy to have that inverse given as the
Legendre transformation for some suitable quantity K on the cotangent
bundle B·C. This means that we would like to have 1/ = aKjapi, so as a
first guess we might set K = 'ip/l But this won't be quite right, because
the i/, via ()- I, will be functions of the p's and so the partial derivatives
will be
aK a.i aL a.i aL
-
api
= i/ + i~n=
I
Pi -q- = 1/
api
+ ~ - .. -q-
i aqJ api
= i/ + - api .

This means that we will get the right quantity H if we set H =K - L.

Th en -aH ..q,i and
- IS
api

This is the formula for the Hamiltonian, previously pulled out of the air
at (9). Now it is no wonder that the differential equations become simpler.
The full picture (for all the points x of the configuration space) is

~l 81 PI
R ...- - - B , C - - - - " B' C-----'--_" R

There is still another twist: In the standard case, the kinetic energy is a
Riemann metric on the configuration space. In more detail, the kinetic
energy in the N particle case is a positive definite quadratic form
8. Tricks versus Ideas 287

(l !2)'i. mi(ili in the velocity coordinates; hence in cases of constraints it

will still be such a form, perhaps T = 'i. g;//qj, with coefficients gij in the
coordinates q' of the tangent space W = BxC. Such a form is at once an
inner product on that space W; we have already observed (in Chapter
VIII) that a Riemann metric on a manifold C may be described as a
smooth way of specifying inner products on the tangent spaces to the
manifold. As a homogeneous quadratic function, T can be expressed by
the formula

2T = ~'i aT ~ ·i
~ q - .i = ~ q Pi ,
i aq i

where we have used the definition of the momenta Pi.

Therefore the Hamiltonian H defined in (9) can be written

H = ~Piqi - L = 2T - (T - V) = T + V. (12)

In other words, in these cases the Hamiltonian is exactly the total energy
T + V.
Moreover an inner product < -, - > in a finite dimensional vector
space W determines a canonical isomorphism cp of W to its dual; indeed,
as in §VII.5, cp sends each vector v in W to the linear function
cpv = < v, - >: W ..... R. In the present case the inner product of vectors
v = (Vi) and w = (Wi) is written as 'i.gijViW j , where the entries gij make
up a positive definite symmetric matrix. Thus the linear function < v, - >,
inner product with v, has as its jth coordinate "2.gij v'. But, except for a
factor 2 and a change of notation (Vi for qi) this is just the quantity

_ aT _ a ~ .i·j
Pi - ail - ail ~ gijqq

used to define the ith component of momentum. In other words, the

Legendre transformation () replacing the qi by the Pi is, on each tangent
space BxC = W, just the canonical isomorphism (§VII.5) from the inner
product space W to its dual W*. In this way, the Riemann metric given by
the kinetic energy determines, via duality, the Legendre transformation.

8. Tricks versus Ideas

Hamilton's elegantly symmetrical equations may be deduced from the
Lagrange equations-either by a trick or by conceptual use of various
underlying ideas. For the trick we merely plug in new coordinates Pi for
the old qi, put down a formula for H and perform a few partial
differentiations and lo-Hamilton's equations. The conceptual analysis is
288 IX. Mechanics

longer but more revealing. First, the new coordinates Pi are physically the
components of momentum, defined by aLI ail Taken together, they are
the components of a differential, the differential dL x of the Lagrangian
along the tangent space at a point x. (It is sometimes therefore called the
"fiber derivative" of L.) The change from the i/ to the Pi is not just a jug-
gling of variables, it is thus a transformation () from the tangent space to
the cotangent space, as determined by L. The inverse transformation is
then given by a different function on the cotangent space-and this func-
tion is the Hamiltonian. Therefore the equation aH I api = dqi Idt (from
Hamilton's equations) just is a statement that H does give the inverse
transformation. Moreover, the kinetic energy is a quadratic form and
therefore an inner product on the tangent space W -and the transforma-
tion () is really just the standard isomorphism of such a space to its dual,
the cotangent space.
These remarks help to understand what is happening and they serve to
connect this development with all sorts of other mathematical ideas, in
particular, ideas from linear algebra and manifold theory. They surely
would not have been formulated, at least in this language, at the time
Hamilton first set up his equations. However, many of the ideas involved,
though not these technical formulations, might well be in the backs of the
minds of people who present the calculation simply as a quick trick.
This case is of considerable interest because of the wide further
development of Hamiltonian dynamics, hinted at below. It also implicitly
raises the question of the possible conceptual background of other tricks.
Analysis is full of ingenious changes of coordinates, clever substitutions,
and astute manipulations. In some of these cases, one can find a concep-
tual background. When so, the ideas so revealed help us to understand
what's what. We submit that this aim of understanding is a vital aspect of
Mathematics.
Understanding is not easy, as I may perhaps indicate by anecdote. It
has taken me over fifty years to understand the derivation of Hamilton's
equations. I first saw them in 1929 in a book on Theoretical Mechanics by
Sir James Jeans, a noted British applied mathematician. That came in a
course taught by Professor E. W. Brown, an expert on celestial dynamics.
He must have felt that the formal presentation by Jeans was inadequate,
for he turned away from his previous steady reliance on Jean's text and
lectured to us from his own handwritten notes. I don't really recall under-
standing the lectures, because I was more impressed with the yellowed
and dog-eared condition of those notes. The very next year I heard the
subject again, this time from Leigh Page, a professor of physics. I can tell
now what he said, because it was all much as in his book Introduction to
Theoretical Physics. He plays the tricks and pulls the Hamiltonian H out
of the air (and a touch of the Calculus of Variations); there was of course
no mention of tangent or cotangent spaces. Nearly 40 years later, I took to
lecturing myself on the subject. My then students took and published
9. The Principal Function 289

careful notes of my lectures. In these notes the trick is duly decked out
with tangent and cotangent spaces. Evidently that didn't satisfy. because
there is then the sentence "We want to understand better how the scheme
produced these equations". There follow two pages of attempts at under-
standing ending "The Hamiltonian function arises from asking the ques-
tion: When is () invertible? (This is probably not the way Hamilton found
it.)". Two years later I found this measure of understanding inadaquate. so
tried again in an article "Hamiltonian Mechanics and Geometry" in the
American Mathematical Monthly Vol. 77 (1970). pp. 57~586. Then after
a twelve year respite the understanding there purveyed seemed formal
(e.g .• in the identification of the tangent spaces to W with W itself) so I
tried again ... to get the presentation above.
The point of this cautionary tale is not the individual events. but the
difficulty of getting to the bottom of it all. Effective or tricky formal ma-
nipulations are introduced by Mathematicians who doubtless have a guid-
ing idea-but it is easier to state the manipulations than it is to formulate
the idea in words. Just as the same idea can be realized in different forms.
so can the same formal success be understood by a variety of ideas. A
perspicuous exposition of a piece of Mathematics would let the ideas
shine through the display of manipulations.

9. The Principal Function

There is another. remarkably fruitful. method for integrating the
differential equations of motion for a conservative system. This method.
suggested by the study of waves in geometrical optics. leads. as does the
theory of waves. to a partial differential equation (P.D.E.) which is useful
in classical mechanics and also in quantum mechanics.
Start with Hamilton's principle. which asserts that the trajectories
qi = qi( t) of the motion in the configuration space C are those paths
q = (ql • ...• t) which render the integral

of the (time-dependent) Lagrangian L stationary-and this in comparison

with other smooth paths starting (at t = 0) at the same point l = a i and
arriving at the same point l = q\ at the same time ti; here we use the
subscript 1 to indicate time and position of arrival. The solutions for the
(2n. first order) Euler equations for this variational problem depend on
the initial values qi = a i and i/= b i of the position and velocity vectors
q and q. These solutions will usually be smooth functions

i=l •... ,n. (1)

290 IX. Mechanics

Now a trajectory of a particle in mechanics is much like a ray of light in

optics, and the latter subject uses a "principal function" W which is essen-
tially the value of the integral above, taken along the actual trajectories,
for all values of the initial conditions ai, bi. In other words the principal
function for the conservative mechanical system is

W(a,b,tl) = So L(q,q,t)dt.
II
(2)

Now Hamilton's principle compares paths with the same starting and
arrival points (not starting velocities b). Hence we wish in (2) to replace
the initial velocities b by the final positions ql. Indeed, the equations (1)
for the trajectories do determined (for t = tl) the coordinates qil of the
point of arrival. The implicit function theorem, under suitable conditions,
states that these equations (with t = tl) can be solved for the initial
velocities b i as functions

i = l , ... ,n (3)

of initial position, final position, and final time (implicitly, also functions
of initial time). Because of the form of Hamilton's principle, we wish to
express "everything" in terms of these quantities ai, q\. and tl, regarded
as 2n + I coordinates on the manifold C X C X I. For example, by
substituting the solution (3) in (1), the trajectories themselves have coordi-
nates given as functions

(4)

of these quantities and time t. Here t is the time "along" the trajectory, so
that the components of velocity are

(5)

Similarly, the principal function W expressed in terms of these variables is

called S; thus

S(a,qJ,tl) = W(a,f(a,ql,tl ),tl)

(6)
(II
= J, L(g,g,t )dt ,
o
with g the vector with components gi as in (4) because the integral is
taken along the trajectory.
Now we calculate some of the partial derivatives of S, considered as a
function of a, ql and tl, so that each partial derivative with "respect" to
one of these quantities is taken while holding the rest of them fixed. By
differentiating (6) under the integral and using (4) and (5) we get
9. The Principal Function 291

Integrating the first terms by parts yields

In the remaining integral the term in brackets is just the ith term of the
Lagrange equations; hence the integral vanishes. At the start
(t = 0), gi = ai is constant, so the partial ag ilag! vanishes; at the end
(t = t]), gi = qL so agi/ag~ = 8) and the final result is simply

as j = l , ... ,n, (7)

- . = Pi'
aqi
because the momenta are defined by Pi = aLla;/! The result is simple,
and so is the proof, provided one pays careful attention to which variable
is where and a function of what-which is why we have here departed
from convention by writing g for the functions representing q along the
trajectory.
Next we consider the partial derivative of S with respect to t]. By (6)
and (1) we may express the function W in terms of S as
W(a,b,t]) = S(a,u(a,b,t] ),t]). (8)

Hence, holding the other variables in W constant, the chain rule gives

But the partial derivative on the left is the derivative of the integral of L
with respect to its upper limit, so is just L. Also as laq\ is Pi by (7).
Hence, solving for aS/at],

The quantity on the right is by definition the negative of the value of the
hamiltonian H at the endpoint t = t]. Thus

-as + H(q,p,t) = 0
at
holds at the endpoint t] of each trajectory. Since any point on the trajec-
tory is the endpoint (of a shorter trajectory) it holds for all t. Moreover,
Pi = as laqi by (7), so the last equation reads
292 IX. Mechanics

-as + H(q I , ... , qn, -as] as

, ... , - , t ) = O. (9)
~ ~ ~n

This is the Hamilton-Jacobi partial differential equation for the principal

function S. It is just as lat plus the Hamiltonian H, with each argument Pi
in H replaced by the corresponding partial derivative as I aqi. This is a
conceptually appropriate replacement, since at each point of the
configuration space C the principal function S, considered just as a func-
tion of the coordinates q of the point, has a differential involving just
these partial derivatives

dS = as] dq] + ... + as dqn

aq aqn '
and hence determines a cross section q f-> (q,dS) of the cotangent bundle
(phase space) B-C, where the coordinates are the q's and p's. The
Hamilton-Jacobi P.D.E. (9) simply uses the value of the Hamiltonian H
on this cross section.
To summarize: A conservative mechanical system with Legrangian L
determines a "principal function" S as the integral of L along the trajec-
tories. As a function of time and the end-point of the trajectory, S satisfies
the Hamilton-Jacobi (HJ) partial differential equation (9). Conversely, as
the next section will indicate, suitable solutions of the HJPDE give the
trajectories, often called the characteristics of the P.D.E. This is a first
indication of the importance of P.D.E. in mechanics (and also in quantum
mechanics) and a starting point of the subject "Differential equations of
Mathematical Physics"-see Frank-v. Mises [1930].
This also reveals the connection between mechanics and geometric
optics. In the latter subject, one wishes to determine the path of a light
ray through a specified medium from a point A to a point B. It is given by
Fermat's principle: The actual path from A to B is the one which makes
the time from A to B (as a function of possible paths) a minimum, more
exactly, stationary. On the other hand the wave fronts, for rays emanating
from a point A are the loci in coordinates q where the corresponding prin-
cipal function S( a,q) is a constant. Both mechanics and optics thus
involve a "least action" principle, which therefore plays a major role in
the calculus of variations (see Caratheodory [1965]). This calculus is a
striking example of a single Mathematical form fitting a variety of
different facts.

10. The Hamilton-Jacobi Equation

Now consider solutions of this P.D.E.; one other partial differential
equation-the wave equation
10. The Hamilton-Jacobi Equation 293

(1)

-has already appeared, in §VI.1l. It has many solutions-for any smooth

function k( x) a wave u = k( x - ct), travelling forward, and for any
g( x) a corresponding wave u = g( x + ct) moving backwards. Typically,
P.D.E. have solutions depending in this way on arbitrary smooth func-
tions such as g and k.
For similar reasons the Hamilton-Jacobi P.D.E.

as as
at + H(q'aq,t) =0 (2)

will have many solutions for S as a function of ql , ... , qn, and t. We may
not be able to identify which one is the principal function of our mechani-
cal system. Instead we search for a solution S which is like the principal
function, in that it depends on n parameters ai, as a function
S(a l , . . . , an,ql , ... , qn,t), and this in such a way that the n X n matrix
with the entries

a2s i,j = I, ... , n, (3)

aaiaql'

is non-singular for all relevant values of the arguments. Such a family of

solutions S is known as a complete solution of the P.D.E. (2). It may not
be the intended principal function for the conservative mechanical system
determined by the Hamiltonian function H. Nevertheless, the Hamilton-
Jacobi theorem asserts that any such complete solution does provide tra-
jectories, by a family of solutions of Hamilton's equations depending on
2n parameters ai and ci , and that this family is constructed in just the way
sketched in §9 for the actual principal function. Specifically we introduce
n new parameters cl , to be considered as the initial values of the
momenta, and use them to set up n equations

i = l , ... ,n (4)

analogous to the equations (9.7). These n equations involve the qi in

S(a,q,t); by the hypothesis on the matrix (3), they can be solved (locally)
for the q's as functions of a, c, and t. The momenta are then defined as in
(9.7) by the n equations

as
aqi = Pi, i = l , ... ,n. (5)

On replacing here the arguments qi in S by the functions qi( a,c,!) just

determined, we obtain the p's expressed also as functions pi(a,c,t). The
294 IX. Mechanics

Hamilton-Jacobi theorem asserts that the ql and Pi so expressed as func-

tions of t will satisfy Hamilton's equations and so will provide trajectories
for the mechanical system specified by the Hamiltonian function H.
We now sketch the proof of this theorem, omitting such details as the
careful specification of the domains of a's and c's where the hypotheses
are to hold. First imagine that the proposed solutions q( a,c,t) are substi-
tuted for the variables q in the H-J P.D.E., and then differentiate (2) par-
tially with respect to (say) a 1. By the usual chain rule for the partial
derivatives of a composite, the result is

The same second partial of S (except for the order of the two
differentiations) can also be found by differentiating (4) with respect to t

We intend to assume S smooth, so that the order of partial differentiations

does not matter. Then taking the difference of these two equations (and
similarly for a 1 replaced by aJ) we get

n a2s aH
-
aqi
= 0, j = l , ... ,n.
i7:
~
aaJaqi api at
But we assumed above that the matrix (3) of coefficients here is non-
singular. Hence, for each i,

aH i = l , ... ,n. (6)

api
This is the first n of Hamilton's equations.
Now return to H-J and differentiate with respect to qi, to get

We can obtain the same second partial of S, up to the order of

differentiation, by differentiating the definition (5) of the momenta Pi rela-
tive to t:

i = l , ... ,n.

In view of (6), the first of Hamilton's equations, the summation here

agrees with that in the previous equation. Hence, subtracting, we have
II. The Spinning Top 295

aH + api = 0 i=l, ... ,n. (7)

aq' at '
This is, as intended, the second set of Hamilton's equations.
Thus we conclude that each complete solution S( a,q,t) of the H-J Par-
tial differential equation does yield a family of solutions qi( t), Pie t) of
Hamilton's e~uations (6) and (7), and that this family depends on the 2n
parameters a , ... , an,e I , . . . , en. There is of course no assurance that
the ai and ei represent the initial values of coordinates qi and momenta Pi
at t = O. We know only that they provide the necessary number (2n) of
constants of integration.
The case when the Lagrangian function L and hence the Hamiltonian
H is independent of time has convenient special properties. The H-J
P.D.E. then has the form

~+H[
at q, ~l=o.
aq (8)

For solutions q(t), pet) of Hamilton's equations one then has

d n aH aqi + aH dPi
dt H( q, p) = i~ = 0,
aqi dt api dt

zero because the two terms in the summation cancel by Hamilton's

equations. This proves that H(p,q) is a constant k along each trajectory.
Since H is the total energy, this conclusion is just a formulation of the
principle of the conservation of energy. Along each trajectory, it follows
that ~~ -k. Hence one may replace S by a function K,

K = K( q I , ... , qn) = S + kt,

which is independent of t. This function K satisfies a somewhat simpler

(and quite useful) version of the Hamilton-Jacobi P.D.E.

11. The Spinning Top

Mechanics developed by the treatment of many specific problems. That
notable child's toy, the spinning top, presents a striking problem in
mechanics which is essentially that of the gyroscope, a grown-up's instru-
ment. As in Osgood's excellent presentation (W. F. Osgood [1937] cf. also
Pars [1965] p. 113) consider only a heavy top symmetrical about an axis,
which is spinning (and tipping) about this axis, but so that the point of
the top at the end of the axis remains fixed. Such a top appears in Figure
1. The configuration space now consists of all the possible positions of the
296 IX. Mechanics

2 201

\ e I Euler's angles and the

~ heavy symmetric top
with one point fixed

Line of nodes

Figure I

top (with its point fixed). These positions can be described in terms of the
angles needed to move from rectangular axes xo, Yo, Zo fixed in space
(origin 0 at the point of the top) to axes x, y, and z fixed in the top, with
the same origin. To describe the position of these axes, think of the point c
at which the axis of the top meets a large sphere centered at the origin O.
Let () be the latitude (measured from the north pole) and cp the longitude
of this point. These two angles () and cp suffice to describe the positions of
the top axis; we need a third angle 'It to specify how much it has rotated
about this axis. Indeed, the top can be brought from an original upright
position to its final position in the following three successive steps (Figure
I):
First rotate the top by the longitude cp about the vertical axis OZo in
space;
Next rotate the top by the latitude () about the top's new Ox axis;
Finally rotate the top about its new Oz axis by an angle 'It.
These three angles

o < cp < 27T, o < () < 7T and 0 < 'It < 27T

suffice to describe completely the position of the top-or the rotation of

any rigid body about a fixed point. They are called Euler's angles. They
are the local coordinates in a configuration space C-but they are not glo-
bal coordinates. (For example, when () = 0 or 7T, the coordinate 'It is not
needed!) This case illustrates well the idea that mechanics often requires
that rectangular coordinates be replaced by other coordinates suited to the
II. The Spinning Top 297

problems at hand-in the present case, coordinates appropriate to the

group of all rotations in 3-space. (This group is also a three dimensional
manifold; it is an example of a Lie group.)
In order to formulate the Hamiltonian for this problem, we need ele-
mentary facts about the kinetic energy of a rotating body. If a solid flat
plate (of any shape) rotates about an axis perpendicular to the plate the
only coordinate needed is the angle 0 of rotation and the corresponding
angular velocity W = dO/dt. A point in the plate with a mass m has coor-
dinates x = r cos 0, y = r sin 0 and therefore kinetic energy

2T = m(x 2 + /) = mr2(sin20 + cOS 20)w2 = mr 2w2. (1)

The total kinetic energy of the plate is then a sum (better an integral) of
all these contributions mr2w2. The appropriate integral of the mr2 is
known as the moment of inertia I of the plate, and the kinetic energy is

T = (l12)Iw 2

(by the way, this provides another occasion for the use of integrals). In the
case of the symmetrical top there are three such moments of inertia, In Iy
and I z , about the respective axes fixed in the top. Also at anyone instant
the top has instantaneous angular velocities wx , wyand W z about these
three axes. Because of the symmetry of the top, these quantities suffice to
determine the whole kinetic energy of the top as

(2)

(without symmetry, there could be cross terms Ixywxwy, etc.) In any case,
symmetry also gives Ix = I y .
Angular velocities may be represented by vectors. Specifically, a rota-
tion about an axis A with angular velocity w may be represented by a vec-
tor along the line of A, suitably directed, and with magnitude w. This
representation is used because it is effective in combining two angular
velocities; by calculating the composite of two rotations one can prove
that the effect of combining two such angular velocities is represented by
the sum ~f ~he two. corresponding vectors. In particular, if the angular
velocities cp, 0, and 'I' for the Euler angles are represented by vectors, one
can add these three vectors by taking their components along the axes
x, y, and z fixed in the top to get (use Figure 1)
. .
Wx = cp sin 0 cos 'I' - 0 sin '1',
Wy = cp sin 0 sin 'I' + 0 cos 'I' ,

Wz = cp cos 0 + '1'.
298 IX. Mechanics

Therefore the kinetic energy T of (2) has the expression

in terms of the Euler angles. The corresponding momentum coordinates

Pi = aT/ailare thus, in an evident notation,

po = Ix O,

pcp = (/xsin20 + Izcos20)~ + Izcos i, ° (4)

Pi' =
.
I z ( cp cos °+ .
'1').

The Hamiltonian H =T + V, as a function of the p's and the q's, then

IS

H = (1/2) Ip~Ix
+ p~
Iz
+ _1 [pcp
Ix ° °]21 +
-.Pi'COS
sm
Mgl cos 0. (5)

Here V = Mgl cos 0, where M is the mass of the top and I is the distance
from the point of the top along its axis to the center of mass, while g is
the acceleration of gravity. The Hamilton-Jacobi P.D.E. then has the
form

1
Ix
[as]2 as ]2 +
ao + I;1 [ a'l' 1
Ixsin20
[as
acp -
as
a'l' cos 0
12
(6)
~ - 2[ ~ + Mgl '"" 0].
To get a complete solution of this P.D.E. for S(t,O,cp, '1'), we use two
devices. The first is "separation of variables": We try to find a solution

°
which is the sum of four separate functions of one variable (t,O,cp, '1')
each. Second, we note that of these variables only appears explicitly (as
cos 0) in the P.D.E. (6), while the others (t,cp and'l') turn up only in the
denominators of partial derivatives. Such variables are said to be "ignor-
able"; the corresponding trick is to make these partial derivatives con-
stant. Thus, all told, we try a solution of the form

(7)

involving three constants ai and a function R. It will satisfy the P.D.E. (6)
if
II. The Spinning Top 299

Thus the separation of variables has replaced the P.D.E. (6) by a first
order ordinary differential equation in three parameters ai for the function
R( 0). In this equation the denominator sin 20 on the right can be rewritten
as I - cos20. The 0 appears only in terms of u = cos 0, so it is natural to
use u as a new variable. Now this equation (8) has the general form

(dRldO)2 = F(u)/(1 - u 2)

where F(u) = F(u,al,a2,a3) is an expression independent of cp and '1'; in

fact it is a polynomial in its four arguments, cubic in u, quadratic in a2
and a3 and linear in ai, with aFlaal = -2IAI - u 2). Thus the func-
tion R is given by an integral
1/2
R = SO [ F(u)2 ] dO = _ So (F(u»:/2 du, (9)
o l-u u l-u

where we have entered the suggested change du = vi(1- u 2 )dO in the

variable of integration. This does yield a formula of sorts for S. However,
we do not need S, but only its partial derivatives. Since
CFlaal = -2IA l-u 2 ), they are

(10)

In the first equation, the integral is now that of the inverse of a square
root of a cubic polynomial in u. It is a so-called elliptic integral-not
expressible in the usual tables of integrals in terms of the elementary
functions, but much studied in classical analysis, in particular in complex
variable theory (Chapter X). Since R does not depend on cp and '1', the
matrix of second derivatives described in (10.3) is essentially

a2s
aa2 a'l'
a2s
aa3 a'l'
it thus follows that we have a complete solution in the three parameters
ai·
We are then instructed to find the trajectories (for all initial conditions)
by taking three new constants c i and solving the equations as laai = - c i
for the coordinates 0, 'I' and cp as functions of time. The first of these
equations reads

t = -c
I rO
- Ix J o ----yy;;'
du
(11)
300 IX. Mechanics

Knowing the elliptic integral, this will determine 0 in terms of time t, the
constants GI, G2 and G3 and c I; indeed -c I can be read as the initial
value of t when 0 = O. The second of these equations will then determine
<I> in terms of t, while the third gives 'I' in terms of 0, <I> and t and thus
ultimately in terms of t. The first two equations are of most interest; for
the axis of the top they provide the latitude 0 and the longitude <I> in
terms of time. A full discussion of the consequences has been presented
by Klein and Sommerfield [1965] in a famous four-volume book [1897-
1910].
One may briefly examine some of the qualitative properties of this solu-
tion. Since F( u) = ku 3 + ... is a cubic polynomial, with a positive
leading coefficient k, the values of F( u) range from - 00 to + 00; more-
over, F( u) = 0 generally has three roots, with two roots
UI and U2 (UI < U2) between -1 and + I (Figure 2). These are the only
roots of physical interest, since the substitution u = cos 0 makes

[:~ r
-1 < <u 1. Nowdtldu = IAF(u))-I!2 as above in (11) makes

= Flu)/l] ,

so the zeros of F( u) are the points where du Idt = 0 (i.e., where 0 = 0). If
one uses the positive square root of F( u) in this (ordinary) differential
equation, one gets a solution for u increasing from UI to U2 in some time
interval. Using the negative square root of F extends this solution, by
reflection, to one decreasing from U2 to UI. All told, this gives a solution
for u of the general form indicated in Figure 3. The function u (and hence
the angle 0) is periodic in t. This is the often observed situation where the
axis of the top bobs up and down in latitude 0 as the top spins (i.e., as <I>
increases ).
For the other two angular coordinates <I> and 'I' the momenta P<t> and
Pi' are given by Hamilton's equations as

dp<t> aH dpi'
dt a<l> ' dt

f(u)

u =-1

Location of turning angles

Figure 2
12. The Form of Mechanics 301

cos /I = u

Figure 3

But the Hamiltonian H does not depend on cp or on '1', so the right hand
sides here are zero. Therefore the (angular) momenta pq, and Py are con-
stants mq, and my. In other words, without friction at the point of the top
(and our formulation has indeed neglected such friction) the top, once
well started, will spin on forever.
The constant momenta mq, and my depend on the a's and the c's. How-
ever, without studying this dependence one can use these momenta to
solve the equation (4) above for the angular velocities ~ and q,. One
finds, for example, that

~ = (mq, - m y u)IIx (l-u 2 ).

This gives the speed of precession of the axis of the top around the vertical
axis: As the axis of the top bobs up and down (i.e., as cos () = u varies
between limits) the speed of precession varies according to this formula.
The magic of mechanics is that it works! Necessarily complex calculations
yield realistic results.

12. The Form of Mechanics

Our summary of theoretical mechanics serves to illustrate some aspects of
formalism in Mathematics. One starts from qualitative and then quantita-
tive descriptions of motion, such as the motion of planets and of projec-
tiles. The quantitative observations can be brought into some order by
empirical laws, such as Kepler's laws. Mathematics proper enters when
one finds out how to deduce the empirical regularities from formal
rules-in this case, from the laws of motion, the inverse square law, and
the formal rules for the calculus and for the manipulation of differential
302 IX. Mechanics

equations. Sometimes this manipulation is combined with knowledgable

approximations-neglecting terms empirically known to be negligable. But
in principle, these deductions do have a properly formal character-given
the initial data and the rules, they proceed to "mechanically" apply these
rules without further attention to the phenomena themselves-and yet the
result of this formal manipulation successfully predicts the phenomena.
While the use of these formal rules can be purely "mechanical" the rules
themselves have a conceptual background; they arise from "ideas" such as
those of rates of change and initial conditions for paths in a vector field.
These ideas, in turn, have arisen from experience. Moreover, the formal
rules and the underlying ideas expand outward into other aspects of
Mathematics, as we have noted for tangent bundles in §4 and for charac-
teristics of P.D.E. in §9.
Newton's laws are just the beginning. For many other problems of
mechanics, additional formal developments are needed. As we have
briefly illustrated, the formalism can get to be quite complicated-and this
is even more the case in important aspects of mechanics which we have
not treated, such as continuum mechanics or the motions of rigid bodies
(for rotations, use Euler's angles!) Also many of the necessary formal
developments seem at first just to be tricks-devices such as the solution of
partial differential equations by "separation of variables". This
apparently ad hoc character is emphasized by the conventional presenta-
tions, which hurry to get on with the formal developments and so often
issue in mystery or ambiguity. Thus one and the same symbol q can
represent the time rate of change of a coordinate q, or a coordinate in its
own right (on a phase space), or a function of some other quantities or
parameters. Indeed, when "parameters" appear it is often not clear
whether they are constants, variables, or functions. Moreover, the usual
use of a symbol for the partial derivative of one variable with respect to
another one often does not specify how the first "variable" is a function
and so leaves the reader uncertain as to which other variables are to be
held constant. The original authors surely know, but this knowledge is
often not well transmitted through a sequence of formulas copied from
book to book.
Nevertheless, behind every elaborate formal manipulation there is, we
claim, a conceptual background which could serve to provide a better
understanding. This we have illustrated in §7 and §8 for the derivation of
Hamilton's equations from Lagrange's equations. In this case, the rela-
tively quick formal change of variables has a deeper meaning, expressed
in the Legendre transformation from the tangent bundle to the cotangent
bundle. The exposition of this meaning, though it takes more space, does
provide a more secure understanding as well as a remarkable connection
to geometric concepts arising in other parts of Mathematics.
For the development of the Hamilton-Jacobi equation we have fol-
lowed more the quick formal exposition, with little attention to the under-
13. Quantum Mechanics 303

lying ideas. They are there and they are well worth the space they would
require. We did note briefly that the use of partial differential equations is
suggested by an analogy with optics, where the light rays are also grouped
into wave fronts.
We have introduced the phase space coordinates p's and q's without
much minding their meaning. This meaning centers on the understanding
of the formal term L pJ/ which cropped up in the definition of the Ham-
iltonian. This term turns out to represent a differential form () = L Pidqi;
a form such as this is present on any even-dimensional manifold which,
like the phase space, is a cotangent bundle-and one can give a
coordinate-free description of this form. For subsequent purposes, it is
better to use the exterior derivative w = d() = L dPidl of this I-form ().
This w is a 2-form (a differential form of degree 2) with a number of use-
ful properties. A manifold with such a form is called a symplectic mani-
fold. Much of Hamiltonian mechanics can be developed most clearly on
such a manifold, with a greater freedom in the choice of coordinates and
with considerable use of certain Poisson brackets of functions defined on
the manifold. This "simplectic" approach frees mechanics from its
apparent dependence on particular coordinates p and q for momenta and
position and explains a classical process of "canonical transformation" by
which one may replace the original p's and q's by new coordinates P and
Q which no longer need represent momenta and position, but which may
be better suited to the problem at hand. This in turn is connected to the
study of differential forms and of Lie groups acting on manifolds. Here a
Lie group means a "continuous" group-such as the group of all rotations
about a point in 3-space, as parametrized, say, by those three Euler
angles. Formally, a Lie group G is a set which is both a group and a man-
ifold, in such a way that the group operations (product and inverse) are
smooth. These operations are therefore differentiable, and this leads from
each Lie group to its associated Lie algebra. This is but one of the many
connections of mechanics with "abstract" mathematics; it illustrates the
way in which abstraction is tied into application. It is to be regretted that
there is no really satisfactory modern exposition of the conceptual
development of the ideas of classical mechanics. A splendid traditional
presentation is given in L.A. Pars [1965].

13. Quantum Mechanics

The Hamiltonian H and the associated Hamilton-Jacobi equation play an
essential role in the formulation of quantum mechanics; we summarize
this role, but very briefly. The essential observation is that certain micro-
scopic systems, such as the hydrogen atom, have discrete characteristic
energy levels represented by the discrete frequencies at which radiation is
304 IX. Mechanics

emitted. These energy levels are explained as the eigenvalues of suitable

energy operators (like H) on an appropriate linear space. Such a space W
may be obtained from the classical configuration space C with coordinates
say q I, q2, q3 by taking W to be the space of all Lebesgue square-
integrable complex valued functions 'I' on C; this W becomes a vector
space (over the complex numbers) with an inner product

(1)

The one-dimensional subspaces of Ware called "pure" states of the sys-

tem, while the observables (such as position or energy) are interpreted as
self-adjoint operators on W.
Now consider the Hamiltonian H for the classical configuration space,
C, expressed as a function of the (rectangular) coordinates qi and the
corresponding momenta Pi' Next introduce Planck's constant h; then in
the function H replace each Pi by the differential operators (h!2'7Ti)a/aqi
and each coordinate qi by the operator "multiply by qi". This is to mean
that functions of q's are to be replaced by the operator "multiply 'I' by
that functions". This process turns the function H into an operator-
perhaps not everywhere defined-on the Hilbert space W. The general
plan is that the eigenvalues of this operator will be the energy levels of
the system.
For example, for an electron of mass m and charge e, subject to the
usual inverse square law of attraction by a point at the origin, the classical
Hamiltonian has the form

By the above convention, this turns into an operator which sends each
function 'I' into

An eigenvector of this operator is then a function 'I' with an eigenvalue E

such that

It turns out that these eigenvalues agree with the observed values of the
spectrum of hydrogen.
This equation is clearly connected to the classical wave equation; this is
one of the reasons that some early versions of quantum mechanics were
known as wave mechanics. The relationship is actually deeper, and is
13. Quantum Mechanics 305

subtly connected with the classical wave equations at the origin of the
Hamilton-Jacobi theorem. Indeed, if one replaces the function 'I' in (4)
by

(for S as in the principal function above), substitutes and takes the limit
as h ->0, the result is essentially the Hamilton-Jacobi partial differential
equation (try it with just one space coordinate). This hints at the precise
sense in which quantum mechanics has classical mechanics as a limit (as
h ->0).
For more details on this very summary sketch refer to Mackey [1978].
306 IX. Mechanics

In terconnections Mathematics - Mechanics

![
Planets Projectiles Oscillations Rotations Spinning Top

K~VSY-A"L"==-----=-
wj" R.," ffCh"" Mom,," / Mom"rum

Energy Calculus Euler's Angles

I
Newton's Laws Angular Momenta - - Conserve Momentum

I
Conserve Energy Diff. Equations - Vector Fields Coordinates

~ ~if"~
~Y'''"Eq"",~ / ~
I \ ~J, ~""dk" ".me,", Roodk"

C,'"',, of V'ri'ki~H,mi"""" Eq"""""' \

T \ / s~p""w M,mf,id,

Light Hamilton-Jacobi

RL / ~ /""'i"ce
PDE Quantum Mechanics - - - - - - - Electrons

ch=,,,i, / ~'" ~ Eigenvalues

/'
Matrices - - - - - - - - Linear Algebra
CHAPTER X

Complex Analysis and

Topology

Since the square of a non-zero real number is always positive, there can
be no real square root of - 1. Inventing such a square root i and adjoin-
ing it to the real numbers, as in §IV.lO, leads to extensive and important
developments. On the one hand, the resulting complex numbers x + iy
represent well the properties of the Euclidean x - y plane and derive part
of their "reality" from the geometric reality of the plane. On the other
hand, well behaved functions f of such a complex number z = x + iY
are those functions f which have a complex derivative, and the properties
of these functions are truly remarkable. The resulting study of "complex
variables", that is, of differentiable functions of a complex number z,
leads to deep mathematical theorems with unexpected practical connec-
tions, for example to electrostatic potential and to the steady flow of fluids
as well as to aerodynamics. This chapter will introduce these concepts of
differentiation and the corresponding integrals and will indicate some of
these connections, all with a view to seeing how the apparently simple
algebraic device of inventing a "number" i with i 2 = -1 has both
geometric and analytic consequences-all a striking instance of the
remarkable interconnections of formal ideas. Just as many geometric ideas
first become apparent in plane geometry (Chapter III), so it is that basic
aspects of differentiation and integration are best exemplified by the com-
plex numbers represented in the plane.

1. Functions of a Complex Variable

All the usual elementary functions of real numbers x work equally well
(or even better) when x is replaced by a complex number (a "complex
variable" z). For example, since the complex numbers form a field, any
real polynomial function is also defined for complex arguments z; indeed,
for any complex numbers ao , ... ,an as coefficients one has the polyno-
mial function
308 X. Complex Analysis and Topology

(1)

Geometrically, it is a mapping sending each point of the complex z-plane

into a point w = f( z) in the same plane (or sometimes, into the point w
in a second complex plane). Just as graphs represent functions of a real
variable, so such maps of planes visualize functions of a complex variable
z in a geometric way. For example, w = z + c translates the z-plane by
the vector c, w = cz with I c I = I rotates the plane through the angle
arg c, and w = z2 sends each point into another with distance from the
origin squared and argument arg z doubled.
Rational functions of z formed from polynomials f and g also yield
mappings

w = f(z)/g(z) (2)

defined for all complex z except for the (finitely many) zeros of the
denominator g. For example, if a,b,c, and d are complex constants with
ad - be =1= 0, then
w = (az + b)/(cz + d) (3)

is a fractional linear transformation of the z-plane to itself, defined, when

c =1= 0, at every point except for the zero z = -die of the denominator. It
is also tempting to say that it sends this point to 00 and sends 00 to a/c.
(Picture this on a sphere by stereographic projection!)
The exponential function also works for complex arguments. Recall that
remarkable real number e and the corresponding function eX character-
ized by e l = e, eO = 1 and de x /dx = eX. This function also turns addi-
tion into multiplication, according to the formula

(4)

In fact, this is why the natural logarithm, 10&, defined as the inverse of
the exponential function, turns multiplication into addition, as in the use
of logarithms for the calculation of products. This formula (4) suggests
that a complex exponential should have eZ = ex+iy = eXe iy . Since we
already know the real exponential eX, this leads us to find the imaginary
exponential eiy ; in other words, to find a complex-valued function of a
real argument y which turns addition in y into multiplication. But this is
precisely what the two addition formulas for cos y and sin y will yield, as
m

(cos y + i sin y)(cos u + i sin u) = cos(y+u) + i sin(y+u).

Therefore we are led to define the complex exponential by

(5)
I. Functions of a Complex Variable 309

This has the desired property that eZe w = e Z + w; it turns out to have all
the other desirable properties; in particular, eZ is its own (complex)
derivative.
By this definition, e Z is never zero. (How could it be; were eZ = 0, then
e + w = Oew = 0 would be always zero!). Also Z I--> e Z maps the infinite
Z

horizontal strip {y I 0 <y < 2'lT} of width 2'lT in the z-plane onto the
whole of the w-plane, omitting only the point w = O. Each horizontal line
(y = const.) becomes a ray (fJ = const.) from the origin in the w-plane;
each vertical interval (x = const.) becomes a circle. All told, each point
w =1= 0 is covered infinitely often (once from each horizontal strip of width
2'lT) by points from the z-plane. Therefore the logarithm function, defined
as the inverse to the exponential, must be "many-valued". Explicitly, write
w =1= 0 in the polar form w = s( cbs 'P + i sin 'P) and observe that the
positive real number s has a real logarithm log., s, while cos 'P and sin 'P
have period 2'lT. Therefore, for any integer k the definition (5) gives

elog., s + i( If' +2kI'T ) = s( COS en

T
+ I. sm
. en)
T = w.
In other words, for each choice of k and for w =1= 0 the (single valued)
function of w,

log, w = log., s + i( 'P + 2k'lT), o < 'P < 2'lT (6)

has elog.,w = w, so is an inverse of the exponential. It is called (in an old-

fashioned terminology) a branch of the "many-valued" function log.,.
Under our formal definition of a function, there is of course no such thing
as a "many-valued" function. However, with Riemann surfaces one can
(geometrically!) paste all these branches together, much as in §VIII.4, so
as to get an honest single-valued function defined not on the plane but on
a Riemann surface.
The definition (5) of the exponential allows us to write the polar form
of any complex number z =1= 0 as

r = I z I, fJ = arg z. (7)

This formula includes the remarkable special case I = e 2I'Ti -a formula

which is really just a consequence of the fact that cos fJ and sin fJ have
period 2'lT! Using the logarithm, one can define a "many-valued"
• Z Z log.,c
exponentIal to any complex base c =1= 0 as c = e
Next, how can one define sin z and cos z for complex z?
Start by noting that the definition (5) of the complex exponential for
x = 0 gives

ed) = cos fJ + i sin fJ, e-;() = cos fJ - i sin fJ (8)

and hence
310 X. Complex Analysis and Topology

But here the right-hand sides have meaning when 0 = z is taken to be

complex; this result we adopt as the definition of cos z and sin z, so that

(9)

By their derivation, the functions so defined agree with the usual ones
when z is real. Moreover, one may verify that cos z and sin z so con-
structed have all the basic properties of the real cosine and sine functions
(period 2'17, addition formulas, cos 2 + sin 2 = 1, etc). We shall presently
see, using the Taylor series, that these are the only definitions which
extend the real functions sin x and cos x to "good" functions of a com-
plex argument.
It remains to explicate which functions of a complex variable are
"good" -and to show that such "good" functions have convergent Taylor
series. Indeed, the familiar Maclaurin series for eX

will also work when x is a complex number z. The new definition fits!
For the moment, we do have the remarkable observation that the
geometric representation of complex numbers and the basic facts of trig-
onometry combine to replace the real variable x by a complex variable in
the functions eX, cos x, sin x, and log x. Analysis, geometry, and trig-
onometry have intertwined in a fashion which becomes clear only when
we use complex numbers.
There are many other interesting functions of complex z. For example,
the infinite series ~ 1In is known to diverge. However, if s = (J + it is a
complex number with real part (J > 1, then the related series

n
s
= eslog. n, log" n real (10)

does converge, and so defines a function in the half-plane (J > 1. It is the

Riemann zeta function, famous for its use in number theory.

2. Pathological Functions
Formally, a function f on the reals is defined (Chapter V) to be any set F
of ordered pairs (x,y) of real numbers containing exactly one pair (x,y)
for each real x. This definition does provide a precise way of saying that
"y depends on x" without using any vague notion of "dependence". How-
ever, it allows for many more functions than might have been intended-
including some which are truly bizarre. For example, the definition
2. Pathological Functions 311

f(x) = 0 when x is ration at

=1 when x is irrational

does produce a legal function f. But this function f jumps incessantly

from value 0 to value 1 and back. One cannot "draw" a graph for it; it is
nowhere continuous. Next, the definition

g(x) = x sin 2rr/x when x =1= 0,

=0 when x = 0

also yields a function g which oscillates faster and faster as x approaches

zero. (Draw a graph!) This function is continuous everywhere, but has no
derivative at x = O. A more elaborate construction yields the function
00

h(x) = ~ 4- nsin 110nx

n =I
which has no first derivative at any point of R. Such nowhere
differentiable continuous functions do seem to violate natural intuitions!
In the calculus, Taylor's formula (§VI.8) suggests that "good" functions
ought all to be expressible by power series. This does work for eX, sin x
and many other familiar functions-but then there are functions like

1(0) = 0, x=foO

given by an explicit formula which have all derivatives zero at the

origin-and so cannot have a Taylor series expansion in powers of x (Wil-
son [1911], p. 66 #9, p. 438 #7). There are other good functions like
1/( 1 - x) defined everywhere except at x = 1 but with an evident
power series (long division!)

I/(l-x) = 1 + x + x2 + x3 + (1)

which converges only when I x I < 1.

There are other more geometrical pathologies. For example, we have
defined a (parametrized) "curve" in the plane as a continuous function
I --+R X R defined (say) on the unit interval I. Intuitively, we think of a
curve as something "one-dimensional"-but Peano succeeded in construct-
ing a curve in the sense of this definition which covers every point in a
square-and which thus is hardly one-dimensional.
This example indicates that all these "correct" formal definitions which
make calculus work rigorously and well also reveal all manner of patho-
logical phenomena. In general, a formal definition of an intuitive idea
such as the idea of "continuity" may easily lead to counter-intuitive
results. The formal is only imperfectly linked to the intuitive!
312 X. Complex Analysis and Topology

Now the familiar formal definitions of continuity and derivative apply

directly to functions of a complex variable z. Indeed, the complex plane is
evidently a metric space; the familiar Euclidean distance between two
points z and z I is given (in complex terms) as the absolute value
I z - Z I I . As for any metric space, the complex plane is then also a
topological space in which the open sets are arbitrary unions of open discs
D( C,f). Here D( C,f) denote the disc with complex center c and positive
real radius f > 0; it consists of all complex numbers z with
I z - c I < f. With this topology (or, equivalently, with this metric) the
standard definitions describe which functions w = f( z) of a complex vari-
able z are continuous (and where). It then turns out that eZ , sin z, cos z
are all continuous, and so is each branch of log., z except when the angle
cp of (1.6) is O. In the same way, a straightforward use of the standard
definition of the first derivative will apply to functions w = f( z) of a
complex z. But here the remarkable fact is that those functions f( z) which
do have a first derivative in the standard sense avoid almost all of the
pathologies we have mentioned! The use of complex derivatives, as we
will see in this chapter, has extensive consequences.

3. Complex Derivatives
As in the calculus, the derivative of a function w = f(z) of a complex
variable z at a point Zo should be the limit

f '() l' f(zo +h) - f(zo)

Zo = h1m h (1)
.... O

as the complex number h approaches O. For this definition to make sense

at Zo both f( zo) and f( Zo + h ) must be defined, the latter for all
sufficiently small (complex) values of the increment h. This means that the
set where the function f is defined must include with each point Zo an
open disc about Zo; this set is thus an open set in the topology of the
plane. Hence we define: A function f: U ~C defined in an open set U of
the complex plane is holomorphic in U (one also says regular analytic in
U) if and only if the limit (1), for h complex, exists at each point Zo in U.
The value of this limit is then the complex derivative!' of f in U. Here
"limit" has its usual formal definition in terms of absolute values (i.e., in
terms of the metric of the plane): For each real f > 0 there is a real
~ > 0 such that for all complex h with I h I < ~ on has

f(zo +h) - f(zo)

h -!'(zo) <f. (2)

The idea behind this definition is the familiar one: the derivative !'(zo)
3. Complex Derivatives 313

measures the instantaneous "rate of change" of f for a change h in the

complex variable z. But there is more to the definition than first meets the
eye, because that change h in Zo, with I h I <~, may be real, purely
imaginary, or a combination of these. In order to see what is going on,
one may write z = x + iy and then w = fez) in terms of its real and
imaginary parts as

fez) = u(x,y) +i v(x,y), (3)

where u and v are now two real-valued functions defined for points (x,y)
in the given open set U. If now we first take the change h in z to be real,
the existence of the limit (1) means that the expression

u(xo+h,yo) - u(xo,yo) . v(xo+h,yo) - v(xo,Yo)

h + I h

must have a limit as the real h approaches O. By the definition of the real
partial derivatives of the functions u and v, this limit must be

(4)

in particular, both these partial derivatives must exist.

On the other hand, if we take the change h in z to be purely imaginary,
say as h = ik, and if we use the fact that i-I = -i, we get for !'(zo) the
limit

-I -
. au (xo,Yo) + -
av (XO,yo). (5)
ay ay
Since the desired complex derivative is given by either of the expressions
(4) or (5), these two expressions must agree. In other words, when
f = u + iv is holomorphic in an open set U, the real and imaginary
parts of f must both have partial derivatives in U which satisfy the equa-
tions

au av au (X,y)EU. (6)
ax ay , ay

These are the Cauchy-Riemann equations. They are necessary if

f = u + iv is to be holomorphic in U.
This can be put differently. If we simply take two well-defined real
valued functions u(x,y) and v(x,y) with partial derivatives, the combina-
tion u(x,y) + iv(x,y) is a function of the complex number z, but it is not
likely to be a holomorphic function. It will be so only if the partial deriva-
tives of u and v satisfy these Cauchy-Riemann equations. But one may
verify that these equations do hold for the exponential function
314 X. Complex Analysis and Topology

eZ = eX ( cos y + i sin y) and also for the functions sin z and cos z as we
have defined them. Hence these functions are indeed holomorphic (in the
whole z-plane).
These equations have further consequences, both in physics and in
geometry. It will presently appear that the existence of the first derivative
1'( z) for a holomorphic function f( z) necessarily implies the existence of
all higher complex derivatives fn l( z) as well. This in turn means that the
real and imaginary parts u and v will have continuous partial derivatives
of all orders. Hence also a2v/axay = a2v/ayax and so, differentiating
both sides of the Cauchy-Riemann equations (6), one has

(7)

everywhere in U. In other words, both u and v satisfy Laplace's equation

l1u = 0, where 11 = a2/ax2 + a2/a/ is the Laplace operator in the
plane. This equation is a two-dimensional version of the Laplace P.D.E.

°
(3) of §VI.lI. Generally, a twice-differentiable function u(x,y) with
l1u = is called a harmonic function (cf. §VI.ll).
Such functions arise in theoretical physics as expressions of both gravi-
tational and electromagnetic potentials. Consider for example the electro-
static field arising from electric charges distributed uniformly along one or
several very long cylinders perpendicular to the (x,y )-plane. The resulting
potential u is then effectively constant in the coordinate perpendicular to
this plane, so it can be considered just as a function of x and y, which
then must satisfy the Laplace equation (7). In other words, every holo-
morphic function f yields such a potential u as its real part. The level
curves u( x,y) = constant are then the "equipotential" curves, while the
imaginary part v of the holomorphic function f gives curves v( x, y) =

°
constant which represent the "lines of force". At all points z = x + iy
where f'(z) =1= these lines of force are orthogonal to the equipotentials.
For a given electrostatic potential of this type, finding the right potential
function means finding a holomorphic function of z for which the real
part u( x, y) has the desired boundary values.
Next consider the geometry, by considering a function w = f( z) holo-
morphic in an open set U as a mapping of the set U by z ~ f(z) into the
w-plane. Since f, as noted above, has all higher derivatives, this will be a
smooth mapping and so must carry smooth curves and their tangents in U
into smooth curves and their tangents in the w-plane. We assert that this
mapping, at any point Zo E U where f'(zo) =1= 0, is conformal in the sense
that it preserves angles between tangents to curves at Zo. For, consider
some smooth curve in the z-plane given in terms of a parameter t as
z = get) = x(t) + iy(t) and passing through the point Zo E U when

°
t = to, so that Zo = g(to). The first derivative has g '( t) =
x '( t) + iy '( t), so when g '(to) =1= the tangent line to the curve at Zo
3. Complex Derivatives 315

makes the angle arg(g '( to» with the x-axis. The image curve in the w-
plane is given by the composite function w = h( t) = f(g( t». By the
Chain Rule for the differentiation of composite functions (which can be
seen to hold here)

h '(t) = /,(z)g'(t).

Thus if both /'(zo) -=1= 0 and g '(to) -=1= 0, the arguments are both defined
and the argument of the product / 'g' is the sum

argh'(to) = arg/'(zo) + argg'(to).

In other words, under the mapping f the tangent line to the curve has
been rotated counterclockwise by the angle arg /,(zo). Hence, given two
parametrized curves meeting at zo, each tangent line is rotated by the
same angle arg /'(zo). Therefore the angle between the curves is
preserved, so the map z f-> f( z) is conformal, as claimed.
However, if /'(zo) = 0 then this deduction gives h '(to) = 0, so the
first derivative /' is not enough to determine the slope of the tangent line
to the image curve. At such a point Zo the mapping z f-> f( z) need not be
a conformal one. For example, the map w = z2 doubles all angles
between curves meeting at the origin!
One may also show that the stereographic projection from the
(Riemann) sphere to the complex plane is conformal, so that the
geometric viewpoint also works well on the sphere.
The preservation of angles at points where /'(zo) -=1= 0 should "really"
be viewed as a consequence of the Cauchy-Riemann equations. Indeed
the mapping z f-> f( z), when written in real coordinates as
(x,y) f-> (u(x,y),v(x,y», induces at each point (x,y) of the plane a linear
mapping of the tangent space with matrix the Jacobian

au au
ax ay
av av
ax ay

The Cauchy-Riemann equations imply that the columns of this matrix

are orthogonal-or, equivalently, that the rows are orthogonal. The matrix
itself is thus just an orthogonal matrix times a real scalar, the square root
of its determinant, which by Cauchy-Riemann again is

au av au av
-- ---
ax ay ay ax

In other words, the linear transformation induced by f on the tangent

spaces preserves angles and multiples lengths by I/'(zo) I .
316 X. Complex Analysis and Topology

This discussion indicates that the use of the idea of tangent spaces,
coming from differential geometry, is a helpful adjunct to complex vari-
able theory-even though many texts on complex variables do not make
this connection.
Holomorphic functions also apply to fluid flow. Consider the flow of a
fluid in the plane (i.e., a "laminar" flow) which is steady in the sense that
the fluid velocity at each point (x,y) is independent of time. Write the
components of this velocity as P(x,y) and Q(x,y). The component of the
velocity along the "infinitesimal" vector (dx,dy) is just the inner product
Pdx + Qdy. Hence the line integral

fcP(x,y)dx + Q(x,y)dy
around a closed curve C represents a physical quantity called the "rota-
tion" of the fluid around that curve. By the Gauss lemma (VI.10J) this
equals the integral of aQ - ap over the interior of C. We assume that
ax ay
this integral is zero for all curves C, meaning that the flow is irrotational
(intuitively, there are no vortices). This assumption implies that
aQjax = ap jay. Hence, under suitable continuity assumptions, there is a
function u(x,y) with the partial derivatives

-
au = P(x,y),
au
ax
ay = Q(x,y). (8)

Also, as in §VI.10, the flow across a curve L is just the line integral of
Qdx - Pdy along L. In particular, the total flow across a closed curve C
in the plane is

fcQ(x,y)dx - P(x,y)dy.

To assume that the fluid is incompressible is to assume that this integral is

zero for all smooth closed curves C, and hence that aQ/ay = - aPlax.
In terms of the function u of (8) above, this states that
a + -a-2u2
2u
--2 = 0,
ay ax
so that u is again a harmonic function. It might thus be regarded as the
real part of a holomorphic function w = f( z). In this way holomorphic
functions represent the steady plane irrotational flows of incompressible
fluids.
All told, we have thus some remarkable connections, all starting from
the requirement that a function z I-> f( z) of a complex variable z have a
complex derivative. This requirement leads to the Cauchy-Riemann
equations which describe maps which are conformal and which in turn
lead to harmonic functions, as they appear both in electrostatic potential
4. Complex Integration 317

and in fluid flow. The deeper ideas of Mathematics do have varied con-
nections!
So much follows from the easy derivation of the Cauchy-Riemann
equations that one might be tempted to try to get even more from the
assumption that z ~ f( z) has a complex derivative-for example, more
by differentiating along some slant lines. There is no more. A theorem of
Looman and Menchoff states that if two real valued functions u(x,y) and
v(x,y) are defined and continuous for (x,y) in an open set U and have
partial derivatives which satisfy the Cauchy-Riemann equations (6) every-
where in U, then the function f( x) = u + iY is indeed holomorphic in
U.
All this is just the beginning of the remarkable properties of holo-
morphic functions f. The definition requires just the existence of a first
derivative, but this will imply the existence (and continuity) of derivatives
of all orders, as well as the presence of suitably convergent Taylor series.
In other words, holomorphic functions of a complex variable escape all
the pathologies attending ordinary functions f of a real variable. The
proof of these remarkable results depends essentially upon the use of
integration (a process already strongly suggested by the applications we
have noted). The integral of f( z) around a closed path in the complex
plane is zero, provided that f is holomorphic in an open set containing
the path and its "interior". This Cauchy integral theorem plays a central
role in complex analysis. To formulate it, we next consider integration and
then properties of the paths of integration (§5).

4. Complex Integration
The complex integral

f h
f( z)dz = f h
f dz (1)

of a holomorphic function f along some path h in the complex plane

should be, as the notation suggests, the infinite sum of values f( z) of f at
points z on the path multiplied by "infinitesimal" increments
dz = z I - z of z along the path. This idea is turned into a precise
definition of the integral as the limit of a suitable finite sum. Let f be
holomorphic in an open set U, while h:I ..... U is a path in U, parametrized
by t, 0< <t I and smooth (i.e., h 'et) is continuous). So divide the unit
interval I at n + 1 points 0 = to < ... < tn = 1. Note that the incre-
ment in z = h(t)

to
I
o
318 X. Complex Analysis and Topology

from ti -I to ti is h( ti) - h( ti- d, choose a point Si in each interval and

evaluate the holomorphic function J at that point. This produces the sum
n
~ J(h(Si))(h(ti) - h(ti_I)). (2)
i=1

The integral (1) is then defined to be the limit of this sum as n approaches
infinity and as the maximum of the ti - ti-I approaches zero. This limit
(and hence this integral) exists if h has a continuous first derivative h'( t)
for each t in I (where the derivative at the endpoints of I is to be defined
by a one-sided limit). The choice of the parameter t for the path is not
important; all that really matters is that the path is smooth and is
"directed". This can be formulated by proving that the integral of J is
independent of the choice of parameter. A new choice would consist in
taking a smooth function k:I ~ I which is monotonic increasing (t < t'
implies k(t) < k(t')) and has k(O) = 0, k(l) = 1; then the integral of J
over the path h is identical to that over the path h·k:I ~ U. (One must
here use the fact that k, continuous on the compact set I, is uniformly
continuous.)
As we will soon see, it is also convenient to be able to integrate over the
path given by the boundary of a rectangle and over other paths with
corners. Such paths are said to be piecewise differentiable-they are
described exactly as the paths obtained by piecing together ("composing")
a finite number of differentiable (i.e., smooth) paths. For such paths h the
limit above and hence the integral exists.
Except for the condition that the function J( z) be holomorphic, this
integral is just the line integral introduced in §VI.lO and motivated there
by various physical applications (for example, to work). Explicitly, if we
write z = x + iy, dz = dx + idy, and J(z) = u(x,y) + iv(x,y) in terms of
real and imaginary parts, the integral just defined is the line integral

fJ(z) dz = f[u(x,y)dx - v(x,y)dy] + i f[v(x,y)dx + u(x,y)dy]. (3)

h h h

Going further, one can substitute the function h(t) describing the path to
reduce this integral to ordinary real integrals in t, from 0 to l. All these
formulas express the age old idea: The whole is the sum of its parts!
The size of the parts yields an evident upper bound for the size of the
integral. Specifically, the function t f-> IJ(h(t)) I is continuous on a
compact set (to wit, the interval 1), hence has a maximum value M there.
On the other hand, the path of integration has a length L h , which can be
defined geometrically as the limit of the sum of the lengths of small
inscribed chords, or equivalently as the integral fal
I h '(t) I dt. Then,
looking at the sums (2) used to define the integral (1), one may prove,
much as in (VI. 4.3), that
4. Complex Integration 319

I ~f(z)dz I < ML h · (4)

This may seem to be just a rough upper estimate of size, but it turns out
to be exceedingly handy. We will see one example of its use (in §6) but
there are many more throughout analysis, which often requires judicious
estimates of the size of all sorts of integrals.
The integral also has useful algebraic properties. It is an additive opera-
tion on holomorphic functions, in the sense that

.J: (j( z) + g(z ) )dz = .J: j( z )dz + f g( z )dz , (5)

it is also additive on paths: If one forms a "composite" path h2 ·h l -first

the path hI, then followed by the path h2, one has

(6)

The direction of a path matters; if one integrates backwards over a path h

(literally, using the path h- I given by h-I(t) = h(l - t» one changes
the signs of all of the terms h(t;) - h( ti-I) in the sum (2) and hence the
sign of the whole integral

.J:-I j(z)dz = -.J: j(z)dz. (7)

Since the complex integral can be reduced to a real integral, we can

also use the fundamental theorem of the integral calculus. Thus, for
example, if h is a path from the point Co = h(O) to the point CI = h( 1),
while j( z) = z3 then

In particular, if h is a closed path (h(O) = h(l» the result is zero: The

integral of a polynomial in z around a closed piecewise smooth path is
zero. That need not be true for other functions such as w = liz. In this
case the integral around a circle, center at the origin with z = rei(J, is

f dz -- f d(rei(J)
'(J -- f 'dO -- 2'TTl..
I (8)
Z re'

This is surely not zero. Here dz / z = d(lo~ z), while the logarithm, as we
saw, is not a single-valued function. Indeed, if we follow a branch of
log.. z continuously around that (counterclockwise) circle, it changes to the
"next" branch by exactly 2'TTi. Moreover, j(z) = liz is not defined at
z = 0; so is not a holomorphic function inside the circle.
These calculations and many like them, suggest that the integral of a
holomorphic function j around a closed path h ought to be zero when
320 X. Complex Analysis and Topology

h = aA is the boundary of some region A in which J is holomorphic, as

in Figure l. In this case the Gauss lemma of §VI.lO states that a line
integral such as f w, for w a differential, may be replaced by a double
integral

I
aA
P dx + Q dy = I I
A
[~-
ax
ap ]dXdY .
ay
(9)

In particular, the line integrals appearing in the formula (3) for f Jdz
become double integrals

f(udx-vdy)=-fI
h A
[~+~ldXdY,
ax ay
(10)

f
h
(vdx + udy) = II [~-
axA
~ldXdY,
ay

If the whole area A is contained in the open set U where J is holo-

morphic, the Cauchy-Riemann equations state that both the integrands
on the right are zero, and hence so is f J dz. This gives a first form of
Cauchy's integral theorem:

Theorem. If w = J( z) is holomorphic with a continuous first derivative

l'(z) in an open set U containing a closed piecewise differentiable path h
and all oj its "interior" area A, then the integral oj J( z) around h vanishes:

f h
J(z)dz = O. (11)

This can also be viewed as a statement that the integral is "independent

of path": If we consider two smooth paths hI and h2 both joining a point
c to a point d, and if hI can be "deformed" into h2 through an area
within which J is holomorphic, as in Figure 2, then

f hI
J(z)dz = f h2
J(z)dz. (12)

Figure 1
4. Complex Integration 321

Figure 2. Deformed path.

This form of Cauchy's theorem is not very satisfactory, on several

counts. First, the Gauss lemma involving those double integrals normally
assumes that the partial derivatives oQ jox and oP joy are continuous
everywhere in the area A; for this reason our statement of the theorem
assumes not just that f is holomorphic, but also that its first derivative f'
is continuous. By other methods, as we will see, it is possible to avoid this
added hypothesis on f. Next, the standard proof of the Gauss lemma (it
was merely indicated in (VJ.1O.5)) involves slicing the area A vertically
and perhaps horizontally; in case the contour h is at all convoluted, some
fussy adjustments are needed to make this slicing process really work-
things may not be as easy as Figure 1 may suggest. Finally, the theorem
describes A as the "interior" of the contour h. Now a smooth oval curve,
such as that of Figure (1), has an easily recognized interior, but this might
not be so with a highly convoluted curve (Figure 3) such as this outline of
a maze. More formally, a path h:! ~P is closed if h(O) = h( I) and simple
if it does not cross itself; that is, if II =1= 12 implies h( II ) =1= h( 12) except
for h(O) = h( I); then the image set h(I) is called a simple closed curve in
the plane C. The Jordan curve theorem asserts that every simple closed
curve divides the plane into two parts: The plane C with h(I) removed is
not connected, but is the union of two connected sets-an "inside" and an
"outside". This theorem is true but is by no means easy to prove. Even for
a polygonal curve it is hard. In the simplest such case, when the curve is
the boundary of a triangle, the Jordan curve theorem rests on geometric
facts such as Pasch's axiom (§III.2). The search for higher dimensional
analogs of the Jordan curve theorem has been a major motive in algebraic
topology.

Figure 3. Inside or outside?

322 X. Complex Analysis and Topology

To summarize: Cauchy's theorem that an integral SJdz of a holo-

morphic function is independent of path is a direct extension of the fun-
damental theorem of the integral calculus, and so is another representa-
tion of the idea that the whole is the sum of its parts. The proof of the
theorem by the Gauss lemma is an illuminating indication of the reasons
behind this theorem, but in any generality it runs into a morass of difficult
geometric and topological considerations. This is one of the reasons that
complex analysis leads directly to general questions about the topology of
the plane-and of other spaces. Nevertheless, we will soon show how a
different approach will give a more direct proof of the Cauchy theorem.

5. Paths in the Plane

The properties of complex integration just presented directly involve
several operations on paths in the plane: The composition of two paths
(one followed by another), the inverse of a path (the path run backwards),
and the deformation of a path, as in the independence of an integral from
the choice of path in the Cauchy integral theorem. This situation has
made it gradually clear that it is useful to consider these operations on
paths in their own right, both for the study of higher-dimensional
integrals and for the analysis of the connectivity of a space (such as a
Riemann surface). For this latter purpose, it is useful to consider all con-
tinuous paths, not just the smooth or piecewise differentiable ones.
A path in the subset S of the complex plane C is thus a continuous map
< <
h: / ..... S of the real unit interval/ = {t I 0 t I} into S. Its starting
point is the point h(O), it ends at h( I). If k is a second path starting at
h( 1), then k·h denotes the composite path: First follow h; then follow k; it
may be described formally as the path k·h:/ ..... C specified by the equa-
tions

(k·h)(t) = h(2t), o < t < 1/2,

(1)
= k(2t-I), 1/2 < t < I.
The resulting function k·h isindeed continuous on / because
h(2(l/2» = h(l) = k(O) = k(2(1!2) - 1); in other words, the two
paths fit together in the middle. As in (4.6), a line integral of a
differentiable form w along the composite path k·h will be the sum of the
f
separate integrals alon k and along h.
The inverse path h - (go backwards along h) is given by the equation

O<t<1. (2)

This is continuous, but it isn't quite an inverse for the composite defined
in (1), because the composite h - I. h is a path that goes over from h (0) to
5. Paths in the Plane 323

h( 1) and back again at the same speed; it is not the identity path, which
ought to be the path staying put at h(O). At best, one can deform h-I·h
into this identity. Also the composite k·h is not really associative; if m is a
third path starting at the end k(l) of k, then the composite m·(k·h) fol-
lows along m for 1/2 < < t 1, while the associated composite (m·k)·h
follows m only for parameter values 3/4 < <
t 1.
What is needed is a process of deforming m·(k·h) to (m·k)·h. This is
provided by the formal concept of a homotopy. If hohl:I ~S are two
paths in the subset S e e with the same endpoints ho(O) = hI (0),
ho (1) = hI (1), then a homotopy holding endpoints fixed (a continuous
deformation) from ho to hI is a continuous map H:I X I ~S with

H(t,O) = ho(t), o < t < 1, (3)

H(t,l) = hI (t), o < t < 1, (4)

H(O,s) = ho(O), H(l,s) = hl(l), (5)

In other words, this homotopy H maps the unit square with coordinates
sand t into S so as to give a "continuous family" of paths hs, each start-
ing at ho (0) and ending at ho (1). The paths are parameterized by s, start-
ing from s = 0 with the first path ho and ending for s = 1 with hI, as in
Figure 1. This figure pictures the deformation of the initial path ho into
the final one, hI. For example, one may deform the closed path given by
the circle of radius 1 about the origin into another closed path given by a
circle of radius 2 by using the homotopy H(s,t) = (1 + s)e 2'/Tit, as in Fig-
ure 2.
A homotopy from ho to hI and a subsequent homotopy from hI to h2
can clearly be "composed" to give a homotopy from ho to h2' while the
"inverse" (replace s by l-s) of a homotopy ho ~ hI from ho to hI is a
homotopy from hI back to h o. Therefore the relation "h o is homotopic to
hI" is reflexive, symmetric, and transitive. Then taking the equivalence
classes for this relation produces for each path ho the homotopy class, call
it rho j, of all paths hI homotopic to ho with the same end points. Since
homotopies ho "" hI and ko '" kl will yield a homotopy ko ·h o "" kl ·h l ,

H
.. ho(O) ~---~----

Figure 1. A homotopy.
324 x. Complex Analysis and Topology

Figure 2

one can define the composite of two homotopy classes as the class
[k)·[h) = [k·h]. This composition of classes is now associative, in view of
a homotopy m·(k·h) ~ (m·k)·h which we can picture as in Figure 3. At
each horizontal level of the deformation, the paths follow h from t = 0 to
t = 1/4 + s/4, then k, then m from t =
1/2 + s/4 to t 1. =
If we consider just the closed paths (the loops) starting and ending at
one and the same point Zo these considerations will prove the

Theorem. For any subset S of C and for any point Zo ( S, the homotopy
classes of closed paths in S from Zo to Zo form a group under composition.

This group is called the Poincare fundamental group 'lTl(S,ZO). Its iden-
tity element is the class of the trivial path, constant at Zo, while the inverse
of any homotopy class [h) of a closed path h is the class [h -I), given by h
run backwards.
This fundamental group construction assigns an algebraic object-the
group 'lTl(S)-tO a geometric object, the subset SeC. For example, if S
is the unit circle S 1 in C while Zo is the point I on that circle, the
corresponding fundamental group 'IT 1(S 1, I) turns out to be the infinite
cyclic group Z. A generator of this infinite cyclic group is the path p going
once around the circle, in a counterclockwise direction. With some care,
we can prove that any closed path h from I to 1 for the circle S 1 is homo-
topic to pn, where the integer n is the net number of times h winds around
the circle in a counterclockwise direction. (This number is also called the
degree of h.) In other words, the algebraic isomorphism 'lTl(SI) == Z
expresses a geometric fact: That a continuous map S 1 ..... S 1 is determined,

h k 3/4 m

Figure 3. Homotopy associativity.

5. Paths in the Plane 325

up to homotopy, by the net number of times the map winds the first circle
about the second one; i.e., by the "winding number". (Observe that a
closed path h:1 ~ S' really amounts to a map S' ~ S', where the first circle
S' is obtained by identifying the ends of the unit interval I.)
If C - {OJ is the entire complex plane with only the origin 0 deleted,
its fundamental group 'IT,(C - {OJ, I) is still the infinite cyclic group.
Indeed, an isomorphism 'IT, (C - {O}, I ) ~ 'IT, (S', I) can be constructed by
taking any closed path h in C which misses the origin and deforming it by
radial projection from the origin into a path running along the unit circle
S'. This same fundamental group also appears in the behavior of the
many-valued function lo~ z = lo~ r + «() + 2'ITk )i, which is defined
for all z = re iO ; i.e., defined in the set C - {OJ. Specifically, starting
with one branch k of this logarithm and following a closed path with
winding number n brings one to the branch k + n.
On the other hand, consider the set B in C consisting of two circles
tangent at one point (say the point I). On B there are two closed paths h
and k running once smoothly in a counterclockwise direction around each
of these two circles (see Figure 4); they give elements x = [h 1 and
y = [k 1in the fundamental group of B. A plausible process will deform
any closed path in B into a path running at a unform rate first around one
circle and again around either the same circle or the other circle and so
on. In this way any element in the fundamental group can be written as a
composite xml/lxm2/2 ... of products of integral powers of x and y.
Any group containing two elements x and y must contain all such pro-
ducts; in the present case, two such products are equal only when they
become equal by formal cancellations; for this reason this fundamental
group is known as the free group on two generators x and y. The complex
plane C - {O,I} with two points deleted has the same fundamental
group, generated by paths x and y going once counterclockwise around
or I, respectively. Then for a function f( z) such as 1/z( z - I) holo-
°
morphic in the set C - {a, I}, the Cauchy integral theorem (in one form
to be discussed) will say for two closed paths hand k that

1: I( z )dz = 1: I( z )dz (6)

- y

Figure 4
326 X. Complex Analysis and Topology

whenever hand k determine the same element in this fundamental group.

In differential geometry (Chapter VIII) smooth functions and smooth
paths at a point give rise to tangent and cotangent vectors at that point,
the two being paired by the combination "directional derivative of the
function along the path at that point". What we now have is somewhat
analogous: holomorphic functions f (or integrands) and paths h in a set
See, paired by the integral fh f dz.
In defining the fundamental group '1T1(S,ZO) we have used only the
closed paths starting and ending at some arbitarily chosen base point Zo in
S. As long as S is pathwise connected, however, this homotopy group is
independent of the choice of the base point Zo, up to an isomorphism.
This can be seen most readily by considering the algebraic system formed
by the homotopy classes [h] of all paths in S, closed or not. Under the
operation of composition these classes do not form a group, because the
composite class [k]· [h] is not always defined. However it is defined when-
ever k starts where h ends, and the triple composite [m]·([k]·[h]) is asso-
ciative whenever it is defined. Moreover, each class [h] has a two sided
inverse. With these properties, the homotopy classes under composition
form an algebraic system called a groupoid, which we will meet again
(under more general auspices) in the next chapter. Using this algebraic
system, one can readily produce the desired isomorphisms
'1T1(S,ZO) == '1T1(S,ZI) for any path-connected set S. Just choose some
path k starting at the first base point Zo and ending at the second ZI. Thus
for any closed path h at Zo the correspondence

(7)

does give the desired isomorphism, by the algebraic properties of

groupoids. Indeed this correspondence (7) is just conjugation in the
groupoid playing much the same role as conjugation in other geometric
examples of transformation groups (§V.6), where conjugation by k carries
the subgroup fixing a point x to the subgroup fixing the point kx.
A holomorphic function f(z) is defined in some open set U of C, usu-
ally in a connected open set. Since U is open, the requirement that it be
connected (not the union of two disjoint non-empty open subsets) is
equivalent to the requirement that it be path connected. To make the
integral ff(z)dz independent of path in U, by Cauchy's theorem, one
may require that U be simply connected in the sense that its fundamental
group reduces to the identity element l. This amounts to the requirement
that every closed path in U is homotopic (with ends fixed) to the constant
path or equivalently that every continuous map h:SI-,> U of the unit circle
S I into U can be extended to a continuous map H: {z I I z I < l} -'> U of
the unit disc into U: Any such extension clearly produces a homotopy
from h to the constant path, as in Figure 5.
5. Paths in the Plane 327

Figure 5

The whole complex plane C is simply connected. This result, though it

may seem related to the Jordan curve theorem, is much easier to prove:
Given a closed path in C, simply deform the path by that motion which
takes each point at a constant velocity on a straight line to the base point.
Now observe that our definition of the fundamental group 'lTI(S,ZO) of a
subset S of the complex plane C really used nothing about the complex
numbers except that they form a topological space under the standard dis-
tance; this implies at once that any subset S is also a topological space
with the same metric (or, equivalently, with all the intersections S n U as
open sets, for any U open in C). Therefore, the same definitions will pro-
vide a fundamental group 'lTI(X,XO) for any topological space X, which
will be, for example, independent of the basepoint Xo as soon as the space
X is path-connected. For instance, such groups can be defined for sur-
faces; in particular for Riemann surfaces, and this is one of the reasons
they are useful in complex analysis. For example, the fundamental group
of the torus SI X SI (an example of a Riemann surface) is just the group
Z X Z-the free abelian group on two generators-one path each around
each of the circles S I in S I X S I. Fundamental groups are also valuable
for all other manifolds and other spaces. (For example, the fundamental
group of the projective plane described in §VIII.4 is the cyclic group of
order 2).
The fundamental group is just a first step in algebraic topology-the
analysis by algebra of connectivity and homotopy properties of geometric
figures which arise in topological spaces. In this section, our attention has
been diverted from complex analysis to topology because complex
analysis was historically one of the sources of algebraic topology. Another
important source of topology was the qualitative study of trajectories in
mechanics. This diversion will thus supplement our too-abrupt introduc-
tion in Chapter I of the notion of a topological space. In every such space
X, it is convenient to call a set C closed when its complement X - C (all
points of X not in C) is open. Closed sets (e.g. circles) abound in the
complex plane. The axioms for a topological space X, stated in §I.10 in
terms of open sets, may be equally well formulated just in terms of closed
328 X. Complex Analysis and Topology

sets; in particular, the intersection of any family of closed sets is closed.

This makes it possible to define for each subset A of a space X a closure .if
(the smallest closed set containing A) as the intersection of all the closed
sets of X which contain A. Then .if is closed. The axioms for a topological
space can also be stated in terms of the unary operations A f-> A, "take
the closure", on all subsets of X. In any topological space X it is also con-
venient to say that a neighborhood of a point x ( X is any open set U of X
which contains x. Then continuity of a function f:X ~ Y can be described
naturally in terms of neighborhoods: f is continuous at x ( X if to every
neighborhood N of f(y) there is a neighborhood M of x with
f(M) eN-note that we do not have to say "arbitrarily small" neighbor-
hood; the requirement for every N suffices!
Thus "topological space" may be defined equally well by way of open
sets, closed sets, closure, or neighborhoods. This again illustrates the gen-
eral thesis that one idea has many formalizations. In this case, the formal-
izations provide a handy language for the study of the many remarkable
and deep properties of topological spaces, in particular of the topology of
the plane.

6. The Cauchy Theorem

We return to the problem of a proof of the Cauchy integral theorem.
A primitive of a function f holomorphic in an open set U of C is
defined much as in the calculus: it is a function F holomorphic in U with
first derivative F' = fin U. When there is such a primitive F, the Cauchy
integral theorem does hold for the integral of f around any closed piece-
wise differentiable path h in U:

~ f(z)dz = O. (1)

Indeed, introducing the real parameter t for the path h and using the
chain rule for differentiation gives

J:f(z)dz = fol f(h(t))h'(t)dt = fol F'(h(t))h'(t)dt

= J(Io ~F(h(t))dt
dt
= F(h(l)) - F(h(O)) = 0,
where zero results because the path is closed (h( I) = h( 0)).
In particular, zn has a primitive zn+ I /(n + I) for each positive n, so the
Cauchy theorem does hold when f(z) is a polynomial. However, this
result still does not tell us how to construct a primitive for more general
holomorphic functions f. In order to do this, we will consider first very
special closed curves: The boundaries of rectangles, taken counterclock-
6. The Cauchy Theorem 329

wise. This will altogether avoid the problems presented by convoluted

boundary curves.

The Cauchy-Goursat Theorem. Let R be a rectangle in C with sides paral-

lel to the real or the imaginary axes, for which the interior of R and the
boundary aR are both contained in the open set U where f is holomorphic.
Then

f aR f(z)dz = O. (2)

The proof will illustrate the use of (- 8 methods in approximating

integrals and will also introduce certain algebraic classes of boundaries
which later develop into homology, a part of algebraic topology. First
record the perimeter p and the diagonal diameter d

p = perimeter R, d = diagonal R (3)

of the rectangle. If the intended integral is not zero, contemplate its abso-
lute value
a = I f aR f( z )dz I . (4)

We intend to successively subdivide till we find a very small rectangle

where the corresponding integral is too big (for f to be holomorphic
there). Specifically, divide the given rectangle into four smaller congruent
rectangles, as in Figure 1. Now integrate over the boundaries of all four
smaller rectangles, always in a counterclockwise direction; then the
integrals over the "inside" boundaries cancel, as shown on the right of the
figure. Therefore, by the additivity of the integral under composition of
paths, the integral of f( z) over aR is the sum of the four integrals over the
boundaries of the four smaller rectangles. There must therefore be at least
one of these four rectangles, call it R J , with an integral of size A /4 or
more

(5)

Pick such an RJ, subdivide it into four subrectangles, choose an offending

one of these as before, and continue indefinitely. The result is a nested
sequence of rectangles

l~~II .~. I
Figure 1
330 X. Complex Analysis and Topology

with

IfaRk J(z)dz I > "A,

4
1 .
penmeter
R
k ~p,
2
diameter Rk = ~d. (6)
2
There is exactly one point c in the complex plane C common to all these
rectangles (which fonn a nested sequence of closed sets); for example, one
can describe the x coordinate of this point c as the least upper bound of
the x-coordinates of the lower left corners of the successive rectangles R k .
Since J is holomorphic in U, it has a first derivative /'( c) at the point c.
This means that for each positive l} > 0 there is a neighborhood of c in
which

J( z) - J( c) _ / '( c ) < l}.

z - c

Since the size of the rectangles approaches zero with k, there is an index k
such that all of the rectangle Rk is in the neighborhood; thus on aRk we
have the estimate
J(z) - J(c) - f'(c)(z - c) = t:(z), I t:( z) I < l} Iz - c I . (7)
Therefore the integral is

f aRk
J(z)dz = f aRk
J(c)dz + f aRk /,(c)(z - c)dz + f aRk t:(z)dz. (8)

On the right, the first two integrals are integrals of polynomials around
closed curves, hence are zero. The remaining integral of t:( z) can then be
bounded by the general formula (4.4) for the size of an integral. By (7),
<
I t:(z) I 8 diameter Rb so with (6) we have
~A
4
< 8 (diameter Rd(perimeter R k ) = 8 ~pd.
4
In other words, A <
l} pd. Since all of this holds for each positive 8, one
must have A = O. This proves the theorem.
The point of this proof is that the nested rectangles close down on one
point c. At this point one can then use the basic idea of the differential
calculus-that the first derivative gives a linear approximation to the func-
tion (here a complex linear approximation (7». For this reason the proof
works assuming just the existence cf the derivative /'(z) and not its con-
tinuity (as was needed in §4 in the argument by the Gauss lemma).
The proof also illustrates several central ideas of topology. Thus the
existence of that point c comes from the fact that in the plane any nested
sequence Cl ::J C2 ::J C3 ::J ... of closed sets has a non-trivial
6. The Cauchy Theorem 331

intersection. This basic fact, about the point-set topology of the plane, is a
statement that the points are all there (the plane is complete); it is com-
parable to the Dedekind cut axiom for the reals.
On the other hand, we have integrated over a sum
hI + h2 + h3 + h4 of the four boundaries hi = aR i of the four smaller
rectangles of Figure 1. Such sums form an abelian group, the group of
"chains"; similarly the Figure 1 suggests a 2-dimensional "chain"
RI + R2 + R3 + R4 • Since hI is the boundary hI = aRI one says that
hI is "homologous" to zero. Also we have subdivided each edge e (of the
rectangle) into two pieces e' and e"; one says that there is a "singular"
homology e ,...., e' + e" because e - e' - e" is a boundary; indeed, a
triangle can be collapsed (vertically in Figure 2) so that its boundary is
e - e' - e". This is the start of an idea of "singular" homology and of
simplicial (Le., triangular) approximation. These ideas of homology and
chains are effective in all dimensions in handling geometric problems of
algebraic topology.
We return to exploit the Cauchy-Goursat theorem. First let f( z) be a
function holomorphic in an open disc

D(c,r) = {z I Iz - c I < r} (9)

with radius r and center the complex number c. Then the Cauchy theorem

f"f(z)dz = 0 (10)

holds for any closed piecewise differentiable path h in the disc (9),
because we can construct a primitive F for f as the integral, for W f D:

F(w) = fWf(z)dz
c

taken along a rectilinear axis-parallel path in the disk from c to w (see

Figure 3). Indeed, by the Cauchy-Goursat theorem, any other axis-
parallel rectilinear path from c to w will give the same integral F( w); from
this one may readily show that F( w) is a primitive of f, so that (1)
applies.

e' e e"

Figure 2
332 X. Complex Analysis and Topology

LJw c

Figure 3

The result is actually more general, since the same argument evidently
works if the disc is replaced by an open set which is convex (i.e., which
contains the line segment joining any two of its points). Indeed, the same
argument can be made to work if the disc D is replaced by any simply
connected and connected open set U. One again wishes to compare two
rectilinear, axis parallel paths in U from c to w, much as in Figure 3.
Since U is simply connected, these paths are homotopic; the trick is to use
suitable approximations to make the homotopy take place through rec-
tangular changes in path.
Another, explicitly homotopy-theoretic version of the Cauchy theorem
asserts that the integrals

1: J(z)dz = 1: J(z)dz
over the paths hand k with the same endpoints are equal when h is
homotopic to k (fixing the endpoints) in the open set U where J is holo-
morphic. There is still another, more subtle version (harder to prove): If h
is a simple closed piecewise differentiable curve in C, while J is holo-
morphic in the interior of k and continuous on h, then 1:J( z )dz = O.
All of these somewhat different formal statements are realizations of the
one idea underlying the Cauchy integral theorem. The care required for
the proof is one of the reasons why complex analysis has served as an
excellent training ground for mathematical rigor. Indeed, the proofs can
all be made rigorous (and most of them perspicuous as well). Historically,
this developed gradually, with Cauchy's foundations for the calculus,
Riemann's emphasis on geometrical insights, and the careful (- 8
methods of Weierstrass in complex analysis, calculus of variations, and
elsewhere. The better understanding and analysis of these methods has in
its turn led to extensive developments in topology.
Though this book does not present many complete proofs, we
emphasize that there is revealed here an objective notion of a rigorous
proof: Each step in the proof is a logical consequence of prior steps or
theorems or axioms (for the real numbers, or for sets, as the case may be);
7. Uniform Convergence 333

moreover, consequence means consequence according to the explicit rules

of inference from mathematical logic (Chapter XI). To be sure, practicing
mathematicians often learn these rules just 'intuitively'-but one should
not take too seriously the sometime response of a student of a noted and
careful professor W, a specialist in complex variables. Asked to tell when
a proof is rigorous, the student responded: "When it is acceptable to Pro-
fessor W".

7. Uniform Convergence
The use of infinite series is the quintessence of the notion of infinitary
Mathematical operations. Thus

(1)

(2)

while for any real r < 1 an "infinite" long division yields the geometric
series which converges to 1I( l-r),

- -1= l + r + r 2 +r 3 +r4 + .... (3)

1- r
To be sure, sometimes an infinite sum gets out of hand, as in the case of
the geometric series (3) for r = 1 or r > 1 and also with the harmonic
series
(4)

Indeed, adding up this series in groups of successive lengths 1,1,2,4, ... ,

one finds that the sum grows faster than 1, P/2, 2, 2'/2, .... It is thus a
divergent series (but it does suggest the Riemann zeta function; see (11.7)
below).
Now, in point of fact, there is no such thing as actually summing an
infinite series; nobody has every actually added up an infinite number of
terms, no matter how conveniently convergent. Here as elsewhere in
Mathematics, the use of the infinite is an intuitively suggestive description
of what is in fact a finite formal process of successive approximations.
This process can be described, in a completely rigorous way, in set-
theoretic and logical terms. For example, an "infinite series"

(5)

of real or complex numbers is determined by giving a function n ~ an on

the set N of natural numbers to the reals R (or to C). This function also
determines the finite partial sums
334 X. Complex Analysis and Topology

Sn = ao + a\ + ... + an,
one for each n. The series (5) is then said to converge (§IV.4) to a number
S as sum when the sequence Sn has a limit lim Sn = s. As in §IV.4, this
n ... oo
statement about a limit means that for each real t: > 0 there exists a
natural number N such that, for all n > N, I Sn - S I < t:; here the
absolute value is that of the real numbers or the complex numbers, as the
case may be. As in other cases, the idea that the limit S is something
approximated (better and better as N increases) by finite sums gets
expressed formally in terms of the logical quantifiers "for all t:" and "there
exists N". In other words, convergence involves repeated finite approxi-
mations (up to t:) to the non-existent infinite sum.
The series (5) is said to converge absolutely when the corresponding
series of absolute values,

I ao I + I a\ I + I a2 I + ...
is convergent. Then absolute convergence implies convergence, as one sees
by applying the triangle law I Z\ + Z\ I < I Z\ I + I Z2 I for absolute
values to show that the sequence Sn satisfies the Cauchy condition for con-
vergence. On the other hand, there are many convergent series such as the
alternating series

which are not absolutely convergent.

In the geometric series (3) the successive terms are functions of a vari-
able r (there r is real, but the same series converges when r = z is a com-
plex number with I z I < 1). More generally, consider a series of func-
tions !n of one or several parameters (here written z and w) in the
"infinite series" form

!o(z,w) + !\(z,w) + fz(z,w) + (6)

Objectively, this means that we are given some function

(n,z,w) f-> !n(z,w) to (say) the complex numbers from the product
N X D X E, where D and E are sets of complex numbers, with z t: D
and w t: E. The partial sums Sn (z, w) of such a series are then also func-
tions s:N X D X E -->C, and the series may converge for some z and w
and diverge for other values. Now in such a series the "speed" of the con-
vergence may very much depend on the values of the parameters z and w.
For example, in the geometric series (3) the partial sums are
Sn = (1 - rn+\)/(l-r); they converse to 1/(1-r) more and more
slowly as r approaches 1. More explicitly, given a level t: > 0 of approxi-
mation desired, one must choose an N which depends on r (and which
7. Uniform Convergence 335

necessarily gets larger as r approaches the number 1) to get

I Sn - 1/(1-r) I = I r n + 1/(I-r) I less than £ for n ~ N; one cannot
to a given £ choose anyone N which works for all r. One says that N can-
not be chosen "uniformly" for all r < 1.
This suggests that we describe formally when convergence is uniform;
this notion will tum out to matter. The definition is straightforward: The
series (6) of functions converges uniformly on D X E to a function
s:D X E ~C when to each £ > 0 there is some N £ N such that, for all
n ~ N and all <z,w> £ D X E, one has
I sn(z,w) - s(z,w) I < £. (7)

This is just the formal assertion that the same N works for all z and w at
issue. If we put it in logistic notation, where V is "for all" and :=1 stands
for "there exists", it reads

y £ > 0 :3 N £ N Y n ~ N Y <z,w> £ D X E (7) holds.

Here again what is crucial is the careful use of the logistic quantifiers Y
and :3, written in the right order (!). For later purposes (in the discus-
sion of quantifiers in the next chapter) we observe that in each quantifier
the variable at issue ranges over a definite set-the variable £ is a positive
real, n is a natural number greater than or equal to N, z is in D and so on.
In the applications of this definition, the set D (and the set E) is usually a
closed set in the topology of C-because convergence over an open set is
less likely to be uniform (as with the set of r with r < 1 for the harmonic
series).
The idea of "uniformity" which is expressed here is really the same idea
as that which appeared in the foundations of the calculus in §VI.7 with
the notion of uniform continuity of a function. There, as here, the formal
expressions of the idea of "uniform approximation" depend on the use of
the quantifiers. This idea of uniformity can be further developed; there is
for example a notion of a "uniform space" (specialized from topological
space). The idea of uniform limits was once (early 1900's) regarded as one
of the most difficult in Mathematical pedagogy. The difficulty now seems
much less.
One point of uniform convergence is that it provides for smooth opera-
tions with infinite series. For example, if each function in(z) is continuous
for z in some open set U and if ~ inez) converges uniformly in every
closed subset of U, then the sum ~ in(z) is a function of z continuous in
U. Similarly, if in the uniformly convergent series (6) each term inez, w) is
complex differentiable in z £ U for all w, so is the sum s(z,w); moreover,
this sum is the limit of the series of derivatives ainlaz. In brief, a uni-
formly convergent series of holomorphic functions may be differentiated
"term-by-term". Similarly, uniformly convergent series may be integrated
term-by-term; we do not enter into the detailed statement.
336 X. Complex Analysis and Topology

There are many tests of convergence. For example, uniform and abso-
lute convergence can be verified by comparison with a known convergent
series of positive terms, such as the geometric series (3). Specifically, in
(6), if there is some positive real r < 1 and some fixed M such that

for all n,z and w.

then (6) converges and the convergence is uniform and absolute.

The properties of infinite series enter vitally in complex analysis by way
of the expression of holomorphic functions in terms of Taylor series. They
are also vital for calculations (for example, of tables of logarithms and of
trigonometric functions).

8. Power Series
The whole subject of complex analysis can now be started over again
from the idea that a "good" function of a complex variable z should be a
function defined by a power series. Historically, this was the point of
view advocated by Weierstrass in his influential lectures in Berlin. One
may still see the contrast between the Weierstrass viewpoint and the more
geometric approach (Riemann and Klein) in the book Funktionen Theorie
by A. Hurwitz and R. Courant. First consider the elementary functions.
For real x, sin x, eX, and cos x can be calculated from their Taylor series,
and the same series converge (uniformly in compact sets) to provide good
definitions of these functions for any complex argument z:

sin z = z - z 3 /3! + z 5 15! -

So consider more generally series in powers of z-c, such as

what about convergence? The essential fact is that for each such series (1)
there is a real number R, with 0 < < R 00, such that (1) converges for
all z with I z-c I < R and diverges for I z-c I > R. (Of course if
R = 00, R is not a real number, but the intent is still clear!) This number
R is called the radius of convergence.
To show this, it suffices to consider the case when c = O. So call a real
number r > 0 "good" if all the terms I an I rn have a common bound;
that is, if there is some real number M, with I an I rn < Mr for all
natural numbers n. Then take R to be the least upper bound of all the
"good" r (this of course might give an R = (0). If I z I < R, then there
8. Power Series 337

is a good r with I z I <r < R, so I anz n I < I an I rn; there is even a

good s with r < s < R. Then for I z I <
r

I G,Z' I .; I G, I"'; I G, I s' [ i r.; i r

M, [

Hence, by comparison with the convergent geometric series in rls, the

power series (1) converges for I z I < R. Indeed, the proof shows that it
converges uniformly in every smaller closed circular disc I z - c I < r for
any r < R. On the other hand, the series cannot converge for
I z I > R-for if it did so converge, its terms anz n must all have a com-
mon bound, contrary to the definition of R.
The circle I z - c I = R is the circle of convergence. At points on this
circle, the series mayor may not converge. For example, }; zn In has
radius of convergence R = I; for z = I it becomes the harmonic series
(7.4), so diverges, but it converges at every other point with I z I = I!
Now we can restart the development of complex analysis. Consider a
function f: U ~C defined on an open subset U of the complex plane. Call
it regular analytic in U if, for each c f U the function f( z) is given in
some nontrivial circle about c by a power series (1), convergent at least in
the circle. As shown above, the convergence is then uniform in any
smaller concentric circle, so the power series (1) can be differentiated
term-by-term to get as derivative a convergent series
!'(z) = a, + 2a2(Z-C)' + 3a3(Z-c)2 + (2)
Therefore, every function f regular analytic in U is holomorphic in U;
moreover the first derivative off at z = c is!'( c) = a,. By taking higher
derivatives we get jn)(c) = n!a n. This determines the coefficients in the
series (l), so the series must have the form

fez) = f(c) + !'(c)(z-c) + ... + (l/n!)jn)(c)(z-ct + .... (3)

This is just the Taylor series for f(z) expanded about the point z = c.
This argument shows that the Taylor series (3) is indeed the only possible
convergent power series at z = c for a regular analytic function. It also
shows that regular analytic functions are holomorphic. Soon we will show
(by the Cauchy integral) that holomorphic functions are necessarily regu-
lar analytic.
The strict power series point of view can be carried out elegantly and
systematically (but not in this book, which argues rather for a variety of
approaches). From the intended perspective, every good function begins
life as a power series, and so is initially defined only inside its circle of
convergence. Thus, we might begin with the function g defined by the
geometric series

g(z) = 1 + z + z2 + z3 + ... ,
338 X. Complex Analysis and Topology

but only for I z I < 1. How does one use series to recover all the rest of
the intended function 1/( 1- z)? We will see in § II that this question leads
to some fundamental ideas about "analytic continuation".

9. The Cauchy Integral Formula

The Cauchy theorem yields a surprising way of representing any function
f holomorphic in U as an integral. For each point w in a simply con-
nected open set U, take some simple closed piecewise differentiable path h
in U which goes once around w in a counterclockwise direction; for exam-
ple, take h to be a small circle about w. Then f( w) is represented as

f(w) = _1. S, f(z) dz. (1)

2m h z - W
Indeed, the right-hand side can be written trivially as the sum of two
integrals,

_1_ f. f( z ) dz = _I S, f( w) dz + _I S, f( z) - f( w) dz .
2'fTi h z - w 2'fTi h z - w 2'fTi h z - W

In the first integral, f( w) is constant along the path h while h can be

deformed into a small circle about w; for this case we have already calcu-
lated the integral f dzj(z-w) to be 2'fTi, as in (4.8). Hence the first term is
just f( w). In the second integral we may again deform the path of integra-
tion to a very small circle, of circumference 2'fTr, about the point w. The
integrand (f(z) - f(w))/(z-w) is just the difference quotient for f: by
definition it approaches the value 1'( w) of the derivative of f at w, as
z -->W. This means that the integrand is bounded by some constant on (and
inside) the small circle, while the length of the circle can be made as small
as desired. Hence, by the standard maximum and length estimate (4.4) for
a complex integral, this second integral can be made as small as desired-
hence zero. This proves (2)-without the explicit display of ('S which the
reader can readily add in.
This Cauchy Integral Formula (1) expresses a remarkable fact about a
holomorphic function: Its values everywhere inside a circle are completely
determined by its values on that circle-and similarly for any other simple
closed boundary curve. This is a tight control on the behavior of holo-
morphic functions which has no analog for functions of a single real vari-
able. It follows also that the values of each derivative of a holomorphic
function f are determined just by the values of f on the boundary-as we
will soon see explicitly in equation (5) below.
Cauchy's formula will also produce a power series for a holomorphic
function f. For example, suppose that f is holomorphic in a simply con-
nected open set containing the origin w = O. In Cauchy's formula (1),
9. The Cauchy Integral Formula 339

take h to be a circle with center 0 of some radius r and lying inside U. In

the integrand the quotient l/(z-w) can be expanded as a geometric
series

!L ~ 1
z
z - w z (z - w) I

r
(2)

;+ [; +

convergent for I w I < I z I =I- O. So choose some r' with 0 < r' < r,
where r is the radius of the circle h; then for all z on this circle and all w
with I w I < r' this series is uniformly convergent because it converges
for all these w at least as fast as the geometric series ~ (r'lr)n. Thus it
can be integrated over z term-by-term to give a convergent series in w

(3)

where the nth coefficient an is an integral

-
an -
I
2'l/'i
f fez)
zn+l dz . (4)

Moreover, the argument shows that the series (3) converges in any circle,
with center at w, inside the given set U where f is holomorphic.
We have thus shown that a function f holomorphic in U is necessarily
regular analytic in U! In other words, the existence of a complex deriva-
tive for f at each point of U is equivalent to the existence of a convergent
power series at each point. Henceforth we thus can (and will) drop the
term "regular analytic" (which is anyhow old fashioned).
Since the convergent series (3) is the unique power series for f around
w = 0, the nth coefficient an in (4) must be exactly the nth coefficient in
the Taylor series (8.3) for f at w = 0 (i.e., the Maclaurin series). Thus, in
(4), an = fn)(O)lnL The corresponding result holds at every point c in U,
so that

fn)(c) = ~ r fez) dz (5)

2'l/'i J h (z - ct+ 1 '

where the path of integration is any sufficiently small circle about z = c.

This is Cauchy's integral formula for the derivatives of f( z). There is an
easy mnemonic for this formula: Take the original Cauchy integral for-
mula (1), differentiate (under the integral sign) n times with respect to w,
and then change w to c.
340 X. Complex Analysis and Topology

The proof has also shown that the Taylor series for a holomorphic func-
tion f at a point z = c will converge inside any circle about c which is
contained in an open set U where f is holomorphic. This leads to the
slogan: The Taylor series for a holomorphic function f converges in the
circle out to the nearest singularity (i.e., the nearest point where f is
undefined or is defined but is not holomorphic). As stated, this is a slogan
and not a theorem, because the function might have been defined in a
different way, for example as the function f with f(z) = z2 when
>
I z I < I and f( z) = I when I z I I. This function is holomorphic
in I z I < I but not in any larger circle about the origin (and not in any
open set containing z = I). Nevertheless, the Taylor series of the function
at the origin is just z2, so converges everywhere; the function "should
have been" z2 to begin with. The slogan can (with a little trouble) be
reformulated to be rigorous, but the interest is already clear. For example,
the geometric series for 1/( I - z) converges for I z I < I and for no
bigger circle about the origin precisely because z = I is the nearest singu-
larity.
There are many familiar functions holomorphic in the whole plane-for
example polynomials, eZ , sin z, and cos z. Such functions are commonly
called entire functions. In the cases just listed, however, each such entire
function (except the constant polynomials) is unbounded in the whole
plane. This is no accident; Liouville's theorem asserts that any function
f( z) holomorphic in the whole plane and bounded there is necessarily a
constant. For, consider the Taylor expansion (3) for such an f around the
origin; it converges in the whole plane. Its coefficients are given by the
formula (4). If M is a bound for f(z) in the plane, so that I f(z) I < M
for all z, the formula (4) with the path h a circle of radius r and cir-
cumference 2'lTr yields the estimate

For n > I we can let r approach infinity, to get I an I = 0 for all such n.
Thus f( w) = ao is indeed constant.
This proof illustrates again the remarkable utility of the basic estimate
(4.4) for the size of an integral.
From Liouville's theorem one readily derives a proof of the fundamen-
tal theorem of algebra (§IV.lO). For let

be a polynomial of degree n > 0 with complex coefficients bi' If g has no

complex zeroes at all, then f( z) = 1/g( z) is defined and holomorphic in
the whole complex plane. One readily shows also that f is bounded there
(because, for large z, the term I bnz n I dominates I g(z) I . Hence, by
Liouville, f( z) is constant-nonsense.
10. Singularities 341

This proof, by an indirect argument, is clearly not constructive, since it

gives no indication of where to find a complex zero of g(z). It is a splen-
did example of an elegant but circuitous proof, since a really elementary
fact about complex polynomials is here proved expeditiously-once one
has in hand the substance of power series and complex integration and
their application to holomorphic functions. There are other, more topo-
logical, proofs of the fundamental theorem and still others in a more alge-
braic style-including some which are constructive.
The Cauchy integral formula has many other striking and useful conse-
quences. For example, any holomorphic function f: U --.C is an open map-
ping in the sense that the image f( V) of each open subset V of U is an
open subset of C. With this one may also establish a maximum principle:
If f is defined in a bounded connected open set U and on its "boundary"
au so as to be holomorphic in U and continuous on the union of U and
au, while M is the maximum value of Ifez) I on the boundary au, then
f is either constant or, for z in U, I fez) I < M. In other words, f does
not take on any maximum value in the "interior" of U. In this formula-
tion, the boundary au of U is defined to be the complement of U in the
closure [J of U in C; this insures that the closure [J is the union of U and
au.
There are a number of other forms of this maximum principle, as well
as corresponding maximum principles for harmonic functions u(x,y).

10. Singularities
Consider a function f with an isolated singularity; it will suffice to con-
sider such a singularity at the origin z = O. This means that the function
f is holomorphic at least in an open set D of all complex z with
o < I z I < R; such a set is a punctured disc of radius R. The Laurent
theorem for such a function states that there are then two convergent
series, one in powers of z and the other in powers of 1/ z, so that f( z), for
any z in the punctured disc, is given as a sum
00 00

fez) = ao + ~anzn + ~a_mz-m. (I)

n=1 m=1

Moreover, the coefficients ai in this "Laurent series" are uniquely deter-

mined by f. The coefficient a_I of z -I is called the residue of fat z = O.
The proof (which we do not give), starts with the concentric circles
I z I = r and I z I = s within the disc and with r < s. (See Figure 1.)
By an argument like that for Cauchy's integral formula one can prove, for
each complex number w between these two circles, that

few) = _1_
27Ti
fI z
fez) dz - _ I
I =s z - w 27Ti
fI z
fez) dz.
I =r z - W
342 x. Complex Analysis and Topology

Figure I. Punctured disk.

For the first integral, I w/z I < 1, so we can get from it a power series in
w, much as in (2) and (3) of §9. For the second integral, I z /w I < 1, so
this yields similarly a power series in w - I; together these give (1), with z
there replaced by w. We can think of the formula (1) as a representation
of f( z ) as the sum of two functions, the first holomorphic inside
I z I < R, the second holomo~hic for I z I > 0.
An example is the function e fz. From the usual power series for eZ one
has

1
e l/z = 1 + -Z1 + - 1
- +
2!z2
+- - +
n!zn
(2)

This is a series like (1) with infinitely many non-zero coefficients a_ m with
negative index. Such a function is said to have an essential singularity at
the origin z = 0. When there are only a finite number of coefficients
a- m 'i= 0, the function f is said to have a pole. It is convenient to write
Vo(f) = -m for the negative of the order of the pole at O.
When there are no negative powers of z present in the Laurent expan-
sion the function f (if not already defined at z = 0) may be defined there
by setting f(O) = ao; it then becomes holomorphic in I z I < R; it is
said to have had a removable singularity at O. If f is defined (and holo-
morphic) near z = 0, it has a zero there if f(O) = ao = O. The smallest n
with coefficient an 'i= 0 is then the order of the zero at the origin; one
writes Vo(f) = n for this order. The operator Vo (order at z 0) has for =
functions f and g the formal properties

Vo(fg) = Vo(f) + Vo(g), Vo(f + g) ~ Min( Vof, VOg)· (3)

These same properties will reappear in number theory (Chapter XII).

Such an o~erator Vo may be called a valuation: if one defines
IIf II = r Vol, the properties (3) of such a valuation take the suggestive
form

Ilfgll Iljllollgll, IIf + gil < Max( Iljll, Ilgll ) ;

10. Singularities 343

the "absolute value" Ilfll is "large" when the function f has a pole of high
order at 0, and "small" when f has a high order zero there.
One may also consider singularities of a function "at 00", where the
point at 00 corresponds under stereographic projection to the north pole
of the Riemann sphere. More explicitly, the point at 00 can be brought to
the origin by the conformal transformation z f-> liz which carries func-
tions holomorphic near 00 (i.e., outside some large circle) into functions
holomorphic in a circle about 0. Hence, by so transforming the Taylor
series at the origin, one sees, as in (2), that a function f holomorphic for
I z I > R and at 00 has a (unique) expansion
00

f(z) = ao + ~ a_mz- m (4)

m=1

convergent for Iz I > R, in powers of liz. The coefficient ao is its value

f( (0); when ao = 0, it has a zero at infinity. The behavior of a holo-
morphic function is largely controlled by the location and orders of its
singularities. For example, a function which is holomorphic in the
extended complex plane (including 00), except for a finite number of
poles, is necessarily a rational function.

it takes on all complex values except °

The function eZ has an essential singularity at 00; notice, however, that
in any neighborhood of 00 (a
neighborhood of infinity contains the outside of a circle). This behavior is
typical of essential singularities; the so-called big Picard theorem states
that a holomorphic function with an essential singularity at a point c will
take on every complex value, with at most one exception, in every neigh-
borhood of c.
For a function such as a_I /z with a pole of order I at 0, we know by
(4.8) that the (counterclockwise) integral along a small circle h around this
pole is

r a_I
J, --dz = (2'1Ti)a_1 ; (5)
h z

hand f a_mz -m = °
it is completely determined by the residue a _I at the pole. On the other
when m > 1. For integrals around several poles
the contributions add. This observation can be made into a proof of the

Residue Theorem. If f(z) is holomorphic in an open set U except for a

finite number of poles at points CI , .•. ,Ck in U, while h is a simple closed
piecewise differentiable curve homotopic to 0 in U and going once (counter-
clockwise) around all these poles, then

i
k
f(z)dz = 2'1Ti ~ Rescj f· (6)
h j=1
344 X. Complex Analysis and Topology

In brief, the integral is 27Ti times the sum of the residues at all poles
inside h. This famous result provides a convenient way to evaluate certain
definite integrals without finding antiderivatives. Often an integral of a
(real valued) function along some interval of the real axis can be made
part of a "contour integral" which can be evaluated by (6), usually so that
the integral along other (nonreal) parts of the contour h approach zero.
The technique can be learned from the appropriate texts; it is another
illustration of the power of the Cauchy integral theorem.

11. Riemann Surfaces

Now we return to the geometric properties of holomorphic functions.
They involve topological ideas, such as that of homeomorphism: The idea
is that two spaces are homeomorphic (topologically the same) if the first
can be deformed into the second without tearing and without identifying
points. Formally, as in §VIII.6, a continuous map f:X ~ Y of topological
spaces is a homeomorphism if it has a two-sided continuous inverse
g: Y ~x. Another idea is that of a "covering" map. A simple example is
the function 8 I-> 58 for 0 < 8 < 27T; it wraps the circle S 1 with coordi-
nate 8 five times around itself, covering the circle evenly, as suggested by
Figure I. This means that S 1 viewed just locally from below looks like a
stack of separate slices; for each point y of the second circle there is a
small neighborhood U, illustrated in Figure 1, such that the inverse image
of U under p consists of disjoint pieces (slices), each one homeomorphic
under p to U. (Here the "inverse image" of U means the set of all points
x with px E U.) There are many such "covering" maps; thus trig-
onometry rests on the covering map x I-> (cos x, sin x) (the line covers
the circle by the wrapping function of §IV.2).

----
----
----
---- l'
----
lp

----
SI
0 p
----y
u

Figure I. Covering space.

II. Riemann Surfaces 345

Another covering map is (x,y) I-> (e 21riX, e 21riy ), which covers the torus
Sl X Sl by the plane R2. Each point on the torus has a small neighbor-
hood which is the image of (denumerably many) small disjoint neighbor-
hoods from the plane, as indicated in Figure 2. This suggests the general
definition of a covering space Y: A covering map p: Y ~X between topo-
logical spaces is a continuous map such that each x E X has a neighbor-
hood V for which the inverse image p -I V is the union of disjoint open
sets Vj, on each of which p restricts to a homeomorphism Vj ~ V.
Covering maps appear prominently with Riemann surfaces. The exam-
ple of Figure I suggests the holomorphic function w I-> w5 = z, which
maps the w-plane Cw onto the z-plane C, multiplying the argument of
each complex number w =1= 0 by 5, and so covering the z-plane (except for
the origin) five times. If C - {OJ is the complex plane with the origin
removed, this map (C w - {O})~(Cz - {O}) is a covering map in the
sense of our definition. Moreover, w5 = z means that w = 5Vz, so each
of the five fifth roots of z =1= 0 appear. Thus Cw - {OJ is the Riemann
surface on which the many-valued function 5Vz becomes single-valued-
like the Riemann surface for Vz discussed in §VIII.4.
The punctured plane C - {OJ has other coverings, an n-fold covering
for wn = z and one with infinitely many sheets. Consider for instance the
exponential function

w= u + iv I-> z = e W = eU(cosv + isinv). (I)

By the definition (1) of the exponential, the horizontal strip with

o ,,;; v < 2'17 of the w-plane is mapped onto the whole z plane (omitting
0), the segment u = 0 along the imaginary v-axis of the strip is mapped
onto the unit circle in the z-plane, while the horizontal lines through
B, C, and D in the w-plane become rays from the origin in the z-plane
(see Figure 3). Each parallel horizontal strip 2'ITk ,,;; r < 2'IT(k+ 1) simi-

0 0 0
(x, y) 0 0 0
0 0 0

Figure 2
346 X. Complex Analysis and Topology

2rri
------E ---------
DI------
CI------
BI------
-----~o----·u --+-----.x
0' =E'

BI-- B',C f--- C',D I-- D'

D'
w-plane
z-plane

Figure 3. The exponential function.

larly covers the whole z-plane, omitting O. This exponential map is a cov-
ering map C w ..... (C z - {O}), since each open disc about a point z =1= 0
and not containing the origin is covered by a denumerable number of dis-
joint open sets from the w-plane, one from each strip. Actually, this cov-
ering space Cw is both connected and simply connected. One can prove
that any connected and simply connected covering space
p: y ..... (CZ - {OD of the punctured plane is necessarily homeomorphic
(preserving the projection p) to this covering space Cw ; for this reason Cw
is called the universal covering space of the punctured plane Cz - {O}.
Moreover, the vertical translations of the w-plane by 2'lTik (sheets to
sheets) are denumerably many, and correspond to the elements of the
(infinite cyclic) fundamental group of the punctured plane. Other cover-
ings of the punctured plane correspond to subgroups of the fundamental
group, so group theory enters here also.
The plane R2 is also, as above, the universal covering space of the torus.
The inverse of the function w f-> e w of (1) is the many-valued logarithm
function w = lo&- z; it is single-valued on the surface Cw , which is thus
the Riemann surface of lo&- z; it can be defined as the manifold of all
pairs (z,w) of complex numbers which satisfy z = e W , so it is just the
manifold Cw of all complex numbers w.
This leads directly to the general definition of a Riemann surface-as a
special kind of surface-one with a complex structure. Recall from §VIII.7
that a 2-dimensional manifold (a surface) S is a topological space which is
the union of open sets U which form an atlas of charts, which are
homeomorphisms

U open in S, V open in R2 (2)

into open sets of the plane R2 such that each overlap of domains U and
U' yields a homeomorphism. Specifically, the charts cp and cp I overlap if
the intersection U n U' of their domains is not empty; then restriction of
cp -, and cp , to maps cp ,-' on cp (U n U') = Wand <p " on U n U'
compose to give the "overlap map"
11. Riemann Surfaces 347

(3)

between two open subsets Wand W' of the Euclidean plane R2. Since this
plane can be identified with the complex plane, we define: A Riemann
surface is just a surface with an atlas in which all the overlap maps are
holomorphic (i.e., each </>'\ • </>\\ of (3) is a holomorphic function on the
open set W C C). In other words, to get a Riemann surface S take a
bunch Ui of topological spaces, each identified by a bijection </>i as in (2)
with some open set Vi of C, and paste the Ui together by suitable holo-
morphic identifications (3). Since the plane C had one complex dimension
(and real dimension 2), one may also say that a Riemann surface is just a
one-dimensional "complex manifold". Higher dimensional such manifolds
also occur, in algebraic geometry and analysis.
Two different such atlases A and A' on the same space S describe the
same Riemann surface just when each overlap map (from a chart of A to
one of A') is holomorphic. Much as for smooth surfaces, one can then
describe the Riemann surface intrinsically by using the maximal such
atlas. Also, any open subset So of a Riemann surface is itself a Riemann
surface Gust restrict each chart to U n So). Often Riemann surfaces (in
particular So) are required to be connected.
The charts serve to replace open sets of the surface S by open sets of C.
For example, they can be used to define when a complex valued function
f:S ...... C on the surface is holomorphic on S: It is so when for each chart </>
of S the composite
.p-I flU
C:=>V-:;.. U-:;.. C (4)

is holomorphic on the codomain V c C of the chart; here flU is the re-

striction of the given function f to U. In particular, by the overlap condi-
tion, each chart </> is itself a holomorphic function s: U ...... V C C on the
domain U of the chart; since the function s is a homeomorphism to V, it
is often called a unijormizing parameter at each point q t: U; by translating
V in C, one can arrange that the parameter value s( q) is 0 at the point q.
Then the definition (4) reads: f is holomorphic in a neighborhood U of q
if it is a holomorphic function g of the uniformizing parameter s there; in
symbols

f(q) = g(sq); (5)

here g = (f I U)</> -\ is the composite function of (4). This is the classical

language; it may reflect the classical view that functions f of abstract
objects (points q) ought to be replaced by functions of "real" objects
(numbers). This is no longer our view; in particular, it is not invariant,
because it depends on a choice of the uniformizing parameter. At each
point, there many such choices.
The definition of holomorphic functions on a Riemann surface by
charts is just like the definition by charts of smooth functions on a
348 X. Complex Analysis and Topology

differentiable manifold, as in (VIII.8.3). The same idea has still other for-
mulations, as for example in manifolds modelled on Banach spaces.
The complex plane, an open disc, any open set in the complex plane,
and the Riemann sphere (with charts given by stereographic projections)
are all Riemann surfaces. So is the torus-with suitable charts. A function
holomorphic on a surface S has a first derivative at each point with
respect to any uniformizing parameter at that point-although the deriva-
tive clearly depends on the choice of uniformizing parameter.
Riemann constructed his surfaces for particular functions ! because
these surfaces could display the way such functions (even when "many
valued") depend essentially just on the location and character of the
singularities. To illustrate, we may, with Weierstrass, start with a function
!(z) defined just in some circle I z I < R by a power series };anz n con-
vergent there. At each point c' in the circle the power series determines
all the derivatives of!, and so gives the Taylor series };/n)( c ')(z -c't In!
which is convergent in some circle I z - c' I < R' about c'; in part, the
circle may extend beyond the initial disc. Continuing from the second cir-
cle one gets a sequence of Taylor series in circular discs which can be
pasted together wherever they match. This gives a connected Riemann
surface S on which the evident extension of the original power series is
defined and holomorphic. Also each point on this surface S comes from a
point (say the point c') in the complex plane, so the surface is equipped
with a holomorphic function p:S ~C; this is usually a covering map (over
part of C). For example, the Riemann surface for V;, constructed as in
§VIII.4, covers C - {O} twice.
This general process of extension, suitably formulated, is called analytic
continuation. For example, one might start with the power series

_1_ =
2 - z 2
~ [1 + ~2 + ~4 + ... + ~
2n
+ ... J (6)

convergent in the disc I z I < 2. At the point z = i the resulting new

power series will converge in a circle of radius Vs-out to the nearest
singularity, the pole at 2. Using the values in this circle one can form the
Taylor series about the point z = 2 + i; it converges in a circle of radius
1. This process, as suggested in Figure 4, will continue until the whole z-
plane except z = 2 is covered. It really yields the intended functions
1/(2 - z) and not something else because of the following readily estab-
lished general theorem:

The Principle of Analytic Continuation. If two functions! and g are holo-

morphic in a connected open set U of C and if! = g on some non-empty
open subset Uo c U, then! = g throughout U.

For a proof, consider the difference h = ! - g and the subset E C U

where all the derivatives h (n) are zero. This set E includes Uo, because any
II. Riemann Surfaces 349

Figure 4. Analytic continuation.

derivative there can be calculated by taking limits of difference quotients

within Vo. E is open, because the Taylor series for h at each point of E is
zero. On the other hand, each h(n) is continuous, so the set of its zeros and
hence their intersection, the set E, is closed. Since E is non-empty and clo-
pen (closed and open, quickly spoken) in a connected set V, it must be all
of U.
This is a typical use of the definition of "connected" (§VIII.6): a set
which is not the union of two disjoint non-empty open sets.
This process of analytic continuation will readily produce Riemann sur-
faces which are not part of the complex plane. Thus if we start at z = I
with the binomial series (i.e., the Taylor series) for
Vz = (I - (1 - z) )112 and continue as above around the origin, we
return to z = I after one circuit with the opposite value of the sguare
root. Thus the family of all such power series continuations of Vz pro-
duces the Riemann surface of this function, just as it has been described
in §VIIIA.
These examples of analytic continuation via power series may not be
persuasive, because they simply reproduce an already known function
1/(2 - z) or Vz out of one of its Taylor series. But the process can be
used to better effect. ,One example is the Riemann zeta function. This
starts from the (divergent) harmonic series I + 1/2 + 1/3 + .... For s
complex with real part s > I the corresponding series

I I I
res) = I + - s + - s + ... + - + (7)
2 3 nS
does converge, where n S means e S log n with real logarithm log n. If one
writes the complex number s as (J + it for (J and t real (this is the stan-
dard notation in connection with n, the series converges absolutely and
uniformly in any closed half-plane (J >C > 1; hence f is holomorphic in

the whole half-plane (J > 1. It can be continued into the whole complex
plane with the exception of a pole (with residue 1) at the point s = 1;
350 X. Complex Analysis and Topology

however, the analytic continuation is best accomplished via suitable

integrals and functional equations and not just by power series. This result
is a striking example of the power of analytic continuation, which here
has produced from the initial definition (7) a holomorphic function
defined in a much larger domain!
The fundamental theorem of arithmetic states that each rational
number n can be written, and uniquely, as a product of prime numbers.
Hence the infinite series on the right of (7) is formally the product of all
the series

1+ ps
+ is
+ ... +pks
- + ...
pS
for all the primes p. This suggests Euler's formula for ns):
1
~(s) = II
p
[ 1 _ _1 1 (8)
pS

as an infinite product. There is a straightforward way to justify such

hypothetical "infinite" products as limits of actual finite products. Thus
formula (8), which is really an analytic restatement of the fundamental
theorem of arithmetic, is just the first indication of the numerous impor-
tant connections of the ~-function with number theory.
The zeta function has no zeros for (J > 1, but has zeros at
s = -2, -4, .... Riemann conjectured that all of the other zeros of
~(s) are on the "critical" line (J = 1/2. This conjecture, if proven, would
have many remarkable number-theoretic consequences.
r
Here's to the zeros of of s
G. Bernard Riemann has made a good guess
They're all on the critical line, said he
And their density's one over 2'lT log t.
This conjecture of Riemann has got them all started
But many a good man from this life has departed
Without ascertaining with suitable rigor
What happens to zeta when log t gets bigger.

This ditty (to the tune of "Sweet Betsy from Pike") goes on to record
that Hardy has at least proved that there are infinitely many zeros on the
critical line. More seriously there are many notable applications of com-
plex analysis to number theory; they constitute the subject known as ana-
lytic number theory. For instance, this theory gives the most perspicuous
(but not the most elementary) proof of the prime number theorem: If '77'( x)
is the number of prime numbers less than or equal to x, then

Lim '77'( x) log x = 1. (9)

12. Germs and Sheaves 351

Covering spaces arise with Riemann surfaces, but their utility extends
well beyond complex analysis. Any "well-behaved" connected space X (in
particular, any connected manifold) has a covering space p:X ->X which is
connected and simply connected, while the elements h of the fundamental
group of the base space X act as continuous transformations h:X ->X with
ph = p; they are then called covering transformations. This covering
space is universal, in the sense that for every other covering space q: Y->X
there is a covering space p':X -> Y with p = qp', as in
- p'

X~-7Y
x
The universal covering space X can be constructed by taking its points to
be homotopy classes of paths from a fixed base point in X, with endpoints
of the path fixed under the homotopy. Both Massey [1967] and Pontryagin
[1939[ give clear expositions of covering spaces.
The covering space serves to visualize the way the fundamental group
of a space "acts" on other objects associated with the space. It is just one
example of the interaction between the quite explicit constructions of
complex analysis and the general development of topology.

12. Germs and Sheaves

The process of analytic continuation by pasting together circles of conver-
gence can be stated more conceptually in terms of germs and sheaves of
germs. Consider two functions f and g holomorphic in neighborhoods
U and V of a point c in C. They are said to have the same germ at c if
there is an open neighborhood W of c with W c Un V and f = g in
W. (With the same words, we can define when two continuous functions
have the same germ at a point in any topological space.) The binary rela-
tion "f and g have the same germ at c" is reflexive, symmetric and transi-
tive; hence we can form for each f its equivalence class [f]c, the germ of
f at c, consisting of all g holomorphic in some neighborhood of c and
having there the same germ as f.
By Taylor's theorem, f and g have the same germ at c if and only if
they have the same Taylor series at c. Therefore a germ at c can be
identified (and so made more "concrete") with a power series in z - c,
convergent in some open circle about c. No such easy identification is pos-
sible for germs of real-valued functions which are merely continuous.
Let Ac denote the set of all germs of holomorphic functions at the point
c, while the union A of all the Ac is the set of all germs at all points of C.
It is in a natural way a topological space, called the sheaf of germs of
holomorphic functions on C. To describe this topology, take for each holo-
morphic f: U ->C the set of germs
352 X. Complex Analysis and Topology

Wf,U = {all germs of f at points of U} = {[f]e ICE U}

and call this set a neighborhood of the germ [f]e in A. In brief, neighbor-
ing germs are those which come from the same holomorphic function.
This informal description of "neighborhoods" can be made formal, by
stating that the sets Wf,u are a base for the open sets of the intended
topological space; this means that an open set of A is any union (possibly
an empty or an infinite union) of sets Wf,u. To show that the axioms for a
topological space hold, it is then enough to show that the intersection of
two such open sets is again open, and for this it is enough to show that
any intersection Wf,u n Wg,v is either empty or a neighborhood. This is
easy: If [f1 = [g]e at some point c E U n V, then f = g in an open
set U' containing c; the desired intersection Wf,u n Wg,v is therefore the
union of such sets Wf,u', This proof that A is a topological space is a typi-
cal use of the idea of neighborhood and the notion of a base for a topol-
ogy. In this particular case, A is actually a Hausdorff space; this would
not be the case for sheaves of germs of simply continuous functions.
There is a natural projection p:A ..... C which sends each germ [f]e to the
indicated point c. This projection is continuous; in fact, it has a much
stronger property: It is a local homeomorphism. To each point a in A there
is a neighborhood W such that p( W) is open in C while p, when
restricted to W, is a homeomorphism W ..... P W. In other words, each germ
a has a neighborhood which, under p, behaves exactly like a neighbor-
hood of pa in the complex plane; the sets Wf,u are such neighborhoods.
Observe that the projection p:x ~X of any covering space is a local
homeomorphism in the sense of this definition. However, the projection
p:A ..... C for germs is by no means a covering map. About each point c in
C one can find holomorphic functions with smaller and smaller radii of
convergence. Hence there is no one neighborhood of c which reappears
around every germ about c. In short, the space A is very large; it can evi-
dently be described as the space of all possible convergent power series
~ an(z-ct, for all c, where each power series is regarded not as a func-
tion of z, but simply as a thing-a point of A. In general topology,
"points" can be almost anything, and "spaces" can be very large and still
conceptually useful. There is also a "very large" continuous function
F:A ..... C, which sends each germ [f]e to the value f( c) of the original f. In
symbols, f([f] c) = f( c).
This one space A with the function F "represents" all possible holo-
morphic functions by way of "cross sections", as follows:

Theorem. For any holomorphic function g: V ..... c on an open subset V of the

complex plane C there is a unique continuous cross section s: V ..... A of P with
F·s = g. Conversely, each continuous cross section s: V ..... A of p over an
open set V of C yields a holomorphic function g on V as the composite
g = F·s.
12. Germs and Sheaves 353

Here a cross section s of p means a map s: V ...... A such that the composite
p·s is the "identity"; that is, the inclusion of V in C as in the diagram
below.

1\\
F
A\ ~ c

c :) 'v
PROOF. Given g, the only choice for s is the mapping with s( c) = [g]c for
each c in V. This definition provides a continuous cross section; moreover
for F as above the composite F·s is, as in the diagram, just the given holo-
morphic function g.
Conversely, let s be any cross section of p over an open set V in C. It
must send each point c ( V to the germ [f]c at c of some function
f = fs,c holomorphic in some open set U containing c. The composite
g = F·s then sends c to F(sc) = f(c); it remains to show g holomorphic.
But the continuity of s at c means that some neighborhood Vo of c in C is
sent by s into the open set WI, u of germs of f. Hence, for c' in Vo,
sc' = [fk' so that gc' = fc '. Therefore g, like f, is holomorphic.
It may at first seem strange that the mere requirement of continuity of a
cross section over V produces a function g on V which is not just continu-
ous, but even holomorphic. This results because the holomorphic charac-
ter is built into the construction of A: Each point of A is the germ of some
function which is holomorphic.
Holomorphic functions on C yield a presheaf H by

H( V) = {g I g: V ...... C is holomorphic} .

Indeed, if VI C V, each holomorphic function on V restricts to one on

VI, and these restriction maps make H( V) a pre sheaf, as defined in
§VIII.ll. Moreover, this presheaf H is a sheaf. If V = u Vj is any cov-
ering of Vee by open sets Vj, while there are corresponding holo-
morphic functions g/ Vj ...... C such that each pair gj and gk agree on the
intersection Vj n Vb then there is a function g ( H( V) which restricts to
gj on each Vj; indeed, g is defined at each point c ( V as the common
value of all the gj( c) for all indices j with c ( V); this g is clearly holo-
morphic.
By the theorem, this sheaf H may also be described by taking each
H( V) to be the set of cross sections over V of the projection p:A ...... C. In
other words, the idea of a holomorphic function can be formulated by a
sheaf or as cross sections of the big space A.
The space A is actually a 2-dimensional manifold and a Riemann sur-
face (but not connected), because each point [f]c of A has a neighborhood
354 X. Complex Analysis and Topology

WI,V which is homeomorphic to the open set U C C, and these

homeomorphisms serve as the charts for the desired Riemann surface
structure. However, A is far from being connected. It does consist of con-
nected pieces, according to the following very general decomposition
theorem.

Theorem. Any n-dimensional manifold M is the union of disjoint connected

subsets Mj , each of which 'is itself a manifold of dimension n and an open
subset of M

The subsets Mj are called the connected components of M; this result is

true for much more general topological spaces; we need it only for mani-
folds, where "connected" is equivalent to "path connected" (see §VIII.6).
To prove the theorem, take any point q in M. The component M j con-
taining q is then defined to be the set of all of those points of M which
can be joined to q by a (continuous) path in M. Since each point q' has
an open neighborhood homeomorphic to an open set of Rn , every path
from q to q' can be prolonged (e.g., by straight segments in Rn) to any
point in a suitable open neighborhood of q'. Therefore, the set M j is
open. It can be described as the equivalence class of a point q under the
reflexive, symmetric, and transitive relation "can be connected by a con-
tinuous path".
In the space A each connected component A j is just a connected
Riemann surface. Indeed, a point q of A is a point c in C together with an
open qisc with center c and a power series fez) = ~ an(z-ct conver-
gent there. A path from q in A then looks like a path from c in C with a
disc and series attached at each point, so that they overlap, as in Figure 1.
The figure shows only a finite number of discs, which suffice to cover
because the interval is compact. It thus follows that the connected com-
ponent of the point q in A consists of all the power series obtained by
analytic continuation of the original series f; hence it is exactly the
Riemann surface of f. This conceptual description of the Riemann surface

Figure 1. Continuation along a curve.

12. Germs and Sheaves 355

as a connected component of the big space A helps to understand exam-

ples like the analytic continuation of l/(z - 2) displayed in Figure 11.2.
This example illustrates how very general abstract concepts can give a
better understanding of concrete examples. Once we have constructed the
big space A and applied to it the easy geometric idea of breaking it up
into components we have all the Riemann surfaces, including for instance
the Riemann surface so meticulously constructed in §VIIIA for the func-
tion V(z2 -1 )(z2 -4).
This cheerful statement is not quite true, but it can be easily repaired.
First, one should enlarge A to include also germs at 00, giving a local
homeomorphism p:Aoo -"S2 to the Riemann sphere. Second, to each com-
ponent of Aoo one should adjoin "branch points", like the point 0 for Vi
and the four branch points z = ± 1, z = ±2 for V(z2-l)(z2-4).
These examples are all branch points of "order" 2. Generally, a branch
point has some positive integer k as order and can be considered as the
"germ" of a convergent power series in a uniformizing parameter
t = (z_c)l/k. Third, when a function f has a pole the "germ" of the
corresponding Laurent series should be adjoined to its Riemann surface:
This is to include poles at branch points. This will insure that the
Riemann surface for a rational function such as 1/( z - 1) is the whole
Riemann sphere, including the pole at z = 1 and the zero at 00. In par-
ticular this Riemann surface is compact.
There is an important monodromy theorem which specifies when ana-
lytic continuation along two different paths to a point c' from a germ at c
yields the same germ at c '. The proof depends on an important topologi-
cal process of "lifting" a homotopy from C to A.
Our conceptual definition of a Riemann surface is just the starting point
for many beautiful and subtle developments. For example, an analytic
isomorphism f:S -"S' of two Riemann surfaces is a holomorphic function
f together with a two-sided hoi om orphic inverse g:S' -"S. The Riemann
mapping theorem asserts that any connected and simply connected open
set U of C which is not all of C is analytically isomorphic to the unit disc.
(One can also describe all the analytic isomorphisms of the disc to itself.)
More generally, the uniJormization theorem asserts that any connected and
simply connected Riemann surface is analytically isomorphic to one of the
following: The unit disc, the complex plane, or the Riemann sphere. The
proof of this theorem requires analytic methods much more subtle than
any we have summarized.
Such analytic isomorphisms also have practical uses. Thus the exterior
of the unit circle in C can be mapped conformally upon the exterior of
the profile of an airplane wing, as in Figure 2. A steady flow of air around
the wing can then be replaced by a flow around the circle (see §3 above).
There are very useful explicit formulas for constructing and approximat-
ing such conformal maps.
356 X. Complex Analysis and Topology

o Figure 2. Airfoil.

13. Analysis, Geometry, and Topology

This chapter has not been limited to complex analysis. It has been
arranged to illustrate the remarkable interaction between analysis and
geometry, in which various immediate and practical problems lead to new
and general concepts. These concepts in turn will help to understand the
solutions of the original problems and turn out to have specific applica-
tions elsewhere; moreover they help develop subtle general theorems like
the big Picard theorem and the uniformization theorem.
Complex analysis starts from the notion of differentiation, as it was
needed to formulate velocity and acceleration in mechanics. Complex
differentiation yields the class of holomorphic functions and their remark-
able properties. At the same time, these functions provide conformal maps
and have application to electrostatic potential and fluid flows. The con-
struction of a variety of holomorphic functions suggests the process of
analytic continuation along a curve, which can then be codified and
understood as the choice of a connected component in the big space A of
germs of holomorphic functions. The explicit Riemann surfaces, used to
handle many-valued functions by way of slitting and pasting planes
together, all fit in the big topological space A.
This is typical of the development of abstract concepts. The original
and tangible notion that a space is something physically visible, consisting
just of points with no extension, gets gradually enlarged, until one arrives
at wildly big topological spaces such as the space A of germs in which a
point is "really" a whole disc carrying a power series. This same general
process-topologize the "manifold" of all the objects under study-has
many other uses in Mathematics.
The notion of function also changes. The holomorphic functions (on
open sets U of C) form a sheaf H which can be represented as the sheaf
of cross sections of a local homeomorphism p:A ~c. This was historically
the first example of a sheaf-a notion which now crops up for all sorts of
manifolds, complex, differentiable, smooth, or algebraic. More generally,
give a sheaf S on any topological space X, one may construct a space A,
whose points are the germs of S, together with a local homeomorphism
p:A ~x. The given sheaf S is then the sheaf of cross sections of p, so that
13. Analysis, Geometry, and Topology 357

S( U) is the set of cross sections over U. In other words, any good class of
functions can be represented simply as continuous cross sections. This
idea has applications even in set theory, as we will note in the next
chapter.
CHAPTER XI

Sets, Logic, and Categories

The rich multiplicity of Mathematical objects and the proofs of theorems

about them can be set out formally with absolute precision on a remark-
ably parsimonious base. Thus almost all the objects of Mathematics can
be described as sets: A natural number is a set of sets (a cardinal), a
rational number is a set of pairs (an equivalence class), a real number is a
set of rationals (a Dedekind cut), and a function is a set of ordered pairs
(a table of values). Similarly the theorems of Mathematics can all be writ-
ten as formulas in a very parsimonious formal language which uses only
set-membership, the basic connectives of logic (or, not, there exists) and
the needed primitive terms of each subject (thus "point" and "line" for
incidence geometry). Finally, most of the proofs of Mathematical
theorems can be stated with absolute rigor as a sequence of inferences,
each an instance of a finite number of basic schemes of inference.
A "foundation" for Mathematics is provided by this remarkable reduc-
tion of Mathematics to sets, well formed formulas, and formal inference.
It does miss a little of Mathematics-formally undecidable theorems and
constructions which need the "universe" of all sets. Since we hold that
Mathematics is based on human activities and scientific problems we have
delayed the discussion of the foundations to this point, when we have at
hand a considerable sample of the things to be "founded".
Almost all Mathematicians now use the convenient language of set
theory to describe Mathematical structures, but very few Mathematicians
can recite the axioms which found set theory, while almost none of them
(except for the logicians) can specify the exact rules of logical inference.
Hence in this chapter we start with the standard Zermelo-Fraenkel (ZF)
axioms of set theory and then the calculi of propositions and of predi-
cates, complete with modus ponens and the other needed rules of infer-
ence. This allows a formal definition of "proof' and a description of the
Godel incompleteness theorem (§7). (In a recursive formal system strong
enough to code arithmetic, there are formal statements which can be nei-
ther proved nor disproved.) Moreover, the Zermelo-Fraenkel axioms do
I. The Hierarchy of Sets 359

not settle the famous continuum hypothesis-there are models of ZF for

which the hypothesis holds and others for which it fails. The demonstra-
tion of the latter fact by Cohen's method of forcing (§8) suggests other
"independence" results and leads to the construction of many alternative
models of set theory. Another result is the introduction of a considerable
variety of axioms meant to supplement ZF.
For these reasons "set" turns out to have many meanings, so that the
purported foundation of all of Mathematics upon set theory totters. But
there are other possibilities. For example, the membership relation for sets
can often be replaced by the composition operation for functions. This
leads to an alternative foundation for Mathematics upon categories-
specifically, on the category of all functions. Now much of Mathematics
is dynamic, in that it deals with morphisms of an object L into another
object of the same kind. Such morphisms (like functions) form categories,
and so the approach via categories fits well with the objective of organiz-
ing and understanding Mathematics. That, in truth, should be the goal of
a proper philosophy of Mathematics.

1. The Hierarchy of Sets

The explicit recognition of sets as objects of serious Mathematical study
came about in the 19th century in at least two ways. On the one hand,
Dedekind used sets to explain Kummer's ideal numbers as suitable sets of
ordinary numbers (§XIIJ) and to explain real numbers as cuts-that is, as
sets of rational numbers. On the other hand, questions about Fourier
series forced the consideration of quite general (and bizarre) sets of real
numbers. Fourier, in studying the conduction of heat (1822) was led to
represent periodic functions, of period say 2'77, as sums of infinite series of
sines and cosines,

j(x) ~ ao /2 + al cos x + b l sin x +

(1)
+ an cos nx + bn sin nx + ....
It was remarkable how many such functions could be so represented
(§VI.ll) but Fourier's methods of showing this were intuitive and unreli-
able. Because of their evident utility not just for the heat equation but for
other partial differential equations, Mathematicians turned to the careful
examination of such series. This led to the lively development of the sub-
ject now known as harmonic analysis, not just for the rotation group of the
circle (which is present but hidden in (1) by the use of the angular coordi-
nates x, 0 <; x < 2'77, for the circle) but for many other groups.
One problem about such trigonometric series, as posed by Riemann, is
the problem of uniqueness: When a function f is the limit of such a series
360 XI. Sets, Logic, and Categories

(I) for all x, does f determine the coefficients an and b n uniquely? George
Cantor found that the answer was "yes"; indeed, he found that he did not
need to assume that the series (1) converged to f(x) for all x, but only for
a great many x. So he was led to the question: For what sets S of reals x
will

~ an cos nx + bn sin nx = 0
n

for all x ES imply that all the coefficients an and bn are zero? It turned
out that some of these sets S required elaborate descriptions. This started
him thinking about general sets; so he became the founder of set theory.
Cantor developed set theory extensively, introducing the cardinal
number, card A, of a set A (§II.8) and the ordinal number of a well-
ordered set (§II.9). Two cardinal numbers are compared by the rule that
Card A < Card B if there is an injection A ~B. The Schroeder-Bernstein
theorem asserts that Card A < Card B and Card B < Card A together
imply Card A = Card B. (For a proof, see Survey, Chapter XIII.) From
this and the axiom of choice it follows that the cardinal numbers are
linearly ordered. Using a "diagonal" argument, Cantor proved that the set
R of real numbers is not denumerable. This means that Card N <
Card R. He then raised the question: Is there any cardinal number prop-
erly between these two? The statement that there is no such is Cantor's
continuum hypothesis. (R is commonly called "the continuum".)
Bigger and bigger sets can be constructed by the power set operation P,
which assigns to each set A its power set P(A), consisting of all subsets of
A. For example, peN) has the same cardinal as R. The diagonal argument
mentioned above shows that the cardinal of P(A) exceeds that of A. For if
f: A ~P(A) were a bijection, one could define a subset S cA as
S = {x I x E A but not x E f(x)}; (2)

here x E A means "X is an element of A", as usual. Since S E P(A) and

f is a bijection, there must be some a in A with f( a) = S. Then a E f( a)
would, by the definition (2), give a tt f( a) = S; while a tt S means by
(2) again that a E f(a)-a contradiction. This is called a "diagonal" argu-
ment because the set S is the complement in A of the "diagonal" set con-
sisting of all those elements x with x E f( x). Recall that the Russell
paradox (II.8.5) also involves a diagonal construction

R = {S I not S E S}. (3)

Inevitably we come to the question: What is a set? Initially, we

described sets as (well defined) collections of Mathematical objects,
without really specifying what a "collection" is or what an "object" is (a
number? a point? a series?). Now note that many constructions produce
I. The Hierarchy of Sets 361

sets all of whose elements are themselves sets: The set P(A) of all subsets
of A; the set Zm of all congruence classes of integers, modulo m, or the
cardinal number of A. This raises the possibility of considering only those
sets whose elements x are themselves sets, the elements t of those ele-
ments x again being themselves sets, and so on. This "and so on" will be
more persuasive if we start from the bottom with the empty set, written as
0. It has only one subset, itself, so the set P( 0 ) of all subsets of 0 is just
the set { 0} whose only element is the empty set. Write these two sets as

Ro = 0, R 1 = P( Ro) = {0}. (4)

Now the one-element set { 0} has two subsets, itself and the empty set, so
its power set is a two-element set

R2 = P(RJ) = {0,{0}}. (5)

This set in its turn has four different subsets, so its power set is

(6)

here {{ 0 }} of course denotes the set whose only element is the set whose
only element is the empty set 0. Iterating this process produces sets Rn
with 2n elements. The union of all these sets Rn is a denumerable set Rw;
here w designates the first infinite ordinal number.
The construction started in (4), (5), and (6) can be formalized as a
recursion over the ordinal numbers. Here an ordinal y =1= 0 which is not
the successor y = a + I of some ordinal a is called a limit ordinal; w is
the first such. The definition of the sets Ra by recursion on the ordinal
number a then reads

Ro = 0, (7)

where the last clause applies when y is a limit ordinal. Since the ordinal
numbers a are well-ordered (i.e., each non-empty set of ordinals has a
first element), one can define the rank of any set x in this hierarchy (7) to
be the first ordinal a with x E: Ra. All the elements of x then have
smaller rank, so that no set x can be a member of itself. Each set x in the
hierarchy has only sets as elements, and can be pictured as a "tree" which
presents the elements of its elements. . .. For example, the set (6) above

{F
is the following tree
~{0.
o {0}
I
o
(8)

o
362 XI. Sets, Logic, and Categories

The whole cumulative hierarchy (7) is intended (or imaged) to produce

all the sets that we need for Mathematics. Though it does not produce
any "universal set" of all sets, it does produce sets of all finite sizes as well
as many infinite sets. Note that the description of this hierarchy uses the
(informal) ideas of empty set, union of a set of sets, subset, and power set,
and it rests also on the intuitive idea of the ordinals (or some equivalent
notion of "stages" of construction). On this basis, the hierarchy provides a
systematic "picture" of all needed sets. It is speculative and not formal
(see also §4 below).

2. Axiomatic Set Theory

The standard foundation for Mathematics intends to build up all
Mathematical concepts out of sets. We have already indicated some of the
steps: Functions and then bijections are defined as sets of ordered pairs of
sets; cardinal numbers are equivalence classes of sets, under bijections;
natural numbers are finite cardinals; integers use pairs of natural
numbers; rational numbers are equivalence classes of pairs of integers;
real numbers are sets (Dedekind cuts) of rationals; functions on the reals
are suitable sets of pairs of reals; complex numbers are pairs of reals;
functions of complex numbers are sets of pairs of complex numbers.
Riemann surfaces are sets of charts where each chart is a suitable function
which again is a set of pairs. And so on. With some care and some pedan-
try, (almost) all Mathematical concepts can in this way be reduced to con-
structions on sets. Of course a class-as in the case of equivalence
classes-is here taken to be a set, while an ordered pair of sets is to be
defined as a certain set, as indicated in §VA.
This approach to Mathematics has the advantage that every concept
can be made absolutely clear and explicit. Thus a Riemann surface is just
a set of charts and not some mysterious configuration formed by ingeni-
ously pasting together sheets cut from the complex plane. But if this
approach is to be dependable, it must avoid the Russell Paradox and its
ilk. That paradox seems to spring from the naive conception of a set as
any collection of identifiable things. To avoid this, one may adopt axioms
for sets which appear to hold in the cumulative hierarchy we have
described. For this purpose, sets are not described as "collections"; one
simply posits that there are sets x,y,z, ... which may stand in the relation
x E y of membership (x is an element of y) and that these sets satisfy
certain axioms formulated strictly in terms of this relation E. Everything
else about sets and the Mathematical objects constructed from them is to
be defined (ultimately) in terms of this membership relation and all
theorems are to be proved from these axioms.
The Zermelo-Fraenkel axioms system (ZF) is standard. One first
defines the inclusion relation c between sets x and y by
2. Axiomatic Set Theory 363

x C Y <=> For all t, (t EX=> t E y). (1)

where <=> is short for "if and only if' while => is short for
"implies". Then the equality of sets is defined by

x =y <=> x C Y & Y ex. (2)

With (l), this amounts to stating that two sets are equal if and only if they
have the same elements-a view basic to the idea of a set.
The Zermelo axioms now come in the following list:

Extensionality, Null set, Pairing, Power set, Union, Infinity,

(3)
Comprehension, and Regularity.
In detail, these axioms read:

Extensionality.

x = Y& Y E z imply x E z. (4)

This states that equals may be substituted for equals on the left in any
statement y E z. The definition (2) of equality has already provided for
substitution on the right in t E y:

x =y and t E Y => t EX.

Sometimes both E and = are taken as (independent) primitive notions

for sets. In this case (4) is just a statement of a basic property of = (an
axiom), while the definition (2) becomes in this version the axiom of
extensionality.

Null set. There is a set 0 with no elements.

By the definition (2), any two such sets are equal.

Pairing. For any sets x and y there is a set u so that, for all t,

t E u <=> (t = x or t = y). (5)

By (2) again, any two such sets u are equal, so one writes u {x,y}. This
axiom provides also for a singleton: A set {x} = {x,x} with just one ele-
ment x. It also produces ordered pairs <x,y> = {{x},{x,y}} as
explained in (VAA).

Power set. For any set x there is a set u with, for all s,

s E u <=> (s ex). (6)

This set u, again unique by (2), is the power set u = Px.

364 XI. Sets, Logic, and Categories

Union. For any set x there is a set u with, for all t,

t E u <¢:;> There is an s with t ESE x.

This set u, again unique, is the union of all the members of x, usually
written u:
x

t E Us<¢:;> (:3s) (t E s & SEx).

SEX

In particular, if x = {y,z}, this set is the union y u z.

Before formulating the next axiom (that of infinity), notice that we can
now construct for each set x the singleton set {x}, the doubleton {x, {x}}
and hence the union of the latter set, which is called the successor of x

s(x) = x U {x}. (7)

In particular, this will yield all the finite ordinal numbers, as in the
definition suggested in (11.9.5):

° = 0, 1= s(0) = {a}, 2 = s(l) = {O,{O}} = {O,l} .... (8)

Each ordinal is thus the set of all smaller ordinals, so the first infinite ordi-
nal w should be the set of all finite ordinals. Hence the existence of at
least one infinite set can be insured by requiring that there be a set con-
taining this w; that is, a set containing this 0 and closed under the opera-
tor s of (7).

Infinity. There exists a set w with 0 E wand such that x E w implies

s(x) E w.

Comprehension. For any set u and any property P of elements of u there

is a set s with

xEs <¢:;> x E u and x has P. (9)

That set s, again unique, is usually written as

s = {x I x E u and x has P}. (10)

We have used this construction repeatedly; for example, it provides for

the intersection u n v of two sets. It also allows one to "cut out" or
separate from a given set u, the subset s of all those elements x of u which
do have the specified property P; hence it is sometimes called the axiom
of separation. This formulation of the axiom does not allow an immediate
construction of Russell's paradoxical set R of (1.3), because it provides
2. Axiomatic Set Theory 365

only for the construction of subsets of a given set u, and not, like R, for
subsets of the universe of all possible sets. This formulation is also in
accord with the way sets are constructed in the cumulative hierarchy of § I
above.
Actually, this axiom (9) is really an infinite family of axioms, one for
each property P; it is thus called an axiom scheme.
These axioms are essentially those formulated by Zermelo in 1908. The
last axiom (9) is not in satisfactory form, because it refers not just to the
one primitive notion E, but also to a "property". With Skolem and
Fraenkel, one must then append here the explanation that a "property" P
of x means something specified by an explicit set-theoretic formula; in
other words, something formulated only in terms of x, membership and
the usual logical connectives. Thus for example, "For all t and s, t E x
and SEX imply t c s" is a property of x.

Regularity. If a set x is not empty it has an element w which has no ele-

ments in common with x.
In other words

x -=1= 0 => (3 w) w E x and x n w = 0. (11)

To see what this axiom means, observe first that it shows that there can
be no set x with x E x (no set is a member of itself). For if this were so,
the singleton set {x} is non-empty, but its only element, to wit x, has an
element (x again) in common with x. Similarly, the axiom insures that
never x EyE x. For if this were so, {x,y} would be a non-empty set
which has elements (x or y) in common with each of its elements. Also,
once we have defined sequences Yn of sets, the axiom of regularity will
insure that there is no infinite "descending" sequence with

... Yn E Yn-I E Yn-2 E '" E YI E Yo, (12)

for then the set x = {YO,YI, ... } violates this axiom. For this reason, the
axiom of regularity is also called the axiom of foundation: Given a set Yo,
there is no infinite regress of elements of its elements.
The intuitive intent is that one should be able to "see" that all these
axioms appear to be valid for the cumulative hierarchy of sets Ra
described in § l. For example, in the case of the regularity axiom, each
non-empty set x has some ordinal rank lX, while all the elements of x have
smaller rank. Among the ordinal ranks of elements t E x there is (by
well-ordering of ordinals) a first ordinal {3. Any element w E x of this
rank {3 has for x the property required in the regularity axiom (II). Put
differently, an infinite descending sequence of sets like (12) would produce
an infinite descending sequence of ordinals. This intuitive argument
depends on a prior intuitive notion of ordinal.
366 XL Sets, Logic, and Categories

These Zermelo axioms are usually supplemented by the axiom of

choice. After functions have been defined, this axiom reads

Choice. For every set x with no non-empty members y there is a function

1 with domain x such that I(y) E Y for each y EX.
In other words, 1 "chooses" for each y =1= 0 in x a particular element
I(y) E y; thus 1 is a "choice function" for x. It is needed only when x is
an infinite set, and many "infinite" arguments in Mathematics use this
axiom, often in one of the equivalent forms (a Zorn lemma). Zermelo
proved that the axiom of choice implies that every set can be well-
ordered.
Just as in the axiomatization of geometry, the operational meaning of
the (Zermelo) axioms for sets is that "everything" about sets should be
formally demonstrable from the axioms without using the intuitive idea
that a set is some sort of collection of things. We now illustrate this pro-
cess briefly, but only for the ordinal and cardinal numbers.
First the intuition. The basic idea, as already used, is that of a well-
ordered set. In particular, the ordinals themselves are well-ordered:

0<1<2<3< .. · <w<w+l< .. ·

Generally, an ordered set is well-ordered when every non-empty subset

has a first element in the order or, equivalently, when there are in the
order no infinite descending sequences (compare the sequence (12). Now
the axioms easily produce sets which are "models" for the finite ordinals:
o = 0, while the successor function s of (7) produces finite ordinals such
as

1 = {OJ, 2 = {O,l}, ... ,n = {0,1,2, ... ,n-I}.

The axiom of infinity suggests that the first infinite ordinal is

w = {0,1,2,3,4, ... }.
On inspection, each of these sets is "transitive". Here a set t is called tran-
sitive when x EyE t implies x E t; in other words, every element y of
t is also a subset of t.
Next the formal development. With von Neumann, one can now define
an ordinal to be a transitive set a such that every element t E a is also
transitive. (This is the case for wand the finite ordinals n.) It then follows
from the axioms that the elements of an ordinal a are linearly ordered by
the membership relation E. With the regularity axiom one can then
prove that this linearly ordered set a is well-ordered-i.e., that each non-
empty subset x C a has a first element in the order of elements of a. For,
by regularity, there is some element y E x with no element in common
with x. In particular, z < y for some z E a means that z E y and so by
2. Axiomatic Set Theory 367

transitivity of x that Z E x, a contradiction to the fact that x and y have

no elements in common. Hence y is the required first element of x.
This proof is one of the critical uses of the regularity axiom.
This formal definition of the von Neumann ordinals IX will lead to all
the needed properties of ordinals; in particular one can prove that every
well-ordered set P is order-isomorphic to a von Neumann ordinal. This in
turn means that the von Neumann ordinals "represent" all the ordinals
given by the older definition (§II.9): An ordinal number is the order type
ord P of some well-ordered set, where

ord P = {P' I P' E V and P ""' P'}, (13)

while P ""' P' means that P and P' are ordinally isomorphic and V is
some "type" or "universe". (Some sort of universe is needed to make the
comprehension axiom apply here; we cannot simply take all sets P' with
P ""' P'.) Whatever the universe, if it is big enough it will contain a von
Neumann ordinal IX with P ""' IX, so the older ordinal, or P, is realized or
"represented" by IX.
(One can describe a universe as a set V such that the elements of V by
themselves satisfy the Zermelo axioms, but that requires an added
axiom-there exists a universe-and might (and has!) led to the use of a
whole succession of universes!)
The naive definition of a cardinal number (§II.8) in terms of the rela-
tion of cardinal equivalence reads

card S = {S' I S == S'} ; (14)

again an illegal use of comprehension. This difficulty can be avoided by

using a universe, much as in (l3). More elegantly, one can first show from
the axioms that one can carry out definitions by recursion over the ordi-
nals, as in the definition (1.7) of the hierarchy Ra. This also allows the
definition from the axioms of the ordinal rank p( T) of each set T, as the
least IX with T E Ra. Then one can describe the cardinal number of a set
S by collecting only the equivalent sets of smallest rank. Formally, one
can legitimately define

card S = {S' I S == S' and VT, T == S ~ p(S') ~ p(T)}. (15)

Alternatively, the troubles with comprehension in (14) can now be

avoided by defining a cardinal number to be an ordinal which is not car-
dinally equivalent to any smaller ordinal. Thus CAl is a cardinal, but CAl + 1
is not. Again, the cardinal numbers so defined provide representatives for
each of the (perhaps non-existent?) equivalence classes (14)-because
every set can be well-ordered.
This treatment of the basic Mathematical idea of number-ordinal and
cardinal-illustrates the clarification of ideas possible by a suitable
368 XI. Sets, Logic, and Categories

axiomatization. Here the initial (and central) idea of a well-ordered set

has come into clearer focus by way of the models provided by the regular-
ity axiom and the notion of a transitive set.
From the infinite set wand the iterated power set axiom one obtains a
tower

of sets which, by the diagonal argument, get larger and larger. To get a
still larger set one would need to form the union of all these sets pn w for
all natural numbers n, and then apply P again. But one cannot form this
union by the axioms unless one knows that there is a set consisting of all
the sets pnw; in other words, that the image of w under the function
n ~ pnw, for nEw, is a set. The final axiom (replacement) does insure
that the image of any set under a function is again a set; here a "func-
tion" x ..... y is to be described by a property R(x,y) specified by a
formula-much as in the case of the comprehension axiom. This again is
an axiom scheme (one axiom for each property)!

Replacement. If R(x,y) is a formula stated in terms of the sets x,y and the
membership relation while u is a set such that for each x E u there is
exactly one y with R(x,y), then there is a set consisting of exactly all these
y.
This set is of course the image of u under the function defined by R.
The replacement axiom scheme implies the comprehension axiom
scheme. When replacement is adjoined to the Zermelo axioms, one gets
the Zermelo-Fraenkel (ZF) axioms for set theory. Normally, one also
includes the axiom of choice, to get ZFC.

3. The Propositional Calculus

To achieve complete clarity, Mathematical statements can be formulated
in a specific symbolic language. This was already needed, for example, to
explain in principle what is meant by a "property" in the comprehension
and replacement axioms of set theory. The use of a language plays a cen-
tral role in logic.
As a start, we consider a language limited to the standard logical con-
nectives "and", "or", and "not". (In symbols, 1\, V, and 7) They are
used to connect statements p, q, and r. For example, these might be state-
ments of arithmetic such as

x < 2, x + 3 = 7, 3 < 4, 3 < 2.

Consider a statement p, like the first two here, which involves just one
"variable" x. It corresponds to a set P; namely, to the set P of all those
3. The Propositional Calculus 369

numbers x for which the statement holds true. Then a conjunction of

statements such as p /\ q (for p & q) corresponds to be intersection
P n Q of the corresponding sets of numbers. In this spirit we have the
following table of connectives.

Name Form Reading Boolean Form Name

Conjunction p /\ q p and q PnQ intersection
disjunction p Vq p or q PuQ union
negation 7p notp P' complement
implication p ::;> q p implies q P'U Q

The last connective, implication, is sometimes called material implication

and sometimes written as p =:J q or p ..... q; it may also be read as "If p,
then q". Other connectives can be defined from these; thus "p if and only
if q" is defined as (p ::;> q) /\ (q ::;> p) and is written p <=> q.
However, we cannot rely upon the verbal reading "if and only if' to
explain what we mean. The whole formal use of these connectives can be
given by truth tables which specify the truth value (+ for true and - for
false) of the result when the truth values of the arguments p and q are
given. These tables are

p q p/\q pVq 7p p::;>q

+ + + + +
+ +
+ + + +
+ +

These tables do reflect the intended meanings. From the third column
we can read off the formula p /\ q as a function

/\: {±} X {±} -'? {±}, {±} = {+,-}

on the product with itself of the two element set {±} of truth values;
::;> and V are also such functions, while 7 is a function of one variable.
There are also composite functions, represented by "formulas".
Specifically, certain strings of letters p, q, r, connectives, and parentheses
are called formulas (sometimes weI/formed formulas). They are defined by
recursion as follows
(i) A letter p,q,r, ... is a formula F;
(ii) If F and G are formulas, so are F /\ G, F V G and F ::;> G;
(iii) If F is a formula, so is 7F.
Formulas arise only in these ways. (Of course, we also insert parentheses,
as in F ::;> (G /\ H), to make a formula unambiguously readable.) By
370 XI. Sets, Logic, and Categories

using the truth tables, any formula F involving n letters PI , ... ,pn is
r
represented by a composite function {± -> {±}. Then a formula F is
called a tautology when all the values of the corresponding function are
+. For example, the formula (p 1\ q) ~ (p V q) is a tautology.
Since p <=> q can be defined in terms of the other connectives, we
can also treat formulas involving" <=> ". Some useful tautologies are

(p ~ q) <=> (7p V q), (1)

(p 1\ q) <=> 7(7p V 7q). (2)

The first is the famous definition of material implication: "p implies q"
means "either not p or q". In particular, a false statement implies any
other statement! Philosophers who find this conclusion uncomfortable
have introduced other versions of propositional logic, such as strict impli-
cation or relevance logic. As yet, these variants have not proved helpful in
the formulation of Mathematics.
Since r <=> s is true precisely when rand s are both true or when
rand s are both false, the tautology (1) states that p ~ q, as a function
of truth values, is identical to the function 7p V q; in other words, this
provides a possible definition of ~ in terms of 7 and V. Similarly (2),
which is one of the de Morgan laws, can be used to define 1\ in terms of
7 and V. These two latter connectives thus would suffice for this calculus.
(It is even possible to replace these two by a single binary connective, but
the result is not illuminating.)
The propositional connectives correspond (according to the first table)
to Boolean operations on sets; more exactly, on subsets of some fixed set
V. In this correspondence, a tautology becomes a Boolean function with
value V for all arguments, while an equivalence (a tautology F <=> G)
becomes an identity in Boolean algebra. Any complete system of axioms
(there are many) for Boolean algebra can be turned into a system of
axioms (with rules of inference) for the propositional calculus. We do not
need such a system, since the consequences of the axioms can be
described directly and simply as "all tautologies".
Thus everything about propositions p turns on truth-but this is not yet
a complete explication of that weighty word!

4. First Order Language

To describe Mathematics objectively, we must describe a "language" in
which all Mathematics can be written down. The quantifiers "there exist"
and "for all", as we have seen, are necessary in any language formulating
Mathematical concepts. Each language is intended to discuss some
domain of individuals using variables x,y,z, . .. , in adequate supply, each
4. First Order Language 371

ranging over the domain. Formulas of the language are built up from
variables and constants (0, I, ... ) using some primitive predicates and
primitive function symbols; we consider as an example the case of a
language L(B,+) with one binary predicate B(x,y) and one binary func-
tion symbol +. On this basis, one first defines terms:

(i) Each variable and each constant is a term;

(ii) If + is a binary function symbol and sand t are terms, then s + t is
a term.

Terms are obtained only in this way. From the terms one builds up for-
mulas by a similar recursion:

(iii) If B is a binary predicate symbol while sand t are terms, then B(s,t)
is a formula;
(iv) If F is a formula, so is 7F;
(v) If F and G are formulas, so is F V G;
(vi) If x is a variable while F is a formula, so are (V x)F and (:3 x )F.

Formulas are obtained only in these ways. The last clause (vi) is usually
employed when F is a formula involving the variable x. This variable is
said to be bound in the formulas (V x)F and (:3 x)F and in any further
formula built up from these. Variables which are not bound are said to be
free. The intent is that a formula F with free variables y,z,w, ... is a state-
ment of some property of the elements of the domain represented by
these free variables, while the property involves other "bound" variables
ranging over the whole domain. For example in a formula such as
[(Vx)B(x,y)] V B(x,z) the first x is bound while the last is free; it is
usually good practise to rename the latter (say as w).
The formulas as defined above use only the two propositional connec-
tives 7 and V; others can be introduced by their definitions, as indicated
in the last section. We have described a language L(B, + ) with just one
binary predicate B and one binary operation +; if B were <, it could be
called a "language of arithmetic". There are languages with several predi-
cates, unary, binary or n-ary; there can also be more function symbols,
unary, binary, ternary, ... or even none. It is a striking observation about
actual Mathematical systems that the primitive functions involved are
almost all unary or binary (successor or product), while the primitive
predicates B are usually unary or binary. There are occasional exceptions,
such as the ternary betweenness relation in the foundations of geometry
(Chapter III); even then one gets rid of "betweenness" as quickly as possi-
ble by defining it in terms of the "less than" relation for real number
coordinates. On the other hand, there is clearly no way in which binary
relations or functions could all be replaced by unary ones. In philosophi-
cal terms, everything cannot be reduced to properties (unary predicates)
of things, as was the apparent intent of Aristoteljan logic.
372 XI. Sets, Logic, and Categories

The neglect of ternary and higher operations is curious. Does it lie in

the nature of Mathematics, or is it just a failure of imagination? For
example, affine geometry is intrinsically based on a ternary operation
(p,q,w) f-> (l-w)p+wq which assigns to two points p and q and a real
number w the weighted average-but one usually quickly chooses an origin
(which is not invariant) and so express the average in vector space terms,
where there are only unary and binary operations. There have been occa-
sional attempts to study objects called n-groups, for example a 3-group is
a set with a ternary operation such as (x,y,z) f-> (x- y+z). These n-
groups have not prospered, although it is no trick to define, as above, a
language to fit this case.
A language L, as we have defined it above, is called a one-sorted first
order predicate language, essentially because the quantifiers Yx and 3x
involve variables x ranging over only one domain. A two-sorted language
would have two types of variables, to range over two different domains
(for the case of vector spaces, domains of scalars and of vectors). A second
order language would be one with additional quantifiers on variables
which range over predicates or over sets. The use of this word "variable"
is doubtless a carry-over from the use of this word in calculus, which was
intended to deal with quantities (times and distances) which really "vary".
Here a variable is just a symbol; its chief role is that one can substitute
other symbols (terms) for a free variable.
The important language L( E) of set theory is a first order language
with one sort (sets), no primitive function symbols, and just one binary
predicate E, that of membership. In this language, one frequently uses
bounded quantifiers, in which the variable x in question is restricted to
range only over the elements of some specified set. Such quantifiers may
be written

(Yx E u)F for (Yx) (x E u ::;> F),

(1)
(::Ix E u)F for (3x) (x E u 1\ F).

In most Mathematical discussions (as the reader of this book has been
asked to note) the quantifiers used are all bounded ones. Quantifiers
without a bound crop up chiefly in the higher reaches of set theory (e.g.,
quantifiers over all ordinals).
With the exact description of the set-theoretic language L( E), we now
have an explicit description of the idea of a "property", as this arose in
the Comprehension and Replacement axiom schemes of ZF. A property
of x and y means a formula of the language L( E) with just two free
variables x and y. In using the comprehension axiom for ordinary parts
of Mathematics it usually suffices to consider only formulas with bounded
quantifiers. This may be stated as follows:
5. The Predicate Calculus 373

Bounded Comprehension (BQ). For any set u and any formula F(x) of the
language of set theory in which all quantifiers are bounded, there is a set s
with

XES ¢::> x E u A F(x). (2)

For most Mathematics, the appropriate axioms for set theory seem to
be ZBQC: The Zermelo axioms with comprehension replaced by bounded
comprehension and with choice added. This approach (which has
nowhere been developed in detail) is a pragmatic choice of axioms: They
suffice (but not for definitions like (2.11 )).
Some set theorists assert that they can "see" that the ZFC axioms are
"true" for the sets described in the cumulative hierarchy of §1. Note that
in this hierarchy at stage a + 1 each new set S E P( Ra) is a subset of
Ra, so consists of elements already present in Ra. But in the description of
the hierarchy there is no real explanation of what is meant by a "subset".
If each subset of Ra is to be described by giving a property of its ele-
ments, it seems plausible that the first order formula expressing this pro-
perty should refer only to those sets already at hand and hence should
involve only quantifiers over Ra. On this ground, bounded comprehension
is much easier to "see" in the hierarchy than unbounded comprehension.
(Much the same argument would apply for "bounded" replacement.) This
again illustrates the observation that an intuitive idea (here, the cumula-
tive hierarchy) has more than one formal realization.
Within ZFC set theory, the cumulative hierarchy can be formally con-
structed. Before axioms are at hand, the hierarchy is a Platonic myth,
clearly visible only to those with a sixth sense for sets.

5. The Predicate Calculus

Complete rules of proof can be formulated in each first order language.
For this, we need substitution (make the variables "vary"): If F or F(x) is
a formula in a predicate language which involves a free variable x while t
is a term of the language we write (as usual) F(t) for the formula
obtained from F by substituting t for each free appearance of x.
A tautology of a predicate language is a formula obtained from a pro-
positional tautology T (§3) by replacing each propositional letter p of T
by a formula of the predicate language. Thus

(F(x) => G(x)) => (7G(x) => 7F(x)), (1)

F(x) ¢::> 77F(x), (2)

are tautologies of the predicate language, which are obtained from

(p => q) => (7q => 7p) and from p ¢::> 77p, respectively.
374 XI. Sets, Logic, and Categories

Now we can define proof. A formal proof in a predicate language is a

finite sequence of formulas of that language in which each formula is
either an axiom or results from prior formulas of the sequence by one of
the rules of inference. Here the axioms (actually, they are axiom
schemes) are
(i) All tautologies of the language.
(ii) For each term t and for each formula F(x) with a free variable, all
formulas of either of the forms
« 'Ix )F(x» => F( t), F(t) => (3x)F(x). (3)

The rules of inference are

(iii) Modus Ponens: From F and F => G infer G.
(iv) Generalization: If the variable x does not occur free in F, then
(iva) From F => G(x) infer F => (Vx)G(x),
(ivb) From G(x) => F infer «3x)G(x» => F.
Conventionally, these rules of inference are presented as two-line figures,
with premises above the line and conclusions below. Thus modus ponens
(think of an Aristotelean syllogism) is the figure

F F=>G
(4)
G
while the generalization rules are the figures (x not free in F)

F => G(x) G(x) => F

(5)
F => (Vx)Gx' (3x)G(x) => F

These rules do correspond to ordinary principles of argumentation: What

holds for an unspecified x holds for all x; what follows from the property
G of an unspecified x follows from the existence of any x whatever with
this property G. By (3) and modus ponens, what holds for all x holds for
any term.
An essential point is that a formal definition of proof requires both
some syntax (how to build formulas) and some inference (how to deduce
new formulas from old). From tautologies and the specified rules of infer-
ence one can derive others; for example a second form of modus ponens:
From F => G and G => H infer F => H.
To see explicitly how these rules work, let us prove
( v x )F(x) ¢:::> 7(::Jx )7F(x), (6)

a result which could be used to define the universal quantifier in terms of

the existential quantifier. In one direction the classical tautology
p ¢:::> 77p yields the following proof:
5. The Predicate Calculus 375

7F(x) => (3x)7F(x), Axiom scheme (ii),

7(:3x)7F(x) => 77F(x), Tautology (1) and Modus Ponens,
7(:3x)7F(x) => F(x), Tautology (2) and Modus Ponens
(Second form),
7(:3x )7F(x) => (Yx )F(x), Generalization rule (iva).

The reader may wish to construct his own formal proof of the reverse
implication, using (ivb) and the other axiom scheme of (ii). These illustra-
tions may support the conviction that such formal proofs are always possi-
ble and always pedantic. There have been elaborate demonstrations (for
example, Principia Mathematica in three volumes by Whitehead and
Russell) which show that the usual proofs of Mathematics can be forced
into this mold.
A predicate calculus with equality is often used. This means that there
is a binary predicate "=" together with appropriate axioms: The
reflexive, symmetric, and transitive laws for equality, plus the axiom stat-
ing that "equals may be substituted for equals anywhere". It suffices to
require the latter for the primitive predicates and function symbols, as in

(x = x' Ay = y') => (B(x,y) ¢::> B(x',y'», (7)

(x=x'Ay=y') => x+y=x'+y', (8)

for the language L( B, + ) of §4. Our surprisingly short list of axioms and
rules turns out to be adequate, both practically and theoretically. Practical
adequacy means that all Mathematical proofs can be formulated in this
way-in the appropriate language, with the appropriate special axioms
adjoined. Thus "all" proofs in set theory can be formulated with these
rules and the ZFC axioms in the language L( E). Theoretical adequacy is
provided by Godels completeness theorem, which refers to sentences (for-
mulas with no free variables): A sentence S in a first order language L
can be proved from a list A of axioms if and only if S holds in every set-
theoretical model of these axioms. Here a model of a first order language
L such as L(B, + ) is (roughly) a set M to represent the intended domain of
the variables, a subset Bo C M X M to represent the binary predicate
symbol B and a function f: M X M -> M to represent the binary function
symbol + of the language, all such that the axioms hold for these sets.
The important compactness theorem asserts that if T is a set of sentences
of a first order language L such that every finite subset of T has a model,
then T has a model. It is called the "compactness" theorem because it
can be interpreted as asserting that a certain topological space is compact
(a remarkable example of the utility of "geometric" ideas such as com-
pactness in this part of Mathematics).
Peano arithmetic (§II.2) can be readily formulated in a first order predi-
cate language with equality, one constant 0 and one unary function sym-
bol s (for successor). These specifications already take care of the first two
376 XI. Sets, Logic, and Categories

Peano axioms (i) and (ii) of §1I.2, stating that 0 is a number and that the
successor of a number is a number. Next come the axioms

(V n) 7 (sn = 0), (9)

(V n)( V m)( sn = sm => n = m), (10)

and the aXIOm scheme for induction: For every formula F( n) of the
language, with one free variable n,

( F( 0) 1\ (V n)[ F( n ) => F( sn ) ]) => (v m )F( m) . (11 )

For an adequate arithmetic one needs addition and multiplication of
natural numbers. The definition of these binary operations by recursion
(as in §1I.2) requires a recursion theorem involving set theory. Hence an
adequate Peano arithmetic independent of set theory requires an enlarged
language with binary operators + and X together with, as axioms, the
recursion equations (11.2.1) and (11.2.2) describing these operators. The
statement (11) provides for induction for all those properties of a natural
number n which can be expressed by a formula of the language. (This is
analogous to the use of formulas in the comprehension axiom.) Since a
formula is just a finite well-formed sequence of symbols, there is only a
denumerable number of formulas in this language. Therefore, the presen-
tation of Peano arithmetic provides only for a denumerable number of
proofs by induction!
This observation has the strange result that these Peano axioms have a
"non-standard" model in ZFC. For consider a language with one addi-
tional constant ~ and the following denumerable list of potential addi-
tional axioms:

o< ~, sO < ~, ss(o) <~, sss(o) <~, ...

Any finite subset of this list has a model. Therefore, by compactness, the
whole list has a model-and in that model of the Peano axioms there is a
"natural number" ~ larger than any sn(o)!
An alternative view embeds Peano arithmetic in set theory. A model is
then a set N with an element 0 E N and a function s: N ~N which
together satisfy (9) and (10) and which have the induction property: For
every subset TeN,

oE TI\(vn)(n E T => sen) E T) => T= N. (12)

In this case, the strength of the induction axiom depends on the set
theory-on how many subsets T are there provided. With ZF or ZBQ set
theory one can then prove the recursion theorem of §II.2 and from this
show that any two such sets N are isomorphic. Located in this way within
a set theory, Peano arithmetic has no non-standard models.
6. Precision and Understanding 377

Some other parts of Mathematics can be readily formulated in the first

order predicate calculus. This is the case, in geometry, for the first four
groups of axioms (Incidence, Order, Congruence, and the Parallel Axiom)
in Hilbert's foundation of plane Euclidean geometry (§1II.2). However,
the formulation of the fifth group (The continuity axiom) uses sets, either
in Hilbert's completeness form (can be embedded in no larger systems;
i.e., no larger set) or in the Dedekind cut form (the cut is a set). Hence it
is convenient to consider the Hilbert plane within set theory as a set of
points and a set of lines, with their various predicates having the proper-
ties stated in the axioms of Chapter III.
The conventional axioms for the real numbers are also formulated
within a set theory; specifically the Dedekind completeness axiom requires
that every bounded set of real numbers has a least upper bound. The
alternative but equivalent Cauchy or Weierstrass completeness axioms
(§IV.5) also involve sets in order to describe what is meant by a sequence.
For group theory there is also an evident first order language with
equality, with variables ranging over the group and with one binary
operation, that of multiplication. This language suffices to formulate the
axioms for a group, but the subsequent development of group theory soon
goes beyond the language. It uses integers (the orders of group elements)
and sets, both cosets and sets upon which the group operates. For groups
and for the other parts of abstract algebra and topology the free use of
sets is a major convenience.

6. Precision and Understanding

Proof in Mathematics is both a means to understand why some result
holds and a way to achieve precision. As to precision, we have now stated
an absolute standard of rigor: A Mathematical proof is rigorous when it is
(or could be) written out in the first order predicate language L( E) as a
sequence of inferences from the axioms ZFC, each inference made
according to one of the stated rules. Here "written out" need not mean
only in the parsimonious language L( E ); all sorts of symbols for derived
concepts may appear, since each occurrence of such a symbol can be
replaced by its (suitably precise) definition. To be sure, practically no one
actually bothers to write out such formal proofs. In practice, a proof is a
sketch, in sufficient detail to make possible a routine translation of this
sketch into a formal proof. When a proof is in doubt, its repair is usually
just a partial approximation of the fully formal version. What is at hand is
not the practice of absolute rigor, but a standard of absolute rigor.
Moreover there are different versions of the standard. There are alter-
native axiomatizations of set theory: ZF, ZFC, Z, ZBQC, or others. The
first order predicate calculus might be expanded by using quantification
over properties to give a second order calculus. An advocate of strict
378 XI. Sets, Logic, and Categories

implication might prefer one of the many systems of model logic. An

intuitionist could assert that an existence statement such as (3x )F(x)
means that one has exhibited such an x; hence this existence statement is
not logically equivalent (as would follow from (5.5)) to 7(vx)7F(x):
The latter statement would give the existence of x only as the result of a
contradiction, using p <==> 77p. Hence the intuitionist gives up the
tautology p <==> 77p, but the standard of rigor remains. There is a
fully formal intuitionist logic, complete with appropriate axioms and rules
of inference. The idea of rigor remains applicable, here and in all the
other variants, except for certain advocates of "constructive" Mathematics
who claim that their views cannot possibly be reduced to any formalism.
One might also except those geometers who hold that future geometric
insights will transcend any present logic. In any other variant of logic, the
test for the correctness of a proposed proof is by formal criteria and not
by reference to the subject matter at issue.
Historically, much of this concept of rigor was already present in Greek
geometry. The full development of rigor came in successive steps in
modern times, with the 19th century rigorous understanding of calculus,
the development of complex analysis, the careful study of the calculus of
variations, the Hilbert foundation of plane geometry, the subsequent
extensive use of axiomatic methods, the formulation of symbolic logic
(Frege, Whitehead-Russell) and the axiomatization of set theory by Zer-
melo. Note that Whitehead and Russell used a theory of "types" (classes,
then classes of classes, and so on) rather than a set theory. Russell's first
publication of his theory of types came in the same year, 1908, as
Zermelo's publication of his axioms for sets.
For the concept of rigor we make a historical claim: That rigor is abso-
lute and here to stay. The future may see additional axioms for sets or
alternatives to set theory or perhaps new more efficient ways of recording
(or discovering) proofs, but the notion of a rigorous proof as a series of
formal steps in accordance with prescribed rules of inference will remain.
Moreover, there are good reasons why Mathematicians do not usually
present their proofs in fully formal style. It is because proofs are not only
a means to certainty, but also a means to understanding. Behind each
substantial formal proof there lies an idea, or perhaps several ideas. The
idea, initially perhaps tenuous, explains why the result holds. The idea
becomes Mathematics only when it can be formally expressed, but that
expression must be so couched as to reveal the idea; it will not do to bury
the idea under the formalism. For example, proofs of the Pythogorean
theorem (§III.l) require formal concepts (area or similarity) but they also
serve to explain why the square on the hypotenuse is a sum....
The proof of Cauchy's integral theorem by rectangular paths (§X.6) is
an example. On the one hand it is rigorous; it replaces a vague slicing up
of areas (Stokes' lemma) by a precise use of rectangular axis-parallel
paths. To be sure the use of rectangles is accidental; much the same proof
7. Godel's Incompleteness Theorems 379

could be done, and has been done, by using triangles instead. However,
the rectangles are a clear vehicle for exposing the underlying reason: that
a holomorphic function has at each point the same derivative in every
direction; in particular, in the two directions exhibited along axis-parallel
rectangular paths used to construct a primitive for the given function. This
is the underlying idea-but it is not enough to wave one's hands about the
idea. Reduced to formal statement, it becomes a proof. Proofs serve both
to convince and to explain-and they should be so presented. Rigorous
proofs serve to avoid errors-an important function.
But proofs have their limitations, as we will now see.

7. Godel's Incompleteness Theorems

If naive set theory were so formalized as to allow the construction of
Russell's paradoxical set R, it would allow the proof of both R EO Rand
7(R EO R); this is a contradiction. In general, a contradiction in a formal
system arises when there is a sentence p of the system and proofs of both
p and 7p. In view of the tautology p 1\ 7p => q, the presence of such
a contradiction allows the proof in the system of any sentence q. Hilbert
aspired to obtain a secure foundation for Mathematics by proving that a
suitable formal system (including at least arithmetic and analysis) is con-
sistent in the sense that it has no contradictions. Such a result would be a
theorem about proofs in the system; that is, a Mathematical theorem in
which the objects of study are not sets or points or numbers but (formal)
proofs. Such a study is called proof theory or metamathematics (The
Mathematical study of Mathematics). It is possible only in the presence of
a fully formal notion of inference and proof. Hilbert intended that this
metamathematics itself should use only "finite" methods, which therefore
would be wholly secure.
In 1931 Kurt G6del, using a subtle diagonal argument, proved an
"incompleteness" theorem which showed that Hilbert's objective can not
be reached, except perhaps by some novel extension of the Hilbert idea of
finite methods. G6del considered a formal theory T which contained ordi-
nary arithmetic, which is consistent (in a strong sense to be explained) and
in which the axioms and rules of inference are either finite in number or
are specified in a recursive way; the latter is the case for all the systems
we have considered. In this system T he showed how to construct a sen-
tence G such that neither G nor 7G could be proved within the system.
Such a G is then undecideable. Its existence is G6del's first incompleteness
theorem.
Since the rules of T are recursive, G6del also showed that one can for-
mulate within the system T a sentence conT which, when interpreted,
states that "T is consistent", he then showed that this sentence could not
be proved within the system. In other words, no such system T is strong
380 XI. Sets, Logic, and Categories

enough to establish its own consistency. This would apply in particular to

ZFC set theory, and shows that the (hoped for) consistency of ZFC can-
not be proved with the means of ZFC. It is this second G6del incomplete-
ness theorem which blocks the Hilbert program. (All "finite" methods can
presumably be formulated within ZFC.)
We will now give a very informal sketch of the proofs of these two
G6del incompleteness theorems. They depend on a coding by "G6del
numbers" which translate statements about proofs of T to statements
about numbers within T. Since T contains ordinary arithmetic, it has con-
stants such as 0, 0' = 1, 0" = 2 for numerals, variables x,y,z, ... for
numbers, equality, a successor function, and symbols + and X for addi-
tion and multiplication, as well as the apparatus of the first order predi-
cate calculus. By a subtle use of the Chinese remainder theorem, in Tone
can define "prime number" and prove in T the fundamental theory or
arithmetic: Every natural number is uniquely (except for order) a product
of primes.
Metamathematics must discuss formulas and proofs in T. These formu-
las (and their proofs) are all finite formal expressions in a denumerable
list of basic symbols, so the collection of all formulas and all proofs is
denumerable. By explicitly enumerating them in some code, we can
replace statements about proofs by statements about the corresponding
numbers, their codes. Here is a code. First label the basic symbols (in
some order) by the primes:

7, V, 3, 0, +, X, x, y, z

2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31.

Each formula is a finite list of symbols, so is coded by a list of numbers

nl,n2,"'; this can be replaced by a single number m = 2nI3n25n3 ....
Conversely, give a natural number m, one can factor it into primes and so
read off the corresponding list of basic symbols. Usually this list will be
meaningless, but sometimes it will represent a well-formed formula. When
this is the case, m has a certain property which we write, in informal arith-
metic, as

form ( m) = "m is the code of a formula of T".

This property is a (complicated) arithmetic statement: "Factor m

and ... ", so there is also within the system T a formula F(x) in a free
variable x which expresses "exactly" this property. (We return below to
the meaning of "exactly".) (Note, however that in the formula F(x) each
quantifier V k or :3 I has a numerical bound, as all k < m or there
exists I < m. For example, to test that m is prime we consider only possi-
ble factors I < m.)
7. Gi:idel's Incompleteness Theorems 381

We also consider proofs in T. Each formal proof is a finite sequence of

formulas, coded by a sequence of numbers m],m2, .... We tum this
sequence into a single number n = 2m'3m15m3 . .. called the code of the
proof. Conversely, given n one can factor it to get the exponents mj and
the corresponding expressions, and then test whether these expressions are
well formed formulas which in that order constitute a proof. When this is
so n has a (complicated) informal property Prf( n ) which is recursive, since
the system T was so given. Hence Prf( n) can also be translated into the
formal system as a statement P(y). Now select a variable w of the formal
system and a corresponding informal variable h. The translation process is
continued as in the following table:

About T Within T
Form (m) = "m is the code of a F(x)
formula",
Prf(n) = "n is the code of a proof', P(y)
Dem(m,n) = Form(m) and Prf(n) and D(x,y)
the proof ends at the
formula with code m.
Sub(m,k) = The code of the following S(x,y) (a term)
formula "Take the formula
(if any) with code m; in it
replace the variable w by
the kth numeral".
Arg( m,n) = Dem(Sub( m,m ),x ) A(x,y) = D(S(x,x),y)
For all n, not Dem(Sub( h,h ),n) (v y)7A(w,y) codep

Now the last formula in the right-hand column, like every formula of the
system, has a code p; let Jf be the corresponding numeral. Plug in p for h
and Jf for w to continue the table as follows:

For all n, not Dem(Sub(p,p ),n) (v y)7A(p;y) = G.

Call the last formula on the right G. It is obtained by taking the formula
just above it in the table, with code p, and substituting therein the
numeral Jf for the variable w. Hence by the definition of Sub above, G is
the formula with code Sub(p,p). However, its translation on the left
states "There is no proof for Sub(p,p)"; that is, there is no proof for G. In
other words, G when translated back to the left says "There is no proof
for me" (in the system).
Our further examination of this startling situation will use the fact that
the translations above (left to right) are "good" in the following sense:

Translation Principle. When a statement such as Arg(m,n) holds in infor-

mal number theory for numbers m and n, then for the corresponding
382 XI. Sets, Logic, and Categories

numbers iff and if there is in T a formal proof for the translation A(fff,if).
Moreover, if Arg(m,n) fails, then there is a formal proof for 7A(fff,if).
This is to apply only to statements such as Arg( m,n) in which quantifiers
like (V k) or (3 k) have numerical bounds; say all k < m.

Lemma. For every number k the informal statement

Arg(p,k) = Dem(Sub(p,p),k)
does not hold.

For suppose it did. This displayed informal statement says that the
number k is the code for a proof of G, whose code is Sub(p,p). In other
words, there is a formal proof (code k) for G. But by the translation prin-
ciple applied to our informal statement there is also a formal proof for
A(p,k) and hence a proof for (3y)A(p,y)-namely, k is the y claimed to
exist. But (3y)A(p,y) is (classically) equivalent to 7( V y)7A(jJ,y),
which is 7G. So there is also a proof of 7G. Contradiction. Since the sys-
tem is consistent, there can be no such contradiction.
Given this lemma, suppose there is a proof for G. That proof has some
code k. Since G has the code Sub(p,p) this means that Dem(Sub(p,p ),k).
By the lemma, this is not so.
On the other hand, suppose there were a proof for 7G; that is, for
7(vy)7A(p,y) and hence for its equivalent (3y)A(p,y). Now accord-
ing to the lemma, for each number k, Dem(Sub(p,p),k) does not hold.
Hence, by the second half of the translation principle, there is a proof for
the negation of the translation; that is, for 7A(p,k) where k is any
numeral. In other words, for the numerals 0, 0', 0", and so on we have
proofs of
7A(p;0), 7A(p;0 '), 7A(p;0") , ... ,
while at the same time we have a proof for 7G; that is of

(3 y)A(p;y).
In other words, this y, proved to exist, can be none of the numerals
0,0 ',0", ... of the formal system. This would be a bizarre situation. It is
excluded if we assume that the formal system Tis w consistent: That there
is no formula H(y) in one free variable and proof of all of

7(vy)H(y); H(O), H(O'), H(O"), ....

(The above case is that where H(y) = 7A(p;y).) The "strong con-
sistency" mentioned above in our first discussion of Godel's first incom-
pleteness theorem is precisely this notion of w-consistency. (There is a
subtle refinement of the construction of G which renders this stronger con-
sistency assumption unnecessary.) This completes our sketch of Godel's
first incompleteness theorem.
8. Independence Results 383

For Godel's second incompleteness theorem one must formalize the

notion of consistency. As we have already noted, any statement q (for
example, 0 = I) can be proved in a system which is inconsistent. Hence
T is consistent if there is no proof of 0 = I, and this becomes a statement
ConT in informal arithmetic which has a translation into a sentence C T of
the formal theory. Now our proof of the first theorem showed that if T is
consistent, there is no proof of G. But "T is consistent" translates to CT ,
while "there is no proof of G" translates to G. Hence, translating the proof
of the first GOdel theorem gives in the system T a proof of

CT :::;> G.

If T could prove its own consistency there would also be a proof in T of

CT. These two and modus ponens yield in T a proof of G-and this we
know is impossible. Therefore a consistent and recursive formal system T
cannot prove its own consistency. This is Godel's second theorem.
This proof heavily uses Sub(p,p); it is thus a "diagonal" argument.
Our discussion above is a simplified summary of the presentation of
Godel's theorems on pages 204-223 of S. C. Kleene's "Introduction of
Metamathematics". That source gives more detail, including a proof of
the translation principle (there called the property "numeralwise expressi-
ble").
The undecideable sentence G is also "true" -because it says "there is no
proof of G". It might thus be added to the system as another (very com-
plicated) axiom-but for this enlarged system T G one could construct a
new undecideable sentence G'. One is thus left with the striking fact that
every such formal system can state sentences which it cannot decide.
The instructions for constructing G are quite indirect; if they were car-
ried out explicitly, the result would be a very long formal sentence,
involving many intercalated references to prime factorizations. Considered
just as a sentence of arithmetic, G is thus not very interesting. Recently J.
Paris has constructed a true and relevant sentence of Peano arithmetic
which is not provable in that arithmetic; for an exposition, see the article
by J. Paris and Leo Harrington on pp. 1133-1142 of the Handbook of
Mathematical Logic (Barwise [1977]).
The Godel theorems are closely related to a well-known theorem of
Tarski, which asserts that the truth (of sentences of arithmetic) cannot be
defined within the language of arithmetic. These theorems provide a pro-
found commentary on the nature of Mathematics and its formalisation.

8. Independence Results
For Zermelo-Fraenkel set theory many sentences-including a number of
interesting ones-are independent of the axioms and so remain undecided
and undecideable. One such is the continuum hypotheses CH, already
384 XI. Sets, Logic, and Categories

mentioned: There is no cardinal number between the cardinal of the

integers and that of the reals. It turns out that CH can be neither proved
nor disproved in ZFC: It is independent of ZFC.
The second half of this result was due to Godel: CH is consistent with
ZFC; that is, its negation cannot be proved from ZFC. To show this, he
started with a model V of ZFC and built up within it a smaller model, the
model L of "constructible" sets; he then showed that L satisfies ZFC and
CH.
Godel's constructible sets [Godel, 1940] are built up in an ordinal
hierarchy like the cumulative hierarchy of § 1. In that case, given the set
Ra of all sets of ordinal rank a, the next step consists in forming the set
Ra+ 1 = P(R a ) of "all" subsets of Ra. But here the notion of "all" possi-
ble subsets of a given set Ra is not wholly clear. How is subset defined?
Perhaps the subsets should be those obtained by some use of the
comprehension axiom. The constructible hierarchy is built by taking as
subsets only those given by "essential" uses of this axiom in producing
subsets of a set B already present. Given B, enlarge the language L( E ) of
set theory by adding to the language one new constant symbol b' for each
element b E B. Use the evident interpretation of formulas F in the
enlarged language and consider only formulas with quantifiers ranging
over B. Now call a set S "define able" from B if S can be described as

S = {x I x E B 1\ F(x)}

for some such formula F( x), with just one free variable x, in the enlarged
language. (Each such formula will involve only a finite number of the
added constants b '.) The collection D(B) of all these subsets S c B is the
set of sets defineable from B. It is a subset of the power set PCB).
The constructible hierarchy La is now defined by recursion over the
ordinals a as

Lo = 0, La+ 1 = D(La), Ly = U Lp, (1)

p<y

where y denotes a limit ordinal. This is parallel to the definition (1.7),

within ZFC, of the cumulative hierarchy Ra; comparison at once shows
that La C Ra for every a. By a careful analysis one can show that the
class L of constructible sets (all those in some La) does satisfy ZFC and
CH. Thus CH is consistent with ZFC.
Subsequently Paul Cohen established the other half of the indepen-
dence of CH. (CH cannot be proved from ZFC.) It was known, via the
important Lowenheim-Skolem theorem, that any model M of ZFC must
contain a denumerable submodel Mo. Starting from Mo, Cohen adjoined
a "generic" set u from M, using it and Mo to build up a larger model MI
containing Mo and u and satisfying ZFC. The assurance that this added
set u was indeed "generic" (had no properties other than those specified)
8. Independence Results 385

was achieved by a method called "forcing". By a subtle argument, it then

could be shown that the larger model Ml did satisfy ZFC but not CH.
CH is thus an example of an "interesting" statement about sets which is
independent of the axioms of ZFC. Similarly, the axiom of choice is
independent of ZF.
The method of forcing has been used to establish many other indepen-
dence results. One striking such result concerns Lebesgue measure, as it
is used to construct the Lebesgue integral of real analysis (§VI.ll). Using
the axiom of choice, one can prove that there are sets of real numbers
which are not Lebesgue measurable. By using forcing, one can construct a
model of ZF (and hence a model of the reals within it) which does not
satisfy choice and in which every set of reals is Lebesgue measurable.
There are also surprising independence results in algebra. One such con-
cerns a problem of lH.C. Whitehead. Let F be a free abelian group (i.e.,
a direct sum of copies of the additive group Z of integers), while g: A-.F
is an epimorphism of abelian groups. By the axiom of choice, g must split
(that is, there is a homomorphism h: F -.A with gh the identity; this h
makes F a direct summand of A). To construct h, send each free generator
x of F to a chosen element a E A with ga = x. Whitehead's problem
now reads: If H is an abelian group such that every epimorphism
g: A -. H with kernel Z splits, is H necessarily free? On first sight, this
seems to be just a question of algebra. However Shelah, using forcing, has
constructed models of set theory which show that it is independent of
ZFC. (For an exposition, see Eklof [1976].)
These results, and others too numerous to mention, show that many
interesting Mathematical questions cannot be settled on the basis of the
Zermelo-Fraenkel axioms for set theory. Various additional axioms have
been proposed, including axioms which insure the existence of some very
large cardinal numbers and an axiom of determinacy (for certain games)
which in its full form contradicts the axiom of choice. This variety and the
undecideability results indicate that set theory is indeterminate in princi-
ple: There is no unique and definitive list of axioms for sets; the intuitive
idea of a set as a collection can lead to wildly different and mutually
inconsistent formulations. On the elementary level, there are options such
as ZFC, ZC, ZBQC or intuitionistic set theory; on the higher level, the
method of forcing provides many alternative models with divergent pro-
perties. The platonic notion that there is somewhere the ideal realm of
sets, not yet fully described, is a glorious illusion.
This situation bears some resemblance to that in geometry after the
discovery of consistency proofs for non-Euclidean geometry showed that
there was not one geometry, but many. This meant that geometries could
be formulated with many different systems of axioms, some of which were
relevant to higher analysis and some to physics (as with non-Euclidean
metrics in relativity). Similarly, the initial idea of a collection leads to sub-
stantially different versions of set theory, some of which (e.g. forcing, see
386 XI. Sets, Logic, and Categories

§ 14 below) have relevance to other parts of Mathematics, though not yet

(?) to Physics.
Moreover, it is thus appropriate to consider alternatives to set theory as
a foundation for Mathematics.

9. Categories and Functors

One way to organize algebra is to look not just at the objects at issue (sets,
groups, or rings) but also at the mappings between them: Functions
between sets, homomorphisms between groups or between rings, and gen-
erally "arrows" between "objects". This approach also fits topology, where
the arrows are continuous maps of spaces, and geometry, with arrows the
smooth maps of manifolds. It leads to the notion of a category.
A category C consists of objects A,B,C, . .. and arrows f,g,h, .... Each
arrow f has an object A as domain and an object B as codomain; one then
writes

f:A -?>B. (1)

To each object A there is an identity arrow lA: A ->A. If the domain of an

arrow g: B -> C is the codomain of f there is defined a composite arrow

(g-j): A -?> C. (2)

These primitive terms are subjected to two axioms. The associative law:
Given arrows

f: A -?> B, g: B -?> C, and h: C -?> D (3)

the two resulting triple composites are equal, as in

h·(g-j) = (h'g)I A -?> D. (4)

The identity law for the identity arrows asserts that for each f: A ->B one
has

f·lA = f = IBI A -?> B. (5)

This whole description could be viewed as a definition within set

theory, in which case a category C is to be regarded as a set of objects and
a set of arrows such that (1)-(5) hold.
Alternatively, this is an independent set of axioms in a first order
language with two sorts, "objects" and "arrows".
There are many models of this simple system of axioms. Here are some
categories, in each of which "composition" is the standard composition of
maps:
9. Categories and Functors 387

Set: Objects are sets, arrows are functions,

Grp: Objects are groups, arrows, homomorphisms of groups,
Ab: Objects are abelian groups with homomorphisms as arrows,
Vect: Objects are vector spaces (over R), arrows linear transformations,
Euclid: Objects are inner product spaces, arrows orthogonal transforma-
tions,
Top: Objects are topological spaces, arrows continuous maps,
Man: Objects are Coo-manifolds, arrows Coo-maps between them.
All these, and many more, are "big" categories. They do not legally
exist within ZFC because the collections of all the objects and all the
arrows in question are not sets but are "classes", like the class of all sets
or the class of all ordinals. These categories also have "small" versions,
such as the category of all finite subsets of N or of all manifolds in Rn.
But note in particular that in each of these categories each function (or
map) is not just a set of ordered pairs of elements, but is also equipped
with both a domain and a codomain. Thus the identity N ~ N and the
inclusion N ~Z count as different arrows. Good reasons for this conven-
tion, will soon appear, in (7) below.
There are many categories which are small (i.e., legal sets). Each group
G is a category in which there is only one object, while the arrows are the
elements of G with group multiplication as composition. Each partially
ordered set S and more generally each pre ordered set is a category (a
preordered set S is one with a reflexive and transitive binary relation
s < t). This category S has objects the elements of S and one arrow s ...... t
exactly when s < t: the transitive law for the order then provides
uniquely for the composition of arrows, while the reflexive law provides
the identity arrows. Conversely, any category in which there is at most
one arrow with given domain and codomain is a preordered set. In partic-
ular, each ordered set is a category, thus the ordinal number
4 = {O,I,2,3} is the category with four objects and the following arrows

~l~
(identities omitted).
o~ / '

--3~
The homotopy classes of paths in a topological space (§X.S) form a
category under composition of paths. In this category every arrow is inver-
tible; that is, has a two-sided inverse. A category with this property is
called a groupoid.
Big categories arise with each new description of Mathematical objects.
Given a type of Mathematical structure, what are the corresponding mor-
ph isms preserving these structures? They are the arrows of a category. The
slogan "What are the morphisms?" applies to categories themselves. A
388 XI. Sets, Logic, and Categories

morphism F: C --> D of categories (usually called a functor) must send

objects to objects, and arrows to arrows, preserving domains, codomains,
identities, and composites. Thus a functor F: C --> D assigns to each object
C of C an object FC of D and to each arrow f: C ..... C' of C an arrow
Ff: FC ..... FC ' of D in such a way that Flc = IFC and
F(g-j) = (Fg).(Ff) whenever the composite g-j is defined. Thus a func-
tor produces a "picture" of C in D:

C FC

Such diagrams as these, in which any two of the arrows or composite

arrows from one vertex to another one are equal, are called commutative
diagrams.
Functors typically arise from many constructions of Mathematical
objects; one gets a functor when the construction also applies suitably to
the morphisms. An example is the tangent bundle T*M of a smooth man-
ifold M. Here any smooth map f: M ..... M I carries tangent vectors at a
point pin M to tangent vectors at the point fp in M ' , thus giving a map
f*: T*M ..... T*M' which is also smooth. Labeling this latter map T*f
specifies the "tangent bundle" functor T* on the category of smooth man-
ifolds to itself.
The fundamental group '17\ is another example. A pointed topological
space X is one with a selected point xo, called its base point. The
corresponding morphisms (X,xo) ..... (Y,yo) are then continuous maps
f: X --> Y which preserve base points in the sense that f( xo) = yo; the
resulting category is called Top •. Now the construction in (§X.5) of the
fundamental group '17\(X,xo) at Xo provides to each such space a group
and to each such morphism a morphism '17\(/): '17\ (X,xo ) ..... '17\( Y,Yo) of
groups, which sends each closed path h in X into the closed path fh in Y.
Thus '17\ is really a functor from pointed spaces to groups. It provides an
algebraic picture, by groups and their homomorphisms, of spaces and
their continuous maps. Since topology is concerned not just with the
spaces themselves but also with the maps between them it can make
effective use of such functors, from topology to algebra. There are many
other such functors: homology groups, cohomology groups, K-groups, and
the like.
This example also illustrates why it is useful to require that each map
or arrow have a specified codomain. Consider a closed circular disc D and
its circumference, a circle S \. Then the identity function lsI and the
inclusion function j,
9. Categories and Functors 389

(7)

have the same effect on each individual point of S I and so (as functions)
are the same sets of ordered pairs. However, if we apply the fundamental
grouf functor 'IT I (with any chosen point of S I as base point) we get
'lT1(S ) = Z and
'lT1(lS) = 1z: 'lT1(SI) -+ 'lT1(SI) = Z,
the identity, while for the inclusion j, 'IT I (j) collapses Z to the unit ele-
ment since 'IT I(D) is the trivial one-point group. Thus the functor 'lT1 has
very different effects on the two arrows of (7); these two functions differ
only in their codomain, but that makes a difference of substantial topolog-
ical interest!
Some functors turn the arrows around. For a vector space W over a
field K the dual space (§VIA) W* consists of all the linear transformations
f: W -'>K. Hence if the arrow T: V -'> W is a linear transformation of
spaces each vector f in W* yields by composition with T a vector
I T: V -'> Kin V*. If we write T*f for this vector I T, while if S: U -'> V is
another linear transformation, then
(T.S)*f = (S*· T*)f,
and also lw*f = f. Thus (T·S)* = S*·T*, so that composition of linear
transformations is reversed under the operation "take the dual". This
operation, written * or V ~ V*, T ~ T*, is thus what is called a contra-
variant functor (on the category of vector spaces over K). A formal
definition of contravariance can be avoided by introducing to each
category C an opposite category COP. This has the same objects as does C
and for each arrow f: A-'>B of C an arrow fOP: B-'>A in COp, with the
same identities and a composition defined (when possible) by
jP.goP = (g.f)°P. A "contravariant" functor G on a category C to a
category D is then just an (ordinary or "covariant") functor Cop -'>D. For
example, the cotangent bundle construction T*M on manifolds M is a
contravariant functor, because a smooth mapping M -'> M' takes each
cotangent vector on M' (determined by some smooth function on M ')
back into a cotangent vector on M. .
In any category C two objects A and B determine the collection

hom(A,B) = {f I f is an arrow of C and f: A -'>B}. (8)

It is called a "hom set", though it might be too big (e.g., a class). C is said
to have small hom sets if these classes are all sets; this is the case even for
all the big categories we have listed above. For each object B in such a
category A ~ hom(A,B) provides a contravariant function COP -'>Sets.
Similarly, A ~ hom(B,A), often written hom(B,-), is a covariant func-
tion on C to Sets.
390 XI. Sets, Logic, and Categories

The open sets U in a topological space X form a partially ordered set

under inclusion and hence a category open(X) with arrows V ..... U the
inclusions. For example, for U open in the complex plane the set H( U) of
all functions holomorphic on U is a contravariant functor on this
category; any inclusion V c U of open sets yields, by restriction of holo-
morphic functions, a map H( U) ..... HV). The statement that H is a functor
means exactly that the composite H( U) ..... H( V) ..... H( W) is H( U) ..... H( W)
whenever W eVe U. Similarly, a presheaf P on any topological space
(§VIII.lI) is just a contravariant functor
P: (Open X)OP ---'7 Sets. (9)

Understanding Mathematical operations leads repeatedly to the forma-

tion of totalities: The collection of all prime numbers, the set of all points
on an ellipse, the manifold of all lines in 3-space, the manifold of all posi-
tions and velocities of a mechanical system, the set of all subsets of a set,
the set of all power series expansions for a function (its Riemannian sur-
face) or the category of all topological spaces. There are no upper limits;
it is useful to consider the "universe" of all sets (a class) or the category
Cat of all small categories as well as CAT, the category of all big
categories. This is the idea of a totality, and these are some of its many
formulations. After each careful delimitation, bigger totalities appear. No
set theory and no category theory can encompass them all-and they are
needed to grasp what Mathematics does.

10. Natural Transformations

Our slogan proclaimed: With each type of Mathematical object, consider
also the morphisms. So, what is a morphism of functors; that is, a mor-
phism from F to G where both F and G are functors F, G: C ..... D between
categories C and D? An example is already at hand in the category V of
all vector spaces V over a fixed field K. Here the formation of the double
dual V f-+ V** is a (covariant) functor V ..... v. With the evaluation map
e: V* XV ..... K each vector v E V gives a function e( -, v): V* ..... K which
is an element of the double dual V**. Hence there is for each Va linear
transformation

Kv: V ---'7 V**; (1)

said to be "natural" (§VII.4) because independent of any choice of basis.

Now think of K as a transformation of the identity functor V f-+ V into
the double dual V f-+ V**. For each linear map T: V ..... W one readily
shows from the definitions that the following diagram commutes
10. Natural Transformations 391

"v
v----~· v'*

r r**
(2)

w----~.. w**
"w
in other words, we get the same result when K is applied before or after T.
We will adopt this property (2) as the general condition under which a
transformation such as K between functors is a morphism of functors (in
the standard terminology, a "natural transformation").
If F, G are two functors C --> D a natural transformation

'T: F ~ G

is thus taken to be any rule which assigns to each object A of C an arrow

'TA: FA -->GA of D in such a way that the diagram

TA
A FA---~"CA

J FJ CJ
(3)

B FB---··CB

commutes for every arrow f: A-->B ofC, in other words Gj-'TA = 'TBoFf
for all f. One can usefully enlarge the picture implicit in this diagram.
The given functors F and G carry each commutative diagram
(A -->B-->C--> ... ) in C into a commutative diagram in D. The natural

(4)

transformation 'T "translates" the first picture on the left into the second,
on the right, so that the whole diagram is commutative.
For example, if H is the functor giving the set H( U) of all holomorphic
functions g on the open set U C C, then the operation g ~ dg/dz, take
the first derivative, is a natural transformation H --> H between functors on
the category open(C).
392 XI. Sets, Logic, and Categories

Now that we have morphisms of functors, we can form a category of

functors, once we note that the composite of two natural transformations
7: F --- G and (J: G --- H is a natural transformation (J. 7: F --- H. More expli-
citly, given categories C and D the junctor category DC has

objects(DC) = {F I F is a functor C ~ D}, (5)

arrows(D c ) = {7 I 7 is natural 7: F ~ G}. (6)

If we regard these as sets, they can be illegitimately big. In particular, for

each category C we can construct SetC, the category of all set-valued func-
tors on C. It is big (at least as big as Set) but not dangerous. It is, rather,
useful; mapping C into Se{ by a hom functor embeds C in a category
with better properties!

11. Universals
The language of categories allows one to give a common description of
certain standard constructions, such as products and pullbacks, which are
used in many different explicit categories. This common description usu-
ally has the form: A certain arrow to or from the constructed object is
"universal" among all possible arrows with some desired property.
First consider products. The "cartesian" product of two sets X and Y is
normally described as are the cartesian coordinates in the plane: It is the
set of all pairs <x,y> of elements x E X and y E Y. This description at
once yields two functions, the projections p: <x,y> I-> x and
q: <x,y> I-> y on the factors X and Y, as in the usual geometric figure
of the product of two intervals:

XXy
(1)

x ________ p~l_______
This also allows the construction of functions into the product X X Y
from separate functions into X and Y. Thus, given two functions

f:Z~X, g:Z ~ Y (2)

from a common domain Z, there is a function h: Z ---X X Y which sends

z E Z into <jz,gz> EX X Y. This function h can also be characterized
in terms of j and g, but without mention of elements z, as that unique
function h: Z ---X X Y with p·h = j and q·h = g.
This last property uses arrows (and their composition) and not ele-
ments, so can be stated in any category as follows.
II. Universals 393

Definition of product. A product of two objects X and Y in a category C is

an object P and two arrows, p and q,
p q
X~P-7Y (3)

such that for each pair of arrows f: Z ..... X, g:Z ..... Y with a common
domain Z there is a unique arrow h: Z ..... P with p·h = J and q·h = g. In
other words, there is a unique h which makes the following diagram com-
mutative:
Z
I
I
I
I
I
I
f Ih g
(4)
I
I
•
I
X 4 P • y
P q

The arrows p and q are the projections of the product, while a category is
said to have p(oducts if there is such a product diagram (3) for every pair
of objects of the category.

This definition has the form: Among the diagrams of the shape (3) for
given ends X and Y, the product diagram (3) is "universal" -every other
such diagram (with middle term Z) maps into the product diagram (3) via
a unique arrow h. As a result, this property determines the product P and
its projections p and q "up to an isomorphism". This means that if there
is in C another diagram
p' q'
X ~ P' -7 Y (5)

which is also a "product" of the given objects X and Y in this sense, then
there is an invertible arrow 0: P' ..... P with pO = p' and qO = q'. (The
universal property of P gives 0; that of P' gives a two sided inverse for 0.)
For the category of groups, we already noted in (V.S.2) that the usual
direct product of two groups is a product in this universal sense. The same
construction by pairs of elements yields the product for abelian groups.
Similarly the (categorical) product V X W of two vector spaces over the
same field consists of pairs <v,w> of vectors with termwise operations;
this product is commonly called a "direct sum". For two topological
spaces X and Y, the set-theoretical product X X Y has a topology in
which the open sets are arbitrary unions of products U X V with U and
V open in X and Y, respectively. This definition of the product topology is
calculated precisely to produce a topology in which both of the projec-
tions p and q are continuous. Thus Top has products.
394 XI. Sets, Logic, and Categories

In all of these cases, the product object consists of pairs of "elements";

in other words, for each type of structure, the product of two sets-with-
structure is formed by putting a structure (of the type at issue) on the
product of the underlying sets. There is a reason for this, which we skip.
(The functor sending set-with-structure to the underlying set has a left
adjoint.)
In any category, binary products (those with two factors) yield by itera-
tion products with any finite number of factors (and described by the
appropriate universal property). In Grp, Set, and Top there exist also prod-
ucts of any (infinite) set of factors. The interested reader may be able to
check that the (Tychonofl) topology on the infinite product of topological
spaces is again chosen exactly so as to make all the projections on the fac-
tors continuous.
Pullbacks are like products. Let there be given three objects X, Y, and
Z and two arrows u, v which form a sort of "comer"
y

x u
l'
~Z
(6)

If the objects are sets, take the set P of all those pairs <x,y> with x E X
and y E Y for which u( x) = v(y); thus P has the evident projections into
X and Y. If the objects are spaces, take the same pairs <x,y>; they form
a subset of the product space X X Y and inherit from it the "subset"
topology. In either category, P fills out a commutative square
p -----'q---~ y

(7)
Pj " j'
x-----·z
with the following universal property: For every other commutative
square

g
w • y

Ij
x u
•z
j' (8)

built on the same comer (6) there is a unique arrow h: w. . . p with

p·h = f and p·h = g; that is, a unique arrow which renders commuta-
tive the diagram
II. Universals 395

w
\~\
\
g

I • y

fx/~
(9)

.1'

In other words, the object P is the universal way of completing the corner
(6) to a square. In any category, a P with this universal property is called
a pullback or a fibered product, written P = X Xz Y. (This notation is
incomplete, since P is not determined just by the objects X, Y, and Z, but
also by the arrows u and v.) One also says that W is the pullback of Y
along u, or is obtained by "changing the base" from Z to X.
Pullbacks are omnipresent. If u and v are inclusions of subsets of Z, the
pullback P is their intersection. If v: T*M -'>M is a tangent bundle, while
u: N -'> M is the inclusion of a sub manifold, the pullback is the tangent
bundle to N.
An object T in a category C is called a terminal object if to every object
X of the category there is a unique arrow X -'> T. As for products, it fol-
lows from this universal property that a terminal object T in C is unique,
up to an isomorphism. In the category of sets, anyone-point set is termi-
nal (and of course any two such sets are isomorphic). Similarly the group
with only one element and the topological space with only one point are
terminal in their respective categories. Given a terminal object T, products
can be constructed from pullbacks, because the product X X Y is just the
pullback X X T Y formed from the necessarily unique arrows X -'> T, Y -'> T.
In fact, given the (binary) pullbacks P and a terminal object, one can con-
struct pullbacks P for all sorts of fancier finite diagrams

Such a pullback is an object P with an arrow p to each object of the

diagram which makes all composites from P equal. Such a pullback is
caHed a finite limit.
Each description of a universal object by means of arrows can be "dual-
ized" by writing the same description with all arrows reversed. Thus an
initial object I (the dual of a terminal one) in a category C is an object
such that there is a unique arrow I -'>X from I to every object X. The
empty set is an initial object in Sets; the one element group is an initial
object in groups. A coproduct of two objects X and Y is a diagram
396 XI. Sets, Logic, and Categories

j
X ---?- Q <E- Y (10)

such that to every similar diagram f: X ...... W <- Y: g there is a unique

h: Q...... W withf = h·i and g = h-j, as in the commutative diagram

x • Q • y
I
I
I
:h
I (11)
f I
g
I
I
I
t
w

The arrows i and j are called the injections of the coproduct. The coprod-
uct of two sets Sand T is their disjoint union, as described in (11.8.1). The
coproduct of two vector spaces V and W is their product (a direct sum)
with the evident inclusions. The coproduct of two groups is their free pro-
duct (§V.8). In particular, the free product of two infinite cyclic groups is
the free group on two generators, which appears in topology as the funda-
mental group of the space consisting of two tangent circles (Fig. X.5.4).
The dual of a pullback can also be formed in these categories; it is
called a pushout-or, in the case of groups, an amalgamated product.
We now tum to the construction of "exponentials". If m and n are
natural numbers, the power mn can be described as the number of
different functions on a set of n things to a set of m things. Hence we
write Z Y for the exponential set or the function set consisting of all func-
tions f: Y ...... Z:

zY = {t I t: Y ---?- Z}. (12)

This function set can also be characterized in a "universal" way by using

a familiar process which turns a function f(x,y) E Z of two variables
x E X and y E Y into a function F(x) of one variable x with values
F( x) which are functions of the second variable y. The formal definition
of this F reads

(F(x))(y) = f(x,y), x E X, Y E Y; (13)

here each value F( x) is a function of y, hence an element of the function

set Z Y, so that F: X ...... Z Y. This is a sort of inference (a bijection) from f to
F

f: X X Y ---?- Z
(14)
F: X ---?- ZY
I I. Universals 397

which is reversible: For each F as in the lower line one may construct the
corresponding f in the upper line by using the same formula (13), read
backwards.
In order to describe the bijection f I-> F as a universal process, we need
to express formally the left hand side of (l3) in which the function
t = F(x) is "evaluated" at the argument y. This uses the evaluation func-
tion e,

e: ZY X Y ~ Z, (t,y) I-> t(y), Y E Y. (15)

Then e( F( x ),y) = (F( x) )(y), so the function F described in terms of ele-

ments x as in (l3) can be described by arrows as follows: To each f there
is a unique F which renders the diagram

zY x Y

Fx·1/
e .. Z

(16)

XXY
commutative. As an equation, this reads f = e( F xI), with 1 = 1Y the
identity arrow of Y. This equation is just a different statement of (l3), not
using elements.
This form of the description of exponentials applies to any category C
which has products. The category is said to "have exponents" if to every
pair of objects Z, Y there is an object Z Y and an arrow e: Z Y X Y ..... Z
such that, whatever the arrow f: X X Y ..... Z, there is a unique arrow F
which makes the diagram (16) commute. In other words, the evaluation
arrow e is universal among arrows f from a product.
Our description has also indicated that the exponential Z r, defined in
this way, exists for any two objects Z and Y in the category of sets. This
holds also in the category of vector spaces (or of abelian groups): Just
take Z Y to be the vector space (or abelian group) of all linear transforma-
tions (or of all homomorphisms) Y ..... Z. However, exponentials do not
exist for all spaces in the category Top; in other words, there is not always
a "good" way to define a topology on the function space Z Y. This obser-
vation has actually led to the suggestion of a suitable restriction (or
perhaps an enlargement) of the notion of a topological space, so as to pro-
vide a "convenient category" of such modified spaces, in which products
and the corresponding exponentials always exist.
The description (16) of the exponential has another formulation, this
time in terms of functors. First observe that in a category C with products
(of objects) one also has products of f X g of arrows, because given
f: X ..... X' and g: Y ..... Y' there is by (14) a unique arrow f X g which
makes the following diagram commutative
 398 XI. Sets, Logic, and Categories

x---,,-P_- xx q- - y
y---
I
I
I
I
f fXgl g
I (17)
I
I
+
X'--p-,-,- - X' x Y' - - q - '- - Y'

In this way the product X X Y is really a functor (of two variables X and Y).
In particular, for each fixed Y, the operator - X Y is a functor of one
variable, the "product with Y", sending X to X X Y and g: X ~X' to
g X 1: X X Y ~X' X Y. Moreover, ZY for this fixed Y is a functor of
Z. The correspondence (14) above states that the functor X f-> X X Y
uniquely determines the functor Z f-> Z Y by the bijection f f-> F. Here
F: X~ZY is an element of the "hom set" hom(X,Z\ while
f E hom(X X Y,Z). Thus the bijection (14), expressing f in terms of F,
can be written as a bijection of hom sets (cf. (9.8»

hom(X X Y,Z) == hom(X,ZY). (18)

Both sides are functors (of the arguments X, Y and Z) to sets, and this
bijection is a natural isomorphism of functors. One says that it makes the
functor ( )Y the right adjoint of the functor - X Y, while the latter func-
tor is a left adjoint (because it is on the left side of the hom set in (18».
Generally a functor P: C ~ X is left adjoint to a functor G: X ~ C when
there is a natural bijection

hom(FC,X) == hom(C,GX) (19)

between functions COP X X ~Sets. There are many examples of adjoint

functors; each universal construction leads to a pair of adjoint functors,
and conversely.

12. Axioms on Functions

The standard "foundation" for Mathematics start with sets and their ele-
ments. It is possible to start differently, by axiomatizing not elements of
sets but functions between sets. This can be done by using the language
of categories and universal constructions. We will now sketch this
approach, in order to emphasize again the observation that the leading
ideas of Mathematics have a variety of alternative formal expressions.
The intent then is to set up an adequate axiom system for functions and
their composition. The functions will be arrows between objects which are
(perhaps) sets, but the axioms will not refer to any elements of these sets;
they will be formulated strictly in terms of objects, arrows, their domains,
12. Axioms on Functions 399

codomains and their composites, as well as identity arrows. The axioms

will thus be stated in a first order language with equality and with two
sorts, objects and arrows.

Axiom I (Category). The objects and arrows form a category.

Axiom II (Finite Limits). In this category there is a terminal object 1 and

a pullback for each "corner" X ~Z <- Y.

This axiom provides for the number 1 and the product of two numbers,
since pullbacks and a terminal give products, as noted in §11. One could
also get zero and addition of numbers by an axiom requiring the existence
of an initial object and pushouts. We do not do this because these proper-
ties turn out to be consequences of the other axioms to follow.

Axiom III (Exponents). For any two objects there is an exponential.

Next we need to handle subsets. Now a subset S of X can be considered

as an inclusion SeX and hence as an arrow m: S ~x. Such inclusions
are injections; they can be characterized as "monic" in the following way:
An arrow m: S ~ T is monic if, for every pair of arrows j,g: S ~ X,
m·j = m·g implies j = g; this is just like the description with elements
x, y of an injective function m: mx = my implies x = y. One might
object that the usual inclusion map is not only an injection, but sends
each element of the subset S to "itself'. This we cannot say. Instead we
call two monics m: S ~ X and m I: S I ~ X equivalent if and only if there is
an invertible arrow 8: S ~ S I with m'· 8 = m, as in the diagram
s

x. (1)

s'
Equivalent subobjects of X will be represented by the same arrow from
X, by taking the arrow to be like a characteristic function 1/;. These func-
tions, much used in probability theory, take only the values 0 and 1; I/;x is
1 when x is in the subset S at issue and 0 otherwise; it thus expresses a
basic fact about a subset S: Either XES or not (x E S). Formally, in
set theory the characteristic function I/;s of a subset SeX is defined by

I/;s(x) = 1 XES,
(2)
= 0 x $. S.
400 XI. Sets, Logic, and Categories

Moreover from the characteristic function I/; one can reconstruct the sub-
set S: It consists of exactly those elements x E X which land on I under
1/;. This means that S is the pullback of the inclusion i: {I} c {a, I}
under 1/;:
s --------+- { I }
I
I
I
I (3)
I
I
+ -.jJ
x - - - - - " {O, I}
This description can be purged of all "elements" by using the "universal"
description of the pullback and by replacing the traditional set {O,l} of
two truth values by an arbitrary object n, as follows.
Axiom IV (Subobject Classifier). There is an object n and an arrow
T: 1 ~n from the terminal object I such that every monic m: S ~X is the
pullback of T along a unique arrow 1/;:
s------+- I

m r •
(4)

x -------n
-.jJ

In this axiom, the arrow S ~ 1 must be the unigue arrow from S to the
terminal 1. Also, since 1 is terminal, any arrow (such as T) from 1 must be
monic. Hence the axiom states: There is an arrow T from I which is
"universal" among monics, in the sense that every monic arrow is a
unique pullback of T. Observe that axioms II and III are also statements
about the existence of universals (i.e., of adjoints).
These four axioms are powerful; a category which satisfies them is
called an elementary topas (elementary because the axioms are first order;
a "topos" because every topological space carries such a topos (of
sheaves; as will appear)). For example, arrows 1 ~ Y act "as if" they were
elements of Y (they would be, in the category of sets). Also X is the prod-
uct X X 1, so there is an isomorphism X X I == X. Hence, starting with
a monic we get successive one-one correspondences to other arrows:
mOllIc,
characteristic function,
X == 1 x X,
exponential law.
In other words, equivalent monics S -4 X correspond to "elements" of nX.
Thus ax is an object which is the "set" of all "subsets" of X, so that we
have here a categorical description of the power set W = PX.
12. Axioms on Functions 401

As in the axioms for set theory, one needs an axiom to assure the
existence of something infinite. In set theory, one axiomatized the proper-
ties of the successor function on finite ordinals. Here we axiomatize the
fact that the successor function, with 0, provides definitions of other
arrows by recursion:

Axiom V (Existence of a "Natural Number Object). There is an object N

and arrows 0: I ..... N, s: N ..... N such that to every diagram I ~ X ~~ X
there is a unique arrow f which makes the following diagram commute:

0 .. N .. N
I I
I I
I I
I I
If (5)
If I
I

,
I I
I I
I
t
x
.. x .. x
h

This diagram (cf. (11.2.5» is just that defining the function f from h by
the recursion

jO = x, f·s = h·f·

Note that this axiom is also a universal construction.

These axioms already allow the development of a good deal of
Mathematics, but they are substantially more general than classical set
theory, chiefly because they do not require that there are only two truth
values, 0 and 1. The logic which they implicitly involve can be intuitionis-
tic, because double negation need not be the identity. This means that the
algebra of subobjects of X (i.e., elements of nx) need not be Boolean.
In set theory, the axiom of choice is equivalent to the statement that
each epimorphism e: X ..... Y has a cross section; that is, a function s: Y ..... X
with e·s = 1. Indeed, the epimorphism e defines for each y E Ya non-
void subset of X, the "inverse image" e - I {y}, and the cross section s
"chooses" in each such set an element sy. In any category, an arrow
e: X ..... Y is said to be epic if, for all arrows f,g: Y ..... Z, fe = ge implies
f = g. Then we can state choice without elements as follows:
Axiom VI (Axiom of Choice). For every epic e: X ..... Y there exists an
s: Y ..... X with e·s = 1.

This axiom implies that all the algebras of subobjects are Boolean.
In this approach to Mathematics the objects X do not have elements. As
noted already, the arrows I ..... X might be regarded as elements of X, but
402 XI. Sets, Logic, and Categories

there may not be "enough" of them to make all the usual arguments. In
set theory two functions f,g: X ..... Yare different only when there is an ele-
ment x with fx i= gx where they differ; here there may be two different
arrows f,g: X ..... Y but no arrow x: I ..... X where they differ; that is, where
f·x i= g·x. One can require a topos to be more like classical set theory by
asking that it be "well-pointed" in the following sense (where 0 designates
an initial object, which can be constructed from the previous axioms).

Axiom VII (Well-Pointed). The objects 0 and I differ. Moreover, to any

two distinct arrows f,g: X ..... Y there is an arrow x: I ..... X with f·x i= g·x.

It is now possible to develop almost all of ordinary Mathematics in a

well-pointed topos with choice and a natural number object (Axioms 1-
VII). The development would seem unfamiliar; it has nowhere been car-
ried out yet in great detail. However, this possibility does demonstrate one
point of philosophic interest: The foundation of Mathematics on the basis
of set theory (ZFC) is by no means the only possible one!
One can also make a direct comparison between axiom systems. Start-
ing with sets given by ZFC, the conventional description of functions
f: X ..... Y as sets of ordered pairs does give a category which is a well-
pointed elementary topos; indeed, this is the process which we have used
to motivate much of the discussion above. Conversely, it is more complex
to construct sets from the topos axioms, because one constructs sets as
trees. More exactly, one can define when a binary relation on an object of
a topos produces a "tree" like the E -tree described in (1.6) above for a
set. The resulting trees (sets) satisfy the Zermelo axioms in a form close
to ZBQ (bounded comprehension) but do not necessarily satisfy the
replacement axiom. This emphasizes again that bounded comprehension
suffices for most of Mathematics.

13. Intuitionistic Logic

Examples of toposes include situations with more than the usual two truth
values. For consider the category D with objects X the functions
t: Xo ..... X I between two sets Xo and XI. An arrow in D is a pair of func-
tions fo and fl such that the following diagram commutes:
fo
Xo----""-----O Yo

(1)
13. Intuitionistic Logic 403

(we use the same letter t for the functions in the two objects X and Y).
Composition of these functions Ji does define composition in D. One may
think of an object X of D as a set "varying with time t"; at the start it is a
set Xo which changes in time to a set XI; thereby different elements of Xo
may coalesce (xo =1= x' 0 but txo = tx' 0) and new elements may appear.
This category is an elementary topos; that is, it satisfies axioms I-IV
above. For example, the product of two objects X and Y is just the object
tXt: Xo X Yo ..... X I X Y I , while an arrow m: S ......X is monic in D if
and only if both mo: So ..... Xo and ml: SI .....X I are injective, as in the ver-
tical arrows of the figure

mr-~--Sj~l (2)
xo_ _ _ _
t----x l
this means that So C Xo and SI C XI are subsets, with mo and ml the
injections. So when does an element Xo E Xo belong to the subobject S
of X? It may be that Xo E So (that is, Xo = moso for some element so);
when that is the case, then also txo E S I. Or, it may be that Xo is not in
So but that txo is in S I. Or, it may be that neither Xo E So nor
txo E SI. Thus there are three possible truth values! We can define a
"characteristic function" 1/; for the subobject S as a function on Xo to
{O,l,oo} for Xo E Xo as follows:

1/;(xo) = ° if Xo E So,

1/;(xo) = 1 if Xo $. So but txo E S I , (3)

1/;(Xo) = 00 if Xo $. So and txo $. SI'

To be more complete, we can define an object Q of D as

t: no = {O,I,oo} -7 QI = {O, oo}; to = 0, tl = 0, too = 00 (4)

and a function 1/;1: XI ..... QI by 1/;1(XI) = 0 if XI E SI and 00 if

XI t/::SI, for all XI E XI' This process defines a diagram of four objects
of D in the form

1 m[~~::~~_____________J;r
~ • to}
1

1 (5)

~ xl--------i--------{ 0, oo}
404 XI. Sets, Logic, and Categories

Here {a} -- {a} is a terminal object in D. The diagram commutes and is a

pullback, so that D does satisfy the axiom IV, with the subobject classifier
the object n described in (4). Thus n has three truth values, for "now",
I for "eventually" and 00 for "never". By a quite similar construction one
°
can form an elementary topos E with objects X the infinite strings of sets
and functions
I I I
X: Xo ----7 XI ----7 X2 ----7 X3 ----7 "', (6)
while the subobject classifier n has infinitely many truth values with
n: no = {O,I,2,3, ... , oo} ----7 ...

and the characteristic function 1/;0: Xo --no

of a subobject SeX is
defined by I/;x = n if n is the least integer with tnx E Sn and I/;x = 00 if
tnx is never in the subset Sn. Briefly, I/;x is the "time till truth".
Classically, the subsets S of a set X form a Boolean algebra under the
operations unions, intersection, and complement V\, V, '); in particular
the double complement of a subset S is S itself, S" = S, in parallel to the
tautology 77p = p of the propositional calculus. The simplest Boolean
algebra is the algebra {O,l} of two truth values, and this algebra is the
subobject classifier n for the topos of sets. The three Boolean operations,
as defined by truth tables, are functions
1\ V 7
n X n ----7 n, n X n ----7 n, n ----7 n.
In an elementary topos one can define corresponding arrows on the
subobject classifier n. They satisfy most but not all of the identities
describing Boolean algebra; the resulting algebra is called a Heyting alge-
bra because it first arose in Heyting's formal description of Brouwer's
intuitionism. In particular, in such an algebra, 77 is not the identity.
We do n"'t give a full description, but we do note a geometric example:
The algebra of all the open subsets U of a topological space X is a Heyt-
ing algebra. Here the intersection or the union of two open sets is again
open, by the axioms of topology, so intersection and union are defined,
with the usual properties. However, the complement 7U of an open set
must be defined to be an open set; it is taken to be the largest open set in
X disjoint from U. As a result, 77U may be larger than U. For example,
if X is the real line and U the open subset of all positive reals except I,
then 7U is the set of all negative reals and so 77U, the set of all positive
reals, is larger than U.
Thus alternative logics are closely related to geometry!

14. Independence by Means of Sheaves

There remains one more surprising connection: The independence proof
for the continuum hypothesis, done by the method of forcing, can be rein-
14. Independence by Means of Sheaves 405

terpreted as a construction of a suitable elementary topos by the

geometric means of sheaf theory.
To begin with, that elementary topos D of "sets through time" (§ 13) has
a different description. Take the category C with only two objects 0 and 1
and only one non-identity arrow T: 0 ...... 1. Then a functor F on C to Sets is
a picture of the category C in Sets as
FT
Fa • Fl

T
01 • I •

Xa • Xl

in other words, such a functor is just an object Xo ...... X I of the category D,

while a natural transformation f: F ...... G between two such functors is just
an arrow (fO,/1 ) in D, as described in (13.1). Thus D is just the functor
category Set C , as defined in (10.5) and (10.6).
This is a general situation: If T is a topos, while C is any category, then
the functor category T C is a topos. In particular, since the category of sets
is a topos, so are all the functor categories Set c . The proof is straightfor-
ward, the essential point being the construction of a subobject classifier Q,
much as in the special case of the Q constructed in §13 for the topos D of
sets through time.
Next, a presheaf P on a topological space X, as in §VIII.1I, is just a
contravariant functor

P: (OpenX)OP ~ Sets,

as in (9.9), on the category of open sets of X. Hence the category of all

presheaves on X is a functor category
Set(Open X)OP

and therefore an elementary topos. In geometry, one passes from the

category of all presheaves to the smaller category of sheaves, defined as in
§VIII.II. Much the same passage is possible in any topos if the "topol-
ogy" is replaced by an arrow j: Q ...... Q with suitable properties. The
double negation arrow 77: Q ...... Q has these properties. Cohen's proof of
the independence of the continuum hypothesis can now be formulated as
follows. Start with a model Set of sets and within it take a (carefully
chosen) partially ordered set P. Regard P as a category P and from it
form the functor category ("presheaves"):

This category is a topos, but its logic is not classical, since its subobject
classifier Q is a Heyting algebra and not Boolean. In this topos, take the
406 XI. Sets, Logic, and Categories

"topology" given by the double negation operator and pass to the

corresponding category of sheaves
pop
Sh77 (Set ).

This category is still a topos (this requires a considerable proof) and

because of the use of double negation in its formation, it is classicaL This
means that the corresponding subobject classifier Q is a Boolean
algebra-but it may be very large. Choosing in it a maximal prime ideal,
one can "divide" it down to be two-valued, like {O,I}. This division pro-
cess still produces a topos, which is now a model of set theory. Because of
the careful choice of the poset P one can show that CH fails there.
This cursory summary is designed just to indicate how geometric ideas
about sheaves, initially introduced with Riemann surfaces, also crop up in
foundational studies about sets. In Mathematics, apparently different for-
mal ideas are often closely connected.
Details of this development are presented in Peter Johnstone's Topos
Theory. For aspects oflogic, we refer to Barwise Handbook of Logic.

15. Foundation or Organization?

Set theory and logic provide a standard "foundation" for Mathematics:
Define all Mathematical objects in the first order language of sets and
prove all Mathematical theorems from ZFC and these definitions using
the rules of inference specified in logic. We have indicated that there can
be alternative foundations, one such being that by elementary topoi. Now
in one sense a foundation is a security blanket: If you meticulously follow
the rules laid down, no paradoxes or contradictions will arise. In reality
there is now no guarantee of this sort of security; we have at hand no
proof that the axioms ZFC for set theory will never yield a contradiction,
while Godel's second theorem tells us that such a consistency proof can-
not be conducted within ZFC. Similarly, there is no proof of the con-
sistency of the axioms for a well-pointed topos; we know only that they
are consistent if ZFC is consistent, because we have provided within ZFC
a model for the topos axioms.
Alternatively, set theory and category theory may be viewed as propo-
sals for the organization of Mathematics. The canons of set theory provide
guides to the formulation of new concepts and emphasize the extensional
character of Mathematics: A "property" is completely determined by
knowing all the elements which have that property. Similarly, the canons
of category theory emphasize the importance of considering not just the
objects but also their morphisms. They also emphasize the use of univer-
sal constructions and their associated adjoint functors.
15. Foundation or Organization? 407

Neither organization is wholly successful. Categories and functors are

everywhere in topology and in parts of algebra, but they do not as yet
relate very well to most of analysis. Set theory is a handy vehicle, but its
constructions are sometimes artificial. Moreover, it is clearly far too gen-
eral; as Hermann Weyl once remarked, it contains far too much sand. We
conclude that there is as yet no simple and adequate way of conceptually
organizing all of Mathematics.
Historically, there have been other proposals for the organization of
Mathematics or of parts of Mathematics. For the Greeks, Mathematics
was geometry, and they formulated real numbers and algebraic operations
only in geometric terms. In the 18th century, Mathematics appeared
largely in the development of all the aspects of the calculus; this was a
natural reflection of the wide opportunities this development offered for
formal manipulations and for extensive applications. Subsequently the
extraordinary fruitful properties of holomorphic functions made complex
variable theory a center about which (much of) Mathematics could
revolve. There were competing organizations. In analysis, rigor was
enshrined under t: and 8. In geometry, Felix Klein proposed that the
many varieties of space provided by non-Euclidean and other geometries
could be classified and hence organized in terms of their groups of
symmetries-the full linear group, the orthogonal group, the projective
group, and others. In a way this organization did include complex analysis
as the study of groups of conformal transformations; this approach
amounts to putting a heavy emphasis on geometric function theory in
contrast to analytic function theory and the overuse of t: - 8 methods.
Currently the remarkable properties of group representations and their
use in arithmetic and physics again suggests that group theory provides an
effective organization.
This variety of proposals for organizations reflects the diversity and
richness of Mathematics.
Figure 1 may indicate some of the interconnections of the subjects
treated in this chapter.
408 XI. Sets, Logic, and Categories

Sets, Functions, and Categories

Property Coll"'oo I I F~~

Set of Pairs------.::

Set of Numbers ~ SET - - FUNCTION - Polynomial fcn

Set of Points ~--~"rn, COdo9~Trigooom,"" 'co

n \
Equivalence Class Probability cat~ \ Algebraic fcn

Cardinal Number Axioms Holomorphic fcn

Dedekind Cuts Universals AXilms \ Morphism

I ~
ZFC Adjoints Topos

~ZB~ "S::::fThOOry
Combinatorics---Large Cardinals Forcing

~Infi~ity
Figure I
CHAPTER XII

The Mathematical Network

From our examination of a variety of mathematical topics, we now return

to the underlying questions: What does this variety say about the nature
of Mathematics? How does it illuminate the philosophical questions as to
Mathematical truth and beauty and does it help to make judgements
about the direction of Mathematical research? In particular, what is the
foundation of Mathematics?
Our central observation is this: the development of Mathematics pro-
vides a tightly connected network of formal rules, concepts, and systems.
Nodes of this network are closely bound to procedures useful in human
activities and to questions arising in science. The transition from activities
to the formal Mathematical systems is guided by a variety of general
insights and ideas. Within the formal network, new developments are
stimulated and guided by conjectures, problems, abstractions, and the
constant desire to understand more. These observations may be summar-
ized by a diagram:

Human Activities
Scientific Questions ] [ Procedures
Problems ] - Ideas

Rules
Definitions [ J
Axioms - Network of Formal Systems
Proofs

We now continue to explicate the portions of this diagram.

410 XII. The Mathematical Network

1. The Formal
The presentation of Mathematics is formal: Calculations are done follow-
ing rules specified in advance; proofs are made from previous axioms and
follow predetermined rules of inference; new concepts recognized as
relevant are introduced by unambiguous definitions; errors and disagree-
ments are cleared up not by dispute but by appeal to the relevant rules.
It is characteristic of any formal procedure that it makes no reference to
the meaning or to the applications, but only to the form. The formalism
may be imperfect and sketchy, but it carries with it perpetually the possi-
bility of perfection. Because of these characteristics, Mathematics (within
its limits) is absolutely precise and independent of persons. The formal
can be communicated well, without ambiguity. It develops in many suc-
cessive stages.
The formal arises first in the rules for arithmetic questions. Given two
integers in base ten form, rules tell us how to form a "sum" or a "prod-
uct". The rules are unambiguous; we can tell when they have been carried
out correctly, whether by hand, by abacus, or by computer. Different cal-
culations of the same product, when done carefully, come to the same
answer. The rules make no mention of what the numbers might mean, or
what decimals are-although in a larger system, still formal, say in Peano
arithmetic, such rules can be derived from the Peano axioms, the recursive
definitions of sum and product, and the description of what decimals are.
This austerity of the rules of arithmetic is the basis of their applicabil-
ity. The use of arithmetic is governed by well understood practical opera-
tions. The operation of counting uses the successor of each decimal and
prescribes how to count, say, the number of matches in a pile. Then the
formal process of adding two decimal integers matches the practical pro-
cess of combining two piles. If the combined count comes out wrong, the
failure is never attributed to the rules, but to miscounts or missed
matches. The very generality of the rules and their manifold prior uses
means that accidental error is never laid at their doorstep. Similarly there
are common sense operations to determine areas, say of rectangles: Pro-
cedures to recognize when a shape is approximately a rectangle and then
ways to measure its length and breadth in feet. The rule for multiplication
then provides the area, which again can be checked by a measurement.
Thus an area of 96 square feet can be covered (more or less exactly) by 96
square tiles, 12 inches by 12 inches each. If it turns out that the tiles must
overlap, then some count or measurement must have slipped, or perhaps
the rectangle was really a parallelogram. It is never the rule for multipli-
cation of decimals which is at fault.
In brief, the formal rules of arithmetic are a firm background for the
occasionally faulty operation of counting.
The Peano axioms provide a second step in formalization. Now the
rules for calculation are deduced from axioms, but again without any
!. The Forma! 411

attention to the meaning of these natural numbers, as cardinals or

perhaps as ordinals. Or, going further, one may adopt the von Neumann
definition of ordinal number, as presented in Chapter XI, but then the
argument deriving the rules is a formal one from the Zermelo-Fraenkel
axioms of set theory and need pay no heed to the idea that an ordinal is
the order-type of a well-ordered set. Thus in arithmetic (as elsewhere in
Mathematics) there are many stages of formalization, connecting precisely
with each other.
Axiom systems provide another prominent example of the formal. The
decisive example is the complete axiomatization of plane Euclidean
geometry. Intuitively we think that we are dealing with plane figures com-
posed of points, lines, triangles, and circles. But with careful formulation
of the (Hilbert) axioms for the plane, all the proofs of theorems can be
done strictly from the axioms, as emphasized in §III.2 "proofs without
figures". We do not need to refer to the straight line as something without
breadth (as it may have been defined by Euclid). We use only the
incidence and congruence axioms. This is fortunate; because of it, there is
no need to assume that there are real straight lines out there somewhere
in the world, nor that there really is a flat plane. Here again the austerity
of the Mathematics is coupled with operational rules for its use. Angles
can be measured by sighting through the telescope on a surveyor's transit
at a fixed vertical rod, then turning the telescope and measuring the
"angle" of turn on a protractor built into the transit. Horizontal distances
can be measured with chain and plumb-bob. If the distances fail to
satisfy (say) Pythagoras, the error is ascribed to the chain man, not to the
proof of the Pythagorean theorem. Or, if we carefully set up three points
and measure from each the angle sub tended by the other two, we might
get three angles which do not add to 180°. If so, that is an error of
observation or perhaps an indication that some non-Euclidean geometry is
in order for this application. Euclidean plane geometry and trigonometry
stand firm on the axioms, immune to the deviations of light rays or the
vagaries of surveyors. It is firm in its formal standards for deductions; this
firmness is known by long experience to be fruitful. Plane geometry is a
deductive system, useful in space, but not itself a science of space.
Euclidean geometry also exhibits the rules of inference from logic. A
proof in geometry does not (except perhaps verbally) refer to the truth or
to successive truths, but only to tautologies and axioms, and to effective
applications of rules of inference. And the description of tautologies in
terms of truth tables is not a way of appealing to philosophical truth, but
only a formal rule for substituting O's and 1's in a formula with results
which have proved to be effective. This observation is most relevant for
the case of quantifiers. In the predicate calculus, a statement (:3 x )F( x)
does not really mean: I have exhibited a point p with the property F(p).
This statement (:3 x )F( x) may be inferred from F(p), or it may be
proved by other means; for example, by an indirect proof. So all told,
 412 XII. The Mathematical Network

( ::J X )F( x) is to be viewed as a logistic formula which can be manipulated

according to the rules of generalization suggested by experience and
specified in §XI.5. This observation does not prevent the construction of
other predicate calculi with more restrictive rules for manipulating
quantifiers, as they might be suggested by the intuitionist's ideas about
existence. But both the classical and some forms of the intuitionist logic
can be fully formal; they then simply specify what is meant by a proof.
Axioms typify the formal, not just in geometry and arithmetic, but every-
where where they appear in Mathematics. Thus the real numbers began
as a source of careful proofs about the reals, giving properties consistent
with the operations used in measurement. Similarly, the Zermelo-
Fraenkel axioms for sets are not a description of real collections or of
totalities of objects; they are formal rules for the manipulation of the
binary relation "member of', and long experience has shown them
effective in constructing Mathematical concepts. Of course these axioms
are suggested by experience, in particular by properties of membership in
the intuitive description of the cumulative hierarchy of sets (§XI.l), but
these suggestions do not in any Mathematical way determine the axioms.
For example, they do not determine whether the comprehension axiom
should use bounded or unbounded quantifiers. This is because the hierar-
chy at each stage forms "all" subsets of the set present at the previous
stage, and there is no formal meaning for the notion "subset" here. Also,
the hierarchy does not determine how far the ordinals go or what large
cardinals may exist. The axioms for set theory have the same status as the
axioms of geometry, which are suggested but in no way determined by the
practices of surveyors.
Definitions also exhibit the formal. In principle, in every argument the
definiendum may be replaced by the definiens. Or consider particular
definitions. A function is defined as a certain kind of set of ordered pairs;
then properties of functions are those which follow from the definition.
The definitions have been chosen so that the properties apply under the
standard operations using functions. Thus when x and y represent physi-
cal quantities and y = f(x), then the measurements of y depend upon
those of x. Or, a function h: U ->C defined on an open set U of the com-
plex plane C is defined to be holomorphic when it has at each point a of
U a complex derivative. Among formal definitions, this one is remarkably
well-chosen, as shown by the amazing properties of such functions. How-
ever, we know that the properties hold because they are all strictly formal
consequences of the definition. To be sure, one might alternatively have
followed Weierstrass in defining a holomorphic function f: U ->C as one
represented at each point a E U by a convergent power series in z -a. But
this, too, is a formal definition, and there is a fully formal proof that it is
equivalent to the definition by the existence of the complex derivative.
Moreover the theory of holomorphic functions requires formal definitions
of other concepts: Integral, piecewise differentiable paths, etc. But all
I. The Formal 413

these cases-and their use in trammg young Mathematicians to under-

stand rigor-exhibit again the clarity which is achieved when all the
relevant concepts are defined firmly and formally.
Newtonian mechanics presents another application of the formal. The
law F = rna for "force is mass times acceleration" is incomplete without a
specification of "force". But when the force is specified by the inverse
square law, one can make a completely formal calculation of the orbit of
a planet. There are then observational operations (what to examine with a
telescope) which tell how to check the result of the formal calculation
against results of observation. The case of an earth-bound projectile or
rocket is similar: Given the basic equation, one calculates formally
without looking at the facts what the trajectory should be-and then com-
pares with observations. This comparison is usually not wholly precise,
but this does not invalidate the formal calculations; it may only suggest
further calculation involving suitable formulas for the forces due to the
resistance of the air.
The formal appears clearly in algebra. For finite groups one proves for-
mally (say) that the order of a subgroup always divides that of the group;
this can then be checked by numerical examples of particular groups or
on figures, using the group-theoretic description of symmetry. In geometry
the expert practitioners usually put the pictorial examples and intuition in
the foreground, but the formal character is really there, as for instance in
the case of the intrinsic geometry defined by a Riemann metric on an
abstract surface, while the actual surface sits there as a surface embedded
somehow in Euclidean space. In this way, the real generality of geometry
depends essentially upon the fully formal definitions of the geometric
objects involved.
The ultimate in formality would be the meticulous deduction of all
Mathematical theorems from definitions in terms of sets and the
Zermelo-Fraenkel axioms. This extreme is never done, but its potential
presence serves to define the possible extent of formality. There are
indeed many other cases where it was historically necessary to develop a
meticulous formalization of parts of Mathematics which had been built up
intuitively. One example is the rigorous foundation of the differential and
integral calculus in the 19th century-a process which was important not
only in itself but also in its contribution to the more conceptual approach
to all of Mathematics. Algebraic geometry provides another example. Its
development in the late nineteenth century was lively (England, Germany,
Italy)-but a bit shaky as reflected in the canard that the highest accom-
plishment in algebraic geometry consisted in the demonstration of a gen-
eral theorem and the simultaneous construction of a (special) counterex-
ample to that theorem. The various geometrical insights of this develop-
ment were subsequently formalized in several successive stages in the
period 1925-1975, first by the more systematic use of ideal theory, then
by the introduction of general valuation theory, then by the study of
414 XII. The Mathematical Network

specializations, sheaves, and schemes. (This last also illustrates to a

remarkable extent the observation that an underlying geometrical idea
can have many different formalizations.)
Arithmetic, algebraic, or trigonometric calculations, as we have
described them, tend to be precise calculations of the answer. Many other
problems need or allow only a rough answer. In other cases, the principles
underlying the calculation may be uncertain and the necessary data is at
best approximate. Hence many calculations have the character of esti-
mates. Side effects are neglected, or terms with probable lower order of
magnitude are dropped, or unmanageable formulas are replaced by more
manageable and (hopefully) adequate alternative formulas. Such pro-
cedures can be formal (in part) but are sometimes material, in that the
choice of approximations depends on the scientific meaning of the data.
Situations such as this prevail in many parts of applied Mathematics.
Sometimes estimates in this sense can be fully formalized-notably in the
estimates used in analytic number theory and in many basic parts of clas-
sical analysis. At other times, estimates in this sense seem to resist reduc-
tion to any strictly axiomatic form. Put differently: Some subjects in
applied Mathematics do not contain any theorems or proofs of theorems;
these subjects thus seem to escape an analysis of Mathematics as formal
or formalizable. Thus the boundary lines between formal Mathematics
and its applications are vague and moveable. This is a necessary result of
the theory that some ideas in Mathematics arise from the applications.
To summarize, the formal in Mathematics appears in many stages:
Rules of calculation (arithmetic calculations, trigonometric identities);
Formulas for differentiation in the calculus;
Rules for estimation and approximation;
Axiom systems for arithmetic and geometry;
Rules for logical inference;
Axioms for abstract algebraic and topological objects;
Axioms for sets and topoi;
The relentlessly full formalization of Mathematics.
Hilbert's doctrine of formalism contends that Mathematics is just the
systematic manipulation of formulas. This was a necessary background for
his attempt to provide a formal consistency proof for all of Mathematics;
it should not be confused with extreme versions of formalism, which
regard Mathematics as simply a game with symbols.
We do not take either of these extreme positions but we do assert that
the formal is essential to Mathematics. Mathematics is not a scientific
study of the facts but a developing analysis of the forms which underly
the facts. It is not a science of time and space, but a formulation of the
ideas needed to understand time, space, and motion. This understanding
depends on ideas. The underlying ideas can be made precise and com-
municable only by being made formal, as exhibited throughout our
chapters.
2. Ideas 415

2. Ideas
Most of the formalizations in Mathematics are based on some underlying
idea-an intuitive notion which gives guidance and purpose. It is not easy
to give a precise description of the nature of an idea; indeed a deeper
idea may be almost impossible to communicate and so may be recognized
only after it has been embodied in some formalization. A number of the
more general ideas of Mathematics are stimulated more or less directly by
human activities. In §I.ll, Table 1 we have given a sample of such ideas.
It was noted then (and in many subsequent examples) that an idea may
have several quite different formal realizations. Thus the idea of "mani-
fold" starts as a geometric locus of all the items or points of some concept.
It becomes formal as a topological manifold or as a differentiable mani-
fold (COO, C 1 or just smooth in some sense) or perhaps as a complex man-
ifold. In all these cases" every point of the manifold has a neighborhood
described by a "good" chart, but the original idea of manifold should also
allow in some way for manifolds with singularities (and I speculate that in
the future geometry will see more formal concepts of this sort). In any
event, the "idea" of manifold is nebulous enough that it has many formal-
izations; moreover these different formalizations allow us to study
separately different aspects of the idea.
Mathematical ideas arise not just from human activities or scientific
questions; they also arise out of the urge to understand prior pieces of
Mathematics. A "set" was initially a collection of points in the line or the
plane, and then any sort of "collection", considered as one thing and con-
sisting of well-defined Mathematical objects. Subsequently the "idea" of a
set was described more specifically in terms of the cumulative hierarchy.
Then we have an idea clear enough to be communicated but still fuzzy:
At each stage Ra of the hierarchy one next forms the set of "all" subsets
of Ra -and there is here no formal description of what subsets are meant.
Ideas, as we use the term, are inherently vague. (They are not to be con-
fused with the Platonic idea of the ideal line or the ideal circle or the
ideal set.)
Since we are not able to give any precise description of the term "idea",
we may simply list examples (see Table 1), this time of ideas which arise
from prior parts of Mathematics; again each idea turns out to have
several formalizations. We have already seen other examples: The idea of
curvature in differential geometry, or the related ideas of eigenvalues and
spectrum for self-adjoint operators.
The development of an idea can be long and involved. Thus the study
of change and motion suggests the idea that change can be "smooth"
rather than sudden. Examples of such smooth changes appear both in
mechanics and in geometry. Eventually, this idea "smooth" divides into
the separate ideas of "continuous" and "differentiable." The difference
between these two, however, is clear only after both have been carefully
formulated by means of the descriptions of limits. Differentiability, so
416 XII. The Mathematical Network

Table 12.1. Some Ideas Arising within Mathematics

Origins Idea Formal Versions

Polynomials Function Formal expression (in a language)

sin, tan, log Complete table of values,
Dependent variables Set of ordered pairs,
Arrow in a category,

Velocity, acceleration Rate of change Derivative

Tangent line Partial derivative
Complex derivative

Differential operator Operation Linear transformation

Matrix Binary operation
Arithmetic operations Unary operations

Linear order Relation Binary relation

Partial order Set of ordered pairs
Congruence (numbers, figures) Ternary relation
Betweenness

Prime factorization Decomposition Prime ideal factorization

Partial fraction expansion Product
Components (of a manifold) Coproduct, etc.

described, then appears in the calculus and in the analysis of the local
structure of curves, surfaces, and manifolds; and then eventually (but not
until the 1930's) in the precise global definition of a smooth manifold.
Continuity, once formalized, is similarly seen to apply to functions of
several variables and to functions of curves, thus gradually leading to the
20th century notion of a metric space. Eventually it becomes clear that
the metric notion of distance can be replaced by the more qualitative
notion of neighborhood, as in the definition of a topological space. Thus it
is that the formal properties of open sets eventually serve to codify the
very idea of continuity.
There are also many less general ideas: An idea how one might prove
some desired theorem. Sometimes the idea fails, and sometimes its suc-
cessful execution involves complications or additional technical tricks.
Again, the proof may be just a routine realization of the idea. In any
event, when the formal proof is at hand the original idea (with supporting
detail) has been realized. Ideas require formalizations.
The idea of putting together separate pieces of a function leads on the
one hand to a concept of a sheaf (as in §VIII.lI); on the other hand,
where the pieces are power series expansions, it becomes the concept of
analytic continuation for holomorphic functions.
Our view that Mathematics is and must be formal is supplemented by
the observation that each formalism rests on some underlying or leading
3. The Network 417

idea. The reverse can also happen, as when geometric ideas are formu-
lated so vaguely that they do not constitute a proof of the intended
results.

3. The Network
We cannot realistically constrain Mathematics to be a single formal sys-
tem; instead we view Mathematics as an elaborate tightly connected net-
work of formal systems, axiom systems, rules, and connections. The net-
work is tied to many sources in human activities and scientific questions.
We have already sketched some pieces of such a network, in §V.lO with a
table of functions, transformations and groups, in §VI.12 with a table of
the interactions of the concepts of the calculus, and in §IX.13 with the
table of the Interconnections of Mathematics and Mechanics.
In the full, dense network of parts of Mathematics, there are many out-
side ties. The most basic subjects (the subjects at the "edges" of the net-
work) are tied closely to connections reaching outside of Mathematics.
The primary ties are those to various human activities: Counting, measur-
ing, moving, observing, changing, and others as listed in §I.8. In some
cases, these ties can also be regarded as connections not to activities, but
to phenomena: Multitude (that which can be counted), extent (that which
can be measured), motion (which can be observed), change (which again
can be observed). By way of these ties and connections, mathematics is
grounded in "reality", at least in whatever reality is represented by these
activities and these phenomena.
The subjects of Mathematics are also tied or connected to other parts of
human knowledge, and most especially to the various sciences. Geometry
is tied to mensuration, architecture, surveying, navigation, and, on a more
sophisticated level, to space, time, and space-time as these enter into phys-
ics. Calculus is tied to mechanics, dynamics, and many other parts of
theoretical physics. Differential equations and Fourier analysis are like-
wise tied to physics-as is also vector analysis, complete with all the
sophistication of dual spaces and tensor products. Calculus is also tied to
economics, for example, in the use of marginal concepts in Mathematical
economics. The number of such connections between Mathematics and
science is very great-and often these connections go to subjects which are
in the "middle" of the network of Mathematics, and not just to the basic
subjects at the edge of the network.
These external connections of Mathematics are numerous and tight, but
they do not wholly describe or determine the Mathematical subjects. Basic
Mathematical concepts may be derived from human activity, but they are
not themselves such activity-nor are they the phenomena involved as the
background of such activity. The subjects of applied Mathematics are
418 XII. The Mathematical Network

closely related to branches of sciences-but I argue that they are usually

not themselves science. A scientific theory can be falsified by factual data,
while a Mathematical theory cannot be so falsified. For example,
Euclidean geometry (as we now view it) would not be falsified by mea-
surements (as, for example, measurements of angle sums in some triangle
of light rays). Were such a sum to deviate substantially from 180·, it would
falsify a scientific theory which asserted that rays of light satisfied the
axioms for lines in Euclidean geometry. However, Euclidean geometry
itself could be falsified at most by a contradiction derived from its axioms.
The axiomatic method is a declaration of independence for Mathematics.
I assert that subjects of Mathematics are extracted from the environ-
ment; that is, from activities, phenomena, or science-and that they are
then later applied to that-or other-environments. Thus number theory is
"extracted" from the activity of counting, and geometry is extracted from
motion and shaping. The exact mechanism of this "extraction" has not
been described in detail here; it will clearly vary considerably from case
to case. I have deliberately chosen this work "extraction" to be close to
the more familiar word "abstraction" -and with the intent that the
Mathematical subject resulting from an extraction is indeed abstract.
Mathematics is not "about" human activity, phenomena, or science. It is
about the extractions and formalization of ideas-and their manifold
consequences.
It is the intimate interconnection of Mathematical ideas which is strik-
ing. We have already remarked in Chapter VIII and IX how the geometric
notions of manifold, tangent bundle, and cotangent bundle also arise in
close parallel in mechanics, with configuration space, velocity phase space,
and momentum phase space. In some sense, this connection is a continua-
tion of the way in which calculus and celestial mechanics came out
together in the hands of Newton. Similarly it was Riemann who con-
nected the manifold ideas of the differential geometry of surfaces with the
complex manifolds now called Riemann surfaces. The full development of
these ideas took nearly a century, and has now issued (in part) in the
notion of a sheaf (the sheaf of germs of holomorphic functions) and its
surprising use in "forcing" methods in set theory.
There are many other deep interconnections not covered in our previ-
ous chapters. One important example is a three-way connection between
prime ideals in algebraic number theory, points on Riemann surfaces and
points on algebraic curves. We sketch these items:
An algebraic number y is a root of an (irreducible) polynomial equation

Yn + an-IY n-I + an -2Y n-2 + ... + alY + ao = 0 (1)

with rational coefficients ai' As in the Galois theory (§V.7) one considers
not just y, but all the rational combinations w of Y and its powers; they
form a field K, called an algebraic number field. The degree of the irredu-
3. The Network 419

cible equation (1) for y is also the dimension of the field K, regarded as a
vector space over the ground field Q of rational numbers; this observation
establishes a useful connection with linear algebra. Every number w in the
field K satisfies a polynominal equation like (1) of degree at most n (use
linear algebra!). When there is such an equation with highest coefficient 1
(the equation is monic) and with all the other coefficients ai rational
integers, the number w is called an algebraic integer. The collection of all
the algebraic integers w in the field K forms a commutative ring tJ; indeed
this important example (with polynomials) is a basic reason for introduc-
ing that concept; see §VI.3. In this way the passage from integers to
rational numbers is extended; tJ passes to K:

Z C Q
n n (2)
tJ C K = Q(y).

This is the start of algebraic number theory and its connections with
Galois theory.
The quadratic case (degree n = 2) appears first. For example, if y = i,
this ring tJ is just the ring of all complex integers m + ni with m,n EZ. In
this ring every integer can be factored, and essentially uniquely, into
prime (i.e. irreducible) complex integers. But there is no such unique fac-
torization for many other quadratic fields. And when y is a higher root of
unity (in which case K is called a cyclotomic field) this need not be the
case, as Kummer discovered in trying to handle the Fermat problem.
However, unique factorization can be restored, by factoring algebraic
integers not into these integers, but into "ideal" factors which are exactly
the ideals used in §7.10 as the kernels of ring homomorphisms; this means
that one must define a product AB of two ideals in tJ-as the ideal gen-
erated by all products ab for a EA, bE B. Also, each element w in tJ deter-
mines an ideal, the set (w) of all its integral multiples. Now the algebraic
number w is prime in tJ when it cannot be factored there; that is, when
for all u, v in tJ, uv E ( w) implies U E ( w) or v E ( w). Similarly an ideal P in
tJ is called prime if uVEP implies UEP or VEP. Now we can state the first
main theorem of algebraic number theory: Every algebraic integer w in tJ
has a unique decomposition

(3)

as a product of powers of distinct prime ideals (with integral exponents

ei). Moreover, the fact that an ideal is the kernel of a homomorphism
tJ .....tJ/ P enters here; the familiar field Z / p of integers modulo P has been
extended to quotient rings (fields) tJ/ P.
Now shift gears and consider an algebraic function w of a complex vari-
able z, as defined by some (irreducible) polynomial equation
420 XII. The Mathematical Network

with coefficients ai(z) which are rational functions of z. This function w

then has a Riemann surface over the z-plane-or better, over the whole
Riemann sphere. All the rational combinations of z and ware also ana-
lytic (holomorphic except for poles) on this Riemann surface. With the
complex numbers C they form a field C(z,w), called an algebraic function
field. Now a rational function of z is determined, up to a constant factor,
by giving its zeros and its poles. Similarly, what matters about each alge-
braic function ware the points P on the Riemann surface where w has a
zero or a pole. One may write Vp w for the order of the zero (or, negative,
for the order of the pole) of the function w at the point P. Then
Vpw = ep =1= 0 for only a finite set of points P. This is vaguely like the
factors in the decomposition (3) of ideals. Also, the function w is entire
(has no poles at points over the finite portion of the Riemann sphere) pre-
cisely when the rational coefficients ai(z) in its irreducible equation (4)
are all polynomials. The entire functions in C( z, w) form a commutative
ring tJ, much like the ring of algebraic integers, and we have an array of
four rings
C[z] c C(z,)
n n (5)
tJ c C(z,w);
here C[ z] is the ring of all polynomials in z.
This suggests a parallel between algebraic numbers and algebraic func-
tions. The parallel can be carried out by developing each theory by the
methods first found for the other. On the one hand, ideals apply to alge-
braic functions. The ring tJ of entire functions on the Riemann surface
has unique decomposition of ideals into products of prime ideals P-the
prime ideals correspond exactly to the points (on the finite part) of the
Riemann surface. Moreover the exponent ep to which P appears in the
decomposition of the ideal (w) of an entire function is precisely the order
of the zero of that function at the point P. The remaining points of the
Riemann surface (those points over 00 on the Riemann sphere) can be
captured by the conformal transformation z I-> liz; that is, by using
prime ideals over the ring C[llz]. In short, the points on this Riemann
surface, which we previously treated geometrically and topologically, can
be described algebraically as prime ideals!
The situation can be reversed, using analytic methods for algebra. For
each point P on the Riemann surface the function Vp (for order of zero)
has the formal properties (X. lOA)

and Vp (c) = 0 for any complex constant c. Such a function on a field is

called a valuation, and every valuation of the field C(z,w) arises from a
3. The Network 421

point P in the Riemann surface, and is Vp , up to a multiple. This is still

another description of "point on a Riemann surface".
Now for an algebraic number w, consider the exponent e to which the
prime ideal P occurs in the decomposition (3) of w. This exponent Vp is a
valuation on the algebraic number field Q( w)-and all the prime ideals of
tJ can be recovered from valuations. Moreover, one can mimic the power
series used in analysis. Each valuation V can be turned into a function

Iw I = 2- Vw

like the absolute values used in analysis. A series in powers of z - a,

Co + cI(z-a) + c2(z-a)2 + ...

with complex coefficients Ci E C[z]j(z-a) is replaced by a series

Co + CIP + C2P
2
+ ... (7)

in powers of p, with coefficients 0 <; Ci < p; that is, coefficients in Zj(p).

This series (7) converges when we turn the valuation at the point Pinto
an "absolute value" by setting

because then I w I is small when Vw is large; i.e., when w is divisible by

a large power of p. The "p-adic" numbers, imported from analysis, have
remarkable algebraic uses. In this way analysis and algebra intersect, to
the profit of both.
There is also a connection to plane algebraic curves defined by
polynomial equations like (4) in coordinates z and w. This extensive
approach eliminates the privileged position of z.
Such ideas arise also with diophantine equations 10 two variables
x and y
n
~aijxiyi = 0 (8)
ij=1

with integer coefficients. If we replace x and y by z and w this equation

becomes complex, and so determines a Riemann surface of some genus g
(the genus is the number of handles, so the torus has genus 1). Some time
ago, Mordell conjectured that a diophantine equation (8) with genus
g > 1 would have at most a finite number of rational solutions (x,y).
Just recently, using a vast extension of the methods we have just sketched,
the conjecture has been established by G. Faltings. In other words, the
connection between number theory and Riemann surfaces is indeed both
tight and effective.
Mathematics is a network... and formal concepts abound (ideal,
valuation, Riemann surfaces).
422 XII. The Mathematical Network

4. Subjects, Specialties, and Subdivisions

Since we have described Mathematics as a network, we must specify more
the nodes of that network-the various special "subjects" into which
Mathematics is (or may be) divided. It will turn out that this notion of a
subject-a branch of Mathematics-is both useful and elusive; the mean-
ing of a "branch" will turn out to be various and variable.
Initially, Mathematics may appear to have a simple subdivision, with
the main branches arising from basic human activities, perhaps as in Fig-
ure 1, where the associated activities are enclosed in boxes.
Here we have the five traditional subject subdivisions (underlined) of
Mathematics. Number Theory is described by its subject matter: The
natural numbers or perhaps the integers; the subject matter is essentially
unique, its structure described say by the Peano postulates (considered as
a definitive description, despite the nonstandard models and the true sen-
tences which cannot be demonstrated from these axioms). Algebra is
described by its methods: The manipulation of formulas using letters for
variables or unknowns which stand for indeterminate numbers, usually
real numbers. Geometry is initially a science: The science of space, as it is
described by the axioms of Euclidean geometry; there is only one space,
and Euclid is its prophet. Calculus is the science of variable quantities as
they arise in measurement and are represented by (real) numbers, it is
also the technique of dealing with such quantities by differentiating and
integrating-or, if one presses deeper, by using infinitesimals and infinite
operations; more precisely by using limits, { - ~ and their careful formu-
lation in terms of quantifiers. Applied Mathematics is (at first) largely the
study of particle and continuum dynamics by the methods of calculus; in
return, it raises new problems for the calculus. Thus the study of heat and
its differential equations leads to trigonometric series, while the proposals
for principles of "least action" -teleology in mechanics-are connected
with the calculus of variations.
This first picture of the network is clearly far too simplistic. In the next
version (Figure 2) there are added inputs from "ideas" (shown in
parentheses).

g5Urin~~haPing

7r
Calculus

\'"

Applied Mathematics
.................
-_
Geometry
I"
I"
---- - - Formulas
-Algebra

/,///
/~
"'"
,
'.
Number Theory

Figure 1
4. Subjects, Specialties, and Subdivisions 423

(Geometric Language) (Many Dimensions)

~ctor Spaces ----------=alytic beometry

Analysis Geometry Algebra

~~RC
R"II:'"~'om')-NOnhl'," i/~:::~::~'"~

T \
Complex Analysis (Curvature) Projective Geometry Fields

Ri",~
C Differential Geometry ~ Geometry
Riom"" 'mi'""' ~ Algebraic Number Theory

~ (Manifolds) /7
(Ideals) ""
Analytic Number Theory

Figure 2

Here the subject "Number Theory" continues, including elementary

number theory, the study of Diophantine equations, quadratic forms and
the like. But two more subjects now split off. Algebraic number theory is
described primarily by its subject matter: Algebraic integers and their
properties, but it is closely tied to algebra by its methods. Analytic
Number theory is characterized by its methods: The use of analytic tech-
niques, especially complex integration, to tackle the hard questions about
the distribution of the primes and the whereabouts of the zeros of zeta of
s. (The important line connecting this subject to complex analysis won't fit
on the figure.)
"Algebra", once the manipulation of formulas about unknown reals,
becomes just the application of the rules for this manipulation to other
number systems-integers modulo m or p-adic numbers from algebraic
number theory or power series treated "formally", that is, without requir-
ing convergence. Hence algebra splits into several subjects each of which
is nominally described as the study of all the set-theoretic models of a list
of axioms on the rational operations involved. Thus ring theory is
described formally by the ring axioms for addition and multiplication; it
subdivides into commutative ring theory, coming from algebraic number
theory and from algebraic geometry, and non-commutative ring theory,
which arises from matrices, from quaternions, and from the search for
more division algebras and which also has connections with algebraic
number theory, because the local behavior of primes can be formulated
by certain linear algebras. Thus in this case a subject (ring theory) is pre-
424 XII. The Mathematical Network

cisely delimited by a formal list of axioms, while its connections in the

Mathematical network exhibit its origins and uses.
Field Theory is again formally described by its axioms (all the good
properties of the four rational operations of addition, multiplication, sub-
traction and division). It arises from the Galois theory-how to understand
the solution of polynomial equations in terms of the symmetries of all
their roots and of the field of all rational combinations of these roots-and
this idea was essential to our discussion in §3 of algebraic number fields
and function fields.
Back at the top of the figure (and historically prior to abstract algebra)
comes "analytic geometry", a subject which can be described as a
method: The remarkable method of so using Cartesian coordinates to
reduce problems of plane and solid geometry to routine algebraic manip-
ulation. This method is well known and widespread; it has been claimed
that the one notion of modern Mathematics familiar to the man in the
street is the idea of coordinates: The tenth floor at 329 West 15th Street.
And this reduction of geometry to the manipulation of coordinates inevit-
ably suggests the idea that things dependent upon many inputs (mechan-
ics, statistical mechanics, etc.) can be described in geometric language, as
if they were things spread out in a "higher dimensional" physical space.
The standard technical formulation of this idea comes in the subject of
vector spaces (also called linear algebra or matrix theory). Its substance is
determined by its many uses (e.g., eigenvalues): its form is determined by
its simple axiomatics: A vector space is a module over some field. It is
thus legally subsumed under the subject Module theory (not on our
figure): A module (§VII.l2) is a ring together with an action of that ring
on an abelian group, and module theory is a necessary complement to
ring theory. In short, the geometric vision is supported by firm and formal
algebraic structure.
The discovery of non-Euclidean geometry and its models showed
decisively that geometry is not a science of actual space but an axiom-
based study of space-like configurations; historically the axioms for
geometry are the very model of what an axiomatic approach can be. The
various non Euclidean geometries can be subsumed under projective
geometry, which has extra-mathematical motivation from the study of
perspective; the perhaps parallel subject of "descriptive geometry" has
never really been recognized as a branch of mathematics, and projective
geometry itself is not now an active field of research. However, the curves,
surfaces, and varieties defined in projective spaces as the loci of polyno-
mial equations in the coordinates form the subject matter of the lively
field of algebraic geometry, even though these varieties are now described
in a context of "schemes" -spaces more abstract than the original projec-
tive spaces. The subject is a striking example of the confluence of
geometric, algebraic, and topological ideas.
Our figure also exhibits the subjects of real and complex analysis-
subjects defined essentially as the study of well behaved functions on
4. Subjects, Specialties, and Subdivisions 425

specific domains (R or C). Here the initial ideas of the calculus (rate of
change, the whole is the sum of its parts) have an extraordinary and for-
mal development. Through the ideas of tangent lines, curvature and tor-
sion there is a deep connection with geometry: Differential geometry as a
subject can be defined either as an application of calculus to geometry or,
more intrinsically, as a study of smoothly curved space-which need not
have an Euclidean context.
Our Figure 2 does not reflect the study of functions or the pervasive
aspects of group theory; this was pictured in part in the Table of §V.10. In
principle, group theory can be described formally as the study of all the
set-models of the (very simple) group axioms. This does not recognize the
variety of interconnections of group theory with other subjects, Figure 3 is
a start, beginning of course with the basic examples of symmetry. Group
theory as a subject quickly splits into different branches: Finite group
theory (which is close to number theoretical ideas) and infinite (combina-
torial) group theory. The addition of the commutative law produces the
subject of abelian groups, which has a quite different flavor and may not
be as "deep" as finite group theory. However abelian groups are extraor-
dinarily useful as topological invariants: they and modules lead to homo-
logical algebra (for example, in abelian categories). In a different direc-
tion, groups with a smooth structure (Lie groups) or with a topology are
closely tied to analysis. Finally the representations of groups by linear
transformations ties group theory back to some of its origins in geometry
and has remarkable further connections: As carriers of symmetry, such
representations can be used to predict the varieties of elementary particles
in Physics.
Figure 2 omitted one side: The connections of analysis with classical
applied Mathematics. Here the various fields are tightly connected to the

Crystals Ornaments Finite Groups

Linear Transformation

~
(Composition) (Multiplication)

Differential Equations
/ Mod\,"
I
Lie Groups - - - - - Groups Abelian Groups

aW"p R'P~'""~~OmOl"""l AI",b"

Topological Groups I Groups in Topology

Quantum Mechanics

Figure 3. Interconnections for group theory.

426 XII. The Mathematical Network

associated scientific problems; they might be arranged as in Figure 4,

beginning with differential equations, ordinary and partial (ODE to
POE!).
In each of these subjects the physical principles involved provide cer-
tain "equations of state", usually partial differential equations. The subse-
quent development is in part a series of formal deductions in a strictly
Mathematical style from these hypothesis (and the appropriate boundary
value conditions) and in part a series of approximations and adjustments
made with awareness of the intended applications. For that matter the
direction taken by the deductions is often guided by the applications. In
this sense, our description of Mathematics as the study of formal systems,
guided by ideas, must be supplemented by an extensive guidance by
scientific questions. The boundary line between science (truths about the
physical world) and Mathematics (deductions from well motivated
axioms) runs right through the middle of some of these subjects. Classical
mechanics, as discussed in Chapter IX, is not quite typical, because it is
older, more carefully axiomatized, and (probably as a result) more closely
tied to analysis and geometry.
There are also many connections back to core Mathematics, as for
example in the use of functional analysis and the various infinite dimen-
sional spaces required for full study of P.O.E. Moreover the quantitative
study of the trajectories of dynamics leads to the consideration of dynami-
cal systems, which in its tum played a large role in the growth of topology
(qualitative geometry).
There are many other subjects in applied Mathematics which we have
hitherto ignored. A small "sample" appears in Figure 5.
Probability as the "frequency" of an event involves serious foundational
questions, avoided by the measure-theoretical axiomatization suggested

Analysis

Dy_~' Sy,'om' / \ LiOOM O""mlo"

~O.D.E. P . D . E . / Hilbert Spaces

/ / \ ~ Fo.oliom' '""y,1i

I ~H=illOO,,"
Mechanics of Particles / ~metric Optics
Moehrni" Cootio"m M"rumi"

Cd"Ii,1 Moehruti" Flo'il Mmli"k, I

Quantum Theory ~ Elasticity

Differential Geometry I
Aerodynamics "" Plasticity

Figure 4
4. Subjects, Specialties, and Subdivisions 427

ICh'T,g I
Icoun<'
)'G-e-ne-ti-cs', Real Analysis

Permutations Probability Theory _(Axioms) _ Measure Theory

/
Algebra _ _ Distributions "'" Integrals
~ Lebesgue
Measure Spaces

(Binomial/Gaussian)

~ Random Process

Figure 5. The probable roots of probability.

here. Probability enters vitally into statistical estimates, but in statistics the
choice of methods (linear regression is overused) is much dependent upon
the application intended, while the foundations are still partially obscure
(Bayesian) or at least controversial. For these reasons we cannot count
statistics as now a part of Mathematics in our sense.
We have thus exhibited several diagrams of (part of) the network of
Mathematics with nodes the various specified "subjects" or "branches".
The description of a subject can be quite various. It is sometimes done by
a concept (holomorphic function) carefully defined in some context. It is
sometimes derived from (our ideas of) a form of experience-as in
geometry. But then the subject may split into more specific forms, some-
times those described by axioms (topological space) and sometimes by
types of formulas (algebraic geometry). Within algebra, different subjects
are often delimited by the appropriate axiom systems, but often the direc-
tion in which the axioms are exploited is suggested by the applications or
by the problems of special interest. When axioms are used, it can easily
happen that one subject is "contained" in another in a formal sense. Thus
a lattice is a particular kind of partially ordered set (one with greatest
lower and least upper bounds of pairs of elements), while a partially
ordered set in its turn is a particular kind of category (one in which, given
objects A and B, there is at most one arrow from A to B, and this when
A < B). Despite these formal inclusions, the subjects in question really
count as different ones, because of the very different ways in which the
axioms are developed.
Many subjects are subdivided into further specialties, reflecting new
techniques, different motivations and sometimes just the pressure of more
and more specialists. Thus Mathematical logic was once neatly subdivided
into four principal branches as shown just below the four boxes in Figure
6. However, one of the four (Proof Theory) did not have equal status and
the other branches proceeded to subdivide again, as indicated in the case
428 XII. The Mathematical Network

Analysis-----f~~_R~l~·go~r~ou:.::s!;F~ou~n~d~at:::io~n-....f-----Geometry
I

Proof Theory

Descriptive Set Theory Axiomatics Combinatorial Set Theory

Independence

Figure 6

of set theory in the figure. Moreover, this neat four-part subdivision

manages to omit some of the smaller activities: Intuitionsistic logic, modal
logic, and algebraic logic for example. There are also extensive ties to
computer science, especially in recursion theory and the study of algo-
rithms.
This particular figure illustrates well the increasing subdivision of
Mathematics attendant upon specialization and a resultant lack of
attention to connections (here to the rest of Mathematics) and neglect of
some of the original objectives (here questions about the foundations of
Mathematics). Many other subdivided subjects show similar effects.
We have thus illustrated many subjects and branches of Mathematics,
together with diagrams of the partial networks in which they appear. The
full network of Mathematics is suggested thereby, but it is far too exten-
sive and entangled with connections to be captured on anyone page of
this book.

5. Problems
The development of Mathematics, as reflected in the multiplication of its
subfields, is driven by various forces: Ideas arising from human activities,
questions posed by science, and problems presented within Mathematics.
Many research mathematicians think that the main guide to Mathematical
research is the desire to solve specific Mathematical problems-not so
much the smaller puzzle-type problems, but the major and famous
Mathematical problems. This includes some of the notable problems
which have recently been solved, such as Hilbert's fifth problem (con-
tinuity assumptions on Lie groups imply analyticity), the negative solution
of the decision problem for Diophantine equations (Hilbert's tenth prob-
5. Problems 429

lem), and the positive solution for the Poincare conjecture in dimension 4
(a simply connected four-dimensional manifold with the homology of a
4-sphere is necessary homeomorphic to a 4-sphere). More especially, one
must count here the many famous problems not yet solved: The Poincare
conjecture in dimension 3, the Riemann hypothesis for the zeros of the
zeta function, the Goldbach hypothesis (every even number is the sum of
two primes), the conjecture that every finite group is the Galois group of a
normal extension of the field Q, and many more. There are other, less
spectacular but equally difficult problems to be found in individual fields
of Mathematics. In all these cases, the continued attempt to solve such
problems can be a major source of new techniques, new ideas, and even
new branches of Mathematics.
Fermat's last theorem (xn + yn = zn has no solution in integers for
n > 2) is a striking example. For the corresponding equation with n = 2
one can find all integral solutions (all integral Pythagorean triangles) by
the factorization

i = (z2 - x 2 ) = (z + x)(z - x)

followed by a routine attention to the prime factors of the right hand side.
In the nineteenth century, Kummer was tempted to try the same for any
n, by the factorization

f = zn - xn = (z-x)(z-f.X)(z-(2x )...(z_(n-l x ), (1)

where ( is a complex primitive nth root of unity. But the algebraic

integers on the right side of (l) do not have a unique factorization into
primes (irreducibles), so the suggested attack on Fermat's last theorem
fails. Kummer's attempt to understand this gap led to the development of
the decomposition into the prime ideals just mentioned in (3.3). This idea
led directly to the whole subject of algebraic number theory (see §3), with
many further results-none of which yet prove Fermat's conjecture. Alge-
braic number theory also extended earlier studies by Gauss of the
classification of quadratic forms such as

ax 2 + 2bxy + ci

with integral coefficients. Given the analogy of algebraic numbers and

algebraic functions, it is likely that algebraic number theory would even-
tually have developed without the Fermat-Kummer impetus. However, it
is still the case that a decisive, well-posed problem-in elementary number
theory-led directly to a major development in the higher ranks of
number theory.
Another example is Burnside's conjecture on finite simple groups: That
every finite simple group which is not cyclic (of prime order) must have
even order. The penetrating Feit-Thompson demonstration of this conjec-
430 XII. The Mathematical Network

ture (§V.9) led to the subsequent elaborate determination of all finite sim-
ple groups and to an extensive revival of the whole subject of finite group
theory. In this case (and in many others) the presence of a pertinent prob-
lem is a major dynamic in the advance of Mathematics. However, this is
not always so. Hilbert's fifth problem (about continuous groups) was
famous until it was solved-and then it dropped out of prominence,
though the general subject of Lie groups continued to be active, little
influenced by the solution of this problem.
There are also all manner of small problems, some natural, some con-
cocted. Problems are a vehicle of competition and display. In some
schools (often in those influenced by Hungarian traditions) the business of
mathematics seems to be just the formulation and the eventual solution of
hard problems, with perhaps more attention to the difficulty than to the
relevance of the problem. Problems in this sense are akin to chess
problems-challenging but not necessarily relevant. Planar graphs provide
an example. Here a graph means an (abstract) finite collection of vertices
and edges, with the prescription that each edge joins exactly two
(different) vertices. Some graphs cannot be drawn in the plane (or,
equivalently, on a sphere) without an unintended crossing of edges. This
is the case in a simple conundrum. Tell the gas, electric and piped-heat
companies how to run their mains (gas etc.) all at the same level, six feet
below ground; to three different houses:

(2)

By simple application of the (deep) Jordan curve theorem, this is impossi-

ble, as suggested in the figure. Or try to tell five secretive embasies,
located in Bern during a war, how to dig tunnels all six feet under so that
they can secretly communicate, each to each:

(3)
6. Understanding Mathematics 431

Jordan again testifies to the inevitable failure of this attempt at secrecy.

Both (2) and (3) are thus non-planar graphs. The story goes that W.L.
Ayres tried to prove that every non-planar graph must contain the
configuration (3), while C. Kuratowski tried to prove that every non-
planar graph must contain (2). Evidently both attempts failed, but the
two people met at some international congress and exchanged notes. Then
Kuratowski went on to prove the striking theorem: Every non-planar
graph contains either the graph (2) or the graph (3).
Both graphs (2) and (3) can be drawn on a torus (or on any surface of
higher genus). Soon Paul Erdos proposed a problem:
Find a list of all minimal non-toral graphs (or similarly for other sur-
faces). What with extensive use of planar electric circuits or computer
chips and the related revival of combinatorial questions, someone recently
proved: There are exactly 103 different graphs which cannot be drawn on
the projective plane and such that at least one of them is contained in any
graph not representable on the projective plane. This evidently cumber-
some result clearly demonstrates that the Erdos problem is not very fruit-
ful: It has not (at least yet) led to an understanding of why a graph can-
not be drawn on the projective plane (or the torus). Mathematics should
be concerned with understanding.
We conclude that problems, like theorems, are sometimes fruitful and
sometimes not. Solving a supposedly famous problem is not necessarily a
famous contribution to Mathematics. However, the emphasis on problems
is an effective counterweight to the emphasis on structure and axiomatics
in our description of Mathematics. They are an essential component in
the dynamic.

6. Understanding Mathematics
A major element in Mathematics is the need to fully understand why the
formulas work and why the theorems hold. In lectures, in courses, and in
conversations Mathematicians try repeatedly to better understand the
received results of each subject. Their methods are various; we now
examine some under the headings analogy, examples, analysis of proofs,
shift of attention, and the search for invariant form:
(a) Analogies. In many respects, 2-dimensional Euclidean space is much
like 3-dimensional space: . Points and lines become points, lines, and
planes; triangles and circles become tetrahedrons and spheres. These and
other analogies strongly suggest a consideration of "spatial" configurations
in higher dimensions. Some initial attempts describe 4-dimensional space
by points, lines, planes, and hyperplanes; as this becomes cumbersome
and as the phenomena requiring still higher dimensions arise there is a
search for more effective methods, first by using coordinates and then by
vector spaces.
432 XII. The Mathematical Network

In the beginning uses of measurements, the analogies between measur-

ing distances, measuring angles, and measuring quantities (also weights)
leads gradually to the formulation of real numbers as the general means
of handling measurements.
The analogy between Boolean operations on sets (intersection, union
and complement) and the logistic operations on propositions (and, or, and
not) plays a central role in logic.
Or consider linear combinations. Every vector in the plane is the sum of
a horizontal and a vertical component. Every complex number is a sum of
real and imaginary parts. Every solution of the differential equation
y" = - Y is a sum of a multiple of cos x and a multiple of sin x. All
these analogous unique linear decompositions lead to the concept of a
basis in a vector space.
Or again, the prime ideals in algebraic number theory are analogous to
the points on Riemann surfaces of algebraic functions (§3). Throughout
mathematics, analogies are powerful sources of development.
(b) The study of examples. The choice of a specific but critical example
may well serve to indicate the direction in which a whole theory might
develop. Thus the Weierstrassian theory of analytic functions as power
series in a complex variable came after his teacher Gudermann had
exposed him to a complicated "combinatorial" study of the particular
power series for eX, sin x, cos x, and the like. Hilbert's intensive study of
relative quadratic extensions of number fields led him to make a series of
conjectures about general abelian extensions of algebraic number fields,
and the demonstration of these conjectures occupied the development of
class field theory for fifty years. Algebraic geometry started from the care-
ful classification of the quadratic, cubic, and quartic curves in the plane.
Riemann surfaces for many-valued functions of a complex variable were
first developed by pasting together slit planes carrying branches of alge-
braic functions, and only later were given a general conceptual formula-
tion.
In a sense, the "general" theories of Mathematics are there in order to
conveniently describe the common features of a multitude of special
cases-and a crucial special case may be the key to a general theory.
(c) The analysis of proofs. There is a continual search to get better
proofs of known theorems-not just shorter proofs, but ones which do
more to reveal why the theorem is true. This study can lead to the
discovery of crucial axioms used in the proof (like the discovery that some
basic arguments in analysis made hidden use of the axiom of choice).
Again, study of a proof may suggest new concepts. For example, the Cau-
chy integral theorem, initially and perhaps carelessly proved by the use of
the Gauss-Green lemma (§X.5), plays a central role in complex analysis.
Study of its proof suggested subdivision arguments, for rectangles and tri-
angle, and so suggested homological methods. Other proofs involve defor-
mations and so bring in homotopy, and still other approaches culminate
6. Understanding Mathematics 433

in the Jordan curve theorem and the resulting deeper study of the topol-
ogy of the plane.
To present a proof, one may often try to put it in a vivid or diagram-
matic form. When this done for the Jordan-Holder theorem (§V.7), it
naturally leads to the diagrammatic presentation of a partially ordered set
and thence to the notion of a lattice. Again, topologists found it suggestive
to denote continuous maps from one space X to another space Y by
arrows X ~ Y. This simple notation soon suggested fertile concepts of
exact sequences and of categories. In brief, proofs are the meat of
mathematics, so chewing them over can produce juicy results!
(d) Shift of attention. The understanding of a proof may progress when
our attention is shifted from the initially emphasized aspect to some other
focus. Thus Galois theory starts as the study of the solutions of a particu-
lar polynomial equation; it then becomes the study of the field of all the
rational combinations of the roots of that equation. This shift makes it
possible to see the Galois group of the equation not as a group of permu-
tations of its roots, but as a group of automorphisms of the corresponding
field. In group theory, the cosets of a normal subgroup N of a group G
become the elements of the factor group G/ N; when, instead, the
emphasis is placed on the function sending each element to the coset
which it generates, there is revealed a homomorphism of the original
group G onto the factor group G/ N -and one then sees that the properties
of this homomorphism, and not the particular properties of the cosets,
provide (§V.6) the more effective description of the factor group. Simi-
larly, shifting from matrices to the corresponding linear transformations
substantially changes the emphasis in linear algebra, from calculation to
visualization-and lays the groundwork for an effective approach to
infinite-dimensional vector spaces.
(e) Invariant formulation. The introduction of analytic geometry by
Descartes provided for algebraic proofs of geometric facts-but the alge-
braic proofs depended on a choice of coordinate systems, while the
geometric facts-for example, the fact that the medians of a triangle are
concurrent-are independent of the choice of coordinates which may be
used in the proof. For this reason, the development of analytic methods in
geometry has constantly seen parallel attempts to make the results (or
even the arguments) independent of such choices. In linear algebra, this
appears in formulations independent of the choice of bases. In tensor
analysis, a tensor, relative to bases, first appeared as a multiply indexed
family tt of scalars t, and the description of a tensor had to be coupled
with a complete description of the effect of a change of basis on these t's;
alternatively, tensors can be defined (§7.9) in a coordinate-free manner as
elements of tensor products of vector spaces. An algebraic variety in affine
n-dimensional space is the locus of a finite list of polynomial equations in
the coordinates-but it can be described more invariantly in terms of the
ideal of all polynomials which vanish on the locus; the given equations
434 XII. The Mathematical Network

are then just one set of generators (among the many possible sets) for that
ideal.
The search for an "invariant" formulation is not limited to questions of
geometry. It can also arise in algebraic contexts. Thus a specific group
may originally be described in terms of specified generators and their rela-
tions; for example, a dihedral group is generated by a rotation (by 360 jn)
0

and a reflection. Alternatively, the same group may be described in more

invariant terms as the group of all symmetries of an n-sided regular
polygon. In universal algebra, a group is defined as an algebraic system
with one binary operation (of multiplication), one nullary operation (iden-
tity), and one unary operation (of inverse). However, there are other
iterated operations intrinsic to a group, such as the ternary operations

For each n, there are similar n-ary operations defined by "words" in n of

the Xi and their inverses. A full and invariant derivation of the "theory"
of groups is given by the totality of all such group operations, suitably
organized (for example, as a category).
Invariant formulations also appear in mechanics; where the principle of
least action presents the laws of mechanics in an essentially coordinate
free fashion.

7. Generalization and Abstraction

Another dynamic leading to the extension of Mathematics is the process
of generalization or abstraction, which can be described as follows under
several headings.
(a) Generalizations from cases refers to the way in which several specific
prior results may be subsumed under a single more general theorem. Thus
the observation of right triangles with given lengths of sides (such as the
triangle with sides 3, 4, 5, or sides 5, 12, 13) leads both to the Pythagorean
theorem about right triangles and to the algorithm which obtains all solu-
tions in integers of the Pythagorean equation x 2 + i = z2. Similarly,
the familiar algebraic expansions of (x + y)2, (x + y)3, and so on are
subsumed in the binomial theorem expanding (x + y)n for any integer n
-complete with the formulas for the binomial coefficients, and general-
ized yet again with infinite binomial series for fractional exponents n.
In these instances, the generalization from cases produced a general
theorem subsuming all that was known in the specific cases. Other, more
involved instances may see the development of a theory which unites
some (but perhaps not all) of the results holding in the given cases. Thus
algebraic number theory started with specific properties of quadratic and
cyclotomic number fields before going on to arbitrary algebraic number
fields, but the deeper results of the class field theory apply only to abelian
7. Generalization and Abstraction 435

extensions (those which are normal with an abelian Galois group (§V.7)
over some ground field, such as Q). Similarly the elaborate study of the
various forms of conics, cubic, and quartic curves in the plane finally
extracted some systematic facts about intersections; for example, a cubic
and a quartic "in general" intersect in 3.4 = 12 points. Generalization
then led to Bezout's theorem: Two plane curves of orders m and n inter-
sect "in general" in mn points. But the validity of this theorem does
require some substantial adjustments. Among the intersections we must
count possible intersections at "infinity" (in the projective plane) as well
as all the intersections at points with complex coordinates-while each
intersection must be counted with a suitable "multiplicity". Here special
cases, such as tangent lines and osculating circles suggest how to go about
defining what is meant by multiplicity; for example, suitable ideals may
be used. Similar notions of multiplicity thus develop for the intersections
of surfaces and manifolds of higher dimension: They playa major role in
algebraic geometry and subsequently in algebraic topology. There they
are transmuted into cup products in cohomology.
There are many other examples of generalizations from cases-for
instance, various cases in number theory led to the unique decomposition
of finite abelian groups into products of cyclic groups (V.8.4).
(b) Generalization by analogous steps refers to situations where a more
general theory runs in parallel to earlier cases without actually replacing
or subsuming those cases. An instance in point is the parallel between real
numbers, complex numbers, and quatemions-followed by the search for
higher dimensional such algebras over the reals, with the discovery that
there were no further finite dimensional division algebras without the loss
(say) of the associativity of multiplication. Another striking example is the
generalization from real analysis to complex analysis.
(c) Generalization by modification refers to instances in which a desir-
able theorem cannot be generalized in a direct way, but in which suitable
concepts are used to modify the theorem so as to make a generalization
possible. Thus the unique prime decomposition valid for the ordinary
integers fails to hold for the ring of integers in most quadratic and
cyclotomic fields of algebraic numbers-but it can be replaced in this and
other rings by a unique decomposition of ideals into products of prime
ideals. Moreover, the original decomposition of an integer n into its prime
factors is reflected in the general theorem by way of the decomposition
into prime ideals of the principal ideal (n) of all multiples of n. Again, the
Jordan curve theorem, valid for simply closed curves in the plane, is not
true for all such simple closed curves on the torus-but the classification of
the ways in which this theorem fails is a major source of the study of
homologous curves and of the introduction of the homology and homo-
topy groups (which, in effect, measure variously the numbers of ways in
which the Jordan curve theorem can fail).
Abstraction. Generalization and abstraction, though closely related, are
best distinguished. A "generalization" is intended to subsume all the prior
436 XII. The Mathematical Network

instances under some common view which includes the major properties
of all those instances. An "abstraction" is intended to pick out certain cen-
tral aspects of the prior instances, and to free them from aspects extrane-
ous to the purpose at hand. Thus abstraction is likely to lead to the
description and analysis of new and more austere or more "abstract"
mathematical concepts. We will describe some types of abstraction under
headings "abstraction by deletion," "abstraction by analogy," and
"abstraction by shift of attention".
(d) Abstraction by deletion is a straightforward process: One carefully
omits parts of the data describing the mathematical concept in question to
obtain the more "abstract" concept. This often leads to a reverse process,
in which it is shown that all (or some) of the abstract objects can have the
deleted data restored, perhaps in more than one way. Such a restoration is
then called a "representation theorem".
For example, if one starts with the notion of a transformation group,
one may delete the elements being transformed but retain the associative,
identity, and inverse laws for the composition of transformations. The
result is the notion of an "abstract" group. The corresponding representa-
tion theorem, due to Cayley, asserts that every abstract group is iso-
morphic to a group of transformations on some set (§V.6).
The algebra of sets (§ 1.9) concerns the algebraic operations of intersec-
tion, union, and complement on the subsets of some given set. If one
deletes all references to the elements of these subsets, but retains the three
operations and (a suitable list of) the basic identities which they satisfy,
one reaches the notion of a Boolean algebra. Conversely, the Stone
representation theorem asserts that every Boolean algebra can be
represented as the algebra of some of the subsets of a suitable set.
Sometimes such a deletion is suggested by the sorts of theorems being
considered. Thus, in functional analysis, certain theorems about the
existence of functions with given properties are like theorems about the
existence of vectors in linear spaces. This use of geometric language then
suggests an abstraction: The deletion of the numbers on which the func-
tions act, so that the functions (for the intended purpose) become just
points or vectors in some Banach or Hilbert space.
(e) Abstraction by analogy occurs when a visible and strong parallel
between two different theories raises the suggestion that there should be
one underlying, less specific theory sufficient to give the common results.
Thus systems of numbers or of functions closed under the usual rational
operations (addition, subtraction, multiplication, and division by non-zero
elements) follow the same algebraic rules and yield the same theorems for
the solutions of quadratic and higher equations (Galois theory). These
parallels then suggest the abstract notion of a field-which soon turns out
to include other instances such as the p-adic numbers and the field of
integers modulo p as well as other finite fields. In the same direction, the
close parallel between algebraic number fields and fields of algebraic
7. Generalization and Abstraction 437

functions suggests the more abstract theory of finite dimensional exten-

sions of any field.
The notion of a modular lattice arose from a similar striking analogy.
On the one hand, the Jordan part of the Jordan-Holder theorem for finite
groups (V.7.3) asserts that any two composition series for such a group
have the same length. On the other hand, for a vector space any two finite
bases have the same number of elements-a number which can be
described as the maximum length of a descending chain of subspaces (the
number is of course the dimension of the total space). Now both sub-
spaces R, S, T of a vector space and normal subgroups of a group form
lattices under inclusion-that is, posets with least upper bounds R U S
and greater lower bounds R n S, as described in §I.9. Moreover these
lattices satisfy an additional condition, the so-called modular law:

R :J T implies R n (S u T) = (R n S) u T. (1)

One is thus led to formulate and prove the theorem asserting that any two
maximal descending chains in a finite modular lattice have the same
length. For subspaces of a vector space, the length is the dimension. How-
ever, this modular lattice theorem does not contain the original Jordan-
Holder theorem, because the latter concerned not chains of normal sub-
groups of the whole group but chains of subgroups, each normal in the
next larger. The modular lattice theorem shows only that any two chief
series (maximal chains of normal subgroups) in a group have the same
length. However, the modular lattice theorem does apply in other cases, as
for example to subspaces of a projective space, and it has infinite-
dimensional generalizations, to the so-called continuous geometries used
by von Neumann to analyse rings of operators on Hilbert spaces. (See
Birkhoff [1967].)
Incidentally, the modular law (1) also holds in a Boolean algebra which
satisfies the stronger distributive law (for all Rand T)

R n (S U T) = (R n S) u (R n T).
This resembles (1), since R :J T there means that T on the right of (1)
may be replaced by R n T.
(f) Abstraction by shift of attention. Some abstractions arise when the
study of a Mathematical situation gradually makes it clear that certain
features of the situation-perhaps features which were originally
obscure-are really the main carriers of the structure. These features, with
their properties, are then suitably abstracted. Thus the Galois theory
begins as a theory of the roots of a given polynomial, and the Galois
group arises first as a group of permutations of those roots-all permuta-
tions leaving invariant all polynomial (and hence, all rational) relations
between the roots. Presently, it appears that this group of permutations
acts on all the rational expressions in these roots, and that the Galois
438 XII. The Mathematical Network

group can thus be described as a group of automorphisms of the field

consisting of all such expressions. With this more abstract view of the
Galois group, the whole theory becomes more perspicuous! Abstraction
supports understanding!
The development of point set topology exhibits several successive such
shifts of attention. Initially, the subject dealt with the properties of sets of
points on the line, the plane, or in three space and with the properties of
functions continuous on such point-sets. Presently, it became clear that
most of the results depended on the internal properties of these point-sets,
and not on the way in which the sets were embedded in a Euclidean
space. All that mattered was the distance between any two points in the
set. This, and the need to handle more general cases arising, say, from the
calculus of variations, led to the abstract notion of a metric space, with
axioms expressing the properties of distance. However, it then appeared
that only "qualitative" properties of the distance were needed in order to
define continuity; this led to the more abstract notion of a topological
space (and the corresponding representation theorem, which arises here as
the problem: When is a topological space metrizable?).
This case illustrates the observation that attention can shift dramatically
from one notion to another. Initially, continuity and limits were defined in
terms of inequalities and an t: - 8 process. It was then recognized that
these numerical equalities amounted to using open intervals (on the line)
or open discs (in the plane). Gradually, Mathematicians came to utilize
the more general open sets-those which are arbitrary unions of open
intervals or open discs-and to the eventual recognition that these open
sets are all that is needed for the definition of continuity. In each
Mathematical situation the basic notions are disentangled by insight and
understanding-which often amounts to a shift of attention, and which
may then call for a corresponding abstraction.
There are many other types and examples of abstraction. Thus, the
"abstract" description of a Riemann surface as a one-dimensional com-
plex manifold helps understand how concrete Riemann surfaces can be
pieced together from sheets with slits. The abstract definition of a sym-
plectic manifold illuminates the role of the p's and q's in the formulation
of Hamiltonian mechanics. The abstract definition of a function as a set of
ordered pairs explains the functional dependence of one quantity upon
another, good whether or not that dependence is expressed by a formula.
Much of the dynamics of Mathematics depends on the constant attempt
to get better and deeper understanding. Generalization and abstraction
are two useful means to such understanding.

8. Novelty
We have seen that problems, generalizations, abstraction and just plain
curiosity are some of the forces driving the development of the
8. Novelty 439

Mathematical network. But this network is also tied to a variety of human

and scientific activities; thus questions so arising can be important sources
of novel Mathematics. Historically, projective geometry was started in part
by considerations of practical perspective, while Fourier's study of heat
led to his series, so widely useful.
There are many other, current, examples. Dirac, in formulating quan-

defined for all real x, zero when x i= 0, non-zero only for x = °

tum mechanics, came to consider his well-known delta function l5(x)-
and such
that its integral from - 00 to 00 is 1. Literally speaking, there is no such
function, so that Dirac's maneuver may have seemed Mathematically
empty. But V.1. Sobolev and Laurent Schwartz observed that the role of
the delta function could be replaced by a suitable operator on functions,
say by a "distribution". This impetus led to the development of a
considerable branch of Mathematics.
The bizarre phenomenon of a single solitary wave was first observed
about 1840, when Scott Russell chased such a wave moving down the
Edinburgh-Glasgow canal. There appeared to be no way to derive the
existence of such a solitary wave from the accepted partial differential
equations for water waves. Finally, several authors did find model equa-
tions which would predict such waves; one such is the Korteweg-de Vries
equation

u, + ux + uux + uxxx = 0, (1)

where u( t,x) is the height of the wave above the standard water level in
the canal at time t and distance x down the canal. In this equation the
basic terms u, + Ux express standard conditions (§VI.ll) for the propaga-
tion of waves such as u = f( x - t); they are modified by the effect of
dispersion represented by the third partial derivative Uxxx and by a term
uU x , non-linear in u. This equation does have explicit travelling wave solu-
tions, appropriate to the observed phenomenon. Moreover, it has surpris-
ing applications elsewhere, for instance in magneto-hydrodynamics.
Computer science brings up other new Mathematical ideas. For exam-
ple there are new algorithms which find the product of two matrices more
quickly than the straightforward application of row-by-column multiplica-
tion. There are many other new algorithms, plus questions of principle
about computational complexity: For a computation of specified size, how
does one estimate the minimum time really necessary; can it be done in
"polynomial time"?
Not all outside influences are really fruitful. For example, one engineer
came up with the notion of a fuzzy set-a set X where a statement x E X

°
of membership may be neither true nor false but lies somewhere in
between, say between and 1. It was hoped that this ingenious notion
would lead to all sorts of fruitful applications, to fuzzy automata, fuzzy
decision theory and elsewhere. However, as yet most of the intended
440 XII. The Mathematical Network

applications tum out to be just extensive exercises, not actually

applicable; there has been a spate of such exercises. After all, if all
Mathematics can be built up from sets, then a whole lot of variant (or,
should we say, deviant) Mathematics can be built by fuzzifying these sets.
All these searches for new ideas and for new formalizations are inevit-
ably adventurous and uncertain. It can tum out that an apparently attrac-
tive new idea is not really fruitful or relevant, or it can tum out that a
proposed new abstract concept or formalization is just too cumbersome or
too peripheral to be productive. On the other hand, some new formal con-
cepts can be of central subsequent importance. What really matters is the
role of the concepts in the network of Mathematical form.

9. Is Mathematics True?
It is customary to ask of a piece of Mathematics: "Is it true?" For exam-
ple, one might ask: Is the Jordan curve theorem true of that simple closed
curve drawn out there in the middle of my complex plane? The theorem
says that the curve separates that plane into an inside and an outside so
that no path can connect an inside point to an outside point without
meeting the curve itself. That particular curve is quite convoluted (Figure
X.4.3), so it's a bit hard to really tell which of the points are inside and
which are outside; but is it true?
The whole thrust of our exhibition and analysis of Mathematics indi-
cates that this issue of truth is a mistaken question. There really isn't an
absolutely fiat Euclidean plane out there, complex or otherwise; there is
only the (very slightly) bumpy surface of the blackboard, which further-
more is very far from extending out to "infinity" in any direction what-
ever. What appears on the blackboard is not really a simple closed curve,
but a wavy and somewhat thick line of chalk marks; perhaps the marks
even skip a little, so that a careful draughtsman might be able to sneak a
path from the "inside" out to some point clearly "outside". Even if the
curve were devoid of gaps, I can't really demonstrate that it is continuous;
for every £ greater than zero I haven't actually chosen a 8 such that. ...
Moreover I am not really interested in the inside and outside of this par-
ticular Jordan curve; I am rather interested in knowing that my plane is
simply connected or that the holomorphic functions I intend to define on
it satisfy Cauchy's integral theorem or that the results of this theorem can
be used, via some conformal map, to design an airplane wing. For this
and all sorts of other reasons the theorems of Mathematics, by Jordan or
others, are not simply statements about the behavior of individual objects
in the physical world.
This point can be put more explicitly. Our survey has indicated that
Mathematics is an extensive network of formal rules, definitions and sys-
tems, tightly tied here and there to activities and to science. This descrip-
9. Is Mathematics True? 441

tion does not in any way provide a physical object for each Mathematical
term, or a physical law corresponding to each Mathematical theorem.
Instead there are many pieces of Mathematics and a variety of accepted
procedures for making practical use of some of these pieces. As a result, it
is simply meaningless to ask, given a relation between Mathematical
terms, whether it is true of the "corresponding" physical objects. The
correspondence is just are not that direct. Instead, our description of
Mathematics indicates that the appropriate questions are different ones:
Is this piece of Mathematics correct? That is, do the calculations follow
the formal rules prescribed, and are the theorems deduced from the stated
axioms by rules of inference on which we have agreed?
Is this piece of Mathematics responsive? That is, does it settle some
problem which had arisen or does it carry further some development
which was incomplete?
Is this piece of Mathematics illuminating? That is, does it help under-
stand what had gone before, either by further analysis or by abstraction or
otherwise?
Is this piece of Mathematics promising? That is, though it is a novel
departure from precedent or fashion, is there a reasonable chance that it
will subsequently fit in the picture?
Is this piece of Mathematics relevant? That is, is it tied to something
which is tied to human activities or to science?
Many of these questions are questions of degree: Is it more or less
relevant? Even that is hard to judge because relevance can be transmitted
all along the network of Mathematics. There are many pieces of
Mathematics which fit well but are not immediately relevant to any
application-and which much later may tum out to be applicable. The
promise of a new idea is also difficult to judge; too often the new may be
dismissed as "off-beat". But before all comes the question: Is it correct?
Centuries ago, when Mathematics dealt largely with arithmetic and
elementary geometry, it was perhaps easier to think of the numbers or the
geometric figures as real objects about which one could make true state-
ments. These objects had been long familiar and they were eminently use-
ful, so it was comfortable to consider them as real. Now, looking carefully,
we see that this comfort was an illusion. Numbers are the means used in
calculating by rules, while figures are the images used to suggest formal
geometric proofs; their eminent utility simply suggests that the formal is
powerful in practise. By now there are so many and various Mathematical
objects that any imputation of their reality offers spectacular cutting
opportunities for Occam's razor.
The real world is understood in terms of many different Mathematical
forms. For example, the fours group Z2 X Z2 is not a single object some-
where in the world. It is rather a form, exemplified many times over in
the world, by the symmetry present in each rectangular shape. It is also
exemplified within Mathematics, say as the Galois group of the field
442 XII. The Mathematical Network

Q( V2,i) generated by i and V2. However, in Mathematics it doesn't exist

all by itself, but as a very small part of group theory, illustrating the
decomposition theorem for finite abelian groups, but overshadowed by
other newer pieces such as the monster group discussed in §V.9. Centuries
ago none of these groups were there, because our ancestors had not then
recognized any groups. In the interim, Mathematics has brought out many
new and relevant forms. They arise by contemplating the world, but of
themselves they are not in the world. They are forms~
This argument rests not just on the few examples here cited, but on the
whole weight of evidence of the character of Mathematics. Our examina-
tion has shown that Mathematics consists of formal rules, formal systems,
and formal definitions of concepts. The proofs of Mathematics do not
depend upon experience, and indeed can often be invented or carried out
by young people with little experience. Hence the results cannot be
checked by asking how they might refer to real experience. Mathematics
is not literally a science because its results cannot be falsified by fact or
experiment. Therefore the question "Is Mathematics true?" is out of place.
To be sure, it is easy and common to think that Mathematics is true.
This is partly because the formal Mathematics realizes intuitive ideas
which themselves can be vivid and forceful. Even more it is because
Mathematics, through the decisive notion of formal proof, can appeal to
an absolute standard of rigor. Moreover, this standard is external and
impersonal: Once the axioms of a system are set, all the statements of the
system are either demonstrable, refutable, or (thanks to Godel) undecid-
able. No collusion, no political influence, no second thoughts can alter the
fact of the matter that the theorem can be proved. Given the straightfor-
ward definition of a finite simple group, the existence of the "monster" is
ineluctable. Mathematics is not true, but its correct results are certain.
This thesis has consequences. The sets used in Mathematics are not
physical totalities or collections; they are instances of an abstract concept
defined by careful axioms-and we may even adjust the axioms (it seems
that we need to!) These axioms mimic what we might operationally do
with a collection of pebbles or of people. But each real collection out
there is finite; even the fanciest collectors of coins or of stamps have col-
lections which are finite. Now the axiom of infinity is necessary, both for
Zermelo-Fraenkel and for categories. It states a formal property of a col-
lection such that, were that collection a real object, it would be an infinite
collection. But there is no such thing. The axiom of infinity is simply a
formal statement (with consequences!) inspired or modeled after the
practical observation that a collector of stamps or coins can always add
one more item to his collection; the axiom imagines that done forever
(which has never yet happened). This is the sense of the infinite in
Mathematics, but all the actual operations of the Mathematician-his
calculations, his definitions, his proofs-are strictly finite. These observa-
tions destroy the position of these finitists (like Kronecker) who objected
to Cantor's use of infinite sets.
9. Is Mathematics True? 443

Mathematical existence is not real existence; it is in no way to be com-

pared with the existence of people, political compaigns, or with the non-
existence of unicorns. Aspects of such "real" existence did lead to
Mathematical (or logical) ideas which in turn were formalized in the
axioms for ( :3 x )F( x) in the predicate calculus. But in that calculus each
theorem (:3 x )F( x) simply means that this formal statement has been
deduced according to those formal rules of inference. By virtue of this
parallel, those existence theorems are often relevant in the applications of
Mathematics. They are also a check on the adequacy of a formalization:
When a differential equation is supposed to model some physical
phenomenon, an existence theorem for solutions of the equation shows at
least that the model is capable of producing some result. The fundamental
theorem of algebra proclaims that there exists a complex root of any poly-
nomial (with complex coefficients): it doesn't of itself produce the root,
though there are some proofs which are constructive in the sense that they
prescribe steps which will calculate approximations to the root. A theorem
of Zermelo's, with an astute proof, asserts that the axiom of choice implies
that the reals can be well-ordered. It does not exhibit a well-ordering.
Nobody has exhibited such. In short, Mathematical theorems about
existence are those that can be proved from the logical rules about
existence; nothing more, nothing less. Indeed, if you want less, it can be
done with intuitionistic logic, but this is equally a formal system.
Brouwer's own earlier, non-formal intuitionism was based on an assump-
tion of the primacy of the natural numbers in Mathematics-and this flies
in the face of the observation that actual Mathematics effectively uses
natural numbers and real numbers as equally primary. The pillarstones
are geometry and arithmetic.
The view that Mathematics is "correct" but not "true" has philosophical
consequences. First, it means that Mathematics makes no ontological
commitments; in other words, Mathematics makes no statement about the
reality of some of its objects. We do not need to know that every Dede-
kind cut in the rationals determines a "real" real number; we do know
only that this description produces forms for real numbers which work
with magnificent success. (This success is perhaps the essential bond
between those pillarstones of geometry and algebra.) Quine has on occa-
sion argued that we make an ontological commitment to the range of a
variable when we quantify over that range. This arises, I suspect, from an
undue concern with logic, as such. For Mathematics, the "laws" of logic
are just those formal rules which it is expedient to adopt in stating
Mathematical proofs. They are (happily) parallel to the laws of logic that
philosophers or lawyers might use in arguing about reality-but
Mathematics itself is concerned not with reality but with rule.
This view means that the philosophy of Mathematics need not involve
questions about epistemology or ontology. If Mathematical theorems do
not assert truths about the world, we need not inquire as to how we know
or would come to know such truths. (We of course do need to inquire how
444 XII. The Mathematical Network

we recognize a correct proof, but getting the recognition is a major part of

advanced education in Mathematics, and is usually not considered as part
of epistemology.) This observation means that the philosophy of
Mathematics cannot be much advanced by many of the books entitled
"Mathematical Knowledge", in view of the observation that such a title
usually covers a book which appears to involve little knowledge of
Mathematics and much discussion of how Mathematicians can (or cannot)
know the truth. This dismissal applies especially to the later (posthumous)
volume of Wittgenstein [1964], where the actual Mathematical content
rarely rises above third grade arithmetic, while the actual concern is less
with Mathematics than with its use to illustrate some strictly philosophical
issue.
But this does not dispose of the hard questions about the philosophy of
Mathematics; they are merely displaced. Mathematics is a formal net-
work, but the concepts and axioms there are based on "ideas". What are
they? We have identified a number of particular such ideas; there are
many more. They are not like Platonic ideal forms, because they are intui-
tive and vague-as we have repeatedly noted, the "same" idea can lead to
a number of different formalizations. Indeed, it seems that we can recog-
nize and name an idea only after it has given rise to one or more formal
expressions. Before that an idea can be pretty nebulous; it is "up in the
air". That may be why it is hard to convey to others some new intuition
about Mathematics. Hence:
Question I. What are the characteristics of a Mathematical idea? How
can an idea be recognized? described? What is the relation between a
(nebulous) idea and a (precise) formulation.
This question may also have historical aspects. Why did the idea of
composition (symmetry, groups, categories) arise so late? Was it that no
applications were visible? Or were the potential uses in geometry just
ignored? Or did it require Galois?
Next comes the conundrum: How do we recognize the presence of for-
mal rules in simple human activities? How do we select the "facts" about
numbers from the informal practises of counting? For that matter, how do
we recognize that some particular scientific question has a formal or
Mathematical component? This is an especially puzzling problem, because
it is a common observation that when a scientist comes to a Mathemati-
cian with a question, he is very often unable to formulate that question is
clear Mathematical terms, so that it may take considerable added inquiry
to get to such a formulation. This gives us
Question II. How does a Mathematical form arise from human activity
or scientific questions? What is it that makes a Mathematical formulation
possible?
This may be a question of epistemology (inverted). How do we know
that there is some Mathematics there to be extracted?
After the Mathematical form is extracted its development continues,
with the dynamics of problems, generalizations, and abstractions which
9. Is Mathematics True? 445

we have described. What is then mysterious is that these abstractions,

inventions, and calculations so often have applications and uses back in
the indubitably real world. It's not just that prime numbers, long con-
sidered a sort of special and esoteric concern of number theory, now sud-
denly can be used in cryptology. It is all those differential equations which
do produce solutions which fit, all those Hilbert spaces which crop up in
the formulation of quantum mechanics, all those tensors in differential
geometry which then can be used in relativity, all those problems in the
calculus of variations which can be used in studying optimal control, and
so on and on. The Mathematician's process of building up his elaborate
formal systems so that they fit together also produces models which (par-
tially) fit the real world. This leads to our third and most basic question,
like that formulated originally by Eugene Wigner:
Question III. How can one account for the unreasonable effectiveness
of Mathematics in providing models for science and knowledge?
Here ontology and reality reappear. And Mathematical knowledge,
apparently just about formulas, turns out to sometimes help produce
knowledge of phenomena.
Is it that the real world is so constituted that parts of it easily fit formal
patterns? Or is it that our very notion of formal pattern has been
developed over millenia so that it is possible to fit the pattern to the fact?
For a strict formalist, for whom Mathematics is just any manipulation
of symbols, this question III cannot be answered.
On our view, the rules, axioms, and definitions of Mathematics are
indeed formal, but they were initially chosen in the light of various
human activities and scientific phenomena, and then extensively
developed to improve understanding. There is therefore reason to expect
that the results might be applicable in return, at least to the original
phenomena. But the resulting application is often to other and different
phenomena!
To fully account for this applicability, the phenomena must in some
sense be ready to fit the formulas. This becomes a question about the
character of reality. All the experience of Mathematics and of the physical
sciences shows that many aspects of reality can be measured, organized,
and then understood by way of theories and concepts which have a large
formal content. We have agreed that the form is chosen to reflect the facts.
It must also be the case that the facts accept the form. We have not
explained why this is so.
Put differently, the elaborate formal development of Mathematics yields
conclusions which often fit the world. Does this mean that the "real
world" is not just a bundle of confusions, but that the variety of events in
the world does fall into patterns which are amenable to sophisticated for-
mal descriptions? If there indeed are such patterns, how is it that man is
able to extract and analyse them so effectively? Numbers in their complex
structure, groups in their variety and symmetry, geometry in its many ver-
sions, and analysis in its apperception of change all deal with patterns
446 XII. The Mathematical Network

which are highly relevant to the world. This requires explanation on the
metaphysical and the epistemological level-and raises thus a variety of
questions to which I have no adequate answer.
Here is a sketch of some tentative answers to these questions.
(i) Various phenomena do have underlying similarities and regularities.
These regularities are somehow propagated from one situation to
another.
(ii) On the basis of millennia of experience, mankind has developed
"ideas" about the phenomena which in turn are used to extract from
the phenomena a conceptual description of some of these similarities
and regularities. Some of the concepts involved can be made strictly
formal, and hence Mathematical.
(iii) Different formal statements in the descriptions, because of the propa-
gation of the regularity of phenomena, can be closely connected with
each other. Because these connections help the practical understand-
ing of the regularities, they have been extensively examined and in
some cases reduced to formal proofs, by rules of inference, from
astutely chosen axioms.
(iv) In many cases, the results of these proofs fit the facts-not just the
particular immediate facts from which the concepts were extracted,
but other facts which, because of the propagation of regularity, also
fit these deductions. By the choice of the successful cases of concept
formation and deductions, various branches of Mathematics are
selected and developed. From time to time the development is sup-
ported by additional formal explanations of the regularity of new
phenomena.
A development of this analysis might account for the "unreasonable
effectiveness of Mathematics in providing methods for science". For
example the concept of a group, originally formulated to analyze certain
symmetries in Galois theory and in geometry, turns out to be relevant to
Mathematical physics, primarily because corresponding symmetry
phenomena arise there.
To summarize: The world has many underlying regularities which, once
extracted, can be analyzed and understood by Mathematical form.
Because it is formal, the same Mathematical notions can apply to widely
different phenomena.
Question IV. What is the boundary between Mathematics and (say)
Physical Science?
Mathematics has been described here as a formal development of ideas
suggested by phenomena. Some portions of this description apply to other
sciences, notably to Theoretical Physics. These, too, make considerable use
of models and forms-and hence of Mathematics. There, however, one
develops only the forms that seem to be useful in fitting a specific type of
phenomena. If it turns out that they don't fit, they are discarded. In
Mathematics, one pursues the forms whichever they may lead.
10. Platonism 447

10. Platonism
Our view of Mathematics provides both for formal concepts and for guid-
ing ideas, and hence accounts for the role of intuition in Mathematics. It
is, however, in sharp contrast to all variants of Platonism. Let us consider
this contrast.
We are not concerned with the actual historical doctrines formulated by
Plato but with the current views going under his name. These views typi-
cally hold that Mathematical concepts are about externally given objects
which therefore impose their nature upon the results of Mathematics.
That Euclidean geometry can lead to Platonism is clear. This geometry is
not "about" the imperfect lines and figures which we may draw on an
uneven blackboard or on crinkly paper. It must be "about" something, so
it is about ideal straight lines, perfectly rounded circles and absolutely flat
planes which have an objective existence. Once formulated so, this doc-
trine extends naturally to encompass ideal numbers and their properties,
functions and their derivatives and finally the whole world of sets. Thus
Platonism in Mathematics views Mathematics as dealing with a domain of
abstract or ideal "objects" which have being independent of our thought
about them and whose being therefore determines what can be truly
thought about them.
There are various versions of this view; I will follow here the careful
distinctions used by Michael Resnik on page 162 of his book Frege and
the Philosophy of Mathematics. First a methodological Platonist is one
whose activities fully endorse all the standard, infinite, and non-
constructive methods of Mathematics, as if Mathematics were "dealing
with a mind-independent infinite domain of abstract entities". An ontolog-
ical Platonist is one who "recognizes the existence of numbers, sets, ... on
a par with ordinary objects". An epistemological Platonist holds that
"knowledge of Mathematics objects is . . . in part based upon a direct
acquaintance with them". Finally a realist "believes that the objects of
Mathematics ... exist independently of us and our mental lives".
These several views are considerably different. First observe that the
practise of Mathematical research does require deep concentration and
complex understanding of what can follow from known axioms and
theorems. This process often may depend on vivid intuitive imagination
that the concepts at issue concern objects which are really "there". This
description of Mathematical practise has been called "mythological Pla-
tonism"; it seems to be essentially the same as the methodological Platon-
ism defined above. It is our view that it is not a philosophical doctrine
about the nature of Mathematics, but a description of the process of
Mathematical research. As such, it is an appropriate start on such a
description. For example, if I wish to make (as I have done) extensive
computations of the cohomology groups of the so-called Eilenberg-Mac
Lane spaces K( '!T,n), I know that such a space is defined to be one in
which there is just one homotopy group '!T in dimension n, and I recall the
448 XII. The Mathematical Network

theorem that this specification completely determines the cohomology and

homology groups of the space. Therefore my computations are much
assisted by an unquestioned belief that the spaces are there before me,
firm and unyielding (even though I also know that there can be several
different ways of realising one such space K( 'IT,n )). When my calculations
are done, I can return to the world of "ordinary" topological spaces.
There I know that such a space X ordinarily has several homotopy
groups-a fundamental group 'IT\(X) and a whole string of higher dimen-
sional homotopy group 'lTn(X), together with certain connections between
them, given by certain "obstruction cocycles". I can then recall that the
Eilenberg-Mac Lane spaces are an abstract but convenient way of
describing all the natural "operations" which one can apply to the coho-
mology of ordinary spaces. In their turn, many of these "ordinary" topo-
logical spaces may be pathological-they are accidental instances of a gen-
eral definition of topology which was really devised in order to get at
quite specific properties of continuous functions and topological mani-
folds. In their turn these manifolds in "reality" are not just topological;
they actually have a complex structure or a Riemann metric or are
embedded in Euclidean space-and we have for the time being forgotten
about that structure in order to concentrate just on the concept "topologi-
cal manifold".
In other words, Mathematics deals with a heaping pile of successive
abstractions, each based on parts of the ones before, referring ultimately
(but at many removes) to human activities or to questions about real
phenomena. The advance of Mathematical understanding depends on the
contemplation of each abstraction in itself. That is why the abstractions
are there-to get rid of the "real" things which are extraneous to the par-
ticular questions at issue. In other words, the elaborate formal nature of
Mathematics which we have observed requires us to deal with highly
abstract concepts "as if" they are real. This is mythological Platonism. It
is a description of the process of Mathematical understanding. It has
nothing to do with ontology.
Ontological Platonism ascribes "reality" to the "objects" of Mathemat-
ics. This assumption first developed long ago when these objects were just
numbers and geometrical figures (lines and circles). But now the objects of
Mathematics come in such profuse abundance that they would hopelessly
overpopulate reality. Moreover, all these new objects were considered not
because they had been discovered "out there" but because they had
proved to be handy means of understanding counting and moving. We
have found no place in the practice of Mathematics and its applications
where there is any use for the notion "real Mathematical object". The
notion "Mathematical form" is completely sufficient.
Epistemological Platonism involves some sort of direct acquaintance
with Mathematical objects, perhaps by some means not included in the
five senses of sight, hearing, taste, smell, and touch. This doctrine seems
to me to be speculation. Nowhere does such acquaintance, sensory or oth-
II. Preferred Directions for Research 449

erwise, enter into the abstract practice of Mathematics as we have

described it. What is there is not acquaintance with objects, but the play
of ideas.
The process of building all of Mathematics out of sets has led to a
related doctrine which might be called Set-theoretic Platonism: There is a
world of real and objective sets, pictured in the cumulative hierarchy,
which we can apprehend. From this apprehension flows the axioms for
sets and the deeper understanding of their properties-and thence, all the
parts of Mathematics. This is a doctrine often espoused by set-theorists.
For example, Kurt Godel held firmly to the idea that sets are real and
that their properties could be known; for example, he held that the axiom
of choice is true, while the continuum hypotheses is false, as is the axiom
V=L of constructibility (every set is constructible, in the sense of §XI.8).
This realist view contrasts sharply with our views on set theory. It seems
to us that realism in sets faces great difficulties. First, it is open to all the
objections made above to ontological and to epistemological Platonism.
Second, that "apprehension" of sets is pretty obscure, since abstract sets
are not apprehended by our usual senses. Third, there are many viable
variants of the Zermelo-Fraenkel axioms, to say nothing of intuitionistic
set theory. Fourth, because of the artificial constructions required this doc-
trine does not account well for the actual "objects" studied by Mathemati-
cians. For example, holomorphic functions of a complex variable are cen-
tral objects of Mathematics-and they are certainly not to be understood
well as "sets of ordered pairs".
In short, save for mythology, all the variants of Platonism shatter on the
actual practice of Mathematics.

11. Preferred Directions for Research

A thorough description or analysis of the form and function of Mathemat-
ics should provide not only insights into the Philosophy of Mathematics
but also some guidance in the effective pursuit of Mathematical research.
The latter is a subtle task, on which we may make a few tentative com-
ments, omitting the psychological aspects (see Hadamard [1954]).
Mathematics, in our description, rests on ideas and problems arising
from human experience and scientific phenomena and consists in many
successive and interconnected steps in formalizing and generalizing these
inputs. Thus Mathematical research can be directed in a wide variety of
overlapping ways:
(a) Extracting ideas and problems from the (scientific) environment,
(b) Formulating ideas;
(c) Solving externally posed problems;
(d) Establishing new connections between of Mathematical concepts;
450 XII. The Mathematical Network

(e) Rigorous formulation of concepts;

(f) Further development of concepts (e.g., new theorems);
(g) Solving (or partially solving) internal Mathematical problems;
(h) Formulating new conjectures and problems;
(i) Understanding aspects of all of the above.

This listing does put special emphasis upon extracting and formulating
ideas and on understanding their import: this is in some contrast to the
traditional view that research consists primarily of finding new theorems
(item (f)). But all of these tasks are hard; the search for new ideas and
new formalizations is inevitably adventurous and uncertain. Here are
some examples.
(a) The idea of using a Fourier series to represent periodic functions
was extracted from the study of heat and eventually led to a better under-
standing of the notion of a function and to an extensive and still continu-
ing development of harmonic analysis (as under item (f)). More recently,
the notions of game theory were extracted from practical concerns: they
have had extensive applications in Mathematical economics, but some-
what less influence in Mathematics proper.
(b) The ideas of complex number, group, and set were first disentan-
gled and formulated in the 19th century. They have each proved to be
extraordinarily important; in particular the study of groups has led to all
sorts of interconnections with other parts of Mathematics, for example in
the harmonic analysis on a group.
The concept of a topological space was developed in its present general
form only after the wide exploration of all sorts of particular examples-
and it has served to establish many new connections with complex
analysis, differential geometry, algebraic geometry, and Galois theory, for
instance. Other formal concepts may be developed "before their time" to
find use only much later. For example, the notions of lattice theory were
found about 1900 by Dedekind and others, but did not at that time find
any noticeable resonance. The same notions, when rediscovered by Gar-
rett Birkhoff and Oystein Ore in the early 1930's, were immediately put to
use in projective geometries, continuous geometries and in the analysis of
subobjects of algebraic systems. It would seem that by 1930 there were at
hand more uses for such abstract notions. Subsequently, after a lively
decade of research, lattice theory became of less central interest to alge-
braic developments-it may be because the principal uses were already
worked out and the remaining problems were artificial. There are many
other examples of failure, success, and partial or temporary success in the
introduction of new mathematical concepts. Exploring the unknown is
bound to be an adventurous and chancy business!
(c) Solving problems arising in science is a major activity of applied
Mathematics. There are many examples not covered in our text-but the
development of the calculus to handle the problems of celestial and ter-
restrial mechanics is an outstanding example.
11. Preferred Directions for Research 451

(d) We have seen all sorts of discoveries of interconnections. Linear

algebra ties in with higher dimensional geometry, linear differential equa-
tions, functional analysis, and Galois theory. Riemannian geometry
develops applications to relativity theory. The phase spaces of mechanics
are tied to the tangent bundles of differential geometry. Each such tie
leads to better understanding.
(e) The rigorous, £-8 formulation of the calculus in the 19th century
was a development of great significance. It was necessary to the calculus
itself, in the understanding of the role of functions; it was a necessaI)
prerequisite to the development of complex analysis, for which clear con-
cepts of convergence and differentiability were essential; it was a prelim-
inary to much wider (unanticipated) subsequent abstract developments.
There have been other analogous steps toward rigor. The first steps
toward algebraic topology (Riemann, Poincare, Schoen flies) were inexact;
their correction (Brouwer, Alexander) made further development possible.
Hilbert's precise treatment of the foundations of Euclidean geometry was
the starting point of a lively extension of axiomatic methods. Algebraic
geometry had a steady development until about 1930, but by that time its
intuitions were so convoluted that they could not be communicated
beyond a group of specialists. Extensive efforts (Krull, van der Waerden,
Zariski, Weil, Chevalley, Serre, Grothendieck) provided a careful and
more extensive foundation on which remarkable further progress has been
possible. Currently, some parts of geometric topology may be in need of
similar reformulation. Geometers are often prone to emphasize the
importance of intuition (= ideas). These geometric ideas are important;
however they cannot stand by themselves, but require rigorous formula-
tion so that they can be clearly communicated and used.
(f) Solving problems can have wide effects. For centuries, Mathemati-
cians tried to prove the parallel axiom from the other axioms of plane
geometry; the negative solution represented by non-Euclidean geometry
not only settled this problem but opened up an array of new possibilities
for the notion of space. Similarly, the attempt to prove the continuum
hypothesis led to a new method of forcing in set theory, with unexpected
connections to geometry. But not all famous problems lead to such exten-
sive effects.
In sum, Mathematical research is not just the proving of theorems.
But what about the choice of direction or speciality? Recently, the
increasing number of active Mathematicians has led to the recognition of
a large number of highly specialized fields of study. An individual choice
of field might depend on many considerations, including for example:

(i) Habit, and talent,

(ii) Authority,
(iii) Fashion,
(iv) Opportunity,
(v) Insight.
452 XII. The Mathematical Network

The influence of habit is clear: it is easiest to work in the field that one
learned first; moreover the field may have been chosen to fit the talent of
the individual. Some Mathematicians are natural analysts (good at
approximations), some are algebraically minded (manipulations of formu-
las), some are inspired by applications, and some have well developed
geometrical intuitions. However, current specializations are much more
specific than these varieties of talent.
Some specialities dry up, but not the mainstream of Mathematics. The
voice of authority recommends work in the "mainstream"-a handy but
inexact label to cover the essential portions of our subject. Our chapters
have described some of the sources of these mainstreams: Number theory,
geometry, calculus, algebra, mechanics, and complex variables (with logic
on the side). Research in mainstream topics is likely to be relevant, but it
is also likely to be difficult-much of the streambed has already been well
raked over. The evidence suggests that it is also useful to keep an eye on
possible new sources!
In the nineteenth century, synthetic projective geometry was in fashion,
and the fashion persisted till at least 1935, when some graduate schools
still required courses in both synthetic and analytic projective geometry.
Today, no one is concerned about the difference in method, and there is
little attempt to find new theorems by either method. Today, graph theory
is in high fashion, perhaps because of uses for computers or purported
applications to social science.
Some fashions are deeper, and depend on new concepts which open
new opportunities. Thus the remarkable properties of holomorphic func-
tions of one complex variable, as revealed in the work of Cauchy and
Riemann, put the study of these functions in a central position in
Mathematics for a considerable period (say 1854-1930). Toward the end
of this period, most of the opportunities had been explored and the sub-
ject became the center of authority; a major (but narrow) objective
appeared to be that of better understanding the big Picard theorem by
finding more elementary proofs for it. By now, the emphasis has shifted,
so that there is relatively much more attention to functions of several
complex variables.
The calculus of variations was another central interest in analysis,
because the precise methods of Weierstrass made possible a careful study
of this calculus and the more general variants (such as the problem of
Bolza) of its standard problems. This emphasis died down about 1930, but
the field was rapidly reinvigorated by new ideas; first the use of topology
(Morse theory) and then the applications to optimal control.
Finite group theory is a striking example of a field developed by a new
opportunity. This field has been quite inactive, and seemed subordinate to
related fields such as algebraic groups. Then powerful new techniques
(Hall-Higman, Thompson and Feit-Thompson) appeared. These tech-
niques suddenly made it possible to imagine a determination of all finite
12. Summary 453

simple groups. This became the specific program of the whole field for the
period 1962-1982. There are many other examples of such programs of
special research-not always so successful.
This is also an example of insight: With these new ideas, such and such
should be possible. There are many smaller and more specific examples of
such insight by individual Mathematicians. There are also major exam-
ples. Riemann's introduction of Riemann surfaces; Hamiiton's use of
canonical coordinates; Galois' recognition of the use of groups; the
emphasis upon power series by Weierstrass. These are examples of the
important role of ideas.
Mathematics develops from ideas in a network of interlocking formal
systems. This accounts for the inevitable specializations of research limited
to one particular node of the network-but it also indicates that such spe-
cialization is by itself not enough. One needs also awareness of the
relevance of other related parts of the network. Understanding Mathemat-
ics goes beyond specializations.

12. Summary
Now we return to the six questions raised in the introduction, pp. 1-4.

Origin. There are many origins for Mathematics: In the practices which
develop in various human activities, leading to procedures, to ideas, and
then to formal rules; in the questions raised by the scientific study of
phenomena old and new; finally, in the human capacity to form ideas, to
extract information, and to make generalizations and abstractions. These
sources continue to supply new material for Mathematical thought.

Organization. The longstanding concern with number, space, time, and

motion leads to corresponding branches of Mathematics: arithmetic,
geometry, calculus and mechanics, respectively. But these branches turn
out to be intimately connected with each other: The same real numbers
arise from completing the rationals and from measuring the line or the
circle. Geometry interacts with calculus in differential geometry. The con-
cepts of calculus are precisely those needed by and forshadowed in
mechanics and its Hamiltonian development. The four branches are sup-
plemented by others. The combinations of numbers or of equations or of
vectors or of differential operators demand the formal algebraic rules of
their manipulation. Hence arises the branch, algebra. With its emphasis
on the formal it becomes clear that Mathematics is no longer the science
of number, space, time, and motion-it is the study of formal concepts
and systems suggested by number, space, time, and motion, and also sug-
gested by other activities: proving, computing, composing, and decompos-
ing. Then geometry is no longer just the analysis of the "usual" space, but
454 XII. The Mathematical Network

becomes the use of spatial intuitions and ideas to analyse surfaces

extracted from space as well as analytic objects which can be understood
by the appropriate geometric forms. The analysis of change, functions,
and summation initiated by the calculus has remarkable extensions to
holomorphic functions and to distributions and measures which are not
functions at all. In short, once Mathematics is recognized for what it is,
the study of ideas expressed by form, it acquires a wide variety of
branches and subfields, some characterized by the objects exhibited. Since
the same form can have many exemplars, there can be many close con-
nections between different subfields. This is what probably accounts for
the remarkably ramified uses of group theory: The forms presented by the
composition of substitutions occurs whenever there are operations to be
composed.
It is for these reasons that our study (especially §4 in subjects, special-
ties, and subdivisions) did not provide any single neat table of organiza-
tion, with explicit lines of command, control, and communication. Some
branches of Mathematics are defined by axioms (ring theory), some by
concepts (holomorphic function) and some by intended application
(mechanics), and they all need to communicate about common formal
concerns.

Formalization. How are the forms in Mathematics derived? Some are sug-
gested by the facts, but they become forms only when extracted from the
facts, next considered vaguely as ideas and then finally pinned down by
meticulous definitions or axioms. Other Mathematical forms are
abstracted from more elaborate pieces of Mathematics, in order the better
to understand those pieces. It may be for this reason that simple forms,
like the forms of group theory, were discovered late in the development of
Mathematics: They required first exemplification by more concrete forms
in algebra and geometry. Mathematical forms are both discovered (within
the examples) and invented (developed by thought and ideas).
It is the formal nature of Mathematics which makes it interpersonal,
objective, and exact.
Dynamics. Mathematics involves repeated new discoveries, but there is no
one simple description of the forces driving these discoveries: Curiosity
about scientific questions, famous problems, or the wish to generalize.
What is common is perhaps the desire to understand: To understand what
that question from science really involves, or to see why that old problem,
though it appears simple, is so subtle, or to understand that related
features apparent in several different situations have a common explana-
tion. In this sense the desire to understand is the most important dynamic
for the advance of Mathematics.
Foundations. Mathematics has access to absolute rigor-because it is about
form and not about fact. However, there is no single and absolute founda-
12. Summary 455

tion for Mathematics. Any such fixed foundation would preclude the
novelty which might result from the discovery of new form. A form is any
development which proceeds by rule rather than by appeal to fact as to
meaning. Among the many contemporary discussions of the philosophy
of Mathematics we cite a number of authors: Bernays [1935], Curry
[1951], Dummett [1977], Godel [1947], Goodman [1979], Kitcher [1983],
Lehman [1979], Mac Lane [1981], Quine [1963], Resnik [1980], Robinson
[1965], Steiner [1975], Weyl [1949], Wilder [1981], and Wittgenstein [1964].
However, none of the usual systematic foundations or philosophies, as
we have listed them in the introduction, seem to us satisfactory. They may
be summarized (too briefly) as follows:

Logicism. Mathematics needs logic to provide firm canons of proof, but

logic does not tell us which proofs to canonize; Mathematicians must
search for the most illuminating proof. What matters most are proofs
which are "deep" -but not in a logistic sense of depth. The logical rules of
inference provide a careful syntax for each proof, but they do not serve to
indicate the steps that are crucial-these points where one understands
why the proof works. For this reason, Mathematics is not just logical
deduction. Moreover, Mathematics requires not just the rules of logic but
also some non-logical axioms; in particular (some version of) the axiom
of infinity. This is a (generally recognized) technical reason why
Mathematics is not a branch of logic, as was originally asserted by Frege
and Bertrand Russell.
Set theory is strong enough to provide a formulation of most of
Mathematics but this provision is often artificial and does not yield a
"coordinate-free" or invariant foundation for Mathematics. Moreover,
there is no unique notion of a "set".
Platonism is a useful mythology and a speculative ontology.
Formalism. Strict formalism can't explain which of many formulas
matter and hasn't yet absorbed the profound consequences of Godel's
incompleteness theorem. The formal aspect is essential to Mathematics,
but the choice of form is determined by ideas and experience.
Intuitionism can be dogmatic, when it endeavors to restrict Mathemati-
cal intuitions to the whole numbers. This simply fails to account for the
many other sources of Mathematical ideas. On the other hand, intuition-
ism leads to important Mathematical structures which are related, through
topos theory, to geometry and sheaf theory. Thus intuitionism has
revealed important Mathematical structure.
Constructivism can be narrowly dogmatic (only constructive existence
proofs are allowed), but can be more liberally interpreted as an emphasis
on the algorithmic aspects of Mathematics.
Finitism recognizes the finite character of Mathematical argument, but
does not account for the extraordinary success of arguments with infinite
sets and limiting processes.
456 XII. The Mathematical Network

Empiricism cannot succeed in exhibiting an empirical exemplar for each

piece of mathematical form; as we have repeatedly seen, the same
Mathematical form can have many different exemplars, and that form can
be many steps removed from the original facts. Mathematics originates
not just in facts, but also in human activities and in scientific conundrums.
The opposite of empiricism is the view that Mathematics is in part a
search for austere forms of beauty, and that Mathematical developments
can be splendid even though they have no conceivable practical use. Such
views are often expressed by the adjective "elegant". For example,
Euclid's proof that there are infinitely many primes or the proof of the
Cauchy-Goursat theorem (§X.6) are said to be elegant because each gets
a "deep" result with a minimum of complexity. Elegance can be viewed
as a desirable property of a well-constructed Mathematical form.
Each of these philosophies illuminates a relevant aspect of Mathemat-
ics, but none of them is remotely adequate as a description or foundation
of the actual extensive network of Mathematics. Instead, our study has
revealed Mathematics as an array of forms, codifying ideas extracted from
human activities and scientific problems and deployed in a network of
formal rules, formal definitions, formal axiom systems, explicit theorems
with their careful proof and the manifold interconnections of these forms.
More briefly, Mathematics aims to understand, to manipulate, to develop,
and to apply those aspects of the universe which are formal. This view, as
expounded in this book, might be called formal functionalism.
Bibliography

Ahlfors, Lars [1966]. Complex Analysis: An introduction to the theory of analytic

functions of one complex variable, 2nd ed. 317 pp. New York: McGraw-Hill
Book Co.
Artin, Emil [1959]. Galois theory, edited by Arthur N. Milgram. 2nd ed. 82 pp.
South Bend, Indiana: University of Notre Dame Press.
Bachmann, Friedrich [1973]. Aufbau der Geometrie aus dem Spiegelungsbegriff. 2nd
en!. ed. 374 pp. Heidelberg-New York: Springer-Verlag.
Barwise, Jon [1977]. Handbook of Mathematical Logic. Studies in Logic and the
foundations of Mathematics 90, 1165 pp. Edited by Jon Barwise. Amsterdam:
North Holland Publishing Co.
Bemays, Paul [1935]. Sur la platonisme dans les mathematiques. L'enseignement
mathematique 34, 52-69.
Birkhoff, Garrett [1967]. Lattice theory, 3rd ed. 418 pp. The American Mathemati-
cal Society Colloquium Publications 25.
Birkhoff, Garrett and Saunders Mac Lane [1977]. A survey of modern algebra, 4th
ed. 500 pp. New York: Macmillan.
Bishop, Errett [1967]. Foundations of constructive analysis. 370 pp. New York:
McGraw-Hill Book Co.
Bott, Raoul and Tu W. Loring [1982]. Differential forms in Algebraic Topology. 331
pp. Heidelberg: Sp~inger- Verlag.
Bourbaki, N. [1946]. Elements de Mathematique. Premier Partie, les structures fon-
damental de l'analyse. Livre III: Topologie generale, chap. 2: structures uni-
form. Actualites Sci. et industrielles, #838. Paris: Hermann & Cie.
Caratheodory, Constantin [1965]. Calculus of variations and partial differential
equations of the first order. Part I [1965], 171 pp. Part II [1967], pp. 175-398.
San Francisco: Holden-Day, Inc.
Curry, Haskell B. [1951]. Outlines of a formalist philosophy of mathematics. Stud-
ies in logic and the foundations of mathematics. 75 pp. Amsterdam: North Hol-
land Publishing Company.
Davis, Philip 1. and Reuben Hersh [1981]. The mathematical experience. 440 pp.
Boston: Birkhauser.
Dieudonne, Jean [1977]. Panorama des Mathematiques pures, Ie choix boubachique.
302 pp. Paris: Gauthiers-Villars.
_ _ [1978]. The difficult birth of mathematical structure 1840-1940. Scientific
culture in the contemporary world. pp. 7-23. Milan: Scientia.
Dodson, C. T. J. and Tim Poston [1978]. Tensor geometry. The geometric
viewpoint and its uses. 598 pp. London and San Francisco: Putnam.
458 Bibliography

Dummett, Michael [1977]. Elements of intuitionism. Written with the assistance of

Robert Minio. 467 pp. Oxford, England: Clarendon Press.
Eklof, Paul C. [1976]. Whitehead's problem is undecidable. A mer. Math. Monthly
83775-788.
Feynman, Richard, Robert Leighton, and Matthew Sands [1964]. The Feynman
lectures on physics. 3 vols. 513 pp., 569 pp., and 365 pp. Reading, Mass.: Addi-
son Wesley.
Frank, Philip and Richard V. Mises [1930]. Die DifJerentialgleichungen und
Integralgleichungen der Mechanik und Physik, 2nd ed. Erster, mathematischer
Teil. 916 pp. Braunschweig: F. Vieweg & Sohn.
Giirding, Lars [1977]. Encounter with mathematics. 270 pp. New York: Springer-
Verlag.
Gleason, Andrew M. [1966]. Fundamentals of abstract analySiS. 404 pp. Reading,
Mass.: Addison-Wesley Pub. Co.
G6del, Durt [1940]. The consistency of the continuum hypothesis,. Annals of
Math. Studies No.3. 66 pp. Princeton University Press.
_ _ [1947]. What is Cantor's continuum problem. A mer. Math. Monthly 54,
515-525.
Goodman, Nicholas D. [1979]. Mathematics as an objective science. Am. Math.
Monthly 86, 540-551.
_ _ [1981]. The experiential foundations of mathematical knowledge. Hist. Phi-
los. LogiC 2, 55-65.
Guggenheim, Heinrich W. [1967]. Plane geometry and its groups. 288 pp. San
Francisco: Holden-Day, Inc.
Hadamard, Jacques [1954]. An essay on the psychology of invention in the
mathematical field. 145 pp. New York: Dover Publications, Inc.
Halmos, Paul R. [1958]. Finite dimensional vector spaces, 2nd ed. D. van Nostrand;
reprinted 1974.200 pp. New York: Springer-Verlag. [1st ed., Princeton Univer-
sity Press, 1942J.
Hamilton, Alan G. [1978]. Logic for mathematics. 224 pp. Cambridge-New York:
Cambridge University Press.
Hardy, G. H. and E. M. Wright [1954]. An introduction to the theory of numbers.
3rd ed. 419 pp. Oxford, England: Clarendon Press.
Hausdorff, Felix [1914]. Grundzuge du Magenlehre. 476 pp. Leipzig: W. A. Gruyter
& Co.; New York: Reprinted Chelsea Pub. Co. 1945.
Hilbert, David [1899; 1971]. Foundations of Geometry, 2nd ed. Translated from the
tenth German edition by Leo Unger. 226 pp. LaSalle, Illinois: Open Court.
Hungerford, Thomas W. [1974]. Algebra. 502 pp. New York: Holt, Rinehart and
Winston.
Hurwitz, A. and R. Courant [1964]. Functionentheorie, 4th ed. 706 pp. Heidelberg:
Springer-Verlag.
Johnstone, Peter [1978]. Topos theory. 367 pp. New York: Academic Press.
_ _ [1982]. Stone spaces. Cambridge studies in advanced mathematics, Vol. 3.
370 pp. Cambridge, England: Cambridge University Press.
Keisler, H. J. [1976]. Elementary calculus. An approach using infinitesimals. 880
pp. Boston: Prindle, Weber & Schmidt.
Kleene, S. C. [1952]. Introduction to metamathematics. New York: Van Nostrand.
Kitcher, Philip [1983]. The nature of mathematical knowledge. 287 pp. New York
and Oxford: Oxford University Press.
Klein, Felix and Arnold Sommerfeld [1965]. Uber die Theorie des Kreisels.
Bibliotheca Math Taubneriana, Vol. I. 966 pp. (Original published in four
volumes 1897, 1898, 1903, 1910.) New York: Johnson Repr. Corp.
Bibliography 459

Kock, Anders [1981]. Synthetic differential geometry. London Math. Soc. Lecture
Notes Series 51. 311 pp. Cambridge and New York: Cambridge University
Press.
Landau, Edmund [1951]. Foundations of analysis. The arithmetic of whole,
rational, irrational, and complex numbers. Translated by F. Steinhardt. 134 pp.
New York: Chelsea Pub. Co.
Lang, Serge [1967]. Introduction to differentiable manifolds. 125 pp. New York:
Interscience (John Wiley & Sons].
Lehman, Hugh [1979]. Introduction to the philosophy of mathematics. 169 pp.
Totowa, New Jersey: Rownan and Littlefield.
Lightstone, A. H. and Abraham Robinson [1975]. Non-archimedean fields and
asymptotic expansions. 204 pp. North Holland Mathematical Library. Vol. 13.
Amsterdam-Oxford: North Holland Publishing Company.
Mackey, George W. [1978]. Unitary group representation in phYSiCS, probability and
number theory. 402 pp. Math lecture notes series # 55, Reading, Mass.
Mac Lane, Saunders [1963]. Homology. Die Grundlehren der Math. Wissenschaf-
ten, Vol. 114. 422 pp. Heidelberg: Springer-Verlag.
_ _ [1971]. Categories for the working mathematician, Graduate texts in
mathematics, Vol. 5. 262 pp. Heidelberg: Springer-Verlag.
_ _ [1981]. Mathematical models: a sketch for the philosophy of mathematics.
Am. Math. Monthly 88: 462-472.
Mac Lane, Saunders and Garrett Birkhoff [1979]. Algebra, 2nd ed. 586 pp. (1st ed.
1967) New York: Macmillan Publishing Co.
Massey, William S. [1967]. Algebraic topology: An introduction. 261 pp. New York:
Springer-Verlag.
Monna, A. F. [1975]. Dirichlet's principle. A mathematical comedy of errors and
its influence on the development of analysis. 138 pp. Utrecht, The Netherlands:
Oosthoek Scheltema & Holkema.
Myhill, John [1972]. What is a real number? Am. Math. Monthly 79: 748-754.
Narasimhan, R. [1985]. Complex analysis in one variable. 216 pp. Boston: Burk-
hauser.
O'Neill, Barrett [1966]. Elementary differential geometry. 411 pp. New York and
London: Academic Press.
Osgood, William Fogg [1937]. Mechanics. 495 pp. New York: The Macmillan Co.
Paige, Leigh [1928). Introduction to theoretical physics. 587 pp. New York: D. van
Nostrand.
Pars, L. A. [1965). A treatise on analytical dynamiCS. 641 pp. New York: John
Wiley & Sons.
Pontryagin, L. S. [1939). Topological groups. Translated from the Russian by
Emma Lehmer. Princeton math series vol. 2. 299 pp. Princeton, New Jersey:
Princeton University Press.
Quine, W. V. O. [1963]. From a logical point of view. Logico-philosophical essays,
2nd ed., rev. 184 pp. New York: Harper and Row.
Resnik, Michael D. [1980]. Frege and the philosophy of mathematics. 243 pp.
Ithaca, New York: Cornell University Press.
Robinson, Abraham [1965]. Formalism 64, pp. 228-246. In Proc. Internat.
Congress for Logic, Methodology, and Philosophy. Jerusalem 1964. Amsterdam:
North Holland Pub. Co.
Russell, Bertrand A. W. [1908]. Mathematical logic as based on the theory of
types. Amer. J. Math. 30: 222-262.
Shoenfield,1. R. [1975). Martin's axiom. Am. Math. Monthly 82: 610-619.
Sondheimer, Ernst and Alan Rogerson [1981). Numbers and infinity. A historical
460 Bibliography

account of mathematical concepts. 172 pp. London and New York: Cambridge
University Press.
Spivak, Michael [1965]. Calculus on manifolds. 144 pp. New York and Amster-
dam: W. A. Benjamin Inc.
Steiner, Mark [1975]. Mathematical knowledge. 164 pp. Ithaca, New York: Cornell
University Press.
Titchmarsh, E. C. [1932]. The theory of functions. 454 pp. Oxford, England: The
Clarendon Press.
Troelstra, A. S. [1972]. Choice sequences. A chapter of intuitionist mathematics.
Oxford logic guides. 170 pp. Oxford Clarendon Press.
Weyl, Hermann [1949]. Philosophy of mathematics and natural science. 311 pp.
Rev. English ed. Trans. by O. Helmer. Princeton, New Jersey: Princeton
University Press.
_ _ [1923]. Raum, Zeit, Materie. Vorlesungen uber allgemeine Relativi-
tatstheorie, 5th ed. 338 pp. Berlin: Springer-Verlag.
Whitehead, A. N. and Bertrand Russell [1910]. Principia mathematica. Vol. 1. 666
pp. 2nd ed. 1925. 674 pp. Cambridge, England: Cambridge University Press,
1925.
Wilder, Raymond L. [1981]. Mathematics as a cultural system. 182 pp. Oxford-
New York: Pergamon Press.
Wilson, Edwin B. [1912]. Advanced calculus. 566 pp. Boston: Ginn & Co.
Wittgenstein, Ludwig [1964]. Remarks on the foundation of mathematics, 2nd ed.
Edited by G. H. von Wright, R. Rhees, and G. E. M. Anscombe. Oxford, Eng-
land: Basil Blackwell.
Zermelo, Ernst. 1908]. Untersuchungen iiber die Grundlagen der Mengenlehre I.
Mathematische Annalen 85: 261-281.
List of Symbols

'Ix for all x, 163,335,371

::Ix there exists x, 163,335, 371
8 delta, 70, 163
epsilon, 70, 163
'IT pi,95
theta }
phi angles, 222
map, morphism, 46, 129, 386
map acting on an element, 123
?'""' implies, 128, 369
<==? if and only if, 26, 128, 369
1\ and,368
V or,368
7 not, 368
< less than, 13
< less than or equal to, 14
c inclusion }
n int~rsection of sets, 27-28
u umon
complement
o null set, 28, 363
{x I ... } all x such that, 8
E member of, 26
C complex numbers, 116
N natural numbers, 43
Q rational numbers, 57
R real numbers, 102
Z integers, 50
Index

abelian alternating group, 134

-category, 425 amalgamation, 396
-group, 23 amplitude, 266
absolute rigor, 377 analogy, 37, 431
absolute value, 99 analysis, 178
abstraction, 38, 435 -of proofs, 432
acceleration, 29, 158, 260 analytic
action (of a group), 85 -continuation, 348
activity, 35, 409 -function, 312, 337
addition, 9, 45, 97 -geometry, 424
addition formula, 112 -number theory, 350
additive, 186 regular-function, 312, 337
adjoint, 201 and,368
adjoint functor, 398 angle, 65, 95
affine geometry, 109, 134 angular momentum, 262
affine transformation, 109, 134 -velocity, 262, 297
aleph zero, 11 anti-symmetric relation, 14
Alexander, 451 applied mathematics, 414, 422, 426
algebra, 1, 422, 423 approximation, 30, 167, 170
algebra of relations, 132 linear-, 164,330
algebraic arc length, 220
-function, 418 archimedian
-integer, 419 -axiom, 66
-number, 94, 418 -law, 99, 154
-structure, 26 arc sin, 124
algorithm, 439 arguing, 35
almost everywhere, 178 argument of complex number, 115
464 Index

arrows, 386 Bourbaki, 5, 166

assignment, 123, 124 bracket notation for sets, 8, 364
associative, 23 branch, 230, 309
-law, 42, 386 -point, 355
atlas, 239, 241 Brouwer, 4, 451
maximal-, 240, 245-246 building, 35
automorphism, 116, 136 Burnside, 25, 429
inner-, 136-137
Ayres, 431
axiom of choice, 366, 401
-of determinacy, 385 calculus, 29, 150ff, 259, 422
axiom scheme, 365 -of variations, 279, 452
axiomatic canonical form, 217
-set theory, 362 Cantor, 106, 360
-structure, 40 cardinal
axiomatization, 39 -equivalence, 54
axioms on functions, 398 -number, 8, 11,54, 132,367
axis of rotation, 89 cartesian
-coordinates, 424
-product, 129, 392
Banach space, 436 cast out, 10, 52
base category, 386
-(of a bundle), 250 opposite-, 389
-field, 139 Cauchy
-point, 388 -condition, 103
-(of a topology), 352 -Goursat, 329, 331
basis, 108, 186 -integral formula, 338
Betti numbers, 217 -Riemann equations, 313
betweenness, 64, 99 -sequence, 101, 105
Bezout theorem, 435 -theorem, 328, 338, 378
bijection, 8, 133 Cayley, 25, 191,436
bilinear, 171, 194, 204 Cayley-Hamilton theorem, 193
-form, 198 Cayley's theorem, 23, 39, 134
binary operation, 26 centrifugal force, 268
binary relation, 14 chain rule, 154, 168
binomial, 222 chains, 331
-series, 101 chance, 35
-theorem, 44, 101,434 change
Birkhoff, 450 -of base, 270
Bolyai,70 -of basis, 189, 191
Bolza,452 -of coordinates, 243
Boolean algebra, 27, 404, 436 -of fields, 207
Borel-Lebesgue theorem, 165 characteristic, 142
bound (least upper), 15, 102 -function, 399
variable-, 371 -polynomial, 192
boundary, 320, 341 chart, 223, 239-241, 244
-values, 181 Chen, 49
bounded, 163 Chevalley, 451
-comprehension, 373 chief series, 437
-quantifiers, 372 Chinese remainder, 380
Index 465

choice -series of a group, 141

axiom of-, 366,410 comprehension axiom, 56, 314
-function, 366 computational complexity, 439
choosing, 35 computing, 35
circle configuration space, 270
-of convergence, 337 conformal, 117,314
circular disc, 31, 312 congruence
class, 387 -axioms, 65
_Ck , 244 -class, 53, 133
_C 244
X
, -theorems, 66
-field theory, 432 conjugate, 80
clockwise, 82 -complex, 116, 138
clopen, 349 conjugation, 86, 326
closed path, 321, 324 conjunction, 369
-set, 327 connected, 238, 349
closure, 327-328 -components, 354
codomain, 129, 382, 386 conservation
Cohen, Paul, 384--385 -{)f energy, 266
collecting, 35 -of momentum, 264
collection, 7, 36 conservative system, 266
commutative consistency, 379
-diagram, 22, 46, 388 constants of integration, 193
-law, 42 constraints, 272
-ring, 98 constructible
compact, 40, 165-166 -hierarchy, 384
compactness theorem (logic), 375 -set, 384
companion matrix, 215 constructivism, 455
comparing, 7, 35 continuity, 30, 32
complement, 27 -axiom, 66
complete continuous geometry, 437
-metric, 179 continuum hypothesis, 360, 383,451
-ordered field, 103 contour lines, 248, 344
-solution, 293 contradiction, 379
-system of invariants, 216, 221 contravariant, 196, 389
completeness, 15 conundrums, 334
axiom of-, 377 convenient category, 397
-theorem, 375 convergence, 100, 173
completion of systems, 37 absolute-, 334
complex conjugation, 116, 138 uniform-, 335
complex derivative, 312ff convergent, 334
complex number, 114 -series, 102
complex variable, 307 convex set, 332
components, 354 coordinates, 108, 188
composite, 18, 25, 121-122, 124, 130 cartesian-, 424
-arrow, 386 local-, 243
-function, 128 Copernicus, 259
-path,319 coproduct, 395
-relation, 132 correspondence, 8
composition, 131, 147-148 coset, 85, 137
-factors, 141 cosine, 110-113,310
466 Index

contangent -form, 174

-bundle, 249 -geometry, 219ff
-space, 171 total-, 174
-vector, 171,249 dihedral group, 434
counterclockwise, 82 dimension, 108, 186
counting, 7 Diophantine equation, 49, 428
covering, 165 Dirac, 439
-map, 345, 351 direct
universal-, 346 -product, 143
Cramer's rule, 189 -sum, 204, 214, 393
cross section, 250, 253, 353, 401 direction, 86
crystallographic groups, 20 directional derivative, 169, 193, 248
cube root, 116 Dirichlet theorem, 49, 180
cumulative-hierarchy, 362 discrete, 237
cup product, 435 disjoint, 9
curvature, 74, 220, 223 -union, 54, 396
curve, 220 disjunction, 369
curved space, 228 disk, 31, 32, 312
cut distance, 16
Dedekind-,46, 103-104 distortion, 236
cycle, 12 distribution, 439
cyclic distributive law, 43, 98, 437
-groups, 23 divergence theorem, 175
-modules, 216 divergent series, 333
cyclomatic field, 419 division
-algorithm, 48
-ring, 98
data type, 26 domain
decimals, 43 -of a chart, 239
decision theory, 439 -of a function, 129-130
Dedekind, 359 -of a map, 382
-axiom, 66, 103, 331 double
--cut, 46, 103 -arrow, 128
defineable set, 384 -dual, 195
deformation, 323 -integral, 174
degree, 139, 324 dual, 171
deletion of axioms, 38 -basis, 194
delta function, 439 -diagrams, 144, 395
de Morgan law, 370 -pairing, 195
dense, 15 -space, 193
denumberable, II dynamics, 259ff
dependence, 127 -of Mathematics, 454
descriptive geometry, 424
determinacy, axiom of, 385
determinant, 190, 213 eccentricity, 263
diagonal argument, 360, 383 econometrics, 168
diagonal matrix, 191 eigenvalue, 192
diagonal subset, 132 eigenvector, 192
differential Eilenberg-Mac Lane spaces, 447
--equation, 161 electromagnetic potentials, 314
Index 467

elementary -derivative, 177

-divisors, 217 extracted, 418
-topos, 400
ellipsoids, 204
elliptic geometry, 73 factor group, 138, 208
empiricism, 4, 456 factorials, 167
empty set, 28, 132, 363 factorization into primes, 419
end (of a path), 322 Feit, 452
endomorphism, 136, 186 Fermat, 2, 36, 419
energy, 266, 282 -principle (mechanics), 292
potential-, 266 -theorem, 2, 36,49
entire functions, 340, 420 fiber derivative, 288
enumeration, 35 fibered product, 395
epic map, 401 field, 98
epimorphism, 136 figure, 17, 62-63
epistemology, 443, 448 finite limit, 395
equality of sets, 128, 363 finitism, 455
equinumerous, 54 first fundamental form, 227
equipotential, 314 first order logic, 370, 372
equivalence, 399 fluent, 158
-dass, 133 fluid flow, 316
-relation, 133 fluxion, 158
equivalent subobjects, 399 focus, 263
Erdos, 431 force, 159, 260, 265, 314
Erlanger program, 134, 407 forcing, 385
essential singularity, 342 formal, 10,410,414
estimating, 35 -functionalism, 456
Euclidean -power series, 102
-algorithm, 49 -proof,374
-geometry, 63ff formalism, 3, 414, 454
Eudoxus, 76 formalization of Math, 454
Euler formula 44, 126, 369, 371
--equations, 261, 289 foundation
-function, 53 -of calculus, 162
Euler's angles, 296 -of Mathematics, 3-4, 358, 365,
evaluating a function, 194 406, 454
evaluation, 194, 397 four group, 13, 19, 441
even permutation, 134 Fourier series, 181,359,450
exact sequence, 144 fractions, 51
examples, study of, 432 fractional linear transformation, 118, 144,
excess of angle sum, 227 308
exists, 130, 163, 371 frames (moving), 222
exponential, 9, 308, 396 free
exponents, 399 -graded algebra, 211
extended plane, 117 -group, 144, 325
extension, 28, 44, 144 -product, 143
group--, 145 Frege, 20, 455
extensionality, 28, 363 friezes, 20
exterior function, 123, 129, 386,412
-algebra, 211, 212 -field,420
468 Index

function (cont.) Grothendieck, 451

-set, 396 group, 23, 135-138
-space, 187 abelian-, 23
functionalism, 456 -action, 85
functor, 388 cyclic-, 135
-category, 392 finite simple-, 429
contravariant-, 389 free-, 144, 325
functorial, 208, 213 linear-, 134
fundamental permutation-, 13
-form, 227 transformation-, 133-135
-group, 324, 388 groupoid, 326, 387
-theorem of algebra, 116, 340 Gudermann, 430
-theorem of calculus, 177
fuzzy set, 439
balf-tum, 79, 87
Hall-Higman, 452
GLn (V), 134 Hamilton
Galois, 25, 37, 87, 141 -coordinates, 452
-group, 138, 210 -equation, 283
-theory, 437 -Jacobi equation, 292
gauge theories, 259, 278 -principle, 279
Gauss, 70, 223 hamiltonian, 282
-lemma, 320 harmonic, 314
Gaussian curvature, 223, 227 -analysis, 180, 359
general linear group, 38, 134, 190 -function, 180
generalization, 434-435 -motion, 180-181
-from cases, 434 -series, 333
-by modification, 435 Hausdorff space, 33, 238
generalized force, 271 Heine-Borel, 40, 165
genus, 239 hermitian, 204
geodesic, 74 Heyting algebra, 402
geometric series, 333 hierarchy, 359, 362, 366, 384
geometry, 61-92, 219-258, 422 Higman, 452
Euclidean-, 61ff Hilbert, 3, 63, 414, 432
Riemannian-, 266 -space, 201, 414, 437
germ of a function, 351, 355 Hilbert's problems, 428, 430
global properties, 236 historical order, 4
Godel, 3, 449 holomorphic function, 312
-incompleteness, 379 holonomic constraints, 272
Goldbach, 49, 429 homeomorphism, 136
graded homogeneous, 186
-algebra, 211 homology, 331
-ring, 211 -group, 217
-vector space, 211 homomorphism, 24, 46, 86, 135, 344
gradient, 170, 249 homotopic, 323
Gram-Schmidt, 198 human activities, 34
graph, 30, 127, 129,430 Hungarian Mathematics, 430
-theory, 452 hyperbolic geometry, 70
gravitation, 260 hyperboloids, 204
Green's theorem, 176 hyperplanes, 91
Index 469

icosahedron, 134 intension of properties, 22

ideal, 210, 419 interior angles, 67
principal-, 435 intersection, 26, 28, 204
ideas, 34, 287, 409,415, 444 -multiplicity, 435
identity, 386 interval, 163
-law, 386 intrinsic geometry, 226-227
-matrix, 192 intuitionism, 4, 402, 454
-transformation, 21 invariance, 37
if and only if, 27 invariant, 221
image, 132, 137, 189 -factors, 216
imaginary, 118 -formulation, 433
implication, 27, 369 inverse, 9, 132
implicit function theorem, 290 -image, 344,401
incidence axiom, 63 one-sided-, 132
inclusion of sets, 28, 363 -path,322
incompressible fluid, 316 -square law, 260
indefinite integral, 156 two-sided-, 132
independent, 384 involution, 87
-axioms, 70 irrational numbers, 75, 94
independence irreducible, 118, 187
linear-, 186 irrotational flow, 316
-theorem, 384 isomorphism, 46, 136
index up, down, 199 --Df groups, 24
induced isotropy group, 85
-homomorphism, 137
-topology, 237
induction, 44
infinite, 11, 117
Jacobian, 275
-dimensional, 187
jet, 250
-series, 101, 333
join, 29
infinitesimal, 152, 154
Jordan
infinity, 116, 117,364
-canonical form, 217
initial condition, 161
-curve, 321,435,440
initial object, 395
-Holder theorem, 141, 437
injection, 21, 132, 396
injective, 21
inner
-automorphism, 137
-product, 113 Kepler's laws, 158, 260
-product space, 113, 197 kernel, 210
inside, 64, 321 -pair, 137
insight, 438 kinetic energy, 265
instantaneous rate, 154 Klein, 87, 134, 407
integer, 50 Korteweg-de Vries equation, 439
integral Kreisel, 5
definite-, 152-153 Kronecker, 442
-domain, 215 -delta, 194
indefinite-, 156 Krull,451
Lebesgue-, 179 Kummer, 359, 412, 419, 429
Riemann-, 153 Kuratowski, 431
470 Index

labelling, 7 linearity, 113

Lagrange, 37, 269 linearly ordered group, 34
-equation, 267ff lines of force, 314
Lagrangian, 273 Liouville theorem, 340
language, 371, 372 Lipschitz condition, 162
Laplace listing, 7
-equation, 180 Lobachevsky, 70
-operator, 314 local, 236
large cardinal numbers, 385 -coordinates, 243
latitude, 221-222 -homeomorphism, 352
lattice, 29, 428 -uniformizing variable, 347
Laurent series, 341 logarithm, 132, 309
law of logicism, 3, 455
-cosines, II3 longitude, 221
-gravitation, 260 Looman, 317
-sines, 113 loops, 324
-the mean, 164 Lowenheim-Skolmen theorem, 384
least action, 282
least upper bound, 15, 28
Lebesgue integral, 179 Maclaurin series, 339
left magnitude, 75, 94
-adjoint, 398 mainstream, 452
-inverse, 132 manifold, 233-236, 239ff
-side of, 83-84 many-sorted algebra, 26
-vector space, 187 many-valued function, 228ff, 309
Legendre transformation, 286 map, 22, 125
length marginal cost, 151
-of a curve, 220 material implication, 369
-of a vector, 197 matrix, 126, 134
less than, 13, 99 -theory, 424
level lines, 248, 344 maximal atlas, 240, 245-246
lexicographic order, 57 maximum, 163
Lie, 87 -principle, 341
-algebra, 303 mean curvature, 223
-groups, 303 measure, 178
lift, 278 measuring, 35
limit, 100, 153 meet, 29
finite-, 395 membership, 26, 128, 362
-ordinal, 361 Menchoff,317
line metamathematics, 379
-integral, 173 metric space, 17, 179,416
linear Mobius band, 235
-algebra, 185, 187,424 model, 375
-approximation, 167, 170 modular lattice, 437
-combination, 108, 185 modular law, 437
-differential equation, 187 module, 214, 424
-independence, 185 modulus, 52
-map, 214 modus ponens, 374
-order, 14 moment of inertia, 297
-transformation, 108, 186 momentum, 262, 264
Index 471

monic -subgroup, 137

-arrow, 399 -vector, 221
-polynomial, 215, 419 novelty, 438-440
monodromy theorem, 355 n-tuple, 188
monoid,26 null
monomorphism, 136 -set, 28, 95, 363
monster group, 146 -space, 189
Mordell conjecture, 421 nullary operation, 26
morphism, 136 nullity, 189
Morse number
-relations, 182 -field,418
-theory, 452 -theory, 48-49, 422
motion, 17,76-78,87
rigid-, 17, 18,76
moving, 35 On' 134
moving frames, 221 object, 386
multiplication, 97 Occam's razor, 441
-table of a group, 136 one-to-one
multiplicity, 204, 435 -correspondence, 8
intersection-, 435 -into, 21
one-sided inverse, 132
ontological platonism, 448
ontology, 443, 445
natural open
-isomorphism, 195 -circular disc, 31, 32
-number, 7, 42 -covering, 238
-number object, 401 -interval, 30
-transformation, 390, 391 -mapping, 341
nearby, 33, 236 -set, 32
nearness, 99 operation, 26
negation, 369 opposite category, 389
neighborhood, 33, 328, 416 optimal control, 452
nested sequence, 330 orbit, 86
network, 409, 417 order, 23
Newton's laws, 158, 260 -isomorphism, 15, 57
next, 34 -of a group, 23, 135
Noether, 39, 120 -of a group element, 136
non-Euclidean geometry, 70ff, 451 -of a pole or zero, 342, 420
non-holonomic, 272 -type, 15, 57
non-linearity, 439 ordered
non-orientable, 241 -field,99
non-singular, 134, 186 linearly-, 14
-matrix, 190 -pair, 54, 128
non-standard -set, 14,28
-analysis, 155 ordinal
-model,376 -hierarchy, 366
-reals, 155 -number, 56, 58, 366
norm, 197 ordinally equivalent, 51
normal Ore, 450
-orthogonal basis, 113, 198 organization, 1, 453
472 Index

organization of Mathematics, 407, 453 perspective, 439

orientability, 247 phase, 266
orientable surface, 239 -plane, 266
orientation, 82, 83, 241 -space, 110, 162, 276
ornaments, 20 Picard theorem, 343, 452
orthogonal piecewise differentiable, 161, 318
-basis, 113 Planck's constant, 304
-group, 200 planets, 259-260
-matrix, 199 platonism, 3, 385,447-449
normal-, 198 Poincare conjecture, 429
-transformation, 134, 199 point set topology, 438
-vectors, 197 pointed space, 388
osculating circle, 220 pointwise, 193
overlap map, 239, 346 poisson bracket, 303
polar coordinates, 115
pole,342
polynomial
P.D.E., 179-180, 291-295, 426 -function, 126
p-adic number, 421, 436 -ring, 214
pair, 54, 128 -time, 439
pairing, 195 Popper, 91
paradox, 54, 360 poset, 28
parallel, 66 positive, 99
-axiom, 66, 67 -definite, 113
parameter, 17, 170,302 potential, 266
parametrized -energy, 266
-curve, 171,219 -equation, 180
-motion, 17,219 power
part, 35 -series, 168, 336-338
partial differential equation, 169, 179- -sets, 56, 360, 363
180, 292, 426 precession of a top, 301
partial order, 28 precision, 377
partial sum, 102, 333 predicate calculus, 373
Pasch's axiom, 64, 321 preorder, 387
patching, 239, 245, 246 pre sheaf, 253
-map, 239 pretzel, 239
path, 171,219,239,247,322 prime, 48
connected-, 239 -ideal,419
pathological functions, 310 -number theorem, 350
-space, 237 primitive, 328
Peano, 187,311 principal
-arithmetic, 44-47 -axes, 203
-postulates, 41, 43 -axis theorem, 204
Pell's equation, 49 -curvature, 224
pendulum, 274 -direction, 224
double-, 233 -function, 289ff
spherical-, 274 -ideal,435
permutation, 11, 12 -ideal domain, 215
-group, 13, 21 probability, 12, 426-427
perpendicular, 16, 197 problems, 36, 428
Index 473

product, 9, 54, 97, 393 -magnitude, 75

-of groups, 24 -number, 51
-of ideals, 419 ratios, 75
-of numbers, 9 ray, 65
-of objects, 393 realist philosopher, 447
projection, 143, 250, 393 rearranging, 35
-of a product, 393 rectangular path, 332
projective geometry, 439, 452 rectifiable curve, 220
-plane, 235, 240 recursion, 45
proof, 374 -theorem, 45
-theory, 379 reflection, 81
proper orthogonal matrices, 200 reflexive, 14
property, 44, 365, 372 regression, 168, 426
proportions, 185 regular, 220
propositional -analytic, 312, 337
-calculus, 368 -path,220
-connectives, 370 -representation, 135
ptolomaic astronomy, 259 regularity axiom, 365
pulled back, 270, 394, 395 relation, 131-132
punctured binary-, 14
-disc, 341 temary-, 67, 371-372
-plane, 345 relativity theory, 228
pure quatemion, 121 relevance logic, 370
pushout, 396 remainder, 52
Pythagorean theorem, 62, 198, 226 -term, 168
removable singularity, 342
representation, 136
-theorem, 436
quadratic differential form, 227 represented, 206
quadratic form, 204 research, directions of, 449
quantifiers, 30, 100, 163,335,371,411 residue, 341
quantity, 247 -theorem, 343
quantum mechanics, 303 Resnik,447
quatemion, 120-121 Riemann, 350, 452
Quine, 443 -integral, 178
quotient, 208 -metric, 117
-group, 138 -sphere, 117
-module, 214 -surface, 230, 346, 438, 453
-ring, 210 Riemannian geometry, 226
-space, 209 right adjoint, 398
right inverse, 132
right vector space, 188
radian measure, 95 rigid motion, 18, 77
radius of convergence, 336 rigor, 332, 377
rank, 189, 361 ring, 39, 98
-{)f tensor, 211 commutative-, 98
rate of change, 29 -{)f integers, 419-420
rational Rolle's theorem, 164, 179
--canonical form, 217 rotation, 78, 316
-function, 308 rule, 127, 410
474 Index

Russell, Bertrand, 375, 439 slide reflection, 20

-paradox, 55, 360 small
Russell, Scott, 439 --category, 387
-hom set, 389
smooth, 170,415
Saccheri, 70 -manifold, 244
scalar product, 264 Sobolev, 439
scale, 94 solitary wave, 439
Schoenfliess,451 solvable group, 148
Schroeder-Bernstein, 360 sort, 382
Schwarz, 439 span, 209
-inequality, 197 sparse matrix, 189
science, 91, 442 special
second-order language, 372 -linear group, 134
sectional curvature, 224 -orthogonal group, 134, 200
segment, 64 spectral theorem, 204
selecting, 35 Spengler, 76
self-adjoint, 202 spherical
semi-direct product, 146 --coordinates, 223
semi-group, 26 -pendulum, 274
sense, 83 split extension, 385
sentence, 375 splitting field, 139
separation, 364 sporadic groups, 146
-axiom, 238 start (of a path), 322
series, 101, 168, 336-337 stationary, 279
infinite-, 101 step function, 178
Serre, 451 stereographic, 117,315
set, 10, 36, 350 Stokes' theorem, 177
axioms for-, 56, 363-368 Stone representation theorem, 148, 436
-theoretic platonism, 449 subdivision of Mathematics, 1, 422
-with structure, 33 strict implication, 370
shape mapping, 223 structure, 26, 33
shaping, 35 -sheaf, 255
sheaf, 254, 351, 405 -theorem, 44
Shelah,385 subgroup, 23, 136
shift of attention, 433 subjects, 422
side of a line, 64 subobject classifier, 400
Sierpinski space, 238 subset, 27, 129,363
similar matrix, 191, 213 subspace topology, 237
simple substitution, 50, 124
-closed curve, 321 successor, 8, 34, 43
-group, 146,229 -ordinal, 363
-harmonic motion, 180 sum, 9, 97
simplectic space, 435 surface, 236, 244
simply connected, 326 surjection, 133
simultaneous equations, 189 surveyor, 411
sine, 110-113, 310 symmetric
singleton, 363 -group, 21, 134
singular homology, 331 -matrix, 202
singularity, 341 symmetry, 18, 19, 113
Index 475

symplectic manifold, 303 transitive, 14, 366

syntax, 127 -group, 85
-relation, 14
translation, 18, 77-78, 80
-principle, 381
table transpose matrix, 190
-of integrals, 156 transpositions, 134
-of values, 127, 129 transversal, 64
tangent, 248 tree, 361,402
-bundle, 249, 250 triangle, 63
-line, 167,435 -axiom, 17
-plane, l71 -inequality, l7, 99, 198
-space, l71 trichotomy, 13
-vector, l71, 249 trigonometric series, 181, 359
Tarski, 383 -functions, 110, 310
tautology, 370, 373 trigonometry, 11 0-113
Taylor's trilinear, 208
-series, 168, 337, 339 truth, 440
-theorem, 167 -tables, 369
teleology, 422 -values, 369, 403-404
tensor, 211 two-body problem, 261
-product, 205 two-sided inverse, 132
term, 371 two-sorted language, 382
-by term, 335 Tychonoff topology, 394
terminal object, 395 types, 56
ternary operation, 26, 109, 372
ternary relation, 64, 371-372
tertium non datur, 4 unary operation, 26
theory unbounded, 15
-of connections, 259 understanding, 38, 377,431
-of groups, 434 uniform
thermodynamics, 169 -continuity, 163
Thompson, 452 -convergence, 333
timing, 35 -space, 335
top, 295-301 uniformization theorem, 355
topological group, 33 uniformizing parameter, 347
topological space, 32, 33, 236--238, union, 26, 28, 364
356 unique factorization, 49, 419
torque, 268 unit
torsion, 221 -of a group, 23
-coefficients, 217 -of magnitude, 74
torus, 135, 232-233, 242 -of multiplication, 98
total differential, 170 -vector, 223
totality, 390 unit normal vector, 223
tractrix, 74-75 unitary
trajectory, l71 -matrix, 200
transcendental, 94 -space, 200
transformation, 21, 134, 188ff universal, 137, 205
-group, 21, 133-135 -algebra, 26, 434
transit, 411 -arrow, 393
476 Index

universal (cont.) Waring's problem, 49

-covering space, 346 wave equation, 180, 292-293
-what-not, 206 Weierstrass, 452, 453
universe, 56 -condition, 103
upper bound, 15 weighted average, 109, 372
least-, 15 Weil,451
well-formed formula, 369
well-order, 58
well-pointed topos, 402
valuation, 342, 420 Weyl,187
van der Waerden, 451 Whitehead, 385
variable, 127 -and Russell, 3, 375
bound-, 371 Wigner, 4, 440, 445
complex-, 307 winding number, 325
free-, 371 Wittgenstein, 1, 444
variation, 280 words, 144
Veblen, 451 work, 174, 264
vector, 78 wrapping function, 95, III
-addition, 78
-field, 173,250,275
-product, 262 Yang-Mills, 278
-space, 37, 39, 108, 185,424
velocity, 29, 158-159, 261
Venn diagrams, 27 Zariski, 451
Vinogradoff, 49 Zermelo, 366, 443
virtual displacement, 273 Zermelo-Fraenkel axioms, 362-368
von Neumann, 437 zeta function, 310, 349,411,429
-ordinals, 366-367 Zorn lemma, 366