DISCRETE MATHEMATICS

GRAPH ALGORITHMS, ALGEBRAIC STRUCTURES,

CODING THEORY, AND CRYPTOGRAPHY

Sriraman Sridharan and R. Balakrishnan

A Chapman & Hall Book

Discrete Mathematics
Graph Algorithms, Algebraic
Structures, Coding Theory, and
Cryptography
Discrete Mathematics
Graph Algorithms, Algebraic
Structures, Coding Theory, and
Cryptography

R. Balakrishnan
Bharathidasan University, Tiruchirappalli, Tamil Nadu, INDIA

Sriraman Sridharan
Laboratoire LAMPS, Département de Mathématiques et
d’Informatique, Université de Perpignan Via Domitia,
Perpignan, FRANCE
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742


c 2020 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-0-8153-4739-2 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access
www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc.
(CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization
that provides licenses and registration for a variety of users. For organizations that have been granted
a photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and
are used only for identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Names: Sridharan, Sriraman, author. | Balakrishnan, R. (Rangaswami), author.

Title: Discrete mathematics : graph algorithms, algebraic structures, coding
theory, and cryptography / Sriraman Sridharan, R. Balakrishnan.
Description: Boca Raton : CRC Press, Taylor & Francis Group, 2019. | Includes
bibliographical references and index.
Identifiers: LCCN 2019011934| ISBN 9780815347392 (hardback : alk. paper) |
ISBN 9780429486326 (ebook)
Subjects: LCSH: Mathematics--Textbooks. | Computer science--Textbooks.
Classification: LCC QA39.3 .S7275 2019 | DDC 511/.1--dc23
LC record available at https://lccn.loc.gov/2019011934

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
 The first author, R. Balakrishnan, dedicates this book
affectionately to his granddaughters:
Amirtha and Aparna

The second author, Sriraman Sridharan, dedicates this book in

memory of his thesis supervisor:
Professor Claude Berge
Contents

List of Figures xiii

List of Tables xvii

Preface xix

Acknowledgment xxiii

Authors xxv

1 Graph Algorithms I 1
1.1 Representation of Graphs . . . . . . . . . . . . . . . . . . . . 1
1.2 Minimum Spanning Tree Algorithms . . . . . . . . . . . . . . 9
1.2.1 Prim’s minimum spanning tree algorithm . . . . . . . 11
1.2.2 Kruskal’s minimum spanning tree algorithm . . . . . . 19
1.2.3 Rooted ordered trees and traversal of trees . . . . . . 22
1.3 Shortest Path Algorithms . . . . . . . . . . . . . . . . . . . . 23
1.3.1 Single-source shortest path algorithm . . . . . . . . . 24
1.4 Dijkstra’s Algorithm for Negative Weighted Arcs . . . . . . . 33
1.5 All-Pairs Shortest Path Algorithm . . . . . . . . . . . . . . . 35
1.5.1 An application of Floyd’s algorithm . . . . . . . . . . 43
1.6 Transitive Closure of a Directed Graph . . . . . . . . . . . . 45
1.7 An O(n3 ) Transitive Closure Algorithm Due to Warshall . . 47
1.8 Navigation in Graphs . . . . . . . . . . . . . . . . . . . . . . 50
1.9 Applications of Depth-First Search . . . . . . . . . . . . . . . 55
1.9.1 Application 1: Finding connected components . . . . . 55
1.9.2 Application 2: Testing acyclic graph . . . . . . . . . . 56
1.9.3 Application 3: Finding biconnected components of a
connected multigraph . . . . . . . . . . . . . . . . . . 58
1.10 Depth-First Search for Directed Graphs . . . . . . . . . . . . 68
1.11 Applications of Depth-First Search for Directed Graphs . . . 70
1.11.1 Application 1: Finding the roots of a directed
graph . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
1.11.2 Application 2: Testing if a digraph is without
circuits . . . . . . . . . . . . . . . . . . . . . . . . . . 72
1.11.3 Application 3: Topological sort . . . . . . . . . . . . . 72
1.11.3.1 An application of topological sort: PERT . . 76

vii
viii Contents

1.11.4 Application 4: Strongly connected components

algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 79
1.12 Traveling Salesman Problem . . . . . . . . . . . . . . . . . . 90
1.12.1 Approximate algorithms for traveling salesman
problem . . . . . . . . . . . . . . . . . . . . . . . . . . 93
1.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

2 Graph Algorithms II 103

2.1 Breadth-First Search . . . . . . . . . . . . . . . . . . . . . . 103
2.2 Applications of bfs Algorithm . . . . . . . . . . . . . . . . . 107
2.3 Matchings in Graphs . . . . . . . . . . . . . . . . . . . . . . 113
2.3.1 An application: (k − 1)-regular subgraphs of k-regular
graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2.4 Matrices and Bipartite Graphs . . . . . . . . . . . . . . . . . 127
2.4.1 Personnel assignment problem or weighted matching
in a bipartite graph . . . . . . . . . . . . . . . . . . . 136
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

3 Algebraic Structures I (Matrices, Groups, Rings,

and Fields) 147
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3.2 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3.3 Operations on Matrices: Addition, Scalar Multiplication,
and Multiplication of Matrices . . . . . . . . . . . . . . . . . 147
3.3.1 Block multiplication of matrices . . . . . . . . . . . . 148
3.3.2 Transpose of a matrix . . . . . . . . . . . . . . . . . . 149
3.3.3 Inverse of a matrix . . . . . . . . . . . . . . . . . . . 149
3.3.4 Symmetric and skew-symmetric matrices . . . . . . . 150
3.3.5 Hermitian and skew-Hermitian matrices . . . . . . . . 150
3.3.6 Orthogonal and unitary matrices . . . . . . . . . . . . 151
3.3.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.4 Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.4.1 Abelian and non-Abelian groups . . . . . . . . . . . . 152
3.4.2 Examples of Abelian groups . . . . . . . . . . . . . . . 154
3.4.3 Examples of non-Abelian groups . . . . . . . . . . . . 155
3.4.4 Group tables . . . . . . . . . . . . . . . . . . . . . . . 155
3.5 A Group of Congruent Transformations (Also called
Symmetries) . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.6 Another Group of Congruent Transformations . . . . . . . . 157
3.7 Subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
3.7.1 Examples of subgroups . . . . . . . . . . . . . . . . . . 158
3.7.2 Subgroup generated by a subset of a group . . . . . . 158
3.8 Cyclic Groups . . . . . . . . . . . . . . . . . . . . . . . . . . 159
3.8.1 Examples of cyclic groups . . . . . . . . . . . . . . . . 159
3.9 Lagrange’s Theorem for Finite Groups . . . . . . . . . . . . 163
3.10 Homomorphisms and Isomorphisms of Groups . . . . . . . . 166
 Contents ix

3.11 Properties of Homomorphisms of Groups . . . . . . . . . . . 168

3.12 Automorphism of Groups . . . . . . . . . . . . . . . . . . . . 170
3.13 Normal Subgroups . . . . . . . . . . . . . . . . . . . . . . . . 171
3.14 Quotient Groups (or Factor Groups) . . . . . . . . . . . . . . 172
3.15 Basic Isomorphism Theorem for Groups . . . . . . . . . . . . 174
3.15.1 Examples of factor groups . . . . . . . . . . . . . . . . 175
3.16 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.17 Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
3.17.1 Rings, definitions and examples . . . . . . . . . . . . . 178
3.17.1.1 Unity element of a ring . . . . . . . . . . . . 180
3.17.2 Units of a ring . . . . . . . . . . . . . . . . . . . . . . 180
3.17.2.1 Units of the ring Zn . . . . . . . . . . . . . . 180
3.17.2.2 Zero divisors . . . . . . . . . . . . . . . . . . 181
3.18 Integral Domains . . . . . . . . . . . . . . . . . . . . . . . . 181
3.19 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
3.20 Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
3.21 Principal Ideals . . . . . . . . . . . . . . . . . . . . . . . . . 184
3.22 Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
3.22.1 Examples of fields . . . . . . . . . . . . . . . . . . . . 188
3.23 Characteristic of a Field . . . . . . . . . . . . . . . . . . . . 188

4 Algebraic Structures II (Vector Spaces and Finite Fields) 191

4.1 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.1.1 Examples of vector spaces . . . . . . . . . . . . . . . . 192
4.2 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
4.2.1 An example of a subspace . . . . . . . . . . . . . . . . 193
4.3 Spanning Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 193
4.4 Linear Independence of Vectors . . . . . . . . . . . . . . . . 195
4.5 Bases of a Vector Space . . . . . . . . . . . . . . . . . . . . . 197
4.6 Dimension of a Vector Space . . . . . . . . . . . . . . . . . . 198
4.7 Solutions of Linear Equations and Rank of a Matrix . . . . . 201
4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
4.9 Solutions of Linear Equations . . . . . . . . . . . . . . . . . 205
4.10 Solutions of Non-Homogeneous Linear Equations . . . . . . 206
4.11 LUP Decomposition . . . . . . . . . . . . . . . . . . . . . . . 208
4.11.1 Computing an LU decomposition . . . . . . . . . . . . 208
4.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
4.13 Finite Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
4.14 Factorization of Polynomials Over Finite Fields . . . . . . . 220
4.15 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
4.16 Mutually Orthogonal Latin Squares . . . . . . . . . . . . . . 223

5 Introduction to Coding Theory 225

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
5.2 Binary Symmetric Channels . . . . . . . . . . . . . . . . . . 226
5.3 Linear Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
x Contents

5.4 Minimum Distance of a Code . . . . . . . . . . . . . . . . . 230

5.5 Hamming Codes . . . . . . . . . . . . . . . . . . . . . . . . 231
5.6 Standard Array Decoding . . . . . . . . . . . . . . . . . . . . 232
5.7 Sphere Packings . . . . . . . . . . . . . . . . . . . . . . . . . 235
5.8 Extended Codes . . . . . . . . . . . . . . . . . . . . . . . . . 236
5.9 Syndrome Decoding . . . . . . . . . . . . . . . . . . . . . . . 237
5.10 Error Detection . . . . . . . . . . . . . . . . . . . . . . . . . 238
5.11 Sphere Packing Bound or Hamming Bound . . . . . . . . . . 238
5.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
5.13 Cyclic Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
5.14 Dual Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
5.15 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

6 Cryptography 249
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.2 Some Classical Cryptosystems . . . . . . . . . . . . . . . . . 249
6.2.1 Caesar cryptosystem . . . . . . . . . . . . . . . . . . . 249
6.2.2 Affine cryptosystem . . . . . . . . . . . . . . . . . . . 250
6.2.3 Private key cryptosystems . . . . . . . . . . . . . . . . 251
6.2.4 Hacking an affine cryptosystem . . . . . . . . . . . . . 252
6.3 Encryption Using Matrices . . . . . . . . . . . . . . . . . . . 253
6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.5 Other Private Key Cryptosystems . . . . . . . . . . . . . . . 255
6.5.1 Vigenere cipher . . . . . . . . . . . . . . . . . . . . . . 255
6.5.2 The one-time pad . . . . . . . . . . . . . . . . . . . . 256
6.6 Public Key Cryptography . . . . . . . . . . . . . . . . . . . . 256
6.6.1 Working of public key cryptosystems . . . . . . . . . . 257
6.6.1.1 Transmission of messages . . . . . . . . . . . 257
6.6.1.2 Digital signature . . . . . . . . . . . . . . . . 257
6.6.2 RSA public key cryptosystem . . . . . . . . . . . . . . 258
6.6.2.1 Description of RSA . . . . . . . . . . . . . . 258
6.6.3 The ElGamal public key cryptosystem . . . . . . . . . 260
6.6.4 Description of ElGamal system . . . . . . . . . . . . . 261
6.7 Primality Testing . . . . . . . . . . . . . . . . . . . . . . . . 261
6.7.1 Non-trivial square roots (mod n) . . . . . . . . . . . . 261
6.7.2 Prime Number Theorem . . . . . . . . . . . . . . . . . 262
6.7.3 Pseudo-primality testing . . . . . . . . . . . . . . . . . 262
6.7.3.1 Base-2 Pseudo-prime test . . . . . . . . . . . 263
6.7.4 Miller-Rabin Algorithm . . . . . . . . . . . . . . . . . 263
6.7.5 Horner’s method to evaluate a polynomial . . . . . . . 263
6.7.6 Modular exponentiation algorithm based on repeated
squaring . . . . . . . . . . . . . . . . . . . . . . . . . . 265
6.8 The Agrawal-Kayal-Saxena (AKS) Primality Testing
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
6.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 267
 Contents xi

6.8.2 The basis of AKS algorithm . . . . . . . . . . . . . . . 267

6.8.3 Notation and preliminaries . . . . . . . . . . . . . . . 268
6.8.4 The AKS algorithm . . . . . . . . . . . . . . . . . . . 269

Appendix A: Answers to Chapter 1—Graph Algorithms I 279

Appendix B: Answers to Chapter 2—Graph Algorithms II 287

Appendix C: Answers to Chapter 3—Algebraic Structures I 289

Appendix D: Answers to Chapter 4—Algebraic Structures II 295

Appendix E: Answers to Chapter 5—Introduction

to Coding Theory 299

Appendix F: Answers to Chapter 6—Cryptography 303

Bibliography 307

Index 311
List of Figures

1.1 Geometric representation of a graph. . . . . . . . . . . . . . 2

1.2 A multigraph. . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Incidence matrix of a graph. . . . . . . . . . . . . . . . . . . 4
1.4 A graph and its adjacency lists. . . . . . . . . . . . . . . . . 5
1.5 A multigraph and its adjacency lists. . . . . . . . . . . . . . 6
1.6 Linked lists representation of graph of Figure 1.5. . . . . . . 6
1.7 A tree. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.8 A weighted graph G and its spanning tree T . . . . . . . . . 12
1.9 A minimum spanning tree T  of G. . . . . . . . . . . . . . . 14
1.10 A step in the proof of Prim’s algorithm. . . . . . . . . . . . 16
1.11 Kruskal’s minimum spanning tree T  of G. . . . . . . . . . . 20
1.12 Recursive definition of a rooted tree. . . . . . . . . . . . . . 22
1.13 Illustration of tree traversals. . . . . . . . . . . . . . . . . . 23
1.14 A directed weighted graph G. . . . . . . . . . . . . . . . . . 24
1.15 A minimum spanning arborescence A of Figure 1.14. . . . . 25
1.16 A step in the proof of Dijkstra’s algorithm. . . . . . . . . . . 28
1.17 Illustration of Dijkstra’s algorithm for negative weighted arcs. 33
1.18 A step illustrating Floyd’s algorithm. . . . . . . . . . . . . . 37
1.19 A digraph on which Floyd’s algorithm is executed. . . . . . 39
1.20 A tree illustrating different recursive call interpath(4,2). . . 43
1.21 Graph illustrating eccentricities, radius and diameter. . . . 43
1.22 A graph to illustrate transitive closure. . . . . . . . . . . . . 46
1.23 The transitive closure of the graph of Figure 1.22. . . . . . . 46
1.24 A simple graph to illustrate depth-first search. . . . . . . . 51
1.25 Illustration of depth-first search algorithm. . . . . . . . . . . 52
1.26 Illustration of cut points and bridges. . . . . . . . . . . . . . 59
1.27 Illustration of biconnected components of a graph. . . . . . 60
1.28 Depth-first search of the graph of Figure 1.27. . . . . . . . . 61
1.29 Digraph to illustrate the depth-first search. . . . . . . . . . . 69
1.30 Depth-first search of graph of Figure 1.29. . . . . . . . . . . 69
1.31 Illustration to test if a graph is acyclic. . . . . . . . . . . . . 72
1.32 A graph to illustrate topological sort. . . . . . . . . . . . . . 73
1.33 Topological sorting of the graph of Figure 1.32. . . . . . . . 73
1.34 An activity digraph. . . . . . . . . . . . . . . . . . . . . . . 77
1.35 A graph and its strongly connected components. . . . . . . . 80

xiii
xiv List of Figures

1.36 Illustration of strong components algorithm. . . . . . . . . . 81

1.37 dfs: strong components algorithm. . . . . . . . . . . . . . . . 82
1.38 Illustrating the traveling salesman problem. . . . . . . . . . 91
1.39 Illustration of Fleury’s algorithm. . . . . . . . . . . . . . . . 97
1.40 Spanning tree of Figure 1.38. . . . . . . . . . . . . . . . . . 98
1.41 Doubling the edges of the tree of Figure 1.40. . . . . . . . . 98
1.42 A weighted simple graph. . . . . . . . . . . . . . . . . . . . . 100
1.43 A weighted directed graph. . . . . . . . . . . . . . . . . . . . 101
1.44 A simple connected graph. . . . . . . . . . . . . . . . . . . . 102
2.1 A graph G and its breadth-first search, bfs(G). . . . . . . . 104
2.2 Breadth-first search tree of graph G of Figure 2.1. . . . . . . 104
2.3 An example of a geodetic graph. . . . . . . . . . . . . . . . . 107
2.4 Algorithm to test if a connected graph is bipartite. . . . . . 111
2.5 Finding a bipartition of a bipartite graph G of Figure 2.5. . 112
2.6 Illustration of matchings. . . . . . . . . . . . . . . . . . . . . 113
2.7 A 3-regular graph without a perfect matching. . . . . . . . . 114
2.8 Illustration of the dancing problem. . . . . . . . . . . . . . . 116
2.9 A graph and its alternating paths. . . . . . . . . . . . . . . 117
2.10 A graph G and the spanning subgraph induced by
M1 ΔM2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2.11 An outline of maximum matching algorithm. . . . . . . . . . 120
2.12 A graph to illustrate matching algorithm. . . . . . . . . . . 121
2.13 Algorithm to find a (k − 1)-regular subgraph in a k-regular
graph with no induced star K1,3 . . . . . . . . . . . . . . . . 122
2.14 A perfect matching in bipartite graph. . . . . . . . . . . . . 125
2.15 A graph to illustrate the idea of a succinct certificate. . . . . 125
2.16 A bipartite graph and its corresponding (0,1) matrix. . . . . 127
2.17 A bipartite graph G and a partial M -alternating tree. . . . 134
2.18 Informal algorithm for finding an M -augmenting path. . . . 135
2.19 Algorithm to find a matching that saturates all the vertices
of the set X1 or else a succinct certificate. . . . . . . . . . . 136
2.20 The graph Gl , where l is the good label. . . . . . . . . . . . 140
2.21 An M -alternating tree A of Gl . . . . . . . . . . . . . . . . . 141
2.22 An M -alternating tree A1 of Gl1 . . . . . . . . . . . . . . . . 142
2.23 The graph Gl2 . . . . . . . . . . . . . . . . . . . . . . . . . . 143
2.24 Kuhn-Munkres algorithm. . . . . . . . . . . . . . . . . . . . 143
3.1 An example for quotient group. . . . . . . . . . . . . . . . . 174
5.1 Communication channel. . . . . . . . . . . . . . . . . . . . . 226
5.2 Binary symmetric channel. . . . . . . . . . . . . . . . . . . . 226
5.3 Standard array decoding of Z72 . . . . . . . . . . . . . . . . . 233
5.4 Form of the parity-check matrix of Hm (m ≥ 2). . . . . . . . 234
5.5 Case when two spheres intersect. . . . . . . . . . . . . . . . 235
A.1 A weighted simple graph. . . . . . . . . . . . . . . . . . . . . 280
A.2 A weighted directed graph. . . . . . . . . . . . . . . . . . . . 281
 List of Figures xv

A.3 A simple connected graph. . . . . . . . . . . . . . . . . . . . 282

A.4 A dfs drawing of the graph of Figure A.3. . . . . . . . . . . 283
A.5 dfs: strong components algorithm. . . . . . . . . . . . . . . . 284
A.6 dfs: topological sort. . . . . . . . . . . . . . . . . . . . . . . 285
B.1 The Petersen graph. . . . . . . . . . . . . . . . . . . . . . . 288
List of Tables

1.1 Execution of Prim’s algorithm on graph of Figure 1.8 . . . . 14

1.2 Sorting the edges of graph Figure 1.8 in order of
increasing cost . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3 Execution of Kruskal’s algorithm on the graph G of
Figure 1.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Execution of Dijkstra’s algorithm on the graph of
Figure 1.14 . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.5 Execution of Dijkstra’s algorithm on Figure 1.17 . . . . . . 33
1.6 Execution of improved Dijkstra’s algorithm on
Figure 1.17 . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.7 Eccentricity table of graph of Figure 1.21 . . . . . . . . . . . 44
1.8 Depth-first search number table of graph G of
Figure 1.25 . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.9 Depth-first search number (dfsn) table of the graph of
Figure 1.28 . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
1.10 LOW function table of the graph of Figure 1.28 . . . . . . . 62
1.11 Depth-first search number (dfsn) table of graph of
Figure 1.37 . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
1.12 LOWLINK function table of graph of Figure 1.37 . . . . . . 83
1.13 Three tours of Figure 1.38 . . . . . . . . . . . . . . . . . . . 91
1.14 Execution of NN algorithm on Figure 1.38 . . . . . . . . . . 94
1.15 An approximate algorithm for TSP using MST . . . . . . . 96
1.16 Informal Fleury algorithm . . . . . . . . . . . . . . . . . . . 97
1.17 An approximate algorithm for TSP using a minimum
matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.1 Checker board . . . . . . . . . . . . . . . . . . . . . . . . . 115
2.2 Domino . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
2.3 Execution of matching algorithm . . . . . . . . . . . . . . . 121
2.4 Brute-force weighted matching algorithm . . . . . . . . . . . 138
3.1 Group table of Klein’s 4-group . . . . . . . . . . . . . . . . 156
3.2 Group table of the dihedral group D3 . . . . . . . . . . . . . 157
6.1 Numerical equivalents of English characters . . . . . . . . . 250
6.2 Horner’s method . . . . . . . . . . . . . . . . . . . . . . . . 264
6.3 Modular exponentiation . . . . . . . . . . . . . . . . . . . . 265
6.4 Certificate of compositeness . . . . . . . . . . . . . . . . . . 266
6.5 Miller-Rabin primality testing algorithm . . . . . . . . . . . 266
A.1 Execution of Prim’s algorithm on graph of Figure A.1 . . . 280

xvii
xviii List of Tables

A.2 Execution of Dijkstra’s algorithm on the graph of

Figure A.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
A.3 Evaluation of dad array in Dijkstra’s algorithm of
Figure A.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
A.4 Dfsn table of graph of Figure A.4 . . . . . . . . . . . . . . . 282
A.5 LOW function table of the graph of Figure A.4 . . . . . . . 282
A.6 dfsn table of graph of Figure 1.38 . . . . . . . . . . . . . . . 283
A.7 LOWLINK function table of graph of Figure 1.38 . . . . . . 284
B.1 Execution of matching algorithm . . . . . . . . . . . . . . . 288
C.1 Numerical equivalents for English characters . . . . . . . . . 293
Preface

Discrete Mathematics and Programming courses are all-pervasive in Com-

puter Science, Mathematics, and Engineering curricula. This book is intended
for undergraduate/graduate students of Computer Science, Mathematics, and
Engineering. The book is in continuation of our first book, Foundations of
Discrete Mathematics with Algorithms and Programming (CRC Press, 2018).
The student with a certain mathematical maturity and who has programmed
in a high-level language like C (although C is not such a high-level language) or
Pascal can use this book for self-study and reference. A glance at the Table of
Contents will reveal that the book deals with fundamental graph algorithms,
elements of matrix theory, groups, rings, ideals, vector spaces, finite fields,
coding theory, and cryptography. A number of examples have been given to
enhance the understanding of concepts. The programming languages used are
Pascal and C. The aim of the book is to bring together advanced Discrete
Mathematics with Algorithms and Programming for the student community.

Scope of the Book

In Chapter 1 on Graph Algorithms I, we study various algorithms on graphs.
Unfortunately, the graphs are not manipulated by their geometrical repre-
sentation inside a computer. We start with two important representations of
graphs, the adjacency matrix representation, and the adjacency list represen-
tation. Of course, we have to choose a representation so that the operations
of our algorithm can be performed in order to minimize the number of ele-
mentary operations.
After seeing how the graphs are represented inside a computer, two mini-
mum spanning tree algorithms in a connected simple graph with weights asso-
ciated with each edge, one due to Prim and the other invented by Kruskal,
are presented.
Then, shortest path problems in a weighted directed graph are studied.
First, the single-source shortest path problem due to Dijkstra is presented.
Next, the single-source shortest path algorithm for negative edge weights is
given. Then, the all-pairs shortest path problem due to Floyd and the transi-
tive closure algorithm by Warshall are given. Floyd’s algorithm is applied to
find the eccentricities of vertices, radius, and diameter of a graph.

xix
xx Preface

Finally, we study a well-known graph traversal technique, called depth-first

search. As applications of the graph traversal method, we study the algorithms
to find connected components, biconnected components, strongly connected
components, topological sort, and PERT (program evaluation and research
technique). In the last subsection, we study the famous NP-complete problem,
the traveling salesman problem (TSP) and present a brute force algorithm and
two approximate algorithms.
In Chapter 2: Graph Algorithms II, we introduce another systematic way
of searching a graph, known as breadth-first search. Testing if a given graph
is geodetic and finding a bipartition of a bipartite graph are given as applica-
tions. Next, matching theory is studied in detail. Berge’s characterization of
a maximum matching using an alternating chain and the König-Hall the-
orem for bipartite graphs are proved. We consider matrices and bipartite
graphs: Birkhoff-von-Neumann theorem concerning doubly stochastic matri-
ces is proved. Then, the bipartite matching algorithm using the tree-growing
procedure (Hungarian method) is studied. The Kuhn-Munkres algorithm con-
cerning maximum weighted bipartite matching is presented. Flows in trans-
portation networks are studied.
Chapter 3: Algebraic Structures I deals with the basic properties of the
fundamental algebraic structures, namely, groups, rings, and fields. The sec-
tions on matrices deal with the inverse of a non-singular matrix, Hermitian
and skew-Hermitian matrices, as well as orthogonal and unitary matrices. The
sections on groups deal with Abelian and non-Abelian groups, subgroups,
homomorphisms, and the basic isomorphism theorem for groups. We then
pass on to discuss rings, subrings, integral domains, ideals, fields, and their
characteristics. A number of illustrative examples are presented.
In Chapter 4 on Algebraic Structures II, we deal with vector spaces and
their applications to solutions of systems of linear homogeneous and non-
homogeneous equations. To be specific, bases of a vector space, the dimension
of a vector space, and linear independence of vectors are discussed. As applica-
tions of these concepts, solutions of linear equations over the real field are dealt
with. Moreover, the LUP decomposition of a system of homogeneous/non-
homogeneous linear equations is discussed in detail. This is followed by illus-
trative examples. Finally, a detailed discussion on finite fields is presented.
Coding theory has its origin in communication engineering. It has been
greatly influenced by mathematics. Chapter 5, Introduction to Coding The-
ory, provides an introduction to linear codes. In particular, we discuss genera-
tor matrices and parity-check matrices, Hamming codes, sphere packings and
syndrome coding, cyclic codes, and dual codes. The algebraic concepts that
are used in this chapter have all been discussed in Chapters 3 and 4.
Chapter 6 deals with cryptography. Cryptography is the science of trans-
mitting messages in a secured way. Naturally, it has become a major tool
in these days of e-commerce, defense, etc. To start with, we discuss some of
the classical cryptosystems like the Caesar cryptosystem, the affine cryptosys-
tem, cryptosystems using matrices, Vigenere ciphers, and the one-time pad.
The most important and widely used cryptosystem is the RSA. We discuss
 Preface xxi

this system as well as the ElGamal cryptosystem. RSA is built on very large
prime numbers. So, this gives rise to the following natural question: Given a
large positive integer, how do we test if the given number is prime or not?
We briefly discuss the Miller-Rabin primality testing algorithm. This is a ran-
domized probabilistic algorithm to test if a given number is prime or not.
However, the problem of finding a deterministic polynomial-time algorithm to
test if a given number is prime or not remained unsolved until the Agrawal-
Kayal-Saxena (AKS) primality testing algorithm was proposed in 2002. We
present this algorithm, its proof, and some illustrative examples.

Use of the Book

The instructor has a great deal of flexibility in choosing the material from
the book. For example, the chapters on graph algorithms I and II may be
suitable for a course on “graphs and algorithms.” The chapters on algebraic
structures I and II and the chapters on coding theory and cryptography may
form a course on “Applied Algebra.” Many illustrative examples have been
given to help the understanding of the concepts. Algorithms are expressed
as informal pseudo-codes and programs in Pascal/C. Of course, these can be
easily translated into any language like C++, JAVA, etc. Exercises at the end
of each chapter/section test the understanding of the concepts developed in
the text.
We feel that the presentation of these chapters would go a long way in
providing a solid foundation in Discrete Mathematics to the students of Math-
ematics, Computer Science, and Engineering.
Acknowledgment

Parts of this book have been taught in Indian and French universities. The
authors thank N. Sridharan for going through some of the chapters and offer-
ing constructive suggestions, and Jaikumar Radhakrishnan and Arti Pandey
for their inputs on Chapter 6 to cryptography. They also thank A. Anu-
radha, R. Dhanalakshmi, N. Geetha, and G. Janani Jayalakshmi for their
help in typesetting. We take this opportunity to thank our institutions,
Bharathidasan University, Tamil Nadu, India and Université de Perpignan
Via Domitia, France, for their academic support. Our thanks are also due
to the faculties in our departments whose encouragement proved vital to
attain our goal. We also thank the four anonymous reviewers for suggest-
ing some changes on the initial version of the book. Last but not least, we
thank Aastha Sharma and Shikha Garg of CRC Press for their kind under-
standing of our problems and for their patience until the completion of our
manuscript.
The second author (S.S.) expresses his deep gratitude to (Late) Professor
K. R. Parthasarathy of the Indian Institute of Technology, Chennai, India,
for introducing him to Graph Theory and guiding his Ph.D. thesis. He is
also indebted to (Late) Professor Claude Berge, one of the greatest pio-
neers in Graph Theory and Combinatorics, who invited him to CAMS (Cen-
tre d’Analyse et de Mathématique Sociale) and guided his doctoral work in
Paris. Claude Berge had been a source of immense inspiration to him. Spe-
cial thanks are also due to Professors R. Balasubramanian (A. M. Jain Col-
lege, TN, India), Philippe Chrétienne (Université de Pierre et Marie Curie),
Robert Cori (Université de Bordeaux I), Alain Fougère (UPVD), Michel
Las Vergnas (CNRS), UFR secretaries Mme. Fabienne Pontramont (UPVD),
Mme. Dominique Bevilis (UPVD), Mircea Sofonea (Directeur de Laboratoire
LAMPS, UPVD), Michel Ventou (UPVD), Annick Truffert (Dean of the Fac-
ulty of Sciences, UPVD). Many thanks are also due to my students of UPVD
for their feedback. We are responsible for all the remaining errors, but still,
we feel that the initial readers of this book could smoke out some more bugs.
Though it is not in the Hindu custom to explicitly thank family members,
he would like to break this tradition and thank his wife, Dr. Usha Sridharan,
and his daughters, Ramapriya and Sripriya, for putting up with unusually
prolonged absence, as well as Dheeraj.
The entire book was composed using the TEX and LATEX systems developed
by D. E. Knuth and L. Lamport.

xxiii
xxiv Acknowledgment

The authors welcome corrections, comments, and criticisms from read-

ers, which would be gratefully acknowledged. They can be sent by email to:
[email protected].

R. Balakrishnan
Sriraman Sridharan
Tiruchirappalli, Tamil Nadu, India
Perpignan, France
August 2018
Authors

R. Balakrishnan studied at the Annamalai Uni-

versity, India for his Honours degree. He got
his Ph.D. degree from the University of Mary-
land, USA. His fields of interest are combinato-
rial designs, graph colorings, and spectral graph
theory. He is a coauthor for the books A Text-
book of Graph Theory, Springer, Second Edition
(2012) and Foundations of Discrete Mathemat-
ics with Algorithms and Programming, CRC Press
(2018).

Sriraman Sridharan is a member of the LAMPS

Research Laboratory and teaches in the Department
of Computer Science of the Université de Perpignan
Via Domitia, France. He has been with The Univer-
sité de Bordeaux and The Université de Pierre et
Marie Curie, Paris. He is a coauthor of the book Foun-
dations of Discrete Mathematics with Algorithms and
Programming, CRC Press (2018). He earned his Ph.D.
in Paris under the guidance of Professor Claude
Berge.

xxv
Chapter 1
Graph Algorithms I

The aim of physical sciences is not the provision of pictures, but the
discovery of laws governing the phenomena and the application of
these laws to discover new phenomena. If a picture exists, so much the
better. Whether a picture exists or not is only a matter of secondary
importance.

P. Dirac

In this chapter, we study various algorithms on graphs. Unfortunately, the

graphs are not manipulated by their geometrical representation inside a com-
puter. We start with two important representations of graphs: the adjacency
matrix representation and the adjacency list representation. Of course, we
have to choose a representation so that the operations of our algorithm can
be performed in order to minimize the number of elementary operations.
After seeing how the graphs are represented inside a computer, two min-
imum spanning tree algorithms in a connected simple graph with weights
associated to each edge, one due to Prim and the other invented by Kruskal
are presented. Then, shortest path problems in a weighted directed graph are
studied. First, the single-source shortest path problem due to Dijkstra is pre-
sented. As an application, an algorithm is given to test the bipartiteness of a
graph. Next, a single-source shortest path algorithm for negative edge weights
is given. Then, all-pairs shortest path problem due to Floyd and the transitive
closure algorithm by Warshall are given. Floyd’s algorithm is applied to find
eccentricities of vertices, radius and diameter of a graph.
Finally, we study a well-known graph traversal technique, called depth-first
search. As applications of the graph traversal method, we study the algorithms
to find connected components, biconnected components, strongly connected
components, topological sort and program evaluation and research technique
(PERT). In the last subsection, we study the famous NP-complete problem,
traveling salesman problem (TSP) and present the brute-force algorithm and
two approximate algorithms. For basic properties of graphs and digraphs, the
reader may see [6].

1.1 Representation of Graphs

Consider the graph of Figure 1.1 represented geometrically [3].

1
2 Discrete Mathematics

FIGURE 1.1: Geometric representation of a graph.

This graph of Figure 1.1 on 4 vertices is represented by the following 4 × 4

matrix M : (For a formal treatment on matrices, see Chapter 3.).

1 2 3 4
⎛ ⎞
1 1 1 0 1
2⎜0 0 0 0⎟
M= ⎜ ⎟
3⎝0 2 0 0⎠
4 0 0 1 0

Here, 4 is the number of vertices of the graph. The (i, j) entry of the above
matrix M is simply the number of arcs with its initial vertex at i and the
terminal vertex at j. This matrix is called the adjacency matrix of the graph
of figure.
More generally, for a n vertex graph G with vertex set X = {1, 2, . . . , n},
the adjacency matrix of G is the n × n matrix M = (mij ) where

mij = number of arcs of the form (i, j).

Memory space for the adjacency matrix: Since an n × n matrix has exactly n2
entries, the memory space necessary for the adjacency matrix representation
of a graph is of order O(n2 ). The time complexity of initializing a graph by
its adjacency graph is O(n2 ). This may preclude algorithms on graphs whose
complexities are of order strictly less than n2 .
Properties of the adjacency matrix: Let M denote the adjacency matrix of
a graph with vertex set X = {1, 2, . . . , n}. Then, by the definition of the
adjacency matrix, we have the following properties:
1. The sum of the entries of the ith row of M is equal to the out-degree of
the vertex i.

2. The sum of the entries of the jth column of M is equal to the in-degree
of the vertex j.

3. The sum of all the entries of the matrix M is the number of arcs of the
graph.
 Graph Algorithms I 3

FIGURE 1.2: A multigraph.

Adjacency matrix representation of a multigraph:

Consider the following multigraph of Figure 1.2 which is obtained by ignoring
the orientations of the arcs of the graph of Figure 1.1.
Its adjacency matrix is the 4 × 4 matrix M (see below) with its (i, j) entry,
the number of edges from the vertex i to the vertex j.
1 2 3 4
⎛ ⎞
1 1 1 0 1
2⎜1 0 2 0⎟
M= ⎜ ⎟
3⎝0 2 0 1⎠
4 1 0 1 0
More generally, the adjacency matrix of a multigraph with vertex set X =
{1, 2, . . . , n} is an n × n matrix M = (mij ) where
mij = The number of edges between the vertices i and j.
Note that the adjacency matrix of a multigraph is a symmetric matrix.
The following properties are easily proved for the adjacency matrix M of a
simple G.
1. The sum of the entries of the ith row of M is the degree of the vertex i.
2. The sum of all the entries of the matrix M is twice the number of edges.

Incidence matrix representation of a graph:

Consider a directed graph G = (X, U ) with vertex set X = {1, 2, . . . , n} of
n elements and the arc sequence U = {u1 , u2 , . . . , um } of m elements. Then,
the incidence matrix M of the graph G is the n × m matrix (mij )n×m where
⎧
⎪
⎪1 if i is the initial vertex of uj
⎨
−1 if i is the final vertex of uj
mij =
⎪
⎪ 2 if uj is a loop at vertex i
⎩
0 otherwise
Note that the first subscript i in mij indicates the vertex i and the second
subscript j indicates the arc uj .
4 Discrete Mathematics

FIGURE 1.3: Incidence matrix of a graph.

The following example clarifies the definition.

Example 1.1: Incidence Matrix

Consider the directed graph G of Figure 1.3. This graph has four
vertices and six arcs. The incidence matrix of the graph G is

u u2 u3 u4 u5 u6
⎛ 1 ⎞
1 1 2 1 0 0 0
2⎜ 0 0 −1 −1 −1 0 ⎟
M= ⎜ ⎟
3⎝ 0 0 0 1 1 −1 ⎠
4 −1 0 0 0 0 1

The following are properties of the incidence matrix of a directed graph:

1. The sum of entries of every column except the ones which represent
loops is 0.
2. The sum of entries of every column representing a loop is 2.

3. The sum of entries of the ith row (not containing the entry 2) is d+ (i) −
d− (i).

Incidence matrix representation of a multigraph:

Consider a multigraph G = (X, E) with X = {1, 2, . . . , n} and with edge
sequence E = (e1 , e2 , . . . , em ). The incidence matrix M is an n × m matrix
(mij )n×m where
⎧
⎨1 if vertex i is incident with ej
mij = 2 if ej is a loop at vertex i
⎩
0 otherwise
 Graph Algorithms I 5

Example 1.2
Consider the graph G of Figure 1.3 by ignoring the orientations of
the arcs and taking uj = ej . Then, the incidence matrix M of G is

e1 e2 e3 e4 e5 e6
⎛ ⎞
1 1 2 1 0 0 0
2⎜ 0 0 1 1 1 0⎟
M= ⎜ ⎟
3⎝ 0 0 0 1 1 1⎠
4 1 0 0 0 0 1

The following are properties of the incidence matrix of a multigraph:

1. The sum of every column entry of an incidence matrix is 2.

2. The sum of the entries of the i-th row of an incidence matrix is the
degree of the vertex i which is d(i).

Space complexity to represent a graph in the form of incidence matrix : There

are mn entries in the incidence matrix and hence the space complexity is
O(mn).
A formal study of matrices is done in Chapter 3.

Adjacency list representation of a digraph:

In the adjacency matrix representation of a graph G, an (i, j) entry zero of
the matrix represents the absence of arc (i, j) in the graph G. The adja-
cency list representation can be considered as a “condensed form” of the adja-
cency matrix representation omitting the zero entries. Consider the graph of
Figure 1.4.
This graph of 4 vertices is represented by 4 adjacency lists: succ(1), succ(2),
succ(3), succ(4) where succ(1) = (1, 2, 4), succ(2) = (), the null list, succ(3) =
(2, 2), succ(4) = (3). More generally, an n vertex graph G is represented by n
lists succ(1), succ(2), . . . , succ(n) where a vertex j appears k times in the list

FIGURE 1.4: A graph and its adjacency lists.

6 Discrete Mathematics

FIGURE 1.5: A multigraph and its adjacency lists.

succ(i) if and only if there are k arcs with initial vertex i and final vertex j,
that is, the multiplicity of the arc (i, j) is k.

Adjacency list representation of a multigraph:

When the sage points his finger to the Moon, the novice looks at its
finger.
Tamil Proverb

A multigraph of n vertices is represented by n adjacency lists adj(1), adj(2), . . . ,

adj(n) where the list adj(i) consists of all vertices j which are joined to the
vertex i. Note that the vertex j appears k times in the list adj(i) if the vertices
i and j are joined by k multiple edges. Further, if ij is an edge, then the vertex
j is in the list adj(i) and the vertex i is in the list adj(j) (see Figure 1.5).
This graph of 4 vertices is represented by 4 adjacency lists: succ(1), succ(2),
succ(3), succ(4) where succ(1) = (1, 2, 4), succ(2) = (1, 3, 3), succ(3) =
(2, 2, 4), succ(4) = (1, 3).
Figure 1.6 gives a linked lists representation of graph of Figure 1.5.

Space needed for adjacency linked list representation of a digraph:

Consider a graph G with n vertices and m arcs. It is represented by n adjacency
lists succ(1), succ(2), . . . , succ(n). The number of elements in the list succ(i)
is clearly the out-degree of the vertex i. In the linked list representation of
succ(i), each vertex in the list succ(i) is stored in a node consisting of two
fields: a vertex field and a pointer field. Counting each node as two units of
space, one for each field, we need 2d+ (i) + 1 space to represent succ(i) where
d+ (i) is the out-degree of the vertex i. Note that the pointer succ(i) needs
one unit of space. Hence the total space needed to represent the graph is

FIGURE 1.6: Linked lists representation of graph of Figure 1.5.

 Graph Algorithms I 7
n
 n

(2d+ (i) + 1) = n + 2d+ (i) = n + 2 m = O(n + m).
i=1 i=1

(since the sum of the out-degrees of the vertices of a graph is equal to the
number of arcs).
Hence, we see that the adjacency list representation uses only linear space
O(m + n), whereas the adjacency matrix needs a quadratic space O(n2 ).

Which representation to choose: adjacency matrix or adjacency linked list?:

The answer to this question depends on the operations performed by the algo-
rithm at hand. If a graph is “sparse,” that is, the graph does not have too many
arcs, then the adjacency list representation may be suitable. Suppose an algo-
rithm on graph frequently needs to test the existence of an arc from a vertex to
another vertex, then the adjacency matrix will be appropriate since accessing
an arbitrary entry of a matrix can be done in constant time. With the adja-
cency matrix, since the initialization itself takes O(n2 ) time, the importance
of algorithms may be diminished with complexity O(n log n) or O(n3/2 ).
Let us write a program in C which reads a directed graph G represented
by a linked list. The program finally writes the successors of each vertex (see
Figure 1.6).

#include <stdio.h>
#define max_n = 20 /* maximum number of vertices */
#define max_m = 100 /* maximum number of arcs */
struct node
{int v; struct node *next;};/* v for vertex */
int n, m; /*n = number of vertices. m = number of arcs */
struct node *succ[max_n]; /* graph is represented as an
array of pointers */
/* We work with a global graph.Otherwise the graph should be
declared as a variable parameter*/
void adj_list ( )
{
int i, a, b; /* i for the loop.(a,b) is an arc.*/
struct node *t; /* t, a temporary pointer */
printf("Enter two integers for n and m\n");
scanf(%d %d\n", &n, &m); /* read n and m */
/* initialize the graph with n vertices and 0 arcs */
for (i = 1; i <= n; i++) succ[i] = NULL;
/* read the m arcs */
for (i = 1; i <= m; i++)
{
printf("Enter the arc number %d ", i);
scanf("%d %d \n", &a, &b);/* un arc is an ordered pair of
vertices. b will be in succ(a) */
/* create a node referenced by t*/
t = (struct node *) malloc(sizeof *t);
8 Discrete Mathematics

t->v = b; /* assign b to v field of t-> */

/* attach t-> at the head of list succ[a] */
t->next = succ[a];
succ[a] = t;
}/*for*/
}/*adj_list*/

void print_list ( )
{
/* print_list writes the list of successors of
each vertex of the graph G*/
int i; /* i for the loop */
struct node *t; /* t, a temporary pointer */
for (i = 1; i <= n; i++)
{
/* write the list succ[i]*/
t = succ[i];
if (t == NULL)
printf(" No successors of %d\n ", i);
else
{
printf(" The successors of %d are :", i);
/* scan the list succ[i] and write the v fields of nodes */
while (t != NULL)
{
printf("%d ", t->v);
t = t->next; /* move t to next node */
}/*while*/
printf("\n");
}/*else*/
}/*for*/
}/* print_list*/

int main( )
{
adj_list ( ); /* call */
print_list ( ); /* call */
}

Let us write a program in C to represent a multigraph G with adjacency

matrix. The Program finally prints the adjacency matrix.

#include <stdio.h>
#define max_n = 20 /* maximum number of vertices*/
int i, j, n, m;/* n, the number of vertices. m, the number of edges
i, j for the loops*/
 Graph Algorithms I 9

int adj[max_n][max_n];/* The graph is represented by the

matrix adj. We don’t use row 0 and column 0*/
/* We work with a global graph.Otherwise the graph should be
declared as a variable parameter*/
void adj_matrix ( )
{
int a, b; /* ab is an edge. An edge is an unordered pair*/
printf("Enter two integers for n and m\n");
scanf("%d %d\n", &n,&m);
/* initialize the graph with n vertices and 0 edges */
for ( i =1; i <= n; i++)
for ( j= 1; j <= n; j++)
adj[i][j] = 0;
/* end of initialization */
/* read the m edges*/
for (i = 1; i <= m; i++)
{
printf("Enter the edge number %d " , i);
scanf("%d %d \n", &a, &b);
adj[a][b]++;adj[b][a]++;
}/*for*/
}

void print_adj_matrix ( )
{
int i, j;/* i, j for the loops */
for (i = 1; i <= n; i++)
{ for( j =1; j <= n; j++)
printf("%d ", adj[i][j]);
printf("\n");
}
}
int main ( )
{
adj_matrix( );/* call */
print_adj ( ); /* call */
}

1.2 Minimum Spanning Tree Algorithms

A free tree or simply a tree is a connected graph without an elementary
cycle, that is, a tree is a connected acyclic graph. Hence, a tree does not contain
loops or multiple edges, that is, a tree is a simple graph. In the following figure,
a tree with ten vertices is drawn (see Figure 1.7).
10 Discrete Mathematics

FIGURE 1.7: A tree.

In the tree of Figure 1.7 with ten vertices, we observe the following
properties:

1. There is a unique elementary walk, that is, a path between any two
vertices.
2. The number of edges is exactly one less than the number of vertices.

We now prove the above two properties for any general tree.
Proposition 1.1. In any tree, there is a unique path joining any two vertices.

Proof. We shall prove the proposition by contradiction. If not, consider two

different paths P1 and P2 joining the vertices x and y. If the paths P1 and P2
are vertex-disjoint, except of course the initial and final vertices, then clearly
the union of the paths P1 ∪ P2 is an elementary cycle in the tree, which is a
contradiction. Otherwise, let z(= y), z  (possibly x) be the last and next-to-
last vertices common to the paths P1 and P2 , respectively. Then, the union of
the two subpaths of the paths P1 and P2 from z  to z is an elementary cycle,
a contradiction. Hence the result.
For the second property, we need the following simple lemma.
Lemma 1.1. Every simple graph with the degree of each vertex at least two
contains an elementary cycle.
Proof. Consider a longest possible path P in the graph. (The length of a path
is the number of edges in the path.) Let x0 be the initial vertex of the path
P = (x0 , x1 , . . . , xl ) (the length of the path P is l) and consider the vertices of
the graph adjacent to the initial vertex x0 . Since the degree of x0 ≥ 2, there
is a vertex y = x1 such that x0 y is an edge of the graph.
If y = xi for 2 ≤ i ≤ l, then we have a path (y, x0 , x1 , . . . , xl ) whose length
l + 1 exceeds that of P , a contradiction.
If y = xi for some i with 2 ≤ i ≤ l, then (x0 , x1 , . . . , xi , x0 ) (since x0 y is
an edge) is an elementary cycle.
 Graph Algorithms I 11

Proposition 1.2. In any tree with n vertices, the number of edges m is exactly
n − 1.

Proof. The proof proceeds by induction on the number of vertices of the tree.
Basis: If the number of vertices n = 1, then the tree consists of only one vertex
and no edges, that is, the number of edges m = 0. Hence the proposition is
trivially satisfied.
Induction hypothesis: Suppose the proposition is true for all trees with the
number of vertices n at least two.
We shall prove the proposition for trees with n+1 vertices. Consider a tree
T with n + 1 vertices. We claim that the tree T contains a vertex of degree
one. If not, the degree of each vertex is at least two and by Lemma 1.1, T
contains a cycle, a contradiction. Let x be a vertex of degree one in the tree
T. Then, the vertex deleted graph T − x is still a tree. Note that T and T − x
differ by exactly one vertex x and exactly one edge, the edge incident with
x. Now T − x is a tree on n vertices and the number of edges m of the tree
T − x, verifies, by induction hypothesis, the following equation,
m = n − 1.
Adding one on both sides, we get m + 1 = (n + 1) − 1, that is, the number
of edges of T is exactly one less than the number of vertices of T. Thus, the
proof is complete.
Spanning tree of a connected graph:
Consider a connected graph G. A spanning subgraph of G, which is a tree,
is called a spanning tree of G. In a connected graph, a spanning tree always
exists as we shall see below:
Since a tree is acyclic, we intuitively feel that if we remove all the cycles of
the graph G without disconnecting G, we will be left with a spanning tree. In
fact, this is the case.
Consider any elementary cycle of G and delete exactly one of its edges.
This results in a connected spanning subgraph G1 of G, since in an elementary
cycle, between any two vertices of the cycle, there are exactly two different
internally vertex-disjoint paths. If the resulting graph is a tree, then we have
obtained a desired spanning tree. Otherwise, we find an elementary cycle of
G1 and delete one of its edges. This gives us a connected spanning graph G2 .
If G2 is a tree, then G2 is a desired spanning tree of G. Otherwise, we continue
the procedure as before till a spanning tree of G is found. If a graph G contains
a spanning tree, then clearly it must be connected. Thus, we have proved the
following result.
Theorem 1.1. A graph G has a spanning tree if and only if it is connected.

1.2.1 Prim’s minimum spanning tree algorithm

Input: The input of this algorithm is a connected simple graph G with weights
associated to each edge of the graph (see Figure 1.8). Such graphs are also
12 Discrete Mathematics

FIGURE 1.8: A weighted graph G and its spanning tree T .

called networks. The vertices are interpreted as cities and the edges as high-
ways connecting its end vertices. The weight associated with an edge is inter-
preted as the cost of constructing a highway connecting its end vertices (cities).
The total cost of the network is the sum of the costs on each edge of G. For
example, the total cost of the figure G is 55 (see Figure 1.8).

How to represent a weighted graph?:

One way to represent a weighted graph is by weighted adjacency matrix
M = (mij )n×n where
⎧
⎨weight of the edge ij if ij is an edge of the graph
mij = 0 if i = j
⎩
∞ if ij is not an edge of the graph.
In the computer implementation, the ∞ is replaced by, for example,
“INT MAX” in the language C if the weights are integers, “maxint” in the
language Pascal, or a very large number which will not be a “legal” cost of
constructing a highway.
The weighted graph G of Figure 1.8 is represented by its weighted adja-
cency matrix as follows (the weight of a missing edge is taken as 100):
1 2 3 4 5 6 7
⎛ ⎞
1 0 9 100 100 100 2 3
2⎜⎜ 9 0 5 100 100 7 100 ⎟
⎟
3⎜⎜ 100 5 0 8 6 6 100 ⎟
⎟
M = 4⎜⎜ 100 100 8 0 6 100 100 ⎟
⎟
5⎜⎜ 100 100 6 6 0 100 100 ⎟
⎟
6⎝ 2 7 5 100 100 0 4 ⎠
7 3 100 100 100 100 4 0
 Graph Algorithms I 13

The other way of representing a weighted graph is by weighted adjacency lists.

In this representation, we introduce a new field called weight in each node, in
addition to the two existing fields, namely, the field containing the vertex and
the field containing the pointer. In the weight field, the weight of the edge ij
will be stored, if the content of the vertex field is j and the node in question
belongs to the list adj(i).
Output: A minimum spanning tree of G, that is, a spanning tree T of the
graph G such that the sum of the costs of all the edges of the tree T is a
minimum (see Figure 1.8). In Figure 1.8, a spanning tree T of G is depicted
and its total cost is 38.

Algorithm 1.1 (Brute-force method). One way to find a minimum cost

spanning tree in the graph G is the following exhaustive search.

Generate all possible spanning trees of G and find the total cost of each of these
spanning trees. Finally, choose one tree for which the total cost is minimum.
This procedure will work “quickly” if the number of vertices of the input graph
G is sufficiently small. If the number of vertices is not sufficiently small, this
procedure takes an enormous amount of time, that is, exponential time since
by Caley’s theorem [1] there are nn−2 possible non-identical trees on n given
vertices.

Algorithm 1.2 (Prim’s algorithm). This is a polynomial time algorithm.

Prim’s algorithm, though it is greedy, provides us with an exact solution. We
shall now describe what is meant by a greedy algorithm.

A Greedy algorithm: A greedy algorithm is an algorithm which “hopes” to

find a global optimum by choosing at each iteration of the algorithm a local
optimum. A greedy algorithm sees only “short-term” gains. More generally, a
heuristic is an algorithm which “quickly” produces a suboptimal solution but
not necessarily an optimal solution. The word “quickly” means “in polynomial
time.”
For example, greedy is a heuristic. Greedy does not always give an optimal
solution, but it gives “somewhat” of an acceptable solution. But in the case
of Prim’s algorithm, greedy provides us with an exact solution. First of all,
we shall illustrate Prim’s algorithm with our graph of Figure 1.8.

Observation 1.1. Trees can be grown by starting with one vertex (initializa-
tion) and adding edges and vertices one-by-one (iteration).

According to Observation 1.1, we start with any vertex, say the vertex 1. The
vertex 1 forms a tree T by itself. The algorithm chooses an edge e1 of the
graph G such that one end of the edge e1 lies in tree T and the other end
lies outside of the tree T (this is to avoid cycle in T ) with the weight of e1
minimum. Then, we add the edge e1 to the tree T. In our figure, the edge
e1 = 16. Then, we select an edge e2 such that one end of the edge e2 lies
in the tree T under construction, and the other end lies outside of the tree
14 Discrete Mathematics

TABLE 1.1: Execution of Prim’s algorithm on graph of Figure 1.8

Iteration S = X? s t S ← S ∪ {t} T ← T ∪ {st}
number
0 – – – {1} ∅
(Initial)
1 yes 1 6 {1, 6} {16}
2 yes 1 7 {1, 6, 7} {16, 17}
3 yes 6 3 {1, 6, 7, 3} {16, 17, 63}
4 yes 3 2 {1, 6, 7, 3, 2} {16, 17, 63, 32}
5 yes 3 5 {1, 6, 7, 3, 2, 5} {16, 17, 63, 32, 35}
6 yes 5 4 {1, 6, 7, 3, 2, 5, 4} {16, 17, 63, 32, 35, 54}
Exit the no
loop

FIGURE 1.9: A minimum spanning tree T  of G.

T with the weight of e2 minimum. In our case, the edge e2 = 17 and so on.
The different steps of the tree-growing procedure are illustrated in Table 1.1.
In Table 1.1, the set S denotes the vertices of the tree T under construction.
T denotes the set of edges of the tree which is grown by the algorithm, s and
t denote vertices of the graph G with s ∈ S and t ∈ X \ S, and X represents
the set of vertices of the graph G. The procedure ends as soon as the tree T
under construction contains all of the vertices of the graph G. The edges of
a minimum spanning tree obtained are found in the 6th column of iteration
number 6 in Table 1.1. The minimum spanning tree obtained by Table 1.1 is
shown in Figure 1.9 and its cost is the sum of the costs of its edges which
is 27. Now let us write the algorithm in pseudo-language.

procedure prim(var G : Graph; var T: Set of Edges);

(* The procedure prim takes as input a connected weighted
graph G and returns as output a minimum spanning tree T of G.
Note that the representation of the graph G as well as the
tree T are left unspecified*)
var S : Set of Vertices;(* representation of S unspecified *)
s, t : Vertex;
begin(* prim *)
 Graph Algorithms I 15

(* initialization of S and T*)

(* S denotes the set of vertices of tree T grown by prim in a
step by step manner. T denotes the set of edges of tree T *)
S := {1}; T := empty_set;
(* iteration *)
while S not_equal_to X do (* X = {1,2,...,n} denotes
the set of vertices of G *)
begin(* while *)
(1) select an edge st such that s in S and t in X - S
with the weight of st minimum;
(2) add the vertex t to the set S;
(3) add the edge st to the tree T;
end; (* while *)
end;(* prim *)

Complexity of Prim’s algorithm:

Clearly, the initialization takes a constant amount of time O(1), which involves
two assignments. The while loop is executed exactly n − 1 times, because of
the following two reasons:

1. At each iteration, the number of vertices in the set S is incremented by

one unity, thanks to the instruction (2) in the while loop.
2. The loop terminates as soon as S = X and |X| = n.

We shall now consider the instruction number (1) of the while loop of Prim.
This statement can be implemented in O(n) time. Hence, each execution of
the while loop demands O(n) time (Recall that when we compute complexity,
the lower-order terms and multiplicative constants can be neglected [6]).
Therefore, complexity of Prim’s algorithm is (n − 1)O(n) = O(n2 ), which
is a polynomial of degree 2.

Why does Prim’s greedy algorithm give the exact optimum?:

Prim’s algorithm chooses at each iteration what appears “locally” as the best
thing to do. At the end of the iteration, it obtains the best overall! The fol-
lowing result explains the magic.

Minimum spanning tree property:

Property 1.1. Let G be a connected simple graph with weight attached to

each edge of the graph G. Consider a proper subset S of vertices of G and an
edge e having one end in S and the other end outside of S with the weight of
the edge e minimum. Then, there is a spanning tree of minimum weight in G
containing the edge e.

Proof. Since G is a connected graph, by Theorem 1.1, it contains a spanning

tree. Consider a spanning tree T of the graph G. If the tree T contains the
16 Discrete Mathematics

edge e = xy, then we are done. Otherwise, the tree T surely will contain an
edge e = x y  having one end x in S and the other end y  outside of S, for,
if not, the tree T will not be connected. (In fact, more generally, a graph is
connected, if and only if for any partition of the vertex set into two proper
subsets, there is an edge having one end in one subset of the partition and the
other end in the other subset of the partition.) Since the weight of the edge
e is a minimum among all edges joining S and and its complement, we have
the weight of e ≥ the weight of e.
Now consider the spanning subgraph obtained by adding the edge e to
the tree T , that is, consider the graph G = T + e. Then, G contains exactly
one elementary cycle. This is because of the following argument: By Propo-
sition 1.1, in the spanning tree T , there is a unique path P joining the end
vertices x and y of the edge e. But then, G = T + e contains the unique ele-
mentary cycle P + e (A path plus the edge joining the initial and final vertices
of the path is an elementary cycle) (see Figure 1.10).
Now form the graph T  = G − e . T  is still connected, because the end
vertices x and y  of the deleted edge e are connected by the path P + e − e
in the graph T  . Moreover, T  is a tree, because the only cycle P + e of G is
destroyed by the removal of the edge e of this cycle. Not that the trees T and
T  differ by only two edges e and e . Since the weight of the edge e ≤ the weight
of the edge e , we have, the weight of the tree T  ≤ the weight of the tree T .
Since T is a minimum weight spanning tree, we must have equality of the
weights of T and T  . Thus, we have constructed a minimum spanning tree T 
containing the edge e.

We shall now write a computer program in C to implement Prim’s algorithm.

The given weighted graph is represented by its cost matrix M where M [i][j]
is the cost of the edge ij. If there is no edge between the vertices i and j,
we set M [i][j] = ∞. In C, ∞ can be replaced by the constant IN T M AX if
costs of the edges are “small” integers. In fact, in a computer implementation
of Prim’s algorithm, ∞ is replaced by a cost which cannot be a “legal” cost

FIGURE 1.10: A step in the proof of Prim’s algorithm.

 Graph Algorithms I 17

of any edge. We define two arrays dad and cost mins. dad[i] gives the vertex
in S that is currently “closest” (with respect to the cost of edges) to vertex i
in X \ S. cost min[i]= the cost of the edge (i, dad[i]).

Implementation of Prim’s algorithm in C:

#include <stdio.h>
#include <stdlib.h>
#include <limits.h>
/* INT_MAX is defined in this library */
#define n_max 10//maximum vertices

int M[n_max][n_max],n,m;//M, the cost matrix

//n,m the number of vertices, edges
void graph ()

{/*keyboarding the M*/

int i,j,x,y,c;
/*i,j for the loops
x--y is an edge, its cost c */
printf ("enter the values of n,m \n ");
scanf ("%d %d",&n,&m ) ;

/* initialization of M*/

for (i=1;i<=n;i++)
for ( j=1;j<=n;j++)
M[i][j]=INT_MAX;

/*end of initialization */

/* enter the costs of edges */

for (i=1;i<=m;i++)
{
printf ("enter the two ends of edge %d \n",i );
scanf ("%d%d", &x,&y );
printf (" enter the edge cost c : \n");
scanf ("%d",&c);
M[x][y]=c; M[y][x]=c;
}
}
void print_graph ()
{
int i,j;
for (i=1;i<=n;i++)
{
18 Discrete Mathematics

for (j=1;j<=n;j++)
printf ("%5d ",M[i][j]);
printf ("\n\n\n");
}
}

//begin Prim
void prim()
{ int dad[n_max], cost_min[n_max];
//array dad represents the tree under construction
int i,j,k,min;
//initialization
for(i=1;i<=n;i++)

dad[i]=1;//tree consists of only one

//vertex 1
cost_min[i]=M[i][1];
//if there is no edge (1,i), M[1][i]=
//INT_MAX
}

for(i=1;i<=n;i++)
{//find a vertex k outside the tree to be added to the
//tree
k=2;
min=cost_min[2];

for(j=3;j<=n;j++)
if(cost_min[j]<min)
{
k=j;
min=cost_min[j];
}

//print edge
printf("%d%d\n",k,dad[k]);
cost_min[k]=INT_MAX;//k is added to tree
//update arrays cost_min, dad
for(j=2;j<=n;j++)

if((M[k][j]<cost_min[j])&&(cost_min[j]<INT_MAX))
{
cost_min[j]=M[k][j];
dad[j]=k;
}
 Graph Algorithms I 19

}
//end of prim

int main ()
{
graph () ;
print_graph () ;
prim();
return 0;

1.2.2 Kruskal’s minimum spanning tree algorithm

This is again a greedy algorithm which furnishes us an exact optimum. An
acyclic simple graph is often referred to as a forest, that is, each connected
component of a forest is a tree. This algorithm proceeds somewhat in the
“opposite direction” of Prim’s algorithm. We shall now describe Kruskal’s
algorithm.
Input: A weighted connected simple graph, that is, a simple connected graph
with a weight attached to each edge. As in Prim’s algorithm, the vertices
represent different cities and the weight associated with an edge e = xy may
represent the cost of constructing a highway or railway line between the two
cities x and y.
Output: A minimum weight spanning tree of the graph.
Algorithm: We shall first illustrate the algorithm on the graph of Figure 1.8.
First of all, the edges of the given graph are sorted in order of increasing
cost. Initially, we start with a forest F consisting of all the vertices of the graph
but no edges. Let us denote by P X the set vertices of the connected com-
ponents of the forest F which evolves during the execution of the algorithm.
P X defines a partition of the vertex set X of the graph G.
Then, we consider each edge e = xy in turn among the sorted list of edges.
If the end vertices x and y of the edge e under consideration are in distinct
components of the forest F , then we add the edge e to the forest. Note that
this operation reduces the number of components of the forest F by one. If
the vertices x and y are in the same component of the forest F , we discard the
edge e. The algorithm proceeds like this, until the forest F consists of only one
connected component, that is, until F becomes a tree. This terminal condition
can be tested by the equality: |P X| = 1, that is, the number of subsets in the
partition is 1.
20 Discrete Mathematics

TABLE 1.2: Sorting the edges of graph Figure 1.8 in

order of increasing cost
Edges Costs Edges Costs
16 2 35 6
17 3 45 6
67 4 26 7
23 5 34 8
36 5 12 9

TABLE 1.3: Execution of Kruskal’s algorithm on the graph G of

Figure 1.8
Iteration Edges Add or PX F |P X| = 1?
number reject
0(Initial) {1}, {2}, {3}
– – {4}, . . . , {7} ∅ –
1 16 add {1, 6}, {2}, {3} {16} no
{4}, {5}, {7}
2 17 add {1, 6, 7}, {2} {16, 17} no
{3}, {4}, {5}
3 67 reject no
4 23 add {1, 6, 7}, {2, 3} {16, 17, 23} no
{4}, {5}
5 36 add {1, 6, 7, 2, 3} {16, 17, 23, 36} no
{4}, {, 5}
6 35 add {1, 6, 7, 2, 3, 5}, {4} {16, 17, 23, 36, 35} no
7 45 add {1, 6, 7, 2, 3, 5, 4} {16, 17, 23, 36, 35, 45} no
Exit loop yes

The different steps in the execution of the algorithm on the graph G of

Figure 1.8 are shown in the following table.
Let us first sort the edges in order of increasing cost (Table 1.2). The
edges of a minimum spanning tree are available in the fifth column of iteration
number 7 in Table 1.3. The reader can see that the spanning tree obtained by
Kruskal’s algorithm is the same as the one obtained in Prim’s algorithm (see
Figure 1.11).
We shall now write Kruskal’s algorithm in pseudo-code.

FIGURE 1.11: Kruskal’s minimum spanning tree T  of G.

 Graph Algorithms I 21

Kruskal’s Algorithm
procedure Kruskal( var G: Graph; var F: Set of Edges);
(* Kruskal’ takes as input a connected weighted graph G and
outputs a minimum spanning tree F*)
var PX : Partition of the vertex set X;
(* PX denotes the partition of the vertex set X, induced by
the connected components of the forest F*)
e : Edge; L : List of Edges;
x,y : Vertex;(* x, y are the ends of edge e*)
begin (* Kruskal *)
(* initialization of PX and F *)
PX := empty_set; F := empty_set;
Form a list L of the edges of G in increasing order of cost;
(* Forest consists of isolated vertices and no edges *)
for each vertex x of the graph G do
add {x} to PX;
(* Iteration *)
while | PX | > 1 do
begin (* while *)
(1) choose an edge e = x y, an edge of minimum cost from the
list L;
(2) delete the edge e from the list L;
(3) if x and y are in two different sets S1 and S2 in PX
then begin
(4) replace the two sets S1 and S2 by their union S1
U S2;
add the edge x y to F;
end;
end;(* while *)
end;

Complexity of Kruskal’s algorithm:

The “for” loop in the initialization part takes O(n) time where n is the number
of vertices of the graph G. In fact, the edges are not sorted in the initialization
part into a list but are kept in a heap data structure [3]. A heap is an array
a[1], a[2], . . . , a[n] of numbers such that a[i] ≤ a[2i] for 1 ≤ i ≤ n/2 and
a[i] ≤ a[2i + 1] for 1 ≤ i < n/2. This array structure can be conveniently
viewed as a quasi-complete binary tree [3] in which the value at a node is less
than or equal to the values at its children node. This organization of the edges
require O(m) time.
The major cost of the algorithm is dominated by the while loop. In the
while loop, each edge is examined one-by-one in order of increasing cost. Either
the edge examined is added to the forest F under construction or it is rejected.
Hence, the while loop is executed at most m times where m is the number
22 Discrete Mathematics

of edges of the graph G. The instruction (1) choosing an edge of minimum

cost can be performed in O(log m) time. In fact, log m is the initial height
of the heap viewed as a binary tree. The total time needed to execute the
instructions (3) and (4) is O(m) for all “practical” values of m, if we use the
fast disjoint-set union algorithm [3]. The remaining instructions in the while
loop can be performed in O(1) time. Hence, the complexity of the Kruskal
algorithm is O(m log m).

1.2.3 Rooted ordered trees and traversal of trees

A rooted tree T can be defined recursively as follows:

Definition 1.1 (Rooted tree). A single vertex r alone is a tree and r is

its root. (basis of recursion)
If T1 , T2 , . . . , Tk are trees with roots s1 , s2 , . . . , sk , respectively, then we may
construct a new tree T = (r; T1 , T2 , . . . , Tk ) with root r as follows (see Figure
1.12):
T1 , T2 , . . . , Tk are called the subtrees of the root r and s1 , s2 , . . . , sk are
called the sons/daughters of the root r taken in this order. (recursion) T1 is
the leftmost subtree of r, and Tk is the rightmost subtree. Note that the order of
the sons is taken into account. The edges are supposed to be directed downwards
(even though it is not represented in Figure 1.12. If there is a downward path
from a vertex x to a vertex y, then y is called a descendant of x, and x is
called an ancestor of y. A direct descendant is a son/daughter and a direct
ancestor is called the dad or mom. A vertex with no son/daughter is called a
leaf.

Traversal of rooted ordered trees:

We define three tree traversals: (1) Preorder, (2) Inorder, (3) Postorder.
These traversals are again defined recursively. In fact, they are simply
mappings of the two-dimensional tree into a one-dimensional list.
If the tree consists of only one vertex, the root r, T = (r) then the list (r)
consisting of only one element is the preorder, inorder and postorder traversal
of T. That is, preorder(T)=inorder(T)=postorder(T)=(r).(basis).

FIGURE 1.12: Recursive definition of a rooted tree.

 Graph Algorithms I 23

FIGURE 1.13: Illustration of tree traversals.

If T = (r; T1 , T2 , . . . , Tk ), then we define the preorder, inorder, and pos-

torder as the lists of vertices

preorder(T ) = (r, preorder(T1 ), preorder(T2 )), . . . , preorder(Tk ))(recursion).

inorder(T ) = (inorder(T1 ), r, inorder(T2 ) . . . , inorder(Tk ))(recursion).
postorder(T ) = (postorder(T1 ), postorder(T2 ), . . . , postorder(Tk ), r)(recursion).

Note that the root r comes first in preorder, last in postorder. In the inorder,
the root r comes after the listing of vertices of the first subtree of r[3].

A useful trick to find these orders:

We shall illustrate this trick by an example. Consider the tree of Figure 1.13.
Think of a walk around the tree in the counter-clockwise direction, starting
from the root and staying as close to the tree as possible.
For preorder, we write the vertex in the list the very first time we meet it
during the walk.
For postorder, we write the vertex in the list the last time we meet it
during the walk, that is, as we move upwards.
For inorder, we write the leaf (a leaf is vertex of the tree with no son/
daughter) the first time we meet it during the walk and other vertices during
the second time during the walk (in the bay).
Note that each vertex is written exacly once in the list. In the graph
of Figure 1.13, preorder(T ) = (r, 1, 4, 5, 9, 2, 6, 7, 3, 8), postorder(T ) =
(4, 9, 5, 1, 6, 7, 2, 8, 3, r), inorder(T ) = (4, 1, 9, 5, r, 6, 2, 7, 8, 3). The sons/
daughters of the root r are 1, 2, 3 taken in this order: the leftmost subtree
of r is the vertex 1 and all its descendants in the tree T .

1.3 Shortest Path Algorithms

In this section, we may consider without loss of generality directed graphs
with no multiple arcs, that is, 1-graphs, without loops. We shall first study
the single-source shortest path algorithm invented by Dijkstra. This is again a
24 Discrete Mathematics

greedy algorithm. We use “cheapest cost path” as synonym for the “shortest
path.” “Path” always means in this section a “directed path.”

1.3.1 Single-source shortest path algorithm

Input: A directed graph G = (X, U ) where X = {1, 2, . . . , n}. To each arc
is associated a non-negative weight. We interpret the vertices of the graph G
as different cities and the weight of an arc u = (x, y) as the cost of traveling
from the city x to the city y along the arc u. The weight function on the set of
arcs is denoted by w, that is, the weight of the arc u = (x, y) is w(x, y). The
vertex 1 is designated as the source or origin. This vertex 1 may be interpreted
as the capital city.
If there is no arc from a vertex s to a vertex t, we set the weight of the
arc (s, t), w(s, t) = ∞. In the computer implementation of the algorithm,
∞ will be replaced by a large number which cannot be a legal cost of traveling
between two cities. For example, if the weights are non-negative integers, we
may replace ∞ by INT MAX in the language C or maxint in the language
Pascal.
Output: The cheapest cost of traveling from the vertex 1 to all other vertices
of the graph is along a directed path. The cost of a directed path is the sum
of the costs of its arcs (see graph G of Figure 1.14).
The cheapest dipath from the vertex 1 to the vertex 5 is (1, 2, 4, 5) and its
cost is 5 + 4 + 3 = 13 and the cheapest path from the source vertex 1 to the
vertex 3 is the path (1, 3) which consists of the directed edge 13 and its cost
is just 3.
Algorithm: To explain the algorithm, we need a few ideas. A vertex x of a
directed graph G, is called a root of G if there is a path from the root vertex
to all other vertices of G. A root may not always exist.
An arborescence is a graph possessing a root and having no cycles. Note
that the root of an arborescence is unique. The following figure is a spanning
arborescence of the graph of Figure 1.14 (see Figure 1.15). In fact, it is a
“minimum weighted” spanning arborescence.)

FIGURE 1.14: A directed weighted graph G.

 Graph Algorithms I 25

FIGURE 1.15: A minimum spanning arborescence A of Figure 1.14.

In an arborescence, there is a unique path from the root to all the other ver-
tices. In fact, Dijkstra’s algorithm constructs a spanning arborescence of the
given graph G having its root as the source of the graph G and its unique path
from the root to a vertex x as a cheapest path in the graph G. In the algorithm,
we use the variable S denoting the set of vertices of a cheapest arborescence
under construction. (Note that in Prim’s algorithm also the variable S denotes
the set of vertices of a minimum spanning tree under construction.)
The variable S satisfies the following two properties, which can be consid-
ered as loop invariant:

1. For all vertices s in the set S, a cheapest cost of a path (also called
shortest path) from the source 1 to s is known and this path lies entirely
in the set S, that is, all the vertices and the arcs of a shortest path
from 1 to s lies fully in S. This condition implies that the set S must
contain the source vertex 1. In fact, the assignment S := {1} is used as
initialization of the set S.
2. For a vertex t outside of the set S (if such a vertex exists), a path relative
to the set S is a path from the source vertex 1 to the vertex t such that
all the vertices of this path lie completely in the set S but for the last
vertex t. We shall now state the second property of the set S in terms
of paths relative to the set S. For each vertex t outside of the set S, a
shortest path from 1 to t relative to S lies completely in the set S except
for the final vertex t.
Interpretation of two properties of S when S equals the whole of the vertex set
X of G:
According to the first property of S = X, a shortest path from the origin
vertex 1 to any vertex x of the graph is known, and we have solved the
problem! Note that the second property is not applicable, since X \ S = ∅.
Hence, the terminal condition of the algorithm is S = X.
In the Dijkstra’s algorithm, the following elementary observation is used:
A subpath of a shortest path is a shortest path!
26 Discrete Mathematics

Note that the above observation will not be valid if the weights of the arcs
are not non-negative numbers. Another property on which Dijkstra’s algo-
rithm works is the following: This is the point where the algorithm becomes
greedy!
Consider a shortest path P among the shortest paths relative to the set
S. Then, the path P is a shortest path in G itself.
Let us write Dijkstra’s algorithm in pseudo-code. The weighted graph is
represented by its weighted adjacency matrix W = (wij )n×n where
⎧
⎨0 if i = j
wij = weight of the arc ij if ij is an arc
⎩
∞ otherwise.

The vertex set is X = {1, 2, . . . , n}. We use the array D where

The cost of the shortest path from the vertex 1 to i if i ∈ S,
D[i] =
The cost of the shortest path relative to S if i ∈
/ S.
Note that we obtain the final result in the array D as soon as S = X. We use
the types, Matrix and T-Array where Matrix is a two-dimensional n × n array
of real numbers, and T-Array is a one-dimensional array of n numbers in the
algorithm:

procedure Dijkstra( var W : Matrix; var D : T_Array);

(* Dijkstra takes as input a matrix of non-negative weights
and it returns an array D of cheapest cost paths.*)
var S : set of vertices; x, y : vertex;
begin(* Dijkstra*)
(* Initialization of S and D*)
S := {1};
D[1] := 0;(* The cheapest cost from 1 to itself of 0*)
for i := 2 to n do
D[i] := w(1, i);(* The shortest path relative to S = {1}
is the arc (1, i). S verifies the two properties *)
(* iteration *)
while S not equal to X do
begin (* while *)
(1) choose a vertex y in X - S such that D[y] is minimum
(* y is called the pivot vertex*)
(2) add y to the set S;(* The shortest path among the
shortest paths relative to S is a shortest path in G.
This is the statement where "greedy" comes into
play *)
(3) for each x in X - S do
(4) D[x] := min( D[x], D[y] + w(y, x));
(* This "for" loop is the crux of the Dijkstra’s
algorithm.
 Graph Algorithms I 27

In this loop, we adjust the values of D[x] to take

into account the vertex y just added to S, to preserve
the second property of the set S*)
end;(*while*)
end;(*Dijkstra*)

Remark 1.1. The vertex y in Dijkstra’s algorithm is called the pivot vertex.
Of course, the pivot changes at each iteration.

Let us now execute Dijkstra’s algorithm on the graph of Figure 1.14. The
following Table 1.4 traces the different steps of the algorithm.
The final result is found in columns 5, 6, 7, 8, 9 of iteration number 5 in
Table 1.4.

We observe the following properties of Dijkstra’s algorithm:

1. The vertices are added to the set S in order of their increasing costs
of D.
2. Once a vertex x is added to the set S, the value of D[x] will remain
constant till the end of the algorithm. Hence the set S can be interpreted
as the set of Stabilized vertices.

3. The statement number (4) of the algorithm can be described as follows:

Can we reduce the value of D[x] for each x in the set X \ S by using
the newly added vertex y to the set S, that is, is there a cheaper path
relative to S from 1 to x by using the vertex y as the next-to-last vertex
in the path? We shall call the vertex y as the pivot vertex. Of course,
the pivot vertex changes at each iteration.

Complexity of Dijkstra’s algorithm:

The initialization of the set S takes O(1) time and the initialization of the
D array costs O(n) time. Hence the total cost of the initialization step is O(n).

TABLE 1.4: Execution of Dijkstra’s algorithm on the graph of

Figure 1.14
Iteration S = X y S ← S ∪ {y} D[2] D[3] D[4] D[5] D[6]
number
0(Initial) – – {1} 4 3 11 ∞ 7
1 yes 3 {1, 3} 4 3 10 ∞ 7
2 yes 2 {1, 3, 2} 4 3 9 ∞ 6
3 yes 6 {1, 3, 2, 6} 4 3 9 14 6
4 yes 4 {1, 3, 2, 6, 4} 4 3 9 13 6
5 yes 5 {1, 3, 2, 6, 4, 5} 4 3 9 13 6
Exit loop no
28 Discrete Mathematics

The “while” loop is executed exactly n − 1 times, since the number of

vertices in the set S is incremented exactly by one at each iteration and the
loop terminates as soon as |S| = n. Statement (1) requires O(n) steps (finding
a minimum in an array of n integers requires O(n) steps) and similarly, the
“for” loop demands O(n) time. Hence, the complexity of the algorithm is
O(n) + (n − 1)O(n) = O(n2 ). (Recall that while computing the complexity
of an algorithm, we can ignore the lower-order terms and the multiplicative
constants.)

Proof of Dijkstra’s algorithm:

Proof. We have to prove the following two properties (the loop invariant) of
the set S at the beginning of the “while” loop of the algorithm:

1. For each vertex, s in S, the cost D[s] of a cheapest path from the vertex 1
to the vertex s is known, and this path lies entirely within the set S.

2. For all t outside of the set S, the cost D[t] of the cheapest path relative
to S is known.

Proof proceeds by the induction on the number of elements of S.

Basis. |S| = 1. If S = {1}, then clearly, the cheapest path from 1 to itself has
no arcs and hence its cost is zero. Further, to each vertex t outside of the set
S, a path relative to S is simply an arc (1, t) and its cost is w(1, t). Thanks to
the initialization of S and the array D, the two properties of S are satisfied.
Induction hypothesis: Suppose the two properties for the set S with |S| =
k < n, that is, the properties are true after (k − 1)th iteration. We shall prove
the two properties after performing the kth iteration.
During the k-th iteration, the pivot vertex y is added to the set S. We claim
that D[y] is the cost of the shortest dipath from 1 to y. Otherwise, there is
a shorter dipath P to the vertex y. By the choice of the pivot vertex y, this
path P must have a vertex = y outside of S (see Figure 1.16).
In Figure 1.16, illustrating Dijkstra’s algorithm, the wavy lines represent
directed paths. The directed path P starts from the source vertex 1 passing

FIGURE 1.16: A step in the proof of Dijkstra’s algorithm.

 Graph Algorithms I 29

through the vertex s, then goes directly to the vertex z, then again goes
possibly inside the set S and finally reaching the vertex y. The directed path
from the vertex 1 to the vertex y relative to the set S goes through the vertex t
and then directly reaches the vertex y from the vertex t through an arc.
Let z be the first vertex of path P outside of S. Then, the subpath P (1, z)
of P from 1 to z is a path relative to S and its cost D[z] is clearly strictly less
than the length of the path P which is strictly less than the number D[y], by
induction hypothesis. This is impossible since D[y] was a minimum among all
vertices outside of the set S during the (k + 1)-th iteration (statement number
(1) of the procedure Dijkstra). Note that we use here the fact that the costs
of the arcs are non-negative numbers. Otherwise, our argument will not be
valid. Hence the first property of the set S after the iteration number k + 1.
The second property of S remains true even after adding the pivot ver-
tex, y, because of statement number (4) of the procedure Dijkstra of Section
1.3.3, which adjusts the cost of each D[x], x ∈ X \ S to take into account the
arrival of the pivot vertex into the set S.
Recovering paths in Dijkstra’s algorithm:
Dijkstra’s algorithm finds the cost of a cheapest path from the source vertex 1
to all vertices of a weighted digraph. We are now interested in finding paths
realizing the minimum costs.
An arborescence can be conveniently represented by an array “dad” where
dad[i] gives the father of the vertex i. Stated differently, dad[i] = j if and only
if (i, j) is an arc of the arborescence. Initially, the paths relative to S = {1}
are the arcs (1, i). Hence, we add the following loop in the initialization part:

for i := 2 to n do
dad[i] := 1; (* The vertex 1 is the root of the arborescence
and hence has no dad*)

When and where to update the array dad?

Consider statement number (4) of Dijkstra’s algorithm. In this assignment, if
D[y] + w(y, x) < D[x], that is, if we could use the pivot vertex y as the next-
to-last vertex to reach x via a path from the source vertex 1, then we update
dad[x] := y. Hence, we assign and replace the “for” loop numbered (3) and the
statement numbered (4) of the procedure Dijkstra 1.3.1 by the following loop:

for each x in X - S do
if D[y] + w(y, x) < D[x]
then
begin
D[x] := D[y] + w(y, x);
dad[x] := y;(* update dad of x*)
end;
30 Discrete Mathematics

Once we have recorded the dads of different vertices, a shortest path from the
vertex 1 to any vertex i can be printed by “climbing” up the arborescence
from the vertex i till the root 1. The following procedure prints the vertices
of a shortest path from the vertex 1 to the vertex i. The vertices which are
encountered while climbing up the tree will be pushed into the array a and
finally will be popped to get the right order of vertices on the path from the
vertex 1 to the vertex i.

procedure path ( i : vertex);

var a : array[vertex] of vertex;
k : integer;
begin(* path*)
k := 1;
while i < > 1 do
begin
a[k] := i; k := k + 1;
i := dad[i];
end;
a[k] := 1;
(* print the vertices of the path*)
for i := k down to 1 do
write(a[i], ’ ’);
end;(* path *)

We shall now write a computer program in C to implement Dijkstra’s

algorithm. The given weighted graph is represented by its cost matrix M
where M [i][j] is the cost of the arc/directed edge (i, j). If there is no edge
between the vertices i and j, we set M [i][j] = ∞. In C, ∞ can be replaced by
the constant IN T M AX if costs of the edges are “small” integers. In fact, in
a computer implementation of Dijkstra’s algorithm, ∞ is replaced by a cost
which cannot be a “legal” cost of any directed edge. We define two arrays
D (for Distance) and S. D[i] represents the minimum cost of a directed path
from the root vertex 1 to the vertex i with respect to the set S. S[i] = 1 if
i ∈ S and S[i] = 0 if i ∈
/ S. dad[i] is the father/mother of the vertex i in the
arborescence under construction.

Dijkstra’s algorithm in C:
#include<stdio.h>
#include<stdlib.h>
#include<limits.h>
#define nmax 10//maximum vertices
int M[nmax][nmax];//cost matrix
int n,m;//number of vertices and arcs
void dijkstra(){

int D[n+1],S[n+1],dad[n+1],stack[n+1];
//S[i]=1 if i in S, otherwise S[i]=0.
 Graph Algorithms I 31

//D, the distance array Dijkstra

int p,i,j,min,k,head=0;//p, the pivot vertex
// in the pseudo-code we have used y for pivot instead of p
S[1]=1;//arborescence consists of only one vertex
for (i=2; i<=n;i++)
{
S[i]=0;
D[i]=M[1][i];
dad[i]=1;
}
for (i=2;i<=n;i++)
{
k=1;//find the least vertex k outside S
while (S[++k] !=0);
p=k;min=D[p];
for (j=k+1;j<=n;j++)
if(S[j]==0)
if (D[j]<D[p])
{p=j;
min=D[p];
}
S[p]=1;//add p to S
//update arrays D, dad
for (j=2; j<=n;j++)
if (S[j]==0)
if ((D[p]!=INT_MAX) && (M[p][j] !=INT_MAX))
if (D[p] +M[p][j]< D[j])
{ D[j]= (D[p]+M[p][j]);
dad[j]=p;
}}

for (i=2;i<=n;i++)
{printf("%5d ",D[i]);
}

//print a shortest path from 1 to a vertex i

printf("enter a vertex to which you want to print a minimum
path\n");
scanf("%d",&i);
//climb the arborescence from i to the root 1
//and stack the vertices encountered
while(i!=1)
{
stack[++head]=i;
i=dad[i];}
32 Discrete Mathematics

stack[++head]=i;//i=1
//print the path from 1 to i
printf("the path is ");
for(i=head;i>0;i--)
printf("%d ",stack[i]);

void graph(){

int i,j,x,y,c;
printf("enter n",\n);
scanf("%d",&n);
printf("\n");
printf("enter m",\n);
scanf("%d",&m);

for(i=1;i<=n;i++)
for(j=1;j<=n;j++)
{
M[i][j]=INT_MAX;
if(i==j)
M[i][j]=0;
}

for(i=1;i<=m;i++)
{
printf("enter the arc %d",i) ;
scanf("%d %d",&x,&y) ;
printf("enter the cost %d",i) ;
scanf("%d",&c) ;
M[x][y]=c;
}}
void print_graph()
{
int i,j;

for(i=1;i<=n;i++)
{
for(j=1;j<=n;j++)
printf("%5d ",M[i][j]);
printf("\n\n\n");
}

}
 Graph Algorithms I 33

int main()
{
graph();
print_graph();
dijkstra();
return 0;
}

1.4 Dijkstra’s Algorithm for Negative Weighted Arcs

Dijkstra’s algorithm does not work if some of the arcs are associated with
negative weights. We shall consider Figure 1.17.
Clearly, in Figure 1.17, the cheapest cost of a path from the source 1 to the
vertex 3 is 3 − 2 = 1. But, we shall see that Dijkstra’s algorithm as presented
above will give the answer 2, which is false. This is an example where the
greedy method does not always work!
Let us execute the algorithm on the graph of Figure 1.17. The following
table traces the different steps during execution. (Table 1.5).
The result of the execution can be read from columns 5 and 6 of the
iteration number 2. Column 6 of the iteration number 2 says that the cost of
a shortest path from the vertex 1 to the vertex 2 is 2, which is not correct.

FIGURE 1.17: Illustration of Dijkstra’s algorithm for negative weighted

arcs.

TABLE 1.5: Execution of Dijkstra’s algorithm on

Figure 1.17
Iteration S = X y S ← S ∪ {y} D[2] D[3]
number
0(Initial) – – {1} 3 2
1 yes 3 {1, 3} 3 2
2 yes 2 {1, 3, 2} 3 2
Exit loop no
34 Discrete Mathematics

Modification of Dijkstra’s algorithm for negative weight arcs:

Consider a directed graph with weight (possibly negative) associated to each
arc; however, the graph has no negative circuits, that is, the graph does not
possess a circuit with the sum of the weights of its arcs < 0.
Note the following fact:

Fact 1.1. In the case of graphs with negative-weight arcs, the “greedy” state-
ment that the shortest path among the shortest paths relative to S is a shortest
path in G is no longer valid!

The following algorithm computes the cost of a shortest path from the
source vertex 1 to all other vertices of the graph. As in Dijkstra’s algorithm
for graphs with non-negative weights, we use the variables S, the array D,
and the variable vertices x, y and the integer variable i.

procedure Dijkstra( var W : Matrix; var D : T_Array);

(* Dijkstra takes as input a matrix whose entries can be < 0
but no negative cycles and it returns an array D
of cheapest cost paths.*)
var S : set of vertices; x, y : vertex;
begin(* Dijkstra*)
(* Initialization of S and D*)
S := {1};
D[1] := 0;(* The cheapest cost from 1 to itself of 0*)
for i := 2 to n do
D[i] := w(1, i);(* The shortest path relative to S = {1}
is the arc (1, i). *)
(* iteration *)
while S not equal to X do
begin (* while *)
(1) choose a vertex y in X - S such that D[y] is minimum;
(2) add y to the set S;
(3) for each x in X with D[x] > D[y] + w(y, x) do
begin(* for *)
(4) D[x] := D[y] + w(y, x);
(5) S := S - {x};(* x not stabilized. so delete x
from S *)
end;(* for *)
end;(*while*)
end;(* Dijkstra *)

Let us execute the negative-weight version of Dijkstra’s algorithm on the

graph of Figure 1.17. The following table illustrates the different steps during
execution (Table 1.6).
The final result of the execution is found in columns 5 and 6 of the last
iteration number 3, which is the correct answer.
 Graph Algorithms I 35

TABLE 1.6: Execution of improved Dijkstra’s algorithm on

Figure 1.17
Iteration S = X y S ← S ∪ {y} D[2] D[3] S ← S \ {x}
number
0(Initial) – – {1} 3 2
1 yes 3 {1, 3} 3 2 {1, 3}
2 yes 2 {1, 3, 2} 3 1 {1, 2}
3 yes 3 {1, 2, 3} 3 1 {1, 2, 3}
Exit loop no

1.5 All-Pairs Shortest Path Algorithm

In this section, we may consider without loss of generality directed graphs
with no multiple arcs, that is, 1-graphs, without loops. We study all-pairs
shortest path algorithm due to Floyd. Floyd’s algorithm finds the cost of a
cheapest path between any two vertices of a directed graph with a non-negative
weight associated to each arc. We use “cheapest cost path” as synonym for
the “shortest path.” “Path” always means in this section “directed path.”
Input: A directed graph with a non-negative weight associated to each arc.
As in Dijkstra’s algorithm, the vertices represent the different cities and the
weight of an arc (i, j) represents the cost of traveling from the city i to the city
j along the arc (i, j). As in Dijkstra’s algorithm, the given graph is represented
by a weighted matrix W = (wij )n×n as follows:
⎧
⎨0 if i = j
wij = ∞ if ij is not an arc
⎩
weight of the arc ij otherwise

In a computer program implementing Floyd’s algorithm, the symbol ∞ will be

replaced by a large number which cannot be a legal cost of traveling from city i
to city j. For example, if the weights of arcs are integers, we may replace the
symbol ∞ by INT_MAX in the language C or maxint in the language Pascal.
We assume that the vertex set X = {1, 2 . . . , n}.
Output: The cheapest cost of a directed path from city i to city j for all
1 ≤ i, j ≤ n. The cost of a directed path is the sum of the cost of its arcs.
Algorithm 1.1: The first algorithm which comes to our mind is to invoke
Dijkstra’s algorithm n times for an n vertex directed graph by varying the
source vertex among all vertices of the graph.
The complexity of calling n times Dijkstra’s algorithm is nO(n2 ) = O(n3 ),
since the complexity of Dijkstra’s algorithm is O(n2 ).
We will now see one direct algorithm to compute the shortest directed
path between any two vertices of a directed weighted graph.
36 Discrete Mathematics

Algorithm 1.3 (Floyd’s Algorithm). For a directed path from the vertex
i to the vertex j, the set of vertices of the path other than the initial vertex of
the path i and the final vertex of the path j are called its intermediate vertices.
The algorithm constructs a sequence of n + 1 matrices M0 , M1 , . . . , Mn
(k) (k)
where in the matrix Mk = (mij )n×n where mij = The cost of a shortest
directed path from the vertex i to the vertex j not passing through any inter-
mediate vertices k + 1, k + 2, . . . , n, that is, the cost of a cheapest dipath from i
to j where the set of its intermediate vertices is a subset of the set {1, 2, . . . , k}.

The subscript k in the matrix Mk represents the matrix obtained after k

(k)
iterations and the superscript k in the entry mij is the (i, j) entry of the
matrix Mk after k iterations. These subscripts and superscripts will not be
present while expressing Floyd’s algorithm, and we will see that we can work
with only one copy of the matrix M.
With this definition of Mk for 0 ≤ k ≤ n, let us interpret the matrices M0
and Mn .

Interpretation of M0 :
(0) (0)
M0 = (mij )n×n , where mij is the weight of a shortest dipath from the ver-
tex i to the vertex j not passing through any intermediate vertices 1, 2, . . . , n.
(0)
Since the vertex set of the graph is X = {1, 2, . . . , n}, this means that mij
is the weight of the shortest dipath not using any intermediate vertices at
(0)
all, that is, mij is the weight of (i, j), if (i, j) is an arc. Hence, M0 = W,
the weighted adjacency matrix. This assignment is used as initialization in
Floyd’s algorithm.

Interpretation of Mn :
(n) (n)
By definition, Mn = (mij ), where mij is the weight of a shortest path from
the vertex i to the vertex j not passing through any intermediate vertices of
the set {n+1, n+2, . . .}. But this set is the empty set ∅, because the vertex set
(n)
of the graph is X = {1, 2, . . . , n}. Hence, mij is simply the cost of a shortest
dipath whose internal vertices are from the set of vertices {1, 2, . . . , n} of the
(n)
graph G. This means that mij is the cost of a shortest dipath in the graph G.
Therefore, the matrix Mn is the desired result of the algorithm.
(k)
We shall now see how to find the matrix Mk = (mij ), given that the
(k−1)
matrix Mk−1 = (mij ), that is, how to move from the (k − 1)-th iteration
to the k-th iteration?
To do this, we again exploit the definition of the matrices Mk−1 and Mk .
Consider a shortest cost path P(k−1) (i, k) from i to k not passing through any
intermediate vertices k, . . . , n and a shortest cost path P(k−1) (k, j) from the
vertex k to the vertex j not passing through any intermediate vertices k, . . . , n.
For the path P(k−1) (i, k), k is the final vertex and for the path P(k−1) (k, j) the
vertex k is the initial vertex. The concatenation or juxtaposition of the dipaths
P(k−1) (i, k) and P(k−1) (k, j) taken in this order which gives us either a directed
elementary path from i to j with k as an intermediate vertex (see Figure 1.18)
 Graph Algorithms I 37

FIGURE 1.18: A step illustrating Floyd’s algorithm.

or there is no directed elementary path (but there will be a simple directed

path but not elementary, since there is a directed circuit passing through k in
the concatenation) from vertex i to vertex j passing through the vertex k.
If the dipaths P(k−1) (i, k) and P(k−1) (k, j) are internally vertex-disjoint,
then their concatenation will still give us a directed path from the vertex i
to the vertex j for which the set of intermediate vertices is a subset of
{1, 2, . . . , k}, that is, this concatenated path includes the vertex k as an inter-
mediate vertex and other intermediate vertices p satisfy the inequalities 1 ≤
p ≤ k − 1. Hence we derive the following formula to compute the entries of the
(k) (k−1)
matrix Mk = (mij ) in terms of the entries of the matrix Mk−1 = (mij ).
 
(k) (k−1) (k−1) (k−1)
mij = min mij , mik + mkj , for 1 ≤ i, j ≤ n.

In words, the above formula can be stated as follows:

The cost of a shortest path from the vertex i to the vertex j not passing
through any intermediate vertices higher than k is the minimum of the cost
of a shortest path from i to j not passing through any intermediate vertex
higher than k − 1, and the cost of a shortest path from vertex i to k not
passing through any intermediate vertices higher than k − 1, plus the cost
of a shortest path from k to j not passing through any intermediate vertices
higher than k − 1.
(k) (k−1)
By setting i = k in the above formula, we have mkj = mkj since
(k−1) (k) (k−1)
mkk = 0. Similarly, by setting j = k in the formula, we get mik = mik .
This means that during the kth iteration of Floyd’s algorithm, that is, during
the computation of the entries of the matrix Mk , the entries of the kth row
and the entries of the kth column of the matrix Mk−1 will remain unchanged.
Hence, we can carry out our computation in Floyd’s algorithm with only one
copy of the matrix M.
The vertex k during the kth iteration is called the pivot vertex. This vertex
k is equivalent to the pivot vertex y in Dijkstra’s algorithm. We now write the
algorithm:

procedure Floyd(var W, M : Matrix);

(* Floyd takes as input parameter a matrix W which is the
weighted adjacency matrix of a directed weighted graph. The
38 Discrete Mathematics

procedure returns as output parameter the matrix M of the cost

of a shortest directed path of the graph G. *)
var i, j, k : integer; (* i, j, k for the loops *)
begin (* Floyd *)
(* Initialization of M = W *)
for i := 1 to n do
for j := 1 to n do
M[i, j] := W[i, j];
[* iteration *)
for k := 1 to n do
(* compute the matrix M sub k *)
for i := 1 to n do
for j := 1 to n do
if M[i, k] + M[k, j] < M[i, j]
then M[i, j] := M[i, k] + M[k, j];
end;(* Floyd *)

The complexity of Floyd’s algorithm:

The initialization part has two nested loops, and each loop is executed exactly
n times. Hence, the number of steps needed for the initialization part is O(n2 ).
The iteration part is constituted by three nested loops and the conditional
statement inside obviously requires O(1) step. Therefore, the iteration part
requires O(n3 ) steps.
Hence, the time taken by Floyd’s algorithm is O(n2 ) + O(n3 ) = O(n3 ).

Proof of Floyd’s algorithm:

Proof. The proof is by induction on k. Base: k = 0. By the definition of the

matrix M0 , we have the equality M0 = W, where W is the weighted adjacency
matrix of the given graph. Hence, we have correctly initialized the matrix M .
(k−1)
Induction Hypothesis: Suppose the entries of the matrix Mk−1 = (mij )
satisfy the following property:
(k−1)
mij = The cost of a cheapest path from the vertex i to the vertex j
not passing through any intermediate vertex > k − 1, for 1 ≤ i, j ≤ n.
(k)
We shall show that the entries of the matrix Mk = (mij ) obtained after k
(k)
iterations are such that mij is equal to the cost of a cheapest path from the
vertex i to the vertex j not passing through any intermediate vertices > k.
The induction step follows because of the fact that the algorithm simply
implements the formula
 
(k) (k−1) (k−1) (k−1)
mij = min mij , mik + mkj .

Thus, the proof is complete.

 Graph Algorithms I 39

FIGURE 1.19: A digraph on which Floyd’s algorithm is executed.

Let us execute Floyd’s algorithm on a given graph.

Example 1.3
Execution of Floyd’s algorithm:
Consider the graph of Figure 1.19.
The graph is represented by the weighted adjacency matrix W where

1 2 3 4 5
⎛ ⎞
1 0 7 10 ∞ ∞
2⎜
⎜∞ 0 2 ∞ 4 ⎟
⎟
W = 3⎜
⎜ ∞ ∞ 0 ∞ 8 ⎟
⎟.
4⎝∞ ∞ 8 0 ∞⎠
5 3 ∞ ∞ 3 0

The indices in the entry represent the last iteration number that changed
the entry. These indices are useful in finding the paths realizing the min-
imum costs. Initially, M0 = W. Can we use the vertex 1 as an interme-
diate vertex to reduce the entries of the matrix M1 ?
Iteration 1: ⎡ ⎤
0 7 3 ∞ ∞
⎢∞ 0 2 ∞ 4 ⎥
⎢ ⎥
⎢
M1 = ⎢∞ ∞ 0 ∞ 8 ⎥ ⎥.
⎣∞ ∞ 8 0 ∞ ⎦
3 101 131 3 0
Can we use some vertices of the set {1, 2} as intermediate vertices to
reduce the entries of the matrix M1 ?
Iteration 2: ⎡ ⎤
0 7 92 ∞ 112
⎢∞ 0 2 ∞ 4 ⎥
⎢ ⎥
⎢
M2 = ⎢∞ ∞ 0 ∞ 8 ⎥ ⎥.
⎣∞ ∞ 8 0 ∞ ⎦
3 101 122 3 0
40 Discrete Mathematics

Can we use some vertices of the set {1, 2, 3} as intermediate vertices

to reduce the entries of the matrix M2 ?
Iteration 3: ⎡ ⎤
0 7 92 ∞ 112
⎢∞ 0 2 ∞ 4 ⎥
⎢ ⎥
M3 = ⎢ ⎢ ∞ ∞ 0 ∞ 8 ⎥.
⎥
⎣∞ ∞ 8 0 163 ⎦
3 101 122 3 0
Can we use some vertices of the set {1, 2, 3, 4} as intermediate ver-
tices to reduce the entries of the matrix M3 ?
Iteration 4: ⎡ ⎤
0 7 92 ∞ 112
⎢∞ 0 2 ∞ 4 ⎥
⎢ ⎥
⎢
M4 = ⎢∞ ∞ 0 ∞ 8 ⎥ ⎥.
⎣∞ ∞ 8 0 163 ⎦
3 101 114 3 0
Can we use some vertices of the set {1, 2, 3, 4, 5} as intermediate vertices
to reduce the entries of the matrix M4 ?
Iteration 5: ⎡ ⎤
0 7 92 145 112
⎢ 75 0 2 75 4 ⎥
⎢ ⎥
⎢
M5 = ⎢115 185 0 115 8 ⎥ ⎥.
⎣195 265 8 0 163 ⎦
3 101 114 3 0
The final result is made available by the entries of the matrix M5 .

Recovering paths in Floyd’s algorithm:

Let us again refer to Example 1.3. As we have already remarked, the subscripts
in the entries of the final matrix represent the last iteration number which has
modified the corresponding entry. We shall denote the entries corresponding
to the subscripts of the final matrix M5 by INTER (for intermediate vertex)
where

0 if there is no subscript in the entry (i,j) of M5
INTER[i, j] =
subscript of the entry (i,j) of M5

With this definition, the matrix INTER of the graph Figure 1.19 is given by
⎡ ⎤
0 0 2 5 2
⎢5 0 0 5 0 ⎥
⎢ ⎥
INTER = ⎢
⎢5 5 0 5 0 ⎥
⎥
⎣5 5 0 0 3 ⎦
0 1 4 0 0
 Graph Algorithms I 41

This matrix INTER will be used to recover paths realizing the cheapest costs.
Note that if an (i, j) entry of the matrix INTER is 0, then the path realizing
the minimum cost from the vertex i to the vertex j is just the arc (i, j), with
no intermediate vertices.
For example, the cost of a cheapest path from the vertex 4 to the vertex
2 is the (4,2) entry of the matrix M5 , which is 26. What are the intermediate
vertices of a path from 4 to 2 realizing this cost 26? To find the intermediate
vertices, we read first the (4,2) entry of the matrix INTER which is 5. This
means that 5 is an intermediate vertex between 4 and 2. Now we read the
entry (4,5) of the matrix INTER to find an intermediate vertex between 4
and 5 which is 3. Hence, 3 is an intermediate vertex between 4 and 5. Next,
we read the entry (4,3) which 0. This means that there is no intermediate
vertex between 4 and 3. We read the entry (3,5) which is again 0. We read the
entry (5,2) which is 1. This means that 1 is an intermediate vertex between
5 and 2. Now the entries (5,1) and (1,2) are 0 meaning that no intermediate
vertices are found between 5,1 and 1,2.
Hence, the intermediate vertices of a cheapest path from 4 to 2 are 3, 5, 1
and the cheapest cost path is (4, 3, 5, 1, 2).
Let us now write an algorithm to find the matrix INTER. To do this, we
have only to initialize the matrix INTER and update the entries of INTER
at the appropriate point in Floyd’s algorithm.

procedure Floyd( var W, M, INTER : Matrix);

(* Floyd takes as input parameter a matrix W which is the
weighted adjacency matrix of a directed weighted graph G. The
procedure returns as output parameters the matrix M giving the
cost of a shortest directed path of the graph G and the matrix
INTER with its (i, j) entry, the last iteration number k which
has changed the (i, j) entry of M *)
var i, j, k : integer; (* i, j, k for the loops *)
begin (* Floyd *)
(* Initialization of M = W and INTER = 0 *)
for i := 1 to n do
for j := 1 to n do
begin
M[i, j] := W[i, j];
INTER[i, j] := 0;
end;
(* iteration *)
for k := 1 to n do
(* compute the matrix M sub k *)
for i := 1 to n do
for j := 1 to n do
if M[i, k] + M[k, j] < M[i, j]
then
42 Discrete Mathematics

begin(* we use k to reduce the cost*)

M[i, j] := M[i, k] + M[k, j];
INTER[i, j] := k;(* update INTER[i, j]*)
end;
end;(* Floyd *)

We shall now write a recursive algorithm to print the intermediate vertices

of a shortest path from the vertex i to the vertex j using the matrix INTER.

procedure interpath(i, j : integer);

(* print_path writes the intermediate vertices of a shortest path
from the vertex i to the vertex j using the matrix INTER *)
var k : integer;
begin
k := INTER[i, j];
if k = 0 then
return;
interpath(i, k);(* recursive call writing intermediate vertices
between i and k *)
write(k, ’ ’);
interpath(k, j);(* recursive call writing intermediate vertices
between k and j *)
end;

The following example illustrates the call of procedure “interpath(4, 2)”

on the graph Figure 1.19 with the help of the matrix INTER.

Example 1.4
Hand simulation of the procedure call “interpath(4, 2)”:
(see the tree of Figure 1.20).
The call interpath(4, 2) results in two calls: interpath(4, 5) and inter-
path(5, 2).
The call interpath(4, 5) leads to two calls interpath(4, 3) and inter-
path(3, 5). The calls interpath(4, 3) and interpath(3, 5) lead to no calls,
since INTER[4, 3] = INTER[3, 5] = 0. This results in printing of the
vertices : 3, 5
Now interpath(5, 2) calls interpath(5, 1) and interpath(1, 2).
The calls interpath(5, 1) and interpath(1, 2) lead to no calls, since
INTER[5,1] = INTER[1, 2] = 0. This leads to printing of the vertex: 1.
Hence, the intermediate vertices of a shortest path from the vertex
4 to the vertex 2 are 3, 5, 1.
 Graph Algorithms I 43

FIGURE 1.20: A tree illustrating different recursive call interpath(4,2).

1.5.1 An application of Floyd’s algorithm

Radius, diameter, and center of a graph:
Consider a directed graph G = (X, U ) with a non-negative weight associated
with each arc. The distance between two vertices x and y denoted by d(x, y)
is the cost of a shortest directed path from the vertex x to the vertex y. The
cost of a directed path is the sum of the cost of its arcs. If there is no path
from x to y, we set d(x, y) = ∞.
The eccentricity e(v) of a vertex v is the maximum of the lengths of the
paths from the vertex v to any vertex of the graph. Symbolically, e(v) =
max(d(v, w) | w ∈ X).
The radius of the graph G, denoted by r(G), is the minimum taken over
all eccentricities of the vertices of G, that is, r(G) = min(e(v) | v ∈ X). The
diameter d(G) of G is the maximum of all eccentricities of the vertices, that
is, d(G) = max(e(v) | v ∈ X). In other words, the diameter of a graph G is
the maximum of the distances between any two vertices of G.
Finally, the center of the graph G is the set of all vertices whose eccen-
tricities coincide with the radius of the graph G. These notions arise from
the telecommunication problems. The graph of Figure 1.21 illustrates these
notions.

FIGURE 1.21: Graph illustrating eccentricities, radius and diameter.

44 Discrete Mathematics

TABLE 1.7: Eccentricity table of graph of Figure 1.21

1 2 3 4 5
e 5 9 7 4 3

The eccentricities of different vertices of the graph of Figure 1.21 are given
by the Table 1.7

Example 1.5:
Radius, Diameter and Center of graph of Figure 1.21.
The graph is represented by its weighted adjacency matrix W where

1 2 3 4 5
⎛ ⎞
1 0 2 ∞ 3 ∞
2⎜
⎜∞ 0 3 ∞ ∞⎟⎟
W = 3⎜
⎜∞ ∞ 0 3 ∞⎟⎟
4⎝∞ ∞ ∞ 0 2 ⎠
5 1 2 2 3 0

Then, we apply Floyd’s algorithm to the matrix to obtain the matrix

M where
1 2 3 4 5
⎛ ⎞
1 0 2 5 5 ∞
2⎜
⎜9 0 3 6 8 ⎟
⎟
M = 3⎜ ⎜6 7 0 3 5 ⎟
⎟
4⎝3 4 4 0 2 ⎠
5 1 2 2 3 0
Then, the eccentricity e(i) of the vertex i is simply the maximum of
the coefficients of the ith line of the matrix M. Finally, the radius of
the graph is the minimum of the eccentricities and the diameter is the
maximum of the eccentricities. In our example, the radius of the graph
is 3 and the diameter is 9. The center of the graph is the set {5}.

Let us now write an algorithm to the radius, diameter and the center of a
graph.

Algorithm to find the radius, diameter and the center of a given

graph G:
Floyd(W, M); (* call to Floyd’s algorithm *)
(* now M, the cheapest cost matrix of G *)
for i:= 1 to n do
begin
 Graph Algorithms I 45

(* find e[i] *)
e[i] := M[i, 1];(* initialization of e[i] *)
(* iteration *)
for j := 2 to n do
if e[i] < M[i,j] then
e[i] := M[i,j];
end;
(* We have found the eccentricity table e *)
(* Initializing r and d *)
r := e[1]; d := e[1];
(* iteration *)
for i:= 2 to n do
begin
if r > e[i] then
r := e[i];
if d < e[i] then
d := e[i];
end;
(* finding the center C*)
(* initialization of C*)
for i: = 1 to n do
C[i] := 0; (* C is empty *)
(* iteration *)
for i := 1 to n do
if r = e[i] then
C[i] := 1;
(* print center *)
(*C[i] = 1 if i in C, 0 otherwise *)
for i := 1 to n do
if C[i] =1 then write (i, ’ ’);

The complexity of the algorithm for finding the radius, diameter,

and the center:
A call to Floyd requires O(n3 ) steps. We have to scan every entry of the
matrix M to find the radius and diameter. This needs O(n2 ) steps. Finally
finding the center requires only O(n) steps as it scans the array e.
Hence the complexity is O(n3 ) + O(n2 ) + O(n) = O(n3 ).

1.6 Transitive Closure of a Directed Graph

Sometimes, we may be interested in knowing if there is a directed path
from the vertex i to the vertex j of a directed graph. Transitive closure of the
graph helps us answer this question directly.
46 Discrete Mathematics

FIGURE 1.22: A graph to illustrate transitive closure.

FIGURE 1.23: The transitive closure of the graph of Figure 1.22.

Consider a directed graph G = (X, U ). The transitive closure of the

graph G denoted by G+ is a graph whose vertex set is the same as that of
the graph G, that is, X, and two vertices i and j (not necessarily distinct)
of the transitive closure are joined by an arc (i, j) if there is a directed path
of length at least one in the original graph G.

Example 1.6
Transtive closure of a graph: Consider the graph G of Figure 1.22.
This graph is equivalently represented by its adjacency matrix M
where
1 2 3 4 5
⎛ ⎞
1 0 1 0 0 0
2⎜⎜0 0 1 1 0⎟
⎟
M = 3⎜ ⎜ 0 0 0 0 0 ⎟
⎟
4⎝0 1 0 0 1⎠
5 1 0 0 0 0
 Graph Algorithms I 47

The transitive closure of the graph of Figure 1.22 is drawn in Figure

1.23.
The adjacency matrix of the transitive closure of the graph G is the
matrix T M where
1 2 3 4 5
⎛ ⎞
1 1 1 1 1 1
2⎜⎜1 1 1 1 1⎟⎟
TM = 3 ⎜
⎜0 0 0 0 0⎟⎟
4⎝1 1 1 1 1⎠
5 1 1 1 1 1

For example, the (4, 4) entry of T M is 1 because there is a directed path

from the vertex 4 to itself, namely (4, 5, 1, 2, 4), the (5, 3) entry of T M
is one because of the directed path (5, 1, 2, 3), whereas the (3, 1) entry
of T M stays as 0 because there is no directed path in the graph G from
the vertex 3 to the vertex 1.

1.7 An O(n3 ) Transitive Closure Algorithm Due to

Warshall
Input: A directed graph G = (X, U ) with X = {1, 2, . . . , n}. The graph is
represented by its adjacency matrix M.
Output: The adjacency matrix T M of the transitive closure of the graph G.
Algorithm: This algorithm is invented by Warshall. This algorithm is a special
case of Floyd’s algorithm but it predates Floyd’s algorithm.

The algorithm constructs a sequence of matrices M0 , M1 , . . . , Mn where

(k)
the matrix Mk = mij for 1 ≤ k ≤ n is defined as follows:
⎧
⎨1 if there is a path from i to j not passing through any
mij = intermediate vertex > k
⎩
0 otherwise
(0)
Interpretation of M0 = (mij ):
(0)
By the above definition of the entries of the matrix M0 , mij is equal to 1
if there is a directed path from i to j not passing through any intermediate
vertex > 0 and is equal to 0, otherwise. Since the smallest numbered vertex is
(0)
1, this means that mij is simply 1 if there is an arc from i to j, zero otherwise.
This is the definition of the adjacency matrix of the graph G. Hence M0 = M.
This equation is used as initialization in Warshall’s algorithm.
48 Discrete Mathematics

(n)
Interpretation of Mn = (mij ):
(n)
By definition, mij is equal to 1 if there is a directed path from i to j not
going through any intermediate vertex > n and is equal to 0, otherwise. Since
(n)
no vertex of the graph is > n, this means that mij is equal to 1 if there is a
directed path from i to j in G and is equal to 0, otherwise.
This means that Mn is the desired output matrix T M . The subscript k in
Mk represents the iteration number.

Induction leap:
How to find the matrix Mk given the matrix Mk−1 ?
(k−1)
We are given the n2 entries mij of the matrix Mk−1 . We have to express
(k) (k−1) (k−1)
mij in terms of mik and mkj . This is because during the construction
of Mk we have the right to use the vertex k as an intermediate vertex which
is not the case with the matrix Mk−1 .
We must not disturb the entries “one” of the matrix Mk−1 , since a directed
path from i to j not going through any intermediate vertex > k − 1 is also a
directed path not passing through any intermediate vertex > k. We are only
interested in a possible reassignment of zero coefficients of the matrix Mk−1
into 1. Since the concatenation of a path from i to k not going through any
intermediate vertex > k − 1 and a path from k to j not passing through any
intermediate vertex > k − 1 is a path from the vertex i to the vertex j not
passing through any intermediate vertex > k, we have the following formula
(k)
for mij .
 
(k) (k−1) (k−1) (k−1)
mij = max mij , mik × mkj , for 1 ≤ i, j ≤ n.
(k) (k−1) (k−1)
By setting j = k in the above formula, we have mik = max(mik , mik ×
(k−1) (k−1) (k) (k−1)
mkk ) = mik and similarly by setting i = k we get, mkj = mkj .
Hence, we can carry out our computation with only one copy of the
matrix M . Let us describe the above formula in words:
There is a directed path not going through any intermediate vertex > k if:

1. There is already a path from the vertex i to j not going through any
intermediate vertex > k − 1 (this is equivalent to saying: we must not
modify the entries “one” of the matrix Mk−1 ) or
2. There is a path from i to k not passing through any intermediate ver-
tex > k − 1 and a path from k to j not going through any intermediate
vertex > k − 1.

We shall now write Warshall’s algorithm. We use the type “Matrix” to repre-
sent n × n matrices of 0 and 1.

procedure Warshall( var M, TM : Matrix);

(* The input parameter is M, the adjacency matrix of the
 Graph Algorithms I 49

given graph G. The output parameter is TM, the adjacency matrix

of the transitive closure of G. *)
var i, j, k : integer; (* i, j, k are for loops *)
begin
(* initialization of TM = M *)
for i := 1 to n do
for j := 1 to n do
TM[i, j] := M[i, j];
(* iteration *)
for k := 1 to n do (* k-th pass over TM *)
for i : = 1 to n do
for j := 1 to n do
if TM[i, j] = 0 then TM[i, j] := TM[i, k] *
TM[k, j];
end;(* Warshall *)

The complexity of Warshall’s algorithm:

The initialization of the matrix requires O(n2 ) steps, as there are two nested
loops involved. The iteration step takes O(n3 ) steps, as there are three nested
loops. Hence, the complexity of Warshall’s algorithm is O(n2 ) + O(n3 ) =
O(n3 ).
Let us execute Washall’s algorithm on a given graph.

Example 1.7
Execution of Warshall’s algorithm: Consider the graph which is an
elementary circuit of length 4 with vertex set {1, 2, 3, 4} and the arc set
{(1, 2), (2, 3), (3, 4), (4, 1)}. The adjacency matrix of the above graph is

1 2 3 4
⎛ ⎞
1 0 1 0 0
2⎜0 0 1 0⎟
M= ⎜ ⎟.
3⎝0 0 0 1⎠
4 1 0 0 0

Initialization of T M : T M = M. Iteration k = 1. We use 1 as an

intermediate vertex.
1 2 3 4
⎛ ⎞
1 0 1 0 0
2⎜0 0 1 0⎟
TM = ⎜ ⎟
3⎝0 0 0 1⎠
4 1 1 0 0
50 Discrete Mathematics

Iteration k = 2. We may use 1/2 as intermediate vertices.

1 2 3 4
⎛ ⎞
1 0 1 1 0
2⎜0 0 1 0⎟
TM = ⎜ ⎟
3⎝0 0 0 1⎠
4 1 1 1 0

Iteration k = 3. We may use 1/2/3 as intermediate vertices.

1 2 3 4
⎛ ⎞
1 0 1 1 1
2⎜0 0 1 1⎟
TM = ⎜ ⎟
3⎝0 0 0 1⎠
4 1 1 1 0

Iteration k = 4. We may use any vertices as intermediate vertices.

1 2 3 4
⎛ ⎞
1 1 1 1 1
2⎜1 1 1 1⎟
TM = ⎜ ⎟
3⎝1 1 1 1⎠
4 1 1 1 1

1.8 Navigation in Graphs

We are interested in traversing a multigraph in a systematic manner, that
is, “visiting” the vertices and edges of a multigraph. The vertices are normally
variables of type “struct” in the language C or the type “record” in Pascal.
These vertices contain information and we are interested in reaching the infor-
mation stored in the vertices to perform some kind of operation. For example,
we may want to update the information in each vertex or simply write the
contents of a particular field of each vertex. We shall study two important
graph traversal methods: The depth-first search and breadth-first search.

Depth-first search:
The depth-first search of a graph can be viewed as a generalization of preorder
traversal of a tree. If the graph does not possess a cycle, then the depth-first
search (dfs) of a graph coincides with the preorder traversal of a tree.
Let us first explain the dfs in an intuitive manner. Consider a connected
multigraph G. We start with a vertex, called the starting vertex, which will
become the root of a spanning tree to be generated by dfs. From the starting
 Graph Algorithms I 51

FIGURE 1.24: A simple graph to illustrate depth-first search.

vertex, we construct a path as long as possible. Then, we backtrack from the

current vertex, that is, we climb up the tree which is a path at present. The
climbing up continues until we find a vertex at which a side-path, that is, a new
branch, may be constructed to a “new” vertex. If the graph is not connected,
then we have to choose a new vertex of a new connected component as a
starting vertex and perform the dfs etc., until all vertices of the graph are in
some tree induced by dfs.
The dfs can be useful as a model around which one can design various effi-
cient graph algorithms like finding: connected components of a graph, bicon-
nected components, strongly connected components, finding triply connected
components, testing planarity, etc.
Dfs may be used to perform “exhaustive search” to solve TSP, permutation
generation, Knight’s tour in a Chess board, etc.
Let us illustrate the dfs with an example.

Example 1.8
dfs: Consider the following simple graph G of Figure 1.24 consisting
of two connected components.
The graph G is represented by the adjacency lists. In each adjacency
list, the vertices are listed in increasing order. We use an array “mark”
the vertices processed by the dfs. Initially, for each vertex i, we have
mark[i] = 0. We set mark[i] ← 1 as soon as the vertex i is reached
during the dfs. (0 means “unprocessed” or “unvisited” and 1 means
“processed” or “visited.”)
L(1) = (4, 6); L(2) = (5); L(3) = (6, 7); L(4) = (1, 5, 6); L(5) =
(2, 4, 6); L(6) = (1, 3, 4, 5, 7); L(8) = (9, 10), L(9) = (8, 10); L(10) =
(8, 9). Now let us take a vertex, say, the vertex 1, as the starting vertex
and set mark[1] ← 1. The adjacency list of 1, L(1) is scanned to find
a vertex marked 0. The very first vertex 4 in L(1) is chosen and we set
mark[4] ← 1. The traversing of the list L(1) is temporarily suspended
52 Discrete Mathematics

and we start scanning the list L(4) to find a vertex “unprocessed.” The
vertex 5 is found and we set mark[5] ← 1. Then, again the scanning
of the list L(4) is temporarily suspended and we start scanning the list
L(5) and we find the vertex 2 and we assign mark[2] ← 1. We now scan
the list L(2) in search of a unprocessed vertex. No new vertices are found
by scanning L(2) and we climb up the tree to find the vertex 5 which is
the “dad” of vertex 2, that is, dad[2] ← 5. We now restart the scanning
of the list L(5), which was suspended previously. We find the vertex 6,
then the vertex 3 and the vertex 7. Now all the vertices of the connected
component containing the vertex 1 have been “processed.” We now con-
sider the second connected component and “visit” each vertex according
to dfs. (See the forest F of Figure 1.25. The forest F is defined by the
solid edges.)
The edges of the forest F are drawn as continuous edges, and the
edges of the graph G not in the forest are depicted as dotted edges.
The children/sons/daughters of a vertex are drawn from left to right. A
vertex j is called a descendant of a vertex i, if there is a downward path
from the vertex i to the vertex j in a tree of the forest generated by dfs.
In this case, we also say that the vertex i is an ancestor of the vertex
j. We consider a vertex x as an ancestor and descendant of itself. A
descendant or ancestor of a vertex x other than itself is called a proper
descendant or proper ancestor of x.
We associate to each vertex i of the graph, called the depth-first
search number(dfsn), where df sn[i] = j if and only if i is the jth vertex
visited during the dfs. Table 1.8 gives the dfsn of each vertex of the
above forest F.

FIGURE 1.25: Illustration of depth-first search algorithm.

 Graph Algorithms I 53

TABLE 1.8: Depth-first search number table of graph G of Figure 1.25

1 2 3 4 5 6 7 8 9 10
dfsn 1 4 6 2 3 5 7 8 9 10

We now write the dfs algorithm in pseudo-code.

Input: A multigraph G = (X, E) represented by n adjacency lists where the
vertex set X = {1, 2, . . . , n}.
Output: A partition of the edge set E into a set of edges belonging to the
spanning tree/forest and a set of dotted edges, also called the back edges. The
algorithm also assigns to each vertex a dfs number.
Algorithm: The algorithm is described recursively. We use an array “mark”
of vertices where mark[i] = 0 if the vertex i is not yet visited by dfs and
mark[i] = 1 if the vertex i is visited by the dfs. df sn[i] = j if i is the jth
vertex visited during the search.

dfs algorithm:
(* initialization *)
for i := 1 to n do
mark[i] := 0;
F : = empty;(* F, the forest under construction*)
counter : = 1; (* counter is used to assign the dfs number *)
(* end of initialization *)
procedure dfs( i : integer);
(* dfs visits each vertex of a connected component of G
containing i using depth-first search *)
var j : integer;(* vertex j scans L(i) *)
begin
mark[i] := 1;
dfsn[i] := counter;
counter := counter + 1;
(* scan L(i) *)
for each vertex j in L(i) do
if mark[j] = 0 then
begin
add the edge ij to F;
dfs(j); (* recursive call *)
end;
(* iteration *)
for i := 1 to n do
if mark[i] = 0 then
dfs(i); (* invoking dfs *)
54 Discrete Mathematics

The edges of the graph G which are not in the spanning forest induced
by dfs are the back edges. If the graph is connected, then the dfs induces a
spanning tree, called the dfs spanning tree.
Let xy be an edge of the dfs spanning forest, that is, forest induced by
dfs, with y, a son of x. Then, dfs(x) called dfs(y) directly during the dfs. In
other words, the vertex y is marked zero, when we scan the list L(x), the list
of vertices adjacent to x.
We distinguish two types of calls:

1. Direct calls

2. Indirect calls

dfs(x) called directly dfs(y) if and only if the edge xy is an edge of the spanning
forest constructed by dfs, that is, x is the father/mother of y in the spanning
forest.
Note also that when we are at vertex x, scanning the list L(x), and z
in L(x), that is, xz is an edge (and hence, x is also in the list L(z)), with
mark[z] = 1, we cannot simply put the edge xz in the set of back edges,
because the vertex z can be the father/parent of the vertex x.
If xy is a back edge, then neither dfs(x) nor dfs(y) called the other directly,
but one called the other indirectly, that is, dfs(y) called dfs(z), which in turn
called dfs(x), and hence y is an ancestor of x.

Remark 1.2. The dfs spanning forest need not be unique. A dfs forest gen-
erated by the search procedure dfs depends on the starting vertices, that is, the
roots of each tree in the spanning forest and the order of the vertices in each
list L(i).

The complexity of the dfs:

Intuitively, during the traversal of the graph, each vertex is marked only once
and each edge xy is traversed exactly twice – once from x and once from y.
Hence the complexity of the dfs algorithm is O(n + m) = O(max(n, m)) by
the sum rule of the complexity. Here, m denotes the number of edges of the
graph G.
In a formal manner, the initialization step takes O(n) time. The iteration
part is executed n times.
The number of steps in executing the call df s(i), without taking into
account the recursive call to itself is the number of vertices in the list L(i),
which is equal to the degree of the vertex i. Thus, one call to dfs costs O(d(i)),
where d(i) is the degree of the vertex i. By Euler’s theorem, the sum of the
degrees of the vertices of a graph is twice the numberof edges. Hence, the
n
total costs of the “for” loop of the “procedure dfs” is i=1 d(i) = 2m. Note
that df s(i) is called only once, since as soon as we touch the vertex i, we
mark i “visited” and we never call dfs on a vertex with its “mark” set to 1.
Hence, the total number of steps required is O(2n + 2 m) = O(max(n, m)).
 Graph Algorithms I 55

Property of a back edge:

Property 1.2. Consider a connected graph and let T and B be the set of
tree edges and back edges obtained by dfs. Then, each back edge xy joins an
ancestor and a descendant in F. (A vertex is considered as an ancestor and a
descendant of itself.)
Proof. We may assume without loss of generality that the vertex x is “visited”
before the vertex y. This means that the vertex y is marked “unvisited” when
we reach the first time the vertex x. Since xy is an edge of the graph, we have
the vertex y in the list L(x). But then the call df s(x) would not be complete
until all the vertices of L(x) are “visited.” Hence, the vertex y is marked
“visited” between the initiation of the call dfs(x) and the completion of the
call df s(x). All vertices between the initiation of dfs(x) and the end of the
call dfs(x) will become descendants of the vertex x in the dfs spanning tree
constructed by the search. In particular, the vertex y is a proper descendant
of x which is not a son of x in the dfs spanning tree. Hence the property.
Relation “x is to the left of y”:
Let us note that the children of a vertex are in the left-to-right order in
the dfs spanning forest and the different trees of the spanning forest are also
drawn in the left-to-right order. Consider a dfs spanning tree T of a connected
graph and consider two vertices x and y which are not related by the relation
descendant-ancestor. Such vertices can be put in relation called, “is to the left
of.” We say that the vertex x lies to the left of the vertex y or y lies to the
right of x if the following condition is satisfied:
We draw the unique path from the root to the vertex y. Then, all vertices to
the “left” of this path are treated as to the left of the vertex x and all vertices
to the “right” of this path are considered as to the right of the vertex y.
If the vertex i is a proper ancestor of the vertex j in a dfs spanning tree,
then dfsn[i] < dfsn[j] and if i is to the left of j, then also dfsn[i] < dfsn[j].

Remark 1.3. The order in which the different calls terminate in the dfs is
the postorder traversal of the dfs tree.

1.9 Applications of Depth-First Search

1.9.1 Application 1: Finding connected components
Input: A multigraph G = (X, E) with X = {1, 2, . . . , n}.
Output: List of vertices of each connected component and the number of con-
nected components.
Algorithm: Of course, the connected components of a graph can be easily
found by “eyeballing” if the graph is drawn in the geometric representation
and if it is sufficiently small. On the other hand, if we are given a complex
56 Discrete Mathematics

electrical circuit, the question of whether the circuit is connected is not at all
obvious “visually.”
Our problem is to instruct a computer so that it writes different connected
components of a multigraph. The algorithm can be easily written using the
dfs. We use an integer variable “nc” to calculate the number of connected
components of G. Then, we add the following statement in the dfs algorithm.
After the assignment mark[i] := 0 at the beginning of the “procedure dfs,”
add the following print statement: write(’ ’,i); which means that as soon as we
“touch” or “visit” a vertex i, we immediately print the corresponding vertex.
Add the statement nc := 0; in the initialization part of the algorithm.
Finally we rewrite the “iteration part” of the dfs algorithm as follows:

for i := 1 to n do
if mark[i] = 0 then
begin
nc := nc + 1;(* update nc *)
dfs(i); (* call to dfs. i is a root of a sub-tree
in dfs spanning forest which is under
construction *)
writeln;(* print a new line character to separate
components*)
end;
writeln(nc);(* print the number of components *)

Justification of the algorithm: Each call dfs(i) in the “iteration part” corre-
sponds to the root i of a subtree in the dfs spanning forest. Once the call dfs(i)
is initiated, all vertices connected to the vertex i by a “downward” path in
dfs forest are marked with 1. Hence, the validity of the algorithm.

The complexity of finding connected components:

The time needed to find the connected components is the same as that of dfs,
since the instructions added are of time complexity O(1). (Recall that O(1)
represents a constant.)

1.9.2 Application 2: Testing acyclic graph

Input: A multigraph G = (X, E) with vertex set {1, 2, . . . , n}.
Output: “Yes” if the graph is acyclic, that is, if G has no cycles;
“No” otherwise. In this case, print the cycles the dfs encounters.
Algorithm: We shall use the following fact: Every back edge we meet during
the dfs search defines a cycle and conversely if a graph possesses a cycle,
then we must meet a back edge in the course of our dfs search. Why? This
is because of Property 1.2. Note that a back edge ij is encountered the first
 Graph Algorithms I 57

time in the course of the dfs search if we are at vertex i and df sn[i] ≥ df sn[j].
Of course, the same back edge ij = ji is traversed the second time when we
are again at vertex j. A loop is considered as a back edge. Let us now write
the algorithm. We shall use an array “dad” where dad[j] = i if the vertex i is
the father/parent of the vertex j in the spanning dfs forest.

dfs algorithm for cycles:

(* initialization *)
for i := 1 to n do
mark[i] := 0;
counter := 1;(* counter to assign dfs number *)
acyclic : = 0; (* acyclic is 0 if the graph has no cycles.
acyclic is 1, otherwise *)
(* end of initialization *)
procedure dfs(i:integer);
(* dfs visits each vertex of a connected component of G
containing i using depth-first search *)
var j, k : integer;(* vertex j scans L(i) *)
begin
mark[i] := 1;
dfsn[i] := counter;
counter := counter + 1;
(* scan L(i) *)
for each vertex j in L(i) do
if mark[j] = 0 then
begin
dad[j] := i;(* ij is an edge of dfs forest *)
dfs(j); (* recursive call *)
end
else (* j is marked ‘‘visited’’ *)
if ((dad[i] <> j) and (dfsn[i] > dfsn[j]))or(i = j)
(* ij is a back edge possibly a loop *)
then
begin
acyclic := 1;
(* print the cycle *)
if i = j
then
write(i,i,’ is a loop’)
else
begin
k := i;
repeat (* climb up the tree
till we meet the vertex j *)
write(k,’ ’);
k := dad[k];
until k = j;
end
58 Discrete Mathematics

writeln;(* new line to separate the cycles *)

end;
end;(* dfs *)
(* iteration *)
for i := 1 to n do
if mark[i] = 0 then
dfs(i); (* invoking dfs *)
if acyclic = 0
then
writeln (’yes’)
else
wtiteln(’no’);

The complexity of the algorithm testing for cycles:

The complexity is O(n + m), which is the same as that of dfs, since we have
only added the statements to the “else part” in the “procedure dfs” of the
“dfs algorithm.” In the worst case, the added statements cost O(n) steps.
By the sum rule [6] of big Oh notation, we may write the complexity as
O(n + m) = O(max(m, n)). (The sum rule says that, in a sum, it is enough to
take into account the “dominating term” of the sum, ignoring constants and
constant coefficients.)

1.9.3 Application 3: Finding biconnected components of a

connected multigraph
Hopcroft’s algorithm:
Consider a connected multigraph G = (X, E) with X = {1, 2, . . . , n}. We
assume that the graph G has no loops. A vertex x is a cut vertex or articulation
vertex of the graph G if the removal of the vertex x from G results in a
disconnected graph. Note that the removal of the vertex x also results in the
removal of all edges incident with the vertex x.
An edge with the analogous cohesive property is called a bridge or cut edge
or isthmus of the graph G, that is an edge e is a bridge of the graph G if the
removal of the edge e from the graph results in a disconnected graph. Note
that by removing an edge e = xy, the end vertices of the edge e will still be
in the graph G − e. It can be easily seen (by constructing examples) that the
removal of an edge results in a graph with at most two connected components,
whereas the removal of a vertex may result in a graph with several connected
components.

Remark 1.4. If we allow loops in the multigraph, then a vertex x incident

with a loop and another edge which is not a loop is considered as a cut vertex.

In the following graph of Figure 1.26, the vertices 3, 4, and 6 are cut ver-
tices and the bridges are edges 47 and 36. Note that an end vertex of a bridge
is a cut vertex if its degree is strictly more than one.
 Graph Algorithms I 59

FIGURE 1.26: Illustration of cut points and bridges.

Example 1.9: Cut vertices and bridges

Consider a tree T with at least two vertices. Each vertex of degree
one of this tree T (there must be at least two such vertices, otherwise,
δ(T ) ≥ 2 and hence the graph has a cycle [6], which is impossible) is
not a cut vertex of T and all other vertices are cut vertices of T.
On the other hand, every edge of T is a bridge.

A connected graph with at least two vertices is a biconnected graph or a

block if it does not contain a cut vertex. A biconnected component or block
of a connected graph is a maximal induced subgraph without any cut vertex.
In other words, the graph H is a biconnected component of the graph G if
H is biconnected in its own right and for any vertex x lying outside of H,
the subgraph of G induced by the vertices of H ∪ x is either disconnected or
contains a cut vertex.
A graph can be partitioned into mutually disjoint connected components.
In an analogous manner, a given connected graph can be decomposed into
mutually edge-disjoint biconnected components. Note that a graph is bicon-
nected if and only if its unique biconnected component is itself.
The nomenclature “biconnected” is justified because of the following prop-
erty enjoyed by a biconnected graph with at least three vertices:

Theorem 1.2. A connected graph with at least three vertices is biconnected

if and only if between any two distinct vertices of the graph there are at least
two internally vertex-disjoint elementary paths.

In the following graph of Figure 1.27, we have a connected graph, with

four biconnected components.
The list of edges of the three biconnected components of the graph of
Figure 1.27 are:

1. 12, 14, 24, 29, 19, 49

2. 23

3. 36, 37, 67
4. 59, 58, 89
60 Discrete Mathematics

FIGURE 1.27: Illustration of biconnected components of a graph.

The biconnected components listed above satisfy the following proper-

ties:

1. Two distinct biconnected components intersect in at most one vertex.

2. If two distinct biconnected components intersect in exactly one vertex

x, then this vertex must be a cut vertex of the graph.

The above two properties are in general true for any connected graph with
at least two vertices. Intuitively, this means that the different cut vertices of
a graph are the “frontiers” between the biconnected components.
We shall now describe the biconnected components algorithm.
Input: A connected multigraph G = (X, E) with X = {1, 2, . . . , n}.
Output: A list of edges of each biconnected component of the graph G.
Algorithm: We shall first illustrate the algorithm on the graph of Figure 1.27.
Step 1: The graph is represented by its adjacency lists. Let us assume that
the vertices are listed in the increasing order in each list L(i), that is, L(1) =
(2, 4, 9), L(2) = (1, 3, 4, 9), etc., till L(9) = (1, 2, 4, 5, 8). We choose a starting
vertex, say, the vertex 1.
Step 2: The graph is searched according to the algorithm dfs. This search
gives us a partition of the edge set into T and B, where T is the set of edges
of the spanning tree induced by the dfs and B is the set of back edges which
is equal to E \ T. During the search, each vertex is assigned a number called
the dfs number. Recall that dfsn[i] = j if and only if the vertex i is the jth
vertex “visited” in the course of the dfs (see the graph of Figure 1.28 below
obtained by dfs).
The following Table 1.9 gives us the dfsn of different vertices.
 Graph Algorithms I 61

FIGURE 1.28: Depth-first search of the graph of Figure 1.27.

TABLE 1.9: Depth-first search number (dfsn) table of the graph of

Figure 1.28

1 2 3 4 5 6 7 8 9
dfsn 1 2 3 6 8 4 5 9 7

To help in finding the biconnected components, we need to define the fol-

lowing function “LOW” on the set of vertices of the graph.

The function LOW:

LOW[i] is the smallest dfsn less than or equal to dfsn[i] one can reach either
from i or from one of the descendants of i in the dfs spanning tree using only
once a back edge. If all such back edges lead to vertices with dfsn > df sn[i],
or if there is no back edge from i or from one of its descendants of i, then we
define LOW[i] = dfsn[i].
Mathematically, we write

LOW[i] = min({dfsn[i]} ∪ {dfsn[j] | there is a back edge

from a descendant of i to j}).
The function LOW will be computed by the algorithm in postorder traversal
of the depth-first search spanning tree obtained. In our example, the postorder
traversal of the dfs tree is: (7, 6, 3, 8, 5, 9, 4, 2, 1). So, the first vertex for which
LOW is computed is the vertex 7.
LOW[7] = min(df sn[7], df sn[3]) = min(5, 3) = 3 because of the back edge 73
from the vertex 7. LOW[6] = min(df sn[6], df sn[3]) = min(5, 3) = 3 because
of the back edge 73 from the son 7 of the vertex 6.
LOW[3] = 3.
62 Discrete Mathematics

LOW[2] = min(df sn[2], df sn[1], df sn[1], df sn[2]) = 1 because of the back edges
41, 91, 92.
Note that a vertex is considered as a descendant of itself. The following table
gives the LOW function of different vertices of our example graph (Table 1.10).
As we have already remarked, the different cut vertices of the graph G serve
as “frontiers” between biconnected components. So we are first interested in
characterizing the cut vertices of the graph G. Let us again refer to the graph
of Figure 1.28. The vertex 3 is a cut vertex, whereas the vertex 6 is not. Why
is vertex 6 not a cut vertex? This is because of the following fact: There is a
back edge, namely, 73, from the son of 6, that is, from a proper descendant
of 6, to a proper ancestor of 6. This back edge gives us a second path from
3 to 7 and the first path from 3 to 7 is obtained via the tree edges. On the
other hand, there is no back edge from a proper descendant of the vertex 3
to a proper ancestor of 6. Hence, there are no two vertex-disjoint paths from
a proper descendant of 3 to a proper ancestor of 3. This makes the vertex 3,
a cut vertex of the graph G. In a similar manner we can conclude that the
vertices 2 and 9 are cut vertices.
We now state the following theorem whose proof is intuitively clear from
the above discussion. For a formal proof, the reader can consult the book [3].
Theorem 1.3. Consider a connected multigraph G and consider a partition
of the edge set of the graph G into T , a set of tree edges and B, a set of back
edges induced by a search dfs. Then, a vertex i is a cut vertex of G if and only
if it satisfies exactly one of the following properties:
1. i is the root of the tree T and the root i has at least two sons in the
spanning dfs tree T.

2. i is not the root and i possesses at least one son s in the spanning dfs
tree with LOW[s] ≥ df sn[i].

Back to our example illustrating Hopcroft’s biconnected components algo-

rithm: With this characterization of cut vertices, we continue our illustration.
We shall use a stack S of edges. Initially our stack is empty. Each edge tra-
versed the first time in the course of the dfs is pushed into the stack. Recall
that each edge of the graph is encountered twice during the search procedure.
Note that a tree edge xy is pushed into the stack in the order (father, son),
where x is the father of y, whereas a back edge da is pushed into the stack
with the vertex a an ancestor of the vertex d. Note that a loop is considered
as a back edge.

TABLE 1.10: LOW function table of the graph of Figure 1.28

1 2 3 4 5 6 7 8 9
LOW 1 1 3 1 7 3 3 7 1
 Graph Algorithms I 63

In our example, the stack will look as below when we are at vertex 7 before
backtracking from 7 to the father of 7 which is 6 (the stack grows from left
right).

S = (12, 23, 36, 67, 73).

The edge at the top of the stack, namely, edge 73, is a back edge whereas all
other edges in the stack are tree edges. Tree edges occur as a result of direct
recursive calls, that is, xy is a tree edge if and only if df s(x) calls df s(y)
directly during the search.
When we backtrack from the vertex 7, that is, when we climb up the tree
from vertex 7 to his father 6, the following test is performed.
LOW[7] ≥ df sn[6]? If the test is true, then we pop up the stack till the edge
67 (including the edge 67) to emit a biconnected component. In our example,
the test is false, since LOW[7] = 3 < df sn[6] = 4. So no pop-up occurs. We
are at the vertex 6 and the search finds no more new vertices from 6. So we
climb up the tree from 6 to the father of 6 which is 3. The following test is
performed.
LOW[6] ≥ df sn[3]? The inequality is true, since LOW[6] = 3 and
dfsn[3] = 3. Hence, we pop up the stack till and including the edge 36.
Thus, our first biconnected component is

B1 = (73, 67, 36)

and the stack will look as follows:

S = (12, 23).

We are at vertex 2 and no new vertices are found from the vertex 2. Hence
we climb up the tree from 2 to his father 1. The test LOW [2] ≥ df sn[1] turns
out to be true and we pop up the stack till the edge 23.
The second biconnected component found is

B2 = 23

and the stack

S = (12).
We are now at vertex 2. We find a new vertex 4 by scanning the list L(2) and
the edge 24 is pushed. Then, the following edges are pushed into the stack:
41, 49, 91, 92, 95, 58, 89. No new vertices are found from the vertex 8. At this
point, the stack S from left to right is

S = (12, 24, 41, 49, 91, 92, 95, 58, 89).

We backtrack from the vertex 8 to his father 5 and the test LOW[8] ≥
df sn[5] is false. We further backtrack from 5 to his father 9 and the test
LOW[5] ≥ df sn[9] turns out to be true. Hence, the stack is popped till we
arrive at the edge 95.
64 Discrete Mathematics

The third biconnected component found is

B3 = (89, 58, 95)

and the stack

S = (12, 24, 41, 49, 91, 92).
We climb up the tree from 9 to 4 (the test is false) and then from 4 to 2
(the test is false) and finally from 2 to 1. The test LOW[2] ≥ df sn[1]? is true.
Hence, the stack is popped till the edge 12. This gives us the fourth and last
biconnected component B4 where

B4 = (12, 24, 41, 49, 91, 92).

Note that the stack is finally empty.
With this hand simulation as our model, we are now almost ready to write
the Hopcroft’s biconnected components algorithm in pseudo-code. One more
final point remains. We have to express the definition of the function LOW in
a recursive manner. Since a son j of a vertex i in a dfs spanning tree T is a
proper descendant i, we rewrite LOW as follows:
part1 part2 part3
        
LOW[i] = min({df sn[i]} ∪ {LOW [j] | j is a son of i in T } ∪ {df sn[k] | ik ∈ B},
where B is the set of back edges. The search procedure dfs is augmented with
suitable statements to print different biconnected components. We use the fol-
lowing variables: the arrays mark[1..n] of 0..1, df sn[1..n] of integer, LOW[1..n]
of integer, dad[1..n] of integer. We also use a stack S whose objects are edges.
The vertices are 1, 2, . . . , n.

Biconnected components algorithm:

procedure dfs_biconn( i : integer);
var j : integer; (* j is used to scan L(i)*)
LOW, dad, dfsn : array[1..n] of integer;
begin
mark[i] := 1;
dfsn[i] := counter;
counter := counter + 1;
(* initialization of LOW[i] *)
LOW[i] := dfsn[i];(* part 1 of the def. of LOW *)
(* scan L(i) for a new vertex *)
for each j in L(i) do
if mark[j] = 0
then
begin
dad[j] := i;(* add edge ij to dfs tree *)
push ij into S;
 Graph Algorithms I 65

dfs_biconn(j);(* recursive call *)

(* at this point we have computed LOW[j] *)
(* update LOW[i], the part 2 of the def. of LOW *)
LOW[i] := min(LOW[i], LOW[j]);
if LOW[j] >= dfsn[i]
then(* a biconnected component is found *)
begin (*write the edges of biconnected component*)
pop S till(including) the edge ij;
writeln;(* new line *)
end
else(* j already visited*)
(* test if ij is a back edge *)
if (dad[i] <> j) and (dfsn[i] > dfsn[j])
then (* update LOW. part 3 of the def. of LOW *)
begin
LOW[i] := min(LOW[i], dfsn[j]);
push ij into S;
end;
end;(* dfs_biconn *)

(* initialization *)
for i := 1 to n do
mark[i] := 0;
counter := 1:
S := empty;(* empty stack *)
dfs_biconn(1); (* procedure call *)

Complete program in Pascal:

Let us write the complete program for finding biconnected components in
Pascal.

Biconnected components in pascal:

program biconnected;
const maxn = 50; (* maximum number of vertices *)
maxm = 250;(* maximum number of edges *)
type pointer = ^node;
node = record
info : integer;
next : pointer;
end;

var n, m, i, j, x, y, counter : integer;

L : array[1..maxn ] of pointer;
S : array[1..maxm ] of integer;
(*S, stack of edges and an edge
is a pair of vertices *)
66 Discrete Mathematics

top :integer; (* top of the stack S *)

procedure input_graph;
var i, j, x, y :integer;(* xy is an edge *)
t :pointer;
begin(* input_graph *)
write(’ enter the number of vertices ’);
readln(n);
write(’ enter the number of edges ’);
readln(m);

(* initialization. graph with n vertices and 0 edges *)

for i := 1 to n do
L[i] := nil;
(* read the edges *)
for i := 1 to m do
begin
write(’enter the edge ’, i);
readln(x,y);
(* add y at the head of the list L[x] *)
new(t); (* create a node pointed by t *)
t^.info := y;
t^.next := L[x];
L[x] := t;(* attach t at the head of L[x] *)
(* add x at the head of list L[y] *)
new(t);
t^.info := x;
t^.next := L[y];
L[y] := t;
end;
end;(* input_graph *)
procedure output_graph;
var t : pointer;
i : integer;
begin
for i := 1 to n do
begin
t := L[i];
if t = nil
then
write(’no vertices joined to ’, i)
else
begin
write(’the vertices joined to ’,i, ’ are :’);
(* scan the list L[i] *)
while t <> nil do
 Graph Algorithms I 67

begin
write(t^.info,’ ’);
t := t^.next;
end;
writeln;
end;
end;(* output_graph *)

procedure dfs_biconnected;
var t : pointer; j:integer;
dfsn : array[1..maxn] of integer;
LOW : array[1..maxn ] of integer;
dad : array[1..maxn ] of integer;
begin
mark[i] := 1;
dfsn[i] := counter; counter := counter + 1;
(* initialization of LOW: part 1 *)
LOW[i] := dfsn[i];
(* scan L[i] *)
t := L[i];
while t <> nil do
begin
j := t^.info;
if mark[j] = 0
then
begin
dad[t^.info] := i;(* add (i,j) to dfs tree *)
(* push (i, j) into the stack *)
S[top+1] := i;
S[top+2] := j;
top := top +2; (* update top *)
dfs_biconnected(j);(* recursive call *)
(* at this point we have found LOW[j] *)
(* update LOW[i]: part 2 *)
if LOW[i] > LOW[j] then
LOW[i] := LOW[j];
if LOW[j] >= dfsn[i] then
begin
(* a biconnected component is found *)
(* pop S till the edge ij *)
repeat
write(’(’,S[top],’,’,S[top-1],’)’);
top := top - 2;
until S[top+1] = i;
writeln;
68 Discrete Mathematics

end;
end(* then*)
else
(* test if ij is a back edge *)
if (dad[i] <> j) and (dfsn[i] > dfsn[j])
then
begin
S[top + 1] := i;
S[top + 2] := j;
top := top + 2;
(* update LOW[i]: part 3 *)
if LOW[i] > dfsn[j] then
LOW[i] := dfsn[j];
end;
t := t^.next;
end;(* while*)
end;(* dfs_biconnected *)

begin(* main program*)

(* initialization *)
for i := 1 to n do
mark[i] := 0;
counter := 1;
top := 0;(* empty stack S *)
(* end of initialization *)
input_graph;
output_graph;
dfs_biconnected(1);(* call *)
end.

The complexity of the biconnected components algorithm:

The complexity of the algorithm is the same as that of the dfs search, since the
added instructions, namely, calculating the function LOW, pushing each edge
once into the stack and popping each edge out of the stack takes O(m) steps.
Hence, the complexity is O(n + m). But the input is a connected graph and
for a connected graph we always have m ≥ n − 1. Hence the complexity is
O(m) by the rule of sum. (We can ignore lower-order terms while computing
complexity.)

1.10 Depth-First Search for Directed Graphs

The algorithm of the depth-first search for un undirected graph holds
equally well for a directed graph also. In a directed graph if (i, j) is an arc
and (j, i) is not an arc, then we place the vertex j in the list succ[i] but the
 Graph Algorithms I 69

FIGURE 1.29: Digraph to illustrate the depth-first search.

FIGURE 1.30: Depth-first search of graph of Figure 1.29.

vertex i will not be in the list succ[j]. In the case of an undirected graph, if
ij is an edge, then we include the vertex j in the list L[i] and also include the
vertex i in the list L[j].
The search dfs for a directed graph is illustrated by the following example.
Consider the graph of Figure 1.29.
As in the case of an undirected graph, the graph is represented by list of
successors of each vertex.
succ[1] = (2, 5, 6); succ[2] = (3, 4); succ[3] = (1); succ[4] = (3); succ[5] =
(6, 7); succ[6] = (); succ[7] = (6); succ[8] = (5, 9, 10); succ[9] = (); succ[10] =
(8, 9).
Let us now perform the dfs on the graph of Figure 1.30. The first call dfs(1)
“visits” the vertices 1, 2, . . . , 7 and gives us a tree with the root 1. Then, the
second call dfs(8) “visits” the remaining vertices 8, 9, 10 and gives us a second
tree of the dfs spanning forest. The original graph is restructured as in the
following graph Figure 1.30. In this figure, the solid arcs define a spanning
arborescence.
We observe the following properties:

Property 1.3. There are four types of arcs:

1. The tree arcs which are drawn as continuous arcs. If (i, j) is an arc
in the dfs spanning forest induced by the algorithm, then df s(i) called
directly df s(j) during the search (e.g., (2, 4)).
70 Discrete Mathematics

2. Dotted arcs like (1, 6) which go from an ancestor towards a decendant.

These arcs are called descending dotted arcs. If (i, j) is a descending
dotted arc, then df sn[i] < df sn[j].

3. Dotted arcs like (3, 1) which go from an ancestor towards a descendant.

These arcs are called mounting dotted arcs. Note that a loop is consid-
ered as a mounting dotted arc. If (i, j) is a mounting dotted arc, then
df sn[i] > df s[j].

4. Dotted arcs like (4, 3) and (8, 5) which join two vertices that are not in
the relation ancestor-descendant or descendant-ancestor. Such arcs are
called dotted cross arcs.

We shall now observe and prove the following key property of a dotted
cross arc.

Property 1.4. A dotted cross arc always goes from the right to the left. (We
assume that the children of each vertex are drawn from left to right and the
different trees of the dfs spanning forest are also drawn in the left to right
fashion.) In other words, if (i, j) is a dotted cross arc, then df sn[i] > df sn[j].

Proof. Consider a dotted cross arc (i, j). This means that the vertex j is
“visited” before the vertex i is “visited” during the search. Otherwise, suppose
the vertex i is “visited” before the vertex j. That is, when we reach the vertex i
for the first time, the vertex j is still marked “unvisited.” Since (i, j) is an edge
of the graph, we have j in the list of succ[i]. But then the “visit” of i would
not be complete unless we touch the vertex j. Hence, j must be a descendant
of i in the dfs spanning forest. Since (i, j) is a dotted edge, we conclude that
(i, j) must be a descending dotted edge, a contradiction.
Hence j is “visited” before the vertex i. This means that when we reach i for
the first time, j is marked “visited.” Hence, df sn[i] > df sn[j]. (Geometrically,
the vertex i is to the right of the vertex j.)

1.11 Applications of Depth-First Search for Directed

Graphs
1.11.1 Application 1: Finding the roots of a directed graph
Consider a directed graph G = (X, U ) where the vertex set X =
{1, 2, . . . , n}. A root of the directed graph G is a vertex r, from which we
can reach all other vertices of the graph by a directed path from r. A root
does not always exist. A graph may have more than one root. With the help
of the dfs, one can find all the roots of a given directed graph.
Input: A directed graph G = (X, U ) with the vertex set X = {1, 2, . . . , n}.
 Graph Algorithms I 71

The graph is represented by n successor lists, succ[1], succ[2], . . . , succ[n].

Output: Find all the roots of the graph G.
Algorithm: We mark all vertices as “unvisited,” set the “counter” used to
assign dfs number to 1. Then, we call the procedure dfs(1). If we mark as
“visited” all the vertices of the graph with this call, then the vertex 1 is
a root. Then, we repeat the calls with dfs(2), dfs(3), . . . , dfs(n). Let us now
write this algorithm.

procedure dfs( i : integer);

(* dfs visits each vertex of a connected component of G
containing i using depth-first search *)
var j : integer;(* vertex j scans succ[i] *)
begin
mark[i] := 1;
dfsn[i] := counter;
counter := counter + 1;
(* scan succ[i] *)
for each vertex j in succ[i] do
if mark[j] = 0 then
begin
add the arc ij to F;
dfs(j); (* recursive call *)
end;

(* initialization *)
for i := 1 to n do
mark[i] := 0;
counter := 1;
F := empty; (* empty forest *)
(* iteration *)
for i := 1 to n do
begin (* test if i is a root *)
dfs(i): (* call to the procedure dfs *)
if counter = n+1
then write(i,’ is a root’);
(* reinitialize: mark all vertices ‘‘unvisited’’ *)
for i := 1 to n do
mark[i] := 0;
counter := 1;
end;

The complexity of finding all the roots of a graph:

A call to dfs costs O(n + m) and we make n calls, one call with each vertex
of the graph. Hence the time complexity of the algorithm is O(n(n + m)).
72 Discrete Mathematics

FIGURE 1.31: Illustration to test if a graph is acyclic.

1.11.2 Application 2: Testing if a digraph is without

circuits
The existence of a directed circuit in a given digraph can be easily tested
using the dfs.
Input: A digraph G.
Output:
“yes” if the graph has no directed circuit.
“no” if the graph possesses a directed circuit.
Algorithm: By definition, each mounting dotted arc encountered during the
search clearly defines a circuit. Conversely, if the given digraph possesses a
circuit, then one must encounter a mounting dotted arc in the course of the
search.
To prove the converse part, consider a graph with at least one circuit and
perform the depth-first search on the graph. Now choose the vertex x whose
dfsn is the smallest among the vertices belonging to a circuit. Consider an arc
(y, x) on some circuit containing the vertex x. Since x belongs to a circuit,
such an arc must exist ( d− (x) ≥ 1.) Since x and y are on the same circuit, y
must be a descendant of x in the dfs spanning forest. Thus, the arc (y, x)
cannot be a dotted cross arc. Because the dfsn of y > the dfsn of x, (y, x)
cannot be a tree arc or dotted descending arc. Hence, (y, x) must be a dotted
mounting arc as in the figure below. Hence the converse part of the assertion
(see the graph of Figure 1.31).
In the above figure, the vertex x in the above argument is the vertex 2, and
the vertex y is the vertex 4 and the circuit in discussion is the circuit (2, 3, 4, 2).

1.11.3 Application 3: Topological sort

Consider a directed graph without circuits. Geometrically, a topological
sort means laying the vertices of a directed graph without circuits in a hor-
izontal manner from left to right such that each arc of the graph goes from
left to right. In other words, if the graph has n vertices, we are interested in
assigning the integers 1, 2, . . . , n to the vertices of the graph such that if there
is an arc from the vertex i to the vertex j, then we have the inequality i < j.
 Graph Algorithms I 73

FIGURE 1.32: A graph to illustrate topological sort.

We shall shortly show that a topological sort is possible if and only if

the graph is without circuits. Before proving the assertion, let us look at an
example.

Example 1.10
Consider the following graph of Figure 1.32 without circuits.
Let us assign the integer 1 to the vertex e, 2 to a, 3 to d, 4 to b and
finally 5 to the vertex c. This assignment defines a topological sorting
of the vertices of the graph. The same graph is redrawn below in Figure
1.33 to reflect the property of “topological sorting.”

Theorem 1.4. A directed graph G admits a topological sort if and only if the
graph is without circuits.

Proof. One part of the theorem is obvious. If the graph admits a topological
sort, then by definition we can assign the integers 1, 2, . . . , n as new labels to
its vertices in such a way that if (i, j) is an arc of the graph, then i < j. If
the graph possesses a circuit (v1 , v2 , . . . , vs , v1 ), then because of the property
of the new labeling of the vertices, we have v1 < v2 < · · · < vs . But then,
(vs , v1 ) is an arc of the circuit with v1 < vs , a contradiction.
Now for the second part, consider a longest directed path (x0 , x1 , . . . , xd )
in the given graph. Its length is d. We shall first prove that the initial vertex
x0 of our longest path satisfies d− (x0 ) = 0. If not, there is an arc (x, x0 ) in
the graph. We shall distinguish two cases.

FIGURE 1.33: Topological sorting of the graph of Figure 1.32.

74 Discrete Mathematics

Case 1. x = xi for 1 ≤ i ≤ d.
We then have a directed path (x, x0 , x1 , . . . , xd ) of length d +1, a con-
tradiction.

Case 2. x = xi for some i with 1 ≤ i ≤ d.

Then, (x0 , . . . , xi , x0 ) is a circuit of the graph, a contradiction.
Hence there is a vertex, say 1, in the graph with no arc entering into it. Now
consider the vertex deleted subgraph graph G−1. It has still no circuits. Hence
there is a vertex, say, 2, with no arcs entering into it in the graph G − 1. Then,
we consider the subgraph G−1−2, etc. till we arrive at the graph with exactly
one vertex which will be labeled as n.
These vertices 1, 2, . . . , n define a topological ordering of the vertices of the
graph G.
We shall now prove the following result concerning the adjacency matrix
of a digraph without circuits. This result is an immediate consequence of the
topological sort.

Theorem 1.5. Let G be a directed graph on n vertices without circuits. Then,

its vertices can be labeled by the integers 1, 2, . . . , n so that its adjacency matrix
is an upper triangular.
Proof. Since the graph is without circuits, it admits a topological sort. Let
this order be 1, 2 . . . , n. We now relabel the vertices of the graph in this order.
According to this labeling, if (i, j) is an arc, then i < j. This means that, if
i > j, then (i, j) is not an arc of the graph. This means that in the adjacency
matrix M of the graph G, we have M [i, j] = 0 if i > j. Hence the result.
Remark 1.5. According to Theorem 1.5, a graph without circuits can be
represented by an upper triangular matrix. Since the graph has no loops, the
diagonal entries of this matrix are also 0. Hence a graph on n vertices and
without circuits may be represented economically by a one-dimensional array
using only (n2 − n)/2 memory space, (by omitting 0 entries on and below the
main diagonal) instead of n2 space needed to store an n × n matrix.
Algorithm for the topological sort:
Input: A directed graph G without circuits.
Output: Assigning the integers 1, 2, . . . , n to the vertices of the graph in such
a way that if (i, j) is an arc of the graph, then we have i < j.
Algorithm 1.4. Choose a vertex x such that the in-degree of the vertex x is 0.
According to Theorem 1.4 such a vertex exists. Label this vertex as 1. Then,
consider the vertex deleted subgraph G1 = G − 1 and choose a vertex in G1
for which the in-degree is 0. Call this vertex 2. Then, consider G2 = G − 1 − 2
and choose a vertex in G2 with in-degree 0. Continue like this till we arrive
at the graph with only one vertex. This last vertex will be labeled as n. The
sequence 1, 2, . . . , n is the required topological sort.
 Graph Algorithms I 75

Let us write the algorithm in pseudo-code. We use the following variables:

The graph is represented by n successor lists, succ[1], succ[2], . . . , succ[n]. We
use an array id[1..n] where id[i] is the in-degree of the vertex i, that is, the
number of arcs entering into the vertex i. Once a vertex i is processed by
the algorithm, we set id[i] equal to -1 or some value which cannot be a legal
in-degree of a vertex. The variable “counter” is used to count the number of
vertices printed in topological order.

Topological sort1:
(* initialization of counter *)
counter := 0;
while counter < n do
begin
(* scan the array id to find vertices of in-degree zero *)
for i := 1 to n do
if id[i] = 0 then
begin
write(i,’ ’);
counter := counter + 1;
id[i] := -1; (* vertex i is processed *)
(* scan succ[i] to reduce the in-degree by one *)
for each j in succ[i] do
id[j] := id[j] - 1;
end;
end;

The complexity of the topological sort:

The worst case arises when the graph is a directed path (1, 2, . . . , n). The array
id is defined as id[i] = 1 for 0 ≤ i < n and id[n] = 0. The algorithm prints
the vertex n first and the variable counter is set to 1. During the next pass
through the while loop, the algorithm prints the vertex n − 1 and the counter
is assigned the value 2, etc. Hence, the while loop is executed n times for
this graph. Each execution of while loop costs O(n) steps, because of the for
loop inside the while loop. Hence the complexity is O(n2 ) steps. The desired
topological sort is (n, n − 1, . . . , 1).

Algorithm 1.5. This algorithm uses dfs. It is based on the following property:
A directed graph is without circuits if and only if there is no dotted mounting
arc during the dfs. A simple write statement at the end of the dfs procedure
gives us a reverse topological sort. We use a stack S of vertices. We will give
the procedure below:

Topological sort algorithm using dfs:

(* initialization *)
for i := 1 to n do
mark[i] := 0;
76 Discrete Mathematics

top := 0; (* empty stack *)

(* end of initialization *)

procedure toposort( i : integer);

begin
mark[i] := 1;
(* scan the list succ(i) *)
for each vertex j in succ(i) do
if mark[j] = 0 then
dfs(j);(* recursive call *)
(* push i into the stack *)
top := top +1;
S[top] := i;
end;

(* iteration *)
for i := 1 to n do
if mark[i] = 0 then
dfs(i); (* invoking toposort *)
(* print the vertices in topological sort *)
for i: = n downto 1 do
write (S[i],’ ’);

Let us justify the algorithm:

Proof. When the procedure “toposort” finishes scanning entirely the list
succ(i), it pushes the vertex i into the stack. Since the graph is without cir-
cuits, there are are no dotted mounting arcs from i. The only possible arcs
going out of i are the arcs of the forest, dotted descending arcs, and dotted
cross arcs. But these arcs are directed toward vertices whose successor lists
have been completely scanned and therefore already pushed into the stack.
Hence the proof.
The complexity of the topological sort:
We have added only two elementary statements to the procedure dfs, namely,
incrementing the variable “top” and pushing a vertex into the stack. Hence
the complexity is the same as that of the complexity of the dfs procedure
which is O(m + n).

1.11.3.1 An application of topological sort: PERT

PERT and the critical path method: Imagine a project like constructing
a big house. A project like this is often divided into a number of smaller
subprojects like clearing land, building foundation, constructing walls, ceil-
ings, carpentry works, plumbing, electricity work, landscaping, etc. In order
to complete well the entire project within a given time limit, these subprojects
should be carried out in a certain order. For example, land clearing should be
 Graph Algorithms I 77

FIGURE 1.34: An activity digraph.

completed first before the foundation work starts, and plumbing and electric-
ity may be taken later once the walls and ceilings will have been completed,
etc. These subprojects are called the activities of tasks; that is, a project is
constituted of activities. These activities are interrelated heavily by the con-
straints like a certain activity may not start until other activities have been
completed.
Let us illustrate this by an example.

Example 1.11
The vertices of the graph (see Figure 1.34) without circuits repre-
sent different activities and the graph itself represents a project. An arc
from the vertex i to the vertex j means the activity i must have been
completed before initiating the activity j. The weight w(i, j) associated
with an arc (i, j) represents the time needed (e.g., measured in weeks) to
finish the task i. In our example, the task 1 has no predecessor and the
task 9 has no successor. Such vertices always exist in a directed graph
without circuits. For example, the first vertex in a topological sort has
no predecessor, and the last vertex of the sort has no successor. We
introduce two new vertices s and e to the activity digraph. The vertex
s corresponds to the starting of the project and the vertex e to the end
of the entire project. The vertex s is joined to all vertices of the graph
for which the in-degree is 0 and all vertices of the graph for which the
out-degree is 0 are all joined to the vertex e. Finally, we associate the
weight to all arcs of the form (s, i) and all arcs of the form (j, e). In our
example, there is only one vertex of in-degree 0, namely the vertex 1
and the only vertex of out-degree 0 is 9.
We are interested in finding the minimum number of weeks needed to
finish the entire project. This problem is equivalent to finding a directed
path of maximum weight in a graph without circuits. In our example, a
directed path of maximum weight is (1, 2, 3, 4, 7, 9) and its weight is 38.
Note that task 2 can start only after completing task 1, and 3 can be
initiated after finishing 3, etc. This means that to complete the entire
project we need a minimum of 38 weeks. Why? The time required to
complete the entire project cannot be less than 38 weeks, since otherwise
not all of the tasks of the graph could be completed.
78 Discrete Mathematics

To see this, let us write the vertices of the activity digraph in topolog-
ical order. The desired order is (1, 2, 3, 4, 5, 6, 7, 8, 9). We first start and
finish the task 1; then we start and finish the task 2 and then 3. Now we
can start simultaneously the tasks 4, 5 and 6 since these three activities
are not related by the constraints of precedence (technically these tasks
4, 5, and 6 form a stable set, that is, there are no arcs joining these three
vertices in the graph. After finishing 4, we can initiate 7. Note that after
finishing 5, we cannot immediately start the task 7, since we have to wait
for the completion of the task 4 which takes longer than the task 5. (The
task 7 has two predecessors: 4 and 5.) After finishing 6, we start 8. The
task 7 starts before the task 8 but is completed only after the task 8,
because the weight of a longest path to 7 is 28, whereas the one to 8 is 31.
That is, 7 starts two weeks later than 8 but is completed one week earlier.
A directed path of maximum weight is called a critical path. The
different tasks on a longest weighted path are called critical tasks. In
our example, the critical tasks are 1, 2, 3, 4, 7, 9. If the completion of one
of the tasks on a critical path is delayed, say, by one week, then the entire
project will be delayed by one week. Techniques referred to as CPM and
PERT use weighted directed graphs without circuits as models.

Longest path in a directed path without circuits:

We shall now present a polynomial time algorithm to find the maximum weight
of a directed path (also called longest directed path) in a given directed graph
without circuits and with a weight associated to each arc.
There is no known polynomial time algorithm for finding the maximum
weight of a directed path in an arbitrary weighted directed graph. This prob-
lem belongs to a large class of problems called NP-Complete problems.
Input: A weighted directed graph without circuits w(i, j) is the weight of the
arc (i, j).
Output: t[i], the maximum weight of a directed path arriving at the vertex i
for each vertex i.
Algorithm: Let us first perform a topological sort of the vertices of the given
graph. Let this order be 1, 2, . . . , n. The algorithm processes the vertices in
this order. For each vertex i, we list all the predecessors of the vertex i in the
list pred(i). Let us write the algorithm in pseudo-code.

Maximum weight of a directed path arriving at each vertex:

The vertices are ordered as 1,2,...,n according to the topological sort.
(* initialization *)
for i := 1 to n do
t[i] := 0;
(* iteration *)
 Graph Algorithms I 79

for i := 1 to n do
begin
for each vertex j in the list pred(i) do
if t[j] + w(j,i) > t[i] then
t[i] := t[j] + w(j,i);
end;

The complexity of finding a longest dipath in a digraph

without circuits:
A call to topological sort requires (n+m) steps. The initialization of the table t
takes O(n) steps. Now consider the iteration step: It has two nested loops. For
each vertex i, the interior loop is executed exactly d− (i) times where d− (i) is
in-degree of the vertex i, that is, the number of arcs entering into the vertex i.
Since the sum of the in-degrees of all vertices of a graph is the number of
arcs of the graph, the cost of the iteration is O(m). Hence the complexity is
O(n + m) + O(m) + O(m) = O(max(n, m)), by the sum rule of the complexity.

1.11.4 Application 4: Strongly connected components

algorithm
Consider a directed graph G = (X, U ) with vertex set X = {1, 2, . . . , n}.
The graph G is strongly connected if between any two vertices x and y there is
a directed path from x to y and a directed path from y to x. In other words,
a graph is strongly connected, if any two distinct vertices lie on a circuit
(not necessarily elementary). By convention, the singleton sequence (x) is
considered as a directed path of length 0 from x to itself. In this subsection,
edge means directed arc/edge.

Example 1.12: Strongly connected graph

The graph G of Figure 1.35 is strongly connected. We see that any
two vertices lie on a circuit.
The vertices 2 and 5 lie on the circuit (2, 1, 3, 4, 5, 1, 3, 2) which is
not even simple, because the arc (1, 3) is repeated twice.

A graph which is not strongly connected can be decomposed into strongly

connected components. A strongly connected component of a directed graph G
is a maximal induced subgraph H which is strongly connected. That is, H is
an induced subgraph of G with the following properties:

1. The graph H is strongly connected.

2. For any vertex x of G but not in H, the subgraph of G induced by

H ∪ {x} is not strongly connected.
80 Discrete Mathematics

FIGURE 1.35: A graph and its strongly connected components.

Let us illustrate this notion by an example. Note that a graph is strongly

connected if and only if its only strongly connected component is itself.

Example 1.13
Consider the graph of Figure 1.35.
The graph of Figure 1.35 is not strongly connected because there is
no directed path from the vertex 1 to the vertex 6 even though there is a
directed path from the vertex 6 to the vertex 1. The graph possesses three
strongly connected components. These strongly connected components
are induced by the vertex sets {1, 2, 3, 4, 5} and {6, 7, 8} and {9}.
Note that every vertex of the graph G lies in exactly one strongly
connected component, but there are edges like (6, 5) and (7, 9) which are
not in any strongly connected components. In other words, the different
strongly connected components of a graph define a partition of the vertex
set of X of G into a union of mutually disjoint subsets of X.

Tarjan’s algorithm for strong components:

We shall now present an algorithm for finding strong components of a directed
graph. This algorithm is based on the dfs of the given graph G. There is
some similarity between the strong components algorithm and the biconnected
components algorithm we have already studied. Recall that we have defined
a function LOW which helped us to find the biconnected components of a
connected graph using dfs. The function LOW used the dfs spanning tree, the
dfsn of the vertices and the back edges. In a similar manner, we will be defining
 Graph Algorithms I 81

a function LOWLINK using dfs spanning forest, dfsn, dotted mounting arcs
and dotted cross arcs.
In the biconnected components algorithm, we have pushed and popped the
edges of the undirected connected graph, because the different biconnected
components form a partition of the edge set. In the strongly connected com-
ponents algorithm, we will be pushing and popping the vertices of the graph
because the different strong components of a graph form a partition of the
vertex set.
In short, the strong components algorithm and the biconnected compo-
nents are “dual” algorithms.

Même quand l’oiseau marche, on sent qu’il a des ailes.

Lemierre (1733–1793)

This duality will be shortly explained.

With this analogy in mind, we shall study the strongly connected compo-
nents algorithm.
Input: A directed graph G = (X, U ) with vertex set X = {1, 2, . . . , n}.
Output: The list of vertices of each strongly connected component.
Algorithm: Let us first illustrate this algorithm by means of an example. Con-
sider the following graph G of Figure 1.36.
To find the strong components of the graph, we search the graph according
to dfs. The graph is represented by successor list of each vertex: succ[1] =
(2, 4); succ[2] = (3, 4); succ[3] = (2); succ[4] = (3); succ[5] = (1, 6, 9); succ[6] =
(7, 8); succ[8] = (7); succ[9] = (10); succ[10] = (9).
The dfs structure of the graph G of Figure 1.36 is given in Figure 1.37:
The following Table 1.11 gives the dfsn of different vertices.
The function LOWLINK: Like the function “LOW” of the biconnected
components algorithm, we shall define the function “LOWLINK” on the set
of vertices of the graph.

FIGURE 1.36: Illustration of strong components algorithm.

82 Discrete Mathematics

FIGURE 1.37: dfs: strong components algorithm.

TABLE 1.11: Depth-first search number (dfsn) table of graph of

Figure 1.37

1 2 3 4 5 6 7 8 9 10
dfsn 1 2 3 4 5 6 7 8 9 10

LOWLINK[i] is the smallest dfsn ≤ df s[i] one can reach, either from the
vertex i or from one of the descendants of i in the dfs spanning forest, using
only one dotted mounting or cross arc. In the case of utilization of a dotted
cross arc in the calculation of LOWLINK, this dotted cross arc must lie in
some circuit of the graph. If such dotted arcs lead to vertices whose depth-first
search number > df sn[i] or if no such dotted arcs exist from i or from one of
its descendants, then we define LOWLINK[i] = df sn[i].
Mathematically, we write LOWLINK[i] = min({df sn[i]} ∪ B), where
B = {df sn[j] | ∃ an arc ∈ M ∪ C lying in a circuit from a
descendant of i to j},
where M is the set of mounting arcs and C is the set of cross arcs. Let
us calculate the function LOWLINK in our example. In the algorithm, the
different LOWLINKS are computed according to the postorder traversal of
the vertices of the spanning dfs forest. The postorder of our dfs forest is
(3, 4, 2, 1, 7, 8, 6, 10, 9, 5). Hence the first vertex for which LOWLINK is com-
puted is the vertex 3.
LOWLINK[3] is min(df sn[3], df sn[2]) = min(3, 2) = 2, because of
the mounting dotted edge (3, 2) from the vertex 3. LOWLINK[4] is
min(df sn[4], df sn[3]) = min(4, 3) = 3 because of the dotted cross edge (4, 3)
from the vertex 4 and this arc lies in a circuit (4, 3, 2, 4). LOWLINK[2] =
min(df sn[2], df sn[2]) = 2 because of the dotted mounting arc from the son 3
of the vertex 2. LOWLINK[1] = 1. LOWLINK[7] is min(df sn[7], df sn[5] =
min(7, 5) = 5. Note that the dotted cross arc (7, 4) from the vertex 7 is not
taken into account for the calculation of LOWLINK[7] because this cross arc
does not lie in any circuit of the given graph. Similarly, while computing
LOWLINK[5], the dotted cross arc (5, 1) will not be considered because of
the absence of a circuit passing through this arc. On the other hand, while
computing LOWLINK[8], we have to take into consideration the dotted cross
arc (8, 7) because this arc lies in a circuit (8, 7, 5, 6, 8) of the given graph.
 Graph Algorithms I 83

TABLE 1.12: LOWLINK function table of graph of Figure 1.37

1 2 3 4 5 6 7 8 9 10
LOWLINK 1 2 2 3 5 5 5 7 9 9

The following table gives the function LOWLINK (Table 1.12).

In our example, there are four strong components. These components are
induced subgraphs of the graph G generated by the vertex set:

{2, 3, 4}; {1}; {5, 6, 7, 8} and {9, 10}.

We make the following important observation from the example:

The intersection of the arcs of the strong component generated by the set
{2, 3, 4} and the arcs of the dfs spanning forest is a tree of the spanning forest.
This is true for the other three strong components as well. In fact this is true
in general. For a proof, see [3].
The root of the tree obtained by the intersection of the arcs of a strong
component and the dfs spanning forest is called the root of the strongly con-
nected component.
Putting it differently, in each strong component, choose the vertex r whose
dfsn is minimum among all vertices of the strong component. This vertex is
called the root of the strongly connected component.
Back to our example, the root of the strong component induced by the set
{2, 3, 4} is the vertex 2 and the root of the strong component generated by
{1} is 1, {5, 6, 7, 8} is 8 and the one generated by {9, 10} is the vertex 9.
Once we have identified the roots of the strong components, we can easily
find the set of vertices of each strong components. We use a stack S of vertices.
Initially, the stack S is empty. Recall that we have used a stack of edges in
the biconnected components algorithm.
As we traverse the graph by means of the dfs, we push the vertices into
the stack as soon as we “visit” them. When should we pop up the stack to get
the strong components? As in the biconnected components algorithm, when
we backtrack, that is, when we climb up the dfs forest under construction,
if a root r of a strong component is encountered, then we pop up the stack
till (including) the vertex r, to obtain the strong component containing the
vertex r.
In our example, we push into the stack the vertices 1 and 2 and finally the
vertex 3. No more new vertices are found from the vertex 3. So we prepare to
backtrack to the father of the vertex 3 which is 2. At this point, the stack will
look as (stack grows from left to right)

S = (1, 2, 3).
Now we test if the vertex 3 is a root of a strong component. The answer is
“no.” So no pop-up occurs. From the vertex 2, we find a new vertex 4 and
84 Discrete Mathematics

hence we push into the stack the vertex 4. No new vertices are found from the
vertex 4 and we are ready to backtrack from 4 to his father 2. At this point
the stack will look as

S = (1, 2, 3, 4).

We perform the test: Is 4 a root of a strong component? Since this is not the
case, no pop-up occurs. We climb up to the vertex 2. From 2, no new vertices
are found and we perform the test if 2 is a root of a strong component. Since
the answer to the test is “yes,” we pop up the stack S till the vertex 2. This
is the vertex set of the first strong component found by the algorithm.
The first strong component is induced by the set {4, 3, 2}.
We go back to the vertex 1. Since no new vertices are found from 1, we
perform the test if 1 is a root of a strong component. Since 1 is a root, we pop
the stack till the vertex 1.
The second strong component found is induced by the set {1}.
At this point the stack is again empty. The dfs takes a new vertex 5 as a
root of a tree in the spanning forest under construction and initiates the dfs
search from this vertex 5. From 5 we find 6 and from 6 we find the vertex 7. No
new vertices are found from 7. So we backtrack. At this point, the stack will be

S = (5, 6, 7).

Test fails for 7 and we move to 6. From 6, we find 8. No new vertices are
“visited” from 8. We prepare to climb up the tree from 8 to the father of 8
which is 6. At this point, the stack will be S = (5, 6, 7, 8). 8 is not the root.
From 6 no new vertices are “visited.” Since 6 is not a root, we reach the
vertex 5. From 5, we find the new vertex 9, from 9 we find the vertex 10. At
this point, the stack is S = (5, 6, 7, 8, 9, 10). 10 is not a root and we reach back
to 9. No new vertices are found from 9. 9 is a root and we pop up the stack
till the vertex 9.
Hence the third strong component is induced by the set {10, 9}.
The stack S = (5, 6, 7, 8). Finally, we climb up to the vertex 5. Since no
new vertices are found from 5, it is a root. Hence we pop up the stack till the
vertex 5.
Hence the fourth and final strong component is induced by the set
{8, 7, 6, 5}.
Note that the order in which the strong components are found is the same
as that of the termination of their dfs calls. In our example, this order is
2, 1, 9, 5. The call dfs(2) ends before the call dfs(1) ends. The call dfs(1) ends
before the call dfs(9) ends. Finally, the call dfsn(9) ends before the call dfs(5)
ends.
It remains to recognize the different vertices which are the roots of strong
components. From the example, we observe the following property of the roots
of strong components.
 Graph Algorithms I 85

The root r of a strong component satisfies the equality LOWLINK[r] =

df sn[r] and conversely if a vertex x satisfies the equality LOWLINK[x] =
df sn[x], then x is a root of some strong component.
Before proceeding to write the algorithm, we shall express the function
LOWLINK in a recursive manner. In this recursive definition, LOWLINK
of a vertex x is expressed in terms of the LOWLINK of the children of the
vertices of x and dotted mounting arcs and cross arcs from a descendant of
the vertex x.
part1 part2
     
LOWLINK[i] = min({df sn[i]} ∪ {LOWLINK[s] | s, a son of i in dfs forest} ∪C  )

where
part3
  

C = {df sn[j] | ∃ an arc ∈ M ∪ C lying in a circuit from a descendant of i to j},

where M is the set of mounting arcs and C is the set of cross arcs. We are
now ready to describe the algorithm in pseudo-code.
We use the following variables: The array “mark” (mark[i] = 0 if the ver-
tex i is “unvisited” and mark[i] = 1 if the vertex i has been “visited during
the dfs). The array “dfsn” where dfsn[i] is the depth-first search number of
the vertex i and the array “LOWLINK” to compute the LOWLINK of each
vertex. The variable “counter” is used to assign the dfsn. The stack S elements
are the vertices of the graph.

Strongly connected components algorithm:

procedure dfs_strong_comp( i : integer);
var j : integer;
LOWLINK : array[1..n] of integer; (* n is the number of vertices *)
begin
mark[i] := 1;
dfsn[i] := counter; counter := counter + 1;
(* initialization of LOWLINK: part 1 of recursive definition *)
LOWLINK[i] := dfsn[i];
top := top +1; S[top] := i; (* push i into the stack S *)
(* scan the list succ(i) to find a new vertex j *)
for each j in succ(i) do
if mark[j] = 0
then (* j is a son of i in dfs forest *)
begin
dfs_strong_comp(j); (* recursive call *)
(* at this point we have computed LOWLINK[j] *)
(* update LOWLINK: part 2 of recursive definition *)
LOWLINK[i] := min(LOWLINK[i],LOWLINK[j]);
end
86 Discrete Mathematics

else(* j is "visited" *)
(* test if arc (j,i) is a mounting or cross arc *)
if (dfsn[i] > dfsn[j]) and ((j,i) in a circuit)
then
(* update LOWLINK: part 3 of LOWLINK definition *)
LOWLINK[i] := min(LOWLINK[i],dfsn[j]);
(* test if the vertex i is a root of a strong component *)
(* if so pop up the stack till the vertex i *)
if LOWLINK[i] = dfsn[i]
then
repeat
write(S[top],’ ’);
top := top - 1;
until S[top+1] = i;
writeln;(*new line to separate strong components *)
end;

(* initialization *)
for i := 1 to n do
mark[i] := 0;
top := 0; (* empty stack *)
counter := 1; (* to assign dfs number *)
for i := 1 to n do
if mark[i] = 0
then dfs_strong_comp(i); (* call to procedure dfs_
strong_comp*)

The above algorithm is simply an implementation of the function

LOWLINK. In order to write a working program, we have to implement
“part 3” of the recursive definition of LOWLINK. In this part 3 of LOWLINK,
we take into account a dotted cross arc (i, j) to compute LOWLINK if this
arc lies in a circuit of the graph (see part 3 of the recursive definition of
LOWLINK).
There is a circuit passing through the dotted cross arc (i, j) if and only if
the vertex j is still in the stack S.
To illustrate this, let us again refer to our example illustrating strong
components algorithm.
In the example, the order of the roots of the strong components according
to the termination of their dfs call is (2, 1, 9, 5). Note that we have taken into
account the cross arc (4, 3) in the computation of LOWLINK[4], since this arc
lies in a circuit or equivalently the terminal vertex 3 of the cross arc (4, 3) is
in the stack S while we examine this cross arc. Similarly, we have taken into
consideration the cross arc (8, 7) in the computation of LOWLINK[8], when
we are at vertex 8 traversing this cross arc, since the end vertex of the arc,
namely, the vertex 7 is in the stack S.
 Graph Algorithms I 87

On the other hand, the cross arc (5, 1) is not taken into consideration in
the computation of LOWLINK[5], since the end vertex 1 of this cross arc does
not belong to the stack S. Note that the strong component containing the
vertex 1 has already been emitted by the algorithm when we are at vertex 5
examining the cross arc (5, 1). For a similar reason, we discard the cross arc
(7, 4) in the computation of LOWLINK[7].

Strong components program in Pascal:

program strongcomp;
const maxn = 50; (* maximum number of vertices *)
maxm = 250;(* maximum number of edges *)
type pointer = ^node;
node = record
info : integer;
next : pointer;
end;

var n, m, i, j, x, y, counter : integer;

succ : array[1..maxn ] of pointer;
S : array[1..maxn ] of integer;(* stack of vertices *)
top :integer; (* top of the stack S *)
procedure input_graph;
var i, j, x, y :integer;(* (x,y) is an arc *)
t :pointer;
begin(* input_graph *)
write(’ enter the number of vertices ’);
readln(n);
write(’ enter the number of arcs ’);
readln(m);

(* initialization. graph with n vertices and 0 arcs *)

for i := 1 to n do
succ[i] := nil;
(* read the arcs *)
for i := 1 to m do
begin
write(’enter the arc ’, i);
readln(x,y);
(* add y at the head of the list succ[x] *)
new(t); (* create a node pointed by t *)
t^.info := y;
t^.next := succ[x];
succ[x] := t;(* attach t at the head of L[x] *)
end;
end;(* input_graph *)
procedure output_graph;
88 Discrete Mathematics

var t : pointer;
i : integer;
begin
for i := 1 to n do
begin
t := succ[i];
if t = nil
then
write(’no successor to ’, i)
else
begin
write(’the successor to ’,i, ’ are :’);
(* scan the list succ[i] *)
while t <> nil do
begin
write(t^.info,’ ’);
t := t^.next;
end;
writeln;
end;
end;(* output_graph *)
procedure dfs_strong_comp( i : integer);
var t : pointer; j : integer;
dfsn : array[1..maxn] of integer;
LOWLINK : array[1..maxn] of integer;
instack : array[1..maxn] of boolean;
(* instack[i] is true if i is in stack, false otherwise *)
begin
mark[i] := 1;
dfsn[i] := counter; counter := counter + 1;
(* push i into stack *)
top := top + 1; S[top] := i;
isstack[i] := true;
(* initialization of LOWLINK[i] using part 1 *)
LOWLINK[i] := dfsn[i];
(* scan the list succ[i] *)
t := succ[i];
while t < > nil do
begin
j := t^.info;
if mark[j] = 0
then (* j is a new vertex *)
begin
dfs_strong_comp(j); (* recursive call *)
(* at this point we have computed LOWLINK[j]*)
(* update LOWLINK[i] using part 2*)
if LOWLINK[i] > LOWLINK[j]
then LOWLINK[i] := LOWLINK[j];
 Graph Algorithms I 89

end
else
if (isstack[j] = true) and (dfsn[i] > dfsn[j])
then (* (i,j) is cross arc in a circuit *)
(* update LOWLINK[i]: part 3 *)
if LOWLINK[i] > dfsn[j]
then LOWLINK[i] := dfsn[j];
t := t^.next;
end;
(* test if i is a root of a strong component *)
if LOWLINK[i] = dfsn[i]
then(* i is a root of a strong component *)
begin
repeat
write(S(top],’ ’);
isstack[S[top]] := false;
top := top - 1;
until S[top +1] = i;
writeln;(* new line *)
end;
end;
begin (* strongcomp *)
(* initialization *)
for i := 1 to n do
begin
mark[i] := 0;
isstack[i] := false;
end;
top := 0; (* empty stack *)
counter := 1;
(* end of initialization *)
input_graph;
output_graph;
for i := 1 to n do
if mark[i] = 0
then dfs_strong_comp(i); (* call *)
end.

Complexity of strong components algorithm:

The complexity of the strong components algorithm is the same as that
of the dfs procedure. Intuitively, each vertex is “visited” only once and
pushed/popped only once. Each edge is traversed only once. Hence the com-
plexity of Tarjan’s algorithm is O(max(n, m)).

Remark 1.6. We have already said that the biconnected components algo-
rithm and the strongly connected components algorithm are dual algorithms.
This duality actually comes from the following observation:
90 Discrete Mathematics

In a graph, vertices and edges can be considered as dual concepts. The

different biconnected components of an undirected connected graph may be
obtained as equivalence classes induced by a suitable equivalence on the set of
edges E of the graph G. This relation R1 is defined as below.
Two edges e1 and e2 are related by the relation R1 if e1 = e2 or there is
an elementary cycle passing through these two edges. This is an equivalence
relation on the set of edges of the graph G. Reflexivity of R1 is evident by
the definition. Symmetry of R1 is also clear, since the order of appearance of
the edges e1 and e2 on an elementary cycle is not relevant according to the
definition. Now let us prove transitivity of R1 . If e1 R1 e2 and e2 R1 e3 , then
there is an elementary cycle C1 passing through e1 and e2 and an elementary
cycle C2 passing through e2 and e3 .
Then, the symmetric difference (considering the cycles C1 and C2 as sets
of edges) C1 ΔC2 = (C1 ∪C2 )\(C1 ∩C2 ) is an elementary cycle passing through
the edges e1 and e3 .
The subgraphs induced by the edges of the equivalence classes of this rela-
tion R1 are the biconnected components of the graph.
Let us define an equivalence relation R2 on the vertex set of a directed
graph so that the different classes of this equivalence relation induces strong
components of the directed graph. Two vertices x and y are in relation under
R2 if either x = y or there is a directed elementary path from the vertex x to
the vertex y and an elementary directed path from the vertex y to the vertex x.
Then, we can prove easily that the relation R2 is an equivalence relation on the
vertex set of the graph. The subgraphs induced by the vertices of the equivalence
classes of R2 are the strong components of the graph.

1.12 Traveling Salesman Problem

One of the famous NP-complete problems is the traveling salesman prob-
lem (TSP). Let us state this problem.
Input: There are n cities 1, 2, . . . , n and the cost of traveling between any two
cities i, j,(i = j) is cij = c(i, j), (1 ≤ i, j ≤ n). We assume that the cost
function is a symmetric function, that is, c(i, j) = c(j, i).
Output: Find a tour of a salesman starting from the city 1, visiting each city
exactly once and returning to the starting point 1, by minimizing the total
cost of the travel. The cost of a tour is the sum of the costs of the travel
between the cities.
This problem can be cast into a graph problem by representing the cities
by vertices and the intercity routes by the corresponding edges, with the cost
of traveling between the vertices i and j as a weight of the edge ij.
Input: A complete simple graph Kn on n vertices and a weight attached to
each edge ij, 1 ≤ i, j ≤ n.
 Graph Algorithms I 91

FIGURE 1.38: Illustrating the traveling salesman problem.

TABLE 1.13: Three tours of Figure 1.38

Tour number Tour Cost of tour
1 (1, 2, 3, 4, 1) 21
2 (1, 3, 2, 4, 1) 19
3 (1, 3, 4, 2, 1) 20

Output: Find a Hamiltonian cycle of the complete graph Kn with the sum of
the weights of the edges of the cycle as a minimum, that is, find an elementary
cycle passing through each vertex exactly once with the sum of the weights of
the edges as a minimum.
Let us illustrate this problem with an example (see graph of Figure 1.38).
In the graph of Figure 1.38, the different tours and their corresponding costs
are given in Table 1.13. In Table 1.13, there are three tours and the minimum
tour is the tour number 2 whose cost is 19. Note that the tours (1, 2, 3, 4, 1)
and the reverse tour (1, 4, 3, 2, 1) give the same cost. Hence such duplicated
tours are not listed in the table. The following lemma gives the number of
Hamiltonian tours in a complete graph Kn .
Lemma 1.2. The number of Hamiltonian tours in Kn is (n − 1)!/2 (by iden-
tifying the reverse tours, that is, for n = 3, (1, 2, 3, 1) = (1, 3, 2, 1), because
they have the same cost).

Proof. Every tour of Kn defines a cyclic permutation and conversely every

cyclic permutation of the set {1, 2 . . . , n} defines a tour of Kn . But the number
of cyclic permutation of the n set [n] is (n − 1)! (see [6]). By identifying a
tour and its reverse tour (because they have the same total cost), the desired
number is (n − 1)!/2.
Brute-force algorithm to solve TSP:
The first algorithm which comes to our mind is the brute-force algorithm:
Generate all the possible tours of the salesman and find one for which the
total cost is a minimum. But this algorithm takes exponential time, because
there are essentially (n − 1)!/2 tours and the following inequality shows that
factorial n is an exponential function (for a proof, see [6]).
n
nn/2 ≤ n! ≤ ((n + 1)/2) .
92 Discrete Mathematics

For a proof of this inequality, see [6]. The following program implements the
brute-force algorithm. This program does not check for equality between a
tour and its reverse tour. There are four cities 1, 2, 3 and 4. 0 is not a city.
The reader is requested to see the recursive permutation generation program
in [6] and appreciate the similarity with TSP brute-force program.

Brute-force C program for tsp:

#include <stdio.h> //load input-output library
#include <stdlib.h>
#include <limits.h> //INT_MAX is defined here
#define N 4 // N, the number of cities

int v[N+1]={0,1,2,3,4} ; // v= ville, means city in French

int c[N+1][N+1]={{0,0,0,0,0},{0,0,4,3,5},{0,4,0,5,6},{0,3,5,0,7},
{0,5,6,7,0}};
//c, the cost matrix
int t[N+1]; //t, the Hamiltonian tour
int mc=INT_MAX ; //mc, minimum cost

void tsp(int k)

{/*tsp finds the cost of a minimum cost Hamiltonian cycle */

int i,temp,s; // i for the loop, temp for swap and s for the sum

if (k==1) //basis of recurrence

{s=0;
for (i=1;i<N;i++)
s = s+c[v[i]][v[i+1]];
s = s+c[v[N]][v[1]];
if (s<mc)
{mc=s;
//save the Hamiltonian cycle
for (i=1;i<=N;i++)
t[i]=v[i];
}
}
else // recurrence
{ tsp(k-1); //recursive call
for (i=1;i<k;i++)
{//swap v[i] et v[k]
temp=v[i];
v[i]=v[k];
v[k]=temp;
 Graph Algorithms I 93

tsp(k-1); //recursive call

//restore array v

//swap again v[i] et v[k]

temp=v[i];
v[i]=v[k];
v[k]=temp;
}
}
}

int main ( )

{ int i ;
tsp ( N ) ; //call with k = N
//print the result
printf("the minimum cost tour is %d \n", mc );
//print the tour
for (i=1;i<=N;i++)
printf("%3d",t[i]);
printf("%3d",t[1]);
system ("pause");
return 0;
}

1.12.1 Approximate algorithms for traveling salesman

problem
Definition 1.2 (Approximate algorithm). An algorithm A is said to be
an approximate algorithm for the traveling salesman problem (TSP), if for
any instance I of the TSP (an instance of a problem is obtained by specifying
particular values of the parameters of the problem, for example, the graph of
Figure 1.38 is an instance of TSP) if the ratio
r(A) = A(I)/E(I)
is bounded above by a constant c, where A(I) is the value found by the algo-
rithm A on the input I and E(I) is the exact value of the instance. The-
oretically, we may obtain E(I) by executing the brute-force algorithm on the
instance I. Note that c ≥ 1. If the constant c is 1, then A is an exact algorithm.

We shall now study three approximate algorithms for TSP [7].

Nearest neighborhood (NN) algorithm:

This is a heuristic. A heuristic for TSP is an algorithm which gives an accept-
able solution but not necessarily optimal solution for TSP. We shall describe
the heuristic.
94 Discrete Mathematics

TABLE 1.14: Execution of

NN algorithm on Figure 1.38
Partial tour Vertex chosen
1 3
(1, 3) 2
(1, 3, 2) 4
(1, 3, 2, 4) 1
(1, 3, 2, 4, 1)

Input: A complete graph Kn with vertices 1, 2, . . . , n and a cost function c asso-

ciated to each of the n(n − 1)/2 edges of Kn . We further assume the triangle
inequality, that is, for any three cities i, j, k, we have c(i, k) ≤ c(i, j) + c(j, k).
Output: A solution for this instance I such that the ratio

N N (I)/E(I) ≤ 1/2 (log2 n + 1)

where N N (I) is the value obtained by the N N algorithm on the instance I

and E(I) is the exact value on the instance I.
Algorithm: We shall describe in an informal way this algorithm. Choose the
starting city 1 and let c1 = 1. Choose a city c2 = c1 (2 < n) such that
c(c1 , c2 ) is a minimum (ties may be broken by choosing a smallest city c2 ).
Next choose a city c3 (3 < n and c3 = c1 , c2 ) such that c(c2 , c3 ) is a minimum.
More generally, having thus constructed a partial tour c1 , c2 , . . . , cp , (p < n),
we choose cp+1 which is different from ci for i = 1, 2, . . . , p and c(cp+1 , cp ) is
a minimum. Finally, having chosen cn , we choose c1 , the starting city to close
the path. We have thus constructed the complete tour (c1 , c2 , . . . , cn , c1 ). An
algorithm must be seen to be believed. Let us execute the NN algorithm on
an example. Let us refer to the graph of Figure 1.38.

Example 1.14: Execution of NN algorithm

Table 1.14 illustrates the execution.

The complete tour obtained is (1, 3, 2, 4, 1), whose cost is 19, which is exact
in this case. We shall now present a C program to implement the NN algorithm
for TSP.

C program implementing the NN algorithm:

#include <stdio.h> //load the input/output library
#include <stdlib.h>
#include <limits.h> //INT_MAX is defined in limits.h
#define N 4 // N is the number if cities
 Graph Algorithms I 95

//int v[N+1]={0,1,2,3,4} ; // v= villes, means cities in French

int c[N+1][N+1]={{0,0,0,0,0},{0,0,4,3,5},{0,4,0,5,6},{0,3,5,0,7},
{0,5,6,7,0}};
// c, the cost matrix
int t[N+1]; //t, the resulting tour
int mark[N+1]={0,0,0,0,0};
// mark[i]=0 if city i is not yet visited
//mark[i]=1, otherwise

void tsp_nn( )

int i,j,k,min,counter; //counter is the number of cities visited

//min, the nearest cost neighborhood
// k is the name of the next city to include
t[1]=1; //start with 1
mark[1]=1;
counter=1;
i=1; //i, current city

while (counter<N)
{
min=INT_MAX; //initialisation
for ( j=1;j<=N;j++)
if ((mark[j]==0)&&(i!=j)&&(c[i][j]<min))
{
min=c[i][j];
k=j;
}
mark[k]=1; //visit city k
counter=counter+1;
i=k;
t[counter]=k;
}//end while
} //end tsp_nn

int main ()
{
int i,mc=0;//mc, minimum cost tour
tsp_nn() ; //function call
for (i=1;i<N;i++)
cm=cm+c[t[i]][t[i+1]];
cm=cm+c[t[N]][t[1]];

//print the result

printf("the minimum cost is %d\n",cm);
//print tour
for (i=1;i<=N;i++)
printf("%3d",t[i]);
printf("%3d",t[1]);
system("pause");
return 0
}
96 Discrete Mathematics

TABLE 1.15: An approximate algorithm for TSP using MST

Algorithm: We shall describe the algorithm in an informal way.
Step 1: Find a minimum cost spanning tree for the graph Kn , by applying
Prim’s algorithm (see Section 1.2.1)
Step 2: To each edge of the minimum cost tree T , add one more additional
edge. This operation converts the T into an Eulerian graph ET ,
that is a connected graph in which the degree of each vertex is even.
Step 3: Find a Eulerian cycle of the graph ET. (A Eulerian cycle is a cycle
passing through each edge exactly once, but a vertex may be visited
more than once. In an eulerian cycle, each vertex x is visited exactly
d(x) times).
Step 4: Convert the Eulerian cycle into a Hamiltonian cycle, that is,
traveling salesman tour, by taking “shortcuts.”

An approximate algorithm for TSP using minimum spanning tree

and Eulerian cycle (MSTEULER):
Input: A complete graph Kn with vertices 1, 2, . . . , n and a cost function c
associated to each of the n(n−1)/2 edges of Kn . We further assume the triangle
inequality, that is, for any three cities i, j, k, we have c(i, k) ≤ c(i, j) + c(j, k).
Output: A solution for this instance I such that the ratio

MSTEULER(I)/E(I) < 2

where MSTEULER(I) is the value obtained by the MSTEULER algorithm

on the instance I and E(I) is the exact value on the instance I.

Before illustrating the algorithm with an example, let us see an algorithm

to find a Eulerian cycle in a Eulerian graph. We shall recall the definition of
a bridge. In any graph G, not necessarily connected, a bridge is an edge e of
G (e is not a loop), such that the number of connected components of G − e
is strictly greater than the number of components of G.

Algorithm to find a Eulerian cycle in a Eulerian graph:

This algorithm is invented by Fleury.
Fleury’s algorithm is given in Table 1.16.

Execution of Fleury’s algorithm:

Let us execute Fleury’s algorithm on the graph of Figure 1.39.
Start with any vertex, say, the vertex 1. Choose any edge incident with the
vertex 1, say, the edge e3 . Now E1 = {e3 } and E2 = E \ E1 . Now we are at
vertex 3 and we must not choose the edge e2 as our next edge as the edge e2
is a bridge of the edge induced subgraph E2 , because we have an alternative
edge e5 or e5 is incident with 3 which are not bridges of the edge induced
subgraph E2 . We choose the edge, say, e5 as our next edge of the cycle. We
are now at the vertex 4. Again, we must not choose the edge e4 for the same
 Graph Algorithms I 97

TABLE 1.16: Informal Fleury algorithm

Input: A connected graph (loops and multiple edges allowed) G = (X, E)
in which the degree of each vertex is even.
Output: A Eulerian cycle.
Algorithm: Start with any vertex x1 . Traverse any edge incident with
the vertex x1 (this edge may be a loop). Having reached a vertex xk ,
(1 ≤ k ≤ n), we distinguish two types of edges:
Type 1: The edges already in the cycle being constructed by Fleury’s
algorithm.
Type 2: The edges not yet included in the partial Eulerian cycle already
constructed by the algorithm.
Let us denote the set of edges of type 1 by E1 and the set of edges of
type 2 by E2 . Note that E = E1 ∪ E2 and E1 ∩ E2 = ∅. Initially, E1 = ∅
and E2 = E, the entire edge set of G.
At the vertex xk , unless there is no alternative, we select an edge incident
with xk in the edge induced subgraph E2 , which is not a bridge of the edge
induced subgraph E2 . The algorithm terminates as soon as E2 = ∅.

FIGURE 1.39: Illustration of Fleury’s algorithm.

reason and we choose an edge, say, e7 incident with 4. At vertex 6, we must

choose the loop e9 . Now again at vertex 6, since there is no alternative, we
choose the edge e8 even though it is a bridge of the edge induced subgraph
E2 = {e8 , e6 , e4 , e2 , e1 }. Continuing like this, we choose the sequence of edges
e8 , e4 , e2 , e1 (note that there are alternative choices left for these edges even
though they are bridges). Now E2 = ∅ and the algorithm stops. We have thus
the Eulerian cycle,
(e3 , e5 , e7 , e9 , e8 , e6 , e4 , e2 , e1 ).
Now we are ready to execute the MSTEULER algorithm for TSP. Consider
the graph of Figure 1.38. We shall execute the algorithm on this graph of
Figure 1.38.
Step 1. Execution of Prim’s algorithm on the graph of Figure 1.38 gives us
the spanning tree Figure 1.40.

Step 2. Doubling the edges of the graph of the spanning tree Figure 1.40 gives
us the graph of Figure 1.41.

Step 3. A Eulerian cycle of the graph of Figure 1.41 is: (e1 , e2 , e5 , e6 , e3 , e4 ).

Step 4. Taking shortcuts gives the Hamiltonian cycle: (1, 2, 4, 3, 1).
98 Discrete Mathematics

FIGURE 1.40: Spanning tree of Figure 1.38.

FIGURE 1.41: Doubling the edges of the tree of Figure 1.40.

How to take the shortcuts:

We follow the vertices visited by the Eulerian cycle already constructed. The
first edge is e1 whose ends are 1 and 2. Hence the Hamiltonian cycle starts
with the vertex 1 followed by the vertex 2. Now the next edge in the Eulerian
cycle is e2 whose end vertices are 2 and 1 (in the order appearing in the cycle),
but these vertices are already in the Hamiltonian cycle under construction.
Hence we do not include them. We now examine next edge in the Eulerian
cycle which is e5 whose ends are 1 and 4 (in the order) and the vertex 1 is
already in the cycle under construction but the vertex 4 is a new vertex (which
is not in the cycle under construction). Hence we take the shortcut from the
vertex 2 directly to the vertex 4. The partial cycle constructed is (1, 2, 4).
The next edge in the Eulerian cycle is e6 whose ends are 4 and 1. Both are
already in the partial Hamiltonian cycle. So we do not include them. The
next edge we consider is e3 whose ends are 1 and 3. But 3 is not in the cycle.
Hence we take the shortcut from the vertex 4 to the vertex 3. The cycle thus
constructed is (1, 2, 4, 3). Since all the vertices are in the cycle, we complete
the cycle by adding once more initial vertex 1 of the cycle. The cycle obtained
is (1, 2, 4, 3, 1). The cost of the cycle is 20.

Theorem 1.6. For any instance of the TSP, we have

MSTEULER(I)/E(I) < 2

where the MSTEULER(I) is the cost obtained by the MSTEULER algorithm

and E(I) is the exact cost.

Proof. Let us observe that if C is a Hamiltonian cycle and e is an edge of

C, then C \ e is a spanning tree of the graph G. Hence, if T is a minimum
cost spanning tree of G, then w(T ) ≥ w(C \ e) where w(T ) is the sum of
 Graph Algorithms I 99

the weights of the edges of T and w(C \ e) is the sum of the weights of the
edges of C \ e. By doubling each edge of T , we obtain the graph ET , with
w(ET ) = 2w(T ). Since a Eulerian cycle traverses each edge of ET exactly
once, its total weight is also w(ET ). Taking shortcuts will not increase the
cost w(ET ) as:

c(x1 , x2 ) + c(x2 , x3 ) + · · · + c(xk−1 , xk ) ≤ c(x1 , ck )

by triangle inequality, where c is the cost function associated with each edge.
Hence the theorem.

Another approximate algorithm for TSP using a minimum

matching (tspmm):
Input: A complete graph Kn with vertices 1, 2, . . . , n and a cost function c
associated to each of the n(n−1)/2 edges of Kn . We further assume the triangle
inequality, that is, for any three cities i, j, k, we have c(i, k) ≤ c(i, j) + c(j, k).
Output: A solution for this instance I such that the ratio

tspmm(I)/E(I) < 3/2

where tspmm(I) is the value obtained by the MSTEULER algorithm on the

instance I and E(I) is the exact value on the instance I.
Algorithm: The algorithm is given in Table 1.17.

TABLE 1.17: An approximate algorithm for TSP using a minimum

matching
Algorithm: We shall describe the algorithm in an informal way.
Step 1: Find a minimum cost spanning tree T of Kn using Prim’s algorithm
in Section 1.2.1. Hence this step is the same as the first step of the
previous algorithm.
Step 2: Recall that in any graph, the number of vertices whose degrees are
odd integer is always an even integer (see [6]). We find a perfect
matching (a matching is a collection of edges no two of which are
adjacent. A perfect matching is a matching with exactly n/2 edges
for a n vertex graph. Of course, n must be even.) among the
vertices of odd degree of T whose total weight is a minimum.
Step 3: Add the edges of a minimum weight matching found in step 2. The
graph thus obtained M T is a connected graph whose degrees are
even (because of the addition of a matching). Hence M T is an
eulerian graph. Find an eulerian cycle using Fleury’s algorithm 1.16.
Step 4: As in the previous algorithm (see step 4 of Table 1.15), take
shortcuts of the Eulerian cycle obtained in step 3.
100 Discrete Mathematics

1.13 Exercises
1. Consider a directed graph G represented by its adjacency matrix. Write
a program in C to find the vertex-deleted subgraph G − k where k is
any vertex of the graph G. Find the complexity of your program.

2. Consider a directed graph G represented as linked lists. Write a program

in C to find the vertex-deleted subgraph G − x where x is any vertex of
the graph G. Find the complexity of your program.

3. Consider a directed graph G represented by its adjacency matrix. Write

a program in C to find the arc-deleted subgraph G−(x, y) where (x, y) is
any directed edge of the graph G. Find the complexity of your program.
4. Consider a directed graph G represented as linked lists. Write a program
in C to find the arc-deleted subgraph G−(i, j) where (i, j) is any directed
edge of the graph G. Find the complexity of your program.

5. Write a program in C to represent a graph in the form of its incidence

matrix. What is the complexity of your program?
6. Let G be a connected simple graph containing at least one bridge, with
weights associated to each edge of the graph G. Let T be a minimum
cost spanning tree found by Prim’s algorithm/Kruskal’s algorithm.
Then, T contains all the bridges of the graph G. True or false? Justify
your answer.
7. Let G = (X, E) be a simple graph on n vertices with weights associated
to each edge of G and let T be a minimum cost spanning tree obtained
by applying Prim’s algorithm on the graph G. Let c1 , c2 , . . . , cn−1 be
the edges of the tree T arranged in non-decreasing order; that is, c1 ≤
c2 ≤ · · · ≤ cn−1 . Consider any spanning tree T  of G with edge costs
c1 ≤ c2 ≤ · · · ≤ cn−1 .
Show that ci ≤ ci for all i with 1 ≤ i ≤ n − 1.
8. Apply Prim’s algorithm on the graph of Figure 1.42. The weights of
edges are written on each edge. Give the weight of the spanning tree
thus obtained.

FIGURE 1.42: A weighted simple graph.

 Graph Algorithms I 101

9. Apply Kruskal’s algorithm on the graph of Figure 1.42. The weights of

edges are indicated on each edge. Give the weight of the spanning tree
obtained.

10. Apply Dijkstra’s algorithm on the graph of Figure 1.43 by taking the
vertex 1 as the source vertex. The weights of arcs are indicated on each
directed edge. Give the array “dad.” Draw the spanning arborescence
obtained by the algorithm.

11. Execute Warshall’s algorithm on the graph of Figure 1.43 by ignoring

the weights associated to each arc.

12. Execute Floyd’s algorithm on the graph of Figure 1.43. Give the matrix
IN T ER5×5 obtained at the end of the execution. Execute the call inter-
path(1,4) and write the vertices printed by the call.

FIGURE 1.43: A weighted directed graph.

13. Prove that in a directed graph, the number of directed walks of length k
from a vertex i to a vertex j is the (i, j) entry of the matrix M k where M
is the adjacency matrix of the graph. Using this result, design a O(n4 )
algorithm for transitive closure of a graph.

14. Execute Hopcroft’s algorithm to find the biconnected components on

the graph of Figure 1.44. The vertices are arranged in increasing order
in each adjacency list. Process the vertices in increasing order. Draw
clearly the dfs tree together with the back edges. Give the arrays a dfs
number and LOW.
102 Discrete Mathematics

FIGURE 1.44: A simple connected graph.

15. Add suitable instructions in Hopcroft’s biconnected components algo-

rithm to print the cut vertices and bridges of a connected graph with at
least two vertices.
16. Execute Tarjan’s strongly connected components algorithm on the graph
of Figure 1.44 after giving following directions/orientations to the edges
of the graph:
Orient the edges as follows: (1, 2), (2, 3), (3, 1), (3, 7), (7, 8), (8, 3),
(4, 3), (1, 5), (2, 5), (3, 5), (5, 6), (6, 10), (10, 9), (9, 5).
Write the arrays dfsn and LOWLINK. Draw the dfs spanning forest with
forward, backward, and cross dotted arcs.

17. Execute Dijkstra’s negative cost algorithm on the following graph of five
vertices. The graph is represented by its 5 × 5 costs matrix W.
⎛ ⎞
0 7 8 9 10
⎜0 0 8 9 10 ⎟
⎜ ⎟
W =⎜ ⎜ 0 −2 0 9 10 ⎟ ⎟
⎝0 −4 −3 0 10 ⎠
0 −7 −6 −5 0
Can you guess the complexity of Dijkstra’s negative cost algorithm from
this example?

18. Execute the topological sort algorithm on the graph of Figure 1.44 after
assigning the following orientations to the edges of the graph:
Orient the edges as follows: (1, 2), (1, 3), (1, 5), (2, 3), (2, 5), (3, 5),(4, 3),
(5, 6), (6, 10), (10, 9), (5, 9), (3, 7), (7, 8), (3, 8).

19. Write a program in Pascal for Dijkstra’s algorithm.

20. Write a program in C for Dijkstra’s algorithm.

21. Write a program in Pascal for Prim’s algorithm.

22. Write a program in C for Prim’s algorithm.

Chapter 2
Graph Algorithms II

The things we have understood in such a way that they can be

instructed to a computer could be called science. Otherwise they
are only art.

D. E. Knuth

In this chapter on graph algorithms II, we introduce another systematic

way of searching a graph known as breadth-first search (BFS). Testing if a
given graph is geodetic and finding a bipartition of a bipartite graph are given
as applications. Next, matching theory is studied in detail. Berge’s charac-
terization of a maximum matching using alternating chain and König-Hall
theorem for bipartite graphs are proved. We consider matrices and bipartite
graphs: Birkhoff-von Neumann theorem concerning doubly stochastic matrices
is proved. Then, the bipartite matching algorithm using a tree growing proce-
dure (Hungarian method) is studied. The Kuhn-Munkres algorithm concern-
ing maximum weighted bipartite matching is presented. Flow in transporta-
tion networks is studied. The Ford-Fulkerson algorithm to find a maximum
flow and a minimum cut capacity in a transportation network is treated.
As applications of bipartite networks, we give an algorithm to determine if
a given sequence of pairs of integers is realizable by a p-graph (that is, a
graph in which the number of parallel edges/directed arcs from a vertex a
to a vertex b is at most p). Greedy algorithm and approximate algorithm
for a minimum transversal in a graph are studied. Havel-Hakimi’s algorithm
concerning degree sequence of a simple graph is studied. Exact exponential
algorithm and polynomial greedy algorithm for graph coloring are studied.
Finally, the relation between the chromatic number and the examination time
tabling problem is discussed.

2.1 Breadth-First Search

In Chapter 1 on Graph Algorithms I, we have discussed the depth-first
search (dfs). Another systematic way of processing/visiting the vertices of a
graph is the breadth-first search (bfs). In dfs, we go as deeply as possible into
the graph before fanning out to other vertices. On the other hand, during the
bfs, we “fan” out to as many vertices as possible, before going “deeper” into
the graph (see [3]).

103
104 Discrete Mathematics

FIGURE 2.1: A graph G and its breadth-first search, bfs(G).

FIGURE 2.2: Breadth-first search tree of graph G of Figure 2.1.

We shall illustrate this procedure by an example. Let us consider the con-

nected simple graph G of Figure 2.1.
The graph is represented by the adjacency lists: L(a) = (b, c, d); L(b) =
(a, c, e, f ); L(c) = (a, b, d, f, g, h); L(d) = (a, c, h, i); L(e) = (b, f ); L(f ) =
(b, c, e); L(g) = (c, h); L(h) = (c, d, g, i); L(i) = (d, h).
Initially all the vertices are marked “unvisited.” We start the search with
any vertex as the starting vertex, say, the vertex a. This vertex a will be
the root of the bfs tree under construction. We mark the vertex a “visited.”
Then, we scan completely the adjacency list of the vertex a looking for vertices
marked “unvisited.” We mark each of the vertices b, c, d “visited” and add the
edges ab, ac, ad to the bfs tree. We now scan the list L(b) looking for vertices
marked “unvisited.” Two new vertices e, f are found and they are marked
“visited” and the edges be, bf are added to the bfs tree.
Now by scanning the list L(c), we find two new vertices g, h and the edges
cg, ch are added to the tree. Next by scanning the list L(d), only one new
vertex i is found and the edge di is added to the tree. The tree T obtained
by the search is defined by the solid edges in Figure 2.2 and all other edges of
the graph G which are not in the tree T are drawn as dotted edges.
We observe the following properties of the dotted edges.

Observation 2.1. A dotted edge, which is not a loop, is always a cross edge,
that is, a dotted edge joins two vertices x, y where neither is an ancestor or
descendant of the other. This property is just the opposite of the property of
the dotted edges of the dfs of an undirected graph.
 Graph Algorithms II 105

Every cross edge defines an elementary cycle as in the dfs. More precisely,
consider a cross edge xy, and let the vertex z be the closest common ancestor
of the vertices x and y in the bfs tree. Then, the unique path from z to x in
the tree and the unique path from z to y in the tree together with the cross
edge gives an elementary cycle of the graph.
For example, in Figure 2.1, the cross edge hi defines the cycle
(a, c, h, i, d, a), the common ancestor being the root vertex a.
Having seen an example, let us now describe the bfs algorithm in
pseudo-code.
Breadth-first or level-order search algorithm:
Input: An undirected graph G = (X, E) with X = { 1, 2, . . . , n }. The graph
is represented by the n lists L(i), i = 1, 2 . . . , n.
Output: A partition of the edge set E into the edge F , where F is a spanning
forest of G, with F being a vertex disjoint union of trees and the set D of
dotted edges.
Algorithm: An array mark[1..n] is used where mark[i] = 0 if and only if the
vertex i is not yet “visited” during the search. Initially, mark[i] := 0 for each
vertex i. We use a “queue of vertices” Q. Initially, the queue is empty. The bfs
tree is represented by an array of “dad,” that is, dad[j] = i if the vertex i is
the father of the vertex j in the tree, that is, ij is an edge of the bfs tree under
construction. Initially, the set of edges of the bfs tree T is empty. We use an
array bf sn[1..n] where bfsn stands for the breadth-first search number. More
precisely, bf sn[i] = k if i is the kth vertex “visited” during the bfs. Thus, the
bfsn of the root vertex of the first tree in F is 1. The integer variable counter
is initialized to 1. The algorithm is given below in pseudo-code.

procedure bfs(x : vertex);

(* bfs marks by breadth first search, all the vertices
accessible from the vertex x by an elementary path *)
var y,z:vertex; Q: queue of vertex;
begin(*bfs*)
(* initialization of Q *)
Q:= empty;
mark[x]:=1; bfsn[x]:=counter; counter;= counter +1;
enqueue x in Q;
while Q is nonempty do
begin
y:= head of the queue Q;
dequeue;(*remove the front vertex of Q*)
for each vertex z in L(y) do
if mark[z]=0
then begin
mark[z] :=1; bfsn[z]:=counter; counter: counter+1;
enqueue z in the queue Q;
106 Discrete Mathematics

dad[z]:=y;(* add the edge yz to the forest F*)

F:=F + {yz};
end;(*while*)
end;(*bfs*)

begin(* main program*)

(*initialization of array mark , F and counter*)
for each vertex x do
mark[x]:=0;
F:=empty;counter:=1;
for each vertex x do
if mark[x] = 0 then
bfs(x);(* call*)
end;

Note that if the graph is not connected, bfs will be called from the main
program on a vertex of each connected component, that is, there will be as
many calls to bfs as the number of connected components of the graph. Let
us now discuss the complexity of bfs.

Complexity of bfs algorithm:

As in the dfs, each vertex is processed at most a constant amount of time
(e.g., marking, enqueueing, dequeueing) and each edge xy is traversed exactly
twice, once from x and once from y. Hence, the time taken by bfs is O(n+2m)
which is O(max(n, m)) by the sum rule of “big-oh” notation.

Property 2.1 (Level sets: Partition of the vertex set induced by bfs).
Bfs induces a partition of the vertex set of a connected graph with the following
properties:
The vertex set
X = L0 (a) ∪ L1 (a) ∪ · · · ∪ Lk (a)
where a is the root vertex, k is the eccentricity of the vertex a and the set
Li (a) consists of all vertices at distance i from the vertex a. Symbolically,
Li (a) = { x | d(a, x) = i }.
The sets Li (a) are called level sets with respect to the vertex a. By the definition
of level sets, there cannot be an edge of the graph between vertices of the level
sets Li (a) and the level sets Lj (a) if |i − j| > 1.

Example 2.1: Level sets induced by bfs

Let us refer to graph G of Figure 2.1. In this graph, the level sets are
L0 (a) = { a } then L1 (a) = { b, c, d } and finally L2 (a) = { e, f, g, h, i }.
Note that Lk (a) = ∅ for all k ≥ 3.
 Graph Algorithms II 107

This partition can be easily obtained from the bfs algorithm as follows:
An array variable L[1..n] of integer is declared together with an integer
variable k to assign levels to vertices. k is initialized to 1. In the beginning
of the bfs procedure, we add the instruction L[a] = 0 and write the statement
L[z] := k just after the instruction mark[z] := 1. Finally, the variable k is
updated by the statement k := k + 1 just before the end of the while loop.
Once the array L is available, to find the vertices at level i from the root
vertex a, we simply scan the L array to print the vertices x for which L[x] = i.

2.2 Applications of bfs Algorithm

As in the case of the dfs algorithm, bfs can be used directly to find the
connected components of a graph and to test the existence of cycles in graphs.
We shall now see few applications of bfs: Testing if a given graph is geode-
tic, testing if a graph is bipartite and finding a maximum matching in a
graph.

Geodetic graph:
An undirected graph is geodetic if between any two vertices there is a unique
elementary path of shortest length. Let us recall that the length of an elemen-
tary path is the number of edges in the path.
By the definition, a geodetic graph must be a simple connected graph. For
example, a tree is a geodetic graph. Other examples are the famous Petersen
graph of Figure 2.3 and any elementary cycle of odd length.

An algorithm to test if a graph is geodetic:

The algorithm is based on the following simple characterization of geodetic
graphs due to Parthasarathy and Srinivasan (see [11]).

Theorem 2.1 (Characterization of geodetic graphs). Consider a simple

connected graph G and a partition of the vertex set X generated by bfs from an
arbitrary root vertex x. Then, the graph is geodetic if and only if for every root

FIGURE 2.3: An example of a geodetic graph.

108 Discrete Mathematics

vertex x, a vertex y of the ith level set Li (x) is joined by an edge to exactly
one vertex z in the (i − 1)th level set Li−1 (x).
Proof. Let G be a geodetic graph. Then, each vertex y in Li (x) is adjacent to
exactly one vertex z of Li−1 (x). For, if there is a vertex y ∈ Li (x) adjacent to
two vertices z1 , z2 belonging to Li−1 , then by the definition of level sets, there
are two paths of shortest length i from the root vertex x to the vertex z, one
going through z1 and the other passing via z2 , a contradiction.
For the converse part, suppose that every vertex of the ith level set is joined
to exactly one vertex of the (i − 1)th level set, for all root vertex x ∈ X. If G is
not geodetic, then there are vertices x and y such that there are two shortest
paths P1 and P2 from x to y. Let z = y be the last vertex common to both
P1 and P2 while going from x to y. Then, clearly, z = x for, otherwise the two
paths will be identical. Now, if the distance between x and z is i, then in the
bfs with x as the root vertex, the vertex z in the ith level set has two adjacent
vertices in the (i − 1)th level set, a contradiction.
We are now ready to describe the algorithm:
Input: A connected simple graph G.
Output: “1” if G is geodetic, “0” if not.
Algorithm: By the above characterization of geodetic graphs, G is geodetic if
and only if there is no dotted edge between two consecutive level sets, for any
root vertex x ∈ X.
The algorithm is given below in pseudo-code: We declare an integer vari-
able “geo” which is initialized to 1. If there is a dotted edge between two
consecutive level sets, then the variable “geo” is set to 0 and the procedure
“bfs” terminates, thanks to the statement “return” and the control falls back
to the statement which activated the procedure “bfs.” L[s] = L[f ] + 1 if and
only if the vertex f is the father of the vertex s in the bfs tree. Note that the
square brackets in L[x] denote the level number of the vertex x, whereas the
parenthesis in L(x) denotes the list of vertices adjacent to the vertex x.

procedure bfs_goedetic(x : vertex);

var y,z:vertex; Q: queue of vertex;
begin(*bfs_geodetic*)
(* initialization of Q *)
Q:= empty;
mark[x]:=1; L[x]:=0;
enqueue x in Q;
while Q is nonempty do
begin
y:= head of the queue Q;
dequeue;(*remove the front vertex of Q*)
for each vertex z in L(y) do
if mark[z]=0
 Graph Algorithms II 109

then begin
mark[z] :=1;
enqueue z in the queue Q;
dad[z]:=y;(* add the edge yz to the T*)
L[z]:= L[y]+1;(*assign level number to z*)
end
else (*z already visited*)
if L[y]=L[z]+1
then
if dad[y] is not z
(* yz is a dotted edge joining consecutive level sets*)
then begin geo:=0; return; end;
end;(*while*)
end;(*bfs*)

begin(* main program*)

(*initialization of mark ,T,geo*)
for each vertex x do
mark[x]:=0;
geo:=1; T:=empty;
for each vertex x do
begin
bfs_geodetic(x);(*call*)
if geo = 1
then (* reinitialize the array mark and tree T*)
for each vertex x do begin mark[x]:=0; T:= empty; end
else break;(* quit the for loop, not geodetic*)
end;
write(geo);
end;

Remark 2.1. During the bfs, each edge is traversed twice. A dotted edge
joining two vertices of the same level is first traversed from left to right and a
dotted edge xy joining two vertices of consecutive levels with x in level i and y
in level i + 1 is first traversed from x to y, that is, geometrically, assuming
that the sons of a vertex are drawn in the left to right manner, the vertex x
is to the right of the vertex y. Hence in the above algorithm, we can simply
replace the “else” part of the “for loop” in the procedure “bfsgeodetic” by the
following statement:
if L[y] = L[z] − 1 then geo:= 0;

Complexity of the geodetic graph algorithm:

The “for loop” of the main program activates the procedure “bfs geodetic” at
most n times for an n vertex graph. The procedure “bfs geodetic” is activated
exactly n times if and only if the graph is geodetic. Since the cost of each
activation is O(max(n, m)), the total time spent is O(n max(n, m)).
110 Discrete Mathematics

Algorithm for testing if a graph is bipartite using bfs:

Input: An undirected connected graph G = (X, E). If the graph is not con-
nected, we can consider its connected components separately.
Output: “yes” if the graph G is bipartite, “no” otherwise.
Algorithm: The algorithm is given in Figure 2.4. The algorithm is based on
the following property:

Property 2.2. A graph is bipartite if and only if there are no dotted edges
joining two vertices of the same level sets.

Proof. By König’s theorem (see the chapter on Introduction to Graph Theory

in [6]), a graph G is bipartite if and only if G does not contain an odd cycle.
Let us first apply the bfs procedure on G from a root vertex a. Suppose
there is a dotted edge xy joining two vertices of the same level, say, x, y ∈ Li [a].
Now consider the closest common ancestor of the vertices x and y in the bfs
tree. Let s be such an ancestor. Then, the shortest path from s to x and
the shortest path from s to y together with the dotted edge xy will be an
elementary cycle of length d(s, x) + d(s, y) + 1 = 2d(s, x) + 1, which is odd, a
contradiction.
Conversely, suppose G contains an odd elementary cycle. We shall show
that there is a dotted edge between two vertices of the same level, when we
perform bfs of the graph G. Let C = (x1 , x2 , . . . , x2p+1 ) be an elementary odd
cycle of shortest possible length. First of all, we shall prove that the distance
d(x1 , xp+1 ) = p in the graph G (evidently, the distance between x1 and xp+1
along the edges of the cycle C is p). Otherwise, suppose there is a path P in
the graph G between x1 and xp+1 with length strictly less than p. Let the
distance path from x1 to xp+1 along the edges of the cycle be Q. Since the
paths P and Q have the same initial and final vertices and P = Q, their union
P ∪ Q contains at least one elementary cycle. Since the length of the path Q <
the length of the path P , there must be an elementary cycle C  in P ∪ Q with
the following property.
Exactly two vertices xi and xj of the cycle C are in C  ∩ Q and the length
of the subpath of P from xi to xj > the length of the subpath of Q from xi
to xj . In fact, the cycle C  is the union of the subpath of P from xi to xj and
the subpath of Q from xi to xj .
Since p is the length of a smallest possible odd cycle and the length of the
cycle C > the length of the cycle C  , the length of the cycle C  must be even.
This means that the subpaths of P from xi to xj and the subpath of Q from
xi to xj are of the same parity. This implies that the removal of the edges of
the subpath of P from xi to xj and adding the edges of the subpath of Q from
xi to xj gives us an odd cycle whose length is < the length of the cycle C,
a contradiction. Hence, the distance d(x1 , xp+1 ) = p in the graph G. By a
similar argument, it can be proved that the distance between the vertices x1
and xp+2 is p in the graph G.
Now let us perform the bfs of the graph G starting from the root vertex x1
which belongs to the cycle C. By the definition of the level set Lp (x1 ), the
 Graph Algorithms II 111

FIGURE 2.4: Algorithm to test if a connected graph is bipartite.

vertices xp+1 and xp+2 are in the level set Lp (x1 ). But then, the vertices xp+1
and xp+2 are joined by a dotted edge (since they are consecutive vertices of
the cycle C).
Hence the property.
112 Discrete Mathematics

Property 2.3. Consider a connected bipartite graph G. Perform a level order

search of the graph G starting from any vertex x. Then, the bipartition of the
graph G is given by the union of evenly subscripted level sets and the union of
oddly subscripted level sets. Symbolically, the bipartition of the vertex set X
is L0 (x) ∪ L2 (x) ∪ L4 (x) ∪ · · · and L1 (x) ∪ L3 (x) ∪ · · · .
Proof. Since graph G is bipartite, by Property 2.2 there are no cross edges
between two vertices of the same level. By Property 2.2 of the level sets,
there can be no edge between any two oddly/evenly subscripted level sets.
This implies that there can be no edge between any two vertices of the union
e(x) e(x)
∪i=0 L2i (x) and no edge joining two vertices of ∪j=1 L2j−1 , where e(x) denotes
the eccentricity of the vertex x.

Example 2.2: Bipartition of a bipartite graph

Consider graph G of Figure 2.5.
Let us perform a level order search from the vertex 1. Vertices are
listed according to the increasing order in each adjacency list. The level
order search gives us the bfs tree of Figure 2.5.
L0 (1) = { 1 } L1 (1) = { 2, 6, 7 } L2 (1) = { 3, 8, 5, 12 } L3 (1) =
{ 4, 9, 13, 11 } L4 (1) = { 10 }. Hence the bipartition of the graph
is L0 (1) ∪ L  2 (1) ∪ L4 (1) = { 1, 3, 8, 5, 12, 10 } and L1 (1) ∪ L3 (1) =
{ 2, 6, 7, 4, 9, 13, 11 }.

Having seen Properties 2.2, 2.3 and Example 2.2, we can now write the
algorithm to test if a given graph is bipartite.
We perform a level order search from an arbitrary vertex. We use a Boolean
variable “bip” which is initialized to 1 (means true). If we find a dotted
edge joining two vertices in the same level set, then the variable “bip” is set
0 (means false) and the procedure “bfs bipartite” terminates thanks to the
instruction “return.” If the graph is bipartite, the algorithm writes the bipar-
tition by scanning through the array L[1..n] printing the evenly subscripted

FIGURE 2.5: Finding a bipartition of a bipartite graph G of Figure 2.5.

 Graph Algorithms II 113

vertices in one line and oddly subscripted vertices in the next line (n is the
number of vertices of the graph).
Complexity of the algorithm “bfs bipartite”:
Clearly the complexity is the same as that of the bfs which is O(max(m, n)).

2.3 Matchings in Graphs

Consider an undirected graph G = (X, E). A matching M in G is a set
of edges of G satisfying the following two conditions: 1. No loop belongs to
M , 2. No two of the edges of M are adjacent, that is, no two of the edges of
M share a common vertex. The following example illustrates the concept.

Example 2.3
M = { f a, gb, hc, id, je } is a matching of graph G of Figure 2.6.
Another matching is M  = { ab, cd, f j, gh }.

A vertex x of the graph G is saturated by the matching M or M − saturates

the vertex x, if there is an edge in M incident with the vertex x. The set of
all saturated vertices under the matching M is denoted by S(M ). If the edge
xy is in the matching M , we say that the vertex x is matched under M with
the vertex y.
A matching M of the graph G is a maximum matching, if there is no
matching M  of G such that |M  | > |M |. Here, |M | stands for the number
of edges in M . A perfect matching M is a matching which saturates every
vertex of the graph G. Perfect matchings are also called 1 -factors. Every
perfect matching is a maximum matching but the converse is not always true.
Note that the number of edges in a perfect matching is exactly n/2 where
n is the number of vertices. Hence the minimal condition for the existence of
a perfect matching is that the number of vertices of the graph must be even.
A perfect matching in a graph of n vertices may be viewed as a “bijection
along the edges” from a suitable set of n/2 vertices onto its complement set
of vertices.

FIGURE 2.6: Illustration of matchings.

114 Discrete Mathematics

FIGURE 2.7: A 3-regular graph without a perfect matching.

When we discuss the maximum matching problem, we may assume without

loss of generality that the graph is a simple graph, that is, it has no loops and
no multiple edges.
Let us refer to Example 2.3 (see the graph G of Figure 2.6). M is a perfect
matching, whereas M  is not. The graph of Figure 2.7 is an example of a
3-regular graph without a perfect matching.

Example 2.4
We shall see below a few real-world situations in which matchings
arise naturally.
We would like to distribute 64 square chocolates among 32 children.
Let us imagine that the 64 pieces are arranged in the form of an 8 × 8
square, like a Chess board. The first child gets two square pieces which
are at the diagonally opposite corners of the 8 × 8 square.
Question: Is it possible to distribute the remaining 62 square pieces
(corresponding to the truncated Chess board) among the remaining 31
children in such a way that each child gets a 1 × 2 rectangular piece?
The problem is equivalent to finding the existence of a perfect match-
ing in the following graph: The vertices correspond to the 62 squares of
the truncated Chess board and two vertices are joined by an edge if the
squares corresponding to vertices share a common side, not merely a
single point. We can demonstrate that the graph thus defined does not
possess a perfect matching. We will use the following parity argument
(see Tables 2.1 and 2.2).
Color the squares of an 8 × 8 Chess board alternately black and
white, that is, if a square s is colored black, then a square sharing a
common side with the square s should be colored white and vice versa.
Note that the two diagonally opposite corner squares are colored with
the same color. Hence the truncated Chess board has either 32 black
 Graph Algorithms II 115

squares and 30 white squares or 32 white squares and 30 black squares.

But an edge of a matching saturates exactly one black square and one
white square. In order to obtain a perfect matching, we should have the
same number of black and white squares. Hence we cannot distribute
the 62 pieces among the 32 children in the form of a 1×2 rectangle.

TABLE 2.1: Checker board

× B W B W B W B
B W B W B W B W
W B W B W B W B
B W B W B W B W
W B W B W B W B
B W B W B W B W
W B W B W B W B
B W B W B W B ×

Example 2.5: Domino

In Table 2.2, the shape of a dominoe is drawn. The dominoe is drawn
horizontally in the table. It can be also drawn vertically.

TABLE 2.2: Domino

W B

Example 2.6: Personnel assignment problem

In a certain company, there are p technicians t1 , t2 , . . . , tp and q jobs
j1 , j2 , . . . , jq . Each technician is qualified to do one or more jobs.
Question: Is it possible to assign each technician one and only one
job for which he/she is qualified?
We construct a bipartite graph with vertices { t1 , t2 , . . . , tp } ∪
{ j1 , j2 , . . . , jq }. Two vertices tk and jr are joined by an edge if and
only if the technician tk is qualified for the job jr . Now the problem is
to find existence of a matching saturating each vertex tk for 1 ≤ k ≤ p
in the bipartite graph thus constructed.
116 Discrete Mathematics

FIGURE 2.8: Illustration of the dancing problem.

Example 2.7: Dancing problem

In a certain college, each girl student has exactly k > 0 boyfriends
and each boy student has exactly k > 0 girlfriends.
Question: Is it possible to organize a weekend dance in such a way that
each boy dances with exactly one of his girlfriends and each girl dances
with exactly one of her boyfriends? (See Figure 2.8.)
In the graph of Figure 2.8, each boy bi has exactly 2 girlfriends and
each girl gi has exactly 2 boyfriends. Thus, the bipartite graph is regular
of degree 2. The perfect matching M = { b1 g3 , b2 g1 , b3 g2 } corresponds
to a dance in which the boy b1 dances with the girl g3 , b2 with g1 and
b3 with g2 .
Such a dance is always possible. We construct a bipartite graph in
which each vertex corresponds to a student and two vertices are joined
by an edge if one vertex corresponds to a boy b and the other to a
girl g with b a boyfriend of g, g a girlfriend of b. We can show that the
bipartite graph thus constructed possesses a perfect matching. This will
be proved later.

Alternating and augmenting paths with respect to a matching:

Consider a matching M in a graph G = (X, E). An M -alternating path or an
alternating path with respect to M is an elementary path in G whose edges
are alternately in M and E \ M , that is, for any two consecutive edges, one of
the edges should be in M and other must be in E \M . An M -augmenting path
is an M -alternating open path whose initial and final vertices are unsaturated
by the matching M . The following example clarifies the notion (see Figure
2.9). Note that an edge xy of G where the vertices x and y are unsaturated
by a matching M is an augmenting path relative to the matching M . A path
of length z ero, that is, a path consisting of the single vertex (x) is regarded
as an alternating path.
 Graph Algorithms II 117

FIGURE 2.9: A graph and its alternating paths.

Example 2.8: Alternating path

Consider a matching M = { 25, 34, 67, 89 } of the graph G of Figure
2.9. The path P1 = (1, 2, 5, 4, 3) is an alternating path with respect to
the matching M . The path P1 is not an M -augmenting path because
the initial vertex 1 of P1 is not saturated by the matching M .
On the other hand, the path P2 = (1, 2, 5, 8, 9, 10) is an
M -augmenting path, because the initial vertex 1 and the final vertex
10 of P2 are saturated by the matching M .

We make the following observation: Consider an elementary cycle C2p =

(x1 , x2 , . . . , x2p , x1 ) of length 2p. Then, the set of alternating edges of this
cycle M = { x1 x2 , x3 x4 , . . . , x2p−1 x2p } is a matching, as well as the set
M  = { x2 x3 , x4 x5 , . . . , x2p x1 }. In the same manner, alternating edges of an
elementary open path form a matching.

Property 2.4. Consider a matching M in a graph G. Let P be an augmenting

path with respect to the matching M . Then, the length of the path P is an odd
integer.

Proof. By the definition of an augmenting path, the initial and final vertices
of P are distinct and are unsaturated by the matching. Hence the path starts
with an initial edge not in M and ends with an edge also not in M . Hence the
number of edges in the path P must be odd.

Let us recall the operation of symmetric difference of two sets: For any two
sets A and B, the symmetric difference of A and B denoted by AΔB is the
set (A ∪ B) \ (A ∩ B) which is also equal to (A \ B) ∪ (B \ A).

Lemma 2.1 (Berge). Consider a graph G = (X, E) and two matchings M1

and M2 in the graph G. Let H = (X, M1 ΔM2 ) be a spanning subgraph of G
where M1 ΔM2 is the symmetric difference (M1 ∪ M2 ) \ (M1 ∩ M2 ). Then, the
118 Discrete Mathematics

connected components of the graph H can only be one of the following three
types:

Type 1: An isolated vertex x, that is, the degree of the vertex x in the
graph H is zero.

Type 2: An elementary cycle of even length whose edges are alternately

in M1 and M2 .
Type 3: An elementary open path whose edges are alternately in M1 and
M2 whose initial and final vertices are unsaturated by exactly one of the
matchings M1 and M2 .

Proof. Let us first observe that the degree of each vertex of the spanning sub-
graph (X, M1 ) is either 0 or 1, since M1 is a matching. Similar observation
holds for the spanning subgraph (X, M2 ). Hence the degree of each vertex x in
the spanning subgraph H = (M1 ΔM2 ) is at most 2, because x can be incident
with at most one edge of M1 and at most one edge of M2 . This means that
the only possible components of the spanning subgraph are isolated vertices,
an elementary path, and an elementary cycle.

(See Figure 2.10. The spanning subgraph induced by the symmetric dif-
ference of matchings M1 ΔM2 has five connected components of which three
are isolated vertices, one is an elementary path of length 2 and the other is
an elementary cycle of length 4.)
By the definition of M1 ΔM2 = (M1 ∪ M2 ) \ (M1 ∩ M2 ), the edges alternate
in M1 and M2 in each of the component containing at least one edge of H.
Hence the length of the cycle component must be even, otherwise either M1
or M2 will not be a matching.
Finally, for a path component, if the initial edge belongs to the matching
M1 , then the initial vertex of the path component cannot be saturated by
M2 , since otherwise either the degree of the initial vertex will be 2 in H or
the initial edge belongs to both M1 and M2 , a contradiction to the symmetric
difference M1 ΔM2 . A similar argument holds for the final vertex of any path
component of H.

An efficient maximum matching algorithm in a graph hinges on the fol-

lowing theorem due to the eminent graph theorist: Claude Berge.

Theorem 2.2 (Berge). A matching M in a graph G = (X, E) is a maximum

matching if and only if the graph G possesses no M-augmenting path relative
to M .

Proof. Consider a matching M and an augmenting path P relative to M . We

shall show that the matching M is not a maximum matching. By Property 2.4,
the length of such a path must be an odd integer. Let (x1 , x2 , . . . , x2p+2 ) be
an M -augmenting path of length 2p + 1. By the definition of augmenting path
relative to M , this means that the edges xi xi+1 are in the matching M for
 Graph Algorithms II 119

FIGURE 2.10: A graph G and the spanning subgraph induced by M1 ΔM2 .

2 ≤ i ≤ 2p and the other edges of the path are in E \ M with x1 and x2p+1
unsaturated by M . Now define a new matching M  which consists of the edges
of the symmetric difference
(M ∪ P ) \ (M ∩ P ) = { M ∪ { x1 x2 , x2 x3 , . . . , x2p x2p+1 }\
{ x2 x3 , x4 x5 , . . . , x2p x2p+1 }.
(Note that P represents the set of edges of the path P .) Then, M  is a matching
and |M  | = |M | + 1, since we have removed p alternate edges of P from M
and added p + 1 alternate edges of P to M to obtain the new matching M  .
Hence M is not a maximum matching.
For the converse part, consider a matching M which is not maximum. We
shall prove the existence of an M -augmenting path. Let M  be a maximum
matching in G. Then, |M  | > |M |.
Consider the spanning subgraph H = (X, M  ΔM ) induced by the edge
set (M  \ M ) ∪ (M \ M  ). By Lemma 2.1, the connected components of H are
isolated vertices or elementary cycles of even length with edges alternately in
M and M  or an elementary path with edges alternately in M and M  with
initial and final vertices saturated by exactly one of the matchings M and M  .
120 Discrete Mathematics

Since, |M  \M | > |M \M  |, the spanning subgraph H contains at least one

more edge of M  than of M . But the (even) cycle components of H account
for the equal number of edges of M and M  . Hence there must be some path
component P of H such that P contains more edges of M  than of M . Such a
path must start with an edge of M  and end with an edge of M  . The initial and
final vertices of P are M  -saturated in H, and hence must be M -unsaturated
in G. Thus, P is an M augmenting path in G.
We are now ready to present an outline of a maximum matching algorithm:

An outline of a maximum matching algorithm:

Input: An undirected graph G = (X, E).
Output: A maximum matching in G.
Algorithm: The algorithm is given in Figure 2.11.
In Figure 2.11, P denotes the set of edges of the path P in G, and M denotes
the set of edges of the matching in G.
Let us perform the algorithm of Figure 2.11 on a given graph.

Example 2.9: Illustration of matching algorithm

Consider the graph of Figure 2.12. We are interested in finding a
maximum matching in this graph.
Initialization of matching M with an arbitrary edge 34: M = { 34 }.
The only saturated vertices are 3 and 4.
Before beginning each iteration, we perform the following test: Is
there any augmenting path relative to the matching M ? If the answer
is “yes,” we continue the iteration. Otherwise, the algorithm terminates
with a maximum matching M .
To find an augmenting path relative to the matching M , we use our
“eye ball,” that is we try to find an augmenting path visually. Table 2.3
illustrates the different steps.

FIGURE 2.11: An outline of maximum matching algorithm.

 Graph Algorithms II 121

FIGURE 2.12: A graph to illustrate matching algorithm.

TABLE 2.3: Execution of matching algorithm

Iteration Is there an M-augmenting M := (M ∪ P )\
number path P? (M ∩ P )
0 (initialization) not applicable { 34 }
1 “yes,” P = (1, 3, 4, 5) { 13, 45 }
2 “yes,” P = (2, 1, 3, 4, 5, 7) { 21, 34, 57 }
3 “yes,” P = (6, 8) { 21, 34, 57, 68 }
“no,” quit the loop

Hence a maximum matching obtained by the above algorithm is

{ 21, 34, 57, 68 }

The above maximum matching algorithm is incomplete because we do

not know yet how to find in a systematic and progressive manner an
M -augmenting path, if one exists in the graph. This can be achieved by “a tree
growing procedure” which is somewhat equivalent to the bfs. Before present-
ing this procedure, called “tree growing procedure,” we prove the following
good characterization of bipartite graphs not containing a particular type of
maximum matching.

2.3.1 An application: (k − 1)-regular subgraphs of

k-regular graphs
As an application of Berge’s matching algorithm, we solve the following
regular subgraph problem:
Input: A k-regular simple connected graph G on odd number of vertices with
no an induced star K1,3 .
Output: A (k − 1)-regular subgraph of G.
Algorithm (Sriraman Sridharan): The vertex set is { 1, 2, . . . , n }. The graph G
is represented by its adjacency matrix M = (mij )n×n where we set mij = ∞
if vertices i and j are not adjacent. In the implementation, ∞ can be replaced
by a “large” integer. The algorithm uses the following results:
122 Discrete Mathematics

Step 1: Apply Floyd’s algorithm to the matrix M = (mij )n×n (see chapter 1).
The (i,j) entry of M is the length of a shortest path between vertices i and
j in G
In step 2, we find a peripheral vertex of G
Step 2:
for i:=1 to n do
begin
set e[i] = max(mij |1 ≤ j ≤ n)
end
The ith entry of array e gives the eccentricity of vertex i
find a vertex p such that e[p] = max(e[i]|1 ≤ i ≤ n)
Step 3:
Construct the subgraph Gp = G \ p \ Γ(p)
find a perfect matching M of Gp
return G \ { p ∪ M }

FIGURE 2.13: Algorithm to find a (k − 1)-regular subgraph in a k-regular

graph with no induced star K1,3 .

If p is a peripheral vertex, that is, a vertex p such that its eccentricity

e(p) = d(G) where d(G) is the diameter of G, then G\p\Γ(p) is still connected
and a connected graph with no induced K1,3 on even number of vertices has a
perfect matching. For more details of the algorithm, see Sriraman Sridharan,
Polynomial time algorithms for two classes of subgraph problem, RAIRO,
Oper. Res. 42, 291-298, 2008.
The algorithm is given in Figure 2.13. The required subgraph is G\{ p∪M }
which is (k − 1)-regular, where M is a perfect matching of G\p \ Γ(p). The
reader is asked to execute the algorithm by hand on a 4-regular graph to
have an understanding of the algorithm. Complexity: The complexity of the
algorithm can be shown to be O(n3 ).
The following theorem can be viewed as a profound generalization of the
following simple observation:

Observation 2.2. Consider two non-empty sets A and B. There is no injec-

tive function f : A → B if and only if |A| > |B|. Stated differently, there is no
injective function from the set A to the set B, if and only if for every function
f : A → B, there is a subset S of A satisfying the inequality |f (S)| < |S|.
Here, f (S) = { f (s) | s ∈ S }, the image set of S.

Theorem 2.3 (König-Hall). Consider a bipartite graph G = (X1 , X2 ; E)

with bipartition X1 and X2 . Then, the graph G has no matching saturating all
the vertices of the set X1 if and only if there is a subset S of X1 satisfying the
inequality |N (S)| < |S| (in f act, |N (S| = |S|−1). Here, N (S), the neighbor set
 Graph Algorithms II 123

of S, denotes the set of vertices which are collectively adjacent to the vertices
of S.

Proof. We first prove the easy part of the theorem. Suppose there is a set
S of X1 satisfying the inequality |N (S)| < |S|. Then, we will prove that
there is no matching in G saturating each vertex of the set X1 . If M is a
matching saturating each vertex of X1 , then since S is a subset of X1 , there
is a subset M  of M such that M  saturates each vertex of S and |M  | = |S|.
M  is also a matching in G (since a subset of a matching is a matching) and
the vertices in S are matched under M  with distinct vertices of the set N (S).
This implies that |N (S) ≥ |S|, a contradiction.
Now consider the converse part. Suppose the graph G possesses no match-
ing saturating all the vertices of X1 . We shall show that there is a subset S
of X1 satisfying the inequality |S| < |N (S)|. By our assumption, M does not
saturate all the vertices of the set X1 . Hence there must exist a vertex x1 ∈ X1
unsaturated by the maximum matching M . Let R be the set of all vertices of
the graph G reachable by an M -alternating path from the initial vertex x1 .
Note that the path of length zero (x1 ) is vacuously an M -alternating path
and hence the set R is non-empty. Since the matching M is maximum, by
Theorem 2.2, x1 is the only unsaturated vertex in the set R.
Let S be the set of vertices of R belonging to the set X1 and let T be
the set of vertices of R belonging to the set X2 , that is, S = R ∩ X1 and
T = R ∩ X2 . Then, we shall show that

|T | = |S| − 1 and |N (S)| = T.

Consider any M -augmenting path P = (x1 , x2 , x3 , . . . , xp ) from the initial

vertex x1 . Note that an initial portion/section of this path P is also an
M -augmenting path. Since G is bipartite, all the oddly subscripted vertices
of the path P are in S and all the evenly subscripted vertices of P are in T .
By the definition of M -augmenting path, the vertex x2 is saturated under M
with x3 , x4 is saturated under M with x5 , etc.
Hence the vertices of the set S \ x1 are matched under M with the vertices
of the set T . This implies that

|T | = |S| − 1 and N (S) ⊇ T. (2.1)

Every vertex belonging to N (S) is reachable from x1 by an M -augmenting

path. For if s ∈ N (S), then by the definition of N (S), there is an s ∈ S such
that ss is an edge of the graph. We shall show that there is an M -augmenting
path from the vertex x1 to the vertex s .
Since s ∈ S, by the definition of the set S, there is an M -augmenting path
Q from x1 to s. If s is in the path Q, then we take the initial portion of
this path Q ending in s , to have an augmenting path from s to s ; otherwise,
we extend the path Q by adding one more vertex s (and the edge ss ) to Q
to obtain an augmenting path from x1 to s . Since s ∈ S, we have by the
124 Discrete Mathematics

definition of the set T , s ∈ T . Hence we have the equality

N (S) = T.

Hence, by (2.1), we have |N (S)| = |T | = |S| − 1 < |S|.

The following example illustrates the fact that Theorem 2.3 is a good
characterization of a bipartite graph G = (X1 , Y1 ; E) possessing no matching
saturating all the vertices of the set X1 .

Example 2.10: Illustration of good characterization

If a bipartite graph contains a matching saturating all the vertices
of X1 , then we have to “convince or exhibit,” the existence of such a
matching in a “reasonable amount of time,” that is, in a formal manner,
in polynomial time.
Conversely, if a bipartite graph does not possess a matching saturat-
ing all the vertices of X1 , we must be able to “exhibit” the non-existence
of such a matching in a “reasonable amount of time” (technically poly-
nomial time in the size of the input length).
Consider the bipartite graph of Figure 2.14.
To “convince quickly” that the graph G of Figure 2.14 has a matching
saturating all the vertices xi s, we have to “exhibit” such a matching. One
such matching is M = { x1 y1 , x2 y2 , x3 y3 }. Now one can “easily” verify
that the set M is indeed a matching and it saturates all the vertices xi
for i = 1, 2, 3. For example, one can color the edges of M in the graph G
and see immediately that M is a desired matching. Note that the time
taken to find the matching M is irrelevant.
Conversely, consider the bipartite graph Figure 2.15.
To “convince quickly” that the graph of Figure 2.15 has no match-
ing saturating all the vertices xi , we must exhibit a subset S of
{ x1 , x2 , x3 , x4 } such that N (S) = |S| − 1. Such a set is S =
{ x2 , x3 , x4 }. Now one can “easily” verify the equation N (S) = |S| − 1,
because N (S) = { y2 , y3 } and |S| = 3 and |N (S)| = 2. Note again here
that the time to find such a set is not taken into account. We take into
account only the time to find the neighbor set N (S) and the verification
of the equation |N (S)| = |S| − 1.

Definition 2.1 (Succinct certificate). Consider a bipartite graph G =

(X1 , Y1 ; E) having no matching saturating all the vertices of the set X1 . Then,
a set S ⊂ X1 satisfying the equality N (S) = |S| − 1 is called a succinct cer-
tificate to “exhibit” the non-existence of a matching saturating all the vertices
of the set X1 .
 Graph Algorithms II 125

FIGURE 2.14: A perfect matching in a bipartite graph.

FIGURE 2.15: A graph to illustrate the idea of a succinct certificate.

We shall now prove that the dancing problem (see Example 2.7) has always
an affirmative answer:

Corollary 2.1. Consider a k-regular (k > 0) bipartite graph G = (X1 , X2 ; E)

with bipartition X1 and X2 . Then, the graph G contains a perfect matching.
Proof. Note that the necessary condition for the existence of a perfect match-
ing in a bipartite graph is |X1 | = |x2 |. We shall first prove this minimal
condition.
Since G is bipartite, the number of edges of G is the same as the number
of edges of the graph G incident with the vertices of the set X1 (because there
are no edges between two vertices of the set X2 ). Since G is k-regular, we have
m = k|X1 |.
By the same argument, we have m = k|X2 |.
Therefore, m = k|X1 | = k|X2 |. Since k > 0, |X1 | = |X2 |.
We shall now show that there is no subset S of X1 satisfying the inequality
|N (S)| < |S|.
Let S be any subset of X1 and let E1 and E2 be the set of edges incident
respectively with the vertices of S and with the vertices of the neighbor set
N (S). Since G is bipartite, by the definition of the neighbor set N (S), we have
the relation E1 ⊆ E2 (because there may be edges of the graph G joining a
vertex of N (S) and a vertex of X1 \ S).
But |E1 | = k|S| and |E2 | = k|X2 |. Hence,

k|S| ≤ k|N (S)| because |E1 | ≤ |E2 |.

126 Discrete Mathematics

This implies that, |S| ≤ |N (S)| for every subset S of X1 . By the König-Hall
Theorem 2.3, G contains a matching M saturating every vertex of X1 . Since
|X1 | = |X2 |, the matching M must be a perfect matching.

The following corollary asserts that the edge set of a regular bipartite
graph containing at least one edge can be partitioned into edge disjoint union
of perfect matchings.

Corollary 2.2. The edge set of a k-regular bipartite graph (k > 0) can be
partitioned into k edge-disjoint perfect matchings.

Proof. Let G be a k-regular bipartite graph with k > 0. Then, by Corollary

2.1, G contains a perfect matching M1 . Now consider the spanning subgraph of
G−M of G obtained by removing the edges of the set M . Since the removal of
the edges of M reduces the degree of each vertex of G by exactly one, G − M
is a (k − 1)-regular bipartite graph. If k − 1 > 0, then we again apply the
Corollary 2.1 to get a perfect matching M2 of the graph G − M . We continue
like this until we get a bipartite spanning subgraph with no edges. Thus, we
have decomposed the edge set E of G as

E = M 1 ∪ M2 · · · ∪ M k
with Mi ∩ Mj = ∅ for all i, j with 1 ≤ i < j ≤ k.

Corollary 2.3. Consider a (0, 1) n × n square matrix A with the property

that the sum of the entries in each row is equal to the sum of the entries in
each column, each sum being an integer k > 0. Then, the matrix A can be
written as the sum of k permutation matrices.

Proof. Associate a bipartite graph G = (I, J; E) where I is the set of row

indices of the matrix A and J is the set of column indices of A. Two vertices i
and j are joined by an edge if i ∈ I and j ∈ J and the (i, j) entry of A is 1.
The degree of a vertex i is simply the sum of the entries of the row i is i ∈ I.
Similarly, the degree of the vertex j is the sum of the j-th column where j ∈ J.
Hence G is a k-regular bipartite graph. By Corollary 2.2, the edge set can be
partitioned into k disjoint perfect matchings

E = M 1 ∪ M2 · · · ∪ M k .

Each perfect matching Mi corresponds to a permutation matrix and con-

versely. For if M = { i1 j1 , i2 j2 , . . . , in jn } is a perfect matching in the graph G,
then the corresponding n × n permutation matrix P has (is , js ) entry 1 for
1 ≤ s ≤ n and all other entries of P are zeros and conversely. Hence, the given
matrix
A = P1 + P2 + · · · + Pk
where Pi s are permutation matrices corresponding to disjoint perfect
matchings.
 Graph Algorithms II 127

2.4 Matrices and Bipartite Graphs

Corollary 2.2 can be formulated in terms of (0, 1) matrices, that is, matrices
whose entries are either 0 or 1. A permutation matrix is a square matrix having
exactly one entry 1 in each row and in each column and all other entries are
z eros.
To each (0,1) matrix A, we can associate a bipartite graph in the following
manner: The vertices of the graph G correspond to the indices of the rows
and columns of the matrix A and two vertices are joined by an edge if one
corresponds to a row i of A and the other to a column j of A with the (i, j)
entry of the matrix A equal to 1. Conversely, we can associate a (0, 1) matrix
to each bipartite graph.

Example 2.11
The matrix corresponding to the bipartite 2-regular graph of Figure
2.16 is
y y2 y3
⎛ 1 ⎞
x1 0 1 1
x2 ⎝ 1 0 1 ⎠
x3 1 1 0
The perfect matching M = { x1 y2 , x2 y3 , x3 y1 } of the graph of Figure
2.16 corresponds to the permutation matrix P where
⎡ ⎤
0 1 0
P =⎣ 0 0 1 ⎦
1 0 0

A (0, 1) matrix is called a permutation matrix if its entries are either 0

or 1 and it contains exactly one entry 1 in each row and in each column.
Equivalently, a permutation matrix is obtained by permuting the rows
and columns of the identity matrix of order n.

FIGURE 2.16: A bipartite graph and its corresponding (0,1) matrix.

128 Discrete Mathematics

Corollary 2.4. Consider a (0, 1) n × n square matrix A with the property

that the sum of the entries in each row is equal to the sum of the coefficients
in each column which is equal to an integer k > 0. Then, the matrix A can be
written as the sum of k permutation matrices.

Proof. Associate a bipartite graph G = (I, J; E) where I is the set of row

indices of the matrix A and J is the set of column indices of A. Two vertices i
and j are joined by an edge if i ∈ I and j ∈ J and the (i, j) entry of A is 1.
The degree of a vertex i is simply the sum of the entries of the row i if i ∈ I.
Similarly, the degree of the vertex j is the sum of the j-th column if j ∈ J.
Hence G is a k-regular bipartite graph. By Corollary 2.2, the edge set can be
partitioned into k disjoint perfect matchings

E = M 1 ∪ M2 · · · ∪ M k .

Each perfect matching Mi corresponds to a permutation matrix and con-

versely. For if M = { i1 j1 , i2 j2 , . . . , in jn } is a perfect matching in the graph G,
then the corresponding n × n permutation matrix P has (is , js ) entry 1 for
1 ≤ s ≤ n and all other coefficients of P are zeros and conversely. Hence, the
given matrix
A = P1 + P2 + · · · + Pk
where Pi s are permutation matrices corresponding to disjoint perfect
matchings.

Birkhoff-von Neumann theorem on bi-stochastic matrices:

As an application of the König-Hall Theorem 2.3, we shall derive the Birkhoff-
von Neumann theorem on bi-stochastic matrices:
An m × n matrix M = (mij ) is bi-stochastic if the entries are all non-negative
and the sum of the coefficients of each row and each column is equal to 1.

Observation 2.3. Any bi-stochastic matrix is necessarily a square matrix. In

fact, if M is bi-stochastic with m rows and n columns, the sum of all entries of
M is the sum of n columns of M = n (as each column sum is 1). In symbols,
m
n n
m  n

mij = mij = 1 = n.
i=1 j=1 j=1 i=1 j=1

The same sum is also equal to the sum of m rows of M = m (as each row
sum is 1). Symbolically,
⎛ ⎞

m
n m


n m

mij = ⎝ mij ⎠ = 1 = m.
i=1 j=1 i=1 j=1 i=1

Hence m = n.
 Graph Algorithms II 129

Example 2.12: Bi-stochastic matrix

Clearly, every permutation matrix is bi-stochastic.
⎡ ⎤
0 1/2 1/4 1/4
⎢ 1/4 1/4 0 1/2 ⎥
M =⎢⎣ 1/4 1/4 1/4 1/4 ⎦
⎥

1/2 0 1/2 0
M is bi-stochastic.

Determinants:
Recall that for an n × n square matrix M = (mij , with mij reals, the deter-
minant of M denoted by det M is defined as the sum of the products of the
entries of M taken only one entry from each row and each column of the
matrix M . Symbolically,

det M = (−1)σ m1σ(1) m2σ(2) · · · , mnσ(n)

σ∈Sn

where the sum is taken over all possible permutations Sn of the set
{ 1, 2, . . . , n }.
Remark 2.2. Of course, this definition does not give an efficient way to cal-
culate the determinants because there are n! terms in the det M . Surprisingly,
determinants of order n can be calculated in polynomial time, in fact in O(n3 )
time by row reducing the matrix M to triangular form.
Lemma 2.2. Consider an n × n square matrix M = (mij ). In the expansion
of the determinant of the matrix M , each term is zero, if and only if there
is a regular submatrix (a matrix which can be obtained from M by removing
certain rows and columns of M ) of order k ×(n−k +1), (k ≤ n) whose entries
are all zeros.

Proof. We will apply the König-Hall Theorem 2.3 to prove the lemma.
Let us associate a bipartite graph G = (X1 , Y1 ; E) with X1 =
{ x1 , x2 , . . . , xn } and Y1 = { y1 , y2 , . . . , yn } with the edge set E = { xi yj |
mij = 0 for all i, j }.
Each term of the determinant is of the form ±m1i1 m2i2 · · · mnin . But then
this term is non-zero, if and only if each mpip = 0 for all p with 1 ≤ p ≤ n.
This is true if and only if the set of edges { x1 yi1 , x2 yi2 , . . . , xn yin } form a
perfect matching of the corresponding bipartite graph G (see Example 2.11),
that is, if and only if G has a matching saturating each vertex of X1 .
Therefore, every term in the expansion of the det M is zero if and only if
the corresponding bipartite graph contains no matching saturating each vertex
of X1 . By the König-Hall theorem, this is so, if and only if there is a subset S
of the set X1 satisfying the equality N (S) = |S| − 1. Set |S| = k. Then,
130 Discrete Mathematics

|N (S)| = k − 1 and |X2 \ N (S)| = n − k + 1 . Note that by the definition of

N (S), there are no edges between the vertex sets S and Y1 − N (S). Hence the
desired k×(n−k+1) matrix is obtained by removing the rows of the matrix M
corresponding to the vertices of X1 −S and the columns of M corresponding to
the vertices of the set N (S). Symbolically, the matrix M  = (mij )S×(Y1 \N (S))
has all its coefficient zeros.
Lemma 2.3. In a bi-stochastic matrix, there is at least one term in the expan-
sion of its determinant which is non-zero.
Proof. Consider an n × n bi-stochastic matrix M = (mij ). If all the terms
in the expansion of the det M is zero, then by Lemma 2.2, there is a subma-
trix M  of dimension k × (n − 1 + k) all of whose terms are zero. Note that
a permutation of two rows and/or two columns of a bi-stochastic matrix is
again a bi-stochastic matrix. Hence, the given matrix can be restructured and
partitioned into the following form:

1 2 ... k−1 k ... n

⎛ ⎞
1
2 ⎜ ⎟
.. ⎜ ⎟
⎜ ⎟
. ⎜ R ZEROS ⎟
⎜ ⎟
k ⎜ ⎟
⎜− − − −−− −−− −−− −−− −−− −−− ⎟
⎜ ⎟
k+1⎜
⎜
⎟
⎟
.. ⎜ ⎟
. ⎝ T N ⎠
n
In the above partition, set R = (rij )k×(k−1) , T = (tij )(n−k)×(k−1) , N =
(nij )(n−k)×(n−k+1) . The above matrix is still bi-stochastic. Note that the
row indices { 1, 2, . . . , k } correspond to the set S and the column indices
{ 1, 2, . . . , k − 1 } correspond to the set of neighbors N (S) in the König-Hall
Theorem 2.3.
Now the sum of all the entries of the matrix N is equal to n − k + 1 (by
adding the entries of N column-by-column).
But then the sum of the entries of the matrix T is equal to
⎛ ⎞

n k−1

⎝ tij⎠ = n−k −sum of all the entries of N = n−k −(n−k +1) = −1,
i=k+1 j=1

a contradiction because the entries of a bi-stochastic matrix are all non-

negative. This contradiction establishes the result.
We are now ready to prove the Birkhoff-von Neumann theorem on the
decomposition of bi-stochastic matrices in terms of permutation matrices.
To present the theorem, we need the notion of convex combination of vec-
tors. A vector of dimension n is an ordered n-tuple (α1 , α2 , . . . , αn ) where each
 Graph Algorithms II 131

αi is a real number. Two vectors v = (α1 , α2 , . . . , αn ) and u = (β1 , β2 , . . . , βn )

are declared equal, that is u = v if and only if αi = βi for all i satisfying the
inequalities 1 ≤ i ≤ n. The sum of the vectors u and v denoted by u + v is the
vector (α1 + β1 , α2 + β2 , . . . , αn + βn ) and the scalar product αv for a real α
is defined as the vector (αα1 , αα2 , . . . , ααn ).
Finally, the vector v is a convex combination of the vectors v1 , v2 , . . . , vn
if the vector v can be expressed as

v = α1 v1 + α2 v2 + · · · + αn vn

n
where each αi ≥ 0 and i=1 α1 = 1.
Before presenting the theorem, let us see an example.

Example 2.13: Convex combination

Consider the 4 × 4 bi-stochastic matrix
⎡ ⎤
1/4 1/4 1/4 1/4
⎢ ⎥
⎢ 1/2 1/4 0 1/4 ⎥
⎢ ⎥
M =⎢ ⎥.
⎢ 0 1/2 1/4 1/4 ⎥
⎣ ⎦
1/4 0 1/2 1/4

This matrix can be viewed as a vector (r1 , r2 , r3 , r4 ) of dimension

16 = 4 × 4 where ri denotes the ith row of the matrix M . We can
write the matrix M as the convex combination of permutation matrices
as follows:
1 1 1 1
M= P1 + P2 + P3 + P4
4 4 4 4
where

⎡ ⎤ ⎡ ⎤
0 1 0 0 0 0 0 1
⎢ 1 0 0 0 ⎥ ⎢ 1 0 0 0 ⎥
P1 = ⎢
⎣ 0
⎥ P2 = ⎢
⎣ 0
⎥
0 1 0 ⎦ 1 0 0 ⎦
0 0 0 1 0 0 1 0
⎡ ⎤ ⎡ ⎤
1 0 0 0 0 0 1 0
⎢ 0 0 0 1 ⎥ ⎢ 0 1 0 0 ⎥
P3 = ⎢
⎣ 0
⎥ P4 = ⎢
⎣ 0
⎥
1 0 0 ⎦ 0 0 1 ⎦
0 0 1 0 1 0 0 0
132 Discrete Mathematics

Theorem 2.4 (Birkhoff-von Neumann). A matrix is bi-stochastic if and

only if it can be expressed as the convex combination of permutation matrices.

Algorithmic proof. By definition of scalar product, a convex combination of

vectors and permutation matrices, it is easy to see that a convex combination
of permutation matrices is a bi-stochastic matrix.
For the converse part, let M = (mij ) be an n × n bi-stochastic matrix.
Then, we shall prove that M can be expressed as a convex combination of
permutation matrices. In fact, we shall give an algorithm to express M as a
convex combination of permutation matrices.
By Lemma 2.3, in the expansion of the det M , there is a term of the form
m1i1 m2i2 · · · , mnin = 0 where (i1 , i2 , . . . , in ) is a permutation of (1, 2, . . . , n)
(in Example 2.13, the product of the four en-squared terms is a term of the
determinant = zero).
Set α1 = min(m1i1 , m2i2 , . . . , mnin ). Note that α1 > 0 (in Example 2.13,
(1) (1)
α1 = 1/4). Now define a permutation matrix P1 = (pij ) where pkik = 1 and
(1)
pkis = 0 if k = s (in Example 2.13, the corresponding permutation matrix is
P1 ).
(1)
Set R1 = M − α1 P1 = (rij ). The matrix R1 possesses more zero entries
(1)
than the matrix M , because at least the coefficient rkik = 0 with α1 = mk ik >
0 for some k.
Now we apply the same argument to the matrix R1 . If every term in
the expansion of the det R1 is equal to z ero, then the theorem is proved,
because this implies that the matrix R1 is the zero matrix. Otherwise, there
(1) (1) (1)
is a term r1 j1 r2 j2 · · · rn jn in the expansion of det R1 different from zero.
(2)
Let P2 = (pij ) be the permutation matrix defined as

(2) 1 if k = s
pkjs =
0 otherwise

(1) (1) (1)

Set α2 = min(r1 j1 , r2 j2 , . . . , rn jn ). Define the matrix

R2 = R1 − α2 P2 = M − α1 P1 − α2 P2 .

As before, the matrix R2 has more zero entries than the matrix R1 . If every
term in the expansion of the det R2 is zero, then the theorem is proved, because
this implies that the matrix R2 is the zero matrix. Otherwise, we continue the
same argument.
Sooner or later we must have the equation

Rs = M − α1 P1 − α2 P2 − · · · − αs Ps (2.2)

with every term in the expansion of Rs equal to zero. Then, we claim that
the matrix Rs is the z ero matrix. Otherwise, by Equation (2.2), the sum
 Graph Algorithms II 133

of the entries of each row and each column of the matrix Rs is 1 − α1 − α2 −

· · · − αs > 0. Set α = 1 − α1 − α2 − · · · − αs . By dividing every term of
the matrix Rs we obtain a bi-stochastic matrix, that is, the matrix (1/α)Rs
is bi-stochastic. According to Lemma 2.3, the matrix (1/α)Rs has a term
in the expansion of det(1/α)Rs different from zero. But by a property of
determinants, det(1/α)Rs = (1/αn ) det Rs . This implies that there is a
term in the expansion of det Rs different from zero, a contradiction to our
assumption. Hence the matrix Rs is identically equal to zero. Therefore, by
Equation (2.2)

M = α1 P1 + α2 P2 + · · · + αs Ps

and the sum of each row of M = the sum of each column of M = α1 +

α2 + · · · + αs = 1, since M is bi-stochastic. Thus, the proof is complete.

In the maximum matching algorithm of Figure 2.11, it is not indicated

how to find an augmenting path relative to the matching M if such a path
exists in the graph G. Now we shall do this for the special case of the bipartite
graphs.
To present the algorithm for finding an augmenting path relative to a given
matching M , we need the following definition.

Definition 2.2 (An M-alternating bfs tree). Consider a connected bipar-

tite graph G = (X1 , X2 ; E) and a matching M in the graph G with at least one
unsaturated vertex of the set X1 . Perform a level order search of the graph G
starting from an unsaturated vertex x1 ∈ X1 . Let T be a bfs tree obtained by
the level order search starting from the vertex x1 (see graph G of Figure 2.2).
Process the vertices in increasing order. Then, as we have already remarked
(see Example 2.1), the tree T , partitions the vertex set of the graph G as
L0 (x1 ) ∪ L1 (x1 ) ∪ · · · ∪ Lk (x1 ) where Li (x1 ) consists of all the vertices at dis-
tance i from the vertex x1 and the integer k represents the eccentricity of the
vertex x1 .
A full M -alternating bfs tree with the root vertex x1 is obtained by the
removal of all the edges of T \ M joining a vertex of L2i−1 (x1 ) and a ver-
tex of L2i (x1 ) for all i = 1, 2, . . .. In other words, by going down in the bfs
tree T , we remove all the edges joining an oddly subscripted level vertex to
an evenly subscripted level vertex not belonging to the matching M . In other
words, during the level order search, we add only the edges belonging to the
matching M from an odd level to an even level. A partial M -alternating tree
or simply an M -alternating tree is a subtree of the full alternating tree having
the same root.
134 Discrete Mathematics

FIGURE 2.17: A bipartite graph G and a partial M -alternating tree.

The following example clarifies the definition.

Example 2.14: An M -augmenting tree

Consider the bipartite graph G of Figure 2.17 with bipartition X1 =
{ 1, 3, 8, 5, 12, 10 } and X2 = { 2, 6, 7, 4, 9, 13, 11 }.
Consider the matching M = { 23, 65, 78, (4, 10) }. Then, the tree T
of Figure 2.17 is a partial M -alternating bfs tree with the root vertex 1.
Note that between the level 1 and level 2 of the tree T , only the edges
belonging to the matching M are added. A full alternating tree T1 is
also shown in Figure 2.17. Note that between the level 3 and level 4,
only the edge (4, 10) belonging to the matching M is added. The full
tree T1 cannot be grown further. The tree T is a subtree of T1 having
the same root as T .

Algorithm for finding an M -augmenting path in a bipartite graph:

Input: A bipartite graph G = (X1 , Y1 ; E) and an arbitrary matching M .
Output: An M -augmenting path, if such a path exists. “No,” if there is no
M -augmenting path.
Algorithm: see Figure 2.18.
 Graph Algorithms II 135

FIGURE 2.18: Informal algorithm for finding an M -augmenting path.

The complexity of the maximum matching algorithm in a bipartite

graph:
For finding an M -augmenting path, we perform a bfs from an unsaturated
vertex of a matching M . The complexity of the bfs is O(max(n, m)) where
n is the number of vertices and m is the number of edges. Each M augmenting
path leads us to an improved matching M  where |M  | = |M | + 1. Since the
cardinality of a maximum matching is at most n/2, the bfs will be called at
most n/2 times. Hence, the complexity of the maximum matching algorithm
is O((n/2) max(n, m)) = O(n max(n, m)).

What can be done once, may be repeated many times.

Dijkstra

Now we give an algorithm to find a succinct certificate (see Definition 2.1)

in a bipartite graph G = (X1 , Y1 ; E), if it exists. The variable array L (whose
indices are the vertices of the bipartite graph G) records the level of each ver-
tex during the bfs from unsaturated root vertex x. More precisely, L[v] = k if
the vertex v is in the kth level, that is, v ∈ Lk (x).

Algorithm to find a succinct certificate:

Input: A bipartite graph G = (X1 , Y1 ; E).
Output: A matching saturating all the vertices of the set X1 or else a succinct
certificate.
Algorithm: The algorithm is given in Figure 2.11.

Proof of the algorithm of Figure 2.19:

The proof that the algorithm of Figure 2.19 gives either a matching saturating
all the vertices of X1 or else a succinct certificate is the same as the proof of
the necessary part of the König-Hall Theorem 2.3 and the properties (see
Property 2.1 of the level sets induced by the bfs).
136 Discrete Mathematics

FIGURE 2.19: Algorithm to find a matching that saturates all the vertices
of the set X1 or else a succinct certificate.

The complexity of the algorithm of Figure 2.19: The complexity is the

same as that of the maximum matching algorithm which is O(n max(n, m)).

As an application of the maximum matching algorithm, we shall study in

the following subsection an algorithm to find a weighted perfect matching in
a weighted complete bipartite graph.

2.4.1 Personnel assignment problem or weighted matching

in a bipartite graph
Input: A complete bipartite simple graph G = (X1 , Y1 ; E) with |X1 | =
|Y1 | = n. Let us recall that such a bipartite graph is denoted by Kn,n . The
vertices of the set X1 denote k technicians x1 , x2 , . . . , xk and the vertices of
the set X2 stand for n jobs y1 , y2 , . . . , yk . Each edge xi yj for i, j = 1, 2, . . . , n
of the graph Kn,n is assigned a non-negative number w(xi yj ). This number
w(xi yj ) can be considered as the efficiency (measured in terms of the profit
to the company) of the technician xi on the job yj .
The weighted complete bipartite Kn,n is represented as an n × n matrix
W = (wij ) where the rows of the matrix represent the n technicians
t1 , t2 , . . . , tn and the columns represent the n jobs y1 , y2 , . . . , yn with wij =
w(xi yj ).
Output: A maximum weighted perfect matching in the graph G, where the
weight of a matching M is the sum of the weights of its edges.
 Graph Algorithms II 137

Example 2.15: Personnel assignment problem

Consider the weighted complete bipartite graph represented by a
4 × 4 matrix W where
y y2 y3 y4
⎛ 1 ⎞
x1 4 3 2 2
x ⎜2 3 5 1⎟
W = 2⎜ ⎟
x3 ⎝ 3 0 2 0 ⎠
x4 3 4 0 3

For example, the efficiency of the technician x2 on the job y3 is

w23 which is equal to 5. A maximum weighted matching is M =
{ x1 y4 , x2 y3 , x3 y1 , x4 y2 } and its weight w(M ) = w14 + w23 + w31 +
w42 = 14. This maximum matching M assigns the technician x1 to the
job y1 and x2 to the job y3 and so on.

Algorithm (Brute-force): Generate all possible perfect matchings of the

complete bipartite graph Kn,n and choose one of them for which the weight
is a maximum.
The most serious objection to this case-by-case method is that it works far too
rarely. The reason is that there are n! possible perfect matchings possible in
the complete bipartite graph Kn,n and n! is an exponential function because
 n
n+1
nn/2 ≤ n! ≤ .
2
For example, 10! > 3.5 million.

Example 2.16: Brute-force weighted matching algorithm

Let us execute the brute-force weighted matching algorithm on the
complete weighted bipartite graph K3,3 where the weights of the edges
are given by the 3 × 3 matrix
y1 y2 y3
⎛ ⎞
x1 2 4 3
W = x2 ⎝ 1 2 4 ⎠.
x3 3 4 5
The following Table 2.4 gives all the possible 3! = 6 perfect matchings
and their corresponding weights. The last row 6 of the table gives us a
maximum weighted matching whose weight is 11.

We are now ready to present the Kuhn-Munkres algorithm for the weighted
matching problem in bipartite graph.
138 Discrete Mathematics

TABLE 2.4: Brute-force weighted matching algorithm

Number Matchings Weights

1 { x1 y1 , x2 y2 , x3 y3 } 9
2 { x1 y1 , x2 y3 , x3 y2 } 10
3 { x1 y3 , x2 y2 , x3 y1 } 8
4 { x1 y2 , x2 y1 , x3 y3 } 10
5 { x1 y3 , x2 y1 , x3 y2 } 8
6 { x1 y2 , x2 y3 , x3 y1 } 11

Description of the Kuhn-Munkres algorithm:

Definition 2.3 (Good labeling of the vertices of K n,n ). Consider a

weighted bipartite graph Kn,n and the weight matrix W = wij of the edges.
We assign to each vertex v of the graph a non-negative number l(v), called the
label of the vertex v such that the weight of an edge is at most the sum of the
labels of its end vertices. Symbolically,
wij ≤ l(xi ) + l(yj )
for all i, j = 1, 2, . . . , n. Such a labeling is called a good labeling of the weighted
complete bipartite graph.

Example 2.17: Good labeling

Consider the weight matrix W of the complete bipartite graph K4,4
where

y1 y2 y3 y4
⎛ ⎞
x1 4 3 2 2
x2 ⎜
⎜ 2 3 5 1⎟ ⎟.
W =
x3 ⎝ 3 0 2 0⎠
x4 3 4 0 3
A good labeling function l is given as follows: l(x1 ) = 4 l(x2 ) = 5
l(x3 ) = 3 l(x4 ) = 4 and l(y1 ) = l(y2 ) = l(y3 ) = l(y4 ) = 0. In fact,
l(xi ) is defined as the maximum value of the entry in the ith row and
l(yj ) is defined as zero for all j.

More generally, we observe that for a given weighted matrix W = (wij ) of

a complete bipartite graph Kn,n , the function l defined as

max1≤j≤n wij if v = xi
l(v) =
0 if v = yj
is always a good labeling.
 Graph Algorithms II 139

There are n2 edges in the complete bipartite graph Kn,n . The following
lemma on which the algorithm is based says “many” of the n2 edges can
be eliminated from the graph in order to find a maximum weighted perfect
matching.

Lemma 2.4. Consider a weighted matrix W = wij , of a balanced complete

bipartite graph G = Kn,n = (X1 , Y1 ; E) and a good labeling function l of its
vertices. Let Gl be a spanning subgraph of G with vertex set X1 ∪ Y1 and the
edge set El where

El = { xi yj | l(xi ) + l(yj ) = wij }.

In other words, the spanning subgraph Gl consists of all edges of G such that
the weight of the edge is equal to the sum of the labels of its end vertices. If
the graph Gl contains a perfect matching M, then M is a maximum weighted
perfect matching of the graph G.
Proof. Let M ⊂ El be a perfect matching in the graph Gl . Then, we have to
show that M is a maximum weighted perfect matching of the graph G. Since
El ⊂ E, M is also a perfect matching of the graph G. Let M  be a perfect
matching of G. We have to show that w(M  ) ≤ w(M ).
Now,

w(M ) = wij
xi yj ∈M ⊂El

= (l(xi ) + l(yj ))
xi yj ∈M ⊂El

= l(v) (since M saturates all the vertices) (2.3)

v∈X1 ∪Y1

But then,

w(M  ) = wij (2.4)

xi yj ∈M  ⊂E

= (l(xi ) + l(yj ))
xi yj ∈M  ⊂E

≤ l(v) (since l is a good labeling)

v∈X1 ∪Y1

By Equations (2.3) and (2.4), we have w(M  ) ≤ w(M ).

Example 2.18: The graph Gl

Let us refer to Example 2.17. The graph Gl corresponding to the
good labeling l is drawn in Figure 2.20.
140 Discrete Mathematics

FIGURE 2.20: The graph Gl , where l is the good label.

Once the graph Gl is constructed, we apply the maximum matching algo-

rithm of Figure 2.11 to the spanning bipartite graph Gl of the given graph G.
If the graph Gl possesses a perfect matching M , then by the Lemma 2.4,
the matching M is a desired maximum weighted perfect matching in the given
graph G and the algorithm “STOPS.”
Otherwise, the graph Gl has a succinct certificate (see Definition 2.1) Sl .
More precisely, there is a subset Sl ∈ X1 such that the number of vertices
collectively adjacent to the vertices of the set Sl in the graph Gl is exactly
one less than the number of vertices in Sl . Symbolically,

|NGl (Sl )| = |Sl | − 1.

Note that this succinct certificate Sl can be obtained by an M -alternating

tree of the graph Gl where M is a maximum (but not perfect) matching of
the graph Gl . More precisely, if M is a maximum matching of Gl which is
not perfect, then we grow a full M -alternating tree A from an unsaturated
vertex x1 ∈ X1 as a root vertex (see Algorithm of Figure 2.18). Since M is a
maximum matching, by Berge’s Theorem 2.2, we obtain a tree A with only
one unsaturated vertex x1 . Then, the succinct certificate Sl is given by the
intersection X1 ∩ X(A) (which is also the union of all oddly subscripted level
sets of the tree A) where X(A) is the set of vertices of the tree A.
Now, with the help of the set Sl , we shall modify the good labeling l in
such a way that the tree A can still be grown further in the new graph Gl .
The following definition redefines the labeling l.

Definition 2.4 (New labeling from the old one). First of all, define the
number dl as
 
dl = min l(xi ) + l(yj ) − w(xi yj ) . (2.5)
xi ∈Sl ,yj ∈Y1 −Tl

Here, the set Tl denotes the neighbor set N (Sl ) in the graph Gl . Note that
by the definition, the number dl satisfies the inequality dl > 0. Now the new
 Graph Algorithms II 141

FIGURE 2.21: An M -alternating tree A of Gl .

labeling l is defined as follows:

⎧
⎨l(v) − dl if v ∈ Sl
l1 (v) := l(v) + dl if v ∈ Tl
⎩
l(v) otherwise

The above modification of the labeling is illustrated in the following

example.

Example 2.19: New labeling from the old one

Let us refer to the graph Gl of Example 2.18. A maximum match-
ing M in this graph Gl is M = { x2 y3 , x3 y1 , x4 y2 } which is not perfect
and the corresponding M -alternating tree A with the root x1 is drawn
in Figure 2.21:
The succinct certificate Sl is Sl = X1 ∩ X(A) = { x1 , x3 } where
X(A) is the set of vertices of the tree A. The set Tl is Tl = NGl (Sl ) =
{ y1 } and the set Tl = Y1 − T = { y2 , y3 , y4 }. Then, the number
dl = minxi ∈Sl ,yj ∈Tl (l(xi ) + l(yj ) − w(xi yj )). Then, the new labeling l1
is given as follows:
l1 (x1 ) := l(x1 ) − dl = 3 l( x3 ) := l(x3 ) − dl = 2 l1 (y1 ) := l(y1 ) +
dl = 1 and all other vertices v keep their old l-values, that is, l1 (v) := l(v)
for all other vertices. The graph Gl1 corresponding to the labeling l1 is
drawn in Figure 2.22.

Now we construct the new spanning subgraph Gl1 of the graph G. Because
of the definition of the labeling function l1 , in the new graph Gl1 all the edges
between the sets Sl and Tl in the old graph Gl will also belong to the new
graph Gl1 . In addition, at least one new edge (e.g., the edge x1 y2 in the graph
Gl1 of Example 2.19) between the set Sl and the set Y1 − T will appear in
the new graph Gl1 . This new edge enables us to grow the M -alternating tree
still further in the graph Gl1 . An edge of the old graph Gl having one end in
142 Discrete Mathematics

FIGURE 2.22: An M -alternating tree A1 of Gl1 .

X1 − Sl and the other end in the set Tl may disappear in the new graph Gl1 .
All other edges between the sets X1 − Sl and Y1 − Tl of the graph Gl will be
conserved in the new graph Gl1 .
If the new graph Gl1 contains a perfect matching M , then by virtue
of Lemma 2.4, M is a maximum weighted perfect matching in the original
graph G and the algorithm STOPS.
If not, we find a maximum matching (but not perfect) and the corre-
sponding succinct certificate in the graph Gl1 and we continue the algorithm
as before. This continuation is explained in the example that follows:

Example 2.20: Continuation of Example 2.19

Let us refer to graph Gl1 of Example 2.19. A maximum matching
M = { x2 y3 , x3 y1 , x4 y2 } in the graph Gl1 which is the same as the
maximum matching in the graph Gl .
The succinct certificate in the graph Gl1 is Sl1 = X1 ∩X(A1 ) = X1 =
{ x1 , x2 , x3 , x4 } where X(A1 ) is the vertex set of the alternating tree A1
relative to the matching M and Tl1 = { y1 , y2 , y3 } which is the set of
vertices collectively adjacent to the vertices of Sl1 in the graph Gl1 . As
before, we calculate the number dl1 whose value is 1. We calculate the
new labeling function l2 where l2 (x1 ) = 3 − 1 = 2, l2 (x2 ) = 5 − 1 = 4,
l2 (x3 ) = 2 − 1 = 1, l2 (x4 ) = 4 − 1 = 3, l2 (y1 ) = 1 + 1 = 2, l2 (y2 ) =
0 + 1 = 1, l2 (y3 ) = 0 + 1 = 1 and l2 (y4 ) = l1 (y4 ) = 0.
Then, we construct the new graph Gl2 (see Figure 2.23) and apply
the maximum matching Algorithm 2.11 to the graph Gl2 .
This graph possesses a perfect matching M = { x1 y2 , x2 y3 , x3 y1 , x4 y4 }
and by Lemma 2.4, M is also a maximum weighted perfect matching in
the original graph G and the maximum weight is w12 +w23 +w31 +w44 =
3 + 5 + 3 + 3 = 14. Hence the algorithm STOPS.
 Graph Algorithms II 143

FIGURE 2.23: The graph Gl2 .

We now write the Kuhn-Munkres algorithm in pseudo-code: To implement

this algorithm in a programming language, we may use the following data
structures:
The initial weighted bipartite graph G = Kn,n = (X1 , Y1 ; E) with X1 =
{ x1 , x2 , . . . , xn }, Y1 = { y1 , y2 , . . . , yn } is represented by an n × n matrix
W = (wij ) where wij is the weight of the edge xi yj . We shall use an array l of
integer for the labeling function. The indices of the array are the vertices of
the graph G. The graph Gl again is represented by an n × n matrix B = (bij )
where 
1 if l[xi ] + l[yj ] = wij
bij =
0 otherwise
The matching M is represented by an n × n matrix M = mij where

1 if the edge xi yj is in the matching M
mij =
0 otherwise

FIGURE 2.24: Kuhn-Munkres algorithm.

144 Discrete Mathematics

A subset is usually represented by its characteristic function, if the universal

set is known in advance. Hence the succinct certificate Sl may be represented
by a characteristic vector. More precisely, the set Sl is an array whose indices
are the vertices of the set X1 where

1 if xi ∈ Sl
Sl [xi ] =
0 otherwise
In a similar manner, we can represent Tl , the set of neighbors of the set Sl in
the graph Gl by a characteristic vector whose indices are the vertices of the
set Y1 (Figure 2.24).

2.5 Exercises
1. Write a complete program in C for the breadth-first search of a simple
graph. (The graph is represented by its adjacency lists.)
2. Can we use the breadth-first search procedure to find the minimum cost
in an undirected connected graph, from a given source vertex, to all
other vertices, with weights associated with each edge of the graph? If
the answer is “yes,” write a pseudo-code to find the minimum costs of
all paths from the source. The weight of each edge is > 0.

3. Write a complete program in C for the breadth-first search of a simple

graph. (The graph is represented by its adjacency matrix.)

4. Describe an algorithm to find the diameter and the radius of an undi-

rected connected graph using the breadth-first search. The weight of
each edge is 1. Give the complexity of your algorithm.

5. Using the breadth-first search, write a program in C to test if a given

simple graph is bipartite.

6. Can a geodetic graph be a bipartite graph? Justify your answer.

7. Using the breadth-first search, write a program in C to test if a given

simple graph is geodetic.

8. Apply the maximum matching algorithm to the Petersen graph. Initially,

we start with any single edge as a matching. Is the matching obtained
at the end, a perfect matching?
9. Prove that the Petersen graph cannot be written as a union of three
mutually disjoint perfect matchings.

10. If M1 and M2 are two different perfect matchings of a graph, then

describe the spanning graph whose edge set is M1 ΔM2 where Δ is the
 Graph Algorithms II 145

symmetric difference of M1 and M2 . (Here, M1 and M2 are considered

as sets of edges.)

11. Prove that the Petersen graph is not a Hamiltonian graph; that is, it
does not contain an elementary cycle passing through each vertex exactly
once.

12. If G is a 3-regular simple graph on an even number of vertices con-

taining a Hamiltonian cycle, then prove that the edge set of G can be
decomposed into union of three mutually disjoint perfect matchings.

13. Apply the Kuhn-Munkres algorithm on the following matrix W corre-

sponding to the weighted complete bipartite graph K4,4 :

y1 y2 y3 y4
⎛ ⎞
x1 4 4 3 2
x2 ⎜
⎜ 3 5 5 2⎟ ⎟
W =
x3 ⎝ 3 2 2 4⎠
x4 3 4 5 3
14. Write the following bi-stochastic matrix A as a convex combination of
permutation matrices:
⎛ ⎞
1/4 1/2 0 1/4
⎜ 1/2 1/4 1/4 0 ⎟
A=⎜ ⎝ 0
⎟
1/4 1/2 1/4 ⎠
1/4 0 1/4 1/2
15. A diagonal of a real square matrix n×n is a set of n entries of the matrix
no two of which are in the same row or in the same column. The weight
of a diagonal is the sum of the elements of the diagonal. Find a diagonal
of maximum weight in the following 4 × 4 matrix A where
⎛ ⎞
5 3 2 2
⎜ 2 3 5 3 ⎟
A=⎜ ⎝ 3 4 2 3 ⎠
⎟

3 4 2 5
16. Using the König-Hall theorem for perfect matching in a bipartite graph
proved in this chapter, prove the following theorem due to the eminent
graph theorist: Tutte.
Let G = (X,E) be a simple graph. For a subset S ⊂ X, denote by
co (G − S), the number of connected components with odd number of
vertices of the induced subgraph X − S.
Tutte’s theorem: A graph G has a perfect matching if and only if
co (G − S) ≤ |S| for all subsets S ⊂ X
(For a proof, see [11].)
Chapter 3
Algebraic Structures I (Matrices,
Groups, Rings, and Fields)

3.1 Introduction
In this chapter, the properties of the fundamental algebraic structures,
namely, matrices, groups, rings, vector spaces, and fields are presented. In
addition, the properties of finite fields which are so basic to finite geometry,
coding theory, and cryptography are also discussed.

3.2 Matrices
A complex matrix A of type (m, n) or an m by n complex matrix is a
rectangular arrangement of mn complex numbers in m rows and n columns
in the form: ⎛ ⎞
a11 a12 . . . a1n
⎜ a21 a22 . . . a2n ⎟
A=⎜ ⎟
⎝. . . . . . . . . . . . . . .⎠ .
am1 am2 . . . amn

A is usually written in the shortened form A = (aij ), where 1 ≤ i ≤ m and

1 ≤ j ≤ n. Here, aij is the (i, j)-th entry of A, that is, the entry common to
the i-th row and j-th column of A. If m = n, A is a square matrix of order n.
In the latter case, the vector with the entries a11 , a22 , . . . , ann is known as the
principal diagonal or main diagonal of A. All the matrices that we consider in
this chapter are complex matrices.

3.3 Operations on Matrices: Addition, Scalar

Multiplication, and Multiplication of Matrices
If A = (aij ) and B = (bij ) are two m by n matrices, then the sum A + B
is the m by n matrix (aij + bij ), and for a scalar (that is, a complex number)
α, αA = (αaij ). Further, if A = (aij ) is an m by n matrix, and B = (bij ) is

147
148 Discrete Mathematics

an n by p matrix, then the product AB is defined to be the m by p matrix

(cij ), where,

cij = ai1 b1j + ai2 b2j + · · · + ain bnj

= the scalar product of the i-th row vector Ri of A
and the j-th column vector Cj of B.

Thus, cij = Ri · Cj . Both Ri and Cj are vectors of length n. It is easy to see

that the matrix product satisfies both the distributive laws and the associative
law, namely, for matrices A, B and C,

A(B + C) = AB + AC,
(A + B)C = AC + BC, and
(AB)C = A(BC)

whenever these sums and products are defined.

3.3.1 Block multiplication of matrices

Let A be an m by n matrix, and B an n by p matrix so that the matrix
product AB is defined, and it is an m by p matrix. Now partition the rows of
A into m1 , m2 , . . . , mk rows in order where m1 + m2 · · · + mk = m. Further,
partition the columns of A into c1 + c2 + · · · + cl = n columns in order. We
partition the rows and columns of B suitably, so that the product AB can be
obtained by multiplying the partitioned matrices.
We explain it by means of an example. Let A = (aij ) be a 5×4 matrix, and
B = (bij ) be a 4 × 3 matrix so that AB is a 5 × 3 matrix. Partition the rows
of A into two parts, the first part containing the first three rows of A and the
second part containing the remaining two rows of A, keeping the order of the
rows unchanged all the time. Now partition the columns of A into two parts
with the first part having the first two columns in order, while the second part
containing the remaining two columns of A in order. Thus, we can write:

A11 A12
A= ,
A21 A22

where A11 is 3 × 2, A12 is 3 × 2, A21 is 2 × 2 and A22 is 2 × 2.

We now partition B as

B11 B12
B=
B21 B22

where B11 and B21 have two rows each. Hence B12 and B22 have also two
rows each. This forces B11 and B21 to have two columns each while B12 and
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 149

B22 have each one column. This gives

  
A11 A12 B11 B12 A11 B11 + A12 B21 A11 B12 + A12 B22
AB = = ,
A21 A22 B21 B22 A21 B11 + A22 B21 A21 B12 + A22 B22

and we find that the matrix on the right, on simplification, does indeed yield
the product AB.
Note that it is not enough if we simply partition the matrices A and B.
What is important is that the partitioned matrices should be conformable
for multiplication. This means, in the above example, that the 8 products
A11 B11 , A12 B21 , . . . A22 B22 are all defined.
The general case is similar.

3.3.2 Transpose of a matrix

If A = (aij ) is an m by n matrix, then the n by m matrix (bij ), where
bij = aji is called the transpose of A. It is denoted by At . Thus, At is obtained
from A by interchanging the row and column vectors of A. For instance, if
⎛ ⎞ ⎛ ⎞
1 2 3 1 4 7
A = ⎝4 5 6⎠ , then At = ⎝2 5 8⎠ .
7 8 9 3 6 9

It is easy to check that

i. (At )t = A, and
ii. (AB)t = B t At , whenever the product AB is defined

(Note: If A is an m by p matrix, and B is a p by n matrix, then (AB)t is an

n by m matrix, and B t At is also an n by m matrix.)

3.3.3 Inverse of a matrix

Let A = (aij ) be an n by n matrix, that is, a matrix of order n. Let Bij
be the determinant minor of aij in A. Then, Bij is the determinant of order
n − 1 got from A by deleting the i-th row and j-th column of A. Let Aij be
the cofactor of aij in det A(= determinant of A) defined by Aij = (−1)i+j Bij .
Then, the matrix (Aij )t of order n is called the adjoint (or adjugate) of A,
and denoted by adj A.

Theorem 3.1. For any square matrix A of order n,

A(adj A) = (adj A)A = (det A)In ,

where In is the identity matrix of order n. (In is the matrix of order n in

which the n entries of the principal diagonal are 1 and the remaining entries
are 0).
150 Discrete Mathematics

Proof. By a property of determinants, we have

ai1 Aj1 + ai2 Aj2 + · · · + ain Ajn = a1j A1i + a2j A2i + · · · + anj Ani = det A or 0

according to whether i = j or i = j. We note that in adj A, (Aj1 , . . . , Ajn )

is the j-th column vector, and (A1j , . . . , Anj ) is the j-th row vector. Hence,
actual multiplication yields
⎛ ⎞
det A 0 . . . 0
⎜ 0 det A . . . 0 ⎟
A(adj A) = (adj A)A = ⎜ ⎟
⎝. . . . . . . . . . . . . . . . . .⎠ = (det A)In .
0 0 . . . det A

Corollary 3.1. Let A be a non-singular matrix, that is, det A = 0. Set A−1 =
(1/det A)(adj A). Then, AA−1 = A−1 A = In , where n is the order of A.

The matrix A−1 , as defined in Corollary 3.1, is called the inverse of the
(non-singular) matrix A. If A, B are square matrices of the same order with
AB = I, then B = A−1 and A = B −1 . These are easily seen by premultiplying
the equation AB = I by A−1 and postmultiplying it by B −1 . Note that A−1
and B −1 exist since taking determinants of both sides of AB = I, we get

det(AB) = det A · det B = det I = 1,

and hence det A = 0 as well as det B = 0.

3.3.4 Symmetric and skew-symmetric matrices

A matrix A is said to be symmetric if A = At . A is skew-symmetric if
A = −At . Hence if A = (aij ), then A is symmetric if aij = aji for all i and
j; it is skew-symmetric if aij = −aji for all i and j. Clearly, symmetric and
skew-symmetric matrices are square matrices. If A = (aij ) is skew-symmetric,
then aii = −aii , and hence aii = 0 for each i. Thus, in a skew-symmetric
matrix, all the entries of the principal diagonal are zero.

3.3.5 Hermitian and skew-Hermitian matrices

Let H = (hij ) denote a complex matrix. The conjugate H of H is the
matrix (hij ). The conjugate-transpose of H is the matrix H  = (H)t = (H t ) =
(hji ) = (hij ). H is Hermitian if and only if (iff) H  = H; H is skew-Hermitian


1 2+3i
iff H  = −H. For example, the matrix H = 2−3i √5 is Hermitian, while
 −i 1+2i
the matrix S = −1+2i 5i is skew-Hermitian. Note that the diagonal entries
of a skew-Hermitian
√ matrix are all of the form ir, where r is a real number
(and i = −1).
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 151

3.3.6 Orthogonal and unitary matrices

A real matrix (that is, a matrix whose entries are real numbers) P of
order n is called orthogonal if P P t = In . If P P t = In , then P t = P −1 . Thus,
the inverse of an orthogonal matrix is its transpose. Further, as P −1 P = In ,
we also have P t P = In . If R1 , . . . , Rn are the row vectors of P , the relation
P P t = In implies that Ri · Rj = δij , where δij = 1 if i = j, and δij = 0
if i = j. A similar statement
 also applies to the column vectors of P . As an
example, the matrix −cos α sin α
sin α cos α is orthogonal where α is a real number.
Indeed, if (x, y) are the Cartesian coordinates of a point P referred to a pair
of rectangular axes, and if (x , y  ) are the coordinates of the same point P with
reference to a new set of rectangular axes got by rotating the original axes
through an angle α about the origin (in the counter-clockwise direction), then
x = x cos α + y sin α
y  = −x sin α + y cos α,
   
x cos α sin α x
that is, =
y − sin α cos α y
so that rotation is effected by an orthogonal matrix.
Again, if (l1 , m1 , n1 ), (l2 , m2 , n2 ) and (l3 , m3 , n3 ) are the direction cosines
of three mutually orthogonal directions referred toan orthogonal coordinate
l1 m1 n1
system in the Euclidean 3-space, then the matrix l2 m2 n2 is orthogonal.
l3 m3 n3
In passing, we mention that rotation in higher-dimensional Euclidean spaces
is defined by means of an orthogonal matrix.
A complex matrix U of order n is called unitary if U U  = In (Recall
that U  is the conjugate transpose of U ). Again, this means that U  U =
In . Also a real unitary matrix is simply an orthogonal matrix. The unit
matrix is both orthogonal
as well as unitary. For example, the matrix U =
(1/5) −1+2i −4−2i
2−4i −2−i is unitary.

3.3.7 Exercises
  1+2k −4k
1. If A = 31 −4
−1 , prove by induction that A =
k
k 1−2k for any posi-
tive integer k.


cos(nα) sin(nα)
2. If M = −cos α sin α
sin α cos α , prove that M n
= − sin(nα) cos(nα) , n ∈ N .

1 −10

3. Compute the transpose, adjoint and inverse of the matrix 0 1 −1 .
1 0 1
 1 3

4. If A = −2 2 , show that A2 − 3A + 8I = 0. Hence compute A−1 .

5. Give two matrices A and B of order 2, so that

i. AB = BA
ii. (AB)t = AB
152 Discrete Mathematics

6. Prove: (i) (AB)t = B t At ; (ii) If A and B are non-singular, (AB)−1 =

B −1 A−1 .
7. Prove that the product of two symmetric matrices is symmetric iff the
two matrices commute.
8. Prove: (i) (iA) = −iA ; (ii) H is Hermitian iff iH is skew-Hermitian.
9. Show that every real matrix is the unique sum of a symmetric matrix
and a skew-symmetric matrix.

10. Show that every complex square matrix is the unique sum of a Hermitian
and a skew-Hermitian matrix.

3.4 Groups
Groups constitute an important basic algebraic structure that occurs very
naturally not only in mathematics, but also in many other fields such as
physics and chemistry. In this section, we present the basic properties of
groups. In particular, we discuss Abelian and non-Abelian groups, cyclic
groups, permutation groups and homomorphisms and isomorphisms of groups.
We establish Lagrange’s theorem for finite groups and the basic isomorphism
theorem for groups.

3.4.1 Abelian and non-Abelian groups

Definition 3.1. A binary operation on a non-empty set S is a map : S ×S →
S, that is, for every ordered pair (a, b) of elements of S, there is associated
a unique element a · b of S. A binary system is a pair (S, ·), where S is a
non-empty set and · is a binary operation on S. The binary system (S, ·) is
associative if · is an associative operation on S, that is, for all a, b, c in S,
(a · b) · c = a · (b · c)

Definition 3.2. A semigroup is an associative binary system. An element e

of a binary system (S, ·) is an identity element of S if a · e = e · a = a for all
a ∈ S.

We use the following standard notations:

Z = the set of integers (positive integers, negative integers and zero),
Z+ = the set of positive integers,
N = the set of natural numbers {1, 2, 3, . . .} = Z+ ,
Q = the set of rational numbers,
Q+ = the set of positive rational numbers,
Q∗ = the set of non-zero rational numbers,
R = the set of real numbers,
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 153

R∗ = the set of non-zero real numbers,

C = the set of complex numbers, and
C∗ = the set of non-zero complex numbers.

Examples
1. (N, ·) is a semigroup, where · denotes the usual multiplication.

2. The operation subtraction is not a binary operation on N (for example,

3−5∈ / N).

3. (Z, −) is a binary system which is not a semigroup since the associative

law is not valid in (Z, −); for instance, 10 − (5 − 8) = (10 − 5) − 8.

We now give the definition of a group.

Definition 3.3. A group is a binary system (G, ·) such that the following
axioms are satisfied:

(G1 ): The operation · is associative on G, that is, for all a, b, c ∈ G, (a · b) · c =

a · (b · c).
(G2 ): (Existence of identity) There exists an element e ∈ G (called an identity
element of G with respect to the operation ·) such that a · e = e · a = a for
all a ∈ G.

(G3 ): (Existence of inverse) To each element a ∈ G, there exists an element

a−1 ∈ G (called an inverse of a with respect to the operation ·) such that
a · a−1 = a−1 · a = e.

Before proceeding to examples of groups, we show that identity element e,

and inverse element a−1 of a, given in Definition 3.3 are unique.
Suppose G has two identities e and f with respect to the operation ·. Then,

e=e·f (as f is an identity of (G, ·))

=f (as e is an identity of (G, ·)).

Next, let b and c be two inverses of a in (G, ·). Then,

b=b·e
= b · (a · c) (as c is an inverse of a)
= (b · a) · c by the associativity of · in G
= e · c (as b is an inverse of a)
= c.

Thus, henceforth, we can talk of “The identity element e” of the group (G, ·),
and “The inverse element a−1 of a” in (G, ·).
154 Discrete Mathematics

If a ∈ G, then a · a ∈ G; also, a · a · · · (n times) ∈ G. We denote a · a · · · (n

times) by an . Further, if a, b ∈ G, a · b ∈ G, and (a · b)−1 = b−1 · a−1 . (Check
that (a · b)(a · b)−1 = (a · b)−1 (a · b) = e). More generally, if a1 , a2 , . . . , an ∈ G,
−1 −1
then (a1 · a2 · · · an )−1 = a−1
n · an−1 · · · a1 , and hence (a )
n −1
= (a−1 )n =
−n m+n m n
(written as) a . Then, the relation a = a · a holds for all integers m
and n, with a0 = e. In what follows, we drop the group operation · in (G, ·),
and simply write group G, unless the operation is explicitly needed.

Lemma 3.1. In a group, both the cancellation laws are valid, that is, if a, b, c
are elements of a group G with ab = ac, then b = c (left cancellation law), and
if ba = ca, then b = c (right cancellation law).

Proof. If ab = ac, premultiplication by a−1 gives a−1 (ab) = a−1 (ac). So by

the associative law in G, (a−1 a)b = (a−1 a)c, and hence eb = ec. This implies
that b = c. The other cancellation is proved similarly.

Definition 3.4. The order of a group G is the cardinality of G. The order

of an element a of a group G is the least positive integer n such that an = e,
the identity element of G. If no such n exists, the order of a is taken to be
infinity.

Definition 3.5. Abelian Group [After the Norwegian mathematician Abel]

A group G is called Abelian if the group operation of G is commutative, that
is, ab = ba for all a, b ∈ G.

A group G is non-Abelian if it is not Abelian, that is, there exists a pair

of elements x, y in G with xy = yx.

3.4.2 Examples of Abelian groups

1. (Z, +) is an Abelian group, that is, the set Z of integers is an Abelian
group under the usual addition operation. The identity element of this
group is O, and the inverse of a is −a. (Z, +) is often referred to as
the additive group of integers. Similarly, (Q, +), (R, +), (C, +) are all
additive Abelian groups.
2. The sets Q∗ , R∗ and C∗ are groups under the usual multiplication oper-
ation.
3. Let G = C[0, 1], the set of complex-valued continuous functions defined
on [0, 1]. G is an Abelian group under addition. Here, if f, g ∈ C[0, 1],
then f + g is defined by (f + g)(x) = f (x) + g(x), x ∈ [0, 1]. The zero
function O is the identity element of the group while the inverse of f
is −f .

4. For any positive integer n, let Zn = {0, 1, . . . , n − 1}. Define addition +

in Zn as “congruent modulo n” addition, that is, if a, b ∈ Zn , then
a + b = c, where c ∈ Zn and a + b ≡ c (mod n). Then, (Zn , +) is an
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 155

Abelian group. For instance, if n = 5, then in Z5 , 2 + 2 = 4, 2 + 3 = 0,

3 + 3 = 1, etc.

5. Let G = {rα : rα = rotation of the plane about the origin through an

angle α in the counter-clockwise sense}. Then, if we set rβ · rα = rα+β
(that is, rotation through α followed by rotation through β = rotation
through α+β), then (G, ·) is a group. The identity element of (G, ·) is r0 ,
while (rα )−1 = r−α , the rotation of the plane about the origin through
an angle α in the clockwise sense.

3.4.3 Examples of non-Abelian groups

1. Let G = GL(n, R), the set of all n by n non-singular matrices with real
entries. Then, G is an infinite non-Abelian group under multiplication.
2. Let G = SL(n, Z) be the set of matrices of order n with integer entries
having determinant 1. G is again an infinite non-Abelian multiplica-
tive group. Note that if A ∈ SL(n, Z), then A−1 = (1/det A)(adj A) ∈
SL(n, Z) since det A = 1 and all the cofactors of the entries of A are
integers (see Section 3.3.3).

3. Let S4 denote the set of all 1–1 maps f : N4 → N4 , where N4 =

{1, 2, 3, 4}. If · denotes composition of maps, then (S4 , ·) is a non-Abelian
group of order 4! = 24. (See Section 3.8 for more about such groups.)
For instance, let 
1 2 3 4
f= .
4 1 2 3
Here, the parantheses notation signifies the fact that the image under f
of a number in the top row is the corresponding number in the bottom
row. For instance, f (1) = 4, f (2) = 1 and so on. Let
 
1 2 3 4 1 2 3 4
g= Then, g · f = .
3 1 2 4 4 3 1 2
 
Note that (g · f )(1) = g f (1) = g(4) = 4, while (f · g)(1) = f g(1) =
f (3) = 2, and hence f · g = g · f . In other words, S4 is a non-Abelian
group. The identity element of S4 is the map
 
1 2 3 4 −1 1 2 3 4
I= and f = .
1 2 3 4 2 3 4 1

3.4.4 Group tables

The structure of a finite group G can be completely specified by means
of its group table (sometimes called multiplication table). This is formed by
listing the elements of G in some order as {g1 , . . . , gn }, and forming an n by
n double array (gij ), where gij = gi gj , 1 ≤ i ≤ j ≤ n. It is customary to take
g1 = e, the identity element of G.
156 Discrete Mathematics

TABLE 3.1: Group table of Klein’s

4-group
e a b c
e e a b c
a a e c b
b b c e a
c c b a e

Examples Continued
1. [Klein’s 4-group K4 ] This is a group of order 4. If its elements are e, a, b, c,
the group table of K4 is given by Table 3.1.

It is observed from the table that

ab = ba = c, and a(ba)b = a(ab)b

This gives

c2 = (ab)(ab) = a(ba)b = a(ab)b = a2 b2 = ee = e.

Thus, every element of K4 other than e is of order 2.

3.5 A Group of Congruent Transformations (Also

called Symmetries)
We now look at the congruent transformations of an equilateral triangle
ABC. Assume without loss of generality that the side BC of the triangle is
horizontal so that A is in the vertical through the middle point D of BC.
Let us denote by r, the rotation of the triangle about its center through
an angle of 120◦ in the counter-clockwise sense and let f denote the flipping
of the triangle about the vertical through the middle point of the base. f
interchanges the base vertices and leaves all the points of the vertical through
the third vertex unchanged. Then, f r denotes the transformation r followed
by f and so on.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 157

TABLE 3.2: Group table of the dihedral group D3

3
r =e r r2 f rf r2 f
3
r =e e r r2 f rf r2 f
r r r2 e rf r2 f f
r2 r2 e r r2 f f rf
f f fr f r2 e r2 r
rf rf f r2 f r e r2
r2 f r2 f rf f r2 r e

Thus, f r leaves B fixed and flips A and C in ABC. There are six con-
gruent transformations of an equilateral triangle and they form a group as per
group Table 3.2
For instance, r2 f r and rf are obtained as follows:

Thus, r2 f r = rf , and similarly the other products can be verified. The result-
ing group is known as the dihedral group D3 . It is of order 2 · 3 = 6.

3.6 Another Group of Congruent Transformations

Let D4 denote the transformations that leave a square invariant. If r
denotes a rotation of 90◦ about the center of the square in the counter-
clockwise sense, and f denotes the flipping of the square about one of its
diagonals, then the defining relations for D4 are given by:

r4 = e = f 2 = (rf )2 .

D4 is the dihedral group of order 2 × 4 = 8. Both D3 and D4 are non-Abelian

groups. The dihedral group Dn of order 2n is defined in a similar fashion as
158 Discrete Mathematics

the group of congruent transformations of a regular polygon of n sides. The

groups Dn , n ≥ 3, are all non-Abelian. Dn is of order 2n for each n.

3.7 Subgroups
Definition 3.6. A subset H of a group (G, ·) is a subgroup of (G, ·) if (H, ·)
is a group.

Definition 3.6 shows that the group operation of a subgroup H of G is the

same as that of G.

3.7.1 Examples of subgroups

1. Z is a subgroup of (Q, +).

2. 2Z is a subgroup of (Z, +). (Here, 2Z denotes the set of even integers).

3. Q∗ is a subgroup of (R∗ , ·). (Here, · denotes multiplication).

4. Let H be the subset of maps of S4 that fix 1, that is, H = {f ∈ S4 :

f (1) = 1}. Then, H is a subgroup of S4 ).
Note that the set N of natural numbers does not form a subgroup of
(Z, +).

3.7.2 Subgroup generated by a subset of a group

Definition 3.7. Let S be a non-empty subset of a group G. The subgroup
generated by S in G, denoted by
S, is the intersection of all subgroups of G
containing S.

Proposition 3.1. The intersection of any family of subgroups of G is a sub-

group of G.

Proof. Consider a family {Gα }α∈I of subgroups of G, and let H = ∩α∈I Gα .

If a, b ∈ H, then a, b ∈ Gα for each α ∈ I, and since Gα is a subgroup of
G, ab ∈ Gα for each α ∈ I. Therefore, ab ∈ H, and similarly a−1 ∈ H and
e ∈ H. The associative law holds in H, a subset of G, as it holds in G. Thus,
H is a subgroup of G.

Corollary 3.2. Let S be a non-empty subset of a group G. Then,

S is the
smallest subgroup of G containing S.

Proof. By definition,
S is contained in every subgroup H of G containing S.
Since
S is itself a subgroup of G containing S, it is the smallest subgroup of
G containing S.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 159

3.8 Cyclic Groups

Definition 3.8. Let G be a group and a, an element of G. Then, the subgroup
generated by a in G is
{a}, that is, the subgroup generated by the singleton
subset {a}. It is also denoted simply by
a.

By Corollary 3.2,
a is the smallest subgroup of G containing a. As a ∈

a, all the powers of an , n ∈ Z, also belong to

a. But then, as may be
checked easily, the set {an : n ∈ Z} of powers of a is already a subgroup of G.
Hence
a = {an : n ∈ Z}. Note that a0 = e, the identity element of G, and
a−n = (a−1 )n , the inverse of an . This makes am · an = am+n for all integers
m and n. The subgroup A =
a of G is called the subgroup generated by a,
and a is called a generator of A. 
Now since {an : n ∈ Z} = a−1 : n ∈ Z , a−1 is also a generator of

a. Suppose a is of finite order m in

a. This means that m is the
least positive integer with the property that am = e. Then, the elements
a1 = a, a2 , . . . , am−1 , am = e are all distinct. Moreover, for any integer n, by
Euclidean algorithm for integers, there are integers q and r such that
n = qm + r, 0 ≤ r < m.
Then,
an = a(qm+r) = (am )q ar = eq ar = ear = ar , 0 ≤ r < m,
and hence an ∈
a. Thus, in this case,
 

a = a, a2 , . . . , am−1 , am = e .
In the contrary case, there exists no positive integer m such that am = e.
Then, all the powers ar : r ∈ Z are distinct. If not, there exist integers r and
s, r = s, such that ar = as . Suppose r > s. Then, r − s > 0 and the equation
ar = as gives a−s ar = a−s as , and therefore ar−s = a0 = e, a contradiction.
In the first case, the cyclic group
a is of order m while in the latter case, it
is of infinite order.

3.8.1 Examples of cyclic groups

1. The additive group of integers (Z, +) is an infinite cyclic group. It is
generated by 1 as well as −1.

2. The group of n-th roots of unity, n ≥ 1. Let G be the set of n-th roots
of unity so that
 
2π 2π
G = ω, ω 2 , . . . , ω n = 1; ω = cos + i sin .
n n
Then, G is a cyclic group of order n generated by ω, that is, G =< ω >.
In fact ω k , 1 ≤ k ≤ n, also generates k iff (k, n) = 1. Hence, the number
of generators of G is φ(n), where φ is the Euler’s totient function.
160 Discrete Mathematics

If G =
a = {an : n ∈ Z}, then since for any two integers n and m,
a a = an+m = am an , G is Abelian. In other words, every cyclic group
n m

is Abelian. However, the converse is not true. K4 , the Klein’s 4-group (See
Table 3.1 of Section 3.4) is Abelian but not cyclic since K4 has no element of
order 4.

Theorem 3.2. Any subgroup of a cyclic group is cyclic.

Proof. Let G =
a be a cyclic group, and H, a subgroup of G. If H = {e},
then H is trivially cyclic. So assume that H = {e}. As the elements of G are
powers of a, an ∈ H for some non-zero integer n. Then, its inverse a−1 also
belongs to H, and of n and −n at least one of them is a positive integer. Let
s be the least positive integer such that as ∈ H. (recall that H = {e} as per
our assumption). We claim that H =
as , the cyclic subgroup of G generated
by as . To prove this, we have to show that each element of H is a power of as .
Let g be any element of H. As g ∈ G, g = am for some integer m. By division
algorithm,
m = qs + r, 0 ≤ r < s.

Hence ar = am−qs = am (as )−q ∈ H as am ∈ H and as ∈ H. Thus, ar ∈ H.

This however implies, by the choice of s, r = 0 (otherwise ar ∈ H with 0 < r <
s). Hence ar = a0 = e = am (as )−q , and therefore, g = am = (as )q , q ∈ Z.
Thus, every element of H is a power of as and so H ⊂
as . Now since as ∈ H,
all powers of as also ∈ H, and so
as  ⊂ H. Thus, H =
as  and therefore H
is cyclic.

Definition 3.9. Let S be any non-empty set. A permutation on S is a bijective

mapping from S to S.

Lemma 3.2. If σ1 and σ2 are permutations on S, then the map σ = σ1 σ2

defined on S by

σ(s) = σ1 σ2 (s) = σ1 (σ2 (s)) , s∈S

is also a permutation on S.

Proof. Indeed, we have, for s1 , s2 in S, σ(s1 ) = σ(s2 ) gives that σ1 (σ2 s1 ) =

σ1 (σ2 s2 ). This implies, as σ1 is 1 − 1, σ2 s1 = σ2 s2 . Again, as σ2 is 1 − 1, this
gives that s1 = s2 . Thus, σ is 1 − 1. For a similar reason, σ is onto.

Let B denote the set of all bijections on S. Then, it is easy to verify that
(B, ·), where · is the composition map, is a group. The identity element of this
group is the identity function e on S.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 161

The case when S is a finite set is of special significance. So let S =

{1, 2, . . . , n}. The set P of permutations of S forms a group under the compo-
sition operation. Any σ ∈ P can be conveniently represented as:

1 2 ... n
σ: .
σ(1) σ(2) . . . σ(n)

Then, σ −1 is just the permutation


σ(1) σ(2) . . . σ(n)
σ −1 :
1 2 ... n

What is the order of the group P? Clearly σ(1) has n choices, namely, any
one of 1, 2, . . . , n. Having chosen σ(1), σ(2) has n − 1 choices (as σ is 1 − 1,
σ(1) = σ(2)). For a similar reason, σ(3) has n−2 choices and so on, and finally
σ(n) has just one left out choice. Thus, the total number of permutations on
S is
n · (n − 1) · (n − 2) · · · 2 · 1 = n!.

In other words, the group P of permutations on a set of n elements is of order

n! It is denoted by Sn and is called the symmetric group of degree n. Any
subgroup of Sn is called a permutation group of degree n.

Example
Let S = {1, 2, 3, 4, 5}, and let σ and τ ∈ S5 be given by
 
1 2 3 4 5 1 2 3 4 5
σ= , τ= .
2 3 5 4 1 5 2 4 1 3

Then,
 
1 2 3 4 5 2 1 2 3 4 5
σ · τ = στ = , and σ = σ · σ = .
1 3 4 2 5 3 5 1 4 2

Definition 3.10. A cycle in Sn is a permutation σ ∈ Sn that can be repre-

sented in the form (a1 , a2 , . . . , ar ), where the ai , 1 ≤ i ≤ r, r ≤ n, are all in
S, and σ(ai ) = ai+1 , 1 ≤ i ≤ r − 1, and σ(ar ) = a1 , that is, each ai is mapped
cyclically to the next element (or) number ai+1 and σ fixes the remaining ai s.
For example, if

1 3 2 4
σ= ∈ S4 ,
3 2 1 4

then σ can be represented by (132). Here, σ leaves 4 fixed.

162 Discrete Mathematics

Now consider the permutation


1 2 3 4 5 6 7
p= .
3 4 2 1 6 5 7
Clearly, p is the product of the cycles
(1324)(56)(7) = (1324)(56).
Since 7 is left fixed by p, 7 is not written explicitly. In this way, every permu-
tation on n symbols is a product of disjoint cycles. The number of symbols in
a cycle is called the length of the cycle. For example, the cycle (1 3 2 4) is a
cycle of length 4. A cycle of length 2 is called a transposition. For example,
the permutation (1 2) is a transposition. It maps 1 to 2, and 2 to 1. Now
consider the product of transpositions t1 = (13), t2 = (12) and t3 = (14). We
have
   
1 3 1 2 1 4 1 2 3 4
t1 t2 t3 = (13)(12)(14) = = = (1423) .
3 1 2 1 4 1 4 3 1 2
To see this, note that
(t1 t2 t3 )(4) = (t1 t2 )(t3 (4)) = (t1 t2 )(1) = t1 (t2 (1)) = t1 (2) = 2, and so on.
In the same way, any cycle (a1 a2 . . . an ) = (a1 an )(a1 an−1 ) · · · (a1 a2 ) is a prod-
uct of transpositions. Since any permutation is a product of disjoint cycles
and any cycle is a product of transpositions, it is clear that any permutation
is a product of transpositions. Now in the expression of a cycle as a product of
transpositions, the number of transpositions need not be unique. For instance,
(12)(12) = identity permutation, and
(1324) = (14)(12)(13) = (12)(12)(14)(12)(13).
However, this number is always odd or always even.

Theorem 3.3. Let σ be any permutation on n symbols. Then, in whatever

way σ is expressed as a product of transpositions, the number of transpositions
is always odd or always even.
Proof. Assume that σ is a permutation on {1, 2, . . . , n}. Let the product
P =(a1 − a2 )(a1 − a3 ) · · · (a1 − an )
(a2 − a3 ) · · · (a2 − an )
.. ..
. .
(an−1 − an )
 
 1 1 ··· 1 

  a1 a2 ··· an 

= (ai − aj ) = det  a21 a22 ··· a2n  .
1≤i<j≤n  ··· ··· · · · · · · 
 n−1
a an−1 · · · an−1 
1 2 n
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 163

Any transposition (ai aj ) applied to the product P changes P to −P as this

amounts to the interchange of the i-th and j-th columns of the above determi-
nant. Now σ when applied to P has a definite effect, namely, either it changes
P to −P or leaves P unchanged. In case σ changes P to −P , σ must always
be a product of an odd number of transpositions; otherwise, it must be the
product of an even number of transpositions.

(The determinant (call it D) on the right is known as the Vander Monde

determinant. If we set, for i < j, ai = aj in D, then two columns become
identical, and hence D = 0. This means that (ai − aj ) is a factor of D. Further
the coefficient of a2 a23 . . . an−1
n on both sides is 1. Hence the last equality.)

Theorem 3.4. The even permutations in Sn form a subgroup An of Sn , called

the alternating group of degree n.

Proof. If σ1 and σ2 are even permutations in Sn , then each of them is a product

of an even number of transpositions and hence so is their product. Further if
−1 −1 −1
σ = t1 t2 . . . tr , where each ti is a transposition, then σ −1 = t−1
r tr−1 . . . t2 t1 .
Hence the inverse of an even permutation is even. Further the identity per-
mutation is even since for any transposition t, e = t ◦ t = t2 .

Definition 3.11. A permutation is odd or even according to whether it is

expressible as a product of an odd number or even number of transpositions.

Example 3.1


1 2 3 4 5 6 7 8 9
Let σ = .
4 5 1 2 3 7 9 8 6
Then, σ = (14253)(679)(8)
= (13)(15)(12)(14)(69)(67)
= a product of an even number of transpositions.

Hence σ is an even permutation.

3.9 Lagrange’s Theorem for Finite Groups

We now establish the most famous basic theorem on finite groups, namely,
Lagrange’s theorem. For this, we need the notion of left and right cosets of a
subgroup.
164 Discrete Mathematics

Definition 3.12. Let G be a group and H a subgroup of G. For a ∈ G, the

left coset aH of a in G is the subset {ah : h ∈ H} of G. The right coset Ha of
a is defined in an analogous manner.

Lemma 3.3. Any two left cosets of a subgroup H of a group G are equipotent
(that is, they have the same cardinality). Moreover, they are equipotent to H.

Proof. Let aH and bH be two left cosets of H in G. Consider the map

φ : aH → bH

defined by φ(ah) = bh, h ∈ H. φ is 1−1 since φ(ah1 ) = φ(ah2 ), for h1 , h2 ∈ H,

implies that bh1 = bh2 and therefore h1 = h2 . Clearly, φ is onto. Thus, φ is
a bijection of aH onto bH. In other words, aH and bH are equipotent. Since
eH = H, H itself is a left coset of H and so all left cosets of H are equipotent
to H. (Recall that two sets are equipotent if there is a bijection between
them.)

Lemma 3.4. The left coset aH is equal to H iff a ∈ H.

Proof. If a ∈ H, then aH = {ah : h ∈ H} ⊂ H as ah ∈ H. Further if b ∈ H,

then a−1 b ∈ H, and so a(a−1 b) ∈ aH. But a(a−1 b) = b. Hence b ∈ aH, and
therefore aH = H. (In particular if H is a group, then multiplication of the
elements of H by any element a ∈ H just gives a permutation of H.)
Conversely, if aH = H, then a = ae ∈ aH = H, as e ∈ H.

Example 3.2
It is not necessary that aH = Ha for all a ∈ G. For example, consider
S3 , the symmetric group of degree 3. The 3! = 6 permutations of S3 are
given by
⎧      
⎪
⎪ 1 2 3 1 2 3 1 2 3
⎪
⎪
⎨e = 1 2 3 , 1 3 2
= (23),
3 2 1
= (13)
S3 =      
⎪ 1 2 3
⎪ 1 2 3 1 2 3
⎪
⎪
⎩ 2 1 3 = (12), 2 3 1
= (123),
3 1 2
= (132).

Let H be the subgroup {e, (12)}. For a = (123), we have

aH = {(123)e = (123), (123)(12) = (13)} , and

Ha = {e(123) = (123), (12)(123) = (23)} ,

so that aH = Ha.

Proposition 3.2. The left cosets aH and bH are equal iff a−1 b ∈ H.

 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 165

Proof. aH = bH ⇔ a−1 (aH) = a−1 (bH) ⇔ H = (a−1 b)H ⇔ a−1 b ∈ H by

Lemma 3.4.
Lemma 3.5. Any two left cosets of the same subgroup of a group are either
identical or disjoint.

Proof. Suppose aH and bH are two left cosets of the subgroup H of a group G,
where a, b ∈ G. If aH and bH are disjoint, there is nothing to prove. Otherwise,
aH ∩ bH = φ, and therefore, there exist h1 , h2 ∈ H with ah1 = bh2 . This
however means that a−1 b = h1 h−1
2 ∈ H. So by Proposition 3.2, aH = bH.

Example 3.3
For the subgroup H of Example 3.2, we have seen that (123)H =
{(123), (13)}. Now (12)H = {(12)e, (12)(12)} = {(12), e} = H, and
hence (123)H ∩ (12)H = φ. Also (23)H = {(23)e, (23)(12)} = {(23) ,
(132)}, and (13)H = (13){e, (12)} = {(13), (123)} = (123)H. Note that
(13)−1 (123) = (13)(123) = (12) ∈ H (refer to Proposition 3.2).

Theorem 3.5. Lagrange’s Theorem [After the French mathematician

J. L. Lagrange] The order of any subgroup of a finite group G divides the
order of G.

Proof. Let H be a subgroup of the finite group G. We want to show that

o(H)|o(G). We show this by proving that the left cosets of H in G form a
partition of G. First of all, if g is any element of G, g = ge ∈ gH. Hence
every element of G is in some left coset of H. Now by Lemma 3.5, the distinct
cosets of H are pairwise disjoint and hence form a partition of G. Again, by
Lemma 3.3, all the left cosets of H have the same cardinality as H, namely,
o(H). Thus, if there are l left cosets of H in G, we have
l · o(H) = o(G) (3.1)
Consequently, o(H) divides o(G).
Definition 3.13. Let H be a subgroup of a group G. Then, the number (may
be infinite) of left cosets of H in G is called the index of H in G and denoted
by iG (H)
If G is a finite group, then Equation (3.1) in the proof of Theorem 3.5
shows that
o(G) = o(H)iG (H)
Theorem 3.6 (An application of Lagrange’s theorem). If p is a prime,
and n any positive integer, then
n|φ(pn − 1),
where φ is Euler’s totient function. First we prove a lemma.
166 Discrete Mathematics

Lemma 3.6. If m ≥ 2 is a positive integer, and S, the set of all positive

integers less than m and prime to it, then S is a multiplicative group modulo m.

Proof. If (a, m) = 1 and (b, m) = 1, then (ab, m) = 1. For if p is a prime

factor of ab and m, then as p divides ab, p must divide either a or b, say, p|a.
Then, (a, m) ≥ p, a contradiction. Moreover, if ab (mod m) ≡ c, 1 ≤ c < m,
then (c, m) = 1. Thus, ab (mod m) = c ∈ S. Also as (1, m) = 1, 1 ∈ S.
Now for any a ∈ S, by Euclidean algorithm, there exists b ∈ N such that
ab ≡ 1 (mod m). Then, (b, m) = 1 (if not, there exists a prime p with p|b and
p|m, then p|1, a contradiction). Thus, a has an inverse b(mod m) in S. Thus,
S is a multiplicative group modulo m, and o(S) = φ(m).
Proof. (Proof
 of Theorem  3.6) We apply Lemma 3.6 by taking m = pn − 1.
2 n−1
Let H = 1, p, p , . . . , p . All the numbers in H are prime to m and hence
H ⊂ S (where S is as defined in Lemma 3.6). Further pj · pn−j = pn ≡ 1
(mod m). Hence every element of H has an inverse modulo m. Therefore (as
the other group axioms are trivially satisfied by H), H is a subgroup of order
n of S. By Lagrange’s theorem, o(H)|o(S), and so n|φ(pn − 1).
As another application of Lagrange’s theorem, we have the following result.

Theorem 3.7. Any group of prime order is cyclic.

Proof. Let G be a group of prime order p. Suppose H is a subgroup of G with

order q. Then, by Lagrange’s theorem q|p. Hence q = 1 or q = p, as p is prime.
This means that G = {e} or G = H and any element a(= e) generates G since
the cyclic subgroup
a of G is equal to G.

3.10 Homomorphisms and Isomorphisms of Groups

Consider the two groups:

G1 = the multiplicative group of the sixth root of unity


= ω 6 = 1, ω, ω 2 , . . . , ω 5 : ω = a primitive sixth root of unity}
and G2 = the additive group Z6 = {0, 1, 2, . . . , 5}.

G1 is a multiplicative group while G2 is an additive group. However,

structure-wise, they are just the same. By this we mean that if we can make
a suitable identification of the elements of the two groups, then they behave
in the same manner. If we make the correspondence

ω i ←→ i, 0 ≤ i ≤ 5,

we see that ω i ω j ←→ i + j as ω i ω j = ω i+j , when i + j is taken modulo 6. For

instance, in G1 , ω 3 ω 4 = ω 7 = ω 1 , while in G2 , 3 + 4 = 1 as 7 ≡ 1(mod
 6). The
order of ω in G1 = 6 = the (additive) order of 1 in G2 . G1 has 1, ω 2 , ω 4 as
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 167

a subgroup while G2 has {0, 2, 4} as a subgroup and so on. It is clear that we

can replace 6 by any positive integer n and a similar result holds good. In the
above situation, we say G1 and G2 are isomorphic groups. We now formalize
the above concept.

Definition 3.14. Let G and G be groups (distinct or not). A homomorphism

from G to G is a map f : G → G such that

f (ab) = f (a)f (b). (3.2)

In Definition 3.14, the multiplication operation has been used to denote

the group operations in G1 and G2 . If, for instance, G1 is an additive group
and G2 is a multiplicative group, Equation 3.2 should be changed to

f (a + b) = f (a)f (b)

and so on.

Definition 3.15. An isomorphism from a group G to a group G is a bijective

homomorphism from G to G , that is, it is a map f : G → G which is both a
bijection and a group homomorphism.

It is clear that if f : G → G is an isomorphism from G to G , then

−1
f : G → G is an isomorphism from G → G. Hence if there exists a group
isomorphism from G to G , we can say without any ambiguity that G and G
are isomorphic groups. A similar statement cannot be made for group homo-
morphism. If G is isomorphic to G , we write: G  G .

Examples of Groups, Homomorphisms and Isomorphisms

1. Let G = (Z, +), and G = (nZ, +), n = 0 (nZ is the set gotten by
multiplying all integers by n). The map f : G → G defined by f (m) =
mn, m ∈ G, is a group homomorphism from G onto G .
2. Let G = (R, +) and G = (R+ , ·). The map f : G → G defined by
f (x) = ex , x ∈ G, is a group homomorphism from G onto G .
3. Let G = (Z, +), and G = (Z × Z, +). The map f : G → G defined by
f (n) = (0, n), n ∈ Z is a homomorphism from G to G .
4. Let G = R2 = R × R, the real plane with addition + as group operation,
that is (x, y) + (x , y  ) = (x + x , y + y  ), and PX : R2 → R be defined by
PX (x, y) = x, the projection of R2 on the X-axis. PX is a homomorphism
from (R2 , +) to (R, +).

We remark that the homomorphism in the last list of examples is not onto
while those in Examples 1, 2 and 4 are onto. The homomorphisms in Examples
1 and 2 are isomorphisms. The isomorphism in Example 1 is an isomorphism
168 Discrete Mathematics

of G onto a proper subgroup of G. We now check that the map f of Example 2

is an isomorphism. First, it is a group homomorphism since

f (x + y) = ex+y = ex ey = f (x)f (y).

(Note that the group operation in G is addition while in G , it is multiplica-

tion). Next we check that f is 1 − 1. In fact, f (x) = f (y) gives ex = ey , and
therefore ex−y = 1. This means, as the domain of f is R, x − y = 0, and hence
x = y. Finally, f is onto. If y ∈ R+ , then there exists x such that ex = y; in
fact, x = loge y, is the unique preimage of y. Thus, f is a 1 − 1, onto group
homomorphism and hence it is a group isomorphism.

3.11 Properties of Homomorphisms of Groups

Let f : G → G be a group homomorphism. Then, f satisfies the following
properties:

Property 3.1. f (e) = e , that is, the image of the identity element e of G
under f is the identity element e of G .

Proof. For x ∈ G, the equation xe = x in G gives, as f is a group homomor-

phism, f (x)f (e) = f (xe) = f (x) = f (x)e in G . As G is a group, both the
cancellation laws are valid in G . Hence cancellation of f (x) gives f (e) = e .

Property 3.2. The image f (a−1 ) of the inverse of an element a of G is the

inverse of f (a) in G , that is, f (a−1 ) = (f (a))−1 .

Proof. The relation aa−1 = e in G gives f (aa−1 ) = f (e). But by Prop-

erty 3.1, f (e) = e , and as f is a homomorphism, f (aa−1 ) = f (a)f (a−1 ).
Thus, f (a)f (a−1 ) = e in G . This implies that f (a−1 ) = (f (a))−1 .

Property 3.3. The image f (G) ⊂ G is a subgroup of G . In other words, the

homomorphic image of a group is a group.

Proof.

i. Let f (a), f (b) ∈ f (G), where a, b ∈ G. Then, f (a)f (b) = f (ab) ∈ f (G),
as ab ∈ G.
ii. The associative law is valid in f (G). As f (G) ⊂ G and G being a group,
f (G) satisfies the associative law.

iii. By Property 3.1, the element f (e) ∈ f (G) acts as the identity element
of f (G).
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 169

iv. Let f (a) ∈ f (G), a ∈ G. By Property 3.2, (f (a))−1 = f (a−1 ) ∈ f (G), as

a−1 ∈ G.

Thus, f (G) is a subgroup of G .

Theorem 3.8. Let f : G → G be a group homomorphism and K =

{a ∈ G : f (a) = e }, that is, K is the set of all those elements of G that are
mapped by f to the identity element e of G. Then, K is a subgroup of G.

Proof. It is enough to check that if a, b ∈ K, then ab−1 ∈ K. (Then, aa−1 =

e ∈ K, and ea−1 = a−1 ∈ K. Further, as b−1 ∈ K, a(b−1 )−1 = ab ∈ K). Now
f (ab−1 ) = (as f is a group homomorphism)
−1
f (a)f (b−1 ) = f (a) (f (b)) = e (e )−1 = e e = e .

and hence ab−1 ∈ K. Thus, K is a subgroup of G.

Definition 3.16. The subgroup K defined in the statement of Theorem 3.8

is called the kernel of the group homomorphism f .

As before, let f : G → G be a group homomorphism.

Property 3.4. For a, b ∈ G, f (a) = f (b) iff ab−1 ∈ K, the kernel of f .

Proof.
−1
f (a) = f (b) ⇔ f (a) (f (b)) = e , the identity element of G
⇔ f (a)f (b−1 ) = e (By Property 3.2)
−1 
⇔ f (ab )=e
−1
⇔ ab ∈ K.

Property 3.5. f is a 1 − 1 map iff K = {e}.

Proof. Let f be 1 − 1, a ∈ K. Then, by the definition of the kernel, f (a) = e .

But e = f (e), by Property 3.1. Thus, f (a) = f (e), and this implies, as f is
1 − 1, that a = e.
Conversely, assume that K = {e}, and let f (a) = f (b). Then, by Prop-
erty 3.4, ab−1 ∈ K. Thus, ab−1 = e and so a = b. Hence f is 1 − 1.

Property 3.6. A group homomorphism f : G → G is an isomorphism iff

f (G) = G and K(= the kernel of f ) = {e}.

Proof. f is an isomorphism iff f is a 1 − 1, onto homomorphism. Now f is

1 − 1 iff K = {e}, by Property 3.5. Further, f is onto iff f (G) = G .

Property 3.7 (Composition of homomorphisms). Let f : G → G and

g : G → G be group homomorphisms. Then, the composition map h = gof :
G → G is also a group homomorphism.
170 Discrete Mathematics

Proof. h is a group homomorphism iff h(ab) = h(a)h(b) for all a, b ∈ G.

Now h(ab) = (gof )(ab) = g (f (ab))
= g (f (a)f (b)) , as f is a group homomorphism
= g (f (a)gf (b)) , as g is a group homomorphism
= (g · f )(a) · (g · f )(b)
= h(a)h(b).

3.12 Automorphism of Groups

Definition 3.17. An automorphism of a group G is an isomorphism of G
onto itself.

Example 3.4
 0 2

Let G = ω = 1, ω, ω be the group of cube roots of unity, where
ω = cos(2π/3) + i sin(2π/3). Let f : G → G be defined by f (ω) = ω 2 .
To make f a group homomorphism, we have to set f (ω 2 ) = f (ω · ω) =
f (ω)f (ω) = ω 2 ·ω 2 = ω, and f (1) = f (ω 3 ) = (f (ω))3 = (ω 2 )3 = (ω 3 )2 =
13 = 1. In other words, the homomorphism f : G → G is uniquely
defined on G once we set f (ω) = ω 2 . Clearly, f is onto. Further, only
1 is mapped to 1 by f , while the other two elements ω and ω 2 are
interchanged by f . Thus, Ker f = {1}. So by Property 3.7, f is an
isomorphism of G onto G, that is, an automorphism of G.

Our next theorem shows that there is a natural way of generating at least
one set of automorphisms of a group.

Theorem 3.9. Let G be a group and a ∈ G. The map fa : G → G defined by

fa (x) = axa−1 is an automorphism of G.
Proof. First we show that fa is a homomorphism. In fact, for x, y ∈ G,
fa (xy) = a(xy)a−1 by the definition of fa
−1
= a(xa ay)a−1
= (axa−1 )(aya−1 )
= fa (x)fa (y).
Thus, fa is a group homomorphism. Next we show that fa is 1 − 1. Suppose
for x, y ∈ G, fa (x) = fa (y). This gives axa−1 = aya−1 , and so by the two
cancellation laws that are valid in a group, x = y. Finally, if y ∈ G, then
a−1 ya ∈ G, and fa (a−1 ya) = a(a−1 ya)a−1 = (aa−1 )y(aa−1 ) = eye = y, and
so f is onto. Thus, f is an automorphism of the group G.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 171

Definition 3.18. A map of the form fa for some a ∈ G defined by fa (x) =

axa−1 , x ∈ G, is called an inner automorphism of G.

3.13 Normal Subgroups

Definition 3.19. A subgroup N of a group G is called a normal subgroup of
G (equivalently, N is normal in G) if

aN a−1 ⊆ N for each a ∈ G. (3.3)

In other words, N is normal in G if N is left invariant by the inner auto-

morphisms fa for each a ∈ G. We state this observation as a proposition.
Proposition 3.3. The normal subgroups of a groups G are those subgroups
of G that are left invariant by all the inner automorphisms of G.

Now the condition aN a−1 ⊆ N for each a ∈ G shows, by replacing a by

a , a−1 N (a−1 )−1 = a−1 N a ⊆ N . The latter condition is equivalent to
−1

N ⊆ aN a−1 for each a ∈ G. (3.4)

The conditions (3.3) and (3.4) give the following equivalent definition of a
normal subgroup.
Definition 3.20. A subgroup N of a group G is normal in G iff aN a−1 = N
(equivalently, aN = N a) for every a ∈ G.

Examples of normal subgroups

1. Let G = S3 , the group of 3! = 6 permutations on {1, 2, 3}. Let N =
{e, (123), (132)}. Then, N is a normal subgroup of S3 . First of all note
that N is a subgroup of G. In fact, we have (123)2 = (132), (132)2 =
(123), and (123)(132) = e. Hence (123)−1 = (132) and (132)−1 = (123).
Let a ∈ S3 . If a ∈ N , then aN = N = N a (See Lemma 3.4). So
let a ∈ S \ N . Hence a = (12), (23) or (13). If a = (12), then
aN a−1 = {(12)e(12), (12)(123)(12), (12)(132)(12)} = {e, (132), (123)}.
In a similar manner, we have (23)N (23) = N and (13)N (13) = N . Thus,
N is a normal subgroup of S3 .
2. Let H = {e, (12)} ⊂ S3 . Then, H is a subgroup of G that is not normal
in S3 . In fact, if a = (23), we have

aHa−1 = (23) {e, (12)} (23)

= {(23)e(23), (23)(12)(23)}
= {e, (13)} = H

Hence H is not a normal subgroup of S3 .

172 Discrete Mathematics

Definition 3.21. The center of a group G consists of those elements of G

each of which commutes with all the elements of G. It is denoted by C(G).
Thus,
C(G) = {x ∈ G : xa = ax for each a ∈ G}
For example, C(S3 ) = {e}, that is, the center of S3 is trivial. Also, it is easy
to see that the center of an Abelian group G is G itself.
Clearly the trivial subgroup {e} is normal in G and G is normal in G.
(Recall that aG = G for each a ∈ G.)
Proposition 3.4. The center C(G) of a group G is a normal subgroup of G.

Proof. We have for a ∈ G,

 
aC(G)a−1 = aga−1 : g ∈ C(G)
 
= (ag)a−1 : g ∈ C(G)
 
= (ga)a−1 : g ∈ C(G)
 
= g(aa−1 ) : g ∈ C(G)
= {g : g ∈ C(G)}
= C(G)

Theorem 3.10. Let f : G → G be a group homomorphism. Then, K = Ker f

is a normal subgroup of G.
 
Proof. We have, for a ∈ G, aKa−1 = aka−1 : k ∈ K . Now f (aka−1 ) =
f (a)f (k)f (a−1 ) = f (a)e f (a−1 ) = f (a)f (a−1 ) = f (a)(f (a))−1 = e . Hence
aka−1 ∈ K for each k ∈ K and so aKa−1 ⊂ Ker f = K for every a ∈ G. This
implies that K is a normal subgroup of G.

3.14 Quotient Groups (or Factor Groups)

Let G be a group, and H a normal subgroup of G. Let G/H (read G modulo
H) be the set of all left cosets of H. Recall that when H is a normal subgroup
of G, there is no distinction between the left coset aH and the right coset
Ha of H. The fact that H is a normal subgroup of G enables us to define a
group operation in G/H. We set, for any two cosets aH and bH of H in G,
aH · bH = (ab)H. This definition is well defined. By this we mean that if we
take different representative elements instead of a and b to define the cosets
aH and bH, still we end up with the same product. To be precise, let
aH = a1 H, and bH = b1 H (3.5)
−1 −1
Then, a a1 ∈ H and b b1 ∈ H.
Let a a1 = h ∈ H. This gives: (ab) (a1 b1 ) = b−1 (a−1 a1 )b1 = b−1 hb1 =
−1 −1

(b−1 hb)(b−1 b1 ) ∈ H, since b−1 hb ∈ H (H being a normal subgroup of G),

 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 173

and b−1 b1 ∈ H. Consequently, (aH) · (bH) = (a1 H)(b1 H) as the product on

the left is (ab)H while the poduct on the right is (a1 b1 )H. Hence the binary
operation on the set of cosets of H in G is well defined in the sense that it is
independent of the representative elements chosen for the cosets of H in G.
Further eH = H acts as the identity element of G/H as

(aH)(eH) = (ae)H = aH = (ea)H = eH · aH

Finally, the inverse of aH is a−1 H since

(aH)(a−1 H) = (aa−1 )H = eH = H,

and for a similar reason (a−1 H)(aH) = H.

Thus, G/H is a group under this binary operation. G/H is called the quotient
group or factor group of G modulo H.

Example 3.5
We now present an example of a quotient group. Let G = (R2 , +), the
additive group of points of the plane R2 . (If (x1 , y1 ) and (x2 , y2 ) are two
points of R2 , their sum (x1 , y1 ) + (x2 , y2 ) is defined as (x1 + x2 , y1 + y2 ).
The identity element of this group (0, 0) and the inverse of (x, y) is
(−x, −y)). Let H be the subgroup: {(x, 0) : x ∈ R} = X-axis. If (a, b) is
any point of R2 , then

(a, b) + H = {(a, b) + (x, 0) = (a + x, b) : x ∈ R}

=line through (a, b) parallel to X-axis. Clearly if (a, b)+H = (a , b )+H,
then ((a − a), (b − b)) ∈ H) = X-axis and therefore the Y-coordinate
b − b = 0 and so b = b. In other words, the line through (a, b) and the
line through (a , b ), both parallel to the X-axis, are the same iff b = b ,
as is expected (See Figure 3.1). For this reason, this line may be taken
as (0, b) + H. Thus, the cosets of H in R are the lines parallel to the
X-axis and therefore the elements of the quotient group R/H are the
lines parallel to the X-axis. If (a, b)+H and (a , b )+H are two elements
of R/H, we define their sum to be (a + a , b + b ) + H = (0, b + b ) + H,
the line through (0, b + b ) parallel to the X-axis. Note that (R2 , +)
is an Abelian group and so H is a normal subgroup of R2 . Hence the
above sum is well defined. The above addition defines a group structure
on the set of lines parallel to the X-axis, that is, the elements of R/H.
The identity element of the quotient group is the X-axis = H, and the
inverse of (0, b) + H is (0, −b) + H.

Our next result exhibits the importance of factor groups.

174 Discrete Mathematics

FIGURE 3.1: An example for quotient group.

3.15 Basic Isomorphism Theorem for Groups

Theorem 3.11. If there exists a homomorphism f from a group G onto a
group G with kernel K, then G/K  G .
Proof. We have to prove that there exists an isomorphism φ from the factor
group G/K onto G (observe that the factor group G/K is defined, as K is a
normal subgroup of G (see Theorem 3.10)). Define φ : G/K → G by φ(gK) =
f (g) (the mapping φ is pictorially depicted in Figure 3.1. See Example 1
below).
It is possible that gK = g  K with g = g  in G. Hence we need to establish
that φ is a well-defined map. As per our definition of φ, φ(gK) = f (g), while
φ(g  K) = f (g  ). Hence our definition of φ will be valid only if f (g) = f (g  ).
−1 −1
Now gK = g  K implies that g  g ∈ K. Let g  g = k ∈ K. Then, f (k) = e ,
−1 −1
the identity element of G . Moreover, e = f (g  g) = f (g  )f (g) =
 −1
f (g ) f (g) (as f is a group homomorphism, Property 3.2 holds). Thus,
f (g  ) = f (g), and f is well defined. We next show that φ is a group iso-
morphism.

i. φ is a group homomorphism: We have for g1 K, g2 K in G/K,

φ((g1 K)(g2 K)) = φ((g1 g2 )K)

= f (g1 g2 ), by the definition of φ
= f (g1 )f (g2 ), as f is a group homomorphism
= φ(g1 K)φ(g2 K)

Thus, φ is a group homomorphism.

ii. φ is 1 − 1: Suppose φ(g1 K) = φ(g2 K), where g1 K, g2 K ∈ G/K. This

gives that f (g1 ) = f (g2 ), and therefore f (g1 )f (g2 )−1 = e , the identity
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 175

element of G . But f (g1 )f (g2 )−1 = f (g1 )f (g2−1 ) = f (g1 g2−1 ). Hence
g1 g2−1 = e , and so g1 g2−1 ∈ K, and consequently, g1 K = g2 K (by
Property 3.4). Thus, φ is 1 − 1.

iii. φ is onto: Let g  ∈ G . As f is onto, there exists g ∈ G with f (g) = g  .

Now gK ∈ G/K, and φ(gK) = f (g) = g  . Thus, φ is onto and hence φ
is a group isomorphism.

3.15.1 Examples of factor groups

1. Let us see as to what this isomorphism means with regard to the factor
group R2 /K given in Example 3.5. Define f : R2 → R by f (a, b) = (0, b),
the projection of the point (a, b) ∈ R2 on the Y-axis = R. The identity
element of the image group is the origin (0, 0). Clearly K is the set
of all points (a, b) ∈ R2 that are mapped to (0, 0), that is, the set of
those points of R2 whose projections on the Y-axis coincide with the
origin. Thus, K is the X-axis(= R). Now φ : G/K = R2 /R → G is
defined by φ((a, b) + K) = f (a, b) = (0, b). This means that all points
of the line through (a, b) parallel to the X-axis are mapped to their
common projection on the Y-axis, namely, the point (0, b). Thus, the
isomorphism between G/K and G is obtained by mapping each line
parallel to the X-axis to the point where the line meets the Y-axis.
(Note: G/K = R2 /(X-axis), and G = Y -axis.)
2. We now consider another example. Let G = Sn , the symmetric group
of degree n, and G = {−1, 1}, the multiplicative group of two elements
(with multiplication defined in the usual way). 1 is the identity element
of G and the inverse of −1 is −1. Define f : G → G by setting

1 if σ ∈ Sn is an even permutation, that is, if σ ∈ An
f (σ) =
−1 if σ ∈ Sn is an odd permutation.

Recall that An is a subgroup of Sn . Now |Sn | = n! and |An | = n!/2.

Further, if α is an odd permutation in Sn , and σ ∈ An , then ασ is an
odd permutation, and hence |αAn | = n!/2. Let Bn denote the set of odd
permutations in Sn . Then, Sn = An ∪ Bn , An ∩ Bn = φ, and αAn = Bn
for each α ∈ Bn . Thus, Sn /An has exactly two distinct cosets of An ,
namely, An and Sn \ An = Bn .
The mapping f : Sn → {1, −1} defined by f (α) = 1 or −1 according
to whether the permutation α is even or odd clearly defines a group
homomorphism. The kernel K of this homomorphism is An , and we
have Sn \ An  {1, −1} = G . The isomorphism is obtained by mapping
the coset σAn to 1 or −1 according to whether σ is an even or an odd
permutation.
176 Discrete Mathematics

3.16 Exercises
1. Let G = SL(n, C) be the set of all invertible complex matrices A of
order n. If the operation · denotes matrix multiplication, show that G is
a group under ·.

2. Let G denote the set of all real matrices of the form ( ab 01 ) with a = 0.
Show that G is a group under matrix multiplication.
3. Which of the following semigroups are groups?
i. (Q, ·)
ii. (R∗ , ·)
iii. (Q, +)
iv. (R∗ , ·)
v. The set of all 2 by 2 real matrices under matrix multiplication.
vi. The set of all 2 by 2 real matrices of the form ( ab 01 ).
4. Prove that a finite semigroup in which the right and left cancellation laws
are valid is a group, that is, if H is a finite semigroup in which both the
cancellation laws are valid (that is, ax = ay implies that x = y, and
xa = ya implies that x = y, where a, x, y ∈ H), is a group.
5. Prove that in semigroup G in which the equations ax = b and yc = d,
are solvable in G, where a, b, c, d ∈ G, is a group.

6. In the group GL(2, C) of 2 × 2 complex non-singular matrices, find the

order of the following elements:
 0  0 
(i) 10 −i (ii) ( 10 11 ) (iii) 0i −i (iv) −2+3i
1−i
−1+2i
3−2i

7. Let G be a group, and φ : G → G be defined by φ(g) = g −1 , g ∈ G. Show

that φ is an automorphism of G iff G is Abelian.

8. Prove that any group of even order has an element of order 2. (Hint:
a = e and o(a) = 2 iff a = a−1 .) Pair off such elements (a, a−1 ).
9. Give an example of a non-cyclic group where each of its proper subgroup
is cyclic.
10. Show that no group can be the set union of two of its proper subgroups.

11. Show that S = {3, 5} generates the group (Z, +).

12. Give an example of an infinite non-Abelian group.

13. Show that the permutations T = {12345} and S = (25)(34) generate a

subgroup of order 10 in S5 .
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 177

14. Find if the following permutations are odd:


1 2 3 4 5 6 7 8 9
(i) (123)(456) (ii) (1546)(2)(3) (iii)
2 5 4 3 1 7 6 9 8
.
15. If G = {a1 , . . . , an } is a finite Abelian group of order n, show that
(a1 a2 . . . an )2 = e.
16. Let G = {a ∈ R : a = −1}. If the binary operation ∗ is defined in G by
a ∗ b = a + b + ab for all a, b ∈ G, show that (G, ∗) is a group.

17. Let G = {a ∈ R : a = 1}. If the binary operation ∗ is defined on G by

a ∗ b = a + b − ab for all a, b ∈ G, show that (G, ∗) is not a group.
18. Let σ = (i1 i2 . . . ir ) be a cycle in Sn of length r. Show that the order of
σ (= the order of the subgroup generated by σ) is r.
19. Prove that the center of a group is a normal subgroup G.
20. Let α, β, γ be permutations in S4 defined by
  
1 2 3 4 1 2 3 4 1 2 3 4
α= , β= , γ= .
1 4 3 2 2 1 4 3 3 1 2 4

Find (i) α−1 , (ii) α−1 βγ, (iii) βγ −1 .

21. Show that any infinite cyclic group is isomorphic to (Z, +).
22. Show that the set {ein : n ∈ Z} forms a multiplicative group. Show that
this is isomorphic to (Z, +). Is this group cyclic?
23. Find a homomorphism of the additive group of integers to itself that is
not onto.

24. Give an example of a group that is isomorphic to one of its proper

subgroups.
25. Prove that (Z, +) is not isomorphic to (Q, +). (Hint: Suppose ∃ an iso-
morphism φ : Z → Q. Let φ(5) = a ∈ Q. Then, ∃ b ∈ Q with 2b = a. Let
x ∈ Z be the preimage of b. Then, 2x = 5 in Z, which is not true.)
26. Prove that the multiplicative groups R and C are not isomorphic.
27. Give an example of an infinite group in which every element is of finite
order.

28. Give the group table of the group S3 . From the table, find the center of
S3 .
29. Show that if a subgroup H of a group G is generated by a subset S of
G, then H is a normal subgroup iff aSa−1 ⊂
S for each a ∈ G.
30. Show that a group G is Abelian iff the center C(G) of G is G.
178 Discrete Mathematics

31. Let G be a group. Let [G, G] denote the subgroup of G generated by

all elements of G of the form aba−1 b−1 (called the commutator of a
and b) for all pairs of elements a, b ∈ G. Show that [G, G] is a nor-
mal subgroup of G. [Hint: For c ∈ G, we have c(aba−1 b−1 )c−1 =
(cac−1 )(cbc−1 )(cac−1 )−1 (cbc−1 )−1 ∈ [G, G]. Now apply Exercise 29.]
32. Show that a group G is Abelian ⇔ [G, G] = {e}.
Remark: [G, G] is called the commutator subgroup of G. In general, the
commutators of a group G need not form a subgroup of G, and it is for
this reason that we take the subgroup generated by the commutators
of G. There is no known elementary counter-example. For a counter-
example, see for instance, Theory of Groups [34].

33. Let G be the set of all roots of unity, that is, G = {ω ∈ C : ω n = 1 for
some n ∈ N}. Prove that G is an Abelian group that is not cyclic.
34. If A and B are normal subgroups of a group G such that A ∩ B = {e}.
Then, show that for ∀a ∈ A and ∀b ∈ B, ab = ba.

35. If H is the only subgroup of a given finite order in a group G, show that
H is normal in G.

36. Show that any subgroup of a group G of index 2 is normal in G.

37. Prove that the subgroup {e, (13)} of S3 is not normal in S3 .

38. Prove that the subgroup {e, (123), (132)} is a normal subgroup of S3 .

3.17 Rings
The study of commutative rings arose as a natural abstraction of the alge-
braic properties of the set of integers, while that of fields arose out of the sets
of rational, real and complex numbers.
We begin with the definition of a ring and then proceed to establish some
of its basic properties.

3.17.1 Rings, definitions and examples

Definition 3.22. A ring is a non-empty set A with two binary operations,
denoted by + and · (called addition and multiplication, respectively) satisfying
the following axioms:
R1 : (A, +) is an Abelian group. (The identity element of (A, +) is denoted
by 0).

R2 : · is associative, that is, a · (b · c) = (a · b) · c for all a, b, c ∈ A.

 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 179

R3 : For all a, b, c ∈ A,
a · (b + c) = a · b + a · c (left distributive law)
(a + b) · c = a · c + b · c (right distributive law).
It is customary to write ab instead of a · b.

Examples of Rings
1. A = Z, the set of all integers with the usual addition + and the usual
multiplication taken as ·.

2. A = 2Z, the set of even integers with the usual addition and multiplica-
tion.

3. A = Q, R or C with the usual addition and multiplication.

4. A = Zn = {0, 1, 2, . . . , n − 1}, the set of integers modulo n, where +

and · denote addition and multiplication taken modulo n. (For instance,
if A = Z5 , then in Z5 , 3 + 4 = 7 = 2, and 3 · 4 = 12 = 2).
5. A = Z[X], the set of polynomials in the indeterminate X with integer
coefficients with addition + and multiplication · defined in the usual
way.
√  √ 
6. A = Z + i 3Z = a + ib 3 : a, b ∈ Z ⊂ C. Then, A is a ring with the
usual + and · in C.
7. (Ring of Gaussian integers). Let A = Z + iZ = {a + ib : a, b ∈ Z} ⊂ C.
Then, with the usual addition + and multiplication · in C, A is a ring.

8. (A ring of functions). Let A = C[0, 1], the set of all complex-valued

continuous functions on [0, 1]. For t ∈ [0, 1], and f, g ∈ A, set (f +g)(t) =
f (t) + g(t), and (f · g)(t) = f (t)g(t). Then, it is clear that both f + g
and f · g are in A. It is easy to check that A is a ring.

Definition 3.23. A ring A is called commutative if for all a, b ∈ A, ab = ba.

Hence if A is a non-commutative ring, there exists a pair of elements
x, y ∈ A with xy = yx.
All the rings given above in Examples 1 to 8 are commutative rings. We
now present an example of a non-commutative ring.

Example 3.6
Let A = M2 (Z), the set of all 2 by 2 matrices with integers as entries.
A is a ring with the usual matrix addition + and usual
 matrix multipli-

cation ·. It is a non-commutative ring since M = 10 10 , N = 01 01 are
in A, but M N = N M .
180 Discrete Mathematics

3.17.1.1 Unity element of a ring

Definition 3.24. An element e of a ring A is called an identity or unity
element of A if ea = ae = a for all a ∈ A.
An identity element of A, if it exists, must be unique. For, if e and f are
identity elements of A, then,
ef = e as f is an identity element of A,
ef = f as e is an identity element of A.
Therefore e = f . Hence if a ring A has an identity element e, we can refer
to it as the identity element e of A.
For the rings in Examples 1, 3 and 7 above, the number 1 is the identity
element. For the ring C[0, 1] of Example 8 above, the function 1 ∈ C[0, 1]
defined by 1(t) = 1 for all t ∈ [0,1] acts
as the identity element. For the ring
M2 (Z) of Example 9, the matrix 10 01 acts as the identity element. A ring may
not have an identity element. For instance, the ring 2Z in Section 3.17.1 above
has no identity element.

3.17.2 Units of a ring

An element a of a ring A with identity element e is called a unit in A if
there exist elements b and c in A such that ab = e = ca.
Proposition 3.5. If a is a unit in a ring A with identity element e, and if
ab = ca = e, then b = c.

Proof. We have, b = eb = (ca)b = c(ab) = ce = c.

We denote the element b(= c) described in Proposition 3.5 as the inverse of
a and denote it by a−1 . Thus, if a is a unit in A, then there exists an element
a−1 ∈ A such that aa−1 = a−1 a = e. Clearly, a−1 is unique.
Proposition 3.6. The units of a ring A (with identity element) form a group
under multiplication.
Proof. Exercise.

3.17.2.1 Units of the ring Zn

Let a be a unit in the ring Zn . (See Example in Section 3.17.1 above). Then,
there exists an x ∈ Zn such that ax = 1 in Zn , or equivalently, ax ≡ 1(mod n).
But this implies that ax − 1 = bn for some integer b. Hence gcd(a, n) = 1.
(Because if an integer c > 1 divides both a and n, then it should divide 1.)
Conversely, if (a, n) = 1, by Euclidean algorithm, there exist integers x and y
with ax + ny = 1, and therefore ax ≡ 1(mod n). This, however, means that a
is a unit in Zn . Thus, the set U of units of Zn is precisely the set of integers
in Zn that are relatively prime to n and we know that the order of this set is
φ(n), where φ is the Euler function.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 181

3.17.2.2 Zero divisors

In the ring Z of integers, a is a divisor of c if there exists an integer b such
that ab = c. As Z is a commutative ring, we simply say that a is a divisor of c
and not a left divisor or right divisor of c. Taking c = 0, we have the following
more general definition.

Definition 3.25. A left zero divisor in a ring A is a non-zero element a of

A such that there exists a non-zero element b of A with ab = 0 in A. a ∈ A is
a right zero divisor in A if ca = 0 for some c ∈ A, c = 0.

If A is a commutative ring, a left zero divisor a in A is automatically a

right zero divisor in A and vice versa. In this case, we simply call a a zero
divisor in A.

Examples relating to zero divisors

1. If a = 2 in Z4 , then a is a zero divisor in Z4 , as 2 · 2 = 4 = 0 in Z4 .

2. In the ring M2 (Z), the matrix [ 10 10 ] is a right zero divisor as

[ 00 01 ] [ 10 10 ] = [ 00 00 ]

and [ 00 01 ] is not the zero matrix of M2 (Z).

3. If p is a prime, then every non-zero element of Zp is a unit. This follows

from the fact that if 1 ≤ a < p, then (a, p) = 1. Hence no a ∈ Zp , a = 0
is a zero divisor in Zp .

Theorem 3.12. The following statements are true for any ring A.

1. a0 = 0a for any a ∈ A.

2. a(−b) = (−a)b = −(ab) for all a, b ∈ A.

3. (−a)(−b) = ab for all a, b ∈ A.

Proof. Exercise.

3.18 Integral Domains

An integral domain is an abstraction of the algebraic structure of the ring
of integers.

Definition 3.26. An integral domain A is a commutative ring with identity

element having no divisors of zero.
182 Discrete Mathematics

Examples relating to integral domains

1. The ring of integers Z and the ring of Gaussian integers Z + iZ are both
integral domains.

2. The ring 2Z of even integers is not an integral domain even though it

has no zero divisors (Why?).
3. Let Z × Z = {(a, b) : a ∈ Z, b ∈ Z}. For (a, b), (c, d) in Z × Z, define
(a, b) ± (c, d) = (a ± c, b ± d),
and(a, b) · (c, d) = (ac, bd).
Clearly Z×Z is a commutative ring with zero element (0, 0) and identity
element (1, 1) but not an integral domain as it has zero divisors. For
instance, (1, 0) · (0, 1) = (0, 0).

3.19 Exercises
1. Prove that Zn , n ≥ 2, is an integral domain iff n is a prime.

2. Give the proof of Proposition 3.6.

3. Determine the group of units of the rings:

√
(i) Z, (ii) M2 (Z), (iii) Z + iZ, (iv) Z + i 3Z.
4. Let A be a ring, and a, b1 , b2 , . . . , bn ∈ A. Then, show that a(b1 + b2 +
· · · + bn ) = ab1 + ab2 + · · · abn . (Hint: Apply induction on n).
5. Let A be a ring, and a, b ∈ A. Then, show that for any positive integer
n,
n(ab) = (na)b = a(nb).
(na stands for the element a + a + · · · (n times) of A).

6. Show that no unit of a ring A can be a zero divisor in A.

7. (Definition: A subset B of a ring A is a subring of A if B is a ring with

respect to the binary operations + and · of A). Prove:
i. Z is a subring of Q.
ii. Q is a subring of R.
iii. R is a subring of C.

8. Prove that any ring A with identity element and cardinality p, where p is
a prime, is commutative. (Hint: Verify that the elements 1, 1 + 1, . . . , 1 +
1 + · · · + 1 (p times) are all distinct elements of A.)
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 183

3.20 Ideals
One of the special classes of subrings of a ring is the class of ideals. Con-
sider, for example, the set S of all multiples of 3 (positive multiples, negative
multiples, and zero multiple) in the ring Z of integers. Then, it is easy to see
that S is a subring of Z. More than this, if n ∈ Z and a ∈ S, then na ∈ S
as na is also a multiple of 3. We then call S an ideal in Z. We now present
the formal definition of an ideal in a general ring (that is, not necessarily
commutative) A.

Definition 3.27. Let A be any ring, and S ⊂ A. S is called a left ideal in

the ring A if

i. S is a subring of A, and

ii. For a ∈ A, and s ∈ S, as ∈ S.

We call S a left ideal of A because we are multiplying s by a on the left

of s. If we take sa instead of as, and if sa ∈ S for every a ∈ A and s ∈ S, we
call S a right ideal of A. If S is both a left ideal and right ideal of A, it is a
two-sided ideal of A. If A is a commutative ring, it is obvious that we need
make no distinction between left ideals and right ideals.

Example 3.7
Let A be the ring of all 2 × 2 matrices over Z. (Note: A is a non-
commutative ring). Let
 
a 0
S= , a, b ∈ Z .
b 0
Then, as   
a 0 c 0 ac 0
= ,
b 0 d 0 bc 0
it is easy to see that S is a subring of A. Moreover, if
 
a 0 x y
∈ S, and ∈ A,
b 0 z t
we have   
x y a 0 xa + yb 0
= ∈ S.
z t b 0 za + tb 0
Hence S is a left ideal of A. However, S is not a right ideal of A. For
instance,
   
5 0 5 0 1 2 5 10
∈ S, but = ∈
/ S.
6 0 6 0 3 4 6 12
184 Discrete Mathematics

Example 3.8
Let A = Z[x], the ring of all polynomial in x with integer coefficients.
Let S =
2, x, the ideal generated by 2 and x in A = smallest ideal
containing 2 and x in A = Set of all integer polynomials in x with even
constant terms. (Note that the integer 0 is also an integer polynomial.
Also remember that A is a commutative ring.)

Definition 3.28. Let A and B be two rings. A ring homomorphism from A

to B is a map f : A → B such that for all a, b ∈ A,
i. f (a + b) = f (a) + f (b), and
ii. f (ab) = f (a)f (b).

Example 3.9
Let A = Z and B = Z5 consisting of integers modulo 5. Hence Z5 =
{0, 1, 2, 3, 4} where addition and multiplication are taken modulo 5. (For
example, 2 + 3 = 0, and 2.3 = 1 in Z5 .) Clearly Z5 is a ring. Now
consider the map f : Z → Z5 defined by f (a) = a0 where a0 ∈ Z5 , and
a ≡ a0 (mod 5). Then, f is a ring homomorphism.

In the above example, what are all the elements that are mapped to 3?
They are all the numbers n ≡ 3(mod 5) in Z. We denote this set by [3], where
[3] = {. . . , −12, −7, −2, 3, 8, 13, . . . }. We call this set the residue class mod-
ulo 5 defined by 3 in Z. Note that [3] = [8], etc. Hence Z5 = {[0], [1], [2], [3], [4]},
where [m] = [n] iff m ≡ n(mod 5) for any integers m and n iff m − n ∈ (5),
the ideal generated by 5 in Z. The ring Z5 is often referred to as the residue
class ring or quotient ring modulo the ideal (5).
More generally, let S be an ideal in a commutative ring A, and let AS
denote the set of residue classes modulo the ideal S, that is, for a ∈ A, the
residue class defined by a ∈ A, namely [a] = {a + s, s ∈ S} then AS is the
residue class ring or quotient ring defined by the ideal S is the ring A.

3.21 Principal Ideals

Definition 3.29. Let S be a subset of a commutative ring A. S is called a
principal ideal of A if

1. S is an ideal of A, and

2. S is generated by a single element, that is ∃ s ∈ S such that S = {as :

a ∈ A}.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 185

Definition 3.30. A principal ideal ring (or more specifically, a principal ideal
domain (P.I.D.)) is a commutative ring A without zero divisors and with unit
element 1 (that is, an integral domain) and in which every ideal is principal.

Example 3.10
The ring Z of integers is a PID. This is because first Z is an integral
domain. Suppose S is an ideal of A. If S = (0), the zero ideal, S is
principal. So assume that S = 0. If a ∈ S, and a = 0, (−1)a = −a ∈ S.
Of the two numbers a and −a, one is a positive integer. Let s be the
least positive integer belonging to S.
Claim: S =
s, the ideal generated by s. Let b be any element of S.
We have to show that b ∈
s, that is, b is an integral multiple of s. By
division algorithm in Z,

b = qs + r, 0  r < s.

As s ∈ S, qs ∈ S and hence b − qs = r ∈ S. But if r =  0, r < s, and this

contradicts the choice of s. In other words, b ∈
s, and S =
s.

Example 3.11
Let F be a field. Then, the ring F [x] of polynomials in x with coef-
ficients in F is a P.I.D.

Proof. Trivially, F [x] is a P.I.D. Let S be an ideal in F [x]. If S = (0),

trivially S is principal. So assume that S = (0): Let s(x) be the monic
polynomial of least degree r belonging to S. Such a polynomial s(x)
must exist since if 0 = a(x) ∈ S, and if the leading coefficient of a(x) is
k ∈ F, k −1 a(x) also ∈ S (as k −1 ∈ F ⊂ F [x]) ⇒ k −1 a(x) is a monic
polynomial in S. Moreover, such a monic polynomial in S is unique.
For, if there are two such monic polynomials of the same least degree
r in S, their difference is a polynomial in S of degree less than r. This
contradicts the choice of r.

We now imitate the proof of Example 3.10. Let a(x) be any poly-
nomial of F [x] in S. Divide a(x) by s(x) by Euclidean algorithm. This
gives

a(x) = q(x)s(x) + r(x),

186 Discrete Mathematics

where either r(x) = 0 or deg r(x) < deg s(x). As a(x), s(x) are in S,
q(x)s(x) ∈ (a), and a(x) − q(x)s(x) = r(x) ∈ S. By the choice of s(x),
we should have r(x) = 0 ⇒ a(x) = q(x)s(x) ⇒ every polynomial in S is
a multiple of s(x) in F [x] ⇒ S is a principal ideal in F [x] ⇒ F [x] is a
principal ideal domain.

Example 3.12
The ring Z[x], though an integral domain, is not a P.I.D. The ideal
S =
2, x in Z[x] which consists of all polynomials in x with even
constant terms is not a principal ideal. For, suppose S is a principal
ideal in Z[x]. Let S =
a(x). As 2 ∈ S, S must be equal to
2, since
otherwise 2 cannot be a multiple of a non-constant polynomial. But
then every polynomial in S must have even integers as coefficients, a
contradiction. (For instance 2 + 3x ∈ S). Hence Z[x] is not a P.I.D.

Definition 3.31. Let A and B be commutative rings, and let f : A → B

be a ring homomorphism, and let K ⊂ A be the set of those elements of A
that are mapped to the zero element of B. Then, K is called the kernel of the
homomorphism f , and denoted by Ker f .

Theorem 3.13. Let A and B be two commutative rings, and let f → B be a

ring homomorphism f . Then,

i. Kerf is an ideal of A, and

ii. If f is onto, then A/Kerf  B.

Proof.

i. Let x, y ∈ K and a ∈ A so that f (x) = 0 = f (y). Then, trivially,

x ± y, ax all belong to K. This is because, f (x ± y) = f (x) ± f (y) [Note
as y + (−y) = 0, f (−y) = −f (y)], and f (ax) = f (a)f (x) = f (a) · 0 = 0.
Hence K is an ideal of A.
ii. We are assuming that f : A → B is onto. Define the map f¯ : A/K → B
by f¯(a + K) = f (a), a ∈ A. Note that this map is well-defined. Indeed,
if a + K = b + K, then a − b ∈ K and hence f (a − b) = 0. But f (a − b) =
f (a) − f (b) =⇒ f (a) = f (b). Hence f¯(a + K) = f (a) = f (b) = f¯(b + K).
We now verify that f is a ring isomorphism.
A. f¯ is 1 − 1 : f¯(a1 + K) = f¯(a2 + K) for a1 , a2 ∈ A ⇒ f (a1 ) = f (a2 ) ⇒
f (a1 − a2 ) = 0 ∈ B ⇒ a1 − a2 ∈ K ⇒ a1 + K = a2 + K.
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 187

B. f¯ is onto: Let b ∈ B. Then, as f is onto, ∃ a ∈ A with f (a) = b.

Then, f¯(a + K) = f (a) = b ⇒ f¯ is onto.
C. f¯ is a ring homomorphism: For a1 , a2 ∈ K, f¯((a1 + K) + (a2 + K)) =
f¯((a1 + a2 ) + K) = f (a1 + a2 ) = f (a1 ) + f (a2 ) = f¯(a1 + K) +
f¯(a2 + K). In a similar manner, f¯((a1 + K)(a2 + K)) = f¯((a1 a2 ) +
K) = f (a1 a2 ) = f (a1 )f (a2 ) = f¯(a1 + K)f¯(a2 + K). This proves that
f¯ : A/K → B is a ring homomorphism.
Thus, f¯ is a 1 − 1, onto ring homomorphism and hence f¯ is an iso-
morphism.

Corollary 3.3. Let F be a field and let φ : F [x] → F [x]/(xn − 1) be the ring
homomorphism defined by:

φ(f (x)) = f (x) = Residue class of f (x) ∈ F [x]/(xn − 1).

Then, K is the principal ideal (xn − 1) ⊂ F [x].

Consider the special case when F = R, and π is the ring homomorphism

φ : R[x] → R[x]/(x3 − 1).

Then, φ(x5 + x4 + x3 + 1) = the residue class [x2 + x + 1 + 1] = [x2 + x + 2],

the class defined by x2 + x + 2 modulo the principal ideal (x3 − 1) in R[x].
This is because [x5 + x4 + x3 + 1] = [(x3 − 1)(x2 + x + 1) + (x2 + x + 2)] =
[x3 − 1][x2 + x + 1] + [x2 + x + 2] = [0][x2 + x + 1] + [x2 + x + 2] = [x2 + x + 2].
(Note: square brackets stand for residue classes.) More generally, we have:
Remark 3.1. Every element of R[x]/(xn − 1) can be identified with a real
polynomial of degree at most (n − 1).

3.22 Fields
We now discuss the fundamental properties of fields and then go on to
develop in the next chapter the properties of finite fields that are basic to
coding theory and cryptography. If rings are algebraic abstractions of the
set of integers, fields are algebraic abstractions of the sets Q, R and C (as
mentioned already).
Definition 3.32. A field is a commutative ring with identity element in which
every non-zero element is a unit.

Hence if F is a field, and F ∗ = F \ {0}, the set of non-zero elements of

F , then every element of F ∗ is a unit in F . Hence F ∗ is a group under the
multiplication operation of F . Conversely, if F is a commutative ring with
unit element and if F ∗ is a group under the multiplication operation of F ,
188 Discrete Mathematics

then every element of F ∗ is a unit under the multiplication operation of F ,

and hence F is a field. This observation enables one to give an equivalent
definition of a field.

Definition 3.33 (Equivalent definition). A field is a commutative ring F

with unit element in which the set F ∗ of non-zero elements is a group under
the multiplication operation of F .

3.22.1 Examples of fields

1. Q, R and C are all examples of infinite fields.

2. If F is a field, then F (x), the set of all rational functions in x, is an

infinite field (even if F is a finite field). Indeed, if a(x)/b(x) is a non-
zero rational function, so that a(x) is not the zero polynomial, then its
inverse is the rational function b(x)/a(x).

3. If p is a prime, then the ring Zp of integers modulo p is a field with p

elements.

Every field is an integral domain. To see this, all we have to verify is that
F has no zero divisors. Indeed, if ab = 0, a = 0, then as a−1 exists in F , we
have 0 = a−1 (ab) = (a−1 a)b = b in F . However, not every integral domain is a
field. For instance, the ring Z of integers is an integral domain but not a field.
(Recall that the only non-zero integers which are units in Z are 1 and −1.)

3.23 Characteristic of a Field

Definition 3.34. A field F is called finite if |F |, the cardinality of F , is finite;
otherwise, F is an infinite field.

Let F be a field whose zero and identity elements are denoted by 0F and
1F , respectively. A subfield of F is a subset F  of F such that F  is also a
field with the same addition and multiplication operations of F . This of course
means that the zero and unity elements of F  are the same as those of F . It is
clear that the intersection of any family of subfields of F is again a subfield of
F . Let P denote the intersection of the family of all subfields of F . Naturally,
this subfield P is the smallest subfield of F . Because if P  is a subfield of
F that is properly contained in P , then P ⊂ P  ⊂ P , a contradiction. This
=
smallest subfield P of F is called the prime field of F . Necessarily, 0F ∈ P
and 1F ∈ P .
As 1F ∈ P , the elements 1F , 1F + 1F = 2 · 1F , 1F + 1F + 1F = 3 · 1F and, in
general, n · 1F , n ∈ N, all belong to P . There are then two cases to consider:
 Algebraic Structures I (Matrices, Groups, Rings, and Fields) 189

Case 1: The elements n · 1F , n ∈ N, are all distinct. In this case, the subfield
P itself is an infinite field and therefore F is an infinite field.

Case 2: The elements n·1F , n ∈ N, are not all distinct. In this case, there exist
r, s ∈ N with r > s such that r·1F = s·1F , and therefore, (r−s)·1F = 0, where
r−s is a positive integer. Hence, there exists a least positive integer p such that
p·1F = 0. We claim that p is a prime number. If not, p = p1 p2 , where p1 and p2
are positive integers less than p. Then, 0 = p·1F = (p1 p2 )·1F = (p1 ·1F )(p2 ·1F )
gives, as F is a field, either p1 · 1F = 0 or p2 · 1F = 0. But this contradicts the
choice of p. Thus, p is prime.

Definition 3.35. The characteristic of a field F is the least positive integer p

such that p·1F = 0 if such a p exists; otherwise, F is said to be of characteristic
zero.

A field of characteristic zero is necessarily infinite (as its prime field already
is). A finite field is necessarily of prime characteristic. However, there are
infinite fields with prime characteristic. Note that if a field F has characteristic
p, then px = 0 for each x ∈ F .

Examples of integral domains and fields

i. The fields Q, R and C are all of characteristic zero.

ii. The field Zp of integers modulo a prime p is of characteristic p.

iii. For a field F , denote by F [X] the set of all polynomials in X over F , that
is, polynomials whose coefficients are in F . F [X] is an integral domain
and the group of units of F [X] = F ∗ , the set of all non-zero elements
of F .

iv. The field Zp (X) of rational functions of the form a(X)/b(X), where
a(X) and b(X) are polynomials in X over Zp , p being a prime, and
b(X) = 0, is an infinite field of (finite) characteristic p.

Theorem 3.14. Let F be a field of (prime) characteristic p. Then, for all

pn n n n n n
x, y ∈ F , (x ± y) = xp ± y p , and (xy)p = xp y p .

Proof. We apply induction on n. If n = 1, (by binomial theorem which is valid

for any commutative ring with unit element).
 
p p p p−1 p
(x + y) = x + x y + ··· + xy p−1 + y p
1 p−1

p p p
= x + y , since p| , 1 ≤ i ≤ p − 1. (3.3)
i
190 Discrete Mathematics

So assume that
n n n
(x + y)p = xp + y p , n ∈ N.

p
n+1 pn
Then, (x + y)p = (x + y)

n 
n p
= xp + y p (by induction assumption)
n n
= (xp )p + (y p )p (by Equation 3.3)
pn+1 pn+1
=x +y . (3.4)
n
Next we consider (x − y)p . If p = 2, then −y = y and so the result is valid.
If p is an odd prime, change y to −y in Equation 3.4. This gives
n n n
(x − y)p = xp + (−y)p
n n n
= xp + (−1)p y p
n n
= xp − y p ,
n
since (−1)p = −1 when p is odd.
Chapter 4
Algebraic Structures II
(Vector Spaces and Finite Fields)

4.1 Vector Spaces

In Section 3.22, we discussed some of the basic properties of fields. In the
present section, we look at the fundamental properties of vector spaces. We
follow up this discussion with a section on finite fields.
While the three algebraic structures—groups, rings, and fields—are natural
generalizations of integers and real numbers, the algebraic structure vector
space is a natural generalization of the 3-dimensional Euclidean space.
We start with the formal definition of a vector space. To define a vector
space, we need two objects: (i) a set V of vectors, and (ii) a field F of scalars.
In the case of the Euclidean 3-space, V is the set of vectors, each vector being
an ordered triple (x1 , x2 , x3 ) of real numbers and F = R, the field of real
numbers. The axioms for a vector space that are given in Definition 4.1 below
are easily seen to be generalizations of the properties of R3 .

Definition 4.1. A vector space (or linear space) V over a field F is a non-
void set V whose elements satisfy the following axioms:

A. V has the structure of an additive Abelian group.

B. For every pair of elements α and v, where α ∈ F and v ∈ V , there

exists an element αv ∈ V called the product of v by α such that
i. α(βv) = (αβ)v for all α, β ∈ F and v ∈ V , and
ii. 1v = v for each v ∈ V (here 1 is the identity or unity element of
the field F ).
C. i. For α ∈ F , and u, v in V , α(u+v) = αu+αv, that is, multiplication
by elements of F is distributive over addition in V .
ii. For α, β ∈ F and v ∈ V , (α + β)v = αv + βv, that is multiplication
of elements of V by elements of F is distributive over addition in F .
If F = R, V is called a real vector space; if F = C, then
V is called a complex vector space. When an explicit reference
to the field F is not required, we simply say that V is a vec-
tor space (omitting the words “over the field F ”). The product

191
192 Discrete Mathematics

αv, α ∈ F, v ∈ V , is often referred to as scalar multiplication,

α being a scalar.

4.1.1 Examples of vector spaces

1. Let V = R3 , the set of ordered triples (x1 , x2 , x3 ) of real numbers.
Then, R3 is a vector space over R (as mentioned earlier) and hence R3
is a real vector space. More generally, if V = Rn , n ≥ 1, the set of
ordered n-tuples (x1 , . . . , xn ) of real numbers, then Rn is a real vector
space. If x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) ∈ Rn , then x + y =
(x1 + y1 , . . . , xn + yn ), and αx = (αx1 , . . . , αxn ). The zero vector of Rn
is (0, . . . , 0). Rn is known as the real affine space of dimension n.
2. Let V = Cn , the set of ordered n-tuples (z1 , . . . , zn ) of complex num-
bers. Then, V is a vector space over R as well as over C. We note that
the real vector space Cn and the complex vector space Cn are essen-
tially different spaces despite the fact that the underlying set of vectors
is the same in both cases.

3. R is a vector space over Q.

4. F [X], the ring of polynomials in X over the field F , is a vector space

over F .

5. The set of solutions of a homogeneous linear ordinary differential equa-

tion with real coefficients forms a real vector space. The reason is that
any such differential equation has the form

dn y dn−1 y
+ C1 + · · · + Cn−1 y = 0. (4.1)
dxn dxn−1
Clearly, if y1 (x) and y2 (x) are two solutions of the differential Equa-
tion 4.1, then so is y(x) = α1 y1 (x) + α2 y2 (x), α1 , α2 ∈ R. It is now easy
to verify that the axioms of a vector space are all satisfied.

4.2 Subspaces
The notion of a subspace of a vector space is something very similar to the
notions of a subgroup, subring, and subfield.

Definition 4.2. A subspace W of a vector space V over F is a subset W

of V such that W is also a vector space over F with addition and scalar
multiplication as defined for V .
 Algebraic Structures II (Vector Spaces and Finite Fields) 193

Proposition 4.1. A non-void subset W of a vector space V is a subspace of

V iff for all u, v ∈ W and α, β ∈ F ,
αu + βv ∈ W
Proof. If W is a subspace of V , then, as W is a vector space over F (with the
same addition and scalar multiplication as in V ), αu ∈ W and βv ∈ W , and
therefore αu + βv ∈ W .
Conversely, if the condition holds, then it means that W is an additive
subgroup of V . Moreover, taking β = 0, we see that for each α ∈ F , u ∈ W ,
αu ∈ W . As W ⊂ V , all the axioms of a vector space are satisfied by W and
hence W is a subspace of V .

4.2.1 An example of a subspace

Let W = {(a, b, 0) : a, b ∈ R}. Then, W is a subspace of R3 . (To see this,
apply Definition 4.2.) Geometrically, this means that the xy-plane of R3 (that
is, the set of points of R3 with the z-coordinate zero) is a subspace of R3 .
Proposition 4.2. If W1 and W2 are subspaces of a vector space V , then
W1 ∩ W2 is also a subspace of V . More generally, the intersection of any
family of subspaces of a vector space V is also a subspace of V .
Proof. Let u, v ∈ W = W1 ∩ W2 , and α, β ∈ F . Then, αu + βv belongs to
W1 as well as to W2 by Proposition 4.1, and therefore to W . Hence W is a
subspace of V again by Proposition 4.1. The general case is similar.

4.3 Spanning Sets

Definition 4.3. Let S be a subset of a vector space V over F . By the sub-
space spanned by S, denoted by S, we mean the smallest subspace of V that
contains S. If S = V , we call S a spanning set of V .

Clearly, there is at least one subspace of V containing S, namely, V . Let

S denote the collection of all subspaces of V containing S. Then, ∩W ∈S W is
also a subspace of V containing S. Clearly, it is the smallest subspace of V
containing S, and hence S = ∩W ∈S W .

Example 4.1
We shall determine the smallest subspace W of R3 containing the
vectors (1, 2, 1) and (2, 3, 4).
Clearly, W must contain the subspace spanned by (1, 2, 1), that is,
the line joining the origin (0, 0, 0) and (1, 2, 1). Similarly, W must also
194 Discrete Mathematics

contain the line joining (0, 0, 0) and (2, 3, 4). These two distinct lines
meet at the origin and hence define a unique plane through the origin,
and this is the subspace spanned by the two vectors (1, 2, 1) and (2, 3, 4).
(See Proposition 4.3 below.)

Proposition 4.3. LetS be a subset of a vector space V over F . Then, S =

L(S), where L(S) = α1 s1 + α2 s2 + · · · + αr sr : si ∈ S, 1 ≤ i ≤ r and
αi ∈ F, 1 ≤ i ≤ r, r ∈ N = set of all finite linear combinations of vectors of
S over F .

Proof. First, it is easy to check that L(S) is a subspace of V . In fact, let

u, v ∈ L(S) so that

u = α1 s1 + · · · + αr sr , and
v = β1 s1 + · · · + βt st

where si ∈ S, αi ∈ F for each i, and sj ∈ S, βj ∈ F for each j. Hence if

α, β ∈ F , then

αu + βv = (αα1 )s1 + · · · + (ααr )sr + (ββ1 )s1 + · · · + (ββt )st ∈ L(S).

Hence by Proposition 4.1, L(S) is a subspace of V . Further 1·s = s ∈ L(S) for

each s ∈ S, and hence L(S) contains S. But by definition, S is the smallest
subspace of V containing S. Hence S ⊆ L(S).
Now, let W be any subspace of V containing S. Then, any linear combi-
nation of vectors of S is a vector of W , and hence L(S) ⊆ W . In other words,
any subspace of V that contains the set S must contain the subspace L(S).
Once again, as S is the smallest subspace of V containing S, L(S) ⊆ S.
Thus, S = L(S).

Note : If S = {u1 , . . . , un } is a finite set, then S = u1 , . . . , un  = subspace

of linear combinations of u1 , . . . , un over F . In this case, we say that S gener-
ates the subspace S or S is a set of generators for S. Also L(S) is called
the linear span of S in V .

Proposition 4.4. Let u1 , . . . , un and v be vectors of a vector space V . Suppose

that v ∈ u1 , u2 , . . . , un . Then, u1 , . . . , un  = u1 , . . . , un ; v.

Proof. Any element α1 u1 + · · · + αn un , αi ∈ F for each i, of u1 , . . . , un  can

be rewritten as
α1 u1 + · · · + αn un + 0 · v
and hence belongs to u1 , . . . , un ; v. Thus,

u1 , . . . , un  ⊆ u1 , . . . , un ; v
 Algebraic Structures II (Vector Spaces and Finite Fields) 195

Conversely, if

w = α1 u1 + · · · + αn un + βv ∈ u1 , . . . , un ; v,

then as v ∈ u1 , . . . , un , v = γ1 u1 + · · · + γn un , γi ∈ F, and therefore,

w = (α1 u1 + · · · + αn un ) + β (γ1 u1 + · · · + γn un )
n
= (αi + βγi ) ui ∈ u1 , . . . , un .
i=1

Thus, u1 , . . . , un ; v ⊆ u1 , . . . , un  and therefore

u1 , . . . , un  = u1 , . . . , un ; v.

Corollary 4.1. If S is any non-empty subset of a vector space V , and v ∈ S,

then S ∪ {v} = S.
Proof. v ∈ S implies, by virtue of Proposition 4.3, v is a linear combination
of a finite set of vectors in S. Now the rest of the proof is as in the proof of
Proposition 4.4.

4.4 Linear Independence of Vectors

Let V = R3 , and e1 = (1, 0, 0), e2 = (0, 1, 0) and e3 = (0, 0, 1) in R3 . If
v = (x, y, z) is any vector of R3 , then v = xe1 + ye2 + ze3 . Trivially, this is the
only way to express v as a linear combination of e1 , e2 , e3 . For this reason, we
call {e1 , e2 , e3 } a base for R3 . We now formalize these notions.
Definition 4.4. Linear independence and linear dependence of vectors

i. A finite subset S = {v1 , . . . , vn } of vectors of a vector space V over a

field F is said to be linearly independent if the equation

α1 v1 + α2 v2 + · · · + αn vn = 0, αi ∈ F

implies that αi = 0 for each i.

In other words, a linearly independent set of vectors admits only the
trivial linear combination between them, namely,

0 · v1 + 0 · v2 + · · · + 0 · vn = 0.

In this case we also say that the vectors v1 , . . . , vn are linearly indepen-
dent over F . In the above equation, the zero on the right refers to the
zero vector of V , while the zeros on the left refer to the scalar zero, that
is, the zero element of F .
196 Discrete Mathematics

ii. An infinite subset S of V is linearly independent in V if every finite

subset of vectors of S is linearly independent.

iii. A subset S of V is linearly dependent over F if it is not linearly inde-

pendent over F . This means that there exists a finite subset {v1 , . . . , vn }
of S and a set of scalars α1 , . . . , αn , not all zero, in F such that
α1 v1 + · · · + αn vn = 0.
If {v1 , . . . , vn } is linearly independent over F , we also note that the
vectors v1 , . . . , vn are linearly independent over F .
Remark 4.1.

i. The zero vector of V forms a linearly dependent set since it satisfies the
non-trivial equation 1 · 0 = 0, where 1 ∈ F and 0 ∈ V .
ii. Two vectors of V are linearly dependent over F iff one of them is a
scalar multiple of the other.

iii. If v ∈ V and v = 0, then {v} is linearly independent (since for α ∈ F,

αv = 0 implies that α = 0).
iv. The empty set is always taken to be linearly independent.

Proposition 4.5. Any subset T of a linearly independent set S of a vector

space V is linearly independent.
Proof. First assume that S is a finite subset of V . We can take T = {v1 , . . . , vr }
and S = {v1 , . . . , vr ; vr+1 , . . . , vn }, n ≥ r. The relation
α1 v1 + · · · αr vr = 0, αi ∈ F
is equivalent to the condition that
(α1 v1 + · · · αr vr ) + (0 · vr+1 + · · · 0 · vn ) = 0.
But this implies, as S is linearly independent over F , αi = 0, 1 ≤ i ≤ n. Hence
T is linearly independent.
If S is an infinite set and T ⊆ S, then any finite subset of T is a finite
subset of S and hence linearly independent. Hence T is linearly independent
over F .
A restatement of Proposition 4.5 is that any superset in V of a linearly
dependent subset of V is also linearly dependent.

Corollary 4.2. If v ∈ L(S), then S ∪ {v} is linearly dependent.

Proof. By hypothesis, there exist v1 , . . . , vn in S, and α1 , . . . , αn ∈ F such

that v = α1 v1 + · · · αn vn and therefore α1 v1 + · · · αn vn + (−1)v = 0. Hence
{v1 , . . . , vn ; v} is linearly dependent and so by Proposition 4.5, S ∪ {v} is
linearly dependent.
 Algebraic Structures II (Vector Spaces and Finite Fields) 197

Examples of linearly independent set of vectors

1. C is a vector space over R. The vectors 1 and i of C are linearly inde-
pendent over R. In fact, if α, β ∈ R, then

α·1+β·i=0

gives that α + βi = 0, and therefore α = 0 = β. One can check that

{1 + i, 1 − i} is also linearly independent over R, while {2 + i, 1 + i, 1 − i}
is linearly dependent over R. The last assertion follows from the fact
that if u = 2 + i, v = 1 + i and w = 1 − i, then u + w = 3 and v + w = 2
so that
2(u + w) = 3(v + w)
and this gives
2u − 3v − w = 0.
2. The infinite set of polynomials S = {1, X, X 2 , . . .} in the vector space
R[X] of polynomials in X with real coefficients is linearly independent.
Recall that an infinite set S is linearly independent iff every
finite subset of S is linearly independent. So consider a finite subset
{X i1 , X i2 , . . . , X in } of S. The equation

λ1 X i1 + λ2 X i2 + · · · + λn X in = 0, the zero polynomial of R[X], (A)

where the scalars λi , 1 ≤ i ≤ n, all belong to R, implies that the

polynomial on the left side of Equation (A) is the zero polynomial and
hence must be zero for every real value of X. In other words, every real
number is a zero (root) of this polynomial. This is possible only if each
λi is zero. Hence the set S is linearly independent over R.

4.5 Bases of a Vector Space

Definition 4.5. A basis (or base) of a vector space V over a field F is a
subset B of V such that

i. B is linearly independent over F , and

ii. B spans V ; in symbols, B = V .

Condition (ii) implies that every vector of V is a linear combination of (a

finite number of ) vectors of B while condition (i) ensures that the expression
is unique. Indeed, if u = α1 u1 + · · · + αn un = β1 u1 + · · · + βn un , where the ui’s
are all in B, then (α1 −β1 )u1 +· · ·+(αn −βn )un = 0. The linear independence
of the vectors u1 , . . . , un means that αi − βi = 0, that is, αi = βi for each i.
198 Discrete Mathematics

Notice that we have taken the same u1 , . . . , un in both expressions, as we can
always add terms with zero coefficients. For example, if

u = α1 u1 + α2 u2 = β1 u1 + β2 u2 , then

u = α1 u1 + α2 u2 + 0 · u1 + 0 · u2
= 0 · u1 + 0 · u2 + β1 u1 + β2 u2 .)

Example 4.2
The vectors e1 = (1, 0, 0), e2 = (0, 1, 0) and e3 = (0, 0, 1) form a
basis for R3 . This follows from the following two facts.
1. {e1 , e2 , e3 } is linearly independent in R3 . In fact, α1 e1 + α2 e2 +
α3 e3 = 0, αi ∈ R, implies that

α1 (1, 0, 0) + α2 (0, 1, 0) + α3 (0, 0, 1) = (α1 , α2 , α3 ) = (0, 0, 0),

and hence αi = 0, 1 ≤ i ≤ 3.
2. e1 , e2 , e3  = R3 . To see this, any vector of e1 , e2 , e3  is of the
form α1 e1 + α2 e2 + α3 e3 = (α1 , α2 , α3 ) ∈ R3 and, conversely, any
(α1 , α2 , α3 ) in R3 is α1 e1 + α2 e2 + α3 e3 and hence belongs to
e1 , e2 , e3 .

4.6 Dimension of a Vector Space

Definition 4.6. By a finite-dimensional vector space, we mean a vector space
that can be generated (or spanned) by a finite number of vectors in it.

Our immediate goal is to establish that any finite-dimensional vector space

has a finite basis and that any two bases of a finite-dimensional vector space
have the same number of elements.

Lemma 4.1. No finite-dimensional vector space can have an infinite basis.

Proof. Let V be a finite-dimensional vector space with a finite spanning set

S = {v1 , . . . , vn }. Suppose to the contrary that V has an infinite basis B. Then,
as B is a basis, vi is a linear combination of a finite subset Bi , 1 ≤ i ≤ n of
B. Let B  = ∪ni=1 Bi . Then, B  is also a finite subset of B. As B  ⊂ B, B 
is linearly independent and further, as each v ∈ V is a linear combination of
v1 , . . . , vn , v is also a linear combination of the vectors of B  . Hence B  is also
a basis for V . If x ∈ B \ B  , then x ∈ L(B  ) and so B  ∪ {x} is a linearly
dependent subset of the linearly independent set B, a contradiction.
 Algebraic Structures II (Vector Spaces and Finite Fields) 199

Lemma 4.2. A finite sequence {v1 , . . . , vn } of non-zero vectors of a vector

space V is linearly dependent iff for some k, 2 ≤ k ≤ n, vk is a linear combi-
nation of its preceding vectors.

Proof. In one direction, the proof is trivial; if vk ∈ v1 , . . . , vk−1 , then by

Proposition 4.5, {v1 , . . . vk−1 ; vk } is linearly dependent and so is its superset
{v1 , . . . , vn }.
Conversely, assume that {v1 , v2 , . . . , vn } is linearly dependent. As v1 is a
non-zero vector, {v1 } is linearly independent (see (iii) of Remark 4.1). Hence,
there must exist a k, 2 ≤ k ≤ n such that {v1 , . . . , vk−1 } is linearly indepen-
dent while {v1 , . . . , vk } is linearly dependent since at worst k can be n. Hence
there exists a set of scalars α1 , . . . , αk , not all zero, such that

α1 v1 + · · · + αk vk = 0.

Now αk = 0; for if αk = 0, there exists a non-trivial linear relation connecting

v1 , . . . , vk−1 contradicting the fact that {v1 , . . . , vk−1 } is linearly independent.
Thus,
vk = −αk−1 α1 v1 − · · · − αk−1 αk−1 vk−1 ,
and hence vk is a linear combination of its preceding vectors.

Lemma 4.2 implies, by Proposition 4.4, that under the stated conditions
on vk ,
∧
v1 , . . . , vk , . . . , vn  = v1 , . . . , vk−1 , vk+1 , . . . , vn  = v1 , . . . , vk , . . . , vn ,

where the ∧ symbol upon vk indicates that the vector vk should be deleted.
We next prove a very important property of finite-dimensional vector
spaces.

Theorem 4.1. Any finite-dimensional vector space has a basis. Moreover,

any two bases of a finite-dimensional vector space have the same number of
elements.
Proof. Let V be a finite-dimensional vector space. By Lemma 4.1, every basis
of V is finite. Let S = {u1 , . . . , um }, and T = {v1 , . . . , vn } be any two bases
of V . We want to prove that m = n.
Now v1 ∈ V = u1 , . . . , um . Hence the set S1 = {v1 ; u1 , . . . , um } is linearly
dependent. By Lemma 4.2, there exists a vector ui1 ∈ {u1 , . . . , um } such that
∧
v1 ; u1 , . . . , um  = v1 ; u1 , . . . ui1 , . . . , um .

Now consider the set of vectors

∧
S2 = {v2 , v1 ; u1 , . . . , , ui1 , . . . , um } = {v2 , v1 ; u1 , . . . um } \ {ui1 }.
∧
As v2 ∈ V = v1 ; u1 , . . . , ui1 , . . . , um , there exists a vector ui2 ∈ {u1 , . . . ,
∧
ui1 , . . . , um } such that ui2 is a linear combination of the vectors preceding it
200 Discrete Mathematics

in the sequence S2 . (Such a vector cannot be a vector of T as every subset of

T is linearly independent.) Hence if

∧ ∧
S3 = {v1 , v2 ; u1 , . . . , ui1 , . . . ui2 , . . . , um }
 
= {v1 , v2 ; u1 , . . . , ui1 , . . . ui2 , . . . , um } \ ui1 , ui2 ,
S3  = V.

Thus, every time we introduce a vector from T , we are in a position to delete a

vector from S. Hence |T | ≤ |S|, that is, n ≤ m. Interchanging the roles of the
bases S and T , we see, by a similar argument, that m ≤ n. Thus, m = n.

Note that we have actually shown that any finite spanning subset of a
finite-dimensional vector space V does indeed contain a finite basis of V .
Theorem 4.1 makes the following definition unambiguous.

Definition 4.7. The dimension of a finite-dimensional vector space is the

number of elements in any one of its bases.

If V is of dimension n over F , we write dimF V = n or, simply, dim V = n,

when F is known.

Examples for dimension of a vector space

1. Rn is of dimension n over R. In fact, it is easy to check that the set of
vectors

S = {e1 = (1, 0, . . . , 0), e2 = (0, 1, 0, . . . , 0), . . . , en = (0, . . . , 0, 1)}

is a basis for Rn over R.

2. Cn is of dimension n over C.

3. Cn is of dimension 2n over R. In fact, if ek ∈ Cn has 1 in the√ k-th

position and 0 in the remaining positions, and fk ∈ Cn has i = −1 in
the k-th position and 0 in the remaining positions, then

S = {e1 , . . . , en ; f1 , . . . , fn }

forms a basis of Cn over R. (Verify!)

4. Let Pn (X) denote the set of real polynomials
 in X with realcoefficients
of degrees not exceeding n. Then, B = 1, X, X 2 , . . . , X n is a basis
for Pn (X). Hence dimR Pn (X) = n + 1.

5. The vector space R over Q is infinite-dimensional. This can be seen as

follows: Suppose dimQ R = n (finite). Then, R has a basis {v1 , . . . , vn }
 Algebraic Structures II (Vector Spaces and Finite Fields) 201

over Q. Hence

R = {α1 v1 + · · · + αn vn : αi ∈ Q for each i} .

But as Q is countable, the number of such linear combinations is count-

able. This is a contradiction as R is uncountable. Thus, R is infinite-
dimensional over Q.
Proposition 4.6. Any maximal linearly independent subset of a finite-
dimensional vector space V is a basis for V .

Proof. Let B be a maximal linearly independent subset of V , that is, B is not

a proper subset of B , where B is linearly independent in V . Suppose B is
not a basis for V . This means, by the definition of a basis for a vector space,
that there exists a vector x in V such that x ∈ B. Then, B ∪ {x} must be a
linearly independent subset of V ; for, suppose B ∪ {x} is linearly dependent.
Then, there exist v1 , . . . , vn in B satisfying a non-trivial relation

α1 v1 + · · · + αn vn + βx = 0, αi ∈ F, for each i and β ∈ F .

Now, β = 0, as otherwise, {v1 , . . . , vn } would be linearly dependent over F .

Thus, x = −β −1 (α1 v1 + · · · + αn vn ) ∈ B, a contradiction. Thus, B ∪ {x}
is linearly independent. But then the fact that B ∪ {x} ⊃ B violates the
maximality of B. Hence B must be a basis for V .

4.7 Solutions of Linear Equations and Rank of a Matrix

Let ⎡a a12 ... a1n ⎤
a11
21 a22 ... a2n
A = ⎣ .. .. .. ⎦
. . .
am1 am2 ... amn
be an m by n matrix over a field F . To be precise, we take F = R, the field of
real numbers. Let R1 , R2 , . . . , Rm be the row vectors and C1 , C2 , . . . , Cn the
column vectors of A. Then, each Ri ∈ Rn and each Cj ∈ Rm . The row space
of A is the subspace R1 , . . . , Rm  of Rn , and its dimension is the row rank
of A. Clearly, (row rank of A) ≤ m since any m vectors of a vector space span
a subspace of dimension at most m. The column space of A and the column
rank of A(≤ n) are defined in an analogous manner.
We now consider three elementary row transformations (or operations)
defined on the row vectors of A:
i. Rij : interchange of the i-th and j-th rows of A.
ii. kRi : multiplication of the i-th row vector Ri of A by a non-zero scalar
(real number) k.
202 Discrete Mathematics

iii. Ri +cRj : addition to the i-th row of A, c times the j-th row of A, c being
a scalar.

The elementary column transformations are defined in an analogous manner.

The inverse of each of these three transformations is again a transformation
of the same type. For instance, the inverse of Ri + cRj is obtained by adding
to the new i-th row, −c times Rj .
Let A∗ be a matrix obtained by applying a finite sequence of elementary
row transformations to a matrix A. Then, the row space of A = row space
of A∗ , and hence, row rank of A = dim(row space of A) = dim(row space of
A∗ ) = row rank of A∗ . Now a matrix A∗ is said to be in row-reduced echelon
form if:
i. The leading non-zero entry of any non-zero row (if any) of A∗ is 1.
ii. The leading 1s in the non-zero rows of A∗ occur in increasing order of
their columns.
iii. Each column of A∗ containing a leading 1 of a row of A∗ has all its other
entries zero.
iv. The non-zero rows of A∗ precede its zero rows, if any.
Now let D be a square matrix of order n. The three elementary row (respec-
tively column) operations considered above do not change the singular or
non-singular nature of D. In other words, if D∗ is a row-reduced echelon form
of D, then D is singular iff D∗ is singular. In particular, D is non-singular iff
D∗ = In , the identity matrix of order n. Hence if a row-reduced echelon form
A∗ of a matrix A has r non-zero rows, the maximum order of a non-singular
square submatrix of A is r. This number is called the rank of A.
Definition 4.8. The rank of a matrix A is the maximum order of a non-
singular square submatrix of A. Equivalently, it is the maximum order of a
non-vanishing determinant minor of A.

Note that from our earlier discussions, rank of A = row rank of A, = col-
umn rank of A. Consequently, a set of n vectors of Rn is linearly independent
iff the determinant formed by them is not zero; equivalently, the n × n matrix
formed by them is invertible.

Example 4.3
Find the row-reduced echelon form of
⎡ ⎤
1 2 3 −1
⎢2 1 −1 4 ⎥
A=⎢ ⎣3 3 2
⎥.
3⎦
6 6 4 6
 Algebraic Structures II (Vector Spaces and Finite Fields) 203

As the leading entry of R1 is 1, we perform the operations R2 − 2R1 ;

R3 − 3R1 ; R4 − 6R1 (where Ri stands for the i-th row of A). This gives
⎡ ⎤
1 2 3 −1
⎢0 −3 −7 6⎥
A1 = ⎢⎣0 −3 −7
⎥.
6⎦
0 −6 −14 12
Next perform (−1/3)R2 (that is, replace R2 by (−1/3)R2 ). This gives
⎡ ⎤
1 2 3 −1
⎢0 1 7/3 −2⎥
A1 = ⎢
⎣0 −3 −7
⎥.
6⎦
0 −6 −14 12

Now perform R1 − 2R2 (that is, replace R1 by R1 − 2R2 , etc.); R3 + 3R2 ;

R4 + 6R2 . This gives the matrix
⎡ ⎤
1 0 −5/3 3
⎢0 1 7/3 −2⎥
A2 = ⎢
⎣0 0
⎥.
0 0⎦
0 0 0 0

A2 is the row-reduced echelon form of A. Note that A2 is uniquely

determined by A. Since the maximum order of a non-singular submatrix
of A2 is 2, rank of A = 2. Moreover, row space of A2 = R1 , R2 (of A2 ).
Clearly R1 and R2 are linearly independent over R since for α1 , α2 ∈ R,
α1 R1 + α2 R2 = (α1 , α2 , −5α1 /3 + 7α2 /3, 3α1 − 2α2 ) = 0 = (0, 0, 0, 0)
implies that α1 = 0 = α2 . Thus, the row rank of A is 2 and therefore
the column rank of A is also 2.

Remark 4.2. Since the last three rows of A1 are proportional (that is, one
row is a multiple of the other two), any 3 × 3 submatrix of A1 will be singular.

Since A1 has a non-singular submatrix of order 2, (for example, 10 −3 2
), A1
is of rank 2 and we can conclude that A is also of rank 2.

Remark 4.3. The word “echelon” refers to the formation of army troops in
parallel divisions each with its front clear of that in advance.

4.8 Exercises
1. Show that Z is not a vector space over Q.
204 Discrete Mathematics

2. If n ∈ N, show that the set of all real polynomials of degree n ≥ 1

does not form a vector space over R (under usual addition and scalar
multiplication of polynomials).

3. Which of the following are vector spaces over R?

i. V1 = {(x, y, z) ∈ R3 such that y + z = 0}.
ii. V1 = {(x, y, z) ∈ R3 such that y + z = 1}.
iii. V1 = {(x, y, z) ∈ R3 such that y ≥ 0}.
iv. V1 = {(x, y, z) ∈ R3 such that z = 0}.
4. Show that the dimension of the vector space of all m by n real matrices
over R is mn. [Hint: For m = 2, n = 3, the matrices
       
1 0 0 0 1 0 0 0 1 0 0 0
, , ,
0 0 0 0 0 0 0 0 0 1 0 0

   
0 0 0 0 0 0
,
0 1 0 0 0 1

form a basis for the space of all 2 by 3 real matrices. Verify this first.]

5. Prove that a subspace of a finite-dimensional vector space is finite-

dimensional.

6. Show that the vector space of real polynomials in X is infinite-

dimensional over R.

7. Find a basis and the dimension of the subspace of R4 spanned by the

vectors u1 = (1, 2, 2, 0), u2 = (2, 4, 0, 1) and u3 = (4, 8, 4, 1).
8. Find the dimension of the subspace of R3 spanned by v1 = (2, 3, 7),
v2 = (1, 0, −1), v3 = (1, 1, 2) and v4 = (0, 1, 3).
[Hint: v1 = 3v3 − v2 and v4 = v3 − v2 .]
9. State with reasons whether each of the following statements is true or
false:
A. A vector space V can have two disjoint subspaces.
B. Every vector space of dimension n has a subspace of dimension m
for each m ≤ n.
C. A two-dimensional vector space has exactly three subspaces.
D. In a vector space, any two generating (that is, spanning) subsets
are disjoint.
E. If n vectors of a vector space V span a subspace U of V , then
dim U = n.
 Algebraic Structures II (Vector Spaces and Finite Fields) 205

4.9 Solutions of Linear Equations

Consider the system of linear homogeneous equations:

X1 + 2X2 + 3X3 − X4 = 0
2X1 + X2 − X3 + 4X4 = 0
3X1 + 3X2 + 2X3 + 3X4 = 0 (4.2)
6X1 + 6X2 + 4X3 + 6X4 = 0.

These equations are called homogeneous because if (X1 , X2 , X3 , X4 ) is a solu-

tion of these equations, then so is (kX1 , kX2 , kX3 , kX4 ) for any scalar k.
Trivially (0, 0, 0, 0) is a solution. We express these equations in the matrix
form
AX = 0, (4.3)
where
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 2 3 −1 X1 0
⎢2 1 −1 4⎥ ⎢X2 ⎥ ⎢0⎥
A=⎢
⎣3
⎥, X=⎢ ⎥
⎣X3 ⎦ , and 0=⎢ ⎥
⎣0⎦ .
3 2 3⎦
6 6 4 6 X4 0
If X1 and X2 are any two solutions of Equation 4.3, then so is aX1 + bX2 for
scalars a and b since A(aX1 + bX2 ) = a(AX1 ) + b(AX2 ) = a · 0 + b · 0 = 0.
Thus, the set of solutions of Equation 4.3 is (as X ∈ R4 ) a vector subspace of
R4 , where 4 = the number of indeterminates in the Equations 4.2.
It is clear that the three elementary row operations performed on a sys-
tem of homogeneous linear equations do not alter the set of solutions of the
equations. Hence if A∗ is the row reduced echelon form of A, the solution sets
of AX = 0 and A∗ X = 0 are the same.
In Example 4.3, A∗ = A2 . Hence the equations A∗ X = 0 are
5
X1 − X3 + 3X4 = 0, and
3
7
X2 + X3 − 2X4 = 0.
3
These give
5
X1 = X3 − 3X4
3
7
X2 = − X3 + 2X4
3
so that ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
X1 5/3 −3
⎢X2 ⎥ ⎢−7/3⎥ ⎢ 2⎥
X=⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣X3 ⎦ = X3 ⎣ 1 ⎦ + X4 ⎣ 0⎦ .
X4 0 1
206 Discrete Mathematics

Thus, the space of solutions of AX = 0 is spanned by the two linearly inde-

pendent vectors
⎡ ⎤ ⎡ ⎤
5/3 −3
⎢−7/3⎥ ⎢2⎥
⎢ ⎥ ⎢ ⎥
⎣ 1 ⎦ and ⎣0⎦
0 1

and hence is of dimension 2. This number 2 corresponds to the fact that

X1 , X2 , X3 , and X4 are all expressible in terms of X3 and X4 . Here, X1 and
X2 correspond to the identity submatrix of order 2 of A∗ , that is, the rank of
A. Also, dimension 2 = 4 − 2 = (number of indeterminates) − (rank of A).
The general case is clearly similar where the system of equations is given by
ai1 X1 + · · · + ain Xn = 0, 1 ≤ i ≤ m. These are given by the matrix equation
AX = 0, where A is the m by n matrix (aij ) of coefficients and
⎡ ⎤
X1
⎢ ⎥
X = ⎣ ... ⎦ ;
Xn

we state it as a theorem.

Theorem 4.2. The solution space of a system of homogeneous linear equa-

tions is of dimension n − r, where n is the number of unknowns and r is the
rank of the matrix A of coefficients.

4.10 Solutions of Non-homogeneous Linear Equations

A system of non-homogeneous linear equations is of the form

a11 X1 + a12 X2 + · · · + a1n Xn = b1

.. .. .. ..
. . . .
am1 X1 + am2 X2 + · · · + amn Xn = bm ,

with not all bi being zero.

These m equations are equivalent to the single matrix equation

AX = B,

where A = (aij ) is an m by n matrix and B is a non-zero column vector of

length m.
 Algebraic Structures II (Vector Spaces and Finite Fields) 207

It is possible that such a system of equations has no solution at all. For

example, consider the system of equations

X 1 − X2 + X3 = 2
X1 + X 2 − X 3 = 0
3X1 = 6.

From the last equation, we get X1 = 2. This, when substituted in the first
two equations, yields −X2 + X3 = 0, X2 − X3 = −2 which are mutually
contradictory. Such equations are called inconsistent equations.
When are the equations represented by AX = B consistent?

Theorem 4.3. The equations AX = B are consistent if and only if B belongs

to the column space of A.

Proof. The equations are consistent iff there exists a vector

⎡ ⎤
α1
⎢ .. ⎥
X0 = ⎣ . ⎦
αn

such that AX0 = B. But this happens iff α1 C1 + · · · + αn Cn = B, where

C1 , . . . , Cn are the column vectors of A, that is, iff B belongs to the column
space of A.

Corollary 4.3. The equations represented by AX = B are consistent iff rank

of A = rank of (A, B). [(A, B) denotes the matrix obtained from A by adding
one more column vector B at the end. It is called the matrix augmented by B].

Proof. By the above theorem, the equations are consistent iff B ∈

C1 , . . . , Cn . But this is the case, by Proposition 4.4, iff C1 , . . . , Cn  =
C1 , . . . , Cn , B. The latter condition is equivalent to, column rank of
A = column rank of (A, B), and consequently to, rank of A = rank of
(A, B).

We now ask the natural question. If the system AX = B is consistent, how

to solve it?

Theorem 4.4. Let X0 be any particular solution of the equation AX = B.

Then, the set of all solutions of AX = B is given by {X0 + U }, where U varies
over the set of solutions of the auxiliary equation AX = 0.

Proof. If AU = 0, then A(X0 + U ) = AX0 + AU = B + 0 = B, so that X0 + U

is a solution of AX = B.
Conversely, let X1 be an arbitrary solution of AX = B, so that AX1 = B.
Then, AX0 = AX1 = B gives that A(X1 − X0 ) = 0. Setting X1 − X0 = U ,
we get, X1 = X0 + U , and AU = 0.
208 Discrete Mathematics

As in the case of homogeneous linear equations, the set of solutions of

AX = B remains unchanged when we perform any finite number of elementary
row transformations on A, provided we take care to perform simultaneously
the same operations on the matrix B on the right. As before, we row-reduce
A to its echelon form A0 and the equation AX = B gets transformed into its
equivalent equation A∗ X0 = B0 .

4.11 LUP Decomposition

Definition 4.9. By an LUP decomposition of a square matrix A, we mean
an equation of the form
P A = LU (4.4)
where P is a permutation matrix, L a unit-lower triangular matrix, and U an
upper triangular matrix.
Recall that a square matrix M is upper(resp. lower) triangular if all its
entries below(resp. above) its principal diagonal are zero. It is unit upper(resp.
lower) triangular if it is (i) upper(resp. lower) triangular, and (ii) all the entries
of its principal diagonal are equal to 1. Hence all the matrices P, A, L and U
in Equation 4.4 must be of the same order.
Suppose we have determined matrices P, L and U so that Equation 4.4
holds. The equation AX = b is equivalent to (P A)X = P b in that both have
the same set of solutions X. This is because P −1 exists and hence (P A)X = P b
is equivalent to P −1 (P A)X = P −1 (P b), that is, AX = b. Now set P b = b .
Then, P AX = P b gives
LU X = b . (4.5)
Hence if we set U X = Y (a column vector), then (4.5) becomes LY = b .
We know from Section 4.10 as to how to solve LY = b . A solution Y of this
equation, when substituted in U X = Y, gives X, again by the same method.

4.11.1 Computing an LU decomposition

We first consider the case when A = (aij ) is a non-singular matrix of
order n. We begin by obtaining an LU decomposition for A; that is, an LUP
decomposition with P = In in Equation 4.4. The process by which we obtain
the LU decomposition for A is known as Gaussian elimination.
Assume that a11 = 0. We write
⎡ ⎤ ⎡ ⎤
a11 a12 . . . a1n a11 a12 . . . a1n
⎢ a21 a22 . . . a2n ⎥ ⎢ a21 ⎥
⎢ ⎥ ⎢ ⎥
A=⎢ . . . ⎥ =⎢ . ⎥,
⎣ .. .. .. ⎦ ⎣ .. A ⎦
an1 an2 ... ann an1
 Algebraic Structures II (Vector Spaces and Finite Fields) 209

where A is a square matrix of order n − 1. We can now factor A as

⎡ ⎤⎡ ⎤
1 0 ... 0 a11 a12 ... a1n
⎢ a21 /a11 ⎥⎢⎢ 0 ⎥
⎥
⎢ ⎥⎢ . ⎥
⎢ .. ⎥⎢ . vwt ⎥
⎣ . In−1 ⎦⎣ . A − ⎦
a11
an1 /a11 0

where ⎡ ⎤
a21
⎢ ⎥
v = ⎣ ... ⎦ and wt = (a12 . . . a1n ).
an1

Note that vwt is also a matrix of order n − 1. The matrix A1 =

A − (vwt /a11 ) is called the Schur complement of A with respect to a11 .


We now recursively find an LU decomposition of A. If we assume that

A1 = L U  , where L is unit lower-triangular and U  is upper-triangular, then
  
1 0 a11 wt
A=
v/a11 In−1 0 A1
  
1 0 a11 wt
= (4.6)
v/a11 In−1 0 L U 
  t

1 0 a11 w
=
v/a11 L 0 U
= LU,

where    
1 0 a11 wt
L= , and U= .
v/a11 L 0 U
The validity of the two middle equations on the right of Equation 4.6 can be
verified by routine block multiplication of matrices (cf: Section 3.3.1). This
method is based on the supposition that a11 and all the leading entries of
the successive Schur complements are all non-zero. If a11 is zero, we inter-
change the first row of A with a subsequent row having a non-zero first entry.
This amounts to premultiplying both sides by the corresponding permutation
matrix P yielding the matrix P A on the left. We now proceed as with the case
when a11 = 0. If a leading entry of a subsequent Schur complement is zero,
once again we make interchanges of rows—not just the rows of the relevant
Schur complement, but the full rows obtained from A. This again amounts to
premultiplication by a permutation matrix. Since any product of permutation
matrices is a permutation matrix, this process finally ends up with a matrix
P  A, where P  is a permutation matrix of order n.
We now present two examples, one to obtain the LU decomposition when
it is possible and another to determine the LUP decomposition.
210 Discrete Mathematics

Example 4.4
Find the LU decomposition of
⎡ ⎤
2 3 1 2
⎢4 7 4 7⎥
A=⎢ ⎣2 7 13 16⎦ .
⎥

6 10 13 15

Here, ⎡ ⎤
4
a11 = 2, v = ⎣2⎦ , wt = [3, 1, 2] .
6
Therefore
⎡ ⎤ ⎡ ⎤
2 6 2 4
v/a11 = ⎣1⎦ , and so vwt /a11 = ⎣3 1 2⎦ ,
3 9 3 6

where wt denotes the transpose of w.

Hence the Schur complement of A is
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
7 4 7 6 2 4 1 2 3
A1 = ⎣ 7 13 16⎦ − ⎣3 1 2⎦ = ⎣4 12 14⎦ .
10 13 15 9 3 6 1 10 9

Now the Schur complement of A1 is

         
12 14 4  12 14 8 12 4 2
A2 = − 2, 3 = − = .
10 9 1 10 9 2 3 8 6

This gives the Schur complement of A2 as

A3 = [6] − [2][2] = [2]

= [1][2] = L3 U3 ,

where L3 = (1) is lower unit triangular and U3 = (2) is upper triangular.

Tracing back we get,
  
1 0 4 w2t
A2 =
v2 /a11 L3 0 U3
  
1 0 4 2
= = L2 U2 .
2 1 0 2
 Algebraic Structures II (Vector Spaces and Finite Fields) 211

This gives, ⎡ ⎤⎡ ⎤
1 0 0 1 2 3
A1 = ⎣ 4 1 0 ⎦⎣ 0 4 2 ⎦.
1 2 1 0 0 2
Consequently,
⎡ ⎤⎡ ⎤
1 0 0 0 2 3 1 2
⎢ 2 1 0 ⎥ ⎢
0 ⎥⎢ 0 1 2 3 ⎥
A=⎢
⎣ 1
⎥ = LU,
4 1 0 ⎦⎣ 0 0 4 2 ⎦
3 1 2 1 0 0 0 2

where
⎡ ⎤
1 0 0 0
⎢ 2 1 0 0 ⎥
⎥
L=⎢ is unit lower-triangular, and
⎣ 1 4 1 0 ⎦
3 1 2 1
⎡ ⎤
2 3 1 2
⎢ 0 1 2 3 ⎥
U =⎢
⎣ 0
⎥ is upper-triangular.
0 4 2 ⎦
0 0 0 2

Note that A is non-singular as L and U are.

Example 4.5
Find the LUP decomposition of
⎡ ⎤
2 3 1 2
⎢4 6 4 7⎥
A=⎢ ⎣2
⎥.
7 13 16⎦
6 10 13 15
Suppose we proceed as before: The Schur complement of A is
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
6 4 7 2  6 4 7 6 2 4
A1 = ⎣ 7 13 16⎦ − ⎣1⎦ 3 1 2 = ⎣ 7 13 16⎦ − ⎣3 1 2⎦
10 13 15 3 10 13 15 9 3 6
⎡ ⎤
0 2 3
= ⎣4 12 14⎦ .
1 10 9
212 Discrete Mathematics

Since the leading entry is zero, we interchange the first row of A1 with
some other row. Suppose we interchange the first and third rows of A1 .
This amounts to considering the matrix P A instead of A1 , where
⎡ ⎤
1 0 0 0
⎢0 0 0 1⎥
P =⎢ ⎣0 0 1 0⎦ .
⎥

0 1 0 0

Note that the first row of A1 corresponds to the second row of A and
the last row of A1 to the fourth row of A. This means that the Schur
complement of P A (instead of A) is
⎡ ⎤
1 10 9
A1 = ⎣4 12 14⎦ .
0 2 3

We now proceed with A1 as before. The Schur complement of A1 is

         
12 14 4  12 14 40 36 −28 −22
A2 = − 10 9 = − = .
2 3 0 2 3 0 0 2 3

The Schur complement of A2 is

 2  11
 10
A3 = [3] − −28 [−22] = 3 − 7 = 7 = (1)(10/7) = L3 U3 ,

where L3 = [1] is unit lower triangular, and U3 = [10/7] is upper trian-

gular. Hence   
1 0 −28 −22
A2 = −2 10 .
28 1 0 7

This gives ⎡ ⎤⎡ ⎤
1 0 0 1 10 9
A1 = ⎣4 1 0⎦ ⎣0 −28 −22⎦ .
−1 10
0 14 1 0 0 7

Thus, ⎡ ⎤⎡ ⎤
1 0 0 0 2 3 1 2
⎢3 1 0 0⎥ ⎢0 1 10 9 ⎥
⎢ ⎥⎢ ⎥ = L U,
⎣1 4 1 0⎦ ⎣0 0 −28 −22⎦
−1 10
2 0 14 1 0 0 0 7

where L and U are the first and second matrices in the product. Notice
that we have interchanged the second and fourth rows of A while com-
puting L . Interchanging the second and fourth rows of L , we get
⎡ ⎤⎡ ⎤
1 0 0 0 2 3 1 2
⎢2 0 −1 1⎥ ⎢0 1 10 9 ⎥
A=⎢ 14 ⎥⎢
⎣1 4 1 0⎦ ⎣0 0 −28 −22⎦ = LU.
⎥
10
3 1 0 0 0 0 0 7
 Algebraic Structures II (Vector Spaces and Finite Fields) 213

We have assumed, to start with, that A is a non-singular matrix. Even if A

is invertible, it is possible
 that A has no LU decomposition. For example, the
non-singular matrix 0 1 0
1 has no LU decomposition. It is easy to check that
if A is singular, it is possible that not only a column of a Schur complement
but even the full column corresponding to it may be zero. In that case, we
have to interchange columns. This would result in a matrix of the form AP
rather than P A. It is also possible that we get the form P1 AP2 where P1 and
P2 are both permutation matrices.

Example 4.6
Solve the system of linear equations using the LUP decomposition
method:

2X1 + 3X2 + X3 + 2X4 = 10

4X1 + 6X2 + 4X3 + 7X4 = 25
2X1 + 7X2 + 13X3 + 16X4 = 40 (4.7)
6X1 + 10X2 + 13X3 + 15X4 = 50.

These equations are equivalent to AX = B, where A is the matrix

of Example 4.5 and
⎡ ⎤
10
⎢25⎥
B=⎣ ⎥⎢ .
40⎦
50

If we interchange any pair of rows of (A|B), it amounts to inter-

changing the corresponding equations. However, this will in no way
alter the solution. Hence the solutions of Equation 4.7 are the same
as solutions of P AX = P B, where P is the permutation matrix
obtained in Example 4.5. Thus, the solutions are the same as the
solutions of LU X = B, where L and M are again those obtained in
Example 4.5.
Now set U X = Y so that the given equations become LY = B,
where
⎡ ⎤
Y1
⎢Y2 ⎥
Y =⎢ ⎥
⎣Y3 ⎦ .
Y4
214 Discrete Mathematics

This gives ⎡ ⎤
1 0 0 0 ⎡Y ⎤ ⎡10⎤
1
⎢ −1 ⎥ ⎥ ⎢50⎥
⎢2 0 1⎥ ⎢ Y2⎥ ⎢ ⎥
⎢ 14 ⎥⎢⎣ ⎦ = ⎣40⎦ .
⎣1 4 1 0⎦ Y3
3 1 0 0 Y4 25

These are equivalent to

3Y1 = 10
1
2Y1 − Y3 + Y4 = 25
14
Y1 + 4Y2 + Y3 = 40
3Y1 + Y2 = 50,

and we get Y1 = 10, Y2 = 20, Y3 = −50 and Y4 = 10/7. Substituting

these values in U X = Y , we get
⎡ ⎤ ⎡ ⎤
2 3 1 2 ⎡X ⎤ 10
1
⎢0 1 10 9 ⎥ ⎢ 20 ⎥
⎢ ⎥ ⎢X2 ⎥
⎥=⎢ ⎥
⎢0 0 −28 −22⎥ ⎢ ⎣ ⎦ ⎢−50⎥ .
⎣ ⎦ X3 ⎣ ⎦
10 10
0 0 0 X4
7 7
These give

2X1 + 3X2 + X3 + 2X4 = 10

X2 + 10X3 + 9X4 = 20
−28X3 − 22X4 = −50
10 10
X4 =
7 7
Solving backward, we get X1 = 2, X2 = X3 = X4 = 1.

4.12 Exercises
1. Examine if the following equations are consistent.

X1 + X2 + X3 + X4 = 0
2X1 − X2 + 3X3 + 4X4 = 1
3X1 + 4X3 + 5X4 = 2.
 Algebraic Structures II (Vector Spaces and Finite Fields) 215

2. Solve the system of homogeneous linear equations:

4X1 + 4X2 + 3X3 − 5X4 = 0

X1 + X2 + 2X3 − 3X4 = 0
2X1 + 2X2 − X3 = 0
X1 + X2 + 2X3 − 2X4 = 0.

Show that the solution space is of dimension 2.

3. Solve:

X1 + X2 + X3 + X4 = 0
X1 + 3X2 + 2X3 + 4X4 = 0
2X1 + X3 − X4 = 0.

4. Solve by finding LUP decomposition:

i.

2X1 + 3X2 − 5X3 + 4X4 = −8

3X1 + X2 − 4X3 + 5X4 = −8
7X1 + 3X2 − 2X3 + X4 = 56
4X1 + X2 − X3 + 3X4 = 20.

ii.

3X1 − 2X2 + X3 = 7
X1 + X2 + X3 = 12
−X1 + 4X2 − X3 = 3.

iii.

2X1 + 4X2 − 5X3 + X4 = 8

4X1 + −5X3 − X4 = 16
−4X1 + 2X2 + X4 = −5
6X1 + 4X2 − 10X3 + 7X4 = 13.

4.13 Finite Fields

In this section, we discuss the basic properties of finite fields. Finite fields
are fundamental to the study of codes and cryptography.
Recall that a field F is finite if its cardinality |F | is finite. |F | is the order
of F . The characteristic of a finite field F , as seen in Section 3.23, is a prime
216 Discrete Mathematics

number p, and the prime field P of F is a field of p elements. P consists of

the p elements 1F , 2 · 1F = 1F + 1F , . . . , p · 1F = 0F . Clearly, F is a vector
space over P . If the dimension of the finite field F over P is n, then n is finite.
Hence F has a basis {u1 , . . . , un } of n elements over P . This means that each
element v ∈ F is a unique linear combination of u1 , . . . , un , say,
v = α1 u1 + α2 u2 + · · · + αn un , αi ∈ P, 1 ≤ i ≤ n.
For each i, αi can take |P | = p values, and so there are p · p · · · (n times) = pn
distinct elements in F . Thus, we have proved the following result.

Theorem 4.5. The order of a finite field is a power of a prime number.

Finite fields are known as Galois fields after the French mathematician
Évariste Galois (1811–1832) who first studied them. A finite field of order q
is denoted by GF (q).
We now look at the converse of Theorem 4.5. Given a prime power pn
(where p is a prime), does there exist a field of order pn ? The answer to this
question is in the affirmative. We give below two different constructions that
yield a field of order pn .
Theorem 4.6. Given pn (where p is a prime), there exists a field of pn
elements.
n
Construction 1: Consider the polynomial X p − X ∈ Zp [X] of degree pn .
(Recall that Zp [X] stands for the ring of polynomials in X with coefficients
from the field Zp of p elements.) The derivative of this polynomial is
n
pn X p −1
− 1 = −1 ∈ Zp [X],
n n
and is therefore relatively prime to X p − X. Hence the pn roots of X p − X
are all distinct. (Here, though no concept of the limit is involved, the notion
of the derivative has been employed as though it is a real polynomial.) It is
known [12] that the roots of this polynomial lie in an extension field K ⊃ Zp .
n
K is also of characteristic p. If a and b are any two roots of X p − X, then
n n
ap = a, and bp = b.
Now by Theorem 3.14 n n n
(a ± b)p = ap ± bp ,
and, by the commutativity of multiplication in K,
n n n
ap bp = (ab)p ,
n
and so a ± b and ab are also roots of X p − X. Moreover, if a is a non-zero
n n n
root of X p − X, then so is a−1 since (a−1 )p = (ap )−1 = a−1 . Also the
associative and distributive laws are valid for the set of roots since they are
n
all elements of the field K. Finally 0 and 1 are also roots of X p − X. In
n pn
other words, the p roots of the polynomial X − X ∈ Zp [X] form a field of
order pn .
 Algebraic Structures II (Vector Spaces and Finite Fields) 217

Construction 2: Let f (X) = X n +a1 X n−1 +· · ·+an ∈ Zp [X] be a polynomial of

degree n irreducible over Zp . The existence of such an irreducible polynomial
(with leading
 coefficient 1) of degree n is guaranteed by a result (see [12]) in
Algebra. Let F denote the ring of polynomials in Zp [X] reduced modulo f (X)
(that is, if g(X) ∈ Zp [X], divide g(X) by f (X) and take the remainder g1 (X)
which is 0 or of degree less than n). Then, every non-zero polynomial in F is a
polynomial of Zp [X] of degree at most n − 1. Moreover, if a0 X n−1 + · · · + an−1
and b0 X n−1 + · · · + bn−1 are two polynomials in F of degrees at most n − 1,
and if they are equal, then,

(a0 − b0 )X n−1 + · · · + (an−1 − bn−1 )

is the zero polynomial of F , and hence is a multiple of f (X) in Zp [X]. As

degree of f is n, this is possible only if ai = bi , 0 ≤ i ≤ n − 1. Now if
a0 X n−1 + a1 X n−2 + · · · + an−1 is any polynomial of F , ai ∈ Zp and hence ai
has p choices. Hence the number of polynomials of the form

a0 X n−1 + · · · + an−1 ∈ Zp [X]

is pn .
We now show that F is a field. Clearly, F is a commutative ring with unit
element 1(= 0 · X n−1 + · · · + 0 · X + 1). Hence we need only verify that if
a(X) ∈ F is not zero, then there exists b(X) ∈ F with a(X)b(X) = 1. As
a(X) = 0, and f (X) is irreducible over Zp , the gcd (a(X), f (X)) = 1. So by
Euclidean algorithm, there exist polynomials C(X) and g(X) in Zp [X] such
that

a(X)C(X) + f (X)g(X) = 1 (4.8)

in Zp [X]. Now there exists C1 (X) ∈ F with C1 (X) ≡ C(X)(mod f (X)). This
means that there exists a polynomial h(X) in Zp [X] with C(X) − C1 (X) =
h(X)f (X), and hence C(X) = C1 (X)+h(X)f (X). Substituting this in Equa-
tion 4.8 and taking modulo f (X), we get, a(X)C1 (X) = 1 in F . Hence a(X)
has C1 (X) as inverse in F . Thus, every non-zero element of F has a multi-
plicative inverse in F , and so F is a field of pn elements.

We have constructed a field of pn elements in two different ways—one, as the

n
field of roots of the polynomial X p − X ∈ Zp [X], and the other, as the field
of polynomials in Zp [X] reduced modulo the irreducible polynomial f (X) of
degree n over ZP . Essentially, there is not much of a difference between the
two constructions, as our next theorem shows (for a proof, see [18]).

Theorem 4.7. Any two finite fields of the same order are isomorphic under
a field isomorphism.
218 Discrete Mathematics

Example 4.7
Take p = 2 and n = 3. The polynomial X 3 + X + 1 of degree 3
is irreducible over Z2 . (If it is reducible, one of the factors must be of
degree 1, and it must be either X or X + 1 = X − 1 ∈ Z2 [X]. But 0 and
1 are not roots of X 3 + X + 1 ∈ Z2 [X].) The 23 = 8 polynomials over
Z2 reduced modulo X 3 + X + 1 are

0, 1, X, X + 1, X 2 , X 2 + 1, X 2 + X, X 2 + X + 1

and they form a field. (Note that X3 = X + 1, X 3

 + X = 1 and X +
3

X + 1 = 0.) We have, for instance, X 2 + X + 1 + (X + 1) = X 2 and

(X 2 + 1)(X + 1) = X 3 + X 2 + X + 1 = X 2 . Also (X + 1)2 = X 2 + 1.

We know that if F is a field, the set F ∗ of non-zero elements of F is a

group. In the case when F is a finite field, F ∗ has an additional algebraic
structure.

Theorem 4.8. If F is a finite field, F ∗ (the set of non-zero elements of F)

is a cyclic group.

Proof. We know (as F is a field), F ∗ is a group. Hence, we need only show

that F ∗ is generated by a single element.
Let α be an element of the group F ∗ of maximum order, say, k. Necessarily,
k ≤ q −1, where q = |F |. Choose β ∈ F ∗ , β = α, 1. Let o(β) (= the order of β)
= l. Then, l > 1. We first show that l|k. Now o(β (k,l) ) = l/(k, l), where (k, l)
denotes the gcd of k and l. Further, as (o(α), o(β (k,l) )) = (k, (l/(k, l))) = 1,
we have o(αβ (k,l) ) = o(α) · o(β (k,l) ) = k · (l/(k, l)) = [k, l], the lcm of k and l.
But, by our choice, the maximum order of any element of F ∗ is k. Therefore
[k, l] = k which implies that l|k. But l = o(β). Therefore β k = 1. Thus, for
each of the q − 1 elements z of F ∗ , z k = 1 and so z is a root of xk − 1. This
means that, as |F ∗ | = q − 1, k = q − 1. Thus, o(α) = q − 1 and so F ∗ is the
cyclic group generated by α.

Definition 4.10. (Primitive element, Minimal polynomial)

i. A primitive element of a finite field F is a generator of the cyclic

group F ∗ .
ii. A monic polynomial is a polynomial with leading coefficient 1. For exam-
ple, X 2 + 2X + 1 ∈ R[X] is monic while 2X 2 + 1 ∈ R[X] is not.
iii. Let F be a finite field with prime field P . A primitive polynomial in
F [X] over P is the minimal polynomial in P [X] of a primitive element
of F . A minimal polynomial in P [X] of an element α ∈ F is a monic
polynomial of least degree in P [X] having α as a root.
 Algebraic Structures II (Vector Spaces and Finite Fields) 219

Clearly the minimal polynomial of any element of F in P [X] is irreducible

over P .
Let α be a primitive element of F , and f (X) = X n + a1 X n−1 + · · · +
an ∈ P[X] be the primitive polynomial of α. Then, any polynomial in P [α] of
degree n or more can be reduced to a polynomial in P [α] of degree at most
n − 1. Moreover, no two distinct polynomials of P [α] of degree at most n − 1
can be equal; otherwise, α would be a root of a polynomial of degree less than
n over P . Hence all the polynomials of the form

a0 + a1 α + · · · + an−1 αn−1 ,

in P [α] are all distinct and so |P [α]| = pn where p = |P |. These pn elements

constitute a subfield F  of F and α ∈ F  . But then F ⊆ F  and hence F = F  .
Thus, |F | = |F  | = pn . As α is a primitive element of F, this means that
 n

F = 0; α, α2 , . . . , αp −1 = 1 .

Example 4.8
Consider the polynomial X 4 + X + 1 ∈ Z2 [X]. This is irreducible
over Z2 (check that it can have no linear or quadratic factor in Z2 [X]).
Let α be a root (in an extension field of Z2 ) of this polynomial so that
α4 + α + 1 = 0. This means that α4 = α + 1.
We now prove that α is a primitive element of a field of 16 elements
over Z2 by checking that the 15 powers α, α2 , . . . , α15 are all distinct
and that α15 = 1. Indeed, we have

α1 = α
α2 = α2
α3 = α3
α4 = α + 1(as α4 + α + 1 = 0 → α4 = α + 1, as1 = −1)
α5 = αα4 = α(α + 1) = α2 + α
α6 = αα5 = α3 + α2
α7 = αα6 = α4 + α3 = α3 + α + 1
α8 = αα7 = α4 + (α2 + α) = (α + 1) + (α2 + α) = α2 + 1
α9 = αα8 = α3 + α
α10 = αα9 = α4 + α2 = α2 + α + 1
α11 = αα10 = α3 + α2 + α
α12 = αα11 = α4 + (α3 + α2 ) = (α + 1) + (α3 + α2 ) = α3 + α2 + α + 1
220 Discrete Mathematics

α13 = αα12 = α4 + (α3 + α2 + α) = (α + 1) + (α3 + α2 + α)

= α3 + α2 + 1
α14 = αα13 = α4 + (α3 + α) = (α + 1) + (α3 + α) = α3 + 1
α15 = αα14 = α4 + α = (α + 1) + α = 1
 
Thus, F = {0} ∪ F ∗ = 0, α, α2 , . . . , α15 = 1 and so α is a primitive
element of F = GF (24 ).

We observe that a polynomial irreducible over a field F need not be primi-

tive over F . For instance, the polynomial f (X) = X 4 +X 3 +X 2 +X+1 ∈ Z2 [X]
is irreducible over Z2 but it is not primitive. To check that f (X) is irreducible,
verify that F (X) has no linear or quadratic factor over Z2 . Next, for any root α
of f (X),
 check that α5 = 1 so that o(α) < 15, and f (X) is not primitive over
Z2 . Recall that if f (X) were a primitive  polynomial, some root of f (X)
should be a primitive element of GF (24 ).

4.14 Factorization of Polynomials Over Finite Fields

n
Let α be a primitive element of the finite
 field F = GF (p ), where p is
∗ 2 pn −1 n
a prime. Then, F = α, α , . . . , α = 1 , and for any x ∈ F , xp = x.
Hence for each i, 1 ≤ i ≤ pn − 1,
n
(αi )p = αi .
t+1
This shows that there exists a least positive integer t such that αi·p = αi .
Then, set
 
Ci = i, pi, p2 i, . . . , pt i , 0 ≤ i ≤ pn − 1.

The sets Ci are called the cyclotomic cosets modulo p defined with respect to
F and α. Now, corresponding to the coset Ci , 0 ≤ i ≤ pn − 1, consider the
polynomial
2 t
fi (X) = (X − αi )(X − αi·p )(X − αi·p ) · · · (X − αi·p ).
t
The coefficients of fi are elementary symmetric functions of αi , αip , . . . , αip
and if β denotes any of these coefficients, then β satisfies the relation β p = β.
Hence β ∈ Zp and fi (X) ∈ Zp [X] for each i, 0 ≤ i ≤ pn −1. Each element of Ci
determines the same cyclotomic coset, that is, Ci = Cip = Cip2 = · · · = Cipt .
n
/ Ci , Ci ∩ Cj = φ. This gives a factorization of X p − X into
Moreover, if j ∈
 Algebraic Structures II (Vector Spaces and Finite Fields) 221
n n
irreducible factors over Zp . In fact, X p − X = X(X p −1
− 1), and

n n
Xp −1
− 1 = (x − α)(X − α2 ) · · · (X − αp −1
)
⎛ ⎞
 
= ⎝ (X − αj )⎠ ,
i j∈Ci

where the first product is taken over all the distinct cyclotomic cosets. Fur-
thermore, each polynomial fi (X) is irreducible over Zp as shown below. To
see this, assume that

g(X) = a0 + a1 X + · · · + ak X k ∈ F [X].
 p
Then, g(X) = ap0 + ap1 X p + · · · + apk (X k )p (Refer Section 3.23)
= a0 + a1 X p + · · · + ak X kp
= g(X p ).

p
Consequently, if β is a root of g, g(β) = 0, and therefore 0 = (g(β)) = g (β p ),
that is, β p is also a root of g(X). Hence if j ∈ Ci and αj is a root of fi (X),
then all the powers αk , k ∈ Ci , are roots of fi (X). Hence any non-constant
irreducible factor of fi (X) over Zp must contain all the terms (X − αj ), j ∈ Ci
as factors. In other words, g(X) is irreducible over Zp .
Thus, the determination of the cyclotomic cosets yields a simple device to
n
factorize X p − X into irreducible factors over ZP . We illustrate this fact by
an example.

Example 4.9
4
Factorize X 2 − X into irreducible factors over Z2 .
Let α be a primitive element of the field GF (24 ). As a primitive
polynomial of degree 4 over Z2 having α as a root, we can take (see
Example 4.8) X 4 + X + 1.
The cyclotomic cosets modulo 2 w.r.t. GF (24 ) and α are

C0 = {0}
 
C1 = 1, 2, 22 = 4, 23 = 8 (Note: 24 = 16 ≡ 1( mod 15))
C3 = {3, 6, 12, 9}
C5 = {5, 10}
C7 = {7, 14, 13, 11} .
222 Discrete Mathematics

Note that C2 = C1 = C4 , and so on. Thus,

15

16 15
 
X − X = X(X − 1) = X X − αi (4.9)
i=1
  
       
0 i i
=X X −α X −α X −α
i∈C1 i∈C3
  
     
i i
× X −α X −α
i∈C5 i∈C7
  
= X (X + 1) X 4 + X + 1 X 4 + X 3 + X 2 + X + 1
  
× X2 + X + 1 X4 + X3 + 1 .

In computing the products, we have used the relation α4 + α + 1 = 0,

that is, α4 = α + 1. Hence, for instance,
     
X − αi = X − α5 X − α10
i∈C5
 
= X 2 − α5 + α10 X + α15
   
= X 2 + α2 + α + α2 + α + 1 X + α15
(see Example 4.8)
= X 2 + X + 1.

The six factors on the right of Equation 4.9 are all irreducible over
Z2 . The minimal polynomials of α, α3 and α7 are all of degree 4 over
Z2 . However, while α and α7 are primitive elements of GF (24 ) (so that
the polynomials X 4 + X + 1 and X 4 + X 3 + 1 are primitive), α3 is not
(even though its minimal polynomial is also of degree 4).

Primitive polynomials are listed in [19]

4.15 Exercises
1. Construct the following fields:
GF (24 ), GF (25 ) and GF (32 ).
2. Show that GF (25 ) has no GF (23 ) as a subfield.
3 5
3. Factorize X 2 + X and X 2 + X over Z2 .
 Algebraic Structures II (Vector Spaces and Finite Fields) 223
2
4. Factorize X 3 − X over Z3 .
5. Using Theorem 4.8, prove Fermat’s Little Theorem that for any prime p,
ap−1 ≡ 1 (mod p), if a ≡ 0 (mod p).

4.16 Mutually Orthogonal Latin Squares

In this section, we show, as an application of finite fields, the existence of
n − 1 mutually orthogonal Latin Squares (MOLSs) of order n.
A Latin square of order n is a double array L of n rows and n columns
in which the entries belong to a set S of n elements such that no two entries
of the same row or column of L are equal. Usually, we take S to be the set
{1, 2, . . . , n} but this is not always essential.
For example, ⎡ ⎤
  1 2 3
1 2
and ⎣2 3 1⎦
2 1
3 1 2
are Latin squares of orders 2 and 3, respectively. Let L1 = (aij ), and L2 = (bij )
be two Latin squares of order n with entries in S. We say that L1 and L2 are
orthogonal Latin squares if the n2 ordered pairs (aij , bij ) are all distinct. For
example, ⎡ ⎤ ⎡ ⎤
1 2 3 1 2 3
L1 = ⎣2 3 1⎦ , and L2 = ⎣3 1 2⎦
3 1 2 2 3 1
are orthogonal Latin squares of order 3 since the nine ordered pairs (1, 1),
(2, 2), (3, 3); (2, 3), (3, 1), (1, 2); (3, 2), (1, 3), (2, 1) are all distinct. However, if
M1 = [ 12 21 ] and M2 = [ 21 12 ], then the four ordered pairs (1, 2), (2, 1), (2, 1) and
(1, 2) are not all distinct. Hence M1 and M2 are not orthogonal. The study of
orthogonal Latin squares started with Euler, who had proposed the following
problem of 36 officers (See [6]). The problem asks for an arrangement of 36
officers of 6 ranks and from 6 regiments in a square formation of size 6 by 6.
Each row and each column of this arrangement is to contain only one officer
of each rank and only one officer from each regiment. We label the ranks and
the regiments from 1 through 6, and assign to each officer an ordered pair of
integers in 1 through 6. The first component of the ordered pair corresponds
to the rank of the officer and the second component his regiment. Euler’s
problem then reduces to finding a pair of orthogonal Latin squares of order 6.
Euler conjectured in 1782 that there exists no pair of orthogonal Latin squares
of order n ≡ 2(mod 4). Euler himself verified the conjecture for n = 2, while
Tarry in 1900 verified it for n = 6 by a systematic case-by-case analysis. But
the most significant result with regard to the Euler conjecture came from Bose,
Shrikande and Parker who disproved the Euler conjecture by establishing that
224 Discrete Mathematics

if n ≡ 2(mod 4) and n > 6, then there exists a pair of orthogonal Latin squares
of order n (see also [17,35]).
A set {L1 , . . . , Lt } of t Latin squares of order n on S is called a set of MOLS
(Mutually Orthogonal Latin squares) if Li and Lj are orthogonal whenever
i = j. It is easy to see [17] that the number t of MOLS of order n is bounded
by n − 1. Further, any set of n − 1 MOLS of order n is known to be equivalent
to the existence of a finite projective plane of order n [17]. A long standing
conjecture is that if n is not a prime power, then there exists no complete set
of MOLS of order n.
We now show that if n is a prime power, there exists a set of n − 1 MOLS
of order n. (Equivalently, this implies that there exists a projective plane of
any prime power order, though we do not prove this here) (see [17] for more
details on finite projective planes).

Theorem 4.9. Let n = pk , where p is a prime and k is a positive integer.

Then, for n ≥ 3, there exists a complete set of MOLS of order n.
Proof. By Theorem 4.6, we know that there exists a finite field GF (pk ) =
GF (n) = F , say. Denote the elements of F by a0 = 0, a1 = 1, a2 , . . . , an−1 .
Define the n − 1 matrices A1 , . . . , An−1 of order n by

At = (atij ), 0 ≤ i, j ≤ n − 1; and 1 ≤ t ≤ n − 1,

atij = at ai + aj (here atij stands for the (i, j)-th entry of the matrix At ). The
entries atij are all elements of the field F . We claim that each At is a Latin
square. Suppose, for instance, two entries of some i-th row of At , say atij and
atil are equal. This implies that

at ai + aj = at ai + al ,

and hence aj = al . Consequently j = l. Thus, all the entries of the i-th row
of At are distinct. For a similar reason, no two entries of the same column of
At are equal. Hence At is a Latin square for each t.
We next claim that {A1 , . . . , An−1 } is a set of MOLS. Suppose 1 ≤ r <
u ≤ n − 1. Then, Ar and Au are orthogonal. For suppose that
 r u  r 
aij , aij = ai j  , aui j  . (4.10)

This means that

ar ai + aj = ar ai + aj  ,
and au ai + aj = au ai + aj  .

Subtraction gives
(ar − au )ai = (ar − au )ai
and hence, as ar = au , ai = ai . Consequently, i = i and j = j  . Thus, Ar
and Au are orthogonal.
Chapter 5
Introduction to Coding Theory

ASCII (American Standard Code for Information Interchange) was

designed by the American Standards Association (ASA) subcom-
mittee, in order to facilitate the general interchange of information
among information processing systems, communication systems, and
associated equipment. An 8-bit set was considered but the need for
more than 128 codes in general applications was not yet evident.

ASA subcommittee

5.1 Introduction
Coding theory has its origin in communication engineering. With Shan-
non’s seminal paper of 1948 [22], it has been greatly influenced by mathematics
with a variety of mathematical techniques to tackle its problems. Algebraic
coding theory uses a great deal of matrices, groups, rings, fields, vector spaces,
algebraic number theory and, not to speak of, algebraic geometry. In algebraic
coding, each message is regarded as a block of symbols taken from a finite
alphabet. On most occasions, this alphabet is Z2 = {0, 1}. Each message is
then a finite string of 0s and 1s. For example, 00110111 is a message. Usu-
ally, the messages get transmitted through a communication channel. It is
quite possible that such channels are subjected to noises, and consequently,
the messages get changed. The purpose of an error correcting code is to add
redundancy symbols to the message, based of course on some rule so that the
original message could be retrieved even though it is garbled. Each message
is also called a codeword and the set of codewords is a code.
Any communication channel looks as in Figure 5.1. The first box of the
channel indicates the message. It is then transmitted to the encoder, which
adds a certain number of redundancy symbols. In Figure 5.1, these redun-
dancy symbols are 001 which when added to the message 1101 give the coded
message 1101001. Because of channel noise, the coded message gets distorted
and the received message is 0101001. This message then enters the decoder.
The decoder applies the decoding algorithm and retrieves the coded message
using the added redundancy symbols. From this, the original message is read
off in the last box (see Figure 5.1). The decoder has thus corrected a single
error, that is, error in one place.

225
226 Discrete Mathematics

Channel Received Original

Message Encoder message Decoder message
noise
1101 1101001 0101001 1101001 1101

FIGURE 5.1: Communication channel.

The efficiency of a code is the number of errors it can correct. A code

is perfect if it can correct all of its errors. It is k-error-correcting if it can
correct k or fewer errors. The aim of coding theory is to devise efficient codes.
Its importance lies in the fact that erroneous messages could prove to be
disastrous.
It is relatively easier to detect errors than to correct them. Sometimes,
even detection may prove to be helpful as in the case of a feedback channel,
that is, a channel that has a provision for retransmission of the messages.
Suppose the message 1111 is sent through a feedback channel. If a single error
occurs and the received message is 0111, we can ask for a feedback twice and
may get 1111 on both occasions. We then conclude that the received message
should have been 1111. Obviously, this method is not perfect as the original
message could have been 0011. On the other hand, if the channel is two-way,
that is, it can detect errors so that the receiver knows the places where the
errors have occurred and also contains the provision for feedback, then it can
prove to be more effective in decoding the received message.

5.2 Binary Symmetric Channels

One of the simplest channels is the binary symmetric channel (BSC). This
channel has no memory and it simply transmits two symbols 0 and 1. It has
the property that each transmitted symbol has the probability p(≤1/2) of
being received in error, so that the probability that a transmitted symbol is
received correctly is q = 1 − p. This is pictorially represented in Figure 5.2.
Before considering an example of a BSC, we first give the formal definition of
a code.

q
0 0

p p

1 1
q

FIGURE 5.2: Binary symmetric channel.

 Introduction to Coding Theory 227

Definition 5.1. A code C of length n over a finite field F is a set of vectors

in F n , the space of ordered n-tuples over F . Any element of C is called a
codeword of C.

As an example, the set of vectors, C = {10110, 00110, 11001, 11010} is a

code of length 5 over the field Z2 . This code C has 4 codewords.
Suppose a binary codeword of length 4 (that is, a 4-digit codeword) is sent
through a BSC with probability q = 0.9. Then, the probability that the sent
word is received correctly is q 4 = (0.9)4 = 0.6561.
We now consider another code, namely, the Hamming (7, 4)-code. (See
Section 5.5 below). This code has as its words the binary vectors 1000001,
0100011, 0010010, 0001111, of length 7 and all of their linear combinations
over the field Z2 .
The first four positions are information positions and the last three are
the redundancy positions. There are in all 24 = 16 codewords in the code. We
shall see later that this code can correct one error. Hence the probability that
a received vector yields the transmitted vector is q 7 + 7pq 6 , where the first
term corresponds to the case of no error and the term 7pq 6 corresponds to
a single error in each of the seven possible positions. If q = 0.9, q 7 + 7pq 6 =
0.4783 + 0.3720 = 0.8503, which is quite large compared to the probability
0.6561 arrived at earlier in the case of a BSC.
Hamming code is an example of a class of codes called Linear Codes. We
now present some basic facts about linear codes.

5.3 Linear Codes

Definition 5.2. An [n, k]-linear code C over a finite field F is a k-dimensional
subspace of F n , the space of ordered n-tuples over F .

If F has q elements, that is, F = GF (q), the [n, k]-code will have q k
codewords. The codewords of C are all of length n as they are n-tuples over F .
k is called the dimension of C. C is a binary code if F = Z2 .
A linear code C is best represented by any one of its generator matrices.

Definition 5.3. A generator matrix of a linear code C over F is a matrix

whose row vectors form a basis for C over F .

If C is an [n, k]-linear code over F , then a generator matrix of C is a k by

n matrix G over F whose row vectors form a basis for C.
For example, consider the binary code C1 with generator matrix
⎡ ⎤
1 0 0 1 1
G1 = ⎣0 1 0 1 0⎦ .
0 0 1 0 1
228 Discrete Mathematics

Clearly, all the three row vectors of G1 are linearly independent over Z2 . Hence
C1 has 23 = 8 codewords. The first three columns of G1 are linearly indepen-
dent over Z2 . Therefore, the first three positions of any codeword of C1 may
be taken as information positions, and the remaining two as redundancy posi-
tions. In fact, the positions corresponding to any three linearly independent
columns of G1 may be taken as information positions and the rest redundan-
cies. Now any word X of C1 is given by
X = x1 R1 + x2 R2 + x3 R3 , (5.1)
where x1 , x2 , x3 are all in Z2 and R1 , R2 , R3 are the three row vectors of G1
in order. Hence by Equation 5.1, X = (x1 , x2 , x3 , x1 + x2 , x1 + x3 ). If we take
X = (x1 , x2 , x3 , x4 , x5 ), we have the relations
x4 = x1 + x2 , and
(5.2)
x5 = x1 + x3 .
In other words, the first redundancy coordinate of any codeword is the sum
of the first two information coordinates of that word, while the next redun-
dancy coordinate is the sum of the first and third information coordinates.
Equations 5.2 are the parity-check equations of the code C1 . They can be
rewritten as
x1 + x2 − x4 = 0, and
(5.3)
x1 + x3 − x5 = 0.
In the binary case, Equations 5.3 become
x1 + x2 + x4 = 0, and
(5.4)
x1 + x3 + x5 = 0.
In other words, the vector X = (x1 , x2 , x3 , x4 , x5 ) ∈ C1 iff its coordinates
satisfy Equations 5.4. Equivalently, X ∈ C1 iff it is orthogonal to the two
vectors 11010 and 10101. If we take these two vectors as the row vectors of a
matrix H1 , then H1 is the 2 by 5 matrix:
 
1 1 0 1 0
H1 = .
1 0 1 0 1
H1 is called a parity-check matrix of the code C1 . The row vectors of H1 are
orthogonal to the row vectors of G1 . (Recall that two vectors X = (x1 , . . . , xn )
and Y = (y1 , . . . , yn ) of the same length n are orthogonal if their inner product
(= scalar product X, Y  = x1 y1 + · · · + xn yn ) is zero). Now if a vector v is
orthogonal to u1 , . . . , uk , then it is orthogonal to any linear combination of
u1 , . . . , uk . Hence the row vectors of H1 , which are orthogonal to the row
vectors of G1 , are orthogonal to all the vectors of the row space of G1 , that
is, to all the vectors of C1 . Thus,

C1 = X ∈ Z52 : H1 X t = 0 = Null space of the matrix H1 ,
where X t is the transpose of X. The orthogonality relations H1 X t = 0 give the
parity-check conditions for the code C1 . These conditions fix the redundancy
 Introduction to Coding Theory 229

positions, given the message positions of any codeword. A similar result holds
good for any linear code. Thus, any linear code over a field F is either the row
space of one of its generator matrices or the null space of its corresponding
parity-check matrix.
Note that if we form any k linear combinations of the generator matrix
of a linear code which are also linearly independent over the base field, the
resulting k words also form a basis for C. For instance, if


1 0 1 1
G=
0 1 1 1
is a generator matrix of a binary linear code C of length 4, the matrix


1 0 1 1
G =
1 1 0 0
is also a generator matrix of C. The reason is that every row vector of G ∈ C
and rank(G) = rank(G ) = 2.
So far, we have been considering binary linear codes. We now consider
linear codes over an arbitrary finite field F . As mentioned in Definition 5.2,
an [n, k] linear code C over F is a k-dimensional subspace of F n , the space of
all ordered n-tuples over F . If {u1 , . . . , uk } is a basis of C over F , every word
of C is a unique linear combination
α1 u1 + · · · + αk uk , αi ∈ F for each i.
Since αi can take q values for each i, 1 ≤ i ≤ k, C has q · q · · · q(k times) = q k
codewords.
Let G be the k by n matrix over F having u1 , . . . , uk of F n as its row
vectors. Then, as G has k (= dimension of C) rows and all the k rows form a
linearly independent set over F , G is a generator matrix of C. Consequently,
C is the row space of G over F . The null space of C is the space of vectors
X ∈ F n which are orthogonal to all the words of C. In other words, it is the
dual space C ⊥ of C. As C is of dimension k over F , C ⊥ is of dimension n − k
over F . Let {X1 , . . . , Xn−k } be a basis of C ⊥ over F . If H is the matrix whose
row vectors are X1 , . . . , Xn−k , then H is a parity-check matrix of C. It is an
(n − k) by n matrix. Thus,
C = row space of G
= null space of H

= X ∈ F n : HX t = 0 .
Theorem 5.1. Let G = (Ik |A) be a generator matrix of a linear code C over
F , where Ik is the identity matrix of order k over F , and A is a k by (n − k)
matrix over F . Then, a generator matrix of C ⊥ is given by

H = −At |In−k
over F .
230 Discrete Mathematics

Proof. Each row of H is orthogonal to all the rows of G since (by block
multiplication, see Chapter 3, Section 3.3.1),
 
−A
GH t = [Ik |A] = −A + A = 0.
In−k

Recall that in the example following Definition 5.3, k = 3, n = 5, and

⎡ ⎤
1 1  
G1 = (I3 |A) , where A = ⎣1 0⎦ while H1 = −At |I2 = At |I2 over Z2 .
0 1

Corollary 5.1. G = [Ik |A] is a generator matrix of a linear code C of length

n iff H = [−At |In−k ] is a parity-check matrix of C.

5.4 Minimum Distance of a Code

Definition 5.4. The weight wt(v) of a codeword v of a code C is the number of
non-zero coordinates in v. The minimum weight of C is the least of the weights
of its non-zero codewords. The weight of the zero vector of C is naturally zero.

Example 5.1
As an example, consider the binary code C2 with generator matrix
 
1 0 1 1 0
G2 = = [I2 |A] .
0 1 1 0 1

C2 has four codewords. Its three non-zero words are u1 = 10110, u2 =

01101, and u3 = u1 + u2 = 11011. Their weights are 3, 3 and 4, respec-
tively. Hence the minimum weight of C2 is 3. A parity-check matrix of C2 , by
Theorem 5.1, is ( as F = Z2 )
⎡ ⎤
    1 1 1 0 0
H2 = −At |I5−2 = At |I3 = ⎣ 1 0 0 1 0 ⎦.
0 1 0 0 1

Definition 5.5. Let X, Y ∈ F n . The distance d(X, Y ), also called the Ham-
ming distance between X and Y , is defined to be the number of places in which
X and Y differ. Accordingly, d(X, Y ) = wt(X − Y ).
 Introduction to Coding Theory 231

If X and Y are codewords of a linear code C, then X −Y is also in C and has

non-zero coordinates only at the places where X and Y differ. Accordingly, if
X and Y are words of a linear code C, then

d(X, Y ) = wt (X − Y ). (5.5)

As a consequence, we have our next theorem.

Theorem 5.2. The minimum distance of a linear code C is the minimum

weight of a non-zero codeword of C and hence it is equal to the minimum
weight of C.

Thus, for the linear code C2 of Example 5.1, the minimum distance is 3.
The function d(X, Y ) defined in Definition 5.5 does indeed define a distance
function (that is, a metric) on F n . That is to say, it has the following three
properties:
For all X, Y, Z in F n ,

i. d(X, Y ) ≥ 0, and d(X, Y ) = 0 iff X = Y ,

ii. d(X, Y ) = d(Y, X), and

iii. d(X, Z) ≤ d(X, Y ) + d(Y, Z).

We now give another interpretation for the minimum weight of a linear code
C over F∗q

Theorem 5.3. Let H be a parity-check matrix of a linear code C over Fq .

The minimum distance of C is d if and only if every set of d − 1 columns of H
is linearly independent and some set of d columns of H is linearly dependent.

Proof. Recall that C = {X ∈ Fnq : HX t = 0} = {X ∈ Fnq : XH t = 0} ⇔

x1 C1 + x2 C2 + · · · + xn Cn = 0 where C1 , C2 , . . . , Cn are the column vectors
of H and X = (x1 , x2 , . . . , xn ). If wt(X) = k, (for instance, if n = 5, X =
(2, 0, 3, 0, −4) and 2C1 + 3C3 − 4C5 = 0, it means that (2, 0, 3, 0, −4)H t = 0
and hence (2, 0, 3, 0, −4) belongs to code C, and it is of weight 3) then there are
k of the coefficients among x1 , x2 , . . . , xn that are not zero, and hence there
is a linear dependence relation among the corresponding k column vectors,
and in this case wt(X) = k. As X is a codeword, d(C) = d is the minimum k
such that there exists a linear dependence relation between d column vectors
of H.

5.5 Hamming Codes

Hamming codes are binary linear codes. They can be defined either by their
generator matrices or by their parity-check matrices. We prefer the latter. Let
232 Discrete Mathematics

us start by defining the [7, 4]-Hamming code H3 . The seven column vectors of
its parity-check matrix H of H3 are the binary representations of the numbers
1 to 7 written in such a way that the last three of its column vectors form I3 ,
the identity matrix of order 3. Thus,
⎡ ⎤
1 1 1 0 1 0 0
H = ⎣ 1 1 0 1 0 1 0 ⎦.
1 0 1 1 0 0 1
The columns of H are the binary representations of the numbers 7, 6, 5, 3; 4,
2, 1, respectively (7 = 22 + 21 + 20 etc.,). As H is of the form [−At |I3 ], the
generator matrix of H3 is given by
⎡ ⎤
1 0 0 0 1 1 1
⎢ 0 1 0 0 1 1 0 ⎥
G = [I7−3 |A] = [I4 |A] = ⎢
⎣ 0 0 1 0 1 0 1 ⎦.
⎥

0 0 0 1 0 1 1
H3 is of length 23 − 1 = 7, and dimension 4 = 7 − 3 = 23 − 1 − 3. What is
the minimum distance of H3 ? One way of finding it is to list all the 24 − 1
non-zero codewords (see Theorem 5.2). However, a better way of determining
it is the following. The first row of G is of weight 4, while the remaining rows
are of weight 3. The sum of any two or three of these row vectors as well as
the sum of all the four row vectors of G are all of weight at least 3. Hence the
minimum distance of H3 is 3.

5.6 Standard Array Decoding

We now write the coset decomposition of Z72 with respect to the subspace
H3 . (Recall that H3 is a subgroup of the additive group Z72 ). As Z72 has 27
vectors, and H3 has 24 codewords, the number of cosets of H3 in F 7 is 27 /24 =
23 . (See Chapter 3). Each coset is of the form X + H3 = {X + v : v ∈ H3 }
where X ∈ F7 . Any two cosets are either identical or disjoint. The vector X
is a representative of the coset X + H3 . The zero vector is a representative of
the coset H3 . If X and Y are each of weight 1, the coset X + H3 = Y + H3 ,
since X − Y is of weight at most 2 and hence does not belong to H3 (as H3
is of minimum weight 3). Hence the seven vectors of weight 1 in Z72 together
with the zero vector define 8 = 23 pairwise disjoint cosets exhausting all the
23 × 24 = 27 vectors of Z72 . These eight vectors (namely, the seven vectors of
weight one and the zero vector) are called coset leaders.
We now construct a double array (Figure 5.3) of vectors of Z72 with the
cosets defined by the eight coset leaders (mentioned above). The first row is
the coset defined by the zero vector, namely, H3 .
Figure 5.3 gives the standard array for the code H3 . If u is the message
vector (that is, codeword) and v is the received vector, then v − u = e is
Coset leader C= 3

0000000 1000111 0100110 0010101 0001011 1100001 1010010 1001100 0110011 0101101 0011010 0111000 1011001 1101010 1110100 1111111

1000000 0000111 1100110 1010101 1001011 0100001 0010010 0001100 1110011 1101101 1011010 1111000 0011001 0101010 0110100 0111111

0100000 1100111 0000110 0110101 0101011 1000001 1110010 1101100 0010011 0001101 0111010 0011000 1111001 1001010 1010100 1011111

0010000 1010111 0110110 0000101 0011011 1110001 1000010 1011100 0100011 0111101 0001010 0101000 1001001 1111010 1100100 1101111

0001000 1001111 0101110 0011101 0000011 1101001 1011010 1000100 0111011 0100101 0010010 0110000 1010001 1100010 1111100 1110111

0000100 1000011 0100010 0010001 0001111 1100101 1010110 1001000 0110111 0101001 0011110 0111100 1011101 1101110 1110000 1111011

0000010 1000101 0100100 0010111 0001001 1100011 1010000 1001110 0110001 0101111 0011000 0111010 1011011 1101000 1110110 1111101

0000001 1000110 0100111 0010100 0001010 1100000 1010011 1001101 0110010 0101100 0011011 0111001 1011000 1101011 1110101 1111110
Introduction to Coding Theory

FIGURE 5.3: Standard array decoding of Z72 .

233
234 Discrete Mathematics

the error vector. If we assume that v has one or no error, then e is of weight
1 or 0. Accordingly e is a coset leader of the standard array. Hence to get
u from v, we subtract e from v. In the binary case (as −e = e), u = v + e.
For instance, if in Figure 5.3, v = 1100110, then v is present in the second
coset for which the leader is e = 1000000. Hence the message is u = v + e =
0100110. This incidentally shows that H3 can correct single errors. However,
if for instance, u = 0100110 and v = 1000110, then e = 1100000 is of weight 2
and is not a coset leader of the standard array of Figure 5.3. In this case, the
standard array decoding of H3 will not work as it would wrongly decode v as
1000110 − 0000001 = 1000111 ∈ H3 . (Notice that v is present in the last row
of Figure 5.3). The error is due to the fact that v has two errors and not just
one. Standard array decoding is therefore maximum likelihood decoding.
The general Hamming code Hm is defined analogous to H3 . Its parity-check
matrix H has the binary representations of the numbers 1, 2, . . . , 2m − 1 as
its column vectors. Each such vector is a vector of length m. Hence H is an
m by 2m − 1 binary matrix and the dimension of Hm is (2m − 1) − m =
(number of columns in H) − (number of rows in H). In other words, Hm is a
[2m − 1, 2m − 1 − m] linear code over Z2 . Notice that H has rank m since H
contains Im as a submatrix.
The minimum distance
 of Hm is 3, m ≥ 2. This can be seen as follows.
m
Recall that Hm = X ∈ Z22 −1 : HX t = 0 . Let i, j, k denote respectively
the numbers of the columns of H in which the m-tuples (that is, vectors of
length m) 0 . . . 011, 00 . . . 0101 and 0 . . . 0110 are present (Figure 5.4).
Let v be the binary vector of length 2m −1 which has 1 in the i-th, j-th and
k-th positions and zero at other positions. Clearly, v is orthogonal to all the
row vectors of H and hence belongs to Hm . Hence Hm has a word of weight
3. Further, Hm has no word of weight 2 or 1. Suppose Hm has a word u of
weight 2. Let i, j be the two positions where u has 1. As HuT = 0, by the
rule of matrix multiplication, Ci + Cj = 0, where Ci , Cj are the i-th and j-th
columns of H, respectively. This means that Ci = −Cj = Cj (the code being
binary). But this contradicts the fact that the columns of H are all distinct.
For a similar reason, Hm has no codeword of weight 1. Indeed, assume that
Hm has a codeword v of weight 1. Let v have 1 in its, say, k-th position and
0 elsewhere. As v ∈ Hm , Hv = 0, and this gives that the k-th column of H is

⎡ ⎤
··· 0 ··· 0 ··· 0 ···
⎢ .. .. .. ⎥
⎢ . . . ⎥
⎢ ⎥
H:⎢ 0 1 1 ⎥
⎢ ⎥
⎣· · · 1 · · · 0 · · · 0 · · ·⎦
··· 1 ··· 1 ··· 0 ···

FIGURE 5.4: Form of the parity-check matrix of Hm (m ≥ 2).

 Introduction to Coding Theory 235

the zero vector, a contradiction. Thus, Hm has no codeword of weight 1 or 2

but it has a word of weight 3. Hence the minimum weight of Hm is 3.

5.7 Sphere Packings

As before, let F n denote the vector space of all ordered n-tuples over the
field F . Recall (Section 5.4) that F n is a metric space with the Hamming
distance between vectors of F n as a metric.
Definition 5.6. In F n , the sphere with center X and radius r is the set

S(X, r) = {Y ∈ F n : d(X, Y ) ≤ r} ⊆ F n .

Definition 5.7. An r-error-correcting linear code C is perfect if the spheres

of radius r with the codewords of C as centers are pairwise disjoint and their
union is F n .
The above definition is justified because if v is a received vector that has
at most r errors, then v is at a distance at most r from a unique codeword u
of C, and hence v belongs to the unique sphere S(u, r); then, v will be decoded
as u.
Out next theorem shows that the minimum distance d of a linear code stands
as a good error-correcting measure of the code.

Theorem 5.4. If C is a linear code of minimum distance d, then C can correct

t =
(d − 1)/2 or fewer errors.

Proof. It is enough to show that the spheres of radius t centered at the

codewords of C are pairwise disjoint. Indeed, if u and v are in C, and if
z ∈ B(u, t) ∩ B(v, t), that is, if the two spheres intersect, (see Figure 5.5),
then

d(u, v) ≤ d(u, z) + d(z, v)

≤ t + t = 2t ≤ d − 1 < d,

a contradiction to the fact that d is the minimum distance of C. Hence the

two spheres must be disjoint.

z
u v

FIGURE 5.5: Case when two spheres intersect.

236 Discrete Mathematics

Corollary 5.2. The Hamming code Hm is single error correcting.

Proof. As d = 3 for Hm ,
(d − 1)/2 = 1. Now apply Theorem 5.4.

Theorem 5.5. The Hamming code Hm is a single-error-correcting perfect

code.
m
Proof. Hm is a code of dimension 2m −1−m over Z2 and hence has 2(2 −1−m)
words. Now, if v is any codeword of Hm , S(v, 1) contains v (which is at distance
zero from v) and the 2m −1 codewords got from v (which is of length 2m −1) by
altering each position once at a time. Thus, S(v, 1) contains 1 + (2m − 1) = 2m
words of Hm (recall that Hm is single error correcting). As d = 3 for Hm , the
spheres s(v, 1), v ∈ Hm are pairwise disjoint. The cardinality of the union
m m
of the spheres S(v, 1) as v varies over Hm is 22 −1−m · 2m = 22 −1 = the
number of vectors in F n , where n = 2m − 1. Thus the spheres S(v, 1), v ∈ Hm
are pairwise disjoint and cover the whole space F n . Hence Hm is perfect.

5.8 Extended Codes

Let C be a binary linear code of length n. We can extend this code by
adding an overall parity check at the end. This means, we add a zero at the
end of each word of even weight in C, and add 1 at the end of every word of
odd weight. This gives an extended code C  of length n + 1.
For looking at some of the properties of C  , we need a lemma.

Lemma 5.1. Let w denote the weight function of a binary code C. Then,

d(X, Y ) = w(X + Y ) = w(X) + w(Y ) − 2(X  Y ) (5.6)

where X  Y is the number of common 1s in X and Y .

Proof. Let X and Y have common 1s in the i1 , i2 , . . . , ip -th positions so that

X  Y = p. Let X have 1s in the i1 , . . . , ip and j1 , . . . , jq -th positions, and
Y in the i1 , . . . , ip and l1 , . . . , lr -th positions (where no jk is an lt ). Then,
w(X) = p + q, w(Y ) = p + r and w(X + Y ) = q + r. The proof is now
clear.

Coming back to the extended code C  , by definition, every codeword of C 

is of even weight. Hence C  is an even weight code.

Theorem 5.6. Suppose d is odd. Then, a binary [n, k]-linear code with dis-
tance d exists iff a binary [n + 1, k]-linear code with distance d + 1 exists.

Proof. Suppose C is a binary [n, k]-linear code with distance d. C  is the

extended code obtained from C. Since wt(X  ) is even for every codeword X 
 Introduction to Coding Theory 237

of C  , by Lemma 5.1, d(X  , Y  ) is even for all X  , Y  ∈ C  . Therefore d(C  ) is

even. Also d ≤ d(C  ) ≤ d + 1. By assumption, d is odd. Hence d(C  ) = d + 1.
Thus, C  is a binary [n + 1, k]-linear code with distance d + 1.
Suppose C  is an [n + 1, k]-linear code with distance d + 1. Since d(C  ) =
d + 1, there exist two codewords X and Y such that d(X, Y ) = d + 1. Choose
a position in which X and Y differ and delete this from all the codewords.
The result is an [n, k]-linear code C with distance d.
A generator matrix of C  is obtained by adding an overall parity check to
the rows of a generator matrix of C. Thus, a generator matrix of the extended

code H4 is ⎡ ⎤
1 0 0 0 1 1 1 0
⎢ 0 1 0 0 1 1 0 1 ⎥
⎢ ⎥
⎣ 0 0 1 0 1 0 1 1 ⎦.
0 0 0 1 0 1 1 1

5.9 Syndrome Decoding

Let C be an [n, k]-linear code over GF (q) = F . The standard array decod-
ing scheme requires storage of q n vectors of F n and also comparisons of a
received vector with the coset leaders. The number of such comparisons is
at most q n−k , the number of distinct cosets in the standard array. Hence
any method that makes a sizeable reduction in storage and the number of
comparisons is to be welcomed. One such method is the syndrome-decoding
scheme.

Definition 5.8. The syndrome of a vector Y ∈ F n with respect to a linear

[n, k]-code over F with parity-check matrix H is the vector HY t .
As H is an (n − k) by n matrix, the syndrome of Y is a column vector
of length n − k. We denote the syndrome of Y by S(Y ). For instance, the
syndrome of Y = 1110001 with respect to the Hamming code H4 is
⎡ ⎤
1
⎢ 1 ⎥
⎡ ⎤⎢ ⎥ ⎡ ⎤
1 1 1 0 1 0 0 ⎢ ⎥
⎢ 1 ⎥ 1
⎣ 1 1 0 1 0 1 0 ⎦⎢ 0 ⎥ = ⎣ 0 ⎦.
⎢ ⎥
1 0 1 1 0 0 1 ⎢ ⎥
⎢ 0 ⎥ 1
⎣ 0 ⎦
1
Theorem 5.7. Two vectors of F n belong to the same coset in the standard
array decomposition of a linear code C iff they have the same syndrome.

Proof. Let u and v belong to the same coset a + C, a ∈ F n , of C. Then,

u = a + X and v = a + Y , where X, Y are in C. Then, S(u) = S(a + X) =
238 Discrete Mathematics

H(a + X)t = Hat + HX t = S(a) (Recall that as X ∈ C, S(X) = HX t = 0).

Similarly S(v) = S(a). Thus, S(u) = S(v).
Conversely, let S(u) = S(v). Then, Hut = Hv t and therefore H(u − v)t =
0. This means that u − v ∈ C, and hence the cosets u + C and v + C are
equal.

Theorem 5.7 shows that the syndromes of all the vectors of F n are deter-
mined by the syndromes of the coset leaders of the standard array of C. In
case C is an [n, k]-binary linear code, there are 2n−k cosets and therefore the
number of distinct syndromes is 2n−k . Hence in contrast to standard-array
decoding, it is enough to store 2n−k vectors (instead of 2n vectors) in the
syndrome decoding. For instance, if C is a [100, 30]-binary linear code, it is
enough to store the 270 syndromes instead of the 2100 vectors in Z100
2 , a huge
saving indeed.

5.10 Error Detection

Consider the binary code C = {0000, 1100, 1111}. If we change any single
coordinate of a codeword, it does not give another codeword. However, the
same thing is not true if we change two coordinates. For instance, if we replace
both 1s by 0 or both 0s by 1 in 1100, we end up in another codeword. Hence
we say that C can detect one error but not 2 or more.

Definition 5.9. Let C be any code, not necessarily linear. Then, C is t-error
detecting, if any word of C incurs k errors, 1 ≤ k ≤ t, then the resulting word
does not belong to C.

Theorem 5.8. A code C of length n over F of minimum distance d can detect

at most d − 1 errors.

Proof. Suppose X ∈ C, and Y ∈ Fn , Y = X, such that d(X, Y ) ≤ t < d. Then,

Y cannot belong to C. Hence C can detect up to d − 1 errors.

5.11 Sphere Packing Bound or Hamming Bound

We look at the following problems: Given a t-error correcting linear code of
length n over the finite field F = GF (q), where q is a prime power pr (p being
a prime), what is the number of elements of F n = {(x1 , x2 , . . . , xn ) : xi ∈
F, 1 ≤ i ≤ n} that a (closed) sphere S of radius t with center at a codeword
u can contain? Now any vector in S is at distance k from u, where 0 ≤ k ≤ t.
Let nk be the number of vectors in S which are at distance k from u. Clearly,
 Introduction to Coding Theory 239

n0 = 1. Any vector in S which is at distance 1 from u can be obtained by

replacing any one of the n coordinates of u by a new element of GF (q), and
this can be done in (q − 1) ways. Each single coordinate  can be chosen in
n ways, as the vector is of length n. Hence there are n1 (q − 1) vectors in
S which are at a distance 1 from u. Next to find the point in S which are
at distance
 2 from u, we change two of the coordinates of u. But then there
are n2 ways
 of choosing two coordinates out of the n coordinates of u. So
there are n2 (q − 1)2 vectors in S which are at a distance 2 from u. The same
argument extends to the number of points at any distance s, 1 ≤ s ≤ t. As
the code is t-error correcting, we have the following result.

Theorem 5.9. The number of points that a (closed) sphere of radius t in the
space Fqn of n-tuples over the Galois field Fq of q elements is





n n n n
+ (q − 1) + (q − 1)2 + · · · + (q − 1)t .
0 1 2 t
Corollary 5.3. (Sphere-packing bound or Hamming bound.) An (n, M,
2t + 1)-linear code C over GF (q) satisfies the condition



 
n n n
M (q − 1) + (q − 1)2 + · · · + (q − 1)t ≤ q n . . . (∗)
0 2 t
Proof. As the minimum distance of C is d = 2t+1, the closed spheres of radius
t = (d − 1)/2 about the codewords are pairwise disjoint. The total number of
points of Fnq that belong to these spheres is, by virtue of Theorem 5.9, the
expression on the LHS of (*). Certainly, this number cannot exceed the total
number of points of Fnq which is q n . This proves the result.
Corollary 5.4. If the (n, M, 2t + 1)-code is binary, then the sphere packing
bound is given by




n n n
M + + ··· + ≤ 2n . . . (∗∗)
0 1 t
Proof. Take q = 2 in Corollary 5.3.
Consider again the binary case (that is, when q = 2), and let n = 6. Then,
|Z62 | = 26 = 64. We ask: Can an (n, M, d) = (6, 9, 3), code exist in Z62 so that
the closed unit spheres with centers at the 9 codewords contain all the 64
vectors of Z62 ? The answer is “no” since the nine spheres
 with centers
 at each
of the nine vectors can contain at the most M [ n0 + n1 ] = 9[ 60 + 61 ] = 63
vectors and hence not all of the 64 vectors of Z62 . This leads us to the concept
of a perfect code.

Definition 5.10. An (n, M, d)-code over the finite field Fq is called perfect if
the spheres of radius
(d − 1)/2 with the centers at the M codewords cover
all the q n vectors of length n over Fq .
In other words, an (n, M, d)-code is perfect iff equality is attained in (*)
(of Corollary 5.3).
240 Discrete Mathematics

Example of a perfect code Let H be the m by 2m − 1 matrix, in

m
which the columns of H are the 2 − 1 non-zero binary vectors of length m
given in some order. Let Hm be the binary linear code with H as its parity-
check matrix. As H contains Im , the identity matrix of order m as a submatrix,
rank(H) = m. As the number of columns of H is 2m − 1, and Hm is the null
space of H,
dim(Hm ) = 2m − 1 − m. (see [20]).
m  
Taking t = 1, and M = 22 −1−m in (**) above, we get M ( n0 + n1 ) =
m m m
22 −1−m (1 + n) = 22 −1−m (1 + (2m − 1)) = 22 −1 = 2n = the RHS of (**).
We thus conclude:

Theorem 5.10. The binary Hamming code Hm is a single error correcting

perfect code.

Our earlier results show that no binary (6, 9, 3) perfect code exists. (See
also Exercise 9 at the end of this section). We now look at the general Ham-
ming code C of length m over the field GF (q). The column vectors of any
parity-check matrix of C consists of non-zero vectors of length m over GF (q).
Now the space spanned by any such vector v over GF (q) is the same as the
space spanned by αv for any non-zero element α of GF (q), that is, v = αv
for 0 = α ∈ GF (q). Hence we choose only one of these q − 1 vectors as a
column vector of H. As there are q m − 1 non-zero vectors of length m over
GF (q), we have (q m − 1)/(q − 1) distinct vectors of length m, no two of which
spanning the same subspace over GF (q). Hence the number of columns of the
parity-check matrix of this generalized Hamming code is (q m − 1)/(q − 1).
Note: When q = 2, the number of columns of the Hamming code is
(2m − 1)/(2 − 1) = 2m − 1, as seen earlier.
Another Example We now construct the Hamming code Hm with m = 3
over the field GF (3). The elements of GF (3) are 0, 1, 2 and addition and
multiplication are taken modulo 3. The number of non-zero vectors of length
3 over GF (3) is 33 − 1 = 26. But then for any vector of x of length 3 over
GF (3), 2x is also such a vector. Hence the number of distinct column vectors
of H = H3 in which no two column vectors span the same space is 26/2 = 13.
Hence H is a 3×13 matrix of rank 3 and it is therefore of dimension 13−3 = 10.
This shows that H3 is a (13, 310 , 3)-ternary perfect code. In fact, the condition
(*) is


 
10 13 13
3 + (3 − 1) = 310 (1 + 13 × 2) = 310 · 33 = 313 ,
0 1

so that equality is present in (*). Finally, we give the parity-check matrix H

of the code as constructed above:
⎡ ⎤
0 0 0 0 1 1 1 1 1 1 1 1 1
⎣ 0 1 1 1 0 0 0 1 1 1 2 2 2 ⎦
1 0 1 2 0 1 2 0 1 2 0 1 2
 Introduction to Coding Theory 241

We observe that the first non-zero entry in each column is 1. This makes H
unique except for the order of the columns. We may as well replace any column
vector v by 2v (where 0 = 2 ∈ GF (3)).

5.12 Exercises
1. Show by means of an example that the syndrome of a vector depends
on the choice of the parity-check matrix.
2.(a) Find allthe codewords of the binary code with generator
1 0 1 1 1
matrix .
1 0 0 1 1
(b) Find a parity-check matrix of the code.
(c) Write down the parity-check equations.
(d) Determine the minimum weight of the code.

3. Decode the received vector 1100011 in H4 using (i) the standard array
decoding, and (ii) syndrome decoding.
4. How many vectors of Z72 are there in S(u, 3), where u ∈ Z72 ?
5. How many vectors of F n are there in S(u, 3), where u ∈ F n , and |F | = q?

6. Show that a t-error-correcting binary perfect [n, k]-linear code satisfies

the relation



n n n
+ + ··· + = 2n−k .
0 1 t

More generally, show that a t-error-correcting perfect [n, k]-linear code

over GF (q) satisfies the relation




n n n
+ + ··· + = q n−k .
0 1 t

7. Show that the function d(X, Y ) defined in Section 5.4 is indeed a metric.

8. Show that it is impossible to find nine binary vectors of length 6 such

that the distance between any two of them is at least 3.

9. Let C be a linear code of length n over a field F , and let C be of minimum

weight 5. Can there be two vectors of Fn2 , each of weight 2, belonging to
the same coset of C in Fn ?
10. Determine the number of errors that the linear code C over F5 with
generators 01234, 12044, 13223 can detect.
242 Discrete Mathematics

11. Show by means of an example that the generator matrix of a linear code
in Z73 (that is a code of length 7 over the field Z3 ) need not be unique.
12. Let C ⊆ Z33 be given by:
C = {(x1 , x2 , x3 ) ∈ Z33 : x1 + x2 + 2x3 = 0}.
A. List all the codewords of C.
B. Give a generator matrix of C.
C. Give another generator matrix of C (different from what you
gave in (ii)).
D. Calculate the minimum distance d(C) of C.
E. How many errors can C correct?
F. How many errors can C detect?
G. Give a parity-check matrix for C using (c) above.

13. Show by means of an example that the coset leader of a coset of a linear
code need not be unique.

We close this chapter with a brief discussion on a special class of linear codes,
namely, cyclic codes.

5.13 Cyclic Codes

Definition 5.11. A linear code C over a field F is called cyclic if every cyclic
shift of a codeword in C also belongs to C.
To be precise, (a0 , a1 , a2 , . . . , an−1 ) ∈ C ⇒ (an−1 , a0 , a1 . . . , an−2 ) ∈ C.
This definition of course implies that (an−2 , an−1 , a0 , . . . , an−3 ) ∈ C, etc.
For example, if n = 4, and (a0 , a1 , a2 , a3 ) ∈ C, then (a3 , a0 , a1 , a2 ), (a2 , a3 , a0 , a1 ),
(a1 , a2 , a3 , a0 ) are all in C.
We now identify the codeword (a0 , a1 , a2 , . . . , an−1 ) of C with the polyno-
mial (See Section 3.20)

g(x) = a0 + a1 x + a2 x2 + · · · + an−1 xn−1 ∈ F [x]/(xn − 1). (5.7)

It is more precise to say that we identify the codeword (a0 , a1 , . . . , an−1 ) with
the residue class

a0 + a1 x · · · + an−1 xn−1 + (xn − 1).

But then as mentioned in Remark 3.1, we can identify the residue class with
the corresponding polynomial and do arithmetic modulo (xn − 1). The identi-
fication given by Equation 5.7 makes xg(x) = x(a0 + a1 x + · · · + an−1 xn−1 ) =
a0 x + a1 x2 + a2 x3 + · · · + an−1 xn = an−1 + a0 x + a1 x2 + a2 x3 + · · · + an−2 xn−1
 Introduction to Coding Theory 243

to correspond to the word (an−1 , a0 , a1 , · · · + an−2 ), x2 g(x) = an−2 + an−1 x +

a0 x2 +. . . , an−3 xn−1 to correspond to the word (an−2 , an−1 , a0 , . . . , an−3 ) and
so on. Moreover, if g(x) ∈ C, kg(x) also belongs to C for every k ∈ F . Hence if
f (x) ∈ F [x], f (x)g(x) ∈ C (under the identification). In other words, C is an
ideal in the commutative ring F [x]/(xn − 1). Recall that (Theorem 3.13), any
ideal in F [x] is principal and it is generated by the unique monic polynomial
of least degree present in C.

Example 5.2
Let V be the space of all binary 3-tuples of R. Then, the cyclic codes
in V are the following:
Code Codewords in C Corresponding polynomials in R3
C1 (0,0,0) 0
C2 (0,0,0) 0
(1,1,1) 1 + x + x2
C3 (0,0,0) 0
(1,1,0) 1+x
(0,1,1) x + x2
(1,0,1) 1 + x2
C4 All of V All of R3


Note: Not every linear code is cyclic. For instance, the binary code C =
{(0, 0, 0), (1, 0, 1)} is the code generated by (1, 0, 1) but it is not cyclic (as

(1, 1, 0) ∈
/ C ).
To summarize, we have the following result.

Theorem 5.11. Let C be a cyclic code of length n over a field F , and let
Rn = F [x]/(xn − 1). Then, the following are true:
i. There exists a unique monic polynomial g(x) of least degree k ≤ n − 1
in C.

ii. C = g(x), the ideal generated by g(x) ∈ Rn .

iii. g(x)/(xn − 1).
(For a proof of (iii), see Theorem 3.13.) We observe that the constant term
in g(x) = 0. For assume that g(x) = a0 + a1 x + · · · + an−k xn−k , and a0 = 0.
Then, xn−1 g(x) ∈ C ⇒ xn−1 (a1 x + · · · + an−k xn−k ) ∈ C (C being an ideal in
Rn ) ⇒ a1 + a2 x + · · · + ak xn−k−1 ∈ C, contradicting (i).
We now determine the generator polynomial of a cyclic code.
244 Discrete Mathematics

Theorem 5.12. Let C = g(x) be a cyclic code of length n, equivalently an

ideal in Rn = F [x]/(xn − 1). Let g(x) = a0 + a1 x + · · · + an−k xn−k , an−k = 0.
Then, a generator matrix of C is

⎡ ⎤
a0 a1 a2 . an−k 0 0 0 . 0
⎢0 a0 a1 a2 . an−k 0 0 . 0 ⎥
⎢ ⎥
⎢0 . . . . . . . . . ⎥
G:⎢
⎢.
⎥.
⎥
⎢ . . . . . . . . . ⎥
⎣0 0 0 0 a0 a1 . . an−k 0 ⎦
0 0 0 0 0 a0 a1 . . an−k

Consequently dim C = k.

Proof. (a0 , a1 , a2 , . . . , an−1 ) ∈ C is equivalent to: g(x) = a0 + a1 x + · · · +

an−1 xn−1 ∈ (ideal) C ⊆∈ Rn . Hence g(x), xg(x), . . . , xk−1 g(x) all belong to
C.
Claim: The above polynomials are linearly independent over F . Indeed,
c0 g(x) + c1 xg(x) + · · · + ck−1 xk−1 g(x) = 0, where ci ∈ F , for each i,
0 ≤ i ≤ k − 1,
⇒ (c0 + c1 x + · · · + ck−1 xk−1 )g(x) = 0
⇒ (c0 + c1 x + · · · + ck−1 xk−1 )(a0 + a1 x + · · · + an−1 xn−1 ) = 0 (the zero poly-
nomial)
⇒ c0 a0 = 0, c0 a1 + c1 a0 = 0, c0 a2 + c1 a1 + c0 a0 = 0, etc.
Recall a0 = 0 ⇒ c0 = 0 ⇒ c1 = 0 ⇒ c2 = 0, etc. Hence ci = 0 for each i,
0 ≤ i ≤ k − 1.
Again, the k polynomials span C. To see this, assume that s(x) ∈ C. Then,
s(x) is a multiple of g(x).
⇒ ∃ a polynomial c(x) of degree ≤ (n − 1) − (n − k) = k − 1, say,
c(x) = c0 + c1 x + · · · + ck−1 xk−1 , such that s(x) = c(x)g(x) = c0 g(x) +
c1 xg(x) + · · · + ck−1 xk−1 g(x) ⇒ s(x) belongs to the row space of the matrix
G. This proves the result.

Consider C = g(x), where g(x) = 1 + x2 + x3 . Then, dimC = 7 − 3 = 4. A

generator matrix G of C is obtained by writing the vectors for the polynomials
g(x), xg(x), x2 g(x) and x3 g(x). Thus, as g(x) = 1 + 0.x + 1.x2 + 1.x3 + 0.x4 +
0.x5 + 0.x6 ,

⎡ ⎤
1 0 1 1 0 0 0
⎢0 1 0 1 1 0 0⎥
G=⎢
⎣0
⎥.
0 1 0 1 1 0⎦
0 0 0 1 0 1 1
 Introduction to Coding Theory 245

Hence C is of dimension 4 = number of rows of G. (Note: The last four columns

form a lower triangular non-singular matrix. In the binary case, it is in fact
unit lower triangular).

Example 5.3
Binary cyclic codes of length 7.
If C = g(x) is a binary cyclic code of length 7, g(x) ∈ Z2 [x] and
g(x)/(x7 − 1). Hence, to determine all binary cyclic codes of length
7, we first factorize x7 − 1 into irreducible factors over Z2 . In fact we
have:
x7 − 1 = (1 + x)(1 + x + x3 )(1 + x2 + x3 ).
As there are three irreducible factors on the right, there exist 23 = 8
binary cyclic codes of length 7.

5.14 Dual Codes

Let C be a linear code of length n and dimension k over R. Let C ⊥ denote
the set of vectors in F n which are orthogonal to all the codewords of C. If v
and w are orthogonal to u, then any linear combination of v and w over F is
also orthogonal to u. Hence C ⊥ is also a linear code. It is the null space of a
generator matrix of C and hence its dimension is n − k.
Suppose C is a cyclic code. The natural question is: Is the code C ⊥ also
cyclic? The answer is “yes” as shown below.
Let g(x) be the generator polynomial of C. By Theorem 5.11 g(x)/(xn −1).
Let xn − 1 = g(x)h(x). However, h(x) need not be the generator polynomial
of C ⊥ . For example, consider again the polynomial g(x) = 1 + x2 + x3 of
Example 5.3 which corresponds to the vector (1, 0, 1, 1, 0, 0, 0) ∈ Z2 . Now
h(x) = (x7 − 1)/(1 + x2 + x3 ) = (1 + x)(1 + x + x3 ) = 1 + x2 + x3 + x4 and
this corresponds to the vector (1, 0, 1, 1, 1, 0, 0). The inner product of these
two vectors is 1 and not zero. It turns out that the generator polynomial of
C ⊥ is the reciprocal polynomial of h(x).

Definition 5.12. Let h(x) = c0 + c1 x + · · · + cn−1 xn−1 . then the reciprocal

polynomial hR (x) of h(x) is defined by: hR (x) = cn−1 +cn−2 x+· · ·+c0 xn−1 =
xn−1 (h(1/x)).

Lemma 5.2. Let u(x) = a0 +a1 x+· · ·+an−1 xn−1 and v(x) = b0 +b1 x+· · ·+
bn−1 xn−1 be two polynomials in Rn = F [x]/(xn −1). Then, u(x)v(x) = 0 in Rn
iff the vector u1 = (a0 , a1 , . . . , an−1 ) is orthogonal to v1 = (bn−1 , bn−2 , . . . , b0 )
and to all the cyclic shifts of v1 .
246 Discrete Mathematics

Proof.

u(x)v(x) = 0
⇐⇒ (a0 bn−1 + a1 bn−2 + · · · + an−1 b0 )xn−1 + (a1 bn−1 + a2 bn−2 + . . .
+ a0 b0 )xn + · · · + (an−1 bn−1 + a0 bn−2 + . . . an−2 b0 )xn−2 = 0
⇐⇒ (on rearranging)(a0 bn−1 + a1 bn−2 + · · · + an−1 b0 )xn−1 +
(a0 b0 + a1 bn−1 + a2 bn−2 + · · · + an−1 b1 )xn−2 + · · · + (a0 bn−2
+ a1 bn−3 + · · · + an−2 b0 + an−1 bn−1 )xn−2 = 0
⇐⇒ each coefficient of xi , 0 ≤ i ≤ n − 1, is zero.
⇐⇒ u1 = (a0 , a1 , . . . , an−1 )is orthogonal to v1 = (b0 , b1 , . . . , bn−1 )
and to all the cyclic shifts of v1 .

Corollary 5.5. (with the same notation as in Lemma 5.2) u(x)v(x) = 0 in

Rn iff (b0 , b1 , . . . , bn−1 ) is orthogonal to v1 = (an−1 , an−2 , . . . , a0 ) and all its
cyclic shifts of v1 .

Proof. a(x)b(x) = 0 ⇐⇒ b(x)a(x) = 0(in Rn ). Now apply Lemma 5.2

Theorem 5.13. Let C be a cyclic code of length n with generator polynomial

g(x) = a0 + a1 x + a2 x2 + · · · + an−k xn−k so that C is of dimension k. Suppose
xn − 1 = g(x)h(x), and let hR (x) be the reciprocal polynomial of h(x). If G
and H are respectively the k by n and n − k by n matrices:

⎡ ⎤
a0 a1 a2 . an−k 0 0 0 . 0
⎢ 0 a0 a1 a2 . a 0 0 . 0 ⎥
⎢ n−k ⎥
⎢0 . . . . . . . . . ⎥
⎢
G= ⎢ ⎥, and
. . . . . . . . . . ⎥
⎢ ⎥
⎣0 0 0 0 a0 a1 . . an−k 0 ⎦
0 0 0 0 0 a0 a1 . . an−k
⎡ ⎤
hk hk−1 hk−2 . h0 0 0 0 . 0
⎢0 hk hk−1 . h1 h0 0 0 . 0⎥
⎢ ⎥
⎢0 0 . . . . . . . .⎥
H= ⎢
⎢.
⎥,
⎢ . . . . . . . . .⎥⎥
⎣. . . . . . . . . .⎦
0 0 0 0 0 hk . . h1 h0

then G is a generator matrix of C, and H is a generator matrix of C ⊥ . Con-

sequently, C ⊥ is also a cyclic code.

Proof. That G is a generator matrix of C has already been seen in Theorem

5.12. By hypothesis, g(x)h(x) = xn − 1 and hence g(x)h(x) = 0 in Rn . This
means, by virtue of Lemma 5.2, every row of G is orthogonal to every row
 Introduction to Coding Theory 247

of H. Moreover, rank of G = k = dim (C), and rank of H = n − k. Hence H

is a generator matrix of C ⊥ and dim(C ⊥ ) = (n − k). Further, as the rows of
H are cyclic shifts of its first row, C ⊥ is also cyclic.
Note that of the matrices G and H, the transpose of one is the parity-check
matrix of the linear code defined by the other.

5.15 Exercises
1. Let G be the generator matrix of a binary linear code C of length n and
dimension k. Let G be the matrix obtained by adding one more parity
check at the end of each row vector of G (that is add 0 or 1 according

to whether the row vector is of even or odd weight). Let C be the code
 
generated by G . Show that G is an even-weight code (that is, every
 
word is of even weight). Determine H , a parity-check matrix of C .
2. Let C3 be the code of Example 5.2. Show that C3 is cyclic. Determine
dim(C3 ) and C3⊥ .
3. Determine all the binary cyclic codes of length 8.
4. Determine the binary cyclic code of length n with generator polynomial
1 + x.

5. Factorice x7 −1 over Z2 . Hence determine two cyclic codes each of dimen-

sion 3. Show that these two codes are equivalent.
6. Let C be a binary code of odd length n and d(C) ≥ 3. Prove that C has
no codewords of weight n − 2.
Chapter 6
Cryptography

Three may keep a secret, if two are dead.

Benjamin Franklin
Statesman and Scientist

6.1 Introduction
To make a message secure, the sender usually sends the message in a
disguised form. The intended receiver removes the disguise and then reads
off the original message. The original message of the sender is the plaintext,
and the disguised message is the ciphertext. The plaintext and the ciphertext
are usually written in the same alphabet. The plaintext and the ciphertext
are divided, for the sake of computational convenience, into units of a fixed
length. The process of converting a plaintext to a ciphertext is known as
enciphering or encryption, and the reverse process is known as deciphering
or decryption. A message unit may consist of a single letter or any ordered
k-tuple, k ≥ 2. Each such unit is converted into a number in a suitable arith-
metic and the transformations are then carried out on this set of numbers. An
enciphering transformation f converts a plaintext message unit P (given by
its corresponding number) into a number that represents the corresponding
ciphertext message unit C while its inverse transformation, namely, the deci-
phering transformation just does the opposite by taking C to P . We assume
that there is a 1–1 correspondence between the set of all plaintext units P
and the set of all ciphertext units C. Hence each plaintext unit gives rise to a
unique ciphertext unit and vice versa. This can be represented symbolically by
f f −1
P −→ ξ −→ C.
Such a setup is known as a cryptosystem.

6.2 Some Classical Cryptosystems

6.2.1 Caesar cryptosystem
One of the earliest of the cryptosystems is the Caesar cryptosystem
attributed to the Greek emperor Julius Caesar of the first century B.C. In
this cryptosystem, the alphabet is the set of English characters A, B, C, . . . ,

249
250 Discrete Mathematics

TABLE 6.1: Numerical equivalents of English characters

A B C D E F G H I J K L M
0 1 2 3 4 5 6 7 8 9 10 11 12
N O P Q R S T U V W X Y Z
13 14 15 16 17 18 19 20 21 22 23 24 25

X, Y , Z labelled 0, 1, 2, . . . , 23, 24, 25, respectively, so that 0 corresponds

to A, 1 corresponds to B and so on, and finally 25 corresponds to Z. In
this system, each message unit is of length 1 and hence consists of a single
character. The encryption (transformation) f : P −→ C is given by
f (a) = a + 3 (mod 26), (6.1)
−1
while the decryption (transformation) f : C −→ P is given by
−1
f (b) = b − 3 (mod 26). (6.2)
Table 6.1 gives the 1–1 correspondence between the characters A to Z and
the numbers 0 to 25.
For example, the word “OKAY” corresponds to the number sequence “(14)
(10) (0) (24)” and this gets transformed, by Equation 6.1 to “(17) (13) (3)
(1)” and so the corresponding ciphertext is “RNDB.” The deciphering trans-
formation applied to “RNDB” then gives back the message “OKAY.”

6.2.2 Affine cryptosystem

Suppose we want to encrypt the message “I LIKE IT.” In addition to
the English characters, we have in the message two spaces in between words.
So we add “space” to our alphabet by assigning to it the number 26. We
now do arithmetic modulo 27 instead of 26. Suppose, in addition, each such
message unit is an ordered pair (sometimes called a digraph). Then, each
unit corresponds to a unique number in the interval [0, 272 − 1]. Now, in
the message, “I LIKE IT,” the number of characters including the two spaces
is 9, an odd number. As our message units are ordered pairs, we add an extra
blank space at the end of the message. This makes the number of characters
as 10 and hence the message can be divided into 5 units U1 , U2 , . . . , U5 :
|I−| |LI| |KE| |−I| |T −|
where U1 = I− etc. (here, − stands for space). Now U1 corresponds to the
number (see Table 6.1) 8 · 271 + 26 = 242. Assume that the enciphering
transformation that acts now on ordered pairs is given by
C ≡ aP + b (mod 272 ) (6.3)
where a and b are in the ring Z27 , a = 0 and (a, 27) = 1. In Equation 6.3,
P and C denote a pair of corresponding plaintext and ciphertext units.
 Cryptography 251

The Extended Euclidean Algorithm (see chapter 3 of [25]) ensures that as

(a, 27) = 1, (a, 272 ) = 1 and hence a has a unique inverse a−1 (mod 729) so
that
aa−1 ≡ 1 (mod 272 ).
This enables us to solve for P in terms of C from the Congruence (6.3). Indeed,
we have
P ≡ a−1 (C − b) (mod 272 ). (6.4)
As a specific example, let us take a = 4 and b = 2. Then,

C ≡ 4P + 2 (mod 272 ).

Further as (4, 27) = 1, 4 has a unique inverse (mod 272 ); in fact 4−1 = 547
(mod 729) as 4 · 547 ≡ 1 (mod 272 ). (Indeed, if 4x ≡ 1 (mod 729), 4(−x) ≡
−1 ≡ 728 ⇒ −x ≡ 182 ⇒ x ≡ −182 ≡ 547 (mod 729).) This when substi-
tuted in the congruence (6.4) gives

P ≡ 547 (C − 2) (mod 272 ).

Getting back to P = “I (space)” = 242 in I LIKE IT, we get

C ≡ 4 · 242 + 2 (mod 272 )

≡ 241 (mod 272 ).

Now 241 = 8 · 27 + 25, and therefore it corresponds to the ordered pair IZ in

the ciphertext. (Here, I corresponds to 8 and Z corresponds to 25.) Similarly,
“LI,” “KE” and “(space)I” and “T(space)” correspond to SH, NS, YH and
ZZ, respectively. Thus, the ciphertext that corresponds to the plaintext “I
LIKE IT” is “IZSHNSYHZZ.” To get back the plaintext, we apply the inverse
transformation (6.4). As the numerical equivalent of “IZ” is 241, relation (6.4)
gives P ≡ 547 (241 − 2) ≡ 243 (mod 272 ) and this, as seen earlier, corresponds
to “I (space).” Similarly, the other pairs can be deciphered in order.
An equation of the form C = aP + b is known as an affine transformation.
Hence such cryptosystems are called affine cryptosystems.
In the Ceasar cryptosystem, given by the transformation f (a) ≡ a + 3
(mod 26), 3 is known as the key of the transformation. In the affine transfor-
mation, given by Equation 6.3, there are two keys, namely, a and b.

6.2.3 Private key cryptosystems

In the Caesar cryptosystem and the affine cryptosystem, the keys are
known to the sender and the receiver in advance. That is to say that what-
ever information the sender has with regard to his encryption, it is shared
by the receiver. For this reason, these cryptosystems are called private key
cryptosystems.
252 Discrete Mathematics

6.2.4 Hacking an affine cryptosystem

Suppose an intruder I (that is a person other than the sender A and
the receiver B) who has no knowledge of the private keys wants to hack the
message, that is, decipher the message stealthily. We may suppose that the
intruder knows the type of cryptosystem used by A and B including the unit
length of the system, though not the keys. Such an information may get leaked
out over a passage of time or may be obtained even by spying. How does I go
about hacking? He does it by a method known as frequency analysis.
Assume for a moment that the message units are of length 1. Look at a
long string of the ciphertext and find out the most-repeated character, the next
most-repeated character and so on. Suppose, for the sake of precision, they
are U, V, X, . . . . Now in the English language, the most common characters
of the alphabet of 27 letters consisting of the English characters A to Z and
“space” are known to be “space” and E in order. Then, “space” and E of
the plaintext correspond to U and V of the ciphertext, respectively. If the
cryptosystem used is the affine system given by the equation (see Table 6.1).

C = aP + b (mod 27),
we have 20 = a · 26 + b (mod 27),
and 21 = a · 4 + b (mod 27).

Subtraction yields
22a ≡ −1 (mod 27) (6.5)

As (22, 27) = 1, (6.5) has a unique solution, namely, a = 11. This gives
b = 21 − 4a = 21 − 44 = −23 = 4 (mod 27). The cipher has thus been hacked.

Suppose now the cryptosystem is based on an affine transformation C =

aP + b with unit length 2. If the same alphabet consisting of 27 characters
of this section (namely, A to Z and space) is used, each unit corresponds to
a unique non-negative integer less that 272 . Suppose the frequency analysis
of the ciphertext reveals that the most commonly occurring ordered pairs are
“CA” and “DX” in their decreasing orders of their frequencies. The decryption
transformation is of the form

P ≡ a C + b (mod 272 ) (6.6)

Here, a and b are the enciphering keys and a , b are the deciphering keys.
Now it is known that in the English language, the most frequently occurring
order pairs, in their decreasing orders of their frequencies, are “E(space)” and
“S(space).” Symbolically,

“E(space)” −→ CA, and

“S(space)” −→ DX.
 Cryptography 253

Writing these in terms of their numerical equivalents, we get

(4 × 27) + 26 = 134 −→ (2 × 27) + 0 = 54, and

(18 × 27) + 26 = 512 −→ (3 × 27) + 23 = 104. (6.7)

These, when substituted in Equation 6.5, give the congruences:

134 ≡ 54a + b (mod 729), and

512 ≡ 104a + b (mod 729).

Note that 729 = 272 . Subtraction gives

50a ≡ 378 (mod 729). (6.8)

As (50, 729) = 1, this congruence has a unique solution by the Extended

Euclidean Algorithm. In fact, a = 270 as Equation 6.8 implies that 50a ≡
33 × 14 (mod 36 ) ⇒ a = 33 c, where 50c ≡ 14(mod 33 ) ⇒ c = 10. This gives
that b ≡ 134 − 54a (mod 272 ) ⇒ b ≡ 134(mod 272 ), as 272 |54a ⇒ b =
4 × 27 + 26 ⇒ b = ESpace. Thus, the deciphering keys a and b have been
determined and the cryptosystem has been hacked.
In our case, the gcd(50, 729) happened to be 1 and hence we had no prob-
lem in determining the deciphering keys. If not, we have to try all the possible
solutions for a and take the plaintext that is meaningful. Instead, we can also
continue with our frequency analysis and compare the next most repeated
ordered pairs in the plaintext and ciphertext and get a third congruence and
try for a solution in conjunction with one or both of the earlier congruences.
If these also fail, we may have to adopt ad hoc techniques to determine a
and b .

6.3 Encryption Using Matrices

Assume once again that the message units are ordered pairs in the same
alphabet of size 27 of Section 6.2. We can use 2 by 2 matrices over the ring
Z27 to set up a private key cryptosystem in this case. In fact if A is any 2 by
2 matrix with entries
 from Z27 , and (X, Y ) is any plaintext unit, we encipher
X
it as B = A , where B is again a 2 by 1 matrix and therefore a ciphertext
Y
 
X
unit of length 2. If B = , we have the equations
Y
   
X X
=A ,
Y Y
   
X X
and = A−1 . (6.9)
Y Y
254 Discrete Mathematics

The first equation of (6.9) gives the encryption, while the second gives the
decryption. Notice that A−1 must be taken in Z27 . For A−1 to exist, we must
have gcd(det A, 27) = 1. If this were not the case, we may have to try once
again ad hoc methods.  
2 1
As an example, take A = . Then, det A = 2, and gcd(det A, 27) =
4 3
gcd(2, 27) = 1. Hence 2−1 exists: in fact, 2−1 = 14 ∈ Z27 . This gives (Recall
A−1 = (1/det A)(adjA)), where (adjA) = adjugate of the matrix A = (Bij ).
Here, Bij = (−1)i+j Aji where Aji is the cofactor of aji in A = (aij ) (see
Chapter 3). Hence
     
−1 3 −1 42 −14 15 13
A = 14 = = over Z27 . (6.10)
−4 2 −56 28 25 1

(The reader can verify that AA−1 = A−1 A = I2 (mod 27).)

Suppose, for instance, we want to encipher “HEAD” using the above matrix 
7
transformation. We proceed as follows: “HE” corresponds to the vector ,
4
 
0
and “AD” to the vector . Hence the enciphering transformation gives the
3
corresponding ciphertext as
     
7 0 7 0
A =A , A (mod 27)
4 3 4 3
     
2 1 7 2 1 0
= , (mod 27)
4 3 4 4 3 3
 
18 3
= (mod 27)
40 9
    
18 3 S D
= =
13 9 N J

Thus, the ciphertext of “HEAD” is “SNDJ.” We can decipher “SNDJ” in

exactly the same manner by taking A−1 in Z27 . This gives the plaintext
     
18 3 15 13
A−1 , A−1 , where A−1 = , as given by (6.10)
13 9 25 1

Therefore the plaintext is

       
450 162 7 0
, (mod 27) = ,
463 84 4 3

and this corresponds to “HEAD.”

 Cryptography 255

6.4 Exercises
 
17 5
1. Find the inverse of A = in Z27 .
8 7
 
12 3
2. Find the inverse of A = in Z29 .
5 17
3. Encipher the word “MATH” using the matrix A of Exercise 1 above as
the enciphering matrix in the alphabet A to Z of size 26. Check your
result by deciphering your ciphertext.

4. Solve the simultaneous congruences

x − y = 4 (mod 26)
7x − 4y = 10 (mod 26).

5. Encipher the word “STRIKES” using the affine transformation C ≡

4P + 7 (mod 272 ) acting on units of length 2 over an alphabet of size
27 consisting of A to Z and the exclamation mark ! with 0 to 25 and
26 as the corresponding numerals.

6. Suppose that we know that our adversary is using a 2 by 2 enciphering

matrix with a 29-letter alphabet, where A to Z have numerical equiv-
alents 0 to 25, (space)=26, ?=27 and !=28. We receive the message

AMGQTZAFJVMHQV

Suppose we know by some means that the last four letters of the plain-
text are our adversary’s signature “MIKE.” Determine the full plain-
text.

6.5 Other Private Key Cryptosystems

We now describe two other private key cryptosystems.

6.5.1 Vigenere cipher

In this cipher, the plaintext is in the English alphabet. The key consists of
an ordered set of d letters for some fixed positive integer d. The plaintext is
divided into message units of length d. The ciphertext is obtained by adding
the key to each message unit using modulo 26 addition.
For example, let d = 3 and the key be XYZ. If the message is “ABAN-
DON,” the ciphertext is obtained by taking the numerical equivalence of the
256 Discrete Mathematics

plaintext, namely,
[0][1][0] [13] [3] [14] [13],
and the addition modulo 26 of the numerical equivalence of “XYZ,” namely,
[23] [24] [25] of the key. This yields

[23] [25] [25] [36] [27] [39] [36] (mod 26)

= [23] [25] [25] [10] [1] [13] [10]
= XZZKBNK

as the ciphertext.

6.5.2 The one-time pad

This was introduced by Frank Miller in 1882 and reinvented by Gilbert
S. Vernam in 1917. The alphabet Σ for the plaintext is the set of 26 English
characters. If the message M is of length N , the key K is generated as a
pseudo-random sequence of characters of Σ also of the same length N . The
ciphertext is then obtained by the equation

C ≡M +K (mod 26)

Notwithstanding the fact that the key K is as long as the message M , the
system has its own drawbacks.

i. There are only standard methods of generating pseudo-random

sequences from out of Σ, and their number is not large.
ii. The long private key K must be communicated to the receiver in
advance.

Despite these drawbacks, this cryptosystem was said to be used in some high-
est levels of communication such as the Washington-Moscow hotline.
There are several other private key cryptosystems. The interested reader
can have a look at these in public domains.

6.6 Public Key Cryptography

All cryptosystems described so far are private key cryptosystems. “This
means that someone who has enough information to encipher messages has
enough information to decipher messages as well.” As a result, in private key
cryptography, any two persons in a group who want to communicate messages
in a secret way must have exchanged keys in a safe way (for instance, through
a trusted courier).
 Cryptography 257

In 1976, the face of cryptography got altered radically with the invention
of public key cryptography by Diffie and Hellman [36]. In this cryptosystem,
the encryption can be done by anyone. But the decryption can be done only
by the intended recipient who alone is in possession of the secret key.
At the heart of this cryptography is the concept of a “one-way func-
tion.” Roughly speaking, a one-way function is a 1–1 function f which is such
that whenever k is given, it is possible to compute f (k) “rapidly” while it is
“extremely difficult” to compute the inverse of f in a “reasonable” amount of
time. There is no way of asserting that such and such a function is a one-way
function since the computations depend on the technology of the day—the
hardware and the software. So what passes for a one-way function today may
fail to be a one-way function a few years later.
As an example of a one-way function, consider two large primes p and q
each having at least 500 digits. Then, it is “easy” to compute their product
n = pq. However, given n, there is no efficient factoring algorithm as on date
that would give p and q in a reasonable amount of time. The same problem
of forming the product pq with p and q having 100 digits had passed for a
one-way function in the 1980s but is no longer so today.

6.6.1 Working of public key cryptosystems

A public key cryptosystem works in the following way: Each person A
in a group has a public key PA and a secret key SA . The public keys are
made public as in a telephone register with PA given against the name A. A
computes his own secret key SA and keeps it within himself. The security of
the system rests on the fact that no person of the group other than A or an
intruder would be able to find out SA . The keys PA and SA are chosen to be
inverses of each other in that for any message M ,

(PA ◦ SA ) M = M = (SA ◦ PA ) M.

6.6.1.1 Transmission of messages

Suppose A wants to send a message M to B in a secured fashion. The
public key of B is SB which is known to everyone. A sends “PB · M ” to
B. Now, to decipher the message, B applies SB to it and gets SB (PB M ) =
(SB · PB )M = M . Note that none other than B can decipher the message sent
by A since B alone is in possession of SB .

6.6.1.2 Digital signature

Suppose A wants to send some instruction to a bank (for instance, transfer
an amount to Mr. C out of his account). If the intended message to the bank
is M , A applies his secret key SA to M and sends SA M to the bank. He
also gives his name for identification. The bank applies A’s public key PA
to it and gets the message PA (SA M ) = (PA ◦ SA )M = M . This procedure
258 Discrete Mathematics

also authenticates A’s digital signature. This is in fact the method adopted in
credit cards.
We now describe two public key cryptosystems. The first is RSA, after their
inventors, Rivest, Shamir and Adleman. In fact, Diffie and Hellman, though
they invented public key cryptography in 1976, did not give the procedure to
implement it. Only Rivest, Shamir and Adleman did it in 1978, two years later.

6.6.2 RSA public key cryptosystem

Suppose there is a group of people who want to communicate among them-
selves secretly. In such a situation, RSA is the most commonly used public key
cryptosystem. The length of the message units is fixed in advance as also the
alphabet in which the cryptosystem is operated. If, for instance, the alphabet
consists of the English characters and the unit length is k, then any message
unit is represented by a number less than 26k .

6.6.2.1 Description of RSA

We now describe RSA.

i. Each person A (traditionally called Alice) chooses two large distinct

primes p and q and computes their product n = pq, where p and q are
so chosen that n > N , where N is a very large positive integer.

ii. Each A chooses a small positive integer e, 1 < e < φ(n), such that
(e, φ(n)) = 1, where the Euler function φ(n) = φ(pq) = φ(p)φ(q) =
(p − 1)(q − 1). (e is odd as φ(n) is even).

iii. As (e, φ(n)) = 1, by Extended Euclidean Algorithm [25], e has a multi-

plicative inverse d modulo φ(n), that is,
ed ≡ 1 (mod φ(n)).
iv. A (Alice) gives the ordered pair (n, e) as her public key and keeps d as
her private (secret) key.
v. Encryption P (M ) of the message unit M is done by
P (M ) ≡ M e (mod n), (6.11)
 
while decryption S(M ) of the cipher text unit M is given by
d
S(M  ) ≡ M  (mod n). (6.12)
Thus, both P and S (of A) act on the ring Zn . Before we establish the cor-
rectness of RSA, we observe that d (which is computed using the Extended
Euclidean Algorithm) can be computed in O(log3 n) time. Further powers of
d
M e and M  modulo n in Equations 6.11 and 6.12 can also be computed in
3
O(log n) time [25]. Thus, all computations in RSA can be done in polynomial
time.
 Cryptography 259

Theorem 6.1 (Correctness of RSA). Equations 6.11 and 6.12 are indeed
inverse transformations.
Proof. We have

S (P (M )) = (S(M e )) ≡ M ed (mod n).

Hence it suffices to show that

M ed ≡ M (mod n).

Now, by the definition of d,

ed ≡ 1 (mod φ(n)).

But φ(n) = φ(pq) = φ(p)φ(q) = (p − 1)(q − 1),

and therefore, ed = 1 + k(p − 1)(q − 1) for some integer k. Hence

M ed = M 1+k(p−1)(q−1)
= M · M k(p−1)(q−1) .

By Fermat’s Little Theorem (FLT), if (M, p) = 1,

M p−1 ≡ 1 (mod p)

and therefore,

M ed = M ·M k(p−1)(q−1) ≡ M ·(M (p−1) )k(q−1) ≡ M ·(1)k(q−1) ≡ M (mod p).

If, however, (M, p) = 1, then (as p is a prime) (M, p) = p, and trivially (as p
is a divisor of M )
M ed ≡ M (mod p).
Hence, in both the cases,

M ed ≡ M (mod p). (6.13)

For a similar reason,

M ed ≡ M (mod q). (6.14)
As p and q are distinct primes, the congruences (6.13) and (6.14) imply that

M ed ≡ M (mod pq),
ed
so that M ≡M (mod n).

The above description shows that if Bob wants to send the message M to
Alice, he will send it as M e (mod n) using the public key of Alice. To decipher
the message, Alice will raise this number to the power d and get M ed ≡ M
(mod n), the original message of Bob.
260 Discrete Mathematics

The security of RSA rests on the supposition that none other than Alice
can determine the private key d of Alice. A person can compute d if he/she
knows φ(n) = (p − 1)(q − 1) = n − (p + q) + 1, that is to say, if he/she knows
the sum p + q. For this, he should know the factors p and q of n. Thus, in
essence, the security of RSA is based on the assumption that factoring a large
number n that is a product of two distinct primes is “difficult.” However, to
quote Koblitz [30], “no one can say with certainty that breaking RSA requires
factoring n. In fact, there is even some indirect evidence that breaking RSA
cryptosystem might not be quite as hard as factoring n. RSA is the public
key cryptosystem that has had by far the most commercial success. But,
increasingly, it is being challenged by elliptic curve cryptography.”

6.6.3 The ElGamal public key cryptosystem

We have seen that RSA is based on the premise that factoring a very
large integer which is a product of two “large” primes p and q is “difficult”
compared to forming their product pq. In other words, given p and q, finding
their product is a one-way function. ElGamal public key cryptosystem uses
a different one-way function, namely, a function that computes the power of
an element of a large finite group G. In other words, given G, g ∈ G, g = e,
and a positive integer a, ElGamal cryptosystem is based on the assumption
that computation of g a = b ∈ G is “easy” while given b ∈ G and g ∈ G, it is
“difficult” to find the exponent a.

Definition 6.1. Let G be a finite group and b ∈ G. If y ∈ G, then the discrete

logarithm of y with respect to base b is any non-negative integer x less than
o(G), the order of G, such that bx = y, and we write logb y = x.

As per the definition, logb y may or may not exist. However, if we take
G = Fq∗ , the group of non-zero elements of a finite field Fq of q elements and g,
a generator of the cyclic group Fq∗ (see [25]), then for any y ∈ Fq∗ , the discrete
logarithm logg y exists.

Example 6.1
∗ ∗
5 is a generator of F17 .
In F17 , the discrete logarithm of 12 with
∗
respect to base 5 is 9. In symbols: log5 12 = 9. In fact, in F17 ,

5
= 51 = 5, 52 = 8, 53 = 6, 54 = 13, 55 = 14, 56 = 2, 57 = 10, 58 = −1,

59 = 12, 510 = 9, 511 = 11, 512 = 4, 513 = 3, 514 = 15, 515 = 7, 516 = 1
This logarithm is called “discrete logarithm,” as it is taken in a finite
group.
 Cryptography 261

6.6.4 Description of ElGamal system

The ElGamal system works in the following way: All the users in the
system agree to work in an already chosen large finite field Fq . A generator g
of Fq∗ is fixed once and for all. Each message unit is then converted into a
number of Fq . For instance, if the alphabet is the set of English characters
and if each message unit is of length 3, then the message unit BCD will have
the numerical equivalent 262 · 1 + 261 · 2 + 3 (mod q). It is clear that in order
that these numerical equivalents of the message units are all distinct, q should
be quite large. In our case, q ≥ 263 . Now each user A in the system randomly
chooses an integer a = aA , 0 < a < q − 1, and keeps it as his or her secret key.
A declares g a ∈ Fq as his public key.
If B wants to send the message unit M to A, he chooses a random positive
integer k, k < q − 1, and sends the ordered pair
 k 
g , M g ak (6.15)
to A. Since B knows k, and since g a is the public key of A, B can compute
g ak . How will A decipher B’s message? She will first raise the first number
of the pair given in Equation 6.15 to the power a and compute it in Fq∗ . She
will then divide the second number M g ak of the pair by g ak and get M . A
can do this as she has a knowledge of a. An intruder who gets to know the
pair (g k , M g ak ) cannot find a = loggk (g ak ) ∈ Fq∗ , since the security of the
system rests on the premise that finding discrete logarithm is “difficult,” that
is, given h and ha in Fq , there is no efficient algorithm to determine a.
There are other public key cryptosystems as well. The interested reader
can refer to [30].

6.7 Primality Testing

6.7.1 Non-trivial square roots (mod n)
We have seen that the most commonly applied public key cryptosystem,
namely, the RSA, is built up on very large prime numbers (numbers having,
say, 500 digits and more). So there arises the natural question: Given a large
positive integer, how do we know that it is a prime or not? A ‘primality test’
is a test that tells if a given number is a prime or not.
Let n be a prime, and a, a positive integer with a2 ≡ 1 (mod n). Then,
a is called a square root mod n. This means that n divides (a − 1)(a + 1),
and so, n|(a − 1) or n|(a + 1); in other words, a ≡ ±1 (mod n). Conversely,
if a ≡ ±1 (mod n), then a2 ≡ 1 (mod n). Hence a prime number has only
the trivial square root 1 and −1 modulo n. However, the converse may not be
true, that is, there exist composite numbers m having only trivial square roots
modulo m; for instance, m = 10 is such a number. On the other hand, 11 is a
non-trivial square root of the composite number 20 since 11 ≡ ±1 (mod 20)
while 112 ≡ 1 (mod 20).
262 Discrete Mathematics

Consider the modular exponentiation algorithm which determines ac

(mod n) in O(log2 n log c) time [25]. At any intermediate stage of the algo-
rithm, the output i is squared, taken modulo n and then multiplied to a or
1 as the case may be. If the square i2 (mod n) of the output i is 1 mod-
ulo n, then already we have determined a non-trivial square root modulo n,
namely, i. Therefore, we can immediately conclude that n is not a prime and
therefore a composite number. This is one of the major steps in the Miller-
Rabin Primality Testing Algorithm to be described below.

6.7.2 Prime Number Theorem

For a positive real number x, let π(x) denote the number of primes less
than or equal to x. The Prime Number Theorem states that π(x) is asymptotic
to x/ log x; in symbols, π(x) ≈ x/ log x. Here, the logarithm is with respect to
base e. Consequently, π(n) ≈ n/ log n, or, equivalently, π(n)/n ≈ 1/log n. In
other words, in order to find a 100-digit prime, one has to examine roughly
loge 10100 ≈ 230 randomly chosen 100-digit numbers for primality (Recall that
any 100-digit number k satisfies the inequalities 1099 ≤ k < 10100 ) (this figure
may drop down by half if we omit even numbers), (cf. [37]).

6.7.3 Pseudo-primality testing

Fermat’s Little Theorem (FLT) states that if n is prime, then for each a,
1 ≤ a ≤ n − 1,
an−1 ≡ 1 (mod n). (6.16)

Note that for any given a, an−1 (mod n) can be computed in polynomial time
using the repeated squaring method [25]. However, the converse of FLT is not
true. This is because of the presence of Carmichael numbers. A Carmichael
number is a composite number n satisfying (6.16) for each a prime to n. They
are sparse but are infinitely many. The first few Carmichael numbers are 561,
1105, 1729 (the smallest 4-digit Ramanujan number).
Since we are interested in checking if a given large number n is prime
or not, n is certainly odd and hence (2, n) = 1. Consequently, if 2n−1 ≡ 1
(mod n), we can conclude with certainty, in view of FLT, that n is composite.
However, if 2n−1 ≡ 1 (mod n), n may be a prime or not. If n is not a prime
but 2n−1 ≡ 1 (mod n), then n is called a pseudo-prime with respect to base b.

Definition 6.2. n is called a pseudo-prime to base a, where (a, n) = 1, if

i. n is composite, and

ii. an−1 ≡ 1 (mod n).

In this case, n is also called a base a pseudo-prime.

 Cryptography 263

6.7.3.1 Base-2 Pseudo-prime test

Given an odd positive integer n, check if 2n−1 ≡ 1 (mod n). If yes, n is
composite. If no, 2n−1 ≡ 1 (mod n) and n may be a prime.
But then there is a chance that n is not a prime. How often does this hap-
pen? For n < 10000, there are only 22 pseudo-primes to base 2. They are 341,
561, 645, 1105, . . . . Using better estimates due to Carl Pomerance (See [27]),
we can conclude that the chance of a randomly chosen 50-digit (resp. 100-digit)
number satisfies (6.16) but fails to be a prime is < 10−6 (resp. < 10−13 ).
More generally, if (a, n) = 1, 1 < a < n, the pseudo-prime test with
reference to base a checks if an−1 ≡ 1 (mod n). If true, a is composite; if not,
a may be a prime.

6.7.4 Miller-Rabin Algorithm

The Miller-Rabin Algorithm is a randomized algorithm for primality test-
ing. A randomized algorithm is an algorithm which during the execution of
the algorithm makes at least one random choice. It is in contrast with the
deterministic algorithms in which every step has a unique subsequent step. A
randomized polynomial time algorithm is a randomized algorithm which runs
in polynomial time (RP for short). We now give the precise definition of a
randomized algorithm.
Definition 6.3 (RP algorithm). Consider a decision problem, that is, a
problem whose answer is either “yes” or “no.” (For example, given an integer
n ≥ 2, we would like to know if n is prime or composite). A randomized
polynomial time algorithm is a randomized algorithm which runs in polynomial
time in the worst case such that for any input I, the probability that the answer
is “yes” for the input I is ≥ 1/2 and the probability that the answer is “no”
is 0. That is, the algorithm errs in case of output “yes” and makes no error
in case of ouput “no.”

A randomized algorithm for primality test can be devised subject to the

following conditions:

1. The existence of a set S which contains a large number of “certificates”

for the proof of primality of a given number p.
2. The set S can be sampled efficiently.

6.7.5 Horner’s method to evaluate a polynomial

The Miller-Rabin method uses the Horner’s method to evaluate a polyno-
mial at a given point. It is a classical algorithm based on parenthesizing. To
evaluate a polynomial of degree n, Horner’s method uses only n additions and
n multiplications unlike the naive method which uses n additions but 2n − 1
multiplications. This method is also called synthetic division. Let us illustrate
this method by an example.
264 Discrete Mathematics

TABLE 6.2: Horner’s

method
t=p[n];
for(i = n − 1;i ≥ 0;−−i)
{ t=t*c; t=t+p[i];
}
printf(“%d”, t)

Example 6.2 Horner’s method

Let us evaluate p(x) = 3x4 − 5x3 + 6x2 + 2x + 7 at x = 2. First we
factorize the polynomial p(x) from right, that is, we write:
p(x) = (((3x − 5)x + 6)x + 2)x + 7.

n
More generally, we write a polynomial p(x) = i=0 pi xi of degree n as

p(x) = (((...(pn x + pn−1 )x + pn−2 )x + pn−3 )x + · · · )x + p0

Note that there are (n − 1) opening parentheses “(” and (n − 1) closing paren-
theses “)” . Let us write a pseudo-code for Horner’s method:
n
Input: A polynomial p(x) = i=0 pi xi of degree n and a constant c.
Output: The value of p(c).
The algorithm is given in Table 6.2.
Clearly, in the loop of the algorithm of Table 6.2, there are n additions
and n multiplications. The Miller-Rabin algorithm is based on the following
result in number theory.

Theorem 6.2. If there is a non-trivial square root of 1 with respect to modulo

n, then n is a composite number. In notation, if the equation

x2 ≡ 1 (mod n)

has a solution other than +1 and −1, then n is a composite number.

For example, the equation x2 ≡ 1 (mod 8) has a solution x = 3 (other

than ±1) because 32 ≡ 1 (mod 8). Here, n = 8, a composite number. In the
algorithm we use the method of converting an integer n into its binary rep-
resentation: Just divide successively n by 2, keeping the remainders, until we
arrive at the quotient 0. To get the binary representation of n, we now sim-
ply write the remainders obtained during the division process, in the reverse
manner. For example, if n = 25, we get the successive remainders: 1,0,0,1,1.
Writing these in reverse order, we get: (11001)2 (the subscript says that the
number is in base 2) as the binary representation of 25.
 Cryptography 265

More generally, the binary representation of a non-negative integer b is

written as

b = (bp bp−1 bp−2 . . . b2 b1 b0 )2

b0 is the least significant digit and bp is the most significant digit. We observe
by Horner’s method b = bp 2p + bp−1 2p−1 + · · · + b2 22 + b1 2 + b0 which is also
equal to (by Horner’s method)

(((...(bp 2 + bp−1 )2 + bp−2 )2 + bp−3 )2 · · · + b1 )2 + b0 .

We need one more algorithm to state Miller-Rabin algorithm: It is called the

modular-exponentiation algorithm:

6.7.6 Modular exponentiation algorithm based on repeated

squaring
Input: Three integers a, b, n where a and b are non-negative integers and n is
a positive integers.
Output: ab mod n.
Algorithm: The naive strategy of multiplying a by itself uses b − 1 multi-
plications which is an exponential algorithm because the number of bits to
represent b in binary is log2 b + 1 and b − 1 = 2log2 b − 1 which is an expo-
nential in log2 b.
Instead, the algorithm we study is a polynomial algorithm based on ‘ ‘repeated
squaring.” The algorithm is given in Table 6.3. It is based on the equation
which uses the Horner’s method:

ab = a(((...(bp 2+bp−1 )2+bp−2 )2+bp−3 )2···+b1 )2+b0

where b = (bp bp−1 bp−2 . . . b2 b1 b0 ) is the binary representation of b. The

repeated multiplication by 2 in the power of the above equation for ab trans-
lates into repeated squaring.
Complexity of the Modular exponentiation: The loop is executed exactly
p + 1 times and p = log2 b + 1, which is polynomial in log2 b.

TABLE 6.3: Modular exponentiation

Write b = (bp bp−1 bp−2 . . . b2 b1 b0 )2 in binary representation.
//initialization of v. At the end of algorithm v = ab .
v=1;//empty product is defined as 1 in the same spirit as 0!=1 or 20 = 1.
for (i = p;i ≥ 0; −−i)
{ v = v ∗ v mod n;//square v
if (bi == 1)
v = v ∗ a mod n //v = a(bp bp−1 ···bi )2 . bi = 0 has no effect on v.
}
return v;
266 Discrete Mathematics

TABLE 6.4: Certificate of compositeness

int certificate(int a, int n)
{Express n − 1 = (bp bp−1 bp−2 . . . b2 b1 b0 )2 in binary notation
v=1;
for (i = p; i ≥ 0;−−i)
t=v; v = v ∗ v mod n;//save in v square v
if ((v == 1)&&(t! = 1)&&(t! = n − 1))// non-trivial square root of 1?
return 1;//1=true
if (bi == 1) v = v ∗ a mod n
}
if (v! = 1) return 1;//1=true because an−1 mod n = 1 by Fermat’s theorem
return 0://0=false
}

TABLE 6.5: Miller-Rabin primality testing algorithm

int miller rabin(int n, int m)
{ for (i = 1;i <= m;+ + i)
{ a=(rand()% (n − 1)+1;
if (certificate(a,n)) return 1;//function call true. surely composite
return 0;//false. Not composite with probability >= 1 − 1/2m
}
}

We are now ready to present the Miller-Rabin algorithm. The algorithm

uses a function called “certificate(a,n).” The function certificate(a,n) returns
the value “true” if and only if a can be used to certify that n is a composite
number, that is, we can use a to prove that n is composite. Table 6.4 gives
the function “certificate(a,n),” where n is assumed to be an odd integer. For
example, a divisor a = 1, n of an integer n is a certificate of compositeness
of n, because one can indeed verify if a divides n by the usual division. The
Miller-Rabin algorithm is the same as the modular exponentiation, except
that it checks if a non-trivial square root of 1 is obtained after every squaring
and if so it stops with output “true,” by Theorem 6.2. In the algorithm, we
use Fermat’s little theorem in the contrapositive form: If there is an integer
a < n such that an−1 mod n = 1 then n is a composite number. Note that
n − 1 ≡ −1 (mod n).
We can now write the Miller-Rabin algorithm (See Table 6.5). The algo-
rithm tests if n is a composite number. It generates randomly m integers a
between 1 and n − 1 and tests if each random integer is a certificate for the
compositeness of n. This algorithm is based on the following fact: By repeating
a randomized algorithm several times, the error probability can be reduced
arbitrarily. Of course, repetition increases the running time. Further the algo-
rithm uses the following result:
 Cryptography 267

Theorem 6.3. For an odd composite number n, the number of certificates for
the proof of compositeness of n is ≥ n−1
2 .

6.8 The Agrawal-Kayal-Saxena (AKS) Primality

Testing Algorithm
6.8.1 Introduction
The Miller-Rabin primality testing algorithm is a probabilistic algorithm
that uses Fermat’s Little Theorem. Another probabilistic algorithm is due
to Solovay and Strassen. It uses the fact that if n is an odd prime, then
a(n−1)/2 ≡ (a/n) (mod n), where (a/n) stands for the Legendre symbol. The
Miller-Rabin primality test is known to be the fastest randomized primal-
ity testing algorithm, to within constant factors. However, the question of
determining a polynomial time algorithm [25] to test if a given number is
prime or not remained unsolved until July 2002. In August 2002, Agrawal,
Kayal and Saxena of the Indian Institute of Technology, Kanpur, India made
a sensational revelation that they have found a polynomial time algorithm for
primality testing which works in Õ (log10.5 n) time (where Õ (f (n)) stands
for O (f (n) · polynomial in log f (n))). (Subsequently, this has been reduced
to Õ (log7.5 n).) It is based on a generalization of Fermat’s Little Theorem
to polynomial rings over finite fields. Notably, the correctness proof of their
algorithm requires only simple tools of algebra. In the following section, we
present the details of the AKS algorithm.

6.8.2 The basis of AKS algorithm

The AKS algorithm is based on the following identity for prime numbers
which is a generalization of Fermat’s Little Theorem.

Lemma 6.1. Let a ∈ Z, n ∈ N, n ≥ 2, and (a, n) = 1. Then, n is prime if

and only if
n
(X + a) ≡ X n + a (mod n). (6.17)

Proof. We have

n

n−1
n

(X + a) = X n + X n−1 ai + an .
n=1
i

 
If n is prime, each term ni , 1 ≤ i ≤ n − 1, is divisible by n. Further, as
(a, n) = 1, by Fermat’s Little Theorem, an ≡ a (mod n). This establishes
(6.17) as we have (X + a)n ≡ X n + an (mod n) ≡ X n + a (mod n).
268 Discrete Mathematics

If n is composite, n has a prime factor

 n−qq < n. Let q k ||n (that is, q k |n but
k+1 n
q  | n). Now consider the term q X a in the expansion of (X + a)n .
q

We have

n n(n − 1) · · · (n − q + 1)
= .
q 1 · 2···q

     

Then, q k  | nq . For if q k  nq , as q k ||n , (n−1) · · · (n−q +1) must be divisible
by q, a contradiction (As q is a prime and q|(n − 1)(n − 2) . . . (n − q + 1),
q must divide at least one of the factors, which is impossible since each term
of (n − 1)(n − 2) . . . (n − q + 1) is of the form n − (q −  k), 1 ≤ k ≤ q − 1).
Hence q k , and therefore n does not divide the term nq X n−q aq . This shows
that (X +a)n −(X n +a) is not identically zero over Zn (note that if an integer
divides a polynomial with integer coefficients, it must divide each coefficient
of the polynomial).

The above identity suggests a simple test for primality: Given input n,
choose an a and test whether the congruence (6.17) is satisfied, However,
this takes time Ω(n) because we need to evaluate n coefficients in the LHS
of congruence 6.17 in the worst case. A simple way to reduce the number of
coefficients is to evaluate both sides of Equation 6.17 modulo a polynomial of
the form X r − 1 for an appropriately chosen small r. In other words, test if
the following equation is satisfied:

n
(X + a) = X n + a (mod X r − 1, n) (6.18)

From Lemma 6.1, it is immediate that all primes n satisfy Equation 6.18 for all
values of a and r. The problem now is that some composites n may also satisfy
Equation 6.18 for a few values of a and r (and indeed they do). However, we
can almost restore the characterization: we show that for an appropriately
chosen r if Equation 6.18 is satisfied for several a’s, then n must be a prime
power. It turns out that the number of such a’s and the appropriate r are both
bounded by a polynomial in log n, and this yields a deterministic polynomial
time algorithm for testing primality.

6.8.3 Notation and preliminaries

Fp denotes the finite field with p elements, where p is prime. Recall that if p
is prime and h(x) is a polynomial of degree d irreducible over Fp , then Fp [X]/
(h(X)) is a finite field of order pd . We will use the notation, f (X) = g(X)
(mod h(X), n) to represent the equation f (X) = g(X) in the ring Zn [X]/
(h(X)), that is, if the coefficients of f (X), g(X) and h(X) are reduced modulo
n, then h(X) divides f (X) − g(X).
 Cryptography 269
 
As mentioned in Section 6.8.1, for
any function f (n) of n, Õ f (n) stands
 
for O f (n) · polynomial in log f (n) . For example,
   
Õ(logk n) = O logk n · poly log logk n
   k 
= O logk n · poly log log n
   
= O logk n · poly k log log n
  
= O logk n · poly log log n

= O logk+ n for any  > 0.

All logarithms in this section are with respect to base 2.

Given r ∈ N, a ∈ Z with (a, r) = 1, the order of a modulo r is the smallest
number k such that ak ≡ 1 (mod r) (note that such a k exists by Euler’s
Theorem [6]). It is denoted Or (a). φ(r) is Euler’s totient function. Since by
Euler’s theorem, aφ(r) ≡ 1 (mod r), and since aOr (a) ≡ 1 (mod r), by the
definition of Or (a), we have Or (a) | φ(r).
We need the following result of Nair (see [33]) on the least common multiple
(lcm) of the first m natural numbers.

Lemma 6.2. Let LCM(m) denote the lcm of the first m natural numbers.
Then, for m ≥ 9, LCM(m) ≥ 2m .

6.8.4 The AKS algorithm

Input, integer n > 1
1 If (n = ab for a ∈ N and b > 1), output COMPOSITE
2 Find the smallest r such that Or (n) > 4 log2 n
3 If 1 < gcd(a, n) < n for some a ≤ r, output COMPOSITE
4 If n ≤ r, outputPRIME
5 For a = 1 to 2 φ(r) · log n, do if (X + a)n = X n + a (mod X r − 1, n),
output COMPOSITE
6 output PRIME
Theorem 6.4. The AKS algorithm returns PRIME iff n is prime.

We now prove Theorem 6.4 through a sequence of lemmas.

Lemma 6.3. If n is prime, then the AKS algorithm returns PRIME.

Proof. If n is prime, we have to show that AKS will not return COMPOSITE
in steps 1, 3 and 5. Certainly, the algorithm will not return COMPOSITE in
step 1 (as no prime n is expressable as ab , b > 1). Also, if n is prime, there
exists no a such that 1 < gcd(a, n) < n, so that the algorithm will not return
COMPOSITE in step 3. By Lemma 6.1, the ‘For loop’ in step 5 cannot return
270 Discrete Mathematics

COMPOSITE. Hence the algorithm will identify n as PRIME either in step 4

or in step 6.

We now consider the steps when the algorithm returns PRIME, namely,
steps 4 and 6. Suppose the algorithm returns PRIME in step 4. Then, n
must be prime. If n were composite, n = n1 n2 , where 1 < n1 , n2 < n.
Then, as n ≤ r, if we take a = n1 , we have a ≤ r. So in step 3, we would
have had 1 < (a, n) = a < n, a ≤ r. Hence the algorithm would have
output COMPOSITE in step 3 itself. Thus, we are left out with only one
case, namely, the case if the algorithm returns PRIME in step 6. For the
purpose of subsequent analysis, we assume this to be the case.
The algorithm has two main steps (namely, 2 and 5). Step 2 finds an
appropriate r and step 5 verifies Equation 6.18 for a number of a’s. We first
bound the magnitude of r.

Lemma 6.4. There exists an r ≤ 16 log5 n + 1 such that Or (n) > 4 log2 n.

Proof. Let r1 , . . . , rt be all the numbers such that Ori (n) ≤ 4 log2 n for each i
and therefore ri divides αi = (nOri (n) −1) for each i (recall that if Ori (n) = ki ,
then nki − 1 is divisible by ri ). Now for each i, αi divides the product

4 log2 n

P = (ni − 1)
i=1

t 2
We now use the fact that i=1 (ni − 1) < nt , the proof of which follows
readily by induction on t. Hence,
4 4 5
P < n16 log n
= (2log n )16 log n
= 216 log n
.

As ri divides αi and αi divides P for each i, 1 ≤ i ≤ t, the lcm of the ri ’s also

5
divides P . Hence (lcm of the ri ’s)< 216 log n . However, by Lemma 6.2,

  log5 n
lcm 1, 2, . . . , 16 log5 n ≥ 216 2 
.
 
Hence there must exist a number r in 1, 2, . . . , 16 log5 n , that is, r ≤
16 log5 n + 1 such that Or (n) > 4 log2 n.

Let p be a prime divisor of n. We must have p > r. For, if p ≤ r, (then

as p < n), n would have been declared COMPOSITE in step 3, while if
p = n ≤ r, n would have been declared PRIME in step 4. This forces that
(n, r) = 1. Otherwise, there exists a prime divisor p of n and r, and hence
p ≤ r, a contradiction as seen above. Hence (p, r) is also
equal to 1. We fix p
and r for the remainder of this section. Also, let l = 2 φ(r) log n.
 Cryptography 271

Step 5 of the algorithm verifies l equations. Since the algorithm does not
output COMPOSITE in this step (recall that we are now examining step 6),
we have
n
(X + a) = X n + a (mod X r − 1, n)
for every a, 1 ≤ a ≤ l. This implies that
n
(X + a) = X n + a (mod X r − 1, p) (6.19)

for every a, 1 ≤ a ≤ l. By Lemma 6.1, we have

p
(X + a) = X p + a (mod X r − 1, p) (6.20)

for 1 ≤ a ≤ l. Comparing Equation 6.19 with Equation 6.20, we notice that n

behaves like prime p. We give a name to this property:
Definition 6.4. For polynomial f (X) and number m ∈ N, m is said to be
introspective for f (X) if
m
[f (X)] = f (X m ) (mod X r − 1, p)

It is clear from Equations 6.19 and 6.20 that both n and p are introspective
for X + a, 1 ≤ a ≤ l. Our next lemma shows that introspective numbers are
closed under multiplication.

Lemma 6.5. If m and m are introspective numbers for f (X), then so

is mm .
Proof. Since m is introspective for f (X), we have
m
f (X m ) = (f (X)) (mod X r − 1, p)
 
m
and hence [f (X m )] = f (X)mm (mod X r − 1, p). (6.21)

Also, since m is introspective for f (X), we have

 

f X m = f (X)m (mod X r − 1, p).

Replacing X by X m in the last equation, we get


 m
f X mm = f (X m ) (mod X mr − 1, p).

 m
and hence f X mm ≡ f (X m ) (mod X p − 1, p) (6.22)

(since X r − 1 divides X mr − 1). Consequently from Equations 6.21 and 6.22


mm 
(f (X)) = f X mm (mod X r − 1, p).

Thus, mm is introspective for f (X).

272 Discrete Mathematics

Next we show that for a given number m, the set of polynomials for which
m is introspective is closed under multiplication.

Lemma 6.6. If m is introspective for both f (X) and g(X), then it is also
introspective for the product f (X)g(X).

Proof. The proof follows from the equation:

m
[f (X) · g(X)] = [f (X)]m [g(X)]m = f (X m )g(X m ) (mod X r − 1, p).

Equations 6.19 and 6.20 together imply that both n and p are intro-
spective for  (X + a). Hence  by Lemmas 6.5 and 6.6, every number in the
set I = ni pj : i, j ≥ 0 is introspective
 for every polynomial in the set
l ea
P = a=1 (X + a) : ea ≥ 0 . We now define two groups based on the
sets I and P that will play a crucial role in the proof.
The first group consists of the set G of all residues of numbers in I mod-
ulo r. Since both n and p are prime to r, so is any number in I. Hence G ⊂ Zr∗ ,
the multiplicative group of residues modulo r that are relatively prime to r.
It is easy to check that G is a group. The only thing that requires verification
is that ni pj has a multiplicative inverse in G. Since nOr (n) ≡ 1 (mod r), there
 
exists i , 0 ≤ i < Or (n) such that ni = ni . Hence inverse of ni (= ni ) is

n(Or (n)−i ) . A similar argument applies for p as pOr (n) = 1. Let |G| = the
order of the group G = t (say). As G is generated by n and p modulo r and
since Or (n) > 4 log2 n, t > 4 log2 n. (Recall that all our logarithms are w.r.t.
base 2.)
To define the second group, we need some basic facts about cyclotomic
polynomials over finite fields. Let Qr (X) be the r-th cyclotomic polynomial
over the field Fp [28]. Then, Qr (X) divides X r − 1 and factors into irreducible
factors of the same degree d = Or (p). Let h(X) be one such irreducible factor
of degree d. Then, F = Fp [X]/(h(X)) is a field. The second group that we
want to consider is the group generated by X + 1, X + 2, . . . , X + l in the
multiplicative group F ∗ of non-zero elements of the field F . Hence it consists
of simply the residues of polynomials in P modulo h(X) and p. Denote this
group by G.
We claim that the order of G is exponential in either t = |G| or l.
 
Lemma 6.7. |G| ≥ min 2t − 1, 2l .

Proof. First note that h(X) | Qr (X) and Qr (X) | (X r − 1). Hence X may be
taken as a primitive r-th root of unity in F = Fp [X]/(h(X)).
We claim:
(*) if f (X) and g(X) are polynomials of degree less than t and if f (X) = g(X)
in P , then their images in F (got by reducing the coefficients modulo p and
then taking modulo h(X)) are distinct.
 Cryptography 273

To see this, assume that f (X) = g(X) in the field F (that is, the images of
f (X) and g(X) in the field F are the same). Let m ∈ I. Recall that every
number of I is introspective with respect to every polynomial in P . Hence m
is introspective with respect to both f (X) and g(X). This means that

f (X)m = f (X m ) (mod X r − 1, p),

and g(X)m = g(X m ) (mod X r − 1, p).
Consequently, f (X m ) = g(X m ) (mod X r − 1, p),

and since h(X) | (X r − 1),

f (X m ) = g(X m ) (mod h(X), p).

In other words f (X m ) = g(X m ) in F , and therefore X m is a root of the

polynomial Q(Y ) = f (Y ) − g(Y ) for each m ∈ G. As X is a primitive r-th
root of unity and for each m ∈ G, (m, r) = 1, X m is also a primitive r-th
root of unity. Since each m ∈ G is reduced modulo r, the powers X m , m ∈ G,
are all distinct. Thus, there are at least t = |G| primitive r-th roots of unity
X m , m ∈ G, and each one of them is a root of Q(Y ). But since f and g are
of degree less than t, Q(Y ) has degree less than t. This contradiction shows
that f (X) = g(X) in F . This establishes (*).
We next observe that the numbers 1, 2, . . . , t are all distinct in Fp . This
is because
 √
l = 2 φ(r) log n < 2 r log n
√
√ t
<2 r (as t > 4 log2 n)
2
< r (as t < r; recall that G is a subgroup of Zr∗ )
< p (by assumption on p).

Hence {1, 2, . . . , l} ⊂ {1, 2, . . . , p − 1}. This shows that the elements X + 1,

X + 2, . . . , X + l are all distinct in Fp [X] and therefore in Fp [X]/h(X) = F .
If t ≤ l, then all the possible products of the polynomials in the set
{X + 1, X + 2, . . . X + t} except the one containing all the t of them are all
distinct and each of them is of degree less than t. Their number is 2t − 1 and
all of them belong to P . By (*), their images in F are all distinct, that is,
|G| ≥ 2t − 1. If t > l, then there exist at least 2l such polynomials (namely,
the product of all subsets of {X + 1, . . . , X + l}). These products are all of
l
 t l andl hence of degree < t. Hence in this case, |G| ≥ 2 . Thus,
degree at most
|G| ≥ min 2 − 1, 2 .

Finally we show that if n is not a prime power, then |G| is bounded above
by a function of t.
√
Lemma 6.8. If n is not a prime power, |G| ≤ (1/2)n2 t
.
274 Discrete Mathematics
 
Proof. Set Iˆ = ni · pj : 0 ≤ i, j ≤ t . If n is not a prime power (recall that
2
p|n), the number of terms in Iˆ = ( t + 1) > t. When reduced mod r, the
ˆ
elements of I give elements of G. But |G| = t. Hence there exist at least two
distinct numbers in Iˆ which become equal when reduced modulo r. Let them
be m1 , m2 with m1 > m2 . So we have (since r divides (m1 − m2 )),

X m1 = X m2 (mod X r − 1) (6.23)

Let f (X) ∈ P . Then,

m1
[f (X)] = f (X m1 ) (mod X r − 1, p)
= f (X m2 ) (mod X r − 1, p) by (6.23)
m
= [f (X)] 2 (mod X r − 1, p)
m
= [f (X)] 2 (mod h(X), p)(since
h(X) | (X r − 1)).
m1 m2
This implies that [f (X)] = [f (X)] in the field F . (6.24)

Now f (X) when reduced modulo (h(X), p) yields an element of G. Thus, every
polynomial of G is a root of the polynomial

Q1 (Y ) = Y m1 − Y m2 over F .

Thus, there are at least |G| distinct roots in F . Naturally, |G| ≤ degree of
Q1 (Y ). Now the degree of Q1 (Y )

= m1 (as m1 > m2 ) (6.25)

√
 t
≤ (n p) , the greatest number in Iˆ (6.26)
2
√t 
n n
< p < , since p | n, and p = n (6.27)
2 2
√ √
n2 t n2 t
= √ < . (6.28)
2 t 2
√
Hence |G| ≤ m1 < (n2 t
/2).

Lemma 6.7 gives a lower bound for |G|, while Lemma 6.8 gives an upper
bound for |G|. These bounds enable us to prove the correctness of the algo-
rithm. Armed with these estimates on the size of G, we are now ready to prove
the correctness of the algorithm.

Lemma 6.9. If the AKS algorithm returns PRIME, then n is prime.

 Cryptography 275

Proof. Suppose that the

 algorithm returns PRIME. Lemma 6.8 implies that
for t = |G| and l = 2 φ(r) log n,

|G| ≥ min{2t − 1, 2l }
 √ 
t log2 φ(r)
≥ min 2 − 1, 2
 
t 1 2√φ(r)
≥ min 2 − 1, n
2
 
1 √
≥ min 2t − 1, n2 t (since t divides φ(r))
2
 √ 
1 √
≥ min 22 t log n , n2 t (since t > (2 log n)2 )
2
1 2√t
≥ n .
2
√
By Lemma 6.8, |G| < (1/2)n2 t if n is not a power of p. Therefore, n = pk for
some k > 0. If k > 1, then the algorithm would have returned COMPOSITE
in step 1. Therefore, n = p, a prime. This completes the proof of Lemma 6.9
and hence of Theorem 6.4.
It is straightforward to calculate the time complexity of the algorithm. In
these calculations, we use the fact that addition, multiplication, and division
operations between two m-bit numbers can be performed in time O (m). Sim-
ilarly, these operations on two degree d polynomials with coefficients at most
m bits in size can be done in time O (d · m) steps.
Theorem 6.5. The asymptotic time complexity of the algorithm is

O(log 10.5
n).

Proof. The first step of the algorithm takes asymptotic time O(log  3
n).
2
In step 2, we find an r with or (n) > 4 log n. This can be done by trying
out successive values of r and testing if nk = 1(mod r) for every k ≤ 4 log2 n.
For a particular r, this will involve at most O(log2 n) multiplications mod-

ulo r and so will take time O(log 2
n log r). By Lemma 6.4, we know that only
5
O(log n) different r’s need to be tried. Thus, the total time complexity of

step 2 is O(log 7
n).
The third step involves computing gcd of r numbers. Each gcd computa-
tion takes time O(log n), and therefore, the time complexity of this step is
O(r log n) = O(log6 n). The time complexity
 of step 4 is just O(log n).
In step 5, we need to verify 2 φ(r) log n equations. Each equation
requires O(log n) multiplications of degree r polynomials with coefficients
of size O(log n). So each equation can be verified  log2 n) steps.
in time O(r

Thus, the time complexity of step 5 is O(r   32 log3 n) =
φ(r) log3 n) = O(r

O(log 10.5
n). This time complexity dominates all the rest and is therefore the
time complexity of the algorithm.
276 Discrete Mathematics

As mentioned earlier, certain improvements on the time complexity of

the algorithm are available (see, for instance, [26,32]). Our presentation is
based on the lecture given by Professor Manindra Agrawal at the Institute
of Mathematical Sciences, Chennai on February 4, 2003 and reproduced in
the Mathematics Newsletter of Vol. 13 (2003) published by the Ramanujan
Mathematical Society.
We now present some illustrative examples. (The computations have been
done in C++. Readers who require the details of the computation may write
to any of the two authors):

Examples using AKS algorithm

1. Check whether the following numbers are prime or composite.
i. 25
Solution: Here, n = 25. According to step 1 of the algorithm, n = ab
for a = 5 and b = 2. Hence, the given number is composite.
ii. 34
Solution: Here, n = 34.
Step 1: Since n = ab ∀ a, b ∈ N and b > 1. Hence, step 1 does not
give anything decisive.
Step 2: Smallest value of r such that Or (n) > 4 log2 n is 113.
Step 3: For a = 2 ≤ r, gcd(a, n) = 2, and hence 1 < gcd(a, n) < n.
It concludes that 34 is composite.
iii. 23
Solution: Here, n = 23.
Step 1: Since n = ab ∀ a, b ∈ N and b > 1. Hence, step 1 does not
give anything decisive.

Step 2: Smallest value of r such that Or (n) > 4 log2 n is 89.

Step 3:
gcd(a, 23) = 1 for all a ≤ 22 and
gcd(a, 23) is either 1 or 23 for all a ≥ 23. Hence, step 3 does not give
anything decisive.

Step 4: n = 23 and r = 89, so n ≤ r. It concludes that 23 is prime.

iv. 31
Solution: Here, n = 31.
Step 1: Since n = ab ∀ a, b ∈ N and b > 1. Hence, step 1 does not
give anything decisive.

Step 2: Smallest value of r such that Or (n) > 4 log2 n is 107.

 Cryptography 277

Step 3:
gcd(a, 31) = 1 for all a ≤ 30 and gcd(a, 31) is either 1 or 31 for all
a ≥ 31. Hence, step 3 does not give anything decisive.

Step 4 : n = 31 and r = 107, so n ≤ r. It concludes that 31 is prime.

v. 271
Solution: Here, n = 271.
Step 1: Since n = ab ∀ a, b ∈ N and b > 1. Hence, step 1 does not
give anything decisive.
Step 2: Smallest value of r such that Or (n) > 4 log2 n is 269.
Step 3: gcd(a, 271) = 1 for all a ≤ 269. Hence, step 3 does not give
anything decisive.

Step 4: n = 271 and r = 269, so n > r. Hence, step 4 does not give
anything decisive.
√
Step 5: φ(r) = 268, and 2 φr · log n = 264.
For all a, where 1 ≤ a ≤ 264,
(X + a)271 = X 271 + a(mod X 269 − 1, 271).
Step 6: The algorithm outputs that 271 is prime.
Appendix A: Answers to
Chapter 1—Graph Algorithms I

Exercises 1.13
Exercise 2: We suppose the graph has been already represented by n linked
lists L[i] of nodes for i = 1, 2, . . . , n where we assume the declaration struct
node {int v; struct node *succ;}; L[i] is the pointer to the list of all successors
of the node i in the graph. Note that a node consists of two fields: v of type
integer representing a vertex and succ is a pointer to a node in linked list.
Here, n is the number of vertices. When we delete the vertex k, we not only
delete the vertex k, but also the arcs going into the vertex k and leaving out
of k. The removed vertex k will point to an artificial node whose “v” field is
−1, which cannot be a vertex. We now write a fragment of a program in C.
The reader is asked to write a complete program in C based on this fragment
and execute on some examples.

/* remove the nodes in the list L[k]*/

struct node *t,*p;/* t,p pointers to node*/
t=L[k];
{while(t!=NULL){p=t; t=t− >succ; free(p)}
t=(struct node *) malloc(sizeof *t);/*create a node pointed by t*/
t− > v=-1; t− >succ=NULL;L[k]=t;
/*scan all other lists to remove the node containing the vertex k*/
int i;
for(i = 1; i <= n; i + +)
if (i!=k)
{t=L[i];while(t!=NULL)if (t− > v==k) {p=t;t=t− >succ; free(p)};else
{t=t− >succ};

Complexity: O(n + m)
Exercise 4: Note that when we remove an arc (directed edge) (i, j), the
vertices i and j remain in the graph. Removing the arc (i, j) means that we
remove the node containing the field j from the liked list L[i]. We now have a
fragment of a program in C. The reader is asked to write a complete program
and execute on some examples.

struct node *t,*p;/* t,p are temporary pointers to node*/

t=L[i];
while(t!=NULL)

279
280 Appendix A

FIGURE A.1: A weighted simple graph.

if (t− > v==j){ p=t;t=t− >succ; free(p)};else {t=t− >succ};

Complexity: O(n)
Exercise 6: True. Suppose a spanning tree T of G does not contain a bridge
e = xy of G. Since, T is a spanning connected graph it contains the vertices
x and y. Since the only elementary path between x and y in G is the edge xy
and this edge is not in T . This means that in the tree T , the vertices x and y
are not connected by any path, which is impossible.
Exercise 8: The reader is asked to draw the tree obtained by Table A.1.

TABLE A.1: Execution of Prim’s algorithm on graph of Figure A.1

Iteration S = X? s t S ← S ∪ {t} T ← T ∪ { st }
Number
0 – – – {1} ∅
(Initial)
1 yes 1 3 { 1, 3 } { 13 }
2 yes 3 5 { 1, 3, 5 } { 13, 35 }
3 yes 1 2 { 1, 3, 5, 2 } { 13, 35, 12 }
4 yes 3 4 { 1, 3, 5, 2, 4 } { 13, 35, 12, 34 }
5 yes 4 6 { 1, 3, 5, 2, 4, 6 } { 13, 35, 12, 34, 46 }
6 yes 6 7 { 1, 3, 5, 2, 4, 6, 7 } { 13, 35, 12, 34, 46, 67 }
7 yes 7 8 { 1, 3, 5, 2, 4, 6, 7, 8 } { 13, 35, 12, 34, 46, 67 }
8 yes 7 9 { 1, 3, 5, 2, 4, 6, 7, 8, 9 } { 13, 35, 12, 34, 46, 67, 79 }
Exit the no
loop

TABLE A.2: Execution of Dijkstra’s algorithm on the graph

of Figure A.2
Iteration S = X y S ← S ∪ { y } D [13] D [14] D [15] D [16]
Number
0(Initial) – – {1} 5 ∞ ∞ 5
1 yes 2 { 1, 2 } 5 12 ∞ 5
2 yes 5 { 1, 2, 5 } 5 11 15 5
3 yes 3 { 1, 2, 5, 3 } 5 11 14 5
4 yes 4 { 1, 2, 5, 3, 4 } 5 11 14 5
Exit loop no
 Appendix A 281

Exercise 10: The execution is given in Tables A.2 and A.3. X =

{ 1, 2, 3, 4, 5 }. The reader is asked to draw the arborescence using the dad
array.

Exercise 12: Initialization:

1 2 3 4 5 1 2 3 4 5
⎛ ⎞ ⎛ ⎞
1 0 5 ∞ ∞ 5 1 0 5 ∞ ∞ 5
2⎜⎜∞ 0 7 ∞ ∞⎟ ⎟ 2⎜⎜∞ 0 7 ∞ ∞⎟ ⎟
M0 = 3 ⎜⎜∞ ∞ 0 3 ∞ ⎟. M1 = 3 ⎜
⎟
⎜∞ ∞ 0 3 ∞⎟ ⎟.
4⎝∞ ∞ ∞ 0 ∞⎠ 4⎝∞ ∞ ∞ 0 ∞⎠
5 6 ∞ 6 10 0 5 6 111 6 10 0
1 2 3 4 5 1 2 3 4 5
⎛ ⎞ ⎛ ⎞
1 0 5 122 ∞ 5 1 0 5 122 153 5
⎜ ⎟ ⎜ ⎟
⎜
2 ∞ 0 7 ∞ ∞ ⎟ 2⎜∞ 0 7 103 ∞ ⎟
⎜ ⎟ ⎜ ⎟
M2 = 3 ⎜
⎜∞ ∞ 0 3 ∞⎟ ⎜
⎟. M3 = 3 ⎜ ∞ ∞ 0 3 ∞⎟ ⎟.
⎜ ⎟ ⎜ ⎟
4⎝∞ ∞ ∞ 0 ∞⎠ 4⎝∞ ∞ ∞ 0 ∞⎠
5 6 111 6 10 0 5 6 111 6 93 0
M4 = M3
+
because d (4) = 0.

FIGURE A.2: A weighted directed graph.

TABLE A.3: Evaluation of dad array in

Dijkstra’s algorithm of Figure A.2
Dad [13] Dad [14] Dad [15] Dad [16]
1 1 1 1
1 2 1 1
1 5 5 1
1 5 3 1
1 5 3 1
282 Appendix A

FIGURE A.3: A simple connected graph.

1 2 3 4 5
⎛ ⎞
1 0 5 115 145 5
2⎜⎜ ∞ 0 7 103 ∞⎟⎟
M5 = 3 ⎜
⎜ ∞ ∞ 0 3 ∞⎟⎟.
4⎝∞ ∞ ∞ 0 ∞⎠
5 6 111 6 93 0
1 2 3 4 5
⎛ ⎞
1 0 0 5 5 0
2⎜⎜0 0 0 3 0⎟⎟
INTER = 3 ⎜
⎜0 0 0 0 0⎟⎟.
4⎝0 0 0 0 0⎠
5 0 1 0 3 0

Call to interpath(1,4) prints the intermediate vertices between 1 and 4 in a

shortest path from 1 to 4. The output is 5, 3.

Exercise 14: The following Table A.4 gives us the dfsn of different vertices
of the graph of Figure A.4.

TABLE A.4: Dfsn table of graph of

Figure A.4
1 2 3 4 5 6 7 8 9 10
dfsn 1 2 3 4 5 6 9 10 8 7

The following Table A.5 gives the LOW function of different vertices.

TABLE A.5: LOW function table of the

graph of Figure A.4
1 2 3 4 5 6 7 8 9 10
low 1 1 1 4 1 5 3 3 5 5
 Appendix A 283

FIGURE A.4: A dfs drawing of the graph of Figure A.3.

The evolution of stack of edges (the rows from left to right):

We push the edges, 12,23,31,34. Stack = (12,23,31,34). Now at the ver-
tex 4, low[4] ≥ df sn[3]. Hence, we pop the edge 34 from the stack,
which is the first biconnected component. Now the stack grows to: stack=
(12, 23, 31, 35, 51, 52, 56, (6, 10), (10, 9), 95). At the vertex 6, low[6] ≥ df sn[5].
Hence, we pop the edges from the stack till and including edge 56, that is,
95,(10,9),(6,10),56 which is the second biconnected component emitted. The
new stack after popping: stack = (12,23,31,35,51,52). The stack grows as: stack
= (12,23,31,35,51,52,37,78,83). At vertex 7, we have low[7] ≥ df sn[3]. Hence,
we pop till the edge 37, that is, 83,78,37 which is the third biconnected com-
ponent. The stack shrinks to stack = (12,23,31,35,51,52). At vertex 2, we have
low[2] ≥ df sn[1]. We pop the edges till the edge 12, that is, 12,23,31,35,51,52
which is the fourth and last biconnected component. Note that the stack is
empty.

Exercise 16: The reader is first asked to draw the graph with direction of
the arcs as given in the exercise. As usual, we process the vertices in increasing
order and we suppose the vertices are listed in increasing order in each list
L[i] for 1 = 1, 2, . . . , 10. The following Table A.6 gives the dfsn of different
vertices.

TABLE A.6: dfsn table of graph of

Figure 1.38
1 2 3 4 5 6 7 8 9 10
dfsn 1 2 3 4 5 6 9 10 8 7

The dfs drawing of the graph is given in Figure A.5.

The following Table A.7 gives the function LOWLINK.
284 Appendix A

FIGURE A.5: dfs: strong components algorithm.

TABLE A.7: LOWLINK function table of

graph of Figure 1.38
1 2 3 4 5 6 7 8 9 10
LOWLINK 1 1 1 4 5 5 3 3 5 5

We use a stack of vertices. The stack grows from left to right.

Stack=(1, 2, 3, 4). At vertex 4, we find LOWLINK[4] = df sn[4]. Hence, we
pop the vertex 4 which forms a strongly connected component consisting of
a single vertex. Then, we push the vertices 5, 6, 10, 9 onto the stack. Hence
the stack is, stack=(1, 2, 3, 5, 6, 10, 9). While climbing the tree, we find at ver-
tex 5, LOWLINK[5] = df sn[5]. Hence we pop the stack till the vertex 5, that
is, we output: 9,10,6,5 which is the second strongly connected component.
The stack now shrinks to: stack=(1, 2, 3). The vertices 7 and 8 are pushed
onto the stack. Stack=(1, 2, 3, 7, 8). We climb the tree till the vertex 1. Since,
LOWLINK[1] = df sn[1], we pop the stack till the vertex 1, that is, we output:
8,7,3,2,1 which is the third and last strongly connected component.

Exercise 18: After assigning the given orientations, we observe that the
resulting directed graph is without circuits. Hence the topological sort is pos-
sible. As usual, we process the vertices in increasing order and in each linked
list L[i], with i = 1, 2, . . . , 10, the vertices are listed in increasing order. We
perform dfs of the graph by drawing only the tree arcs. We obtain two arbores-
cences in Figure A.6 (the reader is asked to perform the dfs):
Note that the second arborescence consists of only one vertex 4. We
now write the vertices of the forest of arborescence in postfix order:
9,10,6,5,8,7,3,2,1,4. We now take the mirror image, that is, write the postfix
order obtained from right to left: 4,1,2,3,7,8,5,6,10,9. This is the topological
 Appendix A 285

FIGURE A.6: dfs: topological sort.

order required. The reader is asked to draw the entire graph by aligning the
vertices horizontally and verify that there are no arcs going from right to left.

Exercise 20: See the text.

Exercise 22: See the text.

Appendix B: Answers to
Chapter 2—Graph Algorithms II

Exercises 2.5
Exercise 2: “Yes.” Let the source vertex be 1. Perform a breadth-first search
(bfs) from 1. This search partitions the vertex set X = { 1, 2, . . . , n } as:

X = L0 (1) ∪ L1 (1) · · · ∪ Le(1) (1)

where L0 (1) = { 1 }, Li (1) = { x|d(1, x) = i }, e(1) is the eccentricity of the

vertex 1. Here, d(1, x) is the minimum length of a path from the vertex 1 to
the vertex x. Let us recall that the length of a path is the number of edges in
it. We use an array D of n reals.
Arrange the vertices of the graph in non-decreasing order of distances from
the vertex 1. For example, write the source vertex 1 first, then the vertices of
L1 (1) (in any order), then the vertices of L2 (1) (in any order), etc., till the
vertices of Le(1) (1). D[1] = 0; //initialization for each vertex x = 1 in non-
decreasing order from the vertex 1 to D[x] = min{ D[y] + c(y, x)|y ∈ Li−1 (1) }
//c(y, x) is the cost associated with the edge yx.

Exercise 4: Perform the bfs from each vertex i, with 1 ≤ i ≤ n. The largest
integer k with Lk (i) = emptyset is the eccentricity of the vertex i. (See the
solution of Exercice 2.) The minimum of the eccentricities is the radius of the
graph, and the maximum of the eccentricities is the diameter of the graph.
Since we perform n bfs, and the cost of a single bfs is O(max(m, n)), the com-
plexity of the algorithm is O(n max(m, n)) where n and m are the number of
vertices and edges, respectively, of the graph.

Exercise 6: A bipartite graph is geodetic if and only if it is a tree.

Exercise 8: Let us first draw the Petersen graph (see Figure B.1). The
execution of the algorithm is illustrated in Table B.1. We find an augmenting
path visually. The reader is asked to draw the edges of the matching during
the execution by a color, say, red, the other edges black. Hence, an augmenting
path is an elementary path with alternating colors black and red, beginning
with a black edge and ending with another black edge. Note that the origin
and the terminus of such a path are both unsaturated by red edges. Note that

287
288 Appendix B

FIGURE B.1: the Petersen graph.

TABLE B.1: Execution of matching algorithm

Iteration Is there an M augmenting M := (M ∪ P )\
number path P? (M ∩ P )
0(initialization) not applicable { 34 }
1 “yes,” P = (2, 3, 4, 5) { 23, 45 }
2 “yes,” P = (7, 2, 3, 4, 5, 1) { 15, 34, 27 }
3 “yes,” P = (9, 7, 2, 1, 5, 10) { 97, 21, (5, 10), 34 }
4 “yes,” P = (6, 8) { 97, 21, (5, 10), 34, 68 }
“no,” quit the loop

a black edge alone with its end vertices unsaturated is an augmenting path.
The matching obtained at the 4th iteration is a perfect matching.

Exercise 10: M1 and M2 are two different perfect matchings of a graph G.

The spanning subgraph whose edge set is M1 δM2 = (M1 ∪ M2 ) − (M1 ∩ M2 )
has connected components either isolated vertices or an elementary cycle of
even length whose edges are alternatively in M1 and M2 . (The reader is asked
to draw a graph with two different perfect matchings M1 and M2 and find the
graph (X, M1 ΔM2 )).

Exercise 12: A 3-regular graph G has always an even number of vertices.

(Because, the number of vertices of odd degree in a graph is always even.)
Since the graph contains a Hamiltonian cycle C, and the number of vertices
in C is even (note that C is a spanning connected subgraph of G with the
degree of each vertex in C is 2), the alternating edges of C form two disjoint
perfect matchings. Now by removing the edges of these two perfect matchings
from the original graph G, we obtain another (third) disjoint perfect match-
ing. Note that these three mutually disjoint perfect matchings exhaust all the
edges of the graph.

Exercise 14: See an example in the text.

Exercise 16: See the reference K. R. Parthasarathy [11]

Appendix C: Answers to
Chapter 3—Algebraic Structures I

Exercises 3.3.7
Exercise 2: The proof is by induction in n. The result is true for n = 1.
Assume the result for n so that
 
cos(nα) sin(nα)
Mn = .
− sin(nα) cos(nα)
Then,
  
n+1 n 1 cos(nα) sin(nα) cos(α) sin(α)
M =M ·M =
− sin(nα) cos(nα) − sin(α) cos(α)
 
cos(nα) cos α − sin(nα) sin α cos(nα) sin(α) + sin(nα) cos α
=
− sin(nα) cos α − cos(nα) sin α − sin(nα) sin α + cos(nα) cos α
 
cos(n + 1)α sin(n + 1)α
= .
− sin(n + 1)α cos(n + 1)α

Exercise 4: If
      
1 3 1 3 1 3 −5 9
A= , A2 = = ,
−2 2 −2 2 −2 2 −6 −2
and hence
       
−5 9 1 3 1 0 0 0
A2 − 3A + 8I = −3 +8 = .
−6 −2 −2 2 0 1 0 0
Now det(A) = 0 and hence A−1 exists. Multiplying by A−1 , we get
  1 −3

−1 −1 −2 3
A − 3I + 8A−1 = 0 ⇒ A−1 = (A − 3I) = = 14 8
.
8 8 −2 −1 4
1
8

Exercise 6:
(i) Take
⎛ ⎞ ⎛ ⎞
a11 a12 ··· a1n b11 b12 ··· b1p
⎜ .. ⎟ ⎜ .. ⎟
A=⎝ . ⎠, B=⎝ . ⎠
am1 am2 ··· amn bn1 bn2 ··· bnp
so that product AB is defined. Now check that (AB) = B At . t t

289
290 Appendix C

(ii) Trivially (AB)(B −1 A−1 ) = A(BB −1 )A−1 = AIA−1 = I. Similarly,

(B −1 )(A−1 )(AB) = I.

Exercise 8:
if A = (aij ), A∗ = (bij ), where brs = āsr . Here, iA = (iars ),
(i) Recall that √
(Recall i = −1) and therefore (iA)∗ = (iasr )∗ = −(iāsr ) = −i(āsr ) =
−iA∗ .
(ii) H is Hermitian ⇔ H ∗ = H. iH is skew-Hermitian ⇔ (iH)∗ = −iH ⇔
−iH ∗ = −iH ⇔ H ∗ = H.
Exercise 10: Let A be a complex square matrix. Then, A = ((A + A∗ )/2)+
((A − A∗ )/2). Here, (A + A∗ )/2 is Hermitian since
(A + A∗ )∗ A∗ + (A∗ )∗ A∗ + A
= = .
2 2 2
Further
 ∗  
A − A∗ A∗ − (A∗ )∗ A∗ − A A − A∗
= = =− .
2 2 2 2

Uniqueness: Suppose A = A1 + A2 = B1 + B2 , where A1 and B1 are Her-

mitian, while B1 and B2 are skew-Hermitian. Then, A1 − B1 = B2 − A2 , and
A1 − B1 is Hermitian while B2 − A2 is skew-Hermitian. But if a matrix is both
Hermitian and skew-Hermitian, it must be the zero matrix (check!). Hence
A1 = B1 and A2 = B2 .

Exercises 3.16
Exercise 2: Closure:
     
a 0 a 0 aa 0
= ∈ G.
b 1 b 1 ba + b 1

Note: As  a = 0, aa = 0). 

a = 0, Associative
 law is true for matrix products.
1 0 a 0
Identity: is of the form with a = 1 and b = 0.
0 1 b 1
Inverse: The inverse of
   1 
a 0 a 0
is −b .
b 1 a 1

Exercise 4: Let S = {x1 , x2 , . . . , xn } be a finite semigroup satisfying both

cancellation laws. Let xi ∈ S. Then, the set {xi x1 , xi x2 , . . . , xi xn } is a per-
mutation of S. This is because xi xj = xi xk ⇒ xj = xk . Hence there exists a
 Appendix C 291

p, 1 ≤ p ≤ n, such that xi xp = xi . We claim that xp is the identity element

of S. This is because if xr ∈ S, xr = xk xi for some k, 1 ≤ k ≤ n. (Note i is
fixed). Hence xr xp = (xk xi )x = (as S is associative) xk (xi xp ) = xk xi = x.
Thus, xr xp = xr , and similarly xp xi = xi . Indeed, as xi xp = xi , for
any i, xi xp xj = xi xj ⇒ (cancelling xi ) xp xj = xj for each j. Hence xp
is the identity element of S. Again, for any i, 1 ≤ i ≤ n, there exists
j, 1 ≤ j ≤ n such that xi xj = xp . Also, if xi xj = xp , xj xi xj = xj xp =
xp xj (as xp is the identity element) ⇒ (by right cancellation law) xj xi = xp .
Thus, xj is the inverse of xi . This proves that S is a group.

Exercise 6: Recall that the order of a non-singular square matrix A is the

least positive integer k such that Ak = I

(i) Order is 4.

(ii) If
     
1 1 2 1 2 3 1 3
A= , then A = , A =
0 1 0 1 0 1
 
k 1 k
and so on, and, in general, A = and hence for no finite k, Ak =
0 1
I. Thus, ◦(A) = zero.

(iii) Order = 4

(iv) If Ak = I, then (detA)k = 1. But this is not the case. Hence ◦(A) = 0.

Exercise 8: If ◦(a) = 2, a2 = e, and therefore a = a−1 . Conversely, if

a = a−1 , and a = e, ◦(a) = 2. Hence if a = a−1 , ◦(a) = 2. Pairing off such
elements (a, a−1 ), we see that since the group is of even order and ◦(e) = 1,
the group must contain an odd number of elements, and hence at least one
element must be of order 2.

Exercise 10: Suppose to the contrary that a group G is the union of two
of its proper subgroups, say, G1 and G2 . Then, there exist x ∈ G \G2 (and
hence x ∈ G1 ) and y ∈ G \ G1 (and hence y ∈ G2 ). Then, look at xy. As
G = G1 ∪ G2 , xy ∈ G1 or G2 . If xy = x1 ∈ G1 , then y = x−1 x1 ∈ G1 , a
contradiction. A similar argument applies if xy ∈ G2 .

Exercise 12: The set of all invertible 2 × 2 real matrices or the set of all
diagonal 2 × 2 matrices with real entries and determinant not equal to zero.

Exercise 14:

(i) (123)(456) = (12)(13)(45)(46), a product of an even number of transpo-

sitions and hence the permutation is even.
292 Appendix C

(ii) (1516)(2)(3) = (15)(1)(16)(2)(3) = (15)(16) is even. [Note: The numbers

that do not appear are all fixed]
(iii)
 
1 2 3 4 5 6 7 8 9
= (125)(34)(67)(89) = (12)(15)(34)(67)(89)
2 5 4 3 1 7 6 9 8

is an odd permutation, as it is a product of an odd number, namely 5,

of transpositions.

Exercise 16: Let a, b ∈ G. Then, a ∗ b also belongs to G as a ∗ b = −1 ⇒

a + b + ab = −1 ⇒ (a + b)(b + 1) = 0 that is, a = −1 or b = −1 which is
not the case. So closure property is satisfied. A routine verification shows that
(a ∗ b) ∗ c = a ∗ (b ∗ c) for any three a, b, c ∈ G. Hence associative property is
true. Further, a ∗ 0 = 0 ∗ a = a. Hence 0 acts as the identity in G. Finally,
given a ∈ G, we have a ∗ ((−a)/(1 + a)) = a − (a/(1 + a)) − (a2 /(1 + a)) = 0,
and hence (−a)/(1 + a) is the inverse of a. Ḡ under ∗ is commutative. Hence
right inverse of a is automatically its left inverse.

Exercise 18: σ = (i1 i2 . . . ir ). Hence σ 2 = (i1 i3 i5 . . . ), σ 3 = (i1 i4 i5 . . . ) and

so σ r = identity. Hence ◦(σ) = r.

Exercise 20:

(i)    
−1 1 4 3 2 1 2 3 4
(α) = =
1 2 3 4 1 4 3 2

(ii)    
−1 1 2 3 4 1 2 3 4 1 2 3 4
α βγ = = δ,
1 4 3 2 2 1 4 3 3 1 2 4
say then
δ(1) = α−1 βγ(1) = α−1 β(3) = α−1 (4) = 2,

(iii)     
1 2 3 4 3 1 2 4 1 2 3 4
βγ −1 = = .
2 1 4 3 1 2 3 4 1 4 2 3

Exercise 22: For m, n ∈ Z, eim · ein = ei(m+n) , m + n ∈ Z. Further ei·0 = 1

is the identity element. ein · e−in = 1 ⇒ e−in is the inverse of ein , −n ∈ Z.
Associative law is trivially true. Hence the first part. Now consider the map
φ : ein → n φ is 1–1. Since φ(ein ) = φ(eim ) ⇒ n = m ⇒ ein = eim .
Trivially, φ is onto. φ is a homomorphism. Since for m, n ∈ Z, φ(ein · eim ) =
φ(ei(n+m) ) = n + m = φ(ein ) + φ(eim ). Hence φ is a 1–1, onto homomorphism
from the multiplicative group {ein : n ∈ Z} to the additive group (Z, +). This
establishes the isomorphism. The given group is cyclic as (Z, +) is.
 Appendix C 293

Exercise 24: The map φ : (Z, +) → (2Z, +) from the additive group of inte-
gers to the additive group of even integers is a group isomorphism. (Check).

Exercise 26: Suppose there exists an isomorphism φ from (R∗ , ·) onto

(C , ·). Then, φ(1) = 1 and hence φ(−1) = −1 since (−1)2 = 1 in R∗ and
∗

C∗ . But then i2 = −1 in C∗ while there exists no element a ∈ R∗ with a2 = −1.

Exercise 28: Note: The center of S3 consists of identity element only

(Table C.1).

TABLE C.1: Numerical equivalents for English

characters
e (12) (23) (31) (123) (132)
e e (12) (23) (31) (123) (132)
(12) (12) e (123) (132) (23) (13)
(213) (23) (132) e (123) (13) (12)
(31) (31) (123) (132) e (12) (23)
(123) (123) (13) (12) (23) (132) –
–

Exercise 30: If G is Abelian, trivially C(G) is G. If C(G) = G, then every

element of G, that is, of C(G) certainly commutes with all elements of G.

Exercise 32: ab = ba ⇔ aba−1 b−1 = e and hence [G, G] = {e}.

Exercise 34: For a ∈ A and b ∈ B, aba−1 b−1 = (aba−1 )b−1 ∈ B (as B

is normal in G and so aba−1 ∈ B). Similarly aba−1 b−1 = a(ba−1 b−1 ) ∈ A
as a−1 ∈ A and as A is a normal subgroup of G, ba−1 b−1 ∈ A and hence
aba−1 b−1 ∈ A. Thus,

aba−1 b−1 ∈ A ∩ B = {e}

⇒ ab = ba for all a, b ∈ G

⇒ G is an Abelian group.

Exercise 36: Let H be a subgroup of index 2 in G. Then, for all h ∈

H, hH = H = Hh, while if g ∈ G \H, G = H ∪ Hg = H ∪ gH, where the
unions are disjoint.

Exercise 38: Follows immediately from Exercise 36.

294 Appendix C

Exercises 3.19
Exercise 2: Routine verification.

Exercise 4: Result is true for n = 2 (from the axioms of a ring). Assume

that a a(b1 + b2 + · · · + bn ) = ab1 + ab2 + · · · abn . Then, for elements
a, b1 , b2 , . . . , bn , bn+1 of A,
a(b1 + · · · + bn + bn+1 ) = a(b1 + · · · + bn ) + (bn+1 )

= a(b1 + · · · + bn ) + abn+1 (by the axioms for a ring)

= (ab1 + · · · + abn ) + abn+1 (by induction hypothesis)
= ab1 + · · · + abn + abn+1 (by the associative property of +).

Exercise 6: Suppose u is a unit and ux = 0. As u−1 exists, u−1 (ux) = 0 ⇒

−1
(u u)x = 0 ⇒ 1A · x = 0 ⇒ x = 0 (1A is the unit element of A). Similarly,
xu = 0 ⇒ x = 0.

Exercise 8: Look at the elements 1, 1 + 1, . . . , 1 + 1 + · · · + 1(p times).

These elements must all be distinct. If not, there exist r, s with 0 < r < s ≤ p
such that 1 + 1 + · · · + (r times) = 1 + 1 + · · · + 1(s times) ⇒ 1 + 1 + · · · +
1(s−r times) = 0. So there must exist a least positive integer k(< p) such that
1+1+· · ·+1(k times) = 0 ⇒ k|p, a contradiction to the fact that p is a prime.
Appendix D: Answers to Chapter
4—Algebraic Structures II

Exercises 4.8
Exercise 2: If f (x) = a0 + a1 x + · · · + an xn ∈ R[x], an = 0, and
g(x) = b0 + b1 x + · · · + bn xn ∈ R[x], bn = 0 are polynomials of degree n,
then bn f (x) − an g(x) is a polynomial in R[x] of degree < n. Hence the result.

Exercise 4: As in the Hint, if

⎛ ⎞ ⎛ ⎞
1 0 0 0 0 0
e11 ⎝
= 0 0 0⎠ , . . . , e33 = ⎝0 0 0⎠ ,
0 0 0 0 0 1

then if
⎛ ⎞
a11 a12 a13
A = ⎝a21 a22 a23 ⎠
a31 a32 a33

3 2
is any real matrix, then A = i,j=1 aij eij . Further, if A = 0, all the 3
coefficients aij are zero. Hence the 32 matrices eij are linearly independent
and span the space of all 3 × 3 real matrices over R. Hence dimension of the
vector space of all real 3 × 3 matrices over R is 32 . Now generalize to m × n
matrices.

Exercise 6: Let V be the vector space of all real polynomials in X. Suppose

V is finite dimensional. Let n be the maximum of the degrees of the polyno-
mials in a basis B of V. As every polynomial of V is a linear combination of
the polynomials in B, each polynomial in V is of degree n, at the most, a
contradiction.

Exercise 8: Note v1 = 3v3 − v2 and v4 = v3 − v2 . Hence the space

v1 , v2 , v3 , v4  = v2 , v3  (here u1 , u2 , . . . , un  stands for the space spanned
by {u1 , u2 , . . . un }). Now v1 and v2 are linearly independent over R since oth-
erwise one would be a scalar multiple of the other, (here the scalars are the
elements of R). Hence dimv1 , v2 , v3 , v4  = 2.

295
296 Appendix D

Exercises 4.12
Exercise 2: The given system of equations is equivalent to
⎛ ⎞⎛ ⎞ ⎛ ⎞
4 4 3 −5 X1 0
⎜1 1 2 −3⎟ ⎜ X2 ⎟ ⎜0⎟
⎜ ⎟⎜ ⎟ = ⎜ ⎟.
⎝2 2 −1 0 ⎠ ⎝ X3 ⎠ ⎝0⎠
1 1 2 −2 4X4 0

Perform the row operations: R1 − (R2 + R3 + R4 ). Followed by R3 − (R2 +

R4 ). These give ⎛ ⎞⎛ ⎞ ⎛ ⎞
0 0 0 0 X1 0
⎜1 1 2 −3⎟ ⎜X2 ⎟ ⎜0⎟
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎝0 0 −5 5 ⎠ ⎝X3 ⎠ = ⎝0⎠ .
1 1 2 −2 X4 0
The matrix on the left is of rank 3 and hence there must be 4(number of
columns)-3(rank)=1 linearly independent solution over R.
The given equations are equivalent to X1 + X2 + 2X3 − 3X4 = 0, −5X3 +
5X4 = 0, X1 + X2 + 2X3 − 2X4 = 0 ⇒ X3 = X4 .
Subtracting the 3rd equation from the 1st, we get X4 = 0. Hence X3 =
X4 = 0. So we can take (1, −1, 0, 0) as a generator of the solution space.

Exercise 4(a): The given set of equations is equivalent to AX = B, where

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
2 3 −5 4 X1 −8
⎜3 1 −4 5⎟ ⎜X2 ⎟ ⎜−8⎟
A=⎜ ⎝7
⎟ , X = ⎜ ⎟ and B = ⎜ ⎟ .
3 −2 1⎠ ⎝X3 ⎠ ⎝ 56 ⎠
4 1 −1 3 X4 20

The Schur complement of A is

⎛ ⎞ ⎛3⎞
1 −4 5 2
⎜ ⎟

A1 = ⎝3 −2 1⎠ − ⎝ 72 ⎠ 3 −5 4
1 −1 3 4
2
⎛ ⎞ ⎛ 9 −15 ⎞ ⎛ −7 7 ⎞
1 −4 5 2 2 6 2 2 −1
= ⎝3 −2 1⎠ − ⎝ 21 2
−35
2 14 ⎠ = ⎝ −15
2
31
2 −13⎠ .
1 −1 3 6 −10 8 −5 9 −5

The Schur complement of A1 is A2 , where

 31  15 
−13 7
7

A2 = 2 − 10 2 −1
9 −5 7

8 −767
= .
4 −257
 Appendix D 297

The Schur complement of A2 is

4
−76

A3 = −25 7 − 8 7
13

= 7

= 1 13 7
= L3 U3

where L3 = 1 is unit lower-triangular, and U3 = 13 7 is upper-triangular.

This gives
  
1 0 8 −767
A2 = 4
8 L3 0 U3
  
1 0 8 −76
7
= 1
2 1 0 137

= L2 U2 .

Therefore,
⎛ ⎞ ⎛ −7 7 ⎞
1 0 0 2 2 −1
A1 = ⎝ 15
7 1 0⎠ ⎝ 0 8 −76 ⎠
7 = L1 U1 .
10 1 13
7 2 1 0 0 7

Consequently,
⎛ ⎞
1 0 0 0 ⎛2 3 −5 4
⎞
⎜ 32 1 0 0⎟ ⎜0 −7 7
−1 ⎟
⎜ ⎟⎜
−76 ⎟
2 2
A= ⎜7 15
0⎟ ⎠ = LU.
⎝2 7 1 ⎠ ⎝0 0 8 7
10 1 13
2 1 0 0 0 7
7 2

Note that L is lower unit-triangular and U is upper-triangular.

Therefore, the given system of linear equations is equivalent to LU X = B.
Set ⎛ ⎞
Y1
⎜Y2 ⎟
UX = Y = ⎜ ⎟
⎝Y3 ⎠ .
Y4
We then have LY = B. We now solve for Y and then for X. Now LY = B
gives
⎛ ⎞⎛ ⎞ ⎛ ⎞
1 0 0 0 Y1 −8
⎜ 32 1 0 0⎟ ⎜Y2 ⎟ ⎜−8⎟
⎜ 7 15 ⎟⎜ ⎟ = ⎜ ⎟.
⎝
2 7 1 0⎠ ⎝Y3 ⎠ ⎝ 56 ⎠
2 10 1
1 Y4 20
7 2
298 Appendix D

3 7 15
⇒ Y1 = −8, Y1 + Y2 = −8, Y1 + Y2 + Y3 = 56 and
2 2 7
10 1
2Y1 + Y2 + Y3 + Y4 = 20
7 2
528 −52
⇒ Y1 = −8, Y2 = 4, Y3 = , and Y4 = .
7 7
We now determine X from the equation U X = Y.
U X = Y ⇒ 2X1 + 3X2 − 5X3 + 4X4 = −8
−7 7
X 2 + X3 − X 4 = 4
2 2
76 528
8X3 − X4 =
7 7
13 −52
X4 =
7 7
Solving backward, we get X4 = −4, 8X3 − (76/7)(−4) = (528/7) ⇒ 8X3 =
(528 − 304)/7 = 32 ⇒ X3 = 4. Hence ((−7)/2)X2 + (7/2)X3 − X4 = 4 gives
((−7)/2)X2 + 14 + 4 = 4 ⇒ X2 = 4.
Finally, 2X1 + 3X2 − 5X3 + 4X4 = 8 gives X1 = 8. Thus,
⎛ ⎞ ⎛ ⎞
X1 8
⎜X2 ⎟ ⎜ 4 ⎟
X=⎜ ⎟ ⎜ ⎟
⎝X3 ⎠ = ⎝ 4 ⎠ .
X4 −4
4(b): Similar to 4(a). Answer:
⎛ ⎞ ⎛ ⎞
X1 2
⎝X2 ⎠ = ⎝3⎠ .
X3 7
4(c): Since there are zeros in the principal diagonal of the matrix A, inter-
change the first two and the last two equations to avoid zeros in the main
diagonal.

Exercises 4.15
Exercise 2: Let F1 = GF (25 ) and F2 = GF (23 ). If F2 is a subfield of
F1 , then F2∗ , the multiplicative group of the non-zero elements of F2 should
be a subgroup of F1∗ . But |F1∗ | = 25 − 1 = 31, and |F2∗ | = 23 − 1 = 7 and 7  31.
2
Exercise 4: X 3 − X = X(X 8 − 1) = X(X 4 − 1)(X 4 + 1) = X(X − 1)(X +
1)(X 2 + 1)(X 4 + 1) = X(X + 2)(X + 1)(X 2 + 1)(X 4 + 1). Note that X 2 + 1
is irreducible over Z3 since neither 1 nor 2 is a root X 2 + 1 in Z3 . A similar
statement applies to X 4 + 1 as well.
Appendix E: Answers to
Chapter 5—Introduction to Coding
Theory

Exercises 5.12
Exercise 2(a): Take
 
1 0 1 1 1
G= over Z2 .
1 0 0 1 1
G is of rank 2, and so H is of rank 5 − 2 = 3 and each row of H is orthogonal
to every row of G. So H can be taken either as
⎛ ⎞ ⎛ ⎞
1 1 0 1 0 1 1 0 0 1
H1 = ⎝1 0 0 0 1⎠ or H2 = ⎝1 0 0 0 1⎠
0 1 0 0 0 0 1 0 1 1
Let ⎛ ⎞
1
⎜0⎟
⎜ ⎟
X=⎜
⎜1⎟
⎟
⎝1⎠
1
Then, S(x) with respect to H1 is
⎛ ⎞
0
H1 X T = ⎝0⎠ ,
0
while S(x) with respect to H2 is
⎛ ⎞
0
H2 X T = ⎝0⎠ .
1
Solutions to 2(b), 2(c) and 2(d) are similar to the steps in Example 5.1.
Exercise 4: u has 7 coordinates. s(u, 3) contains all binary vectors of length


7 which are at distances 0, 1, 2 or 3 from u and hence their number is 1 + 71 +

7
7
7 
2 + 3 = 1 + 7 + 21 + 35 = 64. [Here, 1 arises from vectors which are at a
distance 1 from u etc.]

299
300 Appendix E

Exercise 6: Let C be such a code.Then, |C| = 2k . Further, the spheres with

centers at these 2k codewords and of radii t cover Zn2 , where |Zn2 | = 2n . Each
codeword is of length n. Hence, as in the solution
of Exercise

 2, a
sphere
 of
radius t with center at a codeword will contain 1 + n1 + n2 + · · · + nt vectors
of Zn2 . Hence, as there are 2k spheres,
     
n n n
2k + + ··· + = 2n .
0 1 t

The general case is similar.

Exercise 8: Assume the contray, that is, there exists a set of 9 vectors in
Z62 (the space of binary vectors of length 6) which are pairwise at a distance
at least 3. Look at the last two coordinates of these 9 vectors. They should
be from {00, 01, 10, 11}. As there are 9 vectors, at least one of these pairs
must occur at least 3 times. Consider such a set of 3 vectors which have the
same last two coordinates. These coordinates do not contribute to the dis-
tance between the corresponding three vectors. Hence if we drop these last
two coordinates from these 3 vectors of Z62 , we get 3 binary vectors of length
4 such that the distance between any two of them is at least 3. This is clearly
impossible. This contradiction proves the result.

Exercise 10: Let e1 = 01234, e2 = 12044 and e3 = 13223. Note that

e3 = e1 + e2 over F5 . Hence Spane1 , e2 , e3  = Spane1 , e2 . Now write down
the (5 × 5) − 1 = 24 non-zero codewords of the code and find the minimum
weight d. Then, the code can detect d − 1 errors (Refer to Theorem 5.8).

Exercise 12:

A. Since there is a single constraint dim(C) = 3 − 1 = 2. Hence

|C| = 32 = 9. To get all 9 words, give all possible 32 val-
ues to the pair (x2 , x3 ). These are (0,0),(1,0),(2,0); (0,1),(1,1),(2,1);
(0,2),(1,2),(2,2). These give, as x1 + x2 + 2x3 = 0, the triads,
(0,0,0),(2,1,0),(1,2,0),(1,0,1),(0,1,1), (2,2,1),(2,0,2),(1,1,2),(0,2,2).

B. As dim(C) = 2, any two linearly independent codewords over Z3 would

form a generator matrix for C. For instance,
 
1 0 1
G1 = .
1 1 2

C.  
1 0 1
G2 = .
2 2 1

D. d = minimum weight of a non-zero codeword = 2 (Look at the nine

codewords in (A)).
 Appendix E 301

E. C can correct (d − 1)/2 = 0 error.

F. C can detect d − 1 = 2 − 1 = 1 error.

G. Rank of G2 = 2 = k. Hence rank of the pairity-check matrix H is

3 − 2 = 1. Every row of H is orthogonal to every row of G over Z3 .
Hence we can take H = [112].

Exercises 5.15
Exercise 2: C3 is cyclic since the three non-zero codewords are the cyclic
shifts of any one of them. For example, (1, 1, 0) −→ (0, 1, 1) −→ (1, 0, 1).
Since one of them is a linear combination (over Z2 ) of the other 2, as a gen-
erator matrix for C3 , we can take any two of these codewords. Hence
 
1 1 0
G=
0 1 1

is a generator matrix of C3 . As rank(G) = 2, its null space is of dimension 1.

Hence the dual code of G, namely G⊥ = [1 1 1], and so G⊥ is also cyclic and
dim(G⊥ ) = 1.

Exercise 4: The codeword 1 + x corresponds to the n-vector (1, 1, 0, . . . , 0).

Hence the generator matrix of the binary code generated by 1 + x is
⎛ ⎞
1 1 0 0 ... 0
⎜0 1 1 0 . . . 0⎟
⎜ ⎟
⎝0 0 1 1 . . . 0⎠
0 0 ... 0 1 1

(Note: G is an (n − 1) × n matrix of rank n − 1. Hence the dimension of the

cyclic code is n − 1).

Exercise 6: Suppose C contains a word X of weight (n−2). Then, C contains

exactly two zeros and all the rest are 1s. Cyclically shift X and get the word
X0 = (x0 , x1 , . . . , xn−1 ), where x0 = 0 = xi , and xj = 1 for every j = 0, i.
Cyclically shift X0 until xi moves to the 0-th position. This can be effected in
n − i shifts. These n − i shifts will not move the first zero of X0 to the second
zero of X0 . For this requires i shifts, and this means that i = n − i ⇒ n is
even, a contradiction to our assumption that n is odd. Let X1 be the new word
obtained after these n − i shifts. Clearly, d(X0 , X1 ) = 2, again a contradiction
to our assumption that d(C) ≥ 3.
Appendix F: Answers to
Chapter 6—Cryptography

Exercises 6.4
Exercise 2: det A = 12.7 − 5.3 = 204 − 15 = 189 = 15 in Z29 , and hence it
is prime to 29. Now 15 × 2 ≡ 30 ≡ 1(mod 29). Hence (detA)−1 = 15−1 = 2 in
Z29 . Therefore
   
−1 −1 17 −3 5 23
A = (detA) (adjA) = 2 =
−5 12 19 24

in Z29 .

Exercise 4: The given equations are

     
x 4 1 −1
A = , where A = .
y 10 7 −4

Now detA = 3 which is prime to 26. Hence 3−1 exists in Z26 . In fact 3−1 = 9
as 3 · 9 ≡ 1 (mod 26). Therefore,
   
−4 1 −36 9
A−1 = 9 =
−7 1 −63 9

in Z26 . One can check that AA−1 = A−1 A = I2 in Z26 .

Exercise 6: “MHQV” should be deciphered as “MIKE.” Let the deciphering

matrix be  
a b
.
c d
Then,
         
a b Mn Mn a b Qn Kn
= , and =
c d Hn In c d Vn En
where the suffix n in Mn stands for the numerical equivalence of M ,
namely, 12, etc. (See Table 6.1).Thus, we have
         
a b 12 12 a b 16 10
= and = .
c d 7 8 c d 21 4

303
304 Appendix F

These give a pair of simultaneous linear equations in a, b and c, d, namely,

12a + 7b = 12, 16a + 21b = 10 and

12c + 7d = 8, 16c + 21d = 4.

Note that we have to solve these equations in Z29 , and since 29 is a prime,
Z29 is actually a field. We can add and subtract multiples of 29 to make our
calculations become easier but must avoid multiplication by 29.
Now the first set of equations is equivalent to

12a + 7b = 12,
and 16a − 8b = 10,
⇒ 48a + 28b = 48,
48a − 24b = 30.

Subtraction gives 52b = 18(mod 29) ⇒ −6b ≡ 18 ⇒ b ≡ −3 ≡ 26 (all

congruences are modulo 29). Now 12a + 7b = 12 gives 12a − 21 ≡ 12 ⇒ 12a =
33 ≡ 4 ⇒ 3a ≡ 1 ⇒ a = 10.
We now consider the second set of equations, namely, 12c + 7d = 8 and
16c + 21d = 4. Multiply the first equation by 3, and subtract from it the
second equation. This gives 20c = 20 ⇒ c = 1 and hence d = 16.
Hence the deciphering matrix is
   
a b 10 26
A−1 = = .
c d 1 16

To simplify over-working, we can take

 
10 −3
A−1 = .
1 16

We now start deciphering pairs of letters in succession from left to right in the
ciphertext. The first
  pair in the ciphertext is “AM ,” and its corresponding
0
column vector is . Hence it represents the plaintext
19
    
10 −3 0 −57
≡ (as 19 ≡ −10(mod 29))
1 16 19 −160
 
1
≡ (as − 160 ≡ 14(mod 29))
14
 “BO”.

Thus, “AM ” represents to “BO” in the plaintext. In a similar way, it can be

checked that the following is true:
 Appendix F 305

Ciphertext Corresponding Plaintext Pair

AM BO
GQ MB
TZ !N
AF OW
JV ?(space)
MH MI
QV KE

Hence the plaintext is: BOMB!NOW? MIKE.

Bibliography

1. Knuth D. E., The Art of Computer Programming. Volume 1: Fundamental Algo-

rithms, 2nd edition, Addison-Wesley, Reading, MA, 1973; Volume 2, Seminumer-
ical Algorithms, second edition, Addison-Wesley, Reading, MA, 1981; Volume 3:
Sorting and Searching, second printing, Addison-Wesley, Reading, MA, 1975.

2. Sedgewick R., Algorithms, Addison-Wesley, Reading, MA, 1989; Algorithms in

C, Addison-Wesley, Reading, MA, 1990; Algorithms in C++, Addison-Wesley,
Reading, MA, 1998.

3. Aho A. V., Hopcroft J. E., Ullman J. D., The Design and Analysis of Algorithms,
Addison-Wesley, Reading, MA, 1975; Data Structures and Algorithms, Reading,
MA, 1983

4. Graham R. L., Knuth D. E., Pathashnik O., Concrete Mathematics, Addison-

Wesley, Reading, MA, 1988.

5. Wirth N., Algorithms + Data Structures = Programs, Prentice-Hall, Inc., New

Jersey, 1976; Systematic Programming, An Introduction, Prentice-Hall, Inc.,
New Jersey, 1973.

6. Balakrishnan R., Sriraman Sridharan, Foundations of Discrete Mathematics

with Algorithms and Programming, CRC Press, Taylor & Francis, 2018.

7. M. Garey, D. S. Johnson, Computers and Intractability, A Guide to the Theory

of NP–Completeness, W.H. Freeman Company, New York, 1979.

8. Balakrishnan R., Ranganathan K., A Textbook of Graph Theory, 2nd edition,

Springer, 2012.

9. Berge C., The Theory of Graphs and its Applications, Wiley, New York, 1958;
Graphs and Hypergraphs, North Holland, 1973.

10. Berge C., Graphes, Dunod, Paris, 1970.

11. Parthasarathy K. R., Basic Graph Theory, Tata-McGraw Hill Publishing Com-
pany, New Delhi, 1994.

12. Herstein I. N., Topics in Algebra, Wiley, 2005.

13. Clark A., Elements of Abstract Algebra, Dover Publications Inc., New York,
1984.

14. Fraleigh J. B., A First Course in Abstract Algebra, 7th edition, Wesley Pub. Co.,
2003.

307
308 Bibliography

15. Bhattacharyya P. B., Jain S. K., Nagpaul S. R., Basic Abstract Algebra, 2nd
edition, Cambridge University Press, 1997.

16. Lewis D. W., Matrix Theory, World Scientific Publishing Co: and Allied
Publishers Ltd. (India), 1955.

17. Ryer H. J., Combinatorial Mathematics, 14, Carus Mathematical Monograph,

Wiley, New York, 1963.

18. Singh S., Linear Algebra, Vikas Publishing House, New Delhi, 2000.

19. www.ams.org/mcom/1962-16-079/S0025...1/S0025/S0025-5718-1962-0148256-
1.pdf

20. Hill R., A First Course in Coding Theory, Clarendon Press, Oxford, 2004.

21. Ling S., Coding Theory: A First Course, Cambridge University Press, 2004.

22. Shannon C. E., A mathematical theory of communication, Bell Syst. Tech. J.

27(379–423), 623–656 (1948).

23. van Lint J. H., Introduction to Coding Theory, Springer GTM, 1999.

24. Pless V., Introduction to the Theory of Error Correcting Codes, 3rd edition,
Cambridge University Press, 1998.

25. Agrawal M., Kayal N., Saxena N., PRIMES is in P (http:\www.cse-iitk.ac.in/

users/manindra/primality.ps), August 2002.

26. Agrawal M., Lecture delivered on February 4, 2003 at the Institute of Mathe-
matical Sciences, Chennai on: “PRIMES is in P”. Reproduced in: Mathematics
Newsletter (Published by Ramanujan Mathematical Society), 13, 11–19 (2003).

27. Pomerance C., On the distribution of pseudoprimes, Math. Comput., 37(156),

587–593 (1981).

28. Lidl R., Niederreiter H., Introduction to Finite Fields and Their Applications,
Cambridge University Press, 1986.

29. Motwani R., Raghavan P., Randomized Algorithms, Cambridge University

Press, 1995.

30. Koblitz N., A Course in Number Theory and Cryptography, 2nd Edition, GTM,
Springer, 1994.

31. Stinson D. R., Cryptography: Theory and Practice, Chapman & Hall, CRC, 2003.

32. Salembier R. G., Southerington P., An Implementation of the AKS Primality

Test, CiteSeerX.

33. Nair M., On Chebyshev-type inequalities for primes, Amer. Math. Monthly, 89,
120–129 (1982).

34. Rotman J. J., An Introduction to the Theory of Groups, GTM, Springer, 1999.
 Bibliography 309

35. Bose R. C., Parker E. T., Shrikhande S. S., Further results on the construction of
mutually orthogonal latin squares and the falsity of Euler’s conjecture, Canadian
Journal of Mathematics, 12, 189–203 (1960).

36. Diffie W., Hellman M. E., New directions in cryptography, IEEE Transactions
on Information Theory, II-22(6), 644–654 (1976).

37. Cormen T. H., Leiserson C. E., Rivest R. L., Stein C., Introduction to
Algorithms, 3rd edition, MIT Press, 2009.
Index

(0,1) matrices, 127 extended, 236

1 -factor, 113 generator matrix, 227, 241
Hamming, 227, 231
A full M alternating tree, 133 Hamming weight, see minimum,
a useful trick for traversals, 23 230
activities or tasks, 77 linear, 227, 231
acylic graph, 9 sphere packings, 235
adjacency list, 5 weight
adjacency matrix, 2 minimum, 230
Algebraic Structures, 147–224 Coding theory, 225–247
All-pairs shortest path algorithm, 35 convex combination, 130, 131
ancesstor, 22 critical path, 78
ancestor, 52 critical tasks, 78
approximate algorithm, 93 Cryptography, 249–274
arborescence, 24 cryptosystem, 249
Caesar, 249
back edges, 53 cryptosystems
Benjamin Franklin, 249 affine, 251
bi-stochastic matrix, 128 cut edge, 58
biconnected component, 59 cut vertex or articulation vertex, 58
biconnected graph, 59
block, 59 dad or mom, 22
breadth first search, 103 decode
breadth first search number, 105 standard array, 232
Breadth-first or Level-order search, 105 syndrome, 237
Breadth-First Search, 103 decryption, 249
breadth-first search, 50 depth-first search, 50
bridge, 58 depth-first search number, 52
BSC, see binary symmetric channels, descendant, 22, 52
226 descending dotted arcs, 70
determinant, 129
Caley’s theorem, 13 dfs spanning forest, 54
Carmichal number, 262 dfs spanning tree, 54
center, 43 diameter, 43
channels dotted cross arcs, 70
binary symmetric, 226
characterictic function, 144 eccentricity, 43
characteristic vector, 144 encryption, 249
Claude Berge, 118 equation
code homogeneous, 205
dimension, 227 inconsistent, 207

311
312 Index

equation (Continued ) Lagrange, 152, 165

indeterminates, 205 leaf, 22, 23
linear level sets, 106
non-homogeneous, 206 LUP decomposition, 208
exact algorithm, 93
exponential function, 137 M-alternating path, 116
M-augmenting path, 116
fast disjoint-set union algorithm, matching, 99, 113
22 matrix
Finite Fields, 215–224 adjoint, 149
Fleury, 96 hermitian, 150
forest, 19 inverse, 149
free tree, 9 orthogonal, 151
skew-hermitian, 150
Galois, 216 skew-symmetric, 150
Galois field, 216 symmetric, 150
Gaussian elimination, 208 transpose, 149
geodetic, 107 maximum matching, 107, 113
geodetic graph, 107 maximum matching algorithm, 120
good characterization, 121 minimum spanning tree, 13
good labeling, 138 modular-exponentiation algorithm, 265
greedy algorithm, 13 MOLS, 223
Group mounting dotted arcs, 70
abelian, 154 MSTEULER algorithm for tsp, 96
Automorphism, 170–176
centre, 172
networks, 12
cyclic, 159
NN algorithm, 93
dihedral, 157
NP-complete problem, 90
homomorphism, 167
NP-Complete problems, 78
kernel, 169
isomorphism, 167
Klein’s 4, 156 parity argument, 114
nonabelian, 154 partial M alternating tree, 133
permutation, 161 perfect matching, 99, 113
Quotient, 172 peripheral vertex, 122
subgroup, 171 permutation matrix, 127
table, 155 PERT and the critical path
Groups, 152–170 method, 76
Petersen graph, 107
hamiltonian cycle, 91 pivot vertex, 27, 37
hamiltonian graph, 145 polynomial
heap data structure, 21 cyclotomic, 272
heuristic, 13, 93 monic, 218
Hopcroft’s algorithm, 58 primitive, 218
Horner’s method, 263 postorder, 22
preorder, 22
inorder, 22 program evaluation
isthmus, 58 and research technique, 78
pseudo code dijkstra, 26
Kuhn-Munkres algorithm, 137 pseudo code prim, 14
 Index 313

radius, 43 sum of the vectors, 131

Randomized algorithm, 263 symmetric difference, 117
regular subgraph problem, 121 synthetic division, 263
regular submatrix, 129
repeated squaring, 265 Tarjan’s strong components
Ring, 179 algorithm, 80
commutative, 180 theorem
noncommutative, 180 Fermat’s Little, 262
units, 181 Lagrange’s, 163
unity element, 181 topological sort, 72
root, 24 traveling salesman problem, 51
strong components, 83 Travelling Salesman Problem, 90
Rooted tree, 22 tree, 9
triangle inequality, 94, 96, 99
scalar product, 131 triangular form, 129
Schur complement, 209 tsp using minimum matching, 99
set Tutte’s perfect matching theorem, 145
spanning, 193
single-source shortest vector, 130
path algorithm, 23
son/daughter, 22 weighted adjacency lists, 13
spanning tree, 11 weighted adjacency matrix, 12
Sriraman Sridharan, 122 weighted matching
strongly connected components, 79 problem in bipartite graph, 137
strongly connected graph, 79 weighted perfect
subgroup, 158 matching, 136