Introduction To

Mathematical Analysis

John E. Hutchinson
1994

Revised by
Richard J. Loy
1995/6/7

Department of Mathematics
School of Mathematical Sciences
ANU
Pure mathematics have one peculiar advantage, that they occa-
sion no disputes among wrangling disputants, as in other branches
of knowledge; and the reason is, because the definitions of the
terms are premised, and everybody that reads a proposition has
the same idea of every part of it. Hence it is easy to put an end
to all mathematical controversies by shewing, either that our ad-
versary has not stuck to his definitions, or has not laid down true
premises, or else that he has drawn false conclusions from true
principles; and in case we are able to do neither of these, we must
acknowledge the truth of what he has proved . . .
The mathematics, he [Isaac Barrow] observes, effectually exercise,
not vainly delude, nor vexatiously torment, studious minds with
obscure subtilties; but plainly demonstrate everything within their
reach, draw certain conclusions, instruct by profitable rules, and
unfold pleasant questions. These disciplines likewise enure and
corroborate the mind to a constant diligence in study; they wholly
deliver us from credulous simplicity; most strongly fortify us
against the vanity of scepticism, effectually refrain us from a
rash presumption, most easily incline us to a due assent, per-
fectly subject us to the government of right reason. While the
mind is abstracted and elevated from sensible matter, distinctly
views pure forms, conceives the beauty of ideas and investigates
the harmony of proportion; the manners themselves are sensibly
corrected and improved, the affections composed and rectified,
the fancy calmed and settled, and the understanding raised and
exited to more divine contemplations.

Encyclopædia Britannica [1771]

 i

Philosophy is written in this grand book—I mean the universe—which stands

continually open to our gaze, but it cannot be understood unless one first learns
to comprehend the language and interpret the characters in which it is written.
It is written in the language of mathematics, and its characters are triangles,
circles, and other mathematical figures, without which it is humanly impossible
to understand a single word of it; without these one is wandering about in a dark
labyrinth.
Galileo Galilei Il Saggiatore [1623]
Mathematics is the queen of the sciences.
Carl Friedrich Gauss [1856]
Thus mathematics may be defined as the subject in which we never know what
we are talking about, nor whether what we are saying is true.
Bertrand Russell Recent Work on the Principles of Mathematics,
International Monthly, vol. 4 [1901]
Mathematics takes us still further from what is human, into the region of absolute
necessity, to which not only the actual world, but every possible world, must
conform.
Bertrand Russell The Study of Mathematics [1902]
Mathematics, rightly viewed, possesses not only truth, but supreme beauty—a
beauty cold and austere, like that of a sculpture, without appeal to any part
of our weaker nature, without the gorgeous trappings of painting or music, yet
sublimely pure, and capable of perfection such as only the greatest art can show.
Bertrand Russell The Study of Mathematics [1902]
The study of mathematics is apt to commence in disappointment. . . . We are told
that by its aid the stars are weighed and the billions of molecules in a drop of
water are counted. Yet, like the ghost of Hamlet’s father, this great science eludes
the efforts of our mental weapons to grasp it.
Alfred North Whitehead An Introduction to Mathematics [1911]
The science of pure mathematics, in its modern developments, may claim to be
the most original creation of the human spirit.
Alfred North Whitehead Science and the Modern World [1925]
All the pictures which science now draws of nature and which alone seem capable
of according with observational facts are mathematical pictures . . . . From the
intrinsic evidence of his creation, the Great Architect of the Universe now begins
to appear as a pure mathematician.
Sir James Hopwood Jeans The Mysterious Universe [1930]
A mathematician, like a painter or a poet, is a maker of patterns. If his patterns
are more permanent than theirs, it is because they are made of ideas.
G.H. Hardy A Mathematician’s Apology [1940]
The language of mathematics reveals itself unreasonably effective in the natural
sciences. . . , a wonderful gift which we neither understand nor deserve. We should
be grateful for it and hope that it will remain valid in future research and that it
will extend, for better or for worse, to our pleasure even though perhaps to our
bafflement, to wide branches of learning.
Eugene Wigner [1960]
ii

To instruct someone . . . is not a matter of getting him (sic) to commit results to

mind. Rather, it is to teach him to participate in the process that makes possible
the establishment of knowledge. We teach a subject not to produce little living
libraries on that subject, but rather to get a student to think mathematically for
himself . . . to take part in the knowledge getting. Knowing is a process, not a
product.
J. Bruner Towards a theory of instruction [1966]
The same pathological structures that the mathematicians invented to break loose
from 19-th naturalism turn out to be inherent in familiar objects all around us in
nature.
Freeman Dyson Characterising Irregularity, Science 200 [1978]
Anyone who has been in the least interested in mathematics, or has even observed
other people who are interested in it, is aware that mathematical work is work
with ideas. Symbols are used as aids to thinking just as musical scores are used in
aids to music. The music comes first, the score comes later. Moreover, the score
can never be a full embodiment of the musical thoughts of the composer. Just
so, we know that a set of axioms and definitions is an attempt to describe the
main properties of a mathematical idea. But there may always remain as aspect
of the idea which we use implicitly, which we have not formalized because we have
not yet seen the counterexample that would make us aware of the possibility of
doubting it . . .
Mathematics deals with ideas. Not pencil marks or chalk marks, not physi-
cal triangles or physical sets, but ideas (which may be represented or suggested by
physical objects). What are the main properties of mathematical activity or math-
ematical knowledge, as known to all of us from daily experience? (1) Mathematical
objects are invented or created by humans. (2) They are created, not arbitrarily,
but arise from activity with already existing mathematical objects, and from the
needs of science and daily life. (3) Once created, mathematical objects have prop-
erties which are well-determined, which we may have great difficulty discovering,
but which are possessed independently of our knowledge of them.
Reuben Hersh Advances in Mathematics 31 [1979]
Don’t just read it; fight it! Ask your own questions, look for your own examples,
discover your own proofs. Is the hypothesis necessary? Is the converse true? What
happens in the classical special case? What about the degenerate cases? Where
does the proof use the hypothesis?
Paul Halmos I Want to be a Mathematician [1985]
Mathematics is like a flight of fancy, but one in which the fanciful turns out to
be real and to have been present all along. Doing mathematics has the feel of
fanciful invention, but it is really a process for sharpening our perception so that
we discover patterns that are everywhere around.. . . To share in the delight and
the intellectual experience of mathematics – to fly where before we walked – that
is the goal of mathematical education.
One feature of mathematics which requires special care . . . is its “height”, that
is, the extent to which concepts build on previous concepts. Reasoning in math-
ematics can be very clear and certain, and, once a principle is established, it can
be relied upon. This means that it is possible to build conceptual structures at
once very tall, very reliable, and extremely powerful. The structure is not like a
tree, but more like a scaffolding, with many interconnecting supports. Once the
scaffolding is solidly in place, it is not hard to build up higher, but it is impossible
to build a layer before the previous layers are in place.
William Thurston, Notices Amer. Math. Soc. [1990]
Contents

1 Introduction 1
1.1 Preliminary Remarks . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 History of Calculus . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Why “Prove” Theorems? . . . . . . . . . . . . . . . . . . . . . 2
1.4 “Summary and Problems” Book . . . . . . . . . . . . . . . . . 2
1.5 The approach to be used . . . . . . . . . . . . . . . . . . . . . 3
1.6 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Some Elementary Logic 5

2.1 Mathematical Statements . . . . . . . . . . . . . . . . . . . . 5
2.2 Quantifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Order of Quantifiers . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Connectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 Not . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.2 And . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.3 Or . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.4 Implies . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.5 Iff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Truth Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.1 Proofs of Statements Involving Connectives . . . . . . 16
2.6.2 Proofs of Statements Involving “There Exists” . . . . . 16
2.6.3 Proofs of Statements Involving “For Every” . . . . . . 17
2.6.4 Proof by Cases . . . . . . . . . . . . . . . . . . . . . . 18

3 The Real Number System 19

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Algebraic Axioms . . . . . . . . . . . . . . . . . . . . . . . . . 19

i
ii

3.2.1 Consequences of the Algebraic Axioms . . . . . . . . . 21

3.2.2 Important Sets of Real Numbers . . . . . . . . . . . . . 22
3.2.3 The Order Axioms . . . . . . . . . . . . . . . . . . . . 23
3.2.4 Ordered Fields . . . . . . . . . . . . . . . . . . . . . . 24
3.2.5 Completeness Axiom . . . . . . . . . . . . . . . . . . . 25
3.2.6 Upper and Lower Bounds . . . . . . . . . . . . . . . . 26
3.2.7 *Existence and Uniqueness of the Real Number System 29
3.2.8 The Archimedean Property . . . . . . . . . . . . . . . 30

4 Set Theory 33
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Russell’s Paradox . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Union, Intersection and Difference of Sets . . . . . . . . . . . . 35
4.4 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4.1 Functions as Sets . . . . . . . . . . . . . . . . . . . . . 38
4.4.2 Notation Associated with Functions . . . . . . . . . . . 39
4.4.3 Elementary Properties of Functions . . . . . . . . . . . 40
4.5 Equivalence of Sets . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 Denumerable Sets . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.7 Uncountable Sets . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.8 Cardinal Numbers . . . . . . . . . . . . . . . . . . . . . . . . 46
4.9 More Properties of Sets of Cardinality c and d . . . . . . . . . 50
4.10 *Further Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.10.1 The Axiom of choice . . . . . . . . . . . . . . . . . . . 52
4.10.2 Other Cardinal Numbers . . . . . . . . . . . . . . . . . 53
4.10.3 The Continuum Hypothesis . . . . . . . . . . . . . . . 54
4.10.4 Cardinal Arithmetic . . . . . . . . . . . . . . . . . . . 55
4.10.5 Ordinal numbers . . . . . . . . . . . . . . . . . . . . . 55

5 Vector Space Properties of Rn 57

5.1 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Normed Vector Spaces . . . . . . . . . . . . . . . . . . . . . . 59
5.3 Inner Product Spaces . . . . . . . . . . . . . . . . . . . . . . . 60

6 Metric Spaces 63
6.1 Basic Metric Notions in Rn . . . . . . . . . . . . . . . . . . . 63
 iii

6.2 General Metric Spaces . . . . . . . . . . . . . . . . . . . . . . 63

6.3 Interior, Exterior, Boundary and Closure . . . . . . . . . . . . 66
6.4 Open and Closed Sets . . . . . . . . . . . . . . . . . . . . . . 69
6.5 Metric Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . 73

7 Sequences and Convergence 77

7.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2 Convergence of Sequences . . . . . . . . . . . . . . . . . . . . 77
7.3 Elementary Properties . . . . . . . . . . . . . . . . . . . . . . 79
7.4 Sequences in R . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.5 Sequences and Components in Rk . . . . . . . . . . . . . . . . 83
7.6 Sequences and the Closure of a Set . . . . . . . . . . . . . . . 83
7.7 Algebraic Properties of Limits . . . . . . . . . . . . . . . . . . 84

8 Cauchy Sequences 87
8.1 Cauchy Sequences . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.2 Complete Metric Spaces . . . . . . . . . . . . . . . . . . . . . 90
8.3 Contraction Mapping Theorem . . . . . . . . . . . . . . . . . 92

9 Sequences and Compactness 97

9.1 Subsequences . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
9.2 Existence of Convergent Subsequences . . . . . . . . . . . . . 98
9.3 Compact Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9.4 Nearest Points . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

10 Limits of Functions 103

10.1 Diagrammatic Representation of Functions . . . . . . . . . . . 103
10.2 Definition of Limit . . . . . . . . . . . . . . . . . . . . . . . . 106
10.3 Equivalent Definition . . . . . . . . . . . . . . . . . . . . . . . 111
10.4 Elementary Properties of Limits . . . . . . . . . . . . . . . . . 112

11 Continuity 117
11.1 Continuity at a Point . . . . . . . . . . . . . . . . . . . . . . . 117
11.2 Basic Consequences of Continuity . . . . . . . . . . . . . . . . 119
11.3 Lipschitz and Hölder Functions . . . . . . . . . . . . . . . . . 121
11.4 Another Definition of Continuity . . . . . . . . . . . . . . . . 122
11.5 Continuous Functions on Compact Sets . . . . . . . . . . . . . 124
iv

11.6 Uniform Continuity . . . . . . . . . . . . . . . . . . . . . . . . 125

12 Uniform Convergence of Functions 129

12.1 Discussion and Definitions . . . . . . . . . . . . . . . . . . . . 129
12.2 The Uniform Metric . . . . . . . . . . . . . . . . . . . . . . . 135
12.3 Uniform Convergence and Continuity . . . . . . . . . . . . . . 138
12.4 Uniform Convergence and Integration . . . . . . . . . . . . . . 140
12.5 Uniform Convergence and Differentiation . . . . . . . . . . . . 141

13 First Order Systems of Differential Equations 143

13.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
13.1.1 Predator-Prey Problem . . . . . . . . . . . . . . . . . . 143
13.1.2 A Simple Spring System . . . . . . . . . . . . . . . . . 144
13.2 Reduction to a First Order System . . . . . . . . . . . . . . . 145
13.3 Initial Value Problems . . . . . . . . . . . . . . . . . . . . . . 147
13.4 Heuristic Justification for the
Existence of Solutions . . . . . . . . . . . . . . . . . . . . . . 149
13.5 Phase Space Diagrams . . . . . . . . . . . . . . . . . . . . . . 151
13.6 Examples of Non-Uniqueness
and Non-Existence . . . . . . . . . . . . . . . . . . . . . . . . 153
13.7 A Lipschitz Condition . . . . . . . . . . . . . . . . . . . . . . 155
13.8 Reduction to an Integral Equation . . . . . . . . . . . . . . . . 157
13.9 Local Existence . . . . . . . . . . . . . . . . . . . . . . . . . . 158
13.10Global Existence . . . . . . . . . . . . . . . . . . . . . . . . . 162
13.11Extension of Results to Systems . . . . . . . . . . . . . . . . . 163

14 Fractals 165
14.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
14.1.1 Koch Curve . . . . . . . . . . . . . . . . . . . . . . . . 166
14.1.2 Cantor Set . . . . . . . . . . . . . . . . . . . . . . . . . 167
14.1.3 Sierpinski Sponge . . . . . . . . . . . . . . . . . . . . . 168
14.2 Fractals and Similitudes . . . . . . . . . . . . . . . . . . . . . 170
14.3 Dimension of Fractals . . . . . . . . . . . . . . . . . . . . . . . 171
14.4 Fractals as Fixed Points . . . . . . . . . . . . . . . . . . . . . 174
14.5 *The Metric Space of Compact Subsets of Rn . . . . . . . . . 177
14.6 *Random Fractals . . . . . . . . . . . . . . . . . . . . . . . . . 182
 v

15 Compactness 185
15.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
15.2 Compactness and Sequential Compactness . . . . . . . . . . . 186
15.3 *Lebesgue covering theorem . . . . . . . . . . . . . . . . . . . 189
15.4 Consequences of Compactness . . . . . . . . . . . . . . . . . . 190
15.5 A Criterion for Compactness . . . . . . . . . . . . . . . . . . . 191
15.6 Equicontinuous Families of Functions . . . . . . . . . . . . . . 194
15.7 Arzela-Ascoli Theorem . . . . . . . . . . . . . . . . . . . . . . 197
15.8 Peano’s Existence Theorem . . . . . . . . . . . . . . . . . . . 201

16 Connectedness 207
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
16.2 Connected Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 207
16.3 Connectedness in R . . . . . . . . . . . . . . . . . . . . . . . . 209
16.4 Path Connected Sets . . . . . . . . . . . . . . . . . . . . . . . 210
16.5 Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

17 Differentiation of Real-Valued Functions 213

17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
17.2 Algebraic Preliminaries . . . . . . . . . . . . . . . . . . . . . . 214
17.3 Partial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . 215
17.4 Directional Derivatives . . . . . . . . . . . . . . . . . . . . . . 215
17.5 The Differential (or Derivative) . . . . . . . . . . . . . . . . . 216
17.6 The Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
17.6.1 Geometric Interpretation of the Gradient . . . . . . . . 221
17.6.2 Level Sets and the Gradient . . . . . . . . . . . . . . . 221
17.7 Some Interesting Examples . . . . . . . . . . . . . . . . . . . . 223
17.8 Differentiability Implies Continuity . . . . . . . . . . . . . . . 224
17.9 Mean Value Theorem and Consequences . . . . . . . . . . . . 224
17.10Continuously Differentiable Functions . . . . . . . . . . . . . . 227
17.11Higher-Order Partial Derivatives . . . . . . . . . . . . . . . . . 229
17.12Taylor’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 231

18 Differentiation of Vector-Valued Functions 237

18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
18.2 Paths in Rm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
vi

18.2.1 Arc length . . . . . . . . . . . . . . . . . . . . . . . . . 241

18.3 Partial and Directional Derivatives . . . . . . . . . . . . . . . 242
18.4 The Differential . . . . . . . . . . . . . . . . . . . . . . . . . . 244
18.5 The Chain Rule . . . . . . . . . . . . . . . . . . . . . . . . . . 247

19 The Inverse Function Theorem and its Applications 251

19.1 Inverse Function Theorem . . . . . . . . . . . . . . . . . . . . 251
19.2 Implicit Function Theorem . . . . . . . . . . . . . . . . . . . . 257
19.3 Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
19.4 Tangent and Normal vectors . . . . . . . . . . . . . . . . . . . 267
19.5 Maximum, Minimum, and Critical Points . . . . . . . . . . . . 268
19.6 Lagrange Multipliers . . . . . . . . . . . . . . . . . . . . . . . 269

Bibliography 273
Chapter 1

Introduction

1.1 Preliminary Remarks

These Notes provide an introduction to 20th century mathematics, and

in particular to Mathematical Analysis, which roughly speaking is the “in
depth” study of Calculus. All of the Analysis material from B21H and some
of the material from B30H is included here.
Some of the motivation for the later material in the course will come from
B23H, which is normally a co-requisite for the course. We will not however
formally require this material.
The material we will cover is basic to most of your subsequent mathemat-
ics courses (e.g. differential equations, differential geometry, measure theory,
numerical analysis, to name a few), as well as to much of theoretical physics,
engineering, probability theory and statistics. Various interesting applica-
tions are included; in particular to fractals and to differential and integral
equations.
There are also a few remarks of a general nature concerning logic and the
nature of mathematical proof, and some discussion of set theory.
There are a number of Exercises scattered throughout the text. The
Exercises are usually simple results, and you should do them all as an aid to
your understanding of the material.
Sections, Remarks, etc. marked with a * are non-examinable material,
but you should read them anyway. They often help to set the other material
in context as well as indicating further interesting directions.
There is a list of related books in the Bibliography.
The way to learn mathematics is by doing problems and by thinking very
carefully about the material as you read it. Always ask yourself why the
various assumptions in a theorem are made. It is almost always the case that
if any particular assumption is dropped, then the conclusion of the theorem
will no longer be true. Try to think of examples where the conclusion of the
theorem is no longer valid if the various assumptions are changed. Try to see

1
2

where each assumption is used in the proof of the theorem. Think of various
interesting examples of the theorem.
The dependencies of the various chapters are

14 16 15
4

1 2 3 6 7 8 9 10 11 12 13

5 17 18 19

1.2 History of Calculus

Calculus developed in the seventeenth and eighteenth centuries as a tool to
describe various physical phenomena such as occur in astronomy, mechan-
ics, and electrodynamics. But it was not until the nineteenth century that
a proper understanding was obtained of the fundamental notions of limit,
continuity, derivative, and integral. This understanding is important in both
its own right and as a foundation for further deep applications to all of the
topics mentioned in Section 1.1.

1.3 Why “Prove” Theorems?

A full understanding of a theorem, and in most cases the ability to apply
it and to modify it in other directions as needed, comes only from knowing
what really “makes it work”, i.e. from an understanding of its proof.

1.4 “Summary and Problems” Book

There is an accompanying set of notes which contains a summary of all
definitions, theorems, corollaries, etc. You should look through this at various
stages to gain an overview of the material.
These notes also contains a selection of problems, many of which have
been taken from [F]. Solutions are included.
The problems are at the level of the assignments which you will be re-
quired to do. They are not necessarily in order of difficulty. You should
attempt, or at the very least think about, the problems before you look at
the solutions. You will learn much more this way, and will in fact find the
solutions easier to follow if you have already thought enough about the prob-
lems in order to realise where the main difficulties lie. You should also think
of the solutions as examples of how to set out your own answers to other
problems.
Introduction 3

1.5 The approach to be used

Mathematics can be presented in a precise, logically ordered manner closely
following a text. This may be an efficient way to cover the content, but bears
little resemblance to how mathematics is actually done. In the words of Saun-
ders Maclane (one of the developers of category theory, a rarefied subject to
many, but one which has introduced the language of commutative diagrams
and exact sequences into mathematics) “intuition–trial–error–speculation–
conjecture–proof is a sequence for understanding of mathematics.” It is this
approach which will be taken with B21H, in particular these notes will be
used as a springboard for discussion rather than a prescription for lectures.

1.6 Acknowledgments
Thanks are due to Paulius Stepanas and other members of the 1992 and
1993 B21H and B30H classes, and Simon Stephenson, for suggestions and
corrections from the previous versions of these notes. Thanks also to Paulius
for writing up a first version of Chapter 16, to Jane James and Simon for
some assistance with the typing, and to Maciej Kocan for supplying problems
for some of the later chapters.
The diagram of the Sierpinski Sponge is from [Ma].
4
Chapter 2

Some Elementary Logic

In this Chapter we will discuss in an informal way some notions of logic

and their importance in mathematical proofs. A very good reference is [Mo,
Chapter I].

2.1 Mathematical Statements

In a mathematical proof or discussion one makes various assertions, often
called statements or sentences.1
For example:

1. (x + y)2 = x2 + 2xy + y 2 .

2. 3x2 + 2x − 1 = 0.

3. if n (≥ 3) is an integer then an + bn = cn has no positive integer

solutions.

4. the derivative of the function x2 is 2x.

Although a mathematical statement always has a very precise meaning,

certain things are often assumed from the context in which the statement
is made. For example, depending on the context in which statement (1) is
made, it is probably an abbreviation for the statement

for all real numbers x and y, (x + y)2 = x2 + 2xy + y 2 .

However, it may also be an abbreviation for the statement

for all complex numbers x and y, (x + y)2 = x2 + 2xy + y 2 .

1
Sometimes one makes a distinction between sentences and statements (which are then
certain types of sentences), but we do not do so.

5
6

The precise meaning should always be clear from context; if it is not then
more information should be provided.
Statement (2) probably refers to a particular real number x; although it
is possibly an abbreviation for the (false) statement

for all real numbers x, 3x2 + 2x − 1 = 0.

Again, the precise meaning should be clear from the context in which the
statment occurs.
Statement (3) is known as Fermat’s Last “Theorem”.2 An equivalent
statement is

if n (≥ 3) is an integer and a, b, c are positive integers, then an + bn 6= cn .

Statement (4) is expressed informally. More precisely we interpret it as

saying that the derivative of the function3 x 7→ x2 is the function x 7→ 2x.
Instead of the statement (1), let us again consider the more complete
statement

for all real numbers x and y, (x + y)2 = x2 + 2xy + y 2 .

It is important to note that this has exactly the same meaning as

for all real numbers u and v, (u + v)2 = u2 + 2uv + v 2 ,

or as

for all real numbers x and v, (x + v)2 = x2 + 2xv + v 2 .

In the previous line, the symbols u and v are sometimes called dummy vari-
ables. Note, however, that the statement

for all real numbers x , (x + x)2 = x2 + 2xx + x2

has a different meaning (while it is also true, it gives us “less” information).

In statements (3) and (4) the variables n, a, b, c, x are also dummy vari-
ables; changing them to other variables does not change the meaning of the
statement. However, in statement (2) we are probably (depending on the
context) referring to a particular number which we have denoted by x; and
if we replace x by another variable which represents another number, then
we do change the meaning of the statement.
2
This was probably the best known open problem in mathematics; primarily because
it is very simply stated and yet was incredibly difficult to solve. It was proved recently by
Andrew Wiles.
3
By x 7→ x2 we mean the function f given by f (x) = x2 for all real numbers x. We
read “x 7→ x2 ” as “x maps to x2 ”.
Elementary Logic 7

2.2 Quantifiers
The expression for all (or for every, or for each, or (sometimes) for any), is
called the universal quantifier and is often written ∀.
The following all have the same meaning (and are true)

1. for all x and for all y, (x + y)2 = x2 + 2xy + y 2

2. for any x and y, (x + y)2 = x2 + 2xy + y 2
3. for each x and each y, (x + y)2 = x2 + 2xy + y 2
³ ´
4. ∀x∀y (x + y)2 = x2 + 2xy + y 2

It is implicit in the above that when we say “for all x” or ∀x, we really
mean for all real numbers x, etc. In other words, the quantifier ∀ “ranges
over” the real numbers. More generally, we always quantify over some set of
objects, and often make the abuse of language of suppressing this set when
it is clear from context what is intended. If it is not clear from context, we
can include the set over which the quantifier ranges. Thus we could write
for all x ∈ R and for all y ∈ R, (x + y)2 = x2 + 2xy + y 2 ,
which we abbreviate to
³ ´
∀x ∈ R ∀y ∈ R (x + y)2 = x2 + 2xy + y 2 .

Sometimes statement (1) is written as

(x + y)2 = x2 + 2xy + y 2 for all x and y.
Putting the quantifiers at the end of the statement can be very risky, how-
ever. This is particularly true when there are both existential and universal
quantifiers involved. It is much safer to put quantifiers in front of the part
of the statement to which they refer. See also the next section.

The expression there exists (or there is, or there is at least one, or there
are some), is called the existential quantifier and is often written ∃.
The following statements all have the same meaning (and are true)

1. there exists an irrational number

2. there is at least one irrational number
3. some real number is irrational
4. irrational numbers exist
5. ∃x (x is irrational)

The last statement is read as “there exists x such that x is irrational”.

It is implicit here that when we write ∃x, we mean that there exists a real
number x. In other words, the quantifier ∃ “ranges over” the real numbers.
8

2.3 Order of Quantifiers

The order in which quantifiers occur is often critical. For example, consider
the statements
∀x∃y(x < y) (2.1)
and
∃y∀x(x < y). (2.2)
We read these statements as

for all x there exists y such that x < y

and
there exists y such that for all x, x < y,
respectively. Here (as usual for us) the quantifiers are intended to range over
the real numbers. Note once again that the meaning of these statments is
unchanged if we replace x and y by, say, u and v.4

Statement (2.1) is true. We can justify this as follows5 (in somewhat

more detail than usual!):
Let x be an arbitrary real number.
Then x < x + 1, and so x < y is true if y equals (for example) x + 1.
Hence the statement ∃y(x < y)6 is true.
But x was an arbitrary real number, and so the statement

for all x there exists y such that x < y

is true. That is, (2.1) is true.

On the other hand, statement (2.2) is false.

It asserts that there exists some number y such that ∀x(x < y).
But “∀x(x < y)” means y is an upper bound for the set R.
Thus (2.2) means “there exists y such that y is an upper bound for R.”
We know this last assertion is false.7

Alternatively, we could justify that (2.2) is false as follows:

Let y be an arbitrary real number.
Then y + 1 < y is false.
Hence the statement ∀x(x < y) is false.
Since y is an arbitrary real number, it follows that the statement

there exists y such that for all x, x < y,

4
In this case we could even be perverse and replace x by y and y by x respectively,
without changing the meaning!
5
For more discussion on this type of proof, see the discusion about the arbitrary object
method in Subsection 2.6.3.
6
Which, as usual, we read as “there exists y such that x < y.
7
It is false because no matter which y we choose, the number y + 1 (for example) would
be greater than y, contradicting the fact y is an upper bound for R.
Elementary Logic 9

is false.

There is much more discussion about various methods of proof in Sec-

tion 2.6.3.

We have seen that reversing the order of consecutive existential and uni-
versal quantifiers can change the meaning of a statement. However, changing
the order of consecutive existential quantifiers, or of consecutive universal
quantifiers, does not change the meaning. In particular, if P (x, y) is a state-
ment whose meaning possibly depends on x and y, then

∀x∀yP (x, y) and ∀y∀xP (x, y)

have the same meaning. For example,

∀x∀y∃z(x2 + y 3 = z),

and
∀y∀x∃z(x2 + y 3 = z),
both have the same meaning. Similarly,

∃x∃yP (x, y) and ∃y∃xP (x, y)

have the same meaning.

2.4 Connectives
The logical connectives and the logical quantifiers (already discussed) are used
to build new statements from old. The rigorous study of these concepts falls
within the study of Mathematical Logic or the Foundations of Mathematics.
We now discuss the logical connectives.

2.4.1 Not
If p is a statement, then the negation of p is denoted by

¬p (2.3)

and is read as “not p”.

If p is true then ¬p is false, and if p is false then ¬p is true.
The statement “not (not p)”, i.e. ¬¬p, means the same as “p”.

Negation of Quantifiers
10

³ ´
1. The negation of ∀xP (x), i.e. the statement ¬ ∀xP (x) , is equivalent to
³ ´
∃x ¬P (x) . Likewise, the negation of ∀x ∈ R P (x), i.e. the statement
³ ´ ³ ´
¬ ∀x ∈ R P (x) , is equivalent to ∃x ∈ R ¬P (x) ; etc.
³ ´
2. The negation of ∃xP (x), i.e. the statement ¬ ∃xP (x) , is equivalent to
³ ´
∀x ¬P (x) . Likewise, the negation of ∃x ∈ R P (x), i.e. the statement
³ ´ ³ ´
¬ ∃x ∈ R P (x) , is equivalent to ∀x ∈ R ¬P (x) .

3. If we apply the above rules twice, we see that the negation of

∀x∃yP (x, y)

is equivalent to
∃x∀y¬P (x, y).
Also, the negation of
∃x∀yP (x, y)
is equivalent to
∀x∃y¬P (x, y).
Similar rules apply if the quantifiers range over specified sets; see the
following Examples.

Examples
1 Suppose a is a fixed real number. The negation of

∃x ∈ R (x > a)

is equivalent to
∀x ∈ R ¬(x > a).
From the properties of inequalities, this is equivalent to

∀x ∈ bR (x ≤ a).

2 Theorem 3.2.10 says that

the set N of natural numbers is not bounded above.

The negation of this is the (false) statement

The set N of natural numbers is bounded above.

Putting this in the language of quantifiers, the Theorem says

³ ´
¬ ∃y∀x(x ≤ y) .

The negation is equivalent to

∃y∀x(x ≤ y).
Elementary Logic 11

3 Corollary 3.2.11 says that

if ² > 0 then there is a natural number n such that 0 < 1/n < ².

In the language of quantifiers:

∀² > 0 ∃n ∈ N (0 < 1/n < ²).

The statement 0 < 1/n was only added for emphasis, and follows from the
fact any natural number is positive and so its reciprocal is positive. Thus
the Corollary is equivalent to

∀² > 0 ∃n ∈ N (1/n < ²). (2.4)

The Corollary was proved by assuming it was false, i.e. by assuming the
negation of (2.4), and obtaining a contradiction. Let us go through essentially
the same argument again, but this time using quantifiers. This will take a
little longer, but it enables us to see the logical structure of the proof more
clearly.

Proof: The negation of (2.4) is equivalent to

∃² > 0 ∀n ∈ N ¬(1/n < ²). (2.5)

From the properties of inequalities, and the fact ² and n range over certain
sets of positive numbers, we have

¬(1/n < ²) iff 1/n ≥ ² iff n ≤ 1/².

Thus (2.5) is equivalent to

∃² > 0 ∀n ∈ N (n ≤ 1/²).

But this implies that the set of natural numbers is bounded above by 1/²,
and so is false by Theorem 3.2.10.
Thus we have obtained a contradiction from assuming the negation of (2.4),
and hence (2.4) is true.

4 The negation of

Every differentiable function is continuous (think of ∀f ∈ D C(f ))

is
³ ´
Not (every differentiable function is continuous), i.e. ¬ ∀f ∈ D C(f ) ,

and is equivalent to

Some differentiable function is not continuous, i.e. ∃f ∈ D ¬C(f ).

12

or

There exists a non-continuous differentiable function, which is

also written ∃f ∈ D ¬C(f ).

5 The negation of “all elephants are pink”, i.e. of ∀x ∈ E P (x), is “not

all elephants are pink”, i.e. ¬(∀x ∈ E P (x)), and an equivalent statement is
“there exists an elephant which is not pink”, i.e. ∃x ∈ E ¬P (x).
The negation of “there exists a pink elephant”, i.e. of ∃x ∈ E P (x), is
equivalent to “all elephants are not pink”, i.e. ∀x ∈ E ¬P (x).
This last statement is often confused in³every-day discourse
´ with the
statement “not all elephants are pink”, i.e. ¬ ∀x ∈ E P (x) , although it has
quite a different meaning, and is equivalent to “there is a non-pink elephant”,
i.e. ∃x ∈ E ¬P (x). For example, if there were 50 pink elephants in the world
and 50 white elephants, then the statement “all elephants are not pink”
would be false, but the statement “not all elephants are pink” would be true.

2.4.2 And
If p and q are statements, then the conjunction of p and q is denoted by

p∧q (2.6)

and is read as “p and q”.

If both p and q are true then p ∧ q is true, and otherwise it is false.

2.4.3 Or
If p and q are statements, then the disjunction of p and q is denoted by

p∨q (2.7)

and is read as “p or q”.

If at least one of p and q is true, then p ∨ q is true. If both p and q are
false then p ∨ q is false.
Thus the statement

1 = 1 or 1 = 2

is true. This may seem different from common usage, but consider the fol-
lowing true statement

1 = 1 or I am a pink elephant.
Elementary Logic 13

2.4.4 Implies
This often causes some confusion in mathematics. If p and q are statements,
then the statement
p⇒q (2.8)
is read as “p implies q” or “if p then q”.
Alternatively, one sometimes says “q if p”, “p only if q”, “p” is a sufficient
condition for “q”, or “q” is a necessary condition for “p”. But we will not
usually use these wordings.
If p is true and q is false then p ⇒ q is false, and in all other cases p ⇒ q
is true.
This may seem a bit strange at first, but it is essentially unavoidable.
Consider for example the true statement

∀x(x > 2 ⇒ x > 1).

Since in general we want to be able to say that a statement of the form ∀xP (x)
is true if and only if the statement P (x) is true for every (real number) x,
this leads us to saying that the statement

x>2⇒x>1

is true, for every x. Thus we require that

3 > 2 ⇒ 3 > 1,
1.5 > 2 ⇒ 1.5 > 1,
.5 > 2 ⇒ .5 > 1,

all be true statements. Thus we have examples where p is true and q is true,
where p is false and q is true, and where p is false and q is false; and in all
three cases p ⇒ q is true.
Next, consider the false statement

∀x(x > 1 ⇒ x > 2).

Since in general we want to be able to say that a statement of the form

∀xP (x) is false if and only if the statement P (x) is false for some x, this
leads us into requiring, for example, that

1.5 > 1 ⇒ 1.5 > 2

be false. This is an example where p is true and q is false, and p ⇒ q is true.

In conclusion, if the truth or falsity of the statement p ⇒ q is to depend
only on the truth or falsity of p and q, then we cannot avoid the previous
criterion in italics. See also the truth tables in Section 2.5.
Finally, in this respect, note that the statements
14

If I am not a pink elephant then 1 = 1

If I am a pink elephant then 1 = 1

and

If pigs have wings then cabbages can be kings8

are true statements.

The statement p ⇒ q is equivalent to ¬(p ∧ ¬q), i.e. not(p and not q).
This may seem confusing, and is perhaps best understood by considering the
four different cases corresponding to the truth and/or falsity of p and q.
It follows that the negation of ∀x (P (x) ⇒ Q(x)) is equivalent to the state-
ment ∃x ¬ (P (x) ⇒ Q(x)) which in turn is equivalent to ∃x (P (x) ∧ ¬Q(x)).
As a final remark, note that the statement all elephants are pink can be
written in the form ∀x (E(x) ⇒ P (x)), where E(x) means x is an elephant
and P (x) means x is pink. Previously we wrote it in the form ∀x ∈ E P (x),
where here E is the set of pink elephants, rather than the property of being
a pink elephant.

2.4.5 Iff
If p and q are statements, then the statement

p⇔q (2.9)

is read as “p if and only if q”, and is abbreviated to “p iff q”, or “p is equivalent

to q”.
Alternatively, one can say “p is a necessary and sufficient condition for q”.
If both p and q are true, or if both are false, then p ⇔ q is true. It is false
if (p is true and q is false), and it is also false if (p is false and q is true).

Remark In definitions it is conventional to use “if” where one should more

strictly use “iff”.

2.5 Truth Tables

In mathematics we require that the truth or falsity of ¬p, p ∧ q, p ∨ q, p ⇒ q
and p ⇔ q depend only on the truth or falsity of p and q.
The previous considerations then lead us to the following truth tables.
8
With apologies to Lewis Carroll.
Elementary Logic 15

p q p∧q p∨q p⇒q p⇔q

p ¬p T T T T T T
T F T F F T F F
F T F T F T T F
F F F F T T

Remarks

1. All the connectives can be defined in terms of ¬ and ∧.

2. The statement ¬q ⇒ ¬p is called the contrapositive of p ⇒ q. It has

the same meaning as p ⇒ q.

3. The statement q ⇒ p is the converse of p ⇒ q and it does not have the

same meaning as p ⇒ q.

2.6 Proofs
A mathematical proof of a theorem is a sequence of assertions (mathemati-
cal statements), of which the last assertion is the desired conclusion. Each
assertion

1. is an axiom or previously proved theorem, or

2. is an assumption stated in the theorem, or

3. follows from earlier assertions in the proof in an “obvious” way.

The word “obvious” is a problem. At first you should be very careful to write
out proofs in full detail. Otherwise you will probably write out things which
you think are obvious, but in fact are wrong. After practice, your proofs will
become shorter.
A common mistake of beginning students is to write out the very easy
points in much detail, but to quickly jump over the difficult points in the
proof.
The problem of knowing “how much detail is required” is one which will
become clearer with (much) practice.
In the next few subsections we will discuss various ways of proving math-
ematical statements.
Besides Theorem, we will also use the words Proposition, Lemma and
Corollary. The distiction between these is not a precise one. Generally,
“Theorems” are considered to be more significant or important than “Propo-
sitions”. “Lemmas” are usually not considered to be important in their own
right, but are intermediate results used to prove a later Theorem. “Corollar-
ies” are fairly easy consequences of Theorems.
16

2.6.1 Proofs of Statements Involving Connectives

To prove a theorem whose conclusion is of the form “p and q” we have to
show that both p is true and q is true.
To prove a theorem whose conclusion is of the form “p or q” we have to
show that at least one of the statements p or q is true. Three different ways
of doing this are:
• Assume p is false and use this to show q is true,
• Assume q is false and use this to show p is true,
• Assume p and q are both false and obtain a contradiction.
To prove a theorem of the type “p implies q” we may proceed in one of
the following ways:
• Assume p is true and use this to show q is true,
• Assume q is false and use this to show p is false, i.e. prove the contra-
positive of “p implies q”,
• Assume p is true and q is false and use this to obtain a contradiction.
To prove a theorem of the type “p iff q” we usually
• Show p implies q and show q implies p.

2.6.2 Proofs of Statements Involving “There Exists”

In order to prove a theorem whose conclusion is of the form “there exists x
such that P (x)”, we usually either
• show that for a certain explicit value of x, the statement P (x) is true;
or more commonly
• use an indirect argument to show that some x with property P (x) does
exist.
For example to prove
∃x such that x5 − 5x − 7 = 0
we can argue as follows: Let the function f be defined by
f (x) = x5 − 5x − 7 (for all real x). Then f (1) < 0 and
f (2) > 0; so f (x) = 0 for some x between 1 and 2 by the
Intermediate Value Theorem9 for continuous functions.
An alternative approach would be to
• assume P (x) is false for all x and deduce a contradiction.
9
See later.
Elementary Logic 17

2.6.3 Proofs of Statements Involving “For Every”

Consider the following trivial theorem:

Theorem 2.6.1 For every integer n there exists an integer m such that m >
n.

We cannot prove this theorem by individually examining each integer n.

Instead we proceed as follows:

Proof:

Let n be any integer.

What this really means is—let n be a completely arbitrary integer, so that

anything I prove about n applies equally well to any other integer. We
continue the proof as follows:

Choose the integer

m = n + 1.
Then m > n.

Thus for every integer n there is a greater integer m.

The above proof is an example of the arbitrary object method. We cannot

examine every relevant object individually. So instead, we choose an arbi-
trary object x (integer, real number, etc.) and prove the result for this x.
This is the same as proving the result for every x.
We often combine the arbitrary object method with proof by contradiction.
That is, we often prove a theorem of the type “∀xP (x)” as follows: Choose
an arbitrary x and deduce a contradiction from “¬P (x)”. Hence P (x) is
true, and since x was arbitrary, it follows that “∀xP (x)” is also true.
For example consider the theorem:
√
Theorem 2.6.2 2 is irrational.

From the definition of irrational,

√ this theorem is interpreted as saying:
“for all integers m and n, m/n 6= 2 ”. We prove this equivalent formulation
as follows:

Proof: Let m and n be arbitrary integers with n 6= 0 (as m/n is undefined

if n = 0). Suppose that √
m/n = 2.

By dividing through by any common factors greater than 1, we obtain

√
m∗ /n∗ = 2
18

where m∗ and n∗ have no common factors.

Then
(m∗ )2 = 2(n∗ )2 .
Thus (m∗ )2 is even, and so m∗ must also be even (the square of an odd integer
is odd since (2k + 1)2 = 4k 2 + 4k + 1 = 2(2k 2 + 2k) + 1).
Let m∗ = 2p. Then

4p2 = (m∗ )2 = 2(n∗ )2 ,

and so
2p2 = (n∗ )2 .
Hence (n∗ )2 is even, and so n∗ is even.
Since both m∗ and n∗ are even, √they must have the common factor 2,
which is a contradiction. So m/n 6= 2.

2.6.4 Proof by Cases

We often prove a theorem by considering various possibilities. For example,
suppose we need to prove that a certain result is true for all pairs of integers
m and n. It may be convenient to separately consider the cases m = n,
m < n and m > n.
Chapter 3

The Real Number System

3.1 Introduction
The Real Number System satisfies certain axioms, from which its other prop-
erties can be deduced. There are various slightly different, but equivalent,
formulations.

Definition 3.1.1 The Real Number System is a set1 of objects called Real
Numbers and denoted by R together with two binary operations2 called ad-
dition and multiplication and denoted by + and × respectively (we usually
write xy for x × y), a binary relation called less than and denoted by <,
and two distinct elements called zero and unity and denoted by 0 and 1
respectively.

The axioms satisfied by these fall into three groups and are detailed in the
following sections.

3.2 Algebraic Axioms

Algebraic properties are the properties of the four operations: addition +,
multiplication ×, subtraction −, and division ÷.

Properties of Addition If a, b and c are real numbers then:

A1 a + b = b + a

A2 (a + b) + c = a + (b + c)

A3 a + 0 = 0 + a = a
1
We discuss sets in the next Chapter.
2
To say + is a binary operation means that + is a function such that + : R × R → R.
We write a + b instead of +(a, b). Similar remarks apply to ·.

19
20

A4 there is exactly one real number, denoted by −a, such that

a + (−a) = (−a) + a = 0

Property A1 is called the commutative property of addition; it says it does

not matter how one commutes (interchanges) the order of addition.
Property A2 says that if we add a and b, and then add c to the result,
we get the same as adding a to the result of adding b and c. It is called
the associative property of addition; it does not matter how we associate
(combine) the brackets. The analogous result is not true for subtraction or
division.
Property A3 says there is a certain real number 0, called zero or the
additive identity, which when added to any real number a, gives a.
Property A4 says that for any real number a there is a unique (i.e. exactly
one) real number −a, called the negative or additive inverse of a, which when
added to a gives 0.

Properties of Multiplication If a, b and c are real numbers then:

A5 a × b = b × a

A6 (a × b) × c = a × (b × c)

A7 a × 1 = 1 × a = a,and 1 6= 0.

A8 if a 6= 0 there is exactly one real number, denoted by a−1 , such that

a × a−1 = a−1 × a = 1

Properties A5 and A6 are called the commutative and associative prop-

erties for multiplication.
Property A7 says there is a real number 1 6= 0, called one or the multi-
plicative identity, which when multiplied by any real number a, gives a.
Property A8 says that for any non-zero real number a there is a unique
real number a−1 , called the multiplicative inverse of a, which when multiplied
by a gives 1.
Convention We will often write ab for a × b.

The Distributive Property There is another property which involves

both addition and multiplication:

A9 If a, b and c are real numbers then

a(b + c) = ab + ac

The distributive property says that we can separately distribute multipli-

cation over the two additive terms
Real Number System 21

Algebraic Axioms It turns out that one can prove all the algebraic prop-
erties of the real numbers from properties A1–A9 of addition and multipli-
cation. We will do some of this in the next subsection.
We call A1–A9 the Algebraic Axioms for the real number system.

Equality One can write down various properties of equality. In particular,

for all real numbers a, b and c:

1. a = a

2. a = b ⇒ b = a3

3. a = b and4 b = c ⇒ a = c5

Also, if a = b, then a + c = b + c and ac = bc. More generally, one can always

replace a term in an expression by any other term to which it is equal.
It is possible to write down axioms for “=” and deduce the other prop-
erties of “=” from these axioms; but we do not do this. Instead, we take
“=” to be a logical notion which means “is the same thing as”; the previous
properties of “=” are then true from the meaning of “=”.
When we write a 6= b we will mean that a does not represent the same
number as b; i.e. a represents a different number from b.

Other Logical and Set Theoretic Notions We do not attempt to ax-

iomatise any of the logical notions involved in mathematics, nor do we at-
tempt to axiomatise any of the properties of sets which we will use (see
later). It is possible to do this; and this leads to some very deep and impor-
tant results concerning the nature and foundations of mathematics. See later
courses on the foundations mathematics (also some courses in the philosophy
department).

3.2.1 Consequences of the Algebraic Axioms

Subtraction and Division We first define subtraction in terms of addition
and the additive inverse, by

a − b = a + (−b).

Similarly, if b 6= 0 define
µ ¶
a
a ÷ b = a/b = = ab−1 .
b
3
By ⇒ we mean “implies”. Let P and Q be two statements, then “P ⇒ Q” means “P
implies Q”; or equivalently “if P then Q”.
4
We sometimes write “∧” for “and”.
5
Whenever we write “P ∧ Q ⇒ R”, or “P and Q ⇒ R”, the convention is that we mean
“(P ∧ Q) ⇒ R”, not “P ∧ (Q ⇒ R)”.
22

Some consequences of axioms A1–A9 are as follows. The proofs are given
in the AH1 notes.

Theorem 3.2.1 (Cancellation Law for Addition) If a, b and c are real

numbers and a + c = b + c, then a = b.

Theorem 3.2.2 (Cancellation Law for Multiplication) If a, b and c 6=

0 are real numbers and ac = bc then a = b.

Theorem 3.2.3 If a, b, c, d are real numbers and c 6= 0, d 6= 0 then

1. a0 = 0
2. −(−a) = a
3. (c−1 )−1 = c
4. (−1)a = −a
5. a(−b) = −(ab) = (−a)b
6. (−a) + (−b) = −(a + b)
7. (−a)(−b) = ab
8. (a/c)(b/d) = (ab)/(cd)
9. (a/c) + (b/d) = (ad + bc)/cd

Remark Henceforth (unless we say otherwise) we will assume all the usual
properties of addition, multiplication, subtraction and division. In particular,
we can solve simultaneous linear equations. We will also assume standard
definitions including x2 = x × x, x3 = x × x × x, x−2 = (x−1 ) , etc.
2

3.2.2 Important Sets of Real Numbers

We define
2 = 1 + 1, 3 = 2 + 1 , . . . , 9 = 8 + 1 ,
10 = 9 + 1 , . . . , 19 = 18 + 1 , . . . , 100 = 99 + 1 , . . . .
The set N of natural numbers is defined by
N = {1, 2, 3, . . .}.
The set Z of integers is defined by
Z = {m : −m ∈ N, or m = 0, or m ∈ N}.
The set Q of rational numbers is defined by
Q = {m/n : m ∈ Z, n ∈ N}.
The set of all real numbers is denoted by R.
A real number is irrational if it is not rational.
Real Number System 23

3.2.3 The Order Axioms

As remarked in Section 3.1, the real numbers have a natural ordering. Instead
of writing down axioms directly for this ordering, it is more convenient to
write out some axioms for the set P of positive real numbers. We then define
< in terms of P .

Order Axioms There is a subset6 P of the set of real numbers, called the
set of positive numbers, such that:

A10 For any real number a, exactly one of the following holds:

a = 0 or a ∈ P or −a∈P

A11 If a ∈ P and b ∈ P then a + b ∈ P and ab ∈ P

A number a is called negative when −a is positive.

The “Less Than” Relation We now define a < b to mean b − a ∈ P .

We also define:
a ≤ b to mean b − a ∈ P or a = b;
a > b to mean a − b ∈ P ;
a ≥ b to mean a − b ∈ P or a = b.
It follows that a < b if and only if7 b > a. Similarly, a ≤ b iff8 b ≥ a.

Theorem 3.2.4 If a, b and c are real numbers then

1. a < b and b < c implies a < c

2. exactly one of a < b, a = b and a > b is true

3. a < b implies a + c < b + c

4. a < b and c > 0 implies ac < bc

5. a < b and c < 0 implies ac > bc

6. 0 < 1 and −1 < 0

7. a > 0 implies 1/a > 0

8. 0 < a < b implies 0 < 1/b < 1/a

6
We will use some of the basic notation of set theory. Refer forward to Chapter 4 if
necessary.
7
If A and B are statements, then “A if and only if B” means that A implies B and B
implies A. Another way of expressing this is to say that A and B are either both true or
both false.
8
Iff is an abbreviation for if and only if.
24

Similar properties of ≤ can also be proved.

Remark Henceforth, in addition to assuming all the usual algebraic prop-

erties of the real number system, we will also assume all the standard results
concerning inequalities.

Absolute Value The absolute value of a real number a is defined by

(
a if a ≥ 0
|a| =
−a if a < 0

The following important properties can be deduced from the axioms; but we
will not pause to do so.

Theorem 3.2.5 If a and b are real numbers then:

1. |ab| = |a| |b|

2. |a + b| ≤ |a| + |b|
¯ ¯
3. ¯¯|a| − |b|¯¯ ≤ |a − b|

We will use standard notation for intervals:

[a, b] = {x : a ≤ x ≤ b}, (a, b) = {x : a < x < b},

(a, ∞) = {x : x > a}, [a, ∞) = {x : x ≥ a}

with similar definitions for [a, b), (a, b], (−∞, a], (−∞, a). Note that ∞ is not
a real number and there is no interval of the form (a, ∞].
We only use the symbol ∞ as part of an expression which, when written
out in full, does not refer to ∞.

3.2.4 Ordered Fields

Any set S, together with two operations ⊕ and ⊗ and two members 0⊕ and
0⊗ of S, and a subset P of S, which satisfies the corresponding versions of
A1–A11, is called an ordered field.
Both Q and R are ordered fields, but finite fields are not.
Another example is the field of real algebraic numbers; where a real num-
ber is said to be algebraic if it is a solution of a polynomial equation of the
form
a0 + a1 x + a2 x2 + · · · + an xn = 0
for some integer n > 0 and integers a0 , a1 , . . . , an . Note that
√ any rational
number x = m/n is algebraic, since m − nx = 0, and that 2 is algebraic
since it satisfies the equation 2 − x2 = 0. (As a nice exercise in algebra, show
that the set of real algebraic numbers is indeed a field.)
Real Number System 25

3.2.5 Completeness Axiom

We now come to the property that singles out the real numbers from any
other ordered field. There are a number of versions of this axiom. We take
the following, which is perhaps a little more intuitive. We will later deduce
the version in Adams.

Dedekind Completeness Axiom

A12 Suppose A and B are two (non-empty) sets of real numbers with the
properties:

1. if a ∈ A and b ∈ B then a < b

2. every real number is in either A or B 9
(in symbols; A ∪ B = R).

Then there is a unique real number c such that:

a. if a < c then a ∈ A, and

b. if b > c then b ∈ B

a c b

A B
Note that every number < c belongs to A and every number > c belongs
to B. Moreover, either c ∈ A or c ∈ B by 2. Hence if c ∈ A then A = (−∞, c]
and B = (c, ∞); while if c ∈ B then A = (−∞, c) and B = [c, ∞).
The pair of sets {A, B} is called a Dedekind Cut.
The intuitive idea of A12 is that the Completeness Axiom says there are
no “holes” in the real numbers.

Remark The analogous

√ axiom is not true in the ordered field Q. This is
essentially because 2 is not rational, as we saw in Theorem 2.6.2.
More precisely, let
√ √
A = {x ∈ Q : x < 2}, B = {x ∈ Q : x ≥ 2}.
³
If you do not like to define A and B, which are sets of rational numbers, by
√
using the irrational number 2, you could equivalently define
´
A = {x ∈ Q : x ≤ 0 or (x > 0 and x2 < 2)}, B = {x ∈ Q : x > 0 and x2 ≥ 2}
Suppose c satisfies a and b of A12. Then it follows from algebraic and
order properties10 that c2 ≥ 2 and c2 ≤ 2, hence c2 = 2. But we saw in
Theorem 2.6.2 that c cannot then be rational.
9
It follows that if x is any real number, then x is in exactly one of A and B, since
otherwise we would have x < x from 1.
10
Related arguments are given in a little more detail in the proof of the next Theorem.
26

√
We next use Axiom A12 to prove the existence of 2, i.e. the existence
of a number c such that c2 = 2.

Theorem 3.2.6 There is a (positive)11 real number c such that c2 = 2.

Proof: Let

A = {x ∈ R : x ≤ 0 or (x > 0 and x2 < 2)}, B = {x ∈ R : x > 0 and x2 ≥ 2}

It follows (from the algebraic and order properties of the real numbers; i.e.
A1–A11) that every real number x is in exactly one of A or B, and hence
that the two hypotheses of A12 are satisfied.
By A12 there is a unique real number c such that

1. every number x less than c is either ≤ 0, or is > 0 and satisfies x2 < 2

2. every number x greater than c is > 0 and satisfies x2 ≥ 2.

0 c b

A B
From the Note following A12, either c ∈ A or c ∈ B.
If c ∈ A then c < 0 or (c > 0 and c2 < 2). But then by taking ² > 0
sufficiently small, we would also have c + ² ∈ A (from the definition
of A), which contradicts conclusion b in A12.
Hence c ∈ B, i.e. c > 0 and c2 ≥ 2.
If c2 > 2, then by choosing ² > 0 sufficiently small we would also have
c − ² ∈ B (from the definition of B), which contradicts a in A12.
Hence c2 = 2.

3.2.6 Upper and Lower Bounds

Definition 3.2.7 If S is a set of real numbers, then

1. a is an upper bound for S if x ≤ a for all x ∈ S;

2. b is the least upper bound (or l.u.b. or supremum or sup) for S if b is

an upper bound, and moreover b ≤ a whenever a is any upper bound
for S.

We write
b = l.u.b. S = sup S
One similarly defines lower bound and greatest lower bound (or g.l.b. or infi-
mum or inf ) by replacing “≤” by “≥”.
11
And hence a negative real number c such that c2 = 2; just replace c by −c.
Real Number System 27

A set S is bounded above if it has an upper bound12 and is bounded below

if it has a lower bound.

Note that if the l.u.b. or g.l.b. exists it is unique, since if b1 and b2 are
both l.u.b.’s then b1 ≤ b2 and b2 ≤ b1 , and so b1 = b2 .

Examples

1. If S = [1, ∞) then any a ≤ 1 is a lower bound, and 1 = g.l.b.S. There

is no upper bound for S. The set S is bounded below but not above.

2. If S = [0, 1) then 0 = g.l.b.S ∈ S and 1 = l.u.b.S 6∈ S. The set S is

bounded above and below.

3. If S = {1, 1/2, 1/3, . . . , 1/n, . . .} then 0 = g.l.b.S 6∈ S and 1 = l.u.b.S ∈

S. The set S is bounded above and below.

There is an equivalent form of the Completeness Axiom:

Least Upper Bound Completeness Axiom

A120 Suppose S is a nonempty set of real numbers which is bounded above.
Then S has a l.u.b. in R.

A similar result follows for g.l.b.’s:

Corollary 3.2.8 Suppose S is a nonempty set of real numbers which is

bounded below. Then S has a g.l.b. in R.

Proof: Let
T = {−x : x ∈ S}.
Then it follows that a is a lower bound for S iff −a is an upper bound for T ;
and b is a g.l.b. for S iff −b is a l.u.b. for T .
b 0 -b

( )

T
S

Since S is bounded below, it follows that T is bounded above. Moreover,

T then has a l.u.b. c (say) by A12’, and so −c is a g.l.b. for S.
12
It follows that S has infinitely many upper bounds.
28

Equivalence of A12 and A120

1) Suppose A12 is true. We will deduce A120 .
For this, suppose that S is a nonempty set of real numbers which is
bounded above.
Let
B = {x : x is an upper bound for S}, A = R \ B.13
Note that B 6= ∅; and if x ∈ S then x − 1 is not an upper bound for S so
A 6= ∅.
The first hypothesis in A12 is easy to check: suppose a ∈ A and b ∈ B. If
a ≥ b then a would also be an upper bound for S, which contradicts the
definition of A, hence a < b.
The second hypothesis in A12 is immediate from the definition of A as con-
sisting of every real number not in B.
Let c be the real number given by A12.

We claim that c = l.u.b. S.

If c ∈ A then c is not an upper bound for S and so there exists x ∈ S with
c < x. But then a = (c + x)/2 is not an upper bound for S, i.e. a ∈ A,
contradicting the fact from the conclusion of Axiom A12 that a ≤ c for all
a ∈ A. Hence c ∈ B.
But if c ∈ B then c ≤ b for all b ∈ B; i.e. c is ≤ any upper bound for S. This
proves the claim; and hence proves A120 .

2) Suppose A120 is true. We will deduce A12.

For this, suppose {A, B} is a Dedekind cut.
Then A is bounded above (by any element of B). Let c = l.u.b. A, using
A12’.
We claim that
a < c ⇒ a ∈ A, b > c ⇒ b ∈ B.

Suppose a < c. Now every member of B is an upper bound for A, from

the first property of a Dedekind cut; hence a 6∈ B, as otherwise a would be
an upper bound for A which is less than the least upper bound c. Hence
a ∈ A.
Next suppose b > c. Since c is an upper bound for A (in fact the least
upper bound), it follows we cannot have b ∈ A, and thus b ∈ B.
This proves the claim, and hence A12 is true.

The following is a useful way to characterise the l.u.b. of a set. It says

that b = l.u.b. S iff b is an upper bound for S and there exist members of S
arbitrarily close to b.
13
R \ B is the set of real numbers x not in B
Real Number System 29

Proposition 3.2.9 Suppose S is a nonempty set of real numbers. Then

b = l.u.b. S iff

1. x ≤ b for all x ∈ S, and

2. for each ² > 0 there exist x ∈ S such that x > b − ².

Proof: Suppose S is a nonempty set of real numbers.

First assume b = l.u.b. S. Then 1 is certainly true.
Suppose 2 is not true. Then for some ² > 0 it follows that x ≤ b − ² for
every x ∈ S, i.e. b − ² is an upper bound for S. This contradicts the fact
b = l.u.b. S. Hence 2 is true.
Next assume that 1 and 2 are true. Then b is an upper bound for S
from 1.
Moreover, if b0 < b then from 2 it follows that b0 is not an upper bound of S.
Hence b0 is the least upper bound of S.

We will usually use the version Axiom A120 rather than Axiom A12; and
we will usually refer to either as the Completeness Axiom. Whenever we use
the Completeness axiom in our future developments, we will explictly refer
to it. The Completeness Axiom is essential in proving such results as the
Intermediate Value Theorem14 .
Exercise: Give an example to show that the Intermediate Value Theorem
does not hold in the “world of rational numbers”.

3.2.7 *Existence and Uniqueness of the Real Number

System
We began by assuming that R, together with the operations + and × and
the set of positive numbers P , satisfies Axioms 1–12. But if we begin with
the axioms for set theory, it is possible to prove the existence of a set of
objects satisfying the Axioms 1–12.
This is done by first constructing the natural numbers, then the integers,
then the rationals, and finally the reals. The natural numbers are constructed
as certain types of sets, the negative integers are constructed from the natural
numbers, the rationals are constructed as sets of ordered pairs as in Chapter
II-2 of Birkhoff and MacLane. The reals are then constructed by the method
of Dedekind Cuts as in Chapter IV-5 of Birkhoff and MacLane or Cauchy
Sequences as in Chapter 28 of Spivak.

The structure consisting of the set R, together with the operations + and
× and the set of positive numbers P , is uniquely characterised by Axioms 1–
12, in the sense that any two structures satisfying the axioms are essentially
14
If a continuous real valued function f : [a, b] → R satisfies f (a) < 0 < f (b), then
f (c) = 0 for some c ∈ (a, b).
30

the same. More precisely, the two systems are isomorphic, see Chapter IV-5
of Birkhoff and MacLane or Chapter 29 of Spivak.

3.2.8 The Archimedean Property

The fact that the set N of natural numbers is not bounded above, does
not follow from Axioms 1–11. However, it does follow if we also use the
Completeness Axiom.

Theorem 3.2.10 The set N of natural numbers is not bounded above.

Proof: Recall that N is defined to be the set

N = {1, 1 + 1, 1 + 1 + 1, . . .}.

Assume that N is bounded above.15 Then from the Completeness Axiom

(version A120 ), there is a least upper bound b for N. That is,

n ∈ N implies n ≤ b. (3.1)

It follows that
m ∈ N implies m + 1 ≤ b, (3.2)
since if m ∈ N then m + 1 ∈ N, and so we can now apply (3.1) with n there
replaced by m + 1.
But from (3.2) (and the properties of subtraction and of <) it follows that

m ∈ N implies m ≤ b − 1.

This is a contradiction, since b was taken to be the least upper bound of N.

Thus the assumption “N is bounded above” leads to a contradiction, and so
it is false. Thus N is not bounded above.

The following Corollary is often implicitly used.

Corollary 3.2.11 If ² > 0 then there is a natural number n such that 0 <
1/n < ².16

Proof: Assume there is no natural number n such that 0 < 1/n < ². Then
for every n ∈ N it follows that 1/n ≥ ² and hence n ≤ 1/². Hence 1/² is an
upper bound for N, contradicting the previous Theorem.
Hence there is a natural number n such that 0 < 1/n < ².
15
Logical Point: Our intention is to obtain a contradiction from this assumption, and
hence to deduce that N is not bounded above.
16
We usually use ² and δ to denote numbers that we think of as being small and positive.
Note, however, that the result is true for any real number ²; but it is more “interesting”
if ² is small.
Real Number System 31

We can now prove that between any two real numbers there is a rational
number.

Theorem 3.2.12 For any two reals x and y, if x < y then there exists a
rational number r such that x < r < y.

Proof: (a) First suppose y − x > 1. Then there is an integer k such that
x < k < y.
To see this, let l be the least upper bound of the set S of all integers j
such that j ≤ x. It follows that l itself is a member of S,and so in particular
is an integer.17 ) Hence l + 1 > x, since otherwise l + 1 ≤ x, i.e. l + 1 ∈ S,
contradicting the fact that l = lub S.
Moreover, since l ≤ x and y − x > 1, it follows from the properties of < that
³ l x l+1 y
l + 1 < y. Diagram: )
Thus if k = l + 1 then x < k < y.
(b) Now just assume x < y.
By the previous Corollary choose a natural number n such that 1/n < y − x.
Hence ny −nx > 1 and so by (a) there is an integer k such that nx < k < ny.
Hence x < k/n < y, as required.

A similar result holds for the irrationals.

Theorem 3.2.13 For any two reals x and y, if x < y then there exists an
irrational number r such that x < r < y.

Proof: First
√ suppose a and b are rational
√ and a < b.
Note that 2/2√is irrational (why? ) and 2/2 < 1.√ Hence
a < a + (b − a) 2/2 < b and moreover a + (b − a) 2/2 is irrational18 .
To prove the result for general x < y, use the previous theorem twice to
first choose a rational number a and then another rational number b, such
that x < a < b < y.
By the first paragraph there is a rational number r such that x < a < r <
b < y.
17
The least upper bound b of any set S of integers which is bounded above, must itself
be a member of S. This is fairly clear, using the fact that members of S must be at least
the fixed distance 1 apart.
More precisely, consider the interval [b − 1/2, b]. Since the distance between any two
integers is ≥ 1, there can be at most one member of S in this interval. If there is no
member of S in [b − 1/2, b] then b − 1/2 would also be an upper bound for S, contradicting
the fact b is the least upper bound. Hence there is exactly one member s of S in [b−1/2, b];
it follows s = b as otherwise s would be an upper bound for S which is < b; contradiction.
Note that this argument works for any set S whose members are all at least a fixed
positive distance d > 0 √ apart. Why?
18
Let√r = a + (b − a) 2/2. √
Hence 2 = 2(r − a)/(b − a). So if r were rational then 2 would also be rational, which
we know is not the case.
32

Corollary 3.2.14 For any real number x, and any positive number ² >,
there a rational (resp. irrational) number r (resp.s) such that 0 < |r − x| < ²
(resp. 0 < |s − x| < ²).
Chapter 4

Set Theory

4.1 Introduction
The notion of a set is fundamental to mathematics.
A set is, informally speaking, a collection of objects. We cannot use
this as a definition however, as we then need to define what we mean by a
collection.
The notion of a set is a basic or primitive one, as is membership ∈, which
are not usually defined in terms of other notions. Synonyms for set are
collection, class 1 and family.
It is possible to write down axioms for the theory of sets. To do this
properly, one also needs to formalise the logic involved. We will not follow
such an axiomatic approach to set theory, but will instead proceed in a more
informal manner.
Sets are important as it is possible to formulate all of mathematics in set
theory. This is not done in practice, however, unless one is interested in the
Foundations of Mathematics 2 .

4.2 Russell’s Paradox

It would seem reasonable to assume that for any “property” or “condition”
P , there is a set S consisting of all objects with the given property.
More precisely, if P (x) means that x has property P , then there should
be a set S defined by
S = {x : P (x)} . (4.1)

1
*Although we do not do so, in some studies of set theory, a distinction is made between
set and class.
2
There is a third/fourth year course Logic, Set Theory and the Foundations of
Mathematics.

33
34

This is read as: “S is the set of all x such that P (x) (is true)”3 .
For example, if P (x) is an abbreviation for

x is an integer > 5

or
x is a pink elephant,
then there is a corresponding set (although in the second case it is the so-
called empty set, which has no members) of objects x having property P (x).
However, Bertrand Russell came up with the following property of x:

x is not a member of itself4 ,

or in symbols
x 6∈ x.

Suppose
S = {x : x 6∈ x} .
If there is indeed such a set S, then either S ∈ S or S 6∈ S. But

• if the first case is true, i.e. S is a member of S, then S must satisfy the
defining property of the set S, and so S 6∈ S—contradiction;
• if the second case is true, i.e. S is not a member of S, then S does not
satisfy the defining property of the set S, and so S ∈ S—contradiction.

Thus there is a contradiction in either case.

While this may seem an artificial example, there does arise the important
problem of deciding which properties we should allow in order to describe
sets. This problem is considered in the study of axiomatic set theory. We
will not (hopefully!) be using properties that lead to such paradoxes, our
construction in the above situation will rather be of the form “given a set A,
consider the elements of A satisfying some defining property”.
None-the-less, when the German philosopher and mathematician Gottlob
Frege heard from Bertrand Russell (around the turn of the century) of the
above property, just as the second edition of his two volume work Grundge-
setze der Arithmetik (The Fundamental Laws of Arithmetic) was in press, he
felt obliged to add the following acknowledgment:

A scientist can hardly encounter anything more undesirable than

to have the foundation collapse just as the work is finished. I was
put in this position by a letter from Mr. Bertrand Russell when
the work was almost through the press.
3
Note that this is exactly the same as saying “S is the set of all z such that P (z) (is
true)”.
4
Any x we think of would normally have this property. Can you think of some x which
is a member of itself? What about the set of weird ideas?
Set Theory 35

4.3 Union, Intersection and Difference of Sets

The members of a set are sometimes called elements of the set. If x is a
member of the set S, we write
x ∈ S.
If x is not a member of S we write
x 6∈ S.

A set with a finite number of elements can often be described by explicitly

giving its members. Thus
n o
S = 1, 3, {1, 5} (4.2)

is the set with members 1,3 and {1, 5}. Note that 5 is not a member5 . If we
write the members of a set in a different order, we still have the same set.
If S is the set of all x such that . . . x . . . is true, then we write
S = {x : . . . x . . .} , (4.3)
and read this as “S is the set of x such that . . . x . . .”. For example, if
S = {x : 1 < x ≤ 2}, then S is the interval of real numbers that we also
denote by (1, 2].
Members of a set may themselves be sets, as in (4.2).
If A and B are sets, their union A ∪ B is the set of all objects which
belong to A or belong to B (remember that by the meaning of or this also
includes those objects belonging to both A and B). Thus
A ∪ B = {x : x ∈ A or x ∈ B} .
The intersection A ∩ B of A and B is defined by
A ∩ B = {x : x ∈ A and x ∈ B} .
The difference A \ B of A and B is defined by
A \ B = {x : x ∈ A and x 6∈ B} .
It is sometimes convenient to represent this schematically by means of a Venn
Diagram.

5
However, it is a member of a member; membership is generally not transitive.
36

We can take the union of more than two sets. If F is a family of sets, the
union of all sets in F is defined by
[
F = {x : x ∈ A for at least one A ∈ F} . (4.4)

The intersection of all sets in F is defined by

\
F = {x : x ∈ A for every A ∈ F} . (4.5)

If F is finite, say F = {A1 , . . . , An }, then the union and intersection of

members of F are written
[
n
Ai or A1 ∪ · · · ∪ An (4.6)
i=1

and
\
n
Ai or A1 ∩ · · · ∩ An (4.7)
i=1

respectively. If F is the family of sets {Ai : i = 1, 2, . . .}, then we write

∞
[ ∞
\
Ai and Ai (4.8)
i=1 i=1

respectively. More generally, we may have a family of sets indexed by a set

other than the integers— e.g. {Aλ : λ ∈ J}—in which case we write
[ \
Aλ and Aλ (4.9)
λ∈J λ∈J

for the union and intersection.

Examples
S∞
1. n=1 [0, 1 − 1/n] = [0, 1)
T∞
2. n=1 [0, 1/n] = {0}
T∞
3. n=1 (0, 1/n) =∅

We say two sets A and B are equal iff they have the same members, and
in this case we write
A = B. (4.10)

It is convenient to have a set with no members; it is denoted by

∅. (4.11)

There is only one empty set, since any two empty sets have the same mem-
bers, and so are equal!
Set Theory 37

If every member of the set A is also a member of B, we say A is a subset

of B and write
A ⊂ B. (4.12)
We include the possibility that A = B, and so in some texts this would
be written as A ⊆ B. Notice the distinction between “∈” and “⊂”. Thus
in (4.2) we have 1 ∈ S, 3 ∈ S, {1, 5} ∈ S while {1, 3} ⊂ S, {1} ⊂ S, {3} ⊂ S,
{{1, 5}} ⊂ S, S ⊂ S, ∅ ⊂ S.
We usually prove A = B by proving that A ⊂ B and that B ⊂ A, c.f.
the proof of (4.36) in Section 4.4.3.
If A ⊂ B and A 6= B, we say A is a proper subset of B and write A $ B.
The sets A and B are disjoint if A ∩ B = ∅. The sets belonging to a
family of sets F are pairwise disjoint if any two distinctly indexed sets in F
are disjoint.
The set of all subsets of the set A is called the Power Set of A and is
denoted by
P(A).
In particular, ∅ ∈ P(A) and A ∈ P(A).
The following simple properties of sets are made plausible by considering
a Venn diagram. We will prove some, and you should prove others as an
exercise. Note that the proofs essentially just rely on the meaning of the
logical words and, or, implies etc.

Proposition 4.3.1 Let A, B, C and Bλ (for λ ∈ J) be sets. Then

A∪B =B∪A A∩B =B∩A
A ∪ (B ∪ C) = (A ∪ B) ∪ C A ∩ (B ∩ C) = (A ∩ B) ∩ C
A⊂A∪B A∩B ⊂A
A ⊂ B iff A ∪ B = B A ⊂ B iff A ∩ B = A
A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C) A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C)
[ [ \ \
A∩ Bλ = (A ∩ Bλ ) A∪ Bλ = (A ∪ Bλ )
λ∈J λ∈J λ∈J λ∈J

Proof: We prove A ⊂ B iff A ∩ B = A as an example.

First assume A ⊂ B. We want to prove A ∩ B = A (we will show
A ∩ B ⊂ A and A ⊂ A ∩ B). If x ∈ A ∩ B then certainly x ∈ A, and so
A ∩ B ⊂ A. If x ∈ A then x ∈ B by our assumption, and so x ∈ A ∩ B, and
hence A ⊂ A ∩ B. Thus A ∩ B = A.
Next assume A ∩ B = A. We want to prove A ⊂ B. If x ∈ A, then
x ∈ A ∩ B (as A ∩ B = A) and so in particular x ∈ B. Hence A ⊂ B.

If X is some set which contains all the objects being considered in a

certain context, we sometimes call X a universal set. If A ⊂ X then X \ A
is called the complement of A, and is denoted by
Ac . (4.13)
38

Thus if X is the (set of) reals and A is the (set of) rationals, then the
complement of A is the set of irrationals.
The complement of the union (intersection) of a family of sets is the in-
tersection (union) of the complements; these facts are known as de Morgan’s
laws. More precisely,

Proposition 4.3.2
(A ∪ B)c = Ac ∩ B c and (A ∩ B)c = Ac ∪ B c . (4.14)
More generally,
Ã∞ !c ∞
Ã∞ !c ∞
[ \ \ [
Ai = Aci and Ai = Aci , (4.15)
i=1 i=1 i=1 i=1

and  c  c
[ \ \ [
 Aλ  = Acλ and  Aλ  = Acλ . (4.16)
λ∈J λ∈J λ∈J λ∈J

4.4 Functions
We think of a function f : A → B as a way of assigning to each element a ∈ A
an element f (a) ∈ B. We will make this idea precise by defining functions
as particular kinds of sets.

4.4.1 Functions as Sets

We first need the idea of an ordered pair . If x and y are two objects, the
ordered pair whose first member is x and whose second member is y is
denoted
(x, y). (4.17)
The basic property of ordered pairs is that
(x, y) = (a, b) iff x = a and y = b. (4.18)
Thus (x, y) = (y, x) iff x = y; whereas {x, y} = {y, x} is always true. Any
way of defining the notion of an ordered pair is satisfactory, provided it
satisfies the basic property.

One way to define the notion of an ordered pair in terms of sets

is by setting
(x, y) = {{x}, {x, y}} .
This is natural: {x, y} is the associated set of elements and {x}
is the set containing the first element of the ordered pair. As a
non-trivial problem, you might like to try and prove the basic
property of ordered pairs from this definition. HINT: consider
separately the cases x = y and x 6= y. The proof is in [La, pp.
42-43].
Set Theory 39

If A and B are sets, their Cartesian product is the set of all ordered pairs
(x, y) with x ∈ A and y ∈ B. Thus

A × B = {(x, y) : x ∈ A and y ∈ B} . (4.19)

We can also define n-tuples (a1 , a2 , . . . , an ) such that

(a1 , a2 , . . . , an ) = (b1 , b2 , . . . , bn ) iff a1 = b1 , a2 = b2 , . . . , an = bn . (4.20)

The Cartesian Product of n sets is defined by

A1 × A2 × · · · × An = {(a1 , a2 , . . . , an ) : a1 ∈ A1 , a2 ∈ A2 , . . . , an ∈ An } .
(4.21)
In particular, we write
n
z }| {
R = R × ··· × R.
n
(4.22)

If f is a set of ordered pairs from A × B with the property that for every
x ∈ A there is exactly one y ∈ B such that (x, y) ∈ f , then we say f is a
function (or map or transformation or operator ) from A to B. We write

f : A → B, (4.23)

which we read as: f sends (maps) A into B. If (x, y) ∈ f then y is uniquely

determined by x and for this particular x and y we write

y = f (x). (4.24)

We say y is the value of f at x.

Thus if n o
f = (x, x2 ) : x ∈ R (4.25)
then f : R → R and f is the function usually defined (somewhat loosely) by

f (x) = x2 , (4.26)

where it is understood from context that x is a real number.

Note that it is exactly the same to define the function f by f (x) = x2 for
all x ∈ R as it is to define f by f (y) = y 2 for all y ∈ R.

4.4.2 Notation Associated with Functions

Suppose f : A → B. A is called the domain of f and B is called the co-domain
of f .
The range of f is the set defined by

f [A] = {y : y = f (x) for some x ∈ A} (4.27)

= {f (x) : x ∈ A} . (4.28)
40

Note that f [A] ⊂ B but may not equal B. For example, in (4.26) the range
of f is the set [0, ∞) = {x ∈ R : 0 ≤ x}.
We say f is one-one or injective or univalent if for every y ∈ B there is
at most one x ∈ A such that y = f (x). Thus the function f1 : R → R given
by f1 (x) = x2 for all x ∈ R is not one-one, while the function f2 : R → R
given by f2 (x) = ex for all x ∈ R is one-one.
We say f is onto or surjective if every y ∈ B is of the form f (x) for some
x ∈ A. Thus neither f1 nor f2 is onto. However, f1 maps R onto [0, ∞).
If f is both one-one and onto, then there is an inverse function f −1 :
B → A defined by f (y) = x iff f (x) = y. For example, if f (x) = ex for all
x ∈ R, then f : R → [0, ∞) is one-one and onto, and so has an inverse which
is usually denoted by ln. Note, incidentally, that f : R → R is not onto, and
so strictly speaking does not have an inverse.
If S ⊂ A, then the image of S under f is defined by

f [S] = {f (x) : x ∈ S} . (4.29)

Thus f [S] is a subset of B, and in particular the image of A is the range of f .

If S ⊂ A, the restriction f |S of f to S is the function whose domain is S
and which takes the same values on S as does f . Thus

f |S = {(x, f (x)) : x ∈ S} (4.30)

If T ⊂ B, then the inverse image of T under f is

f −1 [T ] = {x : f (x) ∈ T } . (4.31)

It is a subset of A. Note that f −1 [T ] is defined for any function f : A → B.

It is not necessary that f be one-one and onto, i.e. it is not necessary that
the function f −1 exist.
If f : A → B and g : B → C then the composition function g ◦ f : A → C
is defined by
(g ◦ f )(x) = g(f (x)) ∀x ∈ A. (4.32)
For example, if f (x) = x2 for all x ∈ R and g(x) = sin x for all x ∈ R, then
(g ◦ f )(x) = sin(x2 ) and (f ◦ g)(x) = (sin x)2 .

4.4.3 Elementary Properties of Functions

We have the following elementary properties:

Proposition 4.4.1
h[ i [
f [C ∪ D] = f [C] ∪ f [D] f Aλ = f [Aλ ] (4.33)
λ∈J λ∈J
h\ i \
f [C ∩ D] ⊂ f [C] ∩ f [D] f Cλ ⊂ f [Cλ ] (4.34)
λ∈J λ∈J
Set Theory 41

h[ i [
f −1 [U ∪ V ] = f −1 [U ] ∪ f −1 [V ] f −1 Uλ = f −1 [Uλ ] (4.35)
λ∈J λ∈J
h\ i \
−1 −1 −1 −1
f [U ∩ V ] = f [U ] ∩ f [V ] f Uλ = f −1 [Uλ ] (4.36)
λ∈J λ∈J
−1 c −1 c
(f [U ]) = f [U ] (4.37)
h i
A ⊂ f −1 f [A] (4.38)

Proof: The proofs of the above are straightforward. We prove (4.36) as an

example of how to set out such proofs.
We need to show that f −1 [U ∩V ] ⊂ f −1 [U ]∩f −1 [V ] and f −1 [U ]∩f −1 [V ] ⊂
f −1 [U ∩ V ].
For the first, suppose x ∈ f −1 [U ∩V ]. Then f (x) ∈ U ∩V ; hence f (x) ∈ U
and f (x) ∈ V . Thus x ∈ f −1 [U ] and x ∈ f −1 [V ], so x ∈ f −1 [U ] ∩ f −1 [V ].
Thus f −1 [U ∩ V ] ⊂ f −1 [U ] ∩ f −1 [V ] (since x was an arbitrary member of
f −1 [U ∩ V ]).
Next suppose x ∈ f −1 [U ] ∩ f −1 [V ]. Then x ∈ f −1 [U ] and x ∈ f −1 [V ].
Hence f (x) ∈ U and f (x) ∈ V . This implies f (x) ∈ U ∩ V and so x ∈
f −1 [U ∩ V ]. Hence f −1 [U ] ∩ f −1 [V ] ⊂ f −1 [U ∩ V ].

Exercise Give a simple example to show equality need not hold in (4.34).

4.5 Equivalence of Sets

Definition 4.5.1 Two sets A and B are equivalent or equinumerous if there
exists a function f : A → B which is one-one and onto. We write A ∼ B.

The idea is that the two sets A and B have the same number of elements.
Thus the sets {a, b, c}, {x, y, z} and that in (4.2) are equivalent.
Some immediate consequences are:

Proposition 4.5.2

1. A ∼ A (i.e. ∼ is reflexive).
2. If A ∼ B then B ∼ A (i.e. ∼ is symmetric).
3. If A ∼ B and B ∼ C then A ∼ C (i.e. ∼ is transitive).

Proof: The first claim is clear.

For the second, let f : A → B be one-one and onto. Then the inverse
function f −1 : B → A, is also one-one and onto, as one can check (exercise).
For the third, let f : A → B be one-one and onto, and let g : B → C be
one-one and onto. Then the composition g ◦ f : A → B is also one-one and
onto, as can be checked (exercise).
42

Definition 4.5.3 A set is finite if it is empty or is equivalent to the set

{1, 2, . . . , n} for some natural number n. Otherwise it is infinite.

When we consider infinite sets there are some results which may seem
surprising at first:

• The set E of even natural numbers is equivalent to the set N of natural

numbers.

To see this, let f : E → N be given by f (n) = n/2. Then f

is one-one and onto.

• The open interval (a, b) is equivalent to R (if a < b).

To see this let f1 (x) = (x − a)/(b − a); then f1 : (a, b) → (0, 1)

is one-one and onto, and so (a, b) ∼ (0, 1). Next let f2 (x) =
x/(1 − x); then f2 : (0, 1) → (0, ∞) is one-one and onto6
and so (0, 1) ∼ (0, ∞). Finally, if f3 (x) = (1/x) − x then
f3 : (0, ∞) → R is one-one and onto7 and so (0, ∞) ∼ R.
Putting all this together and using the transitivity of set
equivalence, we obtain the result.

Thus we have examples where an apparently smaller subset of N (respectively

R) is in fact equivalent to N (respectively R).

4.6 Denumerable Sets

Definition 4.6.1 A set is denumerable if it is equivalent to N. A set is
countable if it is finite or denumerable. If a set is denumerable, we say it has
cardinality d or cardinal number d 8 .

Thus a set is denumerable iff it its members can be enumerated in a (non-

terminating) sequence (a1 , a2 , . . . , an , . . .). We show below that this fails to
hold for infinite sets in general.
The following may not seem surprising but it still needs to be proved.

Theorem 4.6.2 Any denumerable set is infinite (i.e. is not finite).

Proof: It is sufficient to show that N is not finite (why?). But in fact any
finite subset of N is bounded, whereas we know that N is not (Chapter 3).
6
This is clear from the graph of f2 . More precisely:
(i) if x ∈ (0, 1) then x/(1 − x) ∈ (0, ∞) follows from elementary properties of inequalities,
(ii) for each y ∈ (0, ∞) there is a unique x ∈ (0, 1) such that y = x/(1 − x), namely
x = y/(1 + y), as follows from elementary algebra and properties of inequalities.
7
As is again clear from the graph of f3 , or by arguments similar to those used for for f2 .
8
See Section 4.8 for a more general discussion of cardinal numbers.
Set Theory 43

We have seen that the set of even integers is denumerable (and similarly
for the set of odd integers). More generally, the following result is straight-
forward (the only problem is setting out the proof in a reasonable way):

Theorem 4.6.3 Any subset of a countable set is countable.

Proof: Let A be countable and let (a1 , a2 , . . . , an ) or (a1 , a2 , . . . , an , . . .) be

an enumeration of A (depending on whether A is finite or denumerable). If
B ⊂ A then we construct a subsequence (ai1 , ai2 , . . . , ain , . . .) enumerating B
by taking aij to be the j’th member of the original sequence which is in B.
Either this process never ends, in which case B is denumerable, or it does
end in a finite number of steps, in which case B is finite.

Remark This proof is rather more subtle than may appear. Why is the
resulting function from N → B onto? We should really prove that every
non-empty set of natural numbers has a least member, but for this we need
to be a little more precise in our definition of N. See [St, pp 13–15] for
details.

More surprising, at least at first, is the following result:

Theorem 4.6.4 The set Q is denumerable.

Proof: We have to show that N is equivalent to Q.

In order to simplify the notation just a little, we first prove that N is
equivalent to the set Q+ of positive rationals. We do this by arranging the
rationals in a sequence, with no repetitions.
Each rational in Q+ can be uniquely written in the reduced form m/n
where m and n are positive integers with no common factor. We write down
a “doubly-infinite” array as follows:

In the first row are listed all positive rationals whose reduced
form is m/1 for some m (this is just the set of natural numbers);
In the second row are all positive rationals whose reduced form
is m/2 for some m;
In the third row are all positive rationals whose reduced form is
m/3 for some m;
...
44

The enumeration we use for Q+ is shown in the following diagram:

1/1 2/1 → 3/1 4/1 → 5/1 . . .

↓ ↑ ↓ ↑ ↓
1/2 → 3/2 5/2 7/2 9/2 . . .
↓ ↑ ↓
1/3 ← 2/3 ← 4/3 5/3 7/3 . . . (4.39)
↓ ↑ ↓
1/4 → 3/4 → 5/4 → 7/4 9/4 . . .
↓
.. .. .. .. .. . .
. . . . . .

Finally, if a1 , a2 , . . . is the enumeration of Q+ then 0, a1 , −a1 , a2 , −a2 , . . .

is an enumeration of Q.

We will see in the next section that not all infinite sets are denumerable.
However denumerable sets are the smallest infinite sets in the following sense:

Theorem 4.6.5 If A is infinite then A contains a denumerable subset.

Proof: Since A 6= ∅ there exists at least one element in A; denote one such
element by a1 . Since A is not finite, A 6= {a1 }, and so there exists a2 , say,
where a2 ∈ A, a2 6= a1 . Similarly there exists a3 , say, where a3 ∈ A, a3 6= a2 ,
a3 6= a1 . This process will never terminate, as otherwise A ∼ {a1 , a2 , . . . , an }
for some natural number n.
Thus we construct a denumerable set B = {a1 , a2 , . . .}9 where B ⊂ A.

4.7 Uncountable Sets

There now arises the question

Are all infinite sets denumerable?

It turns out that the answer is No, as we see from the next theorem. Two
proofs will be given, both are due to Cantor (late nineteenth century), and
the underlying idea is the same.

Theorem 4.7.1 The sets N and (0, 1) are not equivalent.

The first proof is by an ingenious diagonalisation argument. There are a

couple of footnotes which may help you understand the proof.
9
To be precise, we need the so-called Axiom of Choice to justify the construction of B
by means of an infinite number of such choices of the ai . See 4.10.1 below.
Set Theory 45

Proof: 10 We show that for any f : N → (0, 1), the map f cannot be onto.
It follows that there is no one-one and onto map from N to (0, 1).
To see this, let yn = f (n) for each n. If we write out the decimal expansion
for each yn , we obtain a sequence

y1 = .a11 a12 a13 . . . a1i . . .

y2 = .a21 a22 a23 . . . a2i . . .
y3 = .a31 a32 a33 . . . a3i . . . (4.40)
..
.
yi = .ai1 ai2 ai3 . . . aii . . .
..
.

Some rational numbers have two decimal expansions, e.g. .14000 . . . = .13999 . . .
but otherwise the decimal expansion is unique. In order to have uniqueness,
we only consider decimal expansions which do not end in an infinite sequence
of 9’s.
To show that f cannot be onto we construct a real number z not in the
above sequence, i.e. a real number z not in the range of f . To do this define
z = .b1 b2 b3 . . . bi . . . by “going down the diagonal” as follows:

Select b1 6= a11 , b2 6= a22 , b3 6= a33 , . . . ,bi 6= aii ,. . . . We make

sure that the decimal expansion for z does not end in an infinite
sequence of 9’s by also restricting bi 6= 9 for each i; one explicit
construction would be to set bn = ann + 1 mod 9.

It follows that z is not in the sequence (4.40)11 , since for each i it is clear
that z differs from the i’th member of the sequence in the i’th place of z’s
decimal expansion. But this implies that f is not onto.

Here is the second proof.

Proof: Suppose that (an ) is a sequence of real numbers, we show that there
is a real number r ∈ (0, 1) such that r 6= an for every n.
Let I1 be a closed subinterval of (0, 1) with a1 6∈ I1 , I2 a closed subinterval
of I1 such that a2 6∈ I2 . Inductively, we obtain a sequence (In ) of intervals
such that In+1 ⊆ In for all n. Writing In = [αn , βn ], the nesting of the
intervals shows that αn ≤ αn+1 < βn+1 ≤ βn . In particular, (αn ) is bounded
above, (βn ) is bounded below, so that α = supn αn , β = inf n βn are defined.
Further it is clear that [α, β] ⊆ In for all n, and hence excludes all the (an ).
Any r ∈ [α, β] suffices.
10
We will show that any sequence (“list”) of real numbers from (0, 1) cannot include all
numbers from (0, 1). In fact, there will be an uncountable (see Definition 4.7.3) set of real
numbers not in the list — but for the proof we only need to find one such number.
11
First convince yourself that we really have constructed a number z. Then convince
yourself that z is not in the list, i.e. z is not of the form yn for any n.
46

Corollary 4.7.2 N is not equivalent to R.

Proof: If N ∼ R, then since R ∼ (0, 1) (from Section 4.5), it follows that

N ∼ (0, 1) from Proposition (4.5.2). This contradicts the preceding theorem.

A Common Error Suppose that A is an infinite set. Then it is not always

correct to say “let A = {a1 , a2 , . . .}”. The reason is of course that this
implicitly assumes that A is countable.

Definition 4.7.3 A set is uncountable if it is not countable. If a set is

equivalent to R we say it has cardinality c (or cardinal number c)12 .

Another surprising result (again due to Cantor) which we prove in the

next section is that the cardinality of R2 = R × R is also c.

Remark We have seen that the set of rationals has cardinality d. It follows13
that the set of irrationals has cardinality c. Thus there are “more” irrationals
than rationals.
On the other hand, the rational numbers are dense in the reals, in the
sense that between any two distinct real numbers there is a rational number14 .
(It is also true that between any two distinct real numbers there is an irra-
tional number15 .)

4.8 Cardinal Numbers

The following definition extends the idea of the number of elements in a set
from finite sets to infinite sets.

Definition 4.8.1 With every set A we associate a symbol called the cardinal
number of A and denoted by A. Two sets are assigned the same cardinal
number iff they are equivalent16 . Thus A = B iff A ∼ B.
12
c comes from continuum, an old way of referring to the set R.
13
We show in one of the problems for this chapter that if A has cardinality c and B ⊂ A
has cardinality d, then A \ B has cardinality c.
14
Suppose a < b. Choose an integer n such that 1/n < b − a. Then a < m/n < b for
some integer m. √
15
Using the notation of the previous footnote, take the irrational number m/n + 2/N
for some sufficiently large natural number N .
16
We are able to do this precisely because the relation of equivalence is reflexive, sym-
metric and transitive. For example, suppose 10 people are sitting around a round table.
Define a relation between people by A ∼ B iff A is sitting next to B, or A is the same
as B. It is not possible to assign to each person at the table a colour in such a way that
two people have the same colour if and only if they are sitting next to each other. The
problem is that the relation we have defined is reflexive and symmetric, but not transitive.
Set Theory 47

If A = ∅ we write A = 0.
If A = {a1 , . . . , an } (where a1 , . . . , an are all distinct) we write A = n.
If A ∼ N we write A = d (or ℵ0 , called “aleph zero”, where ℵ is the first
letter of the Hebrew alphabet).
If A ∼ R we write A = c.

Definition 4.8.2 Suppose A and B are two sets. We write A ≤ B (or

B ≥ A) if A is equivalent to some subset of B, i.e. if there is a one-one map
from A into B 17 .
If A ≤ B and A 6= B, then we write A < B (or B > A)18 .

Proposition 4.8.3

0 < 1 < 2 < 3 < . . . < d < c. (4.41)

Proof: Consider the sets

{a1 }, {a1 , a2 }, {a1 , a2 , a3 }, . . . , N, R,

where a1 , a2 , a3 , . . . are distinct from one another. There is clearly a one-one

map from any set in this “list” into any later set in the list (why?), and so

1 ≤ 2 ≤ 3 ≤ . . . ≤ d ≤ c. (4.42)

For any integer n we have n 6= d from Theorem 4.6.2, and so n < d

6 c from Corollary 4.7.2, it also follows that d < c
from (4.42). Since d =
from (4.42).
Finally, the fact that 1 6= 2 6= 3 6= . . . (and so 1 < 2 < 3 < . . . from (4.42))
can be proved by induction.

Proposition 4.8.4 Suppose A is non-empty. Then for any set B, there

exists a surjective function g : B → A iff A ≤ B.

Proof: If g is onto, we can choose for every x ∈ A an element y ∈ B such

that g(y) = x. Denote this element by f (x)19 . Thus g(f (x)) = x for all
x ∈ A.
Then f : A → B and f is clearly one-one (since if f (x1 ) = f (x2 ) then
g(f (x1 )) = g(f (x2 )); but g(f (x1 )) = x1 and g(f (x2 )) = x2 , and hence x1 =
x2 ). Hence A ≤ B.
17
This does not depend on the choice of sets A and B. More precisely, suppose A ∼ A0
and B ∼ B 0 , so that A = A0 and B = B 0 . Then A is equivalent to some subset of B iff A0
is equivalent to some subset of B 0 (exercise).
18
This is also independent of the choice of sets A and B in the sense of the previous
footnote. The argument is similar.
19
This argument uses the Axiom of Choice, see Section 4.10.1 below.
48

Conversely, if A ≤ B then there exists a function f : A → B which is

one-one. Since A is non-empty there is an element in A and we denote one
such member by a. Now define g : B → A by
(
x if f (x) = y,
g(y) =
a if there is no such x.

Then g is clearly onto, and so we are done.

We have the following important properties of cardinal numbers, some of

which are trivial, and some of which are surprisingly difficult. Statement 2
is known as the Schröder-Bernstein Theorem.

Theorem 4.8.5 Let A, B and C be cardinal numbers. Then

1. A ≤ A;

2. A ≤ B and B ≤ A implies A = B;

3. A ≤ B and B ≤ C implies A ≤ C;

4. either A ≤ B or B ≤ A.

Proof: The first and the third results are simple. The first follows from
Theorem 4.5.2(1) and the third from Theorem 4.5.2(3).
The other two result are not easy.
*Proof of (2): Since A ≤ B there exists a function f : A → B which is
one-one (but not necessarily onto). Similarly there exists a one-one function
g : B → A since B ≤ A.
If f (x) = y or g(u) = v we say x is a parent of y and u is a parent of v.
Since f and g are one-one, each element has exactly one parent, if it has any.
If y ∈ B and there is a finite sequence x1 , y1 , x2 , y2 , . . . , xn , y or y0 , x1 , y1 ,
x2 , y2 , . . . , xn , y, for some n, such that each member of the sequence is the
parent of the next member, and such that the first member has no parent,
then we say y has an original ancestor, namely x1 or y0 respectively. Notice
that every member in the sequence has the same original ancestor. If y has
no parent, then y is its own original ancestor. Some elements may have no
original ancestor.
Let A = AA ∪ AB ∪ A∞ , where AA is the set of elements in A with original
ancestor in A, AB is the set of elements in A with original ancestor in B,
and A∞ is the set of elements in A with no original ancestor. Similarly let
B = BA ∪ BB ∪ B∞ , where BA is the set of elements in B with original
ancestor in A, BB is the set of elements in B with original ancestor in B,
and B∞ is the set of elements in B with no original ancestor.
Define h : A → B as follows:
Set Theory 49

if x ∈ AA then h(x) = f (x),

if x ∈ AB then h(x) = the parent of x,
if x ∈ A∞ then h(x) = f (x).

Note that every element in AB must have a parent (in B), since if it did
not have a parent in B then the element would belong to AA . It follows that
the definition of h makes sense.
If x ∈ AA , then h(x) ∈ BA , since x and h(x) must have the same original
ancestor (which will be in A). Thus h : AA → BA . Similarly h : AB → BB
and h : A∞ → B∞ .
Note that h is one-one, since f is one-one and since each x ∈ AB has
exactly one parent.
Every element y in BA has a parent in A (and hence in AA ). This parent
is mapped to y by f and hence by h, and so h : AA → BA is onto. A similar
argument shows that h : A∞ → B∞ is onto. Finally, h : AB → BB is onto as
each element y in BB is the image under h of g(y). It follows that h is onto.
Thus h is one-one and onto, as required. End of proof of (2).
*Proof of (4): We do not really have the tools to do this, see Section 4.10.1
below. One lets

F = {f | f : U → V, U ⊂ A, V ⊂ B, f is one-one and onto}.

It follows from Zorn’s Lemma, see 4.10.1 below, that F contains a maximal
element. Either this maximal element is a one-one function from A into B,
or its inverse is a one-one function from B into A.

Corollary 4.8.6 Exactly one of the following holds:

A < B or A = B or B < A. (4.43)

Proof: Suppose A = B. Then the second alternative holds and the first
and third do not.
Suppose A 6= B. Either A ≤ B or B ≤ A from the previous theorem.
Again from the previous theorem exactly one of these possibilities can hold,
as both together would imply A = B. If A ≤ B then in fact A < B since
A 6= B. Similarly, if B ≤ A then B < A.

Corollary 4.8.7 If A ⊂ R and A includes an interval of positive length,

then A has cardinality c.

Proof: Suppose I ⊂ A where I is an interval of positive length. Then

I ≤ A ≤ R. Thus c ≤ A ≤ c, using the result at the end of Section 4.5 on
the cardinality of an interval.
Hence A = c from the Schröder-Bernstein Theorem.
50

NB The converse to Corollary 4.8.7 is false. As an example, consider the

following set
∞
X
S={ an 3−n : an = 0 or 2} (4.44)
n=1
P P
The mapping S → [0, 1] : ∞ n=1 3n 7→
an ∞
n=1 an 2
−n−1
takes S onto [0, 1], so
that S must be uncountable. On the other hand, S contains no interval at
all. To see this, it suffices to show that for any x ∈ S, and any ² > 0 there are
points in [x, x+²] lying outside S. It is a calculation to verify that x+a3−k is
such a point for suitably large k, and suitable choice of a = 1 or 2 (exercise).
The set S above is known as the Cantor ternary set. It has further
important properties which you will come across in topology and measure
theory, see also Section 14.1.2.

We now prove the result promised at the end of the previous Section.

Theorem 4.8.8 The cardinality of R2 = R × R is c.

Proof: Let f : (0, 1) → R be one-one and onto, see Section 4.5. The map
(x, y) 7→ (f (x), f (y)) is thus (exercise) a one-one map from (0, 1)×(0, 1) onto
R × R; thus (0, 1) × (0, 1) ∼ R × R. Since also (0, 1) ∼ R, it is sufficient to
show that (0, 1) ∼ (0, 1) × (0, 1).
Consider the map f : (0, 1) × (0, 1) → (0, 1) given by

(x, y) = (.x1 x2 x3 . . . , .y1 y2 y3 . . .) 7→ .x1 y1 x2 y2 x3 y3 . . . (4.45)

We take the unique decimal expansion for each of x and y given by requiring
that it does not end in an infinite sequence of 9’s. Then f is one-one but
not onto (since the number .191919 . . . for example is not in the range of f ).
Thus (0, 1) × (0, 1) ≤ (0, 1).
On the other hand, there is a one-one map g : (0, 1) → (0, 1) × (0, 1) given
by g(z) = (z, 1/2), for example. Thus (0, 1) ≤ (0, 1) × (0, 1).
Hence (0, 1) = (0, 1) × (0, 1) from the Schröder-Bernstein Theorem, and
the result follows as (0, 1) = c.

The same argument, or induction, shows that Rn has cardinality c for

each n ∈ N. But what about RN = {F : N → R}?

4.9 More Properties of Sets of Cardinality c

and d
Theorem 4.9.1

1. The product of two countable sets is countable.

Set Theory 51

2. The product of two sets of cardinality c has cardinality c.

3. The union of a countable family of countable sets is countable.
4. The union of a cardinality c family of sets each of cardinality c has
cardinality c.

Proof: (1) Let A = (a1 , a2 , . . .) and B = (b1 , b2 , . . .) (assuming A and B are

infinite; the proof is similar if either is finite). Then A×B can be enumerated
as follows (in the same way that we showed the rationals are countable):
(a1 , b1 ) (a1 , b2 ) → (a1 , b3 ) (a1 , b4 ) → (a1 , b5 ) ...
↓ ↑ ↓ ↑ ↓
(a2 , b1 ) → (a2 , b2 ) (a2 , b3 ) (a2 , b4 ) (a2 , b5 ) ...
↓ ↑ ↓
(a3 , b1 ) ← (a3 , b2 ) ← (a3 , b3 ) (a3 , b4 ) (a3 , b5 ) ... (4.46)
↓ ↑ ↓
(a4 , b1 ) → (a4 , b2 ) → (a4 , b3 ) → (a4 , b4 ) (a4 , b5 ) ...
↓
.. .. .. .. .. ...
. . . . .

(2) If the sets A and B have cardinality c then they are in one-one
correspondence20 with R. It follows that A × B is in one-one correspon-
dence with R × R, and so the result follows from Theorem 4.8.8.
(3) Let {Ai }∞
i=1 be a countable family of countable sets. Consider an array
whose first column enumerates the members of A1 , whose second column
enumerates the members of A2 , etc. Then an enumeration similar to that
in (1), but suitably modified to take account of the facts that some columns
may be finite, that the number of columns may be finite, and that some
elements may appear in more than one column, gives the result.
(4) Let {Aα }α∈S be a family of sets each of cardinality c, where the index
set S has cardinality c. Let fα : Aα → R be a one-one and onto function for
each α.
S
Let A = α∈S Aα and define f : A → R × R by f (x) = (α, fα (x)) if
x ∈ Aα (if x ∈ Aα for more than one α, choose one such α 21 ). It follows
that A ≤ R × R, and so A ≤ c from Theorem 4.8.8.
On the other hand there is a one-one map g from R into A (take g equal
to the inverse of fα for some α ∈ S) and so c ≤ A.
The result now follows from the Schröder-Bernstein Theorem.

Remark The phrase “Let fα : Aα → R be a one-one and onto function for

each α” looks like another invocation of the axiom of choice, however one
20
A and B are in one-one correspondence means that there is a one-one map from A
onto B.
21
We are using the Axiom of Choice in simultaneously making such a choice for each
x ∈ A.
52

could interpret the hypothesis on {Aα }α∈S as providing the maps fα . This
has implicitly been done in (3) and (4).

Remark It is clear from the proof that in (4) it is sufficient to assume that
each set is countable or of cardinality c, provided that at least one of the sets
has cardinality c.

4.10 *Further Remarks

4.10.1 The Axiom of choice

For any non-empty set X, there is a function f : P(X) → X such that
f (A) ∈ A for A ∈ P(X)\{∅}.
This axiom has a somewhat controversial history – it has some innocuous
equivalences (see below), but other rather startling consequences such as the
Banach-Tarski paradox22 . It is known to be independent of the other usual
axioms of set theory (it cannot be proved or disproved from the other axioms)
and relatively consistent (neither it, nor its negation, introduce any new
inconsistencies into set theory). Nowadays it is almost universally accepted
and used without any further ado. For example, it is needed to show that
any vector space has a basis or that the infinite product of non-empty sets
is itself non-empty.

Theorem 4.10.1 The following are equivalent to the axiom of choice:

1. If h is a function with domain A, there is a function f with domain A

such that if x ∈ A and h(x) 6= ∅, then f (x) ∈ h(x).

2. If ρ ⊆ A × B is a relation with domain A, then there exists a function

f : A → B with f ⊆ ρ.

3. If g : B → A is onto, then there exists f : A → B such that g ◦ f =

identity on A.

Proof: These are all straightforward; (3) was used in 4.8.4.

For some of the most commonly used equivalent forms we need some
further concepts.

Definition 4.10.2 A relation ≤ on a set X is a partial order on X if, for

all x, y, z ∈ X,

1. (x ≤ y) ∧ (y ≤ x) ⇒ x = y (antisymmetry), and
22
This says that a ball in R3 can be divided into five pieces which can be rearranged by
rigid body motions to give two disjoint balls of the same radius as before!
Set Theory 53

2. (x ≤ y) ∧ (y ≤ z) ⇒ x ≤ z (transitivity), and

3. x ≤ x for all x ∈ X (reflexivity).

An element x ∈ X is maximal if (y ∈ X) ∧ (x ≤ y) ⇒ y = x, x is
maximum (= greatest) if z ≤ x for all z ∈ X. Similar for minimal and
minimum ( = least), and upper and lower bounds.
A subset Y of X such that for any x, y ∈ Y , either x ≤ y or y ≤ x is
called a chain. If X itself is a chain, the partial order is a linear or total
order. A linear order ≤ for which every non-empty subset of X has a least
element is a well order.

Remark Note that if ≤ is a partial order, then ≥, defined by x ≥ y := y ≤

x, is also a partial order. However, if both ≤ and ≥ are well orders, then the
set is finite. (exercise).

With this notation we have the following, the proof of which is not easy
(though some one way implications are).

Theorem 4.10.3 The following are equivalent to the axiom of choice:

1. Zorn’s Lemma A partially ordered set in which any chain has an

upper bound has a maximal element.

2. Hausdorff maximal principle Any partially ordered set contains a

maximal chain.

3. Zermelo well ordering principle Any set admits a well order.

4. Teichmuller/Tukey maximal principle For any property of finite

character on the subsets of a set, there is a maximal subset with the
property23 .

4.10.2 Other Cardinal Numbers

We have examples of infinite sets of cardinality d (e.g. N ) and c (e.g. R).
A natural question is:

Are there other cardinal numbers?

The following theorem implies that the answer is YES.

Theorem 4.10.4 If A is any set, then A < P(A).

23
A property of subsets is of finite character if a subset has the property iff all of its
finite (sub)subsets have the property.
54

Proof: The map a → {a} is a one-one map from A into P(A).

If f : A → P(A), let

X = {a ∈ A : a 6∈ f (a)} . (4.47)

Then X ∈ P(A); suppose X = f (b) for some b in A. If b ∈ X then b 6∈ f (b)

(from the defining property of X), contradiction. If b 6∈ X then b ∈ f (b)
(again from the defining property of X), contradiction. Thus X is not in the
range of f and so f cannot be onto.

Remark Note that the argument is similar to that used to obtain Russell’s
Paradox.

Remark Applying the previous theorem successively to A = R, P(R),

P(P(R)), . . . we obtain an increasing sequence of cardinal numbers. We can
take the union S of all sets thus constructed, and it’s cardinality is larger
still. Then we can repeat the procedure with R replaced by S, etc., etc. And
we have barely scratched the surface!

It is convenient to introduce the notation A ∪· B to indicate the union of

A and B, considering the two sets to be disjoint.

Theorem 4.10.5 The following are equivalent to the axiom of choice:

1. If A and B are two sets then either A ≤ B or B ≤ A.

2. If A and B are two sets, then (A × A = B × B) ⇒ A = B.

3. A × A = A for any infinite set A. ( cf 4.9.1)

4. A × B = A ∪· B for any two infinite sets A and B.

However A ∪· A = A for all infinite sets A 6⇒ AC.

4.10.3 The Continuum Hypothesis

Another natural question is:

Is there a cardinal number between c and d?

More precisely: Is there an infinite set A ⊂ R with no one-one map from

A onto N and no one-one map from A onto R? All infinite subsets of R
that arise “naturally” either have cardinality c or d. The assertion that all
infinite subsets of R have this property is called the Continuum Hypothesis
(CH). More generally the assertion that for every infinite set A there is no
Set Theory 55

cardinal number between A and P(A) is the Generalized Continuum Hypoth-

esis (GCH). It has been proved that the CH is an independent axiom in set
theory, in the sense that it can neither be proved nor disproved from the
other axioms (including the axiom of choice)24 . Most, but by no means all,
mathematicians accept at least CH. The axiom of choice is a consequence of
GCH.

4.10.4 Cardinal Arithmetic

If α = A and β = B are infinite cardinal numbers, we define their sum and
product by

α + β = A ∪· B (4.48)
α × β = A × B, (4.49)

From Theorem 4.10.5 it follows that α + β = α × β = max{α, β}.

More interesting is exponentiation; we define

αβ = {f | f : B → A}. (4.50)

Why is this consistent with the usual definition of mn and Rn where m and
n are natural numbers?
For more information, see [BM, Chapter XII].

4.10.5 Ordinal numbers

Well ordered sets were mentioned briefly in 4.10.1 above. They are precisely
the sets on which one can do (transfinite) induction. Just as cardinal num-
bers were introduced to facilitate the “size” of sets, ordinal numbers may be
introduced as the “order-types” of well-ordered sets. Alternatively they may
be defined explicitly as sets W with the following three properties.

1. every member of W is a subset of W

2. W is well ordered by ⊂

3. no member of W is an member of itself

Then N, and its elements are ordinals, as is N ∪ {N}. Recall that for
n ∈ N, n = {m ∈ N : m < n}. An ordinal number in fact is equal to the set
of ordinal numbers less than itself.

24
We will discuss the Zermelo-Fraenkel axioms for set theory in a later course.
56
Chapter 5

Vector Space Properties of Rn

In this Chapter we briefly review the fact that Rn , together with the usual
definitions of addition and scalar multiplication, is a vector space. With
the usual definition of Euclidean inner product, it becomes an inner product
space.

5.1 Vector Spaces

Definition 5.1.1 A Vector Space (over the reals1 ) is a set V (whose mem-
bers are called vectors), together with two operations called addition and
scalar multiplication, and a particular vector called the zero vector and de-
noted by 0. The sum (addition) of u, v ∈ V is a vector2 in V and is denoted
u + v; the scalar multiple of the scalar (i.e. real number) c ∈ R and u ∈ V
is a vector in V and is denoted cu. The following axioms are satisfied:

1. u + v = v + u for all u, v ∈ V (commutative law)

2. u + (v + w) = (u + v) + w for all u, v, w ∈ V (associative law)

3. u + 0 = u for all u ∈ V (existence of an additive identity)

4. (c + d)u = cu + du, c(u + v) = cu + cv for all c, d ∈ R and u, v ∈ V

(distributive laws)

5. (cd)u = c(du) for all c, d ∈ R and u ∈ V

6. 1u = u for all u ∈ V

1
One can define a vector space over the complex numbers in an analogous manner.
2
It is common to denote vectors in boldface type.

57
58

Examples

1. Recall that Rn is the set of all n-tuples (a1 , . . . , an ) of real numbers.

The sum of two n-tuples is defined by

(a1 , . . . , an ) + (b1 , . . . , bn ) = (a1 + b1 , . . . , an + bn ) 3 . (5.1)

The product of a scalar and an n-tuple is defined by

c(a1 , . . . , an ) = (ca1 , . . . , can ). (5.2)

The zero n-vector is defined to be

(0, . . . , 0). (5.3)

With these definitions it is easily checked that Rn becomes a vector

space.

2. Other very important examples of vector spaces are various spaces of

functions. For example C[a, b], the set of continuous4 real-valued func-
tions defined on the interval [a, b], with the usual addition of functions
and multiplication of a scalar by a function, is a vector space (what is
the zero vector?).

Remarks You should review the following concepts for a general vector
space (see [F, Appendix 1] or [An]):

• linearly independent set of vectors, linearly dependent set of vectors,

• basis for a vector space, dimension of a vector space,

• linear operator between vector spaces.

The standard basis for Rn is defined by

e1 = (1, 0, . . . , 0)
e2 = (0, 1, . . . , 0)
.. (5.4)
.
en = (0, 0, . . . , 1)

Geometric Representation of R2 and R3 The vector x = (x1 , x2 ) ∈ R2

is represented geometrically in the plane either by the arrow from the origin
(0, 0) to the point P with coordinates (x1 , x2 ), or by any parallel arrow of
the same length, or by the point P itself. Similar remarks apply to vectors
in R3 .
3
This is not a circular definition; we are defining addition of n-tuples in terms of
addition of real numbers.
4
We will discuss continuity in a later chapter. Meanwhile we will just use C[a, b] as a
source of examples.
Vector Space Properties of Rn 59

5.2 Normed Vector Spaces

A normed vector space is a vector space together with a notion of magnitude
or length of its members, which satisfies certain axioms. More precisely:

Definition 5.2.1 A normed vector space is a vector space V together with

a real-valued function on V called a norm. The norm of u is denoted by
||u|| (sometimes |u|). The following axioms are satisfied for all u ∈ V and
all α ∈ R:

1. ||u|| ≥ 0 and ||u|| = 0 iff u = 0 (positivity),

2. ||αu|| = |α| ||u|| (homogeneity),

3. ||u + v|| ≤ ||u|| + ||v|| (triangle inequality).

We usually abbreviate normed vector space to normed space.

Easy and important consequences (exercise) of the triangle inequality are

||u|| ≤ ||u − v|| + ||v||, (5.5)

¯ ¯
¯ ¯
¯||u|| − ||v||¯ ≤ ||u − v||. (5.6)

Examples

1. The vector space Rn is a normed space if we define ||(x1 , . . . , xn )||2 =

³ ´1/2
(x1 )2 + · · · +(xn )2 . The only non-obvious part to prove is the
triangle inequality. In the next section we will see that Rn is in fact
an inner product space, that the norm we just defined is the norm
corresponding to this inner product, and we will see that the triangle
inequality is true for the norm in any inner product space.

2. There are other norms which we can define on Rn . For 1 ≤ p < ∞,

à n !1/p
X
||(x1 , . . . , xn )||p = |xi |p (5.7)
i=1

defines a norm on Rn , called the p-norm. It is also easy to check that

||(x1 , . . . , xn )||∞ = max{|x1 |, . . . , |xn |} (5.8)

defines a norm on Rn , called the sup norm. Exercise: Show this nota-
tion is consistent, in the sense that

lim ||x||p = ||x||∞ (5.9)

p→∞
60

3. Similarly, it is easy to check (exercise) that the sup norm on C[a, b]

defined by

||f ||∞ = sup |f | = sup {|f (x)| : a ≤ x ≤ b} (5.10)

is indeed a norm. (Note, incidentally, that since f is continuous, it

follows that the sup on the right side of the previous equality is achieved
at some x ∈ [a, b], and so we could replace sup by max.)

4. A norm on C[a, b] is defined by

Z b
||f ||1 = |f |. (5.11)
a

Exercise: Check that this is a norm.

C[a, b] is also a normed space5 with
ÃZ !1/2
b
||f || = ||f ||2 = f 2
. (5.12)
a

Once again the triangle inequality is not obvious. We will establish it

in the next section.

5. Other examples are the set of all bounded sequences on N:

`∞ (N) = {(xn ) : ||(xn )||∞ = sup |xn | < ∞}. (5.13)

and its subset c0 (N) of those sequences which converge to 0.

6. On the other hand, for RN , which is clearly a vector space under point-
wise operations, has no natural norm. Why?

5.3 Inner Product Spaces

A (real) inner product space is a vector space in which there is a notion of
magnitude and of orthogonality, see Definition 5.3.2. More precisely:

Definition 5.3.1 An inner product space is a vector space V together with

an operation called inner product. The inner product of u, v ∈ V is a real
number denoted by u · v or (u, v)6 . The following axioms are satisfied for all
u, v, w ∈ V :

1. u · u ≥ 0, u · u = 0 iff u = 0 (positivity)

2. u · v = v · u (symmetry)
5
We will see the reason for the || · ||2 notation when we discuss the Lp norm.
6
Other notations are h·, ·i and (·|·).
Vector Space Properties of Rn 61

3. (u + v) · w = u · w + v · w, (cu) · v = c(u · v) (bilinearity)7

Remark In the complex case v · u = u · v. Thus from 2. and 3. The

inner product is linear in the first variable and conjugate linear in the second
variable, that is, it is sesquilinear .

Examples

1. The Euclidean inner product (or dot product or standard inner product)
of two vectors in Rn is defined by

(a1 , . . . , an ) · (b1 , . . . , bn ) = a1 b1 + · · · + an bn (5.14)

It is easily checked that this does indeed satisfy the axioms for an inner
product. The corresponding inner product space is denoted by E n
in [F], but we will abuse notation and use Rn for the set of n-tuples,
for the corresponding vector space, and for the inner product space just
defined.

2. One can define other inner products on Rn , these will be considered in

the algebra part of the course. One simple class of examples is given
by defining

(a1 , . . . , an ) · (b1 , . . . , bn ) = α1 a1 b1 + · · · + αn an bn , (5.15)

where α1 , . . . , αn is any sequence of positive real numbers. Exercise

Check that this defines an inner product.

3. Another important example of an inner

R
product space is C[a, b] with
the inner product defined by f · g = ab f g. Exercise: check that this
defines an inner product.

Definition 5.3.2 In an inner product space we define the length (or norm)
of a vector by
|u| = (u · u)1/2 , (5.16)
and the notion of orthogonality between two vectors by

u is orthogonal to v (written u ⊥ v) iff u · v = 0. (5.17)

Example The functions

1, cos x, sin x, cos 2x, sin 2x, . . . (5.18)

form an important (infinite) set of pairwise orthogonal functions in the inner

product space C[0, 2π], as is easily checked. This is the basic fact in the
theory of Fourier series (you will study this theory at some later stage).
7
Thus an inner product is linear in the first argument. Linearity in the second argument
then follows from 2.
62

Theorem 5.3.3 An inner product on V has the following properties: for

any u, v ∈ V ,

|u · v| ≤ |u| |v| (Cauchy-Schwarz-Bunyakovsky Inequality), (5.19)

and if v 6= 0 then equality holds iff u is a multiple of v.

Moreover, | · | is a norm, and in particular

|u + v| ≤ |u| + |v| (Triangle Inequality). (5.20)

If v 6= 0 then equality holds iff u is a nonnegative multiple of v.

The proof of the inequality is in [F, p. 6]. Although the proof given
there is for the standard inner product in Rn , the same proof applies to any
inner product space. A similar remark applies to the proof of the triangle
inequality in [F, p. 7]. The other two properties of a norm are easy to show.
An orthonormal basis for a finite dimensional inner product space is a
basis {v1 , . . . , vn } such that
(
0 if i 6= j
vi · vj = (5.21)
1 if i = j

Beginning from any basis {x1 , . . . , xn } for an inner product space, one can
construct an orthonormal basis {v1 , . . . , vn } by the Gram-Schmidt process
described in [F, p.10 Question 10]:

First construct v1 of unit length in the subspace generated by

x1 ; then construct v2 of unit length, orthogonal to v1 , and in
the subspace generated by x1 and x2 ; then construct v3 of unit
length, orthogonal to v1 and v2 , and in the subspace generated
by x1 , x2 and x3 ; etc.

If x is a unit vector (i.e. |x| = 1) in an inner product space then the

component of v in the direction of x is v · x. In particular, in Rn the
component of (a1 , . . . , an ) in the direction of ei is ai .
Chapter 6

Metric Spaces

Metric spaces play a fundamental role in Analysis. In this chapter we will

see that Rn is a particular example of a metric space. We will also study
and use other examples of metric spaces.

6.1 Basic Metric Notions in Rn

Definition 6.1.1 The distance between two points x, y ∈ Rn is given by
³ ´1/2
d(x, y) = |x − y| = (x1 − y 1 )2 + · · · + (xn − y n )2 .

Theorem 6.1.2 For all x, y, z ∈ Rn the following hold:

1. d(x, y) ≥ 0, d(x, y) = 0 iff x = y (positivity),

2. d(x, y) = d(y, x) (symmetry),

3. d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality).

Proof: The first two are immediate. For the third we have d(x, y) = |x −
y| = |x−z +z−y| ≤ |x−z|+|z −y| = d(x, z) +d(z, y), where the inequality
comes from version (5.20) of the triangle inequality in Section 5.3.

6.2 General Metric Spaces

We now generalise these ideas as follows:

Definition 6.2.1 A metric space (X, d) is a set X together with a distance

function d : X × X → R such that for all x, y, z ∈ X the following hold:

1. d(x, y) ≥ 0, d(x, y) = 0 iff x = y (positivity),

63
64

2. d(x, y) = d(y, x) (symmetry),

3. d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality).

We often denote the corresponding metric space by (X, d), to indicate that
a metric space is determined by both the set X and the metric d.

Examples

1. We saw in the previous section that Rn together with the distance

function defined by d(x, y) = |x − y| is a metric space. This is called
the standard or Euclidean metric on Rn .
Unless we say otherwise, when referring to the metric space Rn , we
will always intend the Euclidean metric.

2. More generally, any normed space is also a metric space, if we define

d(x, y) = ||x − y||.

The proof is the same as that for Theorem 6.1.2. As examples, the sup
norm on Rn , and both the inner product norm and the sup norm on
C[a, b] (c.f. Section 5.2), induce corresponding metric spaces.

3. An example of a metric space which is not a vector space is a smooth

surface S in R3 , where the distance between two points x, y ∈ S is
defined to be the length of the shortest curve joining x and y and lying
entirely in S. Of course to make this precise we first need to define
smooth, surface, curve, and length, as well as consider whether there
will exist a curve of shortest length (and is this necessary anyway?)

4. French metro, Post Office Let X = {x ∈ R2 : |x| ≤ 1} and define

(
|x − y| if x = ty for some scalar t
d(x, y) =
|x| + |y| otherwise

One can check that this defines a metric—the French metro with Paris
at the centre. The distance between two stations on different lines is
measured by travelling in to Paris and then out again.

5. p-adic metric. Let X = Z, and let p ∈ N be a fixed prime. For

x, y ∈ Z, x 6= y, we have x − y = pk n uniquely for some k ∈ N, and
some n ∈ Z not divisible by p. Define
(
(k + 1)−1 if x 6= y
d(x, y) =
0 if x = y

One can check that this defines a metric which in fact satisfies the
strong triangle inequality (which implies the usual one):

d(x, y) ≤ max{d(x, z), d(z, y)}.

Metric Spaces 65

Members of a general metric space are often called points, although they
may be functions (as in the case of C[a, b]), sets (as in 14.5) or other mathe-
matical objects.

Definition 6.2.2 Let (X, d) be a metric space. The open ball of radius r > 0
centred at x, is defined by

Br (x) = {y ∈ X : d(x, y) < r}. (6.1)

Note that the open balls in R are precisely the intervals of the form (a, b)
(the centre is (a + b)/2 and the radius is (b − a)/2).

Exercise: Draw the open ball of radius 1 about the point (1, 2) ∈ R2 , with
respect to the Euclidean (L2 ), sup (L∞ ) and L1 metrics. What about the
French metro?

It is often convenient to have the following notion 1 .

Definition 6.2.3 Let (X, d) be a metric space. The subset Y ⊂ X is a

neighbourhood of x ∈ X if there is R > 0 such that Br (x) ⊂ Y .

Definition 6.2.4 A subset S of a metric space X is bounded if S ⊂ Br (x)

for some x ∈ X and some r > 0.

Proposition 6.2.5 If S is a bounded subset of a metric space X, then for

every y ∈ X there exists ρ > 0 (ρ depending on y) such that S ⊂ Bρ (y).
In particular, a subset S of a normed space is bounded iff S ⊂ Br (0) for
some r, i.e. iff for some real number r, ||x|| < r for all x ∈ S.

1
Some definitions of neighbourhood require the set to be open (see Section 6.4 below).
66

Proof: Assume S ⊂ Br (x) and y ∈ X. Then Br (x) ⊂ Bρ (y) where ρ =

r+d(x, y); since if z ∈ Br (x) then d(z, y) ≤ d(z, x)+d(x, y) < r +d(x, y) = ρ,
and so z ∈ Bρ (y). Since S ⊂ Br (x) ⊂ Bρ (y), it follows S ⊂ Bρ (y) as required.

The previous proof is a typical application of the triangle inequality in a

metric space.

6.3 Interior, Exterior, Boundary and Closure

Everything from this section, including proofs, unless indicated otherwise and
apart from specific examples, applies with Rn replaced by an arbitrary metric
space (X, d).
The following ideas make precise the notions of a point being strictly
inside, strictly outside, on the boundary of, or having arbitrarily close-by
points from, a set A.

Definition 6.3.1 Suppose that A ⊂ Rn . A point x ∈ Rn is an interior

(respectively exterior ) point of A if some open ball centred at x is a subset
of A (respectively Ac ). If every open ball centred at x contains at least one
point of A and at least one point of Ac , then x is a boundary point of A.
The set of interior (exterior) points of A is called the interior (exterior )
of A and is denoted by A0 or int A (ext A). The set of boundary points of
A is called the boundary of A and is denoted by ∂A.

Proposition 6.3.2 Suppose that A ⊂ Rn .

Rn = int A ∪ ∂A ∪ ext A, (6.2)

ext A = int (Ac ), int A = ext (Ac ), (6.3)

int A ⊂ A, ext A ⊂ Ac . (6.4)
The three sets on the right side of (6.2) are mutually disjoint.

Proof: These all follow immediately from the previous definition, why?

We next make precise the notion of a point for which there are members
of A which are arbitrarily close to that point.

Definition 6.3.3 Suppose that A ⊂ Rn . A point x ∈ Rn is a limit point of

A if every open ball centred at x contains at least one member of A other
than x. A point x ∈ A ⊂ Rn is an isolated point of A if some open ball
centred at x contains no members of A other than x itself.
Metric Spaces 67

NB The terms cluster point and accumulation point are also used here.
However, the usage of these three terms is not universally the same through-
out the literature.

Definition 6.3.4 The closure of A ⊂ Rn is the union of A and the set of

limit points of A, and is denoted by A.

The following proposition follows directly from the previous definitions.

Proposition 6.3.5 Suppose A ⊂ Rn .

1. A limit point of A need not be a member of A.

2. If x is a limit point of A, then every open ball centred at x contains an
infinite number of points from A.
3. A ⊂ A.
4. Every point in A is either a limit point of A or an isolated point of A,
but not both.
5. x ∈ A iff every Br (x) (r > 0) contains a point of A.

Proof: Exercise.

Example 1 If A = {1, 1/2, 1/3, . . . , 1/n, . . .} ⊂ R, then every point in A is

an isolated point. The only limit point is 0.
If A = (0, 1] ⊂ R then there are no isolated points, and the set of limit
points is [0, 1].
Defining fn (t) = tn , set A = {m−1 fn : m, n ∈ N}. Then A has only limit
point 0 in (C[0, 1], k·k∞ ).

Theorem 6.3.6 If A ⊂ Rn then

A = (ext A)c . (6.5)
A = int A ∪ ∂A. (6.6)
A = A ∪ ∂A. (6.7)

Proof: For (6.5) first note that x ∈ A iff every Br (x) (r > 0) contains at
least one member of A. On the other hand, x ∈ ext A iff some Br (x) is a
subset of Ac , and so x ∈ (ext A)c iff it is not the case that some Br (x) is a
subset of Ac , i.e. iff every Br (x) contains at least one member of A.
Equality (6.6) follows from (6.5), (6.2) and the fact that the sets on the
right side of (6.2) are mutually disjoint.
For 6.7 it is sufficient from (6.5) to show A ∪ ∂A = (int A) ∪ ∂A. But
clearly (int A) ∪ ∂A ⊂ A ∪ ∂A.
On the other hand suppose x ∈ A ∪ ∂A. If x ∈ ∂A then x ∈ (int A) ∪ ∂A,
while if x ∈ A then x 6∈ ext A from the definition of exterior, and so x ∈
(int A) ∪ ∂A from (6.2). Thus A ∪ ∂A ⊂ (int A) ∪ ∂A.
68

Example 2

The following proposition shows that we need to be careful in relying too

much on our intuition for Rn when dealing with an arbitrary metric space.

Proposition 6.3.7 Let A = Br (x) ⊂ Rn . Then we have int A = A,

ext A = {y : d(y, x) > r}, ∂A = {y : d(y, x) = r} and A = {y : d(y, x) ≤ r}.
If A = Br (x) ⊂ X where (X, d) is an arbitrary metric space, then
int A = A, ext A ⊃ {y : d(y, x) > r}, ∂A ⊂ {y : d(y, x) = r} and A ⊂
{y : d(y, x) ≤ r}. Equality need not hold in the last three cases.

Proof: We begin with the counterexample to equality. Let X = {0, 1} with

the metric d(0, 1) = 1, d(0, 0) = d(1, 1) = 0. Let A = B1 (0) = {0}. Then
(check) int A = A, ext A = {1}, ∂A = ∅ and A = A.

(int A = A) Since int A ⊂ A, we need to show every point in A is an interior

point. But if y ∈ A, then d(y, x) = s(say) < r and Br−s (y) ⊂ A by
the triangle inequality2 ,

(extA ⊃ {y : d(y, x) > r}) If d(y, x) > r, let d(y, x) = s. Then Bs−r (y) ⊂ Ac
by the triangle inequality (exercise), i.e. y is an exterior point of A.

(ext A = {y : d(y, x) > r} in Rn ) We have ext A ⊃ {y : d(y, x) > r} from

the previous result. If d(y, x) ≤ r then every Bs (y), where s > 0,
contains points in A3 . Hence y 6∈ ext A. The result follows.

(∂A ⊂ {y : d(y, x) = r}, with equality for Rn ) This follows from the previ-
ous results and the fact that ∂A = X \ ((int A) ∪ ext A).

(A ⊂ {y : d(y, x) ≤ r}, with equality for Rn ) This follows from A = A ∪ ∂A

and the previous results.

If z ∈ Bs (y) then d(z, y) < r − s. But d(y, x) = s and so d(z, x) ≤ d(z, y) + d(y, x) <
2

(r − s) + s = r, i.e. d(z, x) < r and so z ∈ Br (x) as required. Draw a diagram in R2 .

3
Why is this true in Rn ? It is not true in an arbitrary metric space, as we see from the
counterexample.
Metric Spaces 69

Example If Q is the set of rationals in R, then int Q = ∅, ∂Q = R, Q = R

and ext Q = ∅ (exercise).

6.4 Open and Closed Sets

Everything in this section apart from specific examples, applies with Rn re-
placed by an arbitrary metric space (X, d).
The concept of an open set is very important in Rn and more generally
is basic to the study of topology4 . We will see later that notions such as
connectedness of a set and continuity of a function can be expressed in terms
of open sets.

Definition 6.4.1 A set A ⊂ Rn is open iff A ⊂ intA.

Remark Thus a set is open iff all its members are interior points. Note
that since always intA ⊂ A, it follows that

A is open iff A = intA.

We usually show a set A is open by proving that for every x ∈ A there

exists r > 0 such that Br (x) ⊂ A (which is the same as showing that every
x ∈ A is an interior point of A). Of course, the value of r will depend on x
in general.
Note that ∅ and Rn are both open sets (for a set to be open, every member
must be an interior point—since the ∅ has no members it is trivially true that
every member of ∅ is an interior point!). Proposition 6.3.7 shows that Br (x)
is open, thus justifying the terminology of open ball.
The following result gives many examples of open sets.

Theorem 6.4.2 If A ⊂ Rn then int A is open, as is ext A.

Proof:

4
There will be courses on elementary topology and algebraic topology in later years.
Topological notions are important in much of contemporary mathematics.
70

Br-s(y)

Br(x)
A y..
.
x
r

Let A ⊂ Rn and consider any x ∈ int A. Then Br (x) ⊂ A for some

r > 0. We claim that Br (x) ⊂ int A, thus proving int A is open.
To establish the claim consider any y ∈ Br (x); we need to show that
y is an interior point of A. Suppose d(y, x) = s (< r). From the triangle
inequality (exercise), Br−s (y) ⊂ Br (x), and so Br−s (y) ⊂ A, thus showing y
is an interior point.
The fact ext A is open now follows from (6.3).

Exercise. Suppose A ⊂ Rn . Prove that the interior of A with respect to the

Euclidean metric, and with respect to the sup metric, are the same. Hint:
First show that each open ball about x ∈ Rn with respect to the Euclidean
metric contains an open ball with respect to the sup metric, and conversely.
Deduce that the open sets corresponding to either metric are the same.

The next result shows that finite intersections and arbitrary unions of
open sets are open. It is not true that an arbitrary intersection of open sets
is open. For example, the intervals (−1/n, 1/n) are open for each positive
T
integer n, but ∞ n=1 (−1/n, 1/n) = {0} which is not open.

Theorem 6.4.3 If A1 , . . . , Ak are finitely many open sets then A1 ∩ · · · ∩ Ak

S
is also open. If {Aλ }λ∈S is a collection of open sets, then λ∈S Aλ is also
open.

Proof: Let A = A1 ∩ · · · ∩ Ak and suppose x ∈ A. Then x ∈ Ai for

i = 1, . . . , k, and for each i there exists ri > 0 such that Bri (x) ⊂ Ai . Let
r = min{r1 , . . . , rn }. Then r > 0 and Br (x) ⊂ A, implying A is open.
S
Next let B = λ∈S Aλ and suppose x ∈ B. Then x ∈ Aλ for some λ. For
some such λ choose r > 0 such that Br (x) ⊂ Aλ . Then certainly Br (x) ⊂ B,
and so B is open.

We next define the notion of a closed set.

Metric Spaces 71

Definition 6.4.4 A set A ⊂ Rn is closed iff its complement is open.

Proposition 6.4.5 A set is open iff its complement is closed.

Proof: Exercise.

We saw before that a set is open iff it is contained in, and hence equals,
its interior. Analogously we have the following result.

Theorem 6.4.6 A set A is closed iff A = A.

Proof: A is closed iff Ac is open iff Ac = int (Ac ) iff Ac = ext A (from (6.3))
iff A = A (taking complements and using (6.5)).

Remark Since A ⊂ A it follows from the previous theorem that A is closed

iff A ⊂ A, i.e. iff A contains all its limit points.

The following result gives many examples of closed sets, analogous to

Theorem (6.4.2).

Theorem 6.4.7 The sets A and ∂A are closed.

Proof: Since A = (ext A)c it follows A is closed, ∂A = (int A ∪ ext A)c , so

that ∂A is closed.

Examples We saw in Proposition 6.3.7 that the set C = {y : |y − x| ≤ r}

in Rn is the closure of Br (x) = {y : |y − x| < r}, and hence is closed. We
also saw that in an arbitrary metric space we only know that Br (x) ⊆ C.
But it is always true that C is closed.
To see this, note that the complement of C is {y : d(y, x) > r}. This is
open since if y is a member of the complement and d(x, y) = s (> r), then
Bs−r (y) ⊂ C c by the triangle inequality (exercise).
Similarly, {y : d(y, x) = r} is always closed; it contains but need not equal
∂Br (x).
In particular, the interval [a, b] is closed in R.
Also ∅ and Rn are both closed, showing that a set can be both open and
closed (these are the only such examples in Rn , why?).

Remark “Most” sets are neither open nor closed. In particular, Q and
(a, b] are neither open nor closed in R.

An analogue of Theorem 6.4.3 holds:

72

Theorem 6.4.8 If A1 , . . . , An are closed sets then A1 ∪· · ·∪An is also closed.

T
If Aλ (λ ∈ S) is a collection of closed sets, then λ∈S Aλ is also closed.

Proof: This follows from the previous theorem by DeMorgan’s rules. More
precisely, if A = A1 ∪ · · · ∪ An then Ac = Ac1 ∩ · · · ∩ Acn and so Ac is open
and hence A is closed. A similar proof applies in the case of arbitrary inter-
sections.

S
Remark The example (0, 1) = ∞ n=1 [1/n, 1 − 1/n] shows that a non-finite
union of closed sets need not be closed.

In R we have the following description of open sets. A similar result is not

true in Rn for n > 1 (with intervals replaced by open balls or open n-cubes).

S
Theorem 6.4.9 A set U ⊂ R is open iff U = i≥1 Ii , where {Ii } is a
countable (finite or denumerable) family of disjoint open intervals.

Proof: *Suppose U is open in R. Let a ∈ U . Since U is open, there

exists an open interval I with a ∈ I ⊂ U . Let Ia be the union of all such
open intervals. Since the union of a family of open intervals with a point
in common is itself an open interval (exercise), it follows that Ia is an open
interval. Clearly Ia ⊂ U .
We next claim that any two such intervals Ia and Ib with a, b ∈ U are
either disjoint or equal. For if they have some element in common, then
Ia ∪ Ib is itself an open interval which is a subset of U and which contains
both a and b, and so Ia ∪ Ib ⊂ Ia and Ia ∪ Ib ⊂ Ib . Thus Ia = Ib .
Thus U is a union of a family F of disjoint open intervals. To see that F
is countable, for each I ∈ F select a rational number in I (this is possible, as
there is a rational number between any two real numbers, but does it require
the axiom of choice?). Different intervals correspond to different rational
numbers, and so the set of intervals in F is in one-one correspondence with
a subset of the rationals. Thus F is countable.

*A Surprising Result Suppose ² is a small positive number (e.g. 10−23 ).

S
Then there exist disjoint open intervals I1 , I2 , . . . such that Q ⊂ ∞i=1 Ii and
P
such that ∞ |I
i=1 i | ≤ ² (where |I i | is the length of I i )!
To see this, let r1 , r2 , . . . be an enumeration of the rationals. About each
S P
ri choose an interval Ji of length ²/2i . Then Q ⊂ i≥1 Ji and i≥1 |Ji | = ².
However, the Ji are not necessarily mutually disjoint.
We say two intervals Ji and Jj are “connectable” if there is a sequence
of intervals Ji1 , . . . , Jin such that i1 = i, in = j and any two consecutive
intervals Jip , Jip+1 have non-zero intersection.
Define I1 to be the union of all intervals connectable to J1 .
Next take the first interval Ji after J1 which is not connectable to J1 and
Metric Spaces 73

define I2 to be the union of all intervals connectable to this Ji .

Next take the first interval Jk after Ji which is not connectable to J1 or Ji
and define I3 to be the union of all intervals connectable to this Jk .
And so on.
Then one can show that the Ii are mutually disjoint intervals and that
P∞ P∞
j=1 |Ii | ≤ i=1 |Ji | = ².

6.5 Metric Subspaces

Definition 6.5.1 Suppose (X, d) is a metric space and S ⊂ X. Then the
metric subspace corresponding to S is the metric space (S, dS ), where
dS (x, y) = d(x, y). (6.8)
The metric dS (often just denoted d) is called the induced metric on S 5 .

It is easy (exercise) to see that the axioms for a metric space do indeed
hold for (S, dS ).

Examples

1. The sets [a, b], (a, b] and Q all define metric subspaces of R.
2. Consider R2 with the usual Euclidean metric. We can identify R with
the “x-axis” in R2 , more precisely with the subset {(x, 0) : x ∈ R}, via
the map x 7→ (x, 0). The Euclidean metric on R then corresponds to
the induced metric on the x-axis.

Since a metric subspace (S, dS ) is a metric space, the definitions of open

ball; of interior, exterior, boundary and closure of a set; and of open set and
closed set; all apply to (S, dS ).
There is a simple relationship between an open ball about a point in a
metric subspace and the corresponding open ball in the original metric space.

Proposition 6.5.2 Suppose (X, d) is a metric space and (S, d) is a metric

subspace. Let a ∈ S. Let the open ball in S of radius r about a be denoted by
BrS (a). Then
BrS (a) = S ∩ Br (a).

Proof:
BrS (a) := {x ∈ S : dS (x, a) < r} = {x ∈ S : d(x, a) < r}
= S ∩ {x ∈ X : d(x, a) < r} = S ∩ Br (a).

5
There is no connection between the notions of a metric subspace and that of a vector
subspace! For example, every subset of Rn defines a metric subspace, but this is certainly
not true for vector subspaces.
74

The symbol “:=” means “by definition, is equal to”.

There is also a simple relationship between the open (closed) sets in a
metric subspace and the open (closed) sets in the original space.

Theorem 6.5.3 Suppose (X, d) is a metric space and (S, d) is a metric sub-
space. Then for any A ⊂ S:

1. A is open in S iff A = S ∩ U for some set U (⊂ X) which is open in X.

2. A is closed in S iff A = S ∩ C for some set C (⊂ X) which is closed

in X.

Proof: (i) Suppose that A = S ∩ U , where U (⊂ X) is open in X. Then

for each a ∈ A (since a ∈ U and U is open in X) there exists r > 0 such that
Br (a) ⊂ U . Hence S ∩ Br (a) ⊂ S ∩ U , i.e. BrS (a) ⊂ A as required.
U
BSr(a) = S∩Br(a)
Br(a)

.
S A = S∩U a

(ii) Next suppose A is open in S. Then for each a ∈ A there exists

S
r = ra > 06 such that BrSa (a) ⊂ A, i.e. S ∩ Bra (a) ⊂ A. Let U = a∈A Bra (a).
Then U is open in X, being a union of open sets.
We claim that A = S ∩ U . Now
[ [ [
S∩U =S∩ Bra (a) = (S ∩ Bra (a)) = BrSa (a).
a∈A a∈A a∈A

But BrSa (a) ⊂ A, and for each a ∈ A we trivially have that a ∈ BrSa (a). Hence
S ∩ U = A as required.

The result for closed sets follow from the results for open sets together
with DeMorgan’s rules.
(iii) First suppose A = S ∩ C, where C (⊂ X) is closed in X. Then
S \ A = S ∩ C c from elementary properties of sets. Since C c is open in X, it
follows from (1) that S \ A is open in S, and so A is closed in S.
6
We use the notation r = ra to indicate that r depends on a.
Metric Spaces 75

(iv) Finally suppose A is closed in S. Then S \ A is open in S, and so

from (1), S \ A = S ∩ U where U ⊂ X is open in X. From elementary
properties of sets it follows that A = S ∩ U c . But U c is closed in X, and so
the required result follows.

Examples

1. Let S = (0, 2]. Then (0, 1) and (1, 2] are both open in S (why?), but
(1, 2] is not open in R. Similarly, (0, 1] and [1, 2] are both closed in S
(why?), but (0, 1] is not closed in R.

2. Consider R as a subset of R2 by identifying x ∈ R with (x, 0) ∈ R2 .

Then R is open and closed as a subset of itself, but is closed (and not
open) as a subset of R2 .
√ √ √ √
3. Note that [− 2, 2]∩Q = (− 2, 2)∩Q. It follows that Q has many
clopen sets.
76
Chapter 7

Sequences and Convergence

In this chapter you should initially think of the cases X = R and X = Rn .

7.1 Notation
If X is a set and xn ∈ X for n = 1, 2, . . ., then (x1 , x2 , . . .) is called a sequence
in X and xn is called the nth term of the sequence. We also write x1 , x2 , . . .,
or (xn )∞
n=1 , or just (xn ), for the sequence.

NB Note the difference between (xn ) and {xn }.

More precisely, a sequence in X is a function f : N → X, where f (n) = xn

with xn as in the previous notation.
We write (xn )∞ n=1 ⊂ X or (xn ) ⊂ X to indicate that all terms of the
sequence are members of X. Sometimes it is convenient to write a sequence
in the form (xp , xp+1 , . . .) for some (possible negative) integer p 6= 1.
Given a sequence (xn ), a subsequence is a sequence (xni ) where (ni ) is a
strictly increasing sequence in N.

7.2 Convergence of Sequences

Definition 7.2.1 Suppose (X, d) is a metric space, (xn ) ⊂ X and x ∈ X.
Then we say the sequence (xn ) converges to x, written xn → x, if for every
r > 0 1 there exists an integer N such that

n ≥ N ⇒ d(xn , x) < r. (7.1)

Thus xn → x if for every open ball Br (x) centred at x the sequence (xn )
is eventually contained in Br (x). The “smaller” the ball, i.e. the smaller the
1
It is sometimes convenient to replace r by ², to remind us that we are interested in
small values of r (or ²).

77
78

value of r, the larger the value of N required for (7.1) to be true, as we

see in the following diagram for three different balls centred at x. Although
for each r > 0 there will be a least value of N such that (7.1) is true, this
particular value of N is rarely of any significance.

Remark The notion of convergence in a metric space can be reduced to

the notion of convergence in R, since the above definition says xn → x iff
d(xn , x) → 0, and the latter is just convergence of a sequence of real numbers.

Examples

1. Let θ ∈ R be fixed, and set

µ ¶
1 1
xn = a + cos nθ, b + sin nθ ∈ R2 .
n n
Then xn → (a, b) as n → ∞. The sequence (xn ) “spirals” around the
point (a, b), with d (xn , (a, b)) = 1/n, and with a rotation by the angle
θ in passing from xn to xn+1 .
2. Let
(x1 , x2 , . . .) = (1, 1, . . .).
Then xn → 1 as n → ∞.
3. Let
1 1 1
(x1 , x2 , . . .) = (1, , 1, , 1, , . . .).
2 3 4
Then it is not the case that xn → 0 and it is not the case that xn → 1.
The sequence (xn ) does not converge.
4. Let A ⊂ R be bounded above and suppose a = lub A. Then there
exists (xn ) ⊂ A such that xn → a.

Proof: Suppose n is a natural number. By the definition of least

upper bound, a − 1/n is not an upper bound for A. Thus there exists
an x ∈ A such that a − 1/n < x ≤ a. Choose some such x and denote
it by xn . Then xn → a since d(xn , a) < 1/n 2 .
2
Note the implicit use of the axiom of choice to form the sequence (xn ).
Sequences and Convergence 79

5. As an indication of “strange” behaviour, for the p-adic metric on Z we

have pn → 0.

P
Series An infinite series ∞ n=1 xn of terms from R (more generally, from R
n

or from a normed space) is just a certain type of sequence. More precisely,

for each i we define the nth partial sum by

sn = x1 + · · · + xn .
P
Then we say the series ∞ n=1 xn converges iff the sequence (of partial sums)
(sn ) converges, and in this case the limit of (sn ) is called the sum of the
series.

NB Note that changing the order (re-indexing) of the (xn ) gives rise to a
possibly different sequence of partial sums (sn ).

Example
P∞
If 0 < r < 1 then the geometric series n=0 rn converges to (1 − r)−1 .

7.3 Elementary Properties

Theorem 7.3.1 A sequence in a metric space can have at most one limit.

Proof: Suppose (X, d) is a metric space, (xn ) ⊂ X, x, y ∈ X, xn → x as

n → ∞, and xn → y as n → ∞.
Supposing x 6= y, let d(x, y) = r > 0. From the definition of convergence
there exist integers N1 and N2 such that

n ≥ N1 ⇒ d(xn , x) < r/4,

n ≥ N2 ⇒ d(xn , y) < r/4.

Let N = max{N1 , N2 }. Then

d(x, y) ≤ d(x, xN ) + d(xN , y)

< r/4 + r/4 = r/2,

i.e. d(x, y) < r/2, which is a contradiction.

Definition 7.3.2 A sequence is bounded if the set of terms from the sequence
is bounded.

Theorem 7.3.3 A convergent sequence in a metric space is bounded.

80

Proof: Suppose that (X, d) is a metric space, (xn )∞

n=1 ⊂ X, x ∈ X and
xn → x. Let N be an integer such that n ≥ N implies d(xn , x) ≤ 1.
Let r = max{d(x1 , x), . . . , d(xN −1 , x), 1} (this is finite since r is the max-
imum of a finite set of numbers). Then d(xn , x) ≤ r for all n and so
(xn )∞
n=1 ⊂ Br+1/10 (x). Thus (xn ) is bounded, as required.

Remark This method, of using convergence to handle the ‘tail’ of the

sequence, and some separate argument for the finitely many terms not in the
tail, is of fundamental importance.

The following result on the distance function is useful. As we will see in

Chapter 11, it says that the distance function is continuous.

Theorem 7.3.4 Let xn → x and yn → y in a metric space (X, d). Then

d(xn , yn ) → d(x, y).

Proof: Two applications of the triangle inequality show that

d(x, y) ≤ d(x, xn ) + d(xn , yn ) + d(yn , y),

and so
d(x, y) − d(xn , yn ) ≤ d(x, xn ) + d(yn , y). (7.2)
Similarly
d(xn , yn ) ≤ d(xn , x) + d(x, y) + d(y, yn ),
and so
d(xn , yn ) − d(x, y) ≤ d(x, xn ) + d(yn , y). (7.3)
It follows from (7.2) and (7.3) that

|d(x, y) − d(xn , yn )| ≤ d(x, xn ) + d(yn , y).

Since d(x, xn ) → 0 and d(y, yn ) → 0, the result follows immediately from

properties of sequences of real numbers (or see the Comparison Test in the
next section).

7.4 Sequences in R
The results in this section are particular to sequences in R. They do not even
make sense in a general metric space.

Definition 7.4.1 A sequence (xn )∞

n=1 ⊂ R is

1. increasing (or non-decreasing) if xn ≤ xn+1 for all n,

Sequences and Convergence 81

2. decreasing (or non-increasing) if xn+1 ≥ xn for all n,

3. strictly increasing if xn < xn+1 for all n,

4. strictly decreasing if xn+1 < xn for all n.

A sequence is monotone if it is either increasing or decreasing.

The following theorem uses the Completeness Axiom in an essential way.

It is not true if we replace R by Q. For example, consider the√sequence of
rational numbers obtained by taking
√ the decimal expansion of 2; i.e. xn is
the decimal approximation to 2 to n decimal places.

Theorem 7.4.2 Every bounded monotone sequence in R has a limit in R.

Proof: Suppose (xn )∞ n=1 ⊂ R and (xn ) is increasing (if (xn ) is decreasing,
the argument is analogous). Since the set of terms {x1 , x2 , . . .} is bounded
above, it has a least upper bound x, say. We claim that xn → x as n → ∞.
To see this, note that xn ≤ x for all n; but if ² > 0 then xk > x − ² for
some k, as otherwise x − ² would be an upper bound. Choose such k = k(²).
Since xk > x − ², then xn > x − ² for all n ≥ k as the sequence is increasing.
Hence
x − ² < xn ≤ x
for all n ≥ k. Thus |x − xn | < ² for n ≥ k, and so xn → x (since ² > 0 is
arbitrary).

It follows that a bounded closed set in R contains its infimum and supre-
mum, which are thus the minimum and maximum respectively.
For sequences (xn ) ⊂ R it is also convenient to define the notions xn → ∞
and xn → −∞ as n → ∞.

Definition 7.4.3 If (xn ) ⊂ R then xn → ∞ (−∞) as n → ∞ if for every

positive real number M there exists an integer N such that

n ≥ N implies xn > M (xn < −M ).

We say (xn ) has limit ∞ (−∞) and write limn→∞ xn = ∞(−∞).

Note When we say a sequence (xn ) converges, we usually mean that xn → x

for some x ∈ R; i.e. we do not allow xn → ∞ or xn → −∞.

The following Comparison Test is easy to prove (exercise). Notice that

the assumptions xn < yn for all n, xn → x and yn → y, do not imply x < y.
For example, let xn = 0 and yn = 1/n for all n.
82

Theorem 7.4.4 (Comparison Test)

1. If 0 ≤ xn ≤ yn for all n ≥ N , and yn → 0 as n → ∞, then xn → 0 as

n → ∞.

2. If xn ≤ yn for all n ≥ N , xn → x as n → ∞ and yn → y as n → ∞,

then x ≤ y.

3. In particular, if xn ≤ a for all n ≥ N and xn → x as n → ∞, then

x ≤ a.

Example Let xm = (1 + 1/m)m and ym = 1 + 1 + 1/2! + · · · + 1/m!

The sequence (ym ) is increasing and for each m ∈ N,
1 1 1
ym ≤ 1 + 1 + + 2 + · · · m < 3.
2 2 2
Thus ym → y0 (say) ≤ 3, from Theorem 7.4.2.
From the binomial theorem,

1 m(m − 1) 1 m(m − 1)(m − 2) 1 m! 1

xm = 1 + m + 2
+ 3
+ ··· + .
m 2! m 3! m m! mm
This can be written as
µ ¶ µ ¶µ ¶
1 1 1 1 2
xm = 1+1+ 1− + 1− 1−
2! m 3! m m
µ ¶µ ¶ µ ¶
1 1 2 m−1
+··· + 1− 1− ··· 1 − .
m! m m m
It follows that xm ≤ xm+1 , since there is one extra term in the expression
for xm+1 and the other terms (after the first two) are larger for xm+1 than
for xm . Clearly xm ≤ ym (≤ y0 ≤ 3). Thus the sequence (xm ) has a limit x0
(say) by Theorem 7.4.2. Moreover, x0 ≤ y0 from the Comparison test.
In fact x0 = y0 and is usually denoted by e(= 2.71828...)3 . It is the base
of the natural logarithms.

P
Example Let zn = nk=1 k1 − log n. Then (zn ) is monotonically decreasing
and 0 < zn < 1 for all n. Thus (zn ) has a limit γ say. This is Euler’s constant,
and γ = 0.577 . . .. It arises when considering the Gamma function:
Z∞ ∞
eγz Y 1
Γ(z) = e−t tz−1 dt = (1 + )−1 ez/n
z 1 n
0

For n ∈ N, Γ(n) = (n − 1)!.

3
See the Problems.
Sequences and Convergence 83

7.5 Sequences and Components in Rk

The result in this section is particular to sequences in Rn and does not apply
(or even make sense) in a general metric space.

Theorem 7.5.1 A sequence (xn )∞ n

n=1 in R converges iff the corresponding
sequences of components (xin ) converge for i = 1, . . . , k. Moreover,
lim x1n , . . . , n→∞
lim xn = (n→∞
n→∞
lim xkn ).

Proof: Suppose (xn ) ⊂ Rn and xn → x. Then |xn − x| → 0, and since

|xin − xi | ≤ |xn − x| it follows from the Comparison Test that xin → xi as
n → ∞, for i = 1, . . . , k.
Conversely, suppose that xin → xi for i = 1, . . . , k. Then for any ² > 0
there exist integers N1 , . . . , Nk such that
n ≥ Ni ⇒ |xin − xi | < ²
for i = 1, . . . , k.
Since
à k ! 12
X
|xn − x| = |xin − xi |2 ,
i=1

it follows that if N = max{N1 , . . . , Nk } then

√ √
n ≥ N ⇒ |xn − x| < k²2 = k².

Since ² > 0 is otherwise arbitrary4 , the result is proved.

7.6 Sequences and the Closure of a Set

The following gives a useful characterisation of the closure of a set in terms
of convergent sequences.

Theorem 7.6.1 Let X be a metric space and let A ⊂ X. Then x ∈ A iff

there is a sequence (xn )∞
n=1 ⊂ A such that xn → x.

Proof: If (xn ) ⊂ A and xn → x, then for every r > 0, Br (x) must contain
some term from the sequence. Thus x ∈ A from Definition (6.3.3) of a limit
point.
Conversely, if x ∈ A then (again from Definition (6.3.3)), B1/n (x) ∩ A 6= ∅
for each n ∈ N . Choose xn ∈ B1/n (x)∩A for n = 1, 2, . . .. Then (xn )∞
n=1 ⊂ A.
Since d(xn , x) ≤ 1/n it follows xn → x as n → ∞.
4
More precisely, to be consistent
√ with the definition
√ of convergence, we could replace ²
throughout the proof by ²/ k and so replace ² k on the last line of the proof by ². We
would not normally bother doing this.
84

Corollary 7.6.2 Let X be a metric space and let A ⊂ X. Then A is closed

in X iff
(xn )∞
n=1 ⊂ A and xn → x implies xn ∈ A. (7.4)

Proof: From the theorem, (7.4) is true iff A = A, i.e. iff A is closed.

Remark Thus in a metric space X the closure of a set A equals the set of
all limits of sequences whose members come from A. And this is generally
not the same as the set of limit points of A (which points will be missed?).
The set A is closed iff it is “closed” under the operation of taking limits of
convergent sequences of elements from A5 .

Exercise Let A = {( m , n ) : m, n ∈ N}. Determine A.

n 1

Exercise Use Corollary 7.6.2 to show directly that the closure of a set is
indeed closed.

7.7 Algebraic Properties of Limits

The important cases in this section are X = R, X = Rn and X is a function
space such as C[a, b]. The proofs are essentially the same as in the case
X = R. We need X to be a normed space (or an inner product space for the
third result in the next theorem) so that the algebraic operations make sense.

Theorem 7.7.1 Let (xn )∞ ∞

n=1 and (yn )n=1 be convergent sequences in a normed
space X, and let α be a scalar. Let limn→∞ xn = x and limn→∞ yn = y. Then
the following limits exist with the values stated:

lim (xn + yn ) = x + y,
n→∞
(7.5)
lim αxn = αx.
n→∞
(7.6)

More generally, if also αn → α, then

lim αn xn = αx.
n→∞
(7.7)

If X is an inner product space then also

lim xn · yn = x · y. (7.8)
n→∞

5
There is a possible inconsistency of terminology here. The sequence (1, 1 + 1, 1/2, 1 +
1/2, 1/3, 1+1/3, . . . , 1/n, 1+1/n, . . .) has no limit; the set A = {1, 1+1, 1/2, 1+1/2, 1/3, 1+
1/3, . . . , 1/n, 1 + 1/n, . . .} has the two limit points 0 and 1; and the closure of A, i.e. the
set of limit points of sequences whose members belong to A, is A ∪ {0, 1}.
Exercise: What is a sequence with members from A converging to 0, to 1, to 1/3?
Sequences and Convergence 85

Proof: Suppose ² > 0. Choose N1 , N2 so that

||xn − x|| < ² if n ≥ N1 ,

||yn − y|| < ² if n ≥ N2 .

Let N = max{N1 , N2 }.
(i) If n ≥ N then

||(xn + yn ) − (x + y)|| = ||(xn − x) + (yn − y)||

≤ ||xn − x|| + ||yn − y||
< 2².

Since ² > 0 is arbitrary, the first result is proved.

(ii) If n ≥ N1 then

||αxn − αx|| = |α| ||xn − x||

≤ |α|².

Since ² > 0 is arbitrary, this proves the second result.

(iii) For the fourth claim, we have for n ≥ N

|xn · yn − x · y| = |(xn − x) · yn + x · (yn − y)|

≤ |(xn − x) · yn | + |x · (yn − y)|
≤ |xn − x| |yn | + |x| |yn − y| by Cauchy-Schwarz
≤ ²|yn | + |x|²,

Since (yn ) is convergent, it follows from Theorem 7.3.3 that |yn | ≤ M1 (say)
for all n. Setting M = max{M1 , |x|} it follows that

|xn · yn − x · y| ≤ 2M ²

for n ≥ N . Again, since ² > 0 is arbitrary, we are done.

(iv) The third claim is proved like the fourth (exercise).
86
Chapter 8

Cauchy Sequences

8.1 Cauchy Sequences

Our definition of convergence of a sequence (xn )∞ n=1 refers not only to the
sequence but also to the limit. But it is often not possible to know the limit
a priori, (see example in Section 7.4). We would like, if possible, a criterion
for convergence which does not depend on the limit itself. We have already
seen that a bounded monotone sequence in R converges, but this is a very
special case.
Theorem 8.1.3 below gives a necessary and sufficient criterion for con-
vergence of a sequence in Rn , due to Cauchy (1789–1857), which does not
refer to the actual limit. We discuss the generalisation to sequences in other
metric spaces in the next section.

Definition 8.1.1 Let (xn )∞n=1 ⊂ X where (X, d) is a metric space. Then
(xn ) is a Cauchy sequence if for every ² > 0 there exists an integer N such
that
m, n ≥ N ⇒ d(xm , xn ) < ².
We sometimes write this as d(xm , xn ) → 0 as m, n → ∞.

Thus a sequence is Cauchy if, for each ² > 0, beyond a certain point in
the sequence all the terms are within distance ² of one another.

Warning This is stronger than claiming that, beyond a certain point in the
sequence, consecutive terms are within distance ² of one another.
√
For example, consider the sequence xn = n. Then
³√
√ √
√ ´ n+1+ n
|xn+1 − xn | = n+1− n √ √
n+1+ n
1
= √ √ .
n+1+ n
(8.1)

87
88

Hence |xn+1 − xn | → 0 as n → ∞.
√ √
But |xm − xn | = m − n if m > n, and so for any N we can choose
n = N and m > N such that |xm − xn | is as large as we wish. Thus the
sequence (xn ) is not Cauchy.

Theorem 8.1.2 In a metric space, every convergent sequence is a Cauchy

sequence.

Proof: Let (xn ) be a convergent sequence in the metric space (X, d), and
suppose x = lim xn .
Given ² > 0, choose N such that

n ≥ N ⇒ d(xn , x) < ².

It follows that for any m, n ≥ N

d(xm , xn ) ≤ d(xm , x) + d(x, xn )

≤ 2².

Thus (xn ) is Cauchy.

Remark The converse of the theorem is true in Rn , as we see in the next

theorem, but is not true in general. For example, a Cauchy sequence from
Q (with the usual metric) will not necessarily converge to a limit in Q (take
the usual example√of the sequence whose nth term is the n place decimal
approximation to 2). We discuss this further in the next section.

Cauchy implies Bounded A Cauchy sequence in any metric space is

bounded. This simple result is proved in a similar manner to the corre-
sponding result for convergent sequences (exercise).

Theorem 8.1.3 (Cauchy) A sequence in Rn converges (to a limit in Rn )

iff it is Cauchy.

Proof: We have already seen that a convergent sequence in any metric

space is Cauchy.
Assume for the converse that the sequence (xn )∞
n=1 ⊂ R is Cauchy.
n

The Case k = 1: We will show that if (xn ) ⊂ R is Cauchy then it is

convergent by showing that it converges to the same limit as an associated
monotone increasing sequence (yn ).
Define
yn = inf{xn , xn+1 , . . .}
for each n ∈ N 1 . It follows that yn+1 ≥ yn since yn+1 is the infimum over a
subset of the set corresponding to yn . Moreover the sequence (yn ) is bounded
1
You may find it useful to think of an example such as xn = (−1)n /n.
Cauchy Sequences and Complete Metric Spaces 89

since the sequence (xn ) is Cauchy and hence bounded (if |xn | ≤ M for all n
then also |yn | ≤ M for all n).
From Theorem 7.4.2 on monotone sequences, yn → a (say) as n → ∞.
We will prove that also xn → a.
2
Suppose ² > 0. Since (xn ) is Cauchy there exists N = N (²) such that

xn − ² ≤ xm ≤ x` + ² (8.2)

for all `, m, n ≥ N .
Claim:
yn − ² ≤ xm ≤ yn + ² (8.3)
for all m, n ≥ N .
To establish the first inequality in (8.3), note that from the first inequality
in (8.2) we immediately have for m, n ≥ N that

y n − ² ≤ xm ,

since yn ≤ xn .
To establish the second inequality in (8.3), note that from the second
inequality in (8.2) we have for `, m ≥ N that

xm ≤ x` + ² .

But yn + ² = inf{x` + ² : ` ≥ n} as is easily seen3 . Thus yn + ² is greater or

equal to any other lower bound for {xn + ², xn+1 + ², . . .}, whence

xm ≤ yn + ².

It now follows from the Claim, by fixing m and letting n → ∞, and from
the Comparison Test (Theorem 7.4.4) that

a − ² ≤ xm ≤ a + ²

for all m ≥ N = N (²).

Since ² > 0 is arbitrary, it follows that xm → a as m → ∞. This finishes
the proof in the case k = 1.
The Case k > 1: If (xn )∞n=1 ⊂ R is Cauchy it follows easily that each
n

sequence of components is also Cauchy, since

|xin − yin | ≤ |xn − yn |

for i = 1, . . . , k. From the case k = 1 it follows xin → ai (say) for i = 1, . . . , k.

Then from Theorem 7.5.1 it follows xn → a where a = (a1 , . . . , an ).
2
The notation N (²) is just a way of noting that N depends on ².
3
If y = inf S, then y + α = inf{x + α : x ∈ S}.
90

Remark In the case k = 1, and for any bounded sequence, the number a
constructed above,
a = sup inf{xi : i ≥ n}
n

is denoted lim inf xn or lim xn . It is the least of the limits of the subse-
quences of (xn )∞ 4
n=1 (why?). One can analogously define lim sup xn or lim xn
(exercise).

8.2 Complete Metric Spaces

Definition 8.2.1 A metric space (X, d) is complete if every Cauchy sequence
in X has a limit in X.
If a normed space is complete with respect to the associated metric, it is
called a complete normed space or a Banach space.

We have seen that Rn is complete, but that Q is not complete. The

next theorem gives a simple criterion for a subset of Rn (with the standard
metric) to be complete.

Examples

1. We will see in Corollary 12.3.5 that C[a, b] (see Section 5.1 for the
definition) is a Banach space with respect to the sup metric. The same
argument works for `∞ (N).
On the other hand, C[a, b] with respect to the L1 norm (see 5.11) is not
complete. For example, let fn ∈ C[−1, 1] be defined by


 0 −1 ≤ x ≤ 0
fn (x) = nx 0 ≤ x ≤ n1


1 1
n
≤x≤1

Then there is no f ∈ C[−1, 1] such that ||fn − f ||L1 → 0, i.e. such that
R1
−1 |fn − f | → 0. (If there were such an f , then we would have to have
f (x) = 0 if −1 ≤ x < 0 and f (x) = 1 if 0 < x ≤ 1 (why?). But such
an f cannot be continuous on [−1, 1].)
The same example shows that C[a, b] with respect to the L2 norm
(see 5.12) is not complete.

2. Take X = R with metric

¯ ¯
¯ x y ¯¯
¯
d(x, y) = ¯ − ¯.
¯ 1 + |x| 1 + |y| ¯

(Exercise) show that this is indeed a metric.

4
Note that the limit of a subsequence of (xn ) may not be a limit point of the set {xn }.
Cauchy Sequences and Complete Metric Spaces 91

If (xn ) ⊂ R with |xn − x| → 0, then certainly d(xn , x) → 0. Not so

obviously, the converse is also true. But whereas (R, | · |) is complete,
(X, d) is not, as consideration of the sequence xn = n easily shows.
Define Y = R ∪ {−∞, ∞}, and extend d by setting
¯ ¯
¯ x ¯
¯
d(±∞, x) = ¯ − (±1)¯¯ , d(−∞, ∞) = 2
¯ 1 + |x| ¯

for x ∈ R. Then (Y, d) is complete (exercise).

Remark The second example here utilizes the following simple fact. Given
a set X, a metric space (Y, dY ) and a suitable function f : X → Y , the
function dX (x, x0 ) = dY (f (x), f (x0 )) is a metric on X. (Exercise : what does
suitable mean here?) The cases X = (0, 1), Y = R2 , f (x) = (x, sin x−1 ) and
f (x) = (cos 2πx, sin 2πx) are of interest.

Theorem 8.2.2 If S ⊂ Rn and S has the induced Euclidean metric, then S

is a complete metric space iff S is closed in Rn .

Proof: Assume that S is complete. From Corollary 7.6.2, in order to

show that S is closed in Rn it is sufficient to show that whenever (xn )∞
n=1
is a sequence in S and xn → x ∈ Rn , then x ∈ S. But (xn ) is Cauchy by
Theorem 8.1.2, and so it converges to a limit in S, which must be x by the
uniqueness of limits in a metric space5 .
Assume S is closed in Rn . Let (xn ) be a Cauchy sequence in S. Then
(xn ) is also a Cauchy sequence in Rn and so xn → x for some x ∈ Rn by
Theorem 8.1.3. But x ∈ S from Corollary 7.6.2. Hence any Cauchy sequence
from S has a limit in S, and so S with the Euclidean metric is complete.

Generalisation If S is a closed subset of a complete metric space (X, d),

then S with the induced metric is also a complete metric space. The proof
is the same.

*Remark A metric space (X, d) fails to be complete because there are

Cauchy sequences from X which do not have any limit in X. It is always
possible to enlarge (X, d) to a complete metric space (X ∗ , d∗ ), where X ⊂ X ∗ ,
d is the restriction of d∗ to X, and every element in X ∗ is the limit of a
sequence from X. We call (X ∗ , d∗ ) the completion of (X, d).
For example, the completion of Q is R, and more generally the completion
of any S ⊂ Rn is the closure of S in Rn .
In outline, the proof of the existence of the completion of (X, d) is as
follows6 :
5
To be more precise, let xn → y in S (and so in Rk ) for some y ∈ S. But we also know
that xn → x in Rk ). Thus by uniqeness of limits in the metric space Rk ) , it follows that
x = y.
6
Note how this proof also gives a construction of the reals from the rationals.
92

Let S be the set of all Cauchy sequences from X. We say

two such sequences (xn ) and (yn ) are equivalent if |xn − yn | → 0
as n → ∞. The idea is that the two sequences are “trying” to
converge to the same element. Let X ∗ be the set of all equivalence
classes from S (i.e. elements of X ∗ are sets of Cauchy sequences,
the Cauchy sequences in any element of X ∗ are equivalent to
one another, and any two equivalent Cauchy sequences are in the
same element of X ∗ ).
Each x ∈ X is “identified” with the set of Cauchy sequences
equivalent to the Cauchy sequence (x, x, . . .) (more precisely one
shows this defines a one-one map from X into X ∗ ). The distance
d∗ between two elements (i.e. equivalence classes) of X ∗ is defined
by
d∗ ((xn ), (yn )) = lim |xn − yn |,
n→∞
where (xn ) is a Cauchy sequence from the first eqivalence class
and (yn ) is a Cauchy sequence from the second eqivalence class.
It is straightforward to check that the limit exists, is independent
of the choice of representatives of the equivalence classes, and
agrees with d when restricted to elements of X ∗ which correspond
to elements of X. Similarly one checks that every element of X ∗
is the limit of a sequence “from” X in the appropriate sense.
Finally it is necessary to check that (X ∗ , d∗ ) is complete. So
suppose that we have a Cauchy sequence from X ∗ . Each member
of this sequence is itself equivalent to a Cauchy sequence from
X. Let the nth member xn of the sequence correspond to a
Cauchy sequence (xn1 , xn2 , xn3 , . . .). Let x be the (equivalence
class corresponding to the) diagonal sequence (x11 , x22 , x33 , . . .)
(of course, this sequence must be shown to be Cauchy in X).
Using the fact that (xn ) is a Cauchy sequence (of equivalence
classes of Cauchy sequences), one can check that xn → x (with
respect to d∗ ). Thus (X ∗ , d∗ ) is complete.

It is important to reiterate that the completion depends crucially on the

metric d, see Example 2 above.

8.3 Contraction Mapping Theorem

Let (X, d) be a metric space and let F : A(⊂ X) → X. We say F is a
contraction if there exists λ where 0 ≤ λ < 1 such that
d(F (x), F (y)) ≤ λd(x, y) (8.4)
for all x, y ∈ X.

Remark It is essential that there is a fixed λ, 0 ≤ λ < 1 in (8.4). The

function f (x) = x2 is a contraction on each interval on [0, a], 0 < a < 0.5,
but is not a contraction on [0, 0.5].
Cauchy Sequences and Complete Metric Spaces 93

A simple example of a contraction map on Rn is the map

x 7→ a + r(x − b), (8.5)
where 0 ≤ r < 1 and a, b ∈ Rn . In this case λ = r, as is easily checked.
If b = 0, then (8.5) is just dilation by the factor r followed by translation
by the vector a. More generally, since
a + r(x − b) = b + r(x − b) + a − b,
we see (8.5) is dilation about b by the factor r, followed by translation by
the vector a − b.
We say z is a fixed point of a map F : A(⊂ X) → X if F (z) = z. In the
preceding example, it is easily checked that the unique fixed point for any
r 6= 1 is (a − rb)/(1 − r).
The following result is known as the Contraction Mapping Theorem or as
the Banach Fixed Point Theorem. It has many important applications; we
will use it to show the existence of solutions of differential equations and of
integral equations, the existence of certain types of fractals, and to prove the
Inverse Function Theorem 19.1.1.
You should first follow the proof in the case X = Rn .

Theorem 8.3.1 (Contraction Mapping Theorem) Let (X, d) be a com-

plete metric space and let F : X → X be a contraction map. Then F has a
unique fixed point7 .

Proof: We will find the fixed point as the limit of a Cauchy sequence.
Let x be any point in X and define a sequence (xn )∞
n=1 by

x1 = F (x), x2 = F (x1 ), x3 = F (x2 ), . . . , xn = F (xn−1 ), . . . .

Let λ be the contraction ratio.
1. Claim: (xn ) is Cauchy.
We have
d(xn , xn+1 ) = d(F (xn−1 ), F (xn )
≤ λd(xn−1 , xn )
= λd(F (xn−2 , F (xn−1 )
≤ λ2 d(xn−2 , xn−1 )
..
.
≤ λn−1 d(x1 , x2 ).
Thus if m > n then
d(xm , xn ) ≤ d(xn , xn+1 ) + · · · + d(xm−1 , xm )
≤ (λn−1 + · · · + λm−2 )d(x1 , x2 ).
7
In other words, F has exactly one fixed point.
94

But

λn−1 + · · · + λm−2 ≤ λn−1 (1 + λ + λ2 + · · ·)

1
= λn−1
1−λ
→ 0 as n → ∞.

It follows (why?) that (xn ) is Cauchy.

Since X is complete, (xn ) has a limit in X, which we denote by x.
2. Claim: x is a fixed point of F .
We claim that F (x) = x, i.e. d(x, F (x)) = 0. Indeed, for any n

d(x, F (x)) ≤ d(x, xn ) + d(xn , F (x))

= d(x, xn ) + d(F (xn−1 ), F (x))
≤ d(x, xn ) + λd(xn−1 , x)
→ 0

as n → ∞. This establishes the claim.

3. Claim: The fixed point is unique.
If x and y are fixed points, then F (x) = x and F (y) = y and so

d(x, y) = d(F (x), F (y)) ≤ λd(x, y).

Since 0 ≤ λ < 1 this implies d(x, y) = 0, i.e. x = y.

Remark Fixed point theorems are of great importance for proving existence
results. The one above is perhaps the simplest, but has the advantage of
giving an algorithm for determining the fixed point. In fact, it also gives an
estimate of how close an iterate is to the fixed point (how?).

In applications, the following Corollary is often used.

Corollary 8.3.2 Let S be a closed subset of a complete metric space (X, d)

and let F : S → S be a contraction map on S. Then F has a unique fixed
point in S.

Proof: We saw following Theorem 8.2.2 that S is a complete metric space

with the induced metric. The result now follows from the preceding Theorem.

Example Take R with the standard metric, and let a > 1. Then the map
1 a
f (x) = (x + )
2 x
takes [1, ∞) into itself, and is contractive with λ = 12 . What is the fixed
point? (This was known to the Babylonians, nearly 4000 years ago.)
Cauchy Sequences and Complete Metric Spaces 95

Somewhat more generally, consider Newton’s method for finding a sim-

ple root of the equation f (x) = 0, given some interval containing the root.
Assume that f 00 is bounded on the interval. Newton’s method is an iteration
of the function
f (x)
g(x) = x − 0 .
f (x)
To see why this could possibly work, suppose that there is a root ξ in the
interval [a, b], and that f 0 > 0 on [a, b]. Since

f (x) 00
g 0 (x) = f (x) ,
(f 0 (x)2

we have |g 0 | < 0.5 on some interval [α, β] containing ξ. Since g(ξ) = ξ, it

follows that g is a contraction on [α, β].

Example Consider the problem of solving Ax = b where x, b ∈ Rn and

A ∈ Mn (R). Let λ ∈ R be fixed. Then the task is equivalent to solving
T x = x where
T x = (I − λA)x + λb.
The role of λ comes when one endeavours to ensure T is a contraction under
some norm on Mn (R):

||T x1 − T x2 ||p = |λ|||A(x1 − x2 )||p ≤ k||x1 − x2 ||p

for some 0 < k < 1 by suitable choice of |λ| sufficiently small(depending on

A and p).

Exercise When S ⊆ R is an interval it is often easy to verify that a function

f : S → S is a contraction by use of the mean value theorem. What about
the following function on [0, 1]?
(
sin(x)
x 6= 0 ≤ 0
f (x) = x
1 x=0
96
Chapter 9

Sequences and Compactness

9.1 Subsequences
Recall that if (xn )∞
n=1 is a sequence in some set X and n1 < n2 < n3 < . . .,
then the sequence (xni )∞ ∞
i=1 is called a subsequence of (xn )n=1 .

The following result is easy.

Theorem 9.1.1 If a sequence in a metric space converges then every subse-

quence converges to the same limit as the original sequence.

Proof: Let xn → x in the metric space (X, d) and let (xni ) be a subse-
quence.
Let ² > 0 and choose N so that
d(xn , x) < ²
for n ≥ N . Since ni ≥ i for all i, it follows
d(xni , x) < ²
for i ≥ N .

Another useful fact is the following.

Theorem 9.1.2 If a Cauchy sequence in a metric space has a convergent

subsequence, then the sequence itself converges.

Proof: Suppose that (xn ) is cauchy in the metric space (X, d). So given
² > 0 there is N such that d(xn , xm ) < ² provided m, n > N . If (xnj ) is a
subsequence convergent to x ∈ X, then there is N 0 such that d(xnj , x) < ²
for j > N 0 . Thus for m > N , take any j > max{N, N 0 } (so that nj > N
certainly), to see that
d(x, xm ) ≤ d(x, xnj ) + d(xnj , xm ) < 2² .

97
98

9.2 Existence of Convergent Subsequences

It is often very important to be able to show that a sequence, even though
it may not be convergent, has a convergent subsequence. The following is a
simple criterion for sequences in Rn to have a convergent subsequence. The
same result is not true in an arbitrary metic space, as we see in Remark 2
following the Theorem.
1
Theorem 9.2.1 (Bolzano-Weierstrass) Every bounded sequence in Rn
has a convergent subsequence.

Let us give two proofs of this significant result.

Proof: By the remark following the proof of Theorem 8.1.3, a bounded

sequence in R has a limit point, and necessarily there is a subsequence which
converges to this limit point.
Suppose then, that the result is true in Rn for 1 ≤ k < m, and take a
bounded sequence (xn ) ⊂ Rm . For each n, write xn = (yn , xm
n ) in the obvious
way. Then (yn ) ⊂ Rm−1 is a bounded sequence, and so has a convergent
subsequence (ynj ) by the inductive hypothesis. But then (xmnj ) is a bounded
m
sequence in R, so has a subsequence (xn0 ) which converges. It follows from
j
Theorem 7.5.1 that the subsequence (xnj )of the original sequence (xn ) is
convergent. Thus the result is true for k = m.

Note the “diagonal” argument in the second proof.

Proof: Let (xn )∞ n

n=1 be a bounded sequence of points in R , which for con-
(1) ∞ (1)
venience we rewrite as (xn )n=1 . All terms (xn ) are contained in a closed
cube n o
I1 = y : |y i | ≤ r, i = 1, . . . , k
for some r > 0.

1
Other texts may have different forms of the Bolzano-Weierstrass Theorem.
Subsequences and Compactness 99

Divide I1 into 2k closed subcubes as indicated in the diagram. At least one

of these cubes must contain an infinite number of terms from the sequence
(xn (1) )∞
n=1 . Choose one such cube and call it I2 . Let the corresponding
subsequence in I2 be denoted by (xn (2) )∞n=1 .

Repeating the argument, divide I2 into 2k closed subcubes. Once again,

at least one of these cubes must contain an infinite number of terms from
the sequence (xn (2) )∞
n=1 . Choose one such cube and call it I3 . Let the corre-
sponding subsequence be denoted by (xn (3) )∞ n=1 .

Continuing in this way we find a decreasing sequence of closed cubes

I1 ⊃ I2 ⊃ I3 ⊃ · · ·
and sequences
(1) (1) (1)
(x1 , x2 , x3 , . . .)
(2) (2) (2)
(x1 , x2 , x3 , . . .)
(3) (3) (3)
(x1 , x2 , x3 , . . .)
..
.
where each sequence is a subsequence of the preceding sequence and the terms
of the ith sequence are all members of the cube Ii .
(i)
We now define the sequence (yi ) by yi = xi for i = 1, 2, . . .. This is a
subsequence of the original sequence.
Notice that for each N , the terms yN , yN +1 , yN +2 , . . . are√
all members of
IN . Since the distance between any two points in IN is ≤ kr/2N −2 → 0
2

as N → ∞, it follows that (yi ) is a Cauchy sequence. Since Rn is complete,

it follows (yi ) converges in Rn . This proves the theorem.

Remark 1 If a sequence in Rn is not bounded, then it need not contain a

convergent subsequence. For example, the sequence (1, 2, 3, . . .) in R does not
contain any convergent subsequence (since the nth term of any subsequence
is ≥ n and so any subsequence is not even bounded).

Remark 2 The Theorem is not true if Rn is replaced by C[0, 1]. For ex-
ample, consider the sequence of functions (fn ) whose graphs are as shown in
the diagram.

fn

1 1
1 1 1
n+1 n 3 2
√
2
The distance
√ between any two points in I1 is ≤ 2 √kr, between any two points in I2
is thus ≤ kr, between any two points in I3 is thus ≤ kr/2, etc.
100

The sequence is bounded since ||fn ||∞ = 1, where we take the sup norm
(and corresponding metric) on C[0, 1].
But if n 6= m then

||fn − fm ||∞ = sup {|fn (x) − fm (x)| : x ∈ [0, 1]} = 1,

as is seen by choosing appropriate x ∈ [0, 1]. Thus no subsequence of (fn )

can converge in the sup norm3 .

We often use the previous theorem in the following form.

Corollary 9.2.2 If S ⊂ Rn , then S is closed and bounded iff every sequence

from S has a subsequence which converges to a limit in S.

Proof: Let S ⊂ Rn , be closed and bounded. Then any sequence from S

is bounded and so has a convergent subsequence by the previous Theorem.
The limit is in S as S is closed.
Conversely, first suppose S is not bounded. Then for every natural num-
ber n there exists xn ∈ S such that |xn | ≥ n. Any subsequence of (xn ) is
unbounded and so cannot converge.
Next suppose S is not closed. Then there exists a sequence (xn ) from S
which converges to x 6∈ S. Any subsequence also converges to x, and so does
not have its limit in S.

Remark The second half of this corollary holds in a general metric space,
the first half does not.

9.3 Compact Sets

Definition 9.3.1 A subset S of a metric space (X, d) is compact if every
sequence from S has a subsequence which converges to an element of S. If
X is compact, we say the metric space itself is compact.

Remark

1. This notion is also called sequential compactness. There is another

definition of compactness in terms of coverings by open sets which
applies to any topological space4 and agrees with the definition here for
metric spaces. We will investigate this more general notion in Chapter
15.
3
We will consider the sup norm on functions in detail in a later chapter. Notice that
fn (x) → 0 as n → ∞ for every x ∈ [0, 1]—we say that (fn ) converges pointwise to the
zero function. Thus here we have convergence pointwise but not in the sup metric. This
notion of pointwise convergence cannot be described by a metric.
4
You will study general topological spaces in a later course.
Subsequences and Compactness 101

2. Compactness turns out to be a stronger condition than completeness,

though in some arguments one notion can be used in place of the other.

Examples

1. From Corollary 9.2.2 the compact subsets of Rn are precisely the closed
bounded subsets. Any such compact subset, with the induced metric,
is a compact metric space. For example, [a, b] with the usual metric is
a compact metric space.

2. The Remarks on C[0, 1] in the previous section show that the closed5
bounded set S = {f ∈ C[0, 1] : ||f ||∞ = 1} is not compact. The set
S is just the “closed unit sphere” in C[0, 1]. (You will find later that
C[0, 1] is not unusual in this regard, the closed unit ball in any infinite
dimensional normed space fails to be compact.)

Relative and Absolute Notions Recall from the Note at the end of
Section (6.4) that if X is a metric space the notion of a set S ⊂ X being
open or closed is a relative one, in that it depends also on X and not just on
the induced metric on S.
However, whether or not S is compact depends only on S itself and the
induced metric, and so we say compactess is an absolute notion. Similarly,
completeness is an absolute notion.

9.4 Nearest Points

We now give a simple application in Rn of the preceding ideas.

Definition 9.4.1 Suppose A ⊂ X and x ∈ X where (X, d) is a metric space.

The distance from x to A is defined by

d(x, A) = inf d(x, y). (9.1)

y∈A

It is not necessarily the case that there exists y ∈ A such that d(x, A) =
d(x, y). For example if A = [0, 1) ⊂ R and x = 2 then d(x, A) = 1, but
d(x, y) > 1 for all y ∈ A.
Moreover, even if d(x, A) = d(x, y) for some y ∈ A, this y may not be
unique. For example, let S = {y ∈ R2 : ||y|| = 1} and let x = (0, 0). Then
d(x, S) = 1 and d(x, y) = 1 for every y ∈ S.
Notice also that if x ∈ A then d(x, A) = 0. But d(x, A) = 0 does not
imply x ∈ A. For example, take A = [0, 1) and x = 1.
5
S is the boundary of the unit ball B1 (0) in the metric space C[0, 1] and is closed as
noted in the Examples following Theorem 6.4.7.
102

However, we do have the following theorem. Note that in the result we

need S to be closed but not necessarily bounded.
The technique used in the proof, of taking a “minimising” sequence and
extracting a convergent subsequence, is a fundamental idea.

Theorem 9.4.2 Let S be a closed subset of Rn , and let x ∈ Rn . Then there

exists y ∈ S such that d(x, y) = d(x, S).

Proof: Let
γ = d(x, S)
and choose a sequence (yn ) in S such that

d(x, yn ) → γ as n → ∞.

This is possible from (9.1) by the definition of inf.

The sequence (yn ) is bounded6 and so has a convergent subsequence which

we also denote by (yn )7 .
Let y be the limit of the convergent subsequence (yn ). Then d(x, yn ) →
d(x, y) by Theorem 7.3.4, but d(x, yn ) → γ since this is also true for the
original sequence. It follows d(x, y) = γ as required.

6
This is fairly obvious and the actual argument is similar to showing that convergent
sequences are bounded, c.f. Theorem 7.3.3. More precisely, we have there exists an integer
N such that d(x, yn ) ≤ γ + 1 for all n ≥ N , by the fact d(x, yn ) → γ. Let

M = max{γ + 1, d(x, y1 ), . . . , d(x, yN −1 )}.

Then d(x, yn ) ≤ M for all n, and so (yn ) is bounded.

7
This abuse of notation in which we use the same notation for the subsequence as
for the original sequence is a common one. It saves using subscripts yni j — which are
particularly messy when we take subsequences of subsequences — and will lead to no
confusion provided we are careful.
Chapter 10

Limits of Functions

10.1 Diagrammatic Representation of Func-

tions

In this Chapter we will consider functions f : A (⊂ X) → Y where X and Y

are metric spaces.
Important cases are

1. f : A (⊂ R) → R,

2. f : A (⊂ Rn ) → R,

3. f : A (⊂ Rn ) → Rm .

You should think of these particular cases when you read the following.
Sometimes we can represent a function by its graph. Of course functions
can be quite complicated, and we should be careful not to be misled by the
simple features of the particular graphs which we are able to sketch. See the
following diagrams.

103
104

R2
R
graph of →R2
f:R→

Sometimes we can sketch the domain and the range of the function, per-
haps also with a coordinate grid and its image. See the following diagram.

Sometimes we can represent a function by drawing a vector at various

Limits of Functions 105

points in its domain to represent f (x).

f:R2→R2

Sometimes we can represent a real-valued function by drawing the level

sets (contours) of the function. See Section 17.6.2.
In other cases the best we can do is to represent the graph, or the domain
and range of the function, in a highly idealised manner. See the following
diagrams.
106

10.2 Definition of Limit

Suppose f : A (⊂ X) → Y where X and Y are metric spaces. In considering
the limit of f (x) as x → a we are interested in the behaviour of f (x) for x
near a. We are not concerned with the value of f at a and indeed f need not
even be defined at a, nor is it necessary that a ∈ A. See Example 1 following
the definition below.
For the above reasons we assume a is a limit point of A, which is equivalent
to the existence of some sequence (xn )∞n=1 ⊂ A \ {a} with xn → a as n → ∞.
In particular, a is not an isolated point of A.
The following definition of limit in terms of sequences is equivalent to
the usual ²–δ definition as we see in the next section. The definition here
is perhaps closer to the usual intuitive notion of a limit. Moreover, with
this definition we can deduce the basic properties of limits directly from the
corresponding properties of sequences, as we will see in Section 10.4.

Definition 10.2.1 Let f : A (⊂ X) → Y where X and Y are metric spaces,

and let a be a limit point of A. Suppose

(xn )∞
n=1 ⊂ A \ {a} and xn → a together imply f (xn ) → b.

Then we say f (x) approaches b as x approaches a, or f has limit b at a and

write
f (x) → b as x → a (x ∈ A),
or
lim f (x) = b,
x→a
x∈A
or
lim f (x) = b
x→a

(where in the last notation the intended domain A is understood from the
context).
Limits of Functions 107

Definition 10.2.2 [One-sided Limits] If in the previous definition X = R

and A is an interval with a as a left or right endpoint, we write

lim f (x) or lim− f (x)

x→a+ x→a

and say the limit as x approaches a from the right or the limit as x approaches
a from the left, respectively.

Example 1 (a) Let A = (−∞, 0) ∪ (0, ∞). Let f : A → R be given by

f (x) = 1 if x ∈ A. Then limx→0 f (x) = 1, even though f is not defined at 0.
(b) Let f : R → R be given by f (x) = 1 if x 6= 0 and f (0) = 0. Then
again limx→0 f (x) = 1.
This example illustrates why in the Definition 10.2.1 we require xn 6= a,
even if a ∈ A.

Example 2 If g : R → R then

g(x) − g(a)
lim
x→a x−a
is (if the limit exists) called the ³derivative of´ g at a. Note that in Defini-
tion 10.2.1 we are taking f (x) = g(x) − g(a) /(x − a) and that f (x) is not
defined at x = a. We take A = R \ {a}, or A = (a − δ, a) ∪ (a, a + δ) for
some δ > 0.

Example 3 (Exercise: Draw a sketch.)

µ ¶
1
lim x sin = 0. (10.1)
x→0 x
To see this take any sequence (xn ) where xn → 0 and xn 6= 0. Now
µ ¶
1
−|xn | ≤ xn sin ≤ |xn |.
xn
Since −|xn | → 0 and |xn | → 0 it follows from the Comparison Test that
xn sin(1/xn ) → 0, and so (10.1) follows.

We can also use the definition to show in some cases that limx→a f (x)
does not exist. For this it is sufficient to show that for some sequence (xn ) ⊂
A \ {a} with xn → a the corresponding sequence (f (xn )) does not have a
limit. Alternatively, it is sufficient to give two sequences (xn ) ⊂ A \ {a} and
(yn ) ⊂ A \ {a} with xn → a and yn → a but lim xn 6= lim yn .

Example 4 (Draw a sketch). limx→0 sin(1/x) does not exist. To see this
consider, for example, the sequences xn = 1/(nπ) and yn = 1/((2n + 1/2)π).
Then sin(1/xn ) = 0 → 0 and sin(1/yn ) = 1 → 1.
108

Example 5

Consider the function g : [0, 1] → [0, 1] defined by



 1/2 if x = 1/2



 1/4 if x = 1/4, 3/4




 1/8 if x = 1/8, 3/8, 5/8, 7/8
g(x) = ...





 1/2k if x = 1/2k , 3/2k , 5/2k , . . . , (2k − 1)/2k




 ..
.
g(x) = 0 otherwise.

In fact, in simpler terms,

(
1/2k x = p/2k for p odd, 1 ≤ p < 2k
g(x) =
0 otherwise

Then we claim limx→a g(x) = 0 for all a ∈ [0, 1].

First define Sk = {1/2k , 2/2k , 3/2k , . . . , (2k − 1)/2k } for k = 1, 2, . . ..
Notice that g(x) will take values 1/2, 1/4, . . . , 1/2k for x ∈ Sk , and

g(x) < 1/2k if x 6∈ Sk . (10.2)

For each a ∈ [0, 1] and k = 1, 2, . . . define the distance from a to Sk \ {a}

(whether or not a ∈ Sk ) by
n o
da,k = min |x − a| : x ∈ Sk \ {a} . (10.3)
Limits of Functions 109

Then da,k > 0, even if a ∈ Sk , since it is the minimum of a finite set of strictly
positive (i.e. > 0) numbers.
Now let (xn ) be any sequence with xn → a and xn 6= a for all n. We need
to show g(xn ) → 0.
Suppose ² > 0 and choose k so 1/2k ≤ ². Then from (10.2), g(xn ) < ² if
xn 6∈ Sk .
On the other hand, 0 < |xn − a| < da,k for all n ≥ N (say), since xn 6= a
for all n and xn → a. It follows from (10.3) that xn 6∈ Sk for n ≥ N . Hence
g(xn ) < ² for n ≥ N .
Since also g(x) ≥ 0 for all x it follows that g(xn ) → 0 as n → ∞. Hence
limx→a g(x) = 0 as claimed.

Example 6 Define h : R → R by
h(x) = m→∞ lim (cos(m!πx))n
lim n→∞

Then h fails to have a limit at every point of R.

Example 7 Let
xy
f (x, y) =
x2 + y2
for (x, y) 6= (0, 0).
If y = ax then f (x, y) = a(1 + a2 )−1 for x 6= 0. Hence
a
lim f (x, y) = .
(x,y)→a 1 + a2
y=ax

Thus we obtain a different limit of f as (x, y) → (0, 0) along different lines.

It follows that
lim f (x, y)
(x,y)→(0,0)

does not exist.

A partial diagram of the graph of f is shown
110

One can also visualize f by sketching level sets 1 of f as shown in the next
diagram. Then you can visualise the graph of f as being swept out by a
straight line rotating around the origin at a height as indicated by the level
sets.

Example 8 Let
x2 y
f (x, y) =
x4 + y 2

for (x, y) 6= (0, 0).

Then

ax3
lim f (x, y) = lim
(x,y)→a x→0 x4 + a2 x2
y=ax
ax
= lim 2
x→0 x + a2
= 0.

Thus the limit of f as (x, y) → (0, 0) along any line y = ax is 0. The limit
along the y-axis x = 0 is also easily seen to be 0.
But it is still not true that lim(x,y)→(0,0) f (x, y) exists. For if we consider
the limit of f as (x, y) → (0, 0) along the parabola y = bx2 we see that
f = b(1 + b2 )−1 on this curve and so the limit is b(1 + b2 )−1 .
You might like to draw level curves (corresponding to parabolas y = bx2 ).

1
A level set of f is a set on which f is constant.
Limits of Functions 111

This example reappears in Chapter 17. Clearly we can make such exam-
ples as complicated as we please.

10.3 Equivalent Definition

In the following theorem, (2) is the usual ²–δ definition of a limit.
The following diagram illustrates the definition of limit corresponding to
(2) of the theorem.

Theorem 10.3.1 Suppose (X, d) and (Y, ρ) are metric spaces, A ⊂ X,

f : A → Y , and a is a limit point of A. Then the following are equivalent:

1. limx→a f (x) = b;

2. For every ² > 0 there is a δ > 0 such that

x ∈ A \ {a}³ and d(x, a)

´ < δ implies ρ(f (x), b) < ²;
i.e. x ∈ A ∩ Bδ (a) \ {a} ⇒ f (x) ∈ B² (b).
112

Proof:
(1) ⇒ (2): Assume (1), so that whenever (xn ) ⊂ A \ {a} and xn → a then
f (xn ) → b.
Suppose (by way of obtaining a contradiction) that (2) is not true. Then
for some ² > 0 there is no δ > 0 such that

x ∈ A \ {a} and d(x, a) < δ implies ρ(f (x), b) < ².

In other words, for some ² > 0 and every δ > 0, there exists an x depending
on δ, with

x ∈ A \ {a} and d(x, a) < δ and ρ(f (x), b) ≥ ². (10.4)

Choose such an ², and for δ = 1/n, n = 1, 2, . . ., choose x = xn satis-

fying (10.4). It follows xn → a and (xn ) ⊂ A \ {a} but f (xn ) 6→ b. This
contradicts (1) and so (2) is true.

(2) ⇒ (1): Assume (2).

In order to prove (1) suppose (xn ) ⊂ A \ {a} and xn → a. We have to
show f (xn ) → b.
In order to do this take ² > 0. By (2) there is a δ > 0 (depending on ²)
such that
∗
z }| {
xn ∈ A \ {a} and d(xn , a) < δ implies ρ(f (xn ), b) < ².

But * is true for all n ≥ N (say, where N depends on δ and hence on ²), and
so ρ(f (xn ), b) < ² for all n ≥ N . Thus f (xn ) → b as required and so (1) is
true.

10.4 Elementary Properties of Limits

Assumption In this section we let f, g : A (⊂ X) → Y where (X, d) and
(Y, ρ) are metric spaces.

The next definition is not surprising.

Definition 10.4.1 The function f is bounded on the set E ⊂ A if f [E] is a

bounded set in Y .

Thus the function f : (0, ∞) → R given by f (x) = x−1 is bounded on [a, ∞)

for any a > 0 but is not bounded on (0, ∞).

Proposition 10.4.2 Assume limx→a f (x) exists. Then for some r > 0, f is
bounded on the set A ∩ Br (a).
Limits of Functions 113

Proof: Let limx→a f (x) = b and let V = B1 (b). V is certainly a bounded

set.
For some r > 0 we have f [(A \ {a}) ∩ Br (a)] ⊂ V from Theorem 10.3.1(2).
Since a subset of a bounded set is bounded, it follows that f [(A \ {a}) ∩ Br (a)]
is bounded, and so f [A ∩ Br (a)] is bounded if a 6∈ A. If a ∈ A then
f [A ∩ Br (a)] = f [(A \ {a}) ∩ Br (a)] ∪ {f (a)}, and so again f [A ∩ Br (a)]
is bounded.

Most of the basic properties of limits of functions follow directly from the
corresponding properties of limits of sequences without the necessity for any
²–δ arguments.

Theorem 10.4.3 Limits are unique; in the sense that if limx→a f (x) = b1
and limx→a f (x) = b2 then b1 = b2 .

Proof: Suppose limx→a f (x) = b1 and limx→a f (x) = b2 . If b1 6= b2 , then

² = |b1 −b
2
2|
> 0. By the defienition of the limit, there are δ1 , δ2 > 0 such that
0 < d(x, a) < δ1 ⇒ ρ(f (x) − b1 ) < ² ,
0 < d(x, a) < δ2 ⇒ ρ(f (x) − b2 ) < ² .
Taking 0 < d(x, a) < min{δ1 , δ − 2} gives a contradiction.

Notation Assume f : A (⊂ X) → Rn . (In applications it is often the case

that X = Rm for some m). We write
f (x) = (f 1 (x), . . . , f k (x)).
Thus each of the f i is just a real-valued function defined on A.

# the linear transformation f : R → R described by the

2 2
For "example,
a b
matrix is given by
c d
f (x) = f (x1 , x2 ) = (ax1 + bx2 , cx1 + dx2 ).
Thus f 1 (x) = ax1 + bx2 and f 2 (x) = cx1 + dx2 .

Theorem 10.4.4 Let f : A (⊂ X) → Rn . Then limx→a f (x) exists iff the

component limits, limx→a f i (x), exist for all i = 1, . . . , k. In this case
lim f (x) = (lim
x→a x→a
f 1 (x), . . . , x→a
lim f k (x)). (10.5)

Proof: Suppose limx→a f (x) exists and equals b = (b1 , . . . , bk ). We want

to show that limx→a f i (x) exists and equals bi for i = 1, . . . , k.
Let (xn )∞
n=1 ⊂ A \ {a} and xn → a. From Definition (10.2.1) we have
that lim f (xn ) = b and it is sufficient to prove that lim f i (xn ) = bi . But this
is immediate from Theorem (7.5.1) on sequences.
Conversely, if limx→a f i (x) exists and equals bi for i = 1, . . . , k, then a
similar argument shows limx→a f (x) exists and equals (b1 , . . . , bk ).
114

More Notation Let f, g : S → V where S is any set (not necessarily a

subset of a metric space) and V is any vector space. In particular, V = R
is an important case. Let α ∈ R. Then we define addition and scalar
multiplication of functions as follows:

(f + g)(x) = f (x) + g(x),

(αf )(x) = αf (x),

for all x ∈ S. That is, f + g is the function defined on S whose value at

each x ∈ S is f (x) + g(x), and similarly for αf . Thus addition of functions
is defined by addition of the values of the functions, and similarly for multi-
plication of a function and a scalar. The zero function is the function whose
value is everywhere 0. (It is easy to check that the set F of all functions
f : S → V is a vector space whose “zero vector” is the zero function.)
If V = R then we define the product and quotient of functions by

(f g)(x) = f (x)g(x),
à !
f f (x)
(x) = .
g g(x)

The domain of f /g is defined to be S \ {x : g(x) = 0}.

If V = X is an inner product space, then we define the inner product of
the functions f and g to be the function f · g : S → R given by

(f · g)(x) = f (x) · g(x).

The following algebraic properties of limits follow easily from the cor-
responding properties of sequences. As usual you should think of the case
X = Rn and V = Rn (in particular, m = 1).

Theorem 10.4.5 Let f, g : A (⊂ X) → V where X is a metric space and V

is a normed space. Let limx→a f (x) and limx→a g(x) exist. Let α ∈ R. Then
the following limits exist and have the stated values:

lim (f + g)(x) = lim f (x) + lim g(x),

x→a x→a x→a
lim (αf )(x) = α x→a
x→a
lim f (x).

If V = R then

lim (f g)(x) = lim f (x) lim g(x),

x→a x→a x→a
à !
f limx→a f (x)
lim (x) = ,
x→a g limx→a g(x)

provided in the last case that g(x) 6= 0 for all x ∈ A\{a}2 and limx→a g(x) 6= 0.
2
It is often convenient to instead just require that g(x) 6= 0 for all x ∈ Br (a) ∩ (A \ {a})
and some r > 0. In this case the function f /g will be defined everywhere in Br (a)∩(A\{a})
and the conclusion still holds.
Limits of Functions 115

If X = V is an inner product space, then

lim (f · g)(x) = lim f (x) · lim g(x).

x→a x→a x→a

Proof: Let limx→a f (x) = b and limx→a g(x) = c.

We prove the result for addition of functions.
Let (xn ) ⊂ A \ {a} and xn → a. From Definition 10.2.1 we have that

f (xn ) → b, g(xn ) → c, (10.6)

and it is sufficient to prove that

(f + g)(xn ) → b + c.

But

(f + g)(xn ) = f (xn ) + g(xn )

→ b+c

from (10.6) and the algebraic properties of limits of sequences, Theorem 7.7.1.
This proves the result.
The others are proved similarly. For the second last we also need the
Problem in Chapter 7 about the ratio of corresponding terms of two conver-
gent sequences.

One usually uses the previous Theorem, rather than going back to the
original definition, in order to compute limits.

Example If P and Q are polynomials then

P (x) P (a)
lim =
x→a Q(x) Q(a)

if Q(a) 6= 0.
To see this, let P (x) = a0 + a1 x + a2 x2 + · · · + an xn . It follows (exer-
cise) from the definition of limit that limx→a c = c (for any real number c)
and limx→a x = a. It then follows by repeated applications of the previous
theorem that limx→a P (x) = P (a). Similarly limx→a Q(x) = Q(a) and so the
result follows by again using the theorem.
116
Chapter 11

Continuity

As usual, unless otherwise clear from context, we consider functions

f : A (⊂ X) → Y , where X and Y are metric spaces.
You should think of the case X = Rn and Y = Rn , and in particular
Y = R.

11.1 Continuity at a Point

We first define the notion of continuity at a point in terms of limits, and then
we give a few useful equivalent definitions.
The idea is that f : A → Y is continuous at a ∈ A if f (x) is arbitrarily
close to f (a) for all x sufficiently close to a. Thus the value of f does not
have a “jump” at a. However, one’s intuition can be misleading, as we see
in the following examples.
If a is a limit point of A, continuity of f at a means limx→a, x∈A f (x) =
f (a). If a is an isolated point of A then limx→a, x∈A f (x) is not defined and
we always define f to be continuous at a in this (uninteresting) case.

Definition 11.1.1 Let f : A (⊂ X) → Y where X and Y are metric spaces

and let a ∈ A. Then f is continuous at a if a is an isolated point of A, or if
a is a limit point of A and limx→a, x∈A f (x) = f (a).
If f is continuous at every a ∈ A then we say f is continuous. The set of
all such continuous functions is denoted by
C(A; Y ),
or by
C(A)
if Y = R.

Example 1 Define
(
x sin(1/x) if x 6= 0,
f (x) =
0 if x = 0.

117
118

From Example 3 of Section 10.2, it follows f is continuous at 0. From the rules

about products and compositions of continuous functions, see Example 1
Section 11.2, it follows that f is continuous everywhere on R.

Example 2 From the Example in Section 10.4 it follows that any rational
function P/Q, and in particular any polynomial, is continuous everywhere it
is defined, i.e. everywhere Q(x) 6= 0.

Example 3 Define (
x if x ∈ Q
f (x) =
−x if x 6∈ Q
Then f is continuous at 0, and only at 0.

Example 4 If g is the function from Example 5 of Section 10.2 then it

follows that g is not continuous at any x of the form k/2m but is continuous
everywhere else.
The similar function
(
0 if x is irrational
h(x) =
1/q x = p/q in simplest terms

is continuous at every irrational and discontinuous at every rational. (*It is

possible to prove that there is no function f : [0, 1] → [0, 1] such that f is
continuous at every rational and discontinuous at every irrational.)

Example 5 Let g : Q → R be given by g(x) = x. Then g is continuous at

every x ∈ Q = dom g and hence is continuous. On the other hand, if f is the
function defined in Example 3, then g agrees with f everywhere in Q, but f
is continuous only at 0. The point is that f and g have different domains.

The following equivalent definitions are often useful. They also have the
advantage that it is not necessary to consider the case of isolated points and
limit points separately. Note that, unlike in Theorem 10.3.1, we allow xn = a
in (2), and we allow x = a in (3).

Theorem 11.1.2 Let f : A (⊂ X) → Y where (X, d) and (Y, ρ) are metric

spaces. Let a ∈ A. Then the following are equivalent.

1. f is continuous at a;

2. whenever (xn )∞
n=1 ⊂ A and xn → a then f (xn ) → f (a);

3. for each ² > 0 there exists δ > 0 such that

x ∈ A and d(x, a)
³ < δ implies
´ ³ρ(f (x),
´ f (a)) < ²;
i.e. f Bδ (a) ⊆ B² f (a) .
Limits of Functions 119

Proof: (1) ⇒ (2): Assume (1). Then in particular for any sequence
(xn ) ⊂ A \ {a} with xn → a, it follows f (xn ) → f (a).
In order to prove (2) suppose we have a sequence (xn ) ⊂ A with xn → a
(where we allow xn = a). If xn = a for all n ≥ some N then f (xn ) = f (a)
for n ≥ N and so trivially f (xn ) → f (a). If this case does not occur then
by deleting any xn with xn = a we obtain a new (infinite) sequence x0n → a
with (x0n ) ⊂ A \ {a}. Since f is continuous at a it follows f (x0n ) → f (a). As
also f (xn ) = f (a) for all terms from (xn ) not in the sequence (x0n ), it follows
that f (xn ) → f (a). This proves (2).

(2) ⇒ (1): This is immediate, since if f (xn ) → f (a) whenever (xn ) ⊂ A

and xn → a, then certainly f (xn ) → f (a) whenever (xn ) ⊂ A \ {a} and
xn → a, i.e. f is continuous at a.

The equivalence of (2) and (3) is proved almost exactly as is the equiva-
lence of the two corresponding conditions (1)and (2) in Theorem 10.3.1. The
only essential difference is that we replace A \ {a} everywhere in the proof
there by A.

Remark Property (3) here is perhaps the simplest to visualize, try giving
a diagram which shows this property.

11.2 Basic Consequences of Continuity

Remark

Note that if f : A → R, f is continuous at a, and f (a) = r > 0, then

f (x) > r/2 for all x sufficiently near a. In particular, f is strictly positive
for all x sufficiently near a. This is an immediate consequence of Theo-
rem 11.1.2 (3), since r/2 < f (x) < 3r/2 if d(x, a) < δ, say. Similar remarks
apply if f (a) < 0.
120

A useful consequence
Rb
of this observation is that if f : [a, b] → R is con-
tinuous, and a |f | = 0, then f = 0. (This fact has already been used in
Section 5.2.)

The following two Theorems are proved using Theorem 11.1.2 (2) in the
same way as are the corresponding properties for limits. The only difference
is that we no longer require sequences xn → a with (xn ) ⊂ A to also satisfy
xn 6= a.

Theorem 11.2.1 Let f : A (⊂ X) → Rn , where X is a metric space. Then

f is continuous at a iff f i is continuous at a for every i = 1, . . . , k.

Proof: As for Theorem 10.4.4

Theorem 11.2.2 Let f, g : A (⊂ X) → V where X is a metric space and V

is a normed space. Let f and g be continuous at a ∈ A. Let α ∈ R. Then
f + g and αf are continuous at a.
If V = R then f g is continuous at a, and moreover f /g is continuous at
a if g(a) 6= 0.
If X = V is an inner product space then f · g is continuous at a.

Proof: Using Theorem 11.1.2 (2), the proofs are as for Theorem 10.4.5,
except that we take sequences (xn ) ⊂ A with possibly xn = a. The only
extra point is that because g(a) 6= 0 and g is continuous at a then from the
remark at the beginning of this section, g(x) 6= 0 for all x sufficiently near a.

The following Theorem implies that the composition of continuous func-

tions is continuous.

Theorem 11.2.3 Let f : A (⊂ X) → B (⊂ Y ) and g : B → Z where X, Y

and Z are metric spaces. If f is continuous at a and g is continuous at f (a),
then g ◦ f is continuous at a.

Proof: Let (xn ) → a with (xn ) ⊂ A. Then f (xn ) → f (a) since f is

continuous at a. Hence g(f (xn )) → g(f (a)) since g is continuous at f (a).

Remark Use of property (3) again gives a simple picture of this result.

Example 1 Recall the function

(
x sin(1/x) if x 6= 0,
f (x) =
0 if x = 0.
Limits of Functions 121

from Section 11.1. We saw there that f is continuous at 0. Assuming the

function x 7→ sin x is everywhere continuous1 and recalling from Example 2
of Section 11.1 that the function x 7→ 1/x is continuous for x 6= 0, it follows
from the previous theorem that x 7→ sin(1/x) is continuous if x 6= 0. Since
the function x is also everywhere continuous and the product of continuous
functions is continuous, it follows that f is continuous at every x 6= 0, and
hence everywhere.

11.3 Lipschitz and Hölder Functions

We now define some classes of functions which, among other things, are very
important in the study of partial differential equations. An important case
to keep in mind is A = [a, b] and Y = R.

Definition 11.3.1 A function f : A (⊂ X) → Y , where (X, d) and (Y, ρ)

are metric spaces, is a Lipschitz continuous function if there is a constant M
with
ρ(f (x), f (x0 )) ≤ M d(x, x0 )
for all x, y ∈ A. The least such M is the Lipschitz constant M of f .

More generally:

Definition 11.3.2 A function f : A (⊂ X) → Y , where (X, d) and (Y, ρ) are

metric spaces, is Hölder continuous with exponent α ∈ (0, 1] if

ρ(f (x), f (x0 )) ≤ M d(x, x0 )α

for all x, x0 ∈ A and some fixed constant M .

Remarks

1. Hölder continuity with exponent α = 1 is just Lipschitz continuity.

³ ´1/α
2. Hölder continuous functions are continuous. Just choose δ = ²
M
in Theorem 11.1.2(3).

3. A contraction map (recall Section 8.3) has Lipschitz constant M < 1,

and conversely.

Examples
1
To prove this we need to first give a proper definition of sin x. This can be done by
means of an infinite series expansion.
122

1. Let f : [a, b] → R be a differentiable function and suppose |f 0 (x)| ≤ M

for all x ∈ [a, b]. If x 6= y are in [a, b] then from the Mean Value
Theorem,
f (y) − f (x)
= f 0 (ξ)
y−x
for some ξ between x and y. It follows |f (y) − f (x)| ≤ M |y − x| and
so f is Lipschitz with Lipschitz constant at most M .
2. An example of a Hölder continuous function
√ defined on [0, 1], which
is not Lipschitz continuous, is f (x) = x. This is Hölder continuous
with exponent 1/2 since
¯√
¯ √ ¯¯ |x − x0 |
¯ x − x0 ¯ = √ √
x + x0
q
q |x − x0 |
= |x − x0 | √ √
x + x0
q
≤ |x − x0 |.

This function is not Lipschitz continuous since

|f (x) − f (0)| 1
=√ ,
|x − 0| x
and the right side is not bounded by any constant independent of x for
x ∈ (0, 1].

11.4 Another Definition of Continuity

The following theorem gives a definition of “continuous function” in terms
only of open (or closed) sets. It does not deal with continuity of a function
at a point.

Theorem 11.4.1 Let f : X → Y , where (X, d) and (Y, ρ) are metric spaces.
Then the following are equivalent:

1. f is continuous;
2. f −1 [E] is open in X whenever E is open in Y ;
3. f −1 [C] is closed in X whenever C is closed Y .

Proof:
(1) ⇒ (2): Assume (1). Let E be open in Y . We wish to show that f −1 [E]
is open (in X).
Let x ∈ f³−1 [E].´ Then f (x) ∈ E, and since E is open there exists r > 0
such that Br f (x) ⊂ E. From Theorem 11.1.2(3) there exists δ > 0 such
Limits of Functions 123

h i ³ ´ h ³ ´i
that f Bδ (x) ⊂ Br f (x) . This implies Bδ (x) ⊂ f −1 Br f (x) . But
h ³ ´i
f −1 Br f (x) ⊂ f −1 [E] and so Bδ (x) ⊂ f −1 [E].
Thus every point x ∈ f −1 [E] is an interior point and so f −1 [E] is open.

(2) ⇔ (3): Assume (2), i.e. f −1 [E] is open in X whenever E is open

in Y . If´ C is closed in Y then C c is open and so f −1 [C c ] is open. But
³ c
f −1 [C] = f −1 [C c ]. Hence f −1 [C] is closed.
We can similarly show (3) ⇒ (2).

(2) ⇒ (1): Assume (2). We will use Theorem 11.1.2(3) to prove (1).
³ ´
Let x ∈ X. In order to prove f is continuous at x take any Br f (x) ⊂
³ ´ h ³ ´i
Y . Since Br f (x) is open it follows that f −1 Br f (x) is open. Since
³ ´ ³ ´
−1
x∈f [Br f (x) ] it follows there exists δ > 0 such that Bδ (x) ⊂ f −1 [Br f (x) ].
h i h h ³ ´ii h h ³ ´ii ³ ´
Hence f Bδ (x) ⊂ f f −1 Br f (x) ; but f f −1 Br f (x) ⊂ Br f (x)
h i ³ ´
(exercise) and so f Bδ (x) ⊂ Br f (x) .
It follows from Theorem 11.1.2(3) that f is continuous at x. Since x ∈ X
was arbitrary, it follows that f is continuous on X.

Corollary 11.4.2 Let f : S (⊂ X) → Y , where (X, d) and (Y, ρ) are metric

spaces. Then the following are equivalent:

1. f is continuous;

2. f −1 [E] is open in S whenever E is open (in Y );

3. f −1 [C] is closed in S whenever C is closed (in Y ).

Proof: Since (S, d) is a metric space, this follows immediately from the
preceding theorem.

Note The function f : R → R given by f (x) = 0 if x is irrational and f (x) =

1 if x is rational is not continuous anywhere, (this is Example 10.2). However,
124

the function g obtained by restricting f to Q is continuous everywhere on

Q.

Applications The previous theorem can be used to show that, loosely

speaking, sets defined by means of continuous functions and the inequalities
< and > are open; while sets defined by means of continuous functions, =,
≤ and ≥, are closed.

1. The half space

H (⊂ Rn ) = {x : z · x < c} ,
where z ∈ Rn and c is a scalar, is open (c.f. the Problems on Chapter
6.) To see this, fix z and define f : Rn → R by
f (x) = z · x.
Then H = f −1 (−∞, c). Since f is continuous2 and (−∞, c) is open, it
follows H is open.
2. The set n o
S (⊂ R2 ) = (x, y) : x ≥ 0 and x2 + y 2 ≤ 1
is closed. To see this let S = S1 ∩ S2 where
n o
S1 = {(x, y) : x ≥ 0} , S2 = (x, y) : x2 + y 2 ≤ 1 .

Then S1 = g −1 [0, ∞) where g(x, y) = x. Since g is continuous and

[0, ∞) is closed, it follows that S1 is closed. Similarly S2 = f −1 [0, 1]
where f (x, y) = x2 + y 2 , and so S2 is closed. Hence S is closed being
the intersection of closed sets.

Remark It is not always true that a continuous image3 of an open set

is open; nor is a continuous image of a closed set always closed. But see
Theorem 11.5.1 below.
For example, if f : R → R is given by f (x) = x2 then f [(−1, 1)] = [0, 1),
so that a continuous image of an open set need not be open. Also, if f (x) = ex
then f [R] = (0, ∞), so that a continuous image of a closed set need not be
closed.

11.5 Continuous Functions on Compact Sets

We saw at the end of the previous section that a continuous image of a closed
set need not be closed. However, the continuous image of a closed bounded
subset of Rn is a closed bounded set.
More generally, for arbitrary metric spaces the continuous image of a
compact set is compact.
2
If xn → x then f xn ) = z · xn → z · x = f (x) from Theorem 7.7.1.
3
By a continuous image we just mean the image under a continuous function.
Limits of Functions 125

Theorem 11.5.1 Let f : K (⊂ X) → Y be a continuous function, where

(X, d) and (Y, ρ) are metric spaces, and K is compact. Then f [K] is compact.

Proof: Let (yn ) be any sequence from f [K]. We want to show that some
subsequence has a limit in f [K].
Let yn = f (xn ) for some xn ∈ K. Since K is compact there is a sub-
sequence (xni ) such that xni → x (say) as i → ∞, where x ∈ K. Hence
yni = f (xni ) → f (x) since f is continuous, and moreover f (x) ∈ f [K] since
x ∈ K. It follows that f [K] is compact.

You know from your earlier courses on Calculus that a continuous function
defined on a closed bounded interval is bounded above and below and has
a maximum value and a minimum value. This is generalised in the next
theorem.

Theorem 11.5.2 Let f : K (⊂ X) → R be a continuous function, where

(X, d) is a metric space and K is compact. Then f is bounded (above and
below) and has a maximum and a minimum value.

Proof: From the previous theorem f [K] is a closed and bounded subset of
R. Since f [K] is bounded it has a least upper bound b (say), i.e. b ≥ f (x)
for all x ∈ K. Since f [K] is closed it follows that b ∈ f [K]4 . Hence b = f (x0 )
for some x0 ∈ K, and so f (x0 ) is the maximum value of f on K.
Similarly, f has a minimum value taken at some point in K.

Remarks The need for K to be compact in the previous theorem is illus-

trated by the following examples:

1. Let f (x) = 1/x for x ∈ (0, 1]. Then f is continuous and (0, 1] is
bounded, but f is not bounded above on the set (0, 1].

2. Let f (x) = x for x ∈ [0, 1). Then f is continuous and is even bounded
above on [0, 1), but does not have a maximum on [0, 1).

3. Let f (x) = 1/x for x ∈ [1, ∞). Then f is continuous and is bounded
below on [1, ∞) but does not have a minimum on [1, ∞).

11.6 Uniform Continuity

In this Section, you should first think of the case X is an interval in R and
Y = R.
4
To see this take a sequence from f [K] which converges to b (the existence of such
a sequence (yn ) follows from the definition of least upper bound by choosing yn ∈ f [K],
yn ≥ b − 1/n.) It follows that b ∈ f [K] since f [K] is closed.
126

Definition 11.6.1 Let (X, d) and (Y, ρ) be metric spaces. The function
f : X → Y is uniformly continuous on X if for each ² > 0 there exists δ > 0
such that
d(x, x0 ) < δ ⇒ ρ(f (x), f (x0 )) < ²,
for all x, x0 ∈ X.

Remark The point is that δ may depend on ², but does not depend on x
or x0 .

Examples

1. Hölder continuous (and in particular Lipschitz continuous) functions

³ ´1/α
are uniformly continuous. To see this, just choose δ = ² in
M
Definition 11.6.1.

2. The function f (x) = 1/x is continuous at every point in (0, 1) and

hence is continuous on (0, 1). But f is not uniformly continuous on
(0, 1).
For example, choose ² = 1 in the definition of uniform continuity. Sup-
pose δ > 0. By choosing x sufficiently close to 0 (e.g. if |x| < δ) it is
clear that there exist x0 with |x − x0 | < δ but |1/x − 1/x0 | ≥ 1. This
contradicts uniform continuity.

3. The function f (x) = sin(1/x) is continuous and bounded, but not

uniformly continuous, on (0, 1).

4. Also, f (x) = x2 is continuous, but not uniformly continuous, on R(why?).

On the other hand, f is uniformly continuous on any bounded interval
from the next Theorem. Exercise: Prove this fact directly.

You should think of the previous examples in relation to the following

theorem.

Theorem 11.6.2 Let f : S → Y be continuous, where X and Y are metric

spaces, S ⊂ X and S is compact. Then f is uniformly continuous.

Proof: If f is not uniformly continuous, then there exists ² > 0 such that
for every δ > 0 there exist x, y ∈ S with

d(x, y) < δ and ρ(f (x), f (y)) ≥ ².

Fix some such ² and using δ = 1/n, choose two sequences (xn ) and (yn )
such that for all n

xn , yn ∈ S, d(xn , yn ) < 1/n, ρ(f (xn ), f (yn )) ≥ ². (11.1)

Limits of Functions 127

Since S is compact, by going to a subsequence if necessary we can suppose

that xn → x for some x ∈ S. Since
d(yn , x) ≤ d((yn , xn ) + d(xn , x),
and both terms on the right side approach 0, it follows that also yn → x.

Since f is continuous at x, there exists τ > 0 such that

z ∈ S, d(z, x) < τ ⇒ ρ(f (z), f (x)) < ²/2. (11.2)

Since xn → x and yn → x, we can choose k so d(xk , x) < τ and d(yk , x) <

τ . It follows from (11.2) that for this k
ρ(f (xk ), f (yk )) ≤ ρ(f (xk ), f (x)) + ρ(f (x), f (yk ))
< ²/2 + ²/2 = ².
But this contradicts (11.1). Hence f is uniformly continuous.

Corollary 11.6.3 A continuous real-valued function defined on a closed bounded

subset of Rn is uniformly continuous.

Proof: This is immediate from the theorem, as closed bounded subsets of

Rn are compact.

Corollary 11.6.4 Let K be a continuous real-valued function on the square

[0, 1]2 . Then the function
Z 1
f (x) = K(x, t)dt
0

is (uniformly) continuous.

Proof: We have
Z 1
|f (x) − f (y)| ≤ |K(x, t) − K(y, t)|dt .
0

Uniform continuity of K on [0, 1]2 means that given ² > 0 there is δ > 0 such
that |K(x, s) − K(y, t)| < ² provided d((x, s), (y, t)) < δ. So if |x − y| < δ
|f (x) − f (y)| < ².

Exercise There is a converse to Corollary 11.6.3: if every function contin-

uous on a subset of R is uniformly continuous, then the set is closed. Must
it also be bounded?
128
Chapter 12

Uniform Convergence of
Functions

12.1 Discussion and Definitions

Consider the following examples of sequences of functions (fn )∞
n=1 , with
graphs as shown. In each case f is in some sense the limit function, as
we discuss subsequently.

1.

f, fn : [−1, 1] → R

for n = 1, 2, . . ., where


 0 −1 ≤ x ≤ 0

 2nx 0 < x ≤ (2n)−1
fn (x) = 

 2 − 2nx (2n)−1 < x ≤ n−1

0 n−1 < x ≤ 1

and
f (x) = 0 for all x.

129
130

2.

f, fn : [−1, 1] → R

for n = 1, 2, . . . and
f (x) = 0 for all x.

3.

f, fn : [−1, 1] → R

for n = 1, 2, . . . and (
1 x=0
f (x) =
6 0
0 x=

4.

f, fn : [−1, 1] → R
Uniform Convergence of Functions 131

for n = 1, 2, . . . and (
0 x<0
f (x) =
1 0≤x
5.

f, fn : R → R
for n = 1, 2, . . . and
f (x) = 0 for all x.
6.

f, fn : R → R
for n = 1, 2, . . . and
f (x) = 0 for all x.
7.
132

f, fn : R → R
for n = 1, 2, . . .,
1
fn (x) = sin nx,
n
and
f (x) = 0 for all x.

In every one of the preceding cases, fn (x) → f (x) as n → ∞ for each

x ∈ domf , where domf is the domain of f .
For example, in 1, consider the cases x ≤ 0 and x > 0 separately. If
x ≤ 0 then fn (x) = 0 for all n, and so it is certainly true that fn (x) → 0 as
n → ∞. On the other hand, if x > 0, then fn (x) = 0 for all n > 1/x 1 , and
in particular fn (x) → 0 as n → ∞.
In all cases we say that fn → f in the pointwise sense. That is, for each
² > 0 and each x there exists N such that
n ≥ N ⇒ |fn (x) − f (x)| < ².
In cases 1–6, N depends on x as well as ²: there is no N which works for
all x.
We can see this by imagining the “²-strip” about the graph of f , which
we define to be the set of all points (x, y) ∈ R2 such that
f (x) − ² < y < f (x) + ².

Then it is not the case for Examples 1–6 that the graph of fn is a subset
of the ²-strip about the graph of f for all sufficiently large n.
However, in Example 7, since
1 1
|fn (x) − f (x)| = | sin nx − 0| ≤ ,
n n
it follows that |fn (x) − f (x)| < ² for all n > 1/². In other words, the graph
of fn is a subset of the ²-strip about the graph of f for all sufficiently large
n. In this case we say that fn → f uniformly.
1
Notice that how large n needs to be depends on x.
Uniform Convergence of Functions 133

Finally we remark that in Examples 5 and 6, if we consider fn and f

restricted to any fixed bounded set B, then it is the case that fn → f
uniformly on B.
Motivated by the preceding examples we now make the following defini-
tions.
In the following think of the case S = [a, b] and Y = R.

Definition 12.1.1 Let f, fn : S → Y for n = 1, 2, . . ., where S is any set and

(Y, ρ) is a metric space.
If fn (x) → f (x) for all x ∈ S then fn → f pointwise.
If for every ² > 0 there exists N such that
³ ´
n ≥ N ⇒ ρ fn (x), f (x) < ²

for all x ∈ S, then fn → f uniformly (on S).

³ ´
Remarks (i) Informally, fn → f uniformly means ρ fn (x), f (x) → 0
“uniformly” in x. See also Proposition 12.2.3.
(ii) If fn → f uniformly then clearly fn → f pointwise, but not necessarily
conversely, as we have seen. However, see the next theorem for a partial
converse.
(iii) Note that fn → f does not converge uniformly iff there exists ² > 0 and
a sequence (xn ) ⊂ S such that |f (x) − fn (xn )| ≥ ² for all n.

It is also convenient to define the notion of a uniformly Cauchy sequence

of functions.

Definition 12.1.2 Let fn : S → Y for n = 1, 2, . . ., where S is any set and

(Y, ρ) is a metric space. Then the sequence (fn ) is uniformly Cauchy if for
every ² > 0 there exists N such that
³ ´
m, n ≥ N ⇒ ρ fn (x), fm (x) < ²

for all x ∈ S.

Remarks (i) Thus (fn ) is uniformly Cauchy iff the following is true: for
each ² > 0, any two functions from the sequence after a certain point (which
will depend on ²) lie within the ²-strip of each other.
³ ´
(ii) Informally, (fn ) is uniformly Cauchy if ρ fn (x), fm (x) → 0 “uniformly”
in x as m, n → ∞.
(iii) We will see in the next section that if Y is complete (e.g. Y = Rn ) then
a sequence (fn ) (where fn : S → Y ) is uniformly Cauchy iff fn → f uniformly
for some function f : S → Y .
134

Theorem 12.1.3 (Dini’s Theorem) Suppose (fn ) is an increasing sequence

(i.e. f1 (x) ≤ f2 (x) ≤ . . . for all x ∈ S) of real-valued continuous functions
defined on the compact subset S of some metric space (X, d). Suppose fn → f
pointwise and f is continuous. Then fn → f uniformly.

Proof: Suppose ² > 0. For each n let

A²n = {x ∈ S : f (x) − fn (x) < ²}.
Since (fn ) is an increasing sequence,
A²1 ⊂ A²2 ⊂ . . . . (12.1)
Since fn (x) → f (x) for all x ∈ S,
∞
[
S= A²n . (12.2)
n=1

Since fn and f are continuous,

A²n is open in S (12.3)
for all n (see Corollary 11.4.2).
In order to prove uniform convergence, it is sufficient (why?) to show
there exists n (depending on ²) such that
S = A²n (12.4)
(note that then S = A²m for all m > n from (12.1)). If no such n exists then
for each n there exists xn such that
xn ∈ S \ A²n (12.5)
for all n. By compactness of S,there is a subsequence xnk with xnk →
x0 (say) ∈ S.
From (12.2) it follows x0 ∈ A²N for some N . From (12.3) it follows xnk ∈
A²N for all nk ≥ M (say), where we may take M ≥ N . But then from (12.1)
we have
xnk ∈ A²nk
for all nk ≥ M . This contradicts (12.5). Hence (12.4) is true for some n,
and so the theorem is proved.

Remarks (i) The same result hold if we replace “increasing” by “decreas-

ing”. The proof is similar; or one can deduce the result directly from the
theorem by replacing fn and f by −fn and −f respectively.
(ii) The proof of the Theorem can be simplified by using the equivalent
definition of compactness given in Definition 15.1.2. See the Exercise after
Theorem 15.2.1.
(iii) Exercise: Give examples to show why it is necessary that the sequence
be increasing (or decreasing) and that f be continuous.
Uniform Convergence of Functions 135

12.2 The Uniform Metric

In the following think of the case S = [a, b] and Y = R.
In order to understand uniform convergence it is useful to introduce the
uniform “metric” du . This is not quite a metric, only because it may take
the value +∞, but it otherwise satisfies the three axioms for a metric.

Definition 12.2.1 Let F(S, Y ) be the set of all functions f : S → Y , where

S is a set and (Y, ρ) is a metric space. Then the uniform “metric” du on
F(S, Y ) is defined by
³ ´
du (f, g) = sup ρ f (x), g(x) .
x∈S

If Y is a normed space (e.g. R), then we define the uniform “norm” by

||f ||u = sup ||f (x)||.

x∈S

(Thus in this case the uniform “metric” is the metric corresponding to the
uniform “norm”, as in the examples following Definition 6.2.1)

In the case S = [a, b] and Y = R it is clear that the ²-strip about f

is precisely the set of functions g such that du (f, g) < ². A similar remark
applies for general S and Y if the “²-strip” is appropriately defined.
The uniform metric (norm) is also known as the sup metric (norm). We
have in fact already defined the sup metric and norm on the set C[a, b] of con-
tinuous real-valued functions; c.f. the examples of Section 5.2 and Section 6.2.
The present definition just generalises this to other classes of functions.
The distance du between two functions can easily be +∞. For example,
let S = [0, 1], Y = R. Let f (x) = 1/x if x 6= 0 and f (0) = 0, and let g be
the zero function. Then clearly du (f, g) = ∞. In applications we will only
be interested in the case du (f, g) is finite, and in fact small.
We now show the three axioms for a metric are satisfied by du , provided
we define

∞ + ∞ = ∞, c ∞ = ∞ if c > 0, c ∞ = 0 if c = 0. (12.6)

Theorem 12.2.2 The uniform “metric” (“norm”) satisfies the axioms for
a metric space (Definition 6.2.1) (Definition 5.2.1) provided we interpret
arithmetic operations on ∞ as in (12.6).

Proof: It is easy to see that du satisfies positivity and symmetry (Exercise).

136

For the triangle inequality let f, g, h : S → Y . Then2

³ ´
du (f, g) = supx∈S ρ f (x), g(x)
h ³ ´ ³ ´i
≤ supx∈S ρ f (x), h(x) + ρ h(x), g(x)
³ ´ ³ ´
≤ supx∈S ρ f (x), h(x) + supx∈S ρ h(x), g(x)
= du (f, h) + du (h, g).

This completes the proof in this case.

Proposition 12.2.3 A sequence of functions is uniformly convergent iff it

is convergent in the uniform metric.

Proof: Let S be a set and (Y, ρ) be a metric space.

Let f, fn ∈ F(S, Y ) for n = 1, 2, . . .. From the definition we have that
fn → f uniformly iff for every ² > 0 there exists N such that
³ ´
n ≥ N ⇒ ρ fn (x), f (x) < ²

for all x ∈ S. This is equivalent to

n ≥ N ⇒ du (fn , f ) < ².

The result follows.

Proposition 12.2.4 A sequence of functions is uniformly Cauchy iff it is

Cauchy in the uniform metric.

Proof: As in previous result.

We next establish the relationship between uniformly convergent, and

uniformly Cauchy, sequences of functions.

Theorem 12.2.5 Let S be a set and (Y, ρ) be a metric space.

If f, fn ∈ F(S, Y ) and fn → f uniformly, then (fn )∞
n=1 is uniformly
Cauchy.
Conversely, if (Y, ρ) is a complete metric space, and (fn )∞
n=1 ⊂ F(S, Y )
is uniformly Cauchy, then fn → f uniformly for some f ∈ F(S, Y ).

Proof: First suppose fn → f uniformly.

Let ² > 0. Then there exists N such that

n ≥ N ⇒ ρ(fn (x), f (x)) < ²

2
The third line uses uses the fact (Exercise) that if u and v are two real-valued functions
defined on the same domain S, then supx∈S (u(x) + v(x)) ≤ supx∈S u(x) + supx∈S v(x).
Uniform Convergence of Functions 137

for all x ∈ S. Since

³ ´ ³ ´ ³ ´
ρ fn (x), fm (x) ≤ ρ fn (x), f (x) + ρ f (x), fm (x) ,

it follows
m, n ≥ N ⇒ ρ(fn (x), fm (x)) < 2²

for all x ∈ S. Thus (fn ) is uniformly Cauchy.

Next assume (Y, ρ) is complete and suppose (fn ) is a uniformly Cauchy

sequence.
³ ´
It follows from the definition of uniformly Cauchy that fn (x) is a
Cauchy sequence for each x ∈ S, and so has a limit in Y since Y is complete.
Define the function f : S → Y by

f (x) = lim fn (x)

n→∞

for each x ∈ S.
We know that fn → f in the pointwise sense, but we need to show that
fn → f uniformly.
So suppose that ² > 0 and, using the fact that (fn ) is uniformly Cauchy,
choose N such that
³ ´
m, n ≥ N ⇒ ρ fn (x), fm (x) < ²

for all x ∈ S. Fixing m ≥ N and letting n → ∞3 , it follows from the

Comparison Test that
³ ´
ρ f (x), fm (x) ≤ ²

for all x ∈ S 4 .
Since this applies to every m ≥ N , we have that
³ ´
m ≥ N ⇒ ρ f (x), fm (x) ≤ ².

for all x ∈ S. Hence fn → f uniformly.

3
This is a commonly used technique; it will probably seem strange at first.
4
In more detail, we argue as follows: Every term in the sequence of real numbers
³ ´ ³ ´ ³ ´
ρ fN (x), fm (x) , ρ fN +1 (x), fm (x) , ρ fN +2 (x), fm (x) , . . .

³ ´
is < ². Since fN +p (x) → f (x) as p → ∞, it follows that ρ fN +p (x), fm (x) →
³ ´
ρ f (x), fm (x) as p → ∞ (this is clear if Y is R or Rk , and follows in general from
³ ´
Theorem 7.3.4). By the Comparison Test it follows that ρ f (x), fm (x) ≤ ².
138

12.3 Uniform Convergence and Continuity

In the following think of the case X = [a, b] and Y = R.

We saw in Examples 3 and 4 of Section 12.1 that a pointwise limit of con-

tinuous functions need not be continuous. The next theorem shows however
that a uniform limit of continuous functions is continuous.

Theorem 12.3.1 Let (X, d) and (Y, ρ) be metric spaces. Let fn : X → Y

for n = 1, 2, . . . be a sequence of continuous functions such that fn → f
uniformly. Then f is continuous.

Proof: Consider any x0 ∈ X; we will show f is continuous at x0 .

Suppose ² > 0. Using the fact that fn → f uniformly, first choose N so
that ³ ´
ρ fN (x), f (x) < ² (12.7)
for all x ∈ X. Next, using the fact that fN is continuous, choose δ > 0 so
that ³ ´
d(x, x0 ) < δ ⇒ ρ fN (x), fN (x0 ) < ². (12.8)

It follows from (12.7) and (12.8) that if d(x, x0 ) < δ then

³ ´ ³ ´ ³ ´ ³ ´
ρ f (x), f (x0 ) ≤ ρ f (x), fN (x) + ρ fN (x), fN (x0 ) + ρ fN (x0 ), f (x0 )
< 3².

Y fN
f
fN (x0)

f(x0 )

fN(x)
f(x)

x x0 X

Hence f is continuous at x0 , and hence continuous as x0 was an arbitrary

point in X.

The next result is very important. We will use it in establishing the

existence of solutions to systems of differential equations.
Uniform Convergence of Functions 139

Recall from Definition 11.1.1 that if X and Y are metric spaces and
A ⊂ X, then the set of all continuous functions f : A → Y is denoted by
C(A; Y ). If A is compact, then we have seen that f [A] is compact and hence
bounded5 , i.e. f is bounded. If A is not compact, then continuous functions
need not be bounded6 .

Definition 12.3.2 The set of bounded continuous functions f : A → Y is

denoted by
BC(A; Y ).

Theorem 12.3.3 Suppose A ⊂ X, (X, d) is a metric space and (Y, ρ) is a

complete metric spaces. Then BC(A; Y ) is a complete metric space with the
uniform metric du .

Proof: It has already been verified in Theorem 12.2.2 that the three axioms
for a metric are satisfied. We need only check that du (f, g) is always finite
for f, g ∈ BC(A; Y ).
But this is immediate. For suppose b ∈ Y . Then since f and g are
bounded on A, it follows there exist K1 and K2 such that ρ(f (x), b) ≤ K1
and ρ(g(x), b) ≤ K2 for all x ∈ A. But then du (f, g) ≤ K1 + K2 from the
definition of du and the triangle inequality. Hence BC(A; Y ) is a metric space
with the uniform metric.
In order to verify completeness, let (fn )∞
n=1 be a Cauchy sequence from
BC(A; Y ). Then (fn ) is uniformly Cauchy, as noted in Proposition 12.2.4.
From Theorem 12.2.5 it follows that fn → f uniformly, for some function
f : A → Y . From Proposition 12.2.3 it follows that fn → f in the uniform
metric.
From Theorem 12.3.1 it follows that f is continuous. It is also clear that
f is bounded7 . Hence f ∈ BC(A; Y ).
We have shown that fn → f in the sense of the uniform metric du , where
f ∈ BC(A; Y ). Hence (BC(A; Y ), du ) is a complete metric space.

Corollary 12.3.4 Let (X, d) be a metric space and (Y, ρ) be a complete met-
ric space. Let A ⊂ X be compact. Then C(A; Y ) is a complete metric space
with the uniform metric du .

Proof: Since A is compact, every continuous function defined on A is

bounded. The result now follows from the Theorem.
5
In Rn , compactness is the same as closed and bounded . This is not true in general,
but it is always true that compact implies closed and bounded . The proof is the same as
in Corollary 9.2.2.
6
Let A = (0, 1) and f (x) = 1/x, or A = R and f (x) = x.
7
Choose N so du (fN , f ) ≤ 1. Choose any b ∈ Y . Since fN is bounded, ρ(fN (x), b) ≤ K
say, for all x ∈ A. It follows ρ(f (x), b) ≤ K + 1 for all x ∈ A.
140

Corollary 12.3.5 The set C[a, b] of continuous real-valued functions defined

on the interval [a, b], and more generally the set C([a, b] : Rn ) of continuous
maps into Rn , are complete metric spaces with the sup metric.

We will use the previous corollary to find solutions to (systems of) differ-
ential equations.

12.4 Uniform Convergence and Integration

It is not necessarily
Rb
true that
Rb
if fn → f pointwise, where f, fn : [a, b] → R are
continuous,
R1
then f
a n → f
Ra 1
. In particular, in Example 2 from Section 12.1,
−1 fn = 1/2 for all n but −1 f = 0. However, integration is better behaved
under uniform convergence.

Theorem 12.4.1 Suppose that f, fn : [a, b] → R for n = 1, 2, . . . are contin-

uous functions and fn → f uniformly. Then
Z b Z b
fn → f.
a a
Rx Rx
Moreover, a fn → a f uniformly for x ∈ [a, b].

Proof: Suppose ² > 0. By uniform convergence, choose N so that

n ≥ N ⇒ |fn (x) − f (x)| < ² (12.9)

for all x ∈ [a, b]

Since fn and f are continuous, they are Riemann integrable. Moreover,8
for n ≥ N ¯R
¯ b R b ¯¯ ¯R ³
¯ b
´¯
¯
¯ a fn − a f ¯ = ¯ a fn − f ¯
R
≤ ab |fn − f |
R
≤ ab ² from (12.9)
= (b − a)².
Rb Rb
It follows that a fn → a f , as required.
Rx Rx
For uniform convergence just note that the same proof gives | a fn − a f|
≤ (x − a)² ≤ (b − a)².

*Remarks
8
For the following, recall ¯Z ¯
¯ b ¯ Z b
¯ ¯
¯ g¯ ≤ |g|,
¯ a ¯ a
Z b Z b
f (x) ≤ g(x) for all x ∈ [a, b] ⇒ f≤ g.
a a
Uniform Convergence of Functions 141

1. More generally, it is not hard to show that the uniform limit of a

sequence of Riemann integrable functions is also Riemann integrable,
and that the corresponding integrals converge. See [Sm, Theorem 4.4,
page 101].
2. There is a much more important notion of integration, called Lebesgue
integration. Lebesgue integration has much nicer properties with re-
spect to convergence than does Riemann integration. See, for exam-
ple, [F], [St] and [Sm].

12.5 Uniform Convergence and Differentia-

tion
Suppose that fn : [a, b] → R, for n = 1, 2, . . ., is a sequence of differentiable
functions, and that fn → f uniformly. It is not true that fn0 (x) → f 0 (x) for
all x ∈ [a, b], in fact it need not even be true that f is differentiable.
For example, let f (x) = |x| for x ∈ [0, 1]. Then f is not differentiable
at 0. But, as indicated in the following diagram, it is easy to find a sequence
(fn ) of differentiable functions such that fn → f uniformly.

In particular, let
(
n 2
2
x+ 1
2n
0 ≤ |x| ≤ n1
fn (x) =
|x| 1
n
≤ |x| ≤ 1
Then the fn are differentiable on [−1, 1] (the only points to check are x =
±1/n), and fn → f uniformly since du (fn , f ) ≤ 1/n.
Example 7 from Section 12.1 gives an example where fn → f uniformly
and f is differentiable, but fn0 does not converge for most x. In fact, fn0 (x) =
cos nx which does not converge (unless x = 2kπ for some k ∈ Z (exercise)).
However, if the derivatives themselves converge uniformly to some limit,
then we have the following theorem.

Theorem 12.5.1 Suppose that fn : [a, b] → R for n = 1, 2, . . . and that the

fn0 exist and are continuous. Suppose fn → f pointwise on [a, b] and (fn0 )
converges uniformly on [a, b].
142

Then f 0 exists and is continuous on [a, b] and fn0 → f 0 uniformly on [a, b].
Moreover, fn → f uniformly on [a, b].

Proof: By the Fundamental Theorem of Calculus,

Z x
fn0 = fn (x) − fn (a) (12.10)
a

for every x ∈ [a, b].

Let fn0 → g(say) uniformly. Then from (12.10), Theorem 12.4.1 and the
hypotheses of the theorem,
Z x
g = f (x) − f (a). (12.11)
a

Since g is continuous, the left side of (12.11) is differentiable on [a, b]

and the derivative equals g 9 . Hence the right side is also differentiable and
moreover
g(x) = f 0 (x)
on [a, b].
Thus f 0 exists and is continuous and fn0 → f 0 uniformly on [a, b].
R Rx
Since fn (x) = ax fn0 and f (x) = a f 0 , uniform convergence of fn to f now
follows from Theorem 12.4.1.

9
Recall that the integral of a continuous function is differentiable, and the derivative
is just the original function.
Chapter 13

First Order Systems of

Differential Equations

The main result in this Chapter is the Existence and Uniqueness Theorem
for first order systems of (ordinary) differential equations. Essentially any
differential equation or system of differential equations can be reduced to a
first-order system, so the result is very general. The Contraction Mapping
Principle is the main ingredient in the proof.
The local Existence and Uniqueness Theorem for a single equation, to-
gether with the necessary preliminaries, is in Sections 13.3, 13.7–13.9. See
Sections 13.10 and 13.11 for the global result and the extension to systems.
These sections are independent of the remaining sections.
In Section 13.1 we give two interesting examples of systems of differential
equations.
In Section 13.2 we show how higher order differential equations (and more
generally higher order systems) can be reduced to first order systems.
In Sections 13.4 and 13.5 we discuss “geometric” ways of analysing and
understanding the solutions to systems of differential equations.
In Section 13.6 we give two examples to show the necessity of the condi-
tions assumed in the Existence and Uniqueness Theorem.

13.1 Examples

13.1.1 Predator-Prey Problem

Suppose there are two species of animals, and let the populations at time t
be x(t) and y(t) respectively. We assume we can approximate x(t) and y(t)
by differentiable functions. Species x is eaten by species y . The rates of

143
144

increase of the species are given by

dx
= ax − bxy − ex2 ,
dt (13.1)
dy
= −cy + dxy − f y 2 .
dt

The quantities a, b, c, d, e, f are constants and depend on the environment

and the particular species.
A quick justification of this model is as follows:

The term ax represents the usual rate of growth of x in the

case of an unlimited food supply and no predators. The term
bxy comes from the number of contacts per unit time between
predator and prey, it is proportional to the populations x and y,
and represents the rate of decrease in species x due to species y
eating it. The term ex2 is similarly due to competition between
members of species x for the limited food supply.
The term −cy represents the natural rate of decrease of species
y if its food supply, species x, were removed. The term dxy is
proportional to the number of contacts per unit time between
predator and prey, and accounts for the growth rate of y in the
absence of other effects. The term f y 2 accounts for competi-
tion between members of species y for the limited food supply
(species x).

We will return to this system later. It is first order, since only first
derivatives occur in the equation, and nonlinear, since some of the terms
involving the unknowns (or dependent variables) x and y occur in a nonlinear
way (namely the terms xy, x2 and y 2 ). It is a system of ordinary differential
equations since there is only one independent variable t, and so we only form
ordinary derivatives; as opposed to differential equations where there are two
or more independent variables, in which case the differential equation(s) will
involve partial derivatives.

13.1.2 A Simple Spring System

Consider a body of mass m connected to a wall by a spring and sliding over

a surface which applies a frictional force, as shown in the following diagram.
First Order Systems 145

Let x(t) be the displacement at time t from the equilibrium position.

From Newton’s second law, the force acting on the mass is given by

Force = mx00 (t).

If the spring obeys Hooke’s law, then the force is proportional to the dis-
placement, but acts in the opposite direction, and so

Force = −kx(t),

for some constant k > 0 which depends on the spring. Thus

mx00 (t) = −kx(t),

i.e.
mx00 (t) + kx(t) = 0.

If there is also a force term, due to friction, and proportional to the

velocity but acting in the opposite direction, then

Force = −kx − cx0 ,

for some constant c > 0, and so

mx00 (t) + cx0 (t) + kx(t) = 0. (13.2)

This is a second order ordinary differential equation, since it contains

second derivatives of the “unknown” x, and is linear since the unknown and
its derivatives occur in a linear manner.

13.2 Reduction to a First Order System

It is usually possible to reduce a higher order ordinary differential equation
or system of ordinary differential equations to a first order system.
For example, in the case of the differential equation (13.2) for the spring
system in Section 13.1.2, we introduce a new variable y corresponding to the
146

velocity x0 , and so obtain the following first order system for the “unknowns”
x and y:
x0 = y
(13.3)
y 0 = −m−1 cy − m−1 kx
This is a first order system (linear in this case).
If x, y is a solution of (13.3) then it is clear that x is a solution of (13.2).
Conversely, if x is a solution of (13.2) and we define y(t) = x0 (t), then x, y is
a solution of (13.3).

An nth order differential equation is a relation between a function x and

its first n derivatives. We can write this in the form
³ ´
F x(n) (t), x(n−1) (t), . . . , x0 (t), x(t), t = 0,
or ³ ´
F x(n) , x(n−1) , . . . , x0 , x, t = 0.
Here t ∈ I for some interval I ⊂ R, where I may be open, closed, or infinite,
at either end. If I = [a, b], say, then we take one-sided derivatives at a and b.
One can usually, in principle, solve this for xn , and so write
³ ´
x(n) = G x(n−1) , . . . , x0 , x, t . (13.4)

In order to reduce this to a first order system, introduce new functions

x , x2 , . . . , xn , where
1

x1 (t) = x(t)
x2 (t) = x0 (t)
x3 (t) = x00 (t)
..
.
x (t) = x(n−1) (t).
n

Then from (13.4) we see (exercise) that x1 , x2 , . . . , xn satisfy the first

order system
dx1
= x2 (t)
dt
dx2
= x3 (t)
dt
dx3
= x4 (t)
dt (13.5)
..
.
dxn−1
= xn (t)
dt
dxn
= G(xn , . . . , x2 , x1 , t)
dt
Conversely, if x1 , x2 , . . . , xn satisfy (13.5) and we let x(t) = x1 (t), then
we can check (exercise) that x satisfies (13.4).
First Order Systems 147

13.3 Initial Value Problems

Notation If x is a real-valued function defined in some interval I, we say
x is continuously differentiable (or C 1 ) if x is differentiable on I and its
derivative is continuous on I. Note that since x is differentiable, it is in
particular continuous. Let
C 1 (I)
denote the set of real-valued continuously differentiable functions defined on
I. Usually I = [a, b] for some a and b, and in this case the derivatives at a
and b are one-sided.
More generally, let x(t) = (x1 (t), . . . , xn (t)) be a vector-valued function
with values in Rn .1 Then we say x is continuously differentiable (or C 1 ) if
each xi (t) is C 1 . Let
C 1 (I; Rn )
denote the set of all such continuously differentiable functions.
In an analogous manner, if a function is continuous, we sometimes say it
is C 0 .

We will consider the general first order system of differential equations in

the form
dx1
= f 1 (t, x1 , x2 , . . . , xn )
dt
dx2
= f 2 (t, x1 , x2 , . . . , xn )
dt
..
.
dxn
= f n (t, x1 , x2 , . . . , xn ),
dt
which we write for short as
dx
= f (t, x).
dt
Here

x = (x1 , . . . , xn )
dx dx1 dxn
= ( ,..., )
dt dt dt
f (t, x) = f (t, x1 , . . . , xn )
³ ´
= f 1 (t, x1 , . . . , xn ), . . . , f n (t, x1 , . . . , xn ) .

It is usually convenient to think of t as representing time, but this is not

necessary.
1
You can think of x(t) as tracing out a curve in Rn .
148

We will always assume f is continuous for all (t, x) ∈ U , where U ⊂

R × Rn = Rn+1 .
By an initial condition is meant a condition of the form

x1 (t0 ) = x10 , x2 (t0 ) = x20 , . . . , xn (t0 ) = xn0

for some given t0 and some given x0 = (x10 , . . . , xn0 ). That is,

x(t0 ) = x0 .

Here, (t0 , x0 ) ∈ U .
The following diagram sketches the situation (schematically in the case
n > 1).

In case n = 2, we have the following diagram.

As an example, in the case of the predator-prey problem (13.1), it is

reasonable to restrict (x, y) to U = {(x, y) : x > 0, y > 0}2 . We might think
2
What happens if one of x or y vanishes at some point in time ?
First Order Systems 149

of restricting t to t ≥ 0, but since the right side of (13.1) is independent of t,

and since in any case the choice of what instant in time should correspond
to t = 0 is arbitrary, it is more reasonable not to make any restrictions on t.
Thus we might take U = R × {(x, y) : x > 0, y > 0} in this example. We
also usually assume for this problem that we know the values of x and y at
some “initial” time t0 .

Definition 13.3.1 [Initial Value Problem] Assume U ⊂ R×Rn = Rn+1 ,

U is open3 and (t0 , x0 ) ∈ U . Assume f (= f (t, x)) : U → R is continu-
ous. Then the following is called an initial value problem, with initial condi-
tion (13.7):

dx
= f (t, x), (13.6)
dt
x(t0 ) = x0 . (13.7)

We say x(t) = (x1 (t), . . . , xn (t)) is a solution of this initial value problem for
t in the interval I if:

1. t0 ∈ I,

2. x(t0 ) = x0 ,

3. (t, x(t)) ∈ U and x(t) is C 1 for t ∈ I,

4. the system of equations (13.6) is satisfied by x(t) for all t ∈ I.

13.4 Heuristic Justification for the

Existence of Solutions
To simplify notation, we consider the case n = 1 in this section. Thus we
consider a single differential equation and an initial value problem of the form

x0 (t) = f (t, x(t)), (13.8)

x(t0 ) = x0 . (13.9)

As usual, assume f is continuous on U , where U is an open set containing

(t0 , x0 ).
It is reasonable to expect that there should exist a (unique) solution
x = x(t) to (13.8) satisfying the initial condition (13.9) and defined for all t
in some time interval I containing t0 . We make this plausible as follows (see
the following diagram).

3
Sometimes it is convenient to allow U to be the closure of an open set.
150

From (13.8) and (13.9) we know x0 (t0 ) = f (t0 , x0 ). It follows that

for small h > 0

x(t0 + h) ≈ x0 + hf (t0 , x0 ) =: x1 4

Similarly

x(t0 + 2h) ≈ x1 + hf (t0 + h, x1 ) =: x2

x(t0 + 3h) ≈ x2 + hf (t0 + 2h, x2 ) =: x3
..
.

Suppose t∗ > t0 . By taking sufficiently many steps, we thus

obtain an approximation to x(t∗ ) (in the diagram we have shown
the case where h is such that t∗ = t0 + 3h). By taking h < 0
we can also find an approximation to x(t∗ ) if t∗ < t0 . By taking
h very small we expect to find an approximation to x(t∗ ) to any
desired degree of accuracy.

In the previous diagram

P = (t0 , x0 )
Q = (t0 + h, x1 )
R = (t0 + 2h, x2 )
S = (t0 + 3h, x3 )

The slope of P Q is f (t0 , x0 ) = f (P ), of QR is f (t0 +h, x1 ) = f (Q),

and of RS is f (t0 + 2h, x2 ) = f (R).
4
a := b means that a, by definition, is equal to b. And a =: b means that b, by definition,
is equal to a.
First Order Systems 151

The method outlined is called the method of Euler polygons. It can be

used to solve differential equations numerically, but there are refinements of
the method which are much more accurate. Euler’s method can also be made
the basis of a rigorous proof of the existence of a solution to the initial value
problem (13.8), (13.9). We will take a different approach, however, and use
the Contraction Mapping Theorem.

Direction field

Direction Field and Solutions of x0 (t) = −x − sin t

Consider again the differential equation (13.8). At each point in the

(t, x) plane, one can draw a line segment with slope f (t, x). The set of all
line segments constructed in this way is called the direction field for the
differential equation. The graph of any solution to (13.8) must have slope
given by the line segment at each point through which it passes. The direction
field thus gives a good idea of the behaviour of the set of solutions to the
differential equation.

13.5 Phase Space Diagrams

A useful way of visualising the behaviour of solutions to a system of differen-
tial equations (13.6) is by means of a phase space diagram. This is nothing
more than a set of paths (solution curves) in Rn (here called phase space)
traced out by various solutions to the system. It is particularly usefu in the
case n = 2 (i.e. two unknowns) and in case the system is autonomous (i.e.
the right side of (13.6) is independent of time).
Note carefully the difference between the graph of a solution, and the path
traced out by a solution in phase space. In particular, see the second diagram
in Section 13.3, where R2 is phase space.
152

We now discuss some general considerations in the context of the following

example.

Competing Species
Consider the case of two species whose populations at time t are x(t) and
y(t). Suppose they have a good food supply but fight each other whenever
they come into contact. By a discussion similar to that in Section 13.1.1,
their populations may be modelled by the equations
dx ³ ´
= ax − bxy = f 1 (x, y) ,
dt
dy ³ ´
= cy − dxy = f 2 (x, y) ,
dt
for suitable a, b, c, d > 0. Consider as an example the case a = 1000, b = 1,
c = 2000 and d = 1.
If a solution x(t), y(t) passes through a point (x, y) in phase space at some
time t, then the “velocity” of the path at this point is (f 1 (x, y), f 2 (x, y)) =
(x(1000 − y), y(2000 − x)). In particular, the path is tangent to the vector
(x(1000 − y), y(2000 − x)) at the point (x, y). The set of all such velocity
vectors (f 1 (x, y), f 2 (x, y)) at the points (x, y) ∈ R2 is called the velocity field
associated to the system of differential equations. Notice that as the example
we are discussing is autonomous, the velocity field is independent of time.

In the previous diagram we have shown some vectors from the velocity
field for the present system of equations. For simplicity, we have only shown
their directions in a few cases, and we have normalised each vector to have
the same length; we sometimes call the resulting vector field a direction field5 .
5
Note the distinction between the direction field in phase space and the direction field
for the graphs of solutions as discussed in the last section.
First Order Systems 153

Once we have drawn the velocity field (or direction field), we have a good
idea of the structure of the set of solutions, since each solution curve must
be tangent to the velocity field at each point through which it passes.
Next note that (f 1 (x, y), f 2 (x, y)) = (0, 0) if (x, y) = (0, 0) or (2000, 1000).
Thus the “velocity” (or rate of change) of a solution passing through either
of these pairs of points is zero. The pair of constant functions given by
x(t) = 2000 and y(t) = 1000 for all t is a solution of the system, and from
Theorem 13.10.1 is the only solution passing through (2000, 1000). Such a
constant solution is called a stationary solution or stationary point. In this
example the other stationary point is (0, 0) (this is not surprising!).
The stationary point (2000, 1000) is unstable in the sense that if we
change either population by a small amount away from these values, then
the populations do not converge back to these values. In this example, one
population will always die out. This is all clear from the diagram.

13.6 Examples of Non-Uniqueness

and Non-Existence
Example 1 (Non-Uniqueness) Consider the initial value problem
dx q
= |x|, (13.10)
dt
x(0) = 0. (13.11)

We use the method of separation of variables, we formally compute from (13.10)

that
dx
q = dt.
|x|
If x > 0 , integration gives
x1/2
= t − a,
1/2
for some a. That is, for x > 0,

x(t) = (t − a)2 /4. (13.12)

We need to justify these formal computations. By differentiating, we

check that (13.12) is indeed a solution of (13.10) provided t ≥ a.
Note also that x(t) = 0 is a solution of (13.10) for all t.
Moreover, we can check that for each a ≥ 0 there is a solution of (13.10)
and (13.11) given by
(
0 t≤a
x(t) =
(t − a) /4 t > a.
2

See the following diagram.

154

Thus we do not have uniqueness for solutions of (13.10), (13.11). There

are even more solutions to (13.10), (13.11), what are they ? (exercise).

We will later prove that we have uniqueness of solutions of the initial

value problem (13.8), (13.9) provided the function f (t, x) is locally Lipschitz
with respect to x, as defined in the next section.

Example 2 (Non-Existence) Let f (t, x) = 1 if x ≤ 1, and f (t, x) = 2 if

x > 1. Notice that f is not continuous. Consider the initial value problem

x0 (t) = f (t, x(t)), (13.13)

x(0) = 0. (13.14)

Then it is natural to take the solution to be

(
t t≤1
x(t) =
2t − 1 t > 1

t t≤1

1
x(t) = { 2t - 1 t > 1

1 t

Notice that x(t) satisfies the initial condition and also satisfies the dif-
ferential equation provided t 6= 1. But x(t) is not differentiable at t = 1.
First Order Systems 155

There is no solution of this initial value problem, in the usual sense of a

solution. It is possible to generalise the notion of a solution, and in this case
the “solution” given is the correct one.

13.7 A Lipschitz Condition

As we saw in Example 1 of Section 13.6, we need to impose a further condition
on f , apart from continuity, if we are to have a unique solution to the Initial
Value Problem (13.8), (13.9). We do this by generalising slightly the notion
of a Lipschitz function as defined in Section 11.3.

Definition 13.7.1 The function f = f (t, x) : A (⊂ R × Rn ) → R is Lipschitz

with respect to x (in A) if there exists a constant K such that

(t, x1 ), (t, x2 ) ∈ A ⇒ |f (t, x1 ) − f (t, x2 )| ≤ K|x1 − x2 |.

If f is Lipschitz with respect to x in Ah,k (t0 , x0 ), for every set Ah,k (t0 , x0 ) ⊂ A
of the form

Ah,k (t0 , x0 ) := {(t, x) : |t − t0 | ≤ h, |x − x0 | ≤ k} , (13.15)

then we say f is locally Lipschitz with respect to x, (see the following dia-
gram).

(t, x )
1

(t, x2)
A

(t 0, x ) h
0

A h.k (t , x )
0 0

We could have replaced the sets Ah,k (t0 , x0 ) by closed balls centred at
(t, x0 ) without affecting the definition, since each such ball contains a set
156

Ah,k (t0 , x0 ) for some h, k > 0, and conversely. We choose sets of the form
Ah,k (t0 , x0 ) for later convenience.
The difference between being Lipschitz with respect to x and being locally
Lipschitz with respect to x is clear from the following Examples.

Example 1 Let n = 1 and A = R × R. Let f (t, x) = t2 + 2 sin x. Then

|f (t, x1 ) − f (t, x2 )| = |2 sin x1 − 2 sin x2 |

= |2 cos ξ| |x1 − x2 |
≤ 2|x1 − x2 |,

for some ξ between x1 and x2 , using the Mean Value Theorem.

Thus f is Lipschitz with respect to x (it is also Lipschitz in the usual
sense).

Example 2 Let n = 1 and A = R × R. Let f (t, x) = t2 + x2 . Then

|f (t, x1 ) − f (t, x2 )| = |x21 − x22 |

= |2ξ| |x1 − x2 |,

for some ξ between x1 and x2 , again using the Mean Value Theorem. If
x1 , x2 ∈ B for some bounded set B, in particular if B is of the form
{(t, x) : |t − t0 | ≤ h, |x − x0 | ≤ k}, then ξ is also bounded, and so f is lo-
cally Lipschitz in A. But f is not Lipschitz in A.

We now give an analogue of the result from Example 1 in Section 11.3.

Theorem 13.7.2 Let U ⊂ R × Rn be open and let f = f (t, x) : U → R. If

∂f
the partial derivatives (t, x) all exist and are continuous in U , then f is
∂xi
locally Lipschitz in U with respect to x.

Proof: Let (t0 , x0 ) ∈ U . Since U is open, there exist h, k > 0 such that

Ah,k (t0 , x0 ) := {(t, x) : |t − t0 | ≤ h, |x − x0 | ≤ k} ⊂ U.

∂f
Since the partial derivatives (t, x) are continuous on the compact set Ah,k ,
∂xi
they are also bounded on Ah,k from Theorem 11.5.2. Suppose
¯ ¯
¯ ∂f ¯
¯ ¯
¯ (t, x)¯ ≤ K, (13.16)
¯ ∂xi ¯

for i = 1, . . . , n and (t, x) ∈ Ah,k .

Let (t, x1 ), (t, x2 ) ∈ Ah,k . To simplify notation, let n = 2 and let x1 =
(x11 , x21 ),
x2 = (x12 , x22 ). Then
(see the following diagram—note that it is in Rn , not in R×Rn ; here n = 2.)
First Order Systems 157

|f (t, x1 ) − f (t, x2 )| = |f (t, x11 , x21 ) − f (t, x12 , x22 )|

≤ |f (t, x11 , x21 ) − f (t, x12 , x21 )| + |f (t, x12 , x21 ) − f (t, x12 , x22 )|
∂f ∂f
= | (ξ1 )| |x12 − x11 | + | (ξ2 )| |x22 − x21 |
∂x1 ∂x2
≤ K|x12 − x11 | + K|x22 − x21 | from (13.16)
≤ 2K|x1 − x2 |,

In the third line, ξ1 is between x1 and x∗ = (x12 , x21 ), and ξ2 is between

x∗ = (x12 , x21 ) and x2 . This uses the usual Mean Value Theorem for a function
of one variable, applied on the interval [x11 , x12 ], and on the interval [x21 , x22 ].
This completes the proof if n = 2. For n > 2 the proof is similar.

13.8 Reduction to an Integral Equation

We again consider the case n = 1 in this section.
Thus we again consider the Initial Value Problem
x0 (t) = f (t, x(t)), (13.17)
x(t0 ) = x0 . (13.18)
As usual, assume f is continuous in U , where U is an open set containing
(t0 , x0 ).
The first step in proving the existence of a solution to (13.17), (13.18) is
to show that the problem is equivalent to solving a certain integral equation.
This follows easily by integrating both sides of (13.17) from t0 to t. More
precisely:

Theorem 13.8.1 Assume the function x satifies (t, x(t)) ∈ U for all t ∈ I,
where I is some closed bounded interval. Assume t0 ∈ I. Then x is a C 1
solution to (13.17), (13.18) in I iff x is a C 0 solution to the integral equation
Z t ³ ´
x(t) = x0 + f s, x(s) ds (13.19)
t0

in I.
158

Proof: First let x be a C 1 solution to (13.17), (13.18) in I. Then the

left side, and hence both sides, of (13.17) are continuous and in particular
integrable. Hence for any t ∈ I we have by integrating (13.17) from t0 to t
that
Z t ³ ´
x(t) − x(t0 ) = f s, x(s) ds.
t0

Since x(t0 ) = x0 , this shows x is a C 1 (and in particular a C 0 ) solution

to (13.19) for t ∈ I.

Conversely, assume x is a C 0 solution to (13.19) for t ∈ I. Since the

functions t 7→ x(t) and t 7→ t are continuous, it follows that the function
s 7→ (s, x(s)) is continuous from Theorem 11.2.1. Hence s 7→ f (s, x(s)) is
continuous from Theorem 11.2.3. It follows from (13.19), using properties of
indefinite integrals of continuous functions6 , that x0 (t) exists and

x0 (t) = f (t, x(t))

for all t ∈ I. In particular, x is C 1 on I. Finally, it follows immediately

from 13.19 that x(t0 ) = x0 . Thus x is a C 1 solution to (13.17), (13.18) in I.

Remark A bootstrap argument shows that the solution x is in fact C ∞

provided f is C ∞ .

13.9 Local Existence

We again consider the case n = 1 in this section.
We first use the Contraction Mapping Theorem to show that the integral
equation (13.19) has a solution on some interval containing t0 .

Theorem 13.9.1 Assume f is continuous, and locally Lipschitz with respect

to the second variable, on the open set U ⊂ R × R. Let (t0 , x0 ) ∈ U . Then
there exists h > 0 such that the integral equation
Z t ³ ´
x(t) = x0 + f t, x(t) dt (13.20)
t0

has a unique C 0 solution for t ∈ [t0 − h, t0 + h].

Proof:
Rt
6
If h exists and is continuous on I, t0 ∈ I and g(t) = t0
h(s) ds for all t ∈ I, then g 0
exists and g 0 = h on I. In particular, g is C 1 .
First Order Systems 159

(t , x ) graph of x(t)
0 0

t -h t t
0 t 0 +h
0

Choose h, k > 0 so that

Ah,k (t0 , x0 ) := {(t, x) : |t − t0 | ≤ h, |x − x0 | ≤ k} ⊂ U.

Since f is continuous, it is bounded on the compact set Ah,k (t0 , x0 ) by

Theorem 11.5.2. Choose M such that

|f (t, x)| ≤ M if (t, x) ∈ Ah,k (t0 , x0 ). (13.21)

Since f is locally Lipschitz with respect to x, there exists K such that

|f (t, x1 ) − f (t, x2 )| ≤ K|x1 − x2 | if (t, x1 ), (t, x2 ) ∈ Ah,k (t0 , x0 ). (13.22)

By decreasing h if necessary, we will require

( )
k 1
h ≤ min , . (13.23)
M 2K

Let C ∗ [t0 − h, t0 + h] be the set of continuous functions defined on [t0 −

h, t0 + h] whose graphs lie in Ah,k (t0 , x0 ). That is,
\
C ∗ [t0 − h, t0 + h] = C[t0 − h, t0 + h]
n o
x(t) : |x(t) − x0 | ≤ k for all t ∈ [t0 − h, t0 + h] .

Now C[t0 − h, t0 + h] is a complete metric space with the uniform metric,

as noted in Example 1 of Section 12.3. Since C ∗ [t0 − h, t0 + h] is a closed
subset7 , it follows from the “generalisation” following Theorem 8.2.2 that
C ∗ [t0 − h, t0 + h] is also a complete metric space with the uniform metric.
We want to solve the integral equation (13.20).
To do this consider the map

T : C ∗ [t0 − h, t0 + h] → C ∗ [t0 − h, t0 + h]
7
If xn → x uniformly and |xn (t)| ≤ k for all t, then |x(t)| ≤ k for all t.
160

defined by
Z t ³ ´
(T x)(t) = x0 + f t, x(t) dt for t ∈ [t0 − h, t0 + h]. (13.24)
t0

Notice that the fixed points of T are precisely the solutions of (13.20).
We check that T is indeed a map into C ∗ [t0 − h, t0 + h] as follows:

(i) Since in (13.24) we are taking the definite integral of a continuous func-
tion, Corollary 11.6.4 shows that T x is a continuous function.

(ii) Using (13.21) and (13.23) we have

¯Z t ³ ´ ¯¯
¯
|(T x)(t) − x0 | = ¯ f t, x(t) dt¯¯
¯
t0
Z t¯ ³ ´¯
¯ ¯
≤ ¯f t, x(t) ¯ dt
t0
≤ hM
≤ k.

It follows from the definition of C ∗ [t0 − h, t0 + h] that T x ∈ C ∗ [t0 − h, t0 + h].

We next check that T is a contraction map. To do this we compute for
x1 , x2 ∈ C ∗ [t0 − h, t0 + h], using (13.22) and (13.23), that
¯Z t µ ³ ´ ³ ´¶ ¯¯
¯
¯
|(T x1 )(t) − (T x2 )(t)| = ¯ f t, x1 (t) − f t, x2 (t) dt¯¯
t0
Z t¯ ³ ´ ³ ´¯
¯ ¯
≤ ¯f t, x1 (t) − f t, x2 (t) ¯ dt
t0
Z t
≤ K|x1 (t) − x2 (t)| dt
t0
≤ Kh sup |x1 (t) − x2 (t)|
t∈[t0 −h,t0 +h]
1
≤ du (x1 , x2 ).
2
Hence
1
du (T x1 , T x2 ) ≤ du (x1 , x2 ).
2
Thus we have shown that T is a contraction map on the complete metric
space C ∗ [t0 − h, t0 + h], and so has a unique fixed point. This completes the
proof of the theorem, since as noted before the fixed points of T are precisely
the solutions of (13.20).

Since the contraction mapping theorem gives an algorithm for finding the
fixed point, this can be used to obtain approximates to the solution of the
differential equation. In fact the argument can be sharpened considerably.
At the step (13.25)
First Order Systems 161

Z t
|(T x1 )(t) − (T x2 )(t)| ≤ K|x1 (t) − x2 (t)| dt
t0
≤ K|t − t0 |du (x1 , x2 ).

Thus applying the next step of the iteration,

Z t
|(T x1 )(t) − (T x2 )(t)| ≤
2 2
K|(T x1 )(t) − (T x2 )(t)| dt
t0
Z t
≤ K2 |t − t0 | dt du (x1 , x2 )
t0
|t − t0 |2
≤ K2 du (x1 , x2 ).
2
Induction gives
r |t − t0 |r
|(T x1 )(t) − (T x2 )(t) ≤ K
r r
du (x1 , x2 ).
r!
Without using the fact that Kh < 1/2, it follows that some power of T
is a contraction. Thus by one of the problems T itself has a unique fixed
point. This observation generally facilitates the obtaining of a larger domain
for the solution.

Example The simple equation x0 (t) = x(t), x(0) = 1 is well known to

have solution the exponential function. Applying the above algorithm, with
x1 (t) = 1 we would have
Z t ³ ´
(T x1 )(t) = 1 + f t, x1 (t) dt = 1 + t,
t0
Z t t2
(T 2 x1 )(t) = 1 + (1 + t)dt = 1 + t +
t0 2
..
.
X
k
ti
(T k x1 )(t) = ,
i=0 i!
giving the exponential series, which in fact converges to the solution uni-
formly on any bounded interval.

Theorem 13.9.2 (Local Existence and Uniqueness) Assume that f (t, x)

is continuous, and locally Lipschitz with respect to x, in the open set U ⊂
R × R. Let (t0 , x0 ) ∈ U . Then there exists h > 0 such that the initial value
problem
x0 (t) = f (t, x(t)),
x(t0 ) = x0 ,
has a unique C 1 solution for t ∈ [t0 − h, t0 + h].

Proof: By Theorem 13.8.1, x is a C 1 solution to this initial value problem iff

it is a C 0 solution to the integral equation (13.20). But the integral equation
has a unique solution in some [t0 − h, t0 + h], by the previous theorem.
162

13.10 Global Existence

Theorem 13.9.2 shows the existence of a (unique) solution to the initial value
problem in some (possibly small) time interval containing t0 . Even if U =
R × R it is not necessarily true that a solution exists for all t.

Example 1 Consider the initial value problem

x0 (t) = x2 ,
x(0) = a,

where for this discussion we take a ≥ 0.

If a = 0, this has the solution

x(t) = 0, (all t).

If a > 0 we use separation of variables to show that the solution is

1
x(t) = , (t < a−1 ).
a−1 −t
It follows from the Existence and Uniqueness Theorem that for each a this
gives the only solution.
Notice that if a > 0, then the solution x(t) → ∞ as t → a−1 from the
left, and x(t) is undefined for t = a−1 . Of course this x also satisfies x0 (t) = x
for t > a−1 , as do all the functions
1
xb (t) = , (0 < b < a).
b−1−t
Thus the presciption of x(0) = a gives zero information about the solution
for t > a−1 .
The following diagram shows the solution for various a.
First Order Systems 163

The following theorem more completely analyses the situation.

Theorem 13.10.1 (Global Existence and Uniqueness) There is a unique

solution x to the initial value problem (13.17), (13.18) and

1. either the solution exists for all t ≥ t0 ,

2. or the solution exists for all t0 ≤ t < T , for some (finite) T > t0 ; in
which case for any closed bounded subset A ⊂ U we have (t, x(t)) 6∈ A
for all t < T sufficiently close to T .

A similar result applies to t ≤ t0 .

Remark* The second alternative in the Theorem just says that the graph
of the solution eventually leaves any closed bounded A ⊂ U . We can think
of it as saying that the graph of the solution either escapes to infinity or
approaches the boundary of U as t → T .

Proof: * (Outline) Let T be the supremum of all t∗ such that a solution

exists for t ∈ [t0 , t∗ ]. If T = ∞, then we are done.
If T is finite, let A ⊂ U where A is compact. If (t, x(t)) does not eventually
leave A, then there exists a sequence tn → T such that (tn , x(tn )) ∈ A. From
the definition of compactness, a subsequence of (tn , x(tn )) must have a limit
in A. Let (tni , x(tni )) → (T, x) ∈ A (note that tni → T since tn → T ). In
particular, x(tni ) → x.
The proof of the Existence and Uniqueness Theorem shows that a solution
beginning at (T, x) exists for some time h > 0, and moreover, that for t0 =
T − h/4, say, the solution beginning at (t0 , x(t0 )) exists for time h/2. But this
then extends the original solution past time T , contradicting the definition
of T .
Hence (t, x(t)) does eventually leave A.

13.11 Extension of Results to Systems

The discussion, proofs and results in Sections 13.4, 13.8, 13.9 and 13.10
generalise to systems, with essentially only notational changes, as we now
sketch.
Thus we consider the following initial value problem for systems:
dx
= f (t, x),
dt
x(t0 ) = x0 .
This is equivalent to the integral equation
Z t ³ ´
x(t) = x0 + f s, x(s) ds.
t0
164

The integral of the vector function on the right side is defined componentwise
in the natural way, i.e.
Z t ³ ´ µZ t ³ ´ Z t ³ ´ ¶
f s, x(s) ds := f 1 s, x(s) ds, . . . , f 2 s, x(s) ds .
t0 t0 t0

The proof of equivalence is essentially the proof in Section 13.8 for the single
equation case, applied to each component separately.
Solutions of the integral equation are precisely the fixed points of the
operator T , where
Z t ³ ´
(T x)(t) = x0 + f s, x(s) ds t ∈ [t0 − h, t0 + h].
t0

As is the proof of Theorem 13.9.1, T is a contraction map on

\
C ∗ ([t0 − h, t0 + h]; Rn ) = C([t0 − h, t0 + h]; Rn )
n o
x(t) : |x(t) − x0 | ≤ k for all t ∈ [t0 − h, t0 + h]

for some I and some k > 0, provided f is locally Lipschitz in x. This

is proved exactly as in Theorem 13.9.1. Thus the integral equation, and
hence the initial value problem has a unique solution in some time interval
containing t0 .
The analogue of Theorem 13.10.1 for global (or long-time) existence is
also valid, with the same proof.
Chapter 14

Fractals

So, naturalists observe, a flea

Hath smaller fleas that on him prey;
And these have smaller still to bite ’em;
And so proceed ad infinitum.
Jonathan Swift On Poetry. A Rhapsody [1733]

Big whorls have little whorls

which feed on their velocity;
And little whorls have lesser whorls,
and so on to viscosity.
Lewis Fry Richardson

Fractals are, loosely speaking, sets which

• have a fractional dimension;

• have certain self-similarity or scale invariance properties.

There is also a notion of a random (or probabilistic) fractal.

Until recently, fractals were considered to be only of mathematical in-
terest. But in the last few years they have been used to model a wide
range of mathematical phenomena—coastline patterns, river tributary pat-
terns, and other geological structures; leaf structure, error distribution in
electronic transmissions, galactic clustering, etc. etc. The theory has been
used to achieve a very high magnitude of data compression in the storage
and generation of computer graphics.
References include the book [Ma], which is written in a somewhat informal
style but has many interesting examples. The books [PJS] and [Ba] provide
an accessible discussion of many aspects of fractals and are quite readable.
The book [BD] has a number of good articles.

165
166

14.1 Examples

14.1.1 Koch Curve

A sequence of approximations A = A(0) , A(1) , A(2) , . . . , A(n) , . . . to the Koch
Curve (or Snowflake Curve) is sketched in the following diagrams.

The actual Koch curve K ⊂ R2 is the limit of these approximations in a

sense which we later make precise.
Notice that
A(1) = S1 [A] ∪ S2 [A] ∪ S3 [A] ∪ S4 [A],
where each Si : R2 → R2 , and Si equals a dilation with dilation ratio 1/3,
followed by a translation and a rotation. For example, S1 is the map given
by dilating with dilation ratio 1/3 about the fixed point P , see the diagram.
Fractals 167

S2 is obtained by composing this map with a suitable translation and then a

rotation through 600 in the anti-clockwise direction. Similarly for S3 and S4 .
Likewise,

A(2) = S1 [A(1) ] ∪ S2 [A(1) ] ∪ S3 [A(1) ] ∪ S4 [A(1) ].

In general,

A(n+1) = S1 [A(n) ] ∪ S2 [A(n) ] ∪ S3 [A(n) ] ∪ S4 [A(n) ].

Moreover, the Koch curve K itself has the property that

K = S1 [K] ∪ S2 [K] ∪ S3 [K] ∪ S4 [K].

This is quite plausible, and will easily follow after we make precise the limiting
process used to define K.

14.1.2 Cantor Set

We next sketch a sequence of approximations A = A(0) , A(1) , A(2) , . . . , A(n) , . . .
to the Cantor Set C.

We can think of C as obtained by first removing the open middle third

(1/3, 2/3) from [0, 1]; then removing the open middle third from each of the
two closed intervals which remain; then removing the open middle third from
each of the four closed interval which remain; etc.
More precisely, let

A == A(0) = [0, 1]
A(1) = [0, 1/3] ∪ [2/3, 1]
A(2) = [0, 1/9] ∪ [2/9, 1/3] ∪ [2/3, 7/9] ∪ [8/9, 1]
..
.
T∞
Let C = n=0 A(n) . Since C is the intersection of a family of closed sets, C
is closed.
168

Note that A(n+1) ⊂ A(n) for all n and so the A(n) form a decreasing family
of sets.
Consider the ternary expansion of numbers x ∈ [0, 1], i.e. write each
x ∈ [0, 1] in the form

a1 a2 an
x = .a1 a2 . . . an . . . = + 2 + ··· + n + ··· (14.1)
3 3 3

where an = 0, 1 or 2. Each number has either one or two such representations,

and the only way x can have two representations is if

x = .a1 a2 . . . an 222 . . . = .a1 a2 . . . an−1 (an +1)000 . . .

for some an = 0 or 1. For example, .210222 . . . = .211000 . . ..

Note the following:

1. x ∈ A(n) iff x has an expansion of the form (14.1) with each of a1 , . . . , an

taking the values 0 or 2.

2. It follows that x ∈ C iff x has an expansion of the form (14.1) with

every an taking the values 0 or 2.

3. Each endpoint of any of the 2n intervals associated with A(n) has an

expansion of the form (14.1) with each of a1 , . . . , an taking the values
0 or 2 and the remaining ai either all taking the value 0 or all taking
the value 2.

Next let
1 1
S1 (x) = x, S2 (x) = 1 + (x − 1).
3 3
Notice that S1 is a dilation with dilation ratio 1/3 and fixed point 0. Similarly,
S2 is a dilation with dilation ratio 1/3 and fixed point 1.
Then
A(n+1) = S1 [A(n) ] ∪ S2 [A(n) ].

Moreover,
C = S1 [C] ∪ S2 [C].

14.1.3 Sierpinski Sponge

The following diagrams show two approximations to the Sierpinski Sponge.

Fractals 169

The Sierpinski Sponge P is obtained by first drilling out from the closed
unit cube A = A(0) = [0, 1]×[0, 1]×[0, 1], the three open, square cross-section,
tubes

(1/3, 2/3) × (1/3, 2/3) × R,

(1/3, 2/3) × R × (1/3, 2/3),
170

R × (1/3, 2/3) × (1/3, 2/3).

The remaining (closed) set A = A(1) is the union of 20 small cubes (8 at the
top, 8 at the bottom, and 4 legs).
From each of these 20 cubes, we again remove three tubes, each of cross-
section equal to one-third that of the cube. The remaining (closed) set is
denoted by A = A(2) .
Repeating this process, we obtain A = A(0) , A(1) , A(2) , . . . , A(n) , . . .; a se-
quence of closed sets such that
A(n+1) ⊂ A(n) ,
for all n. We define \
P = A(n) .
n≥1
Notice that P is also closed, being the intersection of closed sets.

14.2 Fractals and Similitudes

Motivated by the three previous examples we make the following:

Definition 14.2.1 A fractal 1 in Rn is a compact set K such that

[
N
K= Si [K] (14.2)
i=1

for some finite family

S = {S1 , . . . , SN }
of similitudes Si : Rn → Rn .

Similitudes A similitude is any map S : Rn → Rn which is a composition

of dilations 2 , orthonormal transformations 3 , and translations4 .
Note that translations and orthonormal transformations preserve dis-
tances, i.e. |F (x) − F (y)| = |x − y| for all x, y ∈ Rn if F is such a map.
On the other hand, |D(x) − D(y)| = r|x − y| if D is a dilation with dilation
ratio r ≥ 05 . It follows that every similitude S has a well-defined dilation
ratio r ≥ 0, i.e.
|S(x) − S(y)| = r|x − y|,
for all x, y ∈ Rn .
1
The word fractal is often used to denote a wider class of sets, but with analogous
properties to those here.
2
A dilation with fixed point a and dilation ratio r ≥ 0 is a map D : Rn → Rn of the
form D(x) = a + r(x − a).
3
An orthonormal transformation is a linear transformation O : Rn → Rn such that
−1
O = Ot . In R2 and R3 , such maps consist of a rotation, possibly followed by a reflection.
4
A translation is a map T : Rn → Rn of the form T (x) = x + a.
5
There is no need to consider dilation ratios r < 0. Such maps are obtained by com-
posing a positive dilation with the orthonormal transformation −I, where I is the identity
map on Rn .
Fractals 171

Theorem 14.2.2 Every similitude S can be expressed in the form

S = D ◦ T ◦ O,

with D a dilation about 0, T a translation, and O an orthonormal transfor-

mation. In other words,
S(x) = r(Ox + a), (14.3)
for some r ≥ 0, some a ∈ Rn and some orthonormal transformation O.
Moreover, the dilation ratio of the composition of two similitudes is the
product of their dilation ratios.

Proof: Every map of type (14.3) is a similitude.

On the other hand, any dilation, translation or orthonormal transforma-
tion is clearly of type (14.3). To show that a composition of such maps is
also of type (14.3), it is thus sufficient to show that a composition of maps
of type (14.3) is itself of type (14.3). But
µ ³ ´ ¶ µ ³ ´¶
r1 O1 r2 (O2 x + a2 ) + a1 = r1 r2 O1 O2 x + O1 a2 + r2−1 a1 .

This proves the result, including the last statement of the Theorem.

14.3 Dimension of Fractals

A curve has dimension 1, a surface has dimension 2, and a “solid” object
has dimension 3. By the k-volume of a “nice” k-dimensional set we mean its
length if k = 1, area if k = 2, and usual volume if k = 3.
One can in fact define in a rigorous way the so-called Hausdorff dimension
of an arbitrary subset of Rn . The Hausdorff dimension is a real number h
with 0 ≤ h ≤ n. We will not do this here, but you will see it in a later course
in the context of Hausdorff measure. Here, we will give a simple definition of
dimension for fractals, which agrees with the Hausdorff dimension in many
important cases.
Suppose by way of motivation that a k-dimensional set K has the property

K = K1 ∪ · · · ∪ KN ,

where the sets Ki are “almost”6 disjoint. Suppose moreover, that

Ki = Si [K]

where each Si is a similitude with dilation ratio ri > 0. See the following
diagram for a few examples.
6
In the sense that the Hausdorff dimension of the intersection is less than k.
172

K1 K
2
K

K1 K2 K3
K

K1 K3
K1 K
2
K K6 K

K3 K
4

K1
K
K3
K K
2 4

Suppose K is one of the previous examples and K is k-dimensional. Since

dilating a k-dimensional set by the ratio r will multiply the k-volume by rk ,
it follows that
V = r1k V + · · · + rN
k
V,
where V is the k-volume of K. Assume V 6= 0, ∞, which is reasonable if V
is k-dimensional and is certainly the case for the examples in the previous
diagram. It follows
X
N
rik = 1. (14.4)
i=1

In particular, if r1 = . . . = rN = r, say, then

N rk = 1,

and so
log N
k= .
log 1/r
Thus we have a formula for the dimension k in terms of the number N of
“almost disjoint” sets Ki whose union is K, and the dilation ratio r used to
obtain each Ki from K.
Fractals 173

More generally, if the ri are not all equal, the dimension k can be deter-
mined from N and the ri as follows. Define
X
N
g(p) = rip .
i=1

Then g(0) = N (> 1), g is a strictly decreasing function (assuming 0 < ri <
1), and g(p) → 0 as p → ∞. It follows there is a unique value of p such that
g(p) = 1, and from (14.4) this value of p must be the dimension k.
The preceding considerations lead to the following definition:

Definition 14.3.1 Assume K ⊂ Rn is a compact set and

K = S1 [K] ∪ · · · ∪ Sn [K],

where the Si are similitudes with dilation ratios 0 < ri < 1. Then the
similarity dimension of K is the unique real number D such that

X
N
1= riD .
i=1

Remarks This is only a good definition if the sets Si [K] are “almost” dis-
joint in some sense (otherwise different decompositions may lead to different
values of D). In this case one can prove that the similarity dimension and the
Hausdorff dimension are equal. The advantage of the similarity dimension is
that it is easy to calculate.

Examples For the Koch curve,

1
N = 4, r = ,
3
and so
log 4
D= ≈ 1.2619 .
log 3
For the Cantor set,
1
N = 2, r = ,
3
and so
log 2
D= ≈ 0.6309 .
log 3
And for the Sierpinski Sponge,
1
N = 20, r = ,
3
and so
log 20
D= ≈ 2.7268 .
log 3
174

14.4 Fractals as Fixed Points

We defined a fractal in (14.2) to be a compact non-empty set K ⊂ Rn such
that
[
N
K= Si [K], (14.5)
i=1
for some finite family
S = {S1 , . . . , SN }
of similitudes Si : Rn → Rn .
The surprising result is that given any finite family S = {S1 , . . . , SN } of
similitudes with contraction ratios less than 1, there always exists a compact
non-empty set K such that (14.5) is true. Moreover, K is unique.
We can replace the similitudes Si by any contraction map (i.e. Lipschitz
map with Lipschitz constant less than 1)7 . The following Theorem gives the
result.

Theorem 14.4.1 (Existence and Uniqueness of Fractals) Let S

= {S1 , . . . , SN } be a family of contraction maps on Rn . Then there is a
unique compact non-empty set K such that
K = S1 [K] ∪ · · · ∪ SN [K]. (14.6)

Proof: For any compact set A ⊂ Rn , define

S(A) = S1 [A] ∪ · · · ∪ SN [A].
Then S(A) is also a compact subset of Rn 8 .
Let
K = {A : A ⊂ Rn , A 6= ∅, A compact}
denote the family of all compact non-empty subsets of Rn . Then S : K → K,
and K satisfies (14.6) iff K is fixed point of S.
In the next section we will define the Hausdorff metric dH on K, and
show that (K, dH ) is a complete metric space. Moreover, we will show that
S is a contraction mapping on K 9 , and hence has a unique fixed point K,
say. Thus there is a unique compact set K such that (14.6) is true.

A Computational Algorithm From the proof of the Contraction Map-

ping Theorem, we know that if A is any compact subset of Rn , then the
sequence10
A, S(A), S 2 (A), . . . , S k (A), . . .
7
The restriction to similitudes is only to ensure that the similarity and Hausdorff di-
mensions agree under suitable extra hypotheses.
8
Each Si [A] is compact as it is the continuous image of a compact set. Hence S(A) is
compact as it is a finite union of compact sets.
9
Do not confuse this with the fact that the Si are contraction mappings on Rn .
10
Define S 2 (A) := S(S(A)), S 3 (A) := S(S 2 (A)), etc.
Fractals 175

converges to the fractal K (in the Hausdorff metric).

The approximations to the Koch curve which are shown in Section (14.1.1)
were obtained by taking A = A(0) as shown there. We could instead have
taken A = [P, Q], in which case the A shown in the first approximation is
obtained after just one iteration.
The approximations to the Cantor set were obtained by taking A = [0, 1],
and to the Sierpinski sponge by taking A to be the unit cube.
Another convenient choice of A is the set consisting of the N fixed points
of the contraction maps {S1 , . . . , SN }. The advantage of this choice is that
the sets S k (A) are then subsets of the fractal K (exercise).

Variants on the Koch Curve Let K be the Koch curve. We have seen
how we can write
K = S1 [K] ∪ · · · ∪ S4 [K].

It is also clear that we can write

K = S1 [K] ∪ S2 [K]

for suitable other choices of similitudes S1 , S2 . Here S1 [K] is the left side of
the Koch curve, as shown in the next diagram, and S2 [K] is the right side.

The map S1 consists of a reflection in the P Q axis, followed by a dilation

about P with √ the appropriate dilation factor (which a simple calculation
shows to be 1/ 3), followed by a rotation about P such that the final image
of Q is R. Similarly, S2 is a reflection in the P Q axis, followed by a dilation
about Q, followed by a rotation about Q such that the final image of P is R.
The previous diagram was generated with a simple Fortran program by
the previous computational algorithm, using A = [P, Q], and taking 6 itera-
tions.
Simple variations on S = {S1 , S2 } give quite different fractals. If S2 is as
before, and S1 is also as before except that no reflection is performed, then
the following Dragon fractal is obtained:
176

If S1 , S2 are as for the Koch curve, except that no reflection is performed

in either case, then the following Brain fractal is obtained:

If S1 , S2 are as for the previous

√ case except that now S1 maps Q to
(−.15, .6) instead of to R = (0, 1/ 3) ≈ (0, .6), and S2 maps P to (.15, .6),
then the following Clouds are obtained:

An important point worth emphasising is that despite the apparent com-

plexity in the fractals we have just sketched, all the relevant information
is already encoded in the family of generating similitudes. And any such
similitude, as in (14.3), is determined by r "∈ (0, 1), a ∈ Rn#, and the n × n
cos θ − sin θ
orthogonal matrix O. If n = 2, then O = , i.e. O is a ro-
sin θ cos θ
Fractals 177

" #
cos θ − sin θ
tation by θ in an anticlockwise direction, or O = , i.e. O
− sin θ − cos θ
is a rotation by θ in an anticlockwise direction followed by reflection in the
x-axis.
For a given fractal it is often a simple matter to work “backwards” and
find a corresponding family of similitudes. One needs to find S1 , . . . , SN such
that
K = S1 [K] ∪ · · · ∪ SN [K].

If equality is only approximately true, then it is not hard to show that the
fractal generated by S1 , . . . , SN will be approximately equal to K 11 .
In this way, complicated structures can often be encoded in very efficient
ways. The point is to find appropriate S1 , . . . , SN . There is much applied
and commercial work (and venture capital!) going into this problem.

14.5 *The Metric Space of Compact Subsets

of Rn
Let K is the family of compact non-empty subsets of Rn .
In this Section we will define the Hausdorff metric dH on K, show that
(dH , K), is a complete metric space, and prove that the map S : K → K is
a contraction map with respect to dH . This completes the proof of Theo-
rem (14.4.1).
Recall that the distance from x ∈ Rn to A ⊂ Rn was defined (c.f. (9.1))
by
d(x, A) = inf d(x, a). (14.7)
a∈A

If A ∈ K, it follows from Theorem 9.4.2 that the sup is realised, i.e.

d(x, A) = d(x, a) (14.8)

for some a ∈ A. Thus we could replace inf by min in (14.7).

Definition 14.5.1 Let A ⊂ Rn and ² ≥ 0. Then for any ² > 0 the

²-enlargement of A is defined by

A² = {x ∈ Rn : d(x, A) ≤ ²} .

Hence from (14.8), x ∈ A² iff d(x, a) ≤ ² for some a ∈ A.

The following diagram shows the ²-enlargement A² of a set A.

11
In the Hausdorff distance sense, as we will discuss in Section 14.5.
178

Properties of the ²-enlargement

1. A ⊂ B ⇒ A² ⊂ B² .

2. A² is closed. (Exercise: check that the complement is open)

3. A0 is the closure of A. (Exercise)

4. A ⊂ A² for any ² ≥ 0, and A² ⊂ Aγ if ² ≤ γ. (Exercise)

5. \
A² = A δ . (14.9)
²>δ
T
To see this12 first note that Aδ ⊂ ²>δ A² , since Aδ ⊂ A² whenever
T
² > δ. On the other hand, if x ∈ ²>δ A² then x ∈ A² for all ² > δ.
Hence d(x, A) ≤ ² for all ² > δ, and so d(x, A) ≤ δ. That is, x ∈ Aδ .

We regard two sets A and B as being close to each other if A ⊂ B² and

B ⊂ A² for some small ². This leads to the following definition.

Definition 14.5.2 Let A, B ⊂ Rn . Then the (Hausdorff ) distance between

A and B is defined by

dH (A, B) = d(A, B) = inf {² : A ⊂ B² , B ⊂ A² } . (14.10)

We call dH (or just d), the Hausdorff metric on K.

We give some examples in the following diagrams.

12
The result is not completely obvious. Suppose we had instead defined A² =
{x ∈ Rn : d(x, A) < ²}. Let A = [0, 1] ⊂ R. With this changed definition we would
have A² = (−², 1 + ²), and so
\ \
A² = (−², 1 + ²) = [−δ, 1 + δ] 6= Aδ .
²>δ ²>δ
Fractals 179

Remark 1 It is easy to see that the three notions13 of d are consistent, in

the sense that d(x, y) = d(x, {y}) and d(x, y) = d({x}, {y}).

Remark 2 Let δ = d(A, B). Then A ⊂ B² for all ² > δ, and so A ⊂ Bδ

from (14.9). Similarly, B ⊂ Aδ . It follows that the inf in Definition 14.5.2 is
realised, and so we could there replace inf by min.
Notice that if d(A, B) = ², then d(a, B) ≤ ² for every a ∈ A. Similarly,
d(b, A) ≤ ² for every b ∈ B.
13
The distance between two points, the distance between a point and a set, and the
Hausdorff distance between two sets.
180

Elementary Properties of dH

1. (Exercise) If E, F, G, H ⊂ Rn then

d(E ∪ F, G ∪ H) ≤ max{d(E, G), d(F, H)}.

2. (Exercise) If A, B ⊂ Rn and F : Rn → Rn is a Lipschitz map with

Lipschitz constant λ, then

d(F [A], F [B]) ≤ λd(A, B).

The Hausdorff metric is not a metric on the set of all subsets of Rn . For
example, in R we have ³ ´
d (a, b), [a, b] = 0.
Thus the distance between two non-equal sets is 0. But if we restrict to
compact sets, the d is indeed a metric, and moreover it makes K into a
complete metric space.

Theorem 14.5.3 (K, d) is a complete metric space.

Proof: (a) We first prove the three properties of a metric from Defini-
tion 6.2.1. In the following, all sets are compact and non-empty.

1. Clearly d(A, B) ≥ 0. If d(A, B) = 0, then A ⊂ B0 and B ⊂ A0 . But

A0 = A and B0 = B since A and B are closed. This implies A = B.

2. Clearly d(A, B) = d(B, A), i.e. symmetry holds.

3. Finally, suppose d(A, C) = δ1 and d(C, B) = δ2 . We want to show
d(A, B) ≤ δ1 + δ2 , i.e. that the triangle inequality holds.
We first claim A ⊂ Bδ1 +δ2 . To see this consider any a ∈ A. Then
d(a, C) ≤ δ1 and so d(a, c) ≤ d1 for some c ∈ C (by (14.8)). Similarly,
d(c, b) ≤ δ2 for some b ∈ B. Hence d(a, b) ≤ δ1 + δ2 , and so a ∈ Bδ1 +δ2 ,
and so A ⊂ Bδ1 +δ2 , as claimed.
Similarly, B ⊂ Aδ1 +δ2 . Thus d(A, B) ≤ δ1 + δ2 , as required.

(b) Assume (Ai )i≥1 is a Cauchy sequence (of compact non-empty sets)
from K.
Let [
Cj = Ai ,
i≥j

for j = 1, 2, . . .. Then the C j are closed and bounded14 , and hence compact.
Moreover, the sequence (C j ) is decreasing, i.e.

Cj ⊂ Ck,
14
This follows from the fact that (Ak ) is a Cauchy sequence.
Fractals 181

if j ≥ k.
Let \
C= Cj.
j≥1

Then C is also closed and bounded, and hence compact.

Claim: Ak → C in the Hausdorff metric, i.e. d(Ai , C) → 0 as i → ∞.

Suppose that ² > 0. Choose N such that

j, k ≥ N ⇒ d(Aj , Ak ) ≤ ². (14.11)
We claim that
j ≥ N ⇒ d(Aj , C) ≤ ²,
i.e.
j ≥ N ⇒ C ⊂ Aj² . (14.12)
and
j ≥ N ⇒ Aj ⊂ C² (14.13)
To prove (14.12), note from (14.11) that if j ≥ N then
[
Ai ⊂ Aj² .
i≥j

Since Aj² is closed, it follows

[
Cj = Ai ⊂ Aj² .
i≥j

Since C ⊂ C j , this establishes (14.12).

To prove (14.13), assume j ≥ N and suppose
x ∈ Aj .
Then from (14.11), x ∈ Ak² if k ≥ j, and so
 
[ [
k≥j⇒x∈ Ai² ⊂  Ai  ⊂ C²k ,
i≥k i≥k ²
³S ´
where the first “⊂” follows from the fact Ai² ⊂ i≥k A
i
for each i ≥ k. For
²
each k ≥ j, we can then choose xk ∈ C k with
d(x, xk ) ≤ ². (14.14)
Since (xk )k≥j is a bounded sequence, there exists a subsequence converging
to y, say. For each set C k with k ≥ j, all terms of the sequence (xi )i≥j
beyond a certain term are members of C k . Hence y ∈ C k as C k is closed.
T
But C = k≥j C k , and so y ∈ C.
Since y ∈ C and d(x, y) ≤ ² from (14.14), it follows that
x ∈ C²
As x was an arbitrary member of Aj , this proves (14.13)
182

Recall that if S = {S1 , . . . , SN }, where the Si are contractions on Rn ,

then we defined S : K → K by S(K) = S1 [K] ∪ . . . ∪ SN [K].

Theorem 14.5.4 If S is a finite family of contraction maps on Rn , then

the corresponding map S : K → K is a contraction map (in the Hausdorff
metric).

Proof: Let S = {S1 , . . . , SN }, where the Si are contractions on Rn with

Lipschitz constants r1 , . . . , rN < 1.
Consider any A, B ∈ K. From the earlier properties of the Hausdorff
metric it follows
 
[ [
d(S(A), S(B)) = d  Si [A], Si [B]
1≤i≤N 1≤i≤N
≤ max d(Si [A], Si [B])
1≤i≤N
≤ max ri d(A, B),
1≤i≤N

Thus S is a contraction map with Lipschitz constant given by max{r1 , . . . , rn }.

14.6 *Random Fractals

There is also a notion of a random fractal. A random fractal is not a par-
ticular compact set, but is a probability distribution on K, the family of all
compact subsets (of Rn ).
One method of obtaining random fractals is to randomise the Computa-
tional Algorithm in Section 14.4.
As an example, consider the Koch curve. In the discussion, “Variants
on the Koch Curve”, in Section 14.4, we saw how the Koch curve could be
generated from two similitudes S1 and S2 applied to an initial compact set A.
We there took A to be the closed interval [P, Q], where P and Q were the
fixed points of S1 and S2 respectively. The construction can be randomised
by selecting S = {S1 , S2 } at each stage of the iteration according to some
probability distribution.
For example, assume that S1 is always a reflection in the P Q axis, fol-
lowed by a dilation about P and then followed by a rotation, such that the
image of Q is some point R. Assume that S2 is a reflection in the P Q axis,
followed by a dilation about Q and then followed by a rotation about Q,
such that the image of P is the same point R. Finally, assume that R is
chosen according to some probability distribution over R2 (this then gives
a probability distribution on the set of possible S). We √ have chosen R to
be normally distributed in < with mean position (0, 3/3) and variance
2

(.4, .5). The following are three different realisations.

Fractals 183

P
Q

P Q

P Q
184
Chapter 15

Compactness

15.1 Definitions
In Definition 9.3.1 we defined the notion of a compact subset of a metric
space. As noted following that Definition, the notion defined there is usually
called sequential compactness.
We now give another definition of compactness, in terms of coverings
by open sets (which applies to any topological space)1 . We will show that
compactness and sequential compactness agree for metric spaces. (There are
examples to show that neither implies the other in an arbitrary topological
space.)

Definition 15.1.1 A collection Xα ) of subsets of a set X is a cover or

covering of a subset Y of X if ∪α Xα supseteqY .

Definition 15.1.2 A subset K of a metric space (X, d) is compact if when-

ever [
K⊂ U
U ∈F

for some collection F of open sets from X, then

K ⊂ U1 ∪ . . . ∪ UN

for some U1 , . . . , UN ∈ F. That is, every open cover has a finite subcover.
If X is compact, we say the metric space itself is compact.

Remark Exercise: A subset K of a metric space (X, d) is compact in the

sense of the previous definition iff the induced metric space (K, d) is compact.
The main point is that U ⊂ K is open in (K, d) iff U = K ∩ V for some
V ⊂ X which is open in X.
1
You will consider general topological spaces in later courses.

185
186

Examples It is clear that if A ⊂ Rn and A is unbounded, then A is not

compact according to the definition, since
∞
[
A⊂ Bi (0),
i=1

but no finite number of such open balls covers A.

Also B1 (0) is not compact, as
∞
[
B1 (0) = B1−1/i (0),
i=2

and again no finite number of such open balls covers B1 (0).

In Example 1 of Section 9.3 we saw that the sequentially compact subsets
of Rn are precisely the closed bounded subsets of Rn . It then follows from
Theorem 15.2.1 below that the compact subsets of Rn are precisely the closed
bounded subsets of Rn .
In a general metric space, this is not true, as we will see.

There is an equivalent definition of compactness in terms of closed sets,

which is called the finite intersection property.

Theorem 15.1.3 A topological space X is compact iff for every family F of

closed subsets of X,
\
C = ∅ ⇒ C1 ∩ · · · ∩ CN = ∅ for some finite subfamily {C1 , . . . , CN } ⊂ F.
C∈F

Proof: The result follows from De Morgan’s laws (exercise).

15.2 Compactness and Sequential Compact-

ness
We now show that these two notions agree in a metric space.
The following is an example of a non-trivial proof. Think of the case that
X = [0, 1] × [0, 1] with the induced metric. Note that an open ball Br (a) in
X is just the intersection of X with the usual ball Br (a) in R2 .

Theorem 15.2.1 A metric space is compact iff it is sequentially compact.

Proof: First suppose that the metric space (X, d) is compact.

Let (xn )∞
n=1 be a sequence from X. We want to show that some subse-
quence converges to a limit in X.
Compactness 187

Let A = {xn }. Note that A may be finite (in case there are only a finite
number of distinct terms in the sequence).
(1) Claim: If A is finite, then some subsequence of (xn ) converges.
Proof: If A is finite there is only a finite number of distinct terms in the
sequence. Thus there is a subsequence of (xn ) for which all terms are equal.
This subsequence converges to the common value of all its terms.
(2) Claim: If A is infinite, then A has at least one limit point.
Proof: Assume A has no limit points.
It follows that A = A from Definition 6.3.4, and so A is closed by Theo-
rem 6.4.6.
It also follows that each a ∈ A is not a limit point of A, and so from
Definition 6.3.3 there exists a neighbourhood Va of a (take Va to be some
open ball centred at a) such that

Va ∩ A = {a}. (15.1)

In particular, [
X = Ac ∪ Va
a∈A

gives an open cover of X. By compactness, there is a finite subcover. Say

X = Ac ∪ Va1 ∪ · · · ∪ VaN , (15.2)

for some {a1 , . . . , aN } ⊂ A. But this is impossible, as we see by choosing a ∈

A with a 6= a1 , . . . , aN (remember that A is infinite), and noting from (15.1)
that a cannot be a member of the right side of (15.2). This establishes the
claim.
(3) Claim: If x is a limit point of A, then some subsequence of (xn )
converges to x2 .
Proof: Any neighbourhood of x contains an infinite number of points
from A (Proposition 6.3.5) and hence an infinite number of terms from the
sequence (xn ). Construct a subsequence (x0k ) from (xn ) so that, for each k,
d(x0k , x) < 1/k and x0k is a term from the original sequence (xn ) which occurs
later in that sequence than any of the finite number of terms x01 , . . . , x0k−1 .
Then (x0k ) is the required subsequence.
From (1), (2) and (3) we have established that compactness implies se-
quential compactness.

Next assume that (X, d) is sequentially compact.

2
From Theorem 7.5.1 there is a sequence (yn ) consisting of points from A (and hence of
terms from the sequence (xn )) converging to x; but this sequence may not be a subsequence
of (xn ) because the terms may not occur in the right order. So we need to be a little more
careful in order to prove the claim.
188

(4) Claim: 3
For each integer k there is a finite set {x1 , . . . , xN } ⊂ X
such that
x ∈ X ⇒ d(xi , x) < 1/k for some i = 1, . . . , N.

Proof: Choose x1 ; choose x2 so d(x1 , x2 ) ≥ 1/k; choose x3 so d(xi , x3 ) ≥

1/k for i = 1, 2; choose x4 so d(xi , x4 ) ≥ 1/k for i = 1, 2, 3; etc. This
procedure must terminate in a finite number of steps

For if not, we have an infinite sequence (xn ). By sequential

compactness, some subsequence converges and in particular is
Cauchy. But this contradicts the fact that any two members of
the subsequence must be distance at least 1/k apart.

Let x1 , . . . , xN be some such (finite) sequence of maximum length.

It follows that any x ∈ X satisfies d(xi , x) < 1/k for some i = 1, . . . , N .
For if not, we could enlarge the sequence x1 , . . . , xN by adding x, thereby
contradicting its maximality.
(5) Claim: There exists a countable dense4 subset of X.
Proof: Let Ak be the finite set of points constructed in (4). Let A =
S∞
k=1 Ak . Then A is countable. It is also dense, since if x ∈ X then there
exist points in A arbitrarily close to x; i.e. x is in the closure of A.
(6) Claim: Every open cover of X has a countable subcover5 .
Proof: Let F be a cover of X by open sets. For each x ∈ A (where
A is the countable dense set from (5)) and each rational number r > 0, if
Br (x) ⊂ U for some U ∈ F, choose one such set U and denote it by Ux,r . The
collection F ∗ of all such Ux,r is a countable subcollection of F. Moreover, we
claim it is a cover of X.
To see this, suppose y ∈ X and choose U ∈ F with y ∈ U . Choose
s > 0 so Bs (y) ⊂ U . Choose x ∈ A so d(x, y) < s/4 and choose a rational
number r so s/4 < r < s/2. Then y ∈ Br (x) ⊂ Bs (y) ⊂ U . In particular,
Br (x) ⊂ U ∈ F and so there is a set Ux,r ∈ F ∗ (by definition of F ∗ ).
Moreover, y ∈ Br (x) ⊂ Ux,r and so y is a member of the union of all sets
from F ∗ . Since y was an arbitrary member of X, F ∗ is a countable cover of
X.
(7) Claim: Every countable open cover of X has a finite subcover.
Proof: Let G be a countable cover of X by open sets, which we write as
G = {U1 , U2 , . . .}. Let
Vn = U1 ∪ · · · ∪ Un
3
The claim says that X is totally bounded, see Section 15.5.
4
A subset of a topological space is dense if its closure is the entire space. A topological
space is said to be separable if it has a countable dense subset. In particular, the reals are
separable since the rationals form a countable dense subset. Similarly Rn is separable for
any n.
5
This is called the Lindelöf property.
Compactness 189

for n = 1, 2, . . .. Notice that the sequence (Vn )∞

n=1 is an increasing sequence
of sets. We need to show that

X = Vn

for some n.
Suppose not. Then there exists a sequence (xn ) where xn 6∈ Vn for each n.
By assumption of sequential compactness, some subsequence (x0n ) converges
to x, say. Since G is a cover of X, it follows x ∈ UN , say, and so x ∈ VN . But
VN is open and so
x0n ∈ VN for n > M, (15.3)
for some M .
On the other hand, xk 6∈ Vk for all k, and so

xk 6∈ VN (15.4)

for all k ≥ N since the (Vk ) are an increasing sequence of sets.

From 15.3 and 15.4 we have a contradiction. This establishes the claim.
From (6) and (7) it follows that sequential compactness implies compact-
ness.

Exercise Use the definition of compactness in Definition 15.1.2 to simplify

the proof of Dini’s Theorem (Theorem 12.1.3).

15.3 *Lebesgue covering theorem

Definition 15.3.1 The diameter of a subset Y of a metric space (X, d) is

d(Y ) = sup{d(y, y 0 ) : y, y 0 ∈ Y } .

Note this is not necessarily the same as the diameter of the smallest ball
containing the set, however, Y ⊆ Bd(Y ) (y)l for any y ∈ Y .

Theorem 15.3.2 Suppose (Gα ) is a covering of a compact metric space

(X, d) by open sets. Then there exists δ > 0 such that any (non-empty)
subset Y of X whose diameter is less than δ lies in some Gα .

Proof: Supposing the result fails, there are non-empty subsets Cn ⊆ X

with d(Cn ) < n−1 each of which fails to lie in any single Gα . Taking xn ∈ Cn ,
(xn ) has a convergent subsequence, say, xnj → x. Since (Gα ) is a covering,
there is some α such that x ∈ Gα . Now Gα is open, so that B² (x) ⊆ Gα for
some ² > 0. But xnj ∈ B² (x) for all j sufficiently large. Thus for j so large
that nj > 2²−1 , we have Cnj ⊆ Bn−1 (xn)j ) ⊂ B² x, contrary to the definition
j
of (xnj ).
190

15.4 Consequences of Compactness

We review some facts:

1 As we saw in the previous section, compactness and sequential compactness

are the same in a metric space.

2 In a metric space, if a set is compact then it is closed and bounded. The

proof of this is similar to the proof of the corresponding fact for Rn
given in the last two paragraphs of the proof of Corollary 9.2.2. As an
exercise write out the proof.

3 In Rn if a set is closed and bounded then it is compact. This is proved in

Corollary 9.2.2, using the Bolzano Weierstrass Theorem 9.2.1.

4 It is not true in general that closed and bounded sets are compact. In
Remark 2 of Section 9.2 we see that the set

F := C[0, 1] ∩ {f : ||f ||∞ ≤ 1}

is not compact. But it is closed and bounded (exercise).

5 A subset of a compact metric space is compact iff it is closed. Exercise:

prove this directly from the definition of sequential compactness; and
then give another proof directly from the definition of compactness.

We also have the following facts about continuous functions and compact
sets:

6 If f is continuous and K is compact, then f [K] is compact (Theorem 11.5.1).

7 Suppose f : K → R, f is continuous and K is compact. Then f is bounded

above and below and has a maximum and minimum value. (Theo-
rem 11.5.2)

8 Suppose f : K → Y , f is continuous and K is compact. Then f is uni-

formly continuous. (Theorem 11.6.2)

It is not true in general that if f : X → Y is continuous, one-one and

onto, then the inverse of f is continuous.
For example, define the function

f : [0, 2π) → S 1 = {(cos θ, sin θ) : [0, 2π)} ⊂ R2

by
f (θ) = (cos θ, sin θ).
Then f is clearly continuous (assuming the functions cos and sin are continu-
ous), one-one and onto. But f −1 is not continuous, as we easily see by finding
a sequence xn (∈ S 1 ) → (1, 0) (∈ S 1 ) such that f −1 (xn ) 6→ f −1 ((1, 0)) = 0.
Compactness 191

S1

.xn (1,0)
.
.x2
x1

-1
f-1(x1)
. . f. (xn)
0 f-1(x2) 2π

Note that [0, 2π) is not compact (exercise: prove directly from the def-
inition of sequential compactness that it is not sequentially compact, and
directly from the definition of compactness that it is not compact).

Theorem 15.4.1 Let f : X → Y be continuous and bijective. If X is com-

pact then f is a homeomorphism.

Proof: We need to show that the inverse function f −1 : Y → X 6 is con-

tinuous. To do this, we need to show that the inverse image under f −1 of a
closed set C ⊂ X is closed in Y ; equivalently, that the image under f of C
is closed.
But if C is closed then it follows C is compact from remark 5 at the
beginning of this section; hence f [C] is compact by remark 6; and hence
f [C] is closed by remark 2. This completes the proof.

We could also give a proof using sequences (Exercise).

15.5 A Criterion for Compactness

We now give an important necessary and sufficient condition for a metric
space to be compact. This will generalise the Bolzano-Weierstrass Theorem,
Theorem 9.2.1. In fact, the proof of one direction of the present Theorem is
very similar.
The most important application will be to finding compact subsets of
C[a, b].

Definition 15.5.1 Let (X, d) be a metric space. A subset A ⊂ X is totally

bounded iff for every δ > 0 there exist a finite set of points x1 , . . . , xN ∈ X
such that
[
N
A⊂ Bδ (xi ).
i=1
6
The function f −1 exists as f is one-one and onto.
192

Remark If necessary, we can assume that the centres of the balls belong
to A.
To see this, first cover A by balls of radius δ/2, as in the Definition. Let
the centres be x1 , . . . , xN . If the ball Bδ/2 (xi ) contains some point ai ∈ A,
then we replace the ball by the larger ball Bδ (ai ) which contains it. If Bδ/2 (xi )
contains no point from A then we discard this ball. In this way we obtain a
finite cover of A by balls of radius δ with centres in A.

Remark In any metric space, “totally bounded” implies “bounded”. For if

S
A⊂ N i=1 Bδ (xi ), then A ⊂ BR (x1 ) where R = maxi d(xi , x1 ) + δ.

In Rn , we also have that “bounded” implies “totally bounded”. To see

this in R2 , cover the bounded set A by a finite square lattice with grid size
δ. Then A is covered√by the finite number of√open balls with centres at the
vertices and radius δ 2. In Rn take radius δ n.

Note that as the dimension n increases, the number of vertices in a grid

of total side L is of the order (L/δ)n .
It is not true in a general metric space that “bounded” implies “totally
bounded”. The problem, as indicated roughly by the fact that (L/δ)n → ∞
as n → ∞, is that the number of balls of radius δ necessary to cover may be
infinite if A is not a subset of a finite dimensional vector space.
In particular, the set of functions A = {fn }n≥1 in Remark 2 of Section 9.2
is clearly bounded. But it is not totally bounded, since the distance between
any two functions in A is 1, and so no finite number of balls of radius less
than 1/2 can cover A as any such ball can contain at most one member of A.

In the following theorem, first think of the case X = [a, b].

Theorem 15.5.2 A metric space X is compact iff it is complete and totally

bounded.

Proof: (a) First assume X is compact.

In order to prove X is complete, let (xn ) be a Cauchy sequence from
X. Since compactness in a metric space implies sequential compactness by
Compactness 193

Theorem 15.2.1, a subsequence (x0k ) converges to some x ∈ X. We claim the

original sequence also converges to x.

This follows from the fact that

d(xn , x) ≤ d(xn , x0k ) + d(x0k , x).
Given ² > 0, first use convergence of (x0k ) to choose N1 so that
d(x0k , x) < ²/2 if k ≥ N1 . Next use the fact (xn ) is Cauchy to
choose N2 so d(xn , x0k ) < ²/2 if k, n ≥ N2 . Hence d(xn , x) < ² if
n ≥ max{N1 , N2 }.

That X is totally bounded follows from the observation that the set of all
balls Bδ (x), where x ∈ X, is an open cover of X, and so has a finite subcover
by compactness of A.

(b) Next assume X is complete and totally bounded.

Let (xn ) be a sequence from X, which for convenience we rewrite as (x(1)
n ).

Using total boundedness, cover X by a finite number of balls of radius 1.

Then at least one of these balls must contain an (infinite) subsequence of
(x(1) (2)
n ). Denote this subsequence by (xn ).

Repeating the argument, cover X by a finite number of balls of radius

1/2. At least one of these balls must contain an (infinite) subsequence of
(x(2) (3)
n ). Denote this subsequence by (xn ).

Continuing in this way we find sequences

(1) (1) (1)
(x1 , x2 , x3 , . . .)
(2) (2) (2)
(x1 , x2 , x3 , . . .)
(3) (3) (3)
(x1 , x2 , x3 , . . .)
..
.
where each sequence is a subsequence of the preceding sequence and the
terms of the ith sequence are all members of some ball of radius 1/i.
(i)
Define the (diagonal) sequence (yi ) by yi = xi for i = 1, 2, . . .. This is a
subsequence of the original sequence.
Notice that for each i, the terms yi , yi+1 , yi+2 , . . . are all members of the
ith sequence and so lie in a ball of radius 1/i. It follows that (yi ) is a Cauchy
sequence. Since X is complete, it follows (yi ) converges to a limit in X.
This completes the proof of the theorem, since (yi ) is a subsequence of
the original sequence (xn ).

The following is a direct generalisation of the Bolzano-Weierstrass theo-

rem.

Corollary 15.5.3 A subset of a complete metric space is compact iff it is

closed and totally bounded.
194

Proof: Let X be a complete metric space and A be a subset.

If A is closed (in X) then A (with the induced metric) is complete, by
the generalisation following Theorem 8.2.2. Hence A is compact from the
previous theorem.
If A is compact, then A is complete and totally bounded from the previous
theorem. Since A is complete it must be closed7 in X.

15.6 Equicontinuous Families of Functions

Throughout this Section you should think of the case X = [a, b] and Y = R.
We will use the notion of equicontinuity in the next Section in order to
give an important criterion for a family F of continuous functions to be
compact (in the sup metric).

Definition 15.6.1 Let (X, d) and (Y, ρ) be metric spaces. Let F be a family
of functions from X to Y .
Then F is equicontinuous at the point x ∈ X if for every ² > 0 there
exists δ > 0 such that

d(x, x0 ) < δ ⇒ ρ(f (x), f (x0 )) < ²

for all f ∈ F. The family F is equicontinuous if it is equicontinuous at every

x ∈ X.
F is uniformly equicontinuous on X if for every ² > 0 there exists δ > 0
such that
d(x, x0 ) < δ ⇒ ρ(f (x), f (x0 )) < ²
for all x ∈ X and all f ∈ F.

Remarks

1. The members of an equicontinuous family of functions are clearly con-

tinuous; and the members of a uniformly equicontinuous family of func-
tions are clearly uniformly continuous.

2. In case of equicontinuity, δ may depend on ² and x but not on the

particular function f ∈ F. For uniform equicontinuity, δ may depend
on ², but not on x or x0 (provided d(x, x0 ) < δ) and not on f .

3. The most important of these concepts is the case of uniform equiconti-

nuity.
7
To see this suppose (xn ) ⊂ A and xn → x ∈ X. Then (xn ) is Cauchy, and so by
completeness has a limit x0 ∈ A. But then in X we have xn → x0 as well as xn → x. By
uniqueness of limits in X it follows x = x0 , and so x ∈ A.
Compactness 195

Example 1 Let LipM (X; Y ) be the set of Lipschitz functions f : X → Y

with Lipschitz constant at most M . The family LipM (X; Y ) is uniformly
equicontinuous on the set X. This is easy to see since we can take δ = ²/M
in the Definition. Notice that δ does not depend on either x or on the
particular function f ∈ LipM (X; Y ).

Example 2 The family of functions fn (x) = xn for n = 1, 2, . . . and x ∈

[0, 1] is not equicontinuous at 1. To see this just note that if x < 1 then
|fn (1) − fn (x)| = 1 − xn > 1/2, say
for all sufficiently large n. So taking ² = 1/2 there is no δ > 0 such that
|1 − x| < δ implies |fn (1) − fn (x)| < 1/2 for all n.

On the other hand, this family is equicontinuous at each a ∈ [0, 1). In

fact it is uniformly equicontinuous on any interval [0, b] provided b < 1.
To see this, note
|fn (a) − fn (x)| = |an − xn | = |f 0 (ξ)| |a − x| = nξ n−1 |a − x|
for some ξ between x and a. If a, x ≤ b < 1, then ξ ≤ b, and so nξ n−1 is
bounded by a constant c(b) that depends on b but not on n (this is clear
since nξ n−1 ≤ nbn−1 , and nbn−1 → 0 as n → ∞; so we can take c(b) =
maxn≥1 nbn−1 ). Hence
|fn (1) − fn (x)| < ²
provided
²
|a − x| < .
c(b)

Exercise: Prove that the family in Example 2 is uniformly equicontinuous

on [0, b] (if b < 1) by finding a Lipschitz constant independent of n and using
the result in Example 1.

Example 3 In the first example, equicontinuity followed from the fact that
the families of functions had a uniform Lipschitz bound.
More generally, families of Hölder continuous functions with a fixed ex-
ponent α and a fixed constant M (see Definition 11.3.2) are also uniformly
equicontinuous. This follows from the fact that in the definition of uniform
³ ´1/α
equicontinuity we can take δ = ² M
.
196

We saw in Theorem 11.6.2 that a continuous function on a compact metric

space is uniformly continuous. Almost exactly the same proof shows that
an equicontinuous family of functions defined on a compact metric space is
uniformly equicontinuous.

Theorem 15.6.2 Let F be an equicontinuous family of functions f : X → Y ,

where (X, d) is a compact metric space and (Y, ρ) is a metric space. Then F
is uniformly equicontinuous.

Proof: Suppose ² > 0. For each x ∈ X there exists δx > 0 (where δx may
depend on x as well as ²) such that

x0 ∈ Bδx (x) ⇒ ρ(f (x), f (x0 )) < ²

for all f ∈ F.
The family of all balls B(x, δx /2) = Bδx /2 (x) forms an open cover of X.
By compactness there is a finite subcover B1 , . . . , BN by open balls with
centres x1 , . . . , xn and radii δ1 /2 = δx1 /2, . . . , δN /2 = δxN /2, say.
Let
δ = min{δ1 , . . . , δN }.
Take any x, x0 ∈ X with d(x, x0 ) < δ/2.

Then d(xi , x) < δi /2 for some xi since the balls Bi = B(xi , δi /2) cover X.
Moreover,
d(xi , x0 ) ≤ d(xi , x) + d(x, x0 ) < δi /2 + δ/2 ≤ δi .
In particular, both x, x0 ∈ B(xi , δi ).
It follows that for all f ∈ F,

ρ(f (x), f (x0 )) ≤ ρ(f (x), f (xi )) + ρ(f (xi ), f (x0 ))

< ² + ² = 2².

Since ² is arbitrary, this proves F is a uniformly equicontinuous family of

functions.
Compactness 197

15.7 Arzela-Ascoli Theorem

Throughout this Section you should think of the case X = [a, b] and Y = R.

Theorem 15.7.1 (Arzela-Ascoli) Let (X, d) be a compact metric space

and let C(X; Rn ) be the family of continuous functions from X to Rn . Let
F be any subfamily of C(X; Rn ) which is closed, uniformly bounded 8 and
uniformly equicontinuous. Then F is compact in the sup metric.

Remarks

1. Recall from Theorem 15.6.2 that since X is compact, we could replace

uniform equicontinuity by equicontinuity in the statement of the The-
orem.

2. Although we do not prove it now, the converse of the theorem is also

true. That is, F is compact iff it is closed, uniformly bounded, and
uniformly equicontinuous.

3. The Arzela-Ascoli Theorem is one of the most important theorems in

Analysis. It is usually used to show that certain sequences of functions
have a convergent subsequence (in the sup norm). See in particular the
next section.

Example 1 Let CM,K α

(X; Rn ) denote the family of Hölder continuous func-
tions f : X → Rn with exponent α and constant M (as in Definition 11.3.2),
which also satisfy the uniform bound

|f (x)| ≤ K for all x ∈ X.

Claim: CM,K
α
(X; Rn ) is closed, uniformly bounded and uniformly equicon-
tinuous, and hence compact by the Arzela-Ascoli Theorem.
We saw in Example 3 of the previous Section that CM,K
α
(X; Rn ) is equicon-
tinuous.
Boundedness is immediate, since the distance from any f ∈ CM,K
α
(X; Rn )
to the zero function is at most K (in the sup metric).
In order to show closure in C(X; Rn ), suppose that

fn ∈ CM,K
α
(X; Rn )

for n = 1, 2, . . ., and

fn → f uniformly as n → ∞,

(uniform convergence is just convergence in the sup metric). We know f is

continuous by Theorem 12.3.1. We want to show f ∈ CM,K
α
(X; Rn ).
8
That is, bounded in the sup metric.
198

We first need to show

|f (x)| ≤ K for each x ∈ X.
But for any x ∈ X we have |fn (x)| ≤ K, and so the result follows by letting
n → ∞.
We also need to show that
|f (x) − f (y)| ≤ M |x − y|α for all x, y ∈ X.
But for any fixed x, y ∈ X this is true with f replaced by fn , and so is true
for f as we see by letting n → ∞.
This completes the proof that CM,K
α
(X; Rn ) is closed in C(X; Rn ).

Example 2 An important case of the previous example is X = [a, b], Rn =

R, and F = LipM,K [a, b] (the class of real-valued Lipschitz functions with
Lipschitz constant at most M and uniform bound at most K).
You should keep this case in mind when reading the proof of the Theorem.

Remark The Arzela-Ascoli Theorem implies that any sequence from the
class LipM,K [a, b] has a convergent subsequence. This is not true for the set
CK [a, b] of all continuous functions f from C[a, b] merely satisfying sup |f | ≤
K. For example, consider the sequence of functions (fn ) defined by
fn (x) = sin nx x ∈ [0, 2π].

It seems clear that there is no convergent subsequence. More precisely,

one can show that for any m 6= n there exists x ∈ [0, 2π] such that sin mx >
1/2, sin nx < −1/2, and so du (fn , fm ) > 1 (exercise). Thus there is no
uniformly convergent subsequence as the distance (in the sup metric) between
any two members of the sequence is at least 1.
If instead we consider the sequence of functions (gn ) defined by
1
gn (x) = sin nx x ∈ [0, 2π],
n
Compactness 199

then the absolute value of the derivatives, and hence the Lipschitz constants,
are uniformly bounded by 1. In this case the entire sequence converges
uniformly to the zero function, as is easy to see.

Proof of Theorem We need to prove that F is complete and totally

bounded.
(1) Completeness of F.
We know that C(X; Rn ) is complete from Corollary 12.3.4. Since F is
a closed subset, it follows F is complete as remarked in the generalisation
following Theorem 8.2.2.

(2) Total boundedness of F.

Suppose δ > 0.
We need to find a finite set S of functions in C(X; Rn ) such that for any
f ∈ F,

there exists some g ∈ S satisfying max |f − g| < δ. (15.5)

From boundedness of F there is a finite K such that |f (x)| ≤ K for all

x ∈ X and f ∈ F.
By uniform equicontinuity choose δ1 > 0 so that

d(u, v) < δ1 ⇒ |f (u) − f (v)| < δ/4

for all u, v ∈ X and all f ∈ F.

Next by total boundedness of X choose a finite set of points x1 , . . . , xp ∈
X such that for any x ∈ X there exists at least one xi for which

d(x, xi ) < δ1

and hence

|f (x) − f (xi )| < δ/4. (15.6)

Also choose a finite set of points y1 , . . . , yq ∈ Rn so that if y ∈ Rn and

|y| ≤ K then there exists at least one yj for which

|y − yj | < δ/4. (15.7)

200

Consider the set of all functions α where

α : {x1 , . . . , xp } → {y1 , . . . , yq }.

Thus α is a function assigning to each of x1 , . . . , xp one of the values y1 , . . . , yq .

There are only a finite number (in fact q p ) possible such α. For each α, if
there exists a function f ∈ F satisfying

|f (xi ) − α(xi )| < δ/4 for i = 1, . . . , p,

then choose one such f and label it gα . Let S be the set of all gα . Thus

|gα (xi ) − α(xi )| < δ/4 for i = 1, . . . , p. (15.8)

Note that S is a finite set (with at most q p members).

Now consider any f ∈ F. For each i = 1, . . . , p, by (15.7) choose one of
the yj so that
|f (xi ) − yj | < δ/4.
Let α be the function that assigns to each xi this corresponding yj . Thus

|f (xi ) − α(xi )| < δ/4 (15.9)

for i = 1, . . . , p. Note that this implies the function gα defined previously

does exist.
We aim to show du (f, gα ) < δ.
Consider any x ∈ X. By (15.6) choose xi so

d(x, xi ) < δ1 . (15.10)

Compactness 201

Then
|f (x) − gα (x)| ≤ |f (x) − f (xi )| + |f (xi ) − α(xi )|
+|α(xi ) − gα (xi )| + +|gα (xi ) − gα (x)|
δ
≤ 4× from (15.6), (15.9), (15.10) and (15.8)
4
< δ.
This establishes (15.5) since x was an arbitrary member of X. Thus F is
totally bounded.

15.8 Peano’s Existence Theorem

In this Section we consider the initial value problem
x0 (t) = f (t, x(t)),
x(t0 ) = x0 .
For simplicity of notation we consider the case of a single equation, but ev-
erything generalises easily to a system of equations.
Suppose f is continuous, and locally Lipschitz with respect to the first
variable, in some open set U ⊂ R × R, where (t0 , x0 ) ∈ U . Then we saw in
the chapter on Differential Equations that there is a unique solution in some
interval [t0 − h, t0 + h] (and the solution is C 1 ).
If f is continuous, but not locally Lipschitz with respect to the first vari-
able, then there need no longer be a unique solution, as we saw in Example 1
of the Differential Equations Chapter. The Example was
dx q
= |x|,
dt
x(0) = 0.

It turns out that this example is typical. Provided we at least assume

that f is continuous, there will always be a solution. However, it may not√be
unique. Such examples are physically reasonable. In the example, f (x) = x
may be determined by a one dimensional material whose properties do not
change smoothly from the region x < 0 to the region x > 0.
We will prove the following Theorem.

Theorem 15.8.1 (Peano) Assume that f is continuous in the open set

U ⊂ R × R. Let (t0 , x0 ) ∈ U . Then there exists h > 0 such that the
initial value problem
x0 (t) = f (t, x(t)),
x(t0 ) = x0 ,
has a C 1 solution for t ∈ [t0 − h, t0 + h].
202

We saw in Theorem 13.8.1 that every (C 1 ) solution of the initial value

problem is a solution of the integral equation
Z t ³ ´
x(t) = x0 + f s, x(s) ds
t0

(obtained by integrating the differential equation). And conversely, we saw

that every C 0 solution of the integral equation is a solution of the initial value
problem (and the solution must in fact be C 1 ). This Theorem only used the
continuity of f .
Thus Theorem 15.8.1 follows from the following Theorem.

Theorem 15.8.2 Assume f is continuous in the open set U ⊂ R × R. Let

(t0 , x0 ) ∈ U . Then there exists h > 0 such that the integral equation
Z t ³ ´
x(t) = x0 + f s, x(s) ds (15.11)
t0

has a C 0 solution for t ∈ [t0 − h, t0 + h].

Remark We saw in Theorem 13.9.1 that the integral equation does indeed
have a solution, assuming f is also locally Lipschitz in x. The proof used the
Contraction Mapping Principle. But if we merely assume continuity of f ,
then that proof no longer applies (if it did, it would also give uniqueness,
which we have just remarked is not the case).
In the following proof, we show that some subsequence of the sequence
of Euler polygons, first constructed in Section 13.4, converges to a solution
of the integral equation.

Proof of Theorem
(1) Choose h, k > 0 so that

Ah,k (t0 , x0 ) := {(t, x) : |t − t0 | ≤ h, |x − x0 | ≤ k} ⊂ U.

Since f is continuous, it is bounded on the compact set Ah,k (t0 , x0 ).

Choose M such that

|f (t, x)| ≤ M if (t, x) ∈ Ah,k (t0 , x0 ). (15.12)

By decreasing h if necessary, we will require

k
h≤ . (15.13)
M
Compactness 203

(2) (See the diagram for n = 3) For each integer n ≥ 1, let xn (t) be the
piecewise linear function, defined as in Section 13.4, but with step-size h/n.
More precisely, if t ∈ [t0 , t0 + h],

xn (t) = x0 + (t − t0 ) f (t0 , x0 )
" #
h
for t ∈ t0 , t0 +
n
à ! à !
³ h ´ ³ h ´ h ³ h ´
xn (t) = xn t0 + + t − t0 + f t0 + , xn t0 +
n n n n
" #
h h
for t ∈ t0 + , t0 + 2
n n
à ! à !
³ h´ ³ h´ h n³ h´
n n
x (t) = x t0 + 2 + t − t0 + 2 f t0 + 2 , x t0 + 2
n n n n
" #
h h
for t ∈ t0 + 2 , t0 + 3
n n
..
.

Similarly for t ∈ [t0 − h, t0 ].

(3) From (15.12) and (15.13), and as is clear from the diagram, | dtd xn (t)| ≤
M (except at the points t0 , t0 ± nh , t0 ± 2 nh , . . .). It follows (exercise) that xn
is Lipschitz on [t0 − h, t0 + h] with Lipschitz constant at most M .
In particular, since k ≥ M h, the graph of t 7→ xn (t) remains in the closed
rectangle Ah,k (t0 , x0 ) for t ∈ [t0 − h, t0 + h].

(4) From (3), the functions xn belong to the family F of Lipschitz func-
tions
f : [t0 − h, t0 + h] → R
204

such that
Lipf ≤ M
and
|f (t) − x0 | ≤ k for all t ∈ [t0 − h, t0 + h].
But F is closed, uniformly bounded, and uniformly equicontinuous, by the
same argument as used in Example 1 of Section 15.7. It follows from the
0
Arzela-Ascoli Theorem that some subsequence (xn ) of (xn ) converges uni-
formly to a function x ∈ F.
Our aim now is to show that x is a solution of (15.11).

(5) For each point (t, xn (t)) on the graph of xn , let P n (t) ∈ R2 be the
coordinates of the point at the left (right) endpoint of the corresponding line
segment if t ≥ 0 (t ≤ 0). More precisely
à ! " #
h ³ h´ h h
P (t) = t0 + (i − 1) , xn t0 + (i − 1)
n
if t ∈ t0 + (i − 1) , t0 + i
n n n n
for i = 1, . . . , n. A similar formula is true for t ≤ 0.
h ´
Notice that P n (t) is constant for t ∈ t0 + (i − 1) nh , t0 + i nh , (and in
particular P n (t) is of course not continuous in [t0 − h, t0 + h]).
h i
(6) Without loss of generality, suppose t ∈ t0 + (i − 1) nh , t0 + i nh . Then
from (5) and (3)
v
uà !2 à !2
u ³ h´ ³ h´
t
|P (t) − (t, x (t)| ≤
n n
t − t0 + i + xn (t) − xn t0 + i
n n
s
³ h ´2 ³ h ´2
≤ + M
n n
√ h
= 1 + M2 .
n
Thus |P n (t) − (t, xn (t))| → 0, uniformly in t, as n → ∞.

(7) It follows from the definitions of xn and P n , and is clear from the
diagram, that Z t
xn (t) = x0 + f (P n (s)) ds (15.14)
t0

for t ∈ [t0 − h, t0 + h]. Although P n (s) is not continuous in s, the previous

integral still exists (for example, we could define it by considering the integral
over the various segments on which P n (s) and hence f (P n (s)) is constant).

(8) Our intention now is to show (on passing to a suitable subsequence)

that xn (t) → x(t) uniformly, P n (t) → (t, x(t)) uniformly, and to use this
and (15.14) to deduce (15.11).

(9) Since f is continuous on the compact set Ah,k (t0 , x0 ), it is uniformly

continuous there by (8) of Section 15.4.
Compactness 205

Suppose ² > 0. By uniform continuity of f choose δ > 0 so that for any

two points P, Q ∈ Ah,k (t0 , x0 ), if |P − Q| < δ then |f (P ) − f (Q)| < ².
In order to obtain (15.11) from (15.14), we compute

|f (s, x(s)) − f (P n (s))| ≤ |f (s, x(s)) − f (s, xn (s))| + |f (s, xn (s)) − f (P n (s))|.

From (6), |(s, xn (s)) − P n (s)| < δ for all n ≥ N1 (say), independently of
0
s. From uniform convergence (4), |x(s) − xn (s)| < δ for all n0 ≥ N2 (say),
independently of s. By the choice of δ it follows
0
|f (s, x(s)) − f (P n (s))| < 2², (15.15)

for all n0 ≥ N := max{N1 , N2 }.

(10) From (4), the left side of (15.14) converges to the left side of (15.11),
0
for the subsequence (xn ).
From (15.15), the difference of the right sides of (15.14) and (15.11) is
0
bounded by 2²h for members of the subsequence (xn ) such that n0 ≥ N (²).
As ² is arbitrary, it follows that for this subsequence, the right side of (15.14)
converges to the right side of (15.11).
This establishes (15.11), and hence the Theorem.
206
Chapter 16

Connectedness

16.1 Introduction

One intuitive idea of what it means for a set S to be “connected” is that S

cannot be written as the union of two sets which do not “touch” one another.
We make this precise in Definition 16.2.1.
Another informal idea is that any two points in S can be connected by
a “path” which joins the points and which lies entirely in S. We make this
precise in definition 16.4.2.
These two notions are distinct, though they agree on open subsets of
n
R (Theorem 16.4.4 below.)

16.2 Connected Sets

Definition 16.2.1 A metric space (X, d) is connected if there do not exist

two non-empty disjoint open sets U and V such that X = U ∪ V .
The metric space is disconnected if it is not connected, i.e. if there exist
two non-empty disjoint open sets U and V such that X = U ∪ V .
A set S ⊂ X is connected (disconnected) if the metric subspace (S, d) is
connected (disconnected).

Remarks and Examples

1. In the following diagram, S is disconnected. On the other hand, T is

connected; in particular although T = T1 ∪ T2 , any open set contain-
ing T1 will have non-empty intersection with T2 . However, T is not
pathwise connected — see Example 3 in Section 16.4.

207
208

2. The sets U and V in the previous definition are required to be open

in X.
For example, let
A = [0, 1] ∪ (2, 3].
We claim that A is disconnected.
Let U = [0, 1] and V = (2, 3]. Then both these sets are open in the
metric subspace (A, d) (where d is the standard metric induced from R).
To see this, note that both U and V are the intersection of A with sets
which are open in R (see Theorem 6.5.3). It follows from the definition
that X is disconnected.

3. In the definition, the sets U and V cannot be arbitrary disjoint sets.

For example, we will see in Theorem 16.3.2 that R is connected. But
R = U ∪ V where U and V are the disjoint sets (−∞, 0] and (0, ∞)
respectively.

4. Q is disconnected. To see this write

³ √ ´ ³ √ ´
Q = Q ∩ (−∞, 2) ∪ Q ∩ ( 2, ∞) .

The following proposition gives two other definitions of connectedness.

Proposition 16.2.2 A metric space (X, d) is connected

1. iff there do not exist two non-empty disjoint closed sets U and V such
that X = U ∪ V ;

2. iff the only non-empty subset of X which is both open and closed1 is X
itself.

Proof: (1) Suppose X = U ∪ V where U ∩ V = ∅. Then U = X \ V and

V = X \ U . Thus U and V are both open iff they are both closed2 . The first
equivalence follows.
1
Such a set is called clopen.
2
Of course, we mean open, or closed, in X.
Connectedness 209

(2) In order to show the second condition is also equivalent to connected-

ness, first suppose that X is not connected and let U and V be the open sets
given by Definition 16.2.1. Then U = X \ V and so U is also closed. Since
U 6= ∅, X, (2) in the statement of the theorem is not true.
Conversely, if (2) in the statement of the theorem is not true let E ⊂ X
be both open and closed and E 6= ∅, X. Let U = E, V = X \ E. Then U
and V are non-empty disjoint open sets whose union is X, and so X is not
connected.

Example We saw before that if A = [0, 1] ∪ (2, 3] (⊂ R), then A is not

connected. The sets [0, 1] and (2, 3] are both open and both closed in A.

16.3 Connectedness in R
Not surprisingly, the connected sets in R are precisely the intervals in R.
We first need a precise definition of interval.

Definition 16.3.1 A set S ⊂ R is an interval if

a, b ∈ S and a < x < b ⇒ x ∈ S.

Theorem 16.3.2 S ⊂ R is connected iff S is an interval.

Proof: (a) Suppose S is not an interval. Then there exist a, b ∈ S and

there exists x ∈ (a, b) such that x 6∈ S.
Then ³ ´ ³ ´
S = S ∩ (−∞, x) ∪ S ∩ (x, ∞) .
Both sets on the right side are open in S, are disjoint, and are non-empty
(the first contains a, the second contains b). Hence S is not connected.

(b) Suppose S is an interval.

Assume that S is not connected. Then there exist nonempty sets U and
V which are open in S such that

S = U ∪ V, U ∩ V = ∅.

Choose a ∈ U and b ∈ V . Without loss of generality we may assume

a < b. Since S is an interval, [a, b] ⊂ S.
Let
c = sup([a, b] ∩ U ).
Since c ∈ [a, b] ⊂ S it follows c ∈ S, and so either c ∈ U or c ∈ V .
Suppose c ∈ U . Then c 6= b and so a ≤ c < b. Since c ∈ U and U is open,
there exists c0 ∈ (c, b) such that c0 ∈ U . This contradicts the definition of c
as sup([a, b] ∩ U ).
210

Suppose c ∈ V . Then c 6= a and so a < c ≤ b. Since c ∈ V and V is

open, there exists c00 ∈ (a, c) such that [c00 , c] ⊂ V . But this implies that c is
again not the sup. Thus we again have a contradiction.
Hence S is connected.

Remark There is no such simple chararacterisation in Rn for n > 1.

16.4 Path Connected Sets

Definition 16.4.1 A path connecting two points x and y in a metric space
(X, d) is a continuous function f : [0, 1] (⊂ R) → X such that f (0) = x and
f (1) = y.

Definition 16.4.2 A metric space (X, d) is path connected if any two points
in X can be connected by a path in X.
A set S ⊂ X is path connected if the metric subspace (S, d) is path
connected.

The notion of path connected may seem more intuitive than that of con-
nected. However, the latter is usually mathematically easier to work with.
Every path connected set is connected (Theorem 16.4.3). A connected
set need not be path connected (Example (3) below), but for open subsets
of Rn (an important case) the two notions of connectedness are equivalent
(Theorem 16.4.4).

Theorem 16.4.3 If a metric space (X, d) is path connected then it is con-

nected.

Proof: Assume X is not connected3 .

Thus there exist non-empty disjoint open sets U and V such that X =
U ∪V.
Choose x ∈ U , y ∈ V and suppose there is a path from x to y, i.e. suppose
there is a continuous function f : [0, 1] (⊂ R) → X such that f (0) = x and
f (1) = y.
Consider f −1 [U ], f −1 [V ] ⊂ [0, 1]. They are open (continuous inverse im-
ages of open sets), disjoint (since U and V are), non-empty (since 0 ∈ f −1 [U ],
1 ∈ f −1 [V ]), and [0, 1] = f −1 [U ] ∪ f −1 [V ] (since X = U ∪ V ). But this con-
tradicts the connectedness of [0, 1]. Hence there is no such path and so X is
not path connected.

3
We want to show that X is not path connected.
Connectedness 211

Examples

1. Br (x) ⊂ R2 is path connected and hence connected. Since for u, v ∈

Br (x) the path f : [0, 1] → R2 given by f (t) = (1 − t)u + tv is a path in
R2 connecting u and v. The fact that the path does lie in R2 is clear,
and can be checked from the triangle inequality (exercise).
The same argument shows that in any normed space the open balls
Br (x) are path connected, and hence connected. The closed balls {y :
d(y, x) ≤ r} are similarly path connected and hence connected.

2. A = R2 \ {(0, 0), (1, 0), ( 12 , 0) ( 13 , 0), . . . , ( n1 , 0), . . .} is path connected

(take a semicircle joining points in A) and hence connected.

3. Let

A = {(x, y) : x > 0 and y = sin x1 , or x = 0 and y ∈ [0, 1]}.

Then A is connected but not path connected (*exercise).

Theorem 16.4.4 Let U ⊂ Rn be an open set. Then U is connected iff it

is path connected.

Proof: From Theorem 16.4.3 it is sufficient to prove that if U is connected

then it is path connected.
Assume then that U is connected.
The result is trivial if U = ∅ (why? ). So assume U 6= ∅ and choose some
a ∈ U. Let

E = {x ∈ U : there is a path in U from a to x}.

We want to show E = U . Clearly E 6= ∅ since a ∈ E 4 . If we can show

that E is both open and closed, it will follow from Proposition 16.2.2(2) that
E = U 5.

To show that E is open, suppose x ∈ E and choose r > 0 such that

Br (x) ⊂ U . From the previous Example(1), for each y ∈ Br (x) there is a
path in Br (x) from x to y. If we “join” this to the path from a to x, it is
not difficult to obtain a path from a to y6 . Thus y ∈ E and so E is open.
4
A path joining a to a is given by f (t) = a for t ∈ [0, 1].
5
This is a very important technique for showing that every point in a connected set
has a given property.
6
Suppose f : [0, 1] → U is continuous with f (0) = a and f (1) = x, while g : [0, 1] → U
is continuous with g(0) = x and g(1) = y. Define
½
f (2t) 0 ≤ t ≤ 1/2
h(t) =
g(2t − 1) 1/2 ≤ t ≤ 1

Then it is easy to see that h is a continuous path in U from a to y (the main point is to
check what happens at t = 1/2).
212

To show that E is closed in U , suppose (xn )∞n=1 ⊂ E and xn → x ∈ U .

We want to show x ∈ E. Choose r > 0 so Br (x) ⊂ U . Choose n so
xn ∈ Br (x). There is a path in U joining a to xn (since xn ∈ E) and a path
joining xn to x (as Br (x) is path connected). As before, it follows there is a
path in U from a to x. Hence x ∈ E and so E is closed.

Since E is open and closed, it follows as remarked before that E = U ,

and so we are done.

16.5 Basic Results

Theorem 16.5.1 The continuous image of a connected set is connected.

Proof: Let f : X → Y , where X is connected.

Suppose f [X] is not connected (we intend to obtain a contradiction).
Then there exists E ⊂ f [X], E 6= ∅, f [X], and E both open and closed in
f [X]. It follows there exists an open E 0 ⊂ Y and a closed E 00 ⊂ Y such that

E = f [X] ∩ E 0 = f [X] ∩ E 0 .

In particular,
f −1 [E] = f −1 [E 0 ] = f −1 [E 00 ],
and so f −1 [E] is both open and closed in X. Since E 6= ∅, f [X] it follows
that f −1 [E] 6= ∅, X. Hence X is not connected, contradiction.
Thus f [X] is connected.

The next result generalises the usual Intermediate Value Theorem.

Corollary 16.5.2 Suppose f : X → R is continuous, X is connected, and

f takes the values a and b where a < b. Then f takes all values between a
and b.

Proof: By the previous theorem, f [X] is a connected subset of R. Then,

by Theorem 16.3.2, f [X] is an interval. Since a, b ∈ f [X] it then follows
c ∈ f [X] for any c ∈ [a, b].
Chapter 17

Differentiation of Real-Valued
Functions

17.1 Introduction

In this Chapter we discuss the notion of derivative (i.e. differential ) for func-
tions f : D (⊂ Rn ) → R. In the next chapter we consider the case for
functions f : D (⊂ Rn ) → Rn .
We can represent such a function (m = 1) by drawing its graph, as is done
in the first diagrams in Section 10.1 in case n = 1 or n = 2, or as is done
“schematically” in the second last diagram in Section 10.1 for arbitrary n.
In case n = 2 (or perhaps n = 3) we can draw the level sets, as is done in
Section 17.6.

Convention Unless stated otherwise, we will always consider functions

f : D(⊂ Rn ) → R where the domain D is open. This implies that for any
x ∈ D there exists r > 0 such that Br (x) ⊂ D.

x. r
D

Most of the following applies to more general domains D by taking one-

sided, or otherwise restricted, limits. No essentially new ideas are involved.

213
214

17.2 Algebraic Preliminaries

The inner product in Rn is represented by

y · x = y 1 x1 + . . . + y n xn

where y = (y 1 , . . . , y n ) and x = (x1 , . . . , xn ).

For each fixed y ∈ Rn the inner product enables us to define a linear
function
Ly = L : Rn → R
given by
L(x) = y · x.

Conversely, we have the following.

Proposition 17.2.1 For any linear function

L : Rn → R

there exists a unique y ∈ Rn such that

L(x) = y · x ∀x ∈ Rn . (17.1)

The components of y are given by y i = L(ei ).

Proof: Suppose L : Rn → R is linear. Define y = (y 1 , . . . , y n ) by

y i = L(ei ) i = 1, . . . , n.

Then

L(x) = L(x1 e1 + · · · + xn en )
= x1 L(e1 ) + · · · + xn L(en )
= x1 y 1 + · · · + xn y n
= y · x.

This proves the existence of y satisfying (17.1).

The uniqueness of y follows from the fact that if (17.1) is true for some
y, then on choosing x = ei it follows we must have

L(ei ) = y i i = 1, . . . , n.

Note that if L is the zero operator , i.e. if L(x) = 0 for all x ∈ Rn , then
the vector y corresponding to L is the zero vector.
Differential Calculus for Real-Valued Functions 215

17.3 Partial Derivatives

Definition 17.3.1 The ith partial derivative of f at x is defined by
∂f f (x + tei ) − f (x)
(x) = lim (17.2)
∂xi t→0 t
f (x1 , . . . , xi + t, . . . , xn ) − f (x1 , . . . , xi , . . . , xn )
= lim ,
t→0 t
provided the limit exists. The notation ∆i f (x) is also used.

L
. .
x x+tei

∂f
Thus (x) is just the usual derivative at t = 0 of the real-valued func-
∂xi
tion g defined by g(t) = f (x1 , . . . , xi + t, . . . , xn ). Think of g as being defined
along the line L, with t = 0 corresponding to the point x.

17.4 Directional Derivatives

Definition 17.4.1 The directional derivative of f at x in the direction v 6= 0
is defined by
f (x + tv) − f (x)
Dv f (x) = lim , (17.3)
t→0 t
provided the limit exists.

It follows immediately from the definitions that

∂f
(x) = Dei f (x). (17.4)
∂xi

.
. x+tv
x

D L
216

Note that Dv f (x) is just the usual derivative at t = 0 of the real-valued

function g defined by g(t) = f (x + tv). As before, think of the function g as
being defined along the line L in the previous diagram.
Thus we interpret Dv f (x) as the rate of change of f at x in the direction v;
at least in the case v is a unit vector.
Exercise: Show that Dαv f (x) = αDv f (x) for any real number α.

17.5 The Differential (or Derivative)

Motivation Suppose f : I (⊂ R) → R is differentiable at a ∈ I. Then
f 0 (a) can be used to define the best linear approximation to f (x) for x near
a. Namely:
f (x) ≈ f (a) + f 0 (a)(x − a). (17.5)

graph of
x |→ f(x)

graph of
f(x) . x |→ f(a)+f'(a)(x-a)
.
f(a)+f'(a)(x-a)

. .
a x

Note that the right-hand side of (17.5) is linear in x. (More precisely, the
right side is a polynomial in x of degree one.)
The error, or difference between the two sides of (17.5), approaches zero
as x → a, faster than |x − a| → 0. More precisely
¯ ³ ´¯ ¯ ³ ´¯
¯ ¯ ¯ f (x) − f (a) + f 0 (a)(x − a) ¯
¯f (x) − f (a) + f 0 (a)(x − a) ¯ ¯ ¯
= ¯ ¯
¯ ¯
|x − a| ¯ x−a ¯
¯ ¯
¯ f (x) − f (a) ¯
¯
= ¯ − f 0 (a)¯¯
¯ x−a ¯
→ 0 as x → a. (17.6)

We make this the basis for the next definition in the case n > 1.
Differential Calculus for Real-Valued Functions 217

Definition 17.5.1 Suppose f : D (⊂ Rn ) → R. Then f is differentiable at

a ∈ D if there is a linear function L : Rn → R such that
¯ ³ ´¯
¯ ¯
¯f (x) − f (a) + L(x − a) ¯
→ 0 as x → a. (17.7)
|x − a|
The linear function L is denoted by f 0 (a) or df (a) and is called the derivative
or differential of f at a. (We will see in Proposition 17.5.2 that if L exists,
it is uniquely determined by this definition.)

graph of
x |→ f(a)+L(x-a)

graph of
x |→ f(x)

|f(x) - (f(a)+L(x-a))|

x
a

The idea is that the graph

³ of x´ 7→ f (a) + L(x − a) is “tangent” to the
graph of f (x) at the point a, f (a) .
Notation: We write hdf (a), x − ai for L(x − a), and read this as “df at a
applied to x − a”. We think of df (a) as a linear transformation (or function)
which operates on vectors x − a whose “base” is at a.
The next proposition gives the connection between the differential op-
erating on a vector v, and the directional derivative in the direction corre-
sponding to v. In particular, it shows that the differential is uniquely defined
by Definition 17.5.1.
Temporarily, we let df (a) be any linear map satisfying the definition for
the differential of f at a.

Proposition 17.5.2 Let v ∈ Rn and suppose f is differentiable at a.

Then Dv f (a) exists and
hdf (a), vi = Dv f (a).
In particular, the differential is unique.
218

Proof: Let x = a + tv in (17.7). Then

¯ ³ ´¯
¯ ¯
¯f (a + tv) − f (a) + hdf (a), tvi ¯
lim = 0.
t→0 t
Hence
f (a + tv) − f (a)
lim − hdf (a), vi = 0.
t→0 t
Thus
Dv f (a) = hdf (a), vi
as required.

Thus hdf (a), vi is just the directional derivative at a in the direction v.

The next result shows df (a) is the linear map given by the row vector of
partial derivatives of f at a.

Corollary 17.5.3 Suppose f is differentiable at a. Then for any vector v,

X
n
∂f
hdf (a), vi = vi (a).
i=1 ∂xi

Proof:

hdf (a), vi = hdf (a), v 1 e1 + · · · + v n en i

= v 1 hdf (a), e1 i + · · · + v n hdf (a), en i
= v 1 De1 f (a) + · · · + v n Den f (a)
∂f ∂f
= v 1 1 (a) + · · · + v n n (a).
∂x ∂x

Example Let f (x, y, z) = x2 + 3xy 2 + y 3 z + z.

Then
∂f ∂f ∂f
hdf (a), vi = v1 (a) + v2 (a) + v3 (a)
∂x ∂y ∂z
= v1 (2a1 + 3a2 ) + v2 (6a1 a2 + 3a2 2 a3 ) + v3 (a2 3 + 1).
2

Thus df (a) is the linear map corresponding to the row vector (2a1 +3a2 2 , 6a1 a2 +
3a2 2 a3 , a2 3 + 1).
If a = (1, 0, 1) then hdf (a), vi = 2v1 + v3 . Thus df (a) is the linear map
corresponding to the row vector (2, 0, 1).
∂f
If a = (1, 0, 1) and v = e1 then hdf (1, 0, 1), e1 i = (1, 0, 1) = 2.
∂x
Differential Calculus for Real-Valued Functions 219

Rates of Convergence If a function ψ(x) has the property that

|ψ(x)|
→ 0 as x → a,
|x − a|

then we say “|ψ(x)| → 0 as x → a, faster than |x − a| → 0”. We write

o(|x − a|) for ψ(x), and read this as “little oh of |x − a|”.
If
|ψ(x)|
≤M ∀|x − a| < ²,
|x − a|
|ψ(x)|
for some M and some ² > 0, i.e. if is bounded as x → a, then we say
|x − a|
“|ψ(x)| → 0 as x → a, at least as fast as |x − a| → 0”. We write O(|x − a|)
for ψ(x), and read this as “big oh of |x − a|”.
For example, we can write

o(|x − a|) for |x − a|3/2 ,

and
O(|x − a|) for sin(x − a).

Clearly, if ψ(x) can be written as o(|x − a|) then it can also be written as
O(|x − a|), but the converse may not be true as the above example shows.

The next proposition gives an equivalent definition for the differential of

a function.

Proposition 17.5.4 If f is differentiable at a then

f (x) = f (a) + hdf (a), x − ai + ψ(x),

where ψ(x) = o(|x − a|).

Conversely, suppose

f (x) = f (a) + L(x − a) + ψ(x),

where L : Rn → R is linear and ψ(x) = o(|x − a|). Then f is differentiable

at a and df (a) = L.

Proof: Suppose f is differentiable at a. Let

³ ´
ψ(x) = f (x) − f (a) + hdf (a), x − ai .

Then
f (x) = f (a) + hdf (a), x − ai + ψ(x),
and ψ(x) = o(|x − a|) from Definition 17.5.1.
220

Conversely, suppose

f (x) = f (a) + L(x − a) + ψ(x),

where L : Rn → R is linear and ψ(x) = o(|x − a|). Then

³ ´
f (x) − f (a) + L(x − a) ψ(x)
= → 0 as x → a,
|x − a| |x − a|

and so f is differentiable at a and df (a) = L.

Remark The word “differential” is used in [Sw] in an imprecise, and dif-

ferent, way from here.

Finally we have:

Proposition 17.5.5 If f, g : D (⊂ Rn ) → R are differentiable at a ∈ D,

then so are αf and f + g. Moreover,

d(αf )(a) = αdf (a),

d(f + g)(a) = df (a) + dg(a).

Proof: This is straightforward (exercise) from Proposition 17.5.4.

The previous proposition corresponds to the fact that the partial deriva-
tives for f + g are the sum of the partial derivatives corresponding to f and
g respectively. Similarly for αf 1 .

17.6 The Gradient

Strictly speaking, df (a) is a linear operator on vectors in Rn (where, for
convenience, we think of these vectors as having their “base at a”).
We saw in Section 17.2 that every linear operator from Rn to R corre-
sponds to a unique vector in Rn . In particular, the vector corresponding to
the differential at a is called the gradient at a.

Definition 17.6.1 Suppose f is differentiable at a. The vector ∇f (a) ∈ Rn

(uniquely) determined by

∇f (a) · v = hdf (a), vi ∀v ∈ Rn ,

is called the gradient of f at a.

1
We cannot establish the differentiability of f + g (or αf ) this way, since the existence
of the partial derivatives does not imply differentiability.
Differential Calculus for Real-Valued Functions 221

Proposition 17.6.2 If f is differentiable at a, then

à !
∂f ∂f
∇f (a) = (a), . . . , (a) .
∂x1 ∂xn

Proof: It follows from Proposition 17.2.1 that the components of ∇f (a)

∂f
are hdf (a), ei i, i.e. (a).
∂xi

Example For the example in Section 17.5 we have

∇f (a) = (2a1 + 3a2 2 , 6a1 a2 + 3a2 2 a3 , a32 + 1),

∇f (1, 0, 1) = (2, 0, 1).

17.6.1 Geometric Interpretation of the Gradient

Proposition 17.6.3 Suppose f is differentiable at x. Then the directional
derivatives at x are given by

Dv f (x) = v · ∇f (x).

The unit vector v for which this is a maximum is v = ∇f (x)/|∇f (x)|

(assuming |∇f (x)| =
6 0), and the directional derivative in this direction is
|∇f (x)|.

Proof: From Definition 17.6.1 and Proposition 17.5.2 it follows that

∇f (x) · v = hdf (x), vi = Dv f (x)

This proves the first claim.

Now suppose v is a unit vector. From the Cauchy-Schwartz Inequal-
ity (5.19) we have
∇f (x) · v ≤ |∇f (x)|. (17.8)
By the condition for equality in (5.19), equality holds in (17.8) iff v is a
positive multiple of ∇f (x). Since v is a unit vector, this is equivalent to
v = ∇f (x)/|∇f (x)|. The left side of (17.8) is then |∇f (x)|.

17.6.2 Level Sets and the Gradient

Definition 17.6.4 If f : Rn → R then the level set through x is {y: f(y) = f(x) }.

For example, the contour lines on a map are the level sets of the height
function.
222

x12+x 22 = 2.5 x2

∇f(x)
x.

x1
x2

x12+x 22 = .8

x1

graph of f(x)=x12+x22 level sets of f(x)=x12+x 22

x2 x12 - x 22 = -.5

x12 - x 22 = 0 x12 - x 22 = 2

x1

x12 - x22 = 2

level sets of f(x) = x12 - x 22

(the graph of f looks like a saddle)

Definition 17.6.5 A vector v is tangent at x to the level set S through x if

Dv f (x) = 0.

This is a reasonable definition, since f is constant on S, and so the rate

of change of f in any direction tangent to S should be zero.

Proposition 17.6.6 Suppose f is differentiable at x. Then ∇f (x) is or-

thogonal to all vectors which are tangent at x to the level set through x.

Proof: This is immediate from the previous Definition and Proposition 17.6.3.

In the previous proposition, we say ∇f (x) is orthogonal to the level set

through x.
Differential Calculus for Real-Valued Functions 223

17.7 Some Interesting Examples

(1) An example where the partial derivatives exist but the other directional
derivatives do not exist.
Let
f (x, y) = (xy)1/3 .
Then
∂f
1. (0, 0) = 0 since f = 0 on the x-axis;
∂x
∂f
2. (0, 0) = 0 since f = 0 on the y-axis;
∂y
3. Let v be any vector. Then
f (tv) − f (0, 0)
Dv f (0, 0) = lim
t→0 t
t2/3 (v1 v2 )1/3
= lim
t→0 t
(v1 v2 )1/3
= lim .
t→0 t1/3
This limit does not exist, unless v1 = 0 or v2 = 0.

(2) An example where the directional derivatives at some point all exist, but
the function is not differentiable at the point.
Let 

 xy 2
(x, y) 6= (0, 0)
f (x, y) = x2 + y 4

 0 (x, y) = (0, 0)

Let v = (v1 , v2 ) be any non-zero vector. Then

f (tv) − f (0, 0)
Dv f (0, 0) = lim
t→0 t
t3 v1 v2 2
−0
t2 v1 2 + t4 v2 4
= lim
t→0 t
v1 v2 2
= lim 2
t→0 v1 + t2 v2 4
(
v2 2 /v1 v1 6= 0
= (17.9)
0 v1 = 0

Thus the directional derivatives Dv f (0, 0) exist for all v, and are given
by (17.9).
In particular
∂f ∂f
(0, 0) = (0, 0) = 0. (17.10)
∂x ∂y
224

But if f were differentiable at (0, 0), then we could compute any direc-
tional derivative from the partial drivatives. Thus for any vector v we would
have

Dv f (0, 0) = hdf (0, 0), vi

∂f ∂f
= v1 (0, 0) + v2 (0, 0)
∂x ∂y
= 0 from (17.10)

This contradicts (17.9).

(3) An Example where the directional derivatives at a point all exist, but
the function is not continuous at the point
Take the same example as in (2). Approach the origin along the curve
x = λ2 , y = λ. Then

λ4 1
lim f (λ2 , λ) = lim = .
λ→0 λ→0 2λ4 2

But if we approach the origin along any straight line of the form (λv1 , λv2 ),
then we can check that the corresponding limit is 0.
Thus it is impossible to define f at (0, 0) in order to make f continuous
there.

17.8 Differentiability Implies Continuity

Despite Example (3) in Section 17.7, we have the following result.

Proposition 17.8.1 If f is differentiable at a, then it is continuous at a.

Proof: Suppose f is differentiable at a. Then

X
n
∂f
f (x) = f (a) + (a)(xi − ai ) + o(|x − a|).
i=1 ∂xi

Since xi − ai → 0 and o(|x − a|) → 0 as x → a, it follows that f (x) → f (a)

as x → a. That is, f is continuous at a.

17.9 Mean Value Theorem and Consequences

Theorem 17.9.1 Suppose f is continuous at all points on the line segment
L joining a and a + h; and is differentiable at all points on L, except possibly
at the end points.
Differential Calculus for Real-Valued Functions 225

Then

f (a + h) − f (a) = hdf (x), hi (17.11)

Xn
∂f
= i
(x) hi (17.12)
i=1 ∂x

for some x ∈ L, x not an endpoint of L.

L
.
a+h = g(1)
.x
.
a = g(0)

Proof: Note that (17.12) follows immediately from (17.11) by Corollary 17.5.3.
Define the one variable function g by

g(t) = f (a + th).

Then g is continuous on [0,1] (being the composition of the continuous func-

tions t 7→ a + th and x 7→ f (x)). Moreover,

g(0) = f (a), g(1) = f (a + h). (17.13)

We next show that g is differentiable and compute its derivative.

If 0 < t < 1, then f is differentiable at a + th, and so
f (a + th + w) − f (a + th) − hdf (a + th), wi
0 = lim . (17.14)
|w|→0 |w|
Let w = sh where s is a small real number, positive or negative. Since
|w| = ±s|h|, and since we may assume h 6= 0 (as otherwise (17.11) is trivial),
we see from (17.14) that
³ ´
f (a + (t + s)h − f (a + th) − hdf (a + th), shi
0 = lim
s→0
às !
g(t + s) − g(t)
= lim − hdf (a + th), hi ,
s→0 s

using the linearity of df (a + th).

Hence g 0 (t) exists for 0 < t < 1, and moreover

g 0 (t) = hdf (a + th), hi. (17.15)

226

By the usual Mean Value Theorem for a function of one variable, applied
to g, we have
g(1) − g(0) = g 0 (t) (17.16)
for some t ∈ (0, 1).
Substituting (17.13) and (17.15) in (17.16), the required result (17.11)
follows.

If the norm of the gradient vector of f is bounded by M , then it is not

surprising that the difference in value between f (a) and f (a + h) is bounded
by M |h|. More precisely.

Corollary 17.9.2 Assume the hypotheses of the previous theorem and sup-
pose |∇f (x)| ≤ M for all x ∈ L. Then

|f (a + h) − f (a)| ≤ M |h|

Proof: From the previous theorem

|f (a + h) − f (a)| = |hdf (x), hi| for some x ∈ L

= |∇f (x) · h|
≤ |∇f (x)| |h|
≤ M |h|.

Corollary 17.9.3 Suppose Ω ⊂ Rn is open and connected and f : Ω → R.

Suppose f is differentiable in Ω and df (x) = 0 for all x ∈ Ω2 .
Then f is constant on Ω.

Proof: Choose any a ∈ Ω and suppose f (a) = α. Let

E = {x ∈ Ω : f (x) = α}.

Then E is non-empty (as a ∈ E). We will prove E is both open and closed
in Ω. Since Ω is connected, this will imply that E is all of Ω3 . This establishes
the result.
To see E is open 4 , suppose x ∈ E and choose r > 0 so that Br (x) ⊂ Ω.
If y ∈ Br (x), then from (17.11) for some u between x and y,

f (y) − f (x) = hdf (u), y − xi

= 0, by hypothesis.
2
Equivalently, ∇f (x) = 0 in Ω.
3
This is a standard technique for showing that all points in a connected set have a
certain property, c.f. the proof of Theorem 16.4.4.
4
Being open in Ω and being open in Rn is the same for subsets of Ω, since we are
assuming Ω is itself open in Rn .
Differential Calculus for Real-Valued Functions 227

Thus f (y) = f (x) (= α), and so y ∈ E.

Hence Br (x) ⊂ E and so E is open.
To show that E is closed in Ω, it is sufficient to show that E c ={y :
f (x) 6= α} is open in Ω.
From Proposition 17.8.1 we know that f is continuous. Since we have
E = f −1 [R \ {α}] and R \ {α} is open, it follows that E c is open in Ω.
c

Hence E is closed in Ω, as required.

Since E 6= ∅, and E is both open and closed in Ω, it follows E = Ω (as Ω
is connected).
In other words, f is constant (= α) on Ω.

17.10 Continuously Differentiable Functions

We saw in Section 17.7, Example (2), that the partial derivatives (and even
all the directional derivatives) of a function can exist without the function
being differentiable.
However, we do have the following important theorem:

Theorem 17.10.1 Suppose f : Ω (⊂ Rn ) → R where Ω is open. If the

partial derivatives of f exist and are continuous at every point in Ω, then f
is differentiable everywhere in Ω.

Remark: If the partial derivatives of f exist in some neighbourhood of,

and are continuous at, a single point, it does not necessarily follow that f is
differentiable at that point. The hypotheses of the theorem need to hold at
all points in some open set Ω.

Proof: We prove the theorem in case n = 2 (the proof for n > 2 is only
notationally more complicated).
Suppose that the partial derivatives of f exist and are continuous in Ω.
Then if a ∈ Ω and a + h is sufficiently close to a,
f (a1 + h1 , a2 + h2 ) = f (a1 , a2 )
+f (a1 + h1 , a2 ) − f (a1 , a2 )
+f (a1 + h1 , a2 + h2 ) − f (a1 + h1 , a2 )
∂f ∂f
= f (a1 , a2 ) + 1 (ξ 1 , a2 )h1 + 2 (a1 + h1 , ξ 2 )h2 ,
∂x ∂x
for some ξ 1 between a1 and a1 + h1 , and some ξ 2 between a2 and a2 + h2 . The
first partial derivative comes from applying the usual Mean Value Theorem,
for a function of one variable, to the function f (x1 , a2 ) obtained by fixing
a2 and taking x1 as a variable. The second partial derivative is similarly
obtained by considering the function f (a1 + h1 , x2 ), where a1 + h1 is fixed
and x2 is variable.
228

(a1+h1, a2+h2 )
.
. (a +h , ξ )
1 1 2
(ξ1, a2 )
. . .
(a 1, a2) (a 1+h1 , a 2)
Ω

Hence
∂f 1 2 1 ∂f
f (a1 + h1 , a2 + h2 ) = f (a1 , a2 ) + 1
(a , a )h + 2 (a1 , a2 )h2
à ∂x ∂x
!
∂f 1 2 ∂f 1 2
+ (ξ , a ) − 1 (a , a ) h1
∂x1 ∂x
à !
∂f 1 ∂f 1 2
+ 2
(a + h , ξ ) − 2 (a , a ) h2
1 2
∂x ∂x
1 2
= f (a , a ) + L(h) + ψ(h), say.

Here L is the linear map defined by

∂f 1 2 1 ∂f
L(h) = 1
(a , a )h + 2 (a1 , a2 )h2
∂x ∂x
· ¸" 1 #
∂f 1 2 ∂f 1 2 h
= (a , a ) (a , a ) .
∂x 1 ∂x 2 h2

Thus L is represented by the previous 1 × 2 matrix.

We claim that the error term
à ! à !
∂f 1 2 ∂f 1 2 ∂f 1 ∂f 1 2
ψ(h) = (ξ , a ) − (a , a ) h1
+ (a + h1 2
, ξ ) − (a , a ) h2
∂x1 ∂x1 ∂x2 ∂x2

can be written as o(|h|)

This follows from the facts:
∂f 1 2 ∂f 1 2
1. 1
(ξ , a ) → (a , a ) as h → 0 (by continuity of the partial deriva-
∂x ∂x1
tives),
∂f 1 ∂f 1 2
2. 2
(a + h1 , ξ 2 ) → (a , a ) as h → 0 (again by continuity of the
∂x ∂x2
partial derivatives),

3. |h1 | ≤ |h|, |h2 | ≤ |h|.

It now follows from Proposition 17.5.4 that f is differentiable at a, and

the differential of f is given by the previous 1×2 matrix of partial derivatives.
Since a ∈ Ω is arbitrary, this completes the proof.
Differential Calculus for Real-Valued Functions 229

Definition 17.10.2 If the partial derivatives of f exist and are continuous

in the open set Ω, we say f is a C 1 (or continuously differentiable) function
on Ω. One writes f ∈ C 1 (Ω).

It follows from the previous Theorem that if f ∈ C 1 (Ω) then f is indeed

differentiable in Ω. Exercise: The converse may not be true, give a simple
counterexample in R.

17.11 Higher-Order Partial Derivatives

∂f ∂f
Suppose f : Ω (⊂ Rn ) → R. The partial derivatives 1
, . . . , n , if they
∂x ∂x
exist, are also functions from Ω to R, and may themselves have partial deriva-
tives.
∂f
The jth partial derivative of is denoted by
∂xi

∂2f
or fij or Dij f.
∂xj ∂xi

If all first and second partial derivatives of f exist and are continuous in
5
Ω we write
f ∈ C 2 (Ω).

Similar remarks apply to higher order derivatives, and we similarly define

C q (Ω) for any integer q ≥ 0.
Note that
C 0 (Ω) ⊃ C 1 (Ω) ⊃ C 2 (Ω) ⊃ . . .

The usual rules for differentiating a sum, product or quotient of functions

of a single variable apply to partial derivatives. It follows that C k (Ω) is closed
under addition, products and quotients (if the denominator is non-zero).
The next theorem shows that for higher order derivatives, the actual
order of differentiation does not matter, only the number of derivatives with
respect to each variable is important. Thus

∂2f ∂ 2f
= ,
∂xi ∂xj ∂xj ∂xi
and so
∂3f ∂ 3f ∂ 3f
= = , etc.
∂xi ∂xj ∂xk ∂xj ∂xi ∂xk ∂xj ∂xk ∂xi
5
In fact, it is sufficient to assume just that the second partial derivatives are continuous.
For under this assumption, each ∂f /∂xi must be differentiable by Theorem 17.10.1 applied
to ∂f /∂xi . From Proposition 17.8.1 applied to ∂f /∂xi it then follows that ∂f /∂xi is
continuous.
230

Theorem 17.11.1 If f ∈ C 1 (Ω)6 and both fij and fji exist and are contin-
uous (for some i 6= j) in Ω, then fij = fji in Ω.
In particular, if f ∈ C 2 (Ω) then fij = fji for all i 6= j.

Proof: For notational simplicity we take n = 2. The proof for n > 2 is

very similar.
Suppose a ∈ Ω and suppose h > 0 is some sufficiently small real number.
Consider the second difference quotient defined by
µ
1 ³ 1 ´
A(h) = f (a + h, a2
+ h) − f (a1 2
, a + h)
h2 ¶
³ ´
− f (a1 + h, a2 ) − f (a1 , a2 ) (17.17)
1³ 2 ´
= g(a + h) − g(a2
) , (17.18)
h2
where
g(x2 ) = f (a1 + h, x2 ) − f (a1 , x2 ).

(a1, a2+h) (a 1+h, a2 +h)

A B

C D

a = (a1, a2) (a 1+h, a2)

A(h) = ( ( f(B) - f(A) ) - ( f(D) - f(C) ) ) / h2
= ( ( f(B) - f(D) ) - ( f(A) - f(C) ) ) / h2

From the definition of partial differentiation, g 0 (x2 ) exists and

∂f 1 ∂f
g 0 (x2 ) = 2
(a + h, x2 ) − 2 (a1 , x2 ) (17.19)
∂x ∂x

for a2 ≤ x ≤ a2 + h.
Applying the mean value theorem for a function of a single variable
to (17.18), we see from (17.19) that

1 0 2
A(h) = g (ξ ) some ξ 2 ∈ (a2 , a2 + h)
hà !
1 ∂f 1 ∂f 1 2
= (a + h, ξ ) − 2 (a , ξ ) .
2
(17.20)
h ∂x2 ∂x
6
As usual, Ω is assumed to be open.
Differential Calculus for Real-Valued Functions 231

∂f 1 2
Applying the mean value theorem again to the function (x , ξ ), with
∂x2
ξ 2 fixed, we see

∂2f
A(h) = (ξ 1 , ξ 2 ) some ξ 1 ∈ (a1 , a1 + h). (17.21)
∂x1 ∂x2

If we now rewrite (17.17) as

µ
1 ³ 1 ´
A(h) = 2 f (a + h, a2 + h) − f (a1 + h, a2 )
h ¶
³ ´
− f (a1 , a2 + h) − f (a1 + a2 ) (17.22)

and interchange the roles of x1 and x2 in the previous argument, we obtain

∂2f
A(h) = (η 1 , η 2 ) (17.23)
∂x2 ∂x1

for some η 1 ∈ (a1 , a1 + h), η 2 ∈ (a2 , a2 + h).

If we let h → 0 then (ξ 1 , ξ 2 ) and (η 1 , η 2 ) → (a1 , a2 ), and so from (17.21),
(17.23) and the continuity of f12 and f21 at a, it follows that

f12 (a) = f21 (a).

This completes the proof.

17.12 Taylor’s Theorem

If g ∈ C 1 [a, b], then we know
Z b
g(b) = g(a) + g 0 (t) dt
a

This is the case k = 1 of the following version of Taylor’s Theorem for a

function of one variable.

Theorem 17.12.1 (Taylor’s Formula; Single Variable, First Version)

Suppose g ∈ C k [a, b]. Then

1 00
g(b) = g(a) + g 0 (a)(b − a) + g (a)(b − a)2 + · · · (17.24)
2!
Z b
1 (b − t)k−1 (k)
+ g (k−1) (a)(b − a)k−1 + g (t) dt.
(k − 1)! a (k − 1)!
232

Proof: An elegant (but not obvious) proof is to begin by computing:

d ³ (k−1) ´
gϕ − g 0 ϕ(k−2) + g 00 ϕ(k−3) − · · · + (−1)k−1 g (k−1) ϕ
dt ³ ´ ³ ´ ³ ´
= gϕ(k) + g 0 ϕ(k−1) − g 0 ϕ(k−1) + g 00 ϕ(k−2) + g 00 ϕ(k−2) + g 000 ϕ(k−3) −
³ ´
· · · + (−1)k−1 g (k−1) ϕ0 + g (k) ϕ
= gϕ(k) + (−1)k−1 g (k) ϕ. (17.25)

Now choose
(b − t)k−1
ϕ(t) = .
(k − 1)!
Then
(b − t)k−2
ϕ0 (t) = (−1)
(k − 2)!
(b − t)k−3
ϕ00 (t) = (−1)2
(k − 3)!
..
.
(b − t)2
ϕ(k−3) (t) = (−1)k−3
2!
ϕ(k−2) (t) = (−1)k−2 (b − t)
ϕ(k−1) (t) = (−1)k−1
ϕk (t) = 0. (17.26)

Hence from (17.25) we have

à !
d (b − t)2 (b − t)k−1
(−1)k−1 g(t) + g 0 (t)(b − t) + g 00 (t) + · · · + g k−1 (t)
dt 2! (k − 1)!
(b − t)k−1
= (−1)k−1 g (k) (t) .
(k − 1)!

Dividing by (−1)k−1 and integrating both sides from a to b, we get

à !
0 (b − a)2
00 (b − a)k−1
g(b) − g(a) + g (a)(b − a) + g (a) + · · · + g (k−1) (a)
2! (k − 1)!
Z b
(b − t)k−1
= g (k) (t) dt.
a (k − 1)!
This gives formula (17.24).

Theorem 17.12.2 (Taylor’s Formula; Single Variable, Second Version)

Suppose g ∈ C k [a, b]. Then
1 00
g(b) = g(a) + g 0 (a)(b − a) + g (a)(b − a)2 + · · ·
2!
1 1
+ g (k−1) (a)(b − a)k−1 + g (k) (ξ)(b − a)k (17.27)
(k − 1)! k!
for some ξ ∈ (a, b).
Differential Calculus for Real-Valued Functions 233

Proof: We establish (17.27) from (17.24).

Since g (k) is continuous in [a, b], it has a minimum value m, and a maxi-
mum value M , say.
By elementary properties of integrals, it follows that
Z Z Z
b (b − t)k−1 b (b − t)k−1 b (b − t)k−1
m dt ≤ g (k) (t) dt ≤ M dt,
a (k − 1)! a (k − 1)! a (k − 1)!
i.e. Z b (b − t)k−1
g (k) (t) dt
a (k − 1)!
m≤ Z b ≤ M.
(b − t)k−1
dt
a (k − 1)!

By the Intermediate Value Theorem, g (k) takes all values in the range
[m, M ], and so the middle term in the previous inequality must equal g (k) (ξ)
for some ξ ∈ (a, b). Since
Z b (b − t)k−1 (b − a)k
dt = ,
a (k − 1)! k!
it follows Z b (b − t)k−1 (b − a)k (k)
g (k) (t) dt = g (ξ).
a (k − 1)! k!
Formula (17.27) now follows from (17.24).

Remark For a direct proof of (17.27), which does not involve any integra-
tion, see [Sw, pp 582–3] or [F, Appendix A2].

Taylor’s Theorem generalises easily to functions of more than one variable.

Theorem 17.12.3 (Taylor’s Formula; Several Variables)

Suppose f ∈ C k (Ω) where Ω ⊂ Rn , and the line segment joining a and a + h
is a subset of Ω.
Then
X
n
1 X n
f (a + h) = f (a) + Di f (a) hi + Dij f (a) hi hj + · · ·
i−1 2! i,j=1
1 X n
+ Di ...i f (a) hi1 · . . . · hik−1 + Rk (a, h)
(k − 1)! i1 ,···,ik−1 =1 1 k−1

where
X n Z 1
1
Rk (a, h) = (1 − t)k−1 Di1 ...ik f (a + th) dt
(k − 1)! i1 ,...,ik =1 0
1 X n
= Di1 ,...,ik f (a + sh) hi1 · . . . · hik for some s ∈ (0, 1).
k! i1 ,...,ik =1
234

Proof: First note that for any differentiable function F : D (⊂ Rn ) → R

we have
d Xn
F (a + th) = Di F (a + th) hi . (17.28)
dt i=1

This is just a particular case of the chain rule, which we will discuss later.
This particular version follows from (17.15) and Corollary 17.5.3 (with f
there replaced by F ).
Let
g(t) = f (a + th).
Then g : [0, 1] → R. We will apply Taylor’s Theorem for a function of one
variable to g.
From (17.28) we have
X
n
0
g (t) = Di f (a + th) hi . (17.29)
i−1

Differentiating again, and applying (17.28) to Di F , we obtain

 
X
n X
n
g 00 (t) =  Dij f (a + th) hj  hi
i=1 j=1
Xn
= Dij f (a + th) hi hj . (17.30)
i,j=1

Similarly
X
n
g 000 (t) = Dijk f (a + th) hi hj hk , (17.31)
i,j,k=1

etc. In this way, we see g ∈ C k [0, 1] and obtain formulae for the derivatives
of g.
But from (17.24) and (17.27) we have

1 00 1
g(1) = g(0) + g 0 (0) + g (0) + · · · + g (k−1) (0)
2! (k − 1)!
 Z 1

 1


 (1 − t)k−1 g (k) (t) dt
 (k − 1)! 0
+ or



 1 (k)

 g (s) some s ∈ (0, 1).
k!
If we substitute (17.29), (17.30), (17.31) etc. into this, we obtain the required
results.

Remark The first two terms of Taylor’s Formula give the best first order
approximation 7 in h to f (a + h) for h near 0. The first three terms give
7
I.e. constant plus linear term.
Differential Calculus for Real-Valued Functions 235

the best second order approximation 8 in h, the first four terms give the best
third order approximation, etc.
Note that the remainder term Rk (a, h) in Theorem 17.12.3 can be written
as O(|h|k ) (see the Remarks on rates of convergence in Section 17.5), i.e.

Rk (a, h)
is bounded as h → 0.
|h|k

This follows from the second version for the remainder in Theorem 17.12.3
and the facts:

1. Di1 ...ik f (x) is continuous, and hence bounded on compact sets,

2. |hi1 · . . . · hik | ≤ |h|k .

Example Let
f (x, y) = (1 + y 2 )1/2 cos x.
One finds the best second order approximation to f for (x, y) near (0, 1) as
follows.
First note that
f (0, 1) = 21/2 .
Moreover,

f1 = −(1 + y 2 )1/2 sin x; = 0 at (0, 1)

f2 = y(1 + y 2 )−1/2 cos x; = 2−1/2 at (0, 1)
f11 = −(1 + y 2 )1/2 cos x; = −21/2 at (0, 1)
f12 = −y(1 + y 2 )−1/2 sin x; = 0 at (0, 1)
f22 = (1 + y 2 )−3/2 cos x; = 2−3/2 at (0, 1).

Hence
³ ´
f (x, y) = 21/2 + 2−1/2 (y − 1) − 21/2 x2 + 2−3/2 (y − 1)2 + R3 (0, 1), (x, y) ,

where
³ ´ ³ ´ µ³ ´3/2 ¶
R3 (0, 1), (x, y) = O |(x, y) − (0, 1)|3 = O x2 + (y − 1)2 .

8
I.e. constant plus linear term plus quadratic term.
236
Chapter 18

Differentiation of
Vector-Valued Functions

18.1 Introduction
In this chapter we consider functions

f : D (⊂ Rn ) → Rn ,

with m ≥ 1. You should have a look back at Section 10.1.

We write
³ ´
f (x1 , . . . , xn ) = f 1 (x1 , . . . , xn ), . . . , f m (x1 , . . . , xn )

where
f i : D → R i = 1, . . . , m
are real -valued functions.

Example Let
f (x, y, z) = (x2 − y 2 , 2xz + 1).
Then f 1 (x, y, z) = x2 − y 2 and f 2 (x, y, z) = 2xz + 1.

Reduction to Component Functions For many purposes we can reduce

the study of functions f , as above, to the study of the corresponding real -
valued functions f 1 , . . . , f m . However, this is not always a good idea, since
studying the f i involves a choice of coordinates in Rn , and this can obscure
the geometry involved.
In Definitions 18.2.1, 18.3.1 and 18.4.1 we define the notion of partial
derivative, directional derivative, and differential of f without reference to the
component functions. In Propositions 18.2.2, 18.3.2 and 18.4.2 we show these
definitions are equivalent to definitions in terms of the component functions.

237
238

18.2 Paths in Rm
In this section we consider the case corresponding to n = 1 in the notation
of the previous section. This is an important case in its own right and also
helps motivates the case n > 1.

Definition 18.2.1 Let I be an interval in R. If f : I → Rn then the deriva-

tive or tangent vector at t is the vector
f (t + s) − f (t)
f 0 (t) = lim ,
s→0 s
provided the limit exists1 . In this case we say f is differentiable at t. If,
moreover, f 0 (t) 6= 0 then f 0 (t)/|f 0 (t)| is called the unit tangent at t.

Remark Although we say f 0 (t) is the tangent vector at t, we should really

think of f 0 (t) as a vector with its “base” at f (t). See the next diagram.
³ ´
Proposition 18.2.2 Let f (t) = f 1 (t), . . . , f m (t) . Then f is differentiable
at t iff f 1 , . . . , f m are differentiable at t. In this case
³ 0
´
f 0 (t) = f 1 (t), . . . , f m0 (t) .

Proof: Since
à !
f (t + s) − f (t) f 1 (t + s) − f 1 (t) f m (t + s) − f m (t)
= ,..., ,
s s s
The theorem follows by applying Theorem 10.4.4.
³ ´
Definition 18.2.3 If f (t) = f 1 (t), . . . , f m (t) then f is C 1 if each f i is C 1 .

We have the usual rules for differentiating the sum of two functions from I
to <m , and the product of such a function with a real valued function (exer-
cise: formulate and prove such a result). The following rule for differentiating
the inner product of two functions is useful.

Proposition 18.2.4 If f 1 , f 2 : I → Rn are differentable at t then

d³ ´ ³ ´ ³ ´
f 1 (t), f 2 (t) = f 01 (t), f 2 (t) + f 1 (t), f 02 (t) .
dt

Proof: Since
³ ´ X
m
f 1 (t), f 2 (t) = f1i (t)f2i (t),
i=1

the result follows from the usual rule for differentiation sums and products.

1
If t is an endpoint of I then one takes the corresponding one-sided limits.
Differential Calculus for Vector-Valued Functions 239

If f : I → Rn , we can think of f as tracing out a “curve” in Rn (we will

make this precise later). The terminology tangent vector is reasonable, as we
see from the following diagram. Sometimes we speak of the tangent vector
at f (t) rather than at t, but we need to be careful if f is not one-one, as in
the second figure.

f '(t1)
f(t 1)=f(t2)
a path in R2
f(t+s) - f(t)
f '(t2)
s f
f(t+s)
f(t) f '(t)
t1 t2 I

Examples

1. Let
f (t) = (cos t, sin t) t ∈ [0, 2π).
This traces out a circle in R2 and

f 0 (t) = (− sin t, cos t).

2. Let
f (t) = (t, t2 ).
This traces out a parabola in R2 and

f 0 (t) = (1, 2t).

f '(t) = (-sin t, cos t)

f '(t) = (1, 2t)

f(t)=(cos t, sin t)

f(t) = (t, t 2)

Example Consider the functions

1. f 1 (t) = (t, t3 ) t ∈ R,

2. f 2 (t) = (t3 , t9 ) t ∈ R,
240

√
3. f 3 (t) = ( 3 t, t) t ∈ R.

Then each function f i traces out the same “cubic” curve in R2 , (i.e., the
image is the same set of points), and

f 1 (0) = f 2 (0) = f 3 (0) = (0, 0).

However,
f 01 (0) = (1, 0), f 02 (0) = (0, 0), f 03 (0) is undefined.

Intuitively, we will think of a path in Rn as a function f which neither

stops nor reverses direction. It is often convenient to consider the variable t
as representing “time”. We will think of the corresponding curve as the set
of points traced out by f . Many different paths (i.e. functions) will give the
same curve; they correspond to tracing out the curve at different times and
velocities. We make this precise as follows:

Definition 18.2.5 We say f : I → Rn is a path 2 in Rn if f is C 1 and f 0 (t) 6= 0

for t ∈ I. We say the two paths f 1 : I1 → Rn and f 2 : I2 → Rn are equivalent
if there exists a function φ : I1 → I2 such that f 1 = f 2 ◦ φ, where φ is C 1 and
φ0 (t) > 0 for t ∈ I1 .
A curve is an equivalence class of paths. Any path in the equivalence
class is called a parametrisation of the curve.

We can think of φ as giving another way of measuring “time”.

We expect that the unit tangent vector to a curve should depend only on
the curve itself, and not on the particular parametrisation. This is indeed
the case, as is shown by the following Proposition.

f 1'(t) / |f 1'(t)| =
f2'( ϕ(t)) / |f2'( ϕ(t))|

f1(t)=f 2(ϕ(t))

f2
f1

ϕ
I1 I2

Proposition 18.2.6 Suppose f 1 : I1 → Rn and f 2 : I2 → Rn are equivalent

parametrisations; and in particular f 1 = f 2 ◦ φ where φ : I1 → I2 , φ is C 1 and
φ0 (t) > 0 for t ∈ I1 . Then f 1 and f 2 have the same unit tangent vector at t
and φ(t) respectively.
2
Other texts may have different terminology.
Differential Calculus for Vector-Valued Functions 241

Proof: From the chain rule for a function of one variable, we have
³ 0
´
f 01 (t) = f11 (t), . . . , f1m 0 (t)
³ 0
´
= f21 (φ(t)) φ0 (t), . . . , f2m 0 (φ(t)) φ0 (t)
= f 02 (φ(t)) φ0 (t).

Hence, since φ0 (t) > 0,

f 01 (t) f 02 (t)
= .
|f 01 (t)| |f 02 (t)|

Definition 18.2.7 If f is a path in Rn , then the acceleration at t is f 00 (t).

Example If |f 0 (t)| is constant (i.e. the “speed” is constant) then the velocity
and the acceleration are orthogonal.
³ ´
Proof: Since |f (t)|2 = f 0 (t), f 0 (y) is constant, we have from Proposi-
tion 18.2.4 that
d³ 0 ´
0 = f (t), f 0 (y)
dt³ ´
= 2 f 00 (t), f 0 (y) .

This gives the result.

18.2.1 Arc length

Suppose f : [a, b] → Rn is a path in Rn . Let a = t1 < t2 < . . . < tn = b be a
partition of [a, b], where ti − ti−1 = δt for all i.
We think of the length of the curve corresponding to f as being
Z
X
N X
N
|f (ti ) − f (ti−1 )| b
≈ |f (ti ) − f (ti−1 )| = δt ≈ |f 0 (t)| dt.
i=2 i=2 δt a

See the next diagram.

f(t N)

f(t i-1)

f(t i)
f(t 1)
f

t1 t i-1 ti tN

Motivated by this we make the following definition.

242

Definition 18.2.8 Let f : [a, b] → Rn be a path in Rn . Then the length of

the curve corresponding to f is given by
Z b
|f 0 (t)| dt.
a

The next result shows that this definition is independent of the particular
parametrisation chosen for the curve.

Proposition 18.2.9 Suppose f 1 : [a1 , b1 ] → Rn and f 2 : [a2 , b2 ] → Rn are

equivalent parametrisations; and in particular f 1 = f 2 ◦ φ where φ : [a1 , b1 ] →
[a2 , b2 ], φ is C 1 and φ0 (t) > 0 for t ∈ I1 . Then
Z b1 Z b2
|f 01 (t)| dt = |f 02 (s)| ds.
a1 a2

Proof: From the chain rule and then the rule for change of variable of
integration,
Z b1 Z b1
|f 01 (t)| dt = |f 02 (φ(t))| φ0 (t)dt
a1 a1
Z b2
= |f 02 (s)| ds.
a2

18.3 Partial and Directional Derivatives

Analogous to Definitions 17.3.1 and 17.4.1 we have:

Definition 18.3.1 The ith partial derivative of f at x is defined by

∂f ³ ´ f (x + tei ) − f (x)
i
(x) or Di f (x) = lim ,
∂x t→0 t
provided the limit exists. More generally, the directional derivative of f at x
in the direction v is defined by
f (x + tv) − f (x)
Dv f (x) = lim ,
t→0 t
provided the limit exists.

Remarks

1. It follows immediately from the Definitions that

∂f
(x) = Dei f (x).
∂xi
Differential Calculus for Vector-Valued Functions 243

2. The partial and directional derivatives are vectors in Rn . In the ter-

∂f
minology of the previous section, (x) is tangent to the path t 7→
∂xi
f (x + tei ) and Dv f (x) is tangent to the path t 7→ f (x + tv). Note that
the curves corresponding to these paths are subsets of the image of f .

3. As we will discuss later, we may regard the partial derivatives at x as

a basis for the tangent space to the image of f at f (x)3 .

Dv f(a)
∂f
(b)
∂y
R2 f(a)

f(b)

v e2
f
∂f (b)
a b e1 ∂x
R3

Proposition 18.3.2 If f 1 , . . . , f m are the component functions of f then

à !
∂f ∂f 1 ∂f m
(a) = (a), . . . , (a) for i = 1, . . . , n
∂xi ∂xi ∂xi
³ ´
Dv f (a) = Dv f 1 (a), . . . , Dv f m (a)

in the sense that if one side of either equality exists, then so does the other,
and both sides are then equal.

Proof: Essentially the same as for the proof of Proposition 18.2.2.

Example Let f : R2 → R3 be given by

f (x, y) = (x2 − 2xy, x2 + y 3 , sin x).

Then
à !
∂f ∂f 1 ∂f 2 ∂f 3
(x, y) = , , = (2x − 2y, 2x, cos x),
∂x ∂x ∂x ∂x
à !
∂f ∂f 1 ∂f 2 ∂f 3
(x, y) = , , = (−2x, 3y 2 , 0),
∂y ∂y ∂y ∂y

are vectors in R3 .
3
More precisely, if n ≤ m and the differential df (x) has rank n. See later.
244

18.4 The Differential

Analogous to Definition 17.5.1 we have:

Definition 18.4.1 Suppose f : D (⊂ Rn ) → Rn . Then f is differentiable at

a ∈ D if there is a linear transformation L : Rn → Rn such that
¯ ³ ´¯
¯ ¯
¯f (x) − f (a) + L(x − a) ¯
→ 0 as x → a. (18.1)
|x − a|

The linear transformation L is denoted by f 0 (a) or df (a) and is called the

derivative or differential of f at a4 .

A vector-valued function is differentiable iff the corresponding component

functions are differentiable. More precisely:

Proposition 18.4.2 f is differentiable at a iff f 1 , . . . , f m are differentiable

at a. In this case the differential is given by
³ ´
hdf (a), vi = hdf 1 (a), vi, . . . , hdf m (a), vi . (18.2)

In particular, the differential is unique.

Proof: For any linear map L : Rn → Rn , and for

³
each
´i
i = 1, . . . , m, let
L : R → R be the linear map defined by L (v) = L(v) .
i n i

From Theorem 10.4.4 it follows

¯ ³ ´¯
¯ ¯
¯f (x) − f (a) + L(x − a) ¯
→ 0 as x → a
|x − a|
iff
¯ ³ ´¯
¯ i ¯
¯f (x) − f i (a) + Li (x − a) ¯
→ 0 as x → a for i = 1, . . . , m.
|x − a|

Thus f is differentiable at a iff f 1 , . . . , f m are differentiable at a.

In this case we must have

Li = df i (a) i = 1, . . . , m

(by uniqueness of the differential for real -valued functions), and so

³ ´
L(v) = hdf 1 (a), vi, . . . , hdf m (a), vi .

But this says that the differential df (a) is unique and is given by (18.2).
4
It follows from Proposition 18.4.2 that if L exists then it is unique and is given by the
right side of (18.2).
Differential Calculus for Vector-Valued Functions 245

Corollary 18.4.3 If f is differentiable at a then the linear transformation

df (a) is represented by the matrix
 
∂f 1 ∂f 1
 (a) · · · (a) 
 ∂x1 ∂xn 
 .. .. . 
 . . ..  : Rn → Rn (18.3)
 
 ∂f m
∂f m 
1
(a) · · · (a)
∂x ∂xn

Proof: The ith column of the matrix corresponding to df (a) is the vector
hdf (a), ei i5 . From Proposition 18.4.2 this is the column vector corresponding
to ³ ´
hdf 1 (a), ei i, . . . , hdf m (a), ei i ,
i.e. to
³ ∂f 1 ∂f m ´
i
(a), . . . , (a) .
∂x ∂xi
This proves the result.

Remark The jth column is the vector in Rn corresponding to the partial

∂f
derivative (a). The ith row represents df i (a).
∂xj

The following proposition is immediate.

Proposition 18.4.4 If f is differentiable at a then

f (x) = f (a) + hdf (a), x − ai + ψ(x),

where ψ(x) = o(|x − a|).

Conversely, suppose

f (x) = f (a) + L(x − a) + ψ(x),

where L : Rn → Rn is linear and ψ(x) = o(|x − a|). Then f is differentiable

at a and df (a) = L.

Proof: As for Proposition 17.5.4.

Thus as is the case for real-valued functions, the previous proposition

implies f (a) + hdf (a), x − ai gives the best first order approximation to f (x)
for x near a.
For any linear transformation L : Rn → Rm , the ith column of the corresponding
5

matrix is L(ei ).
246

Example Let f : R2 → R2 be given by

f (x, y) = (x2 − 2xy, x2 + y 3 ).

Find the best first order approximation to f (x) for x near (1, 2).
Solution:
" #
−3
f (1, 2) = ,
9
" #
2x − 2y −2x
df (x, y) = ,
2x 3y 2
" #
−2 −2
df (1, 2) = .
2 12

So the best first order approximation near (1, 2) is

f (1, 2) + hdf (1, 2), (x − 1, y − 2)i

" # " #" #
−3 −2 −2 x−1
= +
9 2 12 y−2
" #
−3 − 2(x − 1) − 4(y − 2)
=
9 + 2(x − 1) + 12(y − 2)
" #
7 − 2x − 4y
= .
−17 + 2x + 12y

Alternatively, working with each component separately, the best first or-
der approximation is
µ
∂f 1 ∂f 1
f 1 (1, 2) + (1, 2)(x − 1) + (1, 2)(y − 2),
∂x ∂y
¶
∂f 2 ∂f 2
2
f (1, 2) + (1, 2)(x − 1) + (y − 2)
∂x ∂y
³ ´
= −3 − 2(x − 1) − 4(y − 2), 9 + 2(x − 1) + 12(y − 2)
³ ´
= 7 − 2x − 4y, −17 + 2x + 12y .

Remark One similarly obtains second and higher order approximations by

using Taylor’s formula for each component function.

Proposition 18.4.5 If f , g : D (⊂ Rn ) → Rn are differentiable at a ∈ D,

then so are αf and f + g. Moreover,

d(αf )(a) = αdf (a),

d(f + g)(a) = df (a) + dg(a).

Proof: This is straightforward (exercise) from Proposition 18.4.4.

Differential Calculus for Vector-Valued Functions 247

The previous proposition corresponds to the fact that the partial deriva-
tives for f + g are the sum of the partial derivatives corresponding to f and
g respectively. Similarly for αf .

Higher Derivatives We say f ∈ C k (D) iff f 1 , . . . , f m ∈ C k (D).

It follows from the corresponding results for the component functions that

1. f ∈ C 1 (D) ⇒ f is differentiable in D;
2. C 0 (D) ⊃ C 1 (D) ⊃ C 2 (D) ⊃ . . ..

18.5 The Chain Rule

Motivation The chain rule for the composition of functions of one variable
says that
d ³ ´ ³ ´
g f (x) = g 0 f (x) f 0 (x).
dx
Or to use a more informal notation, if g = g(f ) and f = f (x), then
dg dg df
= .
dx df dx

This is generalised in the following theorem. The theorem says that the
linear approximation to g ◦f (computed at x) is the composition of the linear
approximation to f (computed at x) followed by the linear approximation to
g (computed at f (x)).

A Little Linear Algebra Suppose L : Rn → Rn is a linear map. Then we

define the norm of L by
||L|| = max{|L(x)| : |x| ≤ 1}6 .

A simple result (exercise) is that

|L(x)| ≤ ||L|| |x| (18.4)
for any x ∈ Rn .
It is also easy to check (exercise) that || · || does define a norm on the
vector space of linear maps from Rn into Rn .

Theorem 18.5.1 (Chain Rule) Suppose f : D (⊂ Rn ) → Ω (⊂ Rn ) and

g : Ω (⊂ Rn ) → Rr . Suppose f is differentiable at x and g is differentiable at
f (x). Then g ◦ f is differentiable at x and
d(g ◦ f )(x) = dg(f (x)) ◦ df (x). (18.5)
6
Here |x|, |L(x)| are the usual Euclidean norms on Rn and Rm . Thus ||L|| corresponds
to the maximum value taken by L on the unit ball. The maximum value is achieved, as
L is continuous and {x : |x| ≤ 1} is compact.
248

Schematically:
g◦f
−−−−−−−−f−−−−−−−−−−−→ g
D (⊂ Rn ) −→ Ω (⊂ Rn ) −→Rr
d(g◦f )(x) = dg(f (x)) ◦ df (x)
−−− −−−−−−−−−dg(f
df (x)
−−− −−→
(x))
R n
−→ R n
−→ Rr

Example To see how all this corresponds to other formulations of the chain
rule, suppose we have the following:
f g
R3 −→ R2 −→ R2
(x, y, z) (u, v) (p, q)

Thus coordinates in R3 are denoted by (x, y, z), coordinates in the first copy
of R2 are denoted by (u, v) and coordinates in the second copy of R2 are
denoted by (p, q).
The functions f and g can be written as follows:

f : u = u(x, y, z), v = v(x, y, z),

g : p = p(u, v), q = q(u, v).

Thus we think of u and v as functions of x, y and z; and p and q as functions

of u and v.
We can also represent p and q as functions of x, y and z via
³ ´ ³ ´
p = p u(x, y, z), v(x, y, z) , q = q u(x, y, z), v(x, y, z) .

The usual version of the chain rule in terms of partial derivatives is:
∂p ∂p ∂u ∂p ∂v
= +
∂x ∂u ∂x ∂v ∂x
∂p ∂p ∂u ∂p ∂v
= +
∂x ∂u ∂x ∂v ∂x
..
.
∂q ∂q ∂u ∂q ∂v
= + .
∂z ∂u ∂z ∂v ∂z
∂p ∂p ∂p
In the first equality,´ ∂x
³
is evaluated at (x, y, z), ∂u and ∂v are evaluated at
∂u ∂v
u(x, y, z), v(x, y, z) , and ∂x and ∂x are evaluated at (x, y, z). Similarly for
the other equalities.
In terms of the matrices of partial derivatives:
" ∂p ∂p ∂p # " #" ∂u ∂u ∂u
#
∂p ∂p
∂x ∂y ∂z ∂u ∂v ∂x ∂y ∂z
∂q ∂q ∂q = ∂q ∂q ∂v ∂v ∂v ,
∂x ∂y ∂z ∂u ∂v ∂x ∂y ∂z
| {z } | {z }| {z }
d(g ◦ f )(x) dg(f (x)) df (x)

where x = (x, y, z).

Differential Calculus for Vector-Valued Functions 249

Proof of Chain Rule: We want to show

(f ◦ g)(a + h) = (f ◦ g)(a) + L(h) + o(|h|), (18.6)

where L = df (g(a)) ◦ dg(a).

Now
³ ´
(f ◦ g)(a + h) = f g(a + h)
³ ´
= f g(a) + g(a + h) − g(a)
³ ´ D ³ ´ E
= f g(a) + df g(a) , g(a + h) − g(a)
³ ´
+o |g(a + h) − g(a)|
³
.´. . byD the
³
differentiability
´
of f E
= f g(a) + df g(a) , hdg(a), hi + o(|h|)
³ ´
+o |g(a + h) − g(a)|
³
.´. . byD the
³
differentiability
´
of g
E
= f g(a) + df g(a) , hdg(a), hi
D ³ ´ E ³ ´
+ df g(a) , o(|h|) + o |g(a + h) − g(a)|
= A+B+C +D
D ³ ´ E
But B = df g(a) ◦ dg(a), h , by definition of the “composition” of two
maps. Also C = o(|h|) from (18.4) (exercise). Finally, for D we have
¯ ¯ ¯ ¯
¯ ¯
¯g(a + h) − g(a)¯ = ¯¯hdg(a), hi + o(|h|)¯¯ . . . by differentiability of g
≤ ||dg(a)|| |h| + o(|h|) . . . from (18.4)
= O(|h|) . . . why?

Substituting the above expressions into A + B + C + D, we get

³ ´ D ³ ´ E
(f ◦ g)(a + h) = f g(a) + df g(a)) ◦ dg(a), h + o(|h|). (18.7)

If follows that f ◦ g is differentiable at a, and moreover the differential

equals df (g(a)) ◦ dg(a). This proves the theorem.
250
Chapter 19

The Inverse Function Theorem

and its Applications

19.1 Inverse Function Theorem

Motivation

1. Suppose
f : Ω (⊂ Rn ) → Rn
and f is C 1 . Note that the dimension of the domain and the range are
the same. Suppose f (x0 ) = y0 . Then a good approximation to f (x)
for x near x0 is gven by
x 7→ f (x0 ) + hf 0 (x0 ), x − x0 i. (19.1)

f
x0
f(x 0)

R2 R2
The right curved grid is the image of the left grid under f
The right straight grid is the image of the left grid under the
first order map x |→ f(x 0) + <f'(x0), x-x0>

We expect that if f 0 (x0 ) is a one-one and onto linear map, (which is the
same as det f 0 (x0 ) 6= 0 and which implies the map in (19.1) is one-one
and onto), then f should be one-one and onto near x0 . This is true,
and is called the Inverse Function Theorem.
2. Consider the set of equations
f 1 (x1 , . . . , xn ) = y 1
f 2 (x1 , . . . , xn ) = y 2
..
.
f (x , . . . , x ) = y n ,
n 1 n

251
252

where f 1 , . . . , f n are certain real-valued functions. Suppose that these

equations are satisfied if (x1 , . . . , xn ) = (x10 , . . . , xn0 ) and (y 1 , . . . , y n ) =
(y01 , . . . , y0n ), and that det f 0 (x0 ) 6= 0. Then it follows from the In-
verse Function Theorem that for all (y 1 , . . . , y n ) in some ball centred
at (y01 , . . . , y0n ) the equations have a unique solution (x1 , . . . , xn ) in some
ball centred at (x10 , . . . , xn0 ).

Theorem 19.1.1 (Inverse Function Theorem) Suppose f : Ω (⊂ Rn ) →

Rn is C 1 and Ω is open1 . Suppose f 0 (x0 ) is invertible2 for some x0 ∈ Ω.
Then there exists an open set U 3 x0 and an open set V 3 f (x0 ) such
that

1. f 0 (x) is invertible at every x ∈ U ,

2. f : U → V is one-one and onto, and hence has an inverse g : V → U ,

3. g is C 1 and g 0 (f (x)) = [f 0 (x)]−1 for every x ∈ U .

f
.
x0
.
f(x0)

g V
U

Proof: Step 1 Suppose

y ∗ ∈ Bδ (f (x0 )).

We will choose δ later. (We will take the set V in the theorem to be the open
set Bδ (f (x0 )) )
For each such y, we want to prove the existence of x (= x∗ , say) such that

f (x) = y ∗ . (19.2)

1
Note that the dimensions of the domain and range are equal.
2
That is, the matrix f 0 (x0 ) is one-one and onto, or equivalently, det f 0 (x0 ) 6= 0.
Inverse Function Theorem 253

We write f (x) as a first order function plus an error term. Thus we want
to solve (for x)

f (x0 ) + hf 0 (x0 ), x − x0 i + R(x) = y ∗ , (19.3)

where
R(x) := f (x) − f (x0 ) − hf 0 (x0 ), x − x0 i. (19.4)
In other words, we want to find x such that

hf 0 (x0 ), x − x0 i = y ∗ − f (x0 ) − R(x),

i.e. such that

D E D E
x = x0 + [f 0 (x0 )]−1 , y ∗ − f (x0 ) − [f 0 (x0 )]−1 , R(x) (19.5)

(why?).
The right side of (19.5) is the sum of two terms. The first term, that is
x0 +h[f 0 (x0 )]−1 , y ∗ − f (x0 )i, is the solution of the linear equation y ∗ = f (x0 )+
hf 0 (x0 ), x − x0 i. The second term is the error term − h[f 0 (x0 )]−1 , R(x)i,
which is o(|x − x0 |) because R(x) is o(|x − x0 |) and [f 0 (x0 )]−1 is a fixed
<[f'(x0)]-1, R(x)>
Rn graph of f
R(x)

y*
f(x0)

Rn
x
x0 x*
linear map. x0 + <[f'(x0)]-1, y*-f(x0)>
254

Step 2 Because of (19.5) define

D E D E
Ay∗ (x) := x0 + [f 0 (x0 )]−1 , y ∗ − f (x0 ) − [f 0 (x0 )]−1 , R(x) . (19.6)

Note that x is a fixed point of Ay∗ iff x satisfies (19.5) and hence solves (19.2).
We claim that
Ay∗ : B ² (x0 ) → B² (x0 ), (19.7)
and that Ay∗ is a contraction map, provided ² > 0 is sufficiently small (²
will depend only on x0 and f ) and provided y ∗ ∈ Bδ (y0 ) (where δ > 0 also
depends only on x0 and f ).
To prove the claim, we compute
D E
Ay∗ (x1 ) − Ay∗ (x2 ) = [f 0 (x0 )]−1 , R(x2 ) − R(x1 ) ,
and so
|Ay∗ (x1 ) − Ay∗ (x2 )| ≤ K |R(x1 ) − R(x2 )|, (19.8)
where ° °
K := °°[f 0 (x0 )]−1 °° . (19.9)

From (19.4)
R(x2 ) − R(x1 ) = f (x2 ) − f (x1 ) − hf 0 (x0 ), x2 − x1 i .
We apply the mean value theorem (17.9.1) to each of the components of this
equation to obtain
¯ ¯ ¯D E D E¯
¯ i ¯ ¯ 0 ¯
0
¯R (x2 ) − Ri (x1 )¯ = ¯ f i (ξi ), x2 − x1 − f i (x0 ), x2 − x1 ¯
for i = 1, . . . , n and someE¯ξi ∈ Rn between x1 and x2
¯D 0
¯ 0 ¯
= ¯ f i (ξi ) − f i (x0 ), x2 − x1 ¯
¯ ¯
≤ ¯¯f i (ξi ) − f i (x0 )¯¯ |x2 − x1 |,
0 0

0
by Cauchy-Schwartz, treating f i as a “row vector”.

.. .
x1 ξi

x2

. x0

By the continuity of the derivatives of f , it follows

1
|R(x2 ) − R(x1 )| ≤ |x2 − x1 |, (19.10)
2K
Inverse Function Theorem 255

provided x1 , x2 ∈ B ² (x0 ) for some ² > 0 depending only on f and x0 . Hence

from (19.8)
1
|Ay∗ (x1 ) − Ay∗ (x2 )| ≤ |x1 − x2 |. (19.11)
2
This proves
Ay∗ : B ² (x0 ) → Rn
is a contraction map, but we still need to prove (19.7).
For this we compute
D¯ E¯ D ¯ E¯
¯ ¯ ¯ ¯
|Ay∗ (x) − x0 | ≤ ¯ [f 0 (x0 )]−1 , y ∗ − f (x0 ) ¯ + ¯ [f 0 (x0 )]−1 , R(x) ¯ from (19.6)
≤ K|y ∗ − f (x0 )| + K|R(x)|
= K|y ∗ − f (x0 )| + K|R(x) − R(x0 )| as R(x0 ) = 0
1
≤ K|y ∗ − f (x0 )| + |x − x0 | from (19.10)
2
< ²/2 + ²/2 = ²,

provided x ∈ B ² (x0 ) and y ∗ ∈ Bδ (f (x0 )) (if Kδ < ²). This establishes (19.7)
and completes the proof of the claim.
Step 3 We now know that for each y ∈ Bδ (f (x0 )) there is a unique x ∈
B² (x0 ) such that f (x) = y. Denote this x by g(y). Thus

g : Bδ (f (x0 )) → B² (x0 ).

We claim that this inverse function g is continuous.

To see this let xi = g(yi ) for i = 1, 2. That is, f (xi ) = yi , or equivalently
xi = Ayi (xi ) (recall the remark after (19.6) ). Then

|g(y1 ) − g(y2 )| = |x1 − x2 |

= |Ay1 (x1 ) − Ay2 1 (x2 )|
¯D E¯ ¯D E¯
¯ ¯
¯ ¯
≤ ¯ [f 0 (x0 )]−1 , y1 − y2 ¯ + ¯ [f 0 (x0 )]−1 , R(x1 ) − R(x2 ) ¯ by (19.6)
≤ K |y1 − y2 | + K |R(x1 ) − R(x2 )| from (19.8)
1
≤ K |y1 − y2 | + K |x1 − x2 | from (19.10)
2K
1
= K |y1 − y2 | + |g(y1 ) − g(y2 )|.
2
Thus
1
|g(y1 ) − g(y2 )| ≤ K |y1 − y2 |,
2
and so
|g(y1 ) − g(y2 )| ≤ 2K |y1 − y2 |. (19.12)
In particular, g is Lipschitz and hence continuous.

Step 4 Let
V = Bδ (f (x0 )), U = g [Bδ (f (x0 ))] .
256

Since U = B² (x0 ) ∩ f −1 [V ] (why?), it follows U is open. We have thus

proved the second part of the theorem.
The first part of the theorem is easy. All we need do is first replace Ω by
a smaller open set containing x0 in which f 0 (x) is invertible for all x. This is
possible as det f 0 (x0 ) 6= 0 and the entries in the matrix f 0 (x) are continuous.

Step 5 We claim g is C 1 on V and

g 0 (f (x)) = [f 0 (x)]−1 . (19.13)

To see that g is differentiable at y ∈ V and (19.13) is true, suppose

y, y ∈ V , and let f (x) = y, f (x) = y where x, x ∈ U . Then

|g(y) − g(y) − h[f 0 (x)]−1 , y − yi|

|y − y|
|x − x − h[f (x)]−1 , f (x) − f (x)i|
0
=
|y − y|
¯D E¯
¯ ¯
¯ [f 0 (x)]−1 , hf 0 (x), x − xi − f (x) + f (x) ¯
=
|y − y|
° ° |f (x) − f (x) − hf 0 (x), x − xi| |x − x|
° °
≤ °[f 0 (x)]−1 ° .
|x − x| |y − y|

If we fix y and let y → y, then x is fixed and x → x. Hence the last

line in the previous series of inequalities → 0, since f is differentiable at x
and |x − x|/|y − y| ≤ K/2 by (19.12). Hence g is differentiable at y and the
derivative is given by (19.13).
The fact that g is C 1 follows from (19.13) and the expression for the
inverse of a matrix.

Remark We have
−1
g 0 (y) = [f 0 (g(y))]
Ad [f 0 (g(y))]
= , (19.14)
det[f 0 (g(y))]
where Ad [f 0 (g(y))] is the matrix of cofactors of the matrix [f (g(y))].
If f is C 2 , then since we already know g is C 1 , it follows that the terms
in the matrix (19.14) are algebraic combinations of C 1 functions and so are
C 1 . Hence the terms in the matrix g 0 are C 1 and so g is C 2 .
Similarly, if f is C 3 then since g is C 2 it follows the terms in the ma-
trix (19.14) are C 2 and so g is C 3 .
By induction we have the following Corollary.

Corollary 19.1.2 If in the Inverse Function Theorem the function f is C k

then the local inverse function g is also C k .
Inverse Function Theorem 257

Summary of Proof of Theorem

1. Write the equation f (x∗ ) = y as a perturbation of the first order equa-

tion obtained by linearising around x0 . See (19.3) and (19.4).
Write the solution x as the solution T (y ∗ ) of the linear equation plus
an error term E(x),

x = T (y ∗ ) + E(x) =: Ay∗ (x)

See (19.5).

2. Show Ay∗ (x) is a contraction map on B² (x0 ) (for ² sufficiently small

and y ∗ near y0 ) and hence has a fixed point. It follows that for all y ∗
near y0 there exists a unique x∗ near x0 such that f (x∗ ) = y ∗ . Write
g(y ∗ ) = x∗ .

3. The local inverse function g is close to the inverse T (y ∗ ) of the linear

function. Use this to prove that g is Lipschitz continuous.

4. Wrap up the proof of parts 1 and 2 of the theorem.

5. Write out the difference quotient for the derivative of g and use this
and the differentiability of f to show g is differentiable.

19.2 Implicit Function Theorem

Motivation We can write the equations in the previous “Motivation” sec-
tion as
f (x) = y,
where x = (x1 , . . . , xn ) and y = (y 1 , . . . , y n ).
More generally we may have n equations

f (x, u) = y,

i.e.,

f 1 (x1 , . . . , xn , u1 , . . . , um ) = y 1
f 2 (x1 , . . . , xn , u1 , . . . , um ) = y 2
..
.
f (x , . . . , x , u , . . . , u ) = y n ,
n 1 n 1 m

where we regard the u = (u1 , . . . , um ) as parameters.

Write  
∂f 1 ∂f 1
" #
∂x1
··· ∂xn
∂f  .. .. 
det := det 
 . . .

∂x ∂f n ∂f n
∂x1
··· ∂xn
258

Thus det[∂f /∂x] is the determinant of the derivative of the map f (x1 , . . . , xn ),
where x1 , . . . , xm are taken as the variables and the u1 , . . . , um are taken to
be fixed .
Now suppose that
" #
∂f
f (x0 , u0 ) = y0 , det 6= 0.
∂x (x0 ,u0 )

From the Inverse Function Theorem (still thinking of u1 , . . . , um as fixed),

for y near y0 there exists a unique x near x0 such that

f (x, u0 ) = y.

The Implicit Function Theorem says more generally that for y near y0 and
for u near u0 , there exists a unique x near x0 such that

f (x, u) = y.

In applications we will usually take y = y0 = c(say) to be fixed. Thus we

consider an equation
f (x, u) = c (19.15)
where " #
∂f
f (x0 , u0 ) = c, det 6= 0.
∂x (x0 ,u0 )

Hence for u near u0 there exists a unique x = x(u) near x0 such that

f (x(u), u) = c. (19.16)

In words, suppose we have n equations involving n unknowns x and certain

parameters u. Suppose the equations are satisfied at (x0 , u0 ) and suppose that
the determinant of the matrix of derivatives with respect to the x variables is
non-zero at (x0 , u0 ). Then the equations can be solved for x = x(u) if u is
near u0 .
Moreover, differentiating the ith equation in (19.16) with respect to uj
we obtain
X ∂f i ∂xk ∂f i
j
+ = 0.
k ∂xk ∂u ∂uj
That is " #" # " #
∂f ∂x ∂f
+ = [0],
∂x ∂u ∂u
where the first three matrices are n × n, n × m, and n × m respectively, and
the last matrix is the n × m zero matrix. Since det [∂f /∂x](x0 ,u0 ) 6= 0, it
follows " # " #−1 " #
∂x ∂f ∂f
=− . (19.17)
∂u u0 ∂x (x0 ,u0 ) ∂u (x0 ,u0 )
Inverse Function Theorem 259

Example 1 Consider the circle in R2 described by

x2 + y 2 = 1.

Write
F (x, y) = 1. (19.18)
Thus in (19.15), u is replaced by y and c is replaced by 1.
y

.
(x0,y 0)

. x
(x0,y0)
x2+y2 = 1

Suppose F (x0 , y0 ) = 1 and ∂F/∂x0 |(x0 ,y0 ) 6= 0 (i.e. x0 6= 0). Then for y
√
near y0 there is a unique x near x0 satisfying (19.18). In fact x = ± 1 − y 2
according as x0 > 0 or x0 < 0. See the diagram for two examples of such
points (x0 , y0 ).
Similarly, if ∂F/∂y0 |(x0 ,y0 ) 6= 0, i.e. y0 6= 0, Then for x near x0 there is a
unique y near y0 satisfying (19.18).

Example 2 Suppose a “surface” in R3 is described by

Φ(x, y, z) = 0. (19.19)

Suppose Φ(x0 , y0 , z0 ) = 0 and ∂Φ/∂z (x0 , y0 , z0 ) 6= 0.

z

Φ (x,y,z) = 0
z0

.
y

x (x0,y0)

Then by the Implicit Function Theorem, for (x, y) near (x0 , y0 ) there is
a unique z near z0 such that Φ(x, y, z) = 0. Thus the “surface” can locally3
be written as a graph over the x-y plane
More generally, if ∇Φ(x0 , y0 , z0 ) 6= 0 then at least one of the derivatives
∂Φ/∂x (x0 , y0 , z0 ), ∂Φ/∂y (x0 , y0 , z0 ) or ∂Φ/∂z (x0 , y0 , z0 ) does not equal 0.
The corresponding variable x, y or z can then be solved for in terms of
3
By “locally” we mean in some Br (a) for each point a in the surface.
260

the other two variables and the surface is locally a graph over the plane
corresponding to these two other variables.

Example 3 Suppose a “curve” in R3 is described by

Φ(x, y, z) = 0,
Ψ(x, y, z) = 0.

Φ (x,y,z) = 0
z0 Ψ(x,y,z) = 0

.
y

x (x0,y0)

Suppose (x0 , y0 , z0 ) lies on the curve, i.e. Φ(x0 , y0 , z0 ) = Ψ(x0 , y0 , z0 ) = 0.

Suppose moreover that the matrix
" ∂Φ ∂Φ ∂Φ
#
∂x ∂y ∂z
∂Ψ ∂Ψ ∂Ψ
∂x ∂y ∂z (x0 ,y0 ,z0 )

has rank 2. In other words, two of the three columns must be linearly inde-
pendent. Suppose it is the first two. Then
¯ ¯
¯ ∂Φ ∂Φ ¯
det ¯¯ ∂x ∂y ¯
¯ 6= 0.
¯ ∂Ψ ∂Ψ ¯
∂x ∂y (x0 ,y0 ,z0 )

By the Implicit Function Theorem, we can solve for (x, y) near (x0 , y0 ) in
terms of z near z0 . In other words we can locally write the curve as a graph
over the z axis.

Example 4 Consider the equations

f1 (x1 , x2 , y1 , y2 , y3 ) = 2ex1 + x2 y1 − 4y2 + 3

f2 (x1 , x2 , y1 , y2 , y3 ) = x2 cos x1 − 6x1 + 2y1 − y3 .

Consider the “three dimensional surface in R5 ” given by f1 (x1 , x2 , y1 , y2 , y3 ) =

0, f2 (x1 , x2 , y1 , y2 , y3 ) = 0 4 . We easily check that

f (0, 1, 3, 2, 7) = 0
4
One constraint gives a four dimensional surface, two constraints give a three dimen-
sional surface, etc. Each further constraint reduces the dimension by one.
Inverse Function Theorem 261

and " #
2 3 1 −4 0
f 0 (0, 1, 3, 2, 7) = .
−6 1 2 0 −1
The first two columns are linearly independent and so we can solve for x1 , x2
in terms of y1 , y2 , y3 near (3, 2, 7).
Moreover, from (19.17) we have
" ∂x1 ∂x1 ∂x1
# " #−1 " #
2 3 1 −4 0
∂y1 ∂y2 ∂y3
= −
∂x2
∂y1
∂x2
∂y2
∂x2
∂y3
−6 1 2 0 −1
(3,2,7)
" #" #
1 1 −3 1 −4 0
= −
20 6 2 2 0 −1
" #
1
4
1
5
− 20
3
=
− 12 6
5
1
10

It folows that for (y1 , y2 , y3 ) near (3, 2, 7) we have

1 1 3
x1 ≈ 0 + (y1 − 3) + (y2 − 2) − (y3 − 7)
4 5 20
1 6 1
x2 ≈ 1 − (y1 − 3) + (y2 − 2) + (y3 − 7).
2 5 10
We now give a precise statement and proof of the Implicit Function The-
orem.

Theorem 19.2.1 (Implicit function Theorem) Suppose f : D (⊂ Rn ×

Rk ) → Rn is C 1 and D is open. Suppose f (x0 , u0 ) = y0 where x0 ∈ Rn and
u0 ∈ Rm . Suppose det [∂f /∂x] |(x0 ,u0 ) 6= 0.
Then there exist ², δ > 0 such that for all y ∈ Bδ (y0 ) and all u ∈ Bδ (u0 )
there is a unique x ∈ B² (x0 ) such that
f (x, u) = y.
If we denote this x by g(u, y) then g is C 1 . Moreover,
" # " #−1 " #
∂g ∂f ∂f
=− .
∂u (u0 ,y0 )
∂x (x0 ,u0 )
∂u (x0 ,u0 )

Proof: Define
F : D → Rn × Rm
by ³ ´
F (x, u) = f (x, u), u .

Then clearly5 F is C 1 and

" #
0 ∂f
det F |(x0 ,u0 ) = det .
∂x (x0 ,u0 )
5
Since " h i h i #
∂f ∂f
0 ∂x ∂u
F = ,
O I
where O is the m × n zero matrix and I is the m × m identity matrix.
262

Also
F (x0 , u0 ) = (y0 , u0 ).

From the Inverse Function Theorem, for all (y, u) near (y0 , u0 ) there exists
a unique (x, w) near (x0 , u0 ) such that

F (x, w) = (y, u). (19.20)

Moreover, x and w are C 1 functions of (y, u). But from the definition of F
it follows that (19.20) holds iff w = u and f (x, u) = y. Hence for all (y, u)
near (y0 , u0 ) there exists a unique x = g(u, y) near x0 such that

f (x, u) = y. (19.21)

Moreover, g is a C 1 function of (u, y).

h i
∂g
The expression for ∂u (u0 ,y0 )
follows from differentiating (19.21) precisely
as in the derivation of (19.17).

19.3 Manifolds
Discussion Loosely speaking, M is a k-dimensional manifold in Rn if M
locally6 looks like the graph of a function of k variables. Thus a 2-dimensional
manifold is a surface and a 1-dimensional manifold is a curve.
We will give three different ways to define a manifold and show that they
are equivalent.
We begin by considering manifolds of dimension n − 1 in Rn (e.g. a curve
in R2 or a surface in R3 ). Such a manifold is said to have codimension one.
Suppose
Φ : Rn → R
is C 1 . Let
M = {x : Φ(x) = 0}.
See Examples 1 and 2 in Section 19.2 (where Φ(x, y) = F (x, y) − 1 in Exam-
ple 1).
If ∇Φ(a) 6= 0 for some a ∈ M , then as in Examples 1 and 2 we can write
M locally as the graph of a function of one of the variables xi in terms of the
remaining n − 1 variables.

This leads to the following definition.

Definition 19.3.1 [Manifolds as Level Sets] Suppose M ⊂ Rn and for

each a ∈ M there exists r > 0 and a C 1 function Φ : Br (a) → R such that

M ∩ Br (a) = {x : Φ(x) = 0}.

6
“Locally” means in some neighbourhood for each a ∈ M .
Inverse Function Theorem 263

Suppose also that ∇Φ(x) 6= 0 for each x ∈ Br (a).

Then M is an n − 1 dimensional manifold in Rn . We say M has codi-
mension one.
The one dimensional space spanned by ∇Φ(a) is called the normal space
to M at a and is denoted by Na M 7 .

Na M

∇Φ(a) = 0

Remarks

1. Usually M is described by a single function Φ defined on Rn

2. See Section 17.6 for a discussion of ∇Φ(a) which motivates the defini-
tion of Na M .

3. With the same proof as in Examples 1 and 2 from Section 19.2, we can
locally write M as the graph of a function

xi = f (x1 , . . . , xi−1 , xi+1 , . . . , xn )

for some 1 ≤ i ≤ n.

Higher Codimension Manifolds Suppose more generally that

Φ : Rn → R`

is C 1 and ` ≥ 1. See Example 3 in Section 19.2.

Now
M = M 1 ∩ · · · ∩ M `,
where
M i = {x : Φi (x) = 0}.
7
The space Na M does not depend on the particular Φ used to describe M . We show
this in the next section.
264

Note that each Φi is real-valued. Thus we expect that, under reasonable

conditions, M should have dimension n − ` in some sense. In fact, if
∇Φ1 (x), . . . , ∇Φ` (x)
are linearly independent for each x ∈ M , then the same argument as for
Example 3 in the previous section shows that M is locally the graph of a
function of ` of the variables x1 , . . . , xn in terms of the other n − ` variables.

This leads to the following definition which generalises the previous one.

Definition 19.3.2 [Manifolds as Level Sets] Suppose M ⊂ Rn and for

each a ∈ M there exists r > 0 and a C 1 function Φ : Br (a) → R` such that
M ∩ Br (a) = {x : Φ(x) = 0}.
Suppose also that ∇Φ1 (x), . . . , ∇Φ` (x) are linearly independent for each x ∈
Br (a).
Then M is an n − ` dimensional manifold in Rn . We say M has codi-
mension `.
The ` dimensional space spanned by ∇Φ1 (a), . . . , ∇Φ` (a) is called the
normal space to M at a and is denoted by Na M 8 .

Remarks With the same proof as in Examples 3 from the section on the
Implicit Function Theorem, we can locally write M as the graph of a function
of ` of the variables in terms of the remaining n − ` variables.

Equivalent Definitions There are two other ways to define a manifold.

For simplicity of notation we consider the case M has codimension one, but
the more general case is completely analogous.

Definition 19.3.3 [Manifolds as Graphs] Suppose M ⊂ Rn and that for

each a ∈ M there exist r > 0 and a C 1 function f : Ω (⊂ Rn−1 ) → R such
that for some 1 ≤ i ≤ n
M ∩ Br (a) = {x ∈ Br (a) : xi = f (x1 , . . . , xi−1 , xi+1 , . . . , xn )}.
Then M is an n − 1 dimensional manifold in Rn .

R ~ xi
a

M = graph of f ,
near a

Rn-1 ~ x1,...,x i-1,xi+1,...,xn

8
The space Na M does not depend on the particular Φ used to describe M . We show
this in the next section.
Inverse Function Theorem 265

Equivalence of the Level-Set and Graph Definitions Suppose M is a

manifold as in the Graph Definition. Let

Φ(x) = xi − f (x1 , . . . , xi−1 , xi+1 , . . . , xn ).

Then à !
∂f ∂f ∂f ∂f
∇Φ(x) = − ,...,− , 1, − ,...,−
∂x1 ∂xi−1 ∂xi+1 ∂xn
In particular, ∇Φ(x) 6= 0 and so M is a manifold in the level-set sense.
Conversely, we have already seen (in the Remarks following Definitions 19.3.1
and 19.3.2) that if M is a manifold in the level-set sense then it is also a man-
ifold in the graphical sense.

As an example of the next definition, see the diagram preceding Propo-

sition 18.3.2.

Definition 19.3.4 [Manifolds as Parametrised Sets] Suppose M ⊂ Rn

and that for each a ∈ M there exists r > 0 and a C 1 function

F : Ω (⊂ Rn−1 ) → Rn

such that
M ∩ Br (a) = F [Ω] ∩ Br (a).

Suppose moreover that the vectors

∂F ∂F
(u), . . . , (u)
∂u1 ∂un−1

are linearly independent for each u ∈ Ω.

Then M is an n − 1 dimensional manifold in Rn . We say that (F, Ω) is
a parametrisation of (part of) M .
The n − 1 dimensional space spanned by ∂u
∂F
1
(u), . . . , ∂u∂F
n−1
(u) is called the
9
tangent space to M at a = F (u) and is denoted by Ta M .

Equivalence of the Graph and Parametrisation Definitions Suppose

M is a manifold as in the Parametrisation Definition. We want to show that
M is locally the graph of a C 1 function.
h i
First note that the n × (n − 1) matrix ∂F∂u
(p) has rank n − 1 and so
n − 1 of the rows are linearly independent. Suppose the first n − 1 rows are
linearly independent.
9
The space Ta M does not depend on the particular Φ used to describe M . We show
this in the next section.
266

( F1(u),...,Fn-1(u),Fn (u) )
= ( x1,...,xn-1 , (Fno G)(x 1,...,xn-1 ) ) M

F
⇒

Rn-1

u = (u1,...,u n-1)
= G(x 1,...,xn-1 ) G ( F1(u),...,Fn-1(u) )
⇐ = ( x1,...,xn-1 )

h i
i
It follows that the (n − 1) × (n − 1) matrix ∂F ∂uj
(p) is invert-
1≤i,j≤n−1
ible and hence by the Inverse Function Theorem there is locally a one-one
correspondence between u = (u1 , . . . , un−1 ) and points of the form

(x1 , . . . , xn−1 ) = (F 1 (u), . . . , F n−1 (u)) ∈ Rn−1 ' Rn−1 × {0} (⊂ Rn ),

with C 1 inverse G (so u = G(x1 , . . . , xn−1 )).

Thus points in M can be written in the form
³ ´
F 1 (u), . . . , F n−1 (u), F n (u) = (x1 , . . . , xn−1 , (F n ◦ G)(x1 , . . . , xn−1 )) .

Hence M is locally the graph of the C 1 function F n ◦ G.

Conversely, suppose M is a manifold in the graph sense. Then locally,

after perhaps relabelling coordinates, for some C 1 function f : Ω (⊂ Rn−1 ) →
R,
M = {(x1 , . . . , xn ) : xn = f (x1 , . . . , xn−1 )}.
It follows that M is also locally the image of the C 1 function F : Ω (⊂
Rn−1 ) → Rn defined by

F (x1 , . . . , xn−1 ) = (x1 , . . . , xn−1 , f (x1 , . . . , xn−1 )) .

Moreover,
∂F ∂f
= ei + en
∂xi ∂xi
for i = 1, . . . , n − 1, and so these vectors are linearly independent.

In conclusion, we have established the following theorem.

Theorem 19.3.5 The level-set, graph and parametrisation definitions of a

manifold are equivalent.
Inverse Function Theorem 267

Remark If M is parametrised locally by a function F : Ω (⊂ Rk ) → Rn and

also given locally as the zero-level set of Φ : Rn → R` then it follows that

k + ` = n.

To see this, note that previous arguments show that M is locally the graph
of a function from Rk → Rn−k and also locally the graph of a function from
Rn−` → R` . This makes it very plausible that k = n − `. A strict proof
requires a little topology or measure theory.

19.4 Tangent and Normal vectors

If M is a manifold given as the zero-level set (locally) of Φ : Rn → R` , then we
defined the normal space Na M to be the space spanned by ∇Φ1 (a), . . . , ∇Φ` (a).
If M is parametrised locally by F : Rk → Rn (where k + ` = n), then we
∂F
defined the tangent space Ta M to be the space spanned by ∂u 1
(u), . . . , ∂u∂F
k (u)
,
where F (u) = a.
We next give a definition of Ta M which does not depend on the particular
representation of M . We then show that Na M is the orthogonal complement
of Ta M , and so also Na M does not depend on the particular representation
of M .

.
a = ψ(0) ψ'(0)

Definition 19.4.1 Let M be a manifold in Rn and suppose a ∈ M . Suppose

ψ : I → M is C 1 where 0 ∈ I ⊂ R, I is an interval and ψ(0) = a. Any vector
h of the form
h = ψ 0 (0)
is said to be tangent to M at A. The set of all such vectors is denoted by
Ta M .

Theorem 19.4.2 The set Ta M as defined above is indeed a vector space.

If M is given locally by the parametrisation F : Rk → Rn and F (u) = a
then Ta M is spanned by

∂F ∂F
(u), . . . , (u).10
∂u1 ∂uk
10
As in Definition 19.3.4, these vectors are assumed to be linearly independent.
268

If M is given locally as the zero-level set of Φ : Rn → R` then Ta M is the

orthogonal complement of the space spanned by
∇Φ1 (a), . . . , ∇Φ` (a).

Proof: Step 1 : First suppose h = ψ 0 (0) as in the Definition. Then

Φi (ψ(t)) = 0
for i = 1, . . . , ` and for t near 0. By the chain rule
X
n
∂Φi dψ j
j
(a) (0) for i = 1, . . . , `,
j=1 ∂x dt
i.e.
∇Φi (a) ⊥ ψ 0 (0) for i = 1, . . . , `.
This shows that Ta M (as in Definition 19.4.1) is orthogonal to the space
spanned by ∇Φ1 (a), . . . , ∇Φ` (a), and so is a subset of a space of dimension
n − `.
Step 2 : If M is parametrised by F : Rk → Rn with F (u) = a, then every
vector
Xk
∂F
αi (u)
i=1 ∂ui
is a tangent vector as in Definition 19.4.1. To see this let
ψ(t) = F (u1 + tα1 , . . . , un + tαn ).
Then by the chain rule,
X
k
∂F
ψ 0 (0) = αi (u).
i=1 ∂ui
∂F ∂F
Hence Ta M contains the space spanned by ∂u1
(u), . . . , ∂u k
(u), and so
contains a space of dimension k(= n − `).
Step 3 : From the last line in Steps 1 and 2, it follows that Ta M is a space of
dimension n − `. It follows from Step 1 that Ta M is in fact the orthogonal
complement of the space spanned by ∇Φ1 (a), . . . , ∇Φ` (a), and from Step 2
∂F ∂F
that Ta M is in fact spanned by ∂u 1
(u), . . . , ∂uk
(u).

19.5 Maximum, Minimum, and Critical Points

In this section suppose F : Ω (⊂ Rn ) → R, where Ω is open.

Definition 19.5.1 The point a ∈ Ω is a local minimum point for F if for

some r > 0
F (a) ≤ F (x)
for all x ∈ Br (a).
A similar definition applies for local maximum points.
Inverse Function Theorem 269

Theorem 19.5.2 If F is C 1 and a is a local minimum or maximum point

for F , then
∂F ∂F
(a) = · · · = (a) = 0.
∂x1 ∂xn
Equivalently, ∇F (a) = 0.

Proof: Fix 1 ≤ i ≤ n. Let

g(t) = F (a + tei ) = F (a1 , . . . , ai−1 , ai + t, ai+1 , . . . , an ).
Then g : R → R and g has a local minimum (or maximum) at 0. Hence
g 0 (0) = 0.
But
∂F
g 0 (0) =
(a)
∂xi
by the chain rule, and so the result follows.

Definition 19.5.3 If ∇F (a) = 0 then a is a critical point for F .

Remark Every local maximum or minimum point is a critical point, but

not conversely. In particular, a may correspond to a “saddle point” of the
graph of F .
For example, if F (x, y) = x2 − y 2 , then (0, 0) is a critical point. See the
diagram before Definition 17.6.5.

19.6 Lagrange Multipliers

We are often interested in the problem of investigating the maximum and
minimum points of a real-valued function F restricted to some manifold M
in Rn .

Definition 19.6.1 Suppose M is a manifold in Rn . The function F : Rn →

R has a local minimum (maximum) at a ∈ M when F is constrained to M
if for some r > 0,
F (a) ≤ (≥) F (x)
for all x ∈ Br (a).

If F has a local (constrained) minimum at a ∈ M then it is intuitively

reasonable that the rate of change of F in any direction h in Ta M should be
zero. Since
Dh F (a) = ∇F (a) · h,
this means ∇F (a) is orthogonal to any vector in Ta M and hence belongs to
Na M . We make this precise in the following Theorem.

Theorem 19.6.2 (Method of Lagrange Multipliers) Let M be a man-

ifold in Rn given locally as the zero-level set of Φ : Rn → R` 11 .
11
Thus Φ is C 1 and for each x ∈ M the vectors ∇Φ1 (x), . . . , ∇Φ` (x) are linearly
independent.
270

Suppose
F : Rn → R
is C 1 and F has a constrained minimum (maximum) at a ∈ M . Then

X̀
∇F (a) = λj ∇Φj (a)
j=1

for some λ1 , . . . , λ` ∈ R called Lagrange Multipliers.

Equivalently, let H : Rn+` → R be defined by

H(x1 , . . . , xn , σ1 , . . . , σ` ) = F (x1 , . . . , xn )−σ1 Φ1 (x1 , . . . , xn )−. . .−σ` Φ` (x1 , . . . , xn ).

Then H has a critical point at a1 , . . . , an , λ1 , . . . , λ` for some λ1 , . . . , λ`

Proof: Suppose ψ : I → M where I is an open interval containing 0, ψ(0) =

a and ψ is C 1 .
Then F (ψ(t)) has a local minimum at t = 0 and so by the chain rule

X
n
∂F dψ i
0= (a) (0),
i=1 ∂xi dt

i.e.
∇F (a) ⊥ ψ 0 (0).

Since ψ 0 (0) can be any vector in Ta M , it follows ∇F (a) ∈ Na M . Hence

X̀
∇F (a) = λj ∇Φj (a)
j=1

for some λ1 , . . . , λ` . This proves the first claim.

For the second claim just note that

∂H ∂F X ∂Φj ∂H
= − σj , = −Φj .
∂xi ∂xi j ∂xi ∂σj

Since Φj (a) = 0 it follows that H has a critical point at a1 , . . . , an , λ1 , . . . , λ`

iff
∂F X̀ ∂Φj
(a) = λj (a)
∂xi j=1 ∂xi
for i = 1, . . . , n. That is,

X̀
∇F (a) = λj ∇Φj (a).
j=1
Inverse Function Theorem 271

Example Find the maximum and minimum points of

F (x, y, z) = x + y + 2z

on the ellipsoid
M = {(x, y, z) : x2 + y 2 + 2z 2 = 2}.

Solution: Let
Φ(x, y, z) = x2 + y 2 + 2z 2 − 2.
At a critical point there exists λ such that

∇F = λ∇Φ.

That is

1 = λ(2x)
1 = λ(2y)
2 = λ(4z).

Moreover
x2 + y 2 + 2z 2 = 2.
These four equations give
1 1 1 1 √
x= , y= , z= , = ± 2.
2λ 2λ 2λ λ
Hence
1
(x, y, z) = ± √ (1, 1, 1).
2
Since F is continuous and M is compact,
√ F must have a minimum and a
maximum point. Thus one of ±(1, 1, 1)/ 2 must be the minimum point and
the other the maximum point. A calculation gives
à !
1 √
F √ (1, 1, 1) = 2 2
2
à !
1 √
F √ (1, 1, 1) = −2 2.
− 2
√ √
Thus the minimum and maximum points are −(1, 1, 1)/ 2 and +(1, 1, 1)/ 2
respectively.
272
Bibliography

[An] H. Anton. Elementary Linear Algebra. 6th edn, 1991, Wiley.

[Ba] M. Barnsley. Fractals Everywhere. 1988, Academic Press.

[BM] G. Birkhoff and S. MacLane. A Survey of Modern Algebra. rev. edn,

1953, Macmillan.

[Br] V. Bryant. Metric Spaces, iteration and application. 1985, Cambridge

University Press.

[F] W. Fleming. Calculus of Several Variables. 2nd ed., 1977, Undergrad-

uate Texts in Mathematics, Springer-Verlag.

[La] S. R. Lay. Analysis, an Introduction to Proof. 1986, Prentice-Hall.

[Ma] B. Mandelbrot. The Fractal Geometry of Nature. 1982, Freeman.

[Me] E. Mendelson. Number Systems and the Foundations of Analysis 1973,

Academic Press.

[Ms] E.E. Moise. Introductory Problem Courses in Analysis and Toplogy.

1982, Springer-Verlag.

[Mo] G. P. Monro. Proofs and Problems in Calculus. 2nd ed., 1991, Carslaw
Press.

[PJS] H-O. Peitgen, H. Jürgens, D. Saupe. Fractals for the Classroom. 1992,
Springer-Verlag.

[BD] J. Bélair, S. Dubuc. Fractal Geometry and Analysis. NATO Proceed-

ings, 1989, Kluwer.

[Sm] K. T. Smith. Primer of Modern Analysis. 2nd ed., 1983, Undergradu-

ate Texts in Mathematics, Springer-Verlag.

[Sp] M. Spivak. Calculus. 1967, Addison-Wesley.

[St] K. Stromberg. An Introduction to Classical Real Analysis. 1981,

Wadsworth.

[Sw] E. W. Swokowski. Calculus. 5th ed., 1991, PWS-Kent.

273