Introduction to Algorithms and Data Structures

M ICHAEL J. D INNEEN G EORGY G IMEL’ FARB M ARK C. W ILSON

c
2016

(Fourth edition)
Contents

Contents 2

List of Figures 5

List of Tables 7

Preface 8

I Introduction to Algorithm Analysis 14

1 What is Algorithm Analysis? 15

1.1 Efficiency of algorithms: first examples . . . . . . . . . . . . . . . . . . . 16
1.2 Running time for loops and other computations . . . . . . . . . . . . . 20
1.3 “Big-Oh”, “Big-Theta”, and “Big-Omega” tools . . . . . . . . . . . . . . . 21
1.4 Time complexity of algorithms . . . . . . . . . . . . . . . . . . . . . . . 27
1.5 Basic recurrence relations . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.6 Capabilities and limitations of algorithm analysis . . . . . . . . . . . . . 36
1.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2 Efficiency of Sorting 39
2.1 The problem of sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2 Insertion sort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3 Mergesort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4 Quicksort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5 Heapsort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6 Data selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.7 Lower complexity bound for sorting . . . . . . . . . . . . . . . . . . . . 66
2.8 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3 Efficiency of Searching 69
3.1 The problem of searching . . . . . . . . . . . . . . . . . . . . . . . . . . 69
 3.2 Sorted lists and binary search . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3 Binary search trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.4 Self-balancing binary and multiway search trees . . . . . . . . . . . . . 82
3.5 Hash tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.6 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

II Introduction to Graph Algorithms 103

4 The Graph Abstract Data Type 104

4.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2 Digraphs and data structures . . . . . . . . . . . . . . . . . . . . . . . . 110
4.3 Implementation of digraph ADT operations . . . . . . . . . . . . . . . . 114
4.4 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

5 Graph Traversals and Applications 117

5.1 Generalities on graph traversal . . . . . . . . . . . . . . . . . . . . . . . 117
5.2 DFS and BFS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.3 Additional properties of depth-first search . . . . . . . . . . . . . . . . . 124
5.4 Additional properties of breadth-first search . . . . . . . . . . . . . . . . 129
5.5 Priority-first search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.6 Acyclic digraphs and topological ordering . . . . . . . . . . . . . . . . . 133
5.7 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.8 Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.9 Maximum matchings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.10 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

6 Weighted Digraphs and Optimization Problems 151

6.1 Weighted digraphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.2 Distance and diameter in the unweighted case . . . . . . . . . . . . . . 153
6.3 Single-source shortest path problem . . . . . . . . . . . . . . . . . . . . 154
6.4 All-pairs shortest path problem . . . . . . . . . . . . . . . . . . . . . . . 161
6.5 Minimum spanning tree problem . . . . . . . . . . . . . . . . . . . . . . 164
6.6 Hard graph problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

III Appendices 170

A Java code for Searching and Sorting 171

A.1 Sorting and selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
A.2 Search methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
B Java graph ADT 178
B.1 Java adjacency matrix implementation . . . . . . . . . . . . . . . . . . . 181
B.2 Java adjacency lists implementation . . . . . . . . . . . . . . . . . . . . 186
B.3 Standardized Java graph class . . . . . . . . . . . . . . . . . . . . . . . . 191
B.4 Extended graph classes: weighted edges . . . . . . . . . . . . . . . . . . 192

C Background on Data Structures 201

C.1 Informal discussion of ADTs . . . . . . . . . . . . . . . . . . . . . . . . . 201
C.2 Notes on a more formal approach . . . . . . . . . . . . . . . . . . . . . . 203

D Mathematical Background 204

D.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
D.2 Mathematical induction . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
D.3 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
D.4 Basic rules of logarithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
D.5 L’Hôpital’s rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
D.6 Arithmetic, geometric, and other series . . . . . . . . . . . . . . . . . . . 207
D.7 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

E Solutions to Selected Exercises 211

Bibliography 234

Index 235
List of Figures

1.1 Linear-time algorithm to sum an array. . . . . . . . . . . . . . . . . . . . . . 17

1.2 Quadratic time algorithm to compute sums of an array. . . . . . . . . . . . 18
1.3 Linear-time algorithm to compute sums of an array. . . . . . . . . . . . . . 19
1.4 “Big Oh” property: g(n) is O(n). . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5 Telescoping as a recursive substitution. . . . . . . . . . . . . . . . . . . . . 34

2.1 Insertion sort for arrays. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.2 Recursive mergesort for arrays . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3 Linear time merge for arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4 Basic array-based quicksort. . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.5 Complete binary tree and its array representation. . . . . . . . . . . . . . . 56
2.6 Maximum heap and its array representation. . . . . . . . . . . . . . . . . . 58
2.7 Heapsort. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.8 Basic array-based quickselect. . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.9 Decision tree for n = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.1 A sequential search algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.2 Binary search for the key 42. . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3 Binary tree representation of a sorted array. . . . . . . . . . . . . . . . . . . 75
3.4 Binary search with three-way comparisons. . . . . . . . . . . . . . . . . . . 76
3.5 Faster binary search with two-way comparisons. . . . . . . . . . . . . . . . 76
3.6 Binary trees: only the leftmost tree is a binary search tree. . . . . . . . . . . 77
3.7 Search and insertion in the binary search tree. . . . . . . . . . . . . . . . . 78
3.8 Removal of the node with key 10 from the binary search tree. . . . . . . . . 79
3.9 Binary search trees obtained by permutations of 1, 2, 3, 4. . . . . . . . . . . 80
3.10 Binary search trees of height about log n. . . . . . . . . . . . . . . . . . . . . 81
3.11 Binary search trees of height about n. . . . . . . . . . . . . . . . . . . . . . . 81
3.12 Left and right rotations of a BST. . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.13 Multiway search tree of order m = 4. . . . . . . . . . . . . . . . . . . . . . . 85
3.14 2–4 B-tree with the leaf storage size 7. . . . . . . . . . . . . . . . . . . . . . 87
3.15 Birthday paradox: Pr365 (n). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.1 A graph G1 and a digraph G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 A subdigraph and a spanning subdigraph of G2 . . . . . . . . . . . . . . . . . 108
4.3 The subdigraph of G2 induced by {1, 2, 3}. . . . . . . . . . . . . . . . . . . . 109
4.4 The reverse of digraph G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.5 The underlying graph of G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

5.1 Graph traversal schema. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5.2 Node states in the middle of a digraph traversal. . . . . . . . . . . . . . . . 118
5.3 Decomposition of a digraph in terms of search trees. . . . . . . . . . . . . . 120
5.4 A graph G1 and a digraph G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.5 BFS trees for G1 and G2 , rooted at 0. . . . . . . . . . . . . . . . . . . . . . . . 123
5.6 DFS trees for G1 and G2 , rooted at 0. . . . . . . . . . . . . . . . . . . . . . . 124
5.7 Depth-first search algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.8 Recursive DFS visit algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.9 Breadth-first search algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.10 Priority-first search algorithm (first kind). . . . . . . . . . . . . . . . . . . . 134
5.11 Digraph describing structure of an arithmetic expression. . . . . . . . . . . 135
5.12 Topological orders of some DAGs. . . . . . . . . . . . . . . . . . . . . . . . . 136
5.13 A digraph and its strongly connected components. . . . . . . . . . . . . . . 140
5.14 Structure of a digraph in terms of its strong components. . . . . . . . . . . 140
5.15 Some (di)graphs with different cycle behaviour. . . . . . . . . . . . . . . . . 143
5.16 A bipartite graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
5.17 A maximal and maximum matching in a bipartite graph. . . . . . . . . . . 146
5.18 An algorithm to find an augmenting path. . . . . . . . . . . . . . . . . . . . 147
5.19 Structure of the graph traversal tree for finding augmenting paths. . . . . . 149

6.1 Some weighted (di)graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

6.2 Dijkstra’s algorithm, first version. . . . . . . . . . . . . . . . . . . . . . . . . 155
6.3 Picture for proof of Dijkstra’s algorithm. . . . . . . . . . . . . . . . . . . . . 157
6.4 Dijkstra’s algorithm, PFS version. . . . . . . . . . . . . . . . . . . . . . . . . 158
6.5 Bellman–Ford algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.6 Floyd’s algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.7 Prim’s algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.8 Kruskal’s algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

B.1 Sample output of the graph test program. . . . . . . . . . . . . . . . . . . . 193

D.1 Approximation of an integral by lower rectangles. . . . . . . . . . . . . . . . 209

List of Tables

1.1 Relative growth of linear and quadratic terms in an expression. . . . . . . . 18

1.2 Relative growth of running time T (n) when the input size increases. . . . . 27
1.3 The largest data sizes n that can be processed by an algorithm. . . . . . . . 28

2.1 Sample execution of insertion sort. . . . . . . . . . . . . . . . . . . . . . . . 42

2.2 Number of inversions Ii , comparisons Ci and data moves Mi . . . . . . . . . 44
2.3 Partitioning in quicksort with pivot p = 31. . . . . . . . . . . . . . . . . . . . 54
2.4 Inserting a new node with the key 75 in the heap in Figure 2.6. . . . . . . . 59
2.5 Deletion of the maximum key from the heap in Figure 2.6. . . . . . . . . . 59
2.6 Successive steps of heapsort. . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.1 A map between airport codes and locations. . . . . . . . . . . . . . . . . . . 70

3.2 Height of the optimal m-ary search tree with n nodes. . . . . . . . . . . . . 85
3.3 Open addressing with linear probing (OALP). . . . . . . . . . . . . . . . . . 90
3.4 Open addressing with double hashing (OADH). . . . . . . . . . . . . . . . . 91
3.5 Birthday paradox: Pr365 (n). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.6 Average search time bounds in hash tables with load factor λ. . . . . . . . . 97

4.1 Digraph operations in terms of data structures. . . . . . . . . . . . . . . . . 114

4.2 Comparative worst-case performance of adjacency lists and matrices. . . . 115

6.1 Illustrating Dijkstra’s algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . 156

Introduction to the Fourth Edition

The fourth edition follows the third edition, while incorporates fixes for the errata
discovered in the third edition.

The textbook’s coverpage displays the complete list of 88 connected forbidden

minors for graphs with vertex cover at most 6, computed by M. J. Dinneen and
L. Xiong in 2001.

This work is licensed under a Creative Commons Attribution-NonCommercial

3.0 Unported License.

Michael J. Dinneen
Georgy Gimel’farb
Mark C. Wilson

Department of Computer Science

University of Auckland
Auckland, New Zealand

{mjd, georgy, mcw}@cs.auckland.ac.nz

February 2016
(Minor revisions December 2017)
Introduction to the Third Edition

The focus for this third edition has been to make an electronic version that students
can read on the tablets and laptops that they bring to lectures. The main changes
from the second edition are:

• Incorporated fixes for the errata discovered in the second edition.

• Made e-book friendly by adding cross referencing links to document elements
and formatted pages to fit on most small screen tablets (i.e., width of margins
minimized).
• Limited material for University of Auckland’s CompSci 220 currently taught
topics. Removed the two chapters on formal languages (Part III) and truncated
the book title.
• This work is licensed under a Creative Commons Attribution-NonCommercial
3.0 Unported License.

Michael J. Dinneen
Georgy Gimel’farb
Mark C. Wilson

Department of Computer Science

University of Auckland
Auckland, New Zealand

{mjd, georgy, mcw}@cs.auckland.ac.nz

March 2013
Introduction to the Second Edition

Writing a second edition is a thankless task, as is well known to authors. Much of the
time is spent on small improvements that are not obvious to readers. We have taken
considerable efforts to correct a large number of errors found in the first edition, and
to improve explanation and presentation throughout the book, while retaining the
philosophy behind the original. As far as material goes, the main changes are:
• more exercises and solutions to many of them;
• a new section on maximum matching (Section 5.9);
• a new section on string searching (Part III);
• a Java graph library updated to Java 1.6 and freely available for download.
The web site http://www.cs.auckland.ac.nz/textbookCS220/ for the book pro-
vides additional material including source code. Readers finding errors are encour-
aged to contact us after viewing the errata page at this web site.
In addition to the acknowledgments in the first edition, we thank Sonny Datt for
help with updating the Java graph library, Andrew Hay for help with exercise solu-
tions and Cris Calude for comments. Rob Randtoul (PlasmaDesign.co.uk) kindly
allowed us to use his cube artwork for the book’s cover. Finally, we thank
MJD all students who have struggled to learn from the first edition and have given
us feedback, either positive or negative;
GLG my wife Natasha and all the family for their permanent help and support;
MCW my wife Golbon and sons Yusef and Yahya, for their sacrifices during the writ-
ing of this book, and the joy they bring to my life even in the toughest times.

31 October 2008
Introduction to the First Edition

This book is an expanded, and, we hope, improved version of the coursebook for
the course COMPSCI 220 which we have taught several times in recent years at the
University of Auckland.
We have taken the step of producing this book because there is no single text
available that covers the syllabus of the above course at the level required. Indeed,
we are not aware of any other book that covers all the topics presented here. Our
aim has been to produce a book that is straightforward, concise, and inexpensive,
and suitable for self-study (although a teacher will definitely add value, particularly
where the exercises are concerned). It is an introduction to some key areas at the
theoretical end of computer science, which nevertheless have many practical appli-
cations and are an essential part of any computer science student’s education.
The material in the book is all rather standard. The novelty is in the combina-
tion of topics and some of the presentation. Part I deals with the basics of algorithm
analysis, tools that predict the performance of programs without wasting time im-
plementing them. Part II covers many of the standard fast graph algorithms that
have applications in many different areas of computer science and science in gen-
eral. Part III introduces the theory of formal languages, shifting the focus from what
can be computed quickly to what families of strings can be recognized easily by a
particular type of machine.
The book is designed to be read cover-to-cover. In particular Part I should come
first. However, one can read Part III before Part II with little chance of confusion.
To make best use of the book, one must do the exercises. They vary in difficulty
from routine to tricky. No solutions are provided. This policy may be changed in a
later edition.
The prerequisites for this book are similar to those of the above course, namely
two semesters of programming in a structured language such as Java (currently used
at Auckland). The book contains several appendices which may fill in any gaps in
the reader’s background.
A limited bibliography is given. There are so many texts covering some of the
topics here that to list all of them is pointless. Since we are not claiming novelty
of material, references to research literature are mostly unnecessary and we have
omitted them. More advanced books (some listed in our bibliography) can provide
more references as a student’s knowledge increases.
A few explanatory notes to the reader about this textbook are in order.
We describe algorithms using a pseudocode similar to, but not exactly like, many
structured languages such as Java or C++. Loops and control structures are indented
in fairly traditional fashion. We do not formally define our pseudocode of comment
style (this might make an interesting exercise for a reader who has mastered Part III).
We make considerable use of the idea of ADT (abstract data type). An abstract
data type is a mathematically specified collection of objects together with opera-
tions that can be performed on them, subject to certain rules. An ADT is completely
independent of any computer programming implementation and is a mathematical
structure similar to those studied in pure mathematics. Examples in this book in-
clude digraphs and graphs, along with queues, priority queues, stacks, and lists. A
data structure is simply a higher level entity composed of the elementary memory
addresses related in some way. Examples include arrays, arrays of arrays (matrices),
linked lists, doubly linked lists, etc.
The difference between a data structure and an abstract data type is exemplified
by the difference between a standard linear array and what we call a list. An array is
a basic data structure common to most programming languages, consisting of con-
tiguous memory addresses. To find an element in an array, or insert an element, or
delete an element, we directly use the address of the element. There are no secrets
in an array. By contrast, a list is an ADT. A list is specified by a set S of elements
from some universal set U, together with operations insert, delete, size, isEmpty
and so on (the exact definition depends on who is doing the defining). We denote
the result of the operation as S.isEmpty(), for example. The operations must sat-
isfy certain rules, for example: S.isEmpty() returns a boolean value TRUE or FALSE;
S.insert(x, r) requires that x belong to U and r be an integer between 0 and S.size(),
and returns a list; for any admissible x and r we have S.isEmpty(S.insert(x, r)) =
FALSE, etc. We are not interested in how the operations are to be carried out, only
in what they do. Readers familiar with languages that facilitate object-based and
object-oriented programming will recognize ADTs as, essentially, what are called
classes in Java or C++.
A list can be implemented using an array (to be more efficient, we would also
have an extra integer variable recording the array size). The insert operation, for ex-
ample, can be achieved by accessing the correct memory address of the r-th element
of the array, allocating more space at the end of the array, shifting along some ele-
ments by one, and assigning the element to be inserted to the address vacated by the
shifting. We would also update the size variable by 1. These details are unimportant
in many programming applications. However they are somewhat important when
discussing complexity as we do in Part I. While ADTs allow us to concentrate on algo-
rithms without worrying about details of programming implementation, we cannot
ignore data structures forever, simply because some implementations of ADT oper-
ations are more efficient than others.
In summary, we use ADTs to sweep programming details under the carpet as long
as we can, but we must face them eventually.
A book of this type, written by three authors with different writing styles under
some time pressure, will inevitably contain mistakes. We have been helped to mini-
mize the number of errors by the student participants in the COMPSCI 220 course-
book error-finding competition, and our colleagues Joshua Arulanandham and An-
dre Nies, to whom we are very grateful.
Our presentation has benefitted from the input of our colleagues who have taught
COMPSCI 220 in the recent and past years, with special acknowledgement due to
John Hamer and the late Michael Lennon.

10 February 2004
 Part I

Introduction to Algorithm Analysis

Chapter 1

What is Algorithm Analysis?

Algorithmic problems are of crucial importance in modern life. Loosely speaking,

they are precisely formulated problems that can be solved in a step-by-step, me-
chanical manner.
Definition 1.1 (informal). An algorithm is a list of unambiguous rules that spec-
ify successive steps to solve a problem. A computer program is a clearly specified
sequence of computer instructions implementing the algorithm.
For example, a sufficiently detailed recipe for making a cake could be thought of
as an algorithm. Problems of this sort are not normally considered as part of com-
puter science. In this book, we deal with algorithms for problems of a more abstract
nature. Important examples, all discussed in this book, and all with a huge number
of practical applications, include: sorting a database, finding an entry in a database,
finding a pattern in a text document, finding the shortest path through a network,
scheduling tasks as efficiently as possible, finding the median of a statistical sample.
Strangely enough, it is very difficult to give simple precise mathematical defini-
tions of algorithms and programs. The existing very deep general definitions are too
complex for our purposes. We trust that the reader of this book will obtain a good
idea of what we mean by algorithm from the examples in this and later chapters.
We often wish to compare different algorithms for the same problem, in order
to select the one best suited to our requirements. The main features of interest are:
whether the algorithm is correct (does it solve the problem for all legal inputs), and
how efficient it is (how much time, memory storage, or other resources it uses).
The same algorithm can be implemented by very different programs written in
different programming languages, by programmers of different levels of skill, and
then run on different computer platforms under different operating systems. In
searching for the best algorithm, general features of algorithms must be isolated
from peculiarities of particular platforms and programs.
To analyse computer algorithms in practice, it is usually sufficient to first specify
elementary operations of a “typical” computer and then represent each algorithm
as a sequence of those operations.
Most modern computers and languages build complex programs from ordinary
arithmetic and logical operations such as standard unary and binary arithmetic op-
erations (negation, addition, subtraction, multiplication, division, modulo opera-
tion, or assignment), Boolean operations, binary comparisons (“equals”, “less than”,
or “greater than”), branching operations, and so on. It is quite natural to use these
basic computer instructions as algorithmic operations, which we will call elemen-
tary operations.
It is not always clear what should count as an elementary operation. For example,
addition of two 64-bit integers should definitely count as elementary, since it can be
done in a fixed time for a given implementation. But for some applications, such as
cryptography, we must deal with much larger integers, which must be represented
in another way. Addition of “big” integers takes a time roughly proportional to the
size of the integers, so it is not reasonable to consider it as elementary. From now on
we shall ignore such problems. For most of the examples in this introductory book
they do not arise. However, they must be considered in some situations, and this
should be borne in mind.

Definition 1.2 (informal). The running time (or computing time) of an algorithm is
the number of its elementary operations.

The actual execution time of a program implementing an algorithm is roughly

proportional to its running time, and the scaling factor depends only on the partic-
ular implementation (computer, programming language, operating system, and so
on).
The memory space required for running an algorithm depends on how many
individual variables (input, intermediate, and output data) are involved simultane-
ously at each computing step. Time and space requirements are almost always inde-
pendent of the programming language or style and characterise the algorithm itself.
From here on, we will measure effectiveness of algorithms and programs mostly in
terms of their time requirements. Any real computer has limits on the size and the
number of data items it can handle.

1.1 Efficiency of algorithms: first examples

If the same problem can be solved by different algorithms, then all other things
being equal, the most efficient algorithm uses least computational resources. The
theory of algorithms in modern computer science clarifies basic algorithmic notions
 algorithm linearSum
Input: array a[0..n − 1]
begin
s←0
for i ← 0 step i ← i + 1 until n − 1 do
s ← s + a[i]
end for
return s
end

Figure 1.1: Linear-time algorithm to sum an array.

such as provability, correctness, complexity, randomness, or computability. It stud-

ies whether there exist any algorithms for solving certain problems, and if so, how
fast can they be. In this book, we take only a few small steps into this domain.
To search for the most efficient algorithm, one should mathematically prove cor-
rectness of and determine time/space resources for each algorithm as explicit func-
tions of the size of input data to process. For simplicity of presentation in this book,
we sometimes skip the first step (proof of correctness), although it is very impor-
tant. The focus of this chapter is to introduce methods for estimating the resource
requirements of algorithms.
Example 1.3 (Sum of elements of an array). Let a denote an array of integers where
n−1
the sum s = ∑i=0 a[i] is required. To get the sum s, we have to repeat n times the same
elementary operations (fetching from memory and adding a number). Thus, run-
ning time T (n) is proportional to, or linear in n: T (n) = cn. Such algorithms are also
called linear algorithms. The unknown factor c depends on a particular computer,
programming language, compiler, operating system, etc. But the relative change
in running time is just the same as the change in the data size: T (10) = 10T (1), or
T (1000) = 1000T (1), or T (1000) = 10T (100). The linear algorithm in Figure 1.1 imple-
ments a simple loop.

Example 1.4 (Sums of subarrays). The problem is to compute, for each subarray
a[ j.. j + m − 1] of size m in an array a of size n, the partial sum of its elements s[ j] =
m−1
∑k=0 a[ j + k]; j = 0, . . . , n − m. The total number of these subarrays is n − m + 1. At first
glance, we need to compute n − m + 1 sums, each of m items, so that the running time
is proportional to m(n − m + 1). If m is fixed, the time depends still linearly on n.
But if m is growing with n as a fraction of n, such as m = 2n , then T (n) = c 2n n2 + 1

= 0.25cn2 + 0.5cn. The relative weight of the linear part, 0.5cn, decreases quickly with
respect to the quadratic one as n increases. For example, if T (n) = 0.25n2 + 0.5n, we
see in the last column of Table 1.1 the rapid decrease of the ratio of the two terms.
 Table 1.1: Relative growth of linear and quadratic terms in an expression.

n T (n) 0.25n2 0.5n

value % of quadratic term
10 30 25 5 20.0
50 650 625 25 4.0
100 2550 2500 50 2.0
500 62750 62500 250 0.4
1000 250500 250000 500 0.2
5000 6252500 6250000 2500 0.04

Thus, for large n only the quadratic term becomes important and the running
time is roughly proportional to n2 , or is quadratic in n. Such algorithms are some-
times called quadratic algorithms in terms of relative changes of running time with
respect to changes of the data size: if T (n) ≈ cn2 then T (10) ≈ 100T (1), or T (100) ≈
10000T (1), or T (100) ≈ 100T (10).

algorithm slowSums
Input: array a[0..2m − 1]
begin
array s[0..m]
for i ← 0 to m do
s[i] ← 0
for j ← 0 to m − 1 do
s[i] ← s[i] + a[i + j]
end for
end for
return s
end

Figure 1.2: Quadratic time algorithm to compute sums of an array.

The “brute-force” quadratic algorithm has two nested loops (see Figure 1.2). Let
us analyse it to find out whether it can be simplified. It is easily seen that repeated
computations in the innermost loop are unnecessary. Two successive sums s[i] and
s[i − 1] differ only by two elements: s[i] = s[i − 1] + a[i + m − 1] − a[i − 1]. Thus we need
not repeatedly add m items together after getting the very first sum s[0]. Each next
sum is formed from the current one by using only two elementary operations (ad-
dition and subtraction). Thus T (n) = c(m + 2(n − m)) = c(2n − m). In the first paren-
theses, the first term m relates to computing the first sum s[0], and the second term
2(n − m) reflects that n − m other sums are computed with only two operations per
sum. Therefore, the running time for this better organized computation is always
linear in n for each value m, either fixed or growing with n. The time for comput-
ing all the sums of the contiguous subsequences is less than twice that taken for the
single sum of all n items in Example 1.3
The linear algorithm in Figure 1.3 excludes the innermost loop of the quadratic
algorithm. Now two simple loops, doing m and 2(n − m) elementary operations, re-
spectively, replace the previous nested loop performing m(n − m + 1) operations.

algorithm fastSums
Input: array a[0..2m − 1]
begin
array s[0..m]
s[0] ← 0
for j ← 0 to m − 1 do
s[0] ← s[0] + a[ j]
end for
for i ← 1 to m do
s[i] ← s[i − 1] + a[i + m − 1] − a[i − 1]
end for
return s;
end

Figure 1.3: Linear-time algorithm to compute sums of an array.

Such an outcome is typical for algorithm analysis. In many cases, a careful analy-
sis of the problem allows us to replace a straightforward “brute-force” solution with
much more effective one. But there are no “standard” ways to reach this goal. To ex-
clude unnecessary computation, we have to perform a thorough investigation of the
problem and find hidden relationships between the input data and desired outputs.
In so doing, we should exploit all the tools we have learnt. This book presents many
examples where analysis tools are indeed useful, but knowing how to analyse and
solve each particular problem is still close to an art. The more examples and tools
are mastered, the more the art is learnt.
Exercises
Exercise 1.1.1. A quadratic algorithm with processing time T (n) = cn2 uses 500 ele-
mentary operations for processing 10 data items. How many will it use for processing
1000 data items?
Exercise 1.1.2. Algorithms A and B use exactly TA (n) = cA n lg n and TB (n) = cB n2 ele-
mentary operations, respectively, for a problem of size n. Find the fastest algorithm
for processing n = 220 data items if A and B spend 10 and 1 operations, respectively,
to process 210 ≡ 1024 items.

1.2 Running time for loops and other computations

The above examples show that running time depends considerably on how deeply
the loops are nested and how the loop control variables are changing. Suppose the
control variables change linearly in n, that is, increase or decrease by constant steps.
If the number of elementary operations in the innermost loop is constant, the nested
loops result in polynomial running time T (n) = cnk where k is the highest level of
nesting and c is some constant. The first three values of k have special names: lin-
ear time, quadratic time, and cubic time for k = 1 (a single loop), k = 2 (two nested
loops), and k = 3 (three nested loops), respectively.
When loop control variables change non-linearly, the running time also varies
non-linearly with n.

Example 1.5. An exponential change i = 1, k, k2 , . . . , km−1 of the control variable in

the range 1 ≤ i ≤ n results in logarithmic time for a simple loop. The loop executes
m iterations such that km−1 ≤ n < km . Thus, m − 1 ≤ logk n < m, and T (n) = c⌈logk n⌉.

Additional conditions for executing inner loops only for special values of the
outer variables also decrease running time.

Example 1.6. Let us roughly estimate the running time of the following nested loops:
m←2
for j ← 1 to n do
if j = m then
m ← 2m
for i ← 1 to n do
. . . constant number of elementary operations
end for
end if
end for

The inner loop is executed k times for j = 2, 4, . . . , 2k where k < lg n ≤ k + 1. The

total time for the elementary operations is proportional to kn, that is, T (n) = n⌊lg n⌋.

Conditional and switch operations like if {condition} then {constant running

time T1 } else {constant running time T2 } involve relative frequencies of the groups
of computations. The running time T satisfies T = ftrue T1 + (1 − ftrue )T2 < max{T1 , T2 }
where ftrue is the relative frequency of the true condition value in the if-statement.
 The running time of a function or method call is T = ∑ki=1 Ti where Ti is the run-
ning time of statement i of the function and k is the number of statements.
Exercises
Exercise 1.2.1. Is the running time quadratic or linear for the nested loops below?

m←1
for j ← 1 step j ← j + 1 until n do
if j = m then m ← m · (n − 1)
for i ← 0 step i ← i + 1 until n − 1 do
. . . constant number of elementary operations
end for
end if
end for

Exercise 1.2.2. What is the running time for the following code fragment as a func-
tion of n?
for i ← 1 step i ← 2 ∗ i while i < n do
for j ← 1 step j ← 2 ∗ j while j < n do
if j = 2 ∗ i
for k = 0 step k ← k + 1 while k < n do
. . . constant number of elementary operations
end for
else
for k ← 1 step k ← 3 ∗ k while k < n do
. . . constant number of elementary operations
end for
end if
end for
end for

1.3 “Big-Oh”, “Big-Theta”, and “Big-Omega” tools

Two simple concepts separate properties of an algorithm itself from properties
of a particular computer, operating system, programming language, and compiler
used for its implementation. The concepts, briefly outlined earlier, are as follows:

• The input data size, or the number n of individual data items in a single data
instance to be processed when solving a given problem. Obviously, how to
measure the data size depends on the problem: n means the number of items
to sort (in sorting applications), number of nodes (vertices) or arcs (edges) in
 graph algorithms, number of picture elements (pixels) in image processing,
length of a character string in text processing, and so on.

• The number of elementary operations taken by a particular algorithm, or its

running time. We assume it is a function f (n) of the input data size n. The
function depends on the elementary operations chosen to build the algorithm.

The running time of a program which implements the algorithm is c f (n) where
c is a constant factor depending on a computer, language, operating system, and
compiler. Even if we don’t know the value of the factor c, we are able to answer
the important question: if the input size increases from n = n1 to n = n2 , how does
the relative running time of the program change, all other things being equal? The
(n2 )
answer is obvious: the running time increases by a factor of TT (n = cc ff (n 2) f (n2 )
(n1 ) = f (n1 ) .
1)
As we have already seen, the approximate running time for large input sizes gives
enough information to distinguish between a good and a bad algorithm. Also, the
constant c above can rarely be determined. We need some mathematical notation to
avoid having to say “of the order of . . .” or “roughly proportional to . . .”, and to make
this intuition precise.
The standard mathematical tools “Big Oh” (O), “Big Theta” (Θ), and “Big Omega”
(Ω) do precisely this.
Note. Actually, the above letter O is a capital “omicron” (all letters in this notation
are Greek letters). However, since the Greek omicron and the English “O” are indis-
tinguishable in most fonts, we read O() as “Big Oh” rather than “Big Omicron”.
The algorithms are analysed under the following assumption: if the running time
of an algorithm as a function of n differs only by a constant factor from the running
time for another algorithm, then the two algorithms have essentially the same time
complexity. Functions that measure running time, T (n), have nonnegative values
because time is nonnegative, T (n) ≥ 0. The integer argument n (data size) is also
nonnegative.

Definition 1.7 (Big Oh). Let f (n) and g(n) be nonnegative-valued functions defined
on nonnegative integers n. Then g(n) is O( f (n)) (read “g(n) is Big Oh of f (n)”) iff there
exists a positive real constant c and a positive integer n0 such that g(n) ≤ c f (n) for all
n > n0 .

Note. We use the notation “iff ” as an abbreviation of “if and only if”.
In other words, if g(n) is O( f (n)) then an algorithm with running time g(n) runs for
large n at least as fast, to within a constant factor, as an algorithm with running time
f (n). Usually the term “asymptotically” is used in this context to describe behaviour
of functions for sufficiently large values of n. This term means that g(n) for large n
may approach closer and closer to c · f (n). Thus, O( f (n)) specifies an asymptotic
upper bound.
Note. Sometimes the “Big Oh” property is denoted g(n) = O( f (n)), but we should not
assume that the function g(n) is equal to something called “Big Oh” of f (n). This
notation really means g(n) ∈ O( f (n)), that is, g(n) is a member of the set O( f (n)) of
functions which are increasing, in essence, with the same or lesser rate as n tends to
infinity (n → ∞). In terms of graphs of these functions, g(n) is O( f (n)) iff there exists
a constant c such that the graph of g(n) is always below or at the graph of c f (n) after
a certain point, n0 .

Example 1.8. Function g(n) = 100 log10 n in Figure 1.4 is O(n) because the graph g(n)
is always below the graph of f (n) = n if n > 238 or of f (n) = 0.3n if n > 1000, etc.

T(n) f(n)=n
f(n) = 0.3n
400
g(n)=100 log 10 n

300

200

100

n0 n0
0 200 400 600 800 1000 1200 n

Figure 1.4: “Big Oh” property: g(n) is O(n).

Definition 1.9 (Big Omega). The function g(n) is Ω( f (n)) iff there exists a positive
real constant c and a positive integer n0 such that g(n) ≥ c f (n) for all n > n0 .

“Big Omega” is complementary to “Big Oh” and generalises the concept of “lower
bound” (≥) in the same way as “Big Oh” generalises the concept of “upper bound”
(≤): if g(n) is O( f (n)) then f (n) is Ω(g(n)), and vice versa.

Definition 1.10 (Big Theta). The function g(n) is Θ( f (n)) iff there exist two positive
real constants c1 and c2 and a positive integer n0 such that c1 f (n) ≤ g(n) ≤ c2 f (n) for
all n > n0 .

Whenever two functions, f (n) and g(n), are actually of the same order, g(n) is
Θ( f (n)), they are each “Big Oh” of the other: f (n) is O(g(n)) and g(n) is O( f (n)). In
other words, f (n) is both an asymptotic upper and lower bound for g(n). The “Big
Theta” property means f (n) and g(n) have asymptotically tight bounds and are in
some sense equivalent for our purposes.
In line with the above definitions, g(n) is O( f (n)) iff g(n) grows at most as fast as
f (n) to within a constant factor, g(n) is Ω( f (n)) iff g(n) grows at least as fast as f (n) to
within a constant factor, and g(n) is Θ( f (n)) iff g(n) and f (n) grow at the same rate to
within a constant factor.
“Big Oh”, “Big Theta”, and “Big Omega” notation formally capture two crucial
ideas in comparing algorithms: the exact function, g, is not very important because
it can be multiplied by any arbitrary positive constant, c, and the relative behaviour
of two functions is compared only asymptotically, for large n, but not near the origin
where it may make no sense. Of course, if the constants involved are very large, the
asymptotic behaviour loses practical interest. In most cases, however, the constants
remain fairly small.
In analysing running time, “Big Oh” g(n) ∈ O( f (n)), “Big Omega” g(n) ∈ Ω( f (n)),
and “Big Theta” g(n) ∈ Θ( f (n)) definitions are mostly used with g(n) equal to “exact”
running time on inputs of size n and f (n) equal to a rough approximation to running
time (like log n, n, n2 , and so on).
To prove that some function g(n) is O( f (n)), Ω( f (n)), or Θ( f (n)) using the defi-
nitions we need to find the constants c, n0 or c1 , c2 , n0 specified in Definitions 1.7,
1.9, 1.10. Sometimes the proof is given only by a chain of inequalities, starting with
f (n). In other cases it may involve more intricate techniques, such as mathemati-
cal induction. Usually the manipulations are quite simple. To prove that g(n) is not
O( f (n)), Ω( f (n)), or Θ( f (n)) we have to show the desired constants do not exist, that
is, their assumed existence leads to a contradiction.

Example 1.11. To prove that linear function g(n) = an + b; a > 0, is O(n), we form
the following chain of inequalities: g(n) ≤ an + |b| ≤ (a + |b|)n for all n ≥ 1. Thus,
Definition 1.7 with c = a + |b| and n0 = 1 shows that an + b is O(n).

“Big Oh” hides constant factors so that both 10−10 n and 1010 n are O(n). It is point-
less to write something like O(2n) or O(an + b) because this still means O(n). Also,
only the dominant terms as n → ∞ need be shown as the argument of “Big Oh”, “Big
Omega”, or “Big Theta”.

Example 1.12. The polynomial P5 (n) = a5 n5 + a4 n4 + a3 n3 + a2 n2 + a1 n + a0 ; a5 > 0, is

O(n5 ) because P5 (n) ≤ (a5 + |a4 | + |a3 | + |a2 | + |a1 | + |a0 |)n5 for all n ≥ 1.

Example 1.13. The exponential function g(n) = 2n+k , where k is a constant, is O(2n )
because 2n+k = 2k 2n for all n. Generally, mn+k is O(l n ); l ≥ m > 1, because mn+k ≤ l n+k =
l k l n for any constant k.

Example 1.14. For each m > 1, the logarithmic function g(n) = logm (n) has the same
rate of increase as lg(n) because logm (n) = logm (2) lg(n) for all n > 0. Therefore we may
omit the logarithm base when using the “Big-Oh” and “Big Theta” notation: logm n is
Θ(log n).

Rules for asymptotic notation

Using the definition to prove asymptotic relationships between functions is hard
work. As in calculus, where we soon learn to use various rules (product rule, chain
rule, . . . ) rather than the definition of derivative, we can use some simple rules to
deduce new relationships from old ones.
We present rules for “Big Oh”—similar relationships hold for “Big Omega” and
“Big Theta”.
We will consider the features both informally and formally using the following
notation. Let x and y be functions of a nonnegative integer n. Then z = x + y and
z = xy denote the sum of the functions, z(n) = x(n) + y(n), and the product function:
z(n) = x(n)y(n), respectively, for every value of n. The product function (xy)(n) returns
the product of the values of the functions at n and has nothing in common with the
composition x(y(n)) of the two functions.
Basic arithmetic relationships for “Big Oh” follow from and can be easily proven
with its definition.

Lemma 1.15 (Scaling). For all constants c > 0, c f is O( f ). In particular, f is O( f ).

Proof. The relationship c f (n) ≤ c f (n) obviously holds for all n ≥ 0.

Constant factors are ignored, and only the powers and functions are taken into
account. It is this ignoring of constant factors that motivates such a notation.

Lemma 1.16 (Transitivity). If h is O(g) and g is O( f ), then h is O( f ).

Proof. See Exercise 1.3.6.

Informally, if h grows at most as quickly as g, which grows at most as quickly as f ,

then h grows at most as quickly as f .

Lemma 1.17 (Rule of sums). If g1 is O( f1 ) and g2 is O( f2 ), then g1 +g2 is O(max{ f1 , f2 }).

Proof. See Exercise 1.3.6.

If g is O( f ) and h is O( f ), then is g + h is O( f ). In particular, if g is O( f ), then g + f

is O( f ). Informally, the growth rate of a sum is the growth rate of its fastest-growing
term.

Lemma 1.18 (Rule of products). If g1 is O( f1 ) and g2 is O( f2 ), then g1 g2 is O( f1 f2 ).

Proof. See Exercise 1.3.6.

 In particular, if g is O( f ), then gh is O( f h). Informally, the product of upper
bounds of functions gives an upper bound for the product of the functions.
Using calculus we can obtain a nice time-saving rule.

Lemma 1.19 (Limit Rule). Suppose limn→∞ f (n)/g(n) exists (may be ∞), and denote
the limit by L. Then

• if L = 0, then f is O(g) and f is not Ω(g);

• if 0 < L < ∞ then f is Θ(g);

• if L = ∞ then f is Ω(g) and f is not O(g).

Proof. If L = 0 then from the definition of limit, in particular there is some n0 such
that f (n)/g(n) ≤ 1 for all n ≥ n0 . Thus f (n) ≤ g(n) for all such n, and f (n) is O(g(n)) by
definition. On the other hand, for each c > 0, it is not the case that f (n) ≥ cg(n) for
all n past some threshold value n1 , so that f (n) is not Ω(g(n)). The other two parts are
proved in the analogous way.

To compute the limit if it exists, the standard L’Hôpital’s rule of calculus is useful
(see Section D.5).
More specific relations follow directly from the basic ones.

Example 1.20. Higher powers of n grow more quickly than lower powers: nk is O(nl )
if 0 ≤ k ≤ l. This follows directly from the limit rule since nk /nl = nk−l has limit 1 if
k = l and 0 if k < l.

Example 1.21. The growth rate of a polynomial is given by the growth rate of its
leading term (ignoring the leading coefficient by the scaling feature): if Pk (n) is a
polynomial of exact degree k then Pk (n) is Θ(nk ). This follows easily from the limit
rule as in the preceding example.

Example 1.22. Exponential functions grow more quickly than powers: nk is O(bn ),
for all b > 1, n > 1, and k ≥ 0. The restrictions on b, n, and k merely ensure that
both functions are increasing. This result can be proved by induction or by using the
limit-L’Hôpital approach above.

Example 1.23. Logarithmic functions grow more slowly than powers: logb n is O(nk )
for all b > 1, k > 0. This is the inverse of the preceding feature. Thus, as a result, log n
is O(n) and n log n is O(n2 ).
Exercises
Exercise 1.3.1. Prove that 10n3 − 5n + 15 is not O(n2 ).

Exercise 1.3.2. Prove that 10n3 − 5n + 15 is Θ(n3 ).

Exercise 1.3.3. Prove that 10n3 − 5n + 15 is not Ω(n4 ).

Exercise 1.3.4. Prove that f (n) is Θ(g(n)) if and only if both f (n) is O(g(n) and f (n) is
Ω(g(n)).

Exercise 1.3.5. Using the definition, show that each function f (n) in Table 1.3 stands
in “Big-Oh” relation to the preceding one, that is, n is O(n log n), n log n is O(n1.5 ), and
so forth.

Exercise 1.3.6. Prove Lemmas 1.16–1.18.

Exercise 1.3.7. Decide on how to reformulate the Rule of Sums (Lemma 1.17) for
“Big Omega” and “Big Theta” notation.

Exercise 1.3.8. Reformulate and prove Lemmas 1.15–1.18 for “Big Omega” notation.

1.4 Time complexity of algorithms

Definition 1.24 (Informal). A function f (n) such that the running time T (n) of a
given algorithm is Θ( f (n)) measures the time complexity of the algorithm.

An algorithm is called polynomial time if its running time T (n) is O(nk ) where k is
some fixed positive integer. A computational problem is considered intractable iff
no deterministic algorithm with polynomial time complexity exists for it. But many
problems are classed as intractable only because a polynomial solution is unknown,
and it is a very challenging task to find such a solution for one of them.

Table 1.2: Relative growth of running time T (n) when the input size increases from n = 8 to
n = 1024 provided that T (8) = 1.

Time complexity Input size n Time T (n)

Function Notation 8 32 128 1024
Constant 1 1 1 1 1 1
Logarithmic lg n 1 1.67 2.67 3.33 lg n/3
Log-squared lg2 n 1 2.78 5.44 11.1 lg2 n/9
Linear n 1 4 16 128 n/8
“n log n” n lg n 1 6.67 37.3 427 n lg n/24
Quadratic n2 1 16 256 16384 n2 /64
Cubic n3 1 64 4096 2097152 n3 /512
Exponential 2n 1 224 2120 21016 2n−8
 Table 1.2 shows how the running time T (n) of algorithms having different time
complexity, f (n), grows relatively with the increasing input size n. Time complexity
functions are listed in order such that g is O( f ) if g is above f : for example, the linear
function n is O(n log n) and O(n2 ), etc. The asymptotic growth rate does not depend
on the base of the logarithm, but the exact numbers in the table do — we use log2 = lg
for simplicity.

Table 1.3: The largest data sizes n that can be processed by an algorithm with time complexity
f (n) provided that T (10) = 1 minute.

Length of time to run an algorithm

f (n) 1 minute 1 hour 1 day 1 week 1 year 1 decade
n 10 600 14 400 100 800 5.26 × 106 5.26 × 107
n lg n 10 250 3 997 23 100 883 895 7.64 × 106
n1.5 10 153 1 275 4 666 65 128 302,409
n2 10 77 379 1 003 7 249 22,932
n3 10 39 112 216 807 1,738
2n 10 15 20 23 29 32

Table 1.3 is even more expressive in showing how the time complexity of an algo-
rithm affects the size of problems the algorithm can solve (we again use log2 = lg). A
linear algorithm solving a problem of size n = 10 in exactly one minute will process
about 5.26 million data items per year and 10 times more if we can wait a decade. But
an exponential algorithm with T (10) = 1 minute will deal only with 29 data items af-
ter a year of running and add only 3 more items after a decade. Suppose we have
computers 10, 000 times faster (this is approximately the ratio of a week to a minute).
Then we can solve a problem 10, 000 times, 100 times, or 21.5 times larger than before
if our algorithm is linear, quadratic, or cubic, respectively. But for exponential algo-
rithms, our progress is much worse: we can add only 13 more input values if T (n) is
Θ(2n ).
Therefore, if our algorithm has a constant, logarithmic, log-square, linear, or even
“n log n” time complexity we may be happy and start writing a program with no doubt
that it will meet at least some practical demands. Of course, before taking the plunge,
it is better to check whether the hidden constant c, giving the computation volume
per data item, is sufficiently small in our case. Unfortunately, order relations can be
drastically misleading: for instance, two linear functions 10−4 n and 1010 n are of the
same order O(n), but we should not claim an algorithm with the latter time complex-
ity as a big success.
Therefore, we should follow a simple rule: roughly estimate the computation vol-
ume per data item for the algorithms after comparing their time complexities in a
“Big-Oh” sense! We may estimate the computation volume simply by counting the
number of elementary operations per data item.
In any case we should be very careful even with simple quadratic or cubic algo-
rithms, and especially with exponential algorithms. If the running time is speeded
up in Table 1.3 so that it takes one second per ten data items in all the cases, then we
will still wait about 12 days (220 ≡ 1, 048, 576 seconds) for processing only 30 items by
the exponential algorithm. Estimate yourself whether it is practical to wait until 40
items are processed.
In practice, quadratic and cubic algorithms cannot be used if the input size ex-
ceeds tens of thousands or thousands of items, respectively, and exponential algo-
rithms should be avoided whenever possible unless we always have to process data
of very small size. Because even the most ingenious programming cannot make an
inefficient algorithm fast (we would merely change the value of the hidden constant
c slightly, but not the asymptotic order of the running time), it is better to spend more
time to search for efficient algorithms, even at the expense of a less elegant software
implementation, than to spend time writing a very elegant implementation of an
inefficient algorithm.
Worst-case and average-case performance
We have introduced asymptotic notation in order to measure the running time
of an algorithm. This is expressed in terms of elementary operations. “Big Oh”, “Big
Omega” and “Big Theta” notations allow us to state upper, lower and tight asymp-
totic bounds on running time that are independent of inputs and implementation
details. Thus we can classify algorithms by performance, and search for the “best”
algorithms for solving a particular problem.
However, we have so far neglected one important point. In general, the running
time varies not only according to the size of the input, but the input itself. The ex-
amples in Section 1.4 were unusual in that this was not the case. But later we shall
see many examples where it does occur. For example, some sorting algorithms take
almost no time if the input is already sorted in the desired order, but much longer if
it is not.
If we wish to compare two different algorithms for the same problem, it will be
very complicated to consider their performance on all possible inputs. We need a
simple measure of running time.
The two most common measures of an algorithm are the worst-case running
time, and the average-case running time.
The worst-case running time has several advantages. If we can show, for example,
that our algorithm runs in time O(n log n) no matter what input of size n we consider,
we can be confident that even if we have an “unlucky” input given to our program,
it will not fail to run fairly quickly. For so-called “mission-critical” applications this
is an essential requirement. In addition, an upper bound on the worst-case running
time is usually fairly easy to find.
 The main drawback of the worst-case running time as a measure is that it may be
too pessimistic. The real running time might be much lower than an “upper bound”,
the input data causing the worst case may be unlikely to be met in practice, and the
constants c and n0 of the asymptotic notation are unknown and may not be small.
There are many algorithms for which it is difficult to specify the worst-case input.
But even if it is known, the inputs actually encountered in practice may lead to much
lower running times. We shall see later that the most widely used fast sorting algo-
rithm, quicksort, has worst-case quadratic running time, Θ(n2 ), but its running time
for “random” inputs encountered in practice is Θ(n log n).
By contrast, the average-case running time is not as easy to define. The use of
the word “average” shows us that probability is involved. We need to specify a prob-
ability distribution on the inputs. Sometimes this is not too difficult. Often we can
assume that every input of size n is equally likely, and this makes the mathematical
analysis easier. But sometimes an assumption of this sort may not reflect the inputs
encountered in practice. Even if it does, the average-case analysis may be a rather
difficult mathematical challenge requiring intricate and detailed arguments. And of
course the worst-case complexity may be very bad even if the average case complex-
ity is good, so there may be considerable risk involved in using the algorithm.
Whichever measure we adopt for a given algorithm, our goal is to show that its
running time is Θ( f ) for some function f and there is no algorithm with running
time Θ(g) for any function g that grows more slowly than f when n → ∞. In this case
our algorithm is asymptotically optimal for the given problem.
Proving that no other algorithm can be asymptotically better than ours is usually
a difficult matter: we must carefully construct a formal mathematical model of a
computer and derive a lower bound on the complexity of every algorithm to solve
the given problem. In this book we will not pursue this topic much. If our analysis
does show that an upper bound for our algorithm matches the lower one for the
problem, then we need not try to invent a faster one.
Exercises
Exercise 1.4.1. Add columns to Table 1.3 corresponding to one century (10 decades)
and one millennium (10 centuries).
Exercise 1.4.2. Add rows to Table 1.2 for algorithms with time complexity f (n) =
lg lg n and f (n) = n2 lg n.

1.5 Basic recurrence relations

As we will see later, a great many algorithms are based on the following divide-
and-conquer principle:
• divide a large problem into smaller subproblems and recursively solve each
subproblem, then
• combine solutions of the subproblems to solve the original problem.
Running time of such algorithms is determined by a recurrence relation accounting
for the size and number of the subproblems and for the cost of splitting the problem
into them. The recursive relation defines a function “in terms of itself”, that is, by an
expression that involves the same function. The definition is not circular provided
that the value at a natural number n is defined in terms of values at smaller natural
numbers, and the recursion terminates at some base case below which the function
is not defined.

Example 1.25 (Fibonacci numbers). These are defined by one of the most famous
recurrence relations: F(n) = F(n − 1) + F(n − 2); F(1) = 1, and F(2) = 1. The last two
equations are called the base of the recurrence or initial condition. The recurrence
relation uniquely defines the function F(n) at any number n because any particular
value of the function is easily obtained by generating all the preceding values until
the desired term is produced, for example, F(3) = F(2) + F(1) = 2; F(4) = F(3) +
F(2) = 3, and so forth. Unfortunately, to compute F(10000), we need to perform
9998 additions.

Example 1.26. One more recurrence relation is T (n) = 2T (n − 1) + 1 with the base
condition T (0) = 0. Here, T (1) = 2 · 0 + 1 = 1, T (2) = 2 · 1 + 1 = 3, T (3) = 2 · 3 + 1 = 7,
T (4) = 2 · 7 + 1 = 15, and so on.

Note. A recurrence relation is sometimes simply called a recurrence. In engineering

it is called a difference equation.
We will frequently meet recurrences in algorithm analysis. It is more convenient
to have an explicit expression, (or closed-form expression) for the function in or-
der to compute it quickly for any argument value n and to compare it with other
functions. The closed-form expression for T (n), that is, what is traditionally called a
“formula”, makes the growth of T (n) as a function of n more apparent. The process
of deriving the explicit expression is called “solving the recurrence relation”.
Our consideration will be restricted to only the two simplest techniques for solv-
ing recurrences: (i) guessing a solution from a sequence of values T (0), T (1), T (2),
. . . , and proving it by mathematical induction (a “bottom-up” approach) and (ii)
“telescoping” the recurrence (a “top-down” approach). Both techniques allow us to
obtain closed forms of some important recurrences that describe performance of
sort and search algorithms. For instance, in Example 1.26 we can simply guess the
closed form expression T (n) = 2n −1 by inspecting the first few terms of the sequence
0, 1, 3, 7, 15 because 0 = 1−1, 1 = 2−1, 3 = 4−1, 7 = 8−1, and 15 = 16−1. But in other
cases these techniques may fail and more powerful mathematical tools beyond the
scope of this book, such as using characteristic equations and generating functions,
should be applied.
Guessing to solve a recurrence
There is no formal way to find a closed-form solution. But after we have guessed
the solution, it may be proven to be correct by mathematical induction (see Sec-
tion D.2).
Example 1.27. For the recurrence T (n) = 2T (n−1)+1 with the base condition T (0) =
0 in Example 1.26 we guessed the closed-form relationship T (n) = 2n − 1 by analysing
the starting terms 0, 1, 3, 7, 15. This formula is obviously true for n = 0, because
20 − 1 = 0. Now, by the induction hypothesis,
T (n) = 2T (n − 1) + 1 = 2(2n−1 − 1) + 1 = 2n − 1
and this is exactly what we need to prove.
The Fibonacci sequence provides a sterner test for our guessing abilities.
Example 1.28. The first few terms of the sequence 1, 1, 2, 3, 5, 8, 13, 21, 34, . . . give no
hint regarding the desired explicit form. Thus let us analyse the recurrence F(n) =
F(n − 1) + F(n − 2) itself. F(n) is almost doubled every time, so that F(n) < 2n . The
simplest guess F(n) = c2n with c < 1 fails because for any scaling factor c it leads to
the impossible equality 2n = 2n−1 + 2n−2 , or 4 = 2 + 1. The next guess is that the base 2
of the exponential function should be smaller, φ < 2, that is, F(n) = cφn . The resulting
equation φn = φn−1√+ φn−2 reduces to the 2
√ quadratic one, φ − φ − 1 = 0, with the two
roots: φ1 = 0.5(1 + 5) and φ2 = 0.5(1 − 5). Because each root solves the recurrence,
the same holds for any linear combination of them, so we know that F(n) = c1 φn1 +
c2 φn2 satisfies the recurrence. We choose the constants c1 and c2 to satisfy the base
conditions F(1) = 1 and F(2) = 1: F(1) = c1 φ1 + c2 φ2 = 1 and F(2) = c1 φ21 + c2 φ22 = 1.
Thus c1 = √15 and c2 = − √15 so that
√ !n √ !n
1 1+ 5 1 1− 5 ∼ φn
F(n) = √ −√ =√
5 2 5 2 5
√ √
= 1.618 is the well-known “golden ratio”. The term with (1 − 5)/2 ∼
where φ = 1+2 5 ∼ =
−0.618 tends to zero when n → ∞, and so F(n) is Θ(φn ).
“Telescoping” of a recurrence
This means a recursive substitution of the same implicit relationship in order to
derive the explicit relationship. Let us apply it to the same recurrence T (n) = 2T (n −
1) + 1 with the base condition T (0) = 0 as in Examples 1.26 and 1.27:
Step 0 initial recurrence T (n) = 2T (n − 1) + 1
Step 1 substitute T (n − 1) = 2T (n − 2) + 1, that is, replace T (n − 1):
T (n) = 2(2T (n − 2) + 1) + 1 = 22 T (n − 2) + 2 + 1
Step 2 substitute T (n − 2) = 2T (n − 3) + 1:

T (n) = 22 (2T (n − 3) + 1) + 2 + 1 = 23 T (n − 3) + 22 + 2 + 1

Step 3 substitute T (n − 3) = 2T (n − 4) + 1:

T (n) = 23 (2T (n − 4) + 1) + 22 + 2 + 1
= 24 T (n − 4) + 23 + 22 + 2 + 1

Step . . . . . .
Step n − 2 . . .

T (n) = 2n−1 T (1) + 2n−2 + . . . + 22 + 2 + 1

Step n − 1 substitute T (1) = 2T (0) + 1:

T (n) = 2n−1 (2T (0) + 1) + 2n−2 + . . . + 2 + 1

= 2n T (0) + 2n−1 + 2n−2 + . . . + 2 + 1

Because of the base condition T (0) = 0, the explicit formula is:

T (n) = 2n−1 + 2n−2 + . . . + 2 + 1 ≡ 2n − 1

As shown in Figure 1.5, rather than successively substitute the terms T (n − 1),
T (n − 2), . . . , T (2), T (1), it is more convenient to write down a sequence of the scaled
relationships for T (n), 2T (n − 1), 22 T (n − 2), . . . , 2n−1 T (1), respectively, then individ-
ually sum left and right columns, and eliminate similar terms in the both sums (the
terms are scaled to facilitate their direct elimination). Such solution is called tele-
scoping because the recurrence unfolds like a telescopic tube.
Although telescoping is not a powerful technique, it returns the desired explicit
forms of most of the basic recurrences that we need in this book (see Examples 1.29–
1.32 below). But it is helpless in the case of the Fibonacci recurrence because after
proper scaling of terms and reducing similar terms in the left and right sums, tele-
scoping returns just the same initial recurrence.
Example 1.29. T (n) = T (n − 1) + n; T (0) = 1.
This relation arises when a recursive algorithm loops through the input to elimi-
nate one item and is easily solved by telescoping:

T (n) = T (n − 1) + n
T (n − 1) = T (n − 2) + (n − 1)
...
T (1) = T (0) + 1
 Basic recurrence: an implicit relationship between T(n) and n; the base condition: T(0) = 0

T(n) = 2 T(n−1) + 1 T(n) = 2 T(n−1) + 1

T(n−1) = 2 T(n−2) + 1
substitution

T(n) = 4 T(n−2) + 2 + 1 2 T(n−1) = 4 T(n−2) + 2

T(n−2) = 2 T(n−3) + 1
substitution

T(n) = 8 T(n−3) + 4 + 2 + 1 4 T(n−2) = 8 T(n−3) + 4

... ... ...

T(1) = 2 T(0) + 1
substitution
n
T(n) = 2 T(0) + 2
n−1
+ ... + 4 + 2 + 1 2n−1 T(1) = 2 n T(0) + 2 n−1
left−side sum right−side sum
n
=2 −1
Explicit relationship between T(n) and n by reducing common left− and right−side terms

Figure 1.5: Telescoping as a recursive substitution.

By summing left and right columns and eliminating the similar terms, we obtain that
T (n) = T (0) + 1 + 2 + . . . + (n − 2) + (n − 1) + n = n(n+1)
2 so that T (n) is Θ(n2 ).

Example 1.30. T (n) = T (⌈n/2⌉) + 1; T (1) = 0.

This relation arises for a recursive algorithm that almost halves the input at each
step. Suppose first that n = 2m . Then, the recurrence telescopes as follows:

T (2m ) = T (2m−1 ) + 1
T (2m−1 ) = T (2m−2 ) + 1
...
1
T (2 ) = T (20 ) + 1

so that T (2m ) = m, or T (n) = lg n which is Θ(log n).

For general n, the total number of the halving steps cannot be greater than m =
⌈lg n⌉. Therefore, T (n) ≤ ⌈lg n⌉ for all n. This recurrence is usually called the repeated
halving principle.
Example 1.31. Recurrence T (n) = T (⌈n/2⌉) + n; T (0) = 0.
This relation arises for a recursive algorithm that halves the input after examining
every item in the input for n ≥ 1. Under the same simplifying assumption n = 2m , the
recurrence telescopes as follows:

T (2m ) = T (2m−1 ) + n
T (2m−1 ) = T (2m−2 ) + n/2
T (2m−2 ) = T (2m−3 ) + n/4
...
T (2) = T (1) + 2
T (1) = T (0) + 1

so that T (n) = n + 2n + 4n + . . . + 1 which is Θ(n).

In the general case, the solution is also Θ(n) because each recurrence after halv-
ing an odd-size input may add to the above sum at most 1 and the number of these
extra units is at most ⌈lg n⌉.

Example 1.32. Recurrence T (n) = T (⌈n/2⌉) + T (⌊n/2⌋) + n; T (1) = 0.

This relation arises for a recursive algorithm that makes a linear pass through the
input for n ≥ 2 and splits it into two halves. Under the same simplifying assumption
n = 2m , the recurrence telescopes as follows:

T (2m ) = 2T (2m−1 ) + 2m
m−1
T (2 ) = 2T (2m−2 ) + 2m−1
...
T (2) = 2T (1) + 2

so that
T (2m ) T (2m−1 )
= +1
2m 2m−1
T (2 m−1 ) T (2m−2 )
= +1
2m−1 2m−2
...
T (2) T (1)
= + 1.
2 1

Therefore, T (n) T (1)

n = 1 + m = lg n, so that T (n) = n lg n which is Θ(n log n).
For general n, T (n) is also Θ(n log n) (see Exercise 1.5.2).

There exist very helpful parallels between the differentiation / integration in cal-
culus and recurrence analysis by telescoping.
 • The difference equation T (n)−2T (n−1) = c rewritten as T (n)−T1 (n−1) = T (n−1)+
c resembles the differential equation dTdx(x) = T (x). Telescoping of the difference
equation results in the formula T (n) = c(2n − 1) whereas the integration of the
differential equation produces the analogous exponential one T (x) = cex .

• The difference equation T (n)−T (n−1) = cn has the differential analogue dTdx(x) =
cx, and both equations have similar solutions T (n) = c n(n+1)
2 and T (x) = 2c x2 , re-
spectively.

• Let us change variables by replacing n and x with m = lg n and y = ln x so that

n = 2m and x = ey , respectively. The difference equation T (n) − T ( n2 ) = c where
n = 2m and 2n = 2m−1 reduces to T (m) − T (m − 1) = c. The latter has the differen-
tial analogue dTdy(y) = c. These two equations result in the similar explicit expres-
sions T (m) = cm and T (y) = cy, respectively, so that T (n) = c lg n and T (x) = c ln x.

The parallels between difference and differential equations may help us in deriving
the desired closed-form solutions of complicated recurrences.

Exercise 1.5.1. Show that the solution in Example 1.31 is also in Ω(n) for general n.

Exercise 1.5.2. Show that the solution T (n) to Example 1.32 is no more than n lg n +
n − 1 for every n ≥ 1. Hint: try induction on n.

Exercise 1.5.3. The running time T (n) of

a certain algorithm to process n data items
is given by the recurrence T (n) = kT nk + cn; T (1) = 0 where k is a positive integer
constant and n/k means either ⌈n/k⌉ or ⌊n/k⌋. Derive the explicit expression for T (n)
in terms of c, n, and k assuming n = km with integer m = logk n and k and find time
complexity of this algorithm in the “Big-Oh” sense.

Exercise 1.5.4. The running time T (n) of a slightly different algorithm is given by the
recurrence T (n) = kT nk + ckn; T (1) = 0. Derive the explicit expression for T (n) in

terms of c, n, and k under the same assumption n = km and find time complexity of
this algorithm in the “Big-Oh” sense.

1.6 Capabilities and limitations of algorithm analysis

We should neither overestimate nor underestimate the capabilities of algorithm
analysis. Many existing algorithms have been analysed with much more complex
techniques than used in this book and recommended for practical use on the basis
of these studies. Of course, not all algorithms are worthy of study and we should not
suppose that a rough complexity analysis will result immediately in efficient algo-
rithms. But computational complexity permits us to better evaluate basic ideas in
developing new algorithms.
 To check whether our analysis is correct, we may code a program and see whether
its observed running time fits predictions. But it is very difficult to differentiate be-
tween, say, Θ(n) and Θ(n log n) algorithms using purely empirical evidence. Also,
“Big-Oh” analysis is not appropriate for small amounts of input and hides the con-
stants which may be crucial for a particular task.

Example 1.33. An algorithm A with running time TA = 2n lg n becomes less efficient

than another algorithm B having running time TB = 1000n only when 2n lg n > 1000n,
or lg n > 500, or n > 2500 ∼
= 10150 , and such amounts of input data simply cannot be
met in practice. Thus, although the algorithm B is better in the “Big Oh” sense, in
practice we should use algorithm A.

Large constants have to be taken into account when an algorithm is very com-
plex, or when we must discriminate between cheap or expensive access to input
data items, or when there may be lack of sufficient memory for storing large data
sets, etc. But even when constants and lower-order terms are considered, the per-
formance predicted by our analysis may differ from the empirical results. Recall that
for very large inputs, even the asymptotic analysis may break down, because some
operations (like addition of large numbers) can no longer be considered as elemen-
tary.
In order to analyse algorithm performance we have used a simplified mathemat-
ical model involving elementary operations. In the past, this allowed for fairly accu-
rate analysis of the actual running time of program implementing a given algorithm.
Unfortunately, the situation has become more complicated in recent years. Sophis-
ticated behaviour of computer hardware such as pipelining and caching means that
the time for elementary operations can vary wildly, making these models less useful
for detailed prediction. Nevertheless, the basic distinction between linear, quadratic,
cubic and exponential time is still as relevant as ever. In other words, the crude
differences captured by the Big-Oh notation give us a very good way of comparing
algorithms; comparing two linear time algorithms, for example, will require more
experimentation.
We can use worst-case and average-case analysis to obtain some meaningful es-
timates of possible algorithm performance. But we must remember that both re-
currences and asymptotic “Big-Oh”, “Big-Omega”, and “Big-Theta” notation are just
mathematical tools used to model certain aspects of algorithms. Like all models,
they are not universally valid and so the mathematical model and the real algorithm
may behave quite differently.
Exercises
Exercise 1.6.1. Algorithms A and B use TA (n) = 5n log10 n and TB (n) = 40n elementary
operations, respectively, for a problem of size n. Which algorithm has better per-
formance in the “Big Oh” sense? Work out exact conditions when each algorithm
outperforms the other.
Exercise 1.6.2. We have to choose one of two algorithms, A and B, to process a
database containing 109 records.√ The average running time of the algorithms is
TA (n) = 0.001n and TB (n) = 500 n, respectively. Which algorithm should be used,
assuming the application is such that we can tolerate the risk of an occasional long
running time?

1.7 Notes
The word algorithm relates to the surname of the great mathematician Muham-
mad ibn Musa al-Khwarizmi, whose life spanned approximately the period 780–850.
His works, translated from Arabic into Latin, for the first time exposed Europeans
to new mathematical ideas such as the Hindu positional decimal notation and step-
by-step rules for addition, subtraction, multiplication, and division of decimal num-
bers. The translation converted his surname into “Algorismus”, and the computa-
tional rules took on this name. Of course, mathematical algorithms existed well
before the term itself. For instance, Euclid’s algorithm for computing the greatest
common divisor of two positive integers was devised over 1000 years before.
The Big-Oh notation was used as long ago as 1894 by Paul Bachmann and then
Edmund Landau for use in number theory. However the other asymptotic notations
Big-Omega and Big-Theta were introduced in 1976 by Donald Knuth (at time of writ-
ing, perhaps the world’s greatest living computer scientist).
Algorithms running in Θ(n log n) time are sometimes called linearithmic, to match
“logarithmic”, “linear”, “quadratic”, etc.
The quadratic equation for φ in Example 1.28 is called the characteristic equa-
tion of the recurrence. A similar technique can be used for solving any constant
coefficient linear recurrence of the form F(n) = ∑Kk=1 ak F(n − k) where K is a fixed
positive integer and the ak are constants.
Chapter 2

Efficiency of Sorting

Sorting rearranges input data according to a particular linear order (see Section D.3
for definitions of order and ordering relations). The most common examples are the
usual dictionary (lexicographic) order on strings, and the usual order on integers.
Once data is sorted, many other problems become easier to solve. Some of these
include: finding an item, finding whether any duplicate items exist, finding the fre-
quency of each distinct item, finding order statistics such as the maximum, mini-
mum, median and quartiles. There are many other interesting applications of sort-
ing, and many different sorting algorithms, each with their own strengths and weak-
nesses. In this chapter we describe and analyse some popular sorting algorithms.

2.1 The problem of sorting

The problem of sorting is to rearrange an input list of keys, which can be com-
pared using a total order ≤, into an output list such that if i and j are keys and i
precedes j in the output list, then i ≤ j. Often the key is a data field in a larger object:
for example, we may wish to sort database records of customers, with the key being
their bank balance. If each object has a different key, then we can simply use the
keys as identifiers for the purpose of sorting: rather than moving large objects we
need only keep a pointer from the key to the object.
There are several important attributes of sorting algorithms.

Definition 2.1. A sorting algorithm is called comparison-based if the only way to

gain information about the total order is by comparing a pair of elements at a time
via the order ≤.
 A sorting algorithm is called stable if whenever two objects have the same key in
the input, they appear in the same order in the output.
A sorting algorithm is called in-place if it uses only a fixed additional amount of
working space, independent of the input size.

With a comparison-based sorting algorithm, we cannot use any information about

the keys themselves (such as the fact that they are all small integers, for example),
only their order relation. These algorithms are the most generally applicable and we
shall focus exclusively on them in this book (but see Exercise 2.7.2).
We consider only two elementary operations: a comparison of two items and
a move of an item. The running time of sorting algorithms in practice is usually
dominated by these operations. Every algorithm that we consider will make at most
a constant number of moves for each comparison, so that the asymptotic running
time in terms of elementary operations will be determined by the number of com-
parisons. However, lower order terms will depend on the exact number of moves.
Furthermore, the actual length of time taken by a data move depends on the imple-
mentation of the list. For example, moving an element from the end to the beginning
of an array takes longer than doing the same for a linked list. We shall discuss these
issues later.
The efficiency of a particular sorting algorithm may depend on many factors, for
instance:

• how many items have to be sorted;

• are the items only related by the order relation, or do they have other restric-
tions (for example, are they all integers from the range 1 to 1000);

• to what extent they are pre-sorted;

• can they be placed into an internal (fast) computer memory or must they be
sorted in external (slow) memory, such as on disk (so called external sorting ).

No one algorithm is the best for all possible situations, and so it is important to
understand the strengths and weaknesses of several algorithms.
As far as computer implementation is concerned, sorting makes sense only for
linear data structures. We will consider lists (see Section C.1 for a review of ba-
sic concepts) which have a first element (the head), a last element (the tail) and a
method of accessing the next element in constant time (an iterator). This includes
array-based lists, and singly- and doubly-linked lists. For some applications we will
need a method of accessing the previous element quickly; singly-linked lists do not
provide this. Also, array-based lists allow fast random access. The element at any
given position may be retrieved in constant time, whereas linked list structures do
not allow this.
Exercises
Exercise 2.1.1. The well-known and obvious selection sort algorithm proceeds as
follows. We split the input list into a head and tail sublist. The head (“sorted”) sublist
is initially empty, and the tail (“unsorted”) sublist is the whole list. The algorithm
successively scans through the tail sublist to find the minimum element and moves
it to the end of the head sublist. It terminates when the tail sublist becomes empty.
(Java code for an array implementation is found in Section A.1).
How many comparisons are required by selection sort in order to sort the input
list (6, 4, 2, 5, 3, 1) ?

Exercise 2.1.2. Show that selection sort uses the same number of comparisons on
every input of a fixed size. How many does it use, exactly, for an input of size n?

Exercise 2.1.3. Is selection sort comparison-based? in-place? stable?

Exercise 2.1.4. Give a linear time algorithm to find whether a sorted list contains
any duplicate elements. How would you do this if the list were not sorted?

2.2 Insertion sort

This is the method usually used by cardplayers to sort cards in their hand. Inser-
tion sort is easy to implement, stable, in-place, and works well on small lists and lists
that are close to sorted. However, it is very inefficient for large random lists.
Insertion sort is iterative and works as follows. The algorithm splits a list of size n
into a head (“sorted”) and tail (“unsorted”) sublist.

• The head sublist is initially of size 1.

• Repeat the following step until the tail sublist is empty:

◦ choose the first element x in the tail sublist;

◦ find the last element y in the head sublist not exceeding x;
◦ insert x after y in the head sublist.

Before each step i = 1, 2, . . . , n − 1, the sorted and unsorted parts have i and n −
i elements, respectively. The first element of the unsorted sublist is moved to the
correct position in the sorted sublist by exhaustive backward search, by comparing
it to each element in turn until the right place is reached.

Example 2.2. Table 2.1 shows the execution of insertion sort. Variables Ci and Mi
denote the number of comparisons and number of positions to move backward, re-
spectively, at the ith iteration. Elements in the sorted part are italicized, the currently
sorted element is underlined, and the element to sort next is boldfaced.
 Table 2.1: Sample execution of insertion sort.

i Ci Mi Data to sort
25 8 2 91 70 50 20 31 15 65
1 1 1 8 25 2 91 70 50 20 31 15 65
2 2 2 2 8 25 91 70 50 20 31 15 65
3 1 0 2 8 25 91 70 50 20 31 15 65
4 2 1 2 8 25 70 91 50 20 31 15 65
5 3 2 2 8 25 50 70 91 20 31 15 65
6 5 4 2 8 20 25 50 70 91 31 15 65
7 4 3 2 8 20 25 31 50 70 91 15 65
8 7 6 2 8 15 20 25 31 50 70 91 65
9 3 2 2 8 15 20 25 31 50 65 70 91

Analysis of insertion sort

Insertion sort is easily seen to be correct (see Exercise 2.2.2 for formal proof ),
since the head sublist is always sorted, and eventually expands to include all ele-
ments.
It is not too hard to find the worst case for insertion sort: when the input con-
sists of distinct items in reverse sorted order, every element must be compared with
every element preceding it. The number of moves is also maximized by such input.
The best case for insertion sort is when the input is already sorted, when only n − 1
comparisons are needed.

Lemma 2.3. The worst-case time complexity of insertion sort is Θ(n2 ).

Proof. Fill in the details yourself — see Exercise 2.2.3.

Since the best case is so much better than the worst, we might hope that on aver-
age, for random input, insertion sort would perform well. Unfortunately, this is not
true.

Lemma 2.4. The average-case time complexity of insertion sort is Θ(n2 ).

Proof. We first calculate the average number Ci of comparisons at the ith step. At the
beginning of this step, i elements of the head sublist are already sorted and the next
element has to be inserted into the sorted part. This element will move backward j
steps, for some j with 0 ≤ j ≤ i. If 0 ≤ j ≤ i − 1, the number of comparisons used will
be j + 1. But if j = i (it ends up at the head of the list), there will be only i comparisons
(since no final comparison is needed).
 Assuming all possible inputs are equally likely, the value of j will be equally likely
to take any value 0, . . . , i. Thus the expected number of comparisons will be
 
1 1 i(i + 1) i i
Ci = (1 + 2 + · · · + i − 1 + i + i) = +i = + .
i+1 i+1 2 2 i+1

(see Section D.6 for the simplification of the sum, if necessary).

The above procedure is performed for i = 1, 2, . . . , n − 1, so that the average total
number C of comparisons is as follows:
n−1 n−1 
1 n−1 n−1 i

i i
C= ∑ Ci = ∑ +
2 i+1
= ∑ i+ ∑ i+1
2 i=1
i=1 i=1 i=1

The first sum is equal to (n−1)n i 1

2 . To find the second sum, let us rewrite i+1 as 1 − i+1
so that
n−1 n−1   n−1 n
i 1 1 1
∑ i+1 ∑ = 1 −
i + 1
= n−1− ∑
i + 1
= n − ∑ = n − Hn
i=1 i=1 i=1 i=1 i

where Hn denotes the n-th harmonic number: Hn ≈ ln n when n → ∞.

Therefore, C = (n−1)n
4 + n − Hn . Now the total number of data moves is at least
zero and at most the number of comparisons. Thus the total number of elementary
operations is Θ(n2 ).

The running time of insertion sort is strongly related to inversions. The number
of inversions of a list is one measure of how far it is from being sorted.

Definition 2.5. An inversion in a list a is an ordered pair of positions (i, j) such that
i < j but a[i] > a[ j].

Example 2.6. The list (3, 2, 5) has only one inversion corresponding to the pair (3, 2),
the list (5, 2, 3) has two inversions, namely, (5, 2) and (5, 3), the list (3, 2, 5, 1) has four
inversions (3, 2), (3, 1), (2, 1), and (5, 1), and so on.

Example 2.7. Table 2.2 shows the number of inversions, Ii , for each element a[i] of
the list in Table 2.1 with respect to all preceding elements a[0], . . . , a[i − 1] (Ci and Mi
are the same as in Table 2.1).

Note that Ii = Mi in Table 2.1. This is not merely a coincidence—it is always true.
See Exercise 2.2.4.
n−1
The total number of inversions I = ∑i=1 Ii in a list to be sorted by insertion sort
is equal to the total number of positions an element moves backward during the
n−1
sort. The total number of data comparisons C = ∑i=1 Ci is also equal to the total
number of inversions plus at most n − 1. For the initial list in Tables 2.1 and 2.2,
Table 2.2: Number of inversions Ii , comparisons Ci and data moves Mi for each element a[i] in
sample list.

Index i 0 1 2 3 4 5 6 7 8 9
List element a[i] 25 8 2 91 70 50 20 31 15 65
Ii 1 2 0 1 2 4 3 6 2
Ci 1 2 1 2 3 5 4 7 3
Mi 1 2 0 1 2 4 3 6 2

I = 21, and insertion sort performs C = 28 comparisons and M = 21 data moves: in

total, 49 elementary operations.
Swapping two neighbours that are out of order removes exactly one inversion,
and a sorted list has no inversions. If an original list has I inversions, insertion sort
has to swap I pairs of neighbours. Because of Θ(n) other operations in the algorithm,
its running time is Θ(n + I). Thus, on nearly sorted lists for which I is Θ(n), insertion
sort runs in linear time. Unfortunately this type of list does not occur often, if we
choose one randomly. As we have seen above, the average number of inversions
for a randomly chosen list must be in Θ(n2 ). This shows that more efficient sorting
algorithms must eliminate more than just one inversion between neighbours per
swap. One way to do this is a generalization of insertion sort called Shellsort (see
Exercise 2.2.7).
Implementation of insertion sort
The number of comparisons does not depend on how the list is implemented,
but the number of moves does. The insertion operation can be implemented in a
linked list in constant time, but in an array there is no option but to shift elements to
the right when inserting an element, taking linear time in the worst and average case.
Thus if using an array implementation of a list, we may as well move the element
backward by successive swaps. If using a linked list, we can make fewer swaps by
simply scanning backward. On the other hand, scanning backward is easy in an
array but takes more time in a singly linked list. However, none of these issues affect
the asymptotic Big-Theta running time of the algorithm, just the hidden constants
and lower order terms. The main problem with insertion sort is that it takes too
many comparisons in the worst case, no matter how cleverly we implement it.
Figure 2.1 shows basic pseudocode for arrays.
Exercises
Exercise 2.2.1. Determine the quantities Ci and Mi when insertion sort is run on the
input list (91, 70, 65, 50, 31, 25, 20, 15, 8, 2).

Exercise 2.2.2. Prove by induction that algorithm insertionSort is correct.

 algorithm insertionSort
Input: array a[0..n − 1]
begin
for i ← 1 to n − 1 do
k ← i−1
while k ≥ 0 and a[k] > a[k + 1] do
swap(a, k, k + 1)
k ← k−1
end while
end for
end

Figure 2.1: Insertion sort for arrays.

Exercise 2.2.3. Prove that the worst-case time complexity of insertion sort is Θ(n2 )
and the best case is Θ(n).

Exercise 2.2.4. Prove that the number of inversions, Ii , of an element a[i] with respect
to the preceding elements, a[0], . . . , a[i − 1], in the initial list is equal to the number of
positions moved backward by a[i] in the execution of insertion sort.

Exercise 2.2.5. Suppose a sorting algorithm swaps elements a[i] and a[i + gap] of a
list a which were originally out of order. Prove that the number of inversions in the
list is reduced by at least 1 and at most 2 gap − 1.

Exercise 2.2.6. Bubble sort works as follows to sort an array. There is a sorted left
subarray and unsorted right subarray; the left subarray is initially empty. At each
iteration we step through the right subarray, comparing each pair of neighbours in
turn, and swapping them if they are out of order. At the end of each such pass, the
sorted subarray has increased in size by 1, and we repeat the entire procedure from
the beginning of the unsorted subarray. (Java code is found in Section A.1.)
Prove that the average time complexity of bubble sort is Θ(n2 ), and that bubble
sort never makes fewer comparisons than insertion sort.

Exercise 2.2.7. Shellsort is a generalization of insertion sort that works as follows.

We first choose an increment sequence . . . ht > ht−1 > . . . > h1 = 1. We start with some
value of t so that ht < n. At each step we form the sublists of the input list a consisting
of elements h := ht apart (for example, the first such list has the elements at position
0, h, 2h, . . . , the next has the elements 1, 1 + h, 1 + 2h, . . . , etc). We sort each of these h
lists using insertion sort (we call this the h-sorting step). We then reduce t by 1 and
continue. Note that the last step is always a simple insertion sort.
Explain why Shellsort is not necessarily slower than insertion sort. Give an input
on which Shellsort uses fewer comparions overall than insertion sort.
Exercise 2.2.8. Find the total numbers of comparisons and backward moves per-
formed by Shellsort on the input list (91, 70, 65, 50, 31, 25, 20, 15, 8, 2) and compare the
total number of operations with that for insertion sort in Exercise 2.2.1.

2.3 Mergesort
This algorithm exploits a recursive divide-and-conquer approach resulting in a
worst-case running time of Θ(n log n), the best asymptotic behaviour that we have
seen so far. Its best, worst, and average cases are very similar, making it a very good
choice if predictable runtime is important. Versions of mergesort are particularly
good for sorting data with slow access times, such as data that cannot be held in
internal memory or are stored in linked lists.
Mergesort is based on the following basic idea.

• If the size of the list is 0 or 1, return.

• Otherwise, separate the list into two lists of equal or nearly equal size and re-
cursively sort the first and second halves separately.

• Finally, merge the two sorted halves into one sorted list.

Clearly, almost all the work is in the merge step, which we should make as effi-
cient as possible. Obviously any merge must take at least time that is linear in the
total size of the two lists in the worst case, since every element must be looked at in
order to determine the correct ordering. We can in fact achieve a linear time merge,
as we see in the next section.
Analysis of mergesort
Lemma 2.8. Mergesort is correct.

Proof. We use induction on the size n of the list. If n = 0 or 1, the result is obviously
correct. Otherwise, mergesort calls itself recursively on two sublists each of which
has size less than n. By induction, these lists are correctly sorted. Provided that the
merge step is correct, the top level call of mergesort then returns the correct answer.

Almost all the work occurs in the merge steps, so we need to perform those effi-
ciently.

Theorem 2.9. Two input sorted lists A and B of size nA and nB , respectively, can be
merged into an output sorted list C of size nC = nA + nB in linear time.

Proof. We first show that the number of comparisons needed is linear in n. Let i,
j, and k be pointers to current positions in the lists A, B, and C, respectively. Ini-
tially, the pointers are at the first positions, i = 0, j = 0, and k = 0. Each time the
smaller of the two elements A[i] and B[ j] is copied to the current entry C[k], and the
corresponding pointers k and either i or j are incremented by 1. After one of the
input lists is exhausted, the rest of the other list is directly copied to list C. Each
comparison advances the pointer k so that the maximum number of comparisons is
nA + nB − 1.
All other operations also take linear time.

The above proof can be visualized easily if we think of the lists as piles of playing
cards placed face up. At each step, we choose the smaller of the two top cards and
move it to the temporary pile.

Example 2.10. If a = (2, 8, 25, 70, 91) and b = (15, 20, 31, 50, 65), then merge into c =
(2, 8, 15, 20, 25, 31, 50, 65, 70, 91) as follows.

Step 1 a[0] = 2 and b[0] = 15 are compared, 2 < 15, and 2 is copied to c, that is, c[0] ← 2,
i ← 0 + 1, and k ← 0 + 1.

Step 2 a[1] = 8 and b[0] = 15 are compared to copy 8 to c, that is, c[1] ← 8, i ← 1 + 1,
and k ← 1 + 1.

Step 3 a[2] = 25 and b[0] = 15 are compared and 15 is copied to c so that c[2] ← 15,
j ← 0 + 1, and k ← 2 + 1.

Step 4 a[2] = 25 and b[1] = 20 are compared and 20 is copied to c: c[3] ← 20, j ← 1 + 1,
and k ← 3 + 1.

Step 5 a[2] = 25 and b[2] = 31 are compared, and 25 is copied to c: c[4] ← 25, i ← 2 + 1,
and k ← 4 + 1.

The process continues as follows: comparing a[3] = 70 and b[2] = 31, a[3] = 70 and
b[3] = 50, and a[3] = 70 and b[4] = 65 results in c[5] ← (b[2] = 31), c[6] ← (b[3] = 50),
and c[7] ← (b[4] = 65), respectively. Because the list b is exhausted, the rest of the list
a is then copied to c, c[8] ← (a[3] = 70) and c[9] ← (a[4] = 91).

We can now see that the running time of mergesort is much better asymptotically
than the naive algorithms that we have previously seen.

Theorem 2.11. The running time of mergesort on an input list of size n is Θ(n log n)
in the best, worst, and average case.

Proof. The number of comparisons used by mergesort on an input of size n satisfies

a recurrence of the form T (n) = T (⌈n/2⌉) + T (⌊n/2⌋) + a(n) where 1 ≤ a(n) ≤ n − 1. It
is straightforward to show as in Example 1.32 that T (n) is Θ(n log n).
The other elementary operations in the divide and combine steps depend on the
implementation of the list, but in each case their number is Θ(n). Thus these opera-
tions satisfy a similar recurrence and do not affect the Θ(n log n) answer.
Implementation of mergesort
It is easier to implement the recursive version above for arrays than for linked
lists, since splitting an array in the middle is a constant time operation. Algorithm
merge in Figure 2.3 follows the above description. The first half of the input array, a,
from the leftmost index l to the middle index m acts as A, the second half from m + 1
to the rightmost index r as B, and a separate temporary array t as C. After merging
the halves, the temporary array is copied back to the original one, a.

algorithm mergeSort
Input: array a[0..n − 1]; array indices l, r; array t[0..n − 1]
sorts the subarray a[l..r]
begin
if l < r then
m ← l+r
 
2
mergeSort(a, l, m,t)
mergeSort(a, m + 1, r,t)
merge(a, l, m + 1, r,t)
end if
end

Figure 2.2: Recursive mergesort for arrays

It is easy to see that the recursive version simply divides the list until it reaches
lists of size 1, then merges these repeatedly. We can eliminate the recursion in a
straightforward manner. We first merge lists of size 1 into lists of size 2, then lists of
size 2 into lists of size 4, and so on. This is often called straight mergesort .

Example 2.12. Starting with the input list (1, 5, 7, 3, 6, 4, 2) we merge repeatedly. The
merged sublists are shown with parentheses.

Step 0: (1)(5)(7)(3)(6)(4)(2) Step 2: (1,3,5,7)(2,4,6)

Step 1: (1,5)(3,7)(4, 6)(2) Step 3: (1,2,3,4,5,6,7)

This method works particularly well for linked lists, because the merge steps can
be implemented simply by redefining pointers, without using the extra space re-
quired when using arrays (see Exercise 2.3.4).
Exercises
Exercise 2.3.1. What is the minimum number of comparisons needed when merg-
ing two nonempty sorted lists of total size n into a single list?

Exercise 2.3.2. Give two sorted lists of size 8 whose merging requires the maximum
number of comparisons.
 algorithm merge
Input: array a[0..n − 1]; array indices l, r; array index s; array t[0..n − 1]
merges the two sorted subarrays a[l..s − 1] and a[s..r] into a[l..r]
begin
i ← l; j ← s; k ← l
while i ≤ s − 1 and j ≤ r do
if a[i] ≤ a[ j] then t[k] ← a[i]; k ← k + 1; i ← i + 1
else t[k] ← a[ j]; k ← k + 1; j ← j + 1
end if
end while
while i ≤ s − 1 do copy the rest of the first half
t[k] ← a[i]; k ← k + 1; i ← i + 1
end while
while j ≤ r do copy the rest of the second half
t[k] ← a[ j]; k ← k + 1; j ← j + 1
end while
return a ← t
end

Figure 2.3: Linear time merge for arrays

Exercise 2.3.3. The 2-way merge in this section can be generalized easily to a k-
way merge for any positive integer k. The running time of such a merge is c(k − 1)n.
Assuming that the running time of insertion sort is cn2 with the same scaling factor
c, analyse the asymptotic running time of the following sorting algorithm (you may
assume that n is an exact power of k).

• Split an initial list of size n into k sublists of size nk each.

• Sort each sublist separately by insertion sort.
• Merge the sorted sublists into a final sorted list.

Find the optimum value of k to get the fastest sort and compare its worst/average
case asymptotic running time with that of insertion sort and mergesort.

Exercise 2.3.4. Explain how to merge two sorted linked lists in linear time into a
bigger sorted linked list, using only a constant amount of extra space.

2.4 Quicksort
This algorithm is also based on the divide-and-conquer paradigm. Unlike merge-
sort, quicksort dynamically forms subarrays depending on the input, rather than
sorting and merging predetermined subarrays. Almost all the work of mergesort was
in the combining of solutions to subproblems, whereas with quicksort, almost all the
work is in the division into subproblems.
Quicksort is very fast in practice on “random” data and is widely used in software
libraries. Unfortunately it is not suitable for mission-critical applications, because it
has very bad worst case behaviour, and that behaviour can sometimes be triggered
more often than an analysis based on random input would suggest.
Basic quicksort is recursive and consists of the following four steps.

• If the size of the list is 0 or 1, return the list. Otherwise:

• Choose one of the items in the list as a pivot .

• Next, partition the remaining items into two disjoint sublists: reorder the list
by placing all items greater than the pivot to follow it, and all elements less than
the pivot to precede it.

• Finally, return the result of quicksort of the “head” sublist, followed by the
pivot, followed by the result of quicksort of the “tail” sublist.

The first step takes into account that recursive dynamic partitioning may pro-
duce empty or single-item sublists. The choice of a pivot at the next step is most
critical because the wrong choice may lead to quadratic time complexity while a
good choice of pivot equalizes both sublists in size (and leads to “n log n” time com-
plexity). Note that we must specify in any implementation what to do with items
equal to the pivot. The third step is where the main work of the algorithm is done,
and obviously we need to specify exactly how to achieve the partitioning step (we do
this below). The final step involves two recursive calls to the same algorithm, with
smaller input.
Analysis of quicksort
All analysis depends on assumptions about the pivot selection and partitioning
methods used. In particular, in order to partition a list about a pivot element as
described above, we must compare each element of the list to the pivot, so at least
n − 1 comparisons are required. This is the right order: it turns out that there are
several methods for partitioning that use Θ(n) comparisons (we shall see some of
them below).

Lemma 2.13. Quicksort is correct.

Proof. We use mathematical induction on the size of the list. If the size is 1, the al-
gorithm is clearly correct. Suppose then that n ≥ 2 and the algorithm works correctly
on lists of size smaller than n. Suppose that a is a list of size n, p is the pivot ele-
ment, and i is the position of p after partitioning. Due to the partitioning principle
of quicksort, all elements of the head sublist are no greater than p, and all elements
of the tail sublist are no smaller than p. By the induction hypothesis, the left and
right sublists are sorted correctly, so that the whole list a is sorted correctly.

Unlike mergesort, quicksort does not perform well in the worst case.

Lemma 2.14. The worst-case time complexity of quicksort is Ω(n2 ).

Proof. The worst case for the number of comparisons is when the pivot is always the
smallest or the largest element, one of two sublists is empty, and the second sublist
contains all the elements other than the pivot. Then quicksort is recursively called
only on this second group. We show that a quadratic number of comparisons are
needed; considering data moves or swaps only increases the running time.
Let T (n) denote the worst-case running time for sorting a list containing n ele-
ments. The partitioning step needs (at least) n − 1 comparisons. At each next step
for n ≥ 1, the number of comparisons is one less, so that T (n) satisfies the easy recur-
rence, T (n) = T (n − 1) + n − 1; T (0) = 0, similar to the basic one, T (n) = T (n − 1) + n,
in Example 1.29. This yields that T (n) is Ω(n2 ).

However, quicksort is often used in practice, because it is fast on “random” input.

The proof of this involves a more complicated recurrence than we have seen so far.

Lemma 2.15. The average-case time complexity of quicksort is Θ(n log n).

Proof. Let T (n) denote the average-case running time for sorting a list containing n
elements. In the first step, the time taken to compare all the elements with the pivot
is linear, cn. If i is the final position of the pivot, then two sublists of size i and n− 1− i,
respectively, are quicksorted, and in this particular case T (n) = T (i) + T (n − 1 − i) + cn.
Therefore, the average running time to sort n elements is equal to the partitioning
time plus the average time to sort i and n − 1 − i elements, where i varies from 0 to
n − 1.
All the positions i may be met equiprobably—the pivot picked with the same
chance 1n —as the final pivot position in a sorted array. Hence, the average time of
each recursive call is equal to the average, over all possible subset sizes, of the aver-
age running time of recursive calls on sorting both subsets:

1 n−1 2 n−1
T (n) = ∑
n i=0
(T (i) + T (n − 1 − i)) + cn = ∑ T (i) + cn
n i=0

To solve this more difficult recurrence, let us rewrite it as

n−1
nT (n) = 2 ∑ T (i) + cn2
i=0
and subtract the analogous one with n replaced by n − 1:
n−2
(n − 1)T (n − 1) = 2 ∑ T (i) + c(n − 1)2 .
i=0

After rearranging similar terms, we obtain nT (n) = (n + 1)T (n − 1) + c(2n − 1) Decom-

posing the right side via partial fractions we obtain

T (n) T (n − 1) 3c c
= + − .
n+1 n n+1 n
Telescoping of this recurrence results in the following relationship.
   
T (n) T (0) 1 1 1 1 1 1 1 1 1
= + 3c + + +...+ −c + + + +...+
n+1 1 2 3 4 n+1 1 2 3 4 n
= c(3Hn+1 − 3 − Hn )

where Hn is the n-th harmonic number. Thus (see Section D.6) T (n) is Θ(n log n).

This is the first example we have seen of an algorithm whose worst-case perfor-
mance and average-case performance differ dramatically.
The finer details of the performance of quicksort depend on several implemen-
tation issues, which we discuss next.
Implementation of quicksort
There are many variants of the basic quicksort algorithm, and it makes sense to
speak of “a quicksort”. There are several choices to be made:

• how to implement the list;

• how to choose the pivot;

• how to partition around the pivot.

C HOOSING A PIVOT A passive pivot strategy of choosing a fixed (for example the
first, last, or middle) position in each sublist as the pivot seems reasonable under
the assumption that all inputs are equiprobable. The simplest version of quicksort
just chooses the first element of the list. We call this the naive pivot selection rule.
For such a choice, the likelihood of a random input resulting in quadratic running
time (see Lemma 2.14 above) is very small. However, such a simple strategy is a bad
idea. There are two main reasons:

• (nearly) sorted lists occur in practice rather frequently (think of a huge database
with a relatively small number of updates daily);
 • a malicious adversary may exploit the pivot choice method by deliberately
feeding the algorithm an input designed to cause worst case behaviour (a so-
called “algorithmic complexity attack”).

If the input list is already sorted or reverse sorted, quadratic running time is ob-
tained when actually quicksort “should” do almost no work. We should look for a
better method of pivot selection.
A more reasonable choice for the pivot is the middle element of each sublist. In
this case an already sorted input list produces the perfect pivot at each recursion.
Of course, it is still possible to construct input sequences that result in quadratic
running time for this strategy. They are very unlikely to occur at random, but this
still leaves the door open for a malicious adversary.
As an alternative to passive strategies, an active pivot strategy makes good use of
the input data to choose the pivot. The best active pivot is the exact median of the list
because it partitions the list into (almost) equal sized sublists. But it turns out that
the median cannot be computed quickly enough, and such a choice considerably
slows quicksort down rather than improving it. Thus, we need a reasonably good yet
computationally efficient estimate of the median. (See Section 2.6 for more on the
problem of exact computation of the median of a list.)
A reasonable approximation to the true median is obtained by choosing the me-
dian of the first, middle, and last elements as the pivot (the so-called median-of-
three method). Note that this strategy is also the best for an already-sorted list be-
cause the median of each subarray is recursively used as the pivot.

Example 2.16. Consider the input list of integers (25, 8, 2, 91, 70, 50, 20, 31, 15, 65). Us-
ing the median-of-three rule (with the middle of a size n list defined as the element
at position ⌊n/2⌋), the first pivot chosen is the median of 25, 70 and 65, namely 65.
Thus the left and right sublists after partitioning have sizes 7 and 2 respectively.
The standard choice for pivot would be the left element, namely 25. In this case
the left and right sublists after partitioning have sizes 4 and 5 respectively.

The median-of-three strategy does not completely avoid bad performance, but
the chances of such a case occurring are much less than for a passive strategy.
Finally, another simple method is to choose the pivot randomly. We can show
(by the same calculation as for the average-case running time) that the expected
running time on any given input is Θ(n log n). It is still possible to encounter bad
cases, but these now occur by bad luck, independent of the input. This makes it
impossible for an adversary to force worst-case behaviour by choosing a nasty input
in advance. Of course, in practice an extra overhead is incurred because of the work
needed to generate a “random” number.
A similar idea is to first randomly shuffle the input (which can be done in linear
time), and then use the naive pivot selection method. Again, bad cases still occur,
but by bad luck only.
PARTITIONING There are several ways to partition with respect to the pivot element
in linear time. We present one of them here.
Our partitioning method uses a pointer L starting at the head of the list and an-
other pointer R starting at the end plus one. We first swap the pivot element to the
head of the list. Then, while L < R, we loop through the following procedure:

• Decrement R until it meets an element less than or equal to p.

• Increment L until it meets an element greater than or equal to p.
• Swap the elements pointed to by L and R.

Finally, once L = R, we swap the pivot element with the element pointed to by L.
Example 2.17. Table 2.3 exemplifies the partitioning of a 10-element list. We choose
the pivot p = 31. The bold element is the pivot; elements in italics are those pointed
to by the pointers L and R.

Table 2.3: Partitioning in quicksort with pivot p = 31.

Data to sort Description

25 8 2 91 15 50 20 31 70 65 Initial list
31 8 2 91 15 50 20 25 70 65 Move pivot to head
31 8 2 91 15 50 20 25 70 65 stop R
31 8 2 91 15 50 20 25 70 65 stop L
31 8 2 25 15 50 20 91 70 65 swap elements L and R
31 8 2 25 15 50 20 91 70 65 stop R
31 8 2 25 15 50 20 91 70 65 stop L
31 8 2 25 15 20 50 91 70 65 swap elements L and R
31 8 2 25 15 20 50 91 70 65 stop R
20 8 2 25 15 31 50 91 70 65 swap element L with pivot

Lemma 2.18. The partitioning algorithm described above is correct.

Proof. After each swap, we have the following properties:

• each element to the left of L (including the element at L) is less than or equal
to p;
• each element to the right of R (including the element at R) is greater than or
equal to p.
In the final swap, we swap the pivot element p with an element that is no more than
p. Thus all elements smaller than p are now to its left, and all larger are to its right.
 We finish this section by discussing details associated with implementation of
lists. Quicksort is much easier to program for arrays than for other types of lists. The
pivot selection methods discussed above can be implemented in constant time for
arrays, but not for linked lists (for example, try the median-of-three rules with linked
lists). The partition step can be performed on linked lists, but the method we have
presented requires a doubly-linked list in order to work well, since we need to scan
both forward and backward.
In Figure 2.4 we present pseudocode for array-based quicksort. Here i denotes
the position of the pivot p before the partition step, and j its position after partition-
ing. The algorithm works in-place using the input array.

algorithm quickSort
Input: array a[0..n − 1]; array indices l, r
sorts the subarray a[l..r]
begin
if l < r then
i ← pivot(a, l, r) return position of pivot element
j ← partition(a, l, r, i) return final position of pivot
quickSort(a, l, j − 1) sort left subarray
quickSort(a, j + 1, r) sort right subarray
end if
return a
end

Figure 2.4: Basic array-based quicksort.

Exercises
Exercise 2.4.1. Analyse in the same way as in Example 2.17 the next partitioning of
the left sublist (20, 8, 2, 25, 15) obtained in Table 2.3.
Exercise 2.4.2. Show in detail what happens in the partitioning step if all keys are
equal. What would happen if we change the partition subroutine so that L skipped
over keys equal to the pivot? What if R did? What if both did? (hint: the algorithm
would still be correct, but its performance would differ).
Exercise 2.4.3. Analyse partitioning of the list (25, 8, 8, 8, 8, 8, 2) if the pointers L and
R are advanced while L ≤ p and p ≤ R, respectively, rather than stopping on equality
as in Table 2.3.
Exercise 2.4.4. Suppose that we implement quicksort with the naive pivot choice
rule and the partitioning method of the text. Find an array of size 8, containing each
integer 1, . . . 8 exactly once, that makes quicksort do the maximum number of com-
parisons. Find one that makes it do the minimum number of comparisons.

2.5 Heapsort
This algorithm is an improvement over selection sort that uses a more sophisti-
cated data structure to allow faster selection of the minimum element. It, like merge-
sort, achieves Θ(n log n) in the worst case. Nothing better can be achieved by an al-
gorithm based on pairwise comparisons, as we shall see later, in Section 2.7.
A heap is a special type of tree. See Section D.7 if necessary for general facts about
trees.

Definition 2.19. A complete binary tree is a binary tree which is completely filled at
all levels except, possibly, the bottom level, which is filled from left to right with no
missing nodes.

In such a tree, each leaf is of depth h or h − 1, where h is the tree height, and each
leaf of height h is on the left of each leaf of height h − 1.

Example 2.20. Figure 2.5 demonstrates a complete binary tree with ten nodes. If the
node J had been the right child of the node E, the tree would have not been complete
because of the left child node missed at the bottom level.

A 1

2 C 3
B

4 5 6 7
D E F G

8 9 10
H I J
positions: 1 2 3 4 5 6 7 8 9 10

A B C D E F G H I J

indices: 0 1 2 3 4 5 6 7 8 9

Figure 2.5: Complete binary tree and its array representation.

 A level-order traversal of a complete binary tree is easily stored in a linear array
as shown in Figure 2.5. In this book we mostly use indices 0 to n − 1 for the array
positions 1 to n in order to match popular programming languages such as Java,
C, or C++. But the array representation of a binary tree is described more conve-
niently when both the node numbers and the array positions are changing from 1 to
n. Therefore, when a position p in an array a is mentioned below, we bear in mind an
array element a[i] with the index i = p − 1.
The node in position p has its parent node, left child, and right child in positions
⌊p/2⌋, 2p, and 2p + 1, respectively. The leaves have no children so that for each leaf
node q, the child position 2q exceeds the number of nodes n in the tree: 2q > n.

Example 2.21. In Figure 2.5, the node in position p = 1 is the root with no parent
node. The nodes in positions from 6 to 10 are the leaves. The root has its left child in
position 2 and its right child in position 3. The nodes in positions 2 and 3 have their
left child in position 4 and 6 and their right child in position 5 and 7, respectively.
The node in position 4 has a left child in position 8 and a right child in position 9,
and the node 5 has only a left child, in position 10.

Definition 2.22. A (maximum) heap is a complete binary tree having a key associ-
ated with each node, the key of each parent node being greater than or equal to the
keys of its child nodes.

The heap order provides easy access to the maximum key associated with the
root.

Example 2.23. Figure 2.6 illustrates a maximum heap. Of course, we could just as
easily have a minimum heap where the key of the parent node is less than or equal
to the keys of its child nodes. Then the minimum key is associated with the root.

The heapsort algorithm now works as follows. Given an input list, build a heap
by successively inserting the elements. Then delete the maximum repeatedly (ar-
ranging the elements in the output list in reverse order of deletion) until the heap is
empty. Clearly, this is a variant of selection sort that uses a different data structure.
Analysis of heapsort
Heapsort is clearly correct for the same reason as selection sort. To analyse its
performance, we need to analyse the running time of the insertion and deletion op-
erations.

Lemma 2.24. The height of a complete binary tree with n nodes is at most ⌊lg n⌋.

Proof. Depending on the number of nodes at the bottom level, a complete tree of
height h contains between 2h and 2h+1 − 1 nodes, so that 2h ≤ n < 2h+1 , or h ≤ lg n <
h + 1.
 91 1

3
65 2 70

4 5 6 7
31 8 50 25

8 9 10
20 15 2

1 2 3 4 5 6 7 8 9 10

91 65 70 31 8 50 25 20 15 2

Figure 2.6: Maximum heap and its array representation.

Lemma 2.25. Insertion of a new node into a heap takes logarithmic time.

Proof. To add one more node to a heap of n elements, a new, (n + 1)-st, leaf position
has to be created. The new node with its associated key is placed first in this leaf. If
the inserted key preserves the heap order, the insertion is completed. Otherwise, the
new key has to swap with its parent, and this process of bubbling up, (or percolating
up) the key is repeated toward the root until the heap order is restored. Therefore,
there are at most h swaps where h is the heap height, so that the running time is
O(log n).

Example 2.26. To insert an 11th element, 75, into the heap in Figure 2.6 takes three
steps.

• Position 11 to initially place the new key, 75, is created.

• The new key is swapped with its parent key, 8, in position 5 = ⌊11/2⌋ to restore
the heap order.

• The same type of swap is repeated for the parent key, 65, in position 2 = ⌊5/2⌋.
Because the heap order condition is now satisfied, the process terminates.

This is shown in Table 2.4. The elements moved to restore the heap order are
italicized.
 Table 2.4: Inserting a new node with the key 75 in the heap in Figure 2.6.

Position 1 2 3 4 5 6 7 8 9 10 11
Index 0 1 2 3 4 5 6 7 8 9 10
Array at step 1 91 65 70 31 8 50 25 20 15 2 75
Array at step 2 91 65 70 31 75 50 25 20 15 2 8
Array at step 3 91 75 70 31 65 50 25 20 15 2 8

Lemma 2.27. Deletion of the maximum key from a heap takes logarithmic time in
the worst case.
Proof. The maximum key occupies the tree’s root, that is, position 1 of the array. The
deletion reduces the heap size by 1 so that its last leaf node has to be eliminated.
The key associated with this leaf replaces the deleted key in the root and then is
percolated down the tree. First, the new root key is compared to each child and
swapped with the larger child if at least one child is greater than the parent. This
process is repeated until the order is restored. Therefore, there are h moves in the
worst case where h is the heap height, and the running time is O(log n).
Because we percolate down the previous leaf key, the process usually terminates
at or near the leaves.
Example 2.28. To delete the maximum key, 91, from the heap in Figure 2.6, takes
three steps, as follows.

• Key 2 from the eliminated position 10 is placed at the root.

• The new root key is compared to its children 65 and 70 in positions 2 = 2 · 1 and
3 = 2 · 1 + 1, respectively. To restore the heap order, it is swapped with its larger
child, 70.
• The same swap is repeated for the children 50 and 25 in positions 6 = 2 · 3 and
7 = 2 · 3 + 1. Because the heap order is now correct, the process terminates.

Table 2.5: Deletion of the maximum key from the heap in Figure 2.6.

Position 1 2 3 4 5 6 7 8 9
Index 0 1 2 3 4 5 6 7 8
Array at step 1 2 65 70 31 8 50 25 20 15
Array at step 2 70 65 2 31 8 50 25 20 15
Array at step 3 70 65 50 31 8 2 25 20 15

See also Table 2.5. The leaf key replacing the root key is boldfaced, and the moves
to restore the heap are italicized.
Lemma 2.29. Heapsort runs in time in Θ(n log n) in the best, worst, and average case.

Proof. The heap can be constructed in time O(n log n) (in fact it can be done more
quickly as seen in Lemma 2.31 but this does not affect the result). Heapsort then re-
peats n times the deletion of the maximum key and restoration of the heap property.
In the best, worst, and average case, each restoration is logarithmic, so the total time
is log(n) + log(n − 1) + ... + log(1) = log n! which is Θ(n log n).

Implementation of heapsort
There are several improvements that can be made to the basic idea above. First,
the heap construction phase can be simplified. There is no need to maintain the
heap property as we add each element, since we only require the heap property once
the heap is fully built. A nice recursive approach is shown below. Second, we can
eliminate the recursion. Third, everything can be done in-place starting with an
input array.
We consider each of these in turn.
A heap can be considered as a recursive structure of the form left subheap ← root
→ right subheap, built by a recursive “heapifying” process. The latter assumes that
the heap order exists everywhere except at the root and percolates the root down
to restore the total heap order. Then it is recursively applied to the left and right
subheaps.

Lemma 2.30. A complete binary tree satisfies the heap property if and only if the
maximum key is at the root, and the left and right subtrees of the root also satisfy the
heap property with respect to the same total order.

Proof. Suppose that T is a complete binary tree that satisfies the heap condition.
Then the maximum key is at the root. The left and right subtrees at the root are also
complete binary trees, and they inherit the heap property from T .
Conversely, suppose that T is a complete binary tree with the maximum at the
root and such that the left and right subtrees are themselves heaps. Then the value
at the root is at least as great as that of the keys of the children of the root. For each
other node of T , the same property holds by our hypotheses. Thus the heap property
holds for all nodes in the tree.

For example, in Figure 2.6 we can see this fact clearly.

Lemma 2.31. A heap can be constructed from a list of size n in Θ(n) time.

Proof. Let T (h) denote the worst-case time to build a heap of height at most h. To
construct the heap, each of the two subtrees attached to the root are first trans-
formed into heaps of height at most h − 1 (the left subtree is always of height h − 1,
whereas the right subtree could be of lesser height, h − 2). Then in the worst case
the root percolates down the tree for a distance of at most h steps that takes time
O(h). Thus heap construction is asymptotically described by the recurrence similar
to Example 1.27, T (h) = 2T (h − 1) + ch and so T (h) is O(2h ). Because a heap of size n
is of height h = ⌊lg n⌋, we have 2h ≤ n and thus T (h) is O(n). But since every element
of the input must be inspected, we clearly have a lower bound of Ω(n), which yields
the result.

Now we observe that the recursion above can be eliminated. The key at each po-
sition p percolates down only after all its descendants have been already processed
by the same percolate-down procedure. Therefore, if this procedure is applied to the
nodes in reverse level order, the recursion becomes unnecessary. In this case, when
the node p has to be processed, all its descendants have been already processed.
Because leaves need not percolate down, a non-recursive heapifying process by per-
colating nodes down can start at the non-leaf node with the highest number. This
leads to an extremely simple algorithm for converting an array into a heap (see the
first for loop in Figure 2.7).
Figure 2.7 presents the basic pseudocode for heapsort (for details of the pro-
cedure percolateDown, see the Java code in Section A.1). After each deletion, the
heap size decreases by 1, and the emptied last array position is used to place the just
deleted maximum element. After the last deletion the array contains the keys in as-
cending sorted order. To get them in descending order, we have to build a minimum
heap instead of the above maximum heap.
The first for-loop converts an input array a into a heap by percolating elements
down. The second for-loop swaps each current maximum element to be deleted
with the current last position excluded from the heap and restores the heap by per-
colating each new key from the root down.

algorithm heapSort
Input: array a[0..n − 1]
begin
for i ← ⌊n/2⌋ − 1 while i ≥ 0 step i ← i − 1 do
percolateDown(a, i, n ) build a heap
end for
for i ← n − 1 while i ≥ 1 step i ← i − 1 do
swap(a[0], a[i]) delete the maximum
percolateDown(a, 0, i ) restore the heap
end for
end

Figure 2.7: Heapsort.

Example 2.32. Table 2.6 presents successive steps of heapsort on the input array
a = [70, 65, 50, 20, 2, 91, 25, 31, 15, 8]. In the table, the keys moved are italicized and the
items sorted are boldfaced.
Table 2.6: Successive steps of heapsort.

Position 1 2 3 4 5 6 7 8 9 10
Index 0 1 2 3 4 5 6 7 8 9
Initial array 70 65 50 20 2 91 25 31 15 8
Building max heap 70 65 50 20 8 91 25 31 15 2
70 65 50 31 8 91 25 20 15 2
70 65 91 31 8 50 25 20 15 2
70 65 91 31 8 50 25 20 15 2
91 65 70 31 8 50 25 20 15 2
Max heap 91 65 70 31 8 50 25 20 15 2
Deleting max 1 2 65 70 31 8 50 25 20 15 91
Restoring heap 1-9 70 65 50 31 8 2 25 20 15 91
Deleting max 2 15 65 50 31 8 2 25 20 70 91
Restoring heap 1-8 65 31 50 20 8 2 25 15 70 91
Deleting max 3 15 31 50 20 8 2 25 65 70 91
Restoring heap 1-7 50 31 25 20 8 2 15 65 70 91
Deleting max 4 15 31 25 20 8 2 50 65 70 91
Restoring heap 1-6 31 20 25 15 8 2 50 65 70 91
Deleting max 5 2 20 25 15 8 31 50 65 70 91
Restoring heap 1-5 25 20 2 15 8 31 50 65 70 91
Deleting max 6 8 20 2 15 25 31 50 65 70 91
Restoring heap 1-4 20 15 2 8 25 31 50 65 70 91
Deleting max 7 8 15 2 20 25 31 50 65 70 91
Restoring heap 1-3 15 8 2 20 25 31 50 65 70 91
Deleting max 8 2 8 15 20 25 31 50 65 70 91
Restoring heap 1-2 8 2 15 20 25 31 50 65 70 91
Deleting max 9 2 8 15 20 25 31 50 65 70 91

P RIORITY- QUEUE SORT A heap is a particular implementation of the priority queue

ADT (see Section C.1). There are many other such implementations. From an ab-
stract point of view, heapsort simply builds a priority queue by inserting all the el-
ements of the list to be sorted, then deletes the highest priority element repeatedly.
Any priority queue implementation can be used to sort in the same way.
For example, a very simple implementation of a priority queue is via an unsorted
list. In this case, building the queue is trivial, and the sorting algorithm is exactly
selection sort. Another simple implementation is via a sorted list. In this case, the
algorithm is just insertion sort (we don’t really need to delete the elements since the
resulting list after building the priority queue is the output we are seeking).
There are many more sophisticated implementations of priority queues. They
are useful not only for sorting, but for several important graph algorithms covered
in this book, and also for applications such as discrete event simulation. There is
still active research on finding better priority queue implementations.
Given that we can build a priority queue (such as a heap) in linear time, it is
natural to ask whether we could implement a priority queue in such a way that the
successive deletions can be done in better than O(n log n) time, thus yielding a sorting
algorithm that is asymptotically faster in the worst case than any we have yet seen.
Perhaps surprisingly, the answer is no for comparison-based algorithms, as we see
in the next section.
Exercises
Exercise 2.5.1. Insert a 12th element, 85, into the final heap in Table 2.4.
Exercise 2.5.2. Delete the maximum key from the final heap in Table 2.5 and restore
the heap order.
Exercise 2.5.3. Convert the array [10, 20, 30, 40, 50, 60, 70, 80, 90] into a heap using the
algorithm in Lemma 2.31.
Exercise 2.5.4. Present in the same manner the successive steps of heapsort on the
already sorted input array a = [2, 8, 15, 20, 25, 31, 50, 65, 70, 91].
Exercise 2.5.5. Determine whether heapsort is stable and compare it to insertion
sort, mergesort, and quicksort regarding this property.
Exercise 2.5.6. Determine whether the running time of heapsort on an already sorted
array of size n differs significantly from the average-case running time.

2.6 Data selection

Data selection is closely related to data sorting. The latter arranges an array in
order whereas the former has to find only the kth smallest item (the item of rank
k, also known as the kth order statistic). We have seen (selection sort and heapsort)
that if we can select, then we can sort. The converse is true too: given a list of n items,
we can sort it and then select the desired order statistic easily. The obvious question
is: can we perform selection without sorting, and asymptotically faster?
If for example k = 1 or k = n, then the answer is obviously yes (Exercise 2.6.1).
However, if we require the median (k = ⌈n/2⌉) the answer is not at all obvious. For
example, building a heap and repeatedly extracting the minimum would take time
in Ω(n log n), which is no better than sorting the entire list.
A variation of quicksort works well on this problem, in the average case for ran-
dom data. However it has all the drawbacks of quicksort, and in the worst case its
performance degenerates badly.
 The idea of quickselect is that after the partitioning stage of quicksort, we know
which of the two parts of the partition holds the element we are seeking, and so we
can eliminate one recursive call. In other words, it proceeds as follows.

• If the size of the list is 0, return “not found”; if the size is 1, return the element
of that list. Otherwise:

• Choose one of the items in the list as a pivot.

• Partition the remaining items into two disjoint sublists: reorder the list by plac-
ing all items greater than the pivot to follow it, and all elements less than the
pivot to precede it. Let j be the index of the pivot after partitioning.

• If k < j, then return the result of quickselect on the “head” sublist; otherwise
if k = j, return the element p; otherwise return the result of quickselect on the
“tail” sublist.

Analysis of quickselect
Correctness of quickselect is established just as for quicksort (see Exercise 2.6.2).
In the worst case, the running time can be quadratic; for example, if the input list is
already sorted and we use the naive pivot selection rule, then to find the maximum
element takes quadratic time.

Theorem 2.33. The average-case time complexity of quickselect is Θ(n).

Proof. Let T (n) denote the average time to select the k-th smallest element among
n elements, for fixed k where the average is taken over all possible input sequences.
Partitioning uses no more than cn operations and forms two subarrays, of size i and
n − 1 − i, respectively, where 0 ≤ i < n.
As in quicksort, the final pivot position in the sorted array has equal probability,
1
n , of taking each value of i. Then T (n) averages the average running time for all the
above pairs of the subarrays over all possible sizes. Because only one subarray from
each pair is recursively chosen, the average running time for the pair of size i and
n − 1 − i is (T (i) + T (n − 1 − i))/2 so that

1 n−1 1 n−1
T (n) = ∑
2n i=0
(T (i) + T (n − 1 − i)) + cn = ∑ T (i) + cn
n i=0

n−1
As for quicksort, the above recurrence can be rewritten as nT (n) = ∑i=0 T (i) + cn2
n−2
and subtracting the analogous equation (n − 1)T (n − 1) = ∑i=0 T (i) + c · (n − 1)2 and
rearranging, we are eventually led to the familiar recurrence T (n) = T (n − 1) + c and
can see that T (n) is Θ(n).
Implementation of quickselect
The only change from quicksort is that instead of making two recursive calls on
the left and right subarrays determined by the pivot, quickselect chooses just one of
these subarrays.

algorithm quickSelect
Input: array a[0..n]; array indices l, r; integer k
finds kth smallest element in the subarray a[l..r]
begin
if l ≤ r then
i ← pivot(a, l, r) return position of pivot element
j ← partition(a, l, r, i) return final position of pivot
q ← j − l + 1 the rank of the pivot in a[l..r]
if k = q then return a[ j]
else if k < q then return quickSelect(a, l, j − 1, k)
else return quickSelect(a, j + 1, r, k − q)
end if
else return “not found”
end

Figure 2.8: Basic array-based quickselect.

Figure 2.8 presents a basic array-based implementation. The algorithm processes

a subarray a[l..r], where 0 ≤ l ≤ r ≤ n − 1, of an input array a of size n, assuming that
the desired rank k is in the correct range. A search in the whole array is performed
with the input indices l = 0 and r = n − 1. The search fails only if k is outside the cor-
rect range (including the case where the subarray is empty). Section A.1 contains a
Java implementation that uses median-of-three pivoting.
Exercises
Exercise 2.6.1. Give a simple nonrecursive algorithm to find the maximum of a list
of n elements. How many comparisons does it do in the worst case? Is this the best
we can hope for?

Exercise 2.6.2. Prove that quickselect is correct.

Exercise 2.6.3. To find p keys of fixed ranks, k1 , . . . , k p , in an unordered array, we can

either (i) run quickselect p times or (ii) use quicksort once to order the array and
simply fetch the desired p keys. Find a condition (in terms of p and n) when on the
average the second variant is more time-efficient (time to fetch array elements can
be ignored). Which variant is better if n is 1 million and p = 10?
Exercise 2.6.4. Investigate “heapselect” and its time complexity. Do the same for
“mergeselect”.

2.7 Lower complexity bound for sorting

All the above sorting methods have average and worst time complexity bounds
in Ω(n log n). One of most fundamental theorems in algorithm analysis shows that
no comparison-based sorting algorithm can have a better asymptotic lower bound.
The proof uses the idea of a n-element binary decision tree representation of any
sorting of n items by pairwise comparisons.

1:2
a1 ≥ a2 a1 < a2

2:3 2:3

a2 ≥ a3 a2 < a3 a2 ≥ a3 a2 < a3

321 1:3 1:3 123

a1 ≥ a3 a1 ≥ a3
a1 < a3 a1 < a3

231 213 312 132

Figure 2.9: Decision tree for n = 3.

Example 2.34. A decision tree for three (n = 3) elements a1 , a2 , a3 is shown in Fig-

ure 2.9. Each internal tree node represents a pairwise comparison and decision
made at a particular step of sorting (in Figure 2.9, i : j denotes the comparison be-
tween the elements a[i] and a[ j]). Two downward arcs from the node indicate two
possible results of the comparison: a[i] ≥ a[ j] or a[i] < a[ j]. Leaves represent the
sorted lists. In Figure 2.9, each triple i jk denotes the sorted list (ai , a j , ak ) obtained
from the initial list (a1 , a2 , a3 ). For instance, 123, 213, or 321 mean that the final list is
(a1 , a2 , a3 ), (a2 , a1 , a3 ), or (a3 , a2 , a1 ), respectively. Because any of the n! permutations
of n arbitrary items a1 , a2 , . . . , an may be obtained as the result of the sort, the decision
tree must have at least n! leaves.
The path length from the root to a leaf in the decision tree is equal to the num-
ber of comparisons to be performed in order to obtain the sorted array at the leaf.
Therefore, the longest path (the height of the tree) equals the number of compar-
isons in the worst case. For example, 3 elements can be sorted with no more than 3
comparisons because the height of the 3-element tree in Figure 2.9 is equal to 3.
Theorem 2.35. Every comparison-based sorting algorithm takes Ω(n log n) time in
the worst case.

Proof. We first claim that each binary tree of height h has at most 2h leaves. Once this
claim is established, we proceed as follows. The least value h such that 2h ≥ n! has
the lower bound h ≥ lg(n!) which is in Ω(n log n) (the asymptotic result follows from
Section D.6). This will prove the theorem.
To prove the above claim about tree height, we use mathematical induction on
h. A tree of height 0 has obviously at most 20 leaves. Now suppose that h ≥ 1 and
that each tree of height h − 1 has at most 2h−1 leaves. The root of a decision tree of
height h is linked to two subtrees, being each at most of height h− 1. By the induction
hypothesis, each subtree has at most 2h−1 leaves. The number of leaves in the whole
decision tree is equal to the total number of leaves in its subtrees, that is, at most
2h−1 + 2h−1 = 2h .

This result shows that heapsort and mergesort have asymptotically optimal worst-
case time complexity for comparison-based sorting.
As for average-case complexity, one can also prove the following theorem by us-
ing the decision tree idea. Since we are now at the end of our introductory analysis
of sorting, we omit the proof and refer the reader to the exercises, and to more ad-
vanced books.

Theorem 2.36. Every comparison-based sorting algorithm takes Ω(n log n) time in
the average case.

Proof. Try to prove this yourself (see Exercise 2.7.1).

Exercises
Exercise 2.7.1. Prove Theorem 2.36. The following hints may be useful. First show
that the sum of all depths of leaves in a binary decision tree with k leaves is at least
k lg k. Do this by induction on k, using the recursive structure of these trees. Then
apply the above inequality with k = n!.

Exercise 2.7.2. Consider the following sorting method (often called counting sort )
applied to an array a[n] all of whose entries are integers in the range 1..1000. Intro-
duce a new array t[1000] all of whose entries are initially zero. Scan through the array
a[n] and each time an integer i is found, increment the counter t[i − 1] by 1. Once this
is complete, loop through 0 ≤ i ≤ 999 and print out t[i] copies of integer i + 1 at each
step.
What is the worst-case time complexity of this algorithm? How do we reconcile
this with Theorem 2.35?
2.8 Notes
It was once said that sorting consumes 25% of all CPU time worldwide. What-
ever the true proportion today, sorting clearly remains a fundamental problem to be
solved in a wide variety of applications.
Given the rise of object-oriented programming languages, comparison-based sort-
ing algorithms are perhaps even more important than in the past. In practice the
time taken to perform a basic comparison operation is often much more than that
taken to swap two objects: this differs from the case of, say, 32-bit integers, for which
most analysis was done in the past.
Shellsort was proposed by D. Shell in 1959, quicksort by C. A. R. Hoare in 1960,
mergesort in 1945 by J. von Neumann, and heapsort in 1964 by J. W. J. Williams.
Insertion sort and the other quadratic algorithms are very much older.
At the time of writing, versions of mergesort are used in the standard libraries for
the languages Python, C++ and Java, and a hybrid of quicksort and heapsort is used
by C++.
We have not discussed whether there is an algorithm that will find the median
(or any other given order statistic) in worst-case linear time. For a long time this
was unknown, but the answer was shown to be yes in 1973 by Blum, Floyd, Pratt,
Rivest and Tarjan. The algorithm is covered in more advanced books and is fairly
complicated.
Chapter 3

Efficiency of Searching

Searching in a large database is a fundamental computational task. Usually the in-

formation is partitioned into records and each record has a key to use in the search.
Suppose we have a data structure D of records. The search problem is as follows:
given a search key k, either return the record associated with k in D (a successful
search), or indicate that k is not found, without altering D (an unsuccessful search).
If k occurs more than once, return any occurrence. The purpose of the search is usu-
ally to access data in the record (not merely the key). Simple examples are searching
a phone directory or a dictionary.
In this chapter we discuss searching in a general data structure as above. The
more specialized problem of searching for a pattern in a text string is covered in the
printed version of this textbook.

3.1 The problem of searching

The most general data structure that allows searching is called an associative ar-
ray, table or dictionary.
In the general case, a key and a value are linked by association. An associative
array or dictionary is an abstract data type relating a disjoint set of keys to an arbi-
trary set of values. Keys of entries of an associative array may not have any ordering
relation and may be of unknown range. There is no upper bound on the size of this
structure, so that it can maintain an arbitrary number of different pieces of informa-
tion simultaneously. The analogy with a conventional word dictionary may be mis-
leading, because the latter has a lexicographical order while dictionary structures
need not be ordered.
 Another name for this data type is table. It takes its name from program com-
pilation where a symbol table is used to record variables of a program, together
with their type, address, value, etc. Symbol table operations are also fundamental
to database systems.
Definition 3.1. The table ADT is a set of ordered pairs (table entries) of the form
(k, v) where k is an unique key and v is a data value associated with the key k. Each
key uniquely identifies its entry, and table entries are distinguished by keys because
no two distinct entries have identical keys.
Basic operations defined for this structure are: construct (initialize) an empty ta-
ble, enumerate table entries, search for an entry, insert, delete, retrieve, and update
table entries.
Abstractly, a table is a mapping (function) from keys to values. Given a search
key k, table search has to find the table entry (k, v) containing that key. The found
entry may be retrieved, or removed (deleted) from the table, or its value, v, may be
updated. If the table has no such entry, a new entry with key k may be created and
inserted in the table. Operations on a table also initialize a table to the empty one or
indicate that an entry with the given key is absent. Insertions and deletions modify
the mapping of keys onto values specified by the table.
Example 3.2. Table 3.1 presents a very popular (at least in textbooks on algorithms
and data structures) table having three-letter identifiers of airports as keys and asso-
ciated data such as airport locations, as values. Each identifier has a unique integer
representation k = 262 c0 + 26c1 + c2 where the ci ; i = 0, 1, 2, are ordinal numbers of
letters in the English alphabet (A corresponds to 0, B to 1, . . . , Z to 25). For example,
AKL corresponds to 262 · 0 + 26 · 10 + 11 = 271. In total, there are 263 = 17576 possible
different keys and entries.

Table 3.1: A map between airport codes and locations.

Key Associated value v

Code k City Country State / Place
AKL 271 Auckland New Zealand
DCA 2080 Washington USA District of Columbia (D.C.)
FRA 3822 Frankfurt Germany Hesse
GLA 4342 Glasgow UK Scotland
HKG 4998 Hong Kong China
LAX 7459 Los Angeles USA California
SDF 12251 Louisville USA Kentucky
ORY 9930 Paris France

As can be seen from this example, we may map the keys to integers.
 We deal with both static (where the database is fixed in advance and no inser-
tions, deletions or updates are done) and dynamic (where insertions, deletions or
updates are allowed) implementations of the table ADT.
In all our implementations of the table ADT, we may simplify the analysis as fol-
lows. We use lists and trees as our basic containers. We treat each query or update
of a list element or tree node, or comparison of two of them, as an elementary oper-
ation. The following lemma summarizes some obvious relationships.

Lemma 3.3. Suppose that a table is built up from empty by successive insertions,
and we then search for a key k uniformly at random. Let Tss (k) (respectively Tus (k)) be
the time to perform successful (respectively unsuccessful) search for k. Then

• the time taken to retrieve, delete, or update an element with key k is at least
Tss (k);

• the time taken to insert an element with key k is at least Tus (k);

• Tss (k) ≤ Tus (k).

In addition

• the worst case value for Tss (k) equals the worst case value for Tus (k);

• the average value of Tss (k) equals one plus the average of the times for the un-
successful searches undertaken while building the table.

Proof. To insert a new element, we first try to find where it would be if it were con-
tained in the data structure, and then perform a single insert operation into the con-
tainer. To delete an element, we first find it, and then perform a delete operation
on the container. Analogous statements hold for updating and retrieval. Thus for a
given state of the table formed by insertions from an empty table, the time for suc-
cessful search for a given element is the time that it took for unsuccessful search for
that element, as we built the table, plus one. This means that the time for unsuc-
cessful search is always at least the time for successful search for a given element
(the same in the worst case), and the average time for successful search for an ele-
ment in a table is the average of all the times for unsuccessful searches plus one.

If the data structure used to implement a table arranges the records in a list, the
efficiency of searching depends on whether the list is sorted. In the case of the tele-
phone book, we quickly find the desired phone number (data record) by name (key).
But it is almost hopeless to search directly for a phone number unless we have a spe-
cial reverse directory where the phone number serves as a key. We discuss unsorted
lists in the Exercises below, and sorted lists in the next section.
Exercises
Exercise 3.1.1. The sequential search algorithm simply starts at the head of a list
and examines elements in order until it finds the desired key or reaches the end of
the list. An array-based version is shown in Figure 3.1.

algorithm sequentialSearch
Input: array a[0..n − 1]; key k
begin
for i ← 0 while i < n step i ← i + 1 do
if a[i] = k then return i
end for
return not found
end

Figure 3.1: A sequential search algorithm.

Show that both successful and unsuccessful sequential search in a list of size n
have worst-case and average-case time complexity Θ(n).
Exercise 3.1.2. Show that sequential search is slightly more efficient for sorted lists
than unsorted ones. What is the time complexity of successful and unsuccessful
search?

3.2 Sorted lists and binary search

A sorted list implementation allows for much better search method that uses the
divide-and-conquer paradigm. The basic idea of binary search is simple. Let k be
the desired key for which we want to search.
• If the list is empty, return “not found”. Otherwise:
• Choose the key a[m] of the middle element of the list. If a[m] = k, return its
record; if a[m] > k, make a recursive call on the head sublist; if a[m] < k, make a
recursive call on the tail sublist.
Example 3.4. Figure 3.2 illustrates binary search for the key k = 42 in a list of size 16.
At the first iteration, the search key 42 is compared to the key a[7] = 53 in the middle
position m = ⌊(0 + 15)/2⌋ = 7.
Because 42 < 53, the second iteration explores the first half of the list with posi-
tions 0, . . . , 6. The search key is compared to the key a[3] = 33 in the middle position
m = ⌊(0 + 6)/2⌋ = 3. Because 42 > 33, the third iteration explores the second half of
the current sublist with positions 4, . . . , 6. The comparison to the key a[5] = 49 in the
middle position m = ⌊(4 + 6)/2⌋ = 5 reduces the search range to a single key, a[4] = 42,
because now l = r = m = 4. Because the single entry is equal to the search key, the
algorithm returns the answer 4.
 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 i
7 14 27 33 42 49 51 53 67 70 77 81 89 94 95 99 ai

l=0 m=7 r = 15

7 14 27 33 42 49 51

l=0 m=3 r=6

42 49 51

l=4 r=6
m=5

42

l = r = m = 4; search key = a return 4

4

Figure 3.2: Binary search for the key 42.

Analysis of sorted list-based tables

Binary search is readily seen to be correct by induction on the size of the list (Ex-
ercise 3.2.3).
The performance of binary search on an array is much better than on a linked list
because of the constant time access to a given element.
Lemma 3.5. Consider a sorted list implementation of the table ADT.
Using an array, both successful and unsuccessful search take time in Θ(log n), in
both the average and worst case, as do retrieval and updating. Insertion and deletion
take time in Θ(n) in the worst and average case.
Using a linked list, all the above operations take time in Θ(n).

Proof. Unsuccessful binary search takes ⌈lg(n+1)⌉ comparisons in every case, which
is Θ(log n). By Lemma 3.3, successful search also takes time in Θ(log n) on average
and in the worst case. Insertion and deletion in arrays takes Θ(n) time on average
and in the worst case.
For linked lists, the searches take time in Θ(n) and the list insertion and deletion
take constant time.
 Binary search performs a predetermined sequence of comparisons depending
on the data size n and the search key k. This sequence is better analysed when a
sorted list is represented as a binary search tree. For simplicity of presentation, we
suppress the data records and make all the keys integers. Background information
on trees is found in Section D.7.

Definition 3.6. A binary search tree (BST) is a binary tree that satisfies the follow-
ing ordering relation: for every node ν in the tree, the values of all the keys in the
left subtree are smaller than (or equal, if duplicates allowed) to the key in ν and the
values of all the keys in the right subtree are greater than the key in ν.

In line with the ordering relation, all the keys can be placed in sorted order by
traversing the tree in the following way: recursively visit, for each node, its left sub-
tree, the node itself, then its right subtree (this is the so-called inorder traversal).
The relation is not very desirable for duplicate keys; for exact data duplicates, we
should have one key and attach to it the number of duplicates.
The element in the middle position, m0 = ⌊(n − 1)/2⌋, of a sorted array is the
root of the tree representation. The lower subrange, [0, . . . , m0 − 1], and the upper
subrange, [m0 + 1, . . . , n − 1], of indices are related to the left and right arcs from the
root. The elements in their middle positions, mleft,1 = ⌊(m0 − 1)/2⌋ and mright,1 = ⌊(n +
m0 )/2⌋, become the left and right child of the root, respectively. This process is re-
peated until all the array elements are related to the nodes of the tree. The middle
element of each subarray is the left child if its key is less than the parent key or the
right child otherwise.

Example 3.7. Figure 3.3 shows a binary search tree for the 16 sorted keys in Fig-
ure 3.2. The key a[7] = 53 is the root of the tree. The lower [0..6] and the upper [8..15]
subranges of the search produce the two children of the root: the left child a[3] = 33
and the right child a[11] = 81. All other nodes are built similarly.

The tree representation interprets binary search as a pass through the tree from
the root to a desired key. If a leaf is reached but the key is not found, then the search
is unsuccessful. The number of comparisons to find a key is equal to the number of
nodes along the unique path from the root to the key (the depth of the node, plus
one).
A static binary search always yields a tree that is well-balanced: for each node
in the tree, the left and right subtrees have height that differs by at most 1. Thus all
leaves are on at most two levels. This property is used to define AVL trees in Sec-
tion 3.4.
Implementation of binary search
Algorithm binarySearch in Figure 3.4 searches for key k in a sorted array a.
The search starts from the whole array from l = 0 to r = n − 1. If l and r ever cross,
l > r, then the desired key is absent, indicating an unsuccessful search. Otherwise,
 7

53
0..6 8..15

33 3 11 81

0..2 4..6 8..10 12..15

14 1 49 5 9 70 13 94

0 2 4 6 8 10 12 14..15

7 27 42 51 67 77 89 95
0 2 4 6 8 10 12 14

15

key tree node i array position

l..r range of positions 15 99

Figure 3.3: Binary tree representation of a sorted array.

the middle position

 ofthe odd range or rounded down “almost middle” position of
the even one, m = l+r
2 , is computed. The search key, k, is compared to the key a[m]
in this position:

• If k = a[m], then return the found position m.

• If k < a[m], then the key may be in the range l to m − 1 so that r ← m − 1 at next
step.

• If k > a[m], then it may be in the range m + 1 to r so that l ← m + 1 at next step.

Binary search is slightly accelerated if the test for a successful search is removed
from the inner loop in Figure 3.4 and the search range is reduced by one in all cases.
To determine whether the key k is present or absent, a single test is performed out-
side the loop (see Figure 3.5). If the search key k is not larger than the key a[m] in the
middle position m, then it may be in the range from l to m. The algorithm breaks the
while-loop when the range is 1, that is, l = r, and then tests whether there is a match.
 algorithm binarySearch
Input: array a[0..n − 1]; key k
begin
l ← 0; r ← n − 1
while l ≤ r do
m ← ⌊(l + r)/2⌋
if a[m] < k then l ← m + 1
else if a[m] > k then r ← m − 1 else return m
end if
end while
return not found
end

Figure 3.4: Binary search with three-way comparisons.

algorithm binarySearch2
Input: array a[0..n − 1]; key k
begin
l ← 0; r ← n − 1
while l < r do
m ← ⌊(l + r)/2⌋
if a[m] < k then l ← m + 1
else r ← m
end while
if a[l] = k then return l
else return not found
end

Figure 3.5: Faster binary search with two-way comparisons.

Exercises
Exercise 3.2.1. Perform a search for the key 41 in the situation of Example 3.4.
Exercise 3.2.2. How many comparisons will binary search make in the worst case to
find a key in a sorted array of size n = 106 ?
Exercise 3.2.3. Prove that both given array implementations of binary search cor-
rectly find an element or report that it is absent.
Exercise 3.2.4. If we have more information about a sorted list, interpolation search
can be (but is not always) much faster than binary search. It is the method used
when searching a telephone book: to find a name starting with “C” we open the
book not in the middle, but closer to the beginning.
 k−a[l]
To search for an element between position l and r with l < r, we choose ρ = a[r]−a[l]
and the next guess is m = l + ⌈ρ(r − l)⌉ .
Give an example of a sorted input array of 8 distinct integers for which interpola-
tion search performs no better than sequential search.

Exercise 3.2.5. Determine how many positions interpolation search will examine to
find the key 85 among the ten keys (10, 20, 35, 45, 55, 60, 75, 80, 85, 100).

3.3 Binary search trees

In Section 3.2 we have already used a binary search tree to specify sequences of
comparisons and simplify the performance analysis of binary search. Now we use
a binary search tree not just as a mathematical object to aid our analysis, but as an
explicit data structure that implements the table ADT.

10 10 10

3 15 3 15 3 15

1 5 1 5 1 11

4 8 2 8 4 12

Figure 3.6: Binary trees: only the leftmost tree is a binary search tree.

Example 3.8. In Figure 3.6, two trees are not binary search trees because the key 2
in the middle tree is in the right subtree of the key 3, and the keys 11 and 12 in the
rightmost tree are in the left subtree of the key 10.

Binary search trees implement efficiently the basic search, insert, and remove
operations of the table ADT. In addition a BST allows for more operations such as
sorting records by their keys, finding the minimum key, etc (see Exercises).
Binary search trees are more complex than heaps. Only a root or leaves are added
to or removed from a heap, whereas any node of a binary search tree may be re-
moved.
The search operation resembles usual binary search in that it starts at the root
and moves either left or right along the tree, depending on the result of comparing
the search key to the key in the node.
 10 8<10? 10 7<10?

3 8<3? 15 3 7<3? 15

1 5 8<5? 1 5 7<5?

4 8 8=8 4 8 7<8?

found node

7 inserted node

Figure 3.7: Search and insertion in the binary search tree.

Example 3.9. To find key 8 in the leftmost tree in Figure 3.6, we start at the root 10
and go left because 8 < 10. This move takes us to 3, so we turn right (because 8 > 3)
to 5. Then we go right again (8 > 5) and encounter 8.

Example 3.10. To find key 7, we repeat the same path. But, when we are at node 8,
we cannot turn left because there is no such link. Hence, 7 is not found.

Figure 3.7 illustrates both Examples 3.9 and 3.10. It shows that the node with key
7 can be inserted just at the point at which the unsuccessful search terminates.
The removal operation is the most complex because removing a node may dis-
connect the tree. Reattachment must retain the ordering condition but should not
needlessly increase the tree height. The standard method of removing a node is as
follows.

• A leaf node is simply deleted.

• An internal node with only one child is deleted after its child is linked to its
parent node.

• If the node has two children, then it should be swapped with the node having
the smallest key in its right subtree. The latter node is easily found (see Exer-
cise 3.3.1). After swapping, the node can be removed as in the previous cases,
since it is now in a position where it has at most one child.

This approach appears asymmetric but various modifications do not really im-
prove it. The operation is illustrated in Figure 3.8.
 5 remove 10

3 10

1 4 8 15
5
replace 10
0 2 6 12 18
3 10

1 4 8 15
5

3 12 2
0 6 12 18

minimum
1 4 8 15 in the
right
subtree

0 2 6 18

Figure 3.8: Removal of the node with key 10 from the binary search tree.

Analysis of BST-based tables

Lemma 3.11. The search, retrieval, update, insert and remove operations in a BST
all take time in O(h) in the worst case, where h is the height of the tree.

Proof. The running time of these operations is proportional to the number of nodes
visited. For the find and insert operations, this equals 1 plus the depth of the node,
but for the remove operation it equals the depth of the node plus at most the height
of the node. In each case this is O(h).

For a well-balanced tree, all operations take logarithmic time. The problem is
that insertions and deletions may destroy the balance, and in practice BSTs may be
heavily unbalanced as in Figure 3.11. So in the worst case the search time is linear,
Θ(n), and we have an inferior form of sequential search (because of the extra over-
head in creating the tree arcs).
Figure 3.9 presents all variants of inserting the four keys 1, 2, 3, and 4 into an
empty binary search tree. Because only relative ordering is important, this corre-
sponds to any four keys (a[i] : i = 1, . . . , 4) such that a[1] < a[2] < a[3] < a[4].
 1 1 1 1 1 4 4 4 4 4

4 4 3 2 2 3 3 2 1 1

3 2 2 4 4 3 2 1 1 3 3 2

2 3 1324 3 4 2 4213
1342 1 4231 2 3

1432 1423 1243 1234 4321 4312 4132 4123

3 3 2 2

1 4 2 4 1 3 1 4

2 1 4 3

3124 3214 2134 2143

3142 3241 2314 2413
3412 3421 2341 2431

Figure 3.9: Binary search trees obtained by permutations of 1, 2, 3, 4.

There are in total 4! = 24 possible insertion orders given in Figure 3.9. Some trees
result from several different insertion sequences, and more balanced trees appear
more frequently than unbalanced ones.
Definition 3.12. The total internal path length, Sτ (n), of a binary tree τ is the sum
of the depths of all its nodes.
The average node depth, 1n Sτ (n), gives the average time complexity of a success-
ful search in a particular tree τ. The average-case time complexity of searching is
obtained by averaging the total internal path lengths for all the trees of size n, that
is, for all possible n! insertion orders, assuming that these latter occur with equal
probability, n!1 .
Lemma 3.13. Suppose that a BST is created by n random insertions starting from
an empty tree. Then the expected time for successful and unsuccessful search is
Θ(log n). The same is true for update, retrieval, insertion and deletion.

Proof. Let S(n) denote the average of Sτ over all insertion sequences, each sequence
considered as equally likely. We need to prove that S(n) is Θ(n log n).
It is obvious that S(1) = 0. Furthermore, any n-node tree, n > 1, contains a left
subtree with i nodes, a root at height 0, and a right subtree with (n − i − 1) nodes
where 0 ≤ i ≤ n − 1, each value of i being by assumption equiprobable.
For a fixed i, S(i) is the average total internal path length in the left subtree with re-
spect to its own root and S(n−i−1) is the analogous total path length in the right sub-
tree. The root of the tree adds 1 to the path length of each other node. Because there
are n − 1 such nodes, the following recurrence holds: S(n) = (n − 1) + S(i) + S(n − i − 1)
for 0 ≤ i ≤ n − 1. After summing these recurrences for all i = 0, . . . , n − 1 and averaging,
just the same recurrence as for the average-case quicksort analysis (see Lemma 2.15)
n−1
is obtained: S(n) = (n − 1) + 2n ∑i=0 S(i). Therefore, the average total internal path
length is Θ(n log n). The expected depth of a node is therefore in Θ(log n). This means
that search, update, retrieval and insertion take time in Θ(log n). The same result is
true for deletion (this requires the result that the expected height of a random BST is
Θ(log n), which is harder to prove—see Notes).
Thus in practice, for random input, all BST operations take time about Θ(log n).
However the worst-case linear time complexity, Θ(n), is totally unsuitable in many
applications, and deletions can also destroy balance.
We tried to eliminate the worst-case degradation of quicksort by choosing a pivot
that performs well on random input data and relying on the very low probability of
the worst case. Fortunately, binary search trees allow for a special general technique,
called balancing , that guarantees that the worst cases simply cannot occur. Balanc-
ing restructures each tree after inserting new nodes in such a way as to prevent the
tree height from growing too much. We discuss this more in Section 3.4.

5 5

3 10 4 10

1 4 8 15 3 8 15

Figure 3.10: Binary search trees of height about log n.

15 1 1

10 3 15

8 4 3

5 5 10

4 8 4

3 10 8

1 15 5

Figure 3.11: Binary search trees of height about n.

Implementation of BST
A concrete implementation of a BST requires explicit links between nodes, each
of which contains a data record, and is more complicated than the other implemen-
tations so far. A programming language that supports easy manipulation of objects
makes the coding easier. We do not present any (pseudo)code in this book.
Exercises
Exercise 3.3.1. Show how to find the maximum (minimum) key in a binary search
tree. What is the running time of your algorithm? How could you find the median,
or an arbitrary order statistic?

Exercise 3.3.2. Suppose that we insert the elements 1, 2, 3, 4, 5 into an initially empty
BST. If we do this in all 120 possible orders, which trees occur most often? Which
tree shapes (in other words, ignore the keys) occur most often?

Exercise 3.3.3. Suppose that we insert the elements 1, 2, 3, 4 in some order into an
initially empty BST. Which insertion orders yield a tree of maximum height? Which
yield a tree of minimum height?

Exercise 3.3.4. Show how to output all records in ascending order of their keys, given
a BST. What is the running time of your algorithm?

3.4 Self-balancing binary and multiway search trees

Because balanced binary search trees are more complex than the standard ones,
the insertion and removal operations usually take longer time on the average. But
because of the resulting balanced structure, the worst-case complexity is O(log n). In
other words, the total internal path lengths of the trees are fairly close to the optimal
value.
AVL, red-black, and AA-trees
Definition 3.14. An AVL tree is a binary search tree with the additional balance prop-
erty that, for any node in the tree, the heights of the left and right subtrees can differ
by at most 1.

This balance condition ensures height Θ(log n) for an AVL tree despite being less
restrictive than requiring the tree to be complete. Complete binary trees have too
rigid a balance condition which is difficult to maintain when new nodes are inserted.

Lemma 3.15. The height of an AVL tree with n nodes is Θ(log n).

Proof. Due to the possibly different heights of its subtrees, an AVL tree of height h
may contain fewer than 2h+1 − 1 nodes. We will show that it contains at least ch nodes
for some constant c > 1, so that the maximum height of a tree with n items is logc n.
Let Sh be the size of the smallest AVL tree of height h. It is obvious that S0 = 1 (the
root only) and S1 = 2 (the root and one child). The smallest AVL tree of height h has
subtrees of height h − 1 and h − 2, because at least one subtree is of height h − 1 and
the second height can differ at most by 1 due to the AVL balance condition. These
subtrees must also be the smallest AVL trees for their height, so that Sh = Sh−1 +Sh−2 +
1.
By mathematical induction, we show easily that Sh = Fh+3 − 1 where Fi is the i-
th Fibonacci number (see Example 1.28) because S0 = F3 − 1, S1 = F4 − 1, and Si+1 =
(Fi+3 − 1) + (Fi+2 − 1) + 1 ≡ Fi+4 − 1. Therefore, for each AVL tree with n nodes n ≥ Sh ≈
h+3
φ√
5
− 1 where φ ≈ 1.618. Thus, the height of an AVL tree with n nodes satisfies the
following condition: h ≤ 1.44 · lg(n + 1) − 1.33, and the worst-case height is at most
44% more than the minimum height for binary trees.

Note that AVL trees need in the worst case only about 44% more comparisons
than complete binary trees. They behave even better in the average case. Although
theoretical estimates are unknown, the average-case height of a randomly constructed
AVL tree is close to lg n.
All basic operations in an AVL tree have logarithmic worst-case running time.
The difficulty is, of course, that simple BST insertions and deletions can destroy the
AVL balance condition. We need an efficient way to restore the balance condition
when necessary. All self-balancing binary search trees use the idea of rotation to do
this.
Example 3.16. Figure 3.12 shows a left rotation at the node labelled p and a right
rotation at the node labelled q (the labels are not related to the keys stored in the
BST, which are not shown here). These rotations are mutually inverse operations.
Each rotation involves only local changes to the tree, and only a constant number
of pointers must be updated, independent of the tree size. If, for example, there is a
subtree of large height below a, then the right rotation will decrease the overall tree
height.

q p

p c a q

Right Rotation

a b Left Rotation b c

Figure 3.12: Left and right rotations of a BST.

The precise details of exactly when a rotation is required, and which kind, differ
depending on the type of balanced BST. In each case the programming of the inser-
tion and removal operations is quite complex, as is the analysis. We will not go into
more details here—the reader should consult the recommended references.
 Balancing of AVL trees requires extra memory and heavy computations. This is
why even more relaxed efficient balanced search trees such as red-black trees are
more often used in practice.

Definition 3.17. A red-black tree is a binary search tree such that every node is
coloured either red or black, and every non-leaf node has two children. In addition,
it satisfies the following properties:

• the root is black;

• all children of a red node must be black;

• every path from the root node to a leaf must contain the same number of black
nodes.

Theorem 3.18. If every path from the root to a leaf contains b black nodes, then the
tree contains at least 2b − 1 black nodes.

Proof. The statement holds for b = 1 (in this case the tree contains either the black
root only or the black root and one or two red children). In line with the induction
hypothesis, let the statement hold for all red-black trees with b black nodes in every
path. If a tree contains b + 1 black nodes in every path and has two black children
of the root, then the tree contains two subtrees with b black nodes just under the
root and has in total at least 1+2 · (2b − 1) = 2b+1 − 1 black nodes. If the root has a red
child, the latter has only black children, so that the total number of the black nodes
can become even larger.

Each path cannot contain two consecutive red nodes and increase more than
twice after all the red nodes are inserted. Therefore, the height of a red-black tree is
at most 2⌈lg n⌉, and the search in it is logarithmic, O(log n).
Red-black trees allow for a very fast search. This data structure has still no precise
analysis of its average-case performance. Its properties are found either experimen-
tally or by analysing red-black trees containing random n keys. There are about lg n
comparisons per search on the average and fewer than 2 lg n + 2 comparisons in the
worst case. Restoring the tree after insertion or deletion of single node requires O(1)
rotations and O(log n) colour changes in the worst case.
Another variety of balanced tree, the AA-tree, becomes more efficient than a red-
black tree when node deletions are frequent. An AA-tree has only one extra condi-
tion with respect to a red-black tree, namely that the left child may not be red. This
property simplifies the removal operation considerably.
Balanced B-trees: efficiency of external search
The B-tree is a popular structure for ordered databases in external memory such
as magnetic or optical disks. The previous “Big-Oh” analysis is invalid here because
it assumes all elementary operations have equal time complexity. This does not hold
for disk input / output where one disk access corresponds to hundreds of thousands
of computer instructions and the number of accesses dominates running time. For
a large database of many millions of records, even logarithmic worst-case perfor-
mance of red-black or AA-trees is unacceptable. Each search should involve a very
small number of disk accesses, say, 3–4, even at the expense of reasonably complex
computations (which will still take only a small fraction of a disk access time).
Binary tree search cannot solve the problem because even an optimal tree has
height lg n. To decrease the height, each node must have more branches. The height
of an optimal m-ary search tree (m-way branching) is roughly logm n (see Table 3.2),
or lg m times smaller than with an optimal binary tree (for example, 6.6 times for
m = 100).
Table 3.2: Height of the optimal m-ary search tree with n nodes.

n 105 106 107 108 109

⌈log2 n⌉ 17 20 24 27 30
⌈log10 n⌉ 5 6 7 8 9
⌈log100 n⌉ 3 3 4 4 5
⌈log1000 n⌉ 2 2 3 3 3

4,10

1,3 6,8 14,17,20

0 1 3 4 6 8 10 14 17 20

Figure 3.13: Multiway search tree of order m = 4.

Figure 3.13 shows that the search and the traversal of a multiway search tree gen-
eralize in a straightforward way the binary search tree operations. If in the latter
case the search key is compared to a single key in a node in order to choose one of
two branches or stop in the node, in an m-ary tree, the search key is compared to at
most m − 1 keys in a node to choose one of m branches. The major difference is that
multiple data records are now associated only with leaves although some multiway
trees do not strictly follow this condition. Thus the worst-case and the average-case
search involve the tree height and the average leaf height, respectively.

Example 3.19. Search for a desired key k in Figure 3.13 is guided by thresholds, for
example at the root it goes left if k < 4, down if 4 ≤ k < 10, and right if k ≥ 10. The
analogous comparisons are repeated at every node until the record with the key k
is found at a leaf or its absence is detected. Let k = 17. First, the search goes right
from the root (as 17 > 10), then goes to the third child of the right internal node (as
17 ≤ 17 < 20), and finally finds the desired record in that leaf.

Definition 3.20. A B-tree of order m is an m-ary tree such that:

• the root is either a leaf or it has between 2 and m children inclusive;

• each nonleaf node (except possibly the root) has between ⌈m/2⌉ and m children
inclusive;

• each nonleaf node with µ children has µ − 1 keys, (θ[i] : i = 1, . . . , µ − 1), to guide
the search where θ[i] is the smallest key in subtree i + 1;

• all leaves are at the same depth;

• data items are stored in leaves, each storing between ⌈l/2⌉ and l items, for some
l.

Other definitions of B-trees (mostly with minor changes) also exist, but the above
one is most popular. The first three conditions specify the memory space each node
needs (first of all, for m links and m − 1 keys) and ensure that more than half of it, ex-
cept possibly in the root, will be used. The last two conditions form a well-balanced
tree.
Note. B-trees are usually named by their branching limits, that is, ⌈m/2⌉–m, so that
2–3 and 6–11 trees are B-trees with m = 3 and m = 11, respectively.

Example 3.21. In a 2−4 B-tree in Figure 3.14 all nonleaf nodes have between ⌈4/2⌉ =
2 and 4 children and thus from 1 to 3 keys. The number l of data records associated
with a leaf depends on the capacity of external memory and the record size. In Fig-
ure 3.14, l = 7 and each leaf stores between ⌈7/2⌉ = 4 and 7 data items.

Because the nodes are at least half full, a B-tree with m ≥ 8 cannot be a simple
binary or ternary tree. Simple ternary 2–3 B-trees with only two or three children per
node are sometimes in use for storing ordered symbol tables in internal computer
RAM. But branching limits for B-trees on external disks are considerably greater to
make one node fit in a unit data block on the disk. Then the number of nodes ex-
amined (and hence the number of disk accesses) decreases lg m times or more com-
pared with a binary tree search.
 m = 4 55 75
l = 7

23 31 40 60 71 84 91

11 23 31 40 55 60 71 75 84 91
13 25 33 42 57 61 72 77 85 93
15 27 34 44 58 62 73 78 89 94
16 28 38 50 59 65 74 79 90 95
18 39 51 66 82 98
20 52 68 99
22 70

Figure 3.14: 2–4 B-tree with the leaf storage size 7 (2..4 children per node and 4..7 data items
per leaf ).

In each particular case, the tree order m and the leaf capacity l depend on the disk
block size and the size of records to store. Let one disk block hold d bytes, each key
be of κ bytes, each branch address be of b bytes, and the database contain s records,
each of size r bytes. In a B-tree of order m, each nonleaf node stores at most m − 1
keys and m branch addresses, that is, in total, κ(m − 1) + bm = (κ + b)m − κ bytes. The
largest order m such that one node fits in one disk block, (κ+b)m−κ ≤ d, is m = d+κ

b+κ .
m d
Each internal node, except the root, has at least 2 branches. At most l = r records
l
fit in one block, and each leaf addresses from  2 to l records. Assuming each leaf is
s
full, the total number of the leaves is n = l , so that in the worst case the leaves are
at level ⌈logm/2 n⌉ + 1.

Example 3.22. Suppose the disk block is d = 215 ≡ 32768 bytes, the key size is κ =
26 ≡ 64 bytes, the branch address has b = 8 bytes, and the database contains s =
230 ∼
= 1.07 · 109 records of size r = 210 ≡ 1024 bytes each. Then the B-tree order is
m = ⌊ 32768+64 32832
8+64 ⌋ = ⌊ 72 ⌋ = 456 so that each internal node, except the root, has at least
228 branches. One block contains at most l = 327681024 = 32 records, and the number of
leaves is at least n = 230 /32 = 225 . The worst-case level of the leaves in this B-tree is
 

⌈log228 225 ⌉ + 1 = ⌈3.19⌉ + 1 = 5.

Generally, a search in or an insertion into a B-tree of order m with n data records
requires fewer than ⌈logm/2 n⌉ disk accesses, and this number is practically constant
if m is sufficiently big as shown in Table 3.2. The running time becomes even smaller
if the root and the upper two tree levels are stored in internal RAM and the slow disk
accesses occur only for level 3 or higher. The three-level B-tree with m = 456 can
handle up to 4563 , or 94, 818, 816 entries. If in Example 3.22 each key uses only κ = 24
bytes, then m = 1024, and the three-level tree can handle over 109 entries.
Data insertion into a B-tree is simple if the corresponding leaf is not already full.
A full leaf has to be split into two leaves, both having the minimum number of data
items, and the parent node should be updated. If necessary, the splitting propa-
gates up until it finds a parent that need not be split or reaches the root. Only in
the extremely rare case that the root has to be split, the tree height increases and a
new root with two children (halves of the previous root) is created. Data deletion
is also simple until the leaf is empty and its neighbours must be combined to form
a full leaf. Although the programming is not simple, all changes are well defined.
Algorithm analysis, beyond the scope of this book, shows that both data insertion,
deletion, and retrieval have only about log m n disk accesses in the worst case.
2

Exercises
Exercise 3.4.1. Draw two different red-black trees containing at most two black nodes
along every path from the root to a leaf.

Exercise 3.4.2. Draw two different AVL trees of size n = 7 and compare them to the
complete binary tree of the same size. Is the latter also an AVL tree?

Exercise 3.4.3. Draw an AA-tree containing at most 2 black nodes along every path
from a node to a leaf and differing from the complete binary tree of order n = 7.

Exercise 3.4.4. Draw a binary search tree of minimum size such that a left rotation
reduces the height of the tree.

3.5 Hash tables

There are numerous ways to implement the table ADT. We have already seen that
various search trees will do everything required, provided the keys are from some
totally ordered set. If, say, the keys are dictionary words with the usual ordering,
then it is not necessary to use any integer encoding—keys can be compared directly.
Suppose now that we have a very simple situation where the number of possible
keys is small. Then we can just store the values in an array. One array entry can be
reserved in advance for each possible key, and the key-to-value mapping ends up
as a conventional array address. Searching then has worst-case constant time, as
does insertion and deletion. This implementation of a table works well provided the
number of possible keys is sufficiently small.
However, that nice state of affairs does not occur often (we could use it for the
airport codes in Example 3.2). Usually there exists a very large number of possible
keys although only a tiny fraction of them are actually put into use. For example,
suppose that we have a database where each customer is identified by an 8-digit
telephone number. If we have 10 000 customers, only 0.01% of the array addresses
are filled.
There is another technique to store and search for values in symbol tables, called
hashing , that uses less space and retains many (not all) of the benefits of direct array
addressing.
Definition 3.23. Hashing computes an integer hash code for each object using a
hash function that maps objects (for example, keys) to indices of a given linear array
(the hash table).
Hash functions are designed in such a manner that hash codes are computed
quickly. The computation of an array index with a hash function, or “hashing a key
to an index”, depends only on the key to hash and is independent of other keys in
the table. If h(k) is the value of a hash function for k, then the key k should be placed
at location h(k).
The hash function is chosen so as to always return a valid index for the array. A
perfect hash function maps each key to a different index. Unfortunately, it is difficult
to find such a function in most cases.
Example 3.24. Let us map two-digit integer keys onto the ten array indices [0, 1, . . . ,
9] by a simple hash function h(k) = ⌊k/10⌋. Then the keys 21 and 27 both have the
hash code 2 pointing to the same position in the array. Such a situation in which
two different keys, k1 6= k2 , hash to the same index (table address), h(k1 ) = h(k2 ), is
called a collision. Because both table entries (k1 , v1 ) and (k2 , v2 ) cannot be at the
same address, we need a definite collision resolution policy.
Different keys hashed to the same hash address are called synonyms, and data
items with synonymic keys are frequently also referred to as synonyms.
Collision resolution: OALP, OADH, and SC hash tables
There are many collision resolution policies. The main issues are:

• Do we use extra storage, or not?

• Which element moves when a collision occurs: the incumbent element or the
newcomer (or both)?
• How do we decide where to move the evicted element to?

C HAINING In separate chaining synonyms with the same hash address are stored
in a linked list connected to that address. We still hash the key of each item to obtain
an array index. But if there is a collision, the new item is simply placed in this hash
address, along with all other synonyms. Each array element is a head reference for
the associated linked list, and each node of this list stores not only the key and data
values for a particular table entry but also a link to the next node. The head node of
the list referenced by the array element always contains the last inserted item.
O PEN ADDRESSING Open addressing uses no extra space for collision resolution. In-
stead, we move one of the colliding elements to another slot in the array. We may
use LIFO (last-in, first out — the new element must move), FIFO (first in, first out —
the old element must move), or more complicated methods such as Robin Hood or
cuckoo hashing (see Notes). For our purposes here, we use LIFO.
Each collision resolution policy probes another array slot, and if empty inserts
the currently homeless element. If the probed slot is not empty, we probe again to
find a slot in which to insert the currently homeless element, and so on until we
finish insertion. The probe sequence used can be a simple fixed sequence, or given
by a more complicated rule (but is always deterministic). They all have the property
that they “wrap around” the array when they reach the end. The two most common
probing methods are:
• (Linear probing) always probe the element to the left;
• (Double hashing) probe to the left by an amount determined by the value of a
secondary hash function.
Note. The choice of probing to the left versus probing to the right is clearly a matter
of convention; the reader should note that other books may use rightward probing
in their definitions.

Table 3.3: Open addressing with linear probing (OALP).

Data [key,value] Hash: key/10 Table address Comments

[20,A] 2 2
[15,B] 1 1
[45,C] 4 4
[87,D] 8 8
[39,E] 3 3
[31,F] 3 0 try 3, 2, 1, 0
[24,G] 2 9 try 2, 1, 0, 9

Example 3.25. Table 3.3 shows how OALP fills the hash table of size 10 using the two-
digit keys and the hash function of Example 3.24. The first five insertions have found
empty addresses. However, the key–value pair [31, F] has a collision because the
address h(31) = 3 is already occupied by the pair [39, E] with the same hash address,
h(39) = 3. Thus, the next lower table address, location 2, is probed to see if it is empty,
and in the same way the next locations 1 and 0 are checked. The address 0 is empty
so that the pair [31, F] can be placed there.
A similar collision occurs when we try to insert the next pair, [24, G], because the
hash address h(24) = 2 for the key 24 is already occupied by the previous pair [20, A].
Consequently, we probe successive lower locations 1 and 0, and since they both are
already occupied, we wrap around and continue the search at the highest location
9. Because it is empty, the pair [24, G] is inserted in this location yielding the final
configuration given in Table 3.3.

OALP is simple to implement but the hash table may degenerate due to clus-
tering . A cluster is a sequence of adjacent occupied table entries. OALP tends to
form clusters around the locations where one or more collisions have occurred. Each
collision is resolved using the next empty location available for sequential probing.
Therefore, other collisions become more probable in that neighbourhood, and the
larger the clusters, the faster they grow. As a result, a search for an empty address to
place a collided key may turn into a very long sequential search.
Another probing scheme, double hashing , reduces the likelihood of clustering.
In double hashing, when a collision occurs, the key is moved by an amount deter-
mined by a secondary hash function ∆. Let h denote the primary hash function.
Then for each key k we have the starting probe address i0 = h(k) and the probe decre-
ment ∆(k). Each next successive probe position is it = (it−1 − ∆(k)) mod m; t = 1, 2, . . .
where m is the table size.

Example 3.26. Table 3.4 shows how OADH fills the same hash table as in Exam-
ple 3.25 if the hash function is given by ∆(k) = (h(k) + k) mod 10.

Table 3.4: Open addressing with double hashing (OADH).

Data [key,value] Hash: key/10 Table address Comments

[20,A] 2 2
[15,B] 1 1
[45,C] 4 4
[87,D] 8 8
[39,E] 3 3
[31,F] 3 9 using ∆(31) = 4
[24,G] 2 6 using ∆(24) = 6

Now when we try to place the key–value pair [31, F] into position h(31) = 3, the
collision is resolved by probing the table locations with decrement ∆(31) = 4. The
first position, (3 − 4) mod 10 = 9 is empty so that the pair [31, F] can be placed there.
For the collision of the pair, [24, G], at location 2 the decrement ∆(24) = 6 immedi-
ately leads to the empty location 6. The final table in Figure 3.4 contains three small
clusters instead of one large cluster in Figure 3.3.

Generally, OADH results in more uniform hashing that forms more clusters than
OALP but of smaller sizes. Linear probing extends each cluster from its end with the
lower table address, and nearby clusters join into larger clusters growing even faster.
Double hashing does not extend clusters only at one end and does not tend to join
nearby clusters.
Analysis of hash tables
The time complexity of searching in and inserting items in a hash table of size m
with n already occupied entries is determined by the load factor, λ := mn . In open
addressing, 0 ≤ λ < 1: λ equals the fraction of occupied slots in the array, and cannot
be exactly equal to 1 because a hash table should have at least one empty entry in
order to efficiently terminate the search for a key or the insertion of a new key.
Open addressing and separate chaining require n probes in the worst case, since
all elements of the hash table may be synonyms. However the basic intuition is that
provided the table is not too full, collisions should be rare enough that searching for
an element requires only a constant number of probes on average.
Thus we want a result such as: “Provided the load factor is kept bounded (and
away from 1 in the case of open addressing), all operations in a hash table take Θ(1)
time in the average case.”
In order to have confidence in this result, we need to describe our mathematical
model of hashing. Since a good hash function should scatter keys randomly, and we
have no knowledge of the input data, it is natural to use the “random balls in bins”
model. We assume that we have thrown n balls one at a time into m bins, each ball
independently and uniformly at random.
For our analysis, it will be useful to use the function Q defined below.

Definition 3.27. For each integer m, n with 1 ≤ n ≤ m, we define

m! m m−1 m−n+1
Q(m, n) = = ... .
(m − n)!mn m m m

Note that Q(m, 1) = 1.

B ASIC ANALYSIS OF COLLISIONS It is obvious that if we have more items than the size
of the hash table, at least one collision must occur. But the distinctive feature of
collisions is that they are relatively frequent even in almost empty hash tables.
The birthday paradox refers to the following surprising fact: if there are 23 or
more people in a room, the chance is greater than 50% that two or more of them have
the same birthday. Note: this is not a paradox in the sense of a logical contradiction,
but just a “counter-intuitive” fact that violates “common sense”.
1
More precisely, if each of 365 bins is selected with the same chance 365 , then after
23 entries have been inserted, the probability that at least one collision has occurred
(at least one bin has at least two balls) is more than 50%. Although the table is only
23/365 (∼= 6.3%) full, more than half of our attempts to insert one more entry will
result in a collision!
 Let us see how the birthday paradox occurs. Let m and n denote the size of a table
and the number of items to insert, respectively. Let Prm (n) be the probability of at
least one collision when n balls are randomly placed into m bins.

Lemma 3.28. The probability of no collisions when n balls are thrown indepen-
dently into m boxes uniformly at random is Q(m, n). Thus Prm (n) = 1 − Q(m, n) and
the expected number of balls thrown until the first collision is ∑n≤m Q(m, n).

Proof. Let πm (n) be the probability of no collisions. The “no collision” event after
inserting ν items; ν = 2, . . . , n, is a joint event of “no collision” after inserting the
preceding ν − 1 items and “no collision” after inserting one more item, given ν − 1
positions are already occupied. Thus Prm (ν) = Prm (ν − 1)Pm (no collision | ν − 1) where
Pm (no collision | ν) denotes the conditional probability of no collision for a single item
inserted into the table with m − ν unoccupied positions. This latter probability is
simply m−νm .
This then yields immediately

m m−1 m − n + 1 m(m − 1) · · · (m − n + 1) m!
πm (n) = ... = = n
m m m mn m (m − n)!
m!
Therefore, Prm (n) = 1 − mn (m−n)! = 1 − Q(m, n) which gives the first result.
The number of balls is at least n + 1 with probability Q(m, n). Since the expected
value of a random variable T taking on nonegative integer values can always be com-
puted by E[T ] = ∑n≥1 i Pr(T = i) = ∑ j≥0 Pr(T > j), and these latter probabilities are zero
when j > m, the second result follows.

Table 3.5 presents (to 4 decimal places) some values of Prm (n) for m = 365 and
n = 5 . . . 100. As soon as m = 47 (the table with 365 positions is only 12.9% full),
the probability of collision is greater than 0.95. Thus collisions are frequent even
in sparsely occupied tables.

Table 3.5: Birthday paradox: Pr365 (n).

n 5 10 15 20 22
Pr365 (n) 0.0271 0.1169 0.2529 0.4114 0.4757
n 23 25 30 35 40
Pr365 (n) 0.5073 0.5687 0.7063 0.8144 0.8912
n 45 50 55 60 65
Pr365 (n) 0.9410 0.9704 0.9863 0.9941 0.9977
n 70 75 80 90 100
Pr365 (n) 0.9992 0.9997 0.9999 1.0000 1.0000
 1.0

0.8

0.6

0.4

0.2

0 100 200 300

Figure 3.15: Birthday paradox: Pr365 (n).

Figure 3.15 shows the graph of Pr365 (n) as a function of n. The median of this
distribution occurs around n = 23, as we have said above, and so 23 or more balls
suffice for the probability of a collision to exceed 1/2. Also, the expected number of
balls before the first collision is easily computed to be 25.
When the load factor is much less than 1, the average number of balls per bin is
small. If the load factor exceeds 1 then the average number is large. In each case
analysis is not too difficult. For hashing to be practical, we need to be able to fill
a hash table as much as possible before we spend valuable time rehashing — that
is, allocating more space for the table and reassigning hash codes via a new hash
function. Thus we need to analyse what happens when the load factor is comparable
to 1, that is, when the number of balls is comparable to the number of bins. This also
turns out to be the most interesting case mathematically.
T HEORETICAL ANALYSIS OF HASHING In addition to the basic operations for arrays,
we also consider the computation of the hash address of an item to be an elementary
operation.
Chaining is relatively simple to analyse.
Lemma 3.29. The expected running time for unsuccessful search in a hash table
with load factor λ using separate chaining is given by

Tus (λ) = 1 + λ.

The expected running time for successful search is O(1 + λ/2).

Update, retrieval, insertion and deletion all take time that is O(1 + λ).
Proof. In an unsuccessful search for a key k, the computation of the hash address,
h(k), takes one elementary operation. The average running time to unsuccessfully
search for the key at that address is equal to the average length of the associated
chain, λ = mn . Thus in total Tus (λ) = 1 + λ = 1 + n/m.
The result for successful search and other operations now follows from Lemma 3.3,
since update, retrieval and deletion from a linked list take constant time.

To analyse open addressing, we must make some extra assumptions. We use the
uniform hashing hypothesis: each configuration of n keys in a hash table of size m
is equally likely to occur. This is what we would expect of a truly “random” hash
function, and it seems experimentally to be a good model for double hashing. Note
that this is stronger than just requiring that each key is independently and uniformly
likely to hash initially to each slot before collision resolution (“random balls in bins”).
It also implies that all probe sequences are equally likely.

Lemma 3.30. Assuming the uniform hashing hypothesis holds, the expected num-
ber of probes for search in a hash table satisfy

1
Tus (λ) ≤
1−λ
and
1 1
Tss (λ) ≤ ln .
λ 1−λ
Proof. The average number of probes for an unsuccessful search is Tus (λ) = ∑ni=1 ipm,n (i)
where pm,n (i) denotes the probability of exactly i probes during the search. Obvi-
ously, pm,n (i) = Pr(m, n, i) − Pr(m, n, i + 1) where Pr(m, n, i) is the probability of i or more
probes in the search. By a similar argument to that used in the birthday problem
analysis we have for i ≥ 2

n n−1 n−i+2
Pr(m, n, i) = · ···
m m−1 m−i+2

while Pr(m, n, 1) = 1. Note that clearly Pr(m, n, i) ≤ (n/m)i−1 = λi−1 . Thus

n
Tus (λ) = ∑ i (Pr(n, m, i) − Pr(n, m, i + 1))
i=1
∞ ∞ ∞
1 m
≤ ∑ i (Pr(n, m, i) − Pr(n, m, i + 1)) = ∑ Pr(n, m, i) ≤ ∑ λi−1 = 1 − λ = m − n .
i=1 i=1 i=1
 The result for successful search now follows by averaging. We have

1 n−1 m
Tss (λ) ≤ ∑ m−i
n i=0
m
1 1
=
λ ∑
j=m−n+1 j
Z n    
1 1 m 1 1
≤ dx/x = ln = ln .
λ n−m λ m−n λ 1−λ

Note. It can be shown that the exact values are

m+1 1
Tus (λ) = ≈
m−n+1 1−λ
 
m+1 1 1
Tss (λ) = (Hm+1 − Hm−n+1 ) ≈ ln
n λ 1−λ
so that the upper bounds in the theorem are asymptotically precise as m → ∞.
Good implementations of OADH have been shown in theory and practice to be
well described by the simple uniform hashing model above, so we may safely use the
above results.
However, OALP is not well described by the uniform hashing model, because of
its clustering behaviour. It can be analysed, however, in a similar but more compli-
cated manner.
Lemma 3.31. Assuming uniformly hashed random input, the expected number of
probes for successful, Tss (λ), and unsuccessful, Tus (λ), search in a hash table using
OALP are, respectively,
   2 !
1 1
Tss (λ) ≈ 0.5 1 + and Tus (λ) ≈ 0.5 1 +
1−λ 1−λ

Proof. The proof is beyond the scope of this book (see Notes).

The relationships in Lemma 3.31 and Lemma 3.30 completely fail when λ = 1.
But the latter situation indicates a full hash table, and we should avoid getting close
to it anyway.
Unlike OALP and OADH, the time estimates for separate chaining (SC) remain
valid with data removals. Because each chain may keep several table elements, the
load factor may be more than 1.
Table 3.6 presents the above theoretical estimates of the search time in the OALP,
OADH, and SC hash tables under different load factors. Average time measurements
 Table 3.6: Average search time bounds in hash tables with load factor λ.

λ Successful search: Tss (λ) Unsuccessful search: Tus (λ)

SC OALP OADH SC OALP OADH
0.10 1.05 1.06 1.05 1.10 1.12 1.11
0.25 1.12 1.17 1.15 1.25 1.39 1.33
0.50 1.25 1.50 1.39 1.50 2.50 2.0
0.75 1.37 2.50 1.85 1.75 8.50 4.0
0.90 1.45 5.50 2.56 1.90 50.5 10.0
0.99 1.49 50.5 4.65 1.99 5000.0 100.0

for actual hash tables [12] are close to the estimates for SC tables in the whole range
λ ≤ 0.99 and seem to be valid for larger values of λ, too. The measurements for OADH
tables remain also close to the estimates up to λ = 0.99. But for OALP tables, the
measured time is considerably less than the estimates if λ > 0.90 for a successful
search and λ > 0.75 for an unsuccessful search.
Example 3.32. The expected performance of hashing depends only on the load fac-
tor. If λ = 0.9, OADH double hashing takes on the average 2.56 and 10 probes for
successful and unsuccessful search, respectively. But if λ = 0.5, that is, the same keys
are stored in a roughly twice larger table, the same numbers decrease to 1.39 and 2
probes.
Implementation of hashing
R ESIZING One problem with open addressing is that successive insertions may cause
the table to become almost full, which degrades performance. Eventually we will
need to increase the table size. Doing this each time an element is inserted is very
inefficient. It is better to use an upper bound, say 0.75, on the load factor, and to
double the array size when this threshold is exceeded. This will then require recom-
puting the addresses of each element using a new hash function.
The total time required to resize, when growing a table from 0 to m = 2k elements,
is of order 1 + 2 + 4 + 8 + · · · + 2k−1 = 2k − 1 = m − 1. Since the m insertions take time
of order m (recall the table always has load factor bounded away from 1), the average
insertion time is still Θ(1).
D ELETION It is quite easy to delete a table entry from a hash table with separate
chaining (by mere node deletion from a linked list). However, open addressing en-
counters difficulties. If a particular table entry is physically removed from a OA-hash
table leaving an empty entry in that place, the search for subsequent keys becomes
invalid. This is because the OA-search terminates when the probe sequence en-
counters an empty table entry. Thus if a previously occupied entry is emptied, all
probe sequences that previously travelled through that entry will now terminate be-
fore reaching the right location.
 To avoid this problem, the deleted entry is normally marked in such a way that
insertion and search operations can treat it as an empty and nonempty location,
respectively. Unfortunately, such a policy results in hash tables packed with entries
which are marked as deleted. But in this case the table entries can be rehashed to
preserve only actual data and really delete all marked entries. In any case, the time
to delete a table entry remains O(1) both for SC and OA hash tables.
C HOOSING A HASH FUNCTION Ideally, the hash function, h(k), has to map keys uni-
formly and randomly onto the entire range of hash table addresses. Therefore, the
choice of this function has much in common with the choice of a generator of uni-
formly distributed pseudorandom numbers. A randomly chosen key k has to equiprob-
ably hash to each address so that uniformly distributed keys produce uniformly dis-
tributed indices h(k). A poorly designed hash function distributes table addresses
nonuniformly and tends to cluster indices for contiguous clusters of keys. A well
designed function scatters the keys as to avoid their clustering as much as possible.
If a set of keys is fixed, there always exists a perfect hash function that maps
the set one-to-one onto a set of table indices and thus entirely excludes collisions.
However, the problem is how to design such a function as it should be computed
quickly but without using large tables. There exist techniques to design perfect hash
functions for given sets of keys. But perfect hashing is of very limited interest be-
cause in most applications data sets are not static and the sets of keys cannot be
pre-determined.
Four basic methods for choosing a hash function are division, folding , middle-
squaring , and truncation.
Division assuming the table
  size is a prime number m and keys, k, are integers, the
quotient, q(k, m) = mk , and the remainder, r(k, m) = k mod m, of the integer
division of k by m specify the probe decrement for double hashing and the value
of the hash function h(k), respectively:
h(k) = r(k, m) and ∆(k) = max {1, q(k, m) mod m} .
The probe decrement is put to the range [1, . . . , m − 1] because all decrements
should be nonzero and point to the indices [0, 1, . . . , m − 1] of the table. The
reason that m should be prime is that otherwise some slots may be unreachable
by a probe sequence: for example if m = 12 and ∆(k) = 16, only 3 slots will be
probed before the sequence returns to the starting position.
Folding an integer key k is divided into sections and the value h(k) combines sums,
differences, and products of the sections (for example, a 9-digit decimal key,
such as k = 013402122, can be split into three sections: 013, 402, and 122, to be
added together for getting the value h(k) = 537 in the range [0, . . . , 2997]).
Middle-squaring a middle section of an integer key k, is selected and squared, then
a middle section of the result is the value h(k) (for example, the squared middle
 section, 402, of the above 9-digit key, k = 013402122, results in 161604, and the
middle four digits give the value h(k) = 6160 in the range [0, . . . , 9999]).
Truncation parts of a key are simply cut out and the remaining digits, or bits, or
characters are used as the value h(k) (for example, deletion of all but last three
digits in the above 9-digit key, k = 013402122, gives the value h(k) = 122 in the
range [0, . . . , 999]). While truncation is extremely fast, the keys do not scatter
randomly and uniformly over the hash table indices. This is why truncation is
used together with other methods, but rarely by itself.
Many real-world hash functions combine some of the above methods.
We conclude by discussing the idea of universal hashing. We have seen (in the
section on quicksort) the idea of using randomization to protect against bad worst-
case behaviour. An analogous idea works for hashing.
If a hash table is dynamically changing and its elements are not known in ad-
vance, any fixed hash function can result in very poor performance on certain in-
puts, because of collisions. Universal hashing allows us to reduce the probability of
this occurrence by randomly selecting the hash function at run time from a large
set of such functions. Each selected function still may be bad for a particular input,
but with a low probability which remains equally low if the same input is met once
again. Due to its internal randomisation, universal hashing behaves well even for
totally nonrandom inputs.
Definition 3.33. Let K, m, and F denote a set of keys, a size of a hash table (the range
of indices), and a set of hash functions mapping K to {0, . . . , m−1}, respectively. Then
F is a universal class if any distinct pair of keys k, κ ∈ K collide for no more than m1 of
the functions in the class F, that is,
1   1
 {h ∈ F | h(k) = h(κ)} ≤ .

|F| m
Thus in the universal class all key pairs behave well and the random selection of
a function from the class results in a probability of at most m1 that any pair collide.
One popular universal class of hashing functions is produced by a simple division
method. It assumes the keys are integers and cardinality of the set of keys K is a
prime number larger than the largest actual key. The size m of the hash table can be
arbitrary. This universal class is described by the next theorem.
Theorem 3.34 (Universal Class of Hash Functions). Let K = {0, . . . , p − 1} and |K| = p
be a prime number. For any pair of integers a ∈ {1, . . . , p − 1} and b ∈ {0, . . . , p − 1}, let
ha,b (k) = ((ak + b) mod p) mod m. Then

F = {ha,b | 1 ≤ a < p and 0 ≤ b < p}

is a universal class.
Proof. It is easily shown that the number of collisions in the class F,
 
{h ∈ F | h(k) = h(κ); k, κ ∈ K} ,
 

is the number of distinct numbers (x, y); 0 ≤ x, y < p such that x mod m = y mod m. Let
us denote the latter property: x ≡ y (mod m). It is evident that ha,b (k) = ha,b (κ) iff

(ak + b) mod p ≡ (aκ + b) mod p (mod m) .

Then for any fixed k, κ < p, there is one-to-one correspondence between the pairs
(a, b) such that 0 < a < p and 0 < b < p and ha,b (k) = ha,b (κ), and the pairs of distinct
numbers (x, y) with the property that 0 ≤ x, y < p and x ≡ y (mod n). The correspon-
dence is given in one direction by

x = (ak + b) mod p; y = (aκ + b) mod p

where x 6= y since
{az + b | z = 0, . . . , p − 1} = {0, . . . , p − 1}
when p is prime and a 6= 0. In the other direction the correspondence is given by the
condition that a and b are the unique integers in {0, . . . , p − 1} such that

ak + b ≡ x (mod p) and aκ + b ≡ y (mod p) .

These equations have a unique solution for a and b since p is prime, and a 6= 0 since
x 6= y.
Clearly |F| = p(p − 1). Now let us find out how many pairs of distinct numbers
(x, y) exist
 such that 0 ≤ x, y < p and x ≡ y (mod m). For any fixed s < m there are
at most mp numbers x < p such that x ≡ s (mod m). Since p and m are integers,

 p  p−1 p p−1
m ≤ m + 1. Therefore for each x < p there are no more than m − 1 ≤ m numbers
y < p distinct from x such that x ≡ y (mod m), and the total number of such pairs (x, y)
is at most p(p−1)
m . Hence for any fixed distinct pair of keys (k, κ) the fraction of F that
cause k and κ to collide is at most m1 , so the class F is universal.

This suggests the following strategy for choosing a hash function at run time: (i)
find the current size of the set of keys to hash; (ii) select the next prime number p
larger than the size of the key set found; (iii) randomly choose integers a and b such
that 0 < a < p and 0 ≤ b < p, and (iv) use the function ha,b defined in Theorem 3.34.
Exercises
Exercise 3.5.1. The Java programming language (as of time of writing) uses the fol-
lowing hash function h for character strings. Each character has a Unicode value
represented by an integer (for example, the upper case letters A, B, . . . , Z correspond
to 65, 66, . . . , 90 and the lower case a, b, . . . , z correspond to 97, 98, . . . , 122). Then h is
computed using 32-bit integer addition via

h(s) = s[0] ∗ 31n−1 + s[1] ∗ 31n−2 + · · · + s[n − 1] ∗ 31 + s[n].

Find two 2-letter strings that have the same hash value. How could you use this
to make 2100 different strings all of which have the same hash code?

Exercise 3.5.2. Place the sequence of keys k = 10, 26, 52, 76, 13, 8, 3, 33, 60, 42 into a
hash table of size 13 using the modulo-based hash address i = k mod 13 and linear
probing to resolve collisions.

Exercise 3.5.3. Place the sequence of keys k = 10, 26, 52, 76, 13, 8, 3, 33, 60, 42 into a
hash table of size 13 using the modulo-based hash address i = k mod 13 and double
hashing with the secondary hash function ∆(k) = max{1, k/13} to resolve collisions.

Exercise 3.5.4. Place the sequence of keys k = 10, 26, 52, 76, 13, 8, 3, 33, 60, 42 into a
hash table of size 13 using separate chaining to resolve collisions.

3.6 Notes
Binary search, while apparently simple, is notoriously hard to program correctly
even for professional programmers: see [2] for details.
The expected height of a randomly grown BST was shown to be Θ(log n) by J. M.
Robson in 1979. After much work by many authors it is now known that the average
value is tightly concentrated around α ln n where α is the root of x ln(2e/x) = 1, α ∼
=
4.311.
The historically first balanced binary search tree was proposed in 1962 by G. M.
Adelson-Velskii and E. M. Landis, hence the name AVL tree. Red-black trees were
developed in 1972 by R. Bayer under the name “symmetric binary B-trees” and re-
ceived their present name and definition from L. Guibas and R. Sedgewick in 1978.
AA-trees were proposed by A. Anderson in 1993.
Multiway B-trees were proposed in 1972 by R. Bayer and E. McCreight.
According to D. Knuth, hashing was invented at IBM in early 1950’s simultane-
ously and independently by H. P. Luhn (hash tables with SC) and G. M. Amdahl
(OALP).
The analysis of OALP hashing was first performed by D. Knuth in 1962. This was
the beginning of the modern research field of analysis of algorithms.
The random balls in bins model can be analysed in detail by more advanced
methods than we present in this book (see for example [6]). Some natural questions
are

• When do we expect all bins to have at least one ball?

• What proportion of boxes are expected to be empty when n ≈ m?

 • What is the expected maximum number of balls in a box when n ≈ m?

The answers are applicable to our analysis of chaining: when are all chains expected
to be nonempty? how many chains are empty when the average chain length is Θ(1)?
what is the maximum chain length when the average chain length is Θ(1)? The an-
swers are known to be, respectively: when n ≈ m ln m; about e−λ ; Θ(log n/ log log n).
The last result is much harder to derive than the other two.
 Part II

Introduction to Graph Algorithms

Chapter 4

The Graph Abstract Data Type

Many real-world applications, such as building compilers, finding routing proto-

cols in communication networks, and scheduling processes, are well modelled using
concepts from graph theory. A (directed) graph is an abstract mathematical model:
a set of points and connection relations between them.

4.1 Basic definitions

Graphs can be studied from a purely mathematical point of view (“does some-
thing exist, can something be done?”). In this book we focus on algorithmic aspects
of graph theory (“how do we do it efficiently and systematically?”). However, the
mathematical side is where we must start, with precise definitions. The intuition
behind the definitions is not hard to guess.
We start with the concept of digraph (the word stands for directed graph). A good
example to keep in mind is a street network with one-way streets.

Definition 4.1. A digraph G = (V, E) is a finite nonempty set V of nodes together

with a (possibly empty) set E of ordered pairs of nodes of G called arcs.

Note. In the mathematical language of relations, the definition says that E is a rela-
tion on V . If (u, v) ∈ E, we say that v is adjacent to u, that v is an out-neighbour of u,
and that u is an in-neighbour of v.
We can think of a node as being a point and an arc as an arrow from one node
to another. This allows us to draw pictures that suggest ideas. The pictures cannot
prove anything, however.
 G1 0 G2 0

1 2 1 2

3 4 3 4

Figure 4.1: A graph G1 and a digraph G2 .

Very often the adjacency relation is symmetric (all streets are two-way). There
are two ways to deal with this. We can use a digraph that happens to be symmetric
(in other words, (u, v) is an arc if and only if (v, u) is an arc). However, it is sometimes
simpler to reduce this pair of arcs into a single undirected edge that can be traversed
in either direction.
Definition 4.2. A graph G = (V, E) is a finite nonempty set V of vertices together with
a (possibly empty) set E of unordered pairs of vertices of G called edges.
Note. Since we defined E to be a set, there are no multiple arcs/edges between a
given pair of nodes/vertices.
Non-fluent speakers of English please note: the singular of “vertices” is not “ver-
tice”, but “vertex”.
For a given digraph G we may also denote the set of nodes by V (G) and the set of
arcs by E(G) to lessen any ambiguity.
Example 4.3. We display a graph G1 and a digraph G2 in Figure 4.1. The nodes/vertices
are labelled 0, 1, . . . as in the picture. The arcs and edges are as follows.

E(G1 ) = {{0, 1}, {0, 2}, {1, 2}, {2, 3}, {2, 4}, {3, 4}}

E(G2 ) = {(0, 2), (1, 0), (1, 2), (1, 3), (3, 1), (3, 4), (4, 2)}

Note. Some people like to view a graph as a special type of digraph where every
unordered edge {u, v} is replaced by two directed arcs (u, v) and (v, u). This has the
advantage of allowing us to consider only digraphs, and we shall use this approach
in our Java implementation in Appendix B. It works in most instances.
However, there are disadvantages; for some purposes we must know whether our
object is really a graph or just a symmetric digraph. Whenever there is (in our opin-
ion) a potential ambiguity, we shall point it out.
Example 4.4. Every rooted tree (see Section D.7) can be interpreted as a digraph:
there is an arc from each node to each of its children.
Every free tree is a graph of a very special type (see Appendix D.7).

Note. (Graph terminology) The terminology in this subject is unfortunately not com-
pletely standard. Some authors call a graph by the longer term “undirected graph”
and use the term “graph” to mean what we call a directed graph. However when us-
ing our definition of a graph, it is standard practice to abbreviate the phrase “directed
graph” with the word digraph.
We shall be dealing with both graphs and digraphs throughout these notes. In
order to save writing “(di)graph” too many times, we make the following conven-
tion. We treat the digraph as the fundamental concept. In other words, we shall use
the terminology of digraphs, nodes and arcs, with the understanding that if this is
changed to graphs, edges, and vertices, the resulting statement is still true. However,
if we talk about graphs, edges, and vertices, our statement is not necessarily true for
digraphs. Whenever a result is true for digraphs but not for graphs, we shall say this
explicitly (this happens very rarely).
There is another convention to discuss. An arc that begins and ends at the same
node is called a loop. We make the convention that loops are not allowed in our di-
graphs. Again, other authors may differ. If our conventions are relaxed to allow mul-
tiple arcs and/or loops, many of the algorithms below work with no modification or
with only very minor modification required. However dealing with loops frequently
requires special cases to be considered, and would distract us from our main goal
of introducing the field of graph algorithms. As an example of the problems caused
by loops, suppose that we represent a graph as a symmetric digraph as described
above. How do we represent a loop in the graph?

Definition 4.5. The order of a digraph G = (V, E) is |V |, the number of nodes. The
size of G is |E|, the number of arcs.

We usually use n to denote |V | and m to denote |E|.

For a given n, the value of m can be as low as 0 (a digraph consisting of n totally
disconnected points) and as high as n(n−1) (each node can point to each other node;
recall that we do not allow loops). If m is toward the low end, the digraph is called
sparse, and if m is toward the high end, then the digraph is called dense. These terms
are obviously very informal. For our purposes we will call a class of digraphs sparse
if m is O(n) and dense if m is Ω(n2 ).

Definition 4.6. A walk in a digraph G is a sequence of nodes v0 v1 . . . vl such that, for

each i with 0 ≤ i < l, (vi , vi+1 ) is an arc in G. The length of the walk v0 v1 . . . vl is the
number l (that is, the number of arcs involved).
A path is a walk in which no node is repeated. A cycle is a walk in which v0 = vl
and no other nodes are repeated.
 Thus in a graph, we rule out a walk of the form u v u as a cycle (going back and
forth along the same edge should not count as a cycle). A cycle in a graph must be of
length at least 3.
It is easy to see that if there is a walk from u to v, then (by omitting some steps if
necessary) we can find a path from u to v.

Example 4.7. For the graph G1 of Figure 4.1 the following sequences of vertices are
classified as being walks, paths, or cycles.

vertex sequence walk? path? cycle?

032 no no no
01234 yes yes no
0120 yes no yes
123420 yes no no
010 yes no no

Example 4.8. For the digraph G2 of Figure 4.1 the following sequences of nodes are
classified as being walks, paths, or cycles.

node sequence walk? path? cycle?

01234 no no no
312 yes yes no
131 yes no yes
1312 yes no no
024 no no no

Definition 4.9. In a graph, the degree of a vertex v is the number of edges meeting
v. In a digraph, the outdegree of a node v is the number of out-neighbours of v, and
the indegree of v is the number of in-neighbours of v.
A node of indegree 0 is called a source and a node of outdegree 0 is called a sink.

If the nodes have a natural order, we may simply list the indegrees or outdegrees
in a sequence.

Example 4.10. For our graph G1 , the degree sequence is (2, 2, 4, 2, 2). The in-degree
sequence and out-degree sequence of the digraph G2 are (1, 1, 3, 1, 1) and (1, 3, 0, 2, 1),
respectively. Node 2 is a sink.

Definition 4.11. The distance from u to v in G, denoted by d(u, v), is the minimum
length of a path from u to v. If no path exists, the distance is undefined (or +∞).

For graphs, we have d(u, v) = d(v, u) for all vertices u, v.

Example 4.12. In graph G1 of Figure 4.1, we can see by considering all possibilities
that d(0, 1) = 1, d(0, 2) = 1, d(0, 3) = 2, d(0, 4) = 2, d(1, 2) = 1, d(1, 3) = 2, d(1, 4) = 2,
d(2, 3) = 1, d(2, 4) = 1 and d(3, 4) = 1.
In digraph G2 , we have, for example, d(0, 2) = 1, d(3, 2) = 2. Since node 2 is a sink,
d(2, v) is not defined unless v = 2, in which case the value is 0.

There are several ways to create new digraphs from old ones.
One way is to delete (possibly zero) nodes and arcs in such a way that the result-
ing object is still a digraph (there are no arcs missing any endpoints!).

Definition 4.13. A subdigraph of a digraph G = (V, E) is a digraph G′ = (V ′ , E ′ ) where

V ′ ⊆ V and E ′ ⊆ E. A spanning subdigraph is one with V ′ = V ; that is, it contains all
nodes.

Example 4.14. Figure 4.2 shows (on the left) a subdigraph and (on the right) a span-
ning subdigraph of the digraph G2 of Figure 4.1.

1 2 1 2

3 3 4

Figure 4.2: A subdigraph and a spanning subdigraph of G2 .

Definition 4.15. The subdigraph induced by a subset V ′ of V is the digraph G′ =

(V ′ , E ′ ) where E ′ = {(u, v) ∈ E | u ∈ V ′ and v ∈ V ′ }.

Example 4.16. Figure 4.3 shows the subdigraph of the digraph G2 of Figure 4.1 in-
duced by {1, 2, 3}.

We shall sometimes find it useful to “reverse all the arrows”.

Definition 4.17. The reverse digraph of the digraph G = (V, E), is the digraph Gr =
(V, E ′ ) where (u, v) ∈ E ′ if and only if (v, u) ∈ E.

Example 4.18. Figure 4.4 shows the reverse of the digraph G2 of Figure 4.1.
 1 2

Figure 4.3: The subdigraph of G2 induced by {1, 2, 3}.

1 2

3 4

Figure 4.4: The reverse of digraph G2 .

It is sometimes useful to replace all one-way streets with two-way streets. The
formal definition must take care not to introduce multiple edges. Note below that if
(u, v) and (v, u) belong to E, then only one edge joins u and v in G′ . This is because
{u, v} and {v, u} are equal as sets, so appear only once in the set E ′ .

Definition 4.19. The underlying graph of a digraph G = (V, E) is the graph G′ =

(V, E ′ ) where E ′ = {{u, v} | (u, v) ∈ E}.

Example 4.20. Figure 4.5 shows the underlying graph of the digraph G2 of Figure 4.1.

We may need to combine two or more digraphs G1 , G2 , . . . Gk into a single graph

where the vertices of each Gi are completely disjoint from each other and no arc goes
between the different Gi . The constructed graph G is called the graph union, where
V (G) = V (G1 ) ∪V (G2 ) ∪ . . . ∪V (Gk ) and E(G) = E(G1 ) ∪ E(G2 ) ∪ . . . ∪ E(Gk ).
Exercises
Exercise 4.1.1. Prove that in a digraph, the sum of all outdegrees equals the sum of
all indegrees. What is the analogous statement for a graph?
 0

1 2

3 4

Figure 4.5: The underlying graph of G2 .

Exercise 4.1.2. Let G be a digraph of order n and u, v nodes of G. Show that d(u, v) ≤
n − 1 if there is a walk from u to v.

Exercise 4.1.3. Prove that in a sparse digraph, the average indegree of a node is O(1),
while in a dense digraph, the average indegree of a node is Ω(n).

4.2 Digraphs and data structures

In order to process digraphs by computer we first need to consider how to repre-
sent them in terms of data structures. There are two common computer representa-
tions for digraphs, which we now present. We assume that the digraph has the nodes
given in a fixed order. Our convention is that the vertices are labelled 0, 1, . . . , n − 1.

Definition 4.21. Let G be a digraph of order n. The adjacency matrix of G is the n × n

boolean matrix (often encoded with 0’s and 1’s) such that entry (i, j) is true if and
only if there is an arc from the node i to node j.

Definition 4.22. For a digraph G of order n, an adjacency lists representation is a

sequence of n sequences, L0 , . . . , Ln−1 . Sequence Li contains all nodes of G that are
adjacent to node i.

The sequence Li may or may not be sorted in order of increasing node number.
Our convention is to sort them whenever it is convenient. (However, many imple-
mentations, such as the one given in Appendix B, do not enforce that their adjacency
lists be sorted.)
We can see the structure of these representations more clearly with examples.
Example 4.23. For the graph G1 and digraph G2 of Example 4.3, the adjacency ma-
trices are given below.
   
0 1 1 0 0 0 0 1 0 0
1 0 1 0 0 1 0 1 1 0
   
1 1 0
G1 :  1 1
 0
G2 :  0 0 0 0
0 0 1 0 1 0 1 0 0 1
0 0 1 1 0 0 0 1 0 0

Notice that the number of 1’s in a row (column) is the outdegree (indegree) of the
corresponding node. The corresponding adjacency lists are now given.

1 2 2
0 2 0 2 3
G1 : 0 1 3 4 G2 :
2 4 1 4
2 3 2

Note. Only the out-neighbours are listed in the adjacency lists representation. An
empty sequence can occur (for example, sequence 2 of the digraph G2 ). If the nodes
are not numbered in the usual way (for example, they are numbered 0, . . . , n − 1 or
labelled A, B,C, . . . ), we may include these labels if necessary.
It is often useful to input several digraphs from a single file. Our standard format
is as follows. The file consists of several digraphs one after the other. To distinguish
the beginning of one and the end of the other we have a single line giving the order at
the beginning of each graph. If the order is n then the next n lines give the adjacency
matrix or adjacency lists representation of the digraph. The end of the file is marked
with a line denoting a digraph of order 0.

Example 4.24. Here is a file consisting of 3 digraphs, of orders 4, 3, 0 respectively.

The first contains a sink and hence there is a blank line.
 4
1 2
3

0 1 2
3
2
0
1
0

There are also other specialized (di)graph representations besides the two men-
tioned in this section. These data structures take advantage of special structure for
improved storage or access time, often for families of graphs sharing a common
property. For such specialized purposes they may be better than either the adja-
cency matrix or lists representations.
For example, trees can be stored more efficiently. We have already seen in Sec-
tion 2.5 how a complete binary tree can be stored in an array. A general rooted tree
of n nodes can be stored in an array pred of size n. The value pred[i] gives the parent
of node i. The root is a special case and can be given value −1 (representing a NULL
pointer), for example, if we number nodes from 0 to n − 1 in the usual way. This of
course is a form of adjacency lists representation, where we use in-neighbours in-
stead of out-neighbours.
We will sometimes need to represent ∞ when processing graphs. For example,
it may be more convenient to define d(u, v) = ∞ than to say it is undefined. From a
programming point of view, we can use any positive integer that can not be confused
with any other that might legitimately arise. For example, the distance between 2
nodes in a digraph on n nodes cannot be more than n − 1 (see Exercise 4.1.2). Thus
in this case we may use n to represent the fact that there is no path between a given
pair of nodes. We shall return to this subject in Chapter 6.
Exercises
Exercise 4.2.1. Write down the adjacency matrix of the digraph of order 7 whose ad-
jacency lists representation is given below.
 2
0
0 1
4 5 6
5
3 4 6
1 2

Exercise 4.2.2. Consider the digraph G of order 7 whose adjacency matrix represen-
tation is given below.
 
0 1 0 0 1 1 0
1 0 0 1 0 0 0
 
1 0 0 0 0 0 1
 
1 0 0 0 0 1 0
 
0 0 0 0 0 1 0
 
0 0 0 0 0 0 0
0 0 0 0 0 1 0
Write down the adjacency lists representation of G.

Exercise 4.2.3. Consider the digraph G of order 7 given by the following adjacency
lists representation.
2
0
0 1
4 5 6
5
3 4 6
1 2

Write down the adjacency matrix representation of the reverse digraph Gr .

Exercise 4.2.4. Consider the digraph G whose nodes are the integers from 1 to 12
inclusive and such that (i, j) is an arc if and only if i is a proper divisor of j (that is, i
divides j and i 6= j).
Write down the adjacency matrix representation of G and of Gr .

Exercise 4.2.5. Write the adjacency lists and adjacency matrix representation for a
complete binary tree with 7 vertices, assuming they are ordered 1, . . . , 7 as in Sec-
tion 2.5.
 Table 4.1: Digraph operations in terms of data structures.

Operation Adjacency Matrix Adjacency Lists

arc (i, j) exists? is entry (i, j) 0 or 1 find j in list i
outdegree of i scan row, count 1’s size of list i
indegree of i scan column, count 1’s for j 6= i, find i in list j
add arc (i, j) change entry (i, j) insert j in list i
delete arc (i, j) change entry (i, j) delete j from list i
add node create new row and column add new list at end
delete node i delete row/column i delete list i
shuffle other entries for j 6= i, delete i from list j

4.3 Implementation of digraph ADT operations

In this section we discuss the implementation of basic digraph operations we
have seen in Section 4.1.
A matrix is simply an array of arrays. The lists representation is really a list of lists,
but a list can be implemented in several ways, for example by an array or singly- or
doubly-linked list using pointers. These have different properties (see Section C.1);
for example, accessing the middle element is Θ(1) for an array but Θ(n) for a linked
list. In any case, however, to find a value that may or may not be in the list requires
sequential search and takes Θ(n) time in the worst case.
We now discuss the comparative performance of some basic operations using
the different data structures. In Table 4.1 we show how the basic graph operations
can be described in terms of the adjacency matrix or lists representations.
For example, suppose we wish to check whether arc (i, j) exists. Using the adja-
cency matrix representation, we are simply accessing an array element twice. How-
ever, with the lists representation, we will need to find j in list i. Adding a node in
the lists case is easy, since we just add an empty list at the end, but adding a node
with the matrix representation requires us to allocate an extra row and column of
zeros. Deleting a node (which perhaps necessitates deleting some arcs) is trickier. In
the matrix case, we must delete a row and column, and move up some elements so
there are no gaps in the matrix. In the lists case, we must remove a list and also all
references to the deleted node in other lists. This requires scanning each list for the
offending entry and deleting it.
In Table 4.2 we compare the performance of two different data structures. There
are several ways to implement a list, and hence the square of that number of ways
to implement an adjacency lists representation. We have chosen one, an array of
doubly linked lists, for concreteness.
For example, finding j in list i will take time in the worst case Θ(d), where d is the
 Table 4.2: Comparative worst-case performance of adjacency lists and matrices.

Operation matrix lists

arc (i, j) exists? Θ(1) Θ(d)
outdegree of i Θ(n) Θ(1)
indegree of i Θ(n) Θ(n + m)
add arc (i, j) Θ(1) Θ(1)
delete arc (i, j) Θ(1) Θ(d)
add node Θ(n) Θ(1)
delete node i Θ(n2 ) Θ(n + m)

size of list i, which equals the outdegree of i. In the worst case this might be Θ(m)
since all nodes but i might be sinks. On the other hand, for a sparse digraph, the
average outdegree is Θ(1), so arc lookup can be done on average in constant time.
Note that if we want to print out all arcs of a digraph, this will take time Θ(n + m) in
the lists case and Θ(n2 ) in the matrix case.
Finding outdegree with the lists representation merely requires accessing the
correct list (constant time) plus finding the size of that list (constant time). Find-
ing indegree with the lists representation requires scanning all lists except one, and
this requires us to look at every arc in the worst case, taking time Θ(n + m) (the n is
because we must consider every node’s list even if it is empty). If we wish to compute
just one indegree, this might be acceptable, but if all indegrees are required, this will
be inefficient. It is better to compute the reverse digraph once and then read off the
outdegrees, this last step taking time Θ(n) (see Exercise 4.3.1).
One way around all this work is to use in our definition of adjacency lists repre-
sentation, instead of just the out-neighbours, a list of in-neighbours also. This may
be useful in some contexts but in general requires more space than is needed.
We conclude by discussing space requirements. The adjacency matrix represen-
tation requires Θ(n2 ) storage: we simply need n2 bits. It appears that an adjacency
lists representation requires Θ(n + m) storage, since we must store an endpoint of
each arc, and we need to allocate space for each node’s list. However this is not
strictly true for large graphs. Each node number requires some storage; the number
k requires on average Θ(log k) bits. If, for example, we have a digraph on n nodes
where every possible arc occurs, then the total storage required is of order n2 log n,
worse than with a matrix representation. For small, sparse digraphs, it is true that
lists use less space than a matrix, whereas for small dense digraphs the space re-
quirements are comparable. For large sparse digraphs, a matrix can still be more
efficient, but this happens rarely.
The remarks above show that it is not immediately clear which representation to
use. We will mostly use adjacency lists, which are clearly superior for many common
tasks (such as graph traversals, covered in Chapter 5) and generally better for sparse
digraphs.
Any implementation of an abstract data type (for example as a Java class) must
include objects and “methods”. While most people would include methods for adding
nodes, deleting arcs, and so on, it is not clear where to draw the line. In Appendix B,
one way of writing Java classes to deal with graphs is presented in detail. There are
obviously a lot of different choices one can make. In particular for our Java lists rep-
resentation we use ArrayList<ArrayList<Integer>>.

Exercises
Exercise 4.3.1. Show how to compute the (sorted) adjacency lists representation of
the reverse digraph of G from the (sorted) adjacency lists representation of G itself.
It should take time Θ(n + m).

4.4 Notes
This chapter shows how to represent and process graphs with a computer. Al-
though Appendix B uses the Java programming language, the ideas and algorithms
are applicable to other industrial programming languages. For example, C++ has
several standard graph algorithms libraries such as the Boost Graph Library [11],
LEDA (Library of Efficient Data structures and Algorithms) [9] and GTL (Graph Tem-
plate Library) [3]. All algorithms discussed here are provided in these libraries and
in other mathematical interpreted languages like Mathematica and Maple.
Chapter 5

Graph Traversals and Applications

Many graph problems require us to visit each node of a digraph in a systematic way.
For example, we may want to print out the labels of the nodes in some order, or per-
haps we are in a maze and we have no idea where to find the door. More interesting
examples will be described below. The important requirements are that we must
be systematic (otherwise an algorithm is hard to implement), we must be complete
(visit each node at least once), and we must be efficient (visit each node at most
once).
There are several ways to perform such a traversal. Here we present the two
most common, namely breadth-first search and depth-first search. We also discuss a
more general but also more complicated and slower algorithm, priority-first search.
First we start with general remarks applicable to all graph traversals.

5.1 Generalities on graph traversal

We first discuss a general skeleton common to all graph traversals. In Figure 5.1,
this general skeleton is presented in pseudocode. We first describe it less formally.
Suppose that G is a digraph. Choose a node v as a starting point. We shall visit all
nodes reachable from v in G using an obvious method.
At each stage, there are three possible types of nodes. White nodes are those
that have not yet been visited. Grey nodes (or frontier nodes) are those that have
been visited but may have adjacent nodes that are white. Finally, black nodes are
those nodes that have been visited, along with all their adjacent nodes. This is shown
pictorially in Figure 5.2.
Initially, all nodes are white. The traversal algorithm first colours v grey. At each
subsequent step, it chooses a grey node w and then a white node x adjacent to w
algorithm traverse
Input: digraph G
begin
array colour[0..n − 1], pred[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE
end for
for s ∈ V (G) do
if colour[s] = WHITE then
visit(s)
end if
end for
return pred
end

algorithm visit
Input: node s of digraph G
begin
colour[s] ← GREY; pred[s] ← NULL
while there is a grey node do
choose a grey node u
if there is a white neighbour of u
choose such a neighbour v
colour[v] ← GREY; pred[v] ← u
else colour[u] ← BLACK
end if
end while
end

Figure 5.1: Graph traversal schema.

Figure 5.2: Node states in the middle of a digraph traversal.

(thereby changing the colour of x to grey). This possibly causes some grey nodes
to become black. When all nodes are black, there can be no white nodes reach-
able from the current set of visited nodes. We have reached every node that can be
reached from the root, and the traversal from v terminates.
At this stage we have created a subdigraph of G that is a tree: the nodes are pre-
cisely the black nodes, and the arcs are those arcs connecting a grey node to its white
neighbour at the instant before the white node turns grey.
There may remain unvisited nodes in the digraph. In this case, we may choose
another white node as root and continue. Eventually, we obtain a set of disjoint trees
spanning the digraph, which we call the search forest .
The above is formalized in Figure 5.1. Here traverse simply chooses a new root
s when required. The main work is done by visit. In this procedure, the array pred
stores the parent of each node; each root gets the value NULL, since it has no par-
ent. Each time through the while-loop, a white node is turned grey or a grey node
is turned black. Thus eventually there are no white nodes reachable from s and the
procedure terminates. If a node u is reachable from s then it is visited during the call
to visit with input s, unless it had already been visited in a previous tree.
Now once a traversal has been completed and a search forest obtained, we may
classify the arcs into four distinct types. This classification will be useful later.

Definition 5.1. Suppose we have performed a traversal of a digraph G, resulting in a

search forest F. Let (u, v) ∈ E(G) be an arc. It is called a tree arc if it belongs to one of
the trees of F. If the arc is not a tree arc, there are three possibilities. A non-tree arc
is called a forward arc if u is an ancestor of v in F, a back arc if u is a descendant of v
in F, and a cross arc if neither u nor v is an ancestor of the other in F.

In the first three cases in Definition 5.1, u and v must belong to the same tree of
F. However, a cross arc may join two nodes in the same tree or point from one tree
to another.
The following theorem collects all the basic facts we need for proofs in later sec-
tions. Figure 5.3 illustrates the first part.

Theorem 5.2. Suppose that we have carried out traverse on G, resulting in a search
forest F. Let v, w ∈ V (G).

• Let T1 and T2 be different trees in F and suppose that T1 was explored before T2 .
Then there are no arcs from T1 to T2 .

• Suppose that G is a graph. Then there can be no edges joining different trees
of F.

• Suppose that v is visited before w and w is reachable from v in G. Then v and w

belong to the same tree of F.
 • Suppose that v and w belong to the same tree T in F. Then any path from v to
w in G must have all nodes in T .

Proof. If the first part were not true, then since w is reachable from v, and w has not
been visited before T1 is started, w must be reached in the generation of T1 , contra-
dicting w ∈ T2 . The second part follows immediately for symmetric digraphs and
hence for graphs. Now suppose that v is seen before w. Let r be the root of the tree T
containing v. Then w is reachable from r and so since it has not already been visited
when r is chosen, it belongs to T . Finally, if v and w are in the same tree, then any
path from v to w in G going outside the tree must re-enter it via some arc; either the
leaving or the entering arc will contradict the first part.

T1 T2 Tk

Figure 5.3: Decomposition of a digraph in terms of search trees.

We now turn to the analysis of traverse. The generality of our traversal proce-
dure makes its complexity hard to determine. Its running time is very dependent on
how one chooses the next grey node u and its white neighbour v. It also apparently
depends on how long it takes to determine whether there exist any grey nodes or
whether u has any white neighbours. However, any sensible rule for checking exis-
tence of either type of node should simply return false if there is no such node, and
take no more time in this case than if it does find one. Thus we do not need to take
account of the checking in our analysis.
Since the initialization of the array colour takes time Θ(n), the amount of time
taken by traverse is clearly Θ(n + t), where t is the total time taken by all the calls to
visit.
Each time through the while-loop of visit a grey node is chosen, and either a
white node is turned grey or a grey node is turned black. Note that the same grey
node can be chosen many times. Thus we execute the while-loop in total Θ(n) times
since every node must eventually move from white through grey to black. Let a, A be
lower and upper bounds on the time taken to choose a grey node (note that they may
depend on n and be quite large if the rule used is not very simple). Then the time
taken in choosing grey nodes is O(An) and Ω(an). Now consider the time taken to
find a white neighbour. This will involve examining each neighbour of u and check-
ing whether it is white, then applying a selection rule. If the time taken to apply the
rule is at least b and at most B (which may depend on n), then the total time in choos-
ing white neighbours is O(Bm) and Ω(bm) if adjacency lists are used and O(Bn2 ) and
Ω(bn2 ) if an adjacency matrix is used.
In summary, then, the running time of traverse is O(An + Bm) and Ω(an + bm) if
adjacency lists are used, and O(An + Bn2 ) and Ω(an + bn2 ) if adjacency matrix format
is used.
A more detailed analysis would depend on the rule used. We shall see in Sec-
tion 5.2 that BFS and DFS have a, b, A, B all constant, and so each yields a linear-time
traversal algorithm. In this case, assuming a sparse input digraph, the adjacency list
format seems preferable. On the other hand, for example, suppose that a is at least
of order n (a rather complex rule for grey nodes is being used) and b, B are constant
(for example, the first white node found is chosen). Then asymptotically both repre-
sentations take time Ω(n2 ), so using the adjacency matrix is not clearly ruled out (it
may even be preferable if it makes programming easier).
Exercises
Exercise 5.1.1. Draw a moderately complicated graph representing a maze (corri-
dors are edges and intersections are nodes). Label one node as the start and another
as the end. One rule for getting through a maze is to try to go forward, always make
a right turn when faced with a choice of direction, and back up as little as possible
when faced with a dead end. Apply this method to your example. Interpret what you
do in terms of the procedure traverse.

Exercise 5.1.2. Suppose that in traverse, the grey node is chosen at random and so
is the white node. Find your way through your maze of the previous exercise using
this method.

Exercise 5.1.3. Give an example for each of the following items.

• A search forest in which a cross arc points from one tree to another.
• A search forest in which a cross arc joins two nodes in the same tree.

5.2 DFS and BFS

So far everything has been discussed at a very general level. To proceed fur-
ther we need more analysis of how to guide the traversal: what rule do we use for
choosing the next grey and next white node? Different rules lead to very different
results. The two main rules are breadth-first search (BFS) and depth-first search
(DFS) which we discuss now. We shall also discuss the more complicated priority-
first search (PFS) in Section 5.5.
Breadth-first and depth-first search are dual to each other. In BFS, when we want
to visit a new node, the new grey node chosen is the one that has been grey for the
longest time. By contrast, in DFS, we choose the one that has been grey for the short-
est time.
In BFS we start at a node v and then go to each neighbour of v (in some order),
then each neighbour of a neighbour of v that has not already been visited, and so on.
The search chooses grey nodes across the entire frontier between the white and grey
nodes, and takes us away from the root node as slowly as possible.
By contrast, in DFS we start at a node v, but this time we “deeply” search as far
away from vertex v as possible, until we cannot visit any new nodes, upon which we
backtrack as little as possible. The search keeps us away from the root as long as
possible.
These search concepts are best illustrated with some examples.

Example 5.3. A graph G1 and a digraph G2 are displayed in Figure 5.4.

Breadth-first search trees (originating from node 0) of the graph G1 and digraph
G2 are displayed in Figure 5.5. The dashed arcs indicate the original arcs that are not
part of the BFS trees.
Note that even with DFS or BFS specified there still remains the choice of which
white neighbour of the chosen grey node to visit. While it does not matter what
choice is made, the choice should be algorithmic. Our convention is to choose the
one with lowest index (the nodes being numbered 0, . . . , n − 1).
This convention for choosing white nodes means that we can reconstruct the
progress of the BFS traversal completely. For example, in G1 , node 0 is visited first,
then node 1, node 2, node 3, node 4, node 8, and so on. Thus we can classify the
edges: for example {1, 2} is a cross edge, {2, 8} a tree edge, and so on.
We also display depth-first search trees (originating from node 0) of the graph G1
and digraph G2 in Figure 5.6. Again, the dashed arcs indicate the original arcs that
are not part of the DFS trees.
Here, for example, in G2 we see that the order of visiting nodes is 0, 1, 6, 2, 4, 5, 3.
The arc (5, 0) is a back arc, (3, 5) is a cross arc, there are no forward arcs, etc.

In the examples above, all nodes were reachable from the root, and so there is a
single search tree in the forest. For general digraphs, this may not be true. We should
distinguish between BFS or DFS originating from a given node, or just BFS/DFS run
on a digraph. The former we call BFSvisit/DFSvisit (a special case of visit) and
the latter just BFS/DFS (a special case of traverse). The algorithm DFS, for example,
repeatedly chooses a root and runs DFSvisit from that root, until all nodes have
been visited. Pseudocode for these procedures is given in the next two sections.
 G1 0 G2
0 1

1 2
2 3

3 4

4 5 6
8
5 6 7

Figure 5.4: A graph G1 and a digraph G2 .

1
G1 0 G2 1 2
0 1

2 3
1 2 3
2 4 3

4
3 5 4
5
6 6
4 5 6
7 8 9 8 7
5 6 7

Figure 5.5: BFS trees for G1 and G2 , rooted at 0.

 G1 0 G2 0

1 1 2 3

2 3
6 4

4
5 6
5

Figure 5.6: DFS trees for G1 and G2 , rooted at 0.

Exercises
Exercise 5.2.1. Draw a graph for which DFS and BFS can visit nodes in the same
order. Then draw one for which they must visit nodes in the same order. Make your
examples as large as possible (maximize n + m).

Exercise 5.2.2. A technique called iterative deepening combines features of BFS

and DFS. Starting at a root, for each i = 0, 1, . . . we search all nodes at distance at
most i, using DFS (in other words, we limit the depth of the DFS to i). We do this for
each i until the whole graph is explored or we reach a preassigned limit on i.
Note that we visit nodes near the root many times, so this search is not as efficient
in terms of time. What is the advantage of this search technique when the graphs are
too big to hold in memory?

5.3 Additional properties of depth-first search

In DFS, the next grey node chosen is the last one to be coloured grey thus far. The
data structure for processing grey nodes in this “last in, first out” order is therefore a
stack (see Section C.1 if necessary for the basic facts about stacks). We may store the
nodes in a stack as they are discovered. So the general traverse and visit proce-
dures can be replaced by those in Figure 5.7. Note how the algorithm for visit has
been altered. We loop through the nodes adjacent to the chosen grey node u, and as
soon as we find a white one, we add it to the stack. We also introduce a variable time
and keep track of the time a node was turned grey, and the time it was turned black,
in the arrays seen, done. These will be useful later.
The complexity analysis of DFS is easy. Choosing a grey node u takes constant
time since only the stack pop operation is required. The time taken to apply the
selection rule to a white neighbour is also constant, since we take the first one found
in the for-loop. Our analysis of traverse shows us that DFS runs in time Θ(n + m) if
adjacency lists are used, and Θ(n2 ) using an adjacency matrix. In summary, DFS is a
linear-time traversal algorithm.
One nice feature of depth-first search is its recursive nature. The relationship
between stacks and recursion is, we hope, well known to the reader. We can replace
the DFSvisit procedure in this case by the recursive version in Figure 5.8.
We now note a few important facts about depth-first search that are useful in
proving correctness of various algorithms later on.

Theorem 5.4. The call to recursiveDFSvisit with input s terminates only when all
nodes reachable from s via a path of white nodes have been visited. The descendants
of s in the DFS forest are precisely these nodes.

Proof. See Exercise 5.3.10.

There are not as many possibilities for interleaving of the timestamps as there ap-
pear at first sight. In particular, we cannot have seen[v] < seen[w] < done[v] < done[w].
The following theorem explains why.

Theorem 5.5. Suppose that we have performed DFS on a digraph G, resulting in a

search forest F. Let v, w ∈ V (G) and suppose that seen[v] < seen[w].

• If v is an ancestor of w in F, then

seen[v] < seen[w] < done[w] < done[v] .

• If v is not an ancestor of w in F, then

seen[v] < done[v] < seen[w] < done[w] .

Proof. The first part is clear from the recursive formulation of DFS. Now suppose
that v is not an ancestor of w. Note that w is obviously also not an ancestor of v. Thus
v lives in a subtree that is completely explored before the subtree of w is visited by
recursiveDFSvisit.
algorithm DFS
Input: digraph G
begin
stack S
array colour[0..n − 1], pred[0..n − 1], seen[0..n − 1], done[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE; pred[u] ← NULL
end for
time ← 0
for s ∈ V (G) do
if colour[s] = WHITE then
DFSvisit(s)
end if
end for
return pred, seen, done
end

algorithm DFSvisit
Input: node s
begin
colour[s] ← GREY
seen[s] ← time; time ← time + 1
S.insert(s)
while not S.isEmpty() do
u ← S.peek()
if there is a neighbour v with colour[v] = WHITE then
colour[v] ← GREY; pred[v] ← u
seen[v] ← time; time ← time + 1
S.insert(v)
else
S.delete()
colour[u] ← BLACK
done[u] ← time; time ← time + 1
end if
end while
end

Figure 5.7: Depth-first search algorithm.

 algorithm recursiveDFSvisit
Input: node s
begin
colour[s] ← GREY
seen[s] ← time; time ← time + 1
for each v adjacent to s do
if colour[v] = WHITE then
pred[v] ← s
recursiveDFSvisit(v)
end if
end for
colour[s] ← BLACK
done[s] ← time; time ← time + 1
end

Figure 5.8: Recursive DFS visit algorithm.

All four types of arcs in our search forest classification can arise with DFS. The
different types of non-tree arcs can be easily distinguished while the algorithm is
running. For example, if an arc (u, v) is explored and v is found to be white, then the
arc is a tree arc; if v is grey then the arc is a back arc, and so on (see Exercise 5.3.3).
We can also perform the classification after the algorithm has terminated, just by
looking at the timestamps seen and done (see Exercise 5.3.4).
Exercises
Exercise 5.3.1. Give examples to show that all four types of arcs can arise when DFS
is run on a digraph.

Exercise 5.3.2. Execute depth-first search on the digraph with adjacency lists repre-
sentation given below. Classify each arc as tree, forward, back or cross.

0: 2
1: 0
2: 0 1
3: 4 5 6
4: 5
5: 3 4 6
6: 1 2

Exercise 5.3.3. Explain how to determine, at the time when an arc is first explored
by DFS, whether it is a cross arc or a forward arc.
Exercise 5.3.4. Suppose that we have performed DFS on a digraph G. Let (v, w) ∈
E(G). Show that the following statements are true.

• (v, w) is a tree or forward arc if and only if

seen[v] < seen[w] < done[w] < done[v];

• (v, w) is a back arc if and only if

seen[w] < seen[v] < done[v] < done[w];

• (v, w) is a cross arc if and only if

seen[w] < done[w] < seen[v] < done[v].

Exercise 5.3.5.
Suppose that DFS is run on a digraph G and the following timestamps obtained.

v 0 1 2 3 4 5 6
seen[v] 0 1 2 11 4 3 6
done[v] 13 10 9 12 5 8 7

• List all tree arcs in the DFS forest.

• Suppose that (6, 1) is an arc of G. Which type of arc (tree, forward, back or cross)
is it?

• Is it possible for node 2 to be an ancestor of node 3 in the DFS forest?

• Is it possible that G contains an arc (5, 3)? If so, what type of arc must it be?

• Is it possible that G contains an arc (1, 5)? If so, what type of arc must it be?

Exercise 5.3.6. Is there a way to distinguish tree arcs from non-tree arcs just by look-
ing at timestamps after DFS has finished running?

Exercise 5.3.7. Suppose that DFS is run on a graph G. Prove that cross edges do not
occur.

Exercise 5.3.8. Give an example to show that the following conjecture is not true:
if w is reachable from v and seen[v] < seen[w] then w is a descendant of v in the DFS
forest.
Exercise 5.3.9. DFS allows us to give a so-called pre-order and post-order labelling
to a digraph. The pre-order label indicates the order in which the nodes were turned
grey. The post-order label indicates the order in which the nodes were turned black.
For example, each node of the following tree is labelled with a pair of integers
indicating the pre- and post- orders, respectively, of the layout.

1,8

2,3 5,7

3,2 7,6
6,4

4,1
8,5

post-order traversal
This is obviously strongly related to the values in the arrays seen and done. What
is the exact relationship between the two?
Exercise 5.3.10. Prove Theorem 5.4 by using induction.

5.4 Additional properties of breadth-first search

The first-in first-out processing of the grey nodes in BFS is ideally handled by a
queue. In Figure 5.9 we present the algorithm. The timestamps seen, done of DFS are
of less use here; it is more useful to record the number of steps from the root in the
array d.
A similar analysis to what we did for DFS also holds for BFS: it is also a linear time
traversal algorithm, because the next grey and white node can again be chosen in
constant time.
It is rather obvious that BFS processes all nodes at distance 1, then all nodes at
distance 2, etc, from the root. The formal proof is below.
Theorem 5.6. Suppose we run BFS on a digraph G. Let v ∈ V (G), and let r be the root
of the search tree containing v. Then d[v] = d(r, v).

Proof. Note that since d[v] is the length of a path of tree arcs from r to v, we have
d[v] ≥ d(r, v). We prove the result by induction on the distance. Denote the BFS search
algorithm BFS
Input: digraph G
begin
queue Q
array colour[0..n − 1], pred[0..n − 1], d[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE; pred[u] ← NULL
end for
for s ∈ V (G) do
if colour[s] = WHITE then
BFSvisit(s)
end if
end for
return pred, d
end

algorithm BFSvisit
Input: node s
begin
colour[s] ← GREY; d[s] ← 0
Q.insert(s)
while not Q.isEmpty() do
u ← Q.peek()
for each v adjacent to u do
if colour[v] = WHITE then
colour[v] ← GREY; pred[v] ← u; d[v] ← d[u] + 1
Q.insert(v)
end if
end for
Q.delete()
colour[u] ← BLACK
end while
end

Figure 5.9: Breadth-first search algorithm.

forest by F and let s be the root of a tree in F. Then d[s] = 0 = d(s, s) so the result is
true for distance zero. Suppose it is true for all v for which d(s, v) < k and consider a
node v such that d(s, v) = k ≥ 1. Choose a shortest path from s to v in G and let u be
the penultimate node in the path. Then d(s, u) = k − 1 (it cannot be less, or it would
contradict d(s, v) = k; on the other hand the subpath from s to u must be a shortest
path from s to u, otherwise we could find a shorter one from s to v). By the inductive
hypothesis, d[u] = d(s, u) = k − 1. Now v must be seen after u (otherwise d[v] < k, but
we know d[v] ≥ d(s, v) = k). Thus v is seen in the loop through white neighbours of u,
and so d[v] = d[u] + 1 = k.

We can classify arcs, but the answer is not as nice as with DFS.

Theorem 5.7. Suppose that we are performing BFS on a digraph G. Let (v, w) ∈ E(G)
and suppose that we have just chosen the grey node v. Then

• if (v, w) is a tree arc then colour[w] = WHITE, d[w] = d[v] + 1

• if (v, w) is a back arc, then colour[w] = BLACK, d[w] ≤ d[v] − 1

• There are no forward arcs.

• if (v, w) is a cross arc then one of the following holds:

◦ d[w] < d[v] − 1, and colour[w] = BLACK;

◦ d[w] = d[v], and colour[w] = GREY;
◦ d[w] = d[v], and colour[w] = BLACK;
◦ d[w] = d[v] − 1, and colour[w] = GREY;
◦ d[w] = d[v] − 1, and colour[w] = BLACK.

Proof. The arc is added to the tree if and only if w is white. If the arc is a back arc, then
w is an ancestor of v; the FIFO queue structure means w is black before the adjacency
list of v is scanned.
Now suppose that (x, u) is a forward arc. Then since u is a descendant of x but not
a child in the search forest, Theorem 5.6 yields d[u] ≥ d[x] + 2. But by the last theorem
we have d[u] = d(s, u) ≤ d(s, x) + 1 = d[x] + 1, a contradiction. Hence no such arc exists.
A cross arc may join two nodes on the same level, jump up one level, or jump
up more than one level. In the last case, w is already black before v is seen. In the
second case, w may be seen before v, in which case it is black before v is seen (recall
w is not the parent of v), or it may be seen after v, in which case it is grey when (v, w)
is explored. In the first case, w may be seen before v (in which case it is black before v
is seen), or w may be seen after v (in which case it is grey when (v, w) is explored).

In the special case of graphs we can say more.

Theorem 5.8. Suppose that we have performed BFS on a graph G. Let {v, w} ∈ E(G).
Then exactly one of the following conditions holds.

• {v, w} is a tree edge, |d[w] − d[v]| = 1;

• {v, w} is a cross edge, d[w] = d[v];

• {v, w} is a cross edge, |d[w] − d[v]| = 1.

Proof. By Theorem 5.7 there can be no forward edges, hence no back edges. A cross
edge may not jump up more than one level, else it would also jump down more than
one level, which is impossible by Theorem 5.6.

For a given BFS tree, we can uniquely label the vertices of a digraph based on
the time they were first seen. For the graph G1 of Figure 5.4, we label vertex 0 with
1, vertices {1, 2} with labels {2, 3}, vertices {3, 4, 8} with labels {4, 5, 6}, and the last
vertex level {5, 6, 7} with labels {7, 8, 9}. These are indicated in Figure 5.5.
Exercises
Exercise 5.4.1. Carry out BFS on the digraph with adjacency list given below. Show
the state of the queue after each change in its state.

0: 2
1: 0
2: 0 1
3: 4 5 6
4: 5
5: 3 4 6
6: 1 2

Exercise 5.4.2. How can we distinguish between a back and a cross arc while BFS is
running on a digraph?

Exercise 5.4.3. Explain how to determine whether the root of a BFS tree is contained
in a cycle, while the algorithm is running. You should find a cycle of minimum length
if it exists.

5.5 Priority-first search

Priority-first search is a more sophisticated form of traversal with many applica-
tions. For now, we consider it simply as a common generalization of breadth-first
and depth-first search. Priority-first search may seem a little abstract compared to
the more concrete DFS and BFS. We shall not need it until Chapter 6; however it will
be essential then.
 The important property is that each grey node has associated with it an integer
key. The interpretation of the key is of a priority: the smaller the key, the higher the
priority. The rule for selecting a new grey node is to choose one with the smallest
key.
In the simplest form of PFS, the key value is assigned when the node becomes
grey, and never updated subsequently. More generally, the key may be further up-
dated at other times. We shall see both types in this book. The second type of PFS is
used in optimization problems as we shall discuss in Chapter 6.
The first type of PFS includes both BFS and DFS. In BFS, the key value of v can
be taken as the time v was first coloured grey. Note that this means that a given grey
node can be selected many times—until it becomes black, in fact, it will always have
minimum key among the grey nodes. By contrast, in DFS we can take the key value
to be −seen[v]. Then the last node seen always has minimum key. It cannot be chosen
again until the nodes seen after it have become black.
The running time of PFS depends mostly on how long it takes to find the mini-
mum key value, and how long it takes to update the key values.
In the array implementation mentioned above, finding the minimum key value
takes time of order n at each step, so the quantity a is Ω(n). Thus a PFS of this type
will take time in Ω(n2 ). This is worse than the Θ(n + m) we obtain with BFS and DFS
using adjacency lists and a queue or stack respectively. One reason is that a simple
array is not a particularly good data structure for finding the minimum key. You have
already seen a better one in Part I of this book—the binary heap. In fact PFS is best
described via the priority queue ADT (see Section C.1).
Pseudocode demonstrating the first type of PFS is presented in Figure 5.10. The
subroutine setKey there is the rule for giving the key value when a node is inserted.
We do not include any code for setKey.
We proceed to show some applications of BFS and DFS in the following sections.
Applications of PFS will be discussed in Chapter 6.

5.6 Acyclic digraphs and topological ordering

In this section we show how to arrange, when possible, the nodes of a digraph
into a topological or precedence order. Many computer science applications require
us to find precedences (or dependencies) among events, such as a compiler evalu-
ating sub-expressions of an expression like that shown in Figure 5.11.
Here the compiler would need to compute, for example, both (a + b) and c before
it can compute c − (a + b).
Definition 5.9. Let G be a digraph. A topological sort of G is a linear ordering of
all its vertices such that if G contains an arc (u, v), then u appears before v in the
ordering.
The term topological sort comes from the study of partial orders and is some-
times called a topological order or linear order. If a topological sort exists, then it is
algorithm PFS
Input: digraph G
begin
priority queue Q
array colour[0..n − 1], pred[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE; pred[u] ← NULL
end for
for s ∈ V (G) do
if colour[s] = WHITE then
PFSvisit(s)
end if
end for
return pred
end

algorithm PFSvisit
Input: node s
begin
colour[s] ← GREY
Q.insert(s, setKey (s))
while not Q.isEmpty() do
u ← Q.peek()
if u has a neighbour v with colour[v] = WHITE then
colour[v] ← GREY
Q.insert(v, setKey (v))
else
Q.delete()
colour[u] ← BLACK
end if
end while end

Figure 5.10: Priority-first search algorithm (first kind).

 *

*
+

(a+b)*(c-(a+b))*(-c+d) -
-
+

a b c d
Figure 5.11: Digraph describing structure of an arithmetic expression.

possible to draw a picture of G with all nodes in a straight line, and the arcs “pointing
the same way”.
A digraph without cycles is commonly called a DAG, an abbreviation for directed
acyclic graph. It is much easier for a digraph to be a DAG than for its underlying
graph to be acyclic.
For our arithmetic expression example above, a linear order of the sub-expression
DAG gives us an order (actually the reverse of the order) where we can safely evaluate
the expression.
Clearly if the digraph contains a cycle, it is not possible to find such a linear order-
ing. This corresponds to inconsistencies in the precedences given, and no schedul-
ing of the tasks is possible.

Example 5.10. In Figure 5.12 we list three DAGs and possible topological orders for
each. Note that adding more arcs to a DAG reduces the number of topological orders
it has. This is because each arc (u, v) forces u to occur before v, which restricts the
number of valid permutations of the vertices.

The algorithms for computing all topological orders are more advanced than
what we have time or space for here. We show how to compute one such order,
however.
First we note that if a topological sort of a DAG G is possible, then there must be
a source node in G. The source node can be first in a topological order, and no node
that is not a source can be first (because it has an in-neighbour that must precede it
in the topological order).

Theorem 5.11. A digraph has a topological order if and only if it is a DAG.

 0 0
0

1 1
1 2
5
2
2 3
3 4 3
4
4 5
0,1,2,3,4 6

0,1,2,4,3 0,1,3,2,4,5
0,1,2,4,5,3,6
0,2,1,3,4 0,1,3,2,5,4
0,1,2,3,4,5 0,1,2,5,3,4,6
0,2,1,4,3 0,1,2,5,4,3,6
0,1,2,3,5,4
.. 0,1,5,2,3,4,6
.
0,1,2,5,4,3 0,1,5,2,4,3,6

Figure 5.12: Topological orders of some DAGs.

Proof. First show that every DAG has a source (see exercise 5.6.2). Given this, we
proceed as follows. Deleting a source node creates a digraph that is still a DAG, be-
cause deleting a node and some arcs cannot create a cycle where there was none
previously. Repeatedly doing this gives a topological order.

This theorem then gives an algorithm (zero-indegree sorting ) for topologically

sorting a DAG G.
So far we have not shown how to determine whether a given digraph is a DAG.
But if we apply zero-indegree sorting to a digraph that is not a DAG, eventually it
will stop because no source node can be found at some point (otherwise we would
obtain a topological order and hence the digraph would be a DAG).
There is another way to determine acyclicity and topologically sort a DAG, based
on DFS. If G contains a cycle, DFS must eventually reach a node that points to one
that has been seen before. In other words, we will detect a back arc. The details and
proof that this works now follow.

Theorem 5.12. Suppose that DFS is run on a digraph G. Then G is acyclic if and only
if there are no back arcs.

Proof. Suppose that we run DFS on G. Note that if there is a back arc (v, u), then u
and v belong to the same tree T , with root s say. Then there is a path from s to u, and
there is a path from u to v by definition of back arc. Adding the arc (v, u) gives a cycle.
 Conversely, if there is a cycle v0 v1 . . . vn v0 , we may suppose without loss of gener-
ality that v0 is the first node of the cycle visited by the DFS algorithm. We claim that
(vn , v0 ) is a back arc. To see why this is true, first note that during the DFS v0 is linked
to vn via a path of unvisited nodes (possibly of length shorter than n). We have vn as
a descendant of v0 in the DFS tree and a back arc (vn , v0 ).

One valid topological order is simply the reverse of the DFS finishing times.

Theorem 5.13. Let G be a DAG. Then listing the nodes in reverse order of DFS fin-
ishing times yields a topological order of G.

Proof. Consider any arc (u, v) ∈ E(G). Since G is a DAG, the arc is not a back arc by
Theorem 5.12. In the other three cases, Exercise 5.3.4 shows that done[u] > done[v],
which means u comes before v in the alleged topological order.

We can therefore just run DFS on G, and stop if we find a back arc. Otherwise
printing the nodes in reverse order of finishing time gives a topological order. Note
that printing the nodes in order of finishing time gives a topological order of the
reverse digraph Gr .
Exercises
Exercise 5.6.1. Give an example of a DAG whose underlying graph contains a cycle.
Make your example as small as possible.

Exercise 5.6.2. Prove that every DAG must have at least one source and at least one
sink.

Exercise 5.6.3. Show that the following method for topologically sorting a DAG does
not work in general: print the nodes in order of visiting time.

Exercise 5.6.4. Professor P has the following information taped to his mirror, to help
him to get dressed in the morning.
Socks before shoes; underwear before trousers; trousers before belt; trousers be-
fore shoes; shirt before glasses; shirt before tie; tie before jacket; shirt before hat;
shirt before belt.
Find an acceptable order of dressing for Professor P.

Exercise 5.6.5. What is the time complexity of zero-indegree sorting?

Exercise 5.6.6. Let G be a graph. There is an easy way to show that G is acyclic. It is
not hard to show (see Section D.7) that a graph G is acyclic if and only if G is a forest,
that is, a union of (free) trees.
Give a simple way to check whether a graph G is acyclic. Does the method for
finding a DAG given above work for acyclic graphs also?
5.7 Connectivity
For many purposes it is useful to know whether a digraph is “all in one piece”, and
if not, to decompose it into pieces. We now formalize these notions. The situation
for graphs is easier than that for digraphs.
Definition 5.14. A graph is connected if for each pair of vertices u, v ∈ V (G), there is
a path between them.
In Example 4.3 the graph G1 is connected, as is the underlying graph of G2 .
If a graph is not connected, then it must have more than one “piece”. More for-
mally, we have the following.
Theorem 5.15. Let G be a graph. Then G can be uniquely written as a union of
subgraphs Gi with the following properties:
• each Gi is connected
• if i 6= j, there are no edges from any vertices in Gi to any vertices in G j

Proof. Consider the relation ∼ defined on V (G), given by u ∼ v if and only if there is
a path joining u and v (in other words, u and v are each reachable from the other).
Then ∼ is an equivalence relation and so induces a partition of V (G) into disjoint
subsets. The subgraphs Gi induced by these subsets have no edges joining them by
definition of ∼, and each is connected by definition of ∼.

The subgraphs Gi above are called the connected components of the graph G.
Clearly, a graph is connected if and only if it has exactly one connected component.
Example 5.16. The graph obtained by deleting two edges from a triangle has 2 con-
nected components.
We can determine the connected components of a graph easily by using a traver-
sal algorithm. The following obvious theorem explains why.
Theorem 5.17. Let G be a graph and suppose that DFS or BFS is run on G. Then the
connected components of G are precisely the subgraphs spanned by the trees in the
search forest.

Proof. The result is true for any traversal procedure, as we have already observed in
Theorem 5.2. The trees of the search forest have no edges joining them, and together
they span G.

So we need only run BFS or DFS on the graph, and keep count of the number of
times we choose a root—this is the number of components. We can store or print
the vertices and edges in each component as we explore them. Clearly, this gives a
linear time algorithm for determining the components of a graph.
 So far it may seem that we have been too detailed in our treatment of connect-
edness. After all the above results are all rather obvious. However, now consider the
situation for digraphs. The intuition of “being all in one piece” is not as useful here.
In Example 4.3 the graph G1 is connected, as is the underlying graph of G2 . They
are “all in one piece”, but not the same from the point of view of reachability. For
example, in digraph G2 , node 2 is a sink. This motivates the following definition.
Definition 5.18. A digraph G is strongly connected if for each pair of nodes u, v of G,
there is a path in G from u to v.
Note. In other words, u and v are each reachable from the other.

Suppose that the underlying graph of G is connected (some authors call this be-
ing weakly connected), but G is not strongly connected. Then if G represents a road
network, it is possible to get from any place to any other one, but at least one such
route will be illegal: one must go the wrong way down a one-way street.
A strongly connected digraph must contain many cycles: indeed, if v and w are
different nodes, then there is a path from v to w and a path from w to v, so v and w are
contained in a cycle. Conversely, if each pair of nodes is contained in a cycle, then
the digraph is clearly strongly connected.
Again, we can define strongly connected components in a way that is entirely
analogous to component for graphs. The proof above for connected components
generalizes to this situation.
Theorem 5.19. Let G = (V, E) be a digraph. Then V can be uniquely written as a
union of disjoint subsets Vi , with each corresponding induced subdigraph Gi being
a strongly connected component of G.

Proof. Consider the relation ∼ defined on V , given by u ∼ v if and only if there is a

path joining u and v and a path joining v and u (in other words, u and v are each reach-
able from the other). Then ∼ is an equivalence relation and so induces a partition of
V into disjoint subsets. By definition, each subdigraph Gi is strongly connected and
of maximal order.

Example 5.20. A digraph and its three (uniquely determined) strongly connected
components are displayed in Figure 5.13. Note that there are arcs of the digraph not
included in the strongly connected components.
Note that if the underlying graph of G is connected but G is not strongly con-
nected, then there are strong components C1 and C2 such that it is possible to get
from C1 to C2 but not from C2 to C1 . If C1 and C2 are different strong components,
then any arcs between them must either all point from C1 to C2 or from C2 to C1 . Sup-
pose that we imagine each strong component shrunk to a single node (so we ignore
the internal structure of each component, but keep the arcs between components).
 0 1

0 4

1
2 3

2 3 5

4 5

Figure 5.13: A digraph and its strongly connected components.

Then in the digraph resulting, if v 6= w and we can get from v to w then we cannot
get from w to v. In other words, no pair of nodes can belong to a cycle, and hence
the digraph is acyclic. See Figure 5.14. Note that the converse is also true: if we have
an acyclic digraph G and replace each node by a strongly connected digraph, the
strongly connected components of the resulting digraph are exactly those digraphs
that we inserted.

C1 C2 Cm

Figure 5.14: Structure of a digraph in terms of its strong components.

Note the similarity between this and the search forest decomposition in Figure 5.3.
In that case, if we shrink each search tree to a point, the resulting digraph is also
acyclic.
How to determine the strongly connected components? First we observe that the
previous method for graphs definitely fails (see Exercise 5.7.1). To decide whether a
digraph is strongly connected we could run BFSvisit or DFSvisit originating from
each node in turn and see whether each of the n trees so generated spans the di-
graph. However the running time of such an algorithm is Θ(n2 + nm).
We can do better by using DFS more cleverly.
Consider the reverse Gr . The strong components of Gr are the same as those of G.
Shrinking each strong component to a point, we obtain acyclic digraphs H and Hr .
Consider a sink S1 in Hr . If we run DFS on Gr starting in the strong component S1 , we
will reach every node in that component and no other nodes of Gr . The DFS tree will
exactly span S1 . Now choose the next root to lie in the strong component S2 node of
Hr whose only possible out-neighbour is S1 (this is possible by the same reasoning
used for zero-indegree sort, except here we deal with outdegree). The DFS will visit
all of S2 and no other nodes of Gr because all other possible nodes have already been
visited. Proceed in this way until we have visited all strong components.
We have shown that if we can choose the roots correctly, then we can find all
strong components. Now of course we don’t know these components a priori, so
how do we identify the roots to choose?
Whenever there is a choice for the root of a new search tree, it must correspond
to a new node of the DAG Hr . We want at least a reverse topological order of Hr . This
is simply a topological order for H. Note that in the case where H = G (each strong
component has just one point), then G is a DAG. The method above will work if and
only if we choose the roots so that each tree in the DFS for Gr has only one point. We
just need a topological order for G, so run DFS on G and print the nodes in reverse
order of finishing time. Then choose the roots for the DFS on Gr in the printed order.
It therefore seems reasonable to begin with a DFS of G. Furthermore, an obvi-
ous choice is: in the DFS of Gr , choose each new root from among white nodes that
finished latest in F.
Then each DFS tree in the search of Gr definitely contains the strong component
S of the root r. To see this, note that no node in that strong component could have
been visited before in Gr , otherwise r would have already been visited. By Theo-
rem 5.4, every node in the strong component of r is a descendant of r.
The only thing that could go wrong is that a search tree in Gr might contain more
than one strong component. This cannot happen, as we now prove.
Theorem 5.21. If the following rule for choosing roots is used in the algorithm de-
scribed above, then each tree in the second search forest spans a strong component
of G, and all strong components arise this way.
Rule: use the white node whose finishing time in F was largest.

Proof. Suppose that a search tree in Gr does contain more than one strong compo-
nent. Let S1 be the first strong component seen in Gr and let S2 be another, and let
the roots be r, s respectively. Note that by the rule for choosing nodes r was the first
node of S1 seen in F (by Theorem 5.4, every node of S1 is a descendant of the first
one seen, which therefore has latest finishing time). The analogous statement holds
for s and S2 .
 By the rule for choosing roots, we have done[r] > done[s] in F. We cannot have
seen[s] > seen[r] in F, for then s would be a descendant of r in F and in Gr , so they
would belong to the same strong component. Thus seen[r] > seen[s] in F. Hence S2
was explored in F before r (and hence any node of S1 ) was seen. But then r would
have been reachable from s in G via a path of white nodes, so Theorem 5.4 shows
that r and s are in the same strong component, another contradiction.

The above algorithm runs in linear time with adjacency lists, since each DFS and
the creation of the reverse digraph take linear time. We only need to remember while
performing the first DFS to store the nodes in an array in order of finishing time.
Exercises
Exercise 5.7.1. Give an example to show that a single use of DFS does not in general
find the strongly connected components of a digraph.

Exercise 5.7.2. Carry out the above algorithm by hand on the digraph of Exam-
ple 5.20 and verify that the components given there are correct. Then run it again
on the reverse digraph and verify that the answers are the same.

5.8 Cycles
In this section, we cover three varied topics concerning cycles.
The girth of a graph
The length of the smallest cycle in a graph is an important quantity. For exam-
ple, in a communication network, short cycles are often something to be avoided
because they can slow down message propagation.

Definition 5.22. For a graph (with a cycle), the length of the shortest cycle is called
the girth of the graph. If the graph has no cycles then the girth is undefined but may
be viewed as +∞.

Note. For a digraph we use the term girth for its underlying graph and the (maybe
non-standard) term directed girth for the length of the smallest directed cycle.

Example 5.23. In Figure 5.15 are three (di)graphs. The first has no cycles (it is a free
tree), the second is a DAG of girth 3, and the third has girth 4.

How to compute the girth of a graph? Here is an algorithm for finding the length
of a shortest cycle containing a given vertex v in a graph G. Perform BFSvisit. If we
meet a grey neighbour (that is, we are exploring edge {x, y} from x and we find that y
is already grey), continue only to the end of the current level and then stop. For each
edge {x, y} as above on this level, if v is the lowest common ancestor of x and y in the
BFS tree, then there is a cycle containing x, y, v of length l = d(x) + d(y) + 1. Report the
minimum value of l obtained along the current level.
 0 1 0 0 1

1 2 4
2 3 4
5
2 3

5 6 3
5 6
4
7 8 6 7 8

Figure 5.15: Some (di)graphs with different cycle behaviour.

Theorem 5.24. The above algorithm is correct.

Proof. Suppose that we arrive at vertex x, d(x) = k, and we have just encountered a
grey neighbour y. Then d[y] = d[x] + 1 or d[y] = d[x] (note that d[x] = d[y] + 1 is ruled
out because then x would be a child of y and y would be black). This means that there
is definitely a cycle containing x, y and z, where z is the lowest common ancestor of x
and y in the BFS tree. Note that z 6= x, z 6= y since neither x nor y is an ancestor of the
other. The cycle consists of the tree edges from z to x, the cross edge {x, y} and the
tree edges from z to y. The length of the cycle is l = d[x] + d[y] − 2d[z] + 1.
Conversely, let C be any cycle and let z be the first vertex of the cycle reached by
the search. Let x and y be two vertices in C whose distance from z is greatest, with
d[x] ≤ d[y]. Then the lowest common ancestor of x and y is exactly z, and the cycle
found in the first paragraph is exactly C.
The vertex v belongs to the cycle C if and only if v = z. The length of the cycle
is 2k + 1 if d[x] = d[y] and 2k + 2 if d[y] = d[x] − 1. The minimum length of any cycle
containing v found after this is 2(k + 1) + 1 = 2k + 3, so no better cycle can be found
after the current level k is explored.
Note. An easy-to-implement DFS idea may not work properly. For example, the DFS
tree originating from vertex 0 of the third graph of Figure 5.15 finds only 6-cycles
with the back edges (even though a 4-cycle exists using two of the back edges).
To compute the girth of a graph, we can simply run the above algorithm from
each vertex in turn, and take the minimum cycle length achieved.
Bipartite graphs
Many graphs in applications have two different types of nodes, and no relations
between nodes of the same type (this is a model of sexually reproducing organisms,
for example).
Definition 5.25. A graph G is bipartite if V (G) can be partitioned into two nonempty
disjoint subsets V0 ,V1 such that each edge of G has one endpoint in V0 and one in V1 .
 Figure 5.16: A bipartite graph.

Example 5.26. The graph in Figure 5.16 is bipartite. The isolated vertex could be
placed on either side.
Showing that a graph is bipartite can be done by exhibiting a bipartition (a par-
tition into two subsets as in the definition). Of course finding such a bipartition may
not be easy. Showing a graph is not bipartite seems even harder. In each case, we
certainly do not want to have to test all possible partitions of V into two subsets!
There is a better way.
Definition 5.27. Let k be a positive integer. A graph G has a k-colouring if V (G) can
be partitioned into k nonempty disjoint subsets such that each edge of G joins two
vertices in different subsets.
Example 5.28. The graph in Figure 5.16 has a 2-colouring as indicated.
It is not just a coincidence that our example of a bipartite graph has a 2-colouring.
Theorem 5.29. The following conditions on a graph G are equivalent.
• G is bipartite;
• G has a 2-colouring;
• G does not contain an odd length cycle.

Proof. Given a bipartition, use the same subsets to get a 2-colouring, and vice versa.
This shows the equivalence of the first two conditions. Now suppose G is bipartite.
A cycle must have even length, since the start and end vertices must have the same
colour. Finally suppose that G has no odd length cycle. A 2-colouring is obtained as
follows. Perform BFS and assign each vertex at level i the “colour” i mod 2. If we can
complete this procedure, then by definition each vertex goes from a vertex of one
colour to one of another colour. The only problem could be if we tried to assign a
colour to a node v that was adjacent to a node w of the same colour at the same level
k. But then a cycle of length 2k + 1 is created.

It is now easy to see that we may use the method in the proof above to detect an
odd length cycle if it exists, and otherwise produce a 2-colouring of G. This of course
runs in linear time.
Exercises
Exercise 5.8.1. Give an example to show that in the shortest cycle algorithm, if we
do not continue to the end of the level, but terminate when the first cycle is found,
we may find a cycle whose length is one more than the shortest possible.

Exercise 5.8.2. What is the time complexity of the shortest cycle algorithm?

Exercise 5.8.3. The n-cube (hypercube) is a graph on 2n vertices, each labelled by a

different bit vector of length n. If v = (v0 , . . . , vn−1 ) and w = (w0 , . . . , wn−1 ) are such bit
vectors, then there is an edge between the vertex labelled v and a vertex labelled w
if and only if v and w differ in exactly one component. For which values of n is the
n-cube bipartite?

5.9 Maximum matchings

Next we want to introduce an important graph problem that can be solved in
polynomial time by a clever path augmentation algorithm.

Definition 5.30. A matching in a graph is a set of pairwise non-adjacent edges (that

is, each vertex can be in at most one edge of the matching). A maximal matching
is a matching such that is not a proper subset of any other matching. A maximum
matching is one with the largest possible number of edges (over all possible match-
ings).

Often, for many real-world problems, we want to find a maximum matching in

bipartite graphs as illustrated by the next two examples.

Example 5.31. Suppose we have a set of workers V0 and a set of tasks V1 that need to
be assigned. A given worker of V0 is able to perform a subset of the tasks in V1 . Now
with each worker capable of doing at most one task at a time, the boss would like to
assign (match) as many workers as possible to as many of the tasks.

Example 5.32. Consider the marriage problem where we have a set of men and
women (as vertices) and edges representing compatible relationships. The goal is to
marry as many couples as possible, which is the same as finding a maximum match-
ing in the relationship graph. If there are no homosexual interests then we have a
bipartite graph problem.
 Ann Bob Ann Bob

Cher Doug Cher Doug

Eve Fred Eve Fred

Figure 5.17: A maximal and maximum matching in a bipartite graph.

In Figure 5.17 we illustrate the difference between a maximal and maximum match-
ing in the setting of Example 5.32. The matchings consist of bold-dashed edges (be-
tween females on the left and males on the right) in the drawings.
It is easy to find a maximal matching M in a graph. For example, a simple greedy
approach of iterating over all edges and adding each edge to M if it is non-adjacent
to anything already in M will work. As illustrated in Figure 5.17, a maximal matching
may have fewer edges than a more desirable maximum matching.
Our algorithm to compute a maximum matching will be based on trying to im-
prove an existing matching M by finding certain types of paths in a graph.

Definition 5.33. Given a matching M, an alternating path is a path in which the

edges of the path alternate from being in the matching and not. An augmenting
path is an alternating path that starts from and ends on unmatched vertices.

For example, consider the augmenting path Eve–Doug–Cher–Fred of Figure 5.17.

It contains three edges but only the middle edge is part of a matching on the left case.
We can get the better matching on the right if we add Eve-Doug, remove Doug–Cher,
and add Cher–Fred to the existing matching. Thus, in general, we see that we can
improve a matching if we can find an augmenting path. Note that there is always
one more non-matching edge than matching edge in an augmenting path. Likewise,
it is pretty easy to show that if there is no such augmenting path then we must have
a maximum matching (see Exercise 5.9.1).
We next present a polynomal-time algorithm that finds an augmenting path if
one exists. The basic idea is to start from an unmatched vertex v and build a tree
(via a graph traversal algorithm such as BFS) of alternating paths away from v. If we
reach another unmatched vertex then we have found an augmenting path. Other-
wise, if we visit all vertices (in the same component as v) then we conclude that no
augmenting path exists starting at v. This algorithm is given in Figure 5.18.

Theorem 5.34. There exists a polynomial-time algorithm to find a maximum match-

ing in a bipartite graph.
algorithm findAugmentingPath
Input: bipartite graph G; matching M; unmatched vertex x
begin
queue Q
array status[0..n − 1], pred[0..n − 1]
for each u ∈ V (G) do
status[u] ← WHITE; pred[u] ← NULL
end for
status[x] ← EVEN
Q.insert(x)
while not Q.isEmpty() do
u ← Q.pop()
if status[u] = EVEN then
for each v adjacent to u do
if status[v] = WHITE then
status[v] ← ODD; pred[v] ← u
if v is unmatched in M then
return path x, . . . , pred[pred[u]], pred[u], u, v
else
Q.insert(v)
end if
end if
end for
else that is, status[u] = ODD
v ← matched vertex of u from M
if status[v] = WHITE then
status[v] ← EVEN; pred[v] ← u
Q.insert(v)
end if
end if
end while
return false no augmenting paths containing x
end

Figure 5.18: An algorithm to find an augmenting path, given a matching and an unmatched
starting vertex.
Proof. We first need to show that the algorithm findAugmentingPath, given in Fig-
ure 5.18, does find an augmenting path if one exists. It is sufficient to show that if
there exists at least one augmenting path from vertex x to some other unmatched
vertex that we find any one of them (in our case, by imitating BFS, it will be one of
shortest length). findAugmentingPath builds a traversal tree starting at x using the
following constraints.

• If a reachable vertex u is in the same partition as x (status will be set to EVEN

by the algorithm) then we know (except for x) that there is an alternating path
with the last edge including u being in the matching.

• If a reachable vertex v is not in the same partition as x (status will be set to

ODD if it is matched) then we know that there is an alternating path with the
last edge including v is not in the matching.

This process produces a tree with alternating paths rooted at x as illustrated in Fig-
ure 5.19. The status of the nodes are ODD or EVEN depending if they are an even or
odd distance from x. If a vertex (first seen) is at an odd distance then we have seen
an alternating path where the last edge is not in the matching. If this last vertex is
unmatched then we have found an augmenting path, otherwise we extend the path
by using the matched edge. If a vertex is at an even distance then we have seen an
alternating path where the last edge is in the matching. Since the graph is bipartite
the status of being ODD or EVEN is unambiguous.
Suppose the algorithm findAugmentingPath terminates without finding an aug-
menting path when one does exist. Let x = v0 , v1 , . . . , vk be a counterexample. Con-
sider the first index 0 < i ≤ k such that status[vi ] = WHITE. We know vi−1 was in-
serted in the queue Q. Consider the two cases. If i − 1 is even then since vi is a
neighbour of vi−1 its status would have changed to ODD. If i − 1 is odd then either
(vi−1 , vi ) is in the matching or not. If so, the status of vi would have changed; if not,
a prefix of the counterexample is not an augmenting path. Thus, by contradiction,
findAugmentingPath will find an augmenting path if one exists.
The running time of one invocation of findAugmentingPath is the same as the
running time of BFS since each vertex is added to the queue Q at most once. For
adjacency list representation of graphs this can be carried out in time O(m). If we
find an augmenting path then our best matching increases by one. Since a maxi-
mum matching is bounded by ⌊n/2⌋ we only need to find at most O(n) augmenting
paths. We potentially need to call findAugmentingPath once for each unmatched
vertex, which is bounded by O(n), and repeat the process for each modified match-
ing. Therefore, the total running time to find a maximum matching is at most O(n2 m).

The algorithm presented here can easily be improved to O(nm) by noting that it is
only required to traverse and compute an “alternating path forest” in order to find an
 x
EVEN

ODD

EVEN

ODD

EVEN

pred[u] ODD

u = pred[v] EVEN

v — unmatched

Figure 5.19: Structure of the graph traversal tree for finding augmenting paths.

augmenting path. That is, we do not need to originate a call to findAugmentingPath

for all unmatched vertices. However the correctness is a bit tricky to justify.
There are other algorithms that find maximum matchings more efficiently than
the one
√ presented. One of these is the Hopcroft-Karp algorithm which runs in time
O(m n) and is based on finding a maximal flow in a network (by adding a source
and sink vertex and directing all edges from one vertex partition to the other of the
bipartite graph).
We conclude this section by mentioning that there also exist polynomial-time
algorithms to find maximum matchings in arbitrary (non-bipartite) graphs. The de-
tails are beyond the scope of this book.
Exercises
Exercise 5.9.1. Prove that a matching for a graph is maximum if and only if it does
not have any augmenting paths.
Exercise 5.9.2. Give an example of a bipartite graph of order 12 with a maximal
matching that has an augmenting path of length 5 and a maximum matching of the
same graph with two more edges.
Exercise 5.9.3. Show that the size of a maximum matching in a bipartite graph G =
(V, E) is the same as the size of the smallest vertex cover of G. A vertex cover is a
subset V ′ ⊆ V of vertices such that for all (u, v) ∈ E, at least one of u or v is in V ′ . Does
the equality hold for arbitrary graphs?

5.10 Notes
The linear-time algorithm for finding strong components of a digraph was intro-
duced by R. E. Tarjan in 1971.
One of the early polynomial-time algorithms for finding maximum matchings
in bipartite graphs is based on the Ford–Fulkerson network flow algorithm [7] from
1956. The first polynomial-time algorithm for finding a maximum matching in an
arbitrary graph was presented by J. Edmonds [5] in 1965.
Chapter 6

Weighted Digraphs and Optimization

Problems

So far our digraphs have only encoded information about connection relations be-
tween nodes. For many applications it is important to study not only whether one
can get from A to B, but how much it will cost to do so.
For example, the weight could represent how much it costs to use a link in a com-
munication network, or distance between nodes in a transportation network. We
shall use the terminology of cost and distance interchangeably (so, for example, we
talk about finding a minimum weight path by choosing a shortest edge).
We need a different ADT for this purpose.

6.1 Weighted digraphs

Definition 6.1. A weighted digraph is a pair (G, c) where G is a digraph and c is a
cost function, that is, a function associating a real number to each arc of G.

We interpret c(u, v) as the cost of using arc (u, v). An ordinary digraph can be
thought of as a special type of weighted digraph where the cost of each arc is 1. A
weighted graph may be represented as a symmetric digraph where each of a pair of
antiparallel arcs has the same weight.
In Figure 6.1 we display a classic unweighted graph (called the 3-cube) of diame-
ter 3, a digraph with arc weights, and a graph with edge weights.
There are two obvious ways to represent a weighted digraph on a computer. One
is via a matrix. The adjacency matrix is modified so that each entry of 1 (signifying
that an arc exists) is replaced by the cost of that arc. Another is a double adjacency
 0 1 4
1 0 1
0 1
4 1 2
2 3 2
2 3
2 4 4 3
2 2
5
4 5
3 1
3
4 5
6 7 2

Figure 6.1: Some weighted (di)graphs.

list. In this case, the list associated to a node v contains, alternately, an adjacent node
w and then the cost c(v, w).
If there is no arc between u and v, then in an ordinary adjacency matrix the cor-
responding entry is 0. However, in a weighted adjacency matrix, the “cost” of a non-
existent arc should be set consistantly for most applications. We adopt the following
convention. An entry of null or 0 in a weighted adjacency matrix means that the arc
does not exist, and vice versa. In many of our algorithms below, such entries should
be replaced by the programming equivalent of null for class objects, or ∞ for prim-
itive data types. In the later case, we might use some positive integer greater than
any expected value that might occur during an execution of the program.

Example 6.2. The two weighted (di)graphs of Figure 6.1 are stored as weighted ad-
jacency matrices below.
 
  0 4 1 0 4 0
0 1 4 0  4 0 0 2 3 4 
 
 0 0 0 2   1 0 0 0 3 0 
   
 0 2 0 5   0 2 0 0 0 1 
 
2 0 0 0  4 3 3 0 0 2 
0 4 0 1 2 0
The corresponding weighted adjacency lists representations are

1 1 2 4
3 2
1 2 3 5
0 2
and
1 4 2 1 4 4
0 4 3 2 4 3 5 4
0 1 4 3
1 2 5 1
0 4 1 3 2 3 5 2
1 4 3 1 4 2

See Appendix B.4 for sample Java code for representing the abstract data type of
edge-weighted digraphs.

6.2 Distance and diameter in the unweighted case

One important application for graphs is to model computer networks or parallel
processor connections. There are many properties of such networks that can be
obtained by studying the characteristics of the graph model. For example, how do
we send a message from one computer to another, using the fewest intermediate
nodes? This question is answered by finding a shortest path in the graph. We may
also want to know what the largest number of communication links that may be
required for any two nodes to talk with each other; this is equal to the diameter of
the graph.

Definition 6.3. The diameter of a strongly connected digraph G is the maximum of

d(u, v) over all nodes u, v ∈ V (G).

Note. If the digraph is not strongly connected then the diameter is not defined; the
only “reasonable” thing it could be defined to be would be +∞, or perhaps n (since
no path in G can have length more than n − 1).

Example 6.4. The diameter of the 3-cube in Figure 6.1 is easily seen to be 3. Since
the digraph G2 in Figure 4.1 is not strongly connected, the diameter is undefined.

The problem of computing distances in (ordinary, unweighted) digraphs is rela-

tively easy. We already know from Theorem 5.6 that for each search tree, BFS finds
the distance from the root s to each node in the tree (this distance equals the level of
the node). If v is not in the tree then v is not reachable from s and so d(s, v) = +∞ (or
is undefined, depending on your convention).
It is often useful to have a readily available distance matrix. The (i, j)-entry of
this matrix contains the distance between node i and node j. Such a matrix can be
generated by running BFSvisit from each node in turn; this gives an algorithm with
running time in Θ(n2 + nm).

Example 6.5. An adjacency matrix and a distance matrix for the 3-cube shown in
Figure 6.1 is given below. The maximum entries of value 3 indicate the diameter. The
reader should check these entries by performing a breadth-first search from each
vertex.

   
0 1 1 0 0 0 1 0 0 1 1 2 2 3 1 2

 1 0 0 1 0 0 0 1 


 1 0 2 1 3 2 2 1 


 1 0 0 1 1 0 0 0 


 1 2 0 1 1 2 2 3 


 0 1 1 0 0 1 0 0 


 2 1 1 0 2 1 3 2 


 0 0 1 0 0 1 1 0 


 2 3 1 2 0 1 1 2 


 0 0 0 1 1 0 0 1 


 3 2 2 1 1 0 2 1 

 1 0 0 0 1 0 0 1   1 2 2 3 1 2 0 1 
0 1 0 0 0 1 1 0 2 1 3 2 2 1 1 0

It is more difficult to compute distance in weighted digraphs. In the next two

sections we consider this problem.
Exercises
Exercise 6.2.1. Give an example of a weighted digraph in which the obvious BFS
approach does not find the shortest path from the root to each other node.

Exercise 6.2.2. The eccentricity of a node u in a digraph G is the maximum of d(u, v)

over all v ∈ V (G). The radius of G is the minimum eccentricity of a node. Write an
algorithm to compute the radius of a digraph in time Θ(n2 + nm). How can we read
off the radius from a distance matrix?

6.3 Single-source shortest path problem

The single-source shortest path problem (SSSP) is as follows. We are given a
weighted digraph (G, c) and a source node s. For each node v of G, we must find
the minimum weight of a path from s to v (by the weight of a path we mean the sum
of the weights on the arcs).

Example 6.6. In the weighted digraph of Figure 6.1, we can see by considering all
possibilities that the unique minimum weight path from 0 to 3 is 0 1 3, of weight 3.

We first present an algorithm of Dijkstra that gives an optimal solution to the

SSSP whenever all weights are nonnegative. It does not work in general if some
weights are negative—see Exercise 6.3.5.
Dijkstra’s algorithm is an example of a greedy algorithm. At each step it makes
the best choice involving only local information, and never regrets its past choices.
It is easiest to describe in terms of a set S of nodes which grows to equal V (G).
Initially the only paths available are the one-arc paths from s to v, of weight c(s, v).
At this stage, the set S contains only the single node s. We choose the neighbour u
with c(s, u) minimal and add it to S. Now the fringe nodes adjacent to s and u must
be updated to reflect the new information (it is possible that there exists a path from
 algorithm Dijkstra
Input: weighted digraph (G, c); node s ∈ V (G)
begin
array colour[0..n − 1], dist[0..n − 1]
for u ∈ V (G) do
dist[u] ← c[s, u]; colour[u] ← WHITE
end for
dist[s] ← 0; colour[s] ← BLACK
while there is a white node
find a white node u so that dist[u] is minimum
colour[u] ← BLACK
for x ∈ V (G) do
if colour[x] = WHITE then
dist[x] ← min{dist[x], dist[u] + c[u, x]}
end if
end for
end while
return dist
end

Figure 6.2: Dijkstra’s algorithm, first version.

s to v, passing through u, that is shorter than the direct path from s). Now we choose
the node (at “level” 1 or 2) whose current best distance to s is smallest, and update
again. We continue in this way until all nodes belong to S.
The basic structure of the algorithm is presented in Figure 6.2.
Example 6.7. An application of Dijkstra’s algorithm on the second digraph of Fig-
ure 6.1 is given in Table 6.1 for each starting vertex s.
The table illustrates that the distance vector is updated at most n − 1 times (only
before a new vertex is selected and added to S). Thus we could have omitted the lines
with S = {0, 1, 2, 3} in Table 6.1.
Why does Dijkstra’s algorithm work? The proof of correctness is a little longer
than for previous algorithms. The key observation is the following result. By an S-
path from s to w we mean a path all of whose intermediate nodes belong to S. In
other words, w need not belong to S, but all other nodes in the path do belong to S.
Theorem 6.8. Suppose that all arc weights are nonnegative. Then at the top of the
while loop, we have the following properties:
P1: if x ∈ V (G), then dist[x] is the minimum cost of an S-path from s to x;
P2: if w ∈ S, then dist[w] is the minimum cost of a path from s to w.
 Table 6.1: Illustrating Dijkstra’s algorithm.

current S ⊆ V distance vector dist

{0} 0, 1, 4, ∞
{0, 1} 0, 1, 4, 3
{0, 1, 3} 0, 1, 4, 3
{0, 1, 2, 3} 0, 1, 4, 3
{1} ∞, 0, ∞, 2
{1, 3} 4, 0, ∞, 2
{0, 1, 3} 4, 0, 8, 2
{0, 1, 2, 3} 4, 0, 8, 2
{2} ∞, 2, 0, 5
{1, 2} ∞, 2, 0, 4
{1, 2, 3} 6, 2, 0, 4
{0, 1, 2, 3} 6, 2, 0, 4
{3} 2, ∞, ∞, 0
{0, 3} 2, 3, 6, 0
{0, 1, 3} 2, 3, 6, 0
{0, 1, 2, 3} 2, 3, 6, 0

Note. Assuming the result to be true for a moment, we can see that once a node u
is added to S and dist[u] is updated, dist[u] never changes in subsequent iterations.
When the algorithm terminates, all nodes belong to S and hence dist holds the cor-
rect distance information.

Proof. Note that at every step, dist[x] does contain the length of some path from s to
x; that path is an S-path if x ∈ S. Also, the update formula ensures that dist[x] never
increases.
To prove P1 and P2, we use induction on the number of times k we have been
through the while-loop. Let Sk denote the value of S at this stage. When k = 0, S0 =
{s}, and since dist[s] = 0, P1 and P2 obviously hold. Now suppose that they hold after
k times through the while-loop and let u be the next special node chosen during that
loop. Thus Sk+1 = Sk ∪ {u}.
We first show that P2 holds after k + 1 iterations. Suppose that w ∈ Sk+1 . If w 6= u
then w ∈ S and so P2 trivially holds for w by the inductive hypothesis. On the other
hand, if w = u, consider any Sk+1 -path γ from s to u. We shall show that dist[u] ≤ |γ|
where |γ| denotes the weight of γ. The last node before u is some y ∈ Sk . Let γ1 be the
subpath of γ ending at y. Then dist[u] ≤ dist[y] + c(y, u) by the update formula. Fur-
thermore dist[y] ≤ |γ1 | by the inductive hypothesis applied to y ∈ Sk . Thus, combining
 x
y

S
Figure 6.3: Picture for proof of Dijkstra’s algorithm.

these inequalities, we obtain dist[u] ≤ |γ1 | + c(y, u) = |γ| as required. Hence P2 holds
for every iteration.
Now suppose x ∈ V (G). Let γ be any Sk+1 -path to x. If u is not involved then γ is
an Sk path and so |γ| ≤ dist[x] by the inductive hypothesis. Now suppose that γ does
include u. If γ goes straight from u to x, we let γ1 denote the subpath of γ ending at
u. Then |γ| = |γ1 | + c(u, x) ≥ dist[x] by the update formula. Otherwise, after reaching
u, the path returns into Sk directly, emerging from Sk again, at some node y before
going straight to x (see Figure 6.3). Let γ1 be the subpath of γ ending at y. Since P2
holds for Sk , there is a minimum weight Sk -path β from s to y of length dist[y]. Thus
by the update formula,

|γ| = |γ1 | + c(y, x) ≥ |β| + c(y, x) ≥ dist[y] + c(y, x) ≥ dist[x].

Hence P1 holds for all iterations.

The study of the time complexity of Dijkstra’s algorithm leads to many interesting
topics.
Note that the value of dist[x] will change only if x is adjacent to u. Thus if we use a
weighted adjacency list, the block inside the second for-loop need only be executed
m times. However, if using the adjacency matrix representation, the block inside the
for-loop must still be executed n2 times.
The time complexity is of order an + m if adjacency lists are used, and an + n2
with an adjacency matrix, where a represents the time taken to find the node with
algorithm Dijkstra2
Input: weighted digraph (G, c); node s ∈ V (G)
begin
priority queue Q
array colour[0..n − 1], dist[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE
end for
colour[s] ← GREY
Q.insert(s, 0)
while not Q.isEmpty() do
u ← Q.peek(); t1 ← Q.getKey(u)
for each x adjacent to u do
t2 ← t1 + c(u, x)
if colour[x] = WHITE then
colour[x] ← GREY
Q.insert(x,t2 )
else if colour[x] = GREY and Q.getKey(x) > t2 then
Q.decreaseKey(x, t2 )
end if
end for
Q.delete()
colour[u] ← BLACK
dist[u] ← t1
end while
return dist
end

Figure 6.4: Dijkstra’s algorithm, PFS version.

minimum value of dist. The obvious method of finding the minimum is simply to
scan through array dist sequentially, so that a is of order n, and the running time of
Dijkstra is therefore Θ(n2 ). Dijkstra himself originally used an adjacency matrix and
scanning of the dist array.
The above analysis is strongly reminiscent of our analysis of graph traversals in
Section 5.1, and in fact Dijkstra’s algorithm fits into the priority-first search frame-
work discussed in Section 5.5. The key value associated to a node u is simply the
value dist[u], the current best distance to that node from the root s. In Figure 6.4 we
present Dijkstra’s algorithm in this way.
It is now clear from this formulation that we need to perform n delete-min oper-
ations and at most m decrease-key operations, and that these dominate the running
 algorithm BellmanFord
Input: weighted digraph (G, c); node s
begin
array dist[0..n − 1]
for u ∈ V (G) do
dist[u] ← ∞
end for
dist[s] ← 0
for i from 0 to n − 1 do
for x ∈ V (G) do
for v ∈ V (G) do
dist[v] ← min(dist[v], dist[x] + c(x, v))
end for
end for
end for
return dist
end

Figure 6.5: Bellman–Ford algorithm.

time. Hence using a binary heap (see Section 2.5), we can make Dijkstra’s algorithm
run in time O((n + m) log n). Thus if every node is reachable from the source, it runs
in time O(m log n).
The quest to improve the complexity of algorithms like Dijkstra’s has led to some
very sophisticated data structures that can implement the priority queue in such a
way that the decrease-key operation is faster than in a heap, without sacrificing the
delete-min or other operations. Many such data structures have been found, mostly
complicated variations on heaps; some of them are called Fibonacci heaps and 2–3
heaps. The best complexity bound for Dijkstra’s algorithm, using a Fibonacci heap,
is O(m + n log n).
Bellman–Ford algorithm
This algorithm, unlike Dijkstra’s handles negative weight arcs, but runs slower
than Dijkstra’s when all arcs are nonnegative. The basic idea, as with Dijkstra’s al-
gorithm, is to solve the SSSP under restrictions that become progressively more re-
laxed. Dijkstra’s algorithm solves the problem one node at a time based on their
current distance estimate.
In contrast, the Bellman–Ford algorithm solves the problem for all nodes at “level”
0, 1, . . . , n − 1 in turn. By level we mean the minimum possible number of arcs in a
minimum weight path to that node from the source.

Theorem 6.9. Suppose that G contains no negative weight cycles. Then after the i-th
iteration of the outer for-loop, dist[v] contains the minimum weight of a path to v for
all nodes v with level at most i.

Proof. Note that as for Dijkstra, the update formula is such that dist values never
increase.
We use induction on i. When i = 0 the result is true because of our initialization.
Suppose it is true for i − 1. Let v be a node at level i, and let γ be a minimum weight
path from s to v. Since there are no negative weight cycles, γ has i arcs. If y is the
last node of γ before v, and γ1 the subpath to y, then by the inductive hypothesis
we have dist[y] ≤ |γ1 |. Thus by the update formula we have dist[v] ≤ dist[y] + c(y, v) ≤
|γ1 | + c(y, v) ≤ |γ| as required.

The Bellman–Ford algorithm runs in time Θ(nm) using adjacency lists, since the
statement in the inner for-loop need only be executed if v is adjacent to x, and the
outer loop runs n times. Using an adjacency matrix it runs in time Θ(n3 ).
Exercises
Exercise 6.3.1. Run the Bellman–Ford algorithm on the digraph with weighted adja-
cency matrix given below. Choose each node as the source in turn as in Example 6.7.
 
0 6 0 0 7
0 0
 5 −4 8 
0 −2 0 0 0
 
2 0 7 0 0
0 0 −3 9 0

Exercise 6.3.2. Explain why the SSSP problem makes no sense if we allow digraphs
with cycles of negative total weight.

Exercise 6.3.3. The graph shows minimum legal driving times (in multiples of 5
minutes) between various South Island towns. What is the shortest time to drive
legally from Picton to (a) Wanaka, (b) Queenstown and (c) Invercargill? Explain
which algorithm you use and show your work.
 19
NELSON PICTON

5
23

GREYMOUTH 26 BLENHEIM
28
51
50
MURCHISON CHRISTCHURCH

44
90 OMARAMA
16
WANAKA
13 54
16
8
QUEENSTOWN 10

CROMWELL
28
DUNEDIN
33
INVERCARGILL

Exercise 6.3.4. Suppose the input to the Bellman–Ford algorithm is a digraph with
a negative weight cycle. How does the algorithm detect this, so it can exit gracefully
with an error message?

Exercise 6.3.5. Give an example to show that Dijkstra’s algorithm may fail to give
the correct answer if some weights are negative. Make your example as small as
possible. Then run the Bellman–Ford algorithm on the example and verify that it
gives the correct answer.

Exercise 6.3.6. Where in the proof of Dijkstra’s algorithm do we use the fact that all
the arc weights are nonnegative?

6.4 All-pairs shortest path problem

The problem is as follows: given a weighted digraph (G, c), determine for each
u, v ∈ V (G) (the length of ) a minimum weight path from u to v.
It is easy to present this information in a distance matrix.

Example 6.10. For the digraph of Figure 6.1, we have already calculated the all-pairs
distance matrix in Example 6.7:
 
0 1 4 3
4 0 8 2 
6 2 0 4  .
 

2 3 6 0
 algorithm Floyd
Input: weighted digraph (G, c)
begin
array d[0..n − 1, 0..n − 1]
for u ∈ V (G) do
for v ∈ V (G) do
d[u, v] ← c(u, v)
end for
end for
for x ∈ V (G) do
for u ∈ V (G) do
for v ∈ V (G) do
d[u, v] ← min(d[u, v], d[u, x] + d[x, v])
end for
end for
end for
return d
end

Figure 6.6: Floyd’s algorithm.

Clearly we may compute this matrix as above by solving the single-source short-
est path problem with each node taken as the root in turn. The time complexity is
of course Θ(nA) where A is the complexity of our single-source algorithm. Thus run-
ning the adjacency matrix version of Dijkstra n times gives a Θ(n3 ) algorithm, and
the Bellman–Ford algorithm Θ(n2 m).
There is a simpler method discovered by R. W. Floyd. Like the Bellman-Ford algo-
rithm, it is an example of an algorithm design technique called dynamic program-
ming . This is where smaller, less-difficult subproblems are first solved, and the so-
lutions recorded, before the full problem is solved. Floyd’s algorithm computes a
distance matrix from a cost matrix in time Θ(n3 ). It is faster than repeated Bellman–
Ford for dense digraphs and unlike Dijkstra’s algorithm, it can handle negative costs.
For sparse graphs with positive costs repeated Dijkstra is competitive with Floyd, but
for dense graphs they have the same asymptotic complexity. A key point in favour
of Floyd’s algorithm is its simplicity, as can be seen from the algorithm of Figure 6.6.
Floyd’s algorithm is basically a simple triple for-loop.
Note. Observe that we are altering the value of d[u, v] in the update formula. If we
already have a weighted adjacency matrix d, there is no need for the first double
loop. We simply overwrite entries in d via the update formula, and everything works.
Example 6.11. An application of Floyd’s algorithm on the third graph of Figure 6.1
is given below. The initial cost matrix is as follows.
  
0 4 1 ∞ 4 ∞

 4 0 ∞ 2 3 4 


 1 ∞ 0 ∞ 3 ∞ 


 ∞ 2 ∞ 0 ∞ 1 

 4 3 3 ∞ 0 2 
∞ 4 ∞ 1 2 0
In the matrices below, the index k refers to the number of times we have been
through the outer for-loop.

0 4 1 ∞ 4 ∞ 0 4 1 6 4 8 0 4 1 6 4 8
     
 4 0 5 2 3 4   4 0 5 2 3 4   4 0 5 2 3 4 
1 5 0 ∞ 3 ∞ 1 5 0 7 3 9 1 5 0 7 3 9
     
     
∞ 2 ∞ 0 ∞ 1 6 2 7 0 5 1 6 2 7 0 5 1
     
     
 4 3 3 ∞ 0 2   4 3 3 5 0 2   4 3 3 5 0 2 
∞ 4 ∞ 1 2 0 8 4 9 1 2 0 8 4 9 1 2 0
k=1 k=2 k=3

0 4 1 6 4 7 0 4 1 6 4 6 0 4 1 6 4 6
     
 4 0 5 2 3 3   4 0 5 2 3 3   4 0 5 2 3 3 
1 5 0 7 3 8 1 5 0 7 3 5 1 5 0 6 3 5
     
     
6 2 7 0 5 1 6 2 7 0 5 1 6 2 6 0 3 1
     
     
 4 3 3 5 0 2   4 3 3 5 0 2   4 3 3 3 0 2 
7 3 8 1 2 0 6 3 5 1 2 0 6 3 5 1 2 0
k=4 k=5 k=6

In the above matrices we list the entries that change in bold after each increment
of k. Notice that undirected graphs, as expected, have symmetric distance matrices.
Why does Floyd’s algorithm work? The proof is again by induction.
Theorem 6.12. At the bottom of the outer for loop, for all nodes u and v, d[u, v] con-
tains the minimum length of all paths from u to v that are restricted to using only
intermediate nodes that have been seen in the outer for loop.
Note. Given this fact, when the algorithm terminates, all nodes have been seen in
the outer for loop and so d[u, v] is the length of a shortest path from u to v.

Proof. To establish the above property, we use induction on the outer for-loop. Let
Sk be the set of nodes seen after k times through the outer loop, and define an Sk -path
to be one all of whose intermediate nodes belong to Sk . The corresponding value of d
is denoted dk . We need to show that for all k, after k times through the outer for-loop,
dk [u, v] is the minimum length of an Sk -path from u to v.
When k = 0, S0 = 0/ and the result holds. Suppose it is true after k times through the
outer loop and consider what happens at the end of the (k + 1)-st time through the
outer loop. Suppose that x was the last node seen in the outer loop, so Sk+1 = Sk ∪ {x}.
Fix u, v ∈ V (G) and let L be the minimum length of an Sk+1 -path from u to v. Obviously
L ≤ dk+1 [u, v]; we show that dk+1 [u, v] ≤ L.
Choose an Sk+1 -path γ from u to v of length L. If x is not involved then the result
follows by inductive hypothesis. If x is involved, let γ1 , γ2 be the subpaths from u to x
and x to v respectively. Then γ1 and γ2 are Sk -paths and by the inductive hypothesis,
L ≥ |γ1 | + |γ2 | ≥ dk [u, x] + dk [x, v] ≥ dk+1 [u, v].

The proof does not use the fact that weights are nonnegative—in fact Floyd’s al-
gorithm works for negative weights (provided of course that a negative weight cycle
is not present).
Exercises
Exercise 6.4.1. Run Floyd’s algorithm on the matrix of Exercise 6.3.1 and check your
answer against what was obtained there.
Exercise 6.4.2. Suppose the input to Floyd’s algorithm is a digraph with a negative
weight cycle. How does the algorithm detect this, so it can exit gracefully with an
error message?
Exercise 6.4.3.
The matrix M shows costs of direct flights between towns A, B, C, D, E, F (where ∞,
as usual, means that no direct flight exists). You are given the job of finding the
cheapest route between each pair of towns. Solve this problem. Hint: save your
working.  
0 1 2 6 4 ∞
 1 0 7 4 2 11
 
2 7 0 ∞ 6 4 
M=  .
6 4 ∞ 0 ∞ 1 

4 2 6 ∞ 0 3 
∞ 11 4 1 3 0

The next day, you are told that in towns D, E, F, political troubles mean that no
passenger is allowed to both take off and land there. Solve the problem with this
additional constraint.

6.5 Minimum spanning tree problem

In this section, we use “tree” to mean “free tree” throughout. Recall that a tree is
a connected acyclic graph. A spanning tree of a graph G is a spanning subgraph of
G that is itself a tree.
Definition 6.13. Let G be a weighted graph. A minimum spanning tree (MST) is a
spanning tree for G which has minimum total weight (sum of all edge weights).
Note. If all weights are nonnegative and we only want a spanning subgraph with
minimum total weight, this must be a tree anyway (if not, delete an edge from a
cycle and keep a spanning subgraph).
The problem we are concerned with is this: given a weighted graph G, find a MST
for G. There are obvious practical applications of this idea. For example, how can we
cheaply link sites with communication links so that they are all connected?
Example 6.14. In the third graph of Figure 6.1, the tree determined by the edges
{0, 2}, {1, 3}, {3, 5}, {4, 5}, {2, 4}
has total weight 9. It is a tree and has the 5 smallest weight edges, so must be a MST.
One should not search naively through all possible spanning trees: it is known
that there are nn−2 spanning trees for the complete graph Kn , for example!
In this section we present two efficient algorithms to find a MST that (like Dijk-
stra’s algorithm) fall into the category of greedy algorithms.
Each builds up a MST by iteratively choosing an edge greedily, that is, choosing
one with minimum weight, subject to not obviously ruining our chance of extending
to a spanning tree. It turns out that this simple approach works for the MST problem
(obviously, not for all graph optimization problems!). There are other algorithms
with better theoretical complexity for the MST problem, but none is as simple to
understand.
The algorithms can be described in an informal way very easily. The first, Prim’s
algorithm, starts at a root vertex and chooses at each step an edge of minimum
weight from the remaining edges, subject to: (a) adding the edge does not create a
cycle in the subgraph built so far, and (b) the subgraph built so far is connected. By
contrast, Kruskal’s algorithm does not start at a root: it follows rule (a) and ignores
(b). Prim’s algorithm is perhaps easier to program, but Kruskal’s is easier to perform
by hand.
Each algorithm clearly terminates when no more edges can be found that satisfy
the above condition(s). Since Prim’s algorithm maintains acyclicity and connected-
ness, at each stage it has built a subgraph that is a tree. Kruskal’s algorithm main-
tains acyclicity, so it has a forest at each step, and the different trees merge as the
algorithm progresses.
We might first ask why does each algorithm even form a spanning tree (see Exer-
cise 6.5.2). However, even given that a spanning tree is formed, it is not at all obvious
that this spanning tree has minimum possible weight among all spanning trees of
the graph.
We now show the correctness of these algorithms. We may suppose that the
graph is connected. If it is not, we cannot find a spanning tree anyway, and must be
satisfied with a spanning forest. Prim’s algorithm will terminate when it has explored
the first component and must be restarted from a new root in another component.
Kruskal’s algorithm will find a spanning forest without modification.
Theorem 6.15. Prim’s and Kruskal’s algorithms are correct.
Proof. Define a set of edges to be promising if it can be extended in some way to
a MST. Then the empty set is promising since some MST exists. We claim that at
each step, the algorithms above have chosen a promising set of edges. When they
terminate, no further extension of the set is possible (by rule (a) above), and so we
must have a MST.
To prove the claim efficiently, we need a technical fact, as follows. Suppose that
B is a subset of V (G), not containing all the vertices of G, and T a promising set of
edges such that no edge in T leaves B. In other words, either both endpoints are in B
or neither endpoint is in B. Then if e is a minimum weight edge that does leave B (it
has one endpoint in B and one outside) then T ∪ {e} is also promising.
To see this fact, note that since T is promising, it is contained in some MST, U say.
If e ∈ U there is nothing to prove. Otherwise, when we add e to U we create exactly
one cycle. There must be at least one other edge, say e′ , that leaves B, otherwise the
cycle could not close. If we remove e′ we obtain a new tree that spans G and whose
total weight is no greater than the total weight of U. Thus V is also a MST, and since
it contains T ∪ {e}, that set is promising.
Now to prove the claim, suppose that our algorithm has maintained a promising
set T of edges so far, and it has just chosen edge e = {u, v}. If we take B at each step
to be the set of vertices in the tree (Prim) or the set of vertices in the tree containing
u (Kruskal), then we may apply the fact above to conclude that T ∪ {e} is promising.
This concludes the proof of correctness.
The above informal descriptions of MST algorithms can be converted easily to an
algorithm. In Figure 6.7 we present the algorithm. Note how similar Prim’s algorithm
is to Dijkstra’s. The main difference is in the update formula. We also store the PFS
tree, which we elected not to do for Dijkstra.
In Prim’s algorithm, we checked whether a cycle would be created by adding an
edge in the usual way: when exploring {u, v} from u, if v has already been seen and
is not the parent of u, then adding {u, v} creates a cycle. With Kruskal’s algorithm,
we must use another method, since the above test does not work. Both u and v may
have been seen, but may be in different trees.
Observe that if we try to add an edge both of whose endpoints are in the same
tree in the Kruskal forest, this will create a cycle, so the edge must be rejected. On
the other hand, if the endpoints are in two different trees, a cycle definitely will not
be created; rather, the two trees merge into a single one, and we should accept the
edge. We need a data structure that can handle this efficiently. All we need is to be
able to find the tree containing an endpoint, and to merge two trees. The disjoint
sets or union-find ADT is precisely what is needed. It allows us to perform the find
and union operations efficiently.
The complexity of the algorithm depends to a great extent on the data structures
used. The best known for Prim is the same as for Dijkstra, namely O(m + n log n), and
 algorithm Prim
Input: weighted graph (G, c); vertex s ∈ V (G)
begin
priority queue Q
array colour[0..n − 1], pred[0..n − 1]
for u ∈ V (G) do
colour[u] ← WHITE; pred[u] ← NULL
end for
colour[s] ← GREY
Q.insert(s, 0)
while not Q.isEmpty() do
u ← Q.peek()
for each x adjacent to u do
t ← c(u, x)
if colour[x] = WHITE then
colour[x] ← GREY; pred[x] ← u
Q.insert(x,t)
else if colour[x] = GREY and Q.getKey(x) > t then
Q.decreaseKey(x,t); pred[x] ← u
end if
end for
Q.delete()
colour[u] ← BLACK
end while
return pred
end

Figure 6.7: Prim’s algorithm.

for Kruskal O(m log n). The disjoint sets ADT can be implemented in such a way that
the union and find operations in Kruskal’s algorithm runs in almost linear time (the
exact bound is very complicated). So if the edge weights are presorted, or can be
sorted in linear time (for example, if they are known to be integers in a fixed range),
then Kruskal’s algorithm runs for practical purposes in linear time.
Exercises
Exercise 6.5.1. Carry out each of these algorithms on the weighted graph of Fig-
ure 6.1. Do the two algorithms give the same spanning tree?
Exercise 6.5.2. Prove the assertion made above that when Kruskal’s or Prim’s algo-
rithm terminates, the current set of edges forms a spanning tree.
 algorithm Kruskal
Input: weighted graph (G, c)
begin
disjoint sets ADT A
initialize A with each vertex in its own set
sort the edges in increasing order of cost
for each edge {u, v} in increasing cost order do
if not A.set(u) = A.set(v) then
add this edge
A.union(A.set(u), A.set(v))
end if
end for
return A
end

Figure 6.8: Kruskal’s algorithm.

Exercise 6.5.3. Consider the following algorithm for the MST problem. Repeatedly
delete edges from a connected graph G, at each step choosing the most expensive
edge we can, subject to maintaining connectedness. Does it solve the MST problem
sometimes? always?

6.6 Hard graph problems

We have presented several efficient algorithms for some common digraph prob-
lems. All of these algorithms have running time bounded above by a low degree
polynomial in the size of the input. However, there are many essential problems
that currently do not have known polynomial-time algorithms (so-called NP-hard
problems). Some examples are:
• finding the longest path between two nodes of a digraph;
• finding a k-colouring of a graph, for fixed k ≥ 3;
• finding a cycle that passes through all the vertices of a graph (a Hamiltonian
cycle);
• finding a minimum weight path that passes through all the vertices of a weighted
digraph (the travelling salesperson problem or TSP);
• finding the largest independent set in a graph — that is, a subset of vertices no
two of which are connected by an edge;
• finding the smallest vertex cover of a graph—that is, a special subset of vertices
so that each vertex of the graph is adjacent to one in that subset. (However,
 from Exercise 5.9.3, this problem is polynomial-time solvable when restricted
to bipartite graphs.)

Investigating these problems is an active research area in computer science. In

many cases the only approach known is essentially to try all possibilities, with some
rules that prevent the listing of obviously hopeless ones. In some special cases (for
example, graphs that are planar (they can be drawn in the plane without edges
crossing) or are in some sense “close to” trees, much faster algorithms can be de-
veloped.
Exercises
Exercise 6.6.1. Find a Hamiltonian cycle of the graph in Exercise 6.3.3. Try to solve
the TSP for this graph.

Exercise 6.6.2. What is the exact relation between the independent set and vertex
cover problems?

6.7 Notes
Dijkstra’s algorithm was proposed by E. W. Dijkstra in 1959. The Bellman–Ford al-
gorithm was proposed independently by R. Bellman (1958) and L. R. Ford, Jr (1956).
Floyd’s algorithm was developed in 1962 by R. W. Floyd. Prim’s algorithm was pre-
sented by R. C. Prim in 1957 and reinvented by E. W. Dijkstra in 1959, but had been
previously introduced by V. Jarnik in 1930. Kruskal’s algorithm was introduced by J.
Kruskal in 1956.
 Part III

Appendices
Appendix A

Java code for Searching and Sorting

This appendix contains Java implementations for many of the common search and
sorting algorithms presented in the book.

A.1 Sorting and selection

The Java class below contains class methods for sorting integer arrays and for
selecting an array element of a given rank. These algorithms insertion sort, Shell-
sort, mergesort, quicksort, heapsort, and quickselect are described in detail in Chap-
ter 2. We may place them all in a Java public class or they may be cut-and-pasted, as
needed, into a Java application.
// Insertion sort of an input array a of size n
//
public static void insertionSort( int [] a )
{
for( int i = 1; i < a.length; i++ ) {
int tmp = a [ i ];
int k = i - 1;
while( k >= 0 && tmp < a[ k ] ) {
a [k + 1 ] = a[ k ];
k--;
}
a[ k + 1 ] = tmp;
}
}

// Selection sort of an input array a of size n

//
public static void selectionSort( int [] a )
{
for( int i = 0; i < a.length - 1; i++ ) {
int posMin = i;
for( int k = i + 1; k < a.length; k++ ) {
if ( a[ posMin ] > a[ k ] ) posMin = k;
}
if ( posMin != i ) swap( a, i, posMin );
}
}

// Bubble sort of an input array a of size n

//
public static void bubbleSort( int [] a )
{
for( int i = a.length - 1; i > 0; i-- ) {
for( int k = 0; k < i; k++ ) {
if ( a[ k ] > a[ k + 1 ] ) swap( a, k, k + 1 );
}
}
}

// Insertion sort of an input array a of size n:

// a private method used by quicksort and quickselect:
// sorting between the indices lo and hi: 0 <= lo <= hi < n
//
private static void insertionSort( int [] a, int lo, int hi )
{
if ( lo > hi || lo < 0 || hi >= a.length ) {
lo = 0;
hi = a.length - 1;
}
for ( int i = lo + 1; i <= hi; i++ ) {
int tmp = a[i];
int k = i - 1;
while( k >= lo && tmp < a[ k ] ) {
a[ k + 1 ] = a[ k ];
k--;
}
a[ k + 1 ] = tmp;
}
}

// Shell’s sort of an input array a of size n

// using a sequence of gaps by G. Gonnet
//
public static void shellSort( int [] a )
{
for( int gap = a.length/2; gap > 0;
gap = (gap == 2) ? 1 : (int)(gap/2.2) )
for( int i = gap; i < a.length; i++ ) {
int tmp = a[ i ];
int k = i;
while( k >= gap && tmp < a [ k - gap ] ) {
 a[ k ] = a[ k - gap ];
k -= gap;
}
a[ k ] = tmp;
}
}

// Mergesort of an input array a of size n

// using a temporary array to merge data
//
public static void mergeSort( int [] a)
{
int [] tmp = new int[ a.length ];
mergeSort( a, tmp, 0, a.length - 1 );
}

private static void mergeSort( int [] a, int [] tmp,

int left, int right)
{
if ( left < right ) {
int centre = (left + right) / 2;
mergeSort( a, tmp, left, centre);
mergeSort( a, tmp, centre + 1, right);
merge( a, tmp, left, centre + 1, right );
}
}

private static void merge( int [] a, int [] tmp,

int left, int right, int rend )
{
int lend = right - 1;
int tpos = left;
int lbeg = left;

// Main loop
while( left <= lend && right <= rend )
if ( a[ left ] < a[ right ] )
tmp[ tpos++ ] = a[ left++ ];
else
tmp[ tpos++ ] = a[ right++ ];

// Copy the rest of the first half

while( left <= lend )
tmp[ tpos++ ] = a[ left++ ];

// Copy the rest of the second half

while( right <= rend )
tmp[ tpos++ ] = a[ right++ ];

// Copy tmp array back

for( tpos = lbeg; tpos <= rend; tpos++ )
a[ tpos ] = tmp[ tpos ];
}
// Quicksort of an input array a of size n using a median-of-three pivot
// and insertion sort of subarrays of size less than CUTOFF threshold:
//
static public int CUTOFF = 10;

public static void quickSort( int [] a )

{
quickSort( a, 0, a.length - 1 );
}

private static void quickSort( int [] a, int lo, int hi )

{
if ( lo + CUTOFF > hi )
insertionSort( a, lo, hi );
else {
// Sort low, middle, high
int mi = ( lo + hi ) / 2;
if ( a[ mi ] < a[ lo ] ) swap( a, lo, mi );
if ( a[ hi ] < a[ lo ] ) swap( a, lo, hi );
if ( a[ hi ] < a[ mi ] ) swap( a, mi, hi );

// Place pivot p at position hi - 1

swap( a, mi, hi - 1 );
int p = a[ hi - 1];

// Begin partitioning
int i, j;
for ( i = lo, j = hi - 1; ; ) {
while( a[ ++i ] < p );
while( p < a[ --j ] );
if ( i < j ) swap( a, i, j );
else break;
}

// Restore pivot
swap( a, i, hi - 1 );
// Sort small elements
quickSort( a, lo, i - 1 );
// Sort large elements
quickSort( a, i + 1, hi );
}
}

private static void swap( int [] a, int i, int j )

{
int tmp = a[ i ];
a[ i ] = a[ j ];
a[ j ] = tmp;
}

// Heapsort of an input array a of size n

// using percolateDown() and swap() methods
//
public static void heapSort( int [] a )
{
// build a heap
for ( int i = a.length / 2 - 1; i >= 0; i-- )
percolateDown( a, i, a.length );
// successively delete the max and restore the heap
for( int i = a.length - 1; i >= 1; i-- ) {
swap( a, 0, i );
percolateDown( a, 0, i );
}
}

// Heapifying method to restore a heap a[0],...,a[size-1]

// after changing a[ i ]; a child / parent position is one
// greater than an index of the same array element
private static void percolateDown( int [] a, int i, int size )
{
int child;
int parent = i + 1;

for ( child = parent * 2; child < size; child = parent * 2 ) {

if ( a[ parent - 1 ] < a[ child - 1 ] ||
a[ parent - 1 ] < a[ child ] ) {
if ( a[ child - 1 ] < a[ child] ) {
swap( a, parent - 1, child );
parent = child + 1;
} else {
swap( a, parent - 1, child - 1 );
parent = child;
}
} else break;
}
if ( child == size && a[ parent - 1 ] < a[ child - 1 ] )
swap( a, parent - 1, child - 1 );
}

// Counting sort of an input array a of size n

// with elements such that min <= a[i] <= max
// (assuming that this condition holds)
//
public static void countSort( int [] a, int min, int max )
{
int i, j;
int m = max - min + 1;
int [] accum = new int[ m ];
for ( j = 0; j < m; j++ ) accum[ j ] = 0;
for ( i = 0; i < a.length; i++ ) accum[ a[ i ] ]++;
for ( i = j = 0; j < m; j++ ) {
if ( accum[ j ] == 0 ) continue;
while( ( accum[ j ]-- ) > 0 )
a[ i++ ] = j + min;
}
 }

// Quick select in an input array a of size n:

// returns the k-th smallest element in a[ k - 1 ]
//
public static void quickSelect( int[] a, int k )
{
quickSelect( a, 0, a.length - 1, k );
}

private static void quickSelect( int[] a, int lo, int hi, int k )
{
if ( lo + CUTOFF > hi ) {
insertionSort( a, lo, hi );
} else {
// Sort low, middle, high
int mi = ( lo + hi ) / 2;
if ( a[ mi ] < a[ lo ] ) swap( a, lo, mi );
if ( a[ hi ] < a[ lo ] ) swap( a, lo, hi );
if ( a[ hi ] < a[ mi ] ) swap( a, mi, hi );

// Place the pivot p into the rightmost place

swap( a, mi, hi - 1 );
int p = a[ hi - 1 ];

// Begin partitioning
int i, j;
for( i = lo, j = hi - 1; ; ) {
while( a[ ++i ] < p );
while( p < a[ --j ] );
if ( i < j ) swap( a, i, j ); else break;
}

// Restore pivot
swap( a, i, hi - 1 );

// Selection by recursion (the only changed part!)

if ( k - 1 < i ) quickSelect( a, lo, i - 1, k );
else if ( k - 1 > i ) quickSelect( a, i + 1, hi, k );
}
}

A.2 Search methods

We now present several Java methods for searching in an integer array. The algo-
rithms (sequential search and binary search) are described in detail in Chapter 3.

// Sequential search for key in an array a

//
public static int sequentialSearch( int[] a, int key)
throws ItemNotFound
{
for( int i = 0; i < a.length; i++ )
if ( a[ i ] == key ) return i;
throw new ItemNotFound( ‘‘SequentialSearch fails’’ );
}

// Binary search for key in a sorted array a

//
public static int binarySearch( int[] a, int key)
throws ItemNotFound
{
int lo = 0;
int hi = a.length - 1;
int mi;

while( lo <= hi ) {
mi = ( lo + hi ) / 2;
if ( a[ mi ] < key ) lo = mi + 1;
else if( a[ mi ] > key ) hi = mi - 1;
else return mi;
}
throw new ItemNotFound( "BinarySearch fails" );
}

// Binary search using two-way comparisons

//
public static int binarySearch2( int[] a, int key)
throws ItemNotFound
{
if ( a.length == 0 )
throw new ItemNotFound( "Zero-length array" );

int lo = 0;
int hi = a.length - 1;
int mi;

while( lo < hi ) {
mi = ( lo + hi ) / 2;
if ( a[ mi ] < key ) lo = mi + 1;
else hi = mi;
}
if ( a[ lo ] == key ) return lo;
throw new ItemNotFound( "BinarySearch fails" );
}
Appendix B

Java graph ADT

This appendix presents a simplified abstract class for representing a graph abstract
data type (ADT). Although it is fully functional, it purposely omits most exception
handling and other niceties that should be in any commercial level package. These
details would distract from our overall (introductory) goal of showing how to imple-
ment a basic graph class in Java.

Our plan is to have a common data structure that represents both graphs and
digraphs. A graph will be a digraph with anti-parallel arcs; that is, if (u, v) ∈ E then
(v, u) ∈ E also. The initial abstract class presented below requires a core set of meth-
ods needed for the realized graph ADT. It will be extended with the actual internal
data structure representation in the form of adjacency matrix or adjacency lists (or
whatever the designer picks).
package graphADT;

import java.util.ArrayList;
import java.io.BufferedReader;

/*
* Current Abstract Data Type interface for (di)graph classes.
*/
public interface Graph
{
// Need default, copy and BufferedReader constructors
// (commented since Java doesn’t allow abstract constructors!)
//
// public GraphADT();
 // public GraphADT(GraphADT);
// public GraphADT(BufferedReader in);

Right from the beginning we get in trouble since Java does not allow abstract con-
structors. We will leave these as comments and hope the graph class designer will
abide by them. We want to create graphs from an empty graph, copy an existing
graph, or read in one from some external source. In the case of a BufferedReader
constructor the user has to attach one to a string, file or keyboard. We will see exam-
ples later.

We now proceed by presenting the alteration methods required for our graph
class interface.
// data structure modifiers
//
void addVertices(int i); // Add some vertices
void removeVertex(int i); // Remove vertex

void addArc(int i, int j); // Add directed edge

void removeArc(int i, int j); // Remove directed edge

void addEdge(int i, int j); // Add undirected edge

void removeEdge(int i, int j); // Remove undirected edge

This small set of methods allows one to build the graph. We will soon explicitly
define the methods for adding or deleting edges in terms of the two arc methods. An
extended class can override these to improve efficiency if it wants. We now list a few
methods for extracting information from a graph object.
// data structure queries
//
boolean isArc(int i, int j); // Check for arcs
boolean isEdge(int i, int j); // Check for edges

int degree(int i); // Number of neighbours (outgoing)

int inDegree(int i); // Number of incoming arcs

ArrayList<Integer> neighbours(int i); // List of (out-)neighbours

int order(); // Number of vertices

int size(); // Number of edges

// output (default same as representation)

//
String toString();

} // end of interface Graph

 For our implementation, we want a vertex’s degree to equal the number of ver-
tices returned by the neighbours method, which in our implementations will cor-
respond to degree(i). Also, the method isEdge(i,j) will most likely just check
whether isArc(i,j) && isArc(j,i) is true.
Finally, one nice thing to offer is a method to view/save/print a graph. Tradition-
ally in Java we define a toString method for this. Our two actual implementations
will return human viewable adjacency lists or adjacency matrices, depending on the
internal representation.

We have the toString method as an interface requirement for the derived classes
to define. We want a BufferedReader constructor for a graph class to accept its own
toString output. Two common external graph representations are handled by the
methods given below.
public String toStringAdjMatrix()
{
StringBuffer o = new StringBuffer();
o.append(order()+"\n");

for( int i = 0; i < order(); i++ )

{
for( int j = 0; j < order(); j++ )
{
if ( isArc(i,j) ) o.append("1 ");
else o.append("0 ");
}
o.append("\n");
}
return o.toString();
}

public String toStringAdjLists()

{
StringBuffer o = new StringBuffer();
o.append(order()+"\n");

for( int i = 0; i < order(); i++ )

{
for( int j = 0; j < order(); j++ )
{
if ( isArc(i,j) ) o.append(j+" ");
}
o.append("\n");
}
return o.toString();
}

To make things convenient for ourselves we require that the first line of our (two)
external graph representations contain the number of vertices. Strictly speaking,
this is not needed for an 0/1 adjacency matrix. This makes our parsing job easier
and this format allows us to store more than one graph per input stream. (We can
terminate a stream of graphs with a sentinel graph of order zero.)

B.1 Java adjacency matrix implementation

We now define our first usable graph class based on an adjacency matrix repre-
sentation (for graphs and digraphs). This class extends our graph interface Graph.
package graphADT;

import java.io.*;
import java.util.*;

/* Current implementation uses adjacency matrix form of a graph.

*/
public class GraphAdjMatrix implements Graph
{

// Internal Representation and Constructors

//
protected int order; // Number of vertices
protected boolean adj[][]; // Adjacency matrix of graph

public GraphAdjMatrix() // default constructor

{
order = 0;
}

public GraphAdjMatrix(GraphAdjMatrix G) // copy constructor

{
int n = G.order();
if ( n>0 ) { adj = new boolean[n][n]; order = n; }

for (int i = 0; i < n; i++)

for (int j = 0; j < n; j++)
adj[i][j] = G.adj[i][j];
}
public GraphAdjMatrix(Graph G) // conversion constructor
{
int n = G.order();
if ( n>0 ) { adj = new boolean[n][n]; order = n; }

for (int i = 0; i < n; i++)

for (int j = 0; j < n; j++)
if (G.isArc(i,j)) adj[i][j] = true;
}

The default constructor simply creates an empty graph and thus there is no need
to allocate any space. The two copy constructors simply copy onto a new n-by-n ma-
trix the boolean adjacency values of the old graph. Notice that we want new storage
and not an object reference for the copy.

An alternative implementation (as given in the first edition of this textbook) would
also keep an integer variable space to represent the total space allocated. Whenever
we delete vertices we do not want to reallocate a new matrix but to reshuffle the en-
tries into the upper sub-matrix. Then whenever adding more vertices we just extend
the dimension of the sub-matrix.
Our last input constructor for GraphAdjMatrix is now given.
public GraphAdjMatrix(BufferedReader buffer)
{
try
{
String line = buffer.readLine().trim();
String[] tokens = line.split("\\s+");

if (tokens.length != 1)
{
throw new Error("bad format: number of vertices");
}
int n = order = Integer.parseInt(tokens[0]);

if ( n>0 ) adj = new boolean[n][n];

for (int i = 0; i < n; i++)

{
line = buffer.readLine().trim();
tokens = line.split("\\s+");
if (tokens.length != n)
{
throw new Error("bad format: adjacency matrix");
}

for (int j = 0; j < n; j++)

{
int entry = Integer.parseInt(tokens[j]);
adj[i][j] = entry != 0;
}
}
}
catch (IOException x)
{ throw new Error("bad input stream"); }
}

We have tried to minimize the complexity of this BufferedReader constructor.

We do however throw a couple of errors if something does go wrong. Otherwise, this
method simply reads in an integer n denoting the dimension of the adjacency matrix
and then reads in the 0/1 matrix. Notice how the use of the String.split method to
extract the integer inputs.

We next define several methods for altering this graph data structure. The first
two methods allow the user to add or delete vertices from a graph.
// Mutator Methods
//
public void addVertices(int n)
{
assert(0 <= n );
boolean matrix[][] = new boolean[order+n][order+n];

for (int i = 0; i < order; i++)

{
for (int j = 0; j < order; j++)
{
matrix[i][j] = adj[i][j];
}
}
order += n;
adj = matrix;
}

public void removeVertex(int v)

{
assert(0 <= v && v < order);
order--;

for (int i = 0; i < v; i++)

{
for (int j = v; j < order; j++)
{
adj[i][j] = adj[i][j+1];
}
}

for (int i = v; i < order; i++)

{
for (int j = 0; j < v; j++)
{
adj[i][j] = adj[i+1][j];
}
for (int j = v; j < order; j++)
{
adj[i][j] = adj[i+1][j+1];
}
}
}

The removeVertex method is somewhat complicated in that we have to remove

a row and column from the matrix corresponding to the vertex being deleted. We
decided to do this in two passes. The first pass (for variable i < v) simply shifts all
column indices j ≥ v to the left. The second pass (for variable i ≥ v) has to shift en-
tries up by one while also shifting column indices j ≥ v to the left. The user of the
graph should realize that the indices of the vertices change!

Next, we have four relatively trivial methods for adding and deleting arcs (and
edges). Like the mutator methods for checking for valid vertex indices we add some
important assert statements that can be turned on with an option to the java com-
piler for debugging graph algorithms.
// Mutator Methods (cont.)

public void addArc(int i, int j)

{
assert(0 <= i && i < order);
assert(0 <= j && j < order);
adj[i][j] = true;
}

public void removeArc(int i, int j)

{
assert(0 <= i && i < order);
assert(0 <= j && j < order);
adj[i][j] = false;
}

public void addEdge(int i, int j)

{
assert(0 <= i && i < order);
assert(0 <= j && j < order);
adj[i][j] = adj[j][i] = true;
}

public void removeEdge(int i, int j)

{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
adj[i][j] = adj[j][i] = false;
}

The methods to access properties of the graph are also pretty straightforward.
// Access Methods
//
public boolean isArc(int i, int j)
{
assert(0 <= i && i < order);
assert(0 <= j && j < order);
return adj[i][j];
}
 public boolean isEdge(int i, int j)
{
return isArc(i,j) && isArc(j,i);
}

public int degree(int i) // row count

{
assert(0 <= i && i < order);
int sz = 0;
for (int j = 0; j < order; j++)
{
if (adj[i][j]) sz++;
}
return sz;
}

public int inDegree(int i) // column count

{
assert(0 <= i && i < order);
int sz = 0;
for (int j = 0; j < order; j++)
{
if (adj[j][i]) sz++;
}
return sz;
}

Our constant-time method for checking whether an arc is present in a graph is

given above in the method isArc. Unfortunately, we have to check all neighbours
for computing the in- and out- degrees. Also the method, given below, for returning
a list of neighbours for a vertex will need to scan all potential vertices.
public ArrayList<Integer> neighbours(int i)
{
assert(0 <= i && i < order);
ArrayList<Integer> nbrs = new ArrayList<Integer>();

for (int j = 0; j < order; j++)

{
if (adj[i][j]) nbrs.add(j);
}

return nbrs;
}

public int order()

{
return order;
}
 public int size() // Number of arcs (edges count twice)
{
int sz = 0;
// boolean undirected = true;
for (int i = 0; i< order; i++)
{
for (int j = 0; j< order; j++)
{
if ( adj[i][j]) sz++;
// if ( adj[i][j] != adj[j][i] ) undirected = false;
}
}
return sz; // undirected ? sz / 2 : sz;
}

The order of the graph is stored in an integer variable _order. However, we have
to count all true entries in the boolean adjacency matrix to return the size. Notice
that if we are working with an undirected graph this returns twice the expected num-
ber (since we store each edge as two arcs). If we specialize this class we may want to
uncomment the indicated statements to autodetect undirected graphs (whenever
the matrix is symmetric). It is probably safer to leave it as it is written, with the
understanding that the user knows how size is defined for this implementation of
Graph.
// default output is readable by constructor
//
public String toString() { return toStringAdjMatrix(); }

} // end class GraphAdjMatrix

We finish our implementation by setting our output method toString to return

an adjacency matrix. Recall the method toStringAdjMatrix was presented earlier
on page 180.

B.2 Java adjacency lists implementation

We now present an alternate implementation of our graph ADT using the adja-
cency lists data structure. We will use the Java API class Vector to store these lists.
package graphADT;

import java.io.*;
import java.util.*;

/* Current implementation uses adjacency lists form of a graph.

*/
public class GraphAdjLists implements Graph
{
// Internal Representation and Constructors
//
 protected ArrayList<ArrayList<Integer>> adj;

public GraphAdjLists()
{
adj = new ArrayList<ArrayList<Integer>>();
}

public GraphAdjLists(Graph G)
{
int n = G.order();
adj = new ArrayList<ArrayList<Integer>>();
for (int i = 0; i < n; i++)
{
adj.add(G.neighbours(i));
}
}

We use an ArrayList that contains an ArrayList of Integer for our representa-

tion. We decided that the copy constructor for Graph is sufficient in terms of effi-
ciency so do not need to define a specialized copy constructor for GraphAdjLists,
handled automatically by the Java runtime environment. The default constructor
creates a list of no lists (that is, no vertices). For better efficiency, the copy construc-
tor takes over the role of our allocator and appends the neighbour lists of the graph
parameter G directly onto a new adjacency list.
public GraphAdjLists(BufferedReader buffer)
{
try
{
String line = buffer.readLine().trim();
String[] tokens = line.split("\\s+");

if (tokens.length != 1)
{
throw new Error("bad format: number of vertices");
}

adj = new ArrayList<ArrayList<Integer>>();

int n = Integer.parseInt(tokens[0]);

for (int u = 0; u < n; u++)

{
ArrayList<Integer> current = new ArrayList<Integer>();
line = buffer.readLine().trim();
int limit = 0;
if (!line.equals(""))
{
tokens = line.split("\\s+");
limit = tokens.length;
}
 for (int i = 0; i < limit; i++)
{
current.add(Integer.parseInt(tokens[i]));
}
adj.add(current);
}
}
catch (IOException x)
{ throw new Error("bad input stream"); }
}

Our stream constructor reads in an integer denoting the order n of the graph and
then reads in n lines denoting the adjacency lists. Notice that we do not check for
correctness of the data. For example, a graph of 5 vertices could have erroneous
adjacency lists with numbers outside the range 0 to 4. We leave these robustness
considerations for an extended class to fulfil, if desired. Also note that we do not list
the vertex index in front of the individual lists and we use white space to separate
items. A blank line indicates an empty list (that is, no neighbours) for a vertex.
// Mutator Methods
//
public void addVertices(int n)
{
assert(0 <= n);
if ( n > 0 )
{
for (int i = 0; i < n; i++)
{
adj.add(new ArrayList<Integer>());
}
}
}

public void removeVertex(int i)

{
assert(0 <= i && i < order());
adj.remove(i);
Integer I = new Integer(i);
for (int u = 0; u < order(); u++)
{
ArrayList<Integer> current = adj.get(u);
current.remove(I); // remove i from adj lists
for (Integer num: current)
{
if (num > i) // relabel larger indexed nodes
{
int index = current.indexOf(num);
current.set(index, num-1);
}
}
}
 }

Adding vertices is easy for our adjacency lists representation. Here we just ex-
pand the internal _adj list by appending new empty lists. The removeVertex method
is a little complicated in that we have to scan each list to remove arcs pointing to the
vertex being deleted. We also have chosen to relabel vertices so that there are no
gaps (that is, we want vertex indexed by i to be labeled Integer(i)). A good ques-
tion would be to find a more efficient removeVertex method. One way would be
to also keep an in-neighbour list for each vertex. However, the extra data structure
overhead is not desirable for our simple implementation.
public void addArc(int i, int j)
{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
if (isArc(i,j)) return;
(adj.get(i)).add(j);
}

public void removeArc(int i, int j)

{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
if (!isArc(i,j)) return;
(adj.get(i)).remove(new Integer(j));
}

public void addEdge(int i, int j)

{
addArc(i,j);
addArc(j,i);
}

public void removeEdge(int i, int j)

{
removeArc(i,j);
removeArc(j,i);
}

Adding and removing arcs is easy since the methods to do this exist in the Vector
class. All we have to do is access the appropriate adjacency list. We have decided
to place a safeguard in the addArc method to prevent parallel arcs from being added
between two vertices.
// Access Methods
//
public boolean isArc(int i, int j)
{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
 return (adj.get(i)).contains(new Integer(j));
}

public boolean isEdge(int i, int j)

{
return isArc(i,j) && isArc(j,i);
}

public int inDegree(int i)

{
assert(0 <= i && i < order());
int sz = 0;
for (int j = 0; j < order(); j++)
{
if (isArc(j,i)) sz++;
}
return sz;
}

public int degree(int i)

{
assert(0 <= i && i < order());
return (adj.get(i)).size();
}

Note how we assume that the contains method of a Vector object does a data
equality check and not just a reference check. The outDegree method probably
runs in constant time since we just return the list’s size. However, the inDegree
method has to check all adjacency lists and could have to inspect all arcs of the
graph/digraph.
public ArrayList<Integer> neighbours(int i)
{
assert(0 <= i && i < order());
ArrayList<Integer> nei = new ArrayList<Integer>();
for (Integer vert : adj.get(i))
{
nei.add(vert);
}
return nei;
//return (ArrayList<Integer>)(adj.get(i)).clone();
}

public int order()

{
return adj.size();
}

public int size() // Number of arcs (counts edges twice)

{
int sz = 0;
for (int i=0; i<order(); i++)
 {
sz += (adj.get(i)).size();
}
return sz;
}

We do not want to have any internal references to the graph data structure being
available to non-class members. Thus, we elected to return a clone of the adjacency
list for our neighbours method. We did not want to keep redundant data so the order
of our graph is simply the size of the adj list.
// default output readable by constructor
//
public String toString() { return toStringAdjLists(); }

} // end class GraphAdjLists

Again, we have the default output format for this class be compatible with the
constructor BufferedReader. (The method toStringAdjLists is defined on page 180.)

B.3 Standardized Java graph class

We now have two implementations of a graph class as specified by our interface
(abstract class) Graph. We want to write algorithms that can handle either format.
Since Java is object-oriented we could have all our algorithms take a Graph object
and the run-time dynamic mechanism should ascertain the correct adjacency ma-
trix or adjacency lists methods. For example, we could write a graph coloring algo-
rithm prototyped as public int color(Graph G) and pass it either a GraphAdjMatrix
or a GraphAdjLists. And it should work fine!

We next present a simple test program for how one would use our graph imple-
mentations. We encourage the reader to trace through the steps and to try to obtain
the same output.
import java.io.*; import graphADT.*;

public class test {

public static void main(String argv[])

{
Graph G1 = new GraphAdjLists();

G1.addVertices(5);
G1.addArc(0,2); G1.addArc(0,3); G1.addEdge(1,2);
G1.addArc(2,3); G1.addArc(2,0); G1.addArc(2,4);
G1.addArc(3,2); G1.addEdge(4,1); G1.addArc(4,2);

System.out.println(G1);
 Graph G2 = new GraphAdjMatrix(G1);

G2.removeArc(2,0); G2.removeArc(4,1); G2.removeArc(2,3);

System.out.println(G2);

Graph G3 = new GraphAdjLists(G2);

G3.addVertices(2);
G3.addArc(5,4); G3.addArc(5,2); G3.addArc(5,6);
G3.addArc(2,6); G3.addArc(0,6); G3.addArc(6,0);

System.out.println(G3);

Graph G4 = new GraphAdjLists(G3);

G4.removeVertex(4); G4.removeEdge(5,0); G4.addVertices(1);

G4.addEdge(1,6);

System.out.println(G4);
}
} // test

The expected output, using JDK version 1.6, is given in Figure B.1. Note that the
last version of the digraph G has a vertex of out-degree zero in the adjacency lists.
(To compile our program we type ‘javac test.java’ and to execute it we type ‘java test’
at our command-line prompt ‘$’.)

B.4 Extended graph classes: weighted edges

The graph ADT presented in the previous sections can be easily extended to pro-
vide a customized data type. For example, if one only wants undirected graphs then
a more restrictive class can be developed to prevent arc operations. In this section
we want to illustrate how one can develop an ADT for arc-weighted graphs. We first
want to define a new graph interface that allows for these weights.
public interface wGraph extends Graph
{
class Weight<X>
{
private X value;
public Weight(X arg)
{
value = arg;
}
public X getValue()
{
return value;
}
public void setValue(X arg)
{
$ javac test.java
$ java test
5
2 3
2 4
1 3 0 4
2
1 2

5
0 0 1 1 0
0 0 1 0 1
0 1 0 0 1
0 0 1 0 0
0 0 1 0 0

7
2 3 6
2 4
1 4 6
2
2
4 2 6
0

7
2 3
2 6
1 5
2
2 5

Figure B.1: Sample output of the graph test program.

 value = arg;
}
public String toString()
{
return value.toString();
}
}

void addArc(int i, int j); // overridden with "default weight"

void addArc(int i, int j, Weight weight);

void setArcWeight(int i, int j, Weight weight);

//assumes edge i-j exists; and replaces the weight of edge i-j

Weight getArcWeight(int i, int j);

ArrayList<Weight> neighbourWeights(int i);

// If you call neighbours(i) and neighbourWeights(i) then
// the k-th element of both lists are correlated
}

In the above interface, we define a class to represent arbitrary weights. Usually

one uses Weight<Integer> as the arc attributes. Since wGraph extends Graph, the
existing graph algorithms, written for non-weighted graphs, should still work.
We can implement a wGraph in several ways just like we had GraphAdjMatrix and
GraphAdjLists for the interface Graph. We can then create a wGraph object by creat-
ing an implementation object, such as:
wGraph wG = new wGraphMatrix(Buffer);
An adjacency matrix implementation for this interface is given below. Note that
the BufferedReader constructor assumes weights of type Integer. If one wants a
floating point data type then another constructor (or creation method) is required.
package graphADT;

import java.util.ArrayList; import java.io.*;

public class wGraphMatrix implements wGraph

{
protected int order;
protected Weight[][] adjW; // null entry means no arc

public wGraphMatrix()
{
order = 0;
}

public wGraphMatrix(wGraphMatrix G)
{
int n = order = G.order();
if ( n > 0 )
 {
adjW = new Weight[n][n];
}

for (int i = 0; i < n; i++)

{
for (int j = 0; j < n; j++)
{
adjW[i][j] = G.adjW[i][j];
}
}
}

public wGraphMatrix(wGraph G) //convert implementation

{
int n = order = G.order();
adjW = new Weight[n][n];

for (int i = 0; i < n; i++)

{
ArrayList<Integer> nbrs = G.neighbours(i);
ArrayList<Weight> wNbrs = G.neighbourWeights(i);
for (int j = 0; j < nbrs.size(); j++)
{
int index = nbrs.get(j);
adjW[i][index] = wNbrs.get(j);
}
}
}

public wGraphMatrix(Graph G) // promote and/or copy

{
int n = order = G.order();
if ( n > 0 )
{
adjW = new Weight[n][n];
}

for (int i = 0; i < n; i++)

{
for (int j = 0; j < n; j++)
{
if (G.isArc(i, j))
{
adjW[i][j] = new Weight<Integer>(1);
}
}
}
}

public wGraphMatrix(BufferedReader buffer)

{
try
 {
String line = buffer.readLine().trim();
String[] tokens = line.split("\\s+");

if (tokens.length != 1)
{
throw new Error("bad format: number of vertices");
}
int n = order = Integer.parseInt(tokens[0]);

if ( n > 0 )
{
adjW = new Weight[n][n];
}

for (int i = 0; i < n; i++)

{
line = buffer.readLine().trim();
tokens = line.split("\\s+");
if (tokens.length != n)
{
throw new Error("bad format: adjacency matrix");
}

for (int j = 0; j < n; j++)

{
int entry = Integer.parseInt(tokens[j]);
if (entry != 0)
{
adjW[i][j] = new Weight<Integer>(entry);
}
}
}
}
catch (IOException x)
{
throw new Error("bad input stream");
}

// mutator methods

public void addVertices(int n)

{
assert(0 <= n );
Weight weights[][] = new Weight[order+n][order+n];

for (int i = 0; i < order; i++)

{
for (int j = 0; j < order; j++)
{
weights[i][j] = adjW[i][j];
 }
}
order += n;
adjW = weights;
}

public void removeVertex(int v)

{
assert(0 <= v && v < order);
order--;

for (int i = 0; i < v; i++)

{
for (int j = v; j < order; j++)
{
adjW[i][j] = adjW[i][j+1];
}
}

for (int i = v; i < order; i++)

{
for (int j = 0; j < v; j++)
{
adjW[i][j] = adjW[i+1][j];
}
for (int j = v; j < order; j++)
{
adjW[i][j] = adjW[i+1][j+1];
}
}
}

public void addArc(int i, int j)

{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
adjW[i][j] = new Weight<Integer>(1); //default weight
}

public void removeArc(int i, int j)

{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
adjW[i][j] = null;
}

public void addEdge(int i, int j)

{
addArc(i,j); addArc(j,i);
}

public void removeEdge(int i, int j)

{
 removeArc(i,j); removeArc(j,i);
}

public void addArc(int i, int j, Weight weight)

{
assert(0 <= i && i < order());
assert(0 <= j && j < order());
adjW[i][j] = weight;
}

public void setArcWeight(int i, int j, Weight weight)

{
assert(isArc(i, j));
adjW[i][j] = weight;
}

public Weight<?> getArcWeight(int i, int j)

{
assert(isArc(i, j));
return adjW[i][j];
}

// accessor methods

public boolean isArc(int i, int j)

{
assert(0 <= i && i < order);
assert(0 <= j && j < order);
return adjW[i][j] != null;
}

public boolean isEdge(int i, int j)

{
return isArc(i,j) && isArc(j,i);
}

public int inDegree(int i) // column count

{
assert(0 <= i && i < order);
int sz = 0;
for (int j = 0; j < order; j++)
{
if (adjW[j][i] != null) sz++;
}
return sz;
}

public int degree(int i) // row count

{
assert(0 <= i && i < order);
int sz = 0;
for (int j = 0; j < order; j++)
{
 if (adjW[i][j] != null) sz++;
}
return sz;
}

public int order()

{
return order;
}

public int size() // Number of arcs (edges count twice)

{
int sz = 0;
for (int i = 0; i< order; i++)
{
for (int j = 0; j< order; j++)
{
if (adjW[i][j] != null) sz++;
}
}
return sz; // undirected ? sz / 2 : sz;
}

public ArrayList<Integer> neighbours(int i)

{
assert(0 <= i && i < order);
ArrayList<Integer> nbrs = new ArrayList<Integer>();

for (int j = 0; j < order; j++)

{
if (adjW[i][j] != null) nbrs.add(j);
}
return nbrs;
}

public ArrayList<Weight> neighbourWeights(int i)

{
ArrayList<Weight> nbrsWei = new ArrayList<Weight>();

for (int j = 0; j < order(); j++)

{
if (adjW[i][j] != null)
{
nbrsWei.add(adjW[i][j]); // corresponding weight
}
}
return nbrsWei;
}

public String toString() // print weights in n-by-n matrix

{
StringBuffer o = new StringBuffer();
o.append(order()+"\n");
 for (int i = 0; i < order(); i++)
{
for (int j = 0; j < order(); j++)
{
if (adjW[i][j] != null)
{
o.append(adjW[i][j] + " ");
}
else
{
o.append(0 + " ");
}
}
o.append("\n");
}
return o.toString();
}

One thing to note is that if one wants to output the underlying graph represen-
tation (that is, without weights) one can simply call the toString method of Graph
reference.

We conclude by mentioning that the details for an adjacency lists implementa-

tion, wGraphLists, are included in the graph library accompanying this book. We
note that this adjacency lists version of the ADT is more suitable when one expects
weights of numerical value 0 or has sparse graphs.
Appendix C

Background on Data Structures

We assume that the reader is familiar with basic data structures such as arrays and
with the basic data types built in to most programming languages (such as integer,
floating point, string, etc). Many programming applications require the program-
mer to create complicated combinations of the built-in structures. Some languages
make this easy by allowing the user to define new data types (for example Java or C++
classes), and others do not (for example C, Fortran). These new data types are con-
crete implementations in the given language of abstract data types (ADT s), which
are mathematically specified.

C.1 Informal discussion of ADTs

An ADT consists of a set with certain operations on it. How those operations are
to be carried out is not our concern here. It is up to the programmer to choose an
implementation that suits the given application.
Some of the key ADTs are: list, stack, queue, priority queue, dictionary, disjoint
sets. We discuss them each below in turn, semi-formally.
A container is a collection of objects from some universal set U. Basic operations
are to create a new empty container, insert an element, check whether the container
is empty (the isEmpty operation). We assume that each element when inserted re-
turns a locator that identifies its position uniquely.
We can then try to find an element. Depending on the additional operations
defined, this may be easy or difficult. We may need to enumerate all locators. We
can also find the size of a container by enumerating all locators. Again, this may be
very inefficient, and for certain applications a special size operation may be defined.
Similarly, many container ADTs have a delete operation. Some of these allow quick
removal of an element, while others have to find it first, which may be slower.
Other operations can be expressed in terms of the basic ones. For example, we
can sometimes modify or update an element by finding it, remembering its locator,
deleting it, then inserting the new value at the given location. This procedure is
sometimes very inefficient, so a special update operation may be required in some
situations. We normally try to have as few basic operations as possible, and other
operations such as sorting are expressed in terms of these (this is the aim of “generic
programming”).
As we see a container is very general. Some of the important container ADTs are
listed below.
A list is a container that stores elements in a linear sequence. Some basic oper-
ations are to insert in a given position in the sequence, to delete an element at a
given position. To find an element requires sequential search, enumerating all loca-
tors until the element is found or we run out of locators. The first element of a list is
called the head and the last is the tail. A sublist is a contiguous piece of the list, that
can be traversed by the iterator with no gaps. If we divide the list into two sublists
by choosing an element x and letting the head sublist consist of all elements before
x, and the tail sublist consist of all elements after it (either sublist, or neither, may
contain x, depending on the situation).
The main data structures used for implementing a list are arrays and singly- and
doubly-linked lists. They have different properties. For example, to find the middle
element of a list implemented as an array is a constant-time operation but this is not
true for the linked lists, since one must traverse the list, not having direct access to
the middle element as in an array. On the other hand, to insert an element in a given
(nonempty) position in an array takes longer than with the linked list implementa-
tion.
A sorted list is a container of elements from a totally ordered set U, with the same
basic operations as a list, except that the elements are kept sorted in ascending order.
A stack is a restricted kind of list in which insertion and deletion occur at the
same end (“top”) of the list, and at no other position. The basic operations are insert
(also called push), delete (also called pop), and getTop (also called peek) which re-
turns the element at the top of the list without removing it.
A queue is a restricted kind of list in which insertion occurs at one end (the “tail”)
and deletion occurs at the other end (the “head”). The basic operations are insert
(also called enqueue or push), delete (also called dequeue or pop), and getHead (also
called peek) which returns the element at the head without removing it.
A priority queue is a container of elements from a totally ordered set U that al-
lows us to insert an element, to find the smallest element with peek, and to remove
the smallest element with delete. A more advanced operation is decreaseKey which
finds and makes an element smaller. A general delete operation is not usually de-
fined.
 A priority queue can be well implemented using a binary heap, if the operation
decreaseKey is not required to be particularly efficient. Otherwise, more sophisti-
cated implementations are normally used.
A dictionary (also called table, associative array or map) is a a container with
basic operations find, insert, and delete. It is usually also desired to perform an
update operation many times, since dictionaries are often used for dynamic databases.
Some of these ADTs can be used to simulate other ones. For example, a dictionary
can be implemented using a list. Insertion occurs after the last element, and finding
is via sequential search. For practical situations where a dictionary is used, the find
operation is used a lot, so such an inefficient implementation would not normally
be used. Similarly, one can use a list to simulate a priority queue. Insertion occurs at
one end and finding the maximum element by sequential search. Or we could use a
sorted list, where insertion occurs at the correct point and removing the minimum
element is particularly easy, since it is just removing the first element.

C.2 Notes on a more formal approach

The discussion above still doesn’t define the various ADTs in a completely satis-
factory way. As in abstract algebra, we must specify not only the basic operations
but also their properties using a set of axioms. It can be quite difficult to do this suc-
cinctly. Also, whereas in the case of algebra the basic structures (group, ring, field
. . . ) have been agreed on for many decades, the axiomatic definitions of the main
ADTs do not seem to be completely standardized yet. So we shall not give a com-
plete axiomatic presentation, but limit ourselves to an example.
The stack ADT could be defined as follows.
A stack on a set U is a set S with operations push, pop, peek, isEmpty. There is an
empty stack called ε which corresponds to S being the empty set. These operations
do the following: push takes an ordered pair consisting of a stack and an element of
U as an argument, and returns a stack; pop takes a stack as an argument and returns
an element of U; peek takes a stack as an argument and returns either an element of
u or “ERROR”; isEmpty returns either 0 (false) or 1 (true). The axioms for every stack
S and element x of U are as follows.

• isEmpty(ε) = 1
• pop(push(S, x)) = S

• peek(push(S, x)) = x
Appendix D

Mathematical Background

We collect here some basic useful facts, all of which can be found in standard text-
books on calculus and discrete mathematics, to which the reader should refer for
proofs.

D.1 Sets
A set is a collection of distinguishable objects (its elements) whose order is unim-
portant. Two sets are the same if and only if they have the same elements. We denote
the statement that x is an element of the set X by x ∈ X and the negation of this state-
ment by x 6∈ X. We can list finite sets using the braces notation: for example, the set S
consisting of all integers between 1 and 10 that are divisible by 3 is denoted {3, 6, 9}.
A subset of a set X is a set all of whose elements are also elements of X. Each set has
precisely one subset with zero elements, the empty set which is denoted 0. / A subset
can be described by set-builder notation; for example, the subset of S consisting of
all multiples of 3 between 1 and 7 can be written {s | s ∈ S and s ≤ 7}.
For sets X and Y , the union and intersection of X and Y are, respectively, the sets
defined as follows (note that the “or” is inclusive, so “P or Q” is true if and only if P is
true, Q is true, or both P and Q are true):

X ∪Y = {x | x ∈ X or x ∈ Y }
X ∩Y = {x | x ∈ X and x ∈ Y }.

The set difference X \ Y is defined by X \ Y = {x | x ∈ X and x 6∈ Y }. The complement

of a subset S of X is the subset S = X \ S of S. The number of elements (often called
its cardinality) of a set X is often denoted by |X|.
D.2 Mathematical induction
A common way of proving that a statement P(n) is true for all integers n after
some threshold n0 is as follows. It is called the principle of mathematical induction
and is equivalent to the fact that there are no infinite chains of nonnegative integers
x1 > x2 > . . . .
The principle of mathematical induction states that if

• P(n0 ) is true and

• for each n ≥ n0 , if P(n) is true then P(n + 1) is true

then P(n) is true for all n ≥ n0 .

For example, suppose we wish to prove that there exists some constant c > 0 such
that (n + 1)4 /n4 ≤ c for all n ≥ 1 (see below for motivation). We notice that when n = 1
the ratio is 16 and that it appears to get smaller with increasing n (the next few values
of the ratio are 81/16 ∼ = 5 and 256/81 ∼ = 3). So we may guess that c = 16 will work (or
any larger number). So we have that n0 = 1, P(n) is the statement “(n + 1)4 /n4 ≤ 16 for
all n ≥ 1”.
We already know that P(n0 ) is true. Now suppose that P(n) is true for some n ≥ n0 .
We need to show that P(n + 1) is true. We know that

(n + 1)4 ≤ 16n4

and wish to show that

(n + 2)4 ≤ 16(n + 1)4 .
Taking 4th roots we have n+1 ≤ 2n and we want n+2 ≤ 2(n+1). Now using the fact
that P(n) is true (this is called the inductive hypothesis), we have n + 2 = n + 1 + 1 ≤
2n + 1 < 2(n + 1) so the result follows. Thus P(n) is true for all n by the principle of
mathematical induction.
An alternative form of the principle of mathematical induction, which is equiva-
lent to it, is called complete induction:

• P(n0 ) is true and

• for each n ≥ n0 , if P(i) is true for each i with n0 ≤ i ≤ n, then P(n + 1) is true

then P(n) is true for all n ≥ n0 .

D.3 Relations
A relation on a set S is a set R of ordered pairs of elements of S, that is, a subset
of S × S. If (x, y) is a such a pair, we say that x is related to y and sometimes write x R y.
An example is the relation of divisibility on the positive integers; x R y if and only if y
is a multiple of x. Here 2 R 12, 1 R x for every x, and x R 1 only if x = 1.
 There are some special types of relations that are important for our purposes.
An equivalence relation is a relation that is reflexive, symmetric and transitive.
That is, we have for every x, y, z ∈ S
• xRx
• if xRy then yRx
• if xRy and yRz then xRz
An equivalence relation amounts to the same thing as a partition: a decomposi-
tion of S as a union of disjoint subsets. Each subset consists of all elements that are
related to any one of them, and no elements in different subsets of the partition are
related.
Examples of equivalence relations: “having the same mother” on the set of all
humans; “being divisible by 7” on the set of all positive integers; “being mutually
reachable via a path in a given graph”.
Another important type of relation is a partial order. This is a relation that is
reflexive, antisymmetric and transitive. Antisymmetry means that if x R y and y R x
then x = y. Examples are: x R y if and only if x is a factor of y , where x and y are
positive integers.
A linear order or total order is a partial order where every pair of elements is
related. For example, the usual relation ≤ on the real numbers. The elements of a
finite set with a total order can be arranged in a line so that each is related to the next
and none is related to any preceding element.

D.4 Basic rules of logarithms

For x > 1, y > 0, the logarithm logx y to base x of y satisfies the equality: xlogx y = y
and has the following properties:
• logx (a · b) = logx a + logx b
• logx (a/b) = logx a − logx b
• logx (ab ) = b logx a
Using the definition and the properties of the logarithm, it is easy to show that
the following rules hold for x > 1, y > 0, z > 0.
1
logx y = logy x
logy z
logx z = logy x
zlogx y = ylogx z

The last rule is easily proven by taking logarithm to base x of each side of the equality.
 The notation loge = ln is commonly used, as also is log2 = lg.
Often we want to convert the real values returned from functions like logarithms
to integers. Let x be a real number. The floor ⌊x⌋ of x is the greatest integer not greater
than x, and the ceiling ⌈x⌉ of x is the least integer not less than x. If x is an integer,
then ⌊x⌋ = x = ⌈x⌉.

D.5 L’Hôpital’s rule

This rule was in fact proved by J. Bernoulli and sold to G. de L’Hôpital for inclu-
sion in the first calculus textbook, published in 1696. The form that we need for
asymptotics is as follows.
Theorem D.1. If limx→∞ f (x) = ∞ = limx→∞ g(x) and f , g are positive differentiable
functions for x > 0, then

lim f (x)/g(x) = lim f ′ (x)/g′ (x).

x→∞ x→∞

As an example, to compute the limit of ex /x3 as x → ∞, we use the rule repeatedly:

ex ex ex ex
lim = lim = lim = lim = ∞.
x→∞ x3 x→∞ 3x2 x→∞ 6x x→∞ 6

D.6 Arithmetic, geometric, and other series

A general arithmetic series is specified by the recurrence an = an−1 + c, where c is
a constant. The sum of its n terms is
n n(n − 1)
a1 + a2 + . . . + an = (a1 + an ) = na1 + c .
2 2
When a1 = 1 and c = 1,
n(n + 1)
1+2+···+n =
2
A general geometric series is specified by the recurrence an = can−1 , where c 6= 1
is a constant. The sum of its n terms is
cn − 1
a1 + a2 + . . . + an = a1
c−1
If 0 < c < 1, it is better to rewrite the sum as
1 − cn
a1 + a2 + . . . + an = a1
1−c
When n goes to infinity, the sum of the latter infinite geometric series is
∞
a1
∑ ak = 1 − c
k=1
provided |c| < 1, otherwise the infinite sum does not exist.
The sum of squares has an easily guessed explicit formula
n
n(n + 1)(2n + 1)
∑ i2 = 1 + 22 + 32 + . . . + n2 = 6
i=1

which can be proved by induction. Similar explicit formulae hold for the sum ∑ni=1 i p
where p is a fixed positive integer, but they become rather complicated. More useful
for our purposes is the asymptotic formula for large n
n
∑ i p ∈ Θ(n p+1 )
i=1

which also holds for negative integers p provided p 6= −1. When p = −1 we have an
asymptotic statement about the harmonic numbers Hn = ∑ni=1 i−1 ,
n
∑ 1/i ∈ Θ(log n).
i=1

A similar sum of interest to us is

n
log(n!) = ∑ log i.
i=1

This is in Θ(n log n) for any base of the logarithm.

A more complicated formula (Stirling’s approximation), which we do not prove
here, is √
n! > nn e−n 2πn
and in fact, as n → ∞, we have √
n! ≈ nn e−n 2πn,
where f ≈ g is a stronger result than f ∈ Θ(g), namely limn→∞ f (n)/g(n) = 1.
The above asymptotic results can all be proved in the following way using inte-
gral calculus. The method works for any continuous function that is increasing or
decreasing (a monotone function) for x > 0. For example,
Rn
consider f (x) = x3 which is
increasing. Then we may approximate the integral 0 f (x) dx from below by the sum
of rectangles with base length 1 and height f (i), the sum going from 0 to n − 1. See
Figure D.1. Similarly we can approximate it from above by the same sum from 1 to n.
This gives after rearrangement:
Z n n Z n+1
x3 dx ≤ ∑ i3 ≤ x3 dx
1 i=1 1
 Figure D.1: Approximation of an integral by lower rectangles.

and so we have
n
n4 − 1 (n + 1)4 − 1
≤ ∑ i3 ≤ .
4 i=1 4

This easily yields that ∑ni=1 i3 is Θ(n4 ) since (n + 1)4 /n4 ≤ 16 for n ≥ 1 and n4 − 1 ≥
15n4 /16 for n ≥ 2.

D.7 Trees
A rooted ordered tree is what computer scientists call simply a “tree”. These trees
are defined recursively. An ordered rooted tree is either a single node or a distin-
guished node (the root ) attached to some ordered rooted trees given in a fixed order
(hence such a tree is defined recursively). In a picture, these subtrees are usually
drawn from left to right below the parent node. The parent of a node is defined as
follows. The root has no parent. Otherwise the node was attached in some recursive
call, and the root it was attached to at that time is its parent. The roots of the subtrees
in the definition are the children of the root. A rooted ordered tree can be thought
of as a digraph in which there is an arc from each node to each of its children.
A node with no children is called a leaf . The depth of a node is the distance from
the root to that node (the length of the unique path between them). The height of a
node is the length of a longest path from the node to a leaf. The height of tree is the
height of the root. Note that a tree with a single node has height zero. Some other
books use a definition of height whose value equals the value given by our definition,
plus one.
A binary tree is an ordered rooted tree where the number of children of each
node is always 0, 1, or 2.
 A free tree (what mathematicians call a tree) has no order (so a mirror image of
a picture of a tree is a picture of the same tree) and no distinguished root. Every
free tree can be given a root arbitrarily (in n ways, if the number of nodes is n), and
ordered in many different ways.
A free tree can be thought of as the underlying graph of an ordered rooted tree.
A free tree is a very special type of graph. First, if n is the number of nodes and e the
number of edges, then e = n − 1. To see this, note that in the underlying graph of an
ordered rooted tree, each edge connects a node with its parent. Each node except
one has a parent. Thus there is a one-to-one correspondence between nodes other
than the root and edges, yielding the result.
One can easily show that the following are equivalent for a graph G:
• G is a free tree.

• G is a connected graph with e = n − 1.

• G is an acyclic graph with e = n − 1.

Appendix E

Solutions to Selected Exercises

S OLUTION TO E XERCISE 1.1.1 ON PAGE 19:

The equation T (n) = cn2 has only one unknown, c, to be found from T (10) = c × 102 =
500, so c = 500/100 = 5. Then T (1000) = 5 × (1000)2 = 5 × 106 , that is, the algorithm
takes 5 million elementary operations to process 1000 data items.
In fact we do not need to compute c. We first compute T (1000)/T (10) which
equals 106 c/100c = 104 . Thus the answer is 500 × 104 , or 5 million.
S OLUTION TO E XERCISE 1.1.2 ON PAGE 19:
As above, the constants cA and cB have to be found in order to work out how many
elementary operations each algorithm takes with n = 220 and find the fastest algo-
rithm for processing 220 data items. For n = 210 , TA (210 ) = cA × 210 lg(210 ) = 10, so
cA × 210 × 10 = 10, or cA = 1/210 = 2−10 , and TB (210 ) = cB × (210 )2 = 1, so cB = 2−20 .
Hence,
TA (220 ) = 2−10 × 220 × lg(220 ) = 210 × 20 < 215
TB (220 ) = 2−20 × (220 )2 = 220
and TA (220 ) < TB (220 ), so the algorithm A processes 220 data items the fastest.
S OLUTION TO E XERCISE 1.2.1 ON PAGE 21:
The running time is linear because when j = m = 1 the assignment statement makes
m equal to n − 1. Then when j = n − 1, the assignment statement makes m equal to
(n − 1)2 . As the inner loop runs once every time j = m, it runs in total only two times
and does n operations each loop. The outer loop runs n times. Let ci be the constant
number of elementary operations in the inner loop, and let co be the number of el-
ementary operations in the outer loop other than the operations in the inner loop.
Then T (n) = co n + 2ci n ∈ O(n).
S OLUTION TO E XERCISE 1.3.1 ON PAGE 27:
We have 10n3 − 5n + 15 ∈ O(n2 ) if and only if (iff ) there exist a positive real constant
c and a positive integer n0 such that the inequality 10n3 − 5n + 15 ≤ cn2 holds for all
n ≥ n0 . Reducing it to n − 0.5n−1 + 1.5n−2 ≤ 0.1c shows that for any value of c this
inequality does not hold true for all n > k + 1, where k is the closest integer greater
than 0.1c. Therefore, 10n3 − 5n + 15 ∈
/ O(n2 ).
S OLUTION TO E XERCISE 1.3.2 ON PAGE 27:
As above, 10n3 − 5n + 15 ∈ Θ(n3 ) iff there exist positive real constants c1 and c2 and a
positive integer n0 such that the inequalities c1 n3 ≤ 10n3 −5n+15 ≤ c2 n3 , or what is the
same, c1 ≤ 10 − 5n−2 + 15n−3 ≤ c2 , hold true for all n ≥ n0 . We know that limn→∞ (10 −
5n−2 + 15n−3 ) = 10, so there always will be a value n0 > 3 such that for n > n0 , c1 ≤
10 − 5n−2 + 15n−3 ≤ c2 where c1 ≤ 10 − ε and c2 ≥ 10 − ε where ε = 5n−2 −3
0 − 15n0 > 0,
3
say c1 = 1 and c2 = 20. Therefore, 10n − 5n + 15 ∈ Θ(n ).3

Note that Lemma 1.19, the Limit Rule, can be used instead. In this case f (n) =
f (n)
10n − 5n + 15; g(n) = n3 , and f (n) ∈ Θ(g(n)) because limn→∞ g(n)
3 = 10.
S OLUTION TO E XERCISE 1.3.3 ON PAGE 27:
As above, 10n3 − 5n + 15 ∈ Ω(n4 ) iff there exist a positive real constant c and a posi-
tive integer n0 such that the inequality 10n3 − 5n + 15 ≥ cn4 holds for all n > n0 . We
need to show that for any value of c and n0 this inequality, or what is the same, the
reduced one, 10n−1 − 5n−3 + 15n−4 ≥ c, does not hold true for all n > n0 . We know
limn→∞ (10n−1 − 5n−3 + 15n−4) = 0, so no matter which values c and n0 are picked, the
inequality cannot be true for all n > n0 . Therefore, 10n3 − 5n + 15 ∈
/ Ω(n4 ).
S OLUTION TO E XERCISE 1.3.5 ON PAGE 27:
To show that each function f (n) in Table 1.2 stands in “Big Oh” relation to the preced-
ing one, g(n), that is, f (n) ∈ O(g(n)), it is sufficient to use the Limit Rule (Lemma 1.19)
and show that limn→∞ f (n)/g(n) = 0:
n ∈ O(n log n) because n/(n log n) = (log n)−1 and limn→∞ (log n)−1 = 0;
n log n ∈ O(n1.5 ) because n log n/n1.5 = log n/n0.5 and any positive power of n grows faster
than any logarithm (Example 1.14): limn→∞ log n/n0.5 = 0;
n1.5 ∈ O(n2 ) and n2 ∈ O(n3 ) because higher powers of n grow faster than lower powers
(Example 1.20);
n3 ∈ O(2n ) because exponential functions with base greater than 1 grow faster than
any positive power of n (Example 1.13): so limn→∞ n3 /2n = 0.

S OLUTION TO E XERCISE 1.3.6 ON PAGE 27:

Lemma 1.16
Proof. It follows from h(n) ∈ O(g(n)) and g(n) ∈ O( f (n)) that h(n) ≤ c1 g(n) for all n > n1
and g(n) ≤ c2 f (n) for all n > n2 where c1 and c2 are positive real constants and n1
and n2 are positive integers. Then h(n) ≤ c1 g(n) ≤ c1 c2 f (n) for n ≥ max{n1 , n2 }, so the
relationship h(n) ≤ c f (n) for all n ≥ n0 is also true for c = c1 c2 and n0 = max{n1 , n2 }.
Therefore “Big Oh” is transitive.

Lemma 1.17

Proof. It follows from g1 (n) ∈ O( f1 (n)) and g2 (n) ∈ O( f2 (n)) that g1 (n) ≤ c1 f1 (n) for all
n > n1 and g2 (n) ≤ c2 f2 (n) for all n > n2 , respectively, with positive real constants c1
and c2 and positive integers n1 and n2 . Then for all n ≥ max{n1 , n2 } this is also true:

g1 (n) + g2 (n) ≤ c1 f1 (n) + c2 f2 (n) ≤ max{c1 , c2 } ( f1 (n) + f2 (n))

But f1 (n) + f2 (n) ≤ 2 max{ f1 (n), f2 (n)}, so g1 (n) + g2 (n) ≤ c max{ f1 (n), f2 (n)} where c =
2 max{c1 , c2 }. Therefore, g1 (n) + g2 (n) ∈ O(max{ f1 (n), f2 (n)}), and the rule of sums for
“Big Oh” is true.

Lemma 1.18

Proof. It follows from g1 (n) ∈ O( f1 (n)) and g2 (n) ∈ O( f2 (n)) that g1 (n) ≤ c1 f1 (n) for all
n > n1 and g2 (n) ≤ c2 f2 (n) for all n > n2 , respectively, with positive real constants c1
and c2 and positive integers n1 and n2 . Then for all n ≥ max{n1 , n2 } this is also true:
g1 (n)g2 (n) ≤ c f1 (n) f2 (n) where c = c1 c2 . Therefore the rule of products for “Big Oh” is
true.

S OLUTION TO E XERCISE 1.3.7 ON PAGE 27:

The rule of sums for “Big Omega” and “Big Theta” is similar to that for “Big Oh”,
namely,

If g1 ∈ Ω( f1 ) and g2 ∈ Ω( f2 ), then g1 + g2 ∈ Ω(max{ f1 , f2 }), and

If g1 ∈ Θ( f1 ) and g2 ∈ Θ( f2 ), then g1 + g2 ∈ Θ(max{ f1 , f2 }).

S OLUTION TO E XERCISE 1.3.8 ON PAGE 27:

The Lemmas and proofs are similar to the “Big Oh” ones, except the inequalities are
“greater than” instead of “less than”.
Lemma 1.15 (Scaling). For all constants c > 0, c f ∈ Ω( f ), in particular, f ∈ Ω( f ).

Proof. The relationship c f (n) ≥ c f (n) holds for all n > 0. Thus, constant factors are
ignored.

Lemma 1.16 (Transitivity). If h ∈ Ω(g) and g ∈ Ω( f ), then h ∈ Ω( f ).

Proof. It follows from h(n) ∈ Ω(g(n)) and g(n) ∈ Ω( f (n)) that h(n) ≥ c1 g(n) for all n > n1
and g(n) ≥ c2 f (n) for all n > n2 where c1 and c2 are positive real constants and n1 and
n2 are positive integers. Then h(n) ≥ c1 g(n) ≥ c1 c2 f (n), or h(n) ≥ c f (n) is also true for
c = c1 c2 and all n ≥ n0 = max{n1 , n2 }. Therefore Big Omega is transitive.

Lemma 1.17 (Rule of sums). If g1 ∈ Ω( f1 ) and g2 ∈ Ω( f2 ), then g1 +g2 ∈ Ω(max{ f1 , f2 }).

Proof. It follows from g1 (n) ∈ Ω( f1 (n)) and g2 (n) ∈ Ω( f2 (n)) that g1 (n) ≥ c1 f1 (n) for all
n > n1 and g2 (n) ≥ c2 f2 (n) for all n > n2 , respectively, with positive real constants c1
and c2 and positive integers n1 and n2 . Then for all n ≥ max{n1 , n2 } this is also true:

g1 (n) + g2 (n) ≥ c1 f1 (n) + c2 f2 (n) ≥ min{c1 , c2 } ( f1 (n) + f2 (n))

But f1 (n) + f2 (n) ≥ max{ f1 (n), f2 (n)}, so g1 (n) + g2 (n) ≥ c max{ f1 (n), f2 (n)} where c =
min{c1 , c2 }. Therefore, g1 (n) + g2 (n) ∈ Ω(max{ f1 (n), f2 (n)}), and the rule of sums for
“Big Omega” is true.

Lemma 1.18 (Rule of products). Similar to the above lemmas.

S OLUTION TO E XERCISE 1.4.1 ON PAGE 30:
You can make n the subject of the equation for all f (n) except for n lg n. To work out
n lg n, simply guess n until you find the correct value for 1 millennium.

Length of time to run an algorithm

f (n) 1 century 1 millennium
n 5.26 × 109 5.26 × 109
n lg n 6.72 × 107 5.99 × 108
n1.5 1.40 × 106 6.51 · 106
n2 72 522 229 334
n3 3 746 8 071
2n 35 39

S OLUTION TO E XERCISE 1.4.2 ON PAGE 30:

Time complexity Input size n Time T (n)

Function Notation 10 30 100 1000
“log log n” lg lg n 1 1.23 1.42 1.68 lg lg n/ lg lg 10
2
“n log n” n2 lg n 1 13.29 200 30000 n2 lg n/100 lg 10

S OLUTION TO E XERCISE 1.5.1 ON PAGE 36:

The recurrence T (n) = T (n − 1) + n; T (0) = 0, in Example 1.29 implies that T (n) ≥ 0 for
n ≥ 0, so T (n) ≥ n for all n > 0. Therefore, T (n) ∈ Ω(n) for general n. Similarly, T (n) ≥ 0
for all n ≥ 0, so T (n) ≥ n for all n > 0 and therefore, T (n) ∈ Ω(n) in Examples 1.31
and 1.32.
Conversely, in Example 1.30 T (n) ∈ / Ω(n) because T (n) ≤ ⌈lg n⌉ < lg(n + 1) for all
n ≥ 2, and the logarithmic function grows slower than n (Example 1.14).
S OLUTION TO E XERCISE 1.5.2 ON PAGE 36:
The base case holds for n = 2: T (2) = T (1) + T (1) + 2 = 2 < 3 = 2 lg 2 + 2 − 1. By the
inductive hypothesis, T (m) ≤ m lg m + m − 1 = m(lg m + 1) − 1 for all m < n. For an even
n, ⌈ 2n ⌉ = ⌊ 2n ⌋ = 2n , so

T (n) ≤ 2 2n lg n2 + 1 − 1 = n lg n − 2 ≤ n lg n + n − 1; n ≥ 4

For an odd n, ⌈ n2 ⌉ = n+1

2 ; ⌊ 2n ⌋ = n−1
2 , and lg(n − 1) < lg(n + 1) < lg n + 1, so
n+1 n+1
+ n−1 n−1
+ n+1 n−1



T (n) ≤ 2 lg 2 2 lg 2 2 + 2 −2

n+1
≤ 2 lg(n + 1) + n−1
2 lg(n − 1) − 2

n+1
≤ 2 (lg n + 1) + n−1
2 (lg n + 1) − 1 = n lg n + n − 1; n ≥ 3

Therefore, for all n ≥ 2, T (n) ≤ n lg n + n − 1.

S OLUTION TO E XERCISE 1.5.3 ON PAGE 36:
Substituting n = km into the recurrence T (n) = kT nk + cn; T (1) = 0 gives T (km ) =

kT (km−1 ) + ckm ; T (k0 ) = 0. Telescoping the latter recurrence yields:

T (km ) = kT (km−1 ) + ckm

k kT (km−2 ) + ckm−1 + ckm


=
= k2 T (km−2 ) + 2ckm
T (km ) k2 kT (km−3 ) + ckm−2 + 2ckm


=
= k3 T (km−3 ) + 3ckm
....
T (km ) = km T (1) + cmkm
= cmkm

or, what is the same, T (n) = cn logk n. Therefore, T (n) ∈ O(n log n).
S OLUTION TO E XERCISE 1.5.4 ON PAGE 36:
Just as in the previous solution, substituting n = km into the recurrence T (n) = kT ( nk )+
ckn; T (1) = 0, produces T (km ) = kT (km−1 ) + ckm+1 ; T (1) = 0. Telescoping the latter
recurrence gives:

T (km ) = kT (km−1 ) + ckm+1

k kT (km−2 ) + ckm + ckm+1


=
= k2 T (km−2 ) + 2ckm+1
T (km ) k2 kT (km−3 ) + ckm−1 + 2ckm+1


=
= k3 T (km−3 ) + 3ckm+1
....
T (km ) = km T (1) + cmkm+1
= cmkm+1 = ckmkm

or, what is the same, T (n) = ckn logk n. Therefore, T (n) ∈ O(n log n).
S OLUTION TO E XERCISE 1.6.1 ON PAGE 37:
Because n ∈ O(n log n), in “Big-Oh” sense the linear algorithm B has better perfor-
mance than the “n log n” algorithm A. But for small enough n, the latter algorithm is
faster, e.g. TA (10) = 50 and TB (10) = 400 elementary operations. The cutoff point is
when TA (n) = TB (n), that is: 5n log10 n = 40n, or log10 n = 8, or n = 108 . Therefore, even
though the algorithm B is faster in “Big Oh” sense, this only occurs when more than
100 million data items have to be processed.
S OLUTION TO E XERCISE 1.6.2 ON PAGE 38:
In “Big-Oh” sense,√the average-case time complexity of the linear algorithm A is
larger than of the “ n” algorithm B. But for a database of the given size, TA (109 ) = 106
and TB (109 ) = 1.58 × 107 elementary operations. So in this case the algorithm A is, in
the average, over ten times faster than the algorithm B. Because we can tolerate the
risk of an occasional long running time that might occur more likely with the more
complex algorithm, the algorithm A should be used.
S OLUTION TO E XERCISE 2.1.2 ON PAGE 41:
Regardless of the initial ordering of the list, selection sort searches at each iteration
i through the entire unsorted part of the size n − i and makes n − i − 1 comparisons
to find the minimum element, so in total, ∑i=1 n−1
i = n(n−1)
2 ∈ Θ(n2 ) comparisons in the
worst, average, and best case. The maximum number of data moves is n, because
each iteration moves at most one element into its correct position, and their average
number is n2 . Thus, both the maximum and the average individual time complexity
in selection sort is Θ(n) for data moves and Θ(n2 ) for comparisons.
S OLUTION TO E XERCISE 2.2.1 ON PAGE 44:
Adding up the columns in the next table gives, in total, 90 comparisons plus data
moves.
 i Ci Mi Data to sort
91 70 65 50 31 25 20 15 8 2
1 1 1 70 91 65 50 31 25 20 15 8 2
2 2 2 65 70 91 50 31 25 20 15 8 2
3 3 3 50 65 70 91 31 25 20 15 8 2
4 4 4 31 50 65 70 91 25 20 15 8 2
5 5 5 25 30 50 65 70 91 20 15 8 2
6 6 6 20 25 30 50 65 70 91 15 8 2
7 7 7 15 20 25 30 50 65 70 91 8 2
8 8 8 8 15 20 25 30 50 65 70 91 2
9 9 9 2 8 15 20 25 30 50 65 70 91

S OLUTION TO E XERCISE 2.2.2 ON PAGE 44:

The inductive hypothesis is that each inner-loop iteration i = 1, . . . , n − 1 of insertion
sort increases by one the size of the already sorted part a[0], . . . , a[i − 1]) of size i in
the list under consideration, while keeping it sorted. The base case for the math
induction is for i = 0 when the sorted part (a[0]) of size 1 is sorted by definition. At
iteration i, an element temp from the unsorted part of the list is inserted into the
already sorted part between the elements a[ j − 1] and a[ j] such that a[ j − 1] ≤ temp <
a[ j]. Either the left-hand or right-hand inequality may be absent if j = 0 or j = i,
respectively. The obtained new part of size i + 1 is sorted because all the previous
elements smaller than or equal to temp are before it and stay sorted, and all elements
greater than temp are after it and also stay sorted. So because the iterations terminate
when i > n − 1, insertion sort is correct.
Moreover, duplicates will be lumped together and their relative order in the initial
unsorted array will not change, so insertion sort is stable.
S OLUTION TO E XERCISE 2.2.3 ON PAGE 44:
Insertion sort runs the slowest on the totally reverse ordered list that has the maxi-
mum number of inversions: n2 = n(n−1) ∈ Θ(n2 ). The worst-case time complexity of


2
2
insertion sort is Θ(n ) because each swap removes only one inversion.
S OLUTION TO E XERCISE 2.2.4 ON PAGE 45:
Obviously, sorting of elements that precede a[i] (i.e. with positions less than i) does
not change their inversions with a[i]. So the number ν of inversions between a[i] and
the preceding elements is equal to the total number of elements greater than a[i]
among the elements a[0], . . . , a[i − 1]. Just before the iteration i places the element
a[i] into its correct position, the preceding sorted part will have the ν elements; 0 ≤
ν ≤ i, being greater than a[i] and ordered immediately before a[i] at the positions
i − 1, . . . , i − ν. During execution of insertion sort on an array, every element of the
array that is greater than a[i] must be moved up once. Thus the total number of data
moves to insert a[i] is equal to the total number ν of inversions with respect to the
preceding elements in the initial array.
S OLUTION TO E XERCISE 2.2.5 ON PAGE 45:
The out-of-order elements a[i] and a[i + gap] are not equal one to another, so a[i] >
a[i + gap]. The elements at positions less than i or greater than i + gap do not change
their inversions with respect to either a[i] or a[i + gap] after the latter are swapped.
The swap adds no new inversions with the elements, a[k]; i < k < i + gap, between
these positions because all the already ordered pairs such that a[i] < a[k] and a[k] <
a[i + gap] remain ordered after this swap. Since no inversions are added but one
inversion is removed by placing a[i] and a[i + gap] in order, the minimum number of
the inversions removed is 1.
The maximum number of the inversions is removed if for every pair of positions
(i, k) or (k, i+gap) where i < k < i+gap there was an inversion before, but no inversion
after the swap. There are 2 × (gap − 1) such pairs, so the maximum number of the
inversions removed is 2 × gap − 1. Thus, swapping the out-of-order elements a[i] and
a[i + gap] of a list a removes at least 1 and at most 2 × gap − 1 inversions.
S OLUTION TO E XERCISE 2.2.6 ON PAGE 45:
The while loop runs until there are no data swaps in the inner for-loop, so the while
loop stops just after the list is sorted. Each swap of elements a[i] and a[i − 1] in the
inner for-loop removes exactly one inversion, does not affect their inversions with
the preceding or subsequent elements in the list, and obviously does not create any
new inversion. Because the average number of inversions in a list is 21 n2 = n(n−1)


4 ∈
O(n2 ), the average time complexity of bubble sort is O(n2 ).
S OLUTION TO E XERCISE 2.2.8 ON PAGE 46:
 h i Ci Mi Data to sort
5 91 70 65 50 31 25 20 15 8 2
5 1 1 25 91
6 1 1 20 70
7 1 1 15 65
8 1 1 8 50
9 1 1 2 31
2 25 20 15 8 2 91 70 65 50 31
2 1 1 15 25
3 1 1 8 20
4 2 2 2 15 25
5 1 0 20 91
6 1 0 25 70
7 2 1 20 65 91
8 2 1 25 50 70
9 3 2 20 31 65 91
1 2 8 15 20 25 31 50 65 70 91
1 1 0 2 8
2 1 0 8 15
3 1 0 15 20
4 1 0 20 25
5 1 0 25 31
6 1 0 31 50
7 1 0 50 65
8 1 0 65 70
9 1 0 70 91

By adding up the columns, we get a total of 40 comparisons plus data moves.

This is less than half that of insertion sort’s 90, so even with a low n value, Shellsort is
better than insertion sort.
S OLUTION TO E XERCISE 2.3.3 ON PAGE 49:
Insertion sort runs in Θ(n2 ) in the average and worst case, so the total time for this
algorithm is:
 n 2 n2 n 
T (n) = kc + c(k − 1)n = c + c(k − 1)n = cn + (k − 1)
k k k
Assuming the constant n and the variable k, T (n) is minimal for the same value of
k as the function fn (k) = nk + k − 1. Because 1 ≤ k ≤ n, at the boundaries fn (1) = n
and fn (n) = n, and the function is not equal to n at every other k, at least one local
optimum exists in the interval between k = 1 and k = n. For this optimum, the first
n (k)
√ √
derivative by k equals to 0: d fdk = − kn2 + 1, so kn2 = 1, or k = n. Since fn ( n) =
√ √ √
2 n − 1 < n, it is a local minimum. Thus, when k = n, T (n) = cn(2 n − 1) is minimal,
too.
It is not as fast as mergesort’s Θ(n log n) but quicker than insertion sort’s Θ(n2 ).
S OLUTION TO E XERCISE 2.4.1 ON PAGE 55:
Partitioning of a 5-element list after choosing the pivot p = 15 (the bold element is
the pivot; elements in italics are those pointed to by the pointers L and R).
 Data to sort Description
20 8 2 25 15 Initial array
15 8 2 25 20 move pivot to head
15 8 2 25 20 stop R
15 8 2 25 20 stop L (as L = R)
2 8 15 25 20 swap element L with pivot

S OLUTION TO E XERCISE 2.5.1 ON PAGE 63:

Position 1 2 3 4 5 6 7 8 9 10 11 12
Index 0 1 2 3 4 5 6 7 8 9 10 11
Array at step 1 91 75 70 31 65 50 25 20 15 2 8 85
Array at step 2 91 75 70 31 65 85 25 20 15 2 8 50
Array at step 3 91 75 85 31 65 70 25 20 15 2 8 50

S OLUTION TO E XERCISE 2.5.2 ON PAGE 63:

Restoring the heap after deleting the maximum element:
Position 1 2 3 4 5 6 7 8
Index 0 1 2 3 4 5 6 7
Step 1 15 65 50 31 8 2 25 20
Step 2 65 15 50 31 8 2 25 20
Step 3 65 31 50 15 8 2 25 20
Step 4 65 31 50 20 8 2 25 15

S OLUTION TO E XERCISE 2.5.3 ON PAGE 63:

Position 1 2 3 4 5 6 7 8 9
Index 0 1 2 3 4 5 6 7 8
Initial array 10 20 30 40 50 60 70 80 90
i=3 10 20 30 90 50 60 70 80 40
i=2 10 20 70 90 50 60 30 80 40
i=1 10 90 70 20 50 60 30 80 40
10 90 70 80 50 60 30 20 40
i=0 90 10 70 80 50 60 30 20 40
90 80 70 10 50 60 30 20 40
90 80 70 40 50 60 30 20 10
Max heap 90 80 70 40 50 60 30 20 10

S OLUTION TO E XERCISE 2.5.4 ON PAGE 63:

 Position 1 2 3 4 5 6 7 8 9 10
Index 0 1 2 3 4 5 6 7 8 9
Initial array 2 8 15 20 25 31 50 65 70 91
Building the maximum heap
i=4 2 8 15 20 91 31 50 65 70 25
i=3 2 8 15 70 91 31 50 65 20 25
i=2 2 8 50 70 91 31 15 65 20 25
i=1 2 91 50 70 8 31 15 65 20 25
2 91 50 70 25 31 15 65 20 8
i=0 91 2 50 70 25 31 15 65 20 8
91 70 50 2 25 31 15 65 20 8
91 70 50 65 25 31 15 2 20 8
Max heap 91 70 50 65 25 31 15 2 20 8
Deleting max 1 8 70 50 65 25 31 15 2 20 91
Restoring heap 1-9 70 65 50 20 25 31 15 2 8 91
Deleting max 2 8 65 50 20 25 31 15 2 70 91
Restoring heap 1-8 65 25 50 20 8 31 15 2 70 91
Deleting max 3 2 25 50 20 8 31 15 65 70 91
Restoring heap 1-7 50 25 31 20 8 2 15 65 70 91
Deleting max 4 15 25 31 20 8 2 50 65 70 91
Restoring heap 1-6 31 25 15 20 8 2 50 65 70 91
Deleting max 5 2 25 15 20 8 31 50 65 70 91
Restoring heap 1-5 25 20 15 2 8 31 50 65 70 91
Deleting max 6 8 20 15 2 25 31 50 65 70 91
Restoring heap 1-4 20 8 15 2 25 31 50 65 70 91
Deleting max 7 2 8 15 20 25 31 50 65 70 91
Restoring heap 1-3 15 8 2 20 25 31 50 65 70 91
Deleting max 8 2 8 15 20 25 31 50 65 70 91
Restoring heap 1-2 8 2 15 20 25 31 50 65 70 91
Deleting max 9 2 8 15 20 25 31 50 65 70 91

S OLUTION TO E XERCISE 2.5.5 ON PAGE 63:

No, because the creation of a heap does not preserve the order of equal keys. Simi-
larly, quicksort is also unstable, but insertion sort and mergesort are stable.
S OLUTION TO E XERCISE 2.5.6 ON PAGE 63:
The only significant increase in running time is at the very beginning when the heap
is first created: since data items are in the wrong order, each element has to be per-
colated down to the leaves. But since this step is of O(n) complexity, the Θ(n log n)
running time for the deletion of all the max values dominates. Hence, the running
time does not differ significantly from the average-case one.
S OLUTION TO E XERCISE 2.6.3 ON PAGE 65:
In the average, quickselect does p linear operations, Tselect (n, p) = pc1 n, while quick-
sort does one O(n log n) operation, Tsort (n, p) = c2 n lg n. So quicksort is quicker in find-
ing p keys, Tsort (n, p) < Tselect (n, p), when c2 n lg n < pc1 n, or p > cc21 lg n. Because quick-
select is similar to quicksort except of skipping one half at each step, we can as-
sume that c1 ≈ c2 , so that quicksort is better if p > lg n. When n = 106 and p = 10,
10 < lg n = 19.9. Therefore, the variant with quickselect should be used.
S OLUTION TO E XERCISE 2.6.4 ON PAGE 66:
Both heapselect and mergeselect have to order the list first before selecting the kth
smallest item. So in both the cases the average-case and the worst-case time com-
plexity is Θ(n log n).
S OLUTION TO E XERCISE 2.7.1 ON PAGE 67:
Let us show that the sum of all heights of leaves of a decision tree with k leaves is at
least k lg k. The smallest height is when the tree is balanced so that the number of
leaves on the left subtree is equal to the number of leaves on the right subtree. Let
H(k) be the sum of all heights of k leaves in such a tree. Then the

left and the right
subree attached to the root have 2k leaves each and H(k) = 2H 2k + k because the link
to the root adds one to each height. Therefore, H(k) = k lg k. In any other decision
tree, the sum of heights cannot be smaller than H(k). When k = n!, or the number of
leaves is equal to the number of permutations of an array of n keys, H(n!) = n! log n!.
The average height of a leaf, given that each permutation has equal chance, is ob-
tained by dividing the total of all heights by the total number of leaves:

H(n!)
Havg (n!) = = log n! ≈ n log n − 1.44n
n!
This means that the lower bound of the average-case complexity of sorting n
items by pairwise comparisons is Ω(n log n).
S OLUTION TO E XERCISE 2.7.2 ON PAGE 67:
The time complexity is linear, Θ(n), as it takes n steps to scan through array a, and
then a further n steps to print out the contents of t. Theorem 2.35 says that any
algorithm that sorts by comparing only pairs of elements must use at least ⌈log n!⌉
comparisons in the worst case. This algorithm uses the specific knowledge that the
contents of the array a are integers in the range 1..1000. Thus, this algorithm would
not work if the keys can only be compared to each other because contrary to this
case their absolute values are totally unknown.
S OLUTION TO E XERCISE 3.2.1 ON PAGE 76:
It will be identical to Figure 3.3 except the very last step will not return 4, but find
instead that a[m] > k so r ← m − 1 and l > r, so that the loop will terminate and return
“not found”.
S OLUTION TO E XERCISE 3.2.2 ON PAGE 76:
Binary search halves the array at each step, thus the worst case is when it does not
find the key until there is only one element left in the range. Using the improved
binary search that only does one comparison to split the array, we are looking for
the smallest integer k such that 2k ≥ 106 , or k = ⌈lg 106 ⌉ = 20. Thus 20 comparisons
are needed to reduce the range to 1, and in total there are 21 comparisons as at the
end the comparison to the key is done.
S OLUTION TO E XERCISE 3.2.5 ON PAGE 77:

Index 0 1 2 3 4 5 6 7 8 9 next index

 85−10m 
Step 1 10 20 35 45 55 60 75 80 85 100 8 = 0 + 100−10 ·9
85−85
Step 2 85 100 8 = 8 + ⌈ 100−85 · 1⌉
Step 3 85 return value: 8

Interpolation search will search through three positions.

S OLUTION TO E XERCISE 3.3.1 ON PAGE 82:
In line with the ordering relation of a BST, the search for the maximum key starts at
the root, repeatedly goes right while the right child exists, and stops at the node with
the largest key. The search for the smallest key is similar, except it moves left instead
of right.
The running time of these algorithms for a BST with n nodes depends on the tree
shape: it is Θ(n) in the worst case and Θ(log n) in the best and the average cases.
To find the median or any other rank statistic, notice that the root of a BST be-
haves like the pivot in quickselect or quicksort because all the keys to the left or to
the right are smaller or greater, respectively. In general, the rank of a key is equal to
the rank of its node defined as 1 plus the number of nodes above or to the left in-
cluding the number of nodes in its left subtree. A key of rank k, i.e. the k-th smallest
key, 1 ≤ k ≤ n, is found by a recursive tree search controlled, like in quickselect, by
the rank r of the root. If r = k then the goal key is in the root. Otherwise, it is the
k-th smallest key in the left subtree if k < r or the (k − r)-th smallest key in the right
subtree if k > r. The figure below illustrates this search. The size of the subtree below
a node (including the node itself ) is in italic. The root contains the 6th smallest key,
since its left subtree (not including the root itself ) is of size 5. The 4th smallest key, 3,
has left subtree of size 3. The 9th smallest key, 10, is the 3rd smallest key in the right
subtree (3 = 9 − 5 − 1) and has left subtree of size 2.
0 1 2 3 4 5 6 8 10 12 15 18 key
size: 12 5 1 2 3 4 5 6 7 8 9 10 11 12 rank k
1 2 3 4 5 6 k−S left −1

3 Sleft = 5 10 size: 6 rank in the right subtree

1 3 4 1 8 2 15 3

0 2 6 12 18 1
1 1 1 1

S OLUTION TO E XERCISE 3.3.2 ON PAGE 82:

Under all possible insertion sequences, more balanced trees appear more frequently
than unbalanced ones. Thus the balanced trees (and therefore their shapes) dis-
played below will occur most often.

S OLUTION TO E XERCISE 3.3.3 ON PAGE 82:

As was shown in Figure 3.9, the insertion orders 1423, 1432, 1243, 1234, 4312, 4321,
4132, and 4123 yield a tree of the maximal height 3, while 2134, 2143, 2314, 2341, 2413,
2431, 3124, 3142, 3214, 3241, 3412, and 3421 yield a tree of the minimal height 2.
S OLUTION TO E XERCISE 3.3.4 ON PAGE 82:
Just as in the above solution to Exercise 3.3.1, notice that according to the ordering
relation, the root of a BST behaves like the pivot in quicksort because all the keys to
its left or to its right are smaller or greater, respectively. Therefore, all the keys can
be sorted in ascending order by a recursive tree traversal similar to quicksort in that
the left subtree is visited first, then the root, then the right subtree, so the output of
records in order of visiting the nodes of the BST yields the ascending order of their
keys.
S OLUTION TO E XERCISE 3.4.1 ON PAGE 88:
Several solutions exist, and two red-black trees with two black nodes along “root–
leaf” paths are shown here.

S OLUTION TO E XERCISE 3.4.2 ON PAGE 88:

Several solutions exist, and two AVL trees with 7 nodes are shown below along with
the complete binary tree, which is also an AVL tree.
S OLUTION TO E XERCISE 3.4.3 ON PAGE 88:
One AA tree with two black nodes along “root-leaf” paths, of several possible solu-
tions, is shown below.

S OLUTION TO E XERCISE 3.5.1 ON PAGE 100:

For example, two strings s1 = bC and s2 = ab have just the same hash code 31 × 98 +
67 = 31 × 97 + 98 = 3105. Generally, the same hash code is for a pair of 2-letter strings,
s1 and s2 , such that h(s1 ) = h(s2 ), that is, 31s1 [0] + s1 [1] = 31s2 [0] + s2 [1], or 31(s1 [0] −
s2 [0]) = s2 [1] − s1 [1]. Let for definiteness, s1 [0] > s2 [0]. Positive differences between
the Unicode values of two capitals (A,B,. . . ,Z) or two small letters (a,b,. . . ,z) are in the
range of [0..25], and between the values of a small and a capital letter are in the range
of [7..57]. Therefore, only a single pair of the differences, s1 [0] − s2 [0] = 1 and s2 [1] −
s1 [1] = 31, results in the same hash codes of two different 2-letter strings. Because
there are 25 + 25 = 50 pairs of the first letters such that s2 [0] = s1 [0] − 1 and 25 pairs
of the second letters such that s2 [1] = s1 [1] + 31, in total 50 × 25 = 750 pairs out of all
possible 522 = 2704 2-letter strings have the same hash code.
For the even n, the hash function is reduced to the weighted sum of hash codes
for the successive 2-letter substrings:

h(s) = 31n−2 (31s[0] + s[1]) + 31n−4 (31s[2] + s[3]) + . . . + (31s[n − 2] + s[n − 1])

Let Sm = {(si:0 ; si:1 ) : i = 1, . . . , m} be an arbitrary subset of m different 2-letter pairs

having each the same hash code hi . Then 2m strings s, size of 2m letters each, with
the same hash code h(s) = 312m−2 h0 + 312m−4 h1 + . . . + 31hm−2 + hm−1 can be built by
concatenating the m substrings, one from each pair: s = s1:α1 s2:α2 . . . sm:αm providing
si:0 or si:1 are selected in accord with the i-th binary digit αi of the binary number
between 0 and 2m − 1. To form 2100 such strings, we need to select m = 100 arbitrary
2-letter pairs, S100 , having each the same hash code.
S OLUTION TO E XERCISE 3.5.2 ON PAGE 101:

Hash table index 0 1 2 3 4 5 6 7 8 9 10 11 12

Insert 10 10
Insert 26 26 10
Insert 52, collision at 0 26 10 52
Insert 76 26 10 76 52
Insert 13, collision at 0 26 13 10 76 52
Insert 8 26 8 13 10 76 52
Insert 3 26 3 8 13 10 76 52
Insert 33 26 3 33 8 13 10 76 52
Insert 60, collision at 8 26 3 60 33 8 13 10 76 52
Insert 42, collision at 3 26 42 3 60 33 8 13 10 76 52
Resulting hash table 26 42 3 60 33 8 13 10 76 52

S OLUTION TO E XERCISE 3.5.3 ON PAGE 101:

Hash table index 0 1 2 3 4 5 6 7 8 9 10 11 12

Insert 10 10
Insert 26 26 10
Insert 52, collision at 0, ∆ = 4 26 52 10
Insert 76 26 52 10 76
Insert 13, collision at 0, ∆ = 1 26 52 10 76 13
Insert 8 26 8 52 10 76 13
Insert 3 26 3 8 52 10 76 13
Insert 33 26 3 33 8 52 10 76 13
Insert 60, collision at 8, ∆ = 4 26 3 60 33 8 52 10 76 13
Insert 42, collision at 3, ∆ = 3 26 42 3 60 33 8 52 10 76 13
Resulting hash table 26 42 3 60 33 8 52 10 76 13

S OLUTION TO E XERCISE 3.5.4 ON PAGE 101:

Hash table index 0 1 2 3 4 5 6 7 8 9 10 11 12

Insert 10 10
Insert 26 26 10
Insert 52, collision at 0 {26,52} 10
Insert 76 {26,52} 10 76
Insert 13, collision at 0 {26,52,13} 10 76
Insert 8 {26,52,13} 8 10 76
Insert 3 {26,52,13} 3 8 10 76
Insert 33 {26,52,13} 3 33 8 10 76
Insert 60, collision at 8 {26,52,13} 3 33 {8,60} 10 76
Insert 42, collision at 3 {26,52,13} {3,42} 33 {8,60} 10 76
Resulting hash table {26,52,13} {3,42} 33 {8,60} 10 76

S OLUTION TO E XERCISE 4.1.1 ON PAGE 109:

Consider a set of arcs E and nodes V , we want to construct the digraph G = (V, E ′ )
where E ′ = E. We do this by taking the empty digraph G′ = (V, {}) and for each arc in
E add it to the graph G′ . Each time we add an arc (u, v) ∈ E to G′ , we are adding one
to the outdegree of node u and one to the indegree of v. When we have added all the
arcs we get the graph G and since every time the outdegree of a node increased, the
indegree of a node also increased. Hence, the sum of the outdegrees equal the sum
of the indegrees.
For a graph, an analogous statement would be that the sum of all the degrees of
all vertices is equal to twice the number of edges in the graph. This is because each
edge contains two different vertices.
S OLUTION TO E XERCISE 4.1.2 ON PAGE 110:
Given that the path exists between nodes u and v, then there exists a sequence of
nodes u, v1 , v2 , . . . , vk , v such that no node is repeated. If no node is repeated then
k ≤ n − 2 where n is the number of nodes in the digraph. If k was any greater, than
a node would be repeated, and it would contain a cycle that can be eliminated. As
d(u, v) ≤ k + 1, d(u, v) ≤ n − 1.
S OLUTION TO E XERCISE 4.1.3 ON PAGE 110:
We have defined sparse being if the number of edges being O(n) where n is the num-
ber of nodes. And dense is when the number of edges are Ω(n2 ).
The sum of the indegrees of the digraph is also the number of arcs m. The average
indegree of a node is mn .
For sparse graphs, the number of arcs m ∈ O(n), or m ≤ cn for some fixed constant
c independent of n. Thus, the average indegree is less than or equal to c, which is
O(1).
For dense graphs, the number of edges m ∈ Ω(n2 ), or m ≥ cn2 for some fixed con-
stant c independent of n. Thus, the average is greater than or equal to cn, which is
Ω(n).
S OLUTION TO E XERCISE 4.2.1 ON PAGE 112:
 
0 0 1 0 0 0 0
 1 0 0 0 0 0 0 
 
 1 1 0 0 0 0 0 
 
 0 0 0 0 1 1 1 
 
 0 0 0 0 0 1 0 
 
 0 0 0 1 1 0 1 
0 1 1 0 0 0 0

S OLUTION TO E XERCISE 4.2.2 ON PAGE 113:

 7
1 4 5
0 3
0 6
0 5
5

S OLUTION TO E XERCISE 4.2.3 ON PAGE 113:

 
0 1 1 0 0 0 0
 0 0 1 0 0 0 1 
 
 1 0 0 0 0 0 1 
 
 0 0 0 0 0 1 0 
 
 0 0 0 1 0 1 0 
 
 0 0 0 1 1 0 0 
0 0 0 1 0 1 0

S OLUTION TO E XERCISE 4.2.4 ON PAGE 113:

G:  
0 1 1 1 1 1 1 1 1 1 1 1
 0 0 0 1 0 1 0 1 0 1 0 1 
 
 0 0 0 0 0 1 0 0 1 0 0 1 
 
 0 0 0 0 0 0 0 1 0 0 0 1 
 
 0 0 0 0 0 0 0 0 0 1 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 1 
 
 0 0 0 0 0 0 0 0 0 0 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 0 
 
 0 0 0 0 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0
 Gr :  
0 0 0 0 0 0 0 0 0 0 0 0

 1 0 0 0 0 0 0 0 0 0 0 0 


 1 0 0 0 0 0 0 0 0 0 0 0 


 1 1 0 0 0 0 0 0 0 0 0 0 


 1 0 0 0 0 0 0 0 0 0 0 0 


 1 1 1 0 0 0 0 0 0 0 0 0 


 1 0 0 0 0 0 0 0 0 0 0 0 


 1 1 0 1 0 0 0 0 0 0 0 0 


 1 0 1 0 0 0 0 0 0 0 0 0 


 1 1 0 0 1 0 0 0 0 0 0 0 

 1 0 0 0 0 0 0 0 0 0 0 0 
1 1 1 1 0 1 0 0 0 0 0 0

S OLUTION TO E XERCISE 5.1.2 ON PAGE 121:

At each time we turn right, else if its a dead end back up as little as possible. What
we are doing is applying to the visit algorithm (Figure 5.1) a heuristic, a method of
solving a problem, to the ’choose a grey node u’ statement. The visit algorithm allows
for any way of choosing which grey node to choose next, so the way specified in the
exercise is a valid way that will, eventually, result in finding the exit.
S OLUTION TO E XERCISE 5.3.2 ON PAGE 127:
When DFS is run on the graph, the follow timestamps are obtained.

v 0 1 2 3 4 5 6
seen[v] 0 2 1 6 7 8 9
done[v] 5 3 4 13 12 11 10

Tree arcs: (0,2),(2,1),(3,4),(4,5),(5,6)

Forward arcs: (3,5),(3,6)
Back arcs: (1,0),(2,0),(5,3),(5,4)
Cross arcs: (6,1),(6,2)
S OLUTION TO E XERCISE 5.3.4 ON PAGE 127:
By using the defintions of these arcs and also Theorem 5.5, we can prove each of the
three statements.
Let (v, w) ∈ E(G) be an arc. A forward arc is an arc such that v is an ancestor of w in
the tree and is not a tree arc. If the arc (v, w) is in the tree (a tree arc), then v is still an
ancestor of w. Thus, the arc (v, w) is a tree or forward arc if and only if v is an ancestor
of w.
By Theorem 5.5, v is an ancestor of w is equivalent to:

seen[v] < seen[w] < done[w] < done[v]

 The arc (v, w) is a back arc if w is an ancestor of v in the tree and is not a tree arc. If
w is an ancestor of v then it cannot be a tree arc by the above proof. This means that
the arc is a back arc if and only if w is an ancestor of v.
By Theorem 5.5, w is an ancestor of v is equivalent to:

seen[w] < seen[v] < done[v] < done[w]

The arc (v, w) is a cross arc if neither v nor w are the ancestor of the other. But, we
only need to check if w is not an ancestor of v. This is because of the order that DFS
visits these nodes, if it visits v before w (seen[v] < seen[w]) then this will be a tree or
forward arc. So to be a cross arc, seen[w] < seen[v] must be true. This means that the
arc (v, w) is a cross arc if and only if w is not an ancestor of v.
By Theorem 5.5, w is not an ancestor of v is equivalent to:

seen[w] < done[w] < seen[v] < done[v]

S OLUTION TO E XERCISE 5.3.5 ON PAGE 128:

(i) Simply apply the rules found in Exercise 5.3.4, the table entries are the type of arc
(u, v) would be if the arc existed.

u v
0 1 2 3 4 5 6
0 Tree Forward Tree Forward Forward Forward
1 Back Tree Cross Forward Forward Forward
2 Back Back Cross Forward Tree Forward
3 Back Cross Cross Cross Cross Cross
4 Back Back Back Cross Back Cross
5 Back Back Back Cross Tree Tree
6 Back Back Back Cross Cross Back

Of course, if some of these arcs existed, and DFS was run on the graph, some of
the timestamps would change.

(ii) It would be a back arc.

(iii) No, because they are on different branches of the tree (hence if (3, 2) was an arc
in the graph, it would be a cross arc)

(iv) No, because then when DFS is at time 4, instead of expanding node 4, it would
expand node 3, and the DFS tree would be entirely different.

(v) Yes, because it is DFS the tree would be the same (but not if it was BFS), the arc
would be a forward arc.
S OLUTION TO E XERCISE 5.3.9 ON PAGE 129:
The order of the nodes in the digraph in the seen array is equal to the preorder label-
ing of the nodes, and the order of the nodes in the diagraph in the done array is equal
to the post order labeling of the nodes.
S OLUTION TO E XERCISE 5.3.10 ON PAGE 129:
To prove via induction, we need both a base case and an inductive step.
The base case is when there are no white nodes as neighbours of node s, then the
algorithm does no recursion and returns. In this case recursiveDFSvisit only visits
node s, which is intended.
The inductive step is that given all the white nodes that are neighbours of node s
our theorem is true for them, then it is also true for node s. For each node reachable
from s via a path of white nodes, the start of every path is one of the neighbours
of s. Because the call to recursiveDFSvisit with input s only terminates when the
recursive calls with input of each of the white neighbours of s finish. All the recursive
calls then cover each of the paths from s to each node reachable by a path of white
nodes, and thus satisfies the inductive step.
Finally, because each path cannot have a loop in it, there is a finite number of
recursions and recursiveDFSvisit is guaranteed to terminate.
Therefore, by mathematical induction, Theorem 5.4 is true.
S OLUTION TO E XERCISE 5.6.2 ON PAGE 137:
By Theorem 5.11, every DAG has a topological ordering (v1 , v2 , ...vn ) such that there
are no arcs (vi , v j ) ∈ E(G) such that i < j. This means that there are no arcs going from
right to left in the topological ordering. This means that node v1 has no arcs going
into it and node vn has no nodes going away from it, they are respectively, a source
and sink node. Therefore for every DAG there is at least one source and sink node.
S OLUTION TO E XERCISE 5.6.4 ON PAGE 137:
Shirt, hat, tie, jacket, glasses, underwear, trousers, socks, shoes, belt.
S OLUTION TO E XERCISE 5.6.5 ON PAGE 137:
The standard implementation uses an array of indegrees. This can be computed in
time O(m) from either adjacency lists or adjacency matrices. The algorithm can find
a node v of degree 0 in time O(n) and can decrement the indegrees of the neighbours
of v in constant time for adjacency lists. Since we have at most m decrements of
elements of the array of indegree, the running time is at most O(n2 + m). If a priority
queue is used to extract nodes of indegree 0 the running time slightly improves.
S OLUTION TO E XERCISE 5.6.6 ON PAGE 137:
Simply delete vertices with 0 or 1 edges on them from the graph (including the edges),
if at anytime there are no vertices with the number of edges less than 2, then the
graph has a cycle. Otherwise, if the entire graph can be deleted by only deleting ver-
tices with 0 or 1 edges, then the graph is acyclic.
S OLUTION TO E XERCISE 5.7.1 ON PAGE 142:

Adjacency list: 0

4
1, 2
0 1 2
3
2
3

There are two strongly connected components in this graph, DFS only finds one
tree.
S OLUTION TO E XERCISE 5.8.1 ON PAGE 145:

Adjacency list: 0
5
1, 2, 3
0, 4
1 2 3
0, 4
0, 2
4

If the algorithm does not check to the end of the level, it will return that the short-
est cycle is {0,1,4,2} instead of {0,2,4}.
S OLUTION TO E XERCISE 5.8.3 ON PAGE 145:
We need to find two disjoint subsets. Consider the number of 1’s (it can just as easily
be the number of 0’s) to be k in a bit vector of length n, an edge can only be between
another bit vector with either k − 1 or k + 1 1’s. This is because if the number of 1’s is
less than k − 1 or greater than k + 1 then there will be more than 1 difference in the
bits. Also, two different bit vectors with the same number of 1’s will not have an edge
because they will differ in two places exactly (not the required one).
One way of satisfying this condition is if all the odd number of 1’s are on one side,
and all the even number of 1’s are on the other. This means that for any n-cube you
can find a bipartite consisting of the odd number of 1’s bit vectors in one group, and
the even number of 1’s in the other.
S OLUTION TO E XERCISE 6.2.2 ON PAGE 154:
The running time is the same as the time to compute the distance matrix. The ec-
centricity of a node v is simply the maximum entry of row v of the distance matrix.
The radius is the minimum over all maximum values per row. This can be computed
in time Θ(n2 ) if we have access to a distance matrix.
S OLUTION TO E XERCISE 6.3.2 ON PAGE 160:
If a cycle v1 , v2 , . . . , vk exists with the sum of its arc weights is less than zero then we
can find a walk of total weight as small as we want from v1 to v2 by repeating the cycle
as many times as we want before stopping at v2 .
S OLUTION TO E XERCISE 6.3.6 ON PAGE 161:
Property P2 fails if we allow arcs of negative weight. Suppose u is the next vertex
added to S. If arc (u, w) is of negative weight for some other vertex w that is currently
in S, then the previous distance from s to w, dist[w], may no longer be the smallest.
S OLUTION TO E XERCISE 6.4.2 ON PAGE 164:
If a diagonal entry in the distance matrix ever becomes less than zero when using
Floyd’s algorithm then know that a negative weight cycle has been found.
S OLUTION TO E XERCISE 6.5.1 ON PAGE 167:
For this weighted graph, both Prim’s and Kruskal’s algorithms will find the unique
minimum spanning tree of weight 9.
Bibliography

[1] A. V. Aho, J. D. Ullman. Foundations of Computer Science, Computer Science

Press, 1992.

[2] J. Bentley. Programming Pearls, Second Edition. Addison-Wesley, Inc., 2000. 101

[3] F. J. Brandenburg. The Graph Template Library—GTL,

(see http://www.infosun.fim.uni-passau.de/GTL/) 116

[4] T. H. Cormen, C. E. Leiserson and R. L. Rivest. Introduction to Algorithms,

McGraw-Hill, New York, 1990.
[5] J. Edmonds. “Paths, trees, and flowers,” Canadian Journal of Mathematics 17
(1965), pages 449–467. 150

[6] R. Sedgewick and P. Flajolet. Introduction to the Analysis of Algorithms, Addison-

Wesley, Inc., 1996. 101

[7] L. R. Ford and D. R. Fulkerson “Maximal flow through a network,” Canadian Jour-
nal of Mathematics 8 (1956), pages 399–404. 150

[8] M. T. Goodrich and R. Tamassia. Data Structures and Algorithms in Java, John
Wiley and Sons, Inc., 2001.

[9] K. Mehlhorn and St. Nadher. The LEDA Platform of Combinatorial and Geomet-
ric Computing, Cambridge University Press, 1999.
(see http://www.mpi-sb.mpg.de/LEDA/leda.html) 116

[10] J. Orwant, J. Hietaniemi and J. Macdonald. Mastering Algorithms with Perl,

O’Reilly, August 1999.

[11] J.G. Siek, L-Q. Lee and A. Lumsdaine. The Boost Graph Library: User Guide and
Reference Manual Addison-Wesley, 2001. (see http://www.boost.org) 116

[12] T. A. Standish. Data Structures in JavaT M , Addison-Wesley, 1998. 97

Index

S-path, 155 getTop, 202

k-colouring, 144 heapSort, 61
n-cube, 145 insert, 12, 126, 130, 134, 147, 158,
167, 202, 203
AA-tree, 84 insertionSort, 44, 45
abstract data type, 12, 201 isEmpty, 12, 126, 130, 134, 147, 158,
active pivot strategy, 53 167, 201, 203
adjacency lists, 110 Kruskal, 168
adjacency matrix, 110 linearSum, 17
adjacent, 104 merge, 48, 49
ADT, 201 mergeSort, 48
algorithm, 15 partition, 55, 65
BellmanFord, 159 peek, 126, 130, 134, 158, 167, 202,
BFS, 122, 129–132 203
BFSvisit, 122, 130, 140, 142, 153 percolateDown, 61
binarySearch, 74, 76 PFS, 134
binarySearch2, 76 PFSvisit, 134
decreaseKey, 158, 167, 202, 203 pivot, 55, 65
delete, 12, 126, 130, 134, 158, 167, pop, 147, 202, 203
202, 203 Prim, 167
dequeue, 202 push, 202, 203
DFS, 122, 125, 126, 128 quickSelect, 65
DFSvisit, 122, 125, 126, 140 quickSort, 55
Dijkstra, 155 recursiveDFSvisit, 125, 127
Dijkstra2, 158 sequentialSearch, 72
enqueue, 202 set, 168
fastSums, 19 setKey, 133, 134
find, 166, 201, 203 size, 12, 201
findAugmentingPath, 147–149 slowSums, 18
Floyd, 162 swap, 61
getHead, 202 test, 193
getKey, 158, 167
 traverse, 118–122, 125 complete binary tree, 56
union, 166, 168 complete induction, 205
update, 203 computer program, 15
visit, 118–120, 122, 125 connected, 138
alternating path, 146 connected components, 138
antisymmetric, 206 container, 201
arcs, 104 correct, 15
associative array, 69, 203 cost function, 151
asymptotic upper bound, 22 counting sort, 67
asymptotically, 22 cross arc, 119
asymptotically optimal, 30 cubic time, 20
augmenting path, 146 cycle, 106
average-case running time, 29
AVL tree, 82 DAG, 135
data structure, 12
B-tree, 86 decision tree, 66
back arc, 119 degree, 107
balancing, 81 dense, 106
base of the recurrence, 31 depth, 209
binary search, 72 depth-first search, 121
binary search tree, 74 diameter, 153
binary tree, 209 dictionary, 69, 203
bipartite, 143 difference equation, 31
bipartition, 144 digraph, 104
birthday paradox, 92 directed girth, 142
branching limits, 86 disjoint sets, 166
breadth-first search, 121 distance, 107
Bubble sort, 45 distance matrix, 153
bubbling up, 58 divide-and-conquer, 30
division, 98
cardinality, 204 double hashing, 91
ceiling, 207 dynamic, 71
characteristic equation, 38 dynamic programming, 162
children, 209
closed-form expression, 31 eccentricity, 154
cluster, 91 edges, 105
clustering, 91 elementary operations, 16
collision, 89 elements, 204
collision resolution policy, 89 empty set, 204
comparison, 40 equiprobably, 51
comparison-based, 39 equivalence relation, 206
complement, 204 external sorting, 40
floor, 207 keys, 39
folding, 98 Kruskal’s algorithm, 165
forward arc, 119
free tree, 106, 210 leaf, 209
frontier nodes, 117 length, 106
linear algorithms, 17
girth, 142 linear order, 133, 206
graph, 105 linear time, 20
graph union, 109 linearithmic, 38
greedy algorithm, 154 list, 202
load factor, 92
Hamiltonian cycle, 168 logarithmic time, 20
harmonic number, 43 loop, 106
harmonic numbers, 208
hash code, 89 map, 203
hash function, 89 marriage problem, 145
hash table, 89 matching, 145
hashing, 89 maximal matching, 145
head, 202 maximum matching, 145
heap, 57 median-of-three, 53
heapsort, 57 middle-squaring, 98
height, 209 minimum spanning tree, 164
height of a node, 209 move, 40
hypercube, 145
naive pivot selection rule, 52
iff, 22 nodes, 104
in-neighbour, 104 NP-hard, 168
in-place, 40
increment sequence, 45 order, 106
indegree, 107 order statistic, 63
independent set, 168 out-neighbour, 104
induced, 108 outdegree, 107
inductive hypothesis, 205
initial condition, 31 parent, 209
inorder, 74 partial order, 206
input data size, 21 partition, 50, 206
interpolation search, 76 passive pivot strategy, 52
intersection, 204 path, 106
intractable, 27 percolated down, 59
inversion, 43 perfect hash function, 89, 98
iterative deepening, 124 pivot, 50
planar, 169
key, 69, 133 polynomial time, 27
Prim’s algorithm, 165 spanning tree, 164
principle of mathematical induction, sparse, 106
205 stable, 40
priority queue, 202 stack, 202
priority-first search, 122 static, 71
probe sequence, 90 Stirling’s approximation, 208
probes, 90 straight mergesort, 48
strongly connected, 139
quadratic algorithms, 18 strongly connected components, 139
quadratic time, 20 subdigraph, 108
queue, 202 sublist, 202
quickselect, 64 subset, 204
successful search, 69
radius, 154
synonyms, 89
rank, 63, 223
records, 69 table, 70, 203
recurrence, 31 table search, 70
recurrence relation, 31 tail, 202
red-black tree, 84 telescoping, 33
reflexive, symmetric, 206 time complexity, 27
rehashing, 94 topological order, 133
relation, 205 topological sort, 133
repeated halving principle, 34 total internal path length, 80
reverse digraph, 108 total order, 206
root, 209 transitive, 206
rooted ordered tree, 209 travelling salesperson problem, 168
rooted tree, 106 tree arc, 119
rotation, 83 truncation, 98
running time, 16
underlying graph, 109
search forest, 119 uniform hashing hypothesis, 95
selection sort, 41 union, 204
separate chaining, 89 union-find, 166
sequential search, 72 universal class, 99
set, 204 unsuccessful search, 69
set difference, 204
Shellsort, 45 vertex, 105
single-source shortest path, 154 vertex cover, 150, 168
sink, 107 vertices, 105
size, 106
sorted list, 202 walk, 106
source, 107 weakly connected, 139
spanning, 108 weighted digraph, 151
worst-case running time, 29

zero-indegree sorting, 136