Building Blocks

for
Theoretical Computer Science
(preliminary version 1.2)
Margaret M. Fleck
August 20, 2012
Contents
Preface xii
1 Math review 1
1.1 Some sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Pairs of reals . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Exponentials and logs . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Some handy functions . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Summations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.7 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 9
2 Logic 10
2.1 A bit about style . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Propositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Complex propositions . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Implication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Converse, contrapositive, biconditional . . . . . . . . . . . . . 14
2.6 Complex statements . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Logical Equivalence . . . . . . . . . . . . . . . . . . . . . . . . 16
i
CONTENTS ii
2.8 Some useful logical equivalences . . . . . . . . . . . . . . . . . 18
2.9 Negating propositions . . . . . . . . . . . . . . . . . . . . . . . 18
2.10 Predicates and Variables . . . . . . . . . . . . . . . . . . . . . 19
2.11 Other quantiers . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.12 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.13 Useful notation . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.14 Notation for 2D points . . . . . . . . . . . . . . . . . . . . . . 23
2.15 Negating statements with quantiers . . . . . . . . . . . . . . 24
2.16 Binding and scope . . . . . . . . . . . . . . . . . . . . . . . . 25
2.17 Variations in Notation . . . . . . . . . . . . . . . . . . . . . . 26
3 Proofs 27
3.1 Proving a universal statement . . . . . . . . . . . . . . . . . . 27
3.2 Another example of direct proof involving odd and even . . . . 29
3.3 Direct proof outline . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Proving existential statements . . . . . . . . . . . . . . . . . . 31
3.5 Disproving a universal statement . . . . . . . . . . . . . . . . 31
3.6 Disproving an existential statement . . . . . . . . . . . . . . . 32
3.7 Recap of proof methods . . . . . . . . . . . . . . . . . . . . . 33
3.8 Direct proof: example with two variables . . . . . . . . . . . . 33
3.9 Another example with two variables . . . . . . . . . . . . . . . 34
3.10 Proof by cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.11 Rephrasing claims . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.12 Proof by contrapositive . . . . . . . . . . . . . . . . . . . . . . 37
3.13 Another example of proof by contrapositive . . . . . . . . . . 38
CONTENTS iii
4 Number Theory 39
4.1 Factors and multiples . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Direct proof with divisibility . . . . . . . . . . . . . . . . . . . 40
4.3 Stay in the Set . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4 Prime numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 GCD and LCM . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.6 The division algorithm . . . . . . . . . . . . . . . . . . . . . . 43
4.7 Euclidean algorithm . . . . . . . . . . . . . . . . . . . . . . . 44
4.8 Pseudocode . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.9 A recursive version of gcd . . . . . . . . . . . . . . . . . . . . 46
4.10 Congruence mod k . . . . . . . . . . . . . . . . . . . . . . . . 47
4.11 Proofs with congruence mod k . . . . . . . . . . . . . . . . . . 48
4.12 Equivalence classes . . . . . . . . . . . . . . . . . . . . . . . . 48
4.13 Wider perspective on equivalence . . . . . . . . . . . . . . . . 50
4.14 Variation in Terminology . . . . . . . . . . . . . . . . . . . . . 51
5 Sets 52
5.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.2 Things to be careful about . . . . . . . . . . . . . . . . . . . . 53
5.3 Cardinality, inclusion . . . . . . . . . . . . . . . . . . . . . . . 54
5.4 Vacuous truth . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5 Set operations . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.6 Set identities . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.7 Size of set union . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.8 Product rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.9 Combining these basic rules . . . . . . . . . . . . . . . . . . . 60
CONTENTS iv
5.10 Proving facts about set inclusion . . . . . . . . . . . . . . . . 61
5.11 An abstract example . . . . . . . . . . . . . . . . . . . . . . . 63
5.12 An example with products . . . . . . . . . . . . . . . . . . . . 64
5.13 A proof using sets and contrapositive . . . . . . . . . . . . . . 65
5.14 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 66
6 Relations 67
6.1 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.2 Properties of relations: reexive . . . . . . . . . . . . . . . . . 69
6.3 Symmetric and antisymmetric . . . . . . . . . . . . . . . . . . 70
6.4 Transitive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 Types of relations . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.6 Proving that a relation is an equivalence relation . . . . . . . . 74
6.7 Proving antisymmetry . . . . . . . . . . . . . . . . . . . . . . 75
7 Functions and onto 76
7.1 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.2 When are functions equal? . . . . . . . . . . . . . . . . . . . . 78
7.3 What isnt a function? . . . . . . . . . . . . . . . . . . . . . . 79
7.4 Images and Onto . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.5 Why are some functions not onto? . . . . . . . . . . . . . . . . 81
7.6 Negating onto . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.7 Nested quantiers . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.8 Proving that a function is onto . . . . . . . . . . . . . . . . . 84
7.9 A 2D example . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.10 Composing two functions . . . . . . . . . . . . . . . . . . . . . 86
CONTENTS v
7.11 A proof involving composition . . . . . . . . . . . . . . . . . . 87
7.12 Variation in terminology . . . . . . . . . . . . . . . . . . . . . 87
8 Functions and one-to-one 89
8.1 One-to-one . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
8.2 Bijections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8.3 Pigeonhole Principle . . . . . . . . . . . . . . . . . . . . . . . 91
8.4 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.5 Further applications of permutations . . . . . . . . . . . . . . 93
8.6 Proving that a function is one-to-one . . . . . . . . . . . . . . 94
8.7 Composition and one-to-one . . . . . . . . . . . . . . . . . . . 95
8.8 Strictly increasing functions are one-to-one . . . . . . . . . . . 96
8.9 Making this proof more succinct . . . . . . . . . . . . . . . . . 97
8.10 Variation in terminology . . . . . . . . . . . . . . . . . . . . . 98
9 Graphs 99
9.1 Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.2 Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.3 Complete graphs . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.4 Cycle graphs and wheels . . . . . . . . . . . . . . . . . . . . . 102
9.5 Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
9.6 Subgraphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.7 Walks, paths, and cycles . . . . . . . . . . . . . . . . . . . . . 106
9.8 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.9 Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
9.10 Euler circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
CONTENTS vi
9.11 Bipartite graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.12 Variation in terminology . . . . . . . . . . . . . . . . . . . . . 112
10 2-way Bounding 113
10.1 Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
10.2 Pigeonhole point placement . . . . . . . . . . . . . . . . . . . 114
10.3 Graph coloring . . . . . . . . . . . . . . . . . . . . . . . . . . 116
10.4 Why care about graph coloring? . . . . . . . . . . . . . . . . . 117
10.5 Proving set equality . . . . . . . . . . . . . . . . . . . . . . . . 119
10.6 Variation in terminology . . . . . . . . . . . . . . . . . . . . . 120
11 Induction 121
11.1 Introduction to induction . . . . . . . . . . . . . . . . . . . . . 121
11.2 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
11.3 Why is this legit? . . . . . . . . . . . . . . . . . . . . . . . . . 123
11.4 Building an inductive proof . . . . . . . . . . . . . . . . . . . 124
11.5 Another example . . . . . . . . . . . . . . . . . . . . . . . . . 125
11.6 Some comments about style . . . . . . . . . . . . . . . . . . . 126
11.7 A geometrical example . . . . . . . . . . . . . . . . . . . . . . 127
11.8 Graph coloring . . . . . . . . . . . . . . . . . . . . . . . . . . 128
11.9 Postage example . . . . . . . . . . . . . . . . . . . . . . . . . 130
11.10Nim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
11.11Prime factorization . . . . . . . . . . . . . . . . . . . . . . . . 133
11.12Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 134
12 Recursive Denition 136
12.1 Recursive denitions . . . . . . . . . . . . . . . . . . . . . . . 136
CONTENTS vii
12.2 Finding closed forms . . . . . . . . . . . . . . . . . . . . . . . 137
12.3 Divide and conquer . . . . . . . . . . . . . . . . . . . . . . . . 139
12.4 Hypercubes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
12.5 Proofs with recursive denitions . . . . . . . . . . . . . . . . . 142
12.6 Inductive denition and strong induction . . . . . . . . . . . . 143
12.7 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 144
13 Trees 145
13.1 Why trees? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
13.2 Dening trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
13.3 m-ary trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
13.4 Height vs number of nodes . . . . . . . . . . . . . . . . . . . . 150
13.5 Context-free grammars . . . . . . . . . . . . . . . . . . . . . . 150
13.6 Recursion trees . . . . . . . . . . . . . . . . . . . . . . . . . . 156
13.7 Another recursion tree example . . . . . . . . . . . . . . . . . 158
13.8 Tree induction . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
13.9 Heap example . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
13.10Proof using grammar trees . . . . . . . . . . . . . . . . . . . . 162
13.11Variation in terminology . . . . . . . . . . . . . . . . . . . . . 163
14 Big-O 165
14.1 Running times of programs . . . . . . . . . . . . . . . . . . . . 165
14.2 Function growth: the ideas . . . . . . . . . . . . . . . . . . . . 166
14.3 Primitive functions . . . . . . . . . . . . . . . . . . . . . . . . 167
14.4 Proving a primitive function relationship . . . . . . . . . . . . 169
14.5 The formal denition . . . . . . . . . . . . . . . . . . . . . . . 170
CONTENTS viii
14.6 Applying the denition . . . . . . . . . . . . . . . . . . . . . . 170
14.7 Writing a big-O proof . . . . . . . . . . . . . . . . . . . . . . . 171
14.8 Sample disproof . . . . . . . . . . . . . . . . . . . . . . . . . . 172
14.9 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 173
15 Algorithms 174
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
15.2 Basic data structures . . . . . . . . . . . . . . . . . . . . . . . 174
15.3 Nested loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
15.4 Merging two lists . . . . . . . . . . . . . . . . . . . . . . . . . 177
15.5 A reachability algorithm . . . . . . . . . . . . . . . . . . . . . 178
15.6 Binary search . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
15.7 Mergesort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
15.8 Tower of Hanoi . . . . . . . . . . . . . . . . . . . . . . . . . . 182
15.9 Multiplying big integers . . . . . . . . . . . . . . . . . . . . . 184
16 NP 187
16.1 Finding parse trees . . . . . . . . . . . . . . . . . . . . . . . . 187
16.2 What is NP? . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
16.3 Circuit SAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
16.4 What is NP complete? . . . . . . . . . . . . . . . . . . . . . . 192
16.5 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 193
17 Proof by Contradiction 194
17.1 The method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
17.2

2 is irrational . . . . . . . . . . . . . . . . . . . . . . . . . . 195
17.3 There are innitely many prime numbers . . . . . . . . . . . . 196
CONTENTS ix
17.4 Lossless compression . . . . . . . . . . . . . . . . . . . . . . . 197
17.5 Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
18 Sets of Sets 199
18.1 Sets containing sets . . . . . . . . . . . . . . . . . . . . . . . . 199
18.2 Powersets and set-valued functions . . . . . . . . . . . . . . . 201
18.3 Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
18.4 Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
18.5 Applying the combinations formula . . . . . . . . . . . . . . . 205
18.6 Combinations with repetition . . . . . . . . . . . . . . . . . . 206
18.7 Identities for binomial coecients . . . . . . . . . . . . . . . . 207
18.8 Binomial Theorem . . . . . . . . . . . . . . . . . . . . . . . . 208
18.9 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 209
19 State Diagrams 210
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
19.2 Wolf-goat-cabbage puzzle . . . . . . . . . . . . . . . . . . . . . 212
19.3 Phone lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
19.4 Representing functions . . . . . . . . . . . . . . . . . . . . . . 215
19.5 Transition functions . . . . . . . . . . . . . . . . . . . . . . . . 215
19.6 Shared states . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
19.7 Counting states . . . . . . . . . . . . . . . . . . . . . . . . . . 219
19.8 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 220
20 Countability 221
20.1 The rationals and the reals . . . . . . . . . . . . . . . . . . . . 221
20.2 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
CONTENTS x
20.3 Cardinality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
20.4 Cantor Schroeder Bernstein Theorem . . . . . . . . . . . . . . 223
20.5 More countably innite sets . . . . . . . . . . . . . . . . . . . 224
20.6 P(N) isnt countable . . . . . . . . . . . . . . . . . . . . . . . 225
20.7 More uncountability results . . . . . . . . . . . . . . . . . . . 226
20.8 Uncomputability . . . . . . . . . . . . . . . . . . . . . . . . . 227
20.9 Variation in notation . . . . . . . . . . . . . . . . . . . . . . . 229
21 Planar Graphs 230
21.1 Planar graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
21.2 Faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
21.3 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
21.4 Proof of Eulers formula . . . . . . . . . . . . . . . . . . . . . 233
21.5 Some corollaries of Eulers formula . . . . . . . . . . . . . . . 234
21.6 K
3,3
is not planar . . . . . . . . . . . . . . . . . . . . . . . . . 235
21.7 Kuratowskis Theorem . . . . . . . . . . . . . . . . . . . . . . 236
21.8 Coloring planar graphs . . . . . . . . . . . . . . . . . . . . . . 239
21.9 Application: Platonic solids . . . . . . . . . . . . . . . . . . . 240
A Jargon 243
A.1 Strange technical terms . . . . . . . . . . . . . . . . . . . . . . 243
A.2 Odd uses of normal words . . . . . . . . . . . . . . . . . . . . 244
A.3 Constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
A.4 Unexpectedly normal . . . . . . . . . . . . . . . . . . . . . . . 247
B Acknowledgements and Supplementary Readings 249
CONTENTS xi
C Where did it go? 252
Preface
This book teaches two dierent sorts of things, woven together. It teaches
you how to read and write mathematical proofs. It also provides a survey
of basic mathematical objects, notation, and techniques which will be useful
in later computer science courses. These include propositional and predicate
logic, sets, functions, relations, modular arithmetic, counting, graphs, and
trees.
Why learn formal mathematics?
Formal mathematics is relevant to computer science in several ways. First, it
is used to create theoretical designs for algorithms and prove that they work
correctly. This is especially important for methods that are used frequently
and/or in applications where we dont want failures (aircraft design, Pentagon
security, ecommerce). Only some people do these designs, but many people
use them. The users need to be able to read and understand how the designs
work.
Second, the skills you learn in doing formal mathematics correspond
closely to those required to design and debug programs. Both require keeping
track of what types your variables are. Both use inductive and/or recursive
design. And both require careful proofreading skills. So what you learn in
this class will also help you succeed in practical programming courses.
Third, when you come to design complex real software, youll have to
document what youve done and how it works. This is hard for many peo-
ple to do well, but its critical for the folks using your software. Learning
how to describe mathematical objects clearly will translate into better skills
xii
PREFACE xiii
describing computational objects.
Everyone can do proofs
You probably think about proofs as something created by brilliant young
theoreticians. Some of you are brilliant young theoreticians and youll think
this stu is fun because its what you naturally like to do. However, many
of you are future software and hardware engineers. Some of you may never
think of formal mathematics as fun. Thats ok. We understand. Were
hoping to give you a sense of why its pretty and useful, and enough uency
with it to communicate with the theoretical side of the eld.
Many people think that proofs involve being incredibly clever. That is
true for some steps in some proofs. Thats where it helps to actually be a
brilliant young theoretician. But many proofs are very routine. And many
steps in clever proofs are routine. So theres quite a lot of proofs that all of
us can do. And we can all read, understand, and appreciate the clever bits
that we didnt think up ourselves.
In other words, dont be afraid. You can do proofs at the level required
for this course, and future courses in theoretical computer science.
Is this book right for you?
This book is designed for students who have taken an introductory program-
ming class of the sort intended for scientists or engineers. Algorithms will
be presented in pseudo-code, so it does not matter which programming
language you have used. But you should have written programs that ma-
nipulate the contents of arrays e.g. sort an array of numbers. You should
also have written programs that are recursive, i.e. in which a function (aka
procedure aka method) calls itself.
Well also assume that you have taken one semester of calculus. Though
we will make little direct reference to calculus, we will assume a level of u-
ency with precalculus (algebra, geometry, logarithms, etc) that is typically
developed during while taking calculus. If you already have signicant expe-
PREFACE xiv
rience writing proofs, including inductive proofs, this book may be too easy
for you. You may wish to read some of the books listed as supplementary
readings in Appendix B.
For instructors
This text is designed for a broad range of computer science majors, ranging
from budding young theoreticians to capable programmers with very little
interest in theory. It assumes only limited mathematical maturity, so that it
can be used very early in the major. Therefore, a central goal is to explain
the process of proof construction clearly to students who cant just pick it
up by osmosis.
This book is designed to be succinct, so that students will read and ab-
sorb its contents. Therefore, it includes only core concepts and a selection
of illustrative examples, with the expectation that the instructor will pro-
vide supplementary materials as needed (see Appendix B for useful follow-on
books) and that students can look up a wider range of facts, denitions, and
pictures, e.g. on the internet.
Although the core topics in this book are old and established, terminol-
ogy and notation have changed over the years and vary somewhat between
authors. To avoid overloading students, I have chosen one clean, modern
version of notation, denitions, and terminology to use consistently in the
main text. Common variations are documented at the end of each chapter. If
students understand the underlying concepts, they should have no diculty
adapting when they encounter dierent conventions in other books.
Many traditional textbooks do a careful and exhaustive treatment of each
topic from rst principles, including foundational constructions and theorems
which prove that the conceptual system is well-formed. However, the most
abstract parts of many topics are hard for beginners to absorb and the impor-
tance of the foundational development is lost on most students at this level.
The early parts of this textbook remove most of this clutter, to focus more
clearly on the key concepts and constructs. At the end, we revisit certain
topics to look at more abstract examples and selected foundational issues.
See Appendix C for a quick guide to topic rearrangement.
Chapter 1
Math review
This book assumes that you understood precalculus when you took it. So you
used to know how to do things like factoring polynomials, solving high school
geometry problems, using trigonometric identities. However, you probably
cant remember it all cold. Many useful facts can be looked up (e.g. on the
internet) as you need them. This chapter reviews concepts and notation that
will be used a lot in this book, as well as a few concepts that may be new
but are fairly easy.
1.1 Some sets
Youve all seen sets, though probably a bit informally. Well get back to some
of the formalism in a couple weeks. Meanwhile, here are the names for a few
commonly used sets of numbers:
Z = {. . . , 3, 2, 1, 0, 1, 2, 3, . . .} is the integers.
N = {0, 1, 2, 3, . . .} is the non-negative integers, also known as the
natural numbers.
Z
+
= {1, 2, 3, . . .} is the positive integers.
R is the real numbers
1
CHAPTER 1. MATH REVIEW 2
Q is the rational numbers
C is the complex numbers
Notice that a number is positive if it is greater than zero, so zero is not
positive. If you wish to exclude negative values, but include zero, you must
use the term non-negative. In this book, natural numbers are dened to
include zero. Notice also that the real numbers contain the integers, so a
real number is not required to have a decimal part. And, nally, innity
is not a number in standard mathematics.
Remember that the rational numbers are the set of fractions
p
q
where q
cant be zero and we consider two fractions to be the same number if they
are the same when you reduce them to lowest terms.
The real numbers contain all the rationals, plus irrational numbers such
as

2 and . (Also e, but we dont use e much in this end of computer

science.) Well assume that

x returns (only) the positive square root of x.
The complex numbers are numbers of the form a +bi, where a and b are
real numbers and i is the square root of -1. Some of you from ECE will be
aware that there are many strange and interesting things that can be done
with the complex numbers. We wont be doing that here. Well just use
them for a few examples and expect you to do very basic things, mostly easy
consequences of the fact that i =

1.
If we want to say that the variable x is a real, we write x R. Similarly
y Z means that y is not an integer. In calculus, you dont have to think
much about the types of your variables, because they are largely reals. In
this class, we handle variables of several dierent types, so its important to
state the types explicitly.
It is sometimes useful to select the real numbers in some limited range,
called an interval of the real line. We use the following notation:
closed interval: [a, b] is the set of real numbers from a to b, including a
and b
open interval: (a, b) is the set of real numbers from a to b, not including
a and b
CHAPTER 1. MATH REVIEW 3
half-open intervals: [a, b) and (a, b] include only one of the two end-
points.
1.2 Pairs of reals
The set of all pairs of reals is written R
2
. So it contains pairs like (2.3, 4.7).
Similarly, the set R
3
contains triples of reals such as 8, 7.3, 9. In a computer
program, we implement a pair of reals using an object or struct with two
elds.
Suppose someone presents you with a pair of numbers (x, y) or one of the
corresponding data structures. Your rst question should be: what is the
pair intended to represent? It might be
a point in 2D space
a complex number
a rational number (if the second coordinate isnt zero)
an interval of the real line
The intended meaning aects what operations we can do on the pairs.
For example, if (x, y) is an interval of the real line, its a set. So we can write
things like z (x, y) meaning z is in the interval (x, y). That notation would
be meaningless if (x, y) is a 2D point or a number.
If the pairs are numbers, we can add them, but the result depends on
what they are representing. So (x, y) + (a, b) is (x + a, y + b) for 2D points
and complex numbers. But it is (xb +ya, yb) for rationals.
Suppose we want to multiply (x, y)(a, b) and get another pair as output.
Theres no obvious way to do this for 2D points. For rationals, the formula
would be (xa, yb), but for complex numbers its (xa yb, ya +xb).
Stop back and work out that last one for the non-ECE students, using
more familiar notation. (x+yi)(a+bi) is xa+yai +xbi +byi
2
. But i
2
= 1.
So this reduces to (xa yb) + (ya +xb)i.
CHAPTER 1. MATH REVIEW 4
Oddly enough, you can also multiply two intervals of the real line. This
carves out a rectangular region of the 2D plane, with sides determined by
the two intervals.
1.3 Exponentials and logs
Suppose that b is any real number. We all know how to take integer powers
of b, i,e. b
n
is b multiplied by itself n times. Its not so clear how to precisely
dene b
x
, but weve all got faith that it works (e.g. our calculators produce
values for it) and its a smooth function that grows really fast as the input
gets bigger and agrees with the integer denition on integer inputs.
Here are some special cases to know
b
0
is one for any b.
b
0.5
is

b
b
1
is
1
b
And some handy rules for manipulating exponents:
b
x
b
y
= b
x+y
a
x
b
x
= (ab)
x
(b
x
)
y
= b
xy
b
(x
y
)
= (b
x
)
y
Suppose that b > 1. Then we can invert the function b
x
, to get the func-
tion log
b
x (logarithm of x to the base b). Logarithms appear in computer
science as the running times of particularly fast algorithms. They are also
used to manipulate numbers that have very wide ranges, e.g. probabilities.
Notice that the log function takes only positive numbers as inputs. In this
class, log x with no explicit base always means log
2
x because analysis of
computer algorithms makes such heavy use of base-2 numbers and powers of
2.
CHAPTER 1. MATH REVIEW 5
Useful facts about logarithms include:
b
log
b
(x)
= x
log
b
(xy) = log
b
x + log
b
y
log
b
(x
y
) = y log
b
x
log
b
x = log
a
xlog
b
a
In the change of base formula, its easy to forget whether the last term
should be log
b
a or log
a
b. To gure this out on the y, rst decide which of
log
b
x and log
a
x is larger. You then know whether the last term should be
larger or smaller than one. One of log
b
a and log
a
b is larger than one and
the other is smaller: gure out which is which.
More importantly, notice that the multiplier to change bases is a con-
stant, i.e doesnt depend on x. So it just shifts the curve up and down
without really changing its shape. Thats an extremely important fact that
you should remember, even if you cant reconstruct the precise formula. In
many computer science analyses, we dont care about constant multipliers.
So the fact that base changes simply multiply by a constant means that we
frequently dont have to care what the base actually is. Thus, authors often
write log x and dont specify the base.
1.4 Some handy functions
The factorial function may or may not be familiar to you, but is easy to
explain. Suppose that k is any positive integer. Then k factorial, written k!,
is the product of all the positive integers up to and including k. That is
k! = 1 2 3 . . . (k 1) k
For example, 5! = 1 2 3 4 5 = 120. 0! is dened to be 1.
The max function returns the largest of its inputs. For example, suppose
we dene f(x) = max(x
2
, 7). Then f(3) = 9 but f(2) = 7. The min function
is similar.
CHAPTER 1. MATH REVIEW 6
The oor and ceiling functions are heavily used in computer science,
though not in many areas other of science and engineering. Both functions
take a real number x as input and return an integer near x. The oor func-
tion returns the largest integer no bigger than x. In other words, it converts
x to an integer, rounding down. This is written x. If the input to oor
is already an integer, it is returned unchanged. Notice that oor rounds
downward even for negative numbers. So:
3.75 = 3
3 = 3
3.75 = 4
The ceiling function, written x, is similar, but rounds upwards. For
example:
3.75 = 4
3 = 3
3.75 = 3
Most programming languages have these two functions, plus a function
that rounds to the nearest integer and one that truncates i.e. rounds
towards zero. Round is often used in statistical programs. Truncate isnt
used much in theoretical analyses.
1.5 Summations
If a
i
is some formula that depends on i, then
n

i=1
a
i
= a
1
+a
2
+a
3
+. . . +a
n
For example
n

i=1
1
2
i
=
1
2
+
1
4
+
1
8
. . . +
1
2
n
CHAPTER 1. MATH REVIEW 7
Products can be written with a similar notation, e.g.
n

k=1
1
k
=
1
1

1
2

1
3
. . .
1
n
Certain sums can be re-expressed in closed form i.e. without the sum-
mation notation. For example:
n

i=1
1
2
i
= 1
1
2
n
In Calculus, you may have seen the innite version of this sum, which
converges to 1. In this class, were always dealing with nite sums, not
innite ones.
If you modify the start value, so we start with the zeroth term, we get
the following variation on this summation. Always be careful to check where
your summation starts.
n

i=0
1
2
i
= 2
1
2
n
Many reference books have tables of useful formulas for summations that
have simple closed forms. Well see them again when we cover mathematical
induction and see how to formally prove that they are correct.
Only a very small number of closed forms need to be memorized. One
example is
n

i=1
i =
n(n + 1)
2
Heres one way to convince yourself its right. On graph paper, draw a
box n units high and n+1 units wide. Its area is n(n+1). Fill in the leftmost
part of each row to represent one term of the summation: the left box in the
rst row, the left two boxes in the second row, and so on. This makes a little
triangular pattern which lls exactly half the box.
CHAPTER 1. MATH REVIEW 8
1.6 Strings
Youve probably seen strings in introductory programming, e.g. when pro-
cessing input from the user or a disk le. In computer science, a string is a
nite-length sequence of characters and its length is the number of characters
in it. E.g. pineapple is a string of length 9.
The special symbol is used for the string containing no characters, which
has length 0. Zero-length strings may seem odd, but they appear frequently
in computer programs that manipulate text data. For example, a program
that reads a line of input from the keyboard will typically receive a zero-
length string if the user types ENTER twice.
We specify concatenation of strings, or concatenation of strings and indi-
vidual characters, by writing them next to each other. For example, suppose
that = blue and = cat. Then would be the string bluecat and s
be the string cats.
A bit string is a string consisting of the characters 0 and 1. If A is a
set of characters, then A

is the set of all (nite-length) strings containing

only characters from A. For example, if A contains all lower-case alphabetic
characters, then A

contains strings like e, onion, and kkkkmmmmmbb. It also

contains the empty string .
Sometimes we want to specify a pattern for a set of similar strings. We
often use a shorthand notation called regular expressions. These look much
like normal strings, but we can use two operations:
a | b means either one of the characters a and b.
a

means zero or more copies of the character a.

Parentheses are used to show grouping.
So, for example, ab

species all strings consisting of an a followed by

zero or more bs: a, ab, abb, and so forth. c(b | a)

c species all strings

consisting of one c, followed by zero or more characters that are either as or
bs, followed one c. E.g. cc, cac, cbbac, and so forth.
CHAPTER 1. MATH REVIEW 9
1.7 Variation in notation
Mathematical notation is not entirely standardized and has changed over the
years, so you will nd that dierent subelds and even dierent individual
authors use slightly dierent notation. These variation in notation sections
tell you about synonyms and changes in notation that you are very likely to
encounter elsewhere. Always check carefully which convention an author is
using. When doing problems on homework or exams for a class, follow the
house style used in that particular class.
In particular, authors dier as to whether zero is in the natural numbers.
Moreover,

1 is named j over in ECE and physics, because i is used for

current. Outside of computer science, the default base for logarithms is
usually e, because this makes a number of continuous mathematical formulas
work nicely (just as base 2 makes many computer science formulas work
nicely).
In standard denitions of the real numbers, innity is not a number.
People are sometimes confused because is sometimes used in notation
where numbers also occur, e.g. in dening limits in calculus. There are
non-standard number systems that include innite numbers. And points at
innity are sometimes introduced in discussions of 2D and 3D geometry. We
wont make use of such extensions.
Regular expressions are quite widely used in both theory and practical
applications. As a result there are many variations on the notation. There are
also other operations that make it easy to specify a wider range of patterns.
Chapter 2
Logic
This chapter covers propositional logic and predicate logic at a basic level.
Some deeper issues will be covered later.
2.1 A bit about style
Writing mathematics requires two things. You need to get the logical ow of
ideas correct. And you also need to express yourself in standard style, in a
way that is easy for humans (not computers) to read. Mathematical style is
best taught by example and is similar to what happens in English classes.
Mathematical writing uses a combination of equations and also parts that
look supercially like English. Mathematical English is almost like normal
English, but diers in some crucial ways. You are probably familiar with the
fact that physicists use terms like force dierently from everyone else. Or
the fact that people from England think that paran is a liquid whereas
that word refers to a solid substance in the US. We will try to highlight the
places where mathematical English isnt like normal English.
You will also learn how to make the right choice between an equation and
an equivalent piece of mathematical English. For example, is a shorthand
symbol for and. The shorthand equations are used when we want to look
at a complex structure all at once, e.g. discuss the logical structure of a proof.
When writing the proof itself, its usually better to use the longer English
10
CHAPTER 2. LOGIC 11
equivalents, because the result is easier to read. There is no hard-and-fast
line here, but well help make sure you dont go too far in either direction.
2.2 Propositions
Two systems of logic are commonly used in mathematics: propositional logic
and predicate logic. Well start by covering propositional logic.
A proposition is a statement which is true or false (but never both!). For
example, Urbana is in Illinois or 2 15. It cant be a question. It also
cant contain variables, e.g. x 9 isnt a proposition. Sentence fragments
without verbs (e.g. bright blue owers) or arithmetic expressions (e.g.
5 + 17), arent propositions because they dont state a claim.
The lack of variables prevents propositional logic from being useful for
very much, though it has some applications in circuit analysis, databases, and
articial intelligence. Predicate logic is an upgrade that adds variables. We
will mostly be using predicate logic in this course. We just use propositional
logic to get started.
2.3 Complex propositions
Statements can be joined together to make more complex statements. For
example, Urbana is in Illinois and Margaret was born in Wisconsin. To
talk about complex sequences of statements without making everything too
long, we represent each simple statement by a variable. E.g. if p is Urbana
is in Illinois and q is Margaret was born in Wisconsin, then the whole
long statement would be p and q. Or, using shorthand notation p q.
The statement p q is true when both p and q are true. We can express
this with a truth table:
p q p q
T T T
T F F
F T F
F F F
CHAPTER 2. LOGIC 12
Similarly, p is the shorthand for not p. In our example, p would be
Urbana is not in Illinois. p is true exactly when p is false.
p q is the shorthand for p or q, which is true when either p or q is
true. Notice that it is also true when both p and q are true, i.e. its true table
is:
p q p q
T T T
T F T
F T T
F F F
When mathematicians use or, they always intend it to be understood with
this inclusive reading.
Normal English sometimes follows the same convention, e.g. you need
to wear a green hat or a blue tie allows you to do both. But normal English
is dierent from mathematical English, in that or sometimes excludes the
possibility that both statements are true. For example, Clean up your room
or you wont get desert strongly suggests that if you do clean up your room,
you will get desert. So normal English or sometimes matches mathematical
or and sometimes another operation called exclusive or dened by
p q p q
T T F
T F T
F T T
F F F
Exclusive or has some important applications in computer science, espe-
cially in encoding strings of letters for security reasons. However, we wont
see it much in this class.
2.4 Implication
Two propositions p and q can also be joined into the conditional statement.
if p, then q. The proposition after the if (p in this case) is called the
hypothesis and the proposition after then (q in this example) is called
CHAPTER 2. LOGIC 13
the conclusion. As in normal English, there are a number of alternative
ways to phrase the statement if p, then q, e.g. p implies q or q follows
from p.
The shorthand for this conditional is p q and its truth table is
p q p q
T T T
T F F
F T T
F F T
For example, If Obama is president, then Obama lives in the White
House is true (at least in 2010). But If Obama is president, then 2 > 4
is false. All the examples tend to be a bit articial, because we dont have
variables yet.
In normal English, we tend not to use conditionals in which the if
part is false. E.g. If Bush is president, then Urbana is in Illinois. In
mathematical English, such statements occur more often. And, worse, they
are always considered true, no matter whether the then part is true or
false. For example, this statement is true: If Bush is president, then 2 > 4.
The easiest way to remember the right output values for this operation is
to remember that the value is false in exactly one case: when p is true and
q is false.
Normal English requires that conditional sentences have some sort of
causal connection between the two propositions, i.e. one proposition is true
because the other is true. E.g. If Helen learns to write C++, she will
get a good job. It would seem odd if we said If Urbana is in Illinois, then
Margaret was born in Wisconsin. because theres no reason why one follows
from the other. In mathematical English, this statement is just ne: there
doesnt have to be any causal connection.
In normal English if/then statements, there is frequently a ow of time
involved. Unless we make a special eort to build a model of time, proposi-
tional logic is timeless. This makes the English motivating examples slightly
awkward. Its not a big problem in mathematics, because mathematical
proofs normally discuss a world that is static. It has a cast of characters
(e.g. variables, sets, functions) with a xed set of properties, and we are just
CHAPTER 2. LOGIC 14
reasoning about what those properties are. Only very occasionally do we
talk about taking an object and modifying it.
In computer programming, we often see things that look like conditional
statements, e.g. if x > 0, then increment y. But these are commands for
the computer to do something, changing its little world. whereas the similar-
looking mathematical statements are timeless. Formalizing what it means
for a computer program to do what its supposed to requires modelling
how the world changes over time. Youll see this in later CS classes.
2.5 Converse, contrapositive, biconditional
The converse of p q is q p. The two statements are not equivalent. To
see this, compare the previous truth table with this one:
p q q p
T T T
T F T
F T F
F F T
The converse mostly occurs in two contexts. First, getting the direction
of implication backwards is a common bug in writing proofs. That is, us-
ing the converse rather than the original statement. This is a bug because
implications frequently hold in only one direction.
Second, the phrase p implies q, and conversely means that p and q
are true under exactly the same conditions. The shorthand for this is the
biconditional operator p q.
p q q p
T T T
T F F
F T F
F F T
Another common way to phrase the biconditional is p if and only if q.
The contrapositive of p q is formed by swapping the roles of p and
q and negating both of them to get q p. Unlike the converse, the
CHAPTER 2. LOGIC 15
contrapositive is equivalent to the original statement. Heres a truth table
showing why:
p q q p q p
T T F F T
T F T F F
F T F T T
F F T T T
To gure out the last column of the table, recall that q p will be
false in only one case: when the hypothesis (q) is true and the conclusion
(p) is false.
Lets consider what these variations look like in an English example:
If its below zero, my car wont start.
converse: If my car wont start, its below zero
contrapositive: If my car will start, then its not below zero.
2.6 Complex statements
Very complex statements can be made using combinations of connectives.
E.g. If its below zero or my car does not have gas, then my car wont start
and I cant go get groceries. This example has the form
(p q) (r s)
The shorthand notation is particularly useful for manipulating complicated
statements, e.g. guring out the negative of a statement.
When you try to read a complex set of propositions all glued together
with connectives, there is sometimes a question about which parts to group
together rst. English is a bit vague about the rules. So, for example, in
the previous example, you need to use common sense to gure out that I
cant go get groceries is intended to be part of the conclusion of the if/then
statement.
CHAPTER 2. LOGIC 16
In mathematical shorthand, there are conventions about which parts to
group together rst. In particular, you apply the not operators rst, then
the and and or. Then you take the results and do the implication op-
erations. This is basically similar to the rules in (say) high-school algebra.
Use parentheses if you intend the reader to group things dierently, or if you
arent sure that your reader will group your statement as you intended.
You can build truth tables for complex statements, e.g.
p q r q r (q r) p
T T T T T
T F T F T
F T T T F
F F T F T
T T F F T
T F F F T
F T F F T
F F F F T
Truth tables are a nice way to show equivalence for compound proposi-
tions which use only 2-3 variables. However, if there are k variables, your
table needs 2
k
lines to cover all possible combinations of input truth values.
This is cumbersome for larger numbers of variables.
2.7 Logical Equivalence
Two (simple or compound) propositions p and q are logically equivalent if
they are true for exactly the same input values. The shorthand notation for
this is p q. One way to establish logical equivalence is with a truth table.
For example, we saw that implication has the truth table:
p q p q
T T T
T F F
F T T
F F T
A frequently useful fact is that p q is logically equivalent to p q.
CHAPTER 2. LOGIC 17
To show this, we build the truth table for p q and notice that the output
values exactly match those for p q.
p q p p q
T T F T
T F F F
F T T T
F F T T
Two very well-known equivalences are De Morgans Laws. These state
that (p q) is equivalent to p q. and that (p q) is equivalent to
p q. Similar rules in other domains (e.g. set theory) are also called De
Morgans Laws. They are especially helpful, because they tell you how to
simplify the negation of a complex statement involving and and or.
We can show this easily with another truth table:
p q p q p q (p q) p q
T T F F T F F
T F F T T F F
F T T F T F F
F F T T F T T
T and F are special constant propositions with no variables that are,
respectively, always true and always false. So, since p p is always false,
we have the following equivalence:
p p F
Notice that, in mathematics, the equal operator = can only be applied
to objects such as numbers. When comparing logical expressions that re-
turn true/false values, you must use . If use to create complex logical
equations, use indenting and whitespace to make sure the result is easy to
read.
CHAPTER 2. LOGIC 18
2.8 Some useful logical equivalences
It is easy to nd (e.g. on the internet) long tables of useful logical equiva-
lences. Most of them are commonsense and/or similar to rules from algebra.
For example, and are commutative, e.g. p q q p.
The distributive laws, however, work slightly dierently from those in
algebra. In algebra we have one rule:
a(b +c) = ab +ac
where as in logic we have two rules:
p (q r) (p q) (p r)
p (q r) (p q) (p r)
So, in logic, you can distribute either operator over the other. Also,
arithmetic has a clear rule that multiplication is done rst, so the righthand
side doesnt require parentheses. The order of operations is less clear for the
logic, so more parentheses are required.
2.9 Negating propositions
An important use of logical equivalences is to help you correctly state the
negation of a complex proposition, i.e. what it means for the complex propo-
sition not to be true. This is important when you are trying to prove a claim
false or convert a statement to its contrapositive. Also, looking at the nega-
tion of a denition or claim is often helpful for understanding precisely what
the denition or claim means. Having clear mechanical rules for negation is
important when working with concepts that are new to you, when you have
only limited intuitions about what is correct.
For example, suppose we have a claim like If M is regular, then M is
paracompact or M is not Lindelof. Im sure you have no idea whether this
is even true, because it comes from a math class you are almost certain not
to have taken. However, you can gure out its negation.
CHAPTER 2. LOGIC 19
First, lets convert the claim into shorthand so we can see its structure.
Let r be M is regular, p be M is paracompact, and l be M is Lindelof.
Then the claim would be r (p l).
The negation of r (pl) is (r (pl)). However, to do anything
useful with this negated expression, we normally need to manipulate it into an
equivalent expression that has the negations on the individual propositions.
The key equivalences used in doing this manipulation are:
(p) p
(p q) p q
(p q) p q
(p q) p q.
The last of these follows from an equivalence we saw above: p q pq
plus one of DeMorgans laws.
So we have
(r (p l)) r (p l) r p l r p l
So the negation of our original claim is M is regular and M is not
paracompact and M is Lindelof. Knowing the mechanical rules helps you
handle situations where your logical intuitions arent fully up to the task of
just seeing instinctively what the negation should look like. Mechanical rules
are also very helpful when you are tired or stressed (e.g. during an exam).
Notice that weve also derived a new logical equivalence (r (pl))
r p l by applying a sequence of known equivalences. This is how we
establish new equivalences when truth tables get unwieldy.
2.10 Predicates and Variables
Propositions are a helpful beginning but too rigid to represent most of the
interesting bits of mathematics. To do this, we need predicate logic, which
CHAPTER 2. LOGIC 20
allows variables and predicates that take variables as input. Well get started
with predicate logic now, but delay covering some of the details until they
become relevant to the proofs were looking at.
A predicate is a statement that becomes true or false if you substitute
in values for its variables. For example, x
2
10 or y is winter hardy.
Suppose we call these P(x) and Q(y). Then Q(y) is true if y is mint but
not if y is tomato.
1
If we substitute concrete values for all the variables in a predicate, were
back to having a proposition. That wasnt much use, was it?
The main use of predicates is to make general statements about what
happens when you substitute a variety of values for the variables. For exam-
ple:
P(x) is true for every x
For example, For every integer x, x
2
10 (false).
Consider For all x, 2x x. Is this true or false? This depends on what
values x can have. Is x any integer? In that case the claim is false. But if x
is supposed to be a natural number, then the claim is true.
In order to decide whether a statement involving quantiers is true, you
need to know what types of values are allowed for each variable. Good style
requires that you state the type of the variable explicitly when you introduce
it, e.g. For all natural numbers x, 2x x. Exceptions involve cases
where the type is very, very clear from the context, e.g. when a whole long
discussion is all about (say) the integers. If you arent sure, a redundant type
statement is a minor problem whereas a missing type statement is sometimes
a big problem.
2.11 Other quantiers
The general idea of a quantier is that it expresses how many of the values
in the domain make the claim true. Normal English has a wide range of
1
A winter hardy plant is a plant that can survive the winter in a cold climate, e.g. here
in central Illinois.
CHAPTER 2. LOGIC 21
quantiers which tell you roughly how many of the values work, e.g. some,
a couple, a few, many, most. For example, most students in this
class have taken a programming class.
By contrast, mathematics uses only three quantiers, one of which is used
rarely. Weve seen the universal quantier for all. The other common one
is the existential quantier there exists, as in
There is an integer x such that x
2
= 0.
In normal English, when we say that there is an object with some prop-
erties, this tends to imply that theres only one or perhaps only a couple. If
there were many objects with this property, we normally expect the speaker
to say so. So it would seem odd to say
There is an integer x such that x
2
> 0.
Or
There exists an integer x such that 5 < x < 100.
Mathematicians, however, are happy to say things like that. When they
say there exists an x, with certain properties, they mean that there exists
at least one x with those properties. They are making no claims about how
many such xs there are.
However, it is sometimes important to point out when one and only one x
has some set of properties. The mathematical jargon for this uses the unique
existence quantier, as in:
There is a unique integer x such that x
2
= 0.
Mathematicians use the adjective unique to mean that theres only one
such object (similar to the normal usage but not quite the same).
CHAPTER 2. LOGIC 22
2.12 Notation
The universal quantier has the shorthand notation . For example,
x R, x
2
+ 3 0
In this sentence, is the quantier. x R declares the variable and the set
(R) from which its values can be taken, called its domain or its replacement
set. As computer scientists, we can also think of this as declaring the type
of x, just as in a computer program. Finally, x
2
+ 3 0 is the predicate.
The existential quantier is written , e.g. y R, y =

2. Notice that
we dont write such that when the quantier is in shorthand. The unique
existence quantier is written ! as in !x R, x
2
= 0. When existential
quantiers are written in English, rather than shorthand, we need to add the
phrase such that to make the English sound right, e.g.
There exists a real number y such that y =

2.
Theres no deep reason for adding such that. Its just a quirk about how
mathematical English is written, which you should copy so that your written
mathematics looks professional. Such that is sometimes abbreviated s.t.
2.13 Useful notation
If you want to state a claim about two numbers, you can use two quantiers
as in:
x R, y R, x +y x
This is usually abbreviated to
x, y R, x +y x
This means for all real numbers x and y, x +y x (which isnt true).
CHAPTER 2. LOGIC 23
In such a claim, the two variables x and y might contain dierent values,
but its important to realize that they might also be equal. For example, the
following sentence is true:
x, y Z, x y = 0
We saw above that the statement if p, then q has the contrapositive
if q, then p. This transformation can be extended to statements with a
quantier (typically universal). For example, the statement
x, if p(x), then q(x)
would have a contrapositive
x, if q(x), then p(x)
Notice that the quantier stays the same: we only transform the if/then
statement inside it.
2.14 Notation for 2D points
When writing mathematics that involves 2D points and quantiers, you have
several notational options. You can write something like x, y Z (for any
integers x and y) and then later refer to the pair (x, y). Or you can treat
the pair (x, y) as a single variable, whose replacement set is all 2D points.
For example, the following says that the real plane (R
2
) contains a point on
the unit circle:
(x, y) R
2
, x
2
+y
2
= 1
Another approach is to write something like
p R
2
, p is on the unit circle
When you later need to make precise what it means to be on the unit circle,
you will have to break up p into its two coordinates. At that point, you say
CHAPTER 2. LOGIC 24
that that since p is a point on the plane, it must have the form (x, y), where
x and y are real numbers. This denes the component variables you need to
expand the denition of on the unit circle into the equation x
2
+y
2
= 1.
2.15 Negating statements with quantiers
Suppose we have a universal claim like x R, x
2
0. This claim will be
false if there is at least one real number x such that x
2
< 0. In general, a
statement of the form for all x in A, P(x) is false exactly when there is
some value x in A such that P(x) is false. In other words, when there exists
x in A such that P(x) is not true. In shorthand notation:
(x, P(x)) x, P(x)
Similarly,
(x, P(x)) x, P(x)
So this is a bit like the de Morgans laws: when you move the negation
across the operator, you change it to the other similar operator.
We saw above how to move negation operators from the outside to the
inside of expressions involving , , and the other propositional operators.
Together with these two new rules to handle quantiers, we now have a
mechanical procedure for working out the negation of any random statement
in predicate logic.
So if we have something like
x, P(x) (Q(x) R(x))
Its negation is
CHAPTER 2. LOGIC 25
(x, P(x) (Q(x) R(x))) x, (P(x) (Q(x) R(x)))
x, P(x) (Q(x) R(x)))
x, P(x) (Q(x) R(x))
2.16 Binding and scope
A quantier is said to bind the variable it denes. For example, in the
statement
x R, x
2
+ 3 0
the quantier binds the variable x.
The bound variable in a quantication is only dened for a limited time,
called the scope of the binding. This is usually the end of the quantied
statement or the end of the line, but you sometimes have to use common sense
about what the author intended. Parentheses are often used to indicate that
the author intends a shorter scope.
If a variable hasnt been bound by a quantier, or otherwise given a value
or a set of replacement values, it is called free. Statements containing free
variables dont have a dened truth value, so they cannot be (for example)
a step in a proof.
Variables bound by quantiers are like the dummy variables used in sum-
mations and integrals. For example, the i in

n
i=0
1
i
is only dened while you
are still inside the summation. Variables in computer programs also have a
scope over which their declaration is valid, e.g. the entire program, a single
code le, or local to an individual function/procedure/method. If you try to
use a variable outside the scope of its denition, youll get a compiler error.
When writing mathematics, variables have to be dened, just like vari-
ables in computer programs. Its not polite to start using a variable without
telling the reader where this variable comes from: is it bound by a quantier?
is it being set to a specic value (e.g. let x = 3.1415)? The reader also needs
to know what type of value this variable might contain.
CHAPTER 2. LOGIC 26
2.17 Variations in Notation
Although the core concepts of predicate logic are very standard, a few details
vary from author to author. Please stick to the conventions used above,
because its less confusing if we all use the same notation. However, dont
be surprised if another class or book does things dierently. For example:
There are several conventions about inserting commas after the quan-
tier and/or parentheses around the following predicate. We wont be
picky about this.
Some subelds (but not this class) have a convention that and is
applied before or, so that parentheses around and operations can
be omitted. Well keep the parentheses in such cases.
Some authors use certain variations of or (e.g. either ... or) with an
exclusive meaning, when writing mathematical English. In this class,
always read or as inclusive unless there is a clear explicit indiciation
that its meant to be exclusive. For example, Take some bread or
some cereal should be read as inclusive in a mathematical context. If
we wanted to indicate an exclusive reading, we would write something
like Take bread or some cereal, but not both.
In circuits and certain programming languages, true and false are
represented by 1 and 0. There are a number of dierent shorthand nota-
tions for logical connectives such as and. Finally, programmers often
use the term boolean to refer to true/false values and to expressions
that return true/false values.
Chapter 3
Proofs
Many mathematical proofs use a small range of standard outlines: direct
proof, examples/counter-examples, and proof by contrapositive. These notes
explain these basic proof methods, as well as how to use denitions of new
concepts in proofs. More advanced methods (e.g. proof by induction, proof
by contradiction) will be covered later.
3.1 Proving a universal statement
Now, lets consider how to prove a claim like
For every rational number q, 2q is rational.
First, we need to dene what we mean by rational.
A real number r is rational if there are integers m and n, n = 0, such
that r =
m
n
.
In this denition, notice that the fraction
m
n
does not need to satisfy
conditions like being proper or in lowest terms So, for example, zero is rational
because it can be written as
0
1
. However, its critical that the two numbers
27
CHAPTER 3. PROOFS 28
in the fraction be integers, since even irrational numbers can be written as
fractions with non-integers on the top and/or bottom. E.g. =

1
.
The simplest technique for proving a claim of the form x A, P(x) is
to pick some representative value for x.
1
. Think about sticking your hand
into the set A with your eyes closed and pulling out some random element.
You use the fact that x is an element of A to show that P(x) is true. Heres
what it looks like for our example:
Proof: Let q be any rational number. From the denition of
rational, we know that q =
m
n
where m and n are integers and
n is not zero. So 2q = 2
m
n
=
2m
n
. Since m is an integer, so is
2m. So 2q is also the ratio of two integers and, therefore, 2q is
rational.
At the start of the proof, notice that we expanded the word rational
into what its denition said. At the end of the proof, we went the other
way: noticed that something had the form required by the denition and
then asserted that it must be a rational.
WARNING!! Abuse of notation. Notice that the above denition of
rational used the word if. If you take this literally, it would mean that the
denition could be applied only in one direction. This isnt whats meant.
Denitions are always intended to work in both directions. Technically, I
should have written if and only if (frequently shortened to i). This
little misuse of if in denitions is very, very common.
Notice also that we spelled out the denition of rational but we just
freely used facts from high school algebra as if they were obvious. In general,
when writing proofs, you and your reader come to some agreement about
what parts of math will be considered familiar and obvious, and which re-
quire explicit discussion. In practical terms, this means that, when writing
solutions to homework problems, you should try to mimic the level of detail in
examples presented in lecture and in model solutions to previous homeworks.
1
The formal name for this is universal instantiation.
CHAPTER 3. PROOFS 29
3.2 Another example of direct proof involv-
ing odd and even
Heres another claim that can be proved by direct proof.
Claim 1 For any integer k, if k is odd then k
2
is odd.
This has a slightly dierent form from the previous claim: x Z, if P(x),
then Q(x)
Before doing the actual proof, we rst need to be precise about what we
mean by odd. And, while we are on the topic, what we mean by even.
Denition 1 An integer n is even if there is an integer m such that n = 2m.
Denition 2 An integer n is odd if there is an integer m such that n =
2m+ 1.
Such denitions are sometimes written using the jargon has the form, as
in An integer n is even if it has the form 2m, where m is an integer.
Well assume that its obvious (from our high school algebra) that every
integer is even or odd, and that no integer is both even and odd. You prob-
ably also feel condent that you know which numbers are odd or even. An
exception might be zero: notice that the above denition makes it denitely
even. This is the standard convention in math and computer science.
Using these denitions, we can prove our claim as follows:
Proof of Claim 1: Let k be any integer and suppose that k is odd.
We need to show that k
2
is odd.
Since k is odd, there is an integer j such that k = 2j + 1. Then
we have
k
2
= (2j + 1)
2
= 4j
2
+ 4j + 1 = 2(2j
2
+ 2j) + 1
Since j is an integer, 2j
2
+ 2j is also an integer. Lets call it p.
Then k
2
= 2p + 1. So, by the denition of odd, k
2
is odd.
CHAPTER 3. PROOFS 30
As in the previous proof, we used our key denition twice in the proof:
once at the start to expand a technical term (odd) into its meaning, then
again at the end to summarize our ndings into the appropriate technical
terms.
At the start of the proof, notice that we chose a random (or arbitrary
in math jargon) integer k, like last time. However, we also supposed that
the hypothesis of the if/then statement was true. Its helpful to collect up
all your given information right at the start of the proof, so you know what
you have to work with.
The comment about what we need to show is not necessary to the proof.
Its sometimes included because its helpful to the reader. You may also want
to include it because its helpful to you to remind you of where you need to
get to at the end of the proof.
Similarly, introducing the variable p isnt really necessary with a claim
this simple. However, using new variables to create an exact match to a
denition may help you keep yourself organized.
3.3 Direct proof outline
In both of these proofs, we started from the known information (anything
in the variable declarations and the hypothesis of the if/then statement)
and moved gradually towards the information that needed to be proved (the
conclusion of the if/then statement). This is the standard logical order for
a direct proof. Its the easiest order for a reader to understand.
When working out your proof, you may sometimes need to reason back-
wards from your desired conclusion on your scratch paper. However, when
you write out the nal version, reorder everything so its in logical order.
You will sometimes see proofs that are written partly in backwards order.
This is harder to do well and requires a lot more words of explanation to help
the reader follow what the proof is trying to do. When you are rst starting
out, especially if you dont like writing a lot of comments, its better to stick
to a straightforward logical order.
CHAPTER 3. PROOFS 31
3.4 Proving existential statements
Heres an existential claim:
Claim 2 There is an integer k such that k
2
= 0.
An existential claim such as the following asserts the existence of an object
with some set of properties. So its enough to exhibit some specic concrete
object, of our choosing, with the required properties. So our proof can be
very simple:
Proof: Zero is such an integer. So the statement is true.
We could spell out a bit more detail, but its really not necessary. Proofs
of existential claims are often very short, though there are exceptions.
Notice one dierence from our previous proofs. When we pick a value to
instantiate a universally quantied variable, we have no control over exactly
what the value is. We have to base our reasoning just on what set it belongs
to. But when we are proving an existential claim, we get to pick our own
favorite choice of concrete value, in this case zero.
Dont prove an existential claim using a general argument about why
there must exist numbers with these properties.
2
This is not only overkill,
but very hard to do correctly and harder on the reader. Use a specic,
concrete example.
3.5 Disproving a universal statement
Heres a universal claim that is false:
Claim 3 Every rational number q has a multiplicative inverse.
2
In higher mathematics, you occuasionally must make an abstract argument about
existence because its not feasible to produce a concrete example. We will see one example
towards the end of this course. But this is very rare.
CHAPTER 3. PROOFS 32
Denition 3 If q and r are real numbers, r is a multiplicative inverse for q
if qr = 1.
In general, a statement of the form for all x in A, P(x) is false exactly
when there is some value y in A for which P(y) is false.
3
So, to disprove a
universal claim, we need to prove an existential statement. So its enough to
exhibit one concrete value (a counter-example) for which the claim fails.
In this case, our disproof is very simple:
Disproof of Claim 3: This claim isnt true, because we know from
high school algebra that zero has no inverse.
Dont try to construct a general argument when a single specic coun-
terexample would be sucient.
3.6 Disproving an existential statement
Theres a general pattern here: the negation of x, P(x) is x, P(x). So the
negation of a universal claim is an existential claim. Similarly the negation of
x, P(x) is x, P(x). So the negation of an existential claim is a universal
one.
Suppose we want to disprove an existential claim like:
Claim 4 There is an integer k such that k
2
+ 2k + 1 < 0.
We need to make a general argument that, no matter what value of k we
pick, the equation wont hold. So we need to prove the claim
Claim 5 For every integer k, its not the case that k
2
+ 2k + 1 < 0.
Or, said another way,
3
Notice that for which is occasionally used as a variant of such that. In this case,
it makes the English sound very slightly better.
CHAPTER 3. PROOFS 33
Claim 6 For every integer k, k
2
+ 2k + 1 0.
The proof of this is fairly easy:
Proof: Let k be an integer. Then (k +1)
2
0 because the square
of any real number is non-negative. But (k+1)
2
= k
2
+2k+1. So,
by combining these two equations, we nd that k
2
+ 2k + 1 0.
3.7 Recap of proof methods
So, our general pattern for selecting the proof type is:
prove disprove
universal general argument specic counter-example
existential specic example general argument
Both types of proof start o by picking an element x from the domain
of the quantication. However, for the general arguments, x is a random
element whose identity you dont know. For the proofs requiring specic
examples, you can pick x to be your favorite specic concrete value.
3.8 Direct proof: example with two variables
Lets do another example of direct proof. First, lets dene
Denition 4 An integer n is a perfect square if n = k
2
for some integer k.
And now consider the claim:
Claim 7 For any integers m and n, if m and n are perfect squares, then so
is mn.
Proof: Let m and n be integers and suppose that m and n are
perfect squares.
CHAPTER 3. PROOFS 34
By the denition of perfect square, we know that m = k
2
and
n = j
2
, for some integers k and j. So then mn is k
2
j
2
, which is
equal to (kj)
2
. Since k and j are integers, so is kj. Since mn
is the square of the integer kj, mn is a perfect square, which is
what we needed to show.
Notice that we used a dierent variable name in the two uses of the
denition of perfect square: k the rst time and j the second time. Its
important to use a fresh variable name each time you expand a denition
like this. Otherwise, you could end up forcing two variables (m and n in this
case) to be equal when that isnt (or might not be) true.
Notice that the phrase which is what we needed to show helps tell the
reader that were done with the proof. Its polite to indicate the end in
one way or another. In typed notes, it may be clear from the indentation.
Sometimes, especially in handwritten proofs, we put a box or triangle of dots
or Q.E.D. at the end. Q.E.D. is short for Latin Quod erat demonstrandum,
which is just a translation of what we needed to show.
3.9 Another example with two variables
Heres another example of a claim involving two variables:
Claim 8 For all integers j and k, if j and k are odd, then jk is odd.
A direct proof would look like:
Proof: Let j and k be integers and suppose they are both odd.
Because j is odd, there is an integer p such that j = 2p + 1.
Similarly, there is an integer q such that k = 2q + 1.
So then jk = (2p+1)(2q+1) = 4pq+2p+2q+1 = 2(2pq+p+q)+1.
Since p and q are both integers, so is 2pq + p + q. Lets call it
m. Then jk = 2m+1 and therefore jk is odd, which is what we
needed to show.
CHAPTER 3. PROOFS 35
3.10 Proof by cases
When constructing a proof, sometimes youll nd that part of your given
information has the form p or q. Perhaps your claim was stated using
or, or perhaps you had an equation like |x| > 6 which translates into an
or (x > 6 or x < 6) when you try to manipulate it. When the given
information allows two or more separate possibilities, it is often best to use
a technique called proof by cases.
In proof by cases, you do part of the proof two or more times, once for
each of the possibilities in the or. For example, suppose that our claim is:
Claim 9 For all integers j and k, if j is even or k is even, then jk is even.
We can prove this as follows:
Proof: Let j and k be integers and suppose that j is even or k is
even. There are two cases:
Case 1: j is even. Then j = 2m, where m is an integer. So the
jk = (2m)k = 2(mk). Since m and k are integers, so is mk. So
jk must be even.
Case 2: k is even. Then k = 2n, where n is an integer. So the
jk = j(2n) = 2(nj). Since n and j are integers, so is nj. So jk
must be even.
So jk is even in both cases, which is what we needed to show.
It is ok to have more than two cases. Its also ok if the cases overlap,
e.g. one case might assume that x 0 and another case might assume that
x 0. However, you must be sure that all your cases, taken together, cover
all the possibilities.
In this example, each case involved expanding the denition of even.
We expanded the denition twice but, unlike our earlier examples, only one
expansion is active at a given time. So we could have re-used the variable m
when we expanded even in Case 2. I chose to use a fresh variable name
(n) because this is a safer strategy if you arent absolutely sure when re-use
is ok.
CHAPTER 3. PROOFS 36
3.11 Rephrasing claims
Sometimes youll be asked to prove a claim thats not in a good form for a
direct proof. For example:
Claim 10 There is no integer k such that k is odd and k
2
is even.
Its not clear how to start a proof for a claim like this. What is our given
information and what do we need to show?
In such cases, it is often useful to rephrase your claim using logical equiv-
alences. For example, the above claim is equivalent to
Claim 11 For every integer k, it is not the case that k is odd and k
2
is even.
By DeMorgans laws, this is equivalent to
Claim 12 For every integer k, k is not odd or k
2
is not even.
Since were assuming we all know that even and odd are opposites, this
is the same as
Claim 13 For every integer k, k is not odd or k
2
is odd.
And we can restate this as an implication using the fact that p q is
equivalent to p q:
Claim 14 For every integer k, if k is odd then k
2
is odd.
Our claim is now in a convenient form: a universal if/then statement
whose hypothesis contains positive (not negated) facts. And, in fact, we
proved this claim earlier in these notes.
CHAPTER 3. PROOFS 37
3.12 Proof by contrapositive
A particularly common sort of rephrasing is to replace a claim by its contra-
positive. If the original claim was x, P(x) Q(x) then its contrapositive
is x, Q(x) P(x). Remember that any if/then statement is logically
equivalent to its contrapositive.
Remember that constructing the hypothesis requires swapping the hy-
pothesis with the conclusion AND negating both of them. If you do only
half of this transformation, you get a statement that isnt equivalent to the
original. For example, the converse x, Q(x) P(x) is not equivalent to the
original claim.
For example, suppose that we want to prove
Claim 15 For any integers a and b, if a +b 15, then a 8 or b 8.
This is hard to prove in its original form, because were trying to use
information about a derived quantity to prove something about more basic
quantities. If we rephrase as the contrapositive, we get
Claim 16 For any integers a and b, if its not the case that a 8 or b 8,
then its not the case that a +b 15.
And this is equivalent to:
Claim 17 For any integers a and b, if a < 8 and b < 8, then a +b < 15.
Notice that when we negated the conclusion of the original statement, we
needed to change the or into an and (DeMorgans Law).
When you do this kind of rephrasing, your proof should start by explain-
ing to the reader how you rephrased the claim. Its technically enough to say
that youre proving the contrapositive. But, for a beginning proof writer, its
better to actually write out the contrapositive of the claim. This gives you a
chance to make sure you have constructed the contrapositive correctly. And,
while you are writing the rest of the proof, it helps remind you of exactly
what is given and what you need to show.
So a proof our our original claim might look like:
CHAPTER 3. PROOFS 38
Proof: Well prove the contrapositive of this statement. That is,
for any integers a and b, if a < 8 and b < 8, then a +b < 15.
So, suppose that a and b are integers such that a < 8 and b < 8.
Since they are integers (not e.g. real numbers), this implies that
a 7 and b 7. Adding these two equations together, we nd
that a +b 14. But this implies that a +b < 15.
There is no hard-and-fast rule about when to switch to the contrapositive
of a claim. If you are stuck trying to write a direct proof, write out the
contrapositive of the claim and see whether that version seems easier to
prove.
3.13 Another example of proof by contrapos-
itive
Heres another example where it works well to convert to the contrapositive:
Claim 18 For any integer k, if 3k + 1 is even, then k is odd.
If we rephrase as the contrapositive, we get:
Claim 19 For any integer k, if k is even, 3k + 1 is odd.
So our complete proof would look like:
Proof: We will prove the contrapositive of this claim, i.e. that
for any integer k, if k is even, 3k + 1 is odd.
So, suppose that k is an integer and k is even. Then, k = 2m for
some integer m. Then 3k +1 = 3(2m) +1 = 2(3m) +1. Since m
is an integer, so is 3m. So 3k +1 must be odd, which is what we
needed to show.
Chapter 4
Number Theory
Weve now covered most of the basic techniques for writing proofs. So were
going to start applying them to specic topics in mathematics, starting with
number theory.
Number theory is a branch of mathematics concerned with the behavior
of integers. It has very important applications in cryptography and in the
design of randomized algorithms. Randomization has become an increasingly
important technique for creating very fast algorithms for storing and retriev-
ing objects (e.g. hash tables), testing whether two objects are the same (e.g.
MP3s), and the like. Much of the underlying theory depends on facts about
which numbers evenly divide one another and which numbers are prime.
4.1 Factors and multiples
Youve undoubtedly seen some of the basic ideas (e.g. divisibility) somewhat
informally in earlier math classes. However, you may not be fully clear on
what happens with special cases, e.g. zero, negative numbers. We also need
clear formal denitions in order to write formal proofs. So, lets start with
Denition: Suppose that a and b are integers. Then a divides b
if b = an for some integer n. a is called a factor or divisor of b. b
is called a multiple of a.
39
CHAPTER 4. NUMBER THEORY 40
The shorthand for a divides b is a | b. Be careful about the order. The
divisor is on the left and the multiple is on the right.
Some examples:
7 | 77
77 | 7
7 | 7 because 7 = 7 1
7 | 0 because 0 = 7 0, zero is divisible by any integer.
0 | 7 because 0 n will always give you zero, never 7. Zero is a factor
of only one number: zero.
(3) | 12 because 12 = 3 4
3 | (12) because 12 = 3 4
A number p is even exactly when 2 | p. The fact that zero is even is just
a special case of the fact that zero is divisible by any integer.
4.2 Direct proof with divisibility
We can prove basic facts about divisibility in much the same way we proved
basic facts about even and odd.
Claim 20 For any integers a,b,and c, if a | b and a | c then a | (b +c).
Proof: Let a,b,and c and suppose that a | b and a | c.
Since a | b, there is an integer k such that b = ak (denition of
divides). Similarly, since a | c, there is an integer j such that
c = aj. Adding these two equations, we nd that (b + c) =
ak + aj = a(k + j). Since k and j are integers, so is k + j.
Therefore, by the denition of divides, a | (b +c).
CHAPTER 4. NUMBER THEORY 41
When we expanded the denition of divides for the second time, we used
a fresh variable name. If we had re-used k, then we would have wrongly
forced b and c to be equal.
The following two claims can be proved in a similar way:
Claim 21 For any integers a,b,and c, if a | b and b | c then a | c. (Transi-
tivity of divides.)
Claim 22 For any integers a,b, and c, if a | b, then a | bc.
Youve probably seen transitivity before in the context of inequalities.
E.g. if a < b and b < c, then a < c. Well get back to the general notion of
transitivity later in the term.
4.3 Stay in the Set
Students are sometimes tempted to rephrase the denition of a | b as
b
a
is an
integer. This is not a good idea because it introduces a non-integer rational
number into a problem that only needs to involve integers. Stepping outside
the set of interest in this way is occasionally useful, but more often it leads
to inelegant and/or buggy solutions. This happens for three reasons:
The purely integer proofs are typically simpler.
When constructing math from the ground up, the integers are typically
constructed rst and the rationals built from them. So using rationals
to prove facts about integers can lead to circular proofs.
On computers, integer operations yield exact answers but oating point
operations are only approximate. So implementing an integer calcula-
tion using real numbers often introduces errors.
CHAPTER 4. NUMBER THEORY 42
4.4 Prime numbers
Were all familiar with prime numbers from high school. Firming up the
details:
Denition: an integer q 2 is prime if the only positive factors
of q are q and 1. An integer q 2 is composite if it is not prime.
For example, among the integers no bigger than 20, the primes are 2,
3, 5, 7, 11, 13, 17, and 19. Numbers smaller than 2 are neither prime nor
composite.
A key fact about prime numbers is
Fundanmental Theorem of Arithmetic: Every integer 2 can be
written as the product of one or more prime factors. Except for
the order in which you write the factors, this prime factorization
is unique.
The word unique here means that there is only one way to factor each
integer.
For example, 260 = 2 2 5 13 and 180 = 2 2 3 3 5.
We wont prove this theorem right now, because it requires a proof tech-
nique called induction, which we havent seen yet.
There are quite fast algorithms for testing whether a large integer is prime.
However, even once you know a number is composite, algorithms for factoring
the number are all fairly slow. The diculty of factoring large composite
numbers is the basis for a number of well-known cryptographic algorithms
(e.g. the RSA algorithm).
4.5 GCD and LCM
If c divides both a and b, then c is called a common divisor of a and b. The
largest such number is the greatest common divisor of a and b. Shorthand
for this is gcd(a, b).
CHAPTER 4. NUMBER THEORY 43
You can nd the GCD of two numbers by inspecting their prime factor-
izations and extracting the shared factors. For example, 140 = 2
2
5 7 and
650 = 2 5
2
13. So gcd(140, 6500) is 2 5 = 10.
Similarly, a common multiple of a and b is a number c such that a|c and
b|c. The least common multiple (lcm) is the smallest positive number for
which this is true. The lcm can be computed using the formula:
lcm(a, b) =
ab
gcd(a, b)
For example, lcm(140, 650) =
140650
10
= 9100.
If two integers a and b share no common factors, then gcd(a, b) = 1. Such
a pair of integers are called relatively prime.
If k is a non-zero integer, then k divides zero. the largest common divisor
of k and zero is k. So gcd(k, 0) = gcd(0, k) = k. However, gcd(0, 0) isnt
dened. All integers are common divisors of 0 and 0, so there is no greatest
one.
4.6 The division algorithm
The obvious way to compute the gcd of two integers is to factor both into
primes and extract the shared factors. This is easy for small integers. How-
ever, it quickly becomes very dicult for larger integers, both for humans
and computers. Fortunately, theres a fast method, called the Euclidean al-
gorithm. Before presenting this algorithm, we need some background about
integer division.
In general, when we divide one integer by another, we get a quotient and
a remainder:
Theorem 1 (Division Algorithm) The Division Algorithm: For any in-
tegers a and b, where b is positive, there are unique integers q (the quotient)
and r (the remainder) such that a = bq +r and 0 r < b.
For example, if 13 is divided by 4, the quotient is 3 and the remainder is
CHAPTER 4. NUMBER THEORY 44
1. Notice that the remainder is required to be non-negative. So -10 divided
by 7 has the remainder 4, because (10) = 7 (2) +4. This is the standard
convention in mathematics. The word unique in this theorem simply means
that only one pair of number q and r satisfy the pair of equations.
Now, notice the following non-obvious fact about the gcd:
Claim 23 For any integers a, b, q and r, where b is positive, if a = bq +r,
then gcd(a, b) = gcd(b, r).
Proof: Suppose that n is some integer which divides both a and
b. Then n divides bq and so n divides a bq. (E.g. use various
lemmas about divides from last week.) But a bq is just r. So n
divides r.
By an almost identical line of reasoning, if n divides both b and
r, then n divides a.
So, the set of common divisors of a and b is exactly the same as
the set of common divisors of b and r. But gcd(a, b) and gcd(b, r)
are just the largest numbers in these two sets, so if the sets contain
the same things, the two gcds must be equal.
If r is the remainder when a is divided by b, then a = bq + r, for some
integer q. So we can immediately conclude that:
Corollary: Suppose that a and b are integers and b is positive.
Let r be the remainder when a is divided by b. Then gcd(a, b) =
gcd(b, r).
The term corollary means that this fact is a really easy consequence of
the preceding claim.
4.7 Euclidean algorithm
We can now give a fast algorithm for computing gcd, which dates back to Eu-
clid. Suppose that remainder(a, b) returns the remainder when a is divided
by b. Then we can compute the gcd as follows:
CHAPTER 4. NUMBER THEORY 45
gcd(a,b: positive integers)
x := a
y := b
while (y > 0)
begin
r := remainder(x,y)
x := y
y := r
end
return x
Lets trace this algorithm on inputs a = 105 and b = 252. Traces should
summarize the values of the most important variables.
x y r = remainder(x, y)
105 252 105
252 105 42
105 42 21
42 21 0
21 0
Since x is smaller than y, the rst iteration of the loop swaps x and y.
After that, each iteration reduces the sizes of a and b, because a mod b is
smaller than b. In the last iteration, y has gone to zero, so we output the
value of x which is 21.
To verify that this algorithm is correct, we need to convince ourselves of
two things. First, it must halt, because each iteration reduces the magnitude
of y. Second, by our corollary above, the value of gcd(x, y) does not change
from iteration to iteration. Moreover, gcd(x, 0) is x, for any non-zero integer
x. So the nal output will be the gcd of the two inputs a and b.
This is a genuinely very nice algorithm. Not only is it fast, but it involves
very simple calculations that can be done by hand (without a calculator).
Its much easier than factoring both numbers into primes, especially as the
individual prime factors get larger. Most of us cant quickly see whether a
large number is divisible by, say, 17.
CHAPTER 4. NUMBER THEORY 46
4.8 Pseudocode
Notice that this algorithm is written in pseudocode. Pseudocode is an ab-
stracted type of programming language, used to highlight the important
structure of an algorithm and communicate between researchers who may
not use the same programming language. It borrows many control constructs
(e.g. the while loop) from imperative languages such as C. But details re-
quired only for a mechanical compiler (e.g. type declarations for all variables)
are omitted and equations or words are used to hide details that are easy to
gure out.
If you have taken a programming course, pseudocode is typically easy to
read. Many small details are not standardized, e.g. is the test for equality
written = or ==? However, its usually easy for a human (though not a
computer) to gure out what the author must have intended.
A common question is how much detail to use. Try to use about the same
amount as in the examples shown in the notes. And think about how easily
your pseudocode could be read by a classmate. Actual C or Java code is
almost never acceptable pseudocode, because it is way too detailed.
4.9 A recursive version of gcd
We can also write gcd as a recursive algorithm
procedure gcd(a,b: positive integers)
r := remainder(a,b)
if (r = 0) return b
else return gcd(b,r)
This code is very simple, because this algorithm has a natural recursive
structure. Our corollary allows us to express the gcd of two numbers in terms
of the gcd of a smaller pair of numbers. That is to say, it allows us to reduce
a larger version of the task to a smaller version of the same task.
CHAPTER 4. NUMBER THEORY 47
4.10 Congruence mod k
Many applications of number theory, particularly in computer science, use
modular arithmetic. In modular arithmetic, there are only a nite set of
numbers and addition wraps around from the highest number to the lowest
one. This is true, for example, for the 12 hours on a standard US clock: 3
hours after 11 oclock is 2 oclock, not 14 oclock.
The formal mathematical denitions of modular arithmetic are based on
the notion of congruence. Specically, two integers are congruent mod k
if they dier by a multiple of k. Formally:
Denition: If k is any positive integer, two integers a and b are
congruent mod k (written a b (mod k)) if k | (a b).
Notice that k | (a b) if and only if k | (b a). So it doesnt matter which
number is subtracted from the other.
For example:
17 5 (mod 12) (Familiar to those of us whove had to convert be-
tween US 12-hour clocks and European/military 24-hour clocks.)
3 10 (mod 7)
3 38 (mod 7) (Since 38 3 = 35.)
38 3 (mod 7)
3 4 (mod 7) (Since (3) + 7 = 4.)
3 3 (mod 7)
3 3 (mod 6)
29 13 (mod 8) (Since (13) (29) = 16.)
Congruence mod k is a relaxation of our normal rules for equality of
numbers, in which we agree that certain pairs of numbers will be considered
interchangeable.
CHAPTER 4. NUMBER THEORY 48
4.11 Proofs with congruence mod k
Lets try using our denition to prove a simple fact about modular arithmetic:
Claim 24 For any integers a, b, c, d, and k, k positive, if a b (mod k)
and c d (mod k), then a +c b +d (mod k).
Proof: Let a, b, c, d, and k be integers with k positive. Suppose
that a b (mod k) and c d (mod k).
Since a b (mod k), k | (a b), by the denition of congruence
mod k. Similarly, c d (mod k), k | (c d).
Since k | (a b) and k | (c d), we know by a lemma about
divides (above) that k | (a b) + (c d). So k | (a +c) (b +d)
But then the denition of congruence mod k tells us that a+c
b +d (mod k).
This proof can easily be modied to show that
Claim 25 For any integers a, b, c, d, and k, k positive, if a b (mod k)
and c d (mod k), then ac bd (mod k).
So standard arithmetic operations interact well with our relaxed notation
of equality.
4.12 Equivalence classes
The true power of modular conguence comes when we gather up a group of
conguent numbers and treat them all as a unit. Such a group is known as a
congruence class or an equivalence class. Specically, suppose that we
x a particular value for k. Then, if x is a number, the equivalence class of x
(written [x]) is the set of all numbers congruent to x mod k. Or, equivalently,
the set of numbers that have remainder x when divided by k.
For example, suppose that we x k to be 7. Then
CHAPTER 4. NUMBER THEORY 49
[3] = {3, 10, 4, 17, 11, . . .}
[1] = {1, 8, 6, 15, 13, . . .}
[0] = {0, 7, 7, 14, 14, . . .}
Notice that [4], and [10] are exactly the same set as [3]. That is [4] =
[10] = [3]. So we have one object (the set) with many dierent names (one
per number in it). This is like a student apartment shared by Fred, Emily,
Ali, and Michelle. The supercially dierent phrases Emilys apartment
and Alis apartment actually refer to one and the same apartment.
Having many names for the same object can become confusing, so people
tend to choose a special preferred name for each object. For the k equiva-
lence classes of numbers mod k, mathematicians tend to prefer the names
[0], [1], . . . , [k 1]. Other names (e.g. [30] when k = 7) tend to occur only as
intermediate results in calculations.
Because standard arithmetic operations interact well with modular con-
gruence, we can set up a system of arithmetic on these equivalence classes.
Specically, we dene addition and multiplication on equivalence classes by:
[x] + [y] = [x +y]
[x] [y] = [x y]
So, (still setting k = 7) we can do computations such as
[4] + [10] = [4 + 10] = [14] = [0]
[4] [10] = [4 10] = [40] = [2]
This new set of numbers ([0], [1], . . . , [k 1]), with these modular rules of
arithmetic and equality, is known as the integers mod k or Z
k
for short.
For example, the addition and multiplication tables for Z
4
are:
+
4
[0] [1] [2] [3]
[0] [0] [1] [2] [3]
[1] [1] [2] [3] [0]
[2] [2] [3] [0] [1]
[3] [3] [0] [1] [2]
CHAPTER 4. NUMBER THEORY 50

4
[0] [1] [2] [3]
[0] [0] [0] [0] [0]
[1] [0] [1] [2] [3]
[2] [0] [2] [0] [2]
[3] [0] [3] [2] [1]
People making extensive use of modular arithmetic frequently drop the
square brackets. Were keeping the brackets for the moment to help you un-
derstand more clearly how the modular integers are created from the normal
integers.
4.13 Wider perspective on equivalence
Modular arithmetic is often useful in describing periodic phenomena. For
example, hours of the day form a circular space that is Z
12
. However, the
tradition in time-keeping is to prefer the names [1], [2], . . . , [12] rather than
the [0], [2], . . . , [11] traditional in math.
Low-precision integer storage on computers frequently uses modular arith-
metic. For example, values in digitized pictures are often stored as 8-bit
unsigned numbers. These numbers range from 0 to 255, i.e. the structure is
Z
256
. In Z
256
, if you add [17] to [242], the result will be [3].
Equivalence classes of a more general type are used to construct rational
numbers. That is, rational numbers are basically fractions, except that two
fractions
x
y
and
p
q
are considered equivalent if xq = py. So, for example, the
set [
2
3
] would contain values such as
2
3
,
6
9
, and
4
6
. Most people prefer to
name a rational number using a fraction in lowest terms.
Further aeld, notice that musicians have more than one name for each
note. There are seven note names, each of which can be sharped or atted.
However, there are only 12 dierent notes in a standard Western scale: many
of these 21 names actually refer to the same note.
1
For example, A# is the
same note as Bb. So, using more mathematical notation than the musicians
would nd comfortable, we could say that [A#] contains both A# and Bb.
1
More precisely, this is true on instruments like pianos. More subtle note distinctions
are possible on some other instruments.
CHAPTER 4. NUMBER THEORY 51
4.14 Variation in Terminology
In these notes, a divides b is dened to be true if b = an for some integer
n. There is some variation among authors as to what happens when a is
zero. Clearly, a non-zero number cant be a multiple of zero. But is zero
a multiple of itself? According to our denition, it is, but some authors
explicitly exclude this special case. Fortunately, this is a special case that
one rarely sees in practice. The greatest common divisor is also known as
the highest common factor (HCF).
In the shorthand notation a b (mod k), the notation (mod k) is log-
ically a modier on our notion of equality (). In retrospect, it might have
made more sense to write something like a
k
b. However, a b (mod k)
has become the standard notation and we have to live with it.
There are many variant notations for the quotient and remainder cre-
ated by integer division, particularly if you include the functions built in
to most programming languages. Popular names for the remainder include
mod, modulo, rem, remainder, and %. The behavior on negative inputs dif-
fers from language to language and, in some cases, from implementation to
implementation.
2
This lack of standardization often results in hard-to-nd
program bugs.
Some authors use Z/nZ instead of Z
n
as the shorthand name for the inte-
gers mod n. The notation Z
n
may then refer to a set which is structurally the
same as the integers mod n, but in which multiplication and exponentiation
are used as the basic operations, rather than addition and multiplication.
The equivalence class of n is sometimes written n.
2
The modulo operation entry on wikipedia has a nice table of what happens in
dierent languages.
Chapter 5
Sets
So far, weve been assuming only a basic understanding of sets. Its time to
discuss sets systematically, including a useful range of constructions, opera-
tions, notation, and special cases. Well also see how to compute the sizes of
sets and prove claims involving sets.
5.1 Sets
Sets are an extremely general concept, dened as follows:
Denition: A set is an unordered collection of objects.
For example, the natural numbers are a set. So are the integers between
3 and 7 (inclusive). So are all the planets in this solar system or all the
programs written by students in CS 225 in the last three years. The objects
in a set can be anything you want.
The items in the set are called its elements or members. Weve already
seen the notation for this: x A means that x is a member of the set A.
Theres three basic ways to dene a set:
describe its contents in mathematical English, e.g. the integers be-
tween 3 and 7, inclusive.
52
CHAPTER 5. SETS 53
list all its members, e.g. {3, 4, 5, 6, 7}
use so-called set builder notation, e.g. {x Z | 3 x 7}
Set builder notation has two parts separated with a vertical bar or a
colon. The rst part names a variable (in this case x) that ranges over all
objects in the set. The second part gives one or more constraints that these
objects must satisfy, e.g. 3 x 7. The type of the variable (integer in
our example) can be specied either before or after the vertical bar. The
separator (| or :) is often read such that.
Heres an example of a set containing an innite number of objects
multiples of 7
{. . . 14, 7, 0, 7, 14, 21, 28, . . .}
{x Z | x is a multiple of 7}
We couldnt list all the elements, so we had to use . . .. This is only a
good idea if the pattern will be clear to your reader. If you arent sure, use
one of the other methods.
If we wanted to use shorthand for multiple of, it might be confusing to
have | used for two dierent purposes. So it would probably be best to use
the colon variant of set builder notation:
{x Z : 7 | x}
The notation can be used on sets containing non-numerical objects. For
example, suppose that A = {dog, cat, tree}. Then the set {s : A}
contains the strings dogs, cats, and trees.
5.2 Things to be careful about
A set is an unordered collection. So {1, 2, 3} and {2, 3, 1} are two names for
the same set. Each element occurs only once in a set. Or, alternatively, it
doesnt matter how many times you write it. So {1, 2, 3} and {1, 2, 3, 2} also
name the same set.
CHAPTER 5. SETS 54
Weve seen ordered pairs and triples of numbers, such as (3, 4) and (4, 5, 2).
The general term for an ordered sequence of k numbers is a k-tuple.
1
Tuples
are very dierent from sets, in that the order of values in a tuple matters
and duplicate elements dont magically collapse. So (1, 2, 2, 3) = (1, 2, 3)
and (1, 2, 2, 3) = (2, 2, 1, 3). Therefore, make sure to enclose the elements
of a set in curly brackets and carefully distinguish curly brackets (set) from
parentheses (ordered pair).
A more subtle feature of tuples is that a tuple must contain at least two
elements. In formal mathematics, a 1-dimensional value x is just written as
x, not as (x). And theres no such thing in mathematics as a 0-tuple. So a
tuple is simply a way of grouping two or more values into a single object.
By contrast, a set is like a cardboard box, into which you can put objects.
A kitty is dierent from a box containing a kitty. Similarly, a set containing
a single object is dierent from the object by itself. For example, {57} is not
the same as 57. A set can also contain nothing at all, like an empty box.
The set containing nothing is called the empty set or the null set, and has
the shorthand symbol .
2
The empty set may seem like a pain in the neck. However, computer
science applications are full of empty lists, strings of zero length, and the
like. Its the kind of special case that all of you (even the non-theoreticians)
will spend your life having to watch out for.
Both sets and tuples can contain objects of more than one type, e.g.
(cat, Fluy, 1983) or {a, b, 3, 7}. A set can also contain complex objects,
e.g. {(a, b), (1, 2, 3), 6} is a set containing three objects: an ordered pair, an
ordered triple, and a single number.
5.3 Cardinality, inclusion
If A is a nite set (a set containing only a nite number of objects), then |A|
is the number of (dierent) objects in A. This is also called the cardinality
1
There are latinate terms for longer sequences of numbers, e.g. quadruple, but they
arent used much.
2
Dont write the emptyset as {}. Like a spelling mistake, this will make readers think
less of your mathematical skills.
CHAPTER 5. SETS 55
of A. For example, |{a, b, 3}| = 3. And |{a, b, a, 3}| is also 3, because we
count a group of identical objects only once. The notation of cardinality also
extends to sets with innitely many members (innite sets) such as the
integers, but we wont get into the details of that right now.
Notice that the notation |A| might mean set cardinality or it might be the
more familiar absolute value. To tell which, gure out what type of object
A is. If its a set, the author meant cardinality. If its a number, the author
meant absolute value.
If A and B are sets, then A is a subset of B (written A B) if every
element of A is also in B. Or, if you want it formally: x, x A x B.
For example, Q R, because every member of the rationals is also a member
of the reals.
The notion of subset allows the two sets to be equal. So A A is true for
any set A. So is like . If you want to force the two sets to be dierent (i.e.
like <), you must say that A is a proper subset of B, written A B. Youll
occasionally see reversed versions of these symbols to indicate the opposite
relation, e.g. B A means the same as A B.
5.4 Vacuous truth
If we have a set A, an interesting question is whether the empty set should
be considered a subset of A. To answer this, lets rst back up and look at
one subtlety of mathematical logic.
Consider the following claim:
Claim 26 For all natural numbers n, if 14 +n < 10, then n wood elves will
attack Siebel Center tomorrow.
I claim this is true, a fact which most students nd counter-intuitive. In fact,
it wouldnt be true if n was declared to be an integer.
Notice that this statement has the form n, P(n) Q(n), where P(n) is
the predicate 14+n < 10. Because n is declared to be a natural number, n is
never negative, so n + 14 will always be at least 14. So P(n) is always false.
CHAPTER 5. SETS 56
Therefore, our conventions about the truth values for conditional statements
imply that P(n) Q(n) is true. This argument works for any choice of n.
So n, P(n) Q(n) is true.
Because even mathematicians nd such statements a bit wierd, they typ-
ically say that such a claim is vacuously true, to emphasize to the reader
that it is only true because of this strange convention about the meaning of
conditionals. Vacuously true statements typically occur when you are trying
to apply a denition or theorem to a special case involving an abnormally
small or simple object, such as the empty set or zero or a graph with no
arrows at all.
In particular, this means that the empty set is a subset of any set A. For
to be a subset of A, the denition of subset requires that for every object
x, if x is an element of the empty set, then x is an element of A. But this
if/then statement is considered true because its hypothesis is always false.
5.5 Set operations
Given two sets A and B, the intersection of A and B (A B) is the set
containing all objects that are in both A and B. In set builder notation:
A B = {S | S A and S B}
Lets set up some sample sets:
M = {egg, bread, milk}
P = {milk, egg, our}
Then M P is {milk, egg}.
If the intersection of two sets A and B is the empty set, i.e. the two sets
have no elements in common, then A and B are said to be disjoint.
The union of sets A and B (A B) is the set containing all objects that
are in one (or both) of A and B. So M P is {milk, egg, bread, our}.
CHAPTER 5. SETS 57
The set dierence of A and B (A B) contains all the objects that are
in A but not in B. In this case,
M P = {bread}
The complement of a set A (A) is all the objects that arent in A. For
this to make sense, you need to dene your universal set (often written U).
U contains all the objects of the sort(s) you are discussing. For example, in
some discussions, U might be all real numbers. U doesnt contain everything
you might imagine putting in a set, because constructing a set that inclusive
leads to paradoxes. U is more limited than that. Whenever U is used, you
and your reader need to come to an understanding about whats in it.
So, if our universe is all integers, and A contains all the multiples of 3,
then A is all the integers whose remainder mod 3 is either 1 or 2. Q would
be the irrational numbers if our universe is all real numbers. If we had
been working with complex numbers, it might be the set of all irrational real
numbers plus all the numbers with an imaginary component.
If A and B are two sets, their Cartesian product (A B) contains all
ordered pairs (x, y) where x is in A and y is in B. That is
A B = {(x, y) | x A and y B}
For example, if A = {a, b} and B = {1, 2}, then
A B = {(a, 1), (a, 2), (b, 1), (b, 2)}
B A = {(1, a), (2, a), (1, b), (2, b)}
Notice that these two sets arent equal: order matters for Cartesian prod-
ucts.
We can generalize this denition to create the Cartesian product of three
or more sets. So, for example, if C = {p, q}, then
ABC == {(a, 1, p), (a, 1, q), (a, 2, p), (a, 2, q), (b, 1, p), (b, 1, q), (b, 2, p), (b, 2, q)}
CHAPTER 5. SETS 58
5.6 Set identities
Its easy to nd (e.g. on the net), long lists of identities showing when two
sequences of set operations yield the same output set. For example:
DeMorgans Law: A B = A B
I wont go through these in detail because they are largely identical to
the identities you saw for logical operations, if you make the following corre-
spondences:
is like
is like
A is like P
(the empty set) is like F
U (the universal set) is like T
The two systems arent exactly the same. E.g. set theory doesnt use a
close analog of the operator. But they are very similar.
5.7 Size of set union
Many applications require that we calculate
3
the size of the set created by
applying set operations. These sets are often sets of options for some task.
If two sets A and B dont intersect, then the size of their union is just
the sum of their sizes. That is: |A B| = |A| + |B|. For example, suppose
that its late evening and you want to watch a movie. You have 37 movies
on cable, 57 DVDs on the shelf, and 12 movies stored in I-tunes. If these
three sets of movies dont intersect, you have a total of 37 + 57 + 12 = 106
movies.
3
Or sometimes just estimate.
CHAPTER 5. SETS 59
If your input sets do overlap, then adding up their sizes will double-count
some of the objects. So, to get the right number for the union, we need to
correct for the double-counting. For our movie example, suppose that the
only overlap is that 2 movies are on I-tunes and also on DVD. Then you
would have (37 + 57 + 12) 2 = 104 movies to choose from.
The formal name for this correction is the Inclusion-Exclusion Princi-
ple. Formally, suppose you have two sets A and B. Then
Inclusion-Exclusion Principle: |A B| = |A| +|B| |A B|
We can use this basic 2-set formula to derive the corresponding formula
for three sets A, B, and C:
|A B C| = |A| +|B C| |A (B C)|
= |A| +|B| +|C| |B C| |A (B C)|
= |A| +|B| +|C| |B C| |(A B) (A C)|
= |A| +|B| +|C| |B C| (|A B| +|A C| |(A B) (A C)|)
= |A| +|B| +|C| |B C| |A B| |A C| +|A B C|
In addition to the inclusion-exclusion principle, this derivation uses the
distributive law (third step) and the fact that intersection is commutative
and associative (last step).
5.8 Product rule
Now, suppose that we form the Cartesian product of two sets A and B, where
|A| = n and |B| = q. To form an element (x, y) in the product, we have n
choices for x and q choices for y. So we have nq ways to create an element
of the product. So |AB| = nq.
In general:
CHAPTER 5. SETS 60
The product rule: if you have p choices for one part of a task,
then q choices for a second part, and your options for the second
part dont depend on what you chose for the rst part, then you
have pq options for the whole task.
This rule generalizes in the obvious way to sequences of several choices:
you multiply the number of options for each choice. For example, suppose
that T-shirts come in 4 colors, 5 sizes, 2 sleeve lengths, and 3 styles of
neckline, there are 4 5 2 3 = 120 total types of shirts.
We could represent a specic T-shirt type as a 4-tuple (c, s, l, n): c is its
color, s is its size, l is its sleeve length, and n is its neckline. E.g. one T-shirt
type is (red, small, long, vee) The set of all possible T-shirt types would then
be a 4-way Cartesian product of the set of all colors, the set of all sizes, the
set of all sleeve lengths, and the set of all necklines.
5.9 Combining these basic rules
These two basic counting rules can be combined to solve more complex practi-
cal counting problems. For example, suppose we have a set S which contains
all 5-digit decimal numbers that start with 2 ones or end in 2 zeros, where
we dont allow leading zeros. How large is S, i.e. how many numbers have
this form?
Let T be the set of 5-digit numbers starting in 2 ones. We know the rst
two digits and we have three independent choices (10 options each) for the
last three. So there are 1000 numbers in T.
Let R be the set of 5-digit numbers ending in 2 zeros. We have 9 options
for the rst digit, since it cant be zero. We have 10 options each for the
second and third digits, and the last two are xed. So we have 900 numbers
in R.
Whats the size of T R? Numbers in this set start with 2 ones and
end with 2 zeros, so the only choice is the middle digit. So it contains 10
numbers. So
|S| = |T| +|R| |T R| = 1000 + 900 10 = 1890
CHAPTER 5. SETS 61
5.10 Proving facts about set inclusion
So far in school, most of your proofs or derivations have involved reason-
ing about equality. Inequalities (e.g. involving numbers) have been much
less common. With sets, the situation is reversed. Proofs typically involve
reasoning about subset relations, even when proving two sets to be equal.
Proofs that rely primarily on a chain of set equalities do occur, but they are
much less common. Even when both approaches are possible, the approach
based on subset relations is often easier to write and debug.
As a rst example of a typical set proof, lets suppose that we the following
two sets and wed like to prove that A B
A = {(2, 3) + (1 )(7, 4) | [0, 1]}
B = {(x, y) | x, y R, x 0, and y 0}
When presented with a claim like this, you should rst take a minute to
verify that the claim really is true. Set B is the upper right quadrant of the
plane. To understand the denition of set A, remember how to multiply a
real number by a vector: a(x, y) = (ax, ay). This denition generates all
points of the form (2, 3) +(1 )(7, 4) where is a real number between 0
and 1. Try putting some sample values of into this equation and plotting
them in 2D: what geometrical object is this?
4
Make sure you believe that
this object does live in the upper right quadrant.
Now, remember our denition of : a set A is a subset of a set B if and
only if, for any object x, x A implies that x B. So, to prove our claim,
we need to pick a random object x from A and show that it lives in B. So a
starting sketch of our proof might look like:
Proof: Let sets A and B be dened as above. Let x be an element
of A.
[missing details]
So x is an element of B.
4
Hint: use a pencil and paper. Your plot doesnt need to be neat to see the pattern.
CHAPTER 5. SETS 62
Since x was arbitrarily chosen, weve shown that any element of
A is also an element of B. So A is a subset of B.
We can now use the denition of A to extend forward from the hypotheses
into the missing part. In particular, the denition of A implies that x is a
2D point, so its probably helpful to give names to the two coordinates:
Proof: Let sets A and B be dened as above. Let x be an element
of A. Then x = (2, 3) + (1 )(7, 4) for some [0, 1]. So
x = (p, q) where p = 2 + 7(1 ) and q = 3 + 4(1 )
[missing details]
So x is an element of B.
Since x was arbitrarily chosen, weve shown that any element of
A is also an element of B. So A is a subset of B.
At this point, its worth looking at the end part of the proof. If x = (p, q)
and were trying to show that x is in B, then this translates into showing
that p and q are both non-negative. So we can expand the end of the proof
backwards, narrowing the missing part even further:
Proof: Let sets A and B be dened as above. Let x be an element
of A. Then x = (2, 3) + (1 )(7, 4) for some [0, 1]. So
x = (p, q) where p = 2 + 7(1 ) and q = 3 + 4(1 )
[missing details]
So p 0 and q 0. This means that x = (p, q) is an element of
B.
Since x was arbitrarily chosen, weve shown that any element of
A is also an element of B. So A is a subset of B.
Now we have a straightforward algebra problem: get from the complex
equations dening p and q to the fact that both are non-negative. Working
out the details of that algebra gives us the nal proof:
CHAPTER 5. SETS 63
Proof: Let sets A and B be dened as above. Let x be an element
of A. Then x = (2, 3) + (1 )(7, 4) for some [0, 1]. So
x = (p, q) where p = 2 + 7(1 ) and q = 3 + 4(1 )
Simplifying these equations, we get that p = 2+77) = 75
and q = 3 + 4 4) = 4 Since is in the interval [0, 1],
1. So 7 5 and 4 are both 0.
So p 0 and q 0. This means that x = (p, q) is an element of
B.
Since x was arbitrarily chosen, weve shown that any element of
A is also an element of B. So A is a subset of B.
The last paragraph in this proof is actually is optional. When you rst
start, its a useful recap because you might be a bit fuzzy about what you
needed to prove. Experienced folks often omit it, depending on the reader
to understand the basic outline of a subset inclusion proof. But you will still
see it occasionally at the end of a very long (e.g. multi-page) proof, where
the reader might have been so buried in the details that theyve lost track of
the overall goals of the proof.
5.11 An abstract example
Now, lets try to do a similar proof, but for a claim involving generic sets
rather than specic sets.
Claim 27 For any sets A, B, and C, if A B and B C, then A C.
This property is called transitivity, just like similar properties for (say)
on the real numbers. Both and are examples of a general type of
object called a partial order, for which transitivity is a key dening property.
Lets start our proof by gathering up all the given information in the
hypothesis of the claim:
Proof: Let A, B, and C be sets and suppose that A B and
B C.
CHAPTER 5. SETS 64
Our ultimate goal is to show that A C. This is an if/then statement:
for any x, if x A, then x C. So we need to pick a representative x
and assume the hypothesis is true, then show the conclusion. So our proof
continues:
Let x be an element of A. Since A B and x A, then x B
(denition of subset). Similarly, since x B and B C, x C.
So for any x, if x A, then x C. So A C (denition of
subset again).
5.12 An example with products
Heres another claim, involving Cartesian products:
Claim 28 For any sets A, B, and C, if A B A C and A = , then
B C.
Notice that this claim fails if A = . For example, {1, 2, 3} is a subset
of {a, b}, because both of these sets are empty. However {1, 2, 3} is not
a subset of {a, b}.
This is like dividing both sides of an algebraic equation by a non-zero
number: if xy xz and x = 0 then y z. This doesnt work we allow x
to be zero. Set operations dont always work exactly like operations on real
numbers, but the parallelism is strong enough to suggest special cases that
ought to be investigated.
A general property of proofs is that the proof should use all the informa-
tion in the hypothesis of the claim. If thats not the case, either the proof
has a bug (e.g. on a homework assignment) or the claim could be revised to
make it more interesting (e.g. when doing a research problem, or a buggy
homework problem). Either way, theres an important issue to deal with.
So, in this case, we need to make sure that our proof does use the fact that
A = .
Heres a draft proof:
CHAPTER 5. SETS 65
Proof draft: Suppose that A, B, and C are sets and suppose that
AB AC and A = . We need to show that B C.
So lets choose some x B. . . .
The main fact weve been given is that A B A C. To use it, we
need an element of A B. Right now, we only have an element of B. We
need to nd an element of A to pair it with. To do this, we reach blindly
into A, pull out some random element, and give it a name. But we have to
be careful here: what if A doesnt contain any elements? So we have to use
the assumption that A = .
Proof: Suppose that A, B, and C are sets and suppose that
AB AC and A = . We need to show that B C.
So lets choose some x B. Since A = , we can choose an
element t from A. Then (t, x) A B by the denition of
Cartesian product.
Since (t, x) A B and A B A C, we must have that
(t, x) A C (by the denition of subset). But then (again by
the denition of Cartesian product) x C.
So weve shown that if x B, then x C. So B C, which is
what we needed to show.
5.13 A proof using sets and contrapositive
Heres a claim about sets thats less than obvious:
Claim 29 For any sets A and B, if (A B) (B A) = A B then
A B = .
Notice that the conclusion A B = claims that something does not
exist (i.e. an object thats in both A and B). So this is a good place to apply
proof by contrapositive.
CHAPTER 5. SETS 66
Proof: Lets prove the contrapositive. That is, well prove that if
A B = , then (A B) (B A) = A B.
so, let A and B be sets and suppose that. A B = . Since
AB = , we can choose an element from AB. Lets call it x.
Since x is in A B, x is in both A and B. So x is in A B.
However, since x is in B, x is not in AB. Similarly, since x is in
A, x is not in BA. So x is not a member of (AB) (BA).
This means that (AB) (B A) and A B cannot be equal,
because x is in the second set but not in the rst. .
5.14 Variation in notation
In mathematical writing, it is quite rare to create tuples that contain other
tuples. In principle, tuples containing tuples are formally dierent from
longer tuples. E.g. ((a, b), c) is formally a dierent object from (a, b, c), and
(a, (b, c)) is dierent from both. However, most mathematical writers treat
this dierence as a petty nuisance and view the two objects as interchange-
able.
Linked lists used in computer science behave very dierently from math-
ematical tuples. The linked list ((a, b), c) is completely dierent from the list
(a, b, c) and a linked list can contain a single element or no elements. It is a
bad idea to confuse linked lists and mathatical tuples, even though they are
written with the same notation.
Chapter 6
Relations
Mathematical relations are an extremely general framework for specifying
relationships between pairs of objects. This chapter surveys the types of
relations that can be constructed on a single set A and the properties used
to characterize dierent types of relations.
6.1 Relations
A relation R on a set A is a subset of AA, i.e. R is a set of ordered pairs
of elements from A. For simplicity, we will assume that the base set A is
always non-empty. If R contains the pair (x, y), we say that x is related to
y, or xRy in shorthand. Well write xRy to mean that x is not related to y.
For example, suppose we let A = {2, 3, 4, 5, 6, 7, 8}. We can dene a
relation W on A by xWy if and only if x y x + 2. Then W contains
pairs like (3, 4) and (4, 6), but not the pairs (6, 4) and (3, 6). Under this
relation, each element of A is related to itself. So W also contains pairs like
(5, 5).
We can draw pictures of relations using directed graphs. We draw a
graph node for each element of A and we draw an arrow joining each pair of
elements that are related. So W looks like:
67
CHAPTER 6. RELATIONS 68
2 3 4 5 6 7 8
In fact, theres very little formal dierence between a relation on a set
A and a directed graph, because graph edges can be represented as ordered
pairs of endpoints. They are two ways of describing the same situation.
We can dene another relation S on A by saying that xSy is in S if x y
(mod 3). Then S would look like:
2
5
8
3
6
4
7
Or, suppose that B = {2, 3, 4, 5, 6, 12, 25}. Lets set up a relation T on B
such that xTy if x|y and x = y. Then our picture would look like
25 12
4 6
5 3 2
Mathematical relations can also be used to represent real-world relation-
ships, in which case they often have a less regular structure. For example,
suppose that we have a set of students and student x is related to student y
if x nominated y for ACM president. The graph of this relation (call it Q)
might look like:
CHAPTER 6. RELATIONS 69
Fred
Ginger
Alan Steve
Julie Bill
Henry
Relations can also be dened on innite sets or multi-dimensional objects.
For example, we can dene a relation Z on the real plane R
2
in which (x, y)
is related to (p, q) if and only if x
2
+y
2
= p
2
+q
2
. In other words, two points
are related if they are the same distance from the origin.
For complex relations, the full directed graph picture can get a bit messy.
So there are simplied types of diagrams for certain specic special types of
relations, e.g. the so-called Hasse diagram for partial orders.
6.2 Properties of relations: reexive
Relations are classied by several key properties. The rst is whether an
element of the set is related to itself or not. There are three cases:
Reexive: every element is related to itself.
Irreexive: no element is related to itself.
Neither reexive nor irreexive: some elements are related to them-
selves but some arent.
In our pictures above, elements related to themselves have self-loops. So its
easy to see that W and S are reexive, T is irreexive, and Q is neither.
The familiar relations and = on the real numbers are reexive, but < is
irreexive. Suppose we dene a relation M on the integers by xMy if and
only if x + y = 0. Then 2 isnt related to itself, but 0 is. So M is neither
reexive nor irreexive.
The formal denition states that if R is a relation on a set A then
CHAPTER 6. RELATIONS 70
R is reexive if xRx for all x A.
R is irreexive if xRx for all x A.
Notice that irreexive is not the negation of reexive. The negation of
reexive would be:
not reexive: there is an x A, xRx
6.3 Symmetric and antisymmetric
Another important property of a relation is whether the order matters within
each pair. That is, if xRy is in R, is it always the case that yRx? If this is
true, then the relation is called symmetric.
In a graph picture of a symmetric relation, a pair of elements is either
joined by a pair of arrows going in opposite directions, or no arrows. In our
examples with pictures above, only S is symmetric.
Relations that resemble equality are normally symmetric. For example,
the relation X on the integers dened by xXy i |x| = |y| is symmetric. So
is the relation N on the real plane dened by (x, y)N(p, q)) i (xp)
2
+(y
q)
2
25 (i.e. the two points are no more than 5 units apart).
Relations that put elements into an order, like or divides, have a dier-
ent property called antisymmetry. A relation is antisymmetric if two distinct
1
elements are never related in both directions. In a graph picture of an an-
tisymmetric relation, a pair of points may be joined by a single arrow, or
not joined at all. They are never joined by arrows in both directions. In our
pictures above, W and T are antisymmetric.
As with reexivity, there are mixed relations that have neither property.
So the relation Q above is neither symmetric nor antisymmetric.
If R is a relation on a set A, heres the formal denition of what it means
for R to be symmetric (which doesnt contain anything particularly dicult):
1
Distinct means not equal.
CHAPTER 6. RELATIONS 71
symmetric: for all x, y A, xRy implies yRx
Theres two ways to dene antisymmetric. They are logically equivalent
and you can pick whichever is more convenient for your purposes:
antisymmetric: for all x and y in A with x = y, if xRy, then y Rx
antisymmetric: for all x and y in A, if xRy and yRx, then x = y
To interpret the second denition, remember that when mathematicians
pick two values x and y, they leave open the possibility that the two values
are actually the same. In normal conversational English, if we mention two
objects with dierent names, we normally mean to imply that they are dier-
ent objects. This isnt the case in mathematics. I nd that the rst denition
of antisymmetry is better for understanding the idea, but the second is more
useful for writing proofs.
6.4 Transitive
The nal important property of relations is transitivity. A relation R on a
set A is transitive if
transitive: for all a, b, c A, ifaRb and bRc, then aRc
As weve seen earlier, transitivity holds for a broad range of familiar
numerical relations such as <, =, divides, and set inclusion. For example,
for real numbers, if x < y and y < z, then x < z. Similarly, if x|y and y|z,
then x|z. For sets, X Y and Y Z implies that X Z.
If we look at graph pictures, transitivity means that whenever there is an
indirect path from x to y then there must also be a direct arrow from x to
y. This is true for S and B above, but not for W or Q.
We can also understand this by spelling out what it means for a relation
R on a set A not to be transitive:
CHAPTER 6. RELATIONS 72
not transitive: there are a, b, c A, aRb and bRc and aRc
So, to show that a relation is not transitive, we need to nd one counter-
example, i.e. specic elements a, b, and c such that aRb and bRc but not
aRc. In the graph of a non-transitive relation, you can nd a subsection that
looks like:
a
b
c
It could be that a and c are actually the same element, in which case the
oending subgraph might look like:
a
b
The problem here is that if aRb and bRa, then transitivity would imply
that aRa and bRb.
One subtle point about transitive is that its an if/then statement. So
its ok if some sets of elements just arent connected at all. For example, this
subgraph is consistent with the relation being transitive.
a
b
c
A disgustingly counter-intuitive special case is the relation in which ab-
solutely no elements are related, i.e. no arrows at all in the graph picture.
This relation is transitive. its never possible to satisfy the hypothesis of the
CHAPTER 6. RELATIONS 73
denition of transitive. Its also symmetric, for the same reason. And, oddly
enough, antisymmetric. All of these properties hold via vacuous truth.
Vacuous truth doesnt apply to reexive and irreexive, because they are
unconditional requirements, not if/then statements. So this empty relation
is irreexive and not reexive.
6.5 Types of relations
Now that we have these basic properties dened, we can dene some impor-
tant types of relations. Three of these are ordering relations:
A partial order is a relation that is reexive, antisymmetric, and
transitive.
A linear order (also called a total order) is a partial order in which
every pair of elements are comparable. That is, for every pair x, y,
x y or y x.
A strict partial order is a relation that is irreexive, antisymmetric,
and transitive.
A linear order is like the normal ordering on the integers. A partial
order is similar, but there are some pairs of elements which arent ordered
with respect to each other. For example, the divides relation on the integers
is a partial order but not a linear order, because (for example) 5 is neither a
factor nor a multiple of 7. A strict partial order is like a partial order, except
that objects are not related to themselves, the relation T in section 6.1 is a
strict partial order.
The fourth type of relation is an equivalence relation:
Denition: An equivalence relation is a relation that is reex-
ive, symmetric, and transitive.
These three properties create a well-behaved notation of equivalence or
congruence for a set of objects, like congruence mod k on the set of integers.
CHAPTER 6. RELATIONS 74
In particular, if R is an equivalence relation on a set A and x is an element
of A, we can dene the equivalence class of x to be the set of all elements
related to x. That is
[x]
R
= {y A | xRy}
When the relation is clear from the context (as when we discussed congruence
mod k), we frequently omit the subscript on the square brackets.
For example, we saw the relation Z on the real plane R
2
, where (x, y)Z(p, q)
if and only if x
2
+y
2
= p
2
+q
2
. Then [(0, 1)]
Z
contains all the points related
to (0, 1), i.e. the unit circle.
6.6 Proving that a relation is an equivalence
relation
Lets see how to prove that a new relation is an equivalence relation. These
proofs are usually very mechanical. For example, let F be the set of all
fractions, i.e.
F = {
p
q
| p, q Z and q = 0}
Fractions arent the same thing as rational numbers, because each rational
number is represented by many fractions. We consider two fractions to be
equivalent, i.e. represent the same rational number, if xq = yp. So, we have
an equivalence relation dened by:
x
y

p
q
if and only if xq = yp.
Lets show that is an equivalence relation.
Proof: Reexive: For any x and y, xy = xy. So the denition of
implies that
x
y

x
y
.
Symmetric: if
x
y

p
q
then xq = yp, so yp = xq, so py = qx, which
implies that
p
q

x
y
.
Transitive: Suppose that
x
y

p
q
and
p
q

s
t
. By the denition of
, xq = yp and pt = qs. So xqt = ypt and pty = qsy. Since
CHAPTER 6. RELATIONS 75
ypt = pty, this means that xqt = qsy. Cancelling out the qs, we
get xt = sy. By the denition of , this means that
x
y

s
t
.
Since is reexive, symmetric, and transitive, it is an equivalence
relation.
Notice that the proof has three sub-parts, each showing one of the key
properties. Each part involves using the denition of the relation, plus a
small amount of basic math. The reexive case is very short. The symmetric
case is often short as well. Most of the work is in the transitive case.
6.7 Proving antisymmetry
Heres another example of a relation, this time an order (not an equivalence)
relation. Consider the set of intervals on the real line J = {(a, b) | a, b
R and a < b}. Dene the containment relation C as follows:
(a, b) C (c, d) if and only if a c and d b
To show that C is a partial order, wed need to show that its reex-
ive, antisymmetric, and transitive. Weve seen how to prove two of these
properties. Lets see how to do a proof of antisymmetry.
For proving antisymmetry, its typically easiest to use this form of the
denition of antisymmetry: if xRy and yRx, then x = y. Notice that C is a
relation on intervals, i.e. pairs of numbers, not single numbers. Substituting
the denition of C into the denition of antisymmetry, we need to show that
For any intervals (a, b) and (c, d), if (a, b) C (c, d) and (c, d) C (a, b),
then (a, b) = (c, d).
So, suppose that we have two intervals (a, b) and (c, d) such that (a, b) C (c, d)
and (c, d) C (a, b). By the denition of C, (a, b) C (c, d) implies that a c
and d b. Similarly, (c, d) C (a, b) implies that c a and b d.
Since a c and c a, a = c. Since d b and b d, b = d. So
(a, b) = (c, d).
Chapter 7
Functions and onto
This chapter covers functions, including function composition and what it
means for a function to be onto. In the process, well see what happens when
two dissimilar quantiers are nested.
7.1 Functions
Were all familiar with functions from high school and calculus. However,
these prior math courses concentrate on functions whose inputs and outputs
are numbers, dened by an algebraic formula such as f(x) = 2x+3. Well be
using a broader range of functions, whose input and/or output values may
be integers, strings, characters, and the like.
Suppose that A and B are sets, then a function f from A to B is an
assignment of exactly one element of B (i.e. the output value) to each element
of A (i.e. the input value). A is called the domain of f and B is called the
co-domain. All of this information can be captured in the shorthand type
signature: f : A B. If x is an element of A, then the value f(x) is also
known as the image of x.
For example, suppose P is a set of ve people:
P = {Margaret, Tom, Chen, LaSonya, Emma}
76
CHAPTER 7. FUNCTIONS AND ONTO 77
And suppose C is a set of colors:
C = {red, blue, green, purple, yellow, orange}
We can dene a function f : P C which maps each person to their favorite
color. For example, we might have the following input/output pairs:
f(Margaret) = Blue
f(Tom) = Red
f(LaSonya) = Purple
f(Emma) = Red
f(Chen) = Blue
We also use a bubble diagram to show how f assigns output values to
input values.
Margaret
Tom
Chen
LaSonya
Emma
blue red
green
purple
yellow
orange
Even if A and B are nite sets, there are a very large number of possible
functions from A to B. Suppose that |A| = n, |B| = p. We can write out the
elements of A as x
1
, x
2
, . . . , x
n
. When constructing a function f : A B, we
have p ways to choose the output value for x
1
. The choice of f(x
1
) doesnt
aect our possible choices for f(x
2
): we also have p choices for that value. So
we have p
2
choices for the rst two output values. If we continue this process
for the rest of the elements of A, we have p
n
possible ways to construct our
function f.
CHAPTER 7. FUNCTIONS AND ONTO 78
For any set A, the identity function id
A
maps each value in A to itself.
That is, id
A
: A A and id
A
(x) = x.
7.2 When are functions equal?
Notice that the domain and co-domain are an integral part of the denition
of the function. To be equal, two functions must (obviously) assign the same
output value to each input value. But, in addition, they must have the same
type signature.
For example, suppose D is a slightly smaller set of colors:
D = {red, blue, purple}
Then the function g : P D shown below is not equal to the function f
from section 7.1, even though g(x) = f(x) for every x in P.
Margaret
Tom
Chen
LaSonya
Emma
blue red
purple
Similarly, the following denitions describe quite dierent functions, even
though they are based on the same equation.
f : N N such that f(x) = 2x.
f : R R such that f(x) = 2x.
CHAPTER 7. FUNCTIONS AND ONTO 79
However, the following are all denitions of the same function, because
the three variations have the same type signature and assign the same output
value to each input value.
f : Z Z such that f(x) = |x|.
f : Z Z such that f(x) = max(x, x).
f : Z Z such that f(x) = x if x 0 and f(x) = x if x 0.
Notice that the last denition uses two dierent cases to cover dierent
parts of the domain. This is ne, as long as all the cases, taken together,
provide exactly one output value for each input value.
7.3 What isnt a function?
For each input value, a function must provide one and only one output value.
So the following isnt a function, because one input has no output:
Margaret
Tom
Chen
LaSonya
blue red
purple
green
The following isnt a function, because one input is paired with two out-
puts:
CHAPTER 7. FUNCTIONS AND ONTO 80
Margaret
Tom
LaSonya
blue
red
purple
green
7.4 Images and Onto
The image of the function f : A B is the set of values produced when f
is applied to all elements of A. That is, the image is
f(A) = {f(x) : x A}
For example, suppose M = {a, b, c, d}, N = {1, 2, 3, 4}, and our function
g : M N is as in the following diagram. Then g(A) = {1, 3, 4}.
a
b
c
d
1
2
3
4
A function f : A B is onto if its image is its whole co-domain. Or,
equivalently,
y B, x A, f(x) = y
The function g that we just saw isnt onto, because no input value is mapped
CHAPTER 7. FUNCTIONS AND ONTO 81
onto 2.
Whether a function is onto critically depends on its type signature. Sup-
pose we dene p : Z Z by p(x) = x + 2. If we pick an output value y,
then the input value y 2 maps onto y. So the image of p is all of Z. So this
function is onto.
However, suppose we dene q : N N using the same formula q(x) =
x + 2. q isnt onto, because none of the input values map onto 0 or 1.
7.5 Why are some functions not onto?
You may think many examples of non-onto functions look like they could have
been onto if the author had set up the co-domain more precisely. Sometimes
the co-domain is excessively large simply because the image set is awkward
to specify succinctly. But, also, in some applications, we specically want
certain functions not to be onto.
For example, in graphics or certain engineering applications, we may wish
to map out or draw a curve in 2D space. The whole point of the curve is that
it occupies only part of 2D and it is surrounded by whitespace. These curves
are often specied parametrically, using functions that map into, but not
onto, 2D.
For example, we can specify a (unit) circle as the image of a function
f : [0, 1] R
2
dened by f(x) = (cos 2x, sin 2x). If you think of the input
values as time, then f shows the track of a pen or robot as it goes around
the circle. The cosine and sine are the x and y coordinates of its position.
The 2 multiplier simply puts the input into the right range so that well
sweep exactly once around the circle (assuming that sine and cosine take
their inputs in radians).
7.6 Negating onto
To understand the concept of onto, it may help to think about what it
means for a function not to be onto. This is our rst example of negating a
CHAPTER 7. FUNCTIONS AND ONTO 82
statement involving two nested (dierent) quantiers. Our denition of onto
is
y B, x A, f(x) = y
So a function f is not onto if
y B, x A, f(x) = y
To negate this, we proceed step-by-step, moving the negation inwards.
Youve seen all the identities involved, so this is largely a matter of being
careful.
y B, x A, f(x) = y
y B, x A, f(x) = y
y B, x A, (f(x) = y)
y B, x A, f(x) = y
So, if we want to show that f is not onto, we need to nd some value y
in B, such that no matter which element x you pick from A, f(x) isnt equal
to y.
7.7 Nested quantiers
Notice that the denition of onto combines a universal and an existential
quantier, with the scope of one including the scope of the other. These are
called nested quantiers. When quantiers are nested, the meaning of the
statement depends on the order of the two quantiers.
For example,
For every person p in the Fleck family, there is a toothbrush t
such that p brushes their teeth with t.
CHAPTER 7. FUNCTIONS AND ONTO 83
This sentence asks you to consider some random Fleck. Then, given that
choice, it asserts that they have a toothbrush. The toothbrush is chosen
after weve picked the person, so the choice of toothbrush can depend on the
choice of person. This doesnt absolutely force everyone to pick their own
toothbrush. (For a brief period, two of my sons were using the same one
because they got confused.) However, at least this statement is consistent
with each person having their own toothbrush.
Suppose now that we swap the order of the quantiers, to get
There is a toothbrush t, such that for every person p in the Fleck
family, p brushes their teeth with t.
In this case, were asked to choose a toothbrush t rst. Then were assert-
ing that every Fleck uses this one xed toothbrush t. Eeeuw! That wasnt
what we wanted to say!
We do want the existential quantier rst when theres a single object
thats shared among the various people, as in:
There is a stove s, such that for every person p in the Fleck family,
p cooks his food on s.
Notice that this order issue only appears when a statement a mixture of
existential and universal quantiers. If all the quantiers are existential, or
if all the quantiers are universal, the order doesnt matter.
To take a more mathematical example, lets look back at modular arith-
metic. Two numbers x and y are multiplicative inverses if xy = yx = 1. In
the integers Z, only 1 has a multiplicative inverse. However, in Z
k
, many
other integers have inverses. For example, if k = 7, then [3][5] = [1]. So [3]
and [5] are inverses.
For certain values of k every non-zero element of Z
k
has an inverse.
1
You
can verify that this is true for Z
7
: [3] and [5] are inverses, [2] and [4] are
inverses, [1] is its own inverse, and [6] is its own inverse. So we can say that
1
Can you gure out which values of k have this property?
CHAPTER 7. FUNCTIONS AND ONTO 84
non-zero x Z
7
, y Z
7
, xy = yx = 1
Notice that weve put the universal quantier outside the existential one,
so that each number gets to pick its own inverse. Reversing the order of the
quantifers would give us the following statement:
y Z
7
, non-zero x Z
7
, xy = yx = 1
This version isnt true, because you cant pick one single number that
works as an inverse for all the rest of the non-zero numbers, even in modular
arithmetic.
However, we do want the existential quantier rst in the following claim,
because 0y = y0 = 0 for every y Z
7
.
x Z
7
, y Z
7
, xy = yx = x
7.8 Proving that a function is onto
Now, consider this claim:
Claim 30 Dene the function g from the integers to the integers by the
formula g(x) = x 8. g is onto.
Proof: We need to show that for every integer y, there is an
integer x such that g(x) = y.
So, let y be some arbitrary integer. Choose x to be (y + 8).
x is an integer, since its the sum of two integers. But then
g(x) = (y +8) 8 = y, so weve found the required pre-image for
y and our proof is done.
For some functions, several input values map onto a single output value.
In that case, we can choose any input value in our proof, typically whichever
CHAPTER 7. FUNCTIONS AND ONTO 85
is easiest for the proof-writer. For example, suppose we had g : Z Z such
that g(x) =
x
2
. To show that g is onto, were given an output value x and
need to nd the corresponding input value. The simplest choice would be 2x
itself. But you could also pick 2x + 1.
Suppose we try to build such a proof for a function that isnt onto, e.g.
f : Z Z such that f(x) = 3x + 2.
Proof: We need to show that for every integer y, there is an
integer x such that f(x) = y.
So, let y be some arbitrary integer. Choose x to be
(y2)
3
. . . .
If f was a function from the reals to the reals, wed be ok at this point,
because x would be a good pre-image for y. However, fs inputs are declared
to be integers. For many values of y,
(y2)
3
isnt an integer. So it wont work
as an input value for f.
7.9 A 2D example
Heres a sample function whose domain is 2D. Let f : Z
2
Z be dened by
f(x, y) = x +y. I claim that f is onto.
First, lets make sure we know how to read this denition. f : Z
2
is
shorthand for Z Z, which is the set of pairs of integers. So f maps a pair
of integers to a single integer, which is the sum of the two coordinates.
To prove that f is onto, we need to pick some arbitrary element y in the
co-domain. That is to say, y is an integer. Then we need to nd a sample
value in the domain that maps onto y, i.e. a preimage of y. At this point,
it helps to ddle around on our scratch paper, to come up with a suitable
preimage. In this case, (0, y) will work nicely. So our proof looks like:
Proof: Let y be an element of Z. Then (0, y) is an element of
f : Z
2
and f(0, y) = 0 + y = y. Since this construction will work
for any choice of y, weve shown that f is onto.
CHAPTER 7. FUNCTIONS AND ONTO 86
Notice that this function maps many input values onto each output value.
So, in our proof, we could have used a dierent formula for nding the input
value, e.g. (1, y 1) or (y, 0). A proof writer with a sense of humor might
use (342, y 342).
7.10 Composing two functions
Suppose that f : A B and g : B C are functions. Then g f is the
function from A to C dened by (g f)(x) = g(f(x)). Depending on the
author, this is either called the composition of f and g or the composition of
g and f. The idea is that you take input values from A, run them through
f, and then run the result of that through g to get the nal output value.
Take-home message: when using function composition, look at the au-
thors shorthand notation rather than their mathematical English, to be
clear on which function gets applied rst.
In this denition, notice that g came rst in (g f)(x) and g also comes
rst in g(f(x)). I.e. unlike f(g(x)) where f comes rst. The trick for
remembering this denition is to remember that f and g are in the same
order on the two sides of the dening equation.
For example, suppose we dene two functions f and g from the integers
to the integers by:
f(x) = 3x + 7
g(x) = x 8
Since the domains and co-domains for both functions are the integers, we
can compose the two functions in both orders. But two composition orders
give us dierent functions:
(f g)(x) = f(g(x)) = 3g(x) + 7 = 3(x 8) + 7 = 3x 24 + 7 = 3x 17
(g f)(x) = g(f(x)) = f(x) 8 = (3x + 7) 8 = 3x 1
CHAPTER 7. FUNCTIONS AND ONTO 87
Frequently, the declared domains and co-domains of the two functions
arent all the same, so often you can only compose in one order. For example,
consider the function h : {strings} Z which maps a string x onto its length
in characters. (E.g. h(Margaret) = 8.) Then f h exists but (h f) doesnt
exist because f produces numbers and the inputs to h are supposed to be
strings.
7.11 A proof involving composition
Lets show that onto-ness works well with function composition. Specically:
Claim 31 For any sets A, B, and C and for any functions f : A B and
g : B C, if f and g are onto, then g f is also onto.
Proof: Let A, B, and C be sets. Let f : A B and g : B C
be functions. Suppose that f and g are onto.
We need to show that g f is onto. That is, we need to show
that for any element x in C, there is an element y in A such that
(g f)(y) = x.
So, pick some element x in C. Since g is onto, there is an element
z in B such that g(z) = x. Since f is onto, there is an element y
in A such that f(y) = z.
Substituting the value f(y) = z into the equation g(z) = x, we
get g(f(y)) = x. That is, (g f)(y) = x. So y is the element of
A we needed to nd.
7.12 Variation in terminology
A function is often known as a map and surjective is often used as a
synonym for onto. The image of a function is sometimes written Im(f).
The useful term type signature is not traditional in pure mathematics,
though its in wide use in computer science.
CHAPTER 7. FUNCTIONS AND ONTO 88
The term range is a term that has become out-dated in mathematics.
Depending on the author, it may refer to either the image or the co-domain
of a function, which creates confusion. Avoid using it.
Some authors write gf rather than g f for the composition of two func-
tions. Also, some authors dene g f such that g f(x) = f(g(x)), i.e. the
opposite convention from what were using.
Chapter 8
Functions and one-to-one
In this chapter, well see what it means for a function to be one-to-one and
bijective. This general topic includes counting permutations and comparing
sizes of nite sets (e.g. the pigeonhole principle). Well also see the method
of adding stipulations to a proof without loss of generality as well as the
technique of proving an equality via upper and lower bounds.
8.1 One-to-one
Suppose that f : A B is a function from A to B. If we pick a value y B,
then x A is a pre-image of y if f(x) = y. Notice that I said a pre-image
of y, not the pre-image of y, because y might have more than one preimage.
For example, in the following function, 1 and 2 are pre-images of 1, 4 and 5
are pre-images of 4, 3 is a pre-image of 3, and 2 has no preimages.
89
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 90
1
2
3
4
5
1
2
3
4
A function is one-to-one if it never assigns two input values to the same
output value. Or, said another way, no output value has more than one pre-
image. So the above function isnt one-to-one, because (for example) 4 has
more than one pre-image. If we dene g : Z Z such that g(x) = 2x. Then
g is one-to-one.
As with onto, whether a function is one-to-one frequently depends on its
type signature. For example, the absolute value function |x| is not one-to-
one as a function from the reals to the reals. However, it is one-to-one as a
function from the natural numbers to the natural numbers.
One formal denition of one-to-one is:
x, y A, x = y f(x) = f(y)
For most proofs, its more convenient to use the contrapositive:
x, y A, f(x) = f(y) x = y
When reading this denition, notice that when you set up two variables
x and y, they dont have to have dierent (math jargon: distinct) values.
In normal English, if you give dierent names to two objects, the listener is
expected to understand that they are dierent. By contrast, mathematicians
always mean you to understand that they might be dierent but theres also
the possibility that they might be the same object.
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 91
8.2 Bijections
If a function f is both one-to-one and onto, then each output value has
exactly one pre-image. So we can invert f, to get an inverse function f
1
. A
function that is both one-to-one and onto is called bijective or a bijection. If
f maps from A to B, then f
1
maps from B to A.
Suppose that A and B are nite sets. Constructing an onto function
from A to B is only possible when A has at least as many elements as B.
Constructing a one-to-one function from A to B requires that B have at least
as many values as A. So if there is a bijection between A and B, then the two
sets must contain the same number of elements. As well see later, bijections
are also used to dene what it means for two innite sets to be the same
size.
8.3 Pigeonhole Principle
Suppose that A contains more elements than B. Then its impossible to
construct a function A to B that is one-to-one, because two elements of A
must be mapped to the same element of B. If we rephrase this in terms of
putting labels on objects, we get the following Pigeonhole Principle:
Pigeonhole principle: Suppose you have n objects and assign
k labels to these objects. If n > k, then two objects must get the
same label.
For example, if you have 8 playing cards and the cards come in ve colors,
then at least two of the cards share the same color. These elementary exam-
ples look more interesting when they involve larger numbers. For example, it
would work poorly to assign three letter, case insensitive, usernames for the
undergraduates at the University of Illinois. The Urbana-Champaign cam-
pus alone has over 30,000 undergraduates. But there are only 26
3
= 17, 576
3-letter usernames. So the pigeonhole principle implies that there would be
a pair of undergraduates assigned the same username.
The pigeonhole principle (like many superheros) has a dual identity.
When youre told to apply it to some specic objects and labels, its ob-
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 92
vious how to do so. However, it is often pulled out of nowhere as a clever
trick in proofs, where you would have never suspected that it might be useful.
Such proofs are easy to read, but sometimes hard to come up with.
For example, heres a fact that we can prove with a less obvious applica-
tion of the pigeonhole principle.
Claim 32 Among the rst 100 powers of 17, there are two (distinct) powers
which dier by a multiple of 57.
The trick here is to notice that two numbers dier by a multiple of 57
exactly when they have the same remainder mod 57. But we have 100 distinct
powers of 17 and only 57 dierent possible values for the remaider mod 57.
So our proof might look like:
Proof: The rst 100 powers of 17 are 17
1
, 17
1
, . . . , 17
100
. Consider
their remainders mod 57: r
1
, r
2
, . . . , r
100
. Since there are only 57
possible remainders mod 57, two of the numbers r
1
, r
2
, . . . , r
100
must be equal, by the pigeonhole principle. Lets suppose that
r
j
and r
k
are equal. Then 17
j
and 17
k
have the same remainder
mod 57, and so 17
j
and 17
k
dier by a multiple of 57.
The actual use of the pigeonhole principle almost seems like an afterthought,
after weve come up with the right trick to structuring our analysis of the
problem. This is typical of harder pigeonhole principle proofs.
8.4 Permutations
Now, suppose that |A| = n = |B|. We can construct a number of one-to-one
functions from A to B. How many? Suppose that A = {x
1
, x
2
, . . . , x
n
}.
We have n ways to choose the output value for x
1
. But that choice uses up
one output value, so we have only n 1 choices for the output value of x
2
.
Continuing this pattern, we have n(n 1)(n 2) . . . 2 1 (n! for short) ways
to construct our function.
Similarly, suppose we have a group of n dragon gurines that wed like to
arrange on the mantlepiece. We need to construct a map from positions on
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 93
the mantlepiece to dragons. There are n! ways to do this. An arrangement
of n objects in order is called a permutation of the n objects.
More frequently, we need to select an ordered list of k objects from a
larger set of n objects. For example, we have 30 dragon gurines but space
for only 10 on our mantlepiece. Then we have 30 choices for the rst gurine,
29 for the second, and so forth down to 21 choices for the last one. Thus we
have 30 29 . . . 21 ways to decorate the mantlepiece.
In general, an ordered choice of k objects from a set of n objects is known
as a k-permutation of the n objects. There are n(n1) . . . (nk+1) =
n!
(nk)!
dierent k-permutations of n objects. This number is called P(n, k). P(n, k)
is also the number of one-to-one functions from a set of k objects to a set of
n objects.
8.5 Further applications of permutations
Many real-world counting problems can be solved with the permutations for-
mulas, but some problems require a bit of adaptation. For example, suppose
that 7 adults and 3 kids need to get in line at the airport. Also suppose that
we cant put two children next to each other, because they will ght.
The trick for this problem is to place the 7 adults in line, with gaps
between them. Each gap might be left empty or lled with one kid. There
are 8 gaps, into which we have to put the 3 kids. So, we have 7! ways to
assign adults to positions. Then we have 8 gaps in which we can put kid A,
7 for kid B, and 6 for kid C. That is 7! 8 7 6 ways to line them all up.
Now, lets suppose that we have a set of 7 scrabble tiles and we would
like to gure out how many dierent letter strings we can make out of them.
We could almost use the permutations formula, except that some of the
tiles might contain the same letter. For example, suppose that the tiles are:
C, O, L, L, E, G, E.
First, lets smear some dirt on the duplicate tiles, so we can tell the
two copies apart: C, O, L
1
, L
2
, E
1
, G, E
2
. Then we would calculate 7! per-
mutations of this list. However, in terms of our original problem, this is
double-counting some possibilities, because we dont care about the dier-
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 94
ence between duplicates like L
1
and L
2
. So we need to divide out by the
number of ways we can permute the duplicates. In this case, we have 2! ways
to permute the Ls and 2! ways to permute the Es. So the true number of
orderings of L is
7!
2!2!
.
Similarly, the number of reorderings of J = (a, p, p, l, e, t, r, e, e, s) is
10!
2!3!
.
In general, suppose we have n objects, where n
1
are of type 1, n
2
are of
type 2, and so forth through n
k
are of type k. Then the number of ways to
order our list of objects is
n!
n
1
!n
2
!...n
k
!
.
8.6 Proving that a function is one-to-one
Now, lets move on to examples of how to prove that a specic function is
one-to-one.
Claim 33 Let f : Z Z be dened by f(x) = 3x + 7. f is one-to-one.
Lets prove this using our denition of one-to-one.
Proof: We need to show that for every integers x and y, f(x) =
f(y) x = y.
So, let x and y be integers and suppose that f(x) = f(y). We
need to show that x = y.
We know that f(x) = f(y). So, substituting in our formula for f,
3x + 7 = 3y + 7. So 3x = 3y and therefore x = y, by high school
algebra. This is what we needed to show.
When we pick x and y at the start of the proof, notice that we havent
specied whether they are the same number or not. Mathematical convention
leaves this vague, unlike normal English where the same statement would
strongly suggest that they were dierent.
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 95
8.7 Composition and one-to-one
Like onto, one-to-one works well with function composition. Specically:
Claim 34 For any sets A, B, and C and for any functions f : A B and
g : B C, if f and g are one-to-one, then g f is also one-to-one.
We can prove this with a direct proof, by being systematic about using our
denitions and standard proof outlines. First, lets pick some representative
objects of the right types and assume everything in our hypothesis.
Proof: Let A, B, and C be sets. Let f : A B and g : B C
be functions. Suppose that f and g are one-to-one.
We need to show that g f is one-to-one.
To show that g f is one-to-one, we need to pick two elements x and
y in its domain, assume that their output values are equal, and then show
that x and y must themselves be equal. Lets splice this into our draft proof.
Remember that the domain of g f is A and its co-domain is C.
Proof: Let A, B, and C be sets. Let f : A B and g : B C
be functions. Suppose that f and g are one-to-one.
We need to show that g f is one-to-one. So, choose x and y in
A and suppose that (g f)(x) = (g f)(y)
We need to show that x = y.
Now, we need to apply the denition of function composition and the fact
that f and g are each one-to-one:
Proof: Let A, B, and C be sets. Let f : A B and g : B C
be functions. Suppose that f and g are one-to-one.
We need to show that g f is one-to-one. So, choose x and y in
A and suppose that (g f)(x) = (g f)(y)
Using the denition of function composition, we can rewrite this
as g(f(x)) = g(f(y)). Combining this with the fact that g is
one-to-one, we nd that f(x) = f(y). But, since f is one-to-one,
this implies that x = y, which is what we needed to show.
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 96
8.8 Strictly increasing functions are one-to-
one
Now, lets do a slightly trickier proof. First, a denition. Suppose that A
and B are sets of real numbers (e.g. the reals, the rationals, the integers).
A function f : A B is increasing if, for every x and y in A, x y implies
that f(x) f(y). f is called strictly increasing if, for every x and y in A,
x < y implies that f(x) < f(y).
1
An increasing function can have plateaus
where the output value stays constant, whereas a strictly increasing function
must always increase.
Claim 35 For any sets of real numbers A and B, if f is any strictly increas-
ing function from A to B, then f is one-to-one.
A similar fact applies to strictly decreasing functions.
To prove this, we will restate one-to-one using the alternative, contrapos-
itive version of its denition.
x, y A, x = y f(x) = f(y)
Normally, this wouldnt be a helpful move. The hypothesis involves a
negative fact, whereas the other version of the denition has a positive hy-
pothesis, normally a better place to start. But this is an atypical example
and, in this case, the negative information turns out to be a good starting
point.
Proof: Let A and B be sets of numbers and let f : A B be a
strictly increasing function. Let x and y be distinct elements of
A. We need to show that f(x) = f(y).
Since x = y, theres two possibilities.
Case 1: x < y. Since f is strictly increasing, this implies that
f(x) < f(y). So f(x) = f(y)
1
In math, strictly is often used to exclude the possibility of equality.
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 97
Case 2: y < x. Since f is strictly increasing, this implies that
f(y) < f(x). So f(x) = f(y).
In either case, we have that f(x) = f(y), which is what we needed
to show.
The phrase distinct elements of A is math jargon for x = y.
When we got partway into the proof, we had the fact x = y which isnt
easy to work with. But the trichotomy axiom for real number states that for
any x and y, we have exactly three possibilities: x = y, x < y, or y < x. The
constraint that x = y eliminates one of these possibilities.
8.9 Making this proof more succinct
In this example, the proofs for the two cases are very, very similar. So we can
fold the two cases together. Heres one approach, which I dont recommend
doing in the early part of this course but which will serve you well later on:
Proof: Let A and B be sets of numbers and let f : A B be a
strictly increasing function. Let x and y be distinct elements of
A. We need to show that f(x) = f(y).
Since x = y, theres two possibilities.
Case 1: x < y. Since f is strictly increasing, this implies that
f(x) < f(y). So f(x) = f(y)
Case 2: y < x. Similar to case 1.
In either case, we have that f(x) = f(y), which is what we needed
to show.
This method only works if you, and your reader, both agree that its
obvious that the two cases are very similar and the proof will really be similar.
Dangerous assumption right now. And weve only saved a small amount of
writing, which isnt worth the risk of losing points if the grader doesnt think
it was obvious.
CHAPTER 8. FUNCTIONS AND ONE-TO-ONE 98
But this simplication can be very useful in more complicated situations
where you have may have lots of cases, the proof for each case is long, and
the proofs for dierent cases really are very similar.
Heres another way to simplify our proof:
Proof: Let A and B be sets of numbers and let f : A B be a
strictly increasing function. Let x and y be distinct elements of
A. We need to show that f(x) = f(y).
We know that x = y, so either x < y or y < x. Without loss of
generality, assume that x < y.
Since f is strictly increasing, x < y implies that f(x) < f(y). So
f(x) = f(y), which is what we needed to show.
The phrase without loss of generality means that we are adding an
additional assumption to our proof but we claim it isnt really adding any
more information. In this case, s and t are both arbitrary indices introduced
at the same time. It doesnt really matter which of them is called s and which
is called t. So if we happen to have chosen our names in an inconvenient order,
we could simply swap the two names.
Simplifying a problem without loss of generality is a powerful but po-
tentially dangerous proof technique. Youll probably probably want to see a
few times before you use it yourself. Its actually used fairly often to simplify
proofs, so it has an abbreviation: WOLOG or WLOG.
8.10 Variation in terminology
The term injective is a synonym for one-to-one and one-to-one correspon-
dence is a synonym for bijective. The phrase by symmetry is often used
in place of without loss of generality.
Monotonically increasing is a synonym for increasing. The terms
non-decreasing and weakly increasing are also synonyms for increas-
ing, which are often used when the author worries that the reader might
forget that increasing allows plateaus.
Chapter 9
Graphs
Graphs are a very general class of object, used to formalize a wide variety of
practical problems in computer science. In this chapter, well see the basics
of (nite) undirected graphs, including graph isomorphism and connectivity.
9.1 Graphs
A graph consists of a set of nodes V and a set of edges E. Well sometimes
refer to the graph as a pair of sets (V, E). Each edge in E joins two nodes
in V . Two nodes connected by an edge are called neighbors or adjacent
For example, here is a graph in which the nodes are Illinois cities and the
edges are roads joining them:
Chicago
Springeld
Bloomington
Urbana Danville
Decatur
99
CHAPTER 9. GRAPHS 100
A graph edge can be traversed in both directions, as in this street example,
i.e. the edges are undirected. When discussing relations earlier, we used
directed graphs, in which each edge had a specic direction. Unless we
explicitly state otherwise, a graph will always be undirected. Concepts for
undirected graphs extend in straightforward ways to directed graphs.
When there is only one edge connecting two nodes x and y, we can name
the edge using the pair of nodes. We could call the edge xy or (since order
doesnt matter) yx or {x, y}. So, in the graph above, the Urbana-Danville
edge connects the node Urbana and the node Danville.
In some applications, we need graphs in which two nodes are connected
by multiple edges, i.e. parallel edges with the same endpoints. For exam-
ple, the following graph shows ways to travel among four cities in the San
Francisco Bay Area. It has three edges from San Francisco to Oakland, repre-
senting dierent modes of transportation. When multiple edges are present,
we typically label the edges rather than trying to name edges by their end-
points. This diagram also illustrates a loop edge which connects a node to
itself.
San Francisco Oakland
Fremont Palo Alto
ferry
plane
I-80
tourist cruise
Rt 101
Rt 84
I-880
A graph is called a simple graph if it has neither multiple edges nor
loop edges. Unless we explicitly state otherwise, a graph will always be a
simple graph. Also, well assume that it has at least one node and that it
has only a nite number of edges and nodes. Again, most concepts extend
in a reasonable way to innite and non-simple graphs.
CHAPTER 9. GRAPHS 101
9.2 Degrees
The degree of a node v, written deg(v) is the number of edges which have v as
an endpoint. Self-loops, if you are allowing them, count twice. For example,
in the following graph, a has degree 2, b has degree 6, d has degree 0, and so
forth.
a
b
e
c
d
f
Each edge contributes to two node degrees. So the sum of the degrees of
all the nodes is twice the number of edges. This is called the Handshaking
Theorem and can be written as

vV
deg(v) = 2|E|
This is a slightly dierent version of summation notation. We pick each
node v in the set V , get its degree, and add its value into the sum. Since
V is nite, we could also have given names to the nodes v
1
, . . . , v
n
and then
written
n

k=1
v
k
V deg(v) = 2|E|
The advantage to the rst, set-based, style is that it generalizes well to
situations involving innite sets.
CHAPTER 9. GRAPHS 102
9.3 Complete graphs
Several special types of graphs are useful as examples. First, the complete
graph on n nodes (shorthand name K
n
), is a graph with n nodes in which
every node is connected to every other node. K
5
is shown below.
a
b
c
d
e
To calculate the number of edges in K
n
, think about the situation from
the perspective of the rst node. It is connected to n 1 other nodes. If we
look at the second node, it adds n 2 more connections. And so forth. So
we have

n
k=1
(n k) =

n1
k=0
k =
n(n1)
2
edges.
9.4 Cycle graphs and wheels
Suppose that we have n nodes named v
1
, . . . , v
n
, where n 3. Then the
cycle graph C
n
is the graph with these nodes and edges connecting v
i
to v
i+1
,
plus an additional edge from v
n
to v
1
. That is, the set of edges is:
E = {v
1
v
2
, v
2
v
3
, . . . , v
n1
v
n
, v
n
v
1
}
So C
5
looks like
CHAPTER 9. GRAPHS 103
a
b
c
d
e
C
n
has n nodes and also n edges. Cycle graphs often occur in networking
applications. They could also be used to model games like telephone where
people sit in a circle and communicate only with their neighbors.
The wheel W
n
is just like the cycle graph C
n
except that it has an addi-
tional central hub node which is connected to all the others. Notice that
W
n
has n + 1 nodes (not n nodes). It has 2n edges. For example, W
5
looks
like
a
b
c
d
e
hub
9.5 Isomorphism
In graph theory, we only care about how nodes and edges are connected
together. We dont care about how they are arranged on the page or in
space, how the nodes and edges are named, and whether the edges are drawn
as straight or curvy. We would like to treat graphs as interchangeable if they
have the same abstract connectivity structure.
Specically, suppose that G
1
= (V
1
, E
1
) and G
2
= (V
2
, E
2
) are graphs.
An isomorphism from G
1
to G
2
is a bijection f : V
1
V
2
such that nodes
a and b are joined by an edge if and only if f(a) and f(b) are joined by an
CHAPTER 9. GRAPHS 104
edge. The graphs G
1
and G
2
are isomorphic if there is an isomorphism
from G
1
to G
2
.
For example, the following two graphs are isomorphic. We can prove this
by dening the function f so that it maps 1 to d, 2 to a, 3 to c, and 4 to b.
The reader can then verify that edges exist in the left graph if and only if
the corresponding edges exist in the right graph.
a
b
c
d
1
2
3
4
Graph isomorphism is another example of an equivalence relation. Each
equivalence class contains a group of graphs which are supercially dierent
(e.g. dierent names for the nodes, drawn dierently on the page) but all
represent the same underlying abstract graph.
To prove that two graphs are not isomorphic, we could walk through
all possible functions mapping the nodes of one to the nodes of the other.
However, thats a huge number of functions for graphs of any interesting
size. An exponential number, in fact. Instead, a better technique for many
examples is to notice that a number of graph properties are invariant, i.e.
preserved by isomorphism.
The two graphs must have the same number of nodes and the same
number of edges.
For any node degree k, the two graphs must have the same number of
nodes of degree k. For example, they must have the same number of
nodes with degree 3.
We can prove that two graphs are not isomorphic by giving one example
of a property that is supposed to be invariant but, in fact, diers between
the two graphs. For example, in the following picture, the lefthand graph
has a node of degree 3, but the righthand graph has no nodes of degree 3, so
they cant be isomorphic.
CHAPTER 9. GRAPHS 105
a
b
c
d
1
2
3
4
9.6 Subgraphs
Its not hard to nd a pair of graphs that arent isomorphic but where the
most obvious properties (e.g. node degrees) match. To prove that such a pair
isnt isomorphic, its often helpful to focus on certain specic local features of
one graph that arent present in the other graph. For example, the following
two graphs have the same node degrees: one node of degree 1, three of degree
2, one of degree 3. However, a little experimentation suggests they arent
isomorphic.
a
b
c
d
e
1
2
3
4 5
To make a convincing argument that these graphs arent isomorphic, we
need to dene the notion of a subgraph. If G and G

are graphs, then G

is a subgraph of G if and only if the nodes of G

are a subset of the nodes of

G and the edges of G

are a subset of the edges of G. If two graphs G and

F are isomorphic, then any subgraph of G must have a matching subgraph
somewhere in F.
A graph has a huge number of subgraphs. However, we can usually nd
evidence of non-isomorphism by looking at small subgraphs. For example, in
the graphs above, the lefthand graph has C
3
as a subgraph, but the righthand
graph does not. So they cant be isomorphic.
CHAPTER 9. GRAPHS 106
9.7 Walks, paths, and cycles
In a graph G, a walk of length k from node a to node b is a nite sequence
of nodes a = v
1
, v
2
, . . . , v
n
= b and a nite sequence of edges e
1
, e
2
, . . . , e
n1
in which e
i
connects v
i
and v
i+1
, for all i. Under most circumstances, it isnt
necessary to give both the sequence of nodes and the sequence of edges: one
of the two is usually sucient. The length of a walk is the number of edges
in it. The shortest walks consist of just a single node and have length zero.
A walk is closed if its starting and ending nodes are the same. Otherwise
it is open. A path is a walk in which no node is used more than once. A
cycle is a closed walk with at least three nodes in which no node is used
more than once except that the starting and ending nodes are the same.
For example, in the following graph, there is a length-3 walk from a to
e: ac, cd, de. Another walk of length 3 would have edges: ab, bd, de. These
two walks are also paths. There are also longer walks from a to e, which
arent paths because they re-use nodes, e.g. the walk with a node sequence
a, c, d, b, d, e. If you have a walk between two nodes, you can always creat a
path between the nodes by pruning unnecessary loops from the walk.
a
b
c
d
e
In the following graph, one cycle of length 4 has edges: ab, bd, dc, ca.
Other closely-related cycles go through the same nodes but with a dierent
starting point or in the opposite direction, e.g. dc, bd, ab, ca. (Remember
that cd and dc are two names for the same edge.) Unlike cycles, closed walks
can re-use nodes, e.g. ab, ba, ac, ce, ec, ca is a closed walk but not a cycle.
CHAPTER 9. GRAPHS 107
a
b
c
d
e
f
The following graph is acyclic, i.e. it doesnt contain any cycles.
a
b
c
d
e
f
Notice that the cycle graph C
n
contains 2n dierent cycles. For example,
if the vertices of C
4
are labelled as shown below, then one cycle is ab, bc, cd, da,
another is cd, bc, ab, da, and so forth.
a
b
c
d
9.8 Connectivity
A graph G is connected if there is a walk between every pair of nodes in
G. Our previous examples of graphs were connected. The following graph is
not connected, because there is no walk from (for example), a to g.
CHAPTER 9. GRAPHS 108
a
b
c
d
e
f
g
h
If we have a graph G that might or might not be connected, we can di-
vide G into connected components. Each connected component contains
a maximal (i.e. biggest possible) set of nodes that are all connected to one
another, plus all their edges. So, the above graph has three connected com-
ponents: one containing nodes a,b,c, and d, a second containing nodes e, f,
and g, and a third that contains only the node h.
Sometimes two parts of a graph are connected by only a single edge, so
that the graph would become disconnected if that edge were removed. This
is called a cut edge. For example, in the following graph, the edge ce is a
cut edge. In some applications, cut edges are a problem. E.g. in network-
ing, they are places where the network is vulnerable. In other applications
(e.g. compilers), they represent opportunities to divide a larger problem into
several simpler ones.
a
b
c
d
e
f
g
9.9 Distances
In graphs, distances are based on the lengths of paths connecting pairs of
nodes. Specically, the distance d(a, b) between two nodes a and b is the
CHAPTER 9. GRAPHS 109
length of the shortest path from a to b. The diameter of a graph is the
maximum distance between any pair of nodes in the graph. For example,
the lefthand graph below has diameter 4, because d(f, e) = 4 and no other
pair of nodes is further apart. The righthand graph has diameter 2, because
d(1, 5) = 2 and no other pair of nodes is further apart.
f
a
b
c
d
e
1
2
3
4
5
9.10 Euler circuits
An Euler circuit of a graph G is a closed walk that uses each edge of the
graph exactly once. Notice that G has an Euler circuit means that every
edge of G is in the circuit; its not enough for some subgraph of G to have a
suitable circuit. For example, one Euler circuit of the following graph would
be ac, cd, df, fe, ec, cf, fb, ba.
a
b
c
d
e
f
An Euler circuit is possible exactly when the graph is connected and each
node has even degree. Each node has to have even degree because, in order
to complete the circuit, you have to leave each node that you enter. If the
node has odd degree, you will eventually enter a node but have no unused
edge to go out on.
CHAPTER 9. GRAPHS 110
Fascination with Euler circuits dates back to the 18th century. At that
time, the city of Konigberg, in Prussia, had a set of bridges that looked
roughly as follows:
Folks in the town wondered whether it was possible to take a walk in
which you crossed each bridge exactly once, coming back to the same place
you started. This is the kind of thing that starts long debates late at night in
pubs, or keeps people amused during boring church services. Leonard Euler
was the one who explained clearly why this isnt possible.
For our specic example, the corresponding graph looks as follows. Since
all of the nodes have odd degree, theres no possibility of an Euler circuit.
9.11 Bipartite graphs
Another special type of graph is a bipartite graph. A graph G = (V, E) is
bipartite if we can split V into two non-overlapping subsets V
1
and V
2
such
that every edge in G connects an element of V
1
with an element of V
2
. That
is, no edge connects two nodes from the same part of the division.
CHAPTER 9. GRAPHS 111
For example, the cube graph is bipartite. If we assign nodes to the R and
G groups as shown below, then there are no edges connecting an R node to
an R node, or a G node to a G node.
R
R G
G
G
G R
R
Bipartite graphs often appear in matching problems, where the two sub-
sets represent dierent types of objects. For example, one group of nodes
might be students, the other group of nodes might be workstudy jobs, and
the edges might indicate which jobs each student is interested in.
The complete bipartite graph K
m,n
is a bipartite graph with m nodes in
V
1
, n nodes in V
2
, and which contains all possible edges that are consistent
with the denition of bipartite. The diagram below shows a partial bipartite
graph on a set of 7 nodes, as well as the complete bipartite graph K
3,2
.
The complete bipartite graph K
m,n
has m +n nodes and mn edges.
CHAPTER 9. GRAPHS 112
9.12 Variation in terminology
Although the core ideas of graph theory are quite stable, terminology varies
a lot and there is a huge range of specialized terminology for specic types
of graphs. In particular, nodes are often called vertices Notice that the
singular of this term is vertex (not vertice). A complete graph on some
set of vertices is also known as a clique.
What were calling a walk used to be widely called a path. Authors
who still use this convention would then use the term simple path to ex-
clude repetition of vertices. Terms for closely-related concepts, e.g. cycle,
often change as well.
Chapter 10
2-way Bounding
In high school and early college mathematics, we are often proving equalities
and we often prove them by manipulating a sequence of equalities. For
example
3x
2
5x 2
x 2
=
(x 2)(3x + 1)
x 2
= 3x + 1
On more complex problems, its often necessary to adopt a more exi-
ble approach: bound the quantity of interest from both directions. When
our best upper and lower bounds are equal, weve established an equality.
Otherwise, weve limited the quantity to a known (hopefully small) range.
Because we need to establish a lower bound and also an upper bound, a 2-
way bounding proof contains two sub-proofs. Sometimes the two sub-proofs
are similar. However, the true power of the technique comes from the fact
that the two sub-proofs can use entirely dierent techniques. Therefore, the
method is widely used when proving more dicult results, e.g. in upper-level
computer science and mathematics courses (e.g. real analysis, algorithms).
The most common error building these proofs is to omit one half entirely.
For example, someone intending to show that f(x) = k might prove that
f(x) k but forget to also prove that f(x) k. Do not do this. If one of
the bounds is obvious, say that explicitly.
In this chapter, we will see how this method works on a variety of exam-
113
CHAPTER 10. 2-WAY BOUNDING 114
ples.
10.1 Navigation
Suppose that were exploring the Midwest by car, with a GPS device on-
board, and suppose wed like to nd the driving distance from Urbana, Illinois
to Grand Rapids, Michigan. We could look at the dierence between the GPS
positions of the two locations. This gives us a straight-line distance, which
is a lower bound on the length of the best car route.
Now, suppose that have to go to Grand Rapids on a regular basis. So we
measure our mileage as we drive there from Urbana using dierent choices
of route. The minimum trip mileage found so far is an upper bound on the
length of the best car route.
In this example, we would normally expect our lower bound and our
upper bound to be some distance apart. To bring them together and have
an exact answer for the driving distance, we would need to establish that we
had exhaustively explored all possible routes.
10.2 Pigeonhole point placement
Heres a claim that can be proved using the pigeonhole principle. Like so
many pigeonhole proofs, the proof depends on a non-obvious trick. In this
case, its a trick about dividing up a triangle. Once youve seen the trick,
however, the proof is a classic example of bounding a quantity from above
and below.
Claim 36 Suppose that T is an equilateral triangle with sides of length 2
units We can place a maximum of four points in the triangle such that every
pair of points are more than 1 unit apart.
To show that 4 is the maximum number of points we can place with the
required separations, we need to show that its possible to place 4 points,
but it is not possible to place 5 points. Its easiest to tackle these two sub-
problems separately, using dierent techniques.
CHAPTER 10. 2-WAY BOUNDING 115
Proof: To show that the maximum is at least four, notice that
we can place three points at the corners of the triangle and one
point in the center. The points at the corners are two units apart.
To see that the point in the center is more than one unit from
any corner, notice that the center, the corner, and the midpoint
of the side form a right triangle.
1
The hypotenuse of this triangle connects the center point to the
corner point. Since one leg of the triangle has length 1, the hy-
potenuse must have length greater than 1.
To show that the maximum number of points cannot be greater
than four, divide up the triangle into four smaller equilateral tri-
angles as follows:
Suppose we tried to place ve or more points into the big trian-
gle. Since there are only four small triangles, by the pigeonhole
principle, some small triangle would have to contain at least two
points. But since the small triangle has side length only 1, these
points cant be separated by more than one unit.
CHAPTER 10. 2-WAY BOUNDING 116
10.3 Graph coloring
A coloring of a graph G assigns a color to each node of G, with the restriction
that two adjacent nodes never have the same color. If G can be colored with
k colors, we say that G is k-colorable. The chromatic number of G, written
(G), is the smallest number of colors needed to color G.
For example, only three colors are required for this graph:
R G
G B
R
But the complete graph K
n
requires n colors, because each node is adja-
cent to all the other nodes. E.g. K
4
can be colored as follows:
R G
B Y
To establish that n is the chromatic number for a graph G, we need to
establish two facts:
(G) n: G can be colored with n colors.
(G) n: G cannot be colored with less than n colors
For small nite graphs, the simplest way to show that (G) n is to show
a coloring of G that uses n colors. For a larger class of graphs, we could
describe an algorithm for doing the coloring. For example, we can color a
CHAPTER 10. 2-WAY BOUNDING 117
cyclic graph with an even number of nodes by alternating colors around the
circle.
Showing that (G) n can sometimes be equally straightforward. For
example, if the graph has any edges at all, its chromatic number must be at
least 2. A graph is bipartite if and only if its chromatic number is 2. And G
may contain a copy of K
n
, which cant be colored with less than n colors.
However, ihe following example shows why this process can get tricky.
Its relatively easy to color with 4 colors. But the largest complete graph in
it is K
3
, which gives us a lower bound of only 3. Showing that the chromatic
number is actually 4 requires carefully stepping through all possible ways to
assign three colors to the nodes and explaining why none can end with a
complete coloring.
10.4 Why care about graph coloring?
Graph coloring is required for solving a wide range of practical problems. For
example, there is a coloring algorithm embedded in most compilers. Because
the general problem cant be solved eciently, the implemented algorithms
use limitations or approximations of various sorts so that they can run in a
reasonable amount of time.
For example, catalog and on-line retailers frequently organize their clothes
oerings into collections, e.g. Golng Geezer or Tough but Toucheable
CHAPTER 10. 2-WAY BOUNDING 118
or Uniforms 4 U. All items in a collection are guaranteed to be compatible.
This helps retailers market clothes to folks with no dress sense, as well as
cope with the fact that people are sensitive to ne color distinctions that
arent reproduced accurately in catalogs and on computer screens. How
many collections are required to organize the retailers inventory of clothes?
We can model this problem as graph coloring. Each graph node is an
item of clothing. Edges join pairs of items that arent compatible, e.g. the
bright green pants dont go with the grey-orange shirt, the business formal
blouse doesnt go with the golng pants. The color on each node is the
name of a collection. If we nd that too many collections are required, we
might want to remove selected items of clothing (e.g. the bottle green bell
bottoms that match nothing) from our inventory.
We can model a sudoku puzzle by setting up one node for each square.
The colors are the 9 numbers, and some are pre-assigned to certain nodes.
Two nodes are connected if their squares are in the same block or row or
column. The puzzle is solvable if we can 9-color this graph, respecting the
pre-assigned colors.
We can model exam scheduling as a coloring problem. The exams for two
courses should not be put at the same time if there is a student who is in
both courses. So we can model this as a graph, in which each course is a node
and courses are connected by edges if they share students. The question is
then whether we can color the graph with k colors, where k is the number of
exam times in our schedule.
In the exam scheduling problem, we actually expect the answer to be
no, because eliminating conicts would require an excessive number of
exam times. So the real practical problem is: how few students do we have
to take out of the picture (i.e. give special conict exams to) in order to be
able to solve the coloring problem with a reasonable value for k. We also
have the option of splitting a course (i.e. oering a scheduled conict exam)
to simplify the graph.
A particularly important use of coloring in computer science is register
allocation. A large java or C program contains many named variables. But
a computer has a smallish number (e.g. 32) of fast registers which can feed
basic operations such as addition. So variables must be allocated to specic
registers.
CHAPTER 10. 2-WAY BOUNDING 119
The nodes in this coloring problem are variables. The colors are registers.
Two variables are connected by an edge if they are in use at the same time
and, therefore, cannot share a register. As with the exam scheduling problem,
we actually expect the raw coloring problem to fail. The compiler then uses
so-called spill operations to break up the dependencies and create a graph
we can color with our limited number of registers. The goal is to use as few
spills as possible.
10.5 Proving set equality
Finally, 2-way bounding proofs are often used to prove that two sets A and
B are equal. That is, we show that A B and B A, using separate
subproofs. We can then conclude that A = B. This looks supercially
dierent from the previous examples, because the relation is rather than
. But the main idea is the same: we rst show that A is no larger than B,
and then show that A is no smaller than B.
As an example, lets look at
Claim 37 Let A = {12p + 9q | p, q Z} Then A = {multiples of 3}.
Before you try to prove this, rst put some sample integer values into
the formula 12p + 9q. See if you can convince yourself informally that this
formula really generates all and only multiples of 3.
Using the bounding method, we would write the proof as follows:
Proof:
(1) Show that A {multiples of 3}.
Let x be an element of A. By the denition of A, x = 15p + 9q,
for some integers p and q. But then x = 3(5p +3q). 4p +3q is an
integer, since p and q are integers. So x is a multiple of 3.
(2) Show that {multiples of 3} A.
Notice that (*) 15 (1) + 9 2 = 3.
CHAPTER 10. 2-WAY BOUNDING 120
Let x be a multiple of 3. Then x = 3n for some integer n.
Substituting (*) into this equation, we get x = (15 (1) +9 2)n.
So x = 15 (n) + 9 (2n). So x is an element of A.
Since weve shown that A {multiples of 3} and {multiples of 3}
A, we can conclude that A = {multiples of 3}.
10.6 Variation in terminology
When coloring graphs, we have placed colors on vertices. A closely-related
set of problems involve placing colors on edges. The terms vertex coloring
and edge coloring are used when its necessary to distinguish the two.
Chapter 11
Induction
This chapter covers mathematical induction.
11.1 Introduction to induction
At the start of the term, we saw the following formula for computing the
sum of the rst n integers:
Claim 38 For any positive integer n,
n
i=1
i =
n(n+1)
2
.
At that point, we didnt prove this formula correct, because this is most
easily done using a new proof technique: induction.
Mathematical induction is a technique for showing that a statement P(n)
is true for all natural numbers n, or for some innite subset of the natural
numbers (e.g. all positive even integers). Its a nice way to produce quick,
easy-to-read proofs for a variety of fact that would be awkward to prove with
the techniques youve seen so far. It is particularly well suited to analyzing
the performance of recursive algorithms. Most of you have seen a few of these
in previous programming classes; youll see many more in later classes.
Induction is very handy, but it may strike you as a bit weird. It may
take you some time to get used to it. In fact, you have two tasks which are
somewhat independent:
121
CHAPTER 11. INDUCTION 122
Learn how to write an inductive proof.
Understand why inductive proofs are legitimate.
You can learn to write correct inductive proofs even if you remain some-
what unsure of why the method is legitimate. Over the next few classes,
youll gain condence in the validity of induction and its friend recursion.
11.2 An Example
A proof by induction has the following outline:
Claim: P(n) is true for all positive integers n.
Proof: Well use induction on n.
Base: We need to show that P(1) is true.
Induction: Suppose that P(n) is true for n = 1, 2, . . . , k1. We
need to show that P(k) is true.
The part of the proof labelled induction is a conditional statement. We
assume that P(n) is true for values of n no larger than k1. This assumption
is called the inductive hypothesis. We use this assumption to show that P(k)
is true.
For our formula example, our proposition P(n) is
n
i=1
i =
n(n+1)
2
. Sub-
stituting this denition of P into the outline, we get the following outline for
our specic claim:
Proof: We will show that
n
i=1
i =
n(n+1)
2
for any positive integer
n, using induction on n.
Base: We need to show that the formula holds for n = 1, i.e.

1
i=1
i =
12
2
.
Induction: Suppose that
n
i=1
i =
n(n+1)
2
for n = 1, 2, . . . , k 1.
We need to show that
k
i=1
i =
k(k+1)
2
.
CHAPTER 11. INDUCTION 123
The full proof might then look like:
Proof: We will show that
n
i=1
i =
n(n+1)
2
for any positive integer
n, using induction on n.
Base: We need to show that the formula holds for n = 1.
1
i=1
i =
1. And also
12
2
= 1. So the two are equal for n = 1.
Induction: Suppose that
n
i=1
i =
n(n+1)
2
for n = 1, 2, . . . , k 1.
We need to show that
k
i=1
i =
k(k+1)
2
.
By the denition of summation notation,
k
i=1
i = (
k1
i=1
i) +k
Our inductive hypothesis states that at n = k 1,
k1
i=1
i =
(
(k1)k
2
).
Combining these two formulas, we get that
k
i=1
i = (
(k1)k
2
) +k.
But (
(k1)k
2
) +k = (
(k1)k
2
) +
2k
2
= (
(k1+2)k
2
) =
k(k+1)
2
.
So, combining these equations, we get that
k
i=1
i =
k(k+1)
2
which
is what we needed to show.
One way to think of a proof by induction is that its a template for
building direct proofs. If I give you the specic value n = 47, you could
write a direct proof by starting with the base case and using the inductive
step 46 times to work your way from the n = 1 case up to the n = 47 case.
11.3 Why is this legit?
There are several ways to think about mathematical induction, and under-
stand why its a legitimate proof technique. Dierent people prefer dierent
motivations at this point, so Ill oer several.
Domino Theory: Imagine an innite line of dominoes. The
base step pushes the rst one over. The inductive step claims
that one domino falling down will push over the next domino
in the line. So dominos will start to fall from the beginning all
the way down the line. This process continues forever, because
CHAPTER 11. INDUCTION 124
the line is innitely long. However, if you focus on any specic
domino, it falls after some specic nite delay.
Recursion fairy: The recursion fairy is the mathematicians
version of a programming assistant. Suppose you tell her how to
do the proof for P(1) and also why P(1) up through P(k) implies
P(k + 1). Then suppose you pick any integer (e.g. 1034) then
she can take this recipe and use it to ll in all the details of a
normal direct proof that P holds for this particular integer. That
is, she takes P(1), then uses the inductive step to get from P(1)
to P(2), and so on up to P(1034).
Dening property of the integers: The integers are set up
mathematically so that induction will work. Some formal sets of
axioms dening the integers include a rule saying that induction
works. Other axiom sets include the well-ordering property:
any subset that has a lower bound also has a smallest element.
This is equivalent to an axiom that explicitly states that induction
works. Both of these axioms prevent the integers from having
extra very large elements that cant be reached by repeatedly
adding one to some starting integer. So, for example, is not
an integer.
These arguments dont depend on whether our starting point is 1 or some
other integer, e.g. 0 or 2 or -47. All you need to do is ensure that your base
case covers the rst integer for which the claim is supposed to be true.
11.4 Building an inductive proof
In constructing an inductive proof, youve got two tasks. First, you need
to set up this outline for your problem. This includes identifying a suitable
proposition P and a suitable integer variable n.
Notice that P(n) must be a statement, i.e. something that is either true
or false. For example, it is never just a formula whose value is a number.
Also, notice that P(n) must depend on an integer n. This integer n is known
CHAPTER 11. INDUCTION 125
as our induction variable. The assumption at the start of the inductive step
(P(k) is true) is called the inductive hypothesis.
Your second task is to ll in the middle part of the induction step. That
is, you must gure out how to relate a solution for a larger problem P(k) to
a solution for one or more of the smaller problems P(1) . . . , P(k 1). Most
students want to do this by starting with a small problem, e.g. P(k 1), and
adding something to it. For more complex situations, however, its usually
better to start with the larger problem and try to nd an instance of the
smaller problem inside it.
11.5 Another example
The previous example applied induction to an algebraic formula. We can
also apply induction to other sorts of statements, as long as they involve a
suitable integer n.
Claim 39 For any natural number n, n
3
n is divisible by 3.
In this case, P(n) is n
3
n is divisible by 3.
So the outline of our proof looks like
Proof: By induction on n.
Base: Let n = 0. [prove that n
3
n is divisible by 3 when n = 0]
Induction: Suppose that n
3
n is divisible by 3, for n = 0, 1, . . . , k.
We need to show that (k + 1)
3
(k + 1) is divisible by 3.
Fleshing out the details of the algebra, we get the following full proof:
Proof: By induction on n.
Base: Let n = 0. Then n
3
n = 0
3
0 = 0 which is divisible by
3.
Induction: Suppose that n
3
n is divisible by 3, for n = 0, 1, . . . , k.
We need to show that (k + 1)
3
(k + 1) is divisible by 3.
CHAPTER 11. INDUCTION 126
(k+1)
3
(k+1) = (k
3
+3k
2
+3k+1)(k+1) = (k
3
k)+3(k
2
+k)
From the inductive hypothesis, (k
3
k) is divisible by 3. And
3(k
2
+ k) is divisible by 3 since (k
2
+ k) is an integer. So their
sum is divisible by 3. That is (k + 1)
3
(k + 1) is divisible by 3.

Notice that weve also used a variation on our induction outline, where
the induction hypothesis covers values up through k (instead of k 1) and
we prove the claim at n = k + 1 (instead of at n = k). It doesnt matter
whether your hypothesis goes through n = k 1 or n = k, as long as you
prove the claim for the next larger integer.
Also notice that our hypothesis says that the claim holds for n = 0, 1, . . . , k.
We are using the convention that this notation implies that k 0, not that
k is necessarily any larger (e.g. k 1). That is, the 1 in this expresion is
used only to show the pattern of increase, and it does not imply anything
about the size of k.
The zero base case is technically enough to make the proof solid, but
sometimes a zero base case doesnt provide good intuition or condence. So
youll sometimes see an extra base case written out, e.g. n = 1 in this
example, to help the author or reader see why the claim is plausible.
11.6 Some comments about style
Notice that the start of the proof tells you which variable in your formula (n
in this case) is the induction variable. In this formula, the choice of induc-
tion variable is fairly obvious. But sometimes theres more than one integer
oating around that might make a plausible choice for the induction variable.
Its good style to always mention that you are doing a proof by induction,
say what your induction variable is, and label your base and inductive steps.
Notice that the proof of the base case is very short. In fact, Ive written
about about twice as long as youd normally see it. Almost all the time, the
base case is trivial to prove and fairly obvious to both you and your reader.
Often this step contains only some worked algebra and a check mark at the
end. However, its critical that you do check the base case. And, if your base
CHAPTER 11. INDUCTION 127
case involves an equation, compute the results for both sides (not just one
side) so you can verify they are equal.
The important part of the inductive step is ensuring that you assume
P(1), . . . , P(k 1) and use these facts to show P(k). At the start, you must
spell out your inductive hypothesis, i.e. what P(n) is for your claim. Its
usually helpful to explicitly substitute in some key values for n, e.g. work out
what P(k 1) is. Make sure that you use the information from the inductive
hypothesis in your argument that P(k + 1) holds. If you dont, its not an
inductive proof and its very likely that your proof is buggy.
At the start of the inductive step, its also a good idea to say what you
need to show, i.e. quote what P(k) is.
These style issues are optional in theory, but actually critical for be-
ginners writing inductive proofs. You will lose points if your proof isnt clear
and easy to read. Following these style points (e.g. labelling your base and
inductive steps) is a good way to ensure that it is, and that the logic of your
proof is correct.
11.7 A geometrical example
Lets see another example of the basic induction outline, this time on a
geometrical application. Tiling some area of space with a certain type of
puzzle piece means that you t the puzzle pieces onto that area of space
exactly, with no overlaps or missing areas. A right triomino is a 2-by-2
square minus one of the four squares.
I claim that
CHAPTER 11. INDUCTION 128
Claim 40 For any positive integer n, a 2
n
2
n
checkerboard with any one
square removed can be tiled using right triominoes.
Proof: by induction on n.
Base: Suppose n = 1. Then our 2
n
2
n
checkerboard with one
square removed is exactly one right triomino.
Induction: Suppose that the claim is true for n = 1, . . . , k. That
is a 2
n
2
n
checkerboard with any one square removed can be
tiled using right triominoes as long as n k.
Suppose we have a 2
k+1
2
k+1
checkerboard C with any one
square removed. We can divide C into four 2
k
2
k
sub-checkerboards
P, Q, R, and S. One of these sub-checkerboards is already miss-
ing a square. Suppose without loss of generality that this one is
S. Place a single right triomino in the middle of C so it covers
one square on each of P, Q, and R.
Now look at the areas remaining to be covered. In each of the
sub-checkerboards, exactly one square is missing (S) or already
covered (P, Q, and R). So, by our inductive hypothesis, each
of these sub-checkerboards minus one square can be tiled with
right triominoes. Combining these four tilings with the triomino
we put in the middle, we get a tiling for the whole of the larger
checkerboard C. This is what we needed to construct.
11.8 Graph coloring
We can also use induction to prove a useful general fact about graph col-
orability:
Claim 41 If all nodes in a graph G have degree D, then G can be colored
with D + 1 colors.
The objects involved in this claim are graphs. To apply induction to
objects like graphs, we organize our objects by their size. Each step in the
induction process will show that the claim holds for all objects of a particular
CHAPTER 11. INDUCTION 129
(integer) size. For graphs, the size would typically be either the number
of nodes or the number of edges. For this proof, its most convenient to use
the number of nodes.
Proof: by induction on the number of nodes in G.
Base: The graph with just one node has maximum degree 0 and
can be colored with one color.
Induction: Suppose that any graph with at most k 1 nodes and
maximum node degree D can be colored with D + 1 colors.
Let G be a graph with k nodes and maximum node degree D.
Remove some node v (and its edges) from G to create a smaller
graph G

.
G

has k 1 nodes. Also, the maximum node degree of G

is
no larger than D, because removing a node cant increase the
degree. So, by the inductive hypothesis, G

can be colored with

D + 1 colors.
Because v has at most D neighbors, its neighbors are only using
D of the available colors, leaving a spare color that we can assign
to v. The coloring of G

can be extended to a coloring of G with

D + 1 colors.
We can use this idea to design an algorithm (called the greedy algo-
rithm) for coloring a graph. This algorithm walks through the nodes one-by-
one, giving each node a color without revising any of the previously-assigned
colors. When we get to each node, we see what colors have been assigned to
its neighbors. If there is a previously used color not assigned to a neighbor,
we re-use that color. Otherwise, we deploy a new color. The above theorem
shows that the greedy algorithm will never use more than D + 1 colors.
Notice, however, that D + 1 is only an upper bound on the chromatic
number of the graph. The actual chromatic number of the graph might be a
lot smaller. For example, D +1 would be 7 for the wheel graph W
6
but this
graph actually has chromatic number only three:
CHAPTER 11. INDUCTION 130
G B
G B
B G R
The performance of the greedy algorithm is very sensitive to the order
in which the nodes are considered. For example, suppose we start coloring
W
6
by coloring the center hub, then a node v on the outer ring, and then
the node opposite v. Then our partial coloring might look as shown below.
Completing this coloring will require using four colors.
G G R
Notice that whether we need to deploy a new color to handle a node isnt
actually determined by the degree of the node but, rather, by how many of
its neighbors are already colored. So a useful heuristic is to order nodes by
their degrees and color higher-degree nodes earlier in the process. This tends
to mean that, when we reach a high-degree node, some of its neighbors will
not yet be colored. So we will be able to handle the high-degree nodes with
fewer colors and then extend this partial coloring to all the low-degree nodes.
11.9 Postage example
In the inductive proofs weve seen so far, we didnt actually need the full
information in our inductive hypothesis. Our inductive step assumed that
P(n) was true for all values of n from the base up through k 1, a so-
CHAPTER 11. INDUCTION 131
called strong inductive hypothesis. However, the rest of the inductive step
actually depended only on the information that P(k 1) was true. We
could, in fact, have used a simpler inductive hypothesis, known as a weak
inductive hypothesis, in which we just assumed that P(k 1) was true.
Well now see some examples where a strong inductive hypothesis is es-
sential, beucase the result for n = k depends on the result for some smaller
value of n, but its not the immediately previous value k 1. Heres a classic
example:
Claim 42 Every amount of postage that is at least 12 cents can be made
from 4-cent and 5-cent stamps.
Before trying to prove a claim of this sort, you should try small examples
and verify that the stated base case is correct. That is, in this case, that you
can make postage for some representative integers, 12, but you cant do it
for 11.
For example, 12 cents uses three 4-cent stamps. 13 cents of postage uses
two 4-cent stamps plus a 5-cent stamp. 14 uses one 4-cent stamp plus two
5-cent stamps. If you experiment with small values, you quickly realize that
the formula for making k cents of postage depends on the one for making
k 4 cents of postage. That is, you take the stamps for k 4 cents and add
another 4-cent stamp. We can make this into an inductive proof as follows:
Proof: by induction on the amount of postage.
Base: If the postage is 12 cents, we can make it with three 4-cent
stamps. If the postage is 13 cents, we can make it with two 4-cent
stamps. plus a 5-cent stamp. If it is 14, we use one 4-cent stamp
plus two 5-cent stamps. If it is 15, we use three 5-cent stamps.
Induction: Suppose that we have show how to construct postage
for every value from 12 up through k 1. We need to show how
to construct k cents of postage. Since weve already proved base
cases up through 15 cents, well assume that k 16.
Since k 16, k 4 12. So by the inductive hypothesis, we
can construct postage for k 4 cents using m 4-cent stamps and
CHAPTER 11. INDUCTION 132
n 5-cent stamps, for some natural numbers m and n. In other
words k 4 = 4m+ 5n.
But then k = 4(m + 1) + 5n. So we can construct k cents of
postage using m+1 4-cent stamps and n 5-cent stamps, which is
what we needed to show.
Notice that we needed to directly prove four base cases, since we needed
to reach back four integers in our inductive step. Its not always obvious how
many base cases are needed until you work out the details of your inductive
step. The rst base case was n = 12 because our claim only starts working
consistently for integers 12.
11.10 Nim
In the parlour game Nim, there are two players and two piles of matches. At
each turn, a player removes some (non-zero) number of matches from one of
the piles. The player who removes the last match wins.
1
Claim 43 If the two piles contain the same number of matches at the start
of the game, then the second player can always win.
Heres a winning strategy for the second player. Suppose your opponent
removes m matches from one pile. In your next move, you remove m matches
from the other pile, thus evening up the piles. Lets prove that this strategy
works.
Proof by induction on the number of matches (n) in each pile.
Base: If both piles contain 1 match, the rst player has only one
possible move: remove the last match from one pile. The second
player can then remove the last match from the other pile and
thereby win.
1
Or, in some variations, loses. There seem to be several variations of this game.
CHAPTER 11. INDUCTION 133
Induction: Suppose that the second player can win any game that
starts with two piles of n matches, where n is any value from 1
through k 1. We need to show that this is true if n = k.
So, suppose that both piles contain k matches. A legal move by
the rst player involves removing j matches from one pile, where
1 j k. The piles then contain k matches and k j matches.
The second player can now remove j matches from the other pile.
This leaves us with two piles of k j matches. If j = k, then
the second player wins. If j < k, then were now eectively at
the start of a game with k j matches in each pile. Since j 1,
k j k 1. So, by the induction hypothesis, we know that the
second player can nish the rest of the game with a win.
The induction step in this proof uses the fact that our claim P(n) is
true for a smaller value of n. But since we cant control how many matches
the rst player removes, we dont know how far back we have look in the
sequence of earlier results P(1) . . . P(k). Our previous proof about postage
can be rewritten so as to avoid strong induction. Its less clear how to rewrite
proofs like this Nim example.
11.11 Prime factorization
Early in this course, we saw the Fundamental Theorem of Arithmetic,
which states that every positive integer n, n 2, can be expressed as the
product of one or more prime numbers. Lets prove that this is true.
Recall that a number n is prime if its only positive factors are one and
n. n is composite if its not prime. Since a factor of a number must be no
larger than the number itself, this means that a composite number n always
has a factor larger than 1 but smaller than n. This, in turn, means that we
can write n as ab, where a and b are both larger than 1 but smaller than n.
2
Proof by induction on n.
2
Well leave the details of proving this as an exercise for the reader.
CHAPTER 11. INDUCTION 134
Base: 2 can be written as the product of a single prime number,
2.
Induction: Suppose that every integer between 2 and k can be
written as the product of one or more primes. We need to show
that k + 1 can be written as a product of primes. There are two
cases:
Case 1: k + 1 is prime. Then it is the product of one prime, i.e.
itself.
Case 2: k + 1 is composite. Then k + 1 can be written as ab,
where a and b are integers such that a and b lie in the range
[2, k]. By the induction hypothesis, a can be written as a product
of primes p
1
p
2
. . . p
i
and b can be written as a product of primes
q
1
q
2
. . . q
j
. So then k +1 can be written as the product of primes
p
1
p
2
. . . p
i
q
1
q
2
. . . q
j
.
In both cases k +1 can be written as a product of primes, which
is what we needed to show.
Again, the inductive step needed to reach back some number of steps in
our sequence of results, but we couldnt control how far back we needed to
go.
11.12 Variation in notation
Certain details of the induction outline vary, depending on the individual
preferences of the author and the specic claim being proved. Some folks
prefer to assume the statement is true for k and prove its true for k + 1.
Other assume its true for k 1 and prove its true for k. For a specic
problems, sometimes one or the other choice yields a slightly simpler proofs.
Folks dier as to whether the notation n = 0, 1, . . . , k implies that k is
necessarily at least 0, at least 1, or at least 2.
Some authors prefer to write strong induction hypotheses all the time,
even when a weak hypothesis would be sucient. This saves mental eort,
because you dont have to gure out in advance whether a strong hypothesis
CHAPTER 11. INDUCTION 135
was really required. However, for some problems, a strong hypothesis may
be more complicated to state than a weak one.
Authors writing for more experienced audiences may abbreviate the out-
line somewhat, e.g. packing an entirely short proof into one paragraph with-
out labelling the base and inductive steps separately. However, being careful
about the outline is important when you are still getting used to the tech-
nique.
Chapter 12
Recursive Denition
This chapter covers recursive denition, including nding closed forms.
12.1 Recursive denitions
Thus far, we have dened objects of variable length using semi-formal de-
nitions involving . . .. For example, we dened the summation

n
i=1
i by
n

i=1
i = 1 + 2 + 3 +. . . + (n 1) +n
This method is only ok when the reader can easily see what regular pattern
the . . . is trying to express. When precision is essential, e.g. when the pattern
is less obvious, we need to switch to recursive denitions.
Recursive function denitions in mathematics are basically similar to re-
cursive procedures in programming languages. A recursive denition denes
an object in terms of smaller objects of the same type. Because this pro-
cess has to end at some point, we need to include explicit denitions for the
smallest objects. So a recursive denition always has two parts:
Base case or cases
136
CHAPTER 12. RECURSIVE DEFINITION 137
Recursive formula
For example, the summation

n
i=1
i can be dened as:
g(1) = 1
g(n) = g(n 1) +n, for all n 2
Both the base case and the recursive formula must be present to have a
complete denition. However, it is traditional not to explicitly label these two
pieces. Youre just expected to gure out for yourself which parts are base
case(s) and which is the recursive formula. The input values are normally
assumed to be integers.
The true power of recursive denition is revealed when the result for
n depends on the results for more than one smaller value, as in the strong
induction examples. For example, the famous Fibonacci numbers are dened:
F
0
= 0
F
1
= 1
F
i
= F
i1
+F
i2
, i 2
So F
2
= 1, F
3
= 2, F
4
= 3, F
5
= 5, F
6
= 8, F
7
= 13, F
8
= 21, F
9
= 34. It
isnt at all obvious how to express this pattern non-recursively.
12.2 Finding closed forms
Many recursive numerical formulas have a closed form, i.e. an equivalent
expression that doesnt involve recursion (or summation or the like). Some-
times you can nd the closed form by working out the rst few values of
the function and then guessing the pattern. More often, you need to use
an organized technique. The simplest technique for nding closed forms is
called unrolling.
For example, suppose we have a function T : N Z dened by
CHAPTER 12. RECURSIVE DEFINITION 138
T(1) = 1
T(n) = 2T(n 1) + 3, n 2
The values of this function are T(1) = 1, T(2) = 5, T(3) = 13, T(4) = 29,
T(5) = 61. It isnt so obvious what the pattern is.
The idea behind unrolling is to substitute a recursive denition into itself,
so as to re-express T(n) in terms of T(n 2) rather than T(n 1). We keep
doing this, expressing T(n) in terms of the value of T for smaller and smaller
inputs, until we can see the pattern required to express T(n) in terms of n
and T(0). So, for our example function, we would compute:
T(n) = 2T(n 1) + 3
= 2(2T(n 2) + 3) + 3
= 2(2(2T(n 3) + 3) + 3) + 3
= 2
3
T(n 3) + 2
2
3 + 2 3 + 3
= 2
4
T(n 4) + 2
3
3 + 2
2
3 + 2 3 + 3
. . .
= 2
k
T(n k) + 2
k1
3 +. . . + 2
2
3 + 2 3 + 3
The rst few lines of this are mechanical substitution. To get to the last line,
you have to imagine what the pattern looks like after k substitutions.
We can use summation notation to compactly represent the result of the
kth unrolling step:
T(n) = 2
k
T(n k) + 2
k1
3 +. . . + 2
2
3 + 2 3 + 3
= 2
k
T(n k) + 3(2
k1
+. . . + 2
2
+ 2 + 1)
= 2
k
T(n k) + 3
k1

i=0
(2
i
)
CHAPTER 12. RECURSIVE DEFINITION 139
Now, we need to determine when the input to T will hit the base case. In
our example, the input value is n k and the base case is for an input of 1.
So we hit the base case when n k = 1. i.e. when k = n 1. Substituting
this value for k back into our equation, and using the fact that T(1) = 1, we
get
T(n) = 2
k
T(n k) + 3
k1

i=0
(2
i
)
= 2
n1
T(1) + 3
n2

i=0
(2
i
)
= 2
n1
+ 3
n2

k=0
(2
k
)
= 2
n1
+ 3(2
n1
1) = 4(2
n1
) 3 = 2
n+1
3
So the closed form for this function is T(n) = 2
n+1
3. The unrolling
process isnt a formal proof that our closed form is correct. However, well
see below how to write a formal proof using induction.
12.3 Divide and conquer
Many important algorithms in computer science involve dividing a big prob-
lem of (integer) size n into a sub-problems, each of size n/b. This general
method is called divide and conquer. Analyzing such algorithms involves
recursive denitions that look like:
S(1) = c
S(n) = aS(n/b) +f(n), n 2
The base case takes some constant amount of work c. The term f(n) is
the work involved in dividing up the big problem and/or merging together
CHAPTER 12. RECURSIVE DEFINITION 140
the solutions for the smaller problems. The call to the ceiling function is
required to ensure that the input to S is always an integer.
Handling such denitions in full generality is beyond the scope of this
class.
1
So lets consider a particularly important special case: dividing our
problem into two half-size problems, where the dividing/merging takes time
proportional to the size of the problem. And lets also restrict our input n
to be a power of two, so that we dont need to use the ceiling function. We
then get a recursive denition that looks like:
S(1) = c
S(n) = 2S(n/2) +n, n 2 (n a power of 2)
Unrolling this, we get
S(n) = 2S(n/2) +n
= 2(2S(n/4) +n/2) +n
= 4S(n/4) +n +n
= 8S(n/8) +n +n +n
. . .
= 2
i
S(
n
2
i
) +in
We hit the base case when
n
2
i
= 1 i.e. when i = log n (i.e. log base 2,
which is the normal convention for algorithms applications). Substituting in
this value for i and the base case value S(1) = c, we get
S(n) = 2
i
S(
n
2
i
) +in = 2
log n
c +nlog n = cn +nlog n
So the closed form for S(n) is cn +nlog n.
In real applications, our input n might not be a power of 2, so our actual
recurrence might look like:
1
See any algorithms text for more details.
CHAPTER 12. RECURSIVE DEFINITION 141
S(1) = c
S(n) = 2S(n/2) +n, n 2
We could extend the details of our analysis to handle the input values that
arent powers of 2. In many practical contexts, however, we are only inter-
ested in the overall shape of the function, e.g. is it roughly linear? cubic?
exponential? So it is often sucient to note that S is increasing, so values
of S for inputs that arent powers of 2 will lie between the values of S at the
adjacent powers of 2.
12.4 Hypercubes
Non-numerical objects can also be dened recursively. For example, the
hypercube Q
n
is the graph of the corners and edges of an n-dimensional
cube. It is dened recursively as follows (for any n N):
Q
0
is a single node with no edges
Q
n
consists of two copies of Q
n1
with edges joining corresponding
nodes, for any n 1.
That is, each node v
i
in one copy of Q
n1
is joined by an edge to its clone
copy v

i
in the second copy of Q
n1
. Q
0
, Q
1
, Q
2
, and Q
3
look as follows. The
node labels distinguish the two copies of Qn 1
A A
B
A
A
B
B A
A
A
A
B
B
B
B
The hypercube denes a binary coordinate system. To build this coordi-
nate system, we label nodes with binary numbers, where each binary digit
CHAPTER 12. RECURSIVE DEFINITION 142
corresponds to the value of one coordinate. The edges connect nodes that
dier in exactly one coordinate.
011
010
001
000
111
110
101
100
Q
n
has 2
n
nodes. To compute the number of edges, we set up the following
recursive denition for the number of edges E(n) in the Q
n
:
E(0) = 0
E(n) = 2E(n 1) + 2
n1
, for all n 1
The 2
n1
term is the number of nodes in each copy of Q
n1
, i.e. the number
of edges required to join corresponding nodes. Well leave it as an exercise
to nd a closed form for this recursive denition.
12.5 Proofs with recursive denitions
Recursive denitions are ideally suited to inductive proofs. The main out-
line of the proof often mirrors the structure of the recursive denition. For
example, lets prove the following claim about the Fibonacci numbers:
Claim 44 For any n 0, F
3n
is even.
Lets check some concrete values: F
0
= 0, F
3
= 2, F
6
= 8, F
9
= 34. All
are even. Claim looks good. So, lets build an inductive proof:
Proof: by induction on n.
Base: F
0
= 0, which is even.
CHAPTER 12. RECURSIVE DEFINITION 143
Induction: Suppose that F
3n
is even for n = 0, 1, . . . , k. We need
to show that F
3k
is even. We need to show that that F
3(k+1)
is
even.
F
3(k+1)
= F
3k+3
= F
3k+2
+F
3k+1
But F
3k+2
= F
3k+1
+F
3k
. So, substituting into the above equation,
we get:
F
3(k+1)
= (F
3k+1
+F
3k
) +F
3k+1
= 2F
3k+1
+F
3k
By the inductive hypothesis F
3k
is even. 2F
3k+1
is even because
its 2 times an integer. So their sum must be even. So F
3(k+1)
is
even, which is what we needed to show.
Some people feel a bit uncertain if the base case is a special case like zero.
Its ok to also include a second base case. For this proof, you would check the
case for n = 1 i.e. verify that F
3
is even. The extra base case isnt necessary
for a complete proof, but it doesnt cause any harm and may help the reader.
12.6 Inductive denition and strong induc-
tion
Claims involving recursive denitions often require proofs using a strong
inductive hypothesis. For example, suppose that the function f : N Z is
dened by
f(0) = 2
f(1) = 3
n 1, f(n + 1) = 3f(n) 2f(n 1)
I claim that:
Claim 45 n N, f(n) = 2
n
+ 1
We can prove this claim as follows:
CHAPTER 12. RECURSIVE DEFINITION 144
Proof: by induction on n.
Base: f(0) is dened to be 2. 2
0
+1 = 1+1 = 2. So f(n) = 2
n
+1
when n = 0.
f(1) is dened to be 3. 2
1
+ 1 = 2 + 1 = 3. So f(n) = 2
n
+ 1
when n = 1.
Induction: Suppose that f(n) = 2
n
+ 1 for n = 0, 1, . . . , k.
f(k + 1) = 3f(k) 2f(k 1)
By the induction hypothesis, f(k) = 2
k
+1 and f(k1) = 2
k1
+1.
Substituting these formulas into the previous equation, we get:
f(k+1) = 3(2
k
+1)2(2
k1
+1) = 3 2
k
+32
k
2 = 2 2
k
+1 =
2
k+1
+ 1
So f(k + 1) = 2
k+1
+ 1, which is what we needed to show.
We need to use a strong induction hypothesis, as well as two base cases,
because the inductive step uses the fact that the formula holds for two pre-
vious values of n (k and k 1).
12.7 Variation in notation
Recursive denitions are sometimes called inductive denitions or (especially
for numerical functions) recurrence relations. Folks who call them recur-
rence relations typically use the term initial condition to refer to the base
case.
Chapter 13
Trees
This chapter covers trees and induction on trees.
13.1 Why trees?
Trees are the central structure for storing and organizing data in computer
science. Examples of trees include
Trees which show the organization of real-world data: family/geneaology
trees, taxonomies (e.g. animal subspecies, species, genera, families)
Data structures for eciently storing and retrieving data. The basic
idea is the same one we saw for binary search within an array: sort the
data, so that you can repeatedly cut your search area in half.
Parse trees, which show the structure of a piece of (for example) com-
puter program, so that the compiler can correctly produce the corre-
sponding machine code.
Decision trees, which classify data by asking a series of questions. Each
tree node contains a question, whose answer directs you to one of the
nodes children.
For example, heres a parse tree for the arithmetic expression a c +b d.
145
CHAPTER 13. TREES 146
E
E
E
a
*
c +
b * d
Computer programs that try to understand natural language use parse
trees to represent the structure of sentences. For example, here are two
possible structures for the phrase green eggs and ham. In the rst, but
not the second, the ham would be green.
NP
NP
green eggs
and ham
CHAPTER 13. TREES 147
NP
NP
and ham
green eggs
Heres what a medical decision tree might look like. Decision trees are
also used for engineering classication problems, such as transcribing speech
waveforms into the basic sounds of a natural language.
Cough?
Headache? Fever?
Call doctor Rest Rest Back to work
yes
no
yes
no
yes
no
And here is a tree storing the set of number {2, 8, 10, 32, 47, 108, 200, 327, 400}
CHAPTER 13. TREES 148
32
8 108
-2 10 47 327
200 400
13.2 Dening trees
Formally, a tree is a undirected graph with a special node called the root, in
which every node is connected to the root by exactly one path. When a pair
of nodes are neighbors in the graph, the node nearest the root is called the
parent and the other node is its child. By convention, trees are drawn with
the root at the top. Two children of the same parent are known as siblings.
To keep things simple, we will assume that the set of nodes in a tree is
nite. We will also assume that each set of siblings is ordered from left to
right, because this is common in computer science applications.
A leaf node is a node that has no children. A node that does have
children is known as an internal node. The root is an internal node, except
in the special case of a tree that consists of just one node (and no edges).
The nodes of a tree can be organized into levels, based on how many
edges away from the root they are. The root is dened to be level 0. Its
children are level 1. Their children are level 2, and so forth. The height of a
tree is the maximum level of any of its nodes or, equivalently, the maximum
level of any of its leaves or, equivalently, the maximum length of a path from
the root to a leaf.
CHAPTER 13. TREES 149
If you can get from x to g by following zero or more parent links, then g is
an ancestor of x and x is a descendent of g. So x is an ancestor/descendent
of itself. The ancestors/descendents of x other than x itself are its proper
ancestors/descendents. If you pick some random node a in a tree T, the
subtree rooted at a consists of a (its root), all of as descendents, and all
the edges linking these nodes.
13.3 m-ary trees
Many applications restrict how many children each node can have. A binary
tree (very common!) allows each node to have at most two children. An m-
ary tree allows each node to have up to m children. Trees with fat nodes
with a large bound on the number of children (e.g. 16) occur in some storage
applications.
Important special cases involve trees that are nicely lled out in some
sense. In a full m-ary tree, each node has either zero or m children. Never
an intermediate number. So in a full 3-ary tree, nodes can have zero or three
children, but not one child or two children.
In a complete m-ary tree, all leaves are at the same height. Normally,
wed be interested only in full and complete m-ary trees, where this means
that the whole bottom level is fully populated with leaves.
For restricted types of trees like this, there are strong relationships be-
tween the numbers of dierent types of notes. for example:
Claim 46 A full m-ary tree with i internal nodes has mi + 1 nodes total.
To see why this is true, notice that there are two types of nodes: nodes
with a parent and nodes without a parent. A tree has exactly one node with
no parent. We can count the nodes with a parent by taking the number of
parents in the tree (i) and multiplying by the branching factor m.
Therefore, the number of leaves in a full m-ary tree with i internal nodes
is (mi + 1) i = (m1)i + 1.
CHAPTER 13. TREES 150
13.4 Height vs number of nodes
Suppose that we have a binary tree of height h. How many nodes and how
many leaves does it contain? This clearly cant be an exact formula, since
some trees are more bushy than others. But we can give useful upper and
lower bounds.
To minimize the node counts, consider a tree of height h that has just
one leaf. It contains h + 1 nodes connected into a straight line by h edges.
So the minimum number of leaves is 1 (regardless of h) and the minimum
number of nodes is h + 1.
The node counts are maximized by a tree which is full and complete. For
these trees, the number of leaves is 2
h
. More generally, the number of nodes
at level L is 2
L
. So the total number of nodes n is

h
L=0
2
L
. The closed form
for this summation is 2
h+1
1. So, for full and complete binary trees, the
height is proportional to log
2
n.
Balanced binary trees are binary trees in which all leaves are at approxi-
mately the same height. The exact denition of approximately depends on
the specic algorithms used to keep the tree balanced.
1
Balanced trees are
more exible than full and complete binary trees, but they also have height
proportional to log
2
n, where n is the number of nodes. This means that
data stored in a balanced binary tree can be accessed and modied in log
2
n
time.
13.5 Context-free grammars
Applications involving languages, both human languages and computer lan-
guages, frequently involve parse trees that show the structure of a sequence
of terminal symbols. A terminal symbol might be a word or a charac-
ter, depending on the application. The terminal symbols are stored in the
leaf nodes of the parse tree. The non-leaf nodes contains symbols that help
indicate the structure of the sentence.
For example, the following tree shows one structure for the sentence:
1
See textbooks on data structures and algorithms for further information.
CHAPTER 13. TREES 151
a c +b d. This structure assumes you intend the multiplication to be done
rst, before the addition, i.e. (a c) + (b d). The tree uses eight symbols:
E, V , a, b, c, d, +, .
E
E
E
V
*
V
+
V
*
V
a c
b d
This sort of labelled tree can be conveniently specied by a context-free
grammar. A context-free grammar is a set of rules which specify what sorts
of children are possible for a parent node with each type of label. The lefthand
side of each rule gives the symbol on the parent node and the righthand side
shows one possible pattern for the symbols on its children. If all the nodes of
a tree T have children matching the rules of some grammar G, we say that
the tree T and the terminal sequence stored in its leaves are generated
by G.
For example, the following grammar contains four rules for an E node.
The rst rule allows an E node to have three children, the leftmost one
labelled E, the middle one labelled + and the right one labelled V . According
to these rules, a node labelled V can only have one child, labelled either a,
b, c, or d. The tree shown above follows these grammar rules.
CHAPTER 13. TREES 152
E E +V
E E V
E V +V
E V V
V a
V b
V c
V d
If a grammar contains several patterns for the children of a certain node
label, we can pack them onto one line using a vertical bar to separate the
options. E.g. we can write this grammar more compactly as
E E +V | E V | V +V | V V
V a | b | c | d
To be completely precise about which trees are allowed by a grammar,
we must specify two details beyond the set of grammar rules. First, we must
give the set of terminals, i.e. symbols that are allowed to appear on the
leaf nodes. Second, we must state which symbols, called start symbols, are
allowed to appear on the root node. For example, for the above grammar,
we might stipulate that the terminals are a, b, c, d, +, and and that the
start symbols are E and V . Symbols that can appear on the lefthand side of
rules are often written in capital letters.
Consider the following grammar, with start symbol S and terminals a, b,
and c.
S aSb
S c
CHAPTER 13. TREES 153
S
c
S
a
S b
c
S
a
S b
a
S b
c
S
a
S b
a
S b
a
S b
c
The sequences of terminals for the above trees are (left to right): c, acb,
aacbb, aaacbbb. The sequences from this grammar always have a c in the
middle, with some number of as before it and the same number of bs after
it.
Notice that the left-to-right order of nodes in the tree must match the
order in the grammar rules. So, for example, the following tree doesnt match
the above grammar.
S
b S
a
c
Notice also that we didnt allow S to be a terminal. If we added S to the
set of terminals, our grammar would also allow the following trees:
CHAPTER 13. TREES 154
S S
a
S b
S
a
S b
a
S b
S
a
S b
a
S b
a
S b
Sometimes it is helpful to use relatively few symbols, so as to concentrate
on the possible shapes for the parse tree. Consider, for example, noun com-
pounds such as dump truck driver or wood salad tongs. Rather than
getting into the specics of the words involved, we might represent each noun
with the symbol N. A noun compound structure can look like any binary
tree, so it would have the following grammar, where N is both a start and a
terminal symbol
N N | NN
Here are a number of trees generated by these rules:
N N
N N
N
N N
N N N N
CHAPTER 13. TREES 155
N
N N
N N
N
N N
N N
N
N N
N N N
The bottom left tree shows the structure for dump truck driver and the
bottom middle the structure of wood salad tongs.
In some applications, its convenient to have a branch of the parse tree
which generates nothing in the terminal sequence. This is done by having a
rule whose righthand side is , the standard symbol for an empty string. For
example, here is a grossly-oversimplied grammar of an English sentence,
with start symbol S and terminals: coyotes, rabbits, carrots, eat, kill, wash,
the, all, some, and .
S NPV P
V P V NP | V NPNP
NP DetN
N coyotes | rabbits | carrots
V eat | kill | wash
Det | the | all | some
This will generate terminal sequences such as All coyotes eat some rab-
bits. Every noun phrase (NP) is required to have a determiner (Det), but
this determiner can expand into a word thats invisible in the terminal se-
quence. So we can also generate sequences like Coyotes eat carrots.
CHAPTER 13. TREES 156
S
NP VP
Det N V NP
Det N
coyotes
eat

carrots
It is possible
2
to revise this grammar so that it generates this sentence
without using . However, some theories of the meaning of natural language
sentences make a distinction between items which are completely missing
and items which are present in the underlying structure but missing in the
supercial structure. In this case, one might claim that the above sentence
has a determiner which is understood to be all.
13.6 Recursion trees
One nice application for trees is visualizing the behavior of certain recur-
sive denitions, especially those used to describe algorithm behavior. For
example, consider the following denition, where c is some constant.
S(1) = c
S(n) = 2S(n/2) +n, n 2 (n a power of 2)
We can draw a picture of this denition using a recursion tree. The top
node in the tree represents S(n) and contains everything in the formula for
S(n) except the recursive calls to S. The two nodes below it represent
two copies of the computation of S(n/2). Again, the value in each node
2
Except when the grammar can generate a terminal sequence that contains only invis-
ible symbols.
CHAPTER 13. TREES 157
contains the non-recursive part of the formula for computing S(n/2). The
value of S(n) is then the sum of the values in all the nodes in the recursion
tree.
n
n/2 n/2
n/4 n/4 n/4 n/4
. . . . . . . . . . . . . . . . . . . . . . . .
To sum everything in the tree, we need to ask:
How high is this tree, i.e. how many levels do we need to expand before
we hit the base case n = 1?
For each level of the tree, what is the sum of the values in all nodes at
that level?
How many leaf nodes are there?
In this example, the tree has height log n, i.e. there are log n non-leaf
levels. At each level of the tree, the node values sum to n. So the sum for all
non-leaf nodes is nlog n. There are n leaf nodes, each of which contributes c
to the sum. So the sum of everything in the tree is nlog n +cn, which is the
same closed form we found for this recursive denition in Section 12.3.
Recursion trees are particularly handy when you only need an approx-
imate description of the closed form, e.g. is its leading term quadratic or
cubic?
CHAPTER 13. TREES 158
13.7 Another recursion tree example
Now, lets suppose that we have the following denition, where c is some
constant.
P(1) = c
P(n) = 2P(n/2) +n
2
, n 2 (n a power of 2)
Its recursion tree is
n
2
(n/2)
2
(n/2)
2
(n/4)
2
(n/4)
2
(n/4)
2
(n/4)
2
. . . . . . . . . . . . . . . . . . . . . . . .
The height of the tree is again log n. The sums of all nodes at the top
level is n
2
. The next level down sums to n
2
/2. And then we have sums:
n
2
/4, n
2
/8, n
2
/16, and so forth. So the sum of all nodes at level k is n
2 1
2
k
.
The lowest non-leaf nodes are at level log n 1. So the sum of all the
non-leaf nodes in the tree is
P(n) =
log n1

k=0
n
2
1
2
k
= n
2
log n1

k=0
1
2
k
= n
2
(2
1
2
log n1
) = n
2
(2
2
2
log n
) = n
2
(2
2
n
) = 2n
2
2n
Adding cn to cover the leaf nodes, our nal closed form is 2n
2
+(c 2)n.
CHAPTER 13. TREES 159
13.8 Tree induction
When doing induction on trees, we divide the tree up at the top. That is, we
view a tree as consisting of a root node plus some number of subtrees rooted
at its children. The induction variable is typically the height of the tree.
The child subtrees have height less than that of the parent tree. So, in the
inductive step, well be assuming that the claim is true for these shorter child
subtrees and showing that its true for the taller tree that contains them.
For example, we claimed above that
Claim 47 Let T be a binary tree, with height h and n nodes. Then n
2
h+1
1.
Proof by induction on h, where h is the height of the tree.
Base: The base case is a tree consisting of a single node with
no edges. It has h = 0 and n = 1. Then we work out that
2
h+1
1 = 2
1
1 = 1 = n.
Induction: Suppose that the claim is true for all binary trees of
height < h. Let T be a binary tree of height h (h > 0).
Case 1: T consists of a root plus one subtree X. X has height
h 1. So X contains at most 2
h
1 nodes. T only contains one
more node (its root), so this means T contains at most 2
h
nodes,
which is less than 2
h+1
1.
r
X
Case 2: T consists of a root plus two subtrees X and Y . X and
Y have heights p and q, both of which have to be less than h,
i.e. h 1. X contains at most 2
p+1
1 nodes and Y contains
at most 2
q+1
1 nodes, by the inductive hypothesis. But, since
p and q are less than h, this means that X and Y each contain
2
h
1 nodes.
CHAPTER 13. TREES 160
r
X Y
So the total number of nodes in T is the number of nodes in X
plus the number of nodes in Y plus one (the new root node). This
is 1+(2
p
1)+(2
q
1) 1+2(2
h
1) = 1+2
h+1
2 = 2
h+1
1
So the total number of nodes in T is 2
h+1
1, which is what
we needed to show.
In writing such a proof, its tempting to think that if the full tree has
height h, the child subtrees must have height h 1. This is only true if
the tree is complete. For a tree that isnt necessarily complete, one of the
subtrees must have height h 1 but the other subtree(s) might be shorter
than h 1. So, for induction on trees, it is almost always necessary to use a
strong inductive hypothesis.
In the inductive step, notice that we split up the big tree (T) at its
root, producing two smaller subtrees (X) and (Y ). Some students try to do
induction on trees by grafting stu onto the bottom of the tree. Do not
do this. There are many claims, especially in later classes, for which this
grafting approach will not work and it is essential to divide at the root.
13.9 Heap example
In practical applications, the nodes of a tree are often used to store data.
Algorithm designers often need to prove claims about how this data is ar-
ranged in the tree. For example, suppose we store numbers in the nodes of a
full binary tree. The numbers obey the heap property if, for every node X in
the tree, the value in X is at least as big as the value in each of Xs children.
For example:
CHAPTER 13. TREES 161
32
19 12
18 8 9 1
Notice that the values at one level arent uniformly bigger than the values
at the next lower level. For example, 18 in the bottom level is larger than
12 on the middle level. But values never decrease as you move along a path
from a leaf up to the root.
Trees with the heap property are convenient for applications where you
have to maintain a list of people or tasks with associated priorities. Its easy
to retrieve the person or task with top priority: it lives in the root. And its
easy to restore the heap property if you add or remove a person or task.
I claim that:
Claim 48 If a tree has the heap property, then the value in the root of the
tree is at least as large as the value in any node of the tree.
To keep the proof simple, lets restrict our attention to full binary trees:
Claim 49 If a full binary tree has the heap property, then the value in the
root of the tree is at least as large as the value in any node of the tree.
Lets let v(a) be the value at node a and lets use the recursive structure
of trees to do our proof.
Proof by induction on the tree height h.
CHAPTER 13. TREES 162
Base: h = 0. A tree of height zero contains only one node, so
obviously the largest value in the tree lives in the root!
Induction: Suppose that the claim is true for all full binary trees
of height < h. Let T be a tree of height h (h > 0) which has the
heap property. Since T is a full binary tree, its root r has two
children p and q. Suppose that X is the subtree rooted at p and
Y is the subtree rooted at q.
Both X and Y have height < h. Moreover, notice that X and Y
must have the heap property, because they are subtrees of T and
the heap property is a purely local constraint on node values.
Suppose that x is any node of T. We need to show that v(r)
v(x). There are three cases:
Case 1: x = r. This is obvious.
Case 2: x is any node in the subtree X. Since X has the heap
property and height h, v(p) v(x) by the inductive hypothesis.
But we know that v(r) v(p) because T has the heap property.
So v(r) v(x).
Case 3: x is any node in the subtree Y . Similar to case 2.
So, for any node x in T, v(r) v(x).
13.10 Proof using grammar trees
Consider the following grammar G, with start symbol S and terminals a and
b. I claim that all trees generated by G have the same number of nodes with
label a as with label b.
S ab
S SS
S aSb
We can prove this by induction as follows:
CHAPTER 13. TREES 163
Proof by induction on the tree height h.
Base: Notice that trees from this grammar always have height at
least 1. The only way to produce a tree of height 1 is from the
rst rule, which generates exactly one a node and one b node.
Induction: Suppose that the claim is true for all trees of height
< k, where k 1. Consider a tree T of height k. The root must
be labelled S and the grammar rules give us three possibilities
for what the roots children look like:
Case 1: The roots children are labelled a and b. This is just the
base case.
Case 2: The roots children are both labelled S. The subtrees
rooted at these children have height < k. So, by the inductive
hypothesis, each subtree has equal numbers of a and b nodes. Say
that the left tree has m of each type and the right tree has n of
each type. Then the whole tree has m+n nodes of each type.
Case 3: The root has three children, labelled a, S, and b. Since
the subtree rooted at the middle child has height < k, it has equal
numbers of a and b nodes by the inductive hypothesis. Suppose
it has m nodes of each type. Then the whole tree has m+1 nodes
of each type.
In all three cases, the whole tree T has equal numbers of a and b
nodes.
13.11 Variation in terminology
There are actually two sorts of trees in the mathematical world. This chapter
describes the kind of trees most commonly used in computer science, which
are formally rooted trees with a left-to-right order. In graph theory, the
term tree typically refers to free trees, which are connected acyclic graphs
with no distinguished root node and no clear up/down or left-right directions.
We will see free trees later, when analyzing planar graphs. Variations on these
two types of trees also occur. For example, some data structures applications
use trees that are almost like ours, except that a single child node must be
designated as a left or right child.
CHAPTER 13. TREES 164
Innite trees occur occasionally in computer science applications. They
are an obvious generalization of the nite trees discussed here.
Tree terminology varies a lot, apparently because trees are used for a wide
range of applications with very diverse needs. In particular, some authors use
the term complete binary tree to refer to a tree in which only the leftmost
part of the bottom level is lled in. The term perfect is then used for a
tree whose bottom level is entirely lled. Some authors dont allow a node
to be an ancestor or descendent of itself. A few authors dont consider the
root node to be a leaf.
Many supercial details of context-free grammar notation change with
the application. For example, the BNF notation used in programming lan-
guage specications looks dierent at rst glance because, for example, the
grammar symbols are whole words surrounded by angle brackets.
Conventional presentations of context-free grammars are restricted so
that terminals cant occur on the lefthand side of rules and there is a single
start symbol. The broader denitions here allow us to capture the full range
of examples oating around practical applications, without changing what
sets of terminal sequences can be generated by these grammars.
Chapter 14
Big-O
This chapter covers asymptotic analysis of function growth and big-O nota-
tion.
14.1 Running times of programs
An important aspect of designing a computer programs is guring out how
well it runs, in a range of likely situations. Designers need to estimate how
fast it will run, how much memory it will require, how reliable it will be, and
so forth. In this class, well concentrate on speed issues.
Designers for certain small platforms sometimes develop very detailed
models of running time, because this is critical for making complex appli-
cations work with limited resources. E.g. making God of War run on your
Iphone. However, such programming design is increasingly rare, because
computers are getting fast enough to run most programs without hand opti-
mization.
More typically, the designer has to analyze the behavior of a large C or
Java program. Its not feasible to gure out exactly how long such a pro-
gram will take. The transformation from standard programming languages
to machine code is way too complicated. Only rare programmers have a
clear grasp of what happens within the C or Java compiler. Moreover, a very
detailed analysis for one computer system wont translate to another pro-
165
CHAPTER 14. BIG-O 166
gramming language, another hardware platform, or a computer purchased
a couple years in the future. Its more useful to develop an analysis that
abstracts away from unimportant details, so that it will be portable and
durable.
This abstraction process has two key components:
Ignore multiplicative constants
Ignore behavior on small inputs, concentrating on how well programs
handle large inputs. (Aka asymptotic analysis.)
Multiplicative constants are extremely sensitive to details of the imple-
mentation, hardware platform, etc.
Behavior on small inputs is ignored, because programs typically run fast
enough on small test cases. Or will do so soon, as computers become faster
and faster. Hard-to-address problems more often arise when a programs
use expands to larger examples. For example, a small database program
developed for a community college might have trouble coping if deployed to
handle (say) all registration records for U. Illinois.
14.2 Function growth: the ideas
So, suppose that you model the running time of a program as a function
F(n), where n is some measure of the size of the input problem. E.g. n
might be the number of entries in a database application. For a numerical
program, n might be the magnitude or the number of digits in an input
number. Then, to compare the running times of two programs, we need to
compare the growth rates of the two running time functions.
So, suppose we have two functions f and g, whose inputs and outputs
are real numbers. Which one has bigger outputs?
Suppose that f(x) = x and g(x) = x
2
. For small positive inputs, x
2
is
smaller. For the input 1, they have the same value, and then g gets bigger
and rapidly diverges to become much larger than f. Wed like to say that g
is bigger, because it has bigger outputs for large inputs.
CHAPTER 14. BIG-O 167
Because we are only interested in the running times of algorithms, well
only consider behavior on positive inputs. And well only worry about func-
tions whose output values are positive, or whose output values become pos-
itive as the input value gets big, e.g. the log function.
Because we dont care about constant multipliers, well consider functions
such as 3x
2
, 47x
2
, and 0.03x
2
to all grow at the same rate. Similarly, functions
such as 3x, 47x, and 0.03x will be treated as growing at the same, slower,
rate. The functions in each group dont all have the same slope, but their
graphs have the same shape as they head o towards innity. Thats the
right level of approximation for analyzing most computer programs.
Finally, when a function is the sum of faster and slower-growing terms,
well only be interested in the faster-growing term. For example, 0.3x
2
+
7x + 105 will be treated as equivalent to x
2
. As the input x gets large, the
behavior of the function is dominated by the term with the fastest growth
(the rst term in this case).
14.3 Primitive functions
Lets look at some basic functions and try to put them into growth order.
Were looking for relationships of the form f(x) g(x). But since we only
care about behavior on large inputs, these relationships will only work for
inputs x starting at some lower bound k.
It should be obvious that any constant function grows more slowly than
a linear function (i.e. because a constant function doesnt grow!). For large
enough numbers, a linear polynomial grows more slowly than a quadratic
and a third-order polynomial grows faster than a quadratic.
Youre probably familiar with how fast exponentials grow. Theres a
famous story about a judge imposing a doubling-ne on a borough of New
York, for ignoring the judges orders. It took the borough ocials a few days
to realize that this was serious bad news, at which point a settlement was
reached. Specically, n
2
< 2
n
for any integer n 4. And, in general, for
any exponent k, you can show that n
k
< 2
n
for any n above some suitable
lower bound. Well prove these results only for integer inputs, but they can
be generalized to real inputs intermediate between these integer values.
CHAPTER 14. BIG-O 168
Less obviously, 2
n
n! for every integer n 4. This is true because 2
n
and n! are each the product of n terms. For 2
n
, they are all 2. For n! they
are the rst n integers, and all but the rst two of these are bigger than 2.
So, 2
n
grows faster than any polynomial in n, and n! grows yet faster. If
we use 1 as our sample constant function, we can summarize these facts as:
1 n n
2
n
3
. . . 2
n
n!
Ive used curly because this ordering isnt standard algebraic . The
ordering only works when n is large enough.
For the purpose of designing computer programs, only the rst three of
these running times are actually good news. Third-order polynomials already
grow too fast for most applications, if you expect inputs of non-trivial size.
Exponential algorithms are only worth running on extremely tiny inputs, and
are frequently replaced by faster algorithms (e.g. using statistical sampling)
that return approximate results.
Now, lets look at slow-growing functions, i.e. functions that might be
the running times of ecient programs. Well see that algorithms for nding
entries in large datasets often have running times proportional to log n. If
you draw the log function and ignore its strange values for inputs smaller
than 1, youll see that it grows, but much more slowly than n.
Algorithms for sorting a list of numbers have running times that grow
like nlog n. If n is large enough, 1 < log n < n. So n < nlog n < n
2
. We can
summarize these relationships as:
1 log n n nlog n n
2
Its well worth memorizing the relative orderings of these basic functions,
since youll see them again and again in this and future CS classes.
CHAPTER 14. BIG-O 169
14.4 Proving a primitive function relation-
ship
Lets see how to pin down the formal details of one ordering between primitive
functions, using induction. I claimed above that:
Claim 50 For every positive integer n 4, 2
n
< n!.
In this claim, the proposition P(n) is 2
n
< n!. So an outline of our
inductive proof looks like:
Proof: Suppose that n is an integer and n 4. Well prove that
2
n
< n! using induction on n.
Base: n = 4. [show that the formula works for n = 4]
Induction: Suppose that 2
n
< n! holds for n = 4, 5, . . . , k. That
is, suppose that we have an integer k 4 such that 2
k
< k!.
We need to show that the claim holds for n = k + 1, i.e. that
2
k+1
< (k + 1)!.
When working with inequalities, the required algebra is often far from
obvious. So, its especially important to write down your inductive hypothesis
and the conclusion of your inductive step. You can then work from both ends
to ll in the gap in the middle of the proof.
Proof: Suppose that n is an integer and n 4. Well prove that
2
n
< n! using induction on n.
Base: n = 4. In this case 2
n
= 2
4
= 16. Also n! = 1 2 3 4 = 24.
Since 16 < 24, the formula holds for n = 4.
Induction: Suppose that 2
n
< n! holds for n = 4, 5, . . . , k. That
is, suppose that we have an integer k 4 such that 2
k
< k!. We
need to show that 2
k+1
< (k + 1)!.
2
k+1
= 22
k
. By the inductive hypothesis, 2
k
< k!, so 22
k
< 2k!.
Since k 4, 2 < k + 1. So 2 k! < (k + 1) k! = (k + 1)!.
Putting these equations together, we nd that 2
k+1
< (k + 1)!,
which is what we needed to show.
CHAPTER 14. BIG-O 170
Induction can be used to establish similar relationships between other
pairs of primitive functions, e.g. n
2
and 2
n
.
14.5 The formal denition
Our formal denition needs to include provisions for both lower bounds and
possible constant multipliers. Suppose that f and g are functions whose
domain and co-domain are subsets of the real numbers. Then f(x) is O(g(x))
(read big-O of g) if and only if
There are positive real numbers c and k such that 0 f(x)
cg(x) for every x k.
The constant c in the equation models the fact that we dont care about
multiplicative constants in comparing functions. The restriction that the
equation only holds for x k models the fact that we dont care about the
behavior of the functions on small input values.
So, for example, 3x
2
is O(2
x
). But 3x
2
is not O(x). The big-O relationship
also holds when the two functions grow at the same rate, e.g. 3x
2
is O(x
2
).
So the big-O relationship is like on real numbers, not like <.
When g(x) is O(f(x)) and f(x) is O(g(x)), then f(x) and g(x) must grow
at the same rate. In this case, we say that f(x) is (g(x)) (and also g(x) is
(f(x))).
Big-O is a partial order on the set of all functions from the reals to
the reals. The relationship is an equivalence relation on this same set of
functions. So, for example, under the relation, the equivalence class [x
2
]
contains functions such as x
2
, 57x
2
301, 2x
2
+x + 2, and so forth.
14.6 Applying the denition
To show that a big-O relationship holds, we need to produce suitable values
for c and k. For any particular big-O relationship, there are a wide range
CHAPTER 14. BIG-O 171
of possible choices. First, how you pick the multiplier c aects where the
functions will cross each other and, therefore, what your lower bound k can
be. Second, there is no need to minimize c and k. Since you are just demon-
strating existence of suitable c and k, its entirely appropriate to use overkill
values.
For example, to show that 3x is O(x
2
), we can pick c = 3 and k = 1.
Then 3x cx
2
for every x k translates into 3x 3x
2
for every x 1,
which is clearly true. But we could have also picked c = 100 and k = 100.
Overkill seems less elegant, but its easier to conrm that your chosen
values work properly, especially in situations like exams. Moreover, slightly
overlarge values are often more convincing to the reader, because the reader
can more easily see that they do work.
To take a more complex example, lets show that 3x
2
+ 7x + 2 is O(x
2
).
If we pick c = 3, then our equation would look like 3x
2
+7x +2 3x
2
. This
clearly wont work for large x.
So lets try c = 4. Then we need to nd a lower bound on x that makes
3x
2
+7x +2 4x
2
true. To do this, we need to force 7x +2 x
2
. This will
be true if x is big, e.g. 100. So we can choose k = 100.
To satisfy our formal denition, we also need to make sure that both
functions produce non-negative values for all inputs k. If this isnt already
the case, increase k.
14.7 Writing a big-O proof
In a formal big-O proof, you rst choose values for k and c, then show that
0 f(x) cg(x) for every x k. So the example from the previous section
would look like:
Claim 51 3x
2
+ 7x + 2 is O(x
2
).
Proof: Consider c = 4 and k = 100. Then for any x k, x
2

100x 7x+2. Since x is positive, we also have 0 3x

2
+7x+2.
Therefore, for any x k, 0 3x
2
+7x+2 3x
2
+x
2
= 4x
2
= cx
2
.
So 3x
2
+ 7x + 2 is O(x
2
).
CHAPTER 14. BIG-O 172
Notice that the steps of this proof are in the opposite order from the work
we used to nd values for c and k. This is standard for big-O proofs. Count
on writing such proofs in two drafts: the rst in whatever order, the second
in logical order.
Heres another example of a big-O proof:
Claim 52 Show that 3x
2
+ 8xlog x is O(x
2
).
[On our scratch paper] xlog x x
2
for any x 1. So 3x
2
+ 8xlog x
11x
2
. So if we set c = 11 and k = 1, our denition of big-O is satised.
Writing this out neatly, we get:
Proof: Consider c = 11 and k = 1. Suppose that x k. Then
x 1. So 0 log x x. Since x is positive, this implies that
0 xlog x x
2
. So then 0 3x
2
+ 8xlog x 11x
2
= cx, which
is what we needed to show.
14.8 Sample disproof
Suppose that we want to show that a big-O relationship does not hold. If we
negate the denition of big-O, we nd that f(x) is not O(g(x)) if and only if
For every positive real numbers c and k, there is an x k such
that f(x) is negative or f(x) > cg(x).
For most non-contrived examples in computer science, f(x) will only be
negative for very small values of x. So we would normally be trying to prove
that:
For every positive real numbers c and k, there is an x k such
that f(x) > cg(x).
So we are trying to prove a universal claim and, therefore, we
should pick two random values for c and k and then try to show
the existence of an x with the required properties.
CHAPTER 14. BIG-O 173
For example:
Claim 53 x
3
is not O(7x
2
).
In our proof, well be trying to nd an x such that show that x
3
> c7x
3
.
Working backwards this will be true whenever x > 7c. So our proof might
look like:
Proof: Using the denition of big-O, our claim can be rephrased
as: for every positive real numbers c and k, there is an x k
such that x
3
> cg(x).
Let c and k be positive integers. There are innitely many num-
bers max(k, 7c). So choose x to be any one of them. Since
x > 7c, x
3
> c7x
3
. And clearly x k.
Since c and k were arbitrarily chosen, we have proven our claim.
14.9 Variation in notation
In the denition of big-o, some authors replace 0 f(x) cg(x) with
|f(x)| c|g(x)|. The absolute values and the possibility of negative values
makes this version harder to work with. Some authors state the denition
only for functions f and g with positive output values. This is awkward
because the logarithm function produces negative output values for very small
inputs.
Outside theory classes, computer scientists often say that f(x) is O(g(x))
when they actually mean the (stronger) statement that f(x) is (g(x)). Or
this drives theoreticians nutsthey will say that g(x) is a tight big-O
bound on f(x). In this class, well stick to the proper theory notation, so
that you can learn how to use it. That is, you should use when you mean
to say that two functions grow at the same rate or when you mean to give a
tight bound.
Very, very annoyingly, for historical reasons, the statement f(x) is O(g(x))
is often written as f(x) = O(g(x)). This looks like a sort of equality, but it
isnt. It is actually expressing an inequality. This is badly-designed notation
but, sadly, common.
Chapter 15
Algorithms
This chapter covers how to analyze the running time of algorithms.
15.1 Introduction
The techniques weve developed earlier in this course can be applied to ana-
lyze how much time a computer algorithm requires, as a function of the size
of its input(s). We will see a range of simple algorithms illustrating a variety
of running times. Three methods will be used to analyze the running times:
nested loops, resource consumption, and recursive denitions.
We will gure out only the big-O running time for each algorithm, i.e.
ignoring multiplicative constants and behavior on small inputs. This will
allow us to examine the overall design of the algorithms without excessive
complexity. Being able to cut corners so as to get a quick overview is a
critical skill when you encounter more complex algorithms in later computer
science classes.
15.2 Basic data structures
Many simple algorithms need to store a sequence of objects a
1
, . . . , a
n
. In
pseudocode, the starting index for a sequence is sometimes 0 and sometimes
174
CHAPTER 15. ALGORITHMS 175
1, and you often need to examine the last subscript to nd out the length.
Sequences can be stored using either an array or a linked list. The choice
sometimes aects the algorithm analysis, because these two implementation
methods have slightly dierent features.
An array provides constant-time access to any element. So you can
quickly access elements in any order you choose and the access time does
not depend on the length of the list. However, the length of an array is xed
when the array is built. Changing the array length takes time proportional
to the length of the array, i.e. O(n). Adding or deleting objects in the middle
of the array requires pushing other objects sideways. This can also take O(n)
time. Two-dimensional arrays are similar, except that you need to supply
two subscripts e.g. a
x,y
.
In a linked list, each object points to the next object in the list. An
algorithm has direct access only to the elements at the ends of the list.
1
Objects in the middle of the list can only be accessed by walking element-
by-element from one end, which can take O(n) time. However, the length
of the list is exible and objects can added to, or removed from, the ends of
the list in constant time. Once you are at a position in the middle of a list,
objects can be added or deleted at that position in constant time.
A linked list starts with its head and ends with its tail. For example,
suppose our list is L = (1, 7, 3, 4, 7, 19). Then head(L) is 1 and tail(L) is 19.
The function pop removes and returns the value at the head of a list. I.e.
pop(L) will return 1 leave the list L containing (7, 3, 4, 7, 19).
For some algorithms, the big-O performance does not depend on whether
arrays or linked lists are used. This happens when the number of objects is
xed and the objects are accessed in sequential order. However, remember
that a big-O analysis ignores multiplicative constants. All other things be-
ing equal, array-based implementations tend to have smaller constants and
therefore run faster.
CHAPTER 15. ALGORITHMS 176
01 closestpair(p
1
, . . . , p
n
) : array of 2D points)
02 best1 = p
1
03 best2 = p
2
04 bestdist = dist(p
1
,p
2
)
05 for i = 1 to n
06 for j = 1 to n
07 newdist = dist(p
i
,p
j
)
08 if (i = j and newdist < bestdist)
09 best1 = p
i
10 best2 = p
j
11 bestdist = newdist
12 return (best1, best2)
Figure 15.1: Finding the closest pair using nested loops.
15.3 Nested loops
Algorithms based on nested loops are the easiest to analyze. Suppose that
we have a set of 2D points and we would like to nd the pair that are closest
together. Our code might look as in Figure 15.1, assuming that the function
dist computes the distance between two 2D points.
To analyze this code in big-O terms, rst notice that the start-up code
in lines 1-4 and the ending code in line 12 takes the same amount of time
regardless of the input size n. So well say that it takes constant time or
O(1) time. The block of code inside both loops (lines 7-11) takes a constant
time to execute once. So the big-O running time of this algorithm is entirely
determined by how many times the loops run.
The outer loop runs n times. The inner loop runs n times during each
iteration of the outer loop. So the block of code inside both loops executes
O(n
2
) times.
This code examines each pair of 2D points twice, once in each order.
We could avoid this extra work by having the inner loop (j) run only from
i + 1 to n. In this case, the code inside both loops will execute

n
i=1
(n i)
1
Strictly speaking, this works only for certain types of linked lists. See a data structures
text for all the gory details.
CHAPTER 15. ALGORITHMS 177
times. This is equal to

n1
i=0
i =
n(n1)
2
. This is still O(n
2
): our optimization
improved the constants but not the big-O running time. Improving the big-
O running timeO(nlog n) is possiblerequires restructuring the algorithm
and involves some geometrical detail beyond the scope of this chapter.
15.4 Merging two lists
01 merge(L
1
,L
2
: sorted lists of real numbers)
02 O = emptylist
03 while (L
1
is not empty or L
2
is not empty)
04 if (L
1
is empty)
05 move head(L
2
) to the tail of O
06 else if (L
2
is empty)
07 move head(L
1
) to the tail of O
08 else if (head(L
1
) <= head(L
2
))
09 move head(L
1
) to the tail of O
10 else move head(L
2
) to the tail of O
11 return O
Figure 15.2: Merging two lists
When code contains a while loop, rather than a for loop, it can be less
obvious how many times the loop will run. For example, suppose we have two
sorted lists, a
1
, . . . , a
p
and b
1
, . . . , b
q
. We can merge them very eciently
into a combined sorted list. To do this, we make a new third empty list
to contain our merged output. Then we examine the rst elements of the
two input lists and move the smaller value onto our output list. We keep
looking at the rst elements of both lists until one list is empty. We then
copy over the rest of the non-empty list. Figure 15.2 shows pseudocode for
this algorithm.
For merge, a good measure of the size of the input is the length of the
output array n, which is equal to the sum of the lengths of the two input
arrays (p + q). The merge function has one big while loop. Since the op-
erations within the loop (lines 4-10) all take constant time, we just need to
gure out how many times the loop runs, as a function of n.
CHAPTER 15. ALGORITHMS 178
A good way to analyze while loops is to track a resource that has a known
size and is consumed as the loop runs. In this case, each time the loop runs,
we move one number from L
1
or L
2
onto the output list O. We have n
numbers to move, after while the loop halts. So the loop must run n times.
So merge takes O(n) (aka linear) time. This is called a resource consumption
analysis.
15.5 A reachability algorithm
The function in Figure 15.3 determines whether one node t in a graph is
reachable from another node s, i.e. whether there is a path connecting s and
t. To do this, we start from s and explore outwards by following the edges
of the graph. When we visit a node, we place a marker on it so that we
can avoid examining it again (and thereby getting the code into an innite
loop). The temporary list M contains nodes that we have reached, but whose
neighbors have not yet been explored.
2
01 reachable(G: a graph; s,t: nodes in G)
02 if s = t return true
03 Unmark all nodes of G.
04 M = emptylist
05 Mark node s and add it to M
06 while (M is not empty)
07 p = pop(M)
08 for every node q that is a neighbor of p in G
09 if q = t return true
10 else if q is not marked, mark q and add it to M
11 return false
Figure 15.3: Can node s be reached from node t?
We can use a resource consumption argument to analyze the while loop
in this code. Suppose that G has n nodes and m edges. No node is put onto
2
Line 5 is deliberately vague about which end of M the nodes are added to and,
therefore, vague about the order in which nodes are explored.
CHAPTER 15. ALGORITHMS 179
the list M more than once. So the loop in lines 6-10 runs no more than n
times.
Line 8 starts at a node q and nd all its neighbors p.
3
So it traces all
the edges involving q. During the whole run of the code, a graph edge might
get traced twice, once in each direction. There are m edges in the graph. So
lines 9-10 cannot run more than 2m times.
In total, this algorithm needs O(n +m) time. This is an interesting case
because neither of the two terms n or m dominates the other. It is true
that the number of edges m is no O(n
2
) and thus the connected component
algorithm is O(n
2
). However, in most applications, relatively few of these
potential edges are actually present. So the O(n+m) bound is more helpful.
Notice that there is a wide variety of graphs with n nodes and m edges.
Our analysis was based on the kind of graph that wold causes the algorithm
to run for the longest time, i.e. a graph in which the algorithm reaches every
node and traverses every edge, reaching t last. This is called a worst-case
analysis. On some input graphs, our code might run much more quickly, e.g.
if we encounter t early in the search or if much of the graph is not connected to
s. Unless the author explicitly indicates otherwise, big-O algorithm analyses
are normally understood to be worst-case.
15.6 Binary search
We will now look at a strategy for algorithm design called divide and con-
quer, in which a larger problem is solved by dividing it into several (usually
two) smaller problems. For example, the code in Figure 15.4 uses a technique
called binary search to calculate the square root of its input. For simplicity,
well assume that the input n is quite large, so that we only need the answer
to the nearest integer.
This code operates by dening a range of integers in which the answer
must live, initially between 1 and n. Each recursive call tests whether the
midpoint of the range is higher or lower than the desired answer, and selects
the appropriate half of the range for further exploration. The helper function
3
We assume that the internal computer representation of the graph stores a list of
neighbors for each node.
CHAPTER 15. ALGORITHMS 180
01 squareroot(n: positive integer)
02 p = squarerootrec(n, 1, n)
03 if (n p
2
(p + 1)
2
n)
05 return p
05 else return p + 1
11 squarerootrec(n, bottom, top: positive integers)
12 if (bottom = top) return bottom
13 middle = oor(
bottom+top
2
)
14 if (middle
2
== n)
15 return middle
16 else if (middle
2
n)
17 return squarerootrec(n,middle,top)
18 else
19 return squarerootrec(n,bottom,middle)
Figure 15.4: Binary search for

n
squarerootrec returns when it has found
_
(n). The function then checks
whether this value, or the next higher integer, is the better approximation.
This isnt the fastest way to nd a square root,
4
but this simple method
generalizes well to situations in which we have only a weak model of the
function we are trying to optimize. This method requires only that you can
test whether a candidate value is too low or too high. Suppose, for example,
that you are tuning a guitar string. Many amateur players can tell if the
current tuning is too low or too high, but have a poor model of how far to
turn the tuning knob. Binary search would be a good strategy for optimizing
the tuning.
To analyze how long binary search takes, rst notice that the start-up and
clean-up work in the main squareroot function takes only constant time. So
we can basically ignore its contribution. The function squarerootrec makes
one recursive call to itself and otherwise does a constant amount of work.
The base case requires only a constant amount of work. So if the running
4
A standard faster approach is Newtons method.
CHAPTER 15. ALGORITHMS 181
time of squarerootrec is T(n), we can write the following recursive denition
for T(n), where c and d are constants.
T(1) = c
T(n) = T(n/2) +d
If we unroll this denition k times, we get T(n) = T(
n
2
k
) +kd. Assuming
that n is a power of 2, we hit the base case when k = log n. So T(n) =
c +d log n, which is O(log n).
15.7 Mergesort
Mergesort takes an input linked list of numbers and returns a new linked list
containing the sorted numbers. This is somewhat dierent from bubble and
insertion sort, which rearrange the values within a single array (and dont
return anything).
Mergesort divides its big input list (length n) into two smaller lists of
length n/2. Lists are divided up repeatedly until we have a large number of
very short lists, of length 1 or 2 (depending on the preferences of the code
writer). A length-1 list is necessarily sorted. A length 2 list can be sorted
in constant time. Then, we take all these small sorted lists and merge them
together in pairs, gradually building up longer and longer sorted lists until we
have one sorted list containing all of our original input numbers. Figure 15.5
shows the resulting pseudocode.
01 mergesort(L = a
1
, a
2
, . . . , a
n
: list of real numbers)
02 if (n = 1) then return L
03 else
04 m = n/2
05 L
1
= (a
1
, a
2
, . . . , a
m
)
06 L
2
= (a
m+1
, a
m+2
, . . . , a
n
)
07 return merge(mergesort(L
1
),mergesort(L
2
))
Figure 15.5: Sorting a list using mergesort
CHAPTER 15. ALGORITHMS 182
On an input of length n, mergesort makes two recursive calls to itself. It
also does O(n) work dividing the list in half, because it must walk, element
by element, from the head of the list down to the middle position. And it
does O(n) work merging the two results. So if the running time of mergesort
is T(n), we can write the following recursive denition for T(n), where c and
d are constants.
T(1) = c
T(n) = 2T(n/2) +dn
This recursive denition has the following recursion tree:
dn
dn/2 dn/2
dn/4 dn/4 dn/4 dn/4
. . . . . . . . . . . . . . . . . . . . . . . .
The tree has O(log n) non-leaf levels and the work at each level sums
up to dn. So the work from the non-leaf nodes sums up to O(nlog n). In
addition, there are n leaf nodes (aka base cases for the recursive function),
each of which involves c work. So the total running time is O(nlog n) + cn
which is just O(nlog n).
15.8 Tower of Hanoi
The Tower of Hanoi puzzle was invented by the French mathematician Edouard
Lucas in 1883. It consists of three pegs and a set of k disks of graduated size
that t on them. The disks start out in order on one peg. You are allowed to
CHAPTER 15. ALGORITHMS 183
move only a single disk at a time. The goal is to rebuild the ordered tower
on another peg without ever placing a disk on top of a smaller disk.
The best way to understand the solution is recursively. Suppose that
we know how to move k disks from one peg to another peg, using a third
temporary-storage peg. To move k + 1 disks from peg A to peg B using a
third peg C, we rst move the top k disks from A to C using B as temporary
storage. Then we move the biggest disk from A to B. Then we move the
other k disks from C to B, using A as temporary storage. So our recursive
solver would have pseudocode as in Figure 15.6.
01 hanoi(A,B,C: pegs, d
1
, d
2
. . . d
n
: disks)
02 if (n = 1) move d
1
= d
n
from A to B.
03 else
04 hanoi(A,C,B,d
1
, d
2
, . . . d
n1
)
05 move d
n
from A to B.
06 hanoi(C,B,A,d
1
, d
2
, . . . d
n1
)
Figure 15.6: Solving the Towers of Hanoi problem
The function hanoi breaks up a problem of size n into two problems of
size n1. Alert! Warning bells! This cant be good: the sub-problems arent
much smaller than the original problem!
Anyway, hanoi breaks up a problem of size n into two problems of size
n 1. Other than the two recursive calls, it does only a constant amount of
work. So the running time T(n) for the function hanoi would be given by
the recursive denition (where c and d are constants):
T(1) = c
T(n) = 2T(n 1) +d
If we unroll this denition, we get
CHAPTER 15. ALGORITHMS 184
T(n) = 2T(n 1) +d
= 2 2(T(n 2) +d) +d
= 2 2(2(T(n 3) +d) +d) +d
= 2
3
T(n 3) + 2
2
d + 2d +d
= 2
k
T(n k) +d
k1

i=0
2
i
Well hit the base case when k = n 1. So
T(n) = 2
k
T(n k) +d
k1

i=0
2
i
= 2
n1
c +d
n2

i=0
2
i
= 2
n1
c +d(2
n1
1)
= 2
n1
c + 2
n1
d d
= O(2
n
)
15.9 Multiplying big integers
Suppose we want to multiply two integers. Back in the Bad Old Days of slow
computers, we would need to multiply moderate-sized integers digit-by-digit.
These days, we typically have a CPU that can multiply two moderate-sized
numbers (e.g. 16-bit, 32-bit) as a single operation. But some applications
(e.g. in cryptography) involve multiplying very, very large integers. Each
very long integer must then be broken up into a sequence of 16-bit or 32-bit
integers.
Lets suppose that we are breaking up each integer into individual digits,
and that were working in base 2. The recursive multiplication algorithm
divides each n-digit input number into two n/2-digit halves. Specically,
CHAPTER 15. ALGORITHMS 185
suppose that our input numbers are x and y and they each have 2m digits.
We can then divide them up as
x = x
1
2
m
+x
0
y = y
1
2
m
+y
0
If we multiply x by y in the obvious way, we get
xy = A2
2m
+B2
m
+C
where A = x
1
y
1
, B = x
0
y
1
+x
1
y
0
, and C = x
0
y
0
. Set up this way, computing
xy requires multiplying four numbers with half the number of digits.
In this computation, the operations other than multiplication are fast, i.e.
O(m) = O(n). Adding two numbers can be done in one sweep through the
digits, right to left. Multiplying by 2
m
is also fast, because it just requires
left-shifting the bits in the numbers. Or, if you arent very familiar with
binary yet, notice that multiplying by 10
m
just requires adding m zeros to
the end of the number. Left-shifting in binary is similar.
So, the running time of this naive method has the recursive denition:
T(1) = c
T(n) = 4T(n/2) +O(n)
The closed form for T(n) is O(n
2
) (e.g. use unrolling).
The trick to speeding up this algorithm is to rewrite our algebra for
computing B as follows
B = (x
1
+x
0
)(y
1
+y
0
) AC
This means we can compute B with only one multiplication rather that two.
So, if we use this formula for B, the running time of multiplication has the
recursive denition
CHAPTER 15. ALGORITHMS 186
P(1) = c
P(n) = 3P(n/2) +O(n)
Its not obvious that weve gained anything substantial, but we have. If we
build a recursion tree for P, we discover that the kth level of the tree contains
3
k
problems, each involving n
1
2
k
work. So each non-leaf level requires n(
3
2
)
k
work. The sum of the non-leaf work is dominated by the bottom non-leaf
level.
The tree height is log
2
(n), so the bottom non-leaf level is at log
2
(n) 1.
This level requires n(
3
2
)
log
2
n
work. If you mess with this expression a bit,
using facts about logarithms, you nd that its O(n
log
2
3
) which is approxi-
mately O(n
1.585
).
The number of leaves is 3
log
2
n
and constant work is done at each leaf.
Using log identities, we can show that this expression is also O(n
log
2
3
).
So this trick, due to Anatolii Karatsuba, has improved our algorithms
speed from O(n
2
) to O(n
1.585
) with essentially no change in the constants.
If n = 2
10
= 1024, then the naive algorithm requires (2
10
)
2
= 1, 048, 576
multiplications, whereas Katatsubas method requires 3
10
= 59, 049 multipli-
cations. So this is a noticable improvement and the dierence will widen as
n increases.
There are actually other integer multiplication algorithms with even faster
running times, e.g. Schoohage-Strassens method takes O(nlog nlog log n)
time. But these methods are more involved.
Chapter 16
NP
Some tasks denitely require exponential time. That is, we can not only dis-
play an exponential-time algorithm, but we can also prove that the problem
cannot be solved in anything less than exponential time. For example, weve
seen how to solve the Towers of Hanoi problem using O(2
n
) steps, and it is
known that no better algorithm is possible.
However, there is another large class of tasks where the best known algo-
rithm is exponential, but no one has proved that it is impossible to construct
a polynomial-time algorithm. This group of problems is known as NP.
1
It is
an important unsolved question whether the problems in NP really require
exponential time.
16.1 Finding parse trees
Suppose we are given an input sentence, such as
Scores of famished zombies are roaming the north end of campus.
If the sentence contains n words and we have a context-free grammar G,
it takes O(n
3
) time for a properly-designed parsing algorithm to nd a parse
1
NP is short for non-deterministic polynomial, but that full name makes sense only
when you have additional theory background.
187
CHAPTER 16. NP 188
tree for the sentence or determine that none exists. However, generating all
parse trees for such a sentence takes exponential time.
To see this, consider sentences that end in a large number of prepositional
phrases, e.g.
Prof. Rutenbar killed a zombie with a red hat near the Academic
Oce on the ground oor by the statue with a large sword.
When building the parse tree for such a sentence, each prepositional
phrase needs to be associated with either the main verb or with some preced-
ing noun phrase. For example, with a large sword might be the instrument
used to kill the zombie, in which case it should be linked with killed in the
parse tree. But its also possible that the sword is part of the description of
the statue, in which case with a large sword would be attached to statue
in the tree. No matter what you do with earlier bits of the sentence, there
are at least two possible ways to add each new phrase to the parse tree. So
if there are n prepositional phrases, there are at least 2
n
possible parse trees.
This example may seem silly. However, similar ambiguities occur with
other sorts of modiers. E.g. in wild garlic cheese, which is wild, the garlic
or the cheese? In normal conversation, sentences tend to be short and we
normally have enough context to infer the correct interpretation, and thus
build the correct parse tree. However, explosion in the number of possible
parse trees causes real issues when processing news stories, which tend to
have long sentences with many modiers.
16.2 What is NP?
Now, lets look at a problem that is in NP: graph colorability. Specically,
given a graph G and an integer k, determine whether G can be colored with k
colors. For small graphs, it seems like this problem is quite simple. Moreover,
heuristics such as the greedy algorithm from Chapter 11 work well for many
larger graphs found in practical applications. However, if k 3, for unlucky
choices of input graph, all known algorithms take time that is exponential in
the size of the graph. However, we dont know whether exponential time is
CHAPTER 16. NP 189
really required or whether there is some clever trick that would let us build
a faster algorithm.
To make the theoretical denitions easier to understand, lets upgrade our
colorability algorithm slightly. As stated above, our colorability algorithm
needs to deliver a yes/no answer. Lets ask our algorithm to, in addition,
deliver a justication that its answer is correct.
2
For example, for graph
colorability, a yes answer would also come with an assignment of the k
colors to the nodes of G. Its easy to verify that such an assignment is a
legal coloring using the specied number of colors and, thus, that the yes
answer is correct.
The big dierence between graph colorability and parse tree generation
is this ability to provide succinct, easily checked solutions with justications.
Specically, a computational problem is in the set NP when an algorithm
can provide justications of yes answers, where each justication can be
checked in polynomial time. That is, the justication checking time is poly-
nomial in the size of the input to the algorithm (e.g. the graph G in our
colorability problem). This implies that the justication itself must be suc-
cinct, because the checker must read through it. Graph colorability is in NP.
Parse tree generation is not in NP, because even the raw answer without
justication has exponential length.
For graph colorability, it only seems to be possible to provide succinct
justications for yes answers. For no answers, we can sometimes give a
short and convincing explanation. But, for dicult graphs, we might have to
walk through all exponentially-many possible assignments of colors to nodes,
showing why none of them is a legal coloring. So there doesnt seem to be a
general-purpose way to provide succinct justications for negative answers.
Obviously, the choice of which answer is yes vs. no depends on how
we phrase the problem. Consider the non-colorability problem: given a
graph G and an integer k, is it impossible to color G with k colors? For this
problem, we can give succinct justications for no answers but apparently
not for yes answers. Problems for which we can give polynomial-time
checkable justications for negative answers are in the set co-NP.
There are some dicult problems where we can provide justications for
both negative and positive answers. For example, there are ecient ways
2
Like we force you folks to do, when answering questions on theory exams!
CHAPTER 16. NP 190
to verify whether an integer is prime or not, but factoring integers is much
harder. For example, we do not know of any polynomial-time way to de-
termine whether an integer m has a factor in the range [2, n].
3
However,
having gone to the considerable pain of factoring m, we can provide suc-
cinct, easily-veried justications for both yes and no answers using the
prime factorization of m. So this problem is in both NP and co-NP.
Algorithms that can be solved in polynomial time are in the set P. P is
a subset of both NP and co-NP. Theoreticians strongly suspect that these
three setsP, NP, and co-NPare all dierent. However, no one has found
a way to prove this. So it is possible that all three sets are equal, i.e. that
weve missed some big trick that would let us build more ecient algorithms
for the problems in NP and co-NP.
16.3 Circuit SAT
Another interesting problem in NP is satisability of boolean circuits. A
boolean circuit is a type of graph consisting of a set of gates connected
together by wires. Each wire carries one of the two values 0 (false) and 1
(true). The label on each gate describes how the value on its output wires
(all output wires have the same value) is determined by the values on its
input wires. For example, the output wires of an AND gate have value 1 if
both input wires have value 1, and the output wires have value 0 otherwise.
Like a logical operator, the behavior of a gate can be described by a truth
table. For simplicity, well restrict ourselves to gates with one or two inputs.
A boolean circuit with any possible behavior can be built using the basic
three operators (AND, OR, NOT). However, in practical applications, its
common to also use other gates, e.g. XOR (exclusive OR), NAND (negation
of AND).
For example, here is a very simple boolean circuit. In this picture, inputs
are on the left and outputs on the right. For ease of understanding by
software and theory folks, the type of each gate is indicated with the name
of the operator. For serious circuit design, youd need to memorize the special
node shapes commonly used for dierent gate types.
3
That is, polynomial in the number of digits in the input numbers.
CHAPTER 16. NP 191
NOT
OR
OR
AND
XOR
When designing circuits for practical applications, its important to en-
sure that the circuit really does what it is supposed to do. Or, at least,
that it doesnt make certain critical errors that would lead to tragedy. For
example, the circuit controlling a 4-way intersection must ensure that two
crossing lanes of trac cannot both have a green light at the same time. A
slow-moving 6-legged robot should never raise more than three of its legs at
a time. A heating system should not open the gas jet if the pilot light is o.
Signicant portions of such controllers are well-modelled as boolean circuits.
To nd signicant design aws in a boolean circuit, we need to determine
whether there is a set of n input values could create one of the forbidden
patterns of output values. This problem is called circuit satisability. There
is an obvious exponential algorithm for circuit satisability: if we have n
input wires, generate all 2
n
possible input patterns and see what happens on
the output wires. Thats too slow, so circuit testers must be content with
partial testing. However, if we determine that the bad output is possible, we
can provide a succinct easily-veried justication: show one pattern of input
values that gives rise to the bad output. So this problem is in NP.
Sadly, what wed really like for practical applications would be easily-
veried justications of negative answers. Or, equivalently, positive answers
to the the complementary problem of circuit safety: is the circuit safe from
bad outputs? This seems to be much harder than justications for circuit
satisability, because proving a negative answer seems to require walking
CHAPTER 16. NP 192
through all exponentially many input patterns and showing why none of
them produces the bad output pattern. So circuit safety is in co-NP but
apparently not in NP.
16.4 What is NP complete?
An interesting, and very non-obvious, fact about the harder problems in
NP is that they share a common fate. It can be proven that nding a
polynomial-time algorithm for either circuit satisability or graph colorability
would imply that all problems in NP have polynomial-time algorithms. Such
problems are called NP-complete.
Other examples of NP complete problems include
Propositional logic satisability: given an expression in propositional
logic, is there an assignment of true/false values to the variables which
would make the logic expression true?
Clique: given an integer n and a graph G, does G contain a copy of
K
n
?
Vertex cover: given a graph G, nd a set S of nodes in G such that
each edge of G has at least one of its endpoints in S.
Circuit satisability and propositional logic satisability remain NP-complete
even when these problems are drastically simplied. For example, nothing
changes if we restrict our circuit to have only a single output. We can re-
strict our logic expression to be the AND of some number of 3-variable OR
expressions. Or we can use only the primitive operators/gates AND, OR,
and NOT. Only the most grossly simplied versions of these problems fail to
be NP-complete.
There are two techniques for showing that a problem is NP-complete.
First, we can show directly that the problem is so general that it can simulate
any other problem in NP. The details are beyond the scope of the book. But
notice that circuits are the backbone of computer construction and logic
is the backbone for constructing mathematics, so it makes sense that both
CHAPTER 16. NP 193
mechanisms are general and powerful. Once we have proved that some key
problems are NP-complete, we can then show that other problems are NP-
complete by showing that the new problem can simulate the basic primitives
of a problem already known to be NP-complete. For example, we might
simulate logic expressions using colored graph machinery.
16.5 Variation in notation
A justication for a positive answer can also be called a proof, certicate,
or witness.
Chapter 17
Proof by Contradiction
This chapter covers proof by contradiction. This is a powerful proof technique
that can be extremely useful in the right circumstances. Well need this
method in Chapter 20, when we cover the topic of uncountability. However,
contradiction proofs tend to be less convincing and harder to write than
direct proofs or proofs by contrapositive. So this is a valuable technique
which you should use sparingly.
17.1 The method
In proof by contradiction, we show that a claim P is true by showing that its
negation P leads to a contradiction. If P leads to a contradiction, then
P cant be true, and therefore P must be true. A contradiction can be
any statement that is well-known to be false or a set of statements that are
obviously inconsistent with one another, e.g. n is odd and n is even, or x < 2
and x > 7.
Proof by contradiction is typically used to prove claims that a certain type
of object cannot exist. The negation of the claim then says that an object
of this sort does exist. The existence of an object with specied properties
is often a good starting point for a proof. For example:
Claim 54 There is no largest even integer.
194
CHAPTER 17. PROOF BY CONTRADICTION 195
Proof: Suppose not. That is, suppose that there were a largest
even integer. Lets call it k.
Since k is even, it has the form 2n, where n is an integer. Consider
k + 2. k + 2 = (2n) + 2 = 2(n + 1). So k + 2 is even. But k + 2
is larger than k. This contradicts our assumption that k was the
largest even integer. So our original claim must have been true.

The proof starts by informing the reader that youre about to use proof
by contradiction. The phrase suppose not is one traditional way of doing
this. Next, you should spell out exactly what the negation of the claim is.
Then use mathematical reasoning (e.g. algebra) to work forwards until you
deduce some type of contradiction.
17.2

2 is irrational
One of the best known examples of proof by contradiction is the proof that

2 is irrational. This proof, and consequently knowledge of the existence of

irrational numbers, apparently dates back to the Greek philosopher Hippasus
in the 5th century BC.
We dened a rational number to be a real number that can be written as
a fraction
a
b
, where a and b are integers and b is not zero. If a number can
be written as such a fraction, it can be written as a fraction in lowest terms,
i.e. where a and b have no common factors. If a and b have common factors,
its easy to remove them.
Also, we proved (above) that, for any integer k, if k is odd then k
2
is odd.
So the contrapositive of this statement must also be true: (*) if k
2
is even
then k is even.
Now, we can prove our claim:
Suppose not. That is, suppose that

2 were rational.
Then we can write

2 as a fraction
a
b
where a and b are integers
with no common factors.
CHAPTER 17. PROOF BY CONTRADICTION 196
Since

2 =
a
b
, 2 =
a
2
b
2
. So 2b
2
= a
2
.
By the denition of even, this means a
2
is even. But then a must
be even, by (*) above. So a = 2n for some integer n.
If a = 2n and 2b
2
= a
2
, then 2b
2
= 4n
2
. So b
2
= 2n
2
. This means
that b
2
is even, so b must be even.
We now have a contradiction. a and b were chosen not to have
any common factors. But they are both even, i.e. they are both
divisible by 2.
Because assuming that

2 was rational led to a contradiction, it

must be the case that

2 is irrational.
17.3 There are innitely many prime num-
bers
Contradiction also provides provides a nice proof of a classic theorem about
prime numbers, dating back to Euclid, who lived around 300 B.C.
Euclids Theorem: There are innitely many prime numbers.
This is a lightly disguised type of non-existence claim. The theorem
could be restated as there is no largest prime or there is no nite list of
all primes. So this is a good situation for applying proof by contradiction.
Proof: Suppose not. That is, suppose there were only nitely
many prime numbers. Lets call them p
1
, p
2
, up through p
n
.
Consider Q = p
1
p
2
p
n
+ 1.
If you divide Q by one of the primes on our list, you get a re-
mainder of 1. So Q isnt divisible by any of the primes p
1
, p
2
,
up through p
n
. However, by the Fundamental Theorem of Arith-
metic, Q must have a prime factor (which might be either itself
or some smaller number). This contradicts our assumption that
p
1
, p
2
,. . . p
n
was a list of all the prime numbers.
CHAPTER 17. PROOF BY CONTRADICTION 197
Notice one subtlety. Were not claiming that Q must be prime. Rather,
were making the much weaker claim that Q isnt divisible by any of the rst
n primes. Its possible that Q might be divisible by another prime larger
than p
n
.
17.4 Lossless compression
A nal example concerns le compression. A le compression algorithm
attempts to reduce the size of les, by transforming each input le to an
output le with fewer bits. A lossless algorithm allows you to reconstruct the
original le exactly from its compressed version, whereas a lossy algorithm
only allows you to reconstruct an approximation to the original le. Or, said
another way, a lossless algorithm must convert input les to output les in a
one-to-one-manner, so that two distinct input les are never compressed to
the same output le.
Claim 55 A lossless compression algorithm that makes some les smaller
must make some (other) les larger.
Proof: Suppose not. That is, suppose that we had a lossless
compression algorithm A that makes some les smaller and does
not make any les larger.
Let x be the shortest le whose compressed size is smaller than
its original size. (If there are two such les of the same length,
pick either at random.) Suppose that the input size of x is m
characters.
Suppose that S is the set of distinct les with fewer than m
characters. Because x shrinks, A compresses x to a le in S.
Because no les smaller than x shrink, each le in S compresses
to a le (perhaps the same, perhaps dierent) in S.
Now we have a problem. A is supposed to be lossless, therefore
one-to-one. But A maps a set containing |S| + a les to a set
containing |S| les, so the Pigeonhole Principle states that two
input les must be mapped to the same output le. This is a
contradiction.
CHAPTER 17. PROOF BY CONTRADICTION 198
So, on the face of it, lossless le compression algorithms cant win. How
do they work so well in practice? One secret is that compression algorithms
can ensure that le sizes never increase much. If a le would increase in size,
the algorithm stores the original version unchanged, preceded with a one-bit
marker. This bounds the potential damage if we encounter a bad input
le.
The second secret is that commonly-occurring les are not created at
random but have denite patterns. Text les contain natural language text.
Digitized images contain values that tend to change gradually. Compression
algorithms are tuned so that common types of les shrink. The fact that
some les might get bigger isnt a serious practical problem if those les are
unlikely to occur on your disk.
17.5 Philosophy
Proof by contradiction strikes many people as mysterious, because the argu-
ment starts with an assumption known to be false. The whole proof consists
of building up a fantasy world and then knocking it down. Although the
method is accepted as valid by the vast majority of theoreticians, these proofs
are less satisfying than direct proofs which construct the world as we believe
it to be. The best mathematical style avoids using proof by contradiction
except when it will denitely result in a much simpler argument.
There is, in fact, a minority but long-standing thread within theoretical
mathematics, called constructive mathematics, which does not accept this
proof method. They have shown that most of standard mathematics can
be re-built without it. For example, the irrationality of

2 can be proved
constructively, by showing that there is an error separating

2 from any
chosen fraction
a
b
.
Chapter 18
Sets of Sets
So far, most of our sets have contained atomic elements (such as numbers or
strings) or tuples (e.g. pairs of numbers). Sets can also contain other sets.
For example, {Z, Q} is a set containing two innite sets. {{a, b}, {c}} is a
set containing two nite sets. In this chapter, well see a variety of examples
involving sets that contain other sets. To avoid getting confused, well use
the term collection to refer to a set that contains other sets, and use a script
letter for its variable name.
18.1 Sets containing sets
Sets containing sets arise naturally when an application needs to consider
some or all of the subsets of a base set A. For example, suppose that we
have a set of 6 students:
A = {Ian, Chen, Michelle, Emily, Jose, Anne}
We might divide A up into non-overlapping groups based on what dorm
they live in, to get the collection:
B = {{Ian, Chen, Jose}, {Anne}, {Michelle, Emily}}
199
CHAPTER 18. SETS OF SETS 200
We could also construct a set of overlapping groups, each containing stu-
dents who play a common musical instrument (e.g. perhaps Michelle and
Chen both play the oboe). The collection of these groups might look like:
D = {{Ian, Emily, Jose}, {Anne, Chen, Ian}, {Michelle, Chen}, {Ian}}
Or we could try to list all ways that we could choose a 3-person committee
from this set of students, which would be a rather large collection containing
elements such as {Ian, Emily, Jose} and {Ian, Emily, Michelle}.
When a collection like B is the domain of a function, the function maps
an entire subset to an output value. For example, suppose we have a function
f : B {dorms}. Then f would map each set of students to a dorm. E.g.
f({Michelle, Emily}) = Babcock.
The value of a function on a subset can depend in various ways on what-
ever is in the subset. For example, suppose that we have a collection
D = {{12, 7, 9, 2}, {2, 3, 7}, {10, 3, 10, 4}, {1, 2, 3, 6, 8}}
We might have a function g : D R which maps each subset to some
descriptive statistic. For example, g might map each subset to its mean value.
And then we would have g({12, 7, 9, 2}) = 1.5 and g({1, 2, 3, 6, 9}) = 4.2.
When manipulating sets of sets, its easy to get confused and lose a layer
of structure. To avoid this, imagine each set as a box. Then the collection
F = {{a, b}, {c}, {a, p, q}} is a box containing three boxes. One of the inside
boxes contains a and b, the other contains c, and the third contains a, p, and
q. So the cardinality of F is three.
The empty set, like any other set, can be put into another set. So {}
is a set containing the empty set. Think of it as a box containing an empty
box. The set {, {3, 4}} has two elements: the empty set and the set {3, 4}.
CHAPTER 18. SETS OF SETS 201
18.2 Powersets and set-valued functions
If A is a set, the powerset of A (written P(A) is the collection containing all
subsets of A. For example, suppose that A = {1, 2, 3}. Then
P(A) = {, {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}}
Suppose A is nite and contains n elements. When forming a subset, we
have two choices for each element x: include x in the subset or dont include
it. The choice for each element is independent of the choice we make for the
other elements. So we have 2
n
ways to form a subset and, thus, the powerset
P(A) contains 2
n
elements.
Notice that the powerset of A always contains the empty set, regardless
of whats in A. As a consequence, P() = {}.
Powersets often appear as the co-domain of functions which need to return
a set of values rather than just a single value. For example, suppose that we
have the following graph whose set of nodes is V = {a, b, c, d, e, f, g, h}.
e
a
b
c
d
f
g
h
Now, lets dene the function n so that it takes a node as input and
returns the neighbors of that node. A node might have one neighbor, but
it could have several, and it might have no neighbors. So the outputs of
n cant be individual nodes. They must be sets of nodes. For example,
n(a) = {b, c, e} and N(f) = . Its important to be consistent about the
output type of n: it always returns a set. So n(g) = {h}, not n(g) = h.
Formally, the domain of n is V and the co-domain is P(V ). So the type
signature of n would be n : V P(V ).
Suppose we have the two graphs shown below, with sets of nodes X =
CHAPTER 18. SETS OF SETS 202
{a, b, c, d, e} and Y = {1, 2, 3, 4, 5}. And suppose that were trying to nd
all the possible isomorphisms between the two graphs. We might want a
function f that retrieves likely corresponding nodes. For example, if p is a
node in X, then f(p) might be the set of nodes in Y with the same degree
as p.
a
b
c
d
e
5 1
2
3
4
f cant return a single node, because there might be more than one node
in Y with the same degree. Or, if the two graphs arent isomorphic, no
nodes in Y with the same degree. So well have f return a set of nodes. For
example, f(e) = {1, 5} and f(a) = {2}. The co-domain of f will need to be
P(Y ). So, to declare f, wed write f : X P(Y ).
18.3 Partitions
When we divide a base set A into non-overlapping subsets which include
every element of A, the result is called a partition of A. For example,
suppose that A is the set of nodes in the following graph. The partition
{{a, b, c, d}, {e, f, g}, {h, i, j, k}} groups nodes into the same subset if they
belong to the same connected component.
a
b
c
d
e
f
g
h
i
j
k
Notice that being in the same connected component is an equivalence
relation on the nodes of a graph. In general, each equivalence relation corre-
CHAPTER 18. SETS OF SETS 203
sponds to a partition of its base set, and vice versa. Each set in the partition
is exactly one of the equivalence classes of the relation. For example, con-
gruence mod 4 corresponds to the following partition of the integers:
{{0, 4, 4, 8, 8, . . .}, {1, 5, 3, 9, 7, . . .},
{2, 6, 2, 10, 6, . . .}, {3, 7, 1, 11, 5, . . .}}
We could also write this partition as {[0], [1], [2], [3]} since each equivalence
class is a set of numbers.
Collections of subsets dont always form partitions. For example, consider
the following graph G.
a
b
c
d
f
e
g
Suppose we collect sets of nodes in G that form a cycle. Well get the
following set of subsets. This isnt a partition because some of the subsets
overlap.
{{f, c, d}, {a, b, c, d}, {a, b, c, d, f}, {f, e, g}}
Formally, a partition of a set A is a set of non-empty subsets of A which
cover all the elements of A and which dont overlap. So, if the subsets in the
partition are A
1
, A
2
, . . . A
n
, then they must satisfy three conditions:
1. covers all of A: A
1
A
2
. . . A
n
= A
2. non-empty: A
i
= for all i
3. no overlap: A
i
A
j
= for all i = j.
CHAPTER 18. SETS OF SETS 204
Its possible for a partition of an innite set A to contain innitely many
subsets. For example, we can partition the integers into subsets each of which
contains integers with the same magnitude:
{{0}, {1, 1}, {2, 2}, {3, 3}, . . .}
We need more general notation to cover the possibility of an innite
partition. Suppose that C is a partition of A. Then C must satisfy the
following conditions:
1. covers all of A:

XC
X = A
2. non-empty: X = for all X C
3. no overlap: X Y = for all X, Y C, X = Y
The three dening conditions of an equivalence relation (reexive, sym-
metric, and transitive) were chosen so as to force the equivalence classes
to be a partition. Relations without one of these properties would gener-
ate equivalence classes that might be empty, have partial overlaps, and so
forth.
18.4 Combinations
In many applications, we have an n-element set and need to count all subsets
of a particular size k. A subset of size k is called a k-combination. Notice
the dierence between a permutation and a combination: we care about the
order of elements in a permutation but not in a combination.
For example, how many ways can I select a 7-card hand from a 60-card
deck of Magic cards (assuming no two cards are identical)?
1
One way to analyze this problem is to gure out how many ways we
can select an ordered list of 7 cards, which is P(60, 7). This over-counts the
1
Ok, ok, for those of you who actually play Magic, decks do tend to contain identical
land cards. But maybe we are using lots of special lands or perhaps well treat cards with
dierent artwork as dierent.
CHAPTER 18. SETS OF SETS 205
number of possibilities, so we have to divide by the number of dierent orders
in which the same 7-cards might appear. Thats just 7!. So our total number
of hands is
P(60,7)
7!
This is
60595857565554
765432
. Probably not worth simplifying or
multiplying this out unless you really have to. (Get a computer to do it.)
In general, suppose that we have a set S with n elements and we want to
choose an unordered subset of k elements. We have
n!
(nk)!
ways to choose k
elements in some particular order. Since there are k! ways to put each subset
into an order, we need to divide by k! so that we will only count each subset
once. So the general formula for the number of possible subsets is
n!
k!(nk)!
.
The expression
n!
k!(nk)!
is often written C(n, k) or
_
n
k
_
. This is pronounced
n choose k. It is also sometimes called a binomial coecient, for reasons
that will become obvious shortly. So the shorthand answer to our question
about magic cards would be
_
60
7
_
.
Notice that
_
n
r
_
is only dened when n r 0. What is
_
0
0
_
? This is
0!
0!0!
=
1
11
= 1.
18.5 Applying the combinations formula
The combinations formula is often used when we want to select a set of
locations or positions to contain a specic value. For example, recall that a
bit string is a string of 0s and 1s. Suppose we want to gure out how many
16-digit bit strings contain exactly 5 zeros. Lets think of the string as having
16 positions. We need to choose 5 of these to be the positions containing the
zeros. We can how apply the combinations formula: we have
_
16
5
_
ways to
select these 5 positions.
To take a slightly harder example, lets gure out how many 10-character
strings from the 26-letter ASCII alphabet contain no more than 3 As. Such
strings have to contain 0, 1, 2, or 3 As. To nd the number of strings
containing exactly three As, we rst pick three of the 10 positions to contain
the As. There are
_
10
3
_
ways to do this. Then, we have seven positions to ll
with our choice of any character except A. We have 25
7
ways to do that. So
our total number of strings with 3 As is
_
10
3
_
25
7
.
To get the total number of strings, we do a similar analysis to count
CHAPTER 18. SETS OF SETS 206
the strings with 0, 1, and 2 As. We then add up the counts for the four
possibilities to get a somewhat messy nal answer for the number of strings
with 3 or fewer As:
_
10
3
_
25
7
+
_
10
2
_
25
8
+
_
10
1
_
25
9
+ 25
10
18.6 Combinations with repetition
Suppose I have a set S and I want to select a group of objects of the types
listed in S, but Im allowed to pick more than one of each type of object.
For example, suppose I want to pick 6 plants for my garden and the set of
available plants is S = {thyme, oregano, mint}. The garden store can supply
as many as I want of any type of plant. I could pick 3 thyme and 3 mint. Or
I could pick 2 thyme, 1 oregano, and 3 mint.
Theres a clever way to count the possibilities here. Lets draw a picture of
a selection as follows. Well group all our thymes together, then our oreganos,
then our mints. Between each pair of groups, well put a cardboard separator
#. So 2 thyme, 1 oregano, and 3 mint looks like
T T # O # M M M
And 3 thyme and 3 mint looks like
T T T ## M M M
But this picture is redundant, since the items before the rst separator
are always thymes, the ones between the separators are oreganos, and the
last group are mints. So we can simplify the diagram by using a star for each
object and remembering their types implicitly. Then 2 thyme, 1 oregano,
and 3 mint looks like
** # * # ***
And 3 thyme and 3 mint looks like
CHAPTER 18. SETS OF SETS 207
*** ## ***
To count these pictures, we need to count the number of ways to arrange
6 stars and two #s. That is, we have 8 positions and need to choose 2 to ll
with #s. In other words,
_
8
2
_
.
In general, suppose we are picking a group of k objects (with possible
duplicates) from a list of n types. Then our picture will contain k stars and
n 1 #s. So we have k +n 1 positions in the picture and need to choose
n 1 positions to contain the #s. So the number of possible pictures is
_
k +n 1
n 1
_
.
Notice that this is equal to
_
k + n 1
k
_
because we have an identity
that says so (see above). We could have done our counting by picking a
subset of k positions in the diagram that we would ll with stars (and then
the rest of the positions will get the #s).
If wanted to pick 20 plants and there were ve types available, I would
have
_
24
4
_
=
_
24
20
_
options for how to make my selection.
_
24
4
_
=
24232221
433
= 23 22 21.
18.7 Identities for binomial coecients
There are a large number of useful identities involving binomial coecients,
most of which you can look up as you need them. Two really basic ones are
worth memorizing. First, a simple consequence of the denition of
_
n
k
_
is that
_
n
k
_
=
_
n
n k
_
Pascals identity also shows up frequently. It states that
(Pascals identity)
_
n+1
k
_
=
_
n
k
_
+
_
n
k1
_
CHAPTER 18. SETS OF SETS 208
This is not hard to prove from the denition of
_
n
k
_
. To remember it, suppose
that S is a set with n +1 elements. The lefthand side of the equation is the
number of k-element subsets of S.
Now, x some element a in S. There are two kinds of k-element subsets:
(1) those that dont contain a and (2) those that do contain a. The rst term
on the righthand side counts the subsets in group (1): all k-element subsets
of S {a}. The second term on the righthand side counts the k 1-element
subsets of S {a}. We then add a to each of these to get the subsets in
group (2).
If we have Pascals identity, we can give a recursive denition for the
binomial coecients, for all natural numbers n and k with k n.
Base: For any natural number k,
_
n
0
_
=
_
n
n
_
= 1.
Induction:
_
n
k
_
=
_
n1
k1
_
+
_
n1
k
_
, whenever k < n
18.8 Binomial Theorem
Remember that a binomial is a sum of two terms, e.g. (x + y). Binomial
coecients get their name from the following useful theorem about raising a
binomial to an integer power:
Claim 56 (Binomial Theorem) Let x and y be variables and let n be any
natural number. Then
(x +y)
n
=
n

k=0
_
n
k
_
x
nk
y
k
When we expand the product (x + y)
n
, each term is the product of n
variables, some xs and the rest ys. For example, if n = 5, one term is
yxyyx. So each term is an ordered list of xs and ys.
We can think of our large set of terms as partitioned into subsets, each
containing terms with the same number of xs. For example, the set of terms
with two xs would be
CHAPTER 18. SETS OF SETS 209
[xxyyy] = {xxyyy, xyxyy, xyyxy, xyyyx, yxxyy,
yxyxy, yxyyx, yyxxy, yyxyx, yyyxx}
When we collect terms, the coecient for each term will be the size of this
set of equivalent terms. E.g. the coecient for x
2
y
3
is 10, because [xxyyy]
contains 10 elements. To nd the coecient for x
nk
y
k
, we need to count
how many ways we can make a sequence of n variable names that contains
k ys and n k xs. This amounts to picking a subset of k elements from a
set of n positions in the sequence. In other words, there are
_
n
k
_
such terms.
18.9 Variation in notation
Weve used the notation P(A) for the powerset of A. Another common
notation is 2
A
.
Weve used the term collection to refer to sets containing other sets. A
collection is just a special type of set, so its also ok to just call them sets.
Chapter 19
State Diagrams
In this chapter, well see state diagrams, an example of a dierent way to use
directed graphs.
19.1 Introduction
State diagrams are a type of directed graph, in which the graph nodes rep-
resent states and labels on the graph edges represent actions. For example,
here is a state diagram representing the life cycle of a chicken:
chicken
egg
chick
omelet
hatch
grow
lay
cook
The label on the edge from state A to state B indicates what action
happens as the system moves from state A to state B. In many applications,
all the transitions involve one basic type of action, such as reading a character
210
CHAPTER 19. STATE DIAGRAMS 211
or going though a doorway. In that case, the diagram might simply indicate
the details of this action. For example, the following diagram for a multi-
room computer game shows only the direction of motion on each edge.
entry
dining rm
hall
study
barn
ferry
cellar
north west south
east
east
west
east
west
east
down
down
up
east
Walks (and therefore paths and cycles) in state diagrams must follow the
arrow directions. So, for example, there is no path from the ferry to the
study. Second, an action can result in no change of state, e.g. attempting
to go east from the cellar. Finally, two dierent actions may get you to the
same new state, e.g. going either west or north from the hall gets you to the
dining room.
Remember that the full specication of a walk in a graph contains both
a sequence of nodes and a sequence of edges. For state diagrams, these
correspond to a sequence of states and a sequence of actions. Its often
important to include both sequences, both to avoid ambiguity and because
the states and actions may be important to the end user. For example, for
one walk from the hall to the barn, the full state and action sequences look
like:
states: hall dining room hall cellar barn
actions: west south down up
CHAPTER 19. STATE DIAGRAMS 212
19.2 Wolf-goat-cabbage puzzle
State diagrams are often used to model puzzles or games. For example, one
famous puzzle involves a farmer taking a wolf, a goat, and a cabbage to
market. To do this, he must cross from the east to the west side of a river
using a boat that can only carry him plus one of his three possessions. He
cannot leave the wolf and goat together unsupervised, nor the goat and the
cabbage, because one will eat the other.
We can represent each state of this system by listing the objects that are
on the east bank: w is the wolf, g is the goat, c is the cabbage, and b is the
boat. States like wc and wgb are legal, but wg would not be a legal state.
The diagram of legal states then looks as follows:
wgcb wc
wcb
w
c
wgb
cgb
g
gb

In this diagram, actions arent marked on the edges: youre left to infer
the action from the change in state. The start state (wgcb) where the system
begins is marked by showing an arrow leading into it. The end state () where
nothing is left on the east bank is marked with a double ring.
In this diagram, it is possible for a (directed) walk to loop back on itself,
repeating states. So there are two shortest solutions to this puzzle, but also
an innite number of other solutions that involve undoing and then redoing
some piece of work.
CHAPTER 19. STATE DIAGRAMS 213
19.3 Phone lattices
Another standard application for state diagrams is in modelling pronuncia-
tions of words for speech recognition. In these diagrams, known as phone
lattices, each edge represents the action of reading a single sound (a phone)
from the input speech stream. To make our examples easy to read, well
pretend that English is spelled phonetically, i.e. each letter of English rep-
resents exactly one sound/phone.
1
For example, we could model the set of
words {chat, chop, chip} using the diagram:
1 2 3 4
5
6
7
c h
a
t
o
i
p
As in the wolf-goat-cabbage puzzle, we have a clearly dened start state
(state 1, marked with an incoming arrow). This time there are two end states
(6 and 7) marked with double rings. Its conventional that there is only one
start state but there may be several end states.
Another phone lattice might represent the set of words {cat, cot, cab, cob}.
1 8 9 10
c
a
o
t
b
We can combine these two phone lattices into one large diagram, repre-
senting the union of these two sets of words:
1
This is, of course, totally wrong. But the essential ideas stay the same when you
switch to a phonetic spelling of English.
CHAPTER 19. STATE DIAGRAMS 214
1
2 3 4
5
6
7
8 9 10
c
a
o
t
b
c
h
a
t
o
i
p
Notice that there are two edges leading from state 1, both marked with
the phone c. This indicates that the user (e.g. a speech understanding
program) has two options for how to handle a c on the input stream. If
this was inconvenient for our application, it could be eliminated by merging
states 2 and 8.
Occasionally, its useful for a phone lattice to contain a loop. E.g. peo-
ple often drag out the pronunciation of the word um. So the following
represents all words of the form uu

m, i.e. um, uum, uuuuuum, etc.

1 2 3
u
u m
Many state diagrams are passive representations of a set of possibilities.
For example, a room layout for a computer game merely shows what walks are
possible; the player makes the decisions about which walk to take. Phone
lattices are often used in a more active manner, as a very simple sort of
computer called a nite automaton. The automaton reads an input sequence
of characters, following the edges in the phone lattice. At the end of the
input, it reports whether it has or hasnt successfully reached an end state.
CHAPTER 19. STATE DIAGRAMS 215
19.4 Representing functions
To understand how state diagrams might be represented mathematically
and/or stored in a computer, lets rst look in more detail at how func-
tions are represented. A function f : A B associates each input value
from A with an output value from B. So we can represent f mathematically
as a set of input/output pairs (x, y), where x comes from A and y comes from
B. For example, one particular function from {a, b, c} to {1, 2, 3, 4} might be
represented as
{(a, 4), (b, 1), (c, 4)}
In a computer, this information might be stored as a linked list of pairs.
Each pair might be a short list, or an object or structure. However, nding
a pair in such a list requires time proportional to the length of the list. This
can be very slow if the list is long i.e. if the domain of the function is large.
Another option is to convert input values to small integers. For example,
in C, lowercase integer values (as in the set A above) can be converted to small
integers by subtracting 97 (the ASCII code for a). We can then store the
functions input/output pairs using an array. Each array position represents
an input value. Each array cell contains the corresponding output value.
19.5 Transition functions
Formally, we can dene a state diagram to consist of a set of states S, a set of
actions A, and a transition function that maps out the edges of the graph,
i.e. shows how actions result in state changes. Each edge of the state diagram
shows what happens when you are in a particular state s and execute some
action a, i.e. which new state(s) could you be in.
So an input to is a pair (s, a) of a state and an action. Each output of
is a set of states. So the type signature of looks like : S A P(S).
For many input pairs, produces a set containing a single state. So why
does produce a set of states rather than a single state? This is for two
reasons. First, in some state diagrams, it might not be possible to execute
CHAPTER 19. STATE DIAGRAMS 216
certain actions from certain states, e.g. you cant go up from the study in
our room layout example. returns in such cases. Second, as we saw in
the phone lattice example, it is sometimes convenient to allow the user or the
computer system several choices. In this case, returns a set of possibilities
for the new state.
The hard part of implementing state diagrams in a computer program
is storing the transition function. We could build a 2D array, whose cells
represent all the pairs (s, a). Each cell would then contain a list of output
states. This wouldnt be very ecient, however, because state diagrams tend
to be sparse: most state/action pairs dont produce any new state.
One better approach is to build a 1D array of states. The cell for each
state contains a list of actions possible from that state, together with the
new states for each action. For example, in our nal phone lattice, the
entry for state 1 would be ((c, (2, 8))) and the entry for state 3 would be
((o, (5)), (i, (5)), (a, (4))). This adjacency list style of storage is much more
compact, because we are no longer wasting space representing the large num-
ber of impossible actions.
Another approach is to build a function, called a hash function that maps
each state/action pair to a small integer. We then allocate a 1D array with
one position for each state/action pair. Each array cell then contains a list
of new states for this state/action pair. The details of hash functions are
beyond the scope of this class. However, modern programming languages
often include built-in hash table or dictionary objects that handle the details
for you.
19.6 Shared states
Suppose that each word in our dictionary had its own phone lattice, and
we then merge these individual lattices to form lattices representing sets of
words. We might get a lattice like the following one, which represents the
set of words {cop, cap, cops, caps}.
CHAPTER 19. STATE DIAGRAMS 217
1 5 6 7 8
2 3 4
9 10 11 12
13 14 15
c a
p
s
c
a
p
c
o
p
s c
o
p
Although this lattice encodes the right set of words, it uses a lot more
states than necessary. We can represent the same information using the
following, much more compact, phone lattice.
1 2 3 4 5
c
a
o
p
s
State merger is even more important when states are generated dynam-
ically. For example, suppose that we are trying to nd an optimal strategy
for playing a game like tic-tac-toe. Blindly enumerating sequences of moves
might create state diagrams such as:
X
X
O
X
X O
X O
X
X O
X
X O
X O
X O
X O
Searching through the same sequence of moves twice wastes time as well
as space. If we build an index of states weve already generated, we can detect
when we get to the same state a second time. We can then use the results
of our previous work, rather than repeating it. This trick, called dynamic
CHAPTER 19. STATE DIAGRAMS 218
programming, can signicantly improve the running time of algorithms.
X
X
O
X
X O
X O
X
X O
X O
We can also use state diagrams to model what happens when computer
programs run. Consider the following piece of code
cyclic()
y = 0
x = 0
while (y < 100)
x = remainder(x+1,4)
y = 2x
We can represent the state of this machine by giving the values of x and y.
We can then draw its state diagram as shown below. By recognizing that we
return to the same state after four iterations, we not only keep the diagram
compact but also capture the fact that the code goes into an innite loop.
x=0,y=0 x=1,y=2 x=2,y=4 x=3,y=6
CHAPTER 19. STATE DIAGRAMS 219
19.7 Counting states
The size of state diagrams varies dramatically with the application. For
example, the wolf-goat-cabbage problem has only 2
4
= 16 possible states,
because we have two choices for the position of each of the four key objects
(wolf, goat, cabbage, boat). Of these, six arent legal because they leave a
hungry animal next to its food with the boat on the other side of the river.
For this application, we have no trouble constructing the full state diagram.
Many applications, however, generate quite big state diagrams. For ex-
ample, a current speech understanding system might need to store tens of
thousands of words and millions of short phrases. This amount of data can
be packed into computer memory only by using high-powered storage and
compression tricks.
Eventually, the number of states becomes large enough that it simply
isnt possible to store them explicitly. For example, the board game Go is
played on a 19 by 19 grid of positions. Each position can be empty, lled with
a black stone, or lled with a white stone. So, to a rst approximation, we
have 3
361
possible game positions. Some of these are forbidden by the rules
of the game, but not enough to make the size of the state space tractable.
Some games and theoretical models of computation use an unbounded
amount of memory, so that their state space becomes innite. For example,
Conways Game of Life runs on an 2D grid of cells, in which each cell has
8 neighbors. The game starts with an initial conguration, in which some
set of cells is marked as live, and the rest are marked as dead. At each time
step, the set of live cells is updated using the rules:
A live cell stays live if it 2 or 3 neighbors. Otherwise it becomes dead.
A dead cell with exactly 3 neighbors becomes live.
For some initial congurations, the system goes into a stable state or
oscillates between several congurations. However, even if the initial set of
live cells is nite, the set of live cells can grow without bound as time moves
forwards. So, for some initial congurations, the system has innitely many
states.
CHAPTER 19. STATE DIAGRAMS 220
When a system has an intractably large number of states, whether nite
or innite, we obviously arent going to build its state space explicitly. Anal-
ysis of such systems requires techniques like computing high-level properties
of states, generating states in a smart way, and using heuristics to decide
whether a state is likely to lead to a nal solution.
19.8 Variation in notation
State diagrams of various sorts, and constructs very similar to state diagrams,
are used in a wide range of applications. Therefore, there are many dierent
sets of terminology for equivalent and/or subtly dierent variations on this
idea. In particular, when a state diagram is viewed as an active device, i.e.
as a type of machine or computer, it is often called a state transition or an
automaton.
Start states are also known as initial states. End states can be called end
states or goal states.
There are two choices for setting up the transition function . State
diagrams where the transition function returns a set of states are called non-
deterministic. Those where returns a single state are called deterministic.
Deterministic state diagrams are less exible but easier to implement e-
ciently.
Chapter 20
Countability
This chapter covers innite sets and countability.
20.1 The rationals and the reals
Youre familiar with three basic sets of numbers: the integers, the rationals,
and the reals. The integers are obviously discrete, in that theres a big gap
between successive pairs of integers.
To a rst approximation, the rational numbers and the real numbers seem
pretty similar. The rationals are dense in the reals: if I pick any real number
x and a distance , there is always a rational number within distance of x.
Between any two real numbers, there is always a rational number.
We know that the reals and the rationals are dierent sets, because weve
shown that a few special numbers are not rational, e.g. and

2. However,
these irrational numbers seem like isolated cases. In fact, this intuition is
entirely wrong: the vast majority of real numbers are irrational and the
rationals are quite a small subset of the reals.
221
CHAPTER 20. COUNTABILITY 222
20.2 Completeness
One big dierence between the two sets is that the reals have a so-called
completeness property. It states that any subset of the reals with an upper
bound has a smallest upper bound. (And similarly for lower bounds.) So if I
have a sequence of reals that converges, the limit it converges to is also a real
number. This isnt true for the rationals. We can make a series of rational
numbers that converge (for example) such as
3, 3.1, 3.14, 3.141, 3.1415, 3.14159, 3.141592, 3.1415926, 3.14159265
But there is no rational number equal to .
In fact, the reals are set up precisely to make completeness work. One
way to construct the reals is to construct all convergent sequences of rationals
and add new points to represent the limits of these sequences. Most of the
machinery of calculus depends on the existence of these extra points.
20.3 Cardinality
We know how to calculate and compare the sizes of nite sets. To extend
this idea to innite sets, we use bijections functions to compare the sizes of
sets:
Denition: Two sets A and B have the same cardinality (|A| =
|B|) if and only if there is a bijection from A to B.
Weve seen that there is a bijection between two nite sets exactly when
the sets have the same number of elements. So this denition of cardinality
matches our normal notion of how big a nite set is. But, since we dont
have any numbers of innite size, working with bijections extends better to
innite sets.
The integers and the natural numbers have the same cardinality, because
we can construct a bijection between them. Consider the function f : N Z
where f(n) =
n
2
when n is even and f(n) =
(n+1)
2
when n is odd. f maps
CHAPTER 20. COUNTABILITY 223
the even natural numbers bijectively onto the non-negative integers. It maps
the odd natural numbers bijectively onto the negative integers.
Similarly, we can easily construct bijections between Z or N and various
of their innite subsets. For example, the formula f(n) = 3
n
creates a
bijection from the integers to the powers of 3. If S is any innite subset of
the natural numbers, we can number the elements of S in order of increasing
size: s
0
, s
1
, s
2
, . . .. This creates a bijection between S and N.
Because the integers are so important, theres a special name for sets that
have the same cardinality as the integers:
An innite set A is countably innite if there is a bijection from
N (or equivalently Z) onto A.
The term countable is used to cover both nite sets and sets that are count-
ably innite. All subsets of the integers are countable.
20.4 Cantor Schroeder Bernstein Theorem
For certain countably innite sets, its awkward to directly build a bijection
to the integers or natural numbers. Fortunately, a more general technique is
available. Remember that for nite sets, we could build a one-to-one function
from A to B if and only if |A| |B|. Using this idea, we can dene a partial
order on sets, nite or innite:
Denition: |A| |B| if and only if there is a one-to-one function
from A to B.
It is the case that if |A| |B| and |B| |A|, then |A| = |B|. That
is, if you can build one-to-one functions in both directions, a bijection does
exist. This result is called the Cantor Schroeder Bernstein Theorem.
1
It
allows us to do very slick 2-way bounding proofs that a wide range of sets
are countably innite.
1
See the Fendel and Resek book in the bibliography for a proof, and the Liebeck book
for a much simpler proof for the case where one of the sets is N.
CHAPTER 20. COUNTABILITY 224
For example, consider N
2
, the set of pairs of natural numbers. Its possible
to directly construct a bijection f : N
2
N, but the details are a bit messy.
Instead, lets build one-to-one functions in both directions. Easy direction
rst: dene f
1
: N N

by f
1
(n) = (n, 0). This is one-to-one, so |N| |N
2
|
(which you were probably prepared to consider obvious).
In the opposite direction, consider the following function: f
2
: N

N
such that f
2
(n, m) = 2
n
3
n
. This is one-to-one because prime factorizations
are unique. So |N
2
| |N|. Since we have one-to-one functions in both
directions, Cantor Schroeder Bernstein implies that |N
2
| = |N|. Therefore
N
2
is countably innite.
This construction can be extended to show the countability of any nite
Cartesian product of integers or natural numbers. E.g. the set of 7-tuples of
integers is countable. This also implies that a countable union of countable
sets is countable, because we can use pairs of natural numbers to index the
members of such a union. That is, the kth element of the jthe set in the
union would be associated with the element (j, k) in N
2
.
20.5 More countably innite sets
Suppose that we have a nite set M of characters. For example, M might
be the set of 26 upper-case alphabetic characters. Then the set M

contains
all character strings of various nite lengths, such as I, THIS, FINITE, and
RUMBLESEAT. It also contains the string of zero length. I claim that M

is
countably innite.
To prove this, well build one-to-one functions in both directions. First,
we can create a one-to-one function f from N to M

by mapping each natural

number n to the string consisting of n As. For example, f(0) = , f(2) = AA,
and f(5) = AAAAA.
In the other direction, notice that each letter in M has a 2-digit ASCII
code: A has the code 65, B is 66, and so on up to 90 for Z. We can translate
each string into a sequence of digits by replacing each letter with its ASCII
code. E.g. RUBY becomes 82856689. This doesnt work for the string , so
well translate it specially to the number 0. Weve now created a one-to-one
mapping from strings in A

to natural numbers.
CHAPTER 20. COUNTABILITY 225
We can also show that the non-negative rational numbers are countably
innite. Its easy to make a one-to-one function from the natural numbers
to the non-negative rational numbers: just map each natural number n to
itself. So |N| |Q

|. So now we just need a one-to-one function from the

non-negative rationals to the integers, to show that |Q

| |N|.
To map the non-negative rational numbers to the natural numbers, rst
map each rational number to one representative fraction, e.g. the one in
lowest terms. This isnt a bijection, but it is one-to-one. Then use the
method we saw above to map the fractions, which are just pairs of non-
negative integers, to the natural numbers. We now have the required one-to-
one function from the non-negative rationals to the natural numbers.
This construction can be adapted to also handle negative rational num-
bers. So the set of rational numbers is countably innite. And, more gener-
ally, any subset of the rationals is countable.
20.6 P(N) isnt countable
Before looking at the real numbers, lets rst prove a closely-related result
thats less messy: P(N) isnt countable. Recall that P(N) is the power set of
the natural numbers i.e. the set containing all subsets of the natural numbers.
Suppose that A is a nite set {a
0
, a
1
, a
2
, . . . , a
n
}. We can represent a
subset X of A as a bit vector {b
0
, b
1
, b
2
, . . . , b
n
} where b
i
is 1 if and only if
a
i
is in X. For example, if A = {7, 8, 9, 10, 11}, then the bit-vector 01100
would represent the subset {8, 9} and the bit-vector 10101 would represent
the subset {7, 9, 11}. Similarly, we can represent a subset of the natural
numbers as an innite-length bit vector.
Well use a procedure called diagonalization (due to Georg Cantor) to
show that its impossible to build a bijection from the natural numbers to
these innite bit vectors representing the subsets of the natural numbers
Suppose that there were such a bijection. Then we could put all the bit
vectors into a list, e.g. v
0
would be the rst bit vector, v
1
the second, and
so forth. Our list of bit vectors might look like this, where the kth column
contains the value of the kth digit (b
k
) for all the bit vectors:
CHAPTER 20. COUNTABILITY 226
b
0
b
1
b
2
b
3
b
4
b
5
b
6
b
7
b
8
b
9
. . .
v
0
1 1 0 1 1 0 1 1 1 1 . . .
v
1
1 1 0 0 1 0 1 1 0 0 . . .
v
2
0 0 0 0 1 0 0 1 0 0 . . .
v
3
0 1 1 1 1 0 1 0 0 0 . . .
v
4
0 0 0 0 1 1 1 0 1 1 . . .
v
5
1 1 1 0 1 0 1 0 0 1 . . .
. . . . . .
This is supposed to be a complete list of all the bit vectors. But we can
construct a bit vector x thats not on the list. The value of x
k
, i.e. the kth
bit in our new vector, will be 0 if the k digit of v
k
is 1, and 1 if the k digit
of v
k
is 0. Notice that x is dierent from v
3
because the two vectors dier
in the third position. It cant be v
20
because the two vectors dier in the
twentieth position. And, in general, x cant equal v
k
because the two vectors
dier in the kth position. For the example above, the new vector not in the
list would start out: 0 0 1 0 0 1 . . ..
So, its not possible to put these innite bit vectors into a list indexed
by the natural numbers, because we can always construct a new bit vector
thats not on the list. That is, there cant be a one-to-one function from the
innite bit vectors to the natural numbers. So there cant be a one-to-one
function from the subsets of the natural numbers to the natural numbers. So
P(N) isnt countable. That is, the subsets of the natural numbers are more
numerous than the natural numbers themselves.
20.7 More uncountability results
This same diagonalization trick can be used to show that various other sets
arent countable. For example, we can show that the real numbers arent
countable. To show this, we show that even the numbers in the interval
[0, 1] arent countable. Suppose that the elements of [0, 1] were countable.
Then we could put these real numbers into a list a
1
, a
2
, and so forth. Lets
write out a table of the decimal expansions of the numbers on this list. Now,
examine the digits along the diagonal of this table: a
11
, a
22
, etc. Suppose
we construct a new number b whose kth digit b
k
is 4 when a
kk
is 5, and 5
CHAPTER 20. COUNTABILITY 227
otherwise. Then b wont match any of the numbers in our table, so our table
wasnt a complete list of all the numbers in [0, 1]. So, [0, 1] is not countable,
and therefore the reals cant be countable.
Next, notice that an innite bit vector is a function from the natural
numbers to the set {0, 1}. So weve shown that there are uncountably many
functions from the natural numbers to {0, 1}. So there must be uncountably
many functions from the natural numbers to the natural numbers, or from
the integers to the integers.
Another generalization involves noticing that our diagonalization proof
doesnt really depend on any special properties of the natural numbers. So
it can be adapted to show that, if A is any set, P(A) has a (strictly) larger
cardinality than A. So, not only is P(N) larger than N, but P(P(N)) is even
larger. So there is a whole sequence of larger and larger innite cardinalities.
So, in particular, P(R) is larger than R. However, notice that R
2
has the
same cardinality as R. Lets consider a simpler version of this problem: [0, 1)
2
has the same cardinality as [0, 1). Any element of [0, 1)
2
can be represented
as two innite sequences of decimal digits: 0.a
1
a
2
a
3
a
4
. . . and 0.b
1
b
2
b
3
b
4
. . ..
We can map this to a single real number by interleaving the digits of the two
numbers: 0.a
1
b
1
a
2
b
2
a
3
b
3
. . .. This denes a bijection between the two sets.
This method can be adapted to create a bijection between all of R
2
and R.
20.8 Uncomputability
We can pull some of these facts together into some interesting consequences
for computer science. Notice that a formula for a function is just a nite
string of characters. So the set of formulas is countable. But the set of
functions, even from the integers to the integers, is uncountable. So there
are more functions than formulas, i.e. some functions which have no nite
formula.
Similarly, notice that a computer program is simply a nite string of
ASCII characters. So there are only countably many computer programs.
But there are uncountably many functions. So there are more functions
than programs, i.e. there are functions which cannot be computed by any
program.
CHAPTER 20. COUNTABILITY 228
A nal problem is created by the fact that, although the code for a com-
puter program is nite in length, the trace of the programs execution may
be innite. Specically, program traces fall into three categories
(1) The program eventually halts, so the trace is nite.
(2) The program loops, in the sense of returning back to a previous state.
(3) The program keeps going forever, consuming more and more storage
space.
The second type of behavior is like the decimal expansion of a rational
number: repeating. This third type of behavior is similar to a real numbers
decimal expansion: an innite sequence which doesnt repeat. As with num-
bers, the innite non-repeating sequences create interesting complexity. In
this case, the famous Halting Problem which shows that you cant build a
program that will decide whether other programs halt or not.
A nal example, further aeld but closely related mathematically, involves
tilings of the plane. Periodic tilings, which are made up of many copies of a
single repeated pattern, are like rational numbers or looping programs. Ape-
riodic tilings, whose existence was only proved in 1966, are like real numbers
or the third type of program: they go on forever, but not by repeating a sin-
gle pattern over and over. The original proof involved a set of 20,426 boring
square tiles. In the 1970s, Roger Penrose developed aperiodic sets of tiles
that were small (e.g. 2 tiles) and made pretty patterns. See the internet for
pictures.
Aside from looking cool, aperiodic tilings can have types of symmetries
that periodic tilings cant, e.g. 10-fold rotational symmetry. In 1982, a
material science professor named Dan Shechtman observed such symmetries
in electron diraction patterns. His discovery of these quasicrystals was
initially greeted with skepticism but eventually won him the Nobel prize in
2011.
CHAPTER 20. COUNTABILITY 229
20.9 Variation in notation
Authors dier as to whether the term countable includes nite sets or only
countably innite sets.
Chapter 21
Planar Graphs
This chapter covers special properties of planar graphs.
21.1 Planar graphs
A planar graph is a graph which can be drawn in the plane without any
edges crossing. Some pictures of a planar graph might have crossing edges,
but its possible to redraw the picture to eliminate the crossings. For example,
although the usual pictures of K
4
and Q
3
have crossing edges, its easy to
redraw them so that no edges cross. For example, a planar picture of Q
3
is
shown below. However, if you ddle around with drawings of K
3,3
or K
5
,
there doesnt seem to be any way to eliminate the crossings. Well see how
to prove that these two graphs arent planar.
230
CHAPTER 21. PLANAR GRAPHS 231
a
b
c
d
A B
C D
Why should we care? Planar graphs have some interesting mathematical
properties, e.g. they can be colored with only 4 colors. Also, as well see
later, we can use facts about planar graphs to show that there are only 5
Platonic solids.
There are also many practical applications with a graph structure in which
crossing edges are a nuisance, including design problems for circuits, subways,
utility lines. Two crossing connections normally means that the edges must
be run at dierent heights. This isnt a big issue for electrical wires, but
it creates extra expense for some types of lines e.g. burying one subway
tunnel under another (and therefore deeper than you would ordinarily need).
Circuits, in particular, are easier to manufacture if their connections live in
fewer layers.
21.2 Faces
When a planar graph is drawn with no crossing edges, it divides the plane into
a set of regions, called faces. By convention, we also count the unbounded
area outside the whole graph as one face. The boundary of a face is the
subgraph containing all the edges adjacent to that face and a boundary walk
is a closed walk containing all of those edges. The degree of the face is the
minimum length of a boundary walk. For example, in the gure below, the
lefthand graph has three faces. The boundary of face 2 has edges df, fe, ec, cd,
so this face has degree 4. The boundary of face 3 (the unbounded face) has
edges bd, df, fe, ec, ca, ab, so face 3 has degree 6.
CHAPTER 21. PLANAR GRAPHS 232
1 2 3
a
b
c
d
e
f
1 2
a
b
c
d
e
f
The righthand graph above has a spike edge sticking into the middle of
face 1. The boundary of face 1 has edges bf, fe, ec, cd, ca, ab. However, any
boundary walk must traverse the spike twice, e.g. one possible boundary
walk is bf, fe, ec, cd, cd, ca, ab, in which cd is used twice. So the degree of
face 1 in the righthand graph is 7. In such cases, it may help to think of
walking along inside the face just next to the boundary, rather walking along
the boundaries themselves. Notice that the boundary walk for such a face is
not a cycle.
Suppose that we have a graph with e edges, v nodes, and f faces. We
know that the Handshaking theorem holds, i.e. the sum of node degrees is
2e. For planar graphs, we also have a Handshaking theorem for faces: the
sum of the face degrees is 2e. To see this, notice that a typical edge forms
part of the boundary of two faces, one to each side of it. The exceptions are
edges, such as those involved in a spike, that appear twice on the boundary
of a single face.
Finally, for connected planar graphs, we have Eulers formula: ve+f =
2. Well prove that this formula works.
1
21.3 Trees
Before we try to prove Eulers formula, lets look at one special type of
planar graph: free trees. In graph theory, a free tree is any connected
graph with no cycles. Free trees are somewhat like normal trees, but they
dont have a designated root node and, therefore, they dont have a clear
ancestor-descendent ordering to their notes.
A free tree doesnt divide the plane into multiple faces, because it doesnt
1
You can easily generalize Eulers formula to handle graphs with more than one con-
nected components.
CHAPTER 21. PLANAR GRAPHS 233
contain any cycles. A free tree has only one face: the entire plane surrounding
it. So Eulers theorem reduces to v e = 1, i.e. e = v 1. Lets prove that
this is true, by induction.
Proof by induction on the number of nodes in the graph.
Base: If the graph contains no edges and only a single node, the
formula is clearly true.
Induction: Suppose the formula works for all free trees with up
to n nodes. Let T be a free tree with n + 1 nodes. We need to
show that T has n edges.
Now, we nd a node with degree 1 (only one edge going into it).
To do this start at any node r and follow a walk in any direction,
without repeating edges. Because T has no cycles, this walk cant
return to any node it has already visited. So it must eventually
hit a dead end: the node at the end must have degree 1. Call it
p.
Remove p and the edge coming into it, making a new free tree T

with n nodes. By the inductive hypothesis, T

has n 1 edges.
Since T has one more edge than T

, T has n edges. Therefore our

formula holds for T.
21.4 Proof of Eulers formula
We can now prove Eulers formula (v e + f = 2) works in general, for any
connected planar graph.
Proof: by induction on the number of edges in the graph.
Base: If e = 0, the graph consists of a single node with a single
face surrounding it. So we have 1 0 + 1 = 2 which is clearly
right.
Induction: Suppose the formula works for all graphs with no more
than n edges. Let G be a graph with n + 1 edges.
Case 1: G doesnt contain a cycle. So G is a free tree and we
already know the formula works for free trees.
CHAPTER 21. PLANAR GRAPHS 234
Case 2: G contains at least one cycle. Pick an edge p thats on a
cycle. Remove p to create a new graph G

.
Since the cycle separates the plane into two faces, the faces to
either side of p must be distinct. When we remove the edge p, we
merge these two faces. So G

has one fewer faces than G.

Since G

has n edges, the formula works for G

by the induction
hypothesis. That is v

+ f

= 2. But v

= v, e

= e 1, and
f

= f 1. Substituting, we nd that
v (e 1) + (f 1) = 2
So
v e +f = 2
21.5 Some corollaries of Eulers formula
Corollary 1 Suppose G is a connected planar graph, with v nodes, e edges,
and f faces, where v 3. Then e 3v 6.
Proof: The sum of the degrees of the faces is equal to twice the
number of edges. But each face must have degree 3. So we
have 3f 2e.
Eulers formula says that v e + f = 2, so f = e v + 2 and
thus 3f = 3e 3v + 6. Combining this with 3f 2e, we get
3e 3v + 6 2e. So e 3v 6.
We can also use this formula to show that the graph K
5
isnt planar. K
5
has ve nodes and 10 edges. This isnt consistent with the formula e 3v6.
Unfortunately, this method wont help us with K
3,3
, which isnt planar but
does satisfy this equation.
We can also use this Corollary 1 to derive a useful fact about planar
graphs:
CHAPTER 21. PLANAR GRAPHS 235
Corollary 2 If G is a connected planar graph, G has a node of degree less
than six.
Proof: This is clearly true if G has one or two nodes.
If G has at least three nodes, then suppose that the degree of
each node was at least 6. By the handshaking theorem, 2e equals
the sum of the degrees of the nodes, so we would have 2e 6v.
But corollary 1 says that e 3v 6, so 2e 6v 12. We cant
have both 2e 6v and 2e 6v 12. So there must have been a
node with degree less than six.
If our graph G isnt connected, the result still holds, because we can apply
our proof to each connected component individually. So we have:
Corollary 3 If G is a planar graph, G has a node of degree less than six.
21.6 K
3,3
is not planar
When all the cycles in our graph have at least four nodes, we can get a tighter
relationship between the numbers of nodes and edges.
Corollary 4 Suppose G is a connected planar graph, with v nodes, e edges,
and f faces, where v 3. and if all cycles in G have length 4, then
e 2v 4.
Proof: The sum of the degrees of the faces is equal to twice the
number of edges. But each face must have degree 4 because
all cycles have length 4. So we have 4f 2e, so 2f e.
Eulers formula says that v e + f = 2, so e v + 2 = f, so
2e 2v + 4 = 2f. Combining this with 2f e, we get that
2e 2v + 4 e. So e 2v 4.
This result lets us show that K
3,3
isnt planar. All the cycles in K
3,3
have
at least four nodes. But K
3,3
has 9 edges and 6 nodes, which isnt consistent
with this formula. So K
3,3
cant be planar.
CHAPTER 21. PLANAR GRAPHS 236
21.7 Kuratowskis Theorem
The two example non-planar graphs K
3,3
and K
5
werent picked randomly.
It turns out that any non-planar graph must contain a subgraph closely
related to one of these two graphs. Specically, well say that a graph G is
a subdivision of another graph F if the two graphs are isomorphic or if the
only dierence between the graphs is that G divides up some of Fs edges by
adding extra degree 2 nodes in the middle of the edges.
For example, in the following picture, the righthand graph is a subdivision
of the lefthand graph.
A B
C D
A B
C D
E
F G
We can now state our theorem precisely.
Claim 57 Kuratowskis Theorem: A graph is nonplanar if and only if it
contains a subgraph that is a subdivision of K
3,3
or K
5
.
This was proved in 1930 by Kazimierz Kuratowski, and the proof is ap-
parently somewhat dicult. So well just see how to apply it.
For example, heres a graph known as the Petersen graph (after a Danish
mathematician named Julius Petersen).
CHAPTER 21. PLANAR GRAPHS 237
A
B E
C D
a
b
e
c
d
This isnt planar. The oending subgraph is the whole graph, except for
the node B (and the edges that connect to B):
A
E
C D
a
b
e
c
d
This subgraph is a subdivision of K
3,3
. To see why, rst notice that the
node b is just subdividing the edge from d to e, so we can delete it. Or,
CHAPTER 21. PLANAR GRAPHS 238
formally, the previous graph is a subdivision of this graph:
A
E
C D
a
e
c
d
In the same way, we can remove the nodes A and C, to eliminate unnec-
essary subdivisions:
E
D
a
e
c
d
Now deform the picture a bit and we see that we have K
3,3
.
CHAPTER 21. PLANAR GRAPHS 239
E
D
a e
c
d
21.8 Coloring planar graphs
One application of planar graphs involves coloring maps of countries. Two
countries sharing a border
2
must be given dierent colors. We can turn this
into a graph coloring problem by assigning a graph node to each country.
We then connect two nodes with an edge exactly when their regions share a
border. This graph is called the dual of our original map. Because the maps
are planar, these dual graphs are always planar.
Planar graphs can be colored much more easily than other graphs. For
example, we can prove that they never require more than 6 colors:
Proof: by induction on the number of nodes in G.
Base: The planar graph with just one node has maximum degree
0 and can be colored with one color.
Induction: Suppose that any planar graph with < k nodes can
be colored with 6 colors. Let G be a planar graph with k nodes.
2
Two regions touching at a point are not considered to share a border.
CHAPTER 21. PLANAR GRAPHS 240
By Corollary 3, G has a node of degree less than 6. Lets pick
such a node and call it v.
Remove some node v (and its edges) from G to create a smaller
graph G

. G

is a planar graph with k 1 nodes. So, by the

inductive hypothesis, G

can be colored with 5 colors.

Because v has less than 6 neighbors, its neighbors are only using
5 of the available colors. So there is a spare color to assign to v,
nishing the coloring of G.
Its not hard, but a bit messy, to upgrade this proof to show that planar
graphs require only ve colors. Four colors is much harder. Way back in
1852, Francis Guthrie hypothesized that any planar graph could be colored
with only four colors, but it took 124 years to prove that he was right. Alfred
Kempe thought he had proved it in 1879 and it took 11 years for another
mathematician to nd an error in his proof.
The Four Color Theorem was nally proved by Kenneth Appel and Wolf-
gang Haken at UIUC in 1976. They reduced the problem mathematically, but
were left with 1936 specic graphs that needed to be checked exhaustively,
using a computer program. Not everyone was happy to have a computer
involved in a mathematical proof, but the proof has come to be accepted as
legitimate.
21.9 Application: Platonic solids
A fact dating back to the Greeks is that there are only ve Platonic solids:
cube, dodecahedron, tetrahedron, icosahedron, octahedron. These are con-
vex polyhedra whose faces all have the same number of sides (k) and whose
nodes all have the same number of edges going into them (d).
To turn a Platonic solid into a graph, imagine that its made of a stretchy
material. Make a small hole in one face. Put your ngers into that face and
pull sideways, stretching that face really big and making the whole thing
at. For example, an octahedron (8 triangular sides) turns into the following
graph. Notice that it still has eight faces, one for each face of the original
solid, each with three sides.
CHAPTER 21. PLANAR GRAPHS 241
Graphs of polyhedra are slightly special planar graphs. Polyhedra arent
allowed to have extra nodes partway along edges, so each node in the graph
must have degree at least three. Also, since the faces must be at and the
edges straight, each face needs to be bounded by at least three edges. So,
if G is the graph of a Platonic solid, all the nodes of G must have the same
degree d 3 and all faces must have the same degree k 3.
Now, lets do some algebra to see why there are so few possibilities for
the structure of such a graph.
By the handshaking theorem, the sum of the node degrees is twice the
number of edges. So, since the degrees are equal to d, we have dv = 2e and
therefore
v =
2e
d
By the handshaking theorem for faces, the sum of the face degrees is also
twice the number of edges. That is kf = 2e. So
f =
2e
k
Eulers formula says that v e+f = 2, so v +f = 2+e > e. Substituting
the above equations into this one, we get:
2e
d
+
2e
k
> e
CHAPTER 21. PLANAR GRAPHS 242
Dividing both sides by 2e:
1
d
+
1
k
>
1
2
If we analyze this equation, we discover that d and k cant both be larger
than 3. If they were both 4 or above, the left side of the equation would be

1
2
. Since we know that d and k are both 3, this implies that one of the
two is actually equal to three and the other is some integer that is at least 3.
Suppose we set d to be 3. Then the equation becomes
1
3
+
1
k
>
1
2
. So
1
k
>
1
6
, which means that k cant be any larger than 5. Similarly, if k is 3,
then d cant be any larger than 5.
This leaves us only ve possibilities for the degrees d and k: (3, 3), (3, 4),
(3, 5), (4, 3), and (5, 3). Each of these corresponds to one of the Platonic
solids.
Appendix A
Jargon
Mathematicians write in a dialect of English that diers from everyday En-
glish and from formal scientic English. To read and write mathematics
uently, you need to be aware of the dierences.
A.1 Strange technical terms
Many technical terms, abbreviations, and shorthand symbols are easy to
track down. Perhaps they belong to an obvious topic area. Perhaps they
are used so heavily that no one forgets what they mean. Perhaps they are
easily found on Wikipedia. But others seem to appear mysteriously out of
nowhere.
abuse of notation The author is using a notation that doesnt technically
comply with the rules of mathematical writing, but is very convenient or
conventional. For example, using if in a a denition where we really
mean if and only if. Or writing f(2, 3) for the result of applying f
to the pair (2, 3). Strictly we should have written f((2, 3)), but no one
except a computer program would ever do that.
! This is the unique existence quantier. !x, P(x) means that there is
exactly one x for which P(x) holds.
By symmetry A variant wording for without loss of generality.
243
APPENDIX A. JARGON 244
IH Inductive hypothesis, as in a proof by induction.
QED Short for quod erat demonstrandum. This is just the Latin transla-
tion of which is what we needed to show. The Latin and the English
versions are both polite ways to tell your reader that the proof is n-
ished.
tuple A generalization of pair, triple, etc to larger numbers of items.
A k-tuple is an ordered sequence of k items. One occasionally hears
2-tuple in place of pair.
NTS Need to show, as in we need to show that all horses have four legs.
WLOG, WOLOG, without loss of generality Were stipulating an ex-
tra fact or relationship, but claiming that it doesnt actually add any
new information to the problem. For example, if we have two real vari-
ables x and y, with identical properties so far, its ok to stipulate that
x y because were simply xing which of the two names refers to the
smaller number.
A.2 Odd uses of normal words
In addition to the obvious technical terms (e.g. rational number), some
words are used rather dierently in mathematics than in everyday English.
Clearly See obviously.
Consider The author is about to pull an example or constant or the like out
of thin air. Typically, this part of the proof was constructed backwards,
and the reasons for this step will not become apparent until later in
the proof.
Distinct Not equal. For example, if x and y are chosen to be distinct
elements of a set A, then x and y arent allowed to have the same
value.
General Shorhand for general purpose, i.e. works for all inputs or in all
circumstances.
APPENDIX A. JARGON 245
Graph Might be a picture showing the relationship between two variables,
or it might be a diagram with nodes and links between them.
Obviously Perhaps it really is obvious. Or perhaps the author didnt feel
like working out the details (or forgot them in the heat of lecture). Or,
infamously, perhaps he just wants to intimidate the audience.
Or The connective or and its variations (e.g. either...or) leave open the
possibility that both statements are true. If a mathematician means to
exclude the possibility that both are true (exclusive or), they say so
explicitly. In everyday English, you need to use context to determine
whether an or was meant to be inclusive or exclusive.
Has the form E.g. if x is rational, then x has the form
p
q
where p, q Z
and q = 0. Means that the denition of this type of object forces the
object to have this specic internal structure.
Plainly See obviously.
Proper For example, proper subset or proper subgraph. Used to ex-
clude the possibility of equality. Very similar to strict.
Recall The author is about to tell you something basic that you probably
ought to know, but he realizes that some of you have gaps in your
background or have forgotten things. Hes trying to avoid oending
some of the audience by suggesting they dont know something so basic,
while not embarrassing everyone else by making them confess that they
dont.
Similarly You can easily ll in these details by adapting a previous part of
the proof, and youll just get bored if I spell them out. Occasionally
misused in a manner similar to obviously.
Strict A strict relationship is one that excludes the possibility of equality.
So strictly less than means that the two items cant be equal. A
strictly increasing function never stays on the same value. Compare
proper.
Suppose not A stereotypical way to start a proof by contradiction. Its
shorthand for Suppose that the claim is not true.
APPENDIX A. JARGON 246
Unique There is only one item with the specied properties. There is a
unique real number whose square is zero.
Vacuously true The claim is true, but only because its impossible to sat-
isfy the hypothesis. Recall that in math, an if/then statement is con-
sidered true if its hypothesis is true. Vacuously true statements often
occur when denitions are applied to examples that are very small (e.g.
the empty set) or lacking some important feature (e.g. a graph with
no edges, a function whose domain is the empty set).
Well-dened A mathematical object is well-dened if its denition isnt
buggy. The term usually appears in a context where the denition
might look ok at rst glance but could have a subtle bug.
A.3 Constructions
Mathematicians also use certain syntactic constructions in ways that arent
quite the same as in normal English.
If/then statements An if/then statement with a false hypothesis is con-
sidered true in mathematics. In normal English, such statements are
rare and its not clear if they are true, false, or perhaps neither.
Imperatives A mathematical construction is often written as a recipe that
the author and reader are carrying out together. E.g. We can nd a
real number x such that x
2
< 47. The picture behind this is that the
author is helping you do the mathematics, not doing it by himself while
you watch. Verbs in the imperative form are sometimes actions that
the reader should do by himself (e.g. prove XX on a problem set) but
sometimes commands that the reader should do by following explicit
instructions from the author. E.g. dene f : N N by f(n) = 2n
tells you to dene a function, but you are being told exactly how to
do it. On a problem set, this would be background information, not a
question you need to answer.
Two variables If a mathematician sets up two variables with dierent names,
e.g. x and y, he leaves open the possibility that they might be equal. If
APPENDIX A. JARGON 247
they are intended to have dierent values, they will be explicitly speci-
ed as distinct. In normal English, giving two things dierent names
suggests very strongly that they are dierent objects.
No resetting Do not change the value of a variable in the middle of a proof.
Instead, use a fresh variable name. For example, suppose you know that
p = 3jk + 1 in a proof about divisibility. To simplify the form of the
equation, its tempting to reset j to be three times its original value,
so that you have p = jk +1. This isnt done in standard mathematical
writing. Instead, use a fresh variable name, e.g. s = 3j and then
p = sk + 1.
Variable names can be re-used when the context changes (e.g. youve
started a new proof). When working through a series of examples, e.g.
a bunch of example functions to be discussed, rotate variable names, so
that the new denition appears after the reader has had time to forget
the old denition.
Variable names Variable names are single-letter, e.g. f but not foo. The
letter can be adorned with a wide variety of accents, subscripts, and
superscripts, however.
A.4 Unexpectedly normal
Mathematicians are, of course, underlyingly speakers of normal English. So,
sometimes, they apply the normal rules of conversational English even when
these arent part of the formal mathematical system.
names of variables In principle, any letter can be used as the name for any
variable you need in a proof, with or without accent marks of various
sorts. However, there are strong conventions favoring certain names.
E.g. x is a real number, n is an integer, f is a function, and T is a set.
Observe what names authors use in your specic topic area and dont
stray too far from their conventions.
topics of equations Proofs are conversations about particular objects (e.g.
polynomials, summations). Therefore, equations and other statements
APPENDIX A. JARGON 248
are often about some specic topic object, whose properties are being
described. When this is the case, the topic belongs on the lefthand
side, just like the subject of a normal sentence.
Appendix B
Acknowledgements and
Supplementary Readings
Most of the basic ideas and examples in this book date back many years and
their original source is almost impossible to trace. Ive consulted so many
books and worked with so many helpful teachers, students, and course sta
over the years that the details have blended into a blur. Nevertheless, certain
books have had a particularly strong inuence on my presentation. Most of
them would make excellent supplementary materials, both for instructors
and for students.
Its impossible to even think about discrete mathematics without citing
the classic encyclopedic text by Rosen [Rosen 2007]. It helped me under-
stand what topics would be most relevant to a computer science, as op-
posed to a mathematics, audience. Biggs [Biggs 2002], Matousek and Nesetril
[Matousek and Nesetril 1998] and West [West 2001] have been invaluable ref-
erences for discrete mathematics and graph theory.
From Liebeck [Liebeck 2006], Sipser [Sipser 2006], and Biggs [Biggs 2002],
I learned how to extract core concepts and present them concisely. From
Fendel and Resek [Fendel and Resek 1990], I learned how to explain proof
mechanics and talk about the process of constructing proofs. From my nu-
merous students and course sta, both at Iowa and Illinois, I have learned
to understand why some of these concepts seem so hard to beginners. My
co-instructors Eric Shaer and Viraj Kumar helped mould the curriculum to
249
APPENDIX B. ACKNOWLEDGEMENTS AND SUPPLEMENTARY READINGS250
the needs of our students. Sariel Har-Peled has provided many interesting
examples.
Finally, a special individual citation goes to Je Erickson for the recur-
sion fairy (though my fairy has a slightly dierent job than his) and for
extended arguments about the proper way to present induction (though we
still disagree).
Bibliography
[Biggs 2002] Norman L. Biggs (2002) Discrete Mathematics, second edition,
Oxford University Press, Oxford, UK.
[Fendel and Resek 1990] Daniel Fendel and Diane Resek (1990) Foundations
of Higher Mathematics: Exploration and Proof, Addison-Wesley, Reading
MA.
[Liebeck 2006] Martin Liebeck (2006) A Concise Introduction to Pure Math-
ematics, Chapman Hall/CRC, Boca Raton, FL.
[Matousek and Nesetril 1998] Jiri Matousek and Jaroslav Nesetril (1998) In-
vitation to Discrete Mathematics Oxford University Press.
[Rosen 2007] Kenneth H. Rosen (2007) Discrete Mathematics and Its Appli-
cations, sixth edition, McGraw Hill, New York, NY.
[Sipser 2006] Michael Sipser, Introduction to the Theory of Computation, sec-
ond edition, Thomson, Boston MA.
[West 2001] Douglas B. West (2001) Introduction to Graph Theory, second
edition, Prentice-Hall, Upper Saddle River, NJ.
251
Appendix C
Where did it go?
When teaching the rst few chapters of this book, you may be disconcerted to
nd some important topics apparently missing or breezed through absurdly
fast. In many cases, whats happening is that the missing topics are covered
later. This guide will help you get oriented.
It is traditional to cover most of logic and proof techniques right at the
start of the course. And, similarly, completely cover each topics (e.g. re-
lations) as it appears. This may seem intellectually tidy and it does work
for the small minority who are future theoreticians. However, this method
overwhelms most of the class and they get only a supercial understanding
of key methods. In this book, the most dicult concepts and techniques are
staged in gradually.
Counting material has been integrated with three other topics: Chapter 5
(Sets), Chapter 8 (Functions and One-to-One), and Chapter 18(Sets of Sets).
Chapters 2-3 (Proofs and Logic). Vacuous truth is in Chapter 5 (Sets).
Nested quantiers are in Chapter 7 (Functions and Onto). Connections to
circuits are in Chapter 16(NP). Proof by contradiction is in Chapter 17(Proof
by Contradiction). The non-traditional topic of 2-way bounding is in Chapter
10.
Chapter 4 (Number theory). Equivalence Classes are done only pre-
formally at this stage. Formal details are gradually added in Chapter 6
(Relations) and Chapter 18(Sets of Sets).
252
APPENDIX C. WHERE DID IT GO? 253
Chapter 5 (Sets). Proving a set equality by proving two inequalities is in
Chapter 10 (2-way Bounding). Sets nested inside other sets are covered in
Chapter 18(Sets of sets).
Chapter 6 (Relations). Formal details of equivalence relations will be
covered in Chapter 18(Sets of Sets).
Chapters 7-8 (Functions). Set-valued functions are in Chapter 18(Sets of
Sets). The pairs denition of functions is in Chapter 19(State Diagrams).
Chapter 9 (Graphs). Graph in this book means undirected. Graphs are
introduced early and used as examples in many later chapters. It returns as
a main topic in Chapter 21(Planar Graphs). Directed graphs are covered in
Chapter 6 (Relations) and Chapter 19(State Diagrams). The treatment of
graphs is deliberately representative, not exhaustive.
Chapter 11(Induction). This chapter introduces the basic idea, which is
further developed in Chapter 12(Recursive Denition) and Chapter 13(Trees).
Induction involving inequalities is delayed until Chapter 14(Big-O). This or-
ganization lets us stage induction practice over several problem sets, easier
examples rst.