Algorithmics

I [P
GILLES BRASSARD
PAUL BRATLEY
ALGORITHMICS
ALGORITHMICS
Theory and Practice
Gilles Brassard and Paul Bratley

Departement d'informatique et de recherche operationnelle
Universitd de Montreal
PRENTICE HALL, Englewood Cliffs, New Jersey 07632

Library of Congress Cataloging-in-Publication Data
Brassard, Gilles
Algorithmics : theory and practice.
1. Recursion theory. 2. Algorithms. I. Bratley,
Paul. II. Title.
QA9.6.B73 1987 51 I'.3 88-2326
ISBN 0-13-023243-2
Editorial/production supervision: Editing, Design & Production, Inc.

Cover design: Lundgren Graphics, Ltd.
Manufacturing buyer: Cindy Grant
© 1988 by Prentice-Hall, Inc.

A division of Simon & Schuster
Englewood Cliffs, New Jersey 07632
All rights reserved. No part of this book may be

reproduced, in any form or by any means,
without permission in writing from the publisher.
Printed in the United States of America

10 9 8 7 6 5 4 3 2 1
ISBN 0-13-023243-2
PRENTICE-HALL INTERNATIONAL (UK) LIMITED. London

PRENTICE-HALL OF AUSTRALIA PTY. LIMITED, Sydney
PRENTICE-HALL CANADA INC., Toronto
PRENTICE-HALL HISPANOAMERICANA, S.A., Mexico
PRENTICE-HALL OF INDIA PRIVATE LIMITED, New Delhi
PRENTICE-HALL OF JAPAN, INC., Tokyo
SIMON & SCHUSTER ASIA PTE. LTD., Singapore
EDITORA PRENTICE-HALL DO BRASIL, LTDA., Rio de Janeiro
for Isabelle and Pat
Contents
Preface Xiii
1 Preliminaries 1
1.1. What Is an Algorithm? 1
1.2. Problems and Instances 4

1.3. The Efficiency of Algorithms 5
1.4. Average and Worst-Case Analysis 7
1.5. What Is an Elementary Operation? 9
1.6. Why Do We Need Efficient Algorithms? 11
1.7. Some Practical Examples 12
1.7.1. Sorting, 13
1.7.2. Multiplication of Large Integers, 13
1.7.3. Evaluating Determinants, 14
1.7.4. Calculating the Greatest Common Divisor, 15
1.7.5. Calculating the Fibonacci Sequence, 16
1.7.6. Fourier Transforms, 19
1.8. When Is an Algorithm Specified? 19
1.9. Data Structures 20

1.9.1. Lists, 20
VII
viii Contents
1.9.2. Graphs, 21
1.9.3. Rooted Trees, 23
1.9.4. Heaps, 25
1.9.5. Disjoint Set Structures, 30
1.10. References and Further Reading 35
2 Analysing the Efficiency of Algorithms 37
2.1. Asymptotic Notation 37

2.1.1. A Notation for "the order- of " , 37
2.1.2. Other Asymptotic Notation, 41
2.1.3. Asymptotic Notation with Several Parameters, 43
2.1.4. Operations on Asymptotic Notation, 43
2.1.5. Conditional Asymptotic Notation, 45
2.1.6. Asymptotic Recurrences, 47
2.1.7. Constructive Induction, 48
2.1.8. For Further Reading, 51
2.2. Analysis of Algorithms 52

2.3. Solving Recurrences Using the Characteristic Equation 65
2.3.1. Homogeneous Recurrences, 65
2.3.2. Inhomogeneous Recurrences, 68
2.3.3. Change of Variable, 72
2.3.4. Range Transformations, 75
2.3.5. Supplementary Problems, 76
3 Greedy Algorithms 79
3.1. Introduction 79
3.2. Greedy Algorithms and Graphs 81
3.2.1. Minimal Spanning Trees, 81

3.2.2. Shortest Paths, 87
3.3. Greedy Algorithms for Scheduling 92
3.3.1. Minimizing Time in the System, 92
3.3.2. Scheduling with Deadlines, 95
3.4. Greedy Heuristics 100

3.4.1. Colouring a Graph, 101
3.4.2. The Travelling Salesperson Problem, 102

Contents ix
4 Divide and Conquer 105
4.2. Determining the Threshold 107
4.3. Binary Searching 109
4.4. Sorting by Merging 115
4.5. Quicksort 116
4.6. Selection and the Median 119
4.7. Arithmetic with Large Integers 124
4.8. Exponentiation : An Introduction to Cryptology 128
4.9. Matrix Multiplication 132
4.10. Exchanging Two Sections of an Array 134
4.11. Supplementary Problems 136
5 Dynamic Programming 142
5.2. The World Series 144
5.3. Chained Matrix Multiplication 146
5.4. Shortest Paths 150
5.5. Optimal Search Trees 154
5.6. The Travelling Salesperson Problem 159
5.7. Memory Functions 162
6 Exploring Graphs 169
6.2. Traversing Trees 170
6.3. Depth-First Search : Undirected Graphs 171
6.3.! Articulation Points, 174

x Contents
6.4. Depth-First Search : Directed Graphs 176

6.4.1. Acyclic Graphs: Topological Sorting, 178
6.4.2. Strongly Connected Components, 179
6.5. Breadth-First Search 182
6.6. Implicit Graphs and Trees 184

6.6.1. Backtracking, 185
6.6.2. Graphs and Games: An Introduction, 189
6.6.3. Branch-and-Bound, 199
7 Preconditioning and Precomputation 205
7.1. Preconditioning 205

7.1.1. Introduction, 205
7.1.2. Ancestry in a rooted tree, 207
7.1.3. Repeated Evaluation of a Polynomial, 209
7.2. Precomputation for String-Searching Problems 211

7.2.1. Signatures, 211
7.2.2. The Knuth-Morris-Pratt Algorithm, 213
7.2.3. The Boyer-Moore Algorithm, 216
8 Probabilistic Algorithms 223

8.2. Classification of Probabilistic Algorithms 226
8.3. Numerical Probabilistic Algorithms 228
8.3.1. Buffon's Needle, 228
8.3.2. Numerical Integration, 230
8.3.3. Probabilistic Counting, 232
8.3.4. More Probabilistic Counting, 235
8.3.5. Numerical Problems in Linear Algebra, 237
8.4. Sherwood Algorithms 238

8.4.1. Selection and Sorting, 238
8.4.2. Stochastic Preconditioning, 240
8.4.3. Searching an Ordered List, 242
8.4.4. Universal Hashing, 245
8.5. Las Vegas Algorithms 247

Contents xi
8.5.1. The Eight Queens Problem Revisited, 248

8.5.2. Square Roots Modulo p, 252
8.5.3. Factorizing Integers, 256
8.5.4. Choosing a Leader, 260
8.6. Monte Carlo Algorithms 262
8.6.1. Majority Element in an Array, 268
8.6.2. Probabilistic Primality Testing, 269
8.6.3. A Probabilistic Test for Set Equality, 271
8.6.4. Matrix Multiplication Revisited, 274
9 Transformations of the Domain 277

9.2. The Discrete Fourier Transform 279
9.3. The Inverse Transform 280
9.4. Symbolic Operations on Polynomials 284
9.5. Multiplication of Large Integers 286
10 Introduction to Complexity 292
10.1. Decision Trees 292

10.2. Reduction 300
10.2. 1. Reductions Among Matrix Problems, 302
10.2.2. Reductions Among Graph Problems, 304
10.2.3. Reductions Among Arithmetic and Polynomial Problems, 308
10.3. Introduction to NP-Completeness 315
10.3.1. The Classes P and NP, 316
10.3.2. NP-Complete Problems, 324
10.3.3. Cook's Theorem, 325
10.3.4. Some Reductions, 328
10.3.5. Nondeterminism, 332
Table of Notation 338
Bibliography 341
Index 353
Preface
The explosion in computing we are witnessing is arousing extraordinary interest at

every level of society. As the power of computing machinery grows, calculations once
infeasible become routine. Another factor, however, has had an even more important
effect in extending the frontiers of feasible computation: the use of efficient algorithms.
For instance, today's typical medium-sized computers can easily sort 100,000 items in
30 seconds using a good algorithm, whereas such speed would be impossible, even on
a machine a thousand times faster, using a more naive algorithm. There are other
examples of tasks that can be completed in a small fraction of a second, but that would
require millions of years with less efficient algorithms (read Section 1.7.3 for more
detail).
The Oxford English Dictionary defines algorithm as an "erroneous refashioning
of algorism" and says about algorism that it "passed through many pseudo-
etymological perversions, including a recent algorithm". (This situation is not corrected
in the OED Supplement.) Although the Concise Oxford Dictionary offers a more up-
to-date definition for the word algorithm, quoted in the opening sentence of Chapter 1,
we are aware of no dictionary of the English language that has an entry for algo-
rithmics, the subject matter of this book.
We chose the word algorithmics to translate the more common French term
algorithmique. (Although this word appears in some French dictionaries, the definition
does not correspond to modern usage.) In a nutshell, algorithmics is the systematic
study of the fundamental techniques used to design and analyse efficient algorithms.
The same word was coined independently by several people, sometimes with slightly
different meanings. For instance, Harel (1987) calls algorithmics "the spirit of
xiii
xiv Preface
computing", adopting the wider perspective that it is "the area of human study,
knowledge and expertise that concerns algorithms".
Our book is neither a programming manual nor an account of the proper use of
data structures. Still less is it a "cookbook" containing a long catalogue of programs
ready to be used directly on a machine to solve certain specific problems, but giving at
best a vague idea of the principles involved in their design. On the contrary, the aim of
our book is to give the reader some basic tools needed to develop his or her own algo-
rithms, in whatever field of application they may be required.
Thus we concentrate on the techniques used to design and analyse efficient algo-
rithms. Each technique is first presented in full generality. Thereafter it is illustrated by
concrete examples of algorithms taken from such different applications as optimization,
linear algebra, cryptography, operations research, symbolic computation, artificial intel-
ligence, numerical analysis, computing in the humanities, and so on. Although our
approach is rigorous and theoretical, we do not neglect the needs of practitioners:
besides illustrating the design techniques employed, most of the algorithms presented
also have real-life applications.
To profit fully from this book, you should have some previous programming
experience. However, we use no particular programming language, nor are the exam-
ples for any particular machine. This and the general, fundamental treatment of the
material ensure that the ideas presented here will not lose their relevance. On the other
hand, you should not expect to be able to use the algorithms we give directly: you will
always be obliged to make the necessary effort to transcribe them into some
appropriate programming language. The use of Pascal or similarly structured language
will help reduce this effort to the minimum necessary.
Some basic mathematical knowledge is required to understand this book. Gen-
erally speaking, an introductory undergraduate course in algebra and another in cal-
culus should provide sufficient background. A certain mathematical maturity is more
important still. We take it for granted that the reader is familiar with such notions as
mathematical induction, set notation, and the concept of a graph. From time to time a
passage requires more advanced mathematical knowledge, but such passages can be
skipped on the first reading with no loss of continuity.
Our book is intended as a textbook for an upper-level undergraduate or a lower-
level graduate course in algorithmics. We have used preliminary versions at both the
University de Montreal and the University of California, Berkeley. If used as the basis
for a course at the graduate level, we suggest that the material be supplemented by
attacking some subjects in greater depth, perhaps using the excellent texts by Garey
and Johnson (1979) or Tarjan (1983). Our book can also be used for independent
study: anyone who needs to write better, more efficient algorithms can benefit from it.
Some of the chapters, in particular the one concerned with probabilistic algorithms,
contain original material.
It is unrealistic to hope to cover all the material in this book in an undergraduate
course with 45 hours or so of classes. In making a choice of subjects, the teacher
should bear in mind that the first two chapters are essential to understanding the rest of
Preface xv
the book, although most of Chapter 1 can probably be assigned as independent reading.
The other chapters are to a great extent independent of one another. An elementary
course should certainly cover the first five chapters, without necessarily going over
each and every example given there of how the techniques can be applied. The choice
of the remaining material to be studied depends on the teacher's preferences and incli-
nations.The last three chapters, however, deal with more advanced topics; the teacher
may find it interesting to discuss these briefly in an undergraduate class, perhaps to lay
the ground before going into detail in a subsequent graduate class.
Each chapter ends with suggestions for further reading. The references from each
chapter are combined at the end of the book in an extensive bibliography including
well over 200 items. Although we give the origin of a number of algorithms and ideas,
our primary aim is not historical. You should therefore not be surprised if information
of this kind is sometimes omitted. Our goal is to suggest supplementary reading that
can help you deepen your understanding of the ideas we introduce.
Almost 500 exercises are dispersed throughout the text. It is crucial to read the
problems: their statements form an integral part of the text. Their level of difficulty is
indicated as usual either by the absence of an asterisk (immediate to easy), or by the
presence of one asterisk (takes a little thought) or two asterisks (difficult, maybe even a
research project). The solutions to many of the difficult problems can be found in the
references. No solutions are provided for the other problems, nor do we think it advis-
able to provide a solutions manual. We hope the serious teacher will be pleased to have
available this extensive collection of unsolved problems from which homework assign-
ments can be chosen. Several problems call for an algorithm to be implemented on a
computer so that its efficiency may be measured experimentally and compared to the
efficiency of alternative solutions. It would be a pity to study this material without
carrying out at least one such experiment.
The first printing of this book by Prentice Hall is already in a sense a second edi-
tion. We originally wrote our book in French. In this form it was published by Masson,
Paris. Although less than a year separates the first French and English printings, the
experience gained in using the French version, in particular at an international summer
school in Bayonne, was crucial in improving the presentation of some topics, and in
spotting occasional errors. The numbering of problems and sections, however, is not
always consistent between the French and English versions.
Writing this book would have been impossible without the help of many people.
Our thanks go first to the students who have followed our courses in algorithmics over
the years since 1979, both at the undergraduate and graduate levels. Particular thanks
are due to those who kindly allowed us to copy their course notes: Denis Fortin,
Laurent Langlois, and Sophie Monet in Montreal, and Luis Miguel and Dan Philip in
Berkeley. We are also grateful to those people who used the preliminary versions of
our book, whether they were our own students, or colleagues and students at other
universities. The comments and suggestions we received were most valuable. Our war-
mest thanks, however, must go to those who carefully read and reread several chapters
of the book and who suggested many improvements and corrections: Pierre
xvi Preface
Beauchemin, Andre Chartier, Claude Crepeau, Bennett Fox, Claude Goutier, Pierre
L'Ecuyer, Pierre McKenzie, Santiago Miro, Jean-Marc Robert, and Alan Sherman.
We are also grateful to those who made it possible for us to work intensively on
our book during long periods spent away from Montreal. Paul Bratley thanks Georges
Stamon and the Universite de Franche-Comte. Gilles Brassard thanks Manuel Blum
and the University of California, Berkeley, David Chaum and the CWI, Amsterdam,
and Jean-Jacques Quisquater and Philips Research Laboratory, Bruxelles. He also
thanks John Hopcroft, who taught him so much of the material included in this book,
and Lise DuPlessis who so many times made her country house available; its sylvan
serenity provided the setting and the inspiration for writing a number of chapters.
Denise St.-Michel deserves our special thanks. It was her misfortune to help us
struggle with the text editing system through one translation and countless revisions.
Annette Hall, of Editing, Design, and Production, Inc., was no less misfortuned to help
us struggle with the last stages of production. The heads of the laboratories at the
Universite de Montreal's Departement d'informatique et de recherche operationnelle,
Michel Maksud and Robert Gerin-Lajoie, provided unstinting support. We thank the
entire team at Prentice Hall for their exemplary efficiency and friendliness; we particu-
larly appreciate the help we received from James Fegen. We also thank Eugene L.
Lawler for mentioning our French manuscript to Prentice Hall's representative in
northern California, Dan Joraanstad, even before we plucked up the courage to work on
an English version. The Natural Sciences and Engineering Research Council of Canada
provided generous support.
Last but not least, we owe a considerable debt of gratitude to our wives, Isabelle
and Pat, for their encouragement, understanding, and exemplary patience-in short,
for putting up with us -while we were working on the French and English versions of
this book.
Gilles Brassard
Paul Bratley
ALGORITHMICS
I
Preliminaries
1.1 WHAT IS AN ALGORITHM?
The Concise Oxford Dictionary defines an algorithm as a "process or rules for (esp.
machine) calculation". The execution of an algorithm must not include any subjective
decisions, nor must it require the use of intuition or creativity (although we shall see an
important exception to this rule in Chapter 8). When we talk about algorithms, we
shall mostly be thinking in terms of computers. Nonetheless, other systematic methods
for solving problems could be included. For example, the methods we learn at school
for multiplying and dividing integers are also algorithms. The most famous algorithm
in history dates from the time of the Greeks : this is Euclid's algorithm for calculating
the greatest common divisor of two integers. It is even possible to consider certain
cooking recipes as algorithms, provided they do not include instructions like "Add salt
to taste".
When we set out to solve a problem, it is important to decide which algorithm
for its solution should be used. The answer can depend on many factors : the size of
the instance to be solved, the way in which the problem is presented, the speed and
memory size of the available computing equipment, and so on. Take elementary arith-
metic as an example. Suppose you have to multiply two positive integers using only
pencil and paper. If you were raised in North America, the chances are that you will
multiply the multiplicand successively by each figure of the multiplier, taken from
right to left, that you will write these intermediate results one beneath the other shifting
each line one place left, and that finally you will add all these rows to obtain your
answer. This is the "classic" multiplication algorithm.
1
2 Preliminaries Chap. 1
However, here is quite a different algorithm for doing the same thing, sometimes
called "multiplication a la russe ". Write the multiplier and the multiplicand side by
side. Make two columns, one under each operand, by repeating the following rule
until the number under the multiplier is 1 : divide the number under the multiplier by
2, ignoring any fractions, and double the number under the multiplicand by adding it to
itself. Finally, cross out each row in which the number under the multiplier is even,
and then add up the numbers that remain in the column under the multiplicand. For
example, multiplying 19 by 45 proceeds as in Figure I.I.I. In this example we get
19+76+152+608 = 855. Although this algorithm may seem funny at first, it is essen-
tially the method used in the hardware of many computers. To use it, there is no need
to memorize any multiplication tables : all we need to know is how to add up, and
how to double a number or divide it by 2.
45 19 19
22 38 ---
11 76 76
5 152 152
2 304 -----
1 608 608
855 Figure 1.1.1. Multiplication a la russe.
We shall see in Section 4.7 that there exist more efficient algorithms when the
integers to be multiplied are very large. However, these more sophisticated algorithms
are in fact slower than the simple ones when the operands are not sufficiently large.
At this point it is important to decide how we are going to represent our algo-
rithms. If we try to describe them in English, we rapidly discover that natural
languages are not at all suited to this kind of thing. Even our description of an algo-
rithm as simple as multiplication a la russe is not completely clear. We did not so
much as try to describe the classic multiplication algorithm in any detail. To avoid
confusion, we shall in future specify our algorithms by giving a corresponding pro-
gram. However, we shall not confine ourselves to the use of one particular program-
ming language : in this way, the essential points of an algorithm will not be obscured
by the relatively unimportant programming details.
We shall use phrases in English in our programs whenever this seems to make
for simplicity and clarity. These phrases should not be confused with comments on the
program, which will always be enclosed within braces. Declarations of scalar quanti-
ties (integer, real, or Boolean) are usually omitted. Scalar parameters of functions and
procedures are passed by value unless a different specification is given explicitly, and
arrays are passed by reference.
The notation used to specify that a function or a procedure has an array param-
eter varies from case to case. Sometimes we write, for instance
procedure proc1(T : array)
Sec. 1.1 What Is an Algorithm? 3
or even
procedure proc2(T)
if the type and the dimensions of the array T are unimportant or if they are evident
from the context. In such a case #T denotes the number of elements in the array T. If
the bounds or the type of T are important, we write
procedure proc3(T [ 1 .. n ])
or more generally
procedure proc4(T [a .. b ] : integers)
In such cases n, a, and b should be considered as formal parameters, and their values
are determined by the bounds of the actual parameter corresponding to T when the pro-
cedure is called. These bounds can be specified explicitly, or changed, by a procedure
call of the form
proc3(T [ l .. m ]) .
To avoid proliferation of begin and end statements, the range of a statement such
as if, while, or for, as well as that of a declaration such as procedure, function, or
record, is shown by indenting the statements affected. The statement return marks
the dynamic end of a procedure or a function, and in the latter case it also supplies the
value of the function. The operators div and mod represent integer division (dis-
carding any fractional result) and the remainder of a division, respectively. We assume
that the reader is familiar with the concepts of recursion and of pointers. The latter are
denoted by the symbol " T ". A reader who has some familiarity with Pascal, for
example, will have no difficulty understanding the notation used to describe our algo-
rithms. For instance, here is a formal description of multiplication a la russe.
function russe (A, B )
arrays X, Y
{ initialization }
X[1] -A; Y[1] -B
i (--
I make the two columns }
while X [i] > 1 do
X[i+1]E-X[i]div2
Y[i+l] <-- Y[i] +Y[i]
i -i +1
{ add the appropriate entries }
prod F- 0
while i > 0 do
if X [i ] is odd then prod F prod + Y [i ]
<--
return prod
If you are an experienced programmer, you will probably have noticed that the
arrays X and Y are not really necessary, and that this program could easily be
simplified. However, we preferred to follow blindly the preceding description of the
algorithm, even if this is more suited to a calculation using pencil and paper than to
computation on a machine. The following APL program describes exactly the same
algorithm (although you might reasonably object to a program using logarithms,
exponentiation, and multiplication by powers of 2 to describe an algorithm for multi-
plying two integers ...) .
V R-A RUSAPL B; T
[1] R<-+/(2I LA=T)/BXT-1,2*t[2*A V
On the other hand, the following program, despite a superficial resemblance to the one
given previously, describes quite a different algorithm.
function not -russe (A, B )
arrays X, Y
{ initialization }
X[1] - A ; Y[1] - B
iF1
{ make the two columns)
while X [i ] > 1 do
X[i+1]<-X[i]-1
Y[i+1]-B
i-i+1
{ add the appropriate entries }
prod F-- 0
while i > 0 do
if X [i ] > 0 then prod - prod + Y [i ]
i -i-1
return prod
We see that different algorithms can be used to solve the same problem, and that
different programs can be used to describe the same algorithm. It is important not to
lose sight of the fact that in this book we are interested in algorithms, not in the pro-
grams used to describe them.
1.2 PROBLEMS AND INSTANCES
Multiplication a la russe is not just a way to multiply 45 by 19. It gives a general

solution to the problem of multiplying positive integers. We say that (45, 19) is an
instance of this problem. Most interesting problems ihclude an infinite collection of
instances. Nonetheless, we shall occasionally consider finite problems such as that of
playing a perfect game of chess. An algorithm must work correctly on every instance
of the problem it claims to solve. To show that an algorithm is incorrect, we need only
find one instance of the problem for which it is unable to find a correct answer. On the
Sec. 1.3 The Efficiency of Algorithms 5
other hand, it is usually more difficult to prove the correctness of an algorithm. When
we specify a problem, it is important to define its domain of definition, that is, the set
of instances to be considered. Although multiplication a la russe will not work if the
first operand is negative, this does not invalidate the algorithm since (-45, 19) is not an
instance of the problem being considered.
Any real computing device has a limit on the size of the instances it can handle.
However, this limit cannot be attributed to the algorithm we choose to use. Once again
we see that there is an essential difference between programs and algorithms.
1.3 THE EFFICIENCY OF ALGORITHMS
When we have a problem to solve, it is obviously of interest to find several algorithms

that might be used, so we can choose the best. This raises the question of how to
decide which of several algorithms is preferable. The empirical (or a posteriori)
approach consists of programming the competing algorithms and trying them on dif-
ferent instances with the help of a computer. The theoretical (or a priori) approach,
which we favour in this book, consists of determining mathematically the quantity of
resources (execution time, memory space, etc.) needed by each algorithm as a function
of the size of the instances considered.
The size of an instance x, denoted by I x 1, corresponds formally to the number of
bits needed to represent the instance on a computer, using some precisely defined and
reasonably compact encoding. To make our analyses clearer, however, we often use
the word "size" to mean any integer that in some way measures the number of com-
ponents in an instance. For example, when we talk about sorting (see Section 1.7.1),
an instance involving n items is generally considered to be of size n, even though each
item would take more than one bit when represented on a computer. When we talk
about numerical problems, we sometimes give the efficiency of our algorithms in terms
of the value of the instance being considered, rather than its size (which is the number
of bits needed to represent this value in binary).
The advantage of the theoretical approach is that it depends on neither the com-
puter being used, nor the programming language, nor even the skill of the programmer.
It saves both the time that would have been spent needlessly programming an
inefficient algorithm and the machine time that would have been wasted testing it.
It also allows us to study the efficiency of an algorithm when used on instances of any
size. This is often not the case with the empirical approach, where practical considera-
tions often force us to test our algorithms only on instances of moderate size. This last
point is particularly important since often a newly discovered algorithm may only
begin to perform better than its predecessor when both of them are used on large
instances.
It is also possible to analyse algorithms using a hybrid approach, where the form
of the function describing the algorithm's efficiency is determined theoretically, and
then any required numerical parameters are determined empirically for a particular
program and machine, usually by some form of regression. This approach allows pred-
ictions to be made about the time an actual implementation will take to solve an
instance much larger than those used in the tests. If such an extrapolation is made
solely on the basis of empirical tests, ignoring all theoretical considerations, it is likely
to be less precise, if not plain wrong.
It is natural to ask at this point what unit should be used to express the theoret-
ical efficiency of an algorithm. There can be no question of expressing this efficiency
in seconds, say, since we do not have a standard computer to which all measurements
might refer. An answer to this problem is given by the principle of invariance,
according to which two different implementations of the same algorithm will not differ
in efficiency by more than some multiplicative constant. More precisely, if two imple-
mentations take t i(n) and t2(n) seconds, respectively, to solve an instance of size n ,
then there always exists a positive constant c such that tl(n) <_ ct2(n) whenever n is
sufficiently large. This principle remains true whatever the computer used (provided it
is of a conventional design), regardless of the programming language employed and
regardless of the skill of the programmer (provided that he or she does not actually
modify the algorithm!). Thus, a change of machine may allow us to solve a problem
10 or 100 times faster, but only a change of algorithm will give us an improvement
that gets more and more marked as the size of the instances being solved increases.
Coming back to the question of the unit to be used to express the theoretical
efficiency of an algorithm, there will be no such unit : we shall only express this
efficiency to within a multiplicative constant. We say that an algorithm takes a time in
the order of t (n), for a given function t, if there exist a positive constant c and an
implementation of the algorithm capable of solving every instance of the problem in a
time bounded above by ct (n) seconds, where n is the size (or occasionally the value,
for numerical problems) of the instance considered. The use of seconds in this
definition is obviously quite arbitrary, since we only need change the constant to bound
the time by at (n) years or bt (n) microseconds. By the principle of invariance any
other implementation of the algorithm will have the same property, although the multi-
plicative constant may change from one implementation to another. In the next chapter
we give a more rigorous treatment of this important concept known as the asymptotic
notation. It will be clear from the formal definition why we say "in the order of"
rather than the more usual "of the order of ".
Certain orders occur so frequently that it is worth giving them a name. For
example, if an algorithm takes a time in the order of n, where n is the size of the
instance to be solved, we say that it takes linear time. In this case we also talk about a
linear algorithm. Similarly, an algorithm is quadratic, cubic, polynomial, or exponen-
tial if it takes a time in the order of n 2, n 3, n k , or c, respectively, where k and c are
appropriate constants. Sections 1.6 and 1.7 illustrate the important differences between
these orders of magnitude.
The hidden multiplicative constant used in these definitions gives rise to a certain
danger of misinterpretation. Consider, for example, two algorithms whose im-
plementations on a given machine take respectively n 2 days and n 3 seconds to solve an
Sec. 1.4 Average and Worst-Case Analysis 7
instance of size n. It is only on instances requiring more than 20 million years to

solve that the quadratic algorithm outperforms the cubic algorithm ! Nevertheless,
from a theoretical point of view, the former is asymptotically better than the latter, that
is to say, its performance is better on all sufficiently large instances.
The other resources needed to execute an algorithm, memory space in particular,
can be estimated theoretically in a similar way. It may also be interesting to study the
possibility of a trade-off between time and memory space : using more space some-
times allows us to reduce the computing time, and conversely. In this book, however,
we concentrate on execution time.
Finally, note that logarithms to the base 2 are so frequently used in the analysis
of algorithms that we give them their own special notation : thus "lg n " is an abbrevia-
tion for 1og2n. As is more usual, "In" and "log" denote natural logarithms and loga-
rithms to the base 10, respectively.
1.4 AVERAGE AND WORST-CASE ANALYSIS
The time taken by an algorithm can vary considerably between two different instances
of the same size. To illustrate this, consider two elementary sorting algorithms
inser-:
tion and selection.
procedure insert (T [1 .. n ])
for i F-- 2 to n do
x- T[i]; j- i- 1
while j > 0andx <T[j] doT[j+l] <.-T[j]
j F-j-1
T[j+1]-x
and
procedure select (T [1 .. n ])
for i E-- 1 to n -1 do
minj - i ; minx F- T [i]
for j E-- i + I to n do
if T [ j ] < minx then minj i- j
minx -T[j]
T[minjI -T[i]
T [i] f- minx
Problem 1.4.1. Simulate these two algorithms on the arrays

T = [3, 1, 4, 1, 5, 9, 2, 6, 5, 3] , U = [1, 2, 3, 4, 5, 6] , and V = [6, 5, 4, 3, 2, 1].
Make sure you understand how they work. 0
Let U and V be two arrays of n elements, such that U is already sorted in
ascending order, whereas V is in descending order. Problem 1.4.1 shows that both
these algorithms take more time on V than on U. In fact, V represents the worst case
for these two algorithms : no array of n elements requires more work. Nonetheless,
the time required by the selection sorting algorithm is not very sensitive to the original
order of the array to be sorted : the test "if T [ j ] < minx " is executed exactly the
same number of times in every case. The variation in execution time is only due to the
number of times the assignments in the then part of this test are executed. To verify
this, we programmed this algorithm in Pascal on a DEC VAx 780. We found that the
time required to sort a given number of elements using selection sort does not vary by
more than 15% whatever the initial order of the elements to be sorted. As Example
2.2.1 will show, the time required by select (T) is quadratic, regardless of the initial
order of the elements.
The situation is quite different if we compare the times taken by the insertion
sort algorithm on the arrays U and V. On the one hand, insert (U) is very fast, because
the condition controlling the while loop is always false at the outset. The algorithm
therefore performs in linear time. On the other hand, insert (V) takes quadratic time,
because the while loop is executed i -1 times for each value of i (see Example 2.2.3).
The variation in time is therefore considerable, and moreover, it increases with the
number of elements to be sorted. An implementation in Pascal on the DEC VAx 780
shows that insert (U) takes less than one-fifth of a second if U is an array of 5,000 ele-
ments already in ascending order, whereas insert (V) takes three and a half minutes
when V is an array of 5,000 elements in descending order.
If such large variations can occur, how can we talk about the time taken by an
algorithm solely in terms of the size of the instance to be solved? We usually consider
the worst case of the algorithm, that is, for each size we only consider those instances
of that size on which the algorithm requires the most time. Thus we say that insertion
sorting takes quadratic time in the worst case.
Worst-case analysis is appropriate for an algorithm whose response time is crit-
ical. For example, if it is a question of controlling a nuclear power plant, it is crucial
to know an upper limit on the system's response time, regardless of the particular
instance to be solved. On the other hand, in a situation where an algorithm is to be
used many times on many different instances, it may be more important to know the
average execution time on instances of size n. We saw that the time taken by the
insertion sort algorithm varies between the order of n and the order of n 2. If we can
calculate the average time taken by the algorithm on the n! different ways of initially
ordering n elements (assuming they are all distinct), we shall have an idea of the likely
time taken to sort an array initially in random order. We shall see in Example 2.2.3
that this average time is also in the order of The insertion sorting algorithm thus
n2.
takes quadratic time both on the average and in the worst case, although in certain
cases it can be much faster. In Section 4.5 we shall see another sorting algorithm that
also takes quadratic time in the worst case, but that requires only a time in the order of
n log n on the average. Even though this algorithm has a bad worst case, it is among
the fastest algorithms known on the average.
It is usually harder to analyse the average behaviour of an algorithm than to
analyse its behaviour in the worst case. Also, such an analysis of average behaviour
Sec. 1.5 What Is an Elementary Operation? 9
can be misleading if in fact the instances to be solved are not chosen randomly when
the algorithm is used in practice. For example, it could happen that a sorting algorithm
might be used as an internal procedure in some more complex algorithm, and that for
some reason it might mostly be asked to sort arrays whose elements are already nearly
ordered. In this case, the hypothesis that each of the n! ways of.initially ordering n
elements is equally likely fails. A useful analysis of the average behaviour of an algo-
rithm therefore requires some a priori knowledge of the distribution of the instances to
be solved, and this is normally an unrealistic requirement. In Chapter 8 we shall see
how this difficulty can be circumvented for certain algorithms, and their behaviour
made independent of the specific instances to be solved.
In what follows we shall only be concerned with worst-case analyses unless
stated otherwise.
1.5 WHAT IS AN ELEMENTARY OPERATION?
An elementary operation is an operation whose execution time can be bounded above

by a constant depending only on the particular implementation used (machine, pro-
gramming language, and so on). Since we are only concerned with execution times of
algorithms defined to within a multiplicative constant, it is only the number of elemen-
tary operations executed that matters in the analysis, not the exact time required by
each of them. Equivalently, we say that elementary operations can be executed at unit
cost. In the description of an algorithm it may happen that a line of program
corresponds to a variable number of elementary operations. For example, if T is an
array of n elements, the time required to compute
x -min {T[i]ll <-i <-n }
increases with n, since it is an abbreviation for
x - T[l]
for i F- 2 to n do
if T [i ] < x then x - T [i ]
Similarly, some mathematical operations are too complex to be considered elementary.
If we allowed ourselves to count the evaluation of a factorial and a test for divisibility
at unit cost, regardless of the operand's size, Wilson's theorem would let us test an
integer for primality with astonishing efficiency.
function Wilson (n)
{ returns true if and only if n is prime }
if n divides ((n -1)! + 1) exactly then return true
else return false
Can we consider addition and multiplication to be unit cost operations ? In
theory these operations are not elementary since the time needed to execute them
increases with the length of the operands. In practice, however, it may be sensible to
consider them as elementary operations so long as the operands concerned are of a rea-
sonable size in the instances we expect to encounter. Two examples will illustrate
what we mean.
function Not-Gauss (n)
{ calculates the sum of the integers from 1 to n }
sum E- 0
for i <- l ton do sum E-sum + i
return sum
and
function Fibonacci (n)

{ calculates the n th term of the Fibonacci sequence (see section 1.7.5))
i E- 1; j E- 0
fork E- l to n do j E- i+ j
i E- j - i
return j
In the algorithm called Not-Gauss the value of sum stays quite reasonable for all
the instances that the algorithm can realistically be expected to meet in practice. If we
are using a 32-bit machine, all the additions can be executed directly provided that n is
no greater than 65,535. In theory, however, the algorithm should work for all possible
values of n, so that no real machine can in fact execute these additions at unit cost if n
is chosen sufficiently large. The analysis of the algorithm must therefore depend on its
intended domain of application.
The situation is quite different in the case of Fibonacci. It suffices to take
n = 47 to have the last addition "j E- i + j " cause arithmetic overflow on a 32-bit
machine. As many as 45,496 bits are needed to hold the result corresponding to
n = 65,535. It is therefore not realistic, as a practical matter, to consider that these
additions can be carried out at unit cost ; rather, we must attribute to them a cost pro-
portional to the length of the operands concerned. In Example 2.2.8, this algorithm
(there called fib2) is shown to take quadratic time, even though at first glance its exe-
cution time appears to be linear.
In the case of multiplication, although it may still be reasonable to consider this
an elementary operation for sufficiently small operands, it is even more important to
ensure that arithmetic operations do not overflow : it is easier to produce large
operands by repeated multiplication than by addition. The following problem illus-
trates this danger.
** Problem 1.5.1. Use Wilson's theorem (n is prime if and only if it is a divisor

of (n - 1)! + 1 ), Newton's binomial theorem, and the divide-and-conquer technique dis-
cussed in Chapter 4 to design an algorithm capable of deciding in a time in the order
of log n , given an integer n, whether or not n is prime. In your analysis you
Sec. 1.6 Why Do We Need Efficient Algorithms ? 11
may assume that additions, multiplications, and tests of divisibility by an integer (but
not calculations of factorials or exponentials) can be carried out in unit time, regardless
of the size of the operands involved.
A similar problem can arise when we analyse algorithms involving real numbers
if the required precision increases with the size of the instances to be solved. One typ-
ical example of this phenomenon is the use of De Moivre's formula to calculate values
in the Fibonacci sequence (see Section 1.7.5). In most practical situations, however,
the use of single precision floating point arithmetic proves satisfactory despite the inev-
itable loss of precision. When this is so, it is reasonable to count such arithmetic
operations at unit cost.
To sum up, even deciding whether an instruction as apparently innocent as
"j F i + j " can be considered as elementary or not calls for the use of judgement. In
what follows we count additions, subtractions, multiplications, divisions, modulo
operations, Boolean operations, comparisons, and assignments at unit cost unless expli-
citly stated otherwise.
1.6 WHY DO WE NEED EFFICIENT ALGORITHMS?
As computing equipment gets faster and faster, it may seem hardly worthwhile to
spend our time trying to design more efficient algorithms. Would it not be easier
simply to wait for the next generation of computers? The remarks made in the
preceding sections show that this is not true. Suppose, to illustrate the argument, that
to solve a particular problem you have available an exponential algorithm and a com-
puter capable of running this algorithm on instances of size n in 10-4 x 2' seconds.
Your program can thus solve an instance of size 10 in one-tenth of a second. Solving
an instance of size 20 will take nearly two minutes. To solve an instance of size 30,
even a whole day's computing will not be sufficient. Supposing you were able to run
your computer without interruption for a year, you would only just be able to solve an
instance of size 38.
Since you need to solve bigger instances than this, you buy a new computer one
hundred times faster than the first. With the same algorithm you can now solve an
instance of size n in only 10-6 x 2" seconds. You may feel you have wasted your
money, however, when you figure out that now, when you run your new machine for a
whole year, you cannot even solve an example of size 45. In general, if you were pre-
viously able to solve an instance of size n in some given time, your new machine will
solve instances of size at best n + 7 in the same time.
Suppose you decide instead to invest in algorithmics. You find a cubic algorithm
that can solve your problem. Imagine, for example, that using the original machine
this new algorithm can solve an instance of size n in 10-2 x n 3 seconds. In one day
you can now solve instances whose size is greater than 200; with one year's computa-
tion you can almost reach size 1,500. This is illustrated by Figure 1.6.1.
Figure 1.6.1. Algorithmics versus hardware.
Not only does the new algorithm offer a much greater improvement than the pur-
chase of new machinery, it will also, supposing you are able to afford both, make such
a purchase much more profitable. In fact, thanks to your new algorithm, a machine
one hundred times faster than the old one will allow you to solve instances four or five
times bigger in the same length of time. Nevertheless, the new algorithm should not
be used uncritically on all instances of the problem, in particular on the rather small
ones. On the original machine the new algorithm takes 10 seconds to solve an instance
of size 10, which is one hundred times slower than the old algorithm. The new algo-
rithm is faster only for instances of size 20 or greater. Naturally, it is possible to com-
bine the two algorithms into a third one that looks at the size of the instance to be
solved before deciding which method to use.
1.7 SOME PRACTICAL EXAMPLES
Maybe you are wondering whether it is really possible in practice to accelerate an algo-
rithm to the extent suggested in the previous section. In fact, there have been cases
where even more spectacular improvements have been made, even for well-established
algorithms. Some of the following examples use large integers or real arithmetic.
Unless we explicitly state the contrary, we shall simplify our presentation by ignoring
the problems that may arise because of arithmetic overflow or loss of precision on a
particular machine. Such problems can always be solved by using multiple-precision
arithmetic (see Sections 1.7.2 and 4.7). Additions and multiplications are therefore
generally taken to be elementary operations in the following paragraphs (except, of
course, for Section 1.7.2).
Sec. 1.7 Some Practical Examples 13
1.7.1 Sorting
The sorting problem is of major importance in computer science, and in particular in
algorithmics. We are required to arrange in ascending order a collection of n objects
on which a total ordering is defined. Sorting problems are often found inside more
complex algorithms. We have already seen two classic sorting algorithms in Section
1.4: insertion sorting and selection sorting. Both these algorithms, as we saw, take
quadratic time both in the worst case and on the average.
Although these algorithms are excellent when n is small, other sorting algorithms
are more efficient when n is large. Among others, we might use Williams's heapsort
algorithm (see Example 2.2.4 and Problem 2.2.3), mergesort (see Section 4.4), or
Hoare's quicksort algorithm (see Section 4.5). All these algorithms take a time in the
order of n log n on the average ; the first two take this same amount of time even in
the worst case.
To have a clearer idea of the practical difference between a time in the order of
n 2 and a time in the order of n log n , we programmed insertion sort and quicksort in
Pascal on a DEC VAx 780. The difference in efficiency between the two algorithms is
marginal when the number of elements to be sorted is small. Quicksort is already
almost twice as fast as insertion when sorting 50 elements, and three times as fast
when sorting 100 elements. To sort 1,000 elements, insertion takes more than three
seconds, whereas quicksort requires less than one-fifth of a second. When we have
5,000 elements to sort, the inefficiency of insertion sorting becomes still more pro-
nounced : one and a half minutes are needed on average, compared to little more than
one second for quicksort. In 30 seconds, quicksort can handle 100,000 elements ; our
estimate is that it would take nine and a half hours to carry out the same task using
insertion sorting.
1.7.2 Multiplication of Large Integers

When a calculation requires very large integers to be manipulated, it can happen that
the operands become too long to be held in a single word of the computer in use.
Such operations thereupon cease to be elementary. When this occurs, we can use a
representation such as FORTRAN's "double precision", or, more generally, multiple-
precision arithmetic. In this case, we must ask ourselves how the time necessary to
multiply two large integers increases with the size of the operands. We can measure
this size by either the number of computer words needed to represent the operands on a
machine or the length of their representation in decimal or binary. Since these meas-
ures differ only by a multiplicative constant, this choice does not alter our analysis of
the order of efficiency of the algorithms in question. (This last remark would be false
should we be considering exponential time algorithms-can you see why?)
Suppose two large integers, of sizes m and n, respectively, are to be multiplied.
The classic algorithm of Section 1.1 can easily be transposed to this context. We see
that it multiplies each word of one of the operands by each word of the other, and that
it executes approximately one elementary addition for each of these multiplications.
The time required is therefore in the order of mn. Multiplication a la russe also takes
a time in the order of mn, provided we choose the smaller operand as the multiplier
and the larger as the multiplicand. Thus, there is no reason for preferring it to the
classic algorithm, particularly as the hidden constant is likely to be larger.
Problem 1.7.1. How much time does multiplication a la russe take if the mul-
tiplier is longer than the multiplicand ?
As we mentioned in Section 1.1, more efficient algorithms exist to solve this

problem. The simplest, which we shall study in Section 4.7, takes a time in the order
of nmlg(3/2), or approximately nm0.59, where n is the size of the larger operand and m is
the size of the smaller. If both operands are of size n, the algorithm thus takes a time
in the order of n 1.599 which is preferable to the quadratic time taken by both the classic
algorithm and multiplication a la russe.
The difference between the order of n 2 and the order of n 1.59 is less spectacular
than that between the order of n 2 and the order of n log n , which we saw in the case
of sorting algorithms. To verify this, we programmed the classic algorithm and the
algorithm of Section 4.7 in Pascal on a CDc CYBER 835 and tested them on operands of
different sizes. To take account of the architecture of the machine, we carried out the
calculations in base 220 rather than in base 10. Integers of 20 bits are thus multiplied
directly by the hardware of the machine, yet at the same time space is used quite
efficiently (the machine has 60-bit words). Accordingly, the size of an operand is
measured in terms of the number of 20-bit segments in its binary representation. The
theoretically better algorithm of Section 4.7 gives little real improvement on operands
of size 100 (equivalent to about 602 decimal digits) : it takes about 300 milliseconds,
whereas the classic algorithm takes about 400 milliseconds. For operands ten times
this length, however, the fast algorithm is some three times more efficient than the
classic algorithm : they take about 15 seconds and 40 seconds, respectively. The gain
in efficiency continues to increase as the size of the operands goes up. As we shall see
in Chapter 9, even more sophisticated algorithms exist for much larger operands.
1.7.3 Evaluating Determinants
Let
a1,1 a1,2 ... a1,n

a2,1 a2,2 ... a2,n
M =
I an l an 2 ... an ,n
j
be an n x n matrix. The determinant of the matrix M, denoted by det(M), is often
defined recursively : if M [i, j ] denotes the (n - 1) x (n - 1) submatrix obtained from M
by deleting the i th row and the j th column, then
= n
det(M) Y, (-1)j+'a1,1 det(M[1, j ]) .
j=1
If n = 1, the determinant is defined by det(M) = a 1,1 . Determinants are important in
linear algebra, and we need to know how to calculate them efficiently.
If we use the recursive definition directly, we obtain an algorithm that takes a
time in the order of n! to calculate the determinant of an n x n matrix (see Example
2.2.5). This is even worse than exponential. On the other hand, another classic algo-
rithm, Gauss-Jordan elimination, does the computation in cubic time. We programmed
the two algorithms in Pascal on a Coc CYBER 835. The Gauss-Jordan algorithm finds
the determinant of a 10 x 10 matrix in one-hundredth of a second ; it takes about five
and a half seconds on a 100 x 100 matrix. On the other hand, the recursive algorithm
takes more than 20 seconds on a 5 x 5 matrix and 10 minutes on a 10 x 10 matrix ; we
estimate that it would take more than 10 million years to calculate the determinant of a
20 x 20 matrix, a task accomplished by the Gauss-Jordan algorithm in about one-
twentieth of a second !
You should not conclude from this example that recursive algorithms are neces-
sarily bad. On the contrary, Chapter 4 describes a technique where recursion plays a
fundamental role in the design of efficient algorithms. In particular, Strassen
discovered in 1969 a recursive algorithm that can calculate the determinant of an n x n
matrix in a time in the order of n 1g7 or about n 2.81, thus proving that Gauss-Jordan
elimination is not optimal.
1.7.4 Calculating the Greatest Common Divisor
Let m and n be two positive integers. The greatest common divisor of m and n,
denoted by gcd(m, n), is the largest integer that divides both m and n exactly. When
gcd(m, n) =1, we say that m and n are coprime. For example, gcd(6,15) = 3 and
gcd(10, 21) = 1. The obvious algorithm for calculating gcd(m , n) is obtained directly
from the definition.
function ged (m, n)
i - min(m , n) + 1
repeat i F- i - 1 until i divides both m and n exactly
return i
The time taken by this algorithm is in the order of the difference between the
smaller of the two arguments and their greatest common divisor. When m and n are of
similar size and coprime, it therefore takes a time in the order of n.
A classic algorithm for calculating gcd(m, n) consists of first factorizing m and
n, and then taking the product of the prime factors common to m and n, each prime
factor being raised to the lower of its powers in the two arguments. For example, to
calculate gcd(120, 700) we first factorize 120=2 3 x 3 x 5 and 700=2 2 x5 2 x 7. The
common factors of 120 and 700 are therefore 2 and 5, and their lower powers are 2 and
1, respectively. The greatest common divisor of 120 and 700 is therefore 22 x 51 = 20.
Even though this algorithm is better than the one given previously, it requires us to
factorize m and n, an operation we do not know how to do efficiently.
Nevertheless, there exists a much more efficient algorithm for calculating greatest
common divisors. This is Euclid's famous algorithm.
function Euclid (m, n)
while m > 0 do
tF- nmodm
n F-m
mt
return n
Considering the arithmetic operations to have unit cost, this algorithm takes a time in
the order of the logarithm of its arguments, even in the worst case (see Example 2.2.6),
which is much faster than the preceding algorithms. To be historically exact, Euclid's
original algorithm works using successive subtractions rather than by calculating a
modulo.
1.7.5 Calculating the Fibonacci Sequence
The Fibonacci sequence is defined by the following recurrence :

fo0; fI = 1 and
fn fn -I +fn-2 for n ? 2.
The first ten terms of the sequence are therefore
0, 1, 1, 2, 3, 5, 8, 13, 21, 34.
This sequence has numerous applications in computer science, in mathematics, and in
the theory of games. It is when applied to two consecutive terms of the Fibonacci
sequence that Euclid's algorithm takes the longest time among all instances of compar-
able size. De Moivre proved the following formula (see Example 2.3.2) :
L = _1 [on - (-4)-n
where 4 _ (I +x(5)/2 is the golden V rJatio. Since 4 < 1, the term (-4)-n can be
neglected when n is large, which means that the value of fn is in the order of 0n .
However, De Moivre's formula is of little immediate help in calculating fn exactly,
since the larger n becomes, the greater is the degree of precision required in the values
of 5 and 0. On the CDC CYBER 835, a single-precision computation programmed in
Pascal produces an error for the first time when calculating f66 .
The algorithm obtained directly from the definition of the Fibonacci sequence is
the following.
function fib 1(n)
if n < 2 then return n
else return fibl(n - 1) + fib 1(n-2)
This algorithm is very inefficient because it recalculates the same values many times.
For instance, to calculate fib l (5) we need the values of fib 1(4) and fib l (3) ; but
fib 1(4) also calls for the calculation of fib 1(3). We see that fib 1(3) will be calculated
twice, fibl(2) three times, fibl(1) five times, and fibl(0) three times. In fact, the time
required to calculate f, using this algorithm is in the order of the value of f, itself,
that is to say, in the order of 0" (see Example 2.2.7).
To avoid wastefully recalculating the same values over and over, it is natural to
proceed as in Section 1.5.
function fib2(n)
i E- 1; j - 0
fork -ltondoj -i+j
i 4j-i
return j
This second algorithm takes a time in the order of n, assuming we count each addition
as an elementary operation (see Example 2.2.8). This is much better than the first
algorithm. However, there exists a third algorithm that gives as great an improvement
over the second algorithm as the second does over the first. This third algorithm,
which at first sight appears quite mysterious, takes a time in the order of the logarithm
of n (see Example 2.2.9). It will be explained in Chapter 4.
function fib3(n)
i - 1; j E-0; k -0; h E-1
while n > 0 do
if n is odd then t F-- jh
j - ih+jk+t
i - ik+t
t f- h2
h F- 2kh + t
k E--k2+t
n E-n div 2
return j
Once again, we programmed the three algorithms in Pascal on a CDC CYBER 835
in order to compare their execution times empirically. To avoid problems caused by
arithmetic overflow (the Fibonacci sequence grows very rapidly : fl00 is a number with
21 decimal digits), we carried out all the computations modulo 107, which is to say
that we only obtained the seven least significant figures of the answer. Table 1.7.1 elo-
quently illustrates the difference that the choice of an algorithm can make. (All these
times are approximate. Times greater than two minutes were estimated using the hybrid
approach.) The time required by fib l for n > 50 is so long that we did not bother to
estimate it, with the exception of the case n = 100 on which fib 1 would take well over
109 years ! Note that fib2 is more efficient than fib3 on small instances.
TABLE 1.7.1 PERFORMANCE COMPARISON

BETWEEN MODULO 107 FIBONACCI ALGORITHMS
n 10 20 30 50
fibl 8 msec I sec 2 min 21 days

fib2 I msec
6
! msec
3
I msec
2
1- msec
4
fib3 I msec 2 msec I msec i msec
3 5 2
n 100 10,000 1,000,000 100,000,000
fib2 1 i msec 150 msec 15 sec 25 min

fib3 msec I msec 1 msec 2 msec
2 Z
Using the hybrid approach, we can estimate approximately the time taken by our
implementations of these three algorithms. Writing t, (n) for the time taken by fibi on
the instance n, we find
t1(n) ^ 0,-20 seconds,
t2(n) = 15n microseconds, and

t3(n) = 1logn milliseconds.
It takes a value of n 10,000 times larger to make fib3 take one extra millisecond of
computing time.
We could also have calculated all the figures in the answer using multiple-
precision arithmetic. In this case the advantage that fib3 enjoys over fib2 is less
marked, but their joint advantage over fibi remains just as striking. Table 1.7.2 com-
pares the times taken by these three algorithms when they are used in conjunction with
an efficient implementation of the classic algorithm for multiplying large integers (see
Problem 2.2.9, Example 2.2.8, and Problem 2.2.11).
TABLE 1.7.2. PERFORMANCE COMPARISON

BETWEEN EXACT FIBONACCI ALGORITHMS*
n 5 10 15 20 25
fib l 0.007 0.087 0.941 10.766 118.457

fib2 0.005 0.009 0.011 0.017 0.021
fib3 0.013 0.017 0.019 0.020 0.021
n 100 500 1,000 5,000 10,000
fib2 0.109 1.177 3.581 76.107 298.892

fib3 0.041 0.132 0.348 7.664 29.553
* All times are in seconds.

Sec. 1.8 When Is an Algorithm Specified ? 19
1.7.6 Fourier Transforms
The Fast Fourier Transform algorithm is perhaps the one algorithmic discovery that
had the greatest practical impact in history. We shall come back to this subject in
Chapter 9. For the moment let us only mention that Fourier transforms are of funda-
mental importance in such disparate applications as optics, acoustics, quantum physics,
telecommunications, systems theory, and signal processing including speech recogni-
tion. For years progress in these areas was limited by the fact that the known algo-
rithms for calculating Fourier transforms all took far too long.
The "discovery" by Cooley and Tukey in 1965 of a fast algorithm revolutionized
the situation : problems previously considered to be infeasible could now at last be
tackled. In one early test of the "new" algorithm the Fourier transform was used to
analyse data from an earthquake that had taken place in Alaska in 1964. Although the
classic algorithm took more than 26 minutes of computation, the "new" algorithm was
able to perform the same task in less than two and a half seconds. .
Ironically it turned out that an efficient algorithm had already been published in
1942 by Danielson and Lanczos. Thus the development of numerous applications had
been hindered for no good reason for almost a quarter of a century. And if that were
not sufficient, all the necessary theoretical groundwork for Danielson and Lanczos's
algorithm had already been published by Runge and Konig in 1924!
1.8 WHEN IS AN ALGORITHM SPECIFIED?
At the beginning of this book we said that "the execution of an algorithm must not
include any subjective decisions, nor must it require the use of intuition or creativity".
In this case, can we reasonably maintain that fib3 of Section 1.7.5 describes an algo-
rithm ? The problem arises because it is not realistic to consider that the multiplica-
tions in fib3 are elementary operations. Any practical implementation must take this
into account, probably by using a program package allowing arithmetic operations on
very large integers. Since the exact way in which these multiplications are to be car-
ried out is not specified in fib3, the choice may be considered a subjective decision,
and hence fib3 is not formally speaking an algorithm. That this distinction is not
merely academic is illustrated by Problems 2.2.11 and 4.7.6, which show that indeed
the order of time taken by fib3 depends on the multiplication algorithm used. And
what should we say about De Moivre's formula used as an algorithm ?
Calculation of a determinant by the recursive method of Section 1.7.3 is another
example of an incompletely presented algorithm. How are the recursive calls to be set
up? The obvious approach requires a time in the order of n 2 to be used before each
recursive call. We shall see in Problem 2.2.5 that it is possible to get by with a time in
the order of n to set up not just one, but all the n recursive calls. However, this added
subtlety does not alter the fact that the algorithm takes a time in the order of n! to cal-
culate the determinant of an n x n matrix.
To make life simple, we shall continue to use the word algorithm for certain
incomplete descriptions of this kind. The details will be filled in later should our ana-
lyses require them.
1.9 DATA STRUCTURES
The use of well-chosen data structures is often a crucial factor in the design of efficient
algorithms. Nevertheless, this book is not intended to be a manual on data structures.
We suppose that the reader already has a good working knowledge of such basic
notions as arrays, structures, pointers, and lists. We also suppose that he or she has
already come across the mathematical concepts of directed and undirected graphs, and
knows how to represent these objects efficiently on a computer. After a brief review of
some important points, this section concentrates on the less elementary notions of
heaps and disjoint sets. Chosen because they will be used in subsequent chapters,
these two structures also offer interesting examples of the analysis of algorithms (see
Example 2.2.4, Problem 2.2.3, and Example 2.2.10).
1.9.1 Lists
A list is a collection of nodes or elements of information arranged in a certain order.

The corresponding data structure must allow us to determine efficiently, for example,
which is the first node in the structure, which is the last, and which are the predecessor
and the successor (if they exist) of any given node. Such a structure is frequently
represented graphically by boxes and arrows, as in Figure 1.9.1. The information
attached to a node is shown inside the corresponding box and the arrows show transi-
tions from a node to its successor.
Such lists are subject to a number of operations : we might want to insert an
additional node, to delete a node, to copy a list, to count the number of elements it
contains, and so on. The different computer implementations that are commonly used
differ in the quantity of memory required, and in the greater or less ease of carrying
out certain operations. Here we content ourselves with mentioning the best-known
techniques.
Implemented as an array by the declaration
type tablist = record
counter : 0 .. maxlength
value[ I.. maxlength ] : information
the elements of a list occupy the slots value [ 1 ] to value [counter ], and the order of the
elements is given by the order of their indices in the array. Using this implementation,
alpha l-il beta as gamma 31 delta

Figure 1.9.1. A list.
Sec. 1.9 Data Structures 21
we can find the first and the last elements of the list rapidly, as we can the predecessor
and the successor of a given node. On the other hand, inserting a new element or
deleting one of the existing elements requires a worst-case number of operations in the
order of the current size of the list.
This implementation is particularly efficient for the important structure known as
the stack, which we obtain by restricting the permitted operations on a list : addition
and deletion of elements are allowed only at one particular end of the list. However, it
presents the major disadvantage of requiring that all the memory space potentially
required be reserved from the outset of a program.
On the other hand, if pointers are used to implement a list structure, the nodes
are usually represented by some such structure as
type node = record
value : information
next : T node ,
where each node includes an explicit pointer to its successor. In this case, provided a
suitably powerful programming language is used, the space needed to represent the list
can be allocated and recovered dynamically as the program proceeds.
Even if additional pointers are used to ensure rapid access to the first and last
elements of the list, it is difficult when this representation is used to examine the k th
element, for arbitrary k, without having to follow k pointers and thus to take a time in
the order of k. However, once an element has been found, inserting new nodes or
deleting an existing node can be done rapidly. In our example, a single pointer is used
in each node to designate its successor : it is therefore easy to traverse the list in one
direction, but not in the other. If a higher memory overhead is acceptable, it suffices to
add a second pointer to each node to allow the list to be traversed rapidly in either
direction.
1.9.2 Graphs
Intuitively speaking, a graph is a set of nodes joined by a set of lines or arrows. Con-
sider Figure 1.9.2 for instance. We distinguish directed and undirected graphs. In the
case of a directed graph the nodes are joined by arrows called edges. In the example
of Figure 1.9.2 there exists an edge from alpha to gamma and another from gamma to
alpha; beta and delta, however, are joined only in the direction indicated. In the case
of an undirected graph, the nodes are joined by lines with no direction indicated, also
called edges. In every case, the edges may form paths and cycles.
There are never more than two arrows joining any two given nodes of a directed
graph (and if there are two arrows, then they must go in opposite directions), and there
is never more than one line joining any two given nodes of an undirected graph. For-
mally speaking, a graph is therefore a pair G = < N, A > where N is a set of nodes
and A c N x N is a set of edges. An edge from node a to node b of a directed graph
is denoted by the ordered pair (a, b), whereas an edge joining nodes a and b in an
undirected graph is denoted by the set { a, b } .
alpha
gamma delta
Figure 1.9.2. A directed graph.
There are at least two obvious ways to represent a graph on a computer. The
first is illustrated by
type adjgraph = record

value [ 1 .. nbnodes ] : information
adjacent [ 1 .. nbnodes,1 .. nbnodes ] : Booleans
If there exists an edge from node i of the graph to node j, then adjacent [i , j ] = true ;
otherwise adjacent [i, j ] = false. In the case of an undirected graph, the matrix is
necessarily symmetric.
With this representation it is easy to see whether or not two nodes are connected.
On the other hand, should we wish to examine all the nodes connected to some given
node, we have to scan a complete row in the matrix. This takes a time in the order of
nbnodes, the number of nodes in the graph, independently of the number of edges that
exist involving this particular node. The memory space required is quadratic in the
number of nodes.
A second possible representation is as follows :
type lisgraph = array[ L. nbnodes ] of

record
value : information
neighbours : list .
Here we attach to each node i a list of its neighbours, that is to say of those nodes j
such that an edge from i to j (in the case of a directed graph) or between i and j
(in the case of an undirected graph) exists. If the number of edges in the graph is
small, this representation is preferable from the point of view of the memory space
used. It may also be possible in this case to examine all the neighbours of a given
node in less than nbnodes operations on the average. On the other hand, to determine
whether or not two given nodes i and j are connected directly, we have to scan the list
of neighbours of node i (and possibly of node j, too), which is less efficient than
looking up a Boolean value in an array.
A tree is an acyclic, connected, undirected graph. Equivalently, a tree may be
defined as an undirected graph in which there exists exactly one path between any
given pair of nodes. The same representations used to implement graphs can be used
to implement trees.
1.9.3 Rooted Trees
Let G be a directed graph. If there exists in G a vertex r such that every other vertex
can be reached from r by a unique path, then G is a rooted tree and r is its root. Any
rooted tree with n nodes contains exactly n -1 edges. It is usual to represent a rooted
tree with the root at the top, like a family tree, as in Figure 1.9.3. In this example
alpha is at the root of the tree. (When there is no danger of confusion, we shall use
the simple term "tree" instead of the more correct "rooted tree".) Extending the
analogy with a family tree, we say that beta is the parent of delta and the child of
alpha, that epsilon and zeta are the siblings of delta, that alpha is an ancestor of
epsilon, and so on.
Figure 1.9.3. A rooted tree.
A leaf of a rooted tree is a node with no children; the other nodes are called
internal nodes. Although nothing in the definition indicates this, the branches of a
rooted tree are often considered to be ordered : in the previous example beta is
situated to the left of gamma, and (by analogy with a family tree once again) delta is
the eldest sibling of epsilon and zeta. The two trees in Figure 1.9.4 may therefore be
considered as different.
lambda lambda
Figure 1.9.4. Two distinct rooted trees.
On a computer, any rooted tree may be represented using nodes of the following
type :
type treenode = record
value : information
eldest-child, next-sibling :T treenode
The rooted tree shown in Figure 1.9.3 would be represented as in Figure 1.9.5, where
now the arrows no longer represent the edges of the rooted tree, but rather the pointers
used in the computer representation. As in the case of lists, the use of additional
pointers (for example, to the parent or the eldest sibling of a given node) may speed up
certain operations at the price of an increase in the memory space needed.
alpha
beta gamma
VV
delta
1A
epsilon zeta
M
Figure 1.9.5. Possible computer representation of a rooted tree.
The depth of a node in a rooted tree is the number of edges that need to be
traversed to arrive at the node starting from the root. The height of a node is the
number of edges in the longest path from the node in question to a leaf. The height of
a rooted tree is the height of its root, and thus also the depth of its deepest leaf.
Finally, the level of a node is equal to the height of the tree minus the depth of the
node concerned. For example, gamma has depth 1, height 0, and level 1 in the tree of
Figure 1.9.3.
If each node of a rooted tree can have up to n children, we say it is an n-ary tree.
In this case, the positions occupied by the children are significant. For instance, the
binary trees of Figure 1.9.6 are not the same : in the first case b is the elder child of a
and the younger child is missing, whereas in the second case b is the younger child of
a and the elder child is missing. In the important case of a binary tree, although the
metaphor becomes somewhat strained, we naturally tend to talk about the left-hand
child and the right-hand child.
There are several ways of representing an n-ary tree on a computer. One obvious
representation uses nodes of the type
type n-ary-node = record
value : information
child [ 1 .. n ] : T n-ary-node
Figure 1.9.6. Two distinct binary trees.

In the case of a binary tree we can also define

type binary-node = record
value : information
left-child, right-child : T binary-node
It is also sometimes possible, as we shall see in the following section, to represent a

rooted tree using an array without any explicit pointers.
A binary tree is a search tree if the value contained in every internal node is
larger than or equal to the values contained in its left-hand descendants, and less than
or equal to the values contained in its right-hand descendants. An example of a search
tree is given in Figure 5.5.1. This structure is interesting because it allows efficient
searches for values in the tree.
Problem 1.9.1. Suppose the value sought is held in a node at depth p in a

search tree. Design an algorithm capable of finding this node starting at the root in a
time in the order of p.
It is possible to update a search tree, that is, to delete nodes or to add new values,
without destroying the search tree property. However, if this is done in an uncon-
sidered fashion, it can happen that the resulting tree becomes badly unbalanced, in the
sense that the height of the tree is in the order of the number of nodes it contains.
More sophisticated methods, such as the use of AVL trees or 2-3 trees, allow such
operations as searches and the addition or deletion of nodes in a time in the order of
the logarithm of the number of nodes in the tree in the worst case. These structures
also allow the efficient implementation of several additional operations. Since these
concepts are not used in the rest of this book, here we only mention their existence.
1.9.4 Heaps
A heap is a special kind of rooted tree that can be implemented efficiently in an array
without any explicit pointers. This interesting structure lends itself to numerous appli-
cations, including a remarkable sorting technique, called heapsort (see Problem 2.2.3),
as well as the efficient implementation of certain dynamic priority lists.
A binary tree is essentially complete if each of its internal nodes possesses
exactly two children, one on the left and one on the right, with the possible exception
of a unique special node situated on level 1, which possesses only a left-hand child and
no right-hand child. Moreover, all the leaves are either on level 0, or else they are on
levels 0 and 1, and no leaf is found on level 1 to the left of an internal node at the
same level. The unique special node, if it exists, is to the right of all the other level 1
internal nodes. This kind of tree can be represented using an array T by putting the
nodes of depth k, from left to right, in the positions T [2k], T [2k+1], ... , T[2 k+1_11
(with the possible exception of level 0, which may be incomplete). For instance,
Figure 1.9.7 shows how to represent an essentially complete binary tree containing 10
nodes. The parent of the node represented in T [i] is found in T [i div 2] for i > 1, and
T[4] T[5] T(6] T[7]
T[I0]
Figure 1.9.7. An essentially complete binary tree.
the children of the node represented in T[i] are found in T[2i] and T [2i + 1 ], whenever
they exist. The subtree whose root is in T [i] is also easy to identify.
A heap is an essentially complete binary tree, each of whose nodes includes an
element of information called the value of the node. The heap property is that the
value of each internal node is greater than or equal to the values of its children. Figure
1.9.8 gives an example of a heap. This same heap can be represented by the following
array :
10 7 9 4 7 5 2 2 1 6
The fundamental characteristic of this data structure is that the heap property can
be restored efficiently after modification of the value of a node. If the value of the
node increases, so that it becomes greater than the value of its parent, it suffices to
exchange these two values and then to continue the same process upwards in the tree
until the heap property is restored. We say that the modified value has been percolated
up to its new position (one often encounters the rather strange term sift-up for this pro-
cess). If, on the contrary, the value of a node is decreased so that it becomes less than
the value of at least one of its children, it suffices to exchange the modified value with
Figure 1.9.8. A heap.

the larger of the values in the children, and then to continue this process downwards in
the tree until the heap property is restored. We say that the modified value has been
sifted down to its new position. The following procedures describe more formally the
basic heap manipulation process. For the purpose of clarity, they are written so as to
reflect as closely as possible the preceding discussion. If the reader wishes to make
use of heaps for a "real" application, we encourage him or her to figure out how to
avoid the inefficiency resulting from our use of the "exchange" instruction.
procedure alter-heap (T [ 1 .. n ], i, v )
{ T [I .. n] is a heap ; the value of T[i] is set to v and the
heap property is re-established ; we suppose that 1<_ i<_ n }
x - T[i]
T[i]<-v
if v < x then sift-down (T, i )
else percolate (T, i)
procedure sift-down (T [1 .. n ], i )
{this procedure sifts node i down so as to re-establish the heap
property in T [I .. n ] ; we suppose that T would be a heap if T [i ]
were sufficiently large ; we also suppose that 1 <_ i 5 n }
k F-i
repeat
j *-k
{ find the larger child of node j }
if 2j <_ n and T [2j ] > T [k] then k - 2j
if2j <n andT[2j+1]>T[k]thenk -2j+1
exchange T [ j ] and T [k]
{ if j = k, then the node has arrived at its final position)
until j = k
procedure percolate (T [1 .. n ], i )
{ this procedure percolates node i so as to re-establish the
heap property in T [I .. n I; we suppose that T would be a heap
if T [i] were sufficiently small ; we also suppose that
1 <_ i <_ n ; the parameter n is not used here }
k-i
repeat
j *-k
if j > 1 and T [j div 2] < T [k] then k - j div 2
exchange T [ j ] and T [k]
until j = k
The heap is an ideal data structure for finding the largest element of a set, removing it,
adding a new node, or modifying a node. These are exactly the operations we need to
implement dynamic priority lists efficiently : the value of a node gives the priority of
the corresponding event. The event with highest priority is always found at the root of
the heap, and the priority of an event can be changed dynamically at all times. This is
particularly useful in computer simulations.
function find-max (T [1 .. n ])
{ returns the largest element of the heap T [ 1 .. n
return T[1]
procedure delete-max (T [1 .. n ])
{ removes the largest element of the heap T [I .. n ]
and restores the heap property in T [I .. n -1 ] }
T[1]-T[n]
sift-down (T [ 1 .. n -1 ],1)
procedure insert-node (T [ 1 .. n ], v)
{ adds an element whose value is v to the heap T [ 1 .. n ]
and restores the heap property in T [I .. n + 1 ] }
T[n+1] - v
percolate (T [1 .. n + 1], n + 1)
It remains to be seen how we can create a heap starting from an array T [I .. n ]

of elements in an undefined order. The obvious solution is to start with an empty heap
and to add elements one by one.
procedure slow-make-heap (T [ 1 .. n ])
{ this procedure makes the array T [I .. n ] into a heap,
but rather inefficiently)
for i - 2 to n do percolate (T [1 .. i ], i )
However, this approach is not particularly efficient (see Problem 2.2.2). There exists a
cleverer algorithm for making a heap. Suppose, for example, that our starting point is
the following array :
1 6 9 2 7 5 2 7 4 10
represented by the tree of Figure 1.9.9a. We begin by making each of the subtrees
whose roots are at level 1 into a heap, by sifting down those roots, as illustrated in
Figure 1.9.9b. The subtrees at the next higher level are then transformed into heaps,
also by sifting down their roots. Figure 1.9.9c shows the process for the left-hand sub-
tree. The other subtree at level 2 is already a heap. This results in an essentially com-
plete binary tree corresponding to the array :
1 10 9 7 7 5 2 2 4 6
It only remains to sift down its root in order to obtain the desired heap. The final pro-
cess thus goes as follows :
10 1 9 7 7 5 2 2 4 6
10 7 9 1 7 5 2 2 4 6
10 7 9 4 7 5 2 2 1 6
whose tree representation is shown in the previous example as Figure 1.9.8.

This algorithm can be described formally as follows.
procedure make-heap (T [ 1 .. n ])
{ this procedure makes the array T [ 1 .. n ] into a heap }
for i - (n div 2) downto 1 do sift-down (T, i)
We shall see in Example 2.2.4 that this algorithm allows the creation of a heap in
linear time.
(a) The starting situation.
(b) The level t subtrees are made into heaps.
(c) One level 2 subtree is made into a heap (the other already is a heap).
Figure 1.9.9. Making a heap.

Problem 1.9.2. Let T [I .. 12] be an array such that T [i ] = i for each i <-12.
Exhibit the state of the array after each of the following procedure calls :
make-heap (T)
alter-heap (T, 12, 10)
alter-heap (T, 1, 6)
alter-heap (T, 5, 8) .
Problem 1.9.3. Exhibit a heap T [I .. n ] containing distinct values, such that

the following sequence results in a different heap :
m - find-max (T)
delete-max (T)
insert-node (T [l .. n -1], m)
Draw the heap after each operation.
The basic concept of a heap can be improved in several ways. For applications
that need percolate more often than sift-down (Problems 3.2.12 and 3.2.13) it pays to
have more than two children per internal node : this speeds up percolate (because the
heap is shallower) at the cost of slowing down any operation that must consider every
child at each level. It is still possible to represent such heaps in an array without
explicit pointers, but care is needed to do it correctly.
For applications that have a tendency to sift down the (updated) root almost to
the bottom level, it pays to ignore temporarily the new value stored at the root, treating
this node as an empty location, and to sift it all the way down to a leaf. At this point,
put back the relevant value into the empty leaf and percolate it to its proper position.
The advantage of this procedure is that it requires only one comparison at each level in
order to sift down the empty node, rather than two with the usual procedure (the chil-
dren are compared to each other but not to their father). Experiments have shown this
approach to yield an improvement in the classic heapsort algorithm (Problem 2.2.3).
Finally, the basic heap operations needed to implement a dynamic priority list
can also be handled by data structures completely different from the heap we have con-
sidered so far. In particular, the Fibonacci heap (or lazy binomial queue) can process
each insert-node, find-max, and percolate operation in constant time, and each
delete-max in logarithmic time. As an application, this data structure allows the imple-
mentation of Dijkstra's algorithm in a time in the order of a+ n log n for finding the
length of the shortest path from a designated source node to each of the other nodes of
a graph with n nodes and a edges (Section 3.2.2). The Fibonacci heap allows also the
merging of priority lists in constant time, an operation beyond the (efficient) reach of
classic heaps. (To be precise, the preceding times for Fibonacci heaps are correct in
the amortized sense-a concept not discussed here.)
1.9.5 Disjoint Set Structures
Suppose we have N objects numbered from 1 to N. We wish to group these objects

into disjoint sets, each object being in exactly one set at any given time. In each set
we choose a canonical object, which will serve as a label for the set. Initially, the N
objects are in N different sets, each containing exactly one object, which is necessarily
the label for its set. Thereafter, we execute a series of operations of two kinds :
for a given object, find which set contains it and return the label of this set ; and
given two distinct labels, merge the two corresponding sets.
How can we represent this situation efficiently on a computer?

One possible representation is obvious. Suppose we decide to use the smallest
member of each set as the label : thus the set (7,3, 16,91 will be called "set 3". If we
now declare an array set [ 1 .. N ], it suffices to place the label of the set corresponding
to each object in the appropriate array element. The two operations can be imple-
mented by two procedures.
function find 1(x)

[ finds the label of the set containing object x }
return set [x]
procedure merge 1(a , b)
{merges the sets labelled a and b }
i f- a; j -b
if i > j then exchange i and j
fork E- I to N do
if set [k] = j then set [k] F- i
We wish to know the time required to execute an arbitrary series of n operations

of the type find and merge, starting from the initial situation. If consulting or modi-
fying one element of an array counts as an elementary operation, it is clear that find l
takes constant time and that mergel takes a time in the order of N. A series of n
operations therefore takes a time in the order of nN in the worst case.
We can do better than this. Still using a single array, we can represent each set
as an "inverted" rooted tree. We adopt the following scheme : if set [i ] = i, then i is
both the label of a set and the root of the corresponding tree ; if set [i] =j# i, then j is
the parent of i in some tree. The array
1 2 3 2 1 3 4 3 4
therefore represents the trees given in Figure 1.9.10, which in turn represent the sets
11, 51, (2,4,7,10) and (3,6,8,9). To merge two sets, we need now only change a
single value in the array ; on the other hand, it is harder to find the set to which an
object belongs.
Figure 1.9.10. Tree representation for disjoint sets.

(The figure shows the direction of the pointers in the array, not of the edges in the tree.)
function find2(x)
{ finds the label of the set containing object x }
i -x
while set [i ] 91- i do i - set [i ]
return i
procedure merge2(a, b)
[ merges the sets labelled a and b }
if a < b then set [b] - a
else set [a] - b
Problem 1.9.4. If each consultation or modification of an array element

counts as an elementary operation, prove that the time needed to execute an arbitrary
sequence of n operations find2 or merge2 starting from the initial situation is in the
order of n 2 in the worst case.
In the case when n is comparable to N, we have not gained anything over the use
of findl and mergel. The problem arises because after k calls on merge2, we may find
ourselves confronted by a tree of height k, so that each subsequent call on find2 takes a
time in the order of k. Let us therefore try to limit the height of the trees produced.
So far, we have chosen arbitrarily to use the smallest member of a set as its label.
When we merge two trees whose heights are respectively h 1 and h2, it would be better
to arrange matters so that it is always the root of the tree whose height is least that
becomes a child of the other root. In this way the height of the resulting merged tree
will be max(h 1, h 2) if h 1 #h2, or h 1+1 if h 1=h2- Using this technique, the trees do
not grow as rapidly.
Problem 1.9.5. Prove by mathematical induction that if this tactic is used,

then after an arbitrary sequence of merge operations starting from the initial situation, a
tree containing k nodes will have a height at most Llg kJ .
The height of the trees can be maintained in an additional array height [1 .. N] so

that height [i] gives the height of node i in its current tree. Whenever a is the label of
a set, height [a] therefore gives the height of the corresponding tree. Initially, height [i]
is set to zero for each i. The procedure find2 is still relevant but we must modify
merge accordingly.
procedure merge3(a, b)
{ merges the sets labelled a and b
we suppose that a # b }
if height [a] = height [b]
then
height [a] - height [a] + 1
set [b] <- a
else
if height [a] > height [b]
then set [b] F- a
else set [a] F- b
Problem 1.9.6. Prove that the time needed to execute an arbitrary sequence of
n operations find2 and merge3 starting from the initial situation is in the order of
n log n in the worst case.
By modifying find2, we can make our operations faster still. When we are
trying to determine the set that contains a certain object x, we first traverse the edges
of the tree leading up from x to the root. Once we know the root, we can now traverse
the same edges again, modifying each node encountered on the way to set its pointer
directly to the root. This technique is called path compression. For example, when we
execute the operation find (20) on the tree of Figure 1.9.1 la, the result is the tree of
4 9 4 9 20
11 10 16
1 8 11
1 G )a
(b) after
12 20
21 16
(a) before
Figure 1.9.11. Path compression.

Figure 1.9.1lb: nodes 20, 10, and 9, which lay on the path from node 20 to the root,
now point directly to the root. The pointers of the remaining nodes have not changed.
This technique obviously tends to diminish the height of a tree and thus to accelerate
subsequent find operations. On the other hand, the new find operation takes about
twice as long as before. Is path compression a good idea? The answer is given when
we analyse it in Example 2.2.10.
Using path compression, it is no longer true that the height of a tree whose root
is a is given by height [a]. However, this remains an upper bound on the height. We
call this value the rank of the tree, and change the name of the array accordingly. Our
function becomes
function find3(x )
{ finds the label of the set containing object x }
r Fx
while set [r ] r do r - set [r ]
{ r is the root of the tree }
i <-- x
while i # r do
j E- set [i]
set [i ] - r
i F-j
return r
From now on, when we use this combination of an array and of the procedures find3
and merge3 to deal with disjoint sets of objects, we say we are using a disjoint set
structure.
Problem 1.9.7. A second possible tactic for merging two sets is to ensure that
the root of the tree containing the smaller number of nodes becomes the child of the
other root. Path compression does not change the number of nodes in a tree, so that it
is easy to store this value exactly (whereas we could not efficiently keep track of the
exact height of a tree after path compression).
Write a procedure merge4 to implement this tactic, and give a result cor-
responding to the one in Problem 1.9.5.
** Problem 1.9.8. Analyse the combined efficiency of find3 together with your
merge4 from the previous problem.
Problem 1.9.9. A canonical object has no parent, and we make no use of the
rank of any object that is not canonical. Use this remark to implement a disjoint set
structure that uses only one length N array rather than the two set and rank. (Hint :
use negative values for the ranks.)
Sec. 1.10 References and Further Reading 35
1.10 REFERENCES AND FURTHER READING
We distinguish between three kinds of books on algorithm design. Specific books

cover algorithms that are useful in a particular application area : sorting, searching,
graph theory, Fourier transforms, and so on. General books cover several application
areas : they give algorithms useful in each area. Finally, books on algorithmics con-
centrate on the techniques of algorithm design : they illustrate each technique by
examples of algorithms taken from various applications areas. The excellent book of
Harel (1987) takes a broader view at algorithmics and considers it as no less than "the
spirit of computing".
The most complete collection of algorithms ever proposed is no doubt found in
Knuth (1968, 1969, 1973), originally intended to consist of seven volumes. Several
other well-known general books are worth mentioning : in chronological order Aho,
Hopcroft, and Ullman (1974), Baase (1978), Dromey (1982), Sedgewick (1983), and
Melhorn (1984a, 1984b, 1984c). Specific books will be referred to in the following
chapters whenever they are relevant to our discussion ; we may mention, however,
Nilsson (1971), Brigham (1974), Borodin and Munro (1975), Christofides (1975),
Lawler (1976), Reingold, Nievergelt, and Deo (1977), Gondran and Minoux (1979),
Even (1980), Papadimitriou and Steiglitz (1982), and Tarjan (1983). Besides our own
book and Harel (1987), we are aware of two more works on algorithmics : Horowitz
and Sahni (1978) and Stinson (1985).
For a more popular account of algorithms, see Knuth (1977) and Lewis and
Papadimitriou (1978). Multiplication a la russe is described in Warusfel (1961), a
remarkable little French book of popular mathematics. Although we do not use any
specific programming language in the present book, we suggest that a reader unfamiliar
with Pascal would do well to look at one of the numerous books on this language, such
as Jensen and Wirth (1985) or Lecarme and Nebut (1985).
Knuth (1973) contains a large number of sorting algorithms. Problem 1.5.1
comes from Shamir (1979). The algorithm capable of calculating the determinant of an
n x n matrix in a time in the order of n 2'81 is given in Strassen (1969) and Bunch and
Hopcroft (1974). The fast algorithm for calculating the Fibonacci sequence is
explained in Gries and Levin (1980) and Urbanek (1980). For an introduction to
Fourier transforms, read Demars (1981). We give more references on this subject in
Chapter 9.
To reinforce our remarks in Sections 1.6 and 1.7, we encourage the reader to
look at Bentley (1984), which offers experimental proof that intelligent use of algo-
rithmics may allow a TRs-80 to run rings round a CRAY-1.
For more information about data structures, consult Knuth (1968, 1973), Stone
(1972), Horowitz and Sahni (1976), Standish (1980), Aho, Hopcroft, and Ullman
(1983), and Gonnet (1984). AVL trees come from Adel'son-Vel'skii and Landis
(1962) ; they are described in detail in Knuth (1973) ; 2-3 trees come from Aho, Hop-
croft, and Ullman (1974). Graphs and trees are presented from a mathematical stand-
point in Berge (1958, 1970). The heap was introduced as a data structure for sorting in
Williams (1964). The improvements suggested at the end of the sub-section on heaps
are described in Johnson (1975), Fredman and Tarjan (1984), Gonnet and Munro
(1986), and Carlsson (1986, 1987). Carlsson (1986) also describes a data structure,
which he calls the double-ended heap, or deap, that allows finding efficiently the
largest and the smallest elements of a set. For ideas on building heaps faster, consult
McDiarmid and Reed (1987). In this book, we give only some of the possible uses of
disjoint set structures ; for more applications see Hopcroft and Karp (1971) and Aho,
Hopcroft, and Ullman (1974, 1976).
2
Analysing the Efficiency

of Algorithms
2.1 ASYMPTOTIC NOTATION
As we mentioned in Chapter 1, theoretical analyses of the efficiency of algorithms are

carried out to within a multiplicative constant so as to take account of possible varia-
tions introduced by a change of implementation, of programming language, or of com-
puter. To this end, we now introduce formally the asymptotic notation that will be
used throughout the present book.
2.1.1 A Notation for "the order of "
Let IN and IR represent the set of natural numbers (positive or zero) and the set of real
numbers, respectively. We denote the set of strictly positive natural numbers by IN+,
the set of strictly positive real numbers by IR+, and the set of nonnegative real numbers
by 1R* (the latter being a nonstandard notation). The set { true, false 1 of Boolean con-
stants is denoted by 13.
Let f : IN -* JR* be an arbitrary function. We define
O (f (n)) = { t : IN -4 IR* I (3c E IR+) (2 no E IN) (b'n ? no) [t(n) ! cf (n)] 1.
In other words, 0 (f (n)) (read as "the order of f (n)") is the set of all functions t(n)
bounded above by a positive real multiple of f (n), provided that n is sufficiently large
(greater than some threshold no).
For convenience, we allow ourselves to misuse the notation from time to time.
For instance, we say that t(n) is in the order of f (n) even if t(n) is negative or
undefined for some values n < n 0. Similarly, we talk about the order of f (n) even
37
38 Analysing the Efficiency of Algorithms Chap. 2
when f (n) is negative or undefined for a finite number of values of n ; in this case we
must choose no sufficiently large to be sure that such behaviour does not happen for
n ? no. For example, it is allowable to talk about the order of n / log n , even though
this function is not defined when n =0 or n = 1, and it is correct to write
n3-3n2-n -8 E 0(n).
The principle of invariance mentioned in the previous chapter assures us that if
some implementation of a given algorithm never takes more than t(n) seconds to solve
an instance of size n, then any other implementation of the same algorithm takes a
time in the order of t(n) seconds. We say that such an algorithm takes a time in the
order of f (n) for any function f : IN -4 IR* such that t(n) E O (f (n)). In particular,
since t(n) E O (t(n)), it takes a time in the order of t (n) itself. In general, however, we
try to express the order of the algorithm's running time using the simplest possible
function f such that t(n) E O (f (n)).
Problem 2.1.1. Some implementation of a certain algorithm takes a time

bounded above by
t(n) = 3 seconds - l8n milliseconds + 27n2 microseconds

to solve any instance of size n. (Such behaviour is unlikely since t(n) decreases as n
increases for n <- 333.) Find the simplest possible function f : IN -4 lR* such that the
algorithm takes a time in the order of f (n). Prove that t(n) E O (f (n)).
Problem 2.1.2. Which of the following statements are true? Prove your
answers.
i. n2E0(n3)
ii. n3EO(n2)
iii. 2 ' ' Q ( 2 )
iv. (n + I)! E 0 (n!)
v. for any function f : IN -* IR*, f (n) E O (n) = [f (n)]2 E O (n 2)
vi. for any function f : IN -+ IR*, f (n) E O (n) 2f (n) E O (2n)
Problem 2.1.3. Prove that the following definition of 0 (f (n)) is equivalent to

the preceding one, provided that f (n) is strictly positive for all n E IN:
O(f(n))={t:IN -4 ft*1(3cjR)(b'nEIN)[t(n)<_cf(n)]}.
Sec. 2.1 Asymptotic Notation 39
In other words, the threshold no is not necessary in principle, even though it is often
useful in practice.
Problem 2.1.4. Prove that the relation " E 0 " is transitive : if f (n) E 0 (g(n))
and g(n) E 0 (h (n)), then f (n) E 0 (h (n)). Conclude that if g(n) e 0 (h (n)), then
0 (g(n)) c 0 (h (n))
This asymptotic notation provides a way to define a partial order on functions
and consequently on the relative efficiency of different algorithms to solve a given
problem, as suggested by the following exercises.
Problem 2.1.5. For arbitrary functions f and g : IN -> IR*, prove that
i. 0 (f (n)) = 0 (g(n)) if and only if f (n) E 0 (g(n)) and g(n) E O (f (n)), and
ii. 0 (f (n)) C 0 (g(n)) if and only if f (n) E 0 (g(n)) but g(n) e 0 (f (n)).
Problem 2.1.6. Find two functions f and g : IN -* IN+ such that

f (n) 0 0 (g(n)) and g(n) 0 0 (f (n)). Prove your answer.
Problem 2.1.7. For arbitrary functions f and g : IN -4 IR*, prove that

0 (f (n) + g(n)) = 0 (max(f (n), g(n))), where the sum and the maximum are to be
taken pointwise.
The result of the preceding problem is useful for simplifying asymptotic calcula-
tions. For instance,
n3+3n2+n +8 E O(n3+(3n2+n +8))
= 0 (max(n 3, 3n 2 + n + 8)) = 0 (n 3).
The last equality holds despite the fact that max(n 3, 3n2 + n +8)#n 3 when
0 <- n <- 3, because the asymptotic notation only applies when n is sufficiently large.
However, we do have to ensure that f (n) and g(n) only take nonnegative values (pos-
sibly with a finite number of exceptions) to avoid false arguments like the following :
0 (n2)=O(n3+(n2-n3))
= 0 (max(n 3, n 2- n 3)) = O (n 3).
A little manipulation is sufficient, however, to allow us to conclude that
n3-3n2-n -8 E O(n3)
because
n3-3n2-n -8 E 0(n3-3n2-n -8)

= O(Zn3+(Zn3-3n2-n -8))
= 0(max(fn3, n3-3n2-n - 8))
= O(In3)=0 (n3).
Here again, the fact that n 3 - 3n 2 - n - 8 is negative when 0 <- n <- 6 is of no con-
cern. False arguments ofzthe kind illustrated in the following problem are also to be
avoided.
Problem 2.1.8. Find the error in the following argument :

n
i = 1+2+ +n E 0(1+2+ +n) = 0(max(1,2, ... , n)) = O(n).
Problem 2.1.9. The notion of a limit is a powerful and versatile tool for com-
paring functions. Given f and g : IN -* R+, prove that
i. lim f (n)/g(n) E R+ 0 (f (n)) = 0 (g(n)), and
n->-
ii. lim f (n)/g(n) = 0 = O (f (n)) c O (g(n)) = O (g(n) ± f (n)), but

n-i-
iii. it can happen that 0 (f (n)) = 0 (g(n)) although the limit of f (n)/g(n) does not
exist as n tends to infinity, and
iv. it can happen that 0 (f (n)) C 0 (g(n)) when the limit of f (n)lg(n) does not exist
as n tends to infinity and when it is also not true that 0 (g(n)) = 0 (g(n) -f(n)).
De l'Hopital's rule is often useful in order to apply the preceding problem.

Recall that if lim f (n) = lim g(n) = 0, or if both these limits are infinite, then provided
na- n-+ -
that the domains off and g can be extended to some real interval [no, +oo) in such a
way that the corresponding new functions f and g are differentiable on this interval
and also that g'(x), the derivative of g(x), is never zero for x E [no, +oo), then
lira f (n)lg(n) = lim f '(x)lg'(x),
X-->-
provided that this last limit exists.
Problem 2.1.10. Use de l'Hopital's rule and Problems 2.1.5 and 2.1.9 to
prove that log n e 0 but that e O (log n).
Problem 2.1.11. Let a be an arbitrary real constant, 0 < e < 1. Use the rela-
tions " c " and "_ " to put the orders of the following functions into a sequence :
n log n, n s, n 1+e, (1+E)" , (n 2+8n +Iog3n)4, and n 2/ log n.
Prove your answer.
* Problem 2.1.12. Let f and g : IN -- R+ be two increasing functions, and let c

be a strictly positive real constant such that for every integer n
i. 4g(n) g(2n) <_ 8g(n) and
ii. f (2n) 2f (n) + cg(n).
Prove that
iii. f (n) e O (g(n)).
2.1.2 Other Asymptotic Notation
The notation we have just seen is useful for estimating an upper limit on the time that
some algorithm will take on a given instance. It is also sometimes interesting to esti-
mate a lower limit on this time. The following notation is proposed to this end :
Q(f(n))= {t:IN --> IR* I (2ce1R+)(3noEIN)(Vn ?no)[t(n)?cf(n)] }.

In other words, Q(f (n)), which we read unimaginatively as omega of f (n) is the set of
all the functions t(n) bounded below by a real positive multiple of f(n), provided n is
sufficiently large. The following exercise brings out the symmetry between the nota-
tion 0 and Q.
Problem 2.1.13. For arbitrary functions f and g : IN -a IR*, prove that

f (n) E O (g(n)) if and only if g(n) e Q(f (n)).
In a worst-case analysis there is, however, a fundamental asymmetry between the

notation 0 and Q. If an algorithm takes a time in 0 (f (n)) in the worst case, there
exists a real positive constant c such that a time of cf (n) is sufficient for the algorithm
to solve the worst instance of size n, for each sufficiently large n. This time is obvi-
ously also sufficient to solve all the other instances of size n, since they cannot take
more time than the worst case. Assuming only a finite number of instances of each
given size exists, there can thus be only a finite number of instances, all of size less
than the threshold, on which the algorithm takes a time greater than cf (n). These
instances can all be taken care of as suggested by Problem 2.1.3, by using a bigger
constant. On the other hand, if an algorithm takes a time in Q(f (n)) in the worst case,
there exists a real positive constant d such that the algorithm takes a time longer than
df (n) to solve the worst instance of size n, for each sufficiently large n. This in no
way rules out the possibility that a much shorter time might suffice to solve some other
instances of size n. Thus there can exist an infinity of instances for which the algo-
rithm takes a time less than df (n). Insertion sort, which we saw in Section 1.4, pro-
vides a typical example of such behaviour : it takes a time in S2(n 2) in the worst case,
despite the fact that a time in the order of n is sufficient to solve arbitrarily large
instances in which the items are already sorted.
We shall be happiest if, when we analyse the asymptotic behaviour of an algo-
rithm, its execution time is bounded simultaneously both above and below by positive
real multiples (possibly different) of the same function. For this reason we introduce a
final notation
O(f(n)) = O(f(n)) n f (f(n)),
called the exact order of f (n).
Problem 2.1.14. Prove that f (n) E O(g(n)) if and only if
(3c, d e R+)(3n0EIN)(Vn ? n0) [cg(n) 5 f(n) 5 dg(n)].

The following problem shows that the O notation is no more powerful than the 0
notation for comparing the respective orders of two functions.
Problem 2.1.15. For arbitrary functions f and g : IN -4 R*, prove that the fol-
lowing statements are equivalent :
i. 0 (f (n)) = 0 (g(n)),
ii. O(f (n)) = O(g(n)), and
iii. f (n) E O(g(n)).
Problem 2.1.16. Continuing Problem 2.1.9, prove that if f and g : IN -* IR

are two arbitrary functions, then
I. lim f (n)/g(n) a 1R+ = f (n) E O(g(n)),
H. lim f (n)/g(n) = 0 = f (n) E O (g(n)) but f (n) 0 O(g(n)), and

n-4-
iii. lim f (n)/g(n) = +- = f (n) E f (g(n)) but f (n) 0 O(g(n)).

na00
Problem 2.1.17. Prove the following assertions :
i. log,, n (=_O(logb n) whatever the values of a, b > 1 (so that we generally do not
bother to specify the base of a logarithm in an asymptotic expression), but
210g°n
iii. 0 O(21°g'n) if a # b,
n
iii. ik E O(nk+') for any given integer k >_ 0 (this works even for real k > -1;
r=i
the hidden constant in the 0 notation may depend on the value of k),
iv. log(n!) E O(n log n), and
n
v. i -' E O(log n).
0
2.1.3 Asymptotic Notation with Several Parameters
It may happen when we analyse an algorithm that its execution time depends simul-
taneously on more than one parameter of the instance in question. This situation is
typical of certain algorithms for problems involving graphs, where the time depends on
both the number of vertices and the number of edges. In such cases the notion of the
"size of the instance" that we have used so far may lose much of its meaning. For this
reason the asymptotic notation is generalized in a natural way to functions of several
variables.
Let f : INxIN -* 1R* be an arbitrary function. We define
O(f(m,n))= [t:INxJ-4JR' I(3cElR+)(3mo,noE N)
(Vn?n0)(Vm ?mo)[t(m,n)5cf(m,n)l }
Other generalizations are defined similarly.
There is nevertheless an essential difference between an asymptotic notation with
only one parameter and one with several : unlike the result obtained in Problem 2.1.3,
it can happen that the thresholds mo and no are indispensable. This is explained by the
fact that while there are never more than a finite number of values of n >- 0 such that
n >_ no is not true, there are in general an infinite number of pairs < m , n > such that
m >- 0 and n >- 0 yet such that m >- m o and n >- no are not both true.
* Problem 2.1.18. Give an example of a function f : INxIN 1R+ such that
O(f(m,n))# {t:INx1N-4 1R* 1(3cER+)(`dm,nEIN)[t(m,n)<_cf(m,n)l }.
2.1.4 Operations on Asymptotic Notation
To simplify some calculations, we can manipulate the asymptotic notation using arith-
metic operators. For instance, 0 (f (n))+O (g(n)) represents the set of functions
obtained by adding pointwise any function in 0 (f (n)) to any function in 0 (g(n)).

Intuitively this represents the order of the time taken by an algorithm composed of a
first stage taking a time in the order of f (n) followed by a second stage taking a time
in the order of g(n). The hidden constants that multiply f (n) and g(n) may well be dif-
ferent, but this is of no consequence (see Problem 2.1.19(i)).
More formally, if X and Y are sets of functions from IN into IR* and if op is any
binary operator, then "X op Y " denotes
{ t :IN IR* I (3f EX )(3gEY)(3n0EIN)(b'n>no)[t(n) = f(n) op g(n)] }.
If g is a function from IN into IR*, we stretch the notation by writing X op g to denote
X op { g } . Furthermore, if a E IR*, we use X op a to denote X op Ida, where
Ida : N -f IR* is the constant function Ida (n) = a for every integer n. We also use the
symmetrical notation g op X and a op X, and all this theory of operations on sets is
extended in the obvious way to operators other than binary.
This notation occasionally conflicts with the usual mathematical conventions.
For instance, [0 (f (n))]2 does not denote the set of pairs of functions chosen from the
set 0 (f (n)). Similarly, 0 (f (n)) x 0 (g(n)) does not denote the Cartesian product of
O (f (n)) and 0 (g(n)). On the other hand, if N is the set of vertices of a directed
graph, then N x N denotes as usual the set of possible edges between these vertices. In
every case but one the context removes any potential ambiguity. Still, there is one
case that must be treated cautiously. If the symbol "-" were used to denote the differ-
ence of two sets, a genuinely ambiguous situation would arise : what would
O (n 3) - O (n 2) mean, for example ? To avoid this ambiguity, we use "-" only to
denote arithmetic subtraction, including pointwise subtraction of functions, reserving
the symbol "\' to denote set difference : A \ B = { x E A I x e B 1.
Notice the subtle difference between [O(f (n))]2 and O(f (n)) x O(f (n)).
Although they both denote the same set as 0([f this requires a proof (Problem
(n)]2),
2.1.19(ii)). To belong to [O(f (n))]2, a function g(n) must be the pointwise square of
some function in O(f (n)) ; to belong to O(f (n)) x O(f (n)), however, it suffices for
g(n) to be the pointwise product of two possibly different functions, each a member of
O(f (n)). To understand the first notation, think of it as O(f (n)) exp { Id 2 } , where
" exp " denotes the binary exponentiation operator and " Id 2 " is the constant function
1d2(n)=2 for all n . Similarly, n x O(f (n)) denotes
{ t :IN -4 JR I (3g(n)EO(f(n)))(3noEIN)(Vn ?no) [t(n) = n xg(n)] },
which is not at all the same as yj 1 O(f (n)) = O(f (n))+O(f (n)) + +O(f (n)).
Problem 2.1.19. Let f and g be arbitrary functions from IN into IR*. Prove
the following identities :
i. O(f (n))+O(g(n)) = O(f (n) + g(n)) = O(max(f (n), g(n)))

= max(O(f (n)), O(g(n))) ;
H. O([f(n)]2) _ [0(f(n))]2=O(.f(n))xO(f(n));
iii. [I +0(1)]n = 2°(n), but [0(1)]n # 2°(") ; and
iv. f(n) E Hi=o0(1)
Another kind of operation on the asymptotic notation allows it to be nested. Let

X be a set of functions from IN into R*, possibly defined by some asymptotic notation.
Now 0 (X) denotes
LJ0(f(n)) = { t :IN - R* I (3f (=- X)[te0(f(n))] }.
12(X) and O(X) are defined similarly.
Example 2.1.1. Although this expression can be simplified, the natural way to
express the execution time required by Dixon's integer factorization algorithm (Section
8.5.3) is
0(eo(mi ))
where n is the value of the integer to be factorized.
2.1.5 Conditional Asymptotic Notation
Many algorithms are easier to analyse if initially we only consider instances whose size
satisfies a certain condition, such as being a power of 2. Conditional asymptotic nota-
tion handles this situation. Let f : IN -4 R* be any function and let P : IN -* 113 be a
predicate. We define
O(f(n) I P(n)) t :lN -4 R* I (3cER+)(3noeIN)(Vn ? no)

[P(n) = t(n) <_ cf (n)] },
In other words, 0 (f (n) I P(n)), which we read as the order of f (n) when P(n), is the
set of all functions t(n) bounded above by a real positive multiple of f (n) whenever n
is sufficiently large and provided the condition P(n) holds. The notation 0 (f (n))
defined previously is thus equivalent to 0 (f (n) I P(n)) where P(n) is the predicate
whose value is always true. The notation S2(f (n) I P(n)) and O(f (n) I P(n)) is defined
similarly, as is the notation with several parameters.
The principal reason for using this conditional notation is that it can generally be
eliminated once it has been used to make the analysis of an algorithm easier. You
probably used this idea for solving Problem 2.1.12. A function f : IN -* R* is
eventually nondecreasing if (3 n 0 E IN) ('d n > n 0) [f (n) <_ f (n + 1)], which implies by
mathematical induction that (3 n0 E l) (`d n >_ n0) (V m >_ n) [f (n) <_ f (m)]. Let b ? 2
be any integer. Such a function is b-smooth if, as well as being eventually nonde-
creasing, it satisfies the condition f (bn ) E O (f (n)). It turns out that any function that
is b-smooth for some integer b >_ 2 is also c-smooth for every integer c ? 2 (prove
it!); we shall therefore in future simply refer to such functions as being smooth. The
following problem assembles these ideas.
* Problem 2.1.20. Let b ? 2 be any integer, let f : IN -4 IR* be a smooth func-

tion, and let t : IN --> IR* be an eventually nondecreasing function such that
t(n) EX (f (n) I n is a power of b), where X stands for one of 0, S2, or e. Prove that
t(n) E X (f (n)). Furthermore, if 1(n) E O(f (n)), prove that t (n) is also smooth. Give
two specific examples to illustrate that the conditions "t(n) is eventually nonde-
creasing" and "f (bn ) E O (f (n))" are both necessary to obtain these results. 0
We illustrate this principle using an example suggested by the algorithm for
merge sorting given in Section 4.4.
Example 2.1.2. Let t (n) be defined by the following equation :
t(n) _ Ja ifn = 1
t(Ln/2J)+t([n/21)+bn otherwise,
where a and b are arbitrary real positive constants. The presence of floors and ceilings
makes this equation hard to analyse exactly. However, if we only consider the cases
when n is a power of 2, the equation becomes
t (n)
Ja ifn=l
2t (n / 2) + bn if n > I is a power of 2.
The techniques discussed in Section 2.3, in particular Problem 2.3.6, allow us to infer
immediately that t (n) E O(n log n { n is a power of 2). In order to apply the result of
the previous problem to conclude that t(n) E O(n log n), we need only show that t(n) is
an eventually nondecreasing function and that n log n is smooth.
The proof that (Vn >_ 1) [t(n) <_t(n + 1)] is by mathematical induction. First,
note that t (1) = a <_ 2(a +b) = t (2). Let n be greater than 1. By the induction
hypothesis, assume that (V m < n) [t(m) <_ t (m + 1)]. In particular, t (Ln / 2j)
<_ t (L(n + l )/2]) and t ([n / 21) <_ t (((n + 1)12 ). Therefore
t(n)=t(Ln/2j)+t([n/2])+bn St(L(n+1)12j)+t([(n+l)/21)+b(n+l)=t(n+l).
A word of caution is important here. One might be tempted to claim that t(n) is
eventually nondecreasing because such is obviously the case with n log n . This argu-
mentation is irrelevant and fallacious because the relation between t(n) and n log n has
only been demonstrated thus far when n is a power of 2. The proof that t(n) is nonde-
creasing must use its recursive definition.
Finally, n log n is smooth because it is clearly eventually nondecreasing and

2n log(2n) = 2n (log2 + log n) = (21og2)n + 2n log n
E O (n + n log n) = O (max(n, n log n)) = O (n log n).
2.1.6 Asymptotic Recurrences
When analysing algorithms, we do not always find ourselves faced with equations as
precise as those in Example 2.1.2 in the preceding section. More often we have to deal
with inequalities such as
tI(n) ifn :5 no
t(n) <
t (Ln / 2]) + t (Fn / 21) + cvi otherwise
and simultaneously
t2(n) ifn <_n0

t(n)
t (Ln / 2]) + t ([n / 21) + do otherwise,
for some constants c, d E 1R+, n 0 E IN, and for appropriate initial functions
t 1, t2: IN -f IR+. Our asymptotic notation allows these constraints to be expressed
succinctly as
t(n) E t(Ln/2])+t([n121)+0(n).
To solve such inequalities, it is convenient to convert them first to equalities. To
this end, define f : IN -* IR by
1 ifn = 1
f(n)= f (Ln / 2]) + f ([n / 21) + n otherwise.
We saw in the previous section that f (n) E O(n log n).
Coming back now to the function t(n) satisfying the preceding inequalities, let
u = max(c, max I t I (n) I n Sn0}) and v =min(d,min {t7(n)/f(n) I n Sn0}). It is easy
to prove by mathematical induction that v <_ t(n)/f(n) <_ u for every integer n. We
immediately conclude that t(n) E O(f (n)) = O(n log n).
This change from the original inequalities to a parametrized equation is useful
from two points of view. Obviously it saves having to prove independently both
t (n) c O (n log n) and t (n) E Q (n log n). More importantly, however, it allows us to
confine our analysis in the initial stages to the easier case where n is a power of 2.
It is then possible, using the conditional asymptotic notation and the technique
explained in Problem 2.1.20, to generalize our results automatically to the case where n
is an arbitrary integer. This could not have been done directly with the original in-
equalities, since they do not allow us to conclude that t(n) is eventually nondecreasing,
which in turn prevents us from applying Problem 2.1.20.
2.1.7 Constructive Induction

Mathematical induction is used primarily as a proof technique. Too often it is
employed to prove assertions that seem to have been produced from nowhere like a
rabbit out of a hat. While the truth of these assertions is thus established, their origin
remains mysterious. However, mathematical induction is a tool sufficiently powerful
to allow us to discover not merely the proof of a theorem, but also its exact statement.
By applying this technique, we can simultaneously prove the truth of a partially
specified assertion and discover the missing specifications thanks to which the assertion
is correct. As we shall see in Examples 2.1.4 and 2.2.5, this technique of constructive
induction is especially useful for solving certain recurrences that occur in the context
of the analysis of algorithms. We begin with a simple example.
Example 2.1.3. Let the function f : IN -> IN be defined by the following

recurrence :
0 ifn =0
.f (n) =
n + f (n - 1) otherwise.
It is clear that f (n) = Yi"_o i. Pretend for a moment that you do not know that
f (n) = n (n + 1)/ 2, but that you are looking for some such formula. Obviously
f (n) = YJ%o i <_ J70n = n 2, and so f (n) E O (n 2 ). This suggests that we formulate a
hypothesis that f (n) might be a quadratic polynomial. We therefore try the partially
specified induction hypothesis HI (n) according to which f (n) = an 2 + bn + c. This
hypothesis is partial in the sense that a, b, and c are not yet known. The technique of
constructive induction consists of trying to prove this incomplete hypothesis by
mathematical induction. Along the way we hope to pick up sufficient information
about the constants to determine their values.
Supposing that HI (n - 1) is true for some n ? 1, we know that
f(n)=n +f(n-1)=n +a(n-1)2+b(n-1)+c =an2+(1+b-2a)n +(a-b+c).
If we are to conclude HI (n), it must be the case that f (n) = an 2 + bn + c. By
equating the coefficients of each power of n, we obtain two nontrivial equations for
our three unknowns : 1 + b - 2a = b and a - b + c = c. From these it follows that
a = b = '-f, the value of c being as yet unconstrained. We now have therefore a new,
more complete hypothesis, which we continue to call HI (n) : f (n) = n 2/ 2 + n / 2 + c.
We have just shown that if HI (n -1) is true for some n ? 1, then so is HI (n).
It remains to establish the truth of HI (0) in order to conclude by mathematical induc-
tion that HI (n) is true for every integer n. Now HI (0) says precisely that
f (O) = a 02 + b 0 + c = c. Knowing that f (O) = 0, we conclude that c = 0 and that
f (n) = n 2/ 2 + n / 2 is true for every integer n .
In the case of the preceding example it would have been simpler to determine the
values of a, b, and c by constructing three linear equations using the values of HI (n)
forn =0, 1, and 2.
c=0
a+b+c=1
4a +2b +c=3
Solving this system gives us immediately a = 2, b = Z, and c = 0. However, using
this approach does not prove that f (n) = n 2/ 2 + n / 2, since nothing allows us to assert
a priori that f (n) is in fact given by a quadratic polynomial. Thus once the constants
are determined we must in any case follow this with a proof by mathematical induc-
tion.
Some recurrences are more difficult to solve than the one given in Example
2.1.3. Even the techniques we shall see in Section 2.3 will prove insufficient on occa-
sion. However, in the context of asymptotic notation, an exact solution of the
recurrence equations is generally unnecessary, since we are only interested in estab-
lishing an upper bound on the quantity of interest. In this setting constructive induc-
tion can be exploited to the hilt.
Example 2.1.4. Let the function t : IN+ -> IR+ be given by the recurrence
(a ifn=1
t(n) _
jl bn z + nt(n - 1) otherwise,
where a and b are arbitrary real positive constants. Although this equation is not easy
to solve exactly, it is sufficiently similar to the recurrence that characterizes the fac-
torial (n! = n x(n -1)!) that it is natural to conjecture that t (n) E O(n!). To establish
this, we shall prove independently that t (n) E O (n!) and that t(n) E S2(n!). The tech-
nique of constructive induction is useful in both cases. For simplicity, we begin by
proving that t(n)ES2(n!), that is, there exists a real positive constant u such that
t (n) >- un! for every positive integer n. Suppose by the partially specified induction
hypothesis that t(n -1) >- u (n -1)! for some n > 1. By definition of t(n), we know
that t(n) = bn2 + nt(n -1) >- bn2 + nu (n -1)! = bn2 + un! >- W. Thus we see that
t(n) ? un! is always true, regardless of the value of u, provided that
t(n - 1) ? u (n - 1)!. In order to conclude that t(n) >- un! for every positive integer n,
it suffices to show that this is true for n = 1; that is, t (l) >- u. Since t (l) = a, this is
the same as saying that u <- a. Taking u = a, we have established that t(n) ? an! for
every positive integer n , and thus that t (n) E U(M).
Encouraged by this success, we now try to show that t (n) E O (n!) by proving the
existence of a real positive constant v such that t(n) 5 vn! for every positive integer n.
Suppose by the partially specified induction hypothesis that t(n -1) <- v(n - 1)! for
some n > 1. As usual, this allows us to affirm that t(n) = bn2 + nt(n -1) <- bn2 + vn!.
However our aim is to show that t(n) :- vn!. Unfortunately no positive value of v
allows us to conclude that t(n) <- vn! given that t(n) :- bn2 + vn!. It seems then that
constructive induction has nothing to offer in this context, or perhaps even that the
hypothesis to the effect that t(n) EO (n!) is false.
In fact, it is possible to obtain the result we hoped for. Rather than trying to
prove directly that t(n) <- vn!, we use constructive induction to determine real positive
constants v and w such that t(n) <- vn! - wn for any positive integer n. This idea may
seem odd, since t(n) <- vn! - wn is a stronger statement than t(n) <- vn!, which we
were unable to prove. We may hope for success, however, on the grounds that if the
statement to be proved is stronger, then so too is the induction hypothesis it allows us
to use.
Suppose then by the partially specified induction hypothesis that
t(n -1) < v (n -1)! - w (n -1) for some n > 1. Using the definition of t(n), we con-
clude that
t (n) = bn 2 + nt (n -1) <- bn 2 + n (v (n -1)! - w (n -1)) = vn! + ((b -w )n + w )n.
To conclude that t(n) <- vn! - wn, it is necessary and sufficient that
(b -w )n + w <- -w. This inequality holds if and only if n >- 3 and w > bn /(n -2).
Since n /(n -2):5 3 for every n >- 3, we may in particular choose w = 3b to ensure that
t(n) <- vn! - wn is a consequence of the hypothesis t(n -1) <- v (n - 1)! - w (n -1),
independently of the value of v, provided that n >- 3.
All that remains is to adjust the constant v to take care of the cases n <- 2. When
n = 1, we know that t(l) = a. If we are to conclude that t(n) <- v - 3b, it is necessary
and sufficient that v >- a + 3b. When n = 2, we can apply the recurrence definition of
t(n) to find t(2) = 4b + 2t(1) = 4b + 2a . If we are to conclude that t(n) <- 2v - 6b , it
is necessary and sufficient that v >- a + 5b, which is stronger than the previous condi-
tion. In particular, we may choose v = a + 5b.
The conclusion from all this is that t(n) E 9(n!) since
an! S t(n) S (a + 5b)n! - 3bn
for every positive integer n. If you got lost in the preceding argument, you may wish
to prove this assertion, which is now completely specified, by straightforward
mathematical induction. 11
The following problem is not so easy ; it illustrates well the advantage obtained
by using constructive induction, thanks to which we were able to prove that
t(n) E O(n!) without ever finding an exact expression for t(n).
* Problem 2.1.21. To complete Example 2.1.4, solve exactly the recurrence

defining t(n). Determine in terms of a and b the real positive constant c such that
t(n)
lim = c.
n-*= n!
Verify that a <- c <- a + 5b.
Problem 2.1.22. Let k E IN and a, b E IR+ be arbitrary constants. Let
g : IN+ IR+ be the function defined by the recurrence
a ifn=1
g(n) =
bnk + ng (n - 1) otherwise.
Prove that g(n) e O(n!). 1
2.1.8 For Further Reading
The notation used in this chapter is not universally accepted. You may encounter three
major differences in other books. The most striking is the widespread use of state-
ments such as n 2 = O (n 3) where we would write n 2 E O (n 3). Use of such "one-way
equalities" (for one would not write O (n3) = n2) is hard to defend except on historical
grounds. With this definition we say that the execution time of some algorithm is of
the order of f (n) (or is 0 (f (n))) rather than saying it is in the order of f (n).
The second difference is less striking but more important, since it can lead an
incautious reader astray. Some authors define
Q(f(n)) = { t:IN -> IR* I (3ce1R+)(VnoeIN)(3n ?no) [t(n)?cf(n)] I.

Notice the quantifier reversal. With this definition it suffices that there exist an infinite
number of instances x that force some algorithm to take at least cf ( I x I) steps in order
to conclude that this algorithm takes a time in Q (f (n)). This corresponds more closely
to our intuitive idea of what a lower bound on the performance of an algorithm should
look like. Furthermore, with this definition, the asymmetry between 0 and Q noted
after Problem 2.1.13 is neatly avoided. Unfortunately, the notation becomes difficult to
handle, in particular because Q thus defined is not transitive, and because it makes O
asymmetric.
The third difference concerns the definition of 0. We often find
0(.f(n)) _ { t :IN - IR 1 (3cclR+)(2noEIN)(Vn ? n0)[It(n)I <_ cf(n)] 1
where It(n)I denotes (here only) the absolute value of t(n). Using this definition, one
would write n 3 - n 2 E n 3 + O (n 2). Of course, the meaning of "such-and-such an algo-
rithm takes a time in 0 (n 2)" does not change since algorithms cannot take negative
time. On the other hand, a statement such as O(f (n)) + O(g(n)) = O(max(f (n), g(n)))
is no longer true.
Problem 2.1.23. Why ? 11
When we want the equivalent of this definition of 0 (f (n)) we write ±O (f (n)).

2.2 ANALYSIS OF ALGORITHMS
There is no magic formula for analysing the efficiency of an algorithm. It is largely a

question of judgement, intuition, and experience. Often a first analysis gives rise to a
complicated-looking function, involving summations or recurrences. The next step is
to simplify this function using the asymptotic notation as well as the techniques
explained in Section 2.3. Here are some examples.
Example 2.2.1. Selection sort. Consider the selection sorting algorithm given
in Section 1.4. Most of the execution time is spent carrying out the instructions in the
inner loop, including the implicit control statements for this loop. The time taken by
each trip round the inner loop can be bounded above by a constant a. The complete
execution of the inner loop for a given value of i therefore takes at most a time
b + a (n -i ), where b is a second constant introduced to take account of the time spent
initializing the loop. One trip round the outer loop is therefore bounded above by
c +b + a (n-i ), where c is a third constant, and finally, the complete algorithm takes a
time not greater than d + 7-11 [c + b + a (n -i )], for a fourth constant d. We can
simplify this expression to 2 n 2 + (b +c -a /2)n + (d -c -b ), from which we conclude
that the algorithm takes a time in 0(n2). A similar analysis for the lower bound
shows that in fact it takes a time in O(n 2). 0
In this first example we gave all the details of our argument. Details like the ini-
tialization of the loops are rarely considered explicitly. It is often sufficient to choose
some instruction in the algorithm as a barometer and to count how many times this
instruction is executed. This figure gives us the exact order of the execution time of
the complete algorithm, provided that the time taken to execute the chosen instruction
can itself be bounded above by a constant. In the selection sort example, one possible
barometer is the test in the inner loop, which is executed exactly n (n - 1)/ 2 times
when n items are sorted. The following example shows, however, that such
simplifications should not be made incautiously.
Example 2.2.2. Choosing a barometer. When an algorithm includes several

nested loops, as is the case with selection sort, any instruction of the inner loop can
usually be used as a barometer. However, there are cases where it is necessary to take
account of the implicit control of the loops. Consider the following algorithm (which
is reminiscent of the countsort algorithm discussed in Section 10.1):
k E- 0
for i - 1 to n do
for j f- 1 to T [i ] do
k +-k +T[j] ,
where T is an array of n integers such that 0 <- T[i] <- i for every i <- n. Let s be the
sum of the elements of T. How much time does the algorithm take?
Sec. 2.2 Analysis of Algorithms 53
For each value of i the instruction " k - k + T [ j ] " is executed T [i ] times.

The total number of times it is executed is therefore E° T [i] = s times. If indeed
this instruction could serve as a barometer, we would conclude that the algorithm takes
a time in the exact order of s. A simple example is sufficient to convince us that this
is not the case. Suppose that T [i] = 1 whenever i is a perfect square and that T [i] = 0
otherwise. In this case s = [I J. However, the algorithm clearly takes a time in 11(n)
since each element of T is considered at least once. The problem arises because we
can only neglect the time spent initializing and controlling the loops provided we
include something each time the loop is executed.
The detailed analysis of this algorithm is as follows. Let a be the time taken by
one trip round the inner loop, including the time spent on loop control. To execute the
inner loop completely for a given value of i therefore takes a time b + aT [i], where the
constant b represents the initialization time for the loop. This time is not zero when
T[i]=O. Next, the time taken to execute one trip round the outer loop is
c + b + aT [i ], where c is a new constant. Finally, the complete algorithm takes a time
d + J:i" 1(c + b + aT [i]), for yet another constant d. When simplified, this expression
yields (c +b )n + as + d . The time therefore depends on two independent parameters n
and s and cannot be expressed as a function of just one of these variables. This situa-
tion is typical of algorithms for handling graphs, where the execution time often
depends on both the number of vertices and the number of edges. Coming back to our
algorithm, Problem 2.1.7 says that we can express its execution time in asymptotic
notation in two ways : O(n +s) or O(max(n , s )).
With a little experience the same algorithm can be analysed more succinctly.
The instruction in the inner loop is executed exactly s times. To this we must add n to
take account of the control of the outer loop and of the fact that the inner loop is ini-
tialized n times. The total time taken by the algorithm is therefore in O(n+s).
We now give an example of analysis of the average behaviour of an algorithm.
As Section 1.4 points out, analyses of average behaviour are usually harder than ana-
lyses of the worst case, and they presuppose that we know a priori the probability dis-
tribution of the instances to be solved.
Example 2.2.3. Insertion sort. Consider the insertion sorting algorithm

given in Section 1.4. The time that this algorithm takes to sort n elements depends on
their original order. We use the comparison "x < T [ j ] " as our barometer. The
number of comparisons carried out between elements is a good measure of the com-
plexity of most sorting algorithms, as we shall see in Chapter 10. (Here we do not
count the implicit comparisons involved in control of the for loop, nor the comparison
j>0".)
Suppose for a moment that i is fixed. Let x = T[i], as in the algorithm. The
worst case arises when x is less than T [ j ] for every j between 1 and i - 1, since in this
case we have to compare x to T [i - 1 ] , T [i -2] , ... , T [ 1 ] before we leave the while
loop because j = 0. Thus the algorithm makes i -1 comparisons. This can happen for
every value of i from 2 to n when the array is initially sorted in descending order. The
total number of comparisons is therefore It 2(i - 1) = n (n - 1)/ 2 E O(n 2). In the
worst case insertion sorting thus takes a time in O(n2). Notice that selection sorting
systematically makes the same number of comparisons between elements that insertion
sorting makes in the worst case.
To determine the time needed by the insertion sorting algorithm on the average,
suppose that the n items to be sorted are all distinct and that each permutation of these
items has the same probability of occurrence. If i and k are such that 1 <- k <- i, the
probability that T[i] is the kth largest element among T [1], T [2], ... , T [i] is 1/i
because this happens for [n] (i -1)! (n -i )! = n!/i of the n! possible permutations of n
elements. For a given value of i, T[i] can therefore be situated with equal probability
in any position relative to the items T [I], T [2], ... , T [i -1]. With probability 1/i ,
T[i] < T [i -1 ] is false at the outset, and the first comparison x < T [j ] gets us out of
the while loop. The same probability applies to any given number of comparisons up
to i-2 included. On the other hand, the probability is 2/i that i - 1 comparisons will
be carried out, since this happens both when x < T [ 1 ] and when T [ 1 ] S x < T [2].
The average number of comparisons made for a given value of i is therefore
i-2
ci = - 2(i-1)+ Y, k
k=I
_ (i -1)(i+2) _ i + I _ I
2i 2 i
These events are independent for different values of i. The average number of com-
parisons made by the algorithm when sorting n items is therefore
"
i=2
Eci=z
i=2
" i+1
2
-- 1
- n2+3n -H"
4
E 0(n2).
Here H" = E i "_ I i- 1 , the n th term of the harmonic series, is negligible compared to the
dominant term n 2/4 because H" E O(log n), as shown in Problem 2.1.17.
The insertion sorting algorithm makes on the average about half the number of
comparisons that it makes in the worst case, but this number is still in O(n2).
Although the algorithm takes a time in c2(n2) both on the average and in the worst
case, a time in 0 (n) is sufficient for an infinite number of instances.
When we analyse an algorithm, we often have to evaluate the sum of arithmetic,

geometric, and other series.
Problem 2.2.1. Prove that for any positive integers n and d

d
k=o
Rather than simply proving this formula by mathematical induction, try to see how you
might have discovered it for yourself.
Example 2.2.4. Making a heap. Consider the "make-heap " algorithm given
at the end of Section 1.9.4: this algorithm constructs a heap starting from an array T
of n items. As a barometer we use the instructions in the repeat loop of the algorithm
used to sift down a node. Let m be the largest number of trips round the loop that can
be caused by calling sift-down (T, n, i). Denote by j, the value of j after execution of
the assignment "j - k " on the t th trip round the loop. Obviously j, = i. Moreover,
if 1 < t <- m, then at the end of the (t -1)st trip round the loop we had j # k ; therefore
k >- 2j. This shows that j, ? 2jr_1 for 1 < t <- m. But it is impossible for k (and thus
j ) to exceed n. Consequently
n >_Jm >2Jm-I?4Jm-2? ... >2m-li

Thus 2m n / i, which implies that m <- 1 + lg (n / i ).
The total number of trips round the repeat loop when constructing a heap can
now be bounded above by
Ln/2]
(I +lg(n/i)). (*)
To simplify this expression, notice first that for any k ? 0

2--,
X lg(n/i)<-2klg(n/2k).
% =GA
The interesting part of the sum (*) can therefore be decomposed into sections
corresponding to powers of 2. Let d = Llg (n / 2)].
Ln/2i d
E lg(n/i) <- E2 k lg(n/2k) < 2dllg(n/2d 1)
i=1 k=O
(by Problem 2.2.1). But d = Llg (n / 2)] implies that d + 1 <- lg n and d - I > lg (n / 8).
Hence
Ln/2J
lg(n/i)<-3n
From (*) we thus conclude that Ln / 2j + 3n trips round the repeat loop are enough to
construct a heap, so that this can be done in a time in 0 (n). Since any algorithm for
constructing a heap must look at each element of the array at least once, we obtain our
final result that the construction of a heap of size n can be carried out in a time in
O(n).
A different approach yields the same result. Let t (k) stand for the time needed to
build a heap of height at most k in the worst case. Assume k >- 2. In order to con-
struct the heap, the algorithm first transforms each of the two subtrees attached to the
root into heaps of height at most k - 1 (the right hand subtree could be of height k - 2).
The algorithm then sifts the root down a path whose length is at most k, which takes a
time in the order of k in the worst case. We thus obtain the asymptotic recurrence
t (k) E 2t (k -1) +O (k). The techniques of Section 2.3, in particular Example 2.3.5,
can be used to conclude that t (k) E O (2k). But a heap containing n elements is of
height Llg nj, hence it can be built in at most t (Llg n J) steps, which is in O(n) since
2UgnJ <- n.
Problem 2.2.2. In Section 1.9.4 we saw another algorithm for making a heap
(slow-make-heap ). Analyse the worst case for this algorithm and compare it to the
algorithm analysed in Example 2.2.4.
Problem 2.2.3. Analysis of heapsort. Williams invented the heap to serve as

the underlying data structure for the following sorting algorithm.
procedure heapsort (T [ 1 .. n ])
IT is an array to be sorted)
make-heap (T)
for i F n step -1 to 2 do
exchange T [1] and T [i ]
sift-down (T [1.. i -1], 1)
{ T is sorted)
What is the order of the execution time required by this algorithm in the worst case ?
* Problem 2.2.4. Find the exact order of the execution time for Williams's
heapsort, both in the worst case and on the average. For a given number of elements,
what are the best and the worst ways to arrange the elements initially insofar as the
execution time of the algorithm is concerned?
Example 2.2.5. Recursive calculation of determinants. We now analyse the

algorithm derived from the recursive definition of a determinant (Section 1.7.3). In our
analysis, additions and multiplications are considered to be elementary operations. We
therefore ignore for the time being the problems posed by the fact that the size of the
operands can become very large during the execution of the algorithm.
Let t(n) be the time taken by some implementation of this algorithm working on
an n x n matrix. When n is greater than 1, most of the work done by the algorithm
consists of calling itself recursively n times to work on (n -1) x (n -1) matrices.
Besides this, the matrices for the recursive calls have to be set up and some other
housekeeping done, which takes a time in O(n2) for each of the n recursive calls if we
do this without thinking too much about it (but see Problem 2.2.5). This gives us the
following asymptotic recurrence : t (n) E nt(n - 1) + O(n 3 ). By Problem 2.1.22 the
algorithm therefore takes a time in O(n!) to calculate the determinant of an n x n
matrix.
Problem 2.2.5. Example 2.2.5 supposes that the time needed to compute a
determinant, excluding the time taken by the recursive calls, is in O(n3). Show that
this time can be reduced to O(n). By Problem 2.1.22, however, this does not affect the
fact that the complete algorithm takes a time in O(n!).
* Problem 2.2.6. Analyse the algorithm again, taking account this time of the
fact that the operands may become very large during execution of the algorithm.
Assume that you know how to add two integers of size n in a time in O(n) and that
you can multiply an integer of size m by an integer of size n in a time in O(mn).
Example 2.2.6. Analysis of Euclid's algorithm. Recall that Euclid's algo-

rithm calculates the greatest common divisor of two integers (Section 1.7.4).
function Euclid (m, n)
while m > 0 do
t - n mod m
n <- m
m -t
return n
We first show that for any two integers m and n such that n >_ in, it is always true that
n modm <n12.
If m > n / 2, then 1 <_ n / m < 2, and so Ln / m] = 1, which means that
n modm =n -m < n -n/2=n12.
Ifm<_n/2,then (n mod m)<m <-n/2.
Let k be the number of trips round the loop made by the algorithm working on the
instance < m, n >. For each integer i <- k, let ni and mi be the values of n and m at
the end of the i th trip round the loop. In particular, Mk = 0 causes the algorithm to ter-
minate and mi > 1 for every i < k. The values of mi and ni are defined by the fol-
lowing equations for 1 <- i <- k, where mo and no are the initial values of m and n :
ni =mi_i
mi =ni_i modmi_1.
Clearly n, > m; for each i 1. Using the preceding observation, we have

mi =n,_1 mod m,_1 <n1_1/2=m1_2/2
for every i >- 2. Suppose for the moment that k is odd. Define d by k = 2d + 1. Then
mk_1 < mk_3/2 < Mk-5/4 < . . . < mo/2d.
But recall that Mk-1 ? 1, and so m o >- 2d. Finally, k = 2d + 1 <- 1 + 21g m o . The case
when k is even is handled similarly, remembering that m 1 = no mod m o < m 0.
In conclusion, the number of trips round the loop made by Euclid's algorithm
working on the integers in and n, and therefore the time required by the algorithm, are
in 0 (log m), provided the instructions in the loop can be considered as elementary.
* Problem 2.2.7. Prove that the worst case for Euclid's algorithm arises when
we calculate the greatest common divisor of two consecutive numbers from the
Fibonaccisequence.
Example 2.2.7. Analysis of the algorithm fibl. We now analyse the algo-
rithm fib l of Section 1.7.5, still not taking account of the large size of the operands
involved. Let t(n) be the time taken by some implementation of this algorithm
working on the integer n. We give without explanation the corresponding asymptotic
recurrence: t(n) E t(n-1)+t(n-2)+O(1).
Once again, the recurrence looks so like the one used to define the Fibonacci
sequence that it is tempting to suppose that t(n) must be in 0(f ). However, as in
Example 2.1.4, constructive induction cannot be used directly to find a constant d such
that t(n) <- d,, . On the other hand, it is easy to use this technique to find three
real positive constants a, b, and c such that a f, <- t (n) <- bf - c for any positive
integer n. The algorithm fib 1 therefore takes a time in 0(4") to calculate the
n th term of the Fibonacci sequence, where 0 = (1+' )/ 2.
Problem 2.2.8. Using constructive induction, prove that af <- t(n) <- b,, - c
for appropriate constants a, b, and c, and give values for these constants.
Problem 2.2.9. Prove that the algorithm fibl takes a time in even if we
take into account that we need a time in O(n) to add two integers of size n. (Since the
value of f is in O(4"), its size is in 0(n lg o) = E) (n).)
Example 2.2.8. Analysis of the algorithm fib2. It is clear that the algorithm
fib2 takes a time equal to a + bn on any instance n , for appropriate constants a and b.
This time is therefore in 0(n).
What happens, however, if we take account of the size of the operands involved?
Let a be a constant such that the time to add two numbers of size n is bounded above
by an, and let b be a constant such that the size of fn is bounded above by bn for
every integer n ? 2. Notice first that the values of i and j at the beginning of the k th
trip round the for loop are respectively fk_2 and fk- 1 (where we take f _1 = 1). The
k th trip round the loop therefore consists of calculating fk_2 + fk _ 1 and fk - fk-2 ,
which takes a time bounded above by ab (2k - 1) for k >- 3, plus some constant time c
to carry out the assignments and the loop control. For each of the first two trips round
the loop, the time is bounded above by c +2a. Let d be an appropriate constant to
account for necessary initializations. Then the time taken by ffb2 on an integer n > 2
is bounded above by
n
d + 2(c + 2a) + Y, ab (2k -1) = abn 2 + (d + 2c +4a - 4ab) ,
k=3
which is in 0 (n 2). It is easy to see by symmetry that the algorithm takes a time in
0(n 2).
Example 2.2.9. Analysis of the algorithm fib3. The analysis of fib3 is rela-
tively easy if we do not take account of the size of the operands. To see this, take the
instructions in the while loop as our barometer. To evaluate the number of trips round
the loop, let n, be the value of n at the end of the t th trip ; in particular n 1 = Ln / 2j. It
is obvious that nr = Lnr_1/ 2]<- n,-1/2 for every 2 <- t <_ m. Consequently
nt 5 nr-1/2 5 nr-2/4 <_ ... < n 1/ 2t-1 <- n12' .
Let m = I + LlgnJ. The preceding equation shows that nn 5 n/2m < 1. But nn is a
nonnegative integer, and so nn = 0, which is the condition for ending the loop. We
conclude that the loop is executed at most m times, which implies that the algorithm
fib3 takes a time in 0 (log n).
Problem 2.2.10. Prove that the execution time of the algorithm fib3 on an
integer n is in @(log n) if no account is taken of the size of the operands.
** Problem 2.2.11. Determine the exact order of the execution time of the algo-
rithm fib3 used on an integer n. Assume that addition of two integers of size n takes a
time in O(n) and that multiplication of an integer of size n by an integer of size m
takes a time in O(mn). Compare your result to that obtained in Example 2.2.8.
If you find the result disappointing, look back at the table at the end of Section 1.7.5
and remember that the hidden constants can have practical importance! In Section 4.7
we shall see a multiplication algorithm that can be used to improve the performance of
the algorithm fib3 (Problem 4.7.5), but not, of course, that of fib2 (why not?).
Problem 2.2.12. Consider the following algorithm :
for i F-0ton do
j F- i
while j# 0 do j E- j div 2
Supposing that integer division by 2, assignments, and loop control can all be carried
out at unit cost, it is clear that this algorithm takes a time in U(n) n 0 (n log n). Find
the exact order of its execution time. Prove your answer.
* Problem 2.2.13. Answer the same question as in the preceding problem, this
time for the algorithm
for i -0ton do
j F- i
while j is odd do j F- j div 2
Show a relationship between this algorithm and the act of counting from 0 to n + 1 in
binary.
Example 2.2.10. Analysis of disjoint set structures. It can happen that the
analysis of an algorithm is facilitated by the addition of extra instructions and counters
that have nothing to do with the execution of the algorithm proper. For instance, this
is so when we look at the algorithms find3 and merge3 used to handle the disjoint set
structures introduced in Section 1.9.5. The analysis of these algorithms is the most
complicated case we shall see in this book. We begin by introducing a counter called
global and a new array cost [ 1 .. N J. Their purpose will be explained later. The array
set [1 .. N ] keeps the meaning given to it in algorithms find3 and merge3: set [i ] gives
the parent of node i in its tree, except when set [i ] = i, which indicates that i is the root
of its tree. The array rank [ 1 .. N ] plays the role of height [ 1 .. N ] in algorithm merge3 :
rank [i] denotes the rank of node i (see Section 1.9.5). We also introduce a strictly
increasing function F : IN -* IN (specified later) and its "inverse" G : IN -* N defined
by G (n) = min { m E IN I F (m) >_ n ). Finally, define the group of an element of rank r
as G (r). The algorithms become
procedure init
(initializes the trees)
global <- 0
fori F- 1 to N do set [i] F- i
rank [i] 0
cost [i ] - 0
function find (x)

{ finds the label of the set containing object x I
r -x
while set [r r do r F- set [r]
{ r is the root of the tree)
i E- x
while i * r do
if G (rank [i]) < G (rank [set [i]]) or r = set [i]
then global E- global + 1
else cost [i] F- cost [i ] + 1
j - set [i ]
set [i] F- r
i-j
return r
procedure merge (a, b)
{ merges the sets labelled a and b
we suppose that a b [
if rank [a] = rank [b]
then
rank [a] t- rank [a] + 1
set[b] f- a
else
if rank [a] > rank [b]
then set [b] F a
else set (a] b
With these modifications the time taken by a call on the procedure find can be
reckoned to be in the order of I plus the increase of global + EN cost [i] occasioned
by the call. The time required for a call on the procedure merge can be bounded
above by a constant. Therefore the total time required to execute an arbitrary sequence
of n calls on find and merge, including initialization, is in
N
O (N + n + global + cost [i ]),
i=1
where global and cost [i] refer to the final values of these variables after execution of
the sequence. In order to obtain an upper bound on these values, the following
remarks are relevant :
1. once an element ceases to be the root of a tree, it never becomes a root thereafter
and its rank no longer changes ;
2. the rank of a node that is not a root is always strictly less than the rank of its
parent ;
3. the rank of an element never exceeds the logarithm (to the base 2) of the number
of elements in the corresponding tree ;
4. at every moment and for every value of k, there are not more than N i 2k ele-
ments of rank k ; and
5. at no time does the rank of an element exceed L1gNJ, nor does its group ever
exceed G ( Llg NJ).
Remarks (1) and (2) are obvious if one simply looks at the algorithms. Remark
(3) has a simple proof by mathematical induction, which we leave to the reader.
Remark (5) derives directly from remark (4). To prove the latter, define subk (i) for
each element i and rank k : if node i never attains rank k, subk (i) is the empty set ; oth-
erwise sub, (i) is the set of nodes that are in the tree whose root is i at that precise
moment when the rank of i becomes k. (Note that i is necessarily a root at that
moment, by remark (1).) By remark (3), subk (i) # 0 # subk (i) ? 2k . By remark
(2), i # j subk (i) n subk (j) = 0. Hence, if there were more than N/ 2k elements i
such that subk (i) 0, there would have to be more than N elements in all, which
proves remark (4).
The fact that G is nondecreasing allows us to conclude, using remarks (2) and
(5), that the increase in the value of global caused by a call on the procedure find
cannot exceed 1 + G ( LlgNJ). Consequently, after the execution of a sequence of n
operations, the final value of this variable is in 0 (1 + nG (L1g NJ)). It only remains to
find an upper bound on the final value of cost [i] for each element i in terms of its final
rank.
Note first that cost [i ] remains at zero while i is a root. What is more, the value
of cost [i ] only increases when a path compression causes the parent of node i to be
changed. In this case the rank of the new parent is necessarily greater than the rank of
the old parent by remark (2). But the increase in cost [i ] stops as soon as i becomes
the child of a node whose group is greater than its own. Let r be the rank of i at the
instant when i stops being a root, should this occur. By remark (1) this rank does not
change subsequently. Using all the preceding observations, we see that cost [i ] cannot
increase more than F (G (r)) - F (G (r) -1) -1 times. We conclude from this that the
final value of cost [i] is less than F (G (r)) for every node i (=-final (r), where final (r)
denotes the set of elements that cease to be a root when they have rank r >- 1 (while,
on the other hand, cost [i] remains at zero for those elements that never cease to be a
root or that do so when they have rank zero). Let K = G (Llg NJ) - 1. The rest is
merely manipulation.
N K F(g+l)
cost [i ] = Y E cost [i ]
i=l g=0 r=F(g)+1 iefina((r)
F(g+l)
<I
K
I I F (G (r))
g=0 r=F(g)+I iefinal(r)
K F(g+l)
<_ 1 E (N12`)F(g+1)
g=O r=F(g)+l
K
NEF(g+1)/2F(g)
g=O
It suffices therefore to put F (g + 1) = 2F(g) to balance global and 1"' cost [i] and so
to obtain E 7l cost [i] NG (Llg NJ). The time taken by the sequence of n calls on
find and merge with a universe of N elements, including the initialization time, is
therefore in
N
O(N +n +global + cost[i]) c O(N +n +nG(L1gNJ)+NG(L1gNJ))
=1
= O(max(N,n)(l +G(L1gN]))).
Now that we have decided that F (g + 1) = 2F(g), with the initial condition
F (0) = 0, what can we say about the function G ? This function, which is often
denoted by lg*, can be defined by
G(N)=lg*N =min{k I lglg ...lgN <0)
k times
The function lg* increases very slowly : lg*N <- 5 for every N 5 65,536 and
lg*N <- 6 for every N < 265,536 Notice also that lg*N - lg*(L1gN j) 2, so that
lg*(Llg NJ) E O(lg*N ). The algorithms that we have just analysed can therefore exe-
cute a sequence of n calls on find and merge with a universe of N elements in a time
in 0 (n lg*N ), provided n >_ N, which is to most intents and purposes linear.
This bound can be improved by refining the argument in a way too complex to
give here. We content ourselves with mentioning that the exact analysis involves the
use of Ackermann's function (Problem 5.8.7) and that the time taken by the algorithm
is not linear in the worst case.
4- II-1 ILL 1 I
II II I II
1I Figure 2.2.1
4- I
The towers of Hanoi.

Example 2.2.11. The towers of Hanoi. It is said that after creating the
world, God set on Earth three rods made of diamond and 64 rings of gold. These rings
are all different in size. At the creation they were threaded on one of the rods in order
of size, the largest at the bottom and the smallest at the top. God also created a
monastery close by the rods. The monks' task in life is to transfer all the rings onto
another rod. The only operation permitted consists of moving a single ring from one
rod to another, in such a way that no ring is ever placed on top of another smaller one.
When the monks have finished their task, according to the legend, the world will come
to an end. This is probably the most reassuring prophecy ever made concerning the
end of the world, for if the monks manage to move one ring per second, working night
and day without ever resting nor ever making a mistake, their work will still not be
finished 500,000 million years after they began!
The problem can obviously be generalized to an arbitrary number of rings. For
example, with n = 3, we obtain the solution given in Figure 2.2.1. To solve the gen-
eral problem, we need only realize that to transfer the m smallest rings from rod i to
rod j (where 1 <- i <- 3, 1 <- j <- 3, i # j, and m ? 1), we can first transfer the smallest
m -1 rings from rod i to rod 6 - i -j, next transfer the m th ring from rod i to rod j,
and finally retransfer the m -1 smallest rings from rod 6 - i -j to rod j. Here is a
formal description of this algorithm ; to solve the original instance, all you have to do
(!) is to call it with the arguments (64, 1, 2).
procedure Hanoi(m, i, j)
(moves the m smallest rings from rod i to rod j }
ifm > 0 then Hanoi(m -1, i, 6 - i -j)
write i "-*" j
Hanoi(m - 1, 6 - i j, j)
To analyse the execution time of this algorithm, let us see how often the instruc-
tion write, which we use as a barometer, is executed. The answer is a function of m,
which we denote e (m). We obtain the following recurrence :
1 ifm=1
e(m)=
2e(m-1)+1 ifm > 1,
from which we find that e (m) = 2" - 1 (see Example 2.3.4). The algorithm therefore
takes a time in the exact order of 2" to solve the problem with n rings.
Problem 2.2.14. Prove that the algorithm of Example 2.2.11 is optimal in the
sense that it is impossible with the given constraints to move n rings from one rod to
another in less than 2" - 1 operations. 0
* Problem 2.2.15. Give a nonrecursive algorithm to solve this problem. (It is

cheating simply to rewrite the above algorithm using an explicit stack to simulate the
recursive calls.)
Sec. 2.3 Solving Recurrences Using the Characteristic Equation 65
2.3 SOLVING RECURRENCES USING

THE CHARACTERISTIC EQUATION
We have seen that the indispensable last step when analysing an algorithm is often to
solve a system of recurrences. With a little experience and intuition such recurrences
can often be solved by intelligent guesswork. This approach, which we do not illus-
trate here, generally proceeds in four stages : calculate the first few values of the
recurrence, look for regularity, guess a suitable general form, and finally, prove by
mathematical induction that this form is correct. Fortunately there exists a technique
that can be used to solve certain classes of recurrence almost automatically.
2.3.1 Homogeneous Recurrences
Our starting point is the resolution of homogeneous linear recurrences with constant
coefficients, that is, recurrences of the form
aptn +altn-1 + ... +aktn-k =0 (*)
where
i. the t, are the values we are looking for. The recurrence is linear because it does
not contain terms of the form t, ti+j , t, 2, and so on ;
ii. the coefficients a, are constants ; and
iii. the recurrence is homogeneous because the linear combination of the t, is equal
to zero.
After a while intuition may suggest we look for a solution of the form
tn
=xn
where x is a constant as yet unknown. If we try this solution in (*), we obtain
apx" +a1xn-I + ... +akx"-k = O.
This equation is satisfied if x = 0, a trivial solution of no interest, or else if
aoxk +alxk-I + ... +ak =0.
This equation of degree k in x is called the characteristic equation of the recurrence

M.
Suppose f o r the time being that the k roots r I , r2r 2 ,-. rk of this characteristic
equation are all distinct (they could be complex numbers). It is then easy to verify that
any linear combination
k
t" _ ci rin
i=1
of terms rrn is a solution of the recurrence (*), where the k constants c 1 , c 22 ,-.- Ck
are determined by the initial conditions. (We need exactly k initial conditions to deter-
mine the values of these k constants.) The remarkable fact, which we do not prove
here, is that (*) has only solutions of this form.
Example 2.3.1. Consider the recurrence

to -3tn-1 -4tn_2=0 n >2
subject to to = 0, t i = 1.
The characteristic equation of the recurrence is
x2-3x -4=0
whose roots are -1 and 4. The general solution therefore has the form
=C1(-1)n +C24n.
tn
The initial conditions give
c1+ c2=0 n =0
-c1+4c2=1 n=1
that is, c 1 =- 5. C2 =
1
I
We finally obtain
to = [4n - (-1)n j.
9
Example 2.3.2. Fibonacci. Consider the recurrence
to = to-1 + to-2
n >2
subject to to = 0, t 1 = 1.
(This is the definition of the Fibonacci sequence ; see Section 1.7.5.)
The recurrence can be rewritten in the form to - to-1 - to-2 = 0, so the charac-
teristic equation is
x2-x - 1 =0
whose roots are
2 and
r1=1+5 r2=
The general solution is therefore of the form
to =Cir +c2r2
c1+ c2=0 n =0
rice+r2c2=1 n =1
from which it is easy to obtain
CI=, T5
C2=- I
T5.
Thus t" = 5 (r i - r2" ). To show that this is the same as the result obtained by De
Moivre mentioned in Section 1.7.5, we need to note only that r I = 0 and r2 = - 0-I.
* Problem 2.3.1. Consider the recurrence

t" =2t"_I -2tn-2 n >2
subject to to = 0, t I = 1.
Prove that t, = 2'12sin(n it/4), not by mathematical induction but by using the
characteristic equation.
Now suppose that the roots of the characteristic equation are not all distinct. Let
p(x)=aoxk +alxk- I + ... + ak
be the polynomial in the characteristic equation, and let r be a multiple root. For every
n >- k, consider the n th degree polynomial defined by
h (x) = x [x" -k p (x)]' = a o n x" + a 1(n - 1)x"- I + + ak (n -k )x"-k
Let q (x) be the polynomial such that p (x) = (x -r )2q (x). We have that
n-k )x"-k q (x) + (x-r )2[x"-k q (x)]7
h (x) = x [(x-r )2 x q (x)]' = x [2(x-r
In particular, h (r) = 0. This shows that
aonr" +ai(n-1)r"-I+ +ak(n-k)rn-k =0,
that is, to = nr" is also a solution of (*). More generally, if m is the multiplicity of the
root r, then t" = r" , t" = nr" , t" = n 2r" , t" = n m - I r n are all possible solutions
of (*). The general solution is a linear combination of these terms and of the terms
contributed by the other roots of the characteristic equation. Once again there are k
constants to be determined by the initial conditions.

tn =5tn_I-Stn-2+4tn-3 n >_3
subject to to = 0, t I = 1, t2 = 2.
The recurrence can be written
to - Stn -I + 8tn-2 - 4tn-3 = 0

and so the characteristic equation is
x3-5x2+8x -4=0
or (x-1)(x-2)2 = 0.
The roots are 1 (of multiplicity 1) and 2 (of multiplicity 2). The general solution
is therefore
to =C11n +C22n +C3n2n.
c1+ C2 =0 n =0
c1+2c2+2c3=1 n=1
cl+4c2+8c3=2 n =2
from which we find c 1 = -2, c 2 = 2, c 3 = - z . Therefore
to = 2n +1 - n2n- - 2.
1
2.3.2 Inhomogeneous Recurrences
We now consider recurrences of a slightly more general form.

aotn +a1tn-1+ ... +aktn-k = b n p(n) ( ** )
The left-hand side is the same as (*), but on the right-hand side we have bn p (n),
where
i. b is a constant ; and
H. p (n) is a polynomial in n of degree d.
For example, the recurrence might be

tn -2tn_1=3n.
In this case b = 3 and p (n) = 1, a polynomial of degree 0. A little manipulation
allows us to reduce this example to the form (*). To see this, we first multiply the
recurrence by 3, obtaining
3n + 1
3tn - 6tn - 1 =
If we replace n by n + 1 in the original recurrence, we get
to+l-2tn=3n+1
Finally, subtracting these two equations, we have
to+1 -5th +6th_I =0,

which can be solved by the method of Section 2.3.1. The characteristic equation is
x2-5x +6=0
that is, (x-2)(x-3) = 0.
Intuitively we can see that the factor (x-2) corresponds to the left-hand side of
the original recurrence, whereas the factor (x-3) has appeared as a result of our mani-
pulation to get rid of the right-hand side.
Here is a second example.
th - 2t,-I = (n+5)3n
The necessary manipulation is a little more complicated : we must
a. multiply the recurrence by 9

b. replace n in the recurrence by n +2, and
c. replace n in the recurrence by n + 1 and then multiply by - 6,
obtaining respectively
9th - 18tn_I = (n+5)3n+2
to+2 - 2tn+l = (n+7)3' 2
-6th+I + 12th = -6(n+6)3n+1 .

Adding these three equations, we obtain
th+2 - 8th+1 + 21tn - 18tH-I = 0.
The characteristic equation of this new recurrence is
X3-8X2+21x-18=0
that is, (x-2)(x-3)2 = 0.
Once again, we can see that the factor (x-2) comes from the left-hand side of the ori-
ginal recurrence, whereas the factor (x-3)2 is the result of our manipulation.
Generalizing this approach, we can show that to solve (**) it is sufficient to take
the following characteristic equation :
(aoxk+alxk-I+ ... +ak)(x-b)d+1=0.
Once this equation is obtained, proceed as in the homogeneous case.
Example 2.3.4. The number of movements of a ring required in the Towers of

Hanoi problem (see Example 2.2.11) is given by
tn=2tn_1+I n>?1
subject to to = 0.
to - 2t,, -I = 1,
which is of the form (**) with b = 1 and p (n) = 1, a polynomial of degree 0. The
characteristic equation is therefore
(x-2)(x-1) = 0
where the factor (x-2) comes from the left-hand side and the factor (x-1) comes from
the right-hand side. The roots of this equation are 1 and 2, so the general solution of
the recurrence is
=Ciln +C22n.
to
We need two initial conditions. We know that to = 0; to find a second initial condition
we use the recurrence itself to calculate
t1=2to+I=1.
We finally have
c1+ c2=0 n =0
c1+2c2=1 n=1
from which we obtain the solution
to=2n-1.
If all we want is the order of to , there is no need to calculate the constants in the gen-
eral solution. In the previous example, once we know that
to = C I In + C22n
we can already conclude that to E 8(2n). For this it is sufficient to notice that to , the
number of movements of a ring required, is certainly neither negative nor a constant,
since clearly to >- n. Therefore C2 > 0, and the conclusion follows.
In fact we can obtain a little more. Substituting the general solution back into
the original recurrence, we find
1 = to - 2tn -I
=c1+C22n -2(c1+C22n-')
Whatever the initial condition, it is therefore always the case that c 1 must be equal
to -1.
Problem 2.3.2. There is nothing surprising in the fact that we can determine
one of the constants in the general solution without looking at the initial condition ; on
the contrary! Why?

to = 2tn -I + n.
This can be written
to - 2tn -I = n,
which is of the form (**) with b = 1 and p(n) = n, a polynomial of degree 1. The
characteristic equation is therefore
(x-2)(x-1)2 = 0
with roots 2 (multiplicity 1) and 1 (multiplicity 2). The general solution is
to =c12" +C2In +C3nIn.
In the problems that interest us, we are always looking for a solution where to >- 0 for
every n. If this is so, we can conclude immediately that to must be in 0 (2").
Problem 2.3.3. By substituting the general solution back into the recurrence,
prove that in the preceding example C2 = -2 and C3 = -1 whatever the initial condition.
Conclude that all the interesting solutions of the recurrence must have c i > 0, and
hence that they are all in O(2").
A further generalization of the same type of argument allows us finally to solve

recurrences of the form
aotn +aitn - I + ... +aktn-k = b?p1(n)+bnP2(n)+ ... (***)
where the b, are distinct constants and the p; (n) are polynomials in n respectively of
degree d; . It suffices to write the characteristic equation
(aoxk +aIxk - I+ . .. +ak)(X-b1)d'+I(x-b2)°,+I . . . =0
which contains one factor corresponding to the left-hand side and one factor
corresponding to each term on the right-hand side, and to solve the problem as before.
Example 2.3.6. Solve

tn =2tn_I+n +2" n >- 1
subject to to = 0.
t -2tn_I =n +2n,
which is of the form (***) with b 1 = 1, pi(n) = n , b2 = 2, p2(n) = 1. The degree of

p 1(n) is 1, and p2(n) is of degree 0. The characteristic equation is
(x-2)(x-1)2(x-2) = 0,
which has roots 1 and 2, both of multiplicity 2. The general solution of the recurrence
is therefore of the form
to =C1I" +C2nI" +C32" +Cgn2".
Using the recurrence, we can calculate t 1 = 3, t2 = 12, t3 = 35. We can now determine
C1 , C2, C3 and C4 from
Cl + c3 =0 n =0
c1+ c2+2c3+ 2c4= 3 n=1
c1 +2c2+4c3+ 8c4= 12 n =2
c1+3c2+8c3+24c4=35 n =3
arriving finally at
t, =-2-n +2n+1+n2".
We could obviously have concluded that t o e O (n 2") without calculating the con-
stants.
Problem 2.3.4. Prove that all the solutions of this recurrence are in fact in
O(n 2"), regardless of the initial condition.
Problem 2.3.5. If the characteristic equation of the recurrence (***) is of

degree
m =k +(d1+1)+(d2+1)+ ,
then the general solution contains m constants C 1 , C 2 , cm . How many con-
straints on these constants can be obtained without using the initial conditions ? (See
Problems 2.3.3 and 2.3.4.)
2.3.3 Change of Variable
It is sometimes possible to solve more complicated recurrences by making a change of

variable. In the following examples we write T (n) for the term of a general
recurrence, and tk for the term of a new recurrence obtained by a change of variable.
Example 2.3.7. Here is how we can find the order of T (n) if n is a power of 2
and if
T(n)=4T(n12)+n n > 1.
Replace n by 2k (so that k = lg n) to obtain T (2k) = 4T(2'- 1) + 2k . This can be
written
tk =4tk-1+2k
if tk = T (2k) = T (n). We know how to solve this new recurrence : the characteristic
equation is
(x-4)(x-2) = 0
and hence tk = c 14k + c 22k .
Putting n back instead of k, we find
T(n)=C1n2+C2n.
T (n) is therefore in 0 (n 2 1 n is a power of 2).
Example 2.3.8. Here is how to find the order of T (n) if n is a power of 2 and
if
T(n)=4T(n/2)+n2 n > 1.
Proceeding in the same way, we obtain successively
T(2k)=4T(2k-1)+4k
tk =4tk-I +4k.
The characteristic equation is (x-4)2 = 0, and so
tk =c14k +c2k4k
T(n) =c1n2+C2n2lgn.
Thus T (n) e O (n 2log n I n is a power of 2).
Example 2.3.9. Here is how to find the order of T (n) if n is a power of 2 and
if
T (n) = 2T (n / 2) + n lg n n > 1.
As before, we obtain
T(2k)=2T(2k-1)+k2k
tk =2tk-I+k2k
The characteristic equation is (x-2)3 = 0, and so
tk =c12k +C2k2k +C3k22k
T(n)=cin +c2nlgn +c3nlg2n.

Hence, T (n) E 0 (n log2n I n is a power of 2).
Example 2.3.10. We want to find the order of T (n) if n is a power of 2 and if

T (n) = 3T(n12) + cn (c is constant, n = 2k > 1).
We obtain successively
T(2k)=3T(2k-l)+c2k
tk =3tk_i+c2k.
The characteristic equation is (x-3)(x-2) = 0, and so
tk = c 13k + C22k
T(n)=c13'g" +c2n
and hence since algb = b1g°
T(n)=cInIg3+c2n.
Finally, T (n) E O (n 'g 3 I n is a power of 2). 0
Remark. In Examples 2.3.7 to 2.3.10 the recurrence given for T (n) only applies
when n is a power of 2. It is therefore inevitable that the solution obtained should be
in conditional asymptotic notation. In each of these four cases, however, it is sufficient
to add the condition that T (n) is eventually nondecreasing to be able to conclude that
the asymptotic results obtained apply unconditionally for all values of n. This follows
from problem 2.1.20 since the functions n 2, n 2 log n , n log2n and n I g 3 are smooth.
* Problem 2.3.6. The constants no ? 1, b >- 2 and k >- 0 are integers, whereas a
and c are positive real numbers. Let T : IN -) 1R+ be an eventually nondecreasing
function such that
T(n)=aT(n/b)+cnk n > no
when n In 0 is a power of b. Show that the exact order of T (n) is given by
O(nk) ifa < hk

T (n) E O(n k log n) ifa = b k
'ogn a
O(n ) ifa > b k
Rather than proving this result by constructive induction, obtain it using the techniques
of the characteristic equation and change of variable. This result is generalized in
Problem 2.3.13.
Problem 2.3.7. Solve the following recurrence exactly for n a power of 2:

T(n)=2T(n/2)+lgn n >>-2
subject to T(1) = 1.
Express your solution as simply as possible using the O notation.
22A
Problem 2.3.8. Solve the following recurrence exactly for n of the form
)+lgn n >_4
subject to T(2) = 1.
Express your solution as simply as possible using the 0 notation.
2.3.4 Range Transformations
When we make a change of variable, we transform the domain of the recurrence. It is

sometimes useful to transform the range instead in order to obtain something of the
form (***). We give just one example of this approach. We want to solve
T (n) = n T 2(n / 2) n>1
subject to T (l) = 6 for the case when n is a power of 2. The first step is a change of
variable : put tk = T (2k ), which gives
tk =2ktk2 I k>0
subject to to = 6.
At first glance, none of the techniques we have seen applies to this recurrence
since it is not linear, and furthermore, one of the coefficients is not constant. To
transform the range, we create a new recurrence by putting Vk = lg tk , which yields
-I k>0
subject to V() = lg 6.
The characteristic equation is (x-2)(x-1)2 = 0, and so
Vk =c12k +c2lk +c3klk.
From VO= I + lg 3, V, = 3 + 21g 3, and V2 = 8 + 41g 3 we obtain c i = 3 + lg 3,
C2=-2, and c 3 =-J, and hence
Vk = (3 + lg 3)2k - k - 2.
Finally, using tk = 2t'" and T(n) = tlgn , we obtain
23n-23n
T(n)= n
2.3.5 Supplementary Problems
Problem 2.3.9. Solve the following recurrence exactly :

tn =tn_I +tn_3-tn_4 n >>>4
subject to to = n for 0 5 n < 3. Express your answer as simply as possible using the
O notation.
Problem 2.3.10. Solve the following recurrence exactly for n a power of 2:

T (n) = 5T (n / 2) + (n lg n)2 n >- 2
subject to T(1) = 1. Express your answer as simply as possible using the 0 notation.
Problem 2.3.11. (Multiplication of large integers: see Sections 1.7.2, 4.1,

and 4.7). Consider any constants c E R+ and no E N. Let T AN 4 R* be an eventu-
ally nondecreasing function such that
T(n) ST(Ln/2])+T([n/2])+T(1+[n/21)+cn n > no.
Prove that T (n) E O (n ig 3 ). Hint : observe that T (n) < 3 T (1 + [n / 21) + cn for n > n 0,
make the change of variable T '(n) = T (n+2), use Example 2.3.10 to solve for T '(n)
when n is a power of 2, and use problem 2.1.20 to conclude for T (n).

to = to -1 + 2tn-2 - 2tn-3 n >3
subject to to = 9n 2 - 15n + 106 for 0!5 n :5 2. Express your answer as simply as pos-
sible using the 0 notation.
* Problem 2.3.13. Recurrences arising from the analysis of divide-and-conquer

algorithms (Chapter 4) can usually be handled by Problem 2.3.6. In some cases, how-
ever, a more general result is required (and the technique of the characteristic equation
does not always apply).
The constants no >-1 and b >- 2 are integers, whereas a and d are real positive
constants. Define
X={n ENIlogb(n/no)EIN}={nENNI(3iEIN)[n=nobÌ).
Let f : X - R* be an arbitrary function. Define the function T : X -* R* by the
recurrence
d if n =no
T(n)=
aT (nlb)+ f(n) if nEX,n>no
Let p = loge a . It turns out that the simplest way to express T (n) in asymptotic nota-
tion depends on how f (n) compares to n P . In what follows, all asymptotic notation is
implicitly conditional on n E=X. Prove that
i. If we set f (n o) = d (which is of no consequence for the definition of T ), the

value of T (n) is given by a simple summation when n EX
logn(n In,)
T (n) = E a' f (n l b')
i=o
ii. Let q be any strictly positive real constant ; then
O(no) if f(n) E O(nP/(logn)1+q )

O(f (n) log n log log n) if f (n) E O(n P / log n)
T (n) E <
O(f(n)logn) if f(n) E O(nP(log n)q 1)
Off (n)) if f (n) E O(n p+q) .
Note that the third alternative includes f (n) E O(nP) by choosing q=1.
W. As a special case of the first alternative, T (n) e O(n P) whenever f (n) E O (n' ) for
some real constant r < p .
iv. The last alternative can be generalized to include cases such as
f (n) E O(n P+q log n) or f (n) E O(n P+q / log n) ; we also get T (n) E O(f (n)) if
there exist a function g : X -* IR* and a real constant a> a such that
f(n)EO(g(n)) and g(bn)>-ag(n) for all
v. Prove or disprove that the third alternative can be generalized as follows :
T(n) E O(f (n) log n) whenever there exist two strictly positive real constants
). If you
q 1 <_ q 2 such that f (n) E O (n P (log n)q' 1) and f (n) E Q(n P (log n)'
disprove it, find the simplest but most general additional constraint on f (n) that
suffices to imply T (n) E O(f (n) log n).

t = 1/(4-tn_1) n>1
subject tot 1 = 1/4.
Problem 2.3.15. Solve the following recurrence exactly as a function of the

initial conditions a and b :
T(n+2)=(1+T(n+l))/T(n) n >2
subject to T (O) = a, T (I) = b.

T(n)='T(n/2)-zT(n/4)-1/n n?3
subject to T (1) = 1 and T (2) = 3/2. El
The asymptotic notation has existed for some while in mathematics: see Bachmann
(1894) and de Bruijn (1961). Knuth (1976) gives an account of its history and pro-
poses a standard form for it. Conditional asymptotic notation and its use in Problem
2.1.20 are introduced by Brassard (1985), who also suggests that "one-way inequali-
ties" should be abandoned in favour of a notation based on sets. For information on
calculating limits and on de l'Hopital's rule, consult any book on mathematical
analysis, Rudin (1953), for instance.
The book by Purdom and Brown (1985) presents a number of techniques for ana-
lysing algorithms. The main mathematical aspects of the analysis of algorithms can
also be found in Greene and Knuth (1981).
Example 2.1.1 corresponds to the algorithm of Dixon (1981). Problem 2.2.3
comes from Williams (1964). The analysis of disjoint set structures given in Example
2.2.10 is adapted from Hopcroft and Ullman (1973). The more precise analysis
making use of Ackermann's function can be found in Tarjan (1975, 1983). Buneman
and Levy (1980) and Dewdney (1984) give a solution to Problem 2.2.15.
Several techniques for solving recurrences, including the characteristic equation
and change of variable, are explained in Lueker (1980). For a more rigorous
mathematical treatment see Knuth (1968) or Purdom and Brown (1985). The paper by
Bentley, Haken, and Saxe (1980) is particularly relevant for recurrences occurring from
the analysis of divide-and-conquer algorithms (see Chapter 4).
3
Greedy Algorithms
3.1 INTRODUCTION
Greedy algorithms are usually quite simple. They are typically used to solve optimiza-
tion problems : find the best order to execute a certain set of jobs on a computer, find
the shortest route in a graph, and so on. In the most common situation we have
a set (or a list) of candidates : the jobs to be executed, the nodes of the graph, or
whatever ;
the set of candidates that have already been used ;
a function that checks whether a particular set of candidates provides a solution
to our problem, ignoring questions of optimality for the time being ;
a function that checks whether a set of candidates is feasible, that is, whether or
not it is possible to complete the set in such a way as to obtain at least one solu-
tion (not necessarily optimal) to our problem (we usually expect that the problem
has at least one solution making use of candidates from the set initially avail-
able);
a selection function that indicates at any time which is the most promising of the
candidates not yet used ; and
an objective function that gives the value of a solution (the time needed to exe-
cute all the jobs in the given order, the length of the path we have found, and so
on); this is the function we are trying to optimize.
To solve our optimization problem, we look for a set of candidates constituting a

solution that optimizes (minimizes or maximizes, as the case may be) the value of the
79
80 Greedy Algorithms Chap. 3
objective function. A greedy algorithm proceeds step by step. Initially, the set of
chosen candidates is empty. Then at each step, we try to add to this set the best
remaining candidate, our choice being guided by the selection function. If the enlarged
set of chosen candidates is no longer feasible, we remove the candidate we just added ;
the candidate we tried and removed is never considered again. However, if the
enlarged set is still feasible, then the candidate we just added stays in the set of chosen
candidates from now on. Each time we enlarge the set of chosen candidates, we check
whether the set now constitutes a solution to our problem. When a greedy algorithm
works correctly, the first solution found in this way is always optimal.
function greedy (C : set) : set

{ C is the set of all the candidates)
S f- 0 (S is a set in which we construct the solution)
while not solution (S) and C # 0 do
x - an element of C maximizing select(x)
C-C\{x}
if feasible (S u {x }) then S - S u {x }
if solution (S) then return S
else return "there are no solutions"
It is easy to see why such algorithms are called "greedy": at every step, the pro-
cedure chooses the best morsel it can swallow, without worrying about the future.
It never changes its mind : once a candidate is included in the solution, it is there for
good ; once a candidate is excluded from the solution, it is never reconsidered.
The selection function is usually based on the objective function ; they may even
be identical. However, we shall see in the following examples that at times there may
be several plausible selection functions, so that we have to choose the right one if we
want our algorithm to work properly.
Example 3.1.1. We want to give change to a customer using the smallest pos-
sible number of coins. The elements of the problem are
the candidates: a finite set of coins, representing for instance 1, 5, 10, and 25
units, and containing at least one coin of each type ;
a solution : the total value of the chosen set of coins is exactly the amount we
have to pay ;
a feasible set : the total value of the chosen set does not exceed the amount to be
paid
the selection function : choose the highest-valued coin remaining in the set of
candidates; and
the objective function : the number of coins used in the solution. 0
Sec. 3.2 Greedy Algorithms and Graphs 81
* Problem 3.1.1. Prove that with the values suggested for the coins in the
preceding example the greedy algorithm will always find an optimal solution provided
one exists.
Prove, on the other hand, by giving specific counterexamples, that the greedy
algorithm no longer gives an optimal solution in every case if there also exist 12-unit
coins, or if one type of coin is missing from the initial set. Show that it can even
happen that the greedy algorithm fails to find a solution at all despite the fact that one
exists.
It is obviously more efficient to reject all the remaining 25-unit coins (say) at
once when the remaining amount to be represented falls below this value. Using
integer division is also more efficient than proceeding by successive subtractions.
3.2 GREEDY ALGORITHMS AND GRAPHS
3.2.1 Minimal Spanning Trees
Let G = < N, A > be a connected undirected graph where N is the set of nodes and A
is the set of edges. Each edge has a given non-negative length. The problem is to find
a subset T of the edges of G such that all the nodes remain connected when only the
edges in T are used, and the sum of the lengths of the edges in T is as small as pos-
sible. (Instead of talking about length, we can associate a cost to each edge. In this
case the problem is to find a subset T whose total cost is as small as possible. Obvi-
ously, this change of terminology does not affect the way we solve the problem.)
Problem 3.2.1. Prove that the partial graph < N, T > formed by the nodes of
G and the edges in T is a tree.
The graph < N, T > is called a minimal spanning tree for the graph G. This
problem has many applications. For instance, if the nodes of G represent towns, and
the cost of an edge (a, b } is the cost of building a road from a to b, then a minimal
spanning tree of G shows us how to construct at the lowest possible cost a road system
linking all the towns in question.
We give two greedy algorithms to solve this problem. In the terminology we
have used for greedy algorithms, a set of edges is a solution if it constitutes a spanning
tree, and it is feasible if it does not include a cycle. Moreover, a feasible set of edges
is promising if it can be completed so as to form an optimal solution. In particular, the
empty set is always promising since G is connected. Finally, an edge touches a given
set of nodes if exactly one end of the edge is in the set. The following lemma is cru-
cial for proving the correctness of the forthcoming algorithms.
Lemma 3.2.1. Let G = < N, A > be a connected undirected graph where the
length of each edge is given. Let B C N be a strict subset of the nodes of G. Let
T c A be a promising set of edges such that no edge in T touches B. Let e be the

shortest edge that touches B (or any one of the shortest if ties exist). Then T v f e } is
promising.
Proof. Let U be a minimal spanning tree of G such that T E_ U (such a U must
exist since T is promising by assumption). If e e U, there is nothing to prove. Other-
wise when we add the edge e to U, we create exactly one cycle (this is one of the pro-
perties of a tree). In this cycle, since e touches B, there necessarily exists at least one
other edge, e', say, that also touches B (otherwise the cycle could not close - see
Figure 3.2.1). If we now remove e', the cycle disappears and we obtain a new tree U'
that spans G. But since the length of e is by definition no greater than the length of
e', the total length of the edges in U' does not exceed the total length in U. Therefore
U' is also a minimal spanning tree, and it includes e. To complete the proof, we note
that T c U' because the edge e', which touches B, cannot be in T.
The initial set of candidates is the set of all the edges. A greedy algorithm
selects the edges one by one in some given order. Each edge is either included in the
set that will eventually form the solution or eliminated from further consideration. The
main difference between the various greedy algorithms to solve this problem lies in the
order in which the edges are selected.
Kruskal's algorithm. The set T of edges is initially empty. As the algorithm

progresses, edges are added to T. At every instant the partial graph formed by the
nodes of G and the edges in T consists of several connected components. (Initially,
when T is empty, each node of G forms a distinct trivial connected component.) The
elements of T that are included in a given connected component form a minimal span-
ning tree for the nodes in this component. At the end of the algorithm only one con-
nected component remains, so that T is then a minimal spanning tree for all the nodes
of G.
N\B B
Figure 3.2.1. A cycle is created if we add edge e to U.

To build bigger and bigger connected components, we examine the edges of G in

order of increasing length. If an edge joins two nodes in different connected com-
ponents, we add it to T, and consequently, the two connected components now form
only one component. Otherwise the edge is rejected : it joins two nodes in the same
connected component and cannot therefore be added to T without forming a cycle since
the edges in T form a minimal spanning tree for each component. The algorithm stops
when only one connected component remains.
To illustrate how this algorithm works, consider the graph in figure 3.2.2. In
increasing order of length the edges are : (1, 2) , 12, 3 } , 14,5 } , (6,71, (1,4), (2, 5) ,
(4,7), 13,5), (2,4}, (3,6}, (5,71, (5,6). The algorithm proceeds as follows.
Step Edge Connected components

considered
Initialization
1
-
11,21
(11 (2) (3) (4) 151 (6) f7)
11,21 (3} (4} 151 (61 (71
2 (2,3} (1,2,3} (41 (5} (6) 171
3 (4,5) (1,2,3} (4,51 (61 (7)
4 16,71 (1,2,3) (4,5} (6,7)
5 (1,4} (1,2,3,4,5} (6,7)
6 t2,5) rejected
7 14,7) (1,2,3,4,5,6,7)
T contains the chosen edges (1,2), (2, 31, (4,5), (6,7), (1,4), and (4,7). This
minimal spanning tree is shown by the heavy lines in Figure 3.2.2; its total length
is 17.
Problem 3.2.2. Prove that Kruskal's algorithm works correctly. The proof,
which uses lemma 3.2.1, is by induction on the number of edges selected until now. El
Figure 3.2.2. A graph and its minimal spanning tree.

Problem 3.2.3. A graph may have several different minimal spanning trees.
Is this the case in our example, and if so, where is this possibility reflected in the algo-
rithm ? 0
To implement the algorithm, we have to handle a certain number of sets : the
nodes in each connected component. We have to carry out rapidly the two operations
find (x), which tells us in which component the node x is to be found, and
merge (A , B) to merge two disjoint sets. We therefore use disjoint set structures (Sec-
tion 1.9.5). For this algorithm it is preferable to represent the graph as a vector of
edges with their associated lengths rather than as a matrix of distances. Here is the
algorithm.
function Kruskal (G = < N, A > : graph ; length : A -i IR*) : set of edges
1 initialization)
Sort A by increasing length
n -#N
T - 0 1 will contain the edges of the minimal spanning tree)
initialize n sets, each containing one distinct element of N
{ greedy loop }
repeat
{ u , v) f- shortest edge not yet considered
ucomp F-- find (u )
vcomp f- find (v)
if ucomp # vcomp then
merge (ucomp, vcomp)
T -T u {{u,v}}
until #T = n -1
return T
Problem 3.2.4. What happens if, by mistake, we run the algorithm on a graph
that is not connected?
We can estimate the execution time of the algorithm as follows. On a graph with
n nodes and a edges the number of operations is in
0 (a log a) to sort the edges, which is equivalent to 0 (a log n) since
n-1 <-a Sn(n-1)/2;
0 (n) to initialize the n disjoint sets ;
in the worst case 0 ((2a+n -1)lg*n) for all the find and merge operations, by
the analysis given in example 2.2.10, since there are at most 2a find operations
and n -1 merge operations on a universe containing n elements ; and
at worst, O (a) for the remaining operations.
For a connected graph we know that a >- n - 1. We conclude that the total time
for the algorithm is in 0 (a log n) because 0 (lg*n) c 0 (log n). Although this does not
change the worst-case analysis, it is preferable to keep the edges in a heap (Section
1.9.4 - here the heap property should be inverted so that the value of each internal
node is less than or equal to the values of its children). This allows the initialization to
be carried out in a time in 0 (a), although each search for a minimum in the repeat
loop will now take a time in 0 (log a) = 0 (log n). This is particularly advantageous
in cases when the minimal spanning tree is found at a moment when a considerable
number of edges remain to be tried. In such cases, the original algorithm wastes time
sorting all these useless edges.
Problem 3.2.5. What can you say about the time required by Kruskal's algo-
rithm if, instead of providing a list of edges, the user supplies a matrix of distances,
leaving to the algorithm the job of working out which edges exist ?
Prim's algorithm. In Kruskal's algorithm we choose promising edges

without worrying too much about their connection to previously chosen edges, except
that we are careful never to form a cycle. There results a forest of trees that grows
somewhat haphazardly. In Prim's algorithm, on the other hand, the minimal spanning
tree grows "naturally", starting from an arbitrary root. At each stage we add a new
branch to the tree already constructed, and the algorithm stops when all the nodes have
been reached.
Initially, the set B of nodes contains a single arbitrary node, and the set T of
edges is empty. At each step Prim's algorithm looks for the shortest possible edge
{ u , v J such that u E N \ B and v E B . It then adds u to B and (u, v) to T . In this way
the edges in T form at any instant a minimal spanning tree for the nodes in B. We con-
tinue thus as long as B # N. Here is an informal statement of the algorithm.
function Prim (G = < N, A > : graph ; length : A -> IR*) : set of edges
( initialization)
T F- 0 (will contain the edges of the minimal spanning tree }
B (-- (an arbitrary member of N }
while B # N do
find { u , v } of minimum length such that u E N \ B and v E B
T 6-T u {(u,v}}
B F-B U (u)
return T
Problem 3.2.6. Prove that Prim's algorithm works correctly. The proof,
which again uses Lemma 3.2.1, is by induction on the number of nodes in B.
To illustrate how the algorithm works, consider once again the graph in Figure
3.2.2. We arbitrarily choose node I as the starting node.
gives the length of each directed edge : L [i, j ] ? 0 if the edge (i, j) exists and
L [i, j ] = 00 otherwise. Here is the algorithm.
function Dijkstra (L [1 .. n, I .. n ]) : array[2.. n ]
(initialization)
C F- { 2, 3, ... , n ) { S = N \ C exists only by implication )
for i -2tondoD[i]*-L[l,i]
{ greedy loop }
repeat n -2 times
v - some element of C minimizing D [v ]
C (-- C \ { v) { and implicitly S - S u { v } }
for each w e C do
D[w] f- min(D[w], D [v] +L[v,w])
return D
The algorithm proceeds as follows on the graph in Figure 3.2.3.
Step V C D
Initialization
1
-5
(2,3,4,5)
(2,3,4)
[50,30,100,10]
[50,30,20,10]
2 4 12,31 [40,30,20, 10]
3 3 {2) (35,30,20, 10]
Clearly, D would not change if we did one more iteration to remove the last element of
C, which is why the main loop is only repeated n - 2 times.
If we want not only to know the length of the shortest paths but also where they
pass, it suffices to add a second array P [2.. n ], where P [v ] contains the number of the
node that precedes v in the shortest path. To find the complete path, simply follow the
pointers P backwards from a destination to the source. The modifications to the algo-
rithm are simple :
Figure 3.2.3 A directed graph.

initialize P [i] to 1 for i = 2, 3, ... , n ;

replace the contents of the inner for loop by
ifD[w]>D[v]+L[v,w] then D[w] <-- D[v] +L[v,w]

P[w]-v.
Problem 3.2.10. Show how the modified algorithm works on the graph of
Figure 3.2.3.
Proof of correctness. We prove by mathematical induction that
i. if a node i is in S, then D [i] gives the length of the shortest path from the source
to i
H. if a node i is not in S, then D [i] gives the length of the shortest special path
from the source to i.
Look at the initialization of D and S to convince yourself that these two condi-
tions hold at the outset ; the base for our induction is thus obtained. Next, consider the
inductive step, and suppose by the induction hypothesis that these two conditions hold
just before we add a new node v to S.
I. This follows immediately from the induction hypothesis for each node i that was
already in S before the addition of v. As for node v, it will now belong to S.
We must therefore check that D [v ] gives the length of the shortest path from the
source to v. By the induction hypothesis D [v ] certainly gives the length of the
shortest special path. We therefore have to verify that the shortest path from the
source to v does not pass through a node that does not belong to S. Suppose the
contrary : when we follow the shortest path from the source to v, the first node
encountered that does not belong to S is some node x distinct from v (see Figure
3.2.4).
The initial section of the path, as far as x, is a special path. Consequently,
the total distance to v via x is
distance to x (since edge lengths are non-negative)
D [x] (by part (ii) of the induction)
D [v] (because the algorithm chose v before x)
and the path via x cannot be shorter than the special path leading to v.
We have thus verified that when v is added to S, part (i) of the induction
remains true.
ii. Consider now a node w 0 S different from v When v is added to S, there are
.
two possibilities for the shortest special path from the source to w : either it does
not change, or else it now passes through v. In the latter case it seems at first
The shortest path
The shortest special path
Figure 3.2.4. The shortest path from the source to v cannot go through node x.
glance that there are again two possibilities : either v is the last node in S visited
before arriving at w or it is not. We have to compare explicitly the length of the
old special path leading to w and the length of the special path that visits v just
before arriving at w ; the algorithm does this. However, we can ignore the possi-
bility (see Figure 3.2.5) that v is visited, but not just before arriving at w : a path
of this type cannot be shorter than the path of length D [x] + L [x, w] that we
examined at a previous step when x was added to S, because D [x] <_ D [v].
Thus the algorithm ensures that part (ii) of the induction also remains true when
a new node v is added to S.
To complete the proof that the algorithm works, we need only note that when its
execution stops all the nodes but one are in S (even though the set S is not constructed
explicitly). At this point it is clear that the shortest path from the source to the
remaining node is a special path.
Problem 3.2.11. Show by giving an explicit example that if the edge lengths
can be negative, then Dijkstra's algorithm does not always work correctly. Is it still
sensible to talk about shortest paths if negative distances are allowed?
Figure 3.2.5 The shortest path from the source to w cannot visit x between v and w.
Analysis of the algorithm. Suppose Dijkstra's algorithm is applied to a

graph having n nodes and a edges. Using the representation suggested up to now, the
instance is given in the form of a matrix L [1 .. n , 1 .. n J. Initialization takes a time in
O (n). In a straightforward implementation, choosing v in the repeat loop requires all
the elements of C to be examined, so that we look at n - 1 , n -2, ... , 2 values of D
on successive iterations, giving a total time in 0 (n 2). The inner for loop does n - 2,
n - 1, ... , 1 iterations, for a total also in 0 (n 2). The time required by this version of
the algorithm is therefore in O (n 2).
If a << n 2, it seems we might be able to avoid looking at the many entries con-
taining - in the matrix L. With this in mind, it could be preferable to represent the
graph by an array of n lists, giving for each node its direct distance to adjacent nodes
(like the type lisgraph of Section 1.9.2). This allows us to save time in the inner for
loop, since we only have to consider those nodes w adjacent to v, but how are we to
avoid taking a time in c(n2) to determine in succession the n -2 values taken by v ?
The answer is to use a heap containing one node for each element v of C,
ordered by the value of D [v]. If we remember to invert the heap, the element v of C
that minimizes D [v] will always be found at the root. Initialization of the heap takes a
time in 0 (n). The instruction " C - C \ { v } " consists of eliminating the root
from the heap, which takes a time in O (log n). As for the inner for loop, it now
consists of looking, for each element w of C adjacent to v, to see whether
D [v] + L [v, w ] < D [w]. If so, we must modify D [w] and percolate w up the heap,
which again takes a time in 0 (log n ). This does not happen more than once for each
edge of the graph.
To sum up, we have to remove the root of the heap exactly n - 2 times and to
percolate at most a nodes, giving a total time in 0 ((a +n) log n). If the graph is con-
nected, a > n -1 and this time is in 0 (a log n). The straightforward implementation
is therefore preferable if the graph is dense, whereas it is preferable to use a heap if the
graph is sparse. If a E O(n 2/ log n), the choice of algorithm may depend on the specific
implementation.
* Problem 3.2.12. In the preceding analysis, we saw that up to a nodes can be

percolated, whereas less than n roots are eliminated. This is interesting when we
remember that eliminating the root has for effect to sift down the node that takes its
place, and that percolating up is somewhat quicker than sifting down (at each level, we
compare the value of a node to the value of its parent rather than making comparisons
with both children). We might therefore consider modifying the definition of a heap
slightly to allow percolation to run faster still, at the cost of slowing down sifting.
i. What modification to the definition of a heap do you suggest ?

ii. Let k = max(2, La In j ). Show how your modification allows you to calculate
the shortest paths from a source to all the other nodes of a graph in a time in
O (a logk n). Problem 2.1.17(i) does not apply here since k is not a constant.
Note that this gives 0 (n 2) if a = n 2 and 0 (a log n) if a = n ; it therefore gives
the best of both worlds. (Still faster algorithms exist.)
Problem 3.2.13. Show that Prim's algorithm to find minimal spanning trees
can also be implemented through the use of heaps. Show that it then takes a time in
0(a log n ), just as Kruskal's algorithm would. Finally, show that the modification
suggested in the previous problem applies just as well to Prim's algorithm.
3.3 GREEDY ALGORITHMS FOR SCHEDULING

3.3.1 Minimizing Time in the System
A single server (a processor, a petrol pump, a cashier in a bank, and so on) has n cus-
tomers to serve. The service time required by each customer is known in advance :
customer i will take time t, , 1 <_ i <_ n. We want to minimize
n
T= (time in system for customer i ).
Since the number of customers is fixed, minimizing the total time in the system is
equivalent to minimizing the average time. For example, if we have three customers
with
t1=5, t2=10, t3=3,
then six orders of service are possible.
Order T
123 : 5 + (5+10) + (5+10+3)=38
132: 5 + (5+3) + (5+3+10)=31
213: 10 + (10+5) + (10+5+3)=43
231 : 10 + (10+3) + (10+3+5) = 41
312: 3 + (3+5) + (3 + 5 + 10) = 29 E- optimal
321: 3 + (3+10) + (3+10+5)=34
In the first case, customer 1 is served immediately, customer 2 waits while customer 1
is served and then gets his turn, and customer 3 waits while both 1 and 2 are served
and then is served himself: the total time passed in the system by the three customers
is 38.
Imagine an algorithm that builds the optimal schedule step by step. Suppose that
after scheduling customers i I , i 2, ... , i,,, we add customer j. The increase in T at
this stage is
Sec. 3.3 Greedy Algorithms for Scheduling 93
+ti,+ +ti +tj .

To minimize this increase, we need only minimize tj . This suggests a simple greedy
algorithm : at each step, add to the end of the schedule the customer requiring the least
service among those who remain. In the preceding example this algorithm gives the
correct answer 3, 1, 2.
We now prove that this algorithm is always optimal. Let I = (i I i 2 ... in ) be any
permutation of the integers { 1 , 2, ... , n 1. If customers are served in the order 1, the
total time passed in the system by all the customers is
T(I)=ti,+(ti,+ti)+(ti,+ti,+
=nti +(n-1)ti,+(n-2)ti +
n
_ k(n-k+1)t; .
k=1
Suppose now that I is such that we can find two integers a and b with a t; in other words, the a th customer is served before the b th customer even
:
though the former needs more service time than the latter (see Figure 3.3.1). If we
exchange the positions of these two customers, we obtain a new order of service I'
obtained from I by interchanging the items is and ib This new order is preferable .
because
n
T(I')=(n-a+1)ti,+(n-b+1)ti + E (n-k+l)t;
k=1
kta,b
Service order 1 2 .. a b n
Served customer 11 i,
(from I)
Service duration t, t,1 ... t' ... t" ... fin
After exchange
of is and !b
I
Service duration l t' ... tab ... tea ... tin
Served customer 1o
(from I')
Figure 3.3.1 Exchanging the positions of two customers.

and thus
(b -a )(ti, - tin )
>0.
We can therefore improve any schedule in which a customer is served before someone
else who requires less service. The only schedules that remain are those obtained by
putting the customers in nondecreasing order of service time. All such schedules are
clearly equivalent, and therefore they are all optimal.
Problem 3.3.1. How much time (use the 0 notation) is required by a greedy
algorithm that accepts n and t [I .. n ] as data and produces an optimal schedule ?
The problem can be generalized to a system with s servers, as can the algorithm.
Without loss of generality, suppose the customers are numbered so that
t 1 <- t2 <- S to . In this context, server i, 1 s , must serve customers number
i9 i +s , i + 2s .... in that order.
Problem 3.3.2. Prove that this algorithm always yields an optimal schedule.
Problem 3.3.3. A magnetic tape contains n programs of length l 1 , l2 1 2 ,--

In . We know how often each program is used : a fraction pi of requests to load a pro-
gram concern program i, 1 <- i <- n. (This of course implies that Ji 1 pi = 1.) Infor-
mation is recorded along the tape at constant density, and the speed of the tape drive is
also constant. Each time a program has been loaded, the tape is rewound to the begin-
ning.
If the programs are held in the order i 1, i 2, ... , i, , the average time required
to load a program is
n I
T =cE Pr Eli
- j=1 k=1
where the constant c depends on the recording density and the speed of the drive. We
want to minimize T.
i. Prove by giving an explicit example that it is not necessarily optimal to hold the
programs in order of increasing values of 1i .
ii. Prove by giving an explicit example that it is not necessarily optimal to hold the
programs in order of decreasing values of pi
iii. Prove that T is minimized if the programs are held in order of decreasing pi /li .
3.3.2 Scheduling with Deadlines
We have a set of n jobs to execute, each of which takes unit time. At any instant
t = 1, 2, ... , we can execute exactly one job. Job i, 1 <_ i <_ n , earns us a profit g; if
and only if it is executed no later than time d, .
For example, with n = 4 and the following values :
i 1 2 3 4
g, 50 10 15 30
d; 2 1 2 1
the schedules to consider and the corresponding profits are
Sequence: 1 Profit: 50
2 10
3 15
4 30
1,3 65
2,1 60
2,3 25
3,1 65
4, 1 80 - optimum
4,3 45.
The sequence 3, 2, for instance, is not considered because job 2 would be executed at
time t = 2, after its deadline d2 = 1. To maximize our profit in this example, we
should execute the schedule 4, 1.
A set of jobs is feasible if there exists at least one sequence (also called feasible)
that allows all the jobs in the set to be executed in time for their respective deadlines.
An obvious greedy algorithm consists of constructing the schedule step by step, adding
at each step the job with the highest value of g; among those not yet considered, pro-
vided that the chosen set of jobs remains feasible.
In the preceding example we first choose job 1. Next, we choose job 4: the set
11, 41 is feasible because it can be executed in the order 4, 1. Next, we try the set
11, 3, 41, which turns out not to be feasible ; job 3 is therefore rejected. Finally we try
(1, 2, 41, which is also infeasible ; so job 2 is also rejected. Our solution - optimal in
this case - is therefore to execute the set of jobs 11,41, which in fact can only be
done in the order 4, 1. It remains to prove that this algorithm always finds an optimal
schedule and to find an efficient way of implementing it.
Let J be a set of k jobs. At first glance it seems we might have to try all the k !
possible permutations of these jobs to see whether J is feasible. Happily this is not the
case.
Lemma 3.3.1. Let J be a set of k jobs, and let a = (s 1s 2 . sk) be a permuta-

.
tion of these jobs such that ds <- ds2 ds, . Then the set J is feasible if and only
if the sequence a is feasible.
Proof. The "if " is obvious. For the "only if " :

If J is feasible, there exists at least one sequence of jobs p = (r 1r2 rk) such
that dr, > i , 1 <- i S k. Suppose a # p. Let a be the smallest index such that Sa # rQ
and let b be defined by rb = sQ ; it is clear that b > a . Also
dr, ? d, (by the construction of a and the minimality of a)
= dr, (by the definition of b).
The job ra could therefore be executed later, at the time when rb is at present
scheduled. Since rb can certainly be executed earlier than planned, we can interchange
the items ra and rb in p. The result is a new feasible sequence, which is the same as a
at least in positions 1, 2, ... , a. Continuing thus, we obtain a series of feasible
sequences, each having at least one more position in agreement with a. Finally, after a
maximum of k -1 steps of this type, we obtain a itself, which is therefore feasible. 11
This shows that it suffices to check a single sequence, in order of increasing

deadlines, to know whether a set of jobs is or is not feasible. We now prove that the
greedy algorithm outlined earlier always finds an optimal schedule.
Proof of optimality. Suppose that the greedy algorithm chooses to execute a

set of jobs I whereas in fact the set J # I is optimal. Consider two feasible sequences
S, and S,, for the two sets of jobs in question. By making appropriate interchanges of
jobs in S, and S,, , we can obtain two feasible sequences S; and S,, such that every job
common to both I and J is scheduled for execution at the same time in the two
sequences. (We may have to leave gaps in the schedule. The necessary interchanges
are easily found if we scan the two sequences S, and S., from right to left. See Figure
3.3.2 for an example.) S; and S are distinct sequences since I # J. Let us consider an
arbitrary time when the task scheduled in Sj is different from that scheduled in S; .
If some task a is scheduled in S, whereas there is a gap in S,, (and therefore task
a does not belong to J ), the set J u { a } is feasible and would be more profitable
than J. This is not possible since J is optimal by assumption.
If some task b is scheduled in S. whereas there is a gap in S; , the set I u { b } is
feasible, hence the greedy algorithm should have included b in I. This is also
impossible since it did not do so.
The only remaining possibility is that some task a is scheduled in S; whereas a
different task b is scheduled in S,, . Again, this implies that a does not appear in
J and that b does not appear in I.
If gQ > gb , one could substitute a for b in J and improve it. This goes
against the optimality of J.
- If g, < gl, , the greedy algorithm should have chosen h before even con-
sidering a since (I \ { a }) u { h } would be feasible. This is not possible
either since it did not include h in I.
- The only remaining possibility is therefore that go = gn
In conclusion, for each time slot, sequences S; and S,, either schedule no tasks,
the same task, or two distinct tasks yielding the same profit. This implies that the total
worth of I is identical with that of the optimal set J, and thus I is optimal as well.
For our first implementation of the algorithm suppose without loss of generality
that the jobs are numbered so that g I >- g 2 ? >- g . To allow us to use sentinels,
suppose further that n > 0 and that di > 0, 1 <_ i <- n .
function sequence (d [0.. n ]) : k, array[ 1 .. k ]
array j [0.. n ]
d [0], j [0] f- 0 (sentinels)
k, j [ l ] - 1 { task 1 is always chosen }
{greedy loop I
for i <-- 2 to n do { in decreasing order of g I
r-k
while d [ j Ir ]] > max(d [i], r) do r F- r -
ifd[j[r]]<-d[i]andd[i]>r then
for/ F-kstep-] tor+lldoj[I+1]E-j[l]
j[r+I] - i
k4-k+I
return k, j[I ..k]
P Y q x r S,
r s r P u V q w Si
after reorganization,
if this task is a
x Y P r q S;
u S P r V q w S"
that one will be b

common tasks
Figure 3.3.2. Rearranging schedules to bring identical tasks together.

The exact values of the g, are unnecessary provided the jobs are correctly numbered in
order of decreasing profit.
Problem 3.3.4. Verify that the algorithm works, and show that it requires
quadratic time in the worst case.
A second, more efficient algorithm is obtained if we use a different technique to

verify whether a given set of jobs is feasible.
Lemma 3.3.2. A set of jobs J is feasible if and only if we can construct a

feasible sequence including all the jobs in J as follows : for each job i e J, execute i at
time t, where t is the largest integer such that 0 < t <_ min(n, di) and the job to be
executed at time t has not yet been decided.
In other words, consider each job i e J :n turn, and add it to the schedule being
built as late as possible, but no later than its deadline. If a job cannot be executed in
time for its deadline, then the set J is infeasible. It may happen that there are gaps in
the schedule thus constructed when no job is to be executed. This obviously does not
affect the feasibility of a set of jobs. If we so wish, the schedule can be compressed
once all the jobs are included.
Problem 3.3.5. Prove Lemma 3.3.2. 11
The lemma suggests that we should consider an algorithm that tries to fill one by
one the positions in a sequence of length l = min(n, max({ di 1 <_ i <_ n 1)). For any
I
position t, define n, = max{k <_ t I position k is free }. Also define certain sets of
positions : two positions i and j are in the same set if ni = nj (see Figure 3.3.3). For a
given set K of positions, let F (K) be the smallest member of K. Finally, define a ficti-
tious position 0, which is always free.
Clearly, as we assign new jobs to vacant positions, these sets will merge to form
larger sets ; disjoint set structures are intended for just this purpose. We obtain an
algorithm whose essential steps are the following :
Positions of the same set
n; = nj i j
Free
position
Occupied
position
Figure 3.3.3. Sets of positions.

i. initialization : each position 0, 1, 2, ... , 1 is in a different set and

F({i})=i,0<-i <<-l;
ii. addition of a job with deadline d :
find the set that contains min(n, d ); let this be set K ;
if F (K) = 0, reject the job ;
ifF(K) 0,
- assign the job to position F (K) ;
- find the set that contains F (K) - 1; let this be set L (it is necessarily
different from K) ;
- merge K and L ; the value of F for this new set is the old value of
F (L).
Example 3.3.1. Consider a problem involving six jobs :
i 1 2 3 4 5 6
gi 20 15 10 7 5 3
d, 3 1 1 3 1 3.
Figures 3.3.4 and 3.3.5 illustrate the workings of the slow and fast algorithms, respec-
tively.
Here is a more precise statement of the fast algorithm. To simplify the descrip-
tion, we have assumed that the label of the set produced by a merge operation is
necessarily the label of one of the sets that were merged.
function sequence 2(d [ i n ]) : k, array[1 .. k ]
array j, F [0.. 1 ]
1 -min(n,max{d[i]I1<i <n1)
{ initialization)
for i F- 0 to l do j [i] E- 0
F[i] - i
initialize set { i }
{greedy loop }
for i F- 1 to n do { in decreasing order of g }
k - find (min(n, d [i]))
m - F [k]
if m * 0 then
j[m]F- i
1 <--find (m -1)
F[k] F-F[1]
merge (k, 1 )
{ it remains to compress the solution }
k -0
for i F- I to 1 do
if j[i]>0then k-k+1
j[k] E- j[i]
return k, j[l ..k]
3 dU[il]
Initialization: Ad
k
Try 2: 2
Try 3: unchanged
3 3
Try 4: 2 4
Try 5: unchanged
Try 6: unchanged
Optimal sequence: 2, 1, 4; value = 42
Figure 3.3.4. Illustration of the slow algorithm.
If the instance is given to us with the jobs already ordered by decreasing profit,
so that an optimal sequence can be obtained merely by calling the preceding algorithm,
most of the time will be spent manipulating disjoint sets. Since there are at most n+1
find operations and I merge operations to execute, and since n > 1, the required time is
in 0 (n lg*1), which is essentially linear. If, on the other hand, the jobs are given to us
in arbitrary order, so that we have to begin by sorting them, we need a time in
0 (n log n) to obtain the initial sequence.
3.4 GREEDY HEURISTICS
Because they are so simple, greedy algorithms are often used as heuristics in situations
where we can (or must) accept an approximate solution instead of an exact optimal
solution. We content ourselves with giving two examples of this technique. These
examples also serve to illustrate that the greedy approach does not always yield an
optimal solution.
Sec. 3.4 Greedy Heuristics 101
Initialization: I = min(6, max(d)) = 3
F= 0 1 2 3
Try 1: d, = 3, assign task 1 to position 3

F= 0 1 2
Fry 2: d2 = 1, assign task 2 to position 1
F= 0 2
Try 3: d3 = 1, no free position available since the Fvalue is 0
Try 4: d4 = 3, assign task 4 to position 2
F= 0
Try 5: d5 = 1, no free position available
Try 6: d6 = 3, no free position available
Optimal sequence: 2, 4, 1; value = 42
Figure 3.3.5. Illustration of the fast algorithm.
3.4.1 Colouring a Graph
Let G = < N, A > be an undirected graph whose nodes are to be coloured. If two
nodes are joined by an edge, then they must be of different colours. Our aim is to use
as few different colours as possible. For instance, the graph in Figure 3.4.1 can be
coloured using only two colours: red for nodes 1, 3 and 4, and blue for nodes 2 and 5.
An obvious greedy algorithm consists of choosing a colour and an arbitrary
starting node, and then considering each other node in turn, painting it with this colour
if possible. When no further nodes can be painted, we choose a new colour and a new
Figure 3.4.1. A graph to be coloured.
starting node that has not yet been painted, we paint as many nodes as we can with this
second colour, and so on.
In our example if node 1 is painted red, we are not allowed to paint node 2 with
the same colour, nodes 3 and 4 can be red, and lastly node 5 may not be painted. If we
start again at node 2 using blue paint, we can colour nodes 2 and 5 and finish the job
using only two colours ; this is an optimal solution. However, if we systematically
consider the nodes in the order 1, 5, 2, 3, 4, we get a different answer: nodes I and 5
are painted red, then node 2 is painted blue, but now nodes 3 and 4 require us to use a
third colour; in this case the result is not optimal.
The algorithm is therefore no more than a heuristic that may possibly, but not
certainly, find a "good" solution. Why should we be interested by such algorithms?
For the colouring problem and many others the answer is that all the exact algorithms
known require exponential computation time. This is an example of the NP-complete
problems that we shall study in Chapter 10. For a large-scale instance these algorithms
cannot be used in practice, and we are forced to make do with an approximate method
(of which this greedy heuristic is among the least effective).
Problem 3.4.1. For a graph G and an ordering a of the nodes of G, let c(G)
be the number of colours used by the greedy algorithm. Let C (G) be the optimal
(smallest) number of colours. Prove the following assertions :
i. (VG)(3(;)[c6(G)=c(G)],
R. (t/(XEIR+)(3G)(3o)[ca(G)1c(G) > a].
In other words, the greedy heuristic may find the optimal solution, but it may also give
an arbitrarily bad answer.
Problem 3:4.2. Find two or three practical problems that can be expressed in
terms of the graph colouring problem.
3.4.2 The Travelling Salesperson Problem
We know the distances between a certain number of towns. The travelling salesperson
wants to leave one of these towns, to visit each other town exactly once, and to arrive
back at the starting point, having travelled the shortest total distance possible. We
Sec. 3.4 Greedy Heuristics 103
assume that the distance between two towns is never negative. As for the previous
problem, all the known exact algorithms for this problem require exponential time (it is
also NP-complete). Hence they are impractical for large instances.
The problem can be represented using a complete undirected graph with n nodes.
(The graph can also be directed if the distance matrix is not symmetric: see Section
5.6.) One obvious greedy algorithm consists of choosing at each step the shortest
remaining edge provided that
i. it does not form a cycle with the edges already chosen (except for the very last
edge chosen, which completes the salesperson's tour) ;
ii. if chosen, it will not be the third chosen edge incident on some node.
For example, if our problem concerns six towns with the following distance matrix :
From To : 2 3 4 5 6
1 3 10 11 7 25
2 6 12 8 26
3 9 4 20
4 5 15
5 18
edges are chosen in the order (1,2), (3,5), (4,5), (2,3), (4,6), (1,6) to make the
circuit (1, 2, 3, 5, 4, 6, 1) whose total length is 58. Edge (1, 5 ), for example, was
not kept when we looked at it because it would have completed a circuit (1, 2, 3, 5,
1), and also because it would have been the third edge incident on node 5. In this
instance the greedy algorithm does not find an optimal tour since the tour (1, 2, 3, 6,
4, 5, 1) has a total length of only 56.
Problem 3.4.3. What happens to this greedy algorithm if the graph is not
complete, that is, if it is not possible to travel directly between certain pairs of towns ?
Problem 3.4.4. In some instances it is possible to find a shorter optimal tour

if the salesperson is allowed to pass through the same town several times. Give an
explicit example illustrating this. On the other hand, a distance matrix is said to be
Euclidean if the triangle inequality holds : for any towns i, j, and k, it is true that
distance (i , j) <_ distance (i , k) + distance (k , j ). Show that in this case it is never
advantageous to pass through the same town several times.
Problem 3.4.5. Give a heuristic greedy algorithm to solve the travelling

salesperson problem in the case when the distance matrix is Euclidean. Your algo-
rithm must find a solution whose length is not more than double the length of an
optimal tour. (Hint : start by constructing a minimal spanning tree and then use the
answer to the previous problem.)
Problem 3.4.6. Invent a heuristic greedy algorithm for the case when the dis-
tance matrix is not symmetric.
Problem 3.4.7. There exists a greedy algorithm (which perhaps would better
be called "abstinent") for solving the problem of the knight's tour on a chessboard : at
each step move the knight to the square that threatens the least possible number of
squares not yet visited. Try it !
Problem 3.4.8. In a directed graph a path is said to be Hamiltonian if it
passes exactly once through each node of the graph, but without coming back to the
starting node. Prove that if a directed graph is complete (that is, if each pair of nodes
is joined in at least one direction) then it has a Hamiltonian path, and give an algorithm
for finding such a path in this case.
A discussion of topics connected with Problem 3.1.1 can be found in Wright (1975)
and Chang and Korsh (1976) ; see also Problem 5.8.5 of this book.
The problem of minimal spanning trees has a long history, which is discussed in
Graham and Hell (1985). The first algorithm proposed (which we have not described)
is due to Boruvka (1926). The algorithm to which Prim's name is attached was
invented by Jarnik (1930) and rediscovered by Prim (1957) and Dijkstra (1959).
Kruskal's algorithm comes from Kruskal (1956). Other more sophisticated algorithms
are described in Yao (1975), Cheriton and Tarjan (1976), and Tarjan (1983).
The implementation of Dijkstra's algorithm that takes a time in 0 (n 2) is from
Dijkstra (1959). The details of the improvement suggested in Problem 3.2.12 can be
found in Johnson (1977). Similar improvement for the minimal spanning tree problem
(Problem 3.2.13) is from Johnson (1975). Faster algorithms for both these problems
are given in Fredman and Tarjan (1984); in particular, use of the Fibonacci heap
allows them to implement Dijkstra's algorithm in a time in 0 (a + n log n ). Other
ideas concerning shortest paths can be found in Tarjan (1983).
The solution to Problem 3.4.5 is given in Christofides (1976) ; the same reference
gives an efficient heuristic for finding a solution to the travelling salesperson problem
with a Euclidean distance matrix that is not more than 50% longer than the optimal
tour.
An important greedy algorithm that we have not discussed is used to derive
optimal Huffman codes; see Schwartz (1964). Other greedy algorithms for a variety of
problems are described in Horowitz and Sahni (1978).
4
Divide-and-Conquer
4.1 INTRODUCTION
Divide-and-conquer is a technique for designing algorithms that consists of decom-

posing the instance to be solved into a number of smaller subinstances of the same
problem, solving successively and independently each of these subinstances, and then
combining the subsolutions thus obtained in such a way as to obtain the solution of the
original instance. The first question that springs to mind is naturally "How should we
solve the subinstances ?". The efficiency of the divide-and-conquer technique lies in
the answer to this question.
Suppose you already have some algorithm A that requires quadratic time. Let c
be a constant such that your particular implementation requires a time tA (n) <_ cn2 to
solve an instance of size n. You discover that it would be possible to solve such an
instance by decomposing it into three subinstances each of size Fn/21, solving these
subinstances, and combining the results. Let d be a constant such that the time needed
to carry out the decomposition and the recombination is t(n) <_ dn. By using both your
old algorithm and your new idea, you obtain a new algorithm B whose implementation
takes time
tB(n)=3tA([n/21)+t(n)<_3c((n+l)/2)2+dn =4cn2+(4c+d)n+4c
The term 4cn 2 dominates the others when n is sufficiently large, which means that
algorithm B is essentially 25% faster than algorithm A. Although this improvement is
not to be sneezed at, nevertheless, you have not managed to change the order of the
time required : algorithm B still takes quadratic time.
105
106 Divide-and-Conquer Chap. 4
To do better than this, we come back to the question posed in the opening para-
graph : how should the subinstances be solved? If they are small, it is possible that
algorithm A may still be the best way to proceed. However, when the subinstances are
sufficiently large, might it not be better to use our new algorithm recursively? The
idea is analogous to profiting from a bank account that compounds interest payments !
We thus obtain a third algorithm C whose implementation runs in time
tA(n) ifn <no
tc (n) _
3tc( [ n/2 1 )+t(n) otherwise
where no is the threshold above which the algorithm is called recursively. This equa-
tion, which is similar to the one in Example 2.3.10, gives us a time in the order of ntg3
which is approximately n 1.59 The improvement compared to the order of n 2 is there-
fore quite substantial, and the bigger n is, the more this improvement is worth having.
We shall see in the following section how to choose no in practice. Although this
choice does not affect the order of the execution time of our algorithm, we are also
concerned to make the hidden constant that multiplies nlg3 as small as possible.
Here then is the general outline of the divide-and-conquer method:
function DQ (x)
{ returns a solution to instance x )
if x is sufficiently small or simple then return ADHOC (x)
decompose x into smaller subinstances x 1, x 2, , xk
fori - l tokdoy1 E-DQ(x,)
recombine the y, 's to obtain a solution y for x
return y ,
where ADHOC, the basic subalgorithm, is used to solve small instances of the
problem in question.
The number of subinstances, k, is usually both small and also independent of the
particular instance to be solved. When k = 1, it is hard to justify calling the technique
divide-and-conquer, and in this case it goes by the name of simplification (see sections
4.3, 4.8, and 4.10). We should also mention that some divide-and-conquer algorithms
do not follow the preceding outline exactly, but instead, they require that the first
subinstance be solved even before the second subinstance is formulated (Section 4.6).
For this approach to be worthwhile a number of conditions are usually required :
it must be possible to decompose an instance into subinstances and to recombine the
subsolutions fairly efficiently, the decision when to use the basic subalgorithm rather
than to make recursive calls must be taken judiciously, and the subinstances should be
as far as possible of about the same size.
After looking at the question of how to choose the optimal threshold, this chapter
shows how divide-and-conquer is used to solve a variety of important problems and
how the resulting algorithms can be analysed. We shall see that it is sometimes pos-
sible to replace the recursivity inherent in divide-and-conquer by an iterative loop.
When implemented in a conventional language such as Pascal on a conventional
Sec. 4.2 Determining the Threshold 107
machine, an iterative algorithm is likely to be somewhat faster than the recursive ver-
sion, although only by a constant multiplicative factor. On the other hand, it may be
possible to save a substantial amount of memory space in this way : for an instance of
size n, the recursive algorithm uses a stack whose depth is often in 1 (log n) and in
bad cases even in S2(n).
4.2 DETERMINING THE THRESHOLD
An algorithm derived by divide-and-conquer must avoid proceeding recursively when

the size of the subinstances no longer justifies this. In this case, it is better to apply the
basic subalgorithm. To illustrate this, consider once again algorithm C from the pre-
vious section, whose execution time is given by
f tA(n) ifn <no

tc (n) _
3tc ([n / 21) + t (n) otherwise,
where to (n) is the time required by the basic subalgorithm, and t (n) is the time taken
to do the decomposition and recombination. To determine the value of the threshold
no that minimizes tc (n), it is not sufficient to know that to (n) E O(n 2) and that
t (n) E O(n).
For instance, consider an implementation for which the values of to (n) and t(n)
are given respectively by n2 and 16n milliseconds. Suppose we have an instance of
size 1024 to solve. If the algorithm proceeds recursively until it obtains subinstances
of size 1, that is, if no = 1, it takes more than half an hour to solve this instance. This
is ridiculous, since the instance can be solved in little more than a quarter of an hour
by using the basic subalgorithm directly, that is, by setting no = oo. Must we conclude
that divide-and-conquer allows us to go from a quadratic algorithm to an algorithm
whose execution time is in 0(n 1g 3 ), but only at the cost of an increase in the hidden
constant so enormous that the new algorithm is never economic on instances that can
be solved in a reasonable time? Fortunately, the answer is no : in our example, the
instance of size 1024 can be solved in less than 8 minutes, provided we choose the
threshold no intelligently.
Problem 4.2.1. Prove that if we set no= 2k for some given integer k >_ 0,
then for all 1 >_ k the implementation considered previously takes
2k 31-k (32+2k) -21 +5 milliseconds
to solve an instance of size 21.
Problem 4.2.2. Find all the values of the threshold that allow an instance of
size 1024 to be solved in less than 8 minutes. 0
This example shows that the choice of threshold can have a considerable
influence on the efficiency of a divide-and-conquer algorithm. Choosing the threshold
is complicated by the fact that the best value does not generally depend only on the
algorithm concerned, but also on the particular implementation. Moreover, the
preceding problem shows that, over a certain range, changes in the value of the thres-
hold may have no effect on the efficiency of the algorithm when only instances of
some specific size are considered. Finally, there is in general no uniformly best value
of the threshold : in our example, a threshold larger than 66 is optimal for instances of
size 67, whereas it is best to use a threshold between 33 and 65 for instances of size
66. We shall in future abuse the term "optimal threshold" to mean nearly optimal.
So how shall we choose no? One easy condition is that we must have no > 1 to
avoid the infinite recursion that results if the solution of an instance of size 1 requires
us first to solve a few other instances of the same size. This remark may appear trivial,
but Section 4.6 describes an algorithm for which the ultimate threshold is less obvious,
as Problem 4.6.8 makes clear.
Given a particular implementation, the optimal threshold can be determined
empirically. We vary the value of the threshold and the size of the instances used for
our tests and time the implementation on a number of cases. Obviously, we must
avoid thresholds below the ultimate threshold. It is often possible to estimate an
optimal threshold simply by tabulating the results of these tests or by drawing a few
diagrams. Problem 4.2.2 makes it clear, however, that it is not usually enough simply
to vary the threshold for an instance whose size remains fixed. This approach may
require considerable amounts of computer time. We once asked the students in an
algorithmics course to implement the algorithm for multiplying large integers given in
Section 4.7, in order to compare it with the classic algorithm from Section 1.1. Several
groups of students tried to estimate the optimal threshold empirically, each group using
in the attempt more than 5,000 (1982) Canadian dollars worth of machine time ! On
the other hand, a purely theoretical calculation of the optimal threshold is rarely pos-
sible, given that it varies from one implementation to another.
The hybrid approach, which we recommend, consists of determining theoretically
the form of the recurrence equations, and then finding empirically the values of the
constants used in these equations for the implementation at hand. The optimal thres-
hold can then be estimated by finding the value of n at which it makes no difference,
for an instance of size n, whether we apply the basic subalgorithm directly or whether
we go on for one more level of recursion.
Coming back to our example, the optimal threshold can be found by solving
to (n) = 3tA ([n / 21) + t (n), because tc ([n / 21) = to ([n / 21) if [n / 21 <_ no. The pres-
ence of a ceiling in this equation complicates things. If we neglect this difficulty, we
obtain n = 64. On the other hand, if we systematically replace In / 21 by (n + 1)/2, we
find n = 70. There is nothing surprising in this, since we saw in Problem 4.2.2 that in
fact no uniformly optimal threshold exists. A reasonable compromise, corresponding
to the fact that the average value of Fn / 21 is (2n + 1)/4, is to choose n o = 67 for our
threshold.
Sec. 4.3 Binary Searching 109
* Problem 4.2.3. Show that this choice of no = 67 has the merit of being subop-
timal for only two values of n in the neighbourhood of the threshold. Furthermore,
prove that there are no instances that take more than I % longer with threshold 67 than
they would with any other threshold.
The following problem shows that a threshold of 64 would be optimal were it

always possible to decompose an instance of size n into three subinstances exactly of
size n / 2, that is, if instances of fractional size were allowed. (Notice that this would
not cause an infinite recursion because the threshold is strictly larger than zero.)
Problem 4.2.4. Let a and b be real positive constants. For each positive real
number s, consider the function f5 : IR* -- 1R* defined by the recurrence
x2 if X <_ s
f(x)= 3fs (x / 2) + bx otherwise.
Prove by mathematical induction that if u = 4b /a and if v is an arbitrary positive real
number, then fu (x) 5 f, (x) for every real number x. Notice that this u is chosen so
that au 2 = 3a (u / 2)2 + bu . (For purists : even if the domain of f, and f is not count-
able, the problem can be solved without recourse to transfinite induction, precisely
because infinite recursion is not a worry.)
In practice, one more complication arises. Supposing, for instance, that to (n) is
quadratic, it may happen that tA (n) = an 2 + bn + c for some constants a, b, and c
depending on the implementation. Although bn + c becomes negligible compared to
an 2 when n is large, the basic subalgorithm is used in fact precisely on instances of
moderate size. It is therefore usually insufficient merely to estimate the constant a.
Instead, measure to (n) a number of times for several different values of n , and then
estimate all the necessary constants, probably using a regression technique.
4.3 BINARY SEARCHING
Binary searching predates computers. In essence, it is the algorithm used to look up a

word in a dictionary or a name in a telephone directory. It is probably the simplest
application of divide-and-conquer.
Let T [I .. n] be an array sorted into increasing order; that is,
1 < j 5 n = T [i] <- T [ j ], and let x be some item. The problem consists of
finding x in the array T if indeed it is there. If the item we are looking for is not in the
array, then instead we want to find the position where it might be inserted. Formally,
we wish to find the index i such that 0 <_ i <_ n and T [i] S x < T [i + I], with the log-
ical convention that T [0] _ - oo and T [n + 1] _ + oo. (By logical convention, we mean
that these values are not in fact present as sentinels in the array.) The obvious
approach to this problem is to look sequentially at each element of T until we either

come to the end of the array or find an item bigger than x.
function sequential (T [ 1 .. n ], x)
[ sequential search for x in array T [
for i F- I to n do
if T[i] >x then return i -1
return n
This algorithm clearly takes a time in O(1 + r), where r is the index returned : this is
S2(n) in the worst case and O (1) in the best case. If we assume that all the elements of
T are distinct, that x is indeed somewhere in the array, and that it is to be found with
equal probability at each possible position, then the average number of trips round the
loop is (n 2+ 3n -2)/2n. On the average, therefore, as well as in the worst case,
sequential search takes a time in O(n).
To speed up the search, divide-and-conquer suggests that we should look for x
either in the first half of the array or in the second half. To find out which of these
searches is appropriate, we compare x to an element in the middle of the array : if
x < T [ 1+ Ln / 2j], then the search for x can be confined to T [I .. [n12]]; otherwise it
is sufficient to search T [1+Ln / 2j .. n J. We obtain the following algorithm.
function binsearch (T [ 1 .. n ], x)
f binary search for x in array T [
if n = 0 or x < T [1] then return 0
return binrec (T, x)
function binrec (T [i .. j ], x)
[ binary search for x in subarray T [i .. j
this procedure is only called if T[i] S x < T [j + 1] and i j[
if i = j then return i
k -(i+j+1)div2
if x < T [k] then return binrec (T [i .. k -1 ], x)
else return binrec (T [k .. j ],x)
Problem 4.3.1. Prove that the function binrec is never called on T [i .. j ]

with j < i . Prove too that when binrec (T [i .. j ], x) makes a recursive call
binrec (T [u .. v ], x), it is always true that v - u < j - i. Conclude from these two
results that a call on binsearch always terminates. Show finally that the values T [0]
and T [n + 1] are never used (except in the comments ! ).
Problem 4.3.2. Show that the algorithm takes a time in 0( log n) to find x in
T [1 .. n ] whatever the position of x in T.
The algorithm in fact executes only one of the two recursive calls, so that techni-
cally it is an example of simplification rather than of divide-and-conquer. Because the
recursive call is situated dynamically at the very end of the algorithm, it is easy to pro-
duce an iterative version.
function iterbin (T [1 .. n],x)

( iterative binary search for x in array T }
iE--1;j<-- n
while i < j do
( T [ i ] Sx <T [j+1] }
k <-- (i+j+1)div2
if x < T [k] then j <-- k-1
else i <- k
return i
Problem 4.3.3. It is easy to go wrong when programming the concept of

binary searching, simple though this is. Show by examples that the preceding algo-
rithm would be incorrect if we replaced
i. "k F(i+j+1)div2" by " k <-- (i+j)div2",

ii. "i k" by " i E--k+l", or
iii. "j F- k -1" by " j <-- k ". 11
A first inspection of this algorithm shows what is apparently an inefficiency.

Suppose T contains 17 distinct elements and that x = T [13]. On the first trip round
the loop, i = 1, j = 17, and k = 9. The comparison between x and T [9] causes the
assignment i - 9 to be executed. On the second trip round the loop i = 9, j = 17,
and k = 13. A comparison is made between x and T [13]. This comparison could
allow us to end the search immediately, but no test is made for equality, and so the
assignment i <- 13 is carried out. Two more trips round the loop are necessary before
we leave with i = j = 13. The following algorithm leaves the loop immediately after
we find the element we are looking for.
function iterbin 2(T [1 .. n ],x)
( variant on iterative binary search }
iE- l;j'- n
while i < j do
(T[i] <-x <T [j+1]}
k <-(i+j)div2
case x< T [k] : k-1
x?T[k+1]:i -- k+1
otherwise : i , j <-- k
return i
Which of these algorithms is better ? The first systematically makes a number of
trips round the loop in O(log n), regardless of the position of x in T, while it is pos-
sible that the variant will only make one or two trips round the loop if x is favourably
situated. On the other hand, a trip round the loop in the variant will take a little longer
to execute on the average than a trip round the loop in the first algorithm. To compare
them, we shall analyse exactly the average number of trips round the loop that each
version makes. Suppose to make life simpler that T contains n distinct elements and
that x is indeed somewhere in T, occupying each possible position with equal proba-
bility. Let A (n) and B (n) be the average number of trips round the loop made by the
first and the second iterative versions, respectively.
Analysis of the First Version. Let k = 1 + Ln / 2j. With probability (k -1)/n ,
x < T [k], which causes the assignment j - k - 1, after which the algorithm starts
over on an instance reduced to k -1 elements. With probability I - (k - 1)/n,
x > T [k], which causes the assignment i E- k, after which the algorithm starts over on
an instance reduced to n -k + 1 elements. One trip round the loop is carried out before
the algorithm starts over, so the average number of trips round the loop is given by the
recurrence
A(n)=1+ Ln/2]A(Ln/2])+ rn/21A(Ln/21), n?2

A(1)=0.
Analysis of the Second Version. In a similar way, taking k = In / 21, we
obtain the recurrence
Fn/ 21
B(n)= 1+ - IB(Fn/21-1)+ Ln/2,B(Ln/2j), n >-3
B(1)=0, B(2)=I.
Define a (n) and b (n) as n A (n) and n B (n), respectively. The equations then
become
a(n)=n +a(Ln/2])+a([n/2l), n >-2

b(n)=n +b([n/21-1)+b(Ln/2j), n >-3 (*)
a(1)=b(1)=0, b(2)=2.
The first equation is easy in the case when n is a power of 2, since it then
reduces to
a(n)=2a(n/2)+n, n ?2
a(1)=0,
which yields a (n) = n lg n using the techniques of Section 2.3. Exact analysis for
arbitrary n is harder. We proceed by constructive induction, guessing the likely form
of the answer and determining the missing parameters in the course of a tentative proof
by mathematical induction. A likely hypothesis, already shown to hold when n is a
power of 2, is that n Llg n j <- a (n) <- n Fig n 1. What might we add to n Llg n j
to arrive at a (n) ? Let n * denote the largest power of 2 that is less than or
equal to n. In particular, Llg n] = lg n *. It seems reasonable to hope that
a (n) = n lg n * + cn + do * + e lg n * +f for appropriate constants c, d, e, and f .

Denote this hypothesis by HI (n).
When n > 1 is not of the form 2'-1, then (Ln / 2j)* _ (rn / 21)* = n */ 2. To
prove HI(n) in this case, using the recurrence (*), HI (Ln / 2]) and HI (rn / 2 1 ), it is
thus necessary and sufficient that
n lgn*+cn +dn*+e lgn*+f = n lgn*+cn +dn*+2e lgn*+(2f -2e)
that is, we need e = 0 and f = 0, c and d being still unconstrained. If our hypothesis
is correct, we therefore know
a (n) = n lg n * + cn + do * .
When n > 1 is of the form 2'-1, then Ln / 2] _ (n -1)/2, (Ln / 2] )* =(n+1)14 and
Fn/21 = (rn / 21)* = n * = (n + 1)/2. To prove HI(n) in this case, it is necessary and
sufficient that
n1gn+1
2
+(c+d)n+d
2 2
= nlgn+l
2
+(c+3d+!)n+(3d+
4 2 4
i);
2
that is
4c + 2d = 4c + 3d + 2 and 2d = 3d+2.
These two equations are not linearly independent. They allow us to conclude that
d = - 2, there still being no constraints on c.
At this point we know that if only we can make the hypothesis
a (n) = n lg n * + cn - 2n *
true for the base n = 1, then we shall have proved by mathematical induction that it is
true for every positive integer n. Our final constraint is therefore
0=a(1)=c -2,
which gives c = 2 and implies that the general solution of the recurrence (*) for a (n)
is
a(n)=nlgn*+2(n-n*).
The average number of trips round the loop executed by the first iterative algorithm for
binary searching, when looking for an element that is in fact present with uniform
probability distribution among the n different elements of an array sorted into
increasing order, is given by A (n) = a (n)l n , that is
A (n) = Llg n] + 2(1- n */ n) .
In particular, our initial guess holds :

Llg nJ <- A (n) <- rig n I
Problem 4.3.4. A seemingly simpler approach to determine the constants

would be to substitute the values 1, 2, 3, and 4 for n in the hypothesis HI (n), thus
obtaining four linear equations in four unknowns, which can easily be solved to give
the same c, d, e, and f . Explain why this is insufficient.
Problem 4.3.5. Using the techniques presented in Section 2.3, solve equation
(*) exactly for b (n) when n is of the form 2' -1.
The general solution of the recurrence for b (n) is more difficult than the one we
have just obtained. It seems reasonable to formulate the same incompletely specified
induction hypothesis : b (n) = n lg n * + cn + do * + e lg n * + f for some new con-
stants c, d, e , and f . Constructive induction yields e = 1 and f = 1 + c to take
account of the case when n >- 3 is not a power of 2. The case when n ? 4 is a power
of 2 obliges us to choose d = - 2. The hypothesis therefore becomes
b (n) = n lg n * + cn - 2n* + lg n * + (1 +c). Unfortunately, b (1) = 0 = c = 2 ,
whereas b (2) = 2 c=3, which is inconsistent and shows that the original
hypothesis was wrong. The problem arises because the two basic cases are incompa-
tible in the sense that it is impossible to obtain them both from the recurrence starting
from some artificial definition of b (0). Nonetheless, our efforts were not entirely in
vain. A simple modification of the argument allows us to conclude that
nlgn*+n/2-2n*+lgn*+3/2 <- b(n) <- nlgn*+2n/3-2n*+lgn*+5/3.
Stated more elegantly :
* Problem 4.3.6. Show that there exists a function it : IN+ 1N+ such that
(n + 1)/3 <- n(n) <- (n + 1)/2 for every positive integer n, and such that the exact solu-
tion of the recurrence is b (n) = n lg n * + n - 2n * + lg n * + 2 - n(n).
* Problem 4.3.7. Show that the function n(n) of the previous exercise is given
by
n */ 2 if 2n < 3n *
lt(n -1) =
jl n-n * otherwise
for all n > 2. Equivalently
lt(n -1) = [n * + On / n *j -2)(2n -3n*)] / 2
We are finally in a position to answer the initial question : which of the two algo-
rithms for binary searching is preferable? By combining the preceding analysis of the
function a (n) with the solution to Problem 4.3.6, we obtain
lt(n)-LngnJ-2 3
A(n)-B(n)=1+
Thus we see that the first algorithm makes on the average less than one and a half trips
round the loop more than the second. Given that the first algorithm takes less time on
the average than the variant to execute one trip round the loop, we conclude that the
Sec. 4.4 Sorting by Merging 115
first algorithm is more efficient than the second on the average whenever n is
sufficiently large. The situation is similar if the element we are looking for is not in
fact in the array. However, the threshold beyond which the first algorithm is preferable
to the variant can be very high for some implementations.
4.4 SORTING BY MERGING
Let T [I .. n ] be an array of n elements for which there exists a total ordering. We are
interested in the problem of sorting these elements into ascending order. We have
already seen that the problem can be solved by selection sorting and insertion sorting
(Section 1.4), or by heapsort (Example 2.2.4 and Problem 2.2.3). Recall that an
analysis both in the worst case and on the average shows that the latter method takes a
time in O(n log n), whereas both the former methods take quadratic time.
The obvious divide-and-conquer approach to this problem consists of separating
the array T into two parts whose sizes are as nearly equal as possible, sorting these
parts by recursive calls, and then merging the solutions for each part, being careful to
preserve the order. We obtain the following algorithm :
procedure mergesort (T [ 1 .. n ])
{ sorts array T into increasing order }
if n is small then insert (T)
else arrays U [ 1 .. n div 2], V [ 1 .. (n + 1) div 2]
U F T[ 1 .. n div 2]
V -T[1+(n div 2) .. n]
mergesort (U) ; mergesort (V )
merge (T,U,V),
where insert (T) is the algorithm for sorting by insertion from Section 1.4, and
merge (T, U, V) merges into a single sorted array T two arrays U and V that are
already sorted.
Problem 4.4.1. Give an algorithm capable of merging two sorted arrays U

and V in linear time, that is, in a time in the exact order of the sum of the lengths of U
and V.
** Problem 4.4.2. Repeat the previous problem, but without using an auxiliary
array : the sections T [I .. k ] and T [k + 1 .. n ] of an array are sorted independently,
and you wish to sort the whole array T [I .. n I. You may only use a fixed number of
working variables to solve the problem, and your algorithm must work in linear time.
This sorting algorithm is a good illustration of all the facets of divide-and-

conquer. When the number of elements to be sorted is small, a relatively simple algo-
rithm is used. On the other hand, when this is justified by the number of elements,
mergesort separates the instance into two subinstances half the size, solves each of
these recursively, and then combines the two sorted half-arrays to obtain the solution to
the original instance.
Let t(n) be the time taken by this algorithm to sort an array of n elements.
Separating T into U and V takes linear time. By the result of Problem 4.4.1, the final
merge also takes linear time. Consequently, t(n) E t(Ln / 2j) + t (Fn / 21) + 0(n). This
equation, which we analysed in Section 2.1.6, allows us to conclude that the time
required by the algorithm for sorting by merging is in O(n log n).
Problem 4.4.3. Rather than separate T into two half-size arrays, we might
choose to separate it into three arrays of size Ln/3], L(n + 1)/3j, and L(n + 2)/3],
respectively, to sort each of these recursively, and then to merge the three sorted
arrays. Give a more formal description of this algorithm, and analyse its execution
time.
** Problem 4.4.4. Following up the previous problem, we might choose to

separate T into about LI arrays, each containing approximately L' elements.
Develop this idea, and analyse its performance.
The merge sorting algorithm we gave, and those suggested by the two previous
problems, have two points in common. The fact that the sum of the sizes of the subin-
stances is equal to the size of the original instance is not typical of algorithms derived
using divide-and-conquer, as we shall see in several subsequent examples. On the
other hand, the fact that the original instance is divided into subinstances whose sizes
are as nearly as possible equal is crucial if we are to arrive at an efficient algorithm.
To see why, look at what happens if instead we decide to separate T into an array U
with n -1 elements and an array V containing only 1 element. Let t'(n) be the time
required by this variant to sort n items. We obtain t'(n) E t'(n - 1) + t'(1) + O(n).
Problem 4.4.5. Show that t'(n) E O(n 2).
Simply forgetting to balance the sizes of the subinstances can therefore be disas-
trous for the efficiency of an algorithm obtained using divide-and-conquer.
Problem 4.4.6. This poor sorting algorithm is very like one we have already
seen in this book. Which one, and why?
4.5 QUICKSORT
The sorting algorithm invented by Hoare, usually known as "quicksort", is also based
on the idea of divide-and-conquer. Unlike sorting by merging, the nonrecursive part of
the work to be done is spent constructing the subinstances rather than combining their
solutions. As a first step, this algorithm chooses one of the items in the array to be
sorted as the pivot. The array is then partitioned on either side of the pivot : elements
Sec. 4.5 Quicksort 117
are moved in such a way that those greater than the pivot are placed on its right,
whereas all the others are moved to its left. If now the two sections of the array on
either side of the pivot are sorted independently by recursive calls of the algorithm, the
final result is a completely sorted array, no subsequent merge step being necessary. To
balance the sizes of the two subinstances to be sorted, we would like to use the median
element as the pivot. (For a definition of the median, see Section 4.6.) Unfortunately,
finding the median takes more time than it is worth. For this reason we simply use the
first element of the array as the pivot. Here is the algorithm.
procedure quicksort (T [i .. j ])
{ sorts array T [i .. j ] into increasing order }
if j - i is small then insert (T [i j ]) { Section 1.4 }
else pivot (T [i .. j ],1)
{ after pivoting, i <- k < 1 T [k] 5 T [l]
and 1 <k 5j = T[k]>T[1]}
quicksort(T [i .. I -1])
quicksort (T [1 +1 .. j ])
Designing a linear time pivoting algorithm is no challenge. It is, however, cru-

cial in practice that the hidden constant be small. Let p = T [i ] be the pivot. One good
way of pivoting consists of scanning the array T [i .. j ] just once, but starting at both
ends. Pointers k and 1 are initialized to i and j + 1, respectively. Pointer k is then
incremented until T [k] >p, and pointer I is decremented until T [1] <- p. Now T [k]
and T [1] are interchanged. This process continues as long as k < 1. Finally, T [i] and
T [1] are interchanged to put the pivot in its correct position.
procedure pivot (T [i .. j ] ; var 1)
{ permutes the elements in array T [i .. j ] in such a way that, at the end,
i <- l <- j, the elements of T [i .. 1-1] are not greater than p,
T[11 =p, and the elements of T J1+1 .. j ] are greater than p,
where p is the initial value of T [i ] }
p -T[i]
kF-i;1<-j+1
repeat k- k + 1 until T [k] > p or k >- j
repeat I E- 1- 1 until T [1] <- p
while k p
repeat 1 f- 1- 1 until T [1] p
interchange T [i] and T [1]
Problem 4.5.1. Invent several examples representing different situations that

might arise, and simulate the pivoting algorithm on these examples.
Quicksort is inefficient if it happens systematically on most recursive calls that

the subinstances T [i .. I - 1] and T [1 + 1 .. j ] to be sorted are severely unbalanced.
Problem 4.5.2. Show that in the worst case quicksort requires quadratic time.
Give an explicit example of an array to be sorted that causes such behaviour.
On the other hand, if the array to be sorted is initially in random order, it is

likely that most of the time the subinstances to be sorted will be sufficiently well bal-
anced. To determine the average time required by quicksort to sort an array of n items,
we assume that all the elements of T are distinct and that each of the n! possible initial
permutations of the elements has the same probability of occurring. Let t (m) be the
average time taken by a call on quicksort (T [a + 1 .. a + m ]) for 0 <- m <- n and
0:5 a <- n -m. The pivot chosen by the algorithm is situated with equal probability in
any position with respect to the other elements of T. The value of I returned by the
pivoting algorithm after the initial call pivot (T [I .. n ],1) can therefore be any integer
between 1 and n, each value having probability 11n. This pivoting operation takes a
time in O(n). It remains to sort recursively two subarrays of size 1 -1 and n -1,
respectively. The average time required to execute these recursive calls is
t(I -1) + t(n -1). Consequently,
t(n) E O(n) + n (t(1-1) + t(n -1)) .

1=1
A little manipulation yields

n-1
t(n)eO(n)+ n t(k)
k =O
To make this more explicit, let d and no be two constants such that
n-1
t (n) <- do + n t (k) for n > n 0
k=0
An equation of this type is more difficult to analyse than the linear recurrences we saw
in Section 2.3. By analogy with sorting by merging, it is, nevertheless, reasonable to
hope that t(n) will be in 0 (n log n) and to apply constructive induction to look for a
constant c such that t (n) S cn lg n .
To use this approach we need an upper bound on i =no+ 1 i lg i. This is
obtained with the help of a simple lemma. (We suggest you find a graphical interpre-
tation of the lemma.) Let a and b be real numbers, a < b , and let f : [a, b ] - IR be
a nondecreasing function. Let j and k be two integers such that a <- j < k <- b. Then
k-1 k
f(i) < J f(x) dx .
In particular, taking f (x) = x lg x, j = no and k = n , we obtain

n-1 n
E i lg i <- f x lg x dx
i=no+1 x=no+1
Sec. 4.6 Selection and the Median 119
[-lx - Igex2
2 4 ,x=no+1
2
<
2 lgn - Ige n2
provided no > 1.
Problem 4.5.3. Complete the proof by mathematical induction that

t (n) <- cn lg n for all n > n o >_ 1, where
2d 4
c= + t (k) .
lge (no+l)2lge k=o
Quicksort can therefore sort an array of n distinct elements in an average time in

O (n log n). The hidden constant is in practice smaller than those involved in heapsort
or in merge sort. If an occasional long execution time can be tolerated, this is the
sorting algorithm to be preferred among all those presented in this book. The proba-
bility of suffering an execution time in U(n 2) can be greatly diminished, at the price of
a small increase in the hidden constant, by choosing as pivot the median of T [i],
T[(i+j)div21 andT[j].
Problem 4.5.4. Show by a simple argument that, whatever the choice of pivot,
quicksort as described here always takes a time in S2(n2) in the worst case. Outline a
modification to the algorithm to avoid this.
By combining the modification hinted at in the previous problem with the linear
algorithm from the following section, we can obtain a version of quicksort that takes a
time in 0 (n log n) even in the worst case. We mention this possibility only to point
out that it should be shunned : the hidden constant associated with the "improved" ver-
sion of quicksort is so large that it results in an algorithm worse than heapsort in every
case.
4.6 SELECTION AND THE MEDIAN
Let T [I .. n ] be an array of integers. What could be easier than to find the smallest
element or to calculate the mean of all the elements ? However, it is not obvious that
the median can be found so easily. Intuitively, the median of T is that element m in T
such that there are as many items in T smaller than m as there are items larger than m.
The formal definition takes care of the possibility that n may be even, or that the
elements of T may not all be distinct. Thus we define m to be the median of T if and
only if m is in T and
# I i E [ I . . n ] I T [ i ] < m } <n/2 and # { i E [ I . . n ] I T [ i ] < m f >_n/2.
The naive algorithm for determining the median of T consists of sorting the array
into ascending order and then extracting the In / 21th entry. If we use heapsort or
merge sort, this algorithm takes a time in 0 (n log n) to determine the median of n ele-
ments. Can we do better? To answer this question, we consider a more general
problem : selection. Let T be an array of n elements, and let k be an integer between
1 and n. The kth smallest element of T is that element m such that
# { i e [ l .. n i l T[i] <m } <k, whereas # { i e [ I .. n] I T [i ] <- m } >_ k . In other
words, it is the k th item in T if the array is sorted into ascending order. For instance,
the median of T is its In / 21th smallest element.
The following algorithm, which is not yet completely specified, solves the selec-
tion problem in a way suggested by quicksort.
function selection(T [ 1 .. n ], k)
{ finds the k th smallest element of T ;
this algorithm assumes that 1 k <_ n }
if n is small then sort T
return T RI
p - some element of T [I .. n ] { to be specified later)
u - # { i e [l .. n ] I T [i] < p }
v #{i a [1..n]IT[i]Sp }
if k S u then
array U[1 .. u ]
U F the elements of T smaller than p
{ the k th smallest element of T is
also the k th smallest element of U }
return selection(U, k)
if k <- v then { got it ! } return p
otherwise { k > v }
array V[1 .. n -v ]
V - the elements of T larger than p
{ the k th smallest element of T is
also the (k-v )th smallest of V }
return selection(V, k -v )
Problem 4.6.1. Generalize the notion of pivoting from Section 4.5 to partition
the array T into three sections, T [I .. i -1 ], T [i .. j], and T [j + 1 .. n ], containing
the elements of T that are smaller than p, equal to p, and greater than p, respectively.
The values i and j should be returned by the pivoting procedure, not calculated before-
hand. Your algorithm should scan T once only, and no auxiliary arrays should be
used.
Problem 4.6.2. Using ideas from the iterative version of binary searching seen
in Section 4.3 and the pivoting procedure of the previous problem, give a nonrecursive
version of the selection algorithm. Do not use any auxiliary arrays. Your algorithm is
allowed to alter the initial order of the elements of T.
Which element of T should we use as the pivot p ? The natural choice is surely
the median of T, so that the sizes of the arrays U and V will be as similar as possible
(even if at most one of these two arrays will actually be used in a recursive call).
Problem 4.6.3. What happens to the selection algorithm if the choice of p is

made using "p <- selection (T, (n + 1) div 2) "?
Suppose first that the median can be obtained by magic at unit cost. For the time
being, therefore, the algorithm works by simplification, not by divide-and-conquer. To
analyse the efficiency of the selection algorithm, notice first that, by definition of the
median, u < [n / 21 and v ? [n / 21. Consequently, n - v <- Ln / 2]. If there is a recur-
sive call, the arrays U and V therefore contain a maximum of Ln / 2] elements. The
remaining operations, still supposing that the median can be obtained magically, take a
time in O (n). Let tm (n) be the time required by this method in the worst case to find
the k th smallest element of an array of at most n elements, independently of the value
of k. We have tm(n)E0(n)+max{ tm(i) I i 5 Ln/2] }.
Problem 4.6.4. Show that tm (n) is in O (n).
Thus we have an algorithm capable of finding the k th smallest element of an

array in a time linear in the size of the array. But what shall we do if there is no
magic way of getting the median? If we are willing to sacrifice speed in the worst
case in order to obtain an algorithm reasonably fast on tf.e average, we can once again
borrow an idea from quicksort and choose simply
p F-T[1].
When we do this, we hope that on the average the sizes of the arrays U and V will not
be too unbalanced, even if occasionally we meet an instance where u = 0, v = 1, and
k > 1, which causes a recursive call on n -1 elements.
* Problem 4.6.5. In this problem, consider the selection algorithm obtained by

choosing "p f- T [I] ". Assume the n elements of the array T are distinct and that
each of the n! permutations of the elements has the same probability of occurring. Let
E (n , k) stand for the expected size of the subarray involved in the first recursive call
produced by a call on selection (T [I .. n ], k), taking the size to be zero if there is no
recursive call. Prove that
E(n,k) < n + k (n-k) < 3n

2 n 4
Assuming that the pivoting algorithm preserves random ordering in the subarrays it
produces for the recursive calls, prove that this selection algorithm takes linear time on
the average, whatever the value of k. (The hidden constant must not depend on k.)
Note that the technique hinted at in this exercise only aprlies because the average time
turns out to be linear : the average time taken on several instances is not otherwise
equal to the time taken on an average instance.
Problem 4.6.6. Show, however, that in the worst case this algorithm requires
quadratic time.
This quadratic worst case can be avoided without sacrificing linear behaviour on
the average : the idea is to find quickly a good approximation to the median. This can
be done with a little cunning. Assuming n ? 5, consider the following algorithm :
function pseudomed (T [1 .. n ])
{ finds an approximation to the median of array T)
s - n div 5
array S [I .. s ]
for i E- 1 to s do S [i] t- adhocmed5(T [5i-4 .. 5i ])
return selection(S, (s + 1) div 2) ,
where adhocmed 5 is an algorithm specially designed to find the median of exactly five
elements. Note that the time taken by adhocmed5 is bounded above by a constant.
We look first at the value of the approximation to the median found by the algo-
rithm pseudomed. Let m be this approximation. Since m is the exact median of the
array S, we have
#{ie[l..s]IS[i]<-m }> Is/21.
But each element of S is the median of five elements of T. Consequently, for every i
such that S [i ] <- m , there are three i 1, i2, i3 between Si - 4 and 5i such that
T [i 1] <_ T [i 2] <_ T [i 3] = S [i ] m. Therefore
# { i E [ 1 .. n ] I T[i]m }>31s/21=31Ln/5]/21>(3n-12)/10.
Problem 4.6.7. Show similarly that

#{iE[I..n]IT[i]<m }<(7n-3)/10.
The conclusion is that although m is perhaps not the exact median of T, yet its
rank is approximately between 3n / 10 and 7n / 10. To visualize how these factors arise,
although nothing in the execution of the algorithm pseudomed really corresponds to
this illustration, imagine that all the elements of T are arranged in five rows, with the
possible exception of one to four elements left aside (Figure 4.6.1). Now suppose that
the middle row is sorted by magic, as is each of the Ln / 5] columns, the smallest ele-
ments going to the left and to the top, respectively. The middle row corresponds to the
array S in the algorithm. Similarly, the element in the circle corresponds to the median
of this array, that is, to the value of m returned by the algorithm. By the transitivity of
":5 ", each of the elements in the box is less than or equal to m. Notice that the box
contains approximately three-fifths of one-half of the elements of T, that is, about
3n / 10 elements.
We now look at the efficiency of the selection algorithm given at the beginning
of this section when we use
p E- pseudomed(T) .
Figure 4.6.1. Visualization of the pseudomedian.
Let n be the number of elements in T, and let t (n) be the time required in the worst
case by this algorithm to find the k th smallest element of T, still independently of
the value of k. At the first step, calculating pseudomed (T) takes a time in
O (n) + t( Ln / 5]), because the array S can be constructed in linear time. Calculating u
and v also takes linear time. Problem 4.6.7 and the preceding discussion show that
u <- (7n - 3)/ 10 and v >- (3n - 12)/10, so n -v <- (7n + 12)/10. The recursive call that
may follow therefore takes a time bounded above by
max { t (i) I i <- (7n + 12)/10 } .
The initial preparation of the arrays U and V takes linear time. Hence, there exists a
constant c such that
t(n)<-t(Ln/5j)+max{t(i)Ii <-(7n+12)/10}+cn
for every sufficiently large n.
Problem 4.6.8. We have to ensure that n > 4 in the preceding equation.

What is it in the algorithm that makes this restriction necessary?
This equation looks quite complicated. First let us solve a more general, yet
simpler problem of the same type.
* Problem 4.6.9. Let p and q be two positive real constants such that p + q < 1,
let no be a positive integer, and let b be some positive real constant. Let f : IN -* IR*
be any function such that
f(n) =.f (LpnJ) +.f (Lqn]) + bn
for every n > n 0. Use constructive induction to prove that f (n) E O(n). l7
Problem 4.6.10. Let t (n) be the time required in the worst case to find the k th
smallest element in an array of n elements using the selection algorithm discussed ear-
lier. Give explicitly a nondecreasing function f(n) defined as in Problem 4.6.9 (with
p =1/5 and q = 3/4) such that t(n) <_ f(n) for every integer n. Conclude that
t(n)eO(n). Argue that t(n)ES2(n), and thus t(n)EO(n).
In particular, it is possible to find the median of an array of n elements in linear

time. The version of the algorithm suggested by Problem 4.6.2 is preferable in prac-
tice, even though this does not constitute an iterative algorithm : it avoids calculating u
and v beforehand and using two auxiliary arrays U and V. To use still less auxiliary
space, we can also construct the array S (needed to calculate the pseudomedian) by
exchanging elements inside the array T itself.
4.7 ARITHMETIC WITH LARGE INTEGERS
In most of the preceding analyses we have taken it for granted that addition and multi-
plication are elementary operations, that is, the time required to execute these opera-
tions is bounded above by a constant that depends only on the speed of the circuits in
the computer being used. This is only reasonable if the size of the operands is such
that they can be handled directly by the hardware. For some applications we have to
consider very large integers. Representing these numbers in floating-point is not useful
unless we are concerned solely with the order of magnitude and a few of the most
significant figures of our results. If results have to be calculated exactly and all the
figures count, we are obliged to implement the arithmetic operations in software.
This was necessary, for instance, when the Japanese calculated the first 134 mil-
lion digits of It in early 1987. (At the very least, this feat constitutes an excellent
aerobic exercice for the computer !) The algorithm developed in this section is not,
alas, sufficiently efficient to be used with such operands (see Chapter 9 for more on
this). From a more practical point of view, large integers are of crucial importance in
cryptology (Section 4.8).
Problem 4.7.1. Design a good data structure for representing large integers on
a computer. Your representation should use a number of bits in O (n) for an integer
that can be expressed in n decimal digits. It must also allow negative numbers to be
represented, and it must be possible to carry out in linear time multiplications and
integer divisions by positive powers of 10 (or another base if you prefer), as well as
additions and subtractions.
Problem 4.7.2. Give an algorithm able to add an integer with m digits and an
integer with n digits in a time in O(m + n).
Problem 4.7.3. Implement your solution to Problem 4.7.2 on a computer

using the representation you invented for Problem 4.7.1. Also implement the classic
algorithm for multiplying large integers (see Section 1.1).
Although an elementary multiplication takes scarcely more time than an addition

on most computers, this is no longer true when the operands involved are very large.
Your solution to Problem 4.7.2 shows how to add two integers in linear time. On the
other hand, the classic algorithm and multiplication a la russe both take quadratic time
to multiply these same operands. Can we do better? Let u and v be two integers of n
decimal digits to be multiplied. Divide-and-conquer suggests that we should separate
Sec. 4.7 Arithmetic with Large Integers 125
each of these operands into two parts of as near the same size as possible :
u = 10S w + x and v =10'y +z, where 0<_ x< 10', 0!5z< 10S , and s = Ln / 2j.
The integers w and y therefore both have Fn/21 digits. See Figure 4.7.1. (For con-
venience, we say that an integer has j digits if it is smaller than 10j, even if it is not
greater than or equal to 10i -1.)
u w x
V y z
Figure 4.7.1. Splitting the operands for

fn/21 Ln/2
large integer multiplication.
The product that interests us is
uv = 102swy + 10S(wz+xy)+xz
We obtain the following algorithm.
function mult (u , v : large-integers) : large-integer

n - smallest integer so that _u and v are of size n
if n is small then multiply u by v using the classic algorithm
return the product thus computed
s f-n div 2
w <-- u div lOs ; x - u mod IOs
y E--v div 101z 4-v mod 10,
return mult (w , y) x 102s
+ (mult (w , z) + mult (x , y )) x 101
+mult(x,z)
Let tb (n) be the time required by this algorithm in the worst case to multiply two
n digit integers. If we use the representation suggested in Problem 4.7.1 and the algo-
rithms of Problem 4.7.2, the integer divisions and multiplications by 102s and 10',
as well as the additions, are executed in linear time. The same is true of the modulo
operations, since these are equivalent to an integer division, a multiplication,
and a subtraction. The last statement of the algorithm consists of four recursive
calls, each of which serves to multiply two integers whose size is about n12. Thus
tb (n) E 3tb ([n / 21) + tb (Ln / 2j) + O(n). This equation becomes tb (n) E 4tb (n / 2)
+ O(n) when n is a power of 2. By Example 2.3.7 the time taken by the preceding
algorithm is therefore quadratic, so we have not made any improvement compared to
the classic algorithm. In fact, we have only managed to increase the hidden constant !
The trick that allows us to speed things up consists of calculating wy, wz +xy,
and xz by executing less than four half-length multiplications, even if this means that
we have to do more additions. This is sensible because addition is much faster than
multiplication when the operands are large. Consider the product
r =(w+x)(y+z)=wy +(wz+xy)+xz .
After only one multiplication this includes the three terms we need in order to calculate
uv. Two other multiplications are needed to isolate these terms. This suggests we
should replace the last statement of the algorithm by
r E- mult (w +x, y +z)
p <- mult(w,y);q -mult(x,z)
return 102sp + l0s (r -p -q) + q
Let t(n) be the time required by the modified algorithm to multiply two integers
of size at most n . Taking account of the fact that w + x and y + z may have up to
1 + Fn / 21 digits, we find that there exist constants c E IIt+ and no E IN such that
t(n) <- t(Ln12.J)+t([n/21)+t(l+Fn121)+cn

for every n ? no. This resembles the equation tc (n) E 3tc J n / 21) + 0 (n), which we
met at the beginning of the chapter. Using Problem 2.3.11 (notice that t(n) is nonde-
creasing by definition), we conclude that t (n) E 0 (n 1g3).
It is thus possible to multiply two n digit integers in a time in 0 (n 1g3), which is
in 0 (n 1.59). However, the hidden constants are such that this algorithm only becomes
interesting in practice when n is quite large. A good implementation will probably not
use base 10, but rather, the largest base for which the hardware allows two "digits" to
be multiplied directly. Recall that the performance of this algorithm and of the classic
algorithm are compared empirically at the end of Section 1.7.2.
Example 4.7.1. We require to multiply u = 2,345 and v = 6,789. The initial

decomposition of the operands gives n = 4, s = 2, w = 23, x = 45, y = 67, and z = 89.
We obtain successively p = 23 x 67 = 1541, q = 45 x 89 = 4,005, and
r = (23 + 45)(67 + 89) = 68 x 156 = 10,608. Finally, the required product uv is
obtained by calculating
1,541x104 + (10,608 -1,541-4,005)x102 + 4,005
= 15,410,000 + 506,200 + 405 = 15,920,205 .
Of course, our example is so small that it would be quicker to use the classic multipli-
cation algorithm in this case.
Problem 4.7.4. Let u and v be integers of exactly m and n digits, respectively.

Suppose without loss of generality that m <- n. The classic algorithm obtains the pro-
duct of u by v in a time in 0 (mn). Even when m # n, the algorithm we have just seen
can multiply u and v, since it will simply treat u as though it too were of size n. This
gives an execution time in 0 (n'3), which is unacceptable if m is very much smaller
Sec. 4.7 Arithmetic with Large Integers 127
than n. Show that it is possible in this case to multiply u and v in a time in

O (nm lg(3i2))
** Problem 4.7.5. Rework Problem 2.2.11 (analysis of the algorithm fib3) in the
context of our new multiplication algorithm, and compare your answer once again
to Example 2.2.8.
* Problem 4.7.6. Following up Problem 4.7.3, implement on your machine the

algorithm we have just discussed. Your algorithm must allow for the possibility that
the operands may differ in size. Compare empirically this implementation and the one
you made for Problem 4.7.3.
* Problem 4.7.7. Show that it is possible to separate each of the operands to be

multiplied into three parts rather than two, so as to obtain the required product using
five multiplications of integers about one-third as long (not nine as would seem neces-
sary at first sight). Analyse the efficiency of the algorithm suggested by this idea.
(Remark: Integer division of an n digit integer by an arbitrary constant k can be carried
out in a time in O(n), although the value of the hidden constant may depend on k.)
** Problem 4.7.8. Generalize the algorithm suggested by Problem 4.7.7. by

showing that there exists, for every real number a > 1, an algorithm Aa, that can mul-
tiply two n digit integers in a time in the order of n a.
Problem 4.7.9. Show by a simple argument that the preceding problem is

impossible if we insist that the algorithm A a must take a time in the exact order of n a.
Problem 4.7.10. Following up Problem 4.7.8, consider the following algo-

rithm for multiplying large integers.
function supermul (u, v : large-integers) : large-integer
n - smallest integer so that u and v are of size n
if n is small then multiply u by v using the classic algorithm
elsea - 1+(lglgn)/lgn
multiply u and v using algorithm A a
return the product thus computed
At first glance this algorithm seems to multiply two n digit numbers in a time in the
order of n a, where a = 1 + (lg lg n) / lg n, that is, in a time in 0 (n log n). Find two
fundamental errors in this analysis of supermul.
Although the idea tried in Problem 4.7.10 does not work, it is nevertheless pos-
sible to multiply two n digit integers in a time in 0 (n log n log log n) by separating
each operand to be multiplied into about parts of about the same size and using
Fast Fourier Transforms (Section 9.5).
Multiplication is not the only interesting operation involving large integers.

Integer division, modulo operations, and the calculation of the integer part of a square
root can all be carried out in a time whose order is the same as that required for multi-
plication (Section 10.2.3). Some other important operations, such as calculating the
greatest common divisor, may well be inherently harder to compute ; they are not
treated here.
4.8 EXPONENTIATION :
AN INTRODUCTION TO CRYPTOLOGY
Alice and Bob do not initially share any common secret information. For some reason
they wish to establish such a secret. Their problem is complicated by the fact that the
only way they can communicate is by using a telephone, which they suspect is being
tapped by Eve, a malevolent third party. They do not want Eve to be privy to their
newly exchanged secret. To simplify the problem, we assume that, although Eve can
overhear conversations, she can neither add nor modify messages on the communica-
tions line.
** Problem 4.8.1. Find a protocol by which Alice and Bob can attain their ends.
(If you wish to think about the problem, delay reading the rest of this section!)
A first solution to this problem was given in 1976 by Diffie and Hellman.
Several other protocols have been proposed since. As a first step, Alice and Bob agree
openly on some integer p with a few hundred decimal digits, and on some other integer
g between 2 and p - 1. The security of the secret they intend to establish is not
compromised should Eve learn these two numbers.
At the second step Alice and Bob choose randomly and independently of each
other two positive integers A and B less than p. Next Alice computes a =.g A mod p
and transmits this result to Bob ; similarly, Bob sends Alice the value b = g B mod p.
Finally, Alice computes x = b A mod p and Bob calculates y = a B mod p. Now x = y
since both are equal to g' mod p. This value is therefore a piece of information
shared by Alice and Bob. Clearly, neither of them can control directly what this value
will be. They cannot therefore use this protocol to exchange directly a message chosen
beforehand by one or the other of them. Nevertheless, the secret value exchanged can
now be used as the key in a conventional cryptographic system.
* Problem 4.8.2. Let p be an odd prime number. The cyclic multiplicative

group ZP is defined as { x E N I 1 <- x < p } under multiplication modulo p. An
integer g in this group is called a generator if each member of the group can be
obtained as some integral power of g. Such a generator always exists. Clearly, the
condition that g be a generator is necessary if Alice and Bob require that the secret
Sec. 4.8 Exponentiation : An Introduction to Cryptology 129
exchanged by the protocol could take on any value between 1 and p - 1. Prove, how-
ever, that regardless of the choice of p and g some secrets are more likely to be chosen
than others, even if A and B are chosen randomly with uniform probability between 1
andp -1.
At the end of the exchange Eve has been able to obtain directly the values of p,
g, a, and b only. One way for her to deduce x would be to find an integer A' such
that a = g A'mod p , and then to proceed like Alice to calculate x' = b A' mod p . If p is
an odd prime, g a generator of ZP , and 1 <- A' < p , then A' is necessarily equal to A,
and sox =x and the secret is correctly computed by Eve in this case.
Problem 4.8.3. Show that even if A #A', still b A mod p =b A' mod p pro-
vided that g A mod p = g A' mod p and that there exists a B such that b = g B mod p.
The value x' calculated by Eve in this way is therefore always equal to the value x
shared by Alice and Bob.
Calculating A' from p, g and a is called the problem of the discrete logarithm.
There exists an obvious algorithm to solve it. (If the logarithm does not exist, the
algorithm returns the value p. For instance, there is no integer A such that
3=2Amod7.)
function dlog (g, a, p)
A <- O;x 4- l
repeat
A<-A+1
x 4- xg
until (a = x mod p) or (A = p)
return A
This algorithm takes an unacceptable amount of time, since it makes p 12 trips round
the loop on the average when the conditions of Problem 4.8.2 hold. If each trip round
the loop takes 1 microsecond, this average time is more than the age of Earth even if p
only has two dozen decimal digits. Although there exist other more efficient algo-
rithms for calculating discrete logarithms, none of them is able to solve a randomly
chosen instance in a reasonable amount of time when p is a prime with several hundred
decimal digits. Furthermore, there is no known way of recovering x from p, g, a, and
b that does not involve calculating a discrete logarithm. For the time being, it seems
therefore that this method of providing Alice and Bob with a shared secret is sound,
although no one has yet been able to prove this.
An attentive reader may wonder whether we are pulling his (or her) leg. If Eve
needs to be able to calculate discrete logarithms efficiently to discover the secret shared
by Alice and Bob, it is equally true that Alice and Bob must be able to calculate
efficiently exponentiations of the form a = g A mod p. The obvious algorithm for this
is no more subtle or efficient than the one for discrete logarithms.
function dexpo l (g , A , p )
a<-- 1
for i F- 1 to A do a E- ag
return a mod p
The fact that x y z mod p = ((x y mod p) x z ) mod p for every x, y, z, and p allows us
to avoid accumulation of extremely large integers in the loop. (The same improvement
can be made in dlog, which is necessary if we hope to execute each trip round the loop
in 1 microsecond.)
function dexpo2(g, A , p )
a 4- 1
for i - 1 to A do a F- ag mod p
return a
* Problem 4.8.4. Analyse and compare the execution times of dexpol and
dexpo2 as a function of the value of A and of the size of p. For simplicity, suppose
that g is approximately equal to p/2. Use the classic algorithm for multiplying large
integers. Repeat the problem using the divide-and-conquer algorithm from Section 4.7
for the multiplications. In both cases, assume that calculating a modulo takes a time in
the exact order of that required for multiplication. 0
Happily for Alice and Bob, there exists a more efficient algorithm for computing
the exponentiation. An example will make the basic idea clear.
x25 = (((x2x)2)2)2x
Thus x25 can be obtained with just two multiplications and four squarings. We leave
the reader to work out the connection between 25 and the sequence of bits 11001
obtained from the expression (((x2x)2x 1)2x1)2x by replacing every x by a 1 and every
1bya0.
The preceding formula for x 25 arises because x 25 = x 24x , x 24 = (x 12)2, and so on.
This idea can be generalized to obtain a divide-and-conquer algorithm.
function dexpo (g , A , p )
if A = 0 then return 1
if A is odd then a - dexpo(g,A-l,p)
return (ag mod p)
else a f- dexpo(g,A/2,p)
return (a2 mod p)
Let h (A) be the number of multiplications modulo p carried out when we calcu-
late dexpo (g , A , p ), including the squarings. These operations dominate the execution
time of the algorithm, which consequently takes a time in O (h (A) x M (p )), where
M (p) is an upper bound on the time required to multiply two positive integers less
than p and to reduce the result modulo p. By inspection of the algorithm we find
Sec. 4.8 Exponentiation : An Introduction to Cryptology 131
0 ifA =0
h(A)= l + h (A -1) if A is odd
1 + h (A /2) otherwise .
Problem 4.8.5. Find an explicit formula for h (A) and prove your answer by
mathematical induction. (Do not try to use characteristic equations.)
Without answering Problem 4.8.5, let us just say that h (A) is situated between
once and twice the length of the binary representation of A, provided A >_ 1. This
means that Alice and Bob can use numbers p, A and B of 200 decimal digits each and
still finish the protocol after less than 3,000 multiplications of 200-digit numbers and
3,000 computations of a 400-digit number modulo a 200-digit number, which is
entirely reasonable. More generally, the computation of dexpo (g , A , P) takes a time
in 0 (M (p ) x log A ). As was the case for binary searching, the algorithm dexpo only
requires one recursive call on a smaller instance. It is therefore an example of
simplification rather than of divide-and-conquer. This recursive call is not at the
dynamic end of the algorithm, which makes it harder to find an iterative version.
Nonetheless, there exists a similar iterative algorithm, which corresponds intuitively to
calculating x25 as x 16x8x 1.
function dexpoiter (g , A , p )
n F- A;y F- g;a - 1
while n > 0 do
if n is odd then a - ay mod p
y <__ y2modp
n(-ndiv2
return a
Problem 4.8.6. The algorithms dexpo and dexpoiter do not minimize the
number of multiplications (including squarings) required. For example, dexpo calcu-
lates x 15 as (((lx)2x)2x)2x, that is, with seven multiplications. On the other hand, dex-
poiter calculates x 15 as 1 x x 2x 4x 8, which involves eight multiplications (the last being
a useless computation of x 16). In both cases the number of multiplications can easily
be reduced to six by avoiding pointless multiplications by the constant 1 and the last
squaring carried out by dexpoiter. Show that in fact x 15 can be calculated with only
five multiplications.
* Problem 4.8.7. By suppressing all the reductions modulo p in the preceding

algorithms, we obtain algorithms for handling large integers capable of calculating
efficiently all the digits of g', for an arbitrary base g and an arbitrary exponent A.
The efficiency of these algorithms depends in turn on the efficiency of the algorithm
used to multiply large integers. As a function of the size of the base and of the value
of the exponent, how much time do the algorithms corresponding to dexpo2 and dexpo
take when the classic multiplication algorithm is used? Rework the problem using the
divide-and-conquer multiplication algorithm from Section 4.7.
The preceding problem shows that it is sometimes not sufficient to be only half-
clever !
4.9 MATRIX MULTIPLICATION
Let A and B be two n x n matrices to be multiplied, and let C be their product. The
classic algorithm comes directly from the definition :
Cij = E Aik Bkj

k=1
Each entry in C is calculated in a time in O(n), assuming that scalar addition and mul-
tiplication are elementary operations. Since there are n2 entries to compute in order to
obtain C, the product of A and B can be calculated in a time in 8(n 3).
Towards the end of the 1960s, Strassen caused a considerable stir by improving
this algorithm. The basic idea is similar to that used in the divide-and-conquer algo-
rithm of Section 4.7 for multiplying large integers. First we show that two 2 x 2
matrices can be multiplied using less than the eight scalar multiplications apparently
required by the definition. Let
all a12 b11 b12
A= and B=
a21 a22 b21 b22
be two matrices to be multiplied. Consider the following operations :
m1 =(a21 +a22-a11)(b22-b12+b11)
m2=a11b11
M3 = a 12b21
m4=(a11-a21)(b22-b12)
m5=(a21+a22)(b12-b11)
m6=(a12-a21 +all -a22)b22
m7=a22(b11+b22-b12-b21).
We leave the reader to verify that the required product AB is given by the fol-
lowing matrix :
m2+m3 m1+m2+m5+m6
m 1+m2+m4-m7 m 1+m2+m4+m5
J
Sec. 4.9 Matrix Multiplication 133
It is therefore possible to multiply two 2 x 2 matrices using only seven scalar multipli-
cations, At first glance, this algorithm does not look very interesting : it uses a large
number of additions and subtractions compared to the four additions that are sufficient
for the classic algorithm.
If we now replace each entry of A and B by an n x n matrix, we obtain an algo-
rithm that can multiply two 2n x 2n matrices by carrying out seven multiplications of
n x n matrices, as well as a number of additions and subtractions of n x n matrices.
This is possible because the basic algorithm does not rely on the commutativity of
scalar multiplication. Given that matrix additions can be executed much faster than
matrix multiplications, the few additional additions compared to the classic algorithm
are more than compensated by saving one multiplication, provided n is sufficiently
large.
Problem 4.9.1. Building on the preceding discussion, show that it is possible

to multiply two n x n matrices in a time in 0 (n 2.81). What do you do about matrices
whose size is not a power of 2 ?
Problem 4.9.2. The number of additions and subtractions needed to calculate

the product of two 2 x 2 matrices using this method seems to be 24. Show that this can
be reduced to 15 by using auxiliary variables to avoid recalculating terms such as
m1+m2+m4. Strassen's original algorithm takes 18 additions and subtractions as
well as 7 multiplications. The algorithm discussed in this section is a variant
discovered subsequently by Shmuel Winograd.
* Problem 4.9.3. Assuming that n is a power of 2, find the exact number of

scalar additions and multiplications executed by your algorithm for Problem 4.9.1,
taking account of the idea from Problem 4.9.2. Your answer will depend on the thres-
hold used to stop making recursive calls. Bearing in mind what you learnt from Sec-
tion 4.2, propose a threshold that will minimize the number of scalar operations.
Following publication of Strassen's algorithm, a number of researchers tried to

improve the constant w such that it is possible to multiply two n x n matrices in a time
in 0 (n °). Almost a decade passed before Pan discovered a more efficient algorithm,
once again based on divide-and-conquer: he found a way to multiply two 70 x 70
matrices that involves only 143,640 scalar multiplications. (Compare this to
703 = 343,000 and to 702.81,
which exceeds 150,000.) Numerous algorithms, asymptot-
ically more and more efficient, have been discovered subsequently. The asymptotically
fastest matrix multiplication algorithm known at the time of this writing can multiply
two n x n matrices in a time in 0 (n 2.376) ; it was discovered by Coppersmith and
Winograd in September 1986. Because of the hidden constants, however, none of the
algorithms found after Strassen's is of much practical use.
4.10 EXCHANGING TWO SECTIONS OF AN ARRAY
For this additional example of an algorithm based on simplification, we ignore the

recursive formulation and give only an iterative version. Let T be an array of n ele-
ments. We wish to interchange the first k elements and the last n -k, without making
use of an auxiliary array. For instance, if T is the array
a d e f g h i j k
and k = 3, the required result in T is
d e f g i
0 k a b c
It is easy to invent an algorithm exchange (i , j , m) to interchange the elements

T [i .. i+m -1] and T [ j .. j +m -1] in a time in 0(m), provided that
m < i +m <- j <- n -m + 1 (see Fig. 4.10.1). With its help, we can solve our problem
as illustrated in Figure 4.10.2. Here the arrows indicate the part of the array where
there are still some changes to make. After each exchange this part is smaller than
before : thus we can affirm that each exchange simplifies the solution of the instance.
The general algorithm is as follows.
procedure exchange (i, j, m )
forp *- Otom - l do
interchange T [i +p ] and T [ j +p ]
procedure transpose (T [I .. n ], k)
i - k; j -n -k;k 4-k+1
while i # j do
if i > j then
exchange (k -i , k , j )
i 4i -j
else
j4-j-i
exchange (k -i , k+j, i)
exchange (k -i , k , i )
The analysis of this algorithm is interesting. Let T (i , j) be the number of ele-

mentary exchanges that have to be made to transpose a block of i elements and a block
of j elements. Then
i ifi=j
T(i,j)= j + T(ij,j) if i > j
i +T(i,j-i) if i < j
For instance, Figure 4.10.2 shows how a block of three elements and a block of eight
elements are transposed. We have
Sec. 4.10 Exchanging Two Sections of an Array 135
T(3, 8) = 3 + T(3,5) = 6 + T(3,2) = 8+T(l,2)=9+T(1, 1) = 10.

The progression of the parameters of the function T recalls an application of Euclid's
algorithm and leads to the following result :
Problem 4.10.1. Prove that
T(i,j)=i + j -gcd(i,j),
where gcd(i, j) denotes the greatest common divisor of i and j (Section 1.7.4).
m elements in elements
exchange
Figure 4.10.1. Effect of the exchange algorithm.
a c d e I g h i j k
exchange (1, 9, 3)
J k d e I g h a b c
exchange (1, 6, 3)
I h J k a b c
exchange (1, 4, 2)
d e I h I a c
exchange (3, 5, 1)
e h a c
exchange (3, 4, 1)
d e g h i b c
Figure 4.10.2. Progress of transpose (T, 3).

Problem 4.10.2. Can we do better if we are allowed unlimited auxiliary

space ?
4.11 SUPPLEMENTARY PROBLEMS
Problem 4.11.1. The Fibonacci sequence. Consider the matrix
F=(011.
Ill 1
Let i and j be any two integers. What is the product of the vector (i , j) and the matrix
F? What happens if i and j are two consecutive numbers from the Fibonacci
sequence? Use this idea to invent a divide-and-conquer algorithm to calculate this
sequence. Does this help you to understand the algorithm fib3 of Section 1.7.5 ?
Problem 4.11.2. Polynomial interpolation. Represent the polynomial

2
p (n) = a o + a 1 n + a 2 n + + ad n d of degree d by an array P [0 .. d ] containing
its coefficients. Suppose you already have an algorithm capable of multiplying a poly-
nomial of degree i by a polynomial of degree 1 in a time in 0 (i ), as well as another
algorithm capable of multiplying two polynomials of degree i in a time in 0 (i log i )
- see Chapter 9. Let n 1, n 2, ... , nd be any integers. Give an efficient algorithm
based on divide-and-conquer to find the unique monic polynomial p (n) of degree d
such that p (n 1) = p (n 2) = = p (nd) = 0. (A polynomial is monic if its coefficient
of highest degree ad = 1.) Analyse the efficiency of your algorithm.
Problem 4.11.3. Smallest and largest elements. Let T [1 .. n ] be an array of

n elements. It is easy to find the largest element of T by making exactly n -1 com-
parisons between elements.
max FT[1];ind t-1
for i +- 2 to n do
if max < T [i ] then max 4- T [i ], ind F i
We only count comparisons between elements ; we exclude implicit comparisons in the
control of the for loop. We could subsequently find the smallest element of T by
making n - 2 more comparisons.
T[ind]*-T[1]
min *- T [2]
for i F-3tondo
if min > T [i] then min T[i]
Find an algorithm that can find both the largest and the smallest elements of an array
of n elements by making less than 2n -3 comparisons between elements. You may
Sec. 4.11 Supplementary Problems 137
assume that n is a power of 2. Exactly how many comparisons does your algorithm
require? How would you handle the situation when n is not a power of 2 ?
Problem 4.11.4. Let T [I .. n ] be a sorted array of integers, some of which

may be negative but all of which are different. Give an algorithm that is able to find
an index i such that I <_ i <_ n and T [i] = i, provided such an index exists. Your algo-
rithm should take a time in 0 (log n) in the worst case.
Problem 4.11.5. Majority element. Let T [1 .. n ] be an array of n elements.

An element x is said to be a majority element in T if # [ i T [i ] = x } > n 12. Give an
algorithm that can decide whether an array T [I .. n ] includes a majority element (it
cannot have more than one), and if so find it. Your algorithm must run in linear time.
* Problem 4.11.6. Rework Problem 4.11.5 with the supplementary constraint

that the only comparisons allowed between elements are tests of equality. You may
therefore not assume that an order relation exists between the elements.
Problem 4.11.7. If you could not manage the previous problem, try again, but
allow your algorithm to take a time in 0 (n log n).
Problem 4.11.8. Tally circuit. An n-tally is a circuit that takes n bits as

inputs and produces 1 + LIg n] bits as output. It counts (in binary) the number of bits
equal to 1 among the inputs. For example, if n = 9 and the inputs are 011001011, the
output is 0101. An (i , j)-adder is a circuit that has one i bit input, one j bit input, and
one [I+ max(i, j)]-bit output. It adds its two inputs in binary. For example, if i = 3,
j = 5, and the inputs are 101 and 10111 respectively, the output is 011100. It is
always possible to construct an (i , j)-adder using exactly max(i, j) 3-tallies. For this
reason the 3-tally is often called a full adder.
i. Using full adders and (i , j )-adders as primitive elements, show how to build an
efficient n-tally. You may not suppose that n has any special form.
ii. Give the recurrence, including the initial conditions, for the number of
3-tallies needed to build your n-tally. Do not forget to count the 3-tallies that are
part of any (i , j )-adders you might have used.
iii. Using the O notation, give the simplest expression you can for the number of
3-tallies needed in the construction of your n-tally. Justify your answer.
* Problem 4.11.9. Telephone switching. A switch is a circuit with two inputs,

a control, and two outputs. It connects input A with output A and input B with output
B, or input A with output B and input B with output A, depending on the position of
the control - see Figure 4.11.1. Use these switches to construct a network with n
inputs and n outputs able to implement any of the n! possible permutations of
the inputs. The number of switches used must be in 0 (n log n ).
4 N
A A A A
B B B B
Figure 4.11.1. Telephone switches.
* Problem 4.11.10. Merge circuit. A comparator is a circuit with two inputs

and two outputs. The smaller input appears on the upper output, and the larger input
appears on the lower output. For a given integer n, a merge circuit F has two groups
of n inputs and a single group of 2n outputs. Provided each of the two groups of
inputs is already sorted, then each input appears on one of the outputs, and the outputs
are also sorted. For instance, Figure 4.11.2 shows an F4 circuit, illustrating how the
inputs are transmitted to the outputs. Each rectangle represents a comparator. By con-
vention, inputs are on the left and outputs are on the right.
There are two ways to measure the complexity of such a circuit : the size of the
circuit is the number of comparators it includes, and the depth is the largest number of
comparators an input may have to pass before it reaches the corresponding output. The
depth is interesting because it determines the reaction time of the circuit. For example,
the F4 merge circuit shown in Figure 4.11.2 has size 9 and depth 3.
For n a power of 2, show how to construct a merge circuit F whose size and
depth are exactly 1 + n lg n and 1 + lg n , respectively.
Problem 4.11.11. Batcher's sorting circuit. Following up the previous

problem, a sorting circuit S, has n inputs and n outputs ; it sorts the inputs presented
to it. Figure 4.11.3 gives an example of S4, which is of size 5 and depth 3.
For n a power of 2, show how to construct an efficient sorting circuit S . By
"efficient" we mean that the depth of your circuit must be significantly less than n
3 1
first 1 3 3 3 3 2
sorted 2
group 6 4 4 4 2 L37:
of inputs
6 4 It 4
sorted
2 5 5 outputs
5
second 2 5 5 5 6 6
sorted 2 6
group 7 7 7 7 7 7
4 7
of inputs
Figure 4.11.2. A merge circuit.

3 2 2 1
3 1
2 3 1 2 2 2
2 2
4
4 1 /\3-3 3 3
3
1 4 4 4
1 4
Figure 4.11.3. A sorting circuit.
whenever n is sufficiently large. You may use merge circuits to your heart's content,
but their depth and size must then be taken into account. Give recurrences, including
the initial conditions, for the size T and the depth P of your circuit S, . Solve these
equations exactly, and express T and P in O notation as simply as possible. 0
** Problem 4.11.12. Continuing the two previous problems, show that it is pos-
sible to construct a sorting circuit for n elements whose size and depth are in
0(n log n) and 0(logn), respectively. O
Problem 4.11.13. Tournaments. You are to organize a tournament involving

n competitors. Each competitor must play exactly once against each of his opponents.
Player
1 2 3 4 5
n=5 Day
1 2 1 - 5 4
2 3 5 1 - 2
3 4 3 2 1 -
4 5 - 4 3 1
- 4 5 2 3
Player
1 2 3 4 5 6
n=6 Day
1 2 1 6 5 4 3
2 3 5 1 6 2 4
3 4 3 2 1 6 5
4 5 6 4 3 1 2
5 6 4 5 2 3 1
Figure 4.11.4. Timetables for five and six players.

Moreover, each competitor must play exactly one match every day, with the possible
exception of a single day when he does not play at all.
1. If n is a power of 2, give an algorithm to construct a timetable allowing the tour-

nament to be finished in n -1 days.
2. For any integer n > 1 give an algorithm to construct a timetable allowing the
tournament to be finished in n -1 days if n is even, or in n days if n is odd. For
example, Figure 4.11.4 gives possible timetables for tournaments involving five
and six players.
** Problem 4.11.14. Closest pair of points. You are given the coordinates of n
points in the plane. Give an algorithm capable of finding the closest pair
of points in a time in O (n log n ).
Quicksort is from Hoare (1962). Mergesort and quicksort are discussed in detail in
Knuth (1973). Problem 4.4.2 was solved by Kronrod; see the solution to Exercise 18
of Section 5.2.4 of Knuth (1973). The algorithm linear in the worst case for selection
and for finding the median is due to Blum, Floyd, Pratt, Rivest, and Tarjan (1972).
The algorithm for multiplying large integers in a time in 0 (n 1.59) is attributed to
Karatsuba and Ofman (1962). The answer to Problems 4.7.7 and 4.7.8 is discussed in
Knuth (1969). The survey article by Brassard, Monet, and Zuffellato (1986) covers
computation with very large integers.
The algorithm that multiplies two n x n matrices in a time in O (n 2.81) comes
from Strassen (1969). Subsequent efforts to do better than Strassen's algorithm began
with the proof by Hopcroft and Kerr (1971) that seven multiplications are necessary to
multiply two 2x2 matrices in a non-commutative structure ; the first positive success
was obtained by Pan (1978), and the algorithm that is asymptotically the most efficient
known at present is by Coppersmith and Winograd (1987). The algorithm of Section
4.10 comes from Gries (1981).
The original solution to Problem 4.8.1 is due to Diffie and Hellman (1976). The
importance for cryptology of the arithmetic of large integers and of the theory of
numbers was pointed out by Rivest, Shamir, and Adleman (1978). For more informa-
tion about cryptology, consult the introductory papers by Gardner (1977) and Hellman
(1980) and the books by Kahn (1967), Denning (1983), Kranakis (1986), and Brassard
(1988). Bear in mind, however, that the cryptosystem based on the knapsack problem,
as described in Hellman (1980), has since been broken. Notice too that the integers
involved in these applications are not sufficiently large for the algorithms of Section
4.7 to be worthwhile. On the other hand, efficient exponentiation as described in Sec-
tion 4.8 is crucial. The natural generalization of Problem 4.8.6 is examined in Knuth
(1969).
The solution to Problem 4.11.1 can be found in Gries and Levin (1980) and
Urbanek (1980). Problem 4.11.3 is discussed in Pohl (1972) and Stinson (1985).
Problems 4.11.10 and 4.11.11 are solved in Batcher (1968). Problem 4.11.12 is
solved, at least in principle, in Ajtai, Komlos, and Szemeredi (1983). Problem 4.11.14
is solved in Bentley and Shamos (1976), but consult Section 8.7 for more on this
problem.
5
Dynamic Programming
5.1 INTRODUCTION
In the last chapter we saw that it is often possible to divide an instance into subin-
stances, to solve the subinstances (perhaps by further dividing them), and then to com-
bine the solutions of the subinstances so as to solve the original instance. It sometimes
happens that the natural way of dividing an instance suggested by the structure of the
problem leads us to consider several overlapping subinstances. If we solve each of
these independently, they will in turn create a large number of identical subinstances.
If we pay no attention to this duplication, it is likely that we will end up with an
inefficient algorithm. If, on the other hand, we take advantage of the duplication and
solve each subinstance only once, saving the solution for later use, then a more
efficient algorithm will result. The underlying idea of dynamic programming is thus
quite simple : avoid calculating the same thing twice, usually by keeping a table of
known results, which we fill up as subinstances are solved.
Dynamic programming is a bottom-up technique. We usually start with the
smallest, and hence the simplest, subinstances. By combining their solutions, we
obtain the answers to subinstances of increasing size, until finally we arrive at the solu-
tion of the original instance. Divide-and-conquer, on the other hand, is a top-down
method. When a problem is solved by divide-and-conquer, we immediately attack the
complete instance, which we then divide into smaller and smaller subinstances as the
algorithm progresses.
142
Sec. 5.1 Introduction 143
Example 5.1.1. Consider calculating the binomial coefficient
+
Inkl j, 0<k <n
n
k
otherwise .
If we calculate [ k] directly by
function C (n , k )
if k = 0 or k = n then return 1
else return C (n - l , k - 1) + C (n - 1, k)
many of the values C (i, j), i < n , j < k are calculated over and over. Since the final
result is obtained by adding up a certain number of Is, the execution time of this algo-
rithm is certainly in S2( [ k 1). We have already met a similar phenomenon in algo-
rithm fib] for calculating the Fibonacci sequence (see Section 1.7.5 and Example
2.2.7).
If, on the other hand, we use a table of intermediate results (this is of course
Pascal's triangle ; see Figure 5.1.1), we obtain a more efficient algorithm. The table
should be filled line by line. In fact, it is not even necessary to store a matrix : it is
sufficient to keep a vector of length k, representing the current line, which we update
from left to right. Thus the algorithm takes a time in 0 (nk) and space in 0 (k), if we
assume that addition is an elementary operation.
0 1 2 3 ... k - I k
0
1 1 1
2 1 2 1
C(n-1,k-1)C(n-1,k)
C(n, k) Figure 5.1.1. Pascal's triangle.
Problem 5.1.1. Prove that the total number of recursive calls made during the
computation of C (n, k) is exactly 2 [ n -2-
Problem 5.1.2. Calculating the Fibonacci sequence affords another example

of this kind of technique. Which of the algorithms presented in Section 1.7.5 uses
dynamic programming?
Dynamic programming is often used to solve optimization problems that satisfy

the principle of optimality: in an optimal sequence of decisions or choices, each subse-
quence must also be optimal. Although this principle may appear obvious, it does not
always apply.
144 Dynamic Programming Chap. 5
Example 5.1.2. If the shortest route from Montreal to Toronto goes via
Kingston, then that part of the journey from Montreal to Kingston must also follow the
shortest route between these two cities : the principle of optimality applies. However,
if the fastest way to drive from Montreal to Toronto takes us first to Kingston, it does
not follow that we should drive from Montreal to Kingston as quickly as possible : if
we use too much petrol on the first half of the trip, maybe we have to stop to fill up
somewhere on the second half, losing more time than we gained by driving hard. The
subtrips Montreal-Kingston and Kingston-Toronto are not independent, and the prin-
ciple of optimality does not apply.
Problem 5.1.3. Show that the principle of optimality does not apply to the
problem of finding the longest simple path between two cities. Argue that this is due
to the fact that one cannot in general splice two simple paths together and expect to
obtain a simple path. (A path is simple if it never passes through the same place twice.
Without this restriction the longest path might be an infinite loop.)
The principle of optimality can be restated as follows for those problems for
which it applies : the optimal solution to any nontrivial instance is a combination of
optimal solutions to some of its subinstances. The difficulty in turning this principle
into an algorithm is that it is not usually obvious which subinstances are relevant to the
instance under consideration. Coming back to Example 5.1.2, it is not immediately
obvious that the subinstance consisting of finding the shortest route from Montreal to
Ottawa is irrelevant to the shortest route from Montreal to Toronto. This difficulty
prevents us from using a divide-and-conquer approach that would start from the ori-
ginal instance and recursively find optimal solutions precisely to those relevant subin-
stances. Instead, dynamic programming efficiently solves every possible subinstance in
order to figure out which are in fact relevant, and only then are these combined into an
optimal solution to the original instance.
5.2 THE WORLD SERIES
As our first example of dynamic programming, let us not worry about the principle of
optimality, but rather concentrate on the control structure and the order of resolution of
the subinstances. For this reason the problem considered in this section is not one of
optimization.
Imagine a competition in which two teams A and B play not more than 2n -1
games, the winner being the first team to achieve n victories. We assume that there
are no tied games, that the results of each match are independent, and that for any
given match there is a constant probability p that team A will be the winner and hence
a constant probability q = 1 -p that team B will win.
Let P (i, j) be the probability that team A will win the series given that they still
need i more victories to achieve this, whereas team B still needs j more victories if
they are to win. For example, before the first game of the series the probability that
Sec. 5.2 The World Series 145
team A will be the overall winner is P (n , n) : both teams still need n victories. If
team A has already won all the matches it needs, then it is of course certain that they
will win the series : P (0, i) = 1, 1 <- i < n. Similarly P (i, 0) = 0, 1 < i <- n. P (0, 0) is
undefined. Finally, since team A wins any given match with probability p and loses it
with probability q,
P(i,j)=pP(i-1,j)+gP(i,j-1) i ? l,j ? 1
Thus we can compute P (i, j) using
function P (i, j)
if i = 0 then return 1
else if j = 0 then return 0
else return pP (i -1, j) + qP (i, j -1)
Let T (k) be the time needed in the worst case to calculate P (i, j), where k = i +j.
With this method, we see that
T(1)=c
T(k):5 2T(k - 1) + d , k>1
where c and d are constants. T (k) is therefore in 0 (2k) = 0 (4") if i = j = n. In
fact, if we look at the way the recursive calls are generated, we find the pattern shown
in Figure 5.2.1, which is identical to that followed by the naive calculation of the bino-
mial coefficient C (i + j , j) = C ((i -1) + j , j) + C (i + (j -1), j -1). The total
number of recursive calls is therefore exactly 2(' ' I - 2 (Problem 5.1.1). To calculate
the probability P (n , n) that team A will win given that the series has not yet started,
the required time is thus in S2( [ 2n ll
n JJ
P(i, j) k matches left

calls
P(i - 1, j) P(i, j - 1) k - 1 matches left
that call
P(i - 2, j) P(i - 1, j - 1) P(i, j - 2) k - 2 matches left
etc.
Figure 5.2.1. Recursive calls made by a call on function P (i , j).
Problem 5.2.1. Prove that (nn, >- 4" /(2n + 1).
Combining these results, we see that the time required to calculate P (n , n) is in

0(4 n ) and Q(4 n In). This method is therefore not practical for large values of n.
(Although sporting competitions with n > 4 are the exception, this problem does have
other applications ! )
* Problem 5.2.2. Prove that in fact the time needed to calculate P (n , n) using
the preceding algorithm is in 0(4"I[ ).
To speed up the algorithm, we proceed more or less as with Pascal's triangle : we

declare an array of the appropriate size and then fill in the entries. This time, however,
instead of filling the array line by line, we work diagonal by diagonal. Here is the
algorithm to calculate P (n , n) .
function series (n , p )
array P[0..n,0..n]
qE-1-p
for s F- I to n do
P[0,s]- 1;P[s,0]-0
fork - 1 to s-1 do
P[k,s-k] <- pP[k-l,s-k] + gP[k,s -k - 1]
for s 1 to n do
fork - 0 to n -s do
P[s+k,n-k] - pP[s+k-1,n-k]+qP[s+k,n-k-1]
return P [n, n ]
Problem 5.2.3. Using this algorithm, calculate the probability that team A
will win the series if p = 0.45 and if four victories are needed to win.
Since in essence the algorithm has to fill up an n x n array, and since a constant
time is required to calculate each entry, its execution time is in 0(n 2).
Problem 5.2.4. Show that a memory space in O(n) is sufficient to implement

this algorithm.
Problem 5.2.5. Show how to compute P (n, n) in a time in O(n). (Hint : use a
completely different approach - see Section 8.6.)
5.3 CHAINED MATRIX MULTIPLICATION
We wish to compute the matrix product

M=M1M2...M"
Matrix multiplication is associative, so we can compute the product in a number of
ways :
Sec. 5.3 Chained Matrix Multiplication 147
M=(...((M1M2)M3)...Mn)
= (M1(M2(M3 ... (Mn-1Mn) ... )))
The choice of a method of computation can have a considerable influence on the time
required.
Problem 5.3.1. Show that calculating the product AB of a p x q matrix A and

a q x r matrix B by the direct method requires pqr scalar multiplications.
Example 5.3.1. We wish to calculate the product ABCD of four matrices : A

is 13 x 5, B is 5 x 89, C is 89 x 3, and D is 3 x 34. To measure the efficiency of the
different methods, we count the number of scalar multiplications that are involved. For
example, if we calculate M = ((AB )C )D, we obtain successively
(AB) 5,785 multiplications
(AB )C 3,471 multiplications
((AB )C )D 1,326 multiplications
for a total of 10,582 multiplications. There are five essentially different ways of calcu-
lating this product. (In the second case that follows, we do not differentiate between
the method that first calculates AB and the one that starts with CD.) In each case,
here is the corresponding number of scalar multiplications.
((AB )C )D 10,582
(AB) (CD) 54,201
(A (BC ))D 2,856
A ((BC )D) 4,055
A (B (CD )) 26,418
The most efficient method is almost 19 times faster than the slowest. 0
To find directly the best way to calculate the product, we could simply
parenthesize the expression in every possible fashion and count each time how many
scalar multiplications will be required. Let T (n) be the number of essentially different
ways to parenthesize a product of n matrices. Suppose we decide to make the first cut
between the i th and the (i + 1)st matrices of the product :
M = (M 1 M 2 ... M, ) (Mi
+ 1 Mi +2
... Mn) .
There are now T (i) ways to parenthesize the left-hand term and T (n - i) ways to
parenthesize the right-hand term. Since i can take any value from 1 to n - 1, we
obtain the following recurrence for T (n) :
nI
T(n)= E T(i)T(n-i).
i=1
Adding the initial condition T (l) = 1, we can thus calculate all the values of T.
Among other values, we find
n 1 2 3 4 5 10 15
T (n) 1 1 2 5 14 4,862 2,674,440.
The values of T (n) are called Catalan numbers.
*Problem 5.3.2. Prove that

1 (2n-2]
T(n)= n n-1
For each way that parentheses can be inserted it takes a time in U(n) to count the
number of scalar multiplications required (at least if we do not try to be subtle). Since
T(n) is in S2(4"/n2) (from Problems 5.2.1 and 5.3.2), finding the best way to calculate
M using the direct approach requires a time in Q(4"/n). This method is therefore
impracticable for large values of n : there are too many ways in which parentheses can
be inserted for us to look at them all.
Fortunately, the principle of optimality applies to this problem. For instance, if
the best way of multiplying all the matrices requires us to make the first cut between
the ith and the (i + l)st matrices of the product, then both the subproducts
M1M2 Mi and Mi + A12 ... Mn must also be calculated in an optimal way. This
suggests that we should consider using dynamic programming. We construct a table
mid , I <- i <- j <- n, where mid gives the optimal solution - that is, the required
number of scalar multiplications - for the part Mi Mi +I MM of the required pro-
duct. The solution to the original problem is thus given by m 1, .
Suppose the dimensions of the matrices Mi are given by a vector di , 0 <- i <- n,
such that the matrix Mi is of dimension di -I by di . We build the table mid diagonal
by diagonal : diagonal s contains the elements mid such that j - i = s. We thus obtain
in succession :
s = 0: mii = 0, i = 1, 2, ... , n
s = 1 : mi, i + 1 = di -1di di + 1, i = 1, 2, ... , n -1 (see Problem 5.3.1)
1 <S <n:mi,i+s= min (mik+mk+l,i+s+di-ldkdi+s), i =1,2,.. ,n-s.
i<k<i+s
The third case represents the fact that to calculate Mi Mi+l ... Mi+s we try all the pos-
sibilities (Mi Mi + 1 . Mk ) (Mk + 1 ... Mi +s) and choose the best for i <- k < i + s . It
is only for clarity that the second case is written out explicitly, as it falls under the gen-
eral case with s = 1.
Example 5.3.2. Continuation of Example 5.3.1. We have d =(I 3,=(13,5,89,3,34).

For s = 1, we find m 12 = 5,785, m23 = 1,335, m34 = 9,078. Next, for s = 2
we obtain
Sec. 5.3 Chained Matrix Multiplication 149
m13 =min(mll +m23+ 13x5x3, m12+m33+ 13x89x3)

= min(1530, 9256) = 1,530
m24=min(m22+m34+5x89x34, m23+m44+5x3x34)
= min(24208, 1845) = 1,845
Finally, for s = 3
m14=min({k=1} mll+m24+13x5x34,
{k=2} m12 m34+ 13x89x34 ,
{k=3} m +m44+ 13x3x34)
= min(4055,5 01, 2856) = 2,856 .
The array m is thus given in figure 5.3.1.

j=1 2 3 4
i=1 0 1 \ 5 785 \ 1 530 . 2 856

s=3
2
s=2
3
S=1
4
s=0
Figure 5.3.1. Example of the chained matrix multiplication algorithm.
Problem 5.3.3. Write the algorithm to calculate m In .
Problem 5.3.4. How must the algorithm be modified if we want not only to
calculate the value of m In , but also to know how to calculate the product M in the
most efficient way ?
For s > 0, there are n - s elements to be computed in the diagonal s ; for each,
we must choose between s possibilities (the different possible values of k). The exe-
cution time of the algorithm is therefore in the exact order of
n-1 n-1 n-I
(n-s)s =n2:s - y s2
s=l s=1 s=l
= n 2 (n - 1)/2 - n (n - 1)(2n - 1)/6
= (n 3-n )/6
The execution time is thus in 0(n 3).

* Problem 5.3.5. Prove that with the following recursive algorithm

function minmat (i, j)
if i = j then return 0
ans - 00
fork <--i toj-1 do
ans F- min(ans, d [i -1 ]d [k]d [ j ] + minmat (i, k) + minmat (k + 1, j))
return ans ,
where the array d [0.. n ] is global, a call on minmat (1, n) takes a time in O(3" ). (Hint:
for the "0 " part, use ccnstructive induction to find constants a and b such that the
time taken by a call on minmat (1, n) is no greater than a 3" - b.)
Although a call on the recursive minmat (1, n) of Problem 5.3.5 is faster than
naively trying all possible ways to parenthesize the desired product, it is still much
slower than the dynamic programming algorithm described previously. This behaviour
illustrates a point made in the first paragraph of this chapter. In order to decide on the
best way to parenthesize the product ABCDEFG, minmat recursively solves 12 subin-
stances, including the overlapping ABCDEF and BCDEFG, both of which recursively
solve BCDEF from scratch. It is this duplication of effort that causes the inefficiency
of minmat.
5.4 SHORTEST PATHS
Let G = <N, A > be a directed graph ; N is the set of nodes and A is the set of
edges. Each edge has an associated nonnegative length. We want to calculate the
length of the shortest path between each pair of nodes. (Compare this to Section 3.2.2,
where we were looking for the length of the shortest paths from one particular node,
the source, to all the others.)
As before, suppose that the nodes of G are numbered from 1 to n,
N = { 1, 2, ... , n }, and that a matrix L gives the length of each edge, with L [i , i ] = 0,
L [i, j] ? 0 if i # j, and L [i, j] = oc if the edge (i, j) does not exist.
The principle of optimality applies : if k is a node on the shortest path from i to
j, then that part of the path from i to k, and that from k to j, must also be optimal.
We construct a matrix D that gives the length of the shortest path between each
pair of nodes. The algorithm initializes D to L. It then does n iterations. After itera-
tion k , D gives the length of the shortest paths that only use nodes in 11, 2, ... , k) as
intermediate nodes. After n iterations we therefore obtain the result we want. At itera-
tion k, the algorithm has to check for each pair of nodes (i, j) whether or not there
exists a path passing through node k that is better than the present optimal path passing
only through nodes in { 1, 2, ... , k -1 }. Let Dk be the matrix D after the k th itera-
tion. The necessary check can be written as
Sec. 5.4 Shortest Paths 151
Dk[i,j] =min(Dk_i[i,j],Dk-i[i,k]+Dk-1[k,j]),
where we make use of the principle of optimality to compute the length of the shortest
path passing through k. We have also implicitly made use of the fact that an optimal
path through k does not visit k twice.
At the k th iteration the values in the k th row and the k th column of D do not
change, since D [k , k ] is always zero. It is therefore not necessary to protect these
values when updating D. This allows us to get away with using only a two-
dimensional matrix D, whereas at first sight a matrix n x n x 2 (or even n x n x n) seems
necessary.
The algorithm, known as Floyd's algorithm, follows.
procedure Floyd(L [1 .. n, I.. n ]) : array[ l .. n, I.. n ]

array D[I..n,l..n]
D -L
fork <-Iton do
for i - 1 to n do
for j E- 1 to n do
D [i, j] - min(D [i, j], D [i, k ] +D [k, j])
return D
Figure 5.4.1 gives an example of the way the algorithm works.

It is obvious that this algorithm takes a time in O(n3). We can also use
Dijkstra's algorithm (Section 3.2.2) to solve the same problem. In this case we have to
apply the algorithm n times, each time choosing a different node as the source. If we
use the version of Dijkstra's algorithm that works with a matrix of distances, the total
computation time is in n x O(n 2), that is, in O(n 3 ). The order is the same as for
Floyd's algorithm, but the simplicity of Floyd's algorithm means that it will probably
be faster in practice. On the other hand, if we use the version of Dijkstra's algorithm
that works with a heap, and hence with lists of the distances to adjacent nodes, the total
time is in n x O ((a +n) log n), that is, in O ((an +n 2) log n), where a is the number of
edges in the graph. If the graph is not very dense (a << n 2), it may be preferable to
use Dijkstra's algorithm n times ; if the graph is dense (a = n 2), it is better to use
Floyd's algorithm.
We usually want to know where the shortest path goes, not just its length. In
this case we use a second matrix P initialized to 0. The innermost loop of the algo-
rithm becomes
ifD[i,k]+D[k, j] <D[i, j] then D[i, j] <--D[i,k]+D[k, j]

P[i,j]-k .
15
0 5 oo o
50 0 15 5
Do=L = 30 00 0 15
15 00 5 0
0 5 00 00 0 5 20 10
50 0 15 5 50 0 15 5
D 1 = 30 35 0 15
D2 = 30 35 0 15
15 20 5 0 15 20 5 0
0 5 20 10 0 5 15 10
45 0 15 5 20 0 10 5
D 3 =
30 35 0 15
D4 = 30 35 0 15
15 20 5 0 15 20 5 0
Figure 5.4.1. Floyd's algorithm at work.
When the algorithm stops, P [i, j ] contains the number of the last iteration that caused
a change in D [i, j]. To recover the shortest path from i to j, look at P [i, j]. If
P [i, j] = 0, the shortest path is directly along the edge (i, j) ; otherwise, if
P [i, j ] = k, the shortest path from i to j passes through k. Look recursively at
P [i, k ] and P [k, j] to find any other intermediate nodes along the shortest path.
Sec. 5.4 Shortest Paths 153
Example 5.4.1. For the graph of Figure 5.4.1, P becomes
0 0 4 2
4 0 4 0
P = 0 1 0 0
0 1 0 0
Since P [1, 3] = 4, the shortest path from 1 to 3 passes through 4. Looking now at
P [1, 4] and P [4, 3], we discover that between 1 and 4 we have to go via 2, but that
from 4 to 3 we proceed directly. Finally we see that the trips from 1 to 2 and from 2
to 4 are also direct. The shortest path from I to 3 is thus 1, 2, 4, 3.
Problem 5.4.1. Suppose we allow edges to have negative lengths. If G

includes cycles whose total length is negative, the notion of "shortest path" loses much
of its meaning : the more often we go round the negative cycle, the shorter our path
will be ! Does Floyd's algorithm work
i. On a graph that includes a negative cycle ?
ii. On a graph that has some edges whose lengths are negative, but that does not
include a negative cycle ?
Prove your answer.
Even if a graph has edges with negative length, the notion of a shortest simple
path still makes sense. No efficient algorithm is known for finding shortest simple
paths in graphs that may have edges of negative length. This is the situation we
encountered in Problem 5.1.3. These two problems are NP-complete (see Section
10.3).
Problem 5.4.2. Warshall's algorithm. In this case, the length of the edges is
of no interest ; only their existence is important. Initially, L [i, j] = true if the edge
(i, j) exists, and L [i , j] = false otherwise. We want to find a matrix D such that
D [i, j] = true if there exists at least one path from i to j, and D [i, j] = false other-
wise. (We are looking for the reflexive transitive closure of the graph G.) Adapt
Floyd's algorithm for this slightly different case. (We shall see an asymptotically
more efficient algorithm for this problem in Section 10.2.2.)
* Problem 5.4.3. Find a significantly better algorithm for Problem 5.4.2 in the
case when the matrix L is symmetric (L [i, j] = L [ j, i ]).
5.5 OPTIMAL SEARCH TREES
We begin by recalling the definition of a binary search tree. A binary tree each of
whose nodes contains a key is a search tree if the value contained in every internal
node is greater than or equal to (numerically or lexicographically) the values contained
in its left-hand descendants, and less than or equal to the values contained in its right-
hand descendants.
Problem 5.5.1. Show by an example that the following definition will not do :
"A binary tree is a search tree if the key contained in each internal node is greater than
or equal to the key of its left-hand child, and less than or equal to the key of its right-
hand child."
Figure 5.5.1 shows an example of a binary search tree containing the keys
A, B, C , .. , H. (For the rest of this section, search trees will be understood to be
binary.) To determine whether a key X is present in the tree, we first examine the key
held in the root. Suppose this key is R. If X=R, we have found the key we want, and
the search stops ; if X < R, we only need look at the left-hand subtree ; and if X > R,
we only need look at the right-hand subtree. A recursive implementation of this tech-
nique is obvious. (It provides an example of simplification : see chapter 4.)
Problem 5.5.2. Write a procedure that looks for a given key in a search tree
and returns true if the key is present and false otherwise.
Figure 5.5.1. A binary search tree.
The nodes may also contain further information related to the keys : in this case a
search procedure does not simply return true or false, but rather, the information
attached to the key we are looking for.
For a given set of keys, several search trees may be possible : for instance, the
tree in Figure 5.5.2 contains the same keys as those in Figure 5.5.1.
Problem 5.5.3. How many different search trees can be made with eight dis-
tinct keys ?
Sec. 5.5 Optimal Search Trees 155
*Problem 5.5.4. If T (n) is the number of different search trees we can make
with n distinct keys, find either an explicit formula for T (n) or else an algorithm to
calculate this value. (Hint: reread Section 5.3.)
In Figure 5.5.1 two comparisons are needed to find the key E ; in Figure 5.5.2,
on the other hand, a single comparison suffices. If all the keys are sought with the
same probability, it takes (2+3+1+3+2+4+3+4)/8 = 22/8 comparisons on the
average to find a key in Figure 5.5.1, and (4+3+2+3+1+3+2+3)/8=21/8 com-
parisons on the average in Figure 5.5.2.
Problem 5.5.5. For the case when the keys are equiprobable, give a tree that
minimizes the average number of comparisons needed. Repeat the problem for the
general case of n equiprobable keys.
In fact, we shall solve a more general problem still. Suppose we have an ordered
set c 1 < c 2 < * < cn of n distinct keys. Let the probability that a request refers to
key c, be pi , i = 1, 2, ... , n . For the time being, suppose that Ein=1 pi = 1, that is,
all the requests refer to keys that are indeed present in the search tree.
Recall that the depth of the root of a tree is 0, the depth of its children is 1, and
so on. If some key ci is held in a node at depth di , then di + 1 comparisons are neces-
sary to find it. For a given tree the average number of comparisons needed is
C= pi (di + 1).
i=1
This is the function we seek to minimize.
Consider the sequence of successive keys ci , ci + i , ... , cj , j >- i. Suppose that
in an optimal tree containing all the n keys this sequence of j-i + 1 keys occupies the
nodes of a subtree. If the key Ck , i < k < j, is held in a node of depth dk in the sub-
tree, the average number of comparisons carried out in this subtree when we look
Figure 5.5.2 Another binary search tree.

for a key in the main tree (the key in question is not necessarily one of those held in
the subtree) is
CY. Pk(dk+l).
k=i
We observe that
this expression has the same form as that for C, and

a change in the subtree does not affect the contribution made to C by other sub-
trees of the main tree disjoint from the one under consideration.
We thus arrive at the principle of optimality : in an optimal tree all the subtrees must
also be optimal with respect to the keys they contain.
Let mij = Yk=i Pk , and let CiJ be the average number of comparisons carried out
in an optimal subtree containing the keys ci , ci + I , ... , cJ when a key is sought in
the main tree. (It is convenient to define CiJ = 0 if j = i - 1.) One of these keys, k,
say, must occupy the root of the subtree. In Figure 5.5.3, L is an optimal subtree con-
taining the keys ci , ci + I , ... , Ck - I and R is an optimal subtree containing
Ck +I , ... , cJ . When we look for a key in the main tree, the probability that it is in
the sequence ci , ci +I , ... , c j is mil . In this case one comparison is made with Ck ,
and others may then be made in L or R. The average number of comparisons carried
out is therefore
CiJ = mil + Ci, k - I+ Ck + I , J
where the three terms are the contributions of the root, L and R, respectively.
Figure 5.5.3. A subtree of the optimal search tree is optimal.
To obtain a dynamic programming scheme, it remains to remark that the root k

is chosen so as to minimize Cij :
CiJ =mil + mm n (Ci,k-I +Ck+I,J) (*)
ISk<J
In particular, Cii = pi .
Example 5.5.1. To find the optimal search tree if the probabilities associated
with five keys c1 to c5 are
Sec. 5.5 Optimal Search Trees 157
i 1 2 3 4 5
Pi 0.30 0.05 0.08 0.45 0.12
we first calculate the matrix m.
0.30 0.35 0.43 0.88 1.00

0.05 0.13 0.58 0.70
m = 0.08 0.53 0.65
0.45 0.57
0.12
Now, we note that Cii = pi , 1 <_ i <_ 5, and next, we use (*) to calculate the other
values of Cii.
C12 = m12 + min(CI0+C22, CII+C32)
= 0.35 + min(0.05, 0.30) = 0.40
Similarly
C23=0.18, C34=0.61, C45=0.69.
Then
C13 = m13 + min(C10+C23, C11+C33, C12+C43)
= 0.43 + min(0.18, 0.38, 0.40) = 0.61
C24 = m24 + min(C21+C34, C22+C44, C23+C54)

= 0.58 + min(0.61, 0.50, 0.18) = 0.76
C35 = m35 + min(C32+C45, C33+C55, C34+C65)

= 0.65 + min(0.69, 0.20, 0.61) = 0.85
C14=m14+min(C1o+C24, C11 +C34, C12+C44, C13+C54)

= 0.88 + min(0.76, 0.91, 0.85, 0.61) = 1.49
C25 = m25 + min(C21 +C35, C22+C45, C23+C55, C24+C65)

= 0.70 + min(0.85, 0.74, 0.30, 0.76) = 1.00
C15=m15+min(C1o+C25, C11+C35, C12+C45, C13+C55, C14+C65)

= 1.00 + min(1.00, 1.15, 1.09, 0.73, 1.49)
=1.73.
The optimal search tree for these keys requires 1.73 comparisons on the average to find
a key (see Figure 5.5.4).
Problem 5.5.6. We know how to find the minimum number of comparisons

necessary in the optimal tree, but how do we find the form of this tree ? 0
Problem 5.5.7. Write an algorithm that accepts the values n and pi ,

i = 1 , 2, ... , n , and that produces a description of the optimal search tree for these
probabilities. (Provided the keys are sorted, we do not need their exact values.)
In this algorithm we calculate the values of Cij first for j -i = 1, then for
j -i = 2, and so on. When j -i = m , there are n -m values of C,, to calculate, each
involving a choice among m + 1 possibilities. The required computation time is there-
fore in
,(In-t (n-m)(m+l)) =8(n3) .
m=1
Problem 5.5.8. Prove this last equality.
Figure 5.5.4. An optimal binary search tree.
* Problem 5.5.9. Generalize the preceding argument to take account of the pos-
sibility that a request may involve a key that is not in fact in the tree. Specifically, let
pi , i = 1, 2, ... , n , be the probability that a request concerns a key ci that is in the
tree, and let qj , i = 0, 1, 2, ... , n, be the probability that it concerns a missing key
situated between ci and ci+1 (with the obvious interpretation for q0 and qn ). We now
have
n n
,pi +Iqi=1.
The optimal tree must minimize the average number of comparisons required to either
find a key, if it is present in the tree, or to ascertain that it is missing.
Give an algorithm that can determine the optimal search tree in this context.
** Problem 5.5.10. For 1 <- i <- j <- n, let

rid = max { k I i <- k <- j and C, =Mid + Ci , k - I+ Ck + 1, ] }
Sec. 5.6 The Travelling Salesperson Problem 159
be the root of an optimal subtree containing c; , c, +I , ... , cj . Write also r;,, -I = i .

Prove that r;,j _i <_rj <r,+1,j for every 1 <_i j Sn.
Problem 5.5.11. Use the result of Problem 5.5.10 to show how to calculate an
optimal search tree in a time in 0(n2). (Problems 5.5.10 and 5.5.11 generalize to the
case discussed in Problem 5.5.9.)
Problem 5.5.12. There is an obvious greedy approach to the problem of con-

structing an optimal search tree : place the most probable key, Ck, say, at the root, and
construct the left- and right-hand subtrees for c i , c 2, ... , ck _ 1 and Ck + i , Ck+2 . .
c, recursively in the same way.
i. How much time does this algorithm take in the worst case, assuming the keys are
already sorted ?
ii. Show with the help of a simple, explicit example that this greedy algorithm does
not always find the optimal search tree. Give an optimal search tree for your
example, and calculate the average number of comparisons needed to find a key
for both the optimal tree and the tree found by the greedy algorithm.
5.6 THE TRAVELLING SALESPERSON PROBLEM
We have already met this problem in Section 3.4.2. Given a graph with nonnegative
lengths attached to the edges, we are required to find the shortest possible circuit that
begins and ends at the same node, after having gone exactly once through each of the
other nodes.
Let G = < N, A > be a directed graph. As usual, we take N = { 1, 2, ... , n
and the lengths of the edges are denoted by L;j , with L [i , i ] = 0, L [i , j ] >_ 0 if i j,
and L [i, j] = oo if the edge (i, j) does not exist.
Suppose without loss of generality that the circuit begins and ends at node 1.
It therefore consists of an edge (1, j), j # 1, followed by a path from j to 1 that passes
exactly once through each node in N \ 11, j }. If the circuit is optimal (as short as pos-
sible), then so is the path from j to 1 : the principle of optimality holds.
Consider a set of nodes S c N \ 111 and a node i EN \ S, with i =1 allowed
only if S = N \ { 1 } . Define g (i, S) as the length of the shortest path from node i to
node 1 that passes exactly once through each node in S. Using this definition,
g (1, N \ { 1 }) is the length of an optimal circuit. By the principle of optimality, we see
that
g(1,N\{1})=2<j<n
min(Ljj +g(j,N\{l,j})). (*)
More generally, if i # 1, S # 0, S # N \ { 1 ), and i 0 S,
g(i,S)=rni (L,j +g(j,S\ { j})).
jES
Furthermore,
g(i,0)=Li1, i =2,3..... n .
The values of g (i, S) are therefore known when S is empty. We can apply (**) to cal-
culate the function g for all the sets S that contain exactly one node (other than 1);
then we can apply (**) again to calculate g for all the sets S that contain two nodes
(other than 1), and so on. Once the value of g ( j, N \ { 1, j }) is known for all the
nodes j except node 1, we can use (*) to calculate g (1, N \ { 11) and solve the
problem.
Example 5.6.1. Let G be the complete graph on four nodes given in Figure
5.6.1:
0 10 15 20
5 0 9 10
L =
6 13 0 12
8 8 9 0
We initialize
g (2, 0) = 5, g (3, 0) = 6, g (4, 0) = 8 .
15
Figure 5.6.1. A directed graph for the travelling salesperson problem.

Sec. 5.6 The Travelling Salesperson Problem 161
Using (**), we obtain

g(2,13 1) =L23+g(3,0)= 15
g(2,{41)=L24+g(4,0)= 18
and similarly
g(3,{2})=18, g(3,14))=20
g(4,{2})= 13, g(4,{3})= 15.
Next, using (**) for sets of two nodes, we have
g (2, { 3,4)) = min(L23 + g (3, (4)), L24 + g (4, { 3) ))
= min(29,25) = 25
g (3, 1 2,4)) = min(L 32 + g (2, { 4) ), L 34 + g (4, { 2 } ))
= min(31,25) = 25
g (4, (2,3 )) = min(L42 + g (2, { 3) ), L43 + g (3,12)))
= min(23, 27) = 23
Finally we apply (*) to obtain
g (1, { 2,3,41) = min(L 12 + g (2, (3,4 } ), L 13 + g (3, (2,4 1), L 14 + g (4, ( 2,3) ))
= min(35, 40, 43) = 35
The optimal circuit in Figure 5.6.1 has length 35.
To know where this circuit goes, we need an extra function : J (i , S) is the value
of j chosen to minimize g at the moment when we apply (*) or (**) to calculate
g(i,S).
Example 5.6.2. (Continuation of Example 5.6.1.) In this example we find
J(2,{3,4)) =4
J(3,(2,4)) =4
J(4,12,31) =2
J(1, {2,3,4)) = 2
and the optimal circuit is
1 -+J(1,{2,3,4))=2
J(2,{3,4})=4
-3J(4,(31)=3
-1
The required computation time can be calculated as follows :
- to calculate g ( j, 0) : n -1 consultations of a table;

- to calculate all the g (i , S) such that 1 :5 #S = k <- n - 2: (n -1) (n k 21 k addi-
tions in all;
- to calculate g (1, N \ (1 }) : n -1 additions.
These operations can be used as a barometer. The computation time is thus in

In
0(2(n -1) + - (n -1)k k 21) = 19(n22") since k (k1 = r 2r-1
k=1 k1
This is considerably better than having a time in SZ(n !), as would be the case if we
simply tried all the possible circuits, but it is still far from offering a practical algo-
rithm. What is more ...
Problem 5.6.1. Verify that the space required to hold the values of g and J is
in Q(n 2" ), which is not very practical either. 13
TABLE 5.6.1. SOLVING THE TRAVELLING SALESPERSON PROBLEM.

Time: Time: Space:
n Direct method Dynamic programming Dynamic programming
n! n22" n 2"
5 120 800 160

10 3,628,800 102,400 10,240
15 1.31 x 1012 7,372,800 491,520
20 2.43x1018 419,430,400 20,971,520
Problem 5.6.2. The preceding analysis assumes that we can find in constant
time a value of g (j , S) that has already been calculated. Since S is a set, which data
structure do you suggest to hold the values of g ? With your suggested structure, how
much time is needed to access one of the values of g ?
Table 5.6.1 illustrates the dramatic increase in the time and space necessary as n
goes up. For instance, 202220 microseconds is less than 7 minutes, whereas 20!
microseconds exceeds 77 thousand years.
5.7 MEMORY FUNCTIONS
If we want to implement the method of Section 5.6 on a computer, it is easy to write a

function that calculates g recursively. For example, consider
Sec. 5.7 Memory Functions 163
function g (i, S )
if S = 0 then return L [i ,1]
ans <- oo
for each j E S do
distviaj F- L [i, j ] + g ( j, S \ { j } )
if distviaj < ans then ans - distviaj
return ans .
Unfortunately, if we calculate g in this top-down way, we come up once more against
the problem outlined at the beginning of this chapter : most values of g are recalcu-
lated many times and the program is very inefficient. (In fact, it ends up back in
S2((n - 1)!).)
So how can we calculate g in the bottom-up way that characterizes dynamic pro-
gramming ? We need an auxiliary program that generates first the empty set, then all
the sets containing just one element from N \ { I), then all the sets containing two ele-
ments from N \ { 1), and so on. Although it is maybe not too hard to write such a gen-
erator, it is not immediately obvious how to set about it.
One easy way to take advantage of the simplicity of a recursive formulation
without losing the efficiency offered by dynamic programming is to use a memory
function. To the recursive function we add a table of the necessary size. Initially, all
the entries in this table hold a special value to indicate that they have not yet been cal-
culated. Thereafter, whenever we call the function, we first look in the table to see
whether it has already been evaluated with the same set of parameters. If so, we return
the value held in the table. If not, we go ahead and calculate the function. Before
returning the calculated value, however, we save it at the appropriate place in the table.
In this way it is never necessary to calculate the function twice for the same values of
its parameters.
For the algorithm of Section 5.6 let gtab be a table all of whose entries are ini-
tialized to -1 (since a distance cannot be negative). Formulated in the following way :
function g (i, S )
if S = 0 then return L [i ,1 ]
if gtab [i, S ] > 0 then return gtab [i, S ]
ans F oo
for each j E S do
distviaj - L [i, j ] + g ( j, S \ { j } )
if distviaj < ans then ans - distviaj
gtab [i, S ] - ans
return ans ,
the function g combines the clarity obtained from a recursive formulation and the
efficiency of dynamic programming.
Problem 5.7.1. Show how to calculate (i) a binomial coefficient and (ii) the
function series (n , p) of Section 5.2 using a memory function.
We sometimes have to pay a price for using this technique. We saw in Section
5.1, for instance, that we can calculate a binomial coefficient (k] using a time in 0 (nk)
and space in 0(k). Implemented using a memory function, the calculation takes the
same amount of time but needs space in Si(nk).
* Problem 5.7.2. If we are willing to use a little more space (the space needed
is only multiplied by a constant factor, however), it is possible to avoid the initializa-
tion time needed to set all the entries of the table to some special value. This is partic-
ularly desirable when in fact only a few values of the function are to be calculated, but
we do not know in advance which ones. (For an example, see Section 6.6.2.) Show
how an array T [1 .. n ] can be virtually initialized with the help of two auxiliary arrays
B [1 .. n ] and P [1 .. n ] and a few pointers. You should write three algorithms.
procedure init
{ virtually initializes T [1 .. n ] }
procedure store (i , v )
{ sets T [i ] to the value v }
function val (i)
{ returns the last value given to T [i ], if any ;
returns a default value (such as -1) otherwise }
A call on any of these procedures or functions (including a call on init !) should take
constant time in the worst case.
* Problem 5.8.1. Let u and v be two strings of characters. We want to

transform u into v with the smallest possible number of operations of the following
types :
delete a character,
add a character,
change a character.
For instance, we can transform abbac into abcbc in three stages.

abbac -* abac (delete b)
ababc (add b )
abcbc (change a into c ).
Show that this transformation is not optimal.
Write a dynamic programming algorithm that finds the minimum number of

operations needed to transform u into v, and that tells us what these operations are.
How much time does your algorithm take as a function of the lengths of u and v ?
Problem 5.8.2. Consider the alphabet E = { a, b, c } . The elements of Y. have

the multiplication table given in Table 5.8.1. Thus ab = b, ba = c, and so on. Note
that the multiplication defined by this table is neither commutative nor associative.
TABLE 5.8.1. AN ABSTRACT MULTIPLICATION TABLE
Right-hand symbol
a b c
Left-hand a b b a
symbol b c b a
c a c C
Find an efficient algorithm that examines a string x = x 1x2 x of characters

of E and decides whether or not it is possible to parenthesize x in such a way that the
value of the resulting expression is a. For instance, if x = bbbba, your algorithm
should return "yes" because (b (bb ))(ba) = a. (This expression is not unique. For
example, (b (b (b (ba )))) = a as well.)
In terms of n, the length of the string x, how much time does your algorithm
take ?
Problem 5.8.3. Modify your algorithm from the previous problem so it

returns the number of different ways of parenthesizing x to obtain a.
Problem 5.8.4. There are N Hudson's Bay Company posts on the River Kok-
soak. At any of these posts you can rent a canoe to be returned at any other post down-
stream. (It is next to impossible to paddle against the current.) For each possible
departure point i and each possible arrival point j the company's tariff gives the cost
of a rental between i and j. However, it can happen that the cost of renting from i to
j is higher than the total cost of a series of shorter rentals, in which case you can
return the first canoe at some post k between i and j and continue the journey in a
second canoe. There is no extra charge if you change canoes in this way.
Find an efficient algorithm to determine the minimum cost of a trip by canoe
from each possible departure point i to each possible arrival point j. In terms of N,
what is the computing time needed by your algorithm ?
Problem 5.8.5. In the introduction to Chapter 3 we saw a greedy algorithm

for making change. This algorithm works correctly in a country where there are coins
worth 1, 5, 10, and 25 units, but it does not always find an optimal solution if there
also exists a coin worth 12 units (see Problem 3.1.1). The general problem can be
solved exactly using dynamic programming. Let n be the number of different coins
that exist, and let T [I .. n ] be an array giving the value of these coins. We suppose
that an unlimited number of coins of each value is available. Let L be a bound on the
sum we wish to obtain.
i. For 1 <- i <_ n and 1 <_ j <_ L, let cij be the minimum number of coins required to
obtain the sum j if we may only use coins of types T [1], T[2],. . ., T[i], or
cii _+00 if the amount j cannot be obtained using just these coins. Give a
recurrence for cij , including the initial conditions.
ii. Give a dynamic programming algorithm that calculates all the c,1 , 1 <_ j <- L.
Your algorithm may use only a single array of length L. As a function of n and
L, how much time does your algorithm take?
iii. Give a greedy algorithm that can make change using the minimum number of
coins for any amount M <_ L once the % have been calculated. Your algorithm
should take a time in 0 (n + provided 0.
* Problem 5.8.6. You have n objects, which you wish to put in order using the
relations "<" and "=". For example, 13 different orderings are possible with three
objects.
A =B =C A =B <C A <B =C A <B <C A <C <B

A =C <B B <A =C B <A <C B <C <A B =C <A
C <A =B C <A <B C <B <A
Give a dynamic programming algorithm that can calculate, as a function of n, the
number of different possible orderings. Your algorithm should take a time in 0 (n 2)
and space in 0 (n).
** Problem 5.8.7. Ackermann's function is defined recursively as follows :
A(0,n)=n+1
A (m, 0) = A (m - 1, 1)ifm >0
A(m,n)=A(m-1,A(m,n-l)) ifm,n >0.
This function grows extremely rapidly.
i. Calculate A (2, 5), A (3, 3), and A (4, 4).

ii. Give a dynamic programming algorithm to calculate A (m, n). Your algorithm
must consist simply of two nested loops (recursion is not allowed). Moreover, it
is restricted to using a space in 0 (m), although some of these memory words can
0
Figure 5.8.1. Cutting a hexagon into triangles.
grow quite large. (Hint : use two arrays val [0.. m ] and ind [0.. m ] such that at
every instant val [i ] = A (i, ind [i ]).) 11
Problem 5.8.8. Prove that the number of ways to cut an n -sided convex
polygon into n - 2 triangles using diagonal lines that do not cross is T (n - 1), the
(n -1)st Catalan number (see Section 5.3). For example, a hexagon can be cut in
14 different ways, as shown in Figure 5.8.1. 11
Several books are concerned with dynamic programming. We mention only Bellman
1957, Bellman and Dreyfus (1962), Nemhauser (1966), and Lauriere (1979). The algo-
rithm in Section 5.3 is described in Godbole (1973) ; a more efficient algorithm, able
to solve the problem of chained matrix multiplications in a time in 0 (n log n), can be
found in Hu and Shing (1982, 1984). Catalan's numbers are discussed in many places,
including Sloane (1973) and Purdom and Brown (1985).
Floyd's algorithm for calculating all shortest paths is due to Floyd (1962). A
theoretically more efficient algorithm is known : Fredman (1976) shows how to solve
the problem in a time in O (n 3(log log n / log n) 1/3 ). The solution to Problem 5.4.2 is
supplied by the algorithm in Warshall (1962). Both Floyd's and Warshall's algorithms
are essentially the same as the one in Kleene (1956) to determine the regular expres-
sion corresponding to a given finite automaton (Hopcroft and Ullman 1979). All these
algorithms (with the exception of Fredman's) are unified in Tarjan (1981).
The algorithm of Section 5.5 for constructing optimal search trees, including the
solution to Problem 5.5.9, comes from Gilbert and Moore (1959). The improvements
suggested by Problems 5.5.10 and 5.5.11 come from Knuth (1971, 1973). A solution
to Problem 5.5.10 that is both simpler and more general is given by Yao (1980); this
paper gives a sufficient condition for certain dynamic programming algorithms that run
in cubic time to be transformable automatically into quadratic algorithms. The optimal
search tree for the 31 most common words in English is compared in Knuth (1973)
with the tree obtained using the greedy algorithm suggested in Problem 5.5.12.
The algorithm for the travelling salesperson problem given in Section 5.6 comes
from Held and Karp (1962). Memory functions are introduced in Michie (1968) ; for
further details see Marsh (1970). Problem 5.7.2, which suggests how to avoid initial-
izing a memory function, comes from Exercise 2.12 in Aho, Hopcroft and Ullman
(1974).
A solution to Problem 5.8.1 is given in Wagner and Fischer (1974). Problem
5.8.5 is discussed in Wright (1975) and Chang and Korsh (1976). Problem 5.8.6 sug-
gested itself to the authors one day when they set an exam including a question resem-
bling Problem 2.1.11: we were curious to know what proportion of all the possible
answers was represented by the 69 different answers suggested by the students (see
also Lemma 10.1.2). Problem 5.8.7 is based on Ackermann (1928). Problem 5.8.8 is
discussed in Sloane (1973). An important dynamic programming algorithm that we
have not mentioned is the one in Kasimi (1965) and Younger (1967), which takes
cubic time to carry out the syntactic analysis of any context-free language (Hopcroft
and Ullman 1979).
6
Exploring Graphs
6.1 INTRODUCTION
A great many problems can be formulated in terms of graphs. We have seen, for
instance, the shortest route problem and the problem of the minimal spanning tree. To
solve such problems, we often need to look at all the nodes, or all the edges, of a
graph. Sometimes, the structure of the problem is such that we need only visit some of
the nodes or edges. Up to now, the algorithms we have seen have implicitly imposed
an order on these visits : it was a case of visiting the nearest node, the shortest edge,
and so on. In this chapter we introduce some general techniques that can be used when
no particular order of visits is required.
We shall use the word "graph" in two different ways. A graph may be a data
structure in the memory of a computer. In this case, the nodes are represented by a
certain number of bytes, and the edges are represented by pointers. The operations to
be carried out are quite concrete : to "mark a node" means to change a bit in memory,
to "find a neighbouring node" means to follow a pointer, and so on.
At other times, the graph exists only implicitly. For instance, we often use
abstract graphs to represent games : each node corresponds to a particular position of
the pieces on the board, and the fact that an edge exists between two nodes means that
it is possible to get from the first to the second of these positions by making a single
legal move. When we explore such a graph, it does not really exist in the memory of
the machine. Most of the time, all we have is a representation of the current position
(that is, of the node we are in the process of visiting) and possibly representations of a
few other positions. In this case to "mark a node" means to take any appropriate meas-
ures that enable us to recognize a position we have already seen, or to avoid arriving at
169
170 Exploring Graphs Chap. 6
the same position twice ; to "find a neighbouring node" means to change the current
position by making a single legal move ; and so on.
However, whether the graph is a data structure or merely an abstraction, the tech-
niques used to traverse it are essentially the same. In this chapter we therefore do not
distinguish the two cases.
6.2 TRAVERSING TREES
We shall not spend long on detailed descriptions of how to explore a tree. We simply
remind the reader that in the case of binary trees three techniques are often used. If at
each node of the tree we visit first the node itself, then all the nodes in the left-hand
subtree, and finally, all the nodes in the right-hand subtree, we are traversing the tree in
preorder ; if we visit first the left-hand subtree, then the node itself, and finally, the
right-hand subtree, we are traversing the tree in inorder; and if we visit first the left-
hand subtree, then the right-hand subtree, and lastly, the node itself, then we are
visiting the tree in postorder. Preorder and postorder generalize in the obvious way to
nonbinary trees.
These three techniques explore the tree from left to right. Three corresponding
techniques explore the tree from right to left. It is obvious how to implement any of
these techniques using recursion.
Lemma 6.2.1. For each of the six techniques mentioned, the time T (n) needed
to explore a binary tree containing n nodes is in O(n).
Proof. Suppose that visiting a node takes a time in O(1), that is, the time
required is bounded above by some constant c. Without loss of generality, we may
suppose that c >- T (0).
Suppose further that we are to explore a tree containing n nodes, n > 0, of
which one node is the root, g nodes are situated in the left-hand subtree, and n -g -1
nodes are in the right-hand subtree. Then
T(n) <- max (T(g) + T(n-g -1) + c) n > 0
O<g <n -1
We prove by mathematical induction that T (n) <- do + c where d is a constant such

that d ? 2c. By the choice of c the hypothesis is true for n = 0. Now suppose that it
is true for all n , 0 <- n < m , for some m > 0. Then
T (m) <_ max (T (g) + T (m -g -1) + c)
0<g <m-1
<_ max
0<g <m-1
(dg +c +d(m-g-1)+c +c)
<- dm +3c -d <- dm +c

Sec. 6.3 Depth-First Search : Undirected Graphs 171
so the hypothesis is also true for n = in. This proves that T (n) 5 do + c for every
n >_ 0, and hence T (n) is in O (n).
On the other hand, it is clear that T (n) is in c2(n) since each of the n nodes is
visited. Therefore T (n) is in O(n).
Problem 6.2.1. Prove that for any of the techniques mentioned, a recursive
implementation takes memory space in S2(n) in the worst case.
* Problem 6.2.2. Show how the preceding exploration techniques can be imple-
mented so as to take only a time in 0(n) and space in 0(1), even when the nodes do
not contain a pointer to their parents (otherwise the problem becomes trivial).
Problem 6.2.3. Show how to generalize the concepts of preorder and post-
order to arbitrary (nonbinary) trees. Assume the trees are represented as in Figure
1.9.5. Prove that both these techniques still run in a time in the order of the number of
nodes in the tree to be traversed.
6.3 DEPTH-FIRST SEARCH : UNDIRECTED GRAPHS
Let G = < N, A > be an undirected graph all of whose nodes we wish to visit. Sup-
pose that it is somehow possible to mark a node to indicate that it has already been
visited. Initially, no nodes are marked.
To carry out a depth-first traversal of the graph, choose any node v E N as the
starting point. Mark this node to show that it has been visited. Next, if there is a node
adjacent to v that has not yet been visited, choose this node as a new starting point and
call the depth-first search procedure recursively. On return from the recursive call, if
there is another node adjacent to v that has not been visited, choose this node as the
next starting point, call the procedure recursively once again, and so on. When all the
nodes adjacent to v have been marked, the search starting at v is finished.
If there remain any nodes of G that have not been visited, choose any one of
them as a new starting point, and call the procedure yet again. Continue in this way
until all the nodes of G have been marked. Here is the recursive algorithm.
procedure search (G)
for each v E N do mark [v] f- not-visited
for each v E N do
if mark [v] # visited then dfs (v )
procedure dfs (v : node )
[ node v has not been visited }
mark [v] - visited
for each node w adjacent to v do
if mark [w] # visited then dfs (w )
The algorithm is called depth-first search since it tries to initiate as many recur-
sive calls as possible before it ever returns from a call. The recursivity is only stopped
when exploration of the graph is blocked and can go no further. At this point the
recursion "unwinds" so that alternative possibilities at higher levels can be explored.
Example 6.3.1. If we suppose that the neighbours of a given node are exam-
ined in numerical order, and that node 1 is the first starting point, a depth-first search
of the graph in Figure 6.3.1 progresses as follows :
dfs (1) initial call

dfs (2) recursive call
dfs (5) recursive call ; progress is blocked
dfs (4) a neighbour of node 1 has not been visited
dfs (8) recursive call; progress is blocked
there are no more nodes to visit. 0
Problem 6.3.1. Show how a depth-first search progresses through the graph in
Figure 6.3.1 if the neighbours of a given node are examined in numerical order but the
initial starting point is node 6.
How much time is needed to explore a graph with n nodes and a edges ? Since
each node is visited exactly once, there are n calls of the procedure dfs. When we
visit a node, we look at the mark on each of its neighbouring nodes. If the graph is
represented in such a way as to make the lists of adjacent nodes directly accessible
(type lisgraph of Section 1.9.2), this work is proportional to a in total. The algorithm
therefore takes a time in 0 (n) for the procedure calls and a time in 0 (a) to inspect the
marks. The execution time is thus in 0 (max(a, n)).
Figure 6.3.1. An undirected graph.

Problem 6.3.2. What happens if the graph is represented by an adjacency

matrix (type adjgraph of Section 1.9.2) rather than by lists of adjacent nodes ?
Problem 6.3.3. Show how depth-first search can be used to find the connected
components of an undirected graph.
A depth-first traversal of a connected graph associates a spanning tree to the

graph. The edges of the tree correspond to the edges used to traverse the graph ; they
are directed from the first node visited to the second. Edges that are not used in the
traversal of the graph have no corresponding edge in the tree. The initial starting point
of the exploration becomes the root of the tree.
Example 6.3.2. (Continuation of Example 6.3.1) The edges used in the

depth-first search of Example 6.3.1 are (1,2), 12, 31, (3,6), (6,5), (1,4), 14,7) and
(7,8). The corresponding directed edges (1, 2), (2, 3), and so on, form a spanning tree
for the graph in Figure 6.3.1. The root of the tree is node 1. See Figure 6.3.2.
If the graph being explored is not connected, a depth-first search associates to it

not merely a single tree, but rather a forest of trees, one for each connected component
of the graph.
A depth-first search also provides a way to number the nodes of the graph being
visited : the first node visited (the root of the tree) is numbered 1, the second is num-
bered 2, and so on. In other words, the nodes of the associated tree are numbered in
preorder. To implement this numbering, we need only add the following two state-
ments at the beginning of the procedure dfs :
pnum F- pnum + 1
prenum [v] E-- pnum
where pnum is a global variable initialized to zero.
Example 6.3.3. (Continuation of Example 6.3.1) The depth-first search illus-

trated by Example 6.3.1 numbers the nodes as follows :
node 1 2 3 4 5 6 7 8
prenum 1 2 3 6 5 4 7 8 11
Of course, the tree and the numbering generated by a depth-first search in a

graph are not unique, but depend on the chosen starting point and on the order in
which neighbours are visited.
Problem 6.3.4. Exhibit the tree and the numbering generated by the search of
Problem 6.3.1.
6.3.1 Articulation Points
A node v of a connected graph is an articulation point if the subgraph obtained by

deleting v and all the edges incident on v is no longer connected. For example, node
1 is an articulation point of the graph in Figure 6.3.1 ; if we delete it, there remain two
connected components (2,3,5,61 and (4,7, 8) . A graph G is biconnected
(or unarticulated) if it is connected and has no articulation points. It is bicoherent (or
isthmus free, or 2-edge-connected) if each articulation point is joined by at least two
edges to each component of the remaining sub-graph. These ideas are important in
practice : if the graph G represents, say, a telecommunications network, then the fact
that it is biconnected assures us that the rest of the network can continue to function
even if the equipment in one of the nodes fails ; if G is bicoherent, we can be sure that
all the nodes will be able to communicate with one another even if one transmission
line stops working.
The following algorithm finds the articulation points of a connected graph G .
a. Carry out a depth-first search in G, starting from any node. Let T be the tree
generated by the depth-first search, and for each node v of the graph, let
prenum [v ] be the number assigned by the search.
b. Traverse the tree T in postorder. For each node v visited, calculate lowest [v] as
the minimum of
i. prenum [v]
ii. prenum [w] for each node w such that there exists an edge [ v, w } in G
that has no corresponding edge in T
W. lowest [x] for every child x of v in T.
c. Articulation points are now determined as follows:
i. The root of T is an articulation point of G if and only if it has more than
one child.
H. A node v other than the root of T is an articulation point of G if and only
if v has a child x such that lowest [x] >_ prenum [v].
Example 6.3.4. (Continuation of Examples 6.3.1, 6.3.2, and 6.3.3) The search
described in Example 6.3.1 generates the tree illustrated in Figure 6.3.2. The edges of
G that have no corresponding edge in T are represented by broken lines. The value of
prenum [v] appears to the left of each node v, and the value of lowest [v] to the right.
The values of lowest are calculated in postorder, that is, for nodes 5, 6, 3, 2, 8, 7, 4,
and 1 successively. The articulation points of G are nodes 1 (by rule c(i)) and 4
(by rule c(ii)).
Problem 6.3.5. Verify that the same articulation points are found if we start
the search at node 6. 11
Figure 6.3.2. A depth-first search tree ; prenum on the left and lowest on the right.
Problem 6.3.6. Prove that an edge of G that has no corresponding edge in T

(a broken line in Figure 6.3.2) necessarily joins some node v to one of its ancestors
in T.
Informally, we can define lowest [v] by

lowest [v] = min(prenum [w] I you can get to w from v by following
down as many solid lines as you like and then
going up at most one broken line) .
If x is a child of v and if lowest [x] < prenum [v], there thus exists a chain of edges
that joins x to the other nodes of the graph even if v is deleted. On the other hand,
there is no chain joining x to the parent of v if v is not the root and if
lowest [x] >_ prenum [v].
Problem 6.3.7. Complete the proof that the algorithm is correct.
Problem 6.3.8. Show how to carry out the operations of steps (a) and (b) in
parallel and write the corresponding algorithm.
* Problem 6.3.9. Write an algorithm that decides whether or not a given con-
nected graph is bicoherent.
* Problem 6.3.10. Write an efficient algorithm that, given an undirected graph

that is connected but not biconnected, finds a set of edges that could be added to make
the graph biconnected. Your algorithm should find the smallest possible set of edges.
Analyse the efficiency of your algorithm.
Problem 6.3.11. Prove or give a counterexample :
i. If a graph is biconnected, then it is bicoherent.

ii. If a graph is bicoherent, then it is biconnected.
Problem 6.3.12. Prove that a node v in a connected graph is an articulation

point if and only if there exist two nodes a and b different from v such that every path
joining a and b passes through v.
Problem 6.3.13. Prove that for every pair of distinct nodes v and w in a
biconnected graph, there exist at least two chains of edges joining v and w that have
no nodes in common (except the starting and ending nodes).
6.4 DEPTH-FIRST SEARCH : DIRECTED GRAPHS
The algorithm is essentially the same as the one for undirected graphs, the difference
being in the interpretation of the word "adjacent". In a directed graph, node w is adja-
cent to node v if the directed edge (v , w) exists. If (v , w) exists and (w , v) does not,
then w is adjacent to v but v is not adjacent to w. With this change of interpretation
the procedures dfs and search from Section 6.3 apply equally well in the case of a
directed graph.
The algorithm behaves quite differently, however. Consider a depth-first search
of the directed graph in Figure 6.4.1. If the neighbours of a given node are examined
in numerical order, the algorithm progresses as follows :
1. dfs(1) initial call

2. dfs (2) recursive call
3. dfs (3) recursive call ; progress is blocked
4. dfs (4) a neighbour of node 1 has not been visited
5. dfs (8) recursive call
7. dfs (5) new starting point
9. there are no more nodes to visit.
Problem 6.4.1. Illustrate the progress of the algorithm if the neighbours of a

given node are examined in decreasing numerical order, and if the starting point is
node 1. 0
An argument identical with the one in Section 6.3 shows that the time taken by
this algorithm is also in 0 (max(a, n)). In this case, however, the edges used to visit
all the nodes of a directed graph G = <N, A > may form a forest of several trees even
Sec. 6.4 Depth-First Search : Directed Graphs 177
Figure 6.4.1. A directed graph.
if G is connected. This happens in our example : the edges used, namely (1, 2), (2, 3),
(1, 4), (4, 8), (8, 7), and (5, 6), form the forest shown by the solid lines in Figure 6.4.2.
(The numbers to the left of each node are explained in Section 6.4.2.)
Let F be the set of edges in the forest. In the case of an undirected graph the
edges of the graph with no corresponding edge in the forest necessarily join some node
to one of its ancestors (Problem 6.3.6). In the case of a directed graph three kinds of
edges can appear in A \ F (these edges are shown by the broken lines in Figure 6.4.2).
i. Those like (3, 1) or (7,4) that lead from a node to one of its ancestors ;
ii. those like (1, 8) that lead from a node to one of its descendants ; and
W. those like (5,2) or (6,3) that join one node to another that is neither its ancestor
nor its descendant. Edges of this type are necessarily directed from right to left.
Problem 6.4.2. Prove that if (v , w) is an edge of the graph that has no

corresponding edge in the forest, and if v is neither an ancestor nor a descendant of w
in the forest, then prenum [v] > prenum [w], where the values of prenum are attributed
as in Section 6.3.
Figure 6.4.2. A depth-first search forest.

6.4.1 Acyclic Graphs: Topological Sorting
Directed acyclic graphs can be used to represent a number of interesting relations.

This class of structures includes trees, but is less general than the class of all directed
graphs. For example, a directed acyclic graph can be used to represent the structure of
an arithmetic expression that includes repeated subexpressions : thus Figure 6.4.3
represents the structure of the expression
(a +b) (c +d) + (a +b) (c - d) .
Such graphs also offer a natural representation for partial orderings (such as the rela-
tion "smaller than" defined on the integers and the set-inclusion relation). Figure 6.4.4
illustrates part of another partial ordering defined on the integers. (What is the partial
ordering in question?) Finally, directed acyclic graphs are often used to represent the
different stages of a complex project : the nodes are different states of the project, from
the initial state to final completion, and the edges correspond to activities that have to
be completed to pass from one state to another. Figure 6.4.5 gives an example of this
type of diagram.
Depth-first search can be used to detect whether a given directed graph is acyclic.
Problem 6.4.3. Let F be the forest generated by a depth-first search on a

directed graph G = <N, A >. Prove that G is acyclic if and only if A \ F includes no
edge of type (i) (that is, from a node of G to one of its ancestors in the forest). 11
Figure 6.4.3 A directed acyclic graph.
Figure 6.4.4. Another directed acyclic graph.

make coffee drink coffee
dress D get out G
bring documents
Figure 6.4.5. Yet another directed acyclic graph.
A topological sort of the nodes of a directed acyclic graph is the operation of

arranging the nodes in order in such a way that if there exists an edge (i, j ), then i
precedes j in the list. For example, for the graph of Figure 6.4.4, the natural order 1,
2, 3, 4, 6, 8, 12, 24 is adequate; but the order 1, 3, 2, 6, 4, 12, 8, 24 is also acceptable,
as are several others. For graphs as in Figure 6.4.5, a topological ordering of the states
can be used to obtain a feasible schedule for the activities involved in the project; in
our example, the order A, B, F, C, E, D, G will serve.
The necessary modification to the procedure dfs to make it into a topological sort
is immediate : if we add a supplementary line
write v
at the end of the procedure, the numbers of the nodes will be printed in reverse topo-
logical order.
Problem 6.4.4. Prove this.
Problem 6.4.5. For the graph of Figure 6.4.4, what is the topological order
obtained if the neighbours of a node are visited in numerical order and if the depth-first
search begins at node I ?
6.4.2 Strongly Connected Components
A directed graph is strongly connected if there exists a path from u to v and also a
path from v to u for every distinct pair of nodes u and v. If a directed graph is not
strongly connected, we are interested in the largest sets of nodes such that the
corresponding subgraphs are strongly connected. Each of these subgraphs is called a
strongly connected component of the original graph. In the graph of Figure 6.4.1, for
instance, nodes 11, 2, 3) and the corresponding edges form a strongly connected com-
ponent. Another component corresponds to the nodes {4,7,8}. Despite the fact that
there exist edges (1,4) and (1,8), it is not possible to merge these two strongly con-
nected components into a single component because there exists no path from node 4
to node 1.
To detect the strongly connected components of a directed graph, we must first

modify the procedure dfs. In Section 6.3 we number each node at the instant when
exploration of the node begins. Here, we number each node at the moment when
exploration of the node has been completed. In other words, the nodes of the tree pro-
duced are numbered in postorder. To do this, we need only add at the end of pro-
cedure dfs the following two statements :
nump F nump + 1
postnum [v] - nump
where nump is a global variable initialized to zero. Figure 6.4.2 shows to the left of
each node the number thus assigned.
The following algorithm finds the strongly connected components of a directed
graph G .
i. Carry out a depth-first search of the graph starting from an arbitrary node. For
each node v of the graph let postnum [v] be the number assigned during the
search.
ii. Construct a new graph G' : G' is the same as G except that the direction of
every edge is reversed.
iii. Carry out a depth-first search in G'. Begin this search at the node w that has
the highest value of postnum. (If G contains n nodes, it follows that
postnum [w] = n.) If the search starting at w does not reach all the nodes,
choose as the second starting point the node that has the highest value of
postnum among all the unvisited nodes ; and so on.
iv. To each tree in the resulting forest there corresponds one strongly connected
component of G.
Example 6.4.1. On the graph of Figure 6.4.1, the first depth-first search
assigns the values of postnum shown to the left of each node in Figure 6.4.2. The
graph G' is illustrated in Figure 6.4.6, with the values of postnum shown to the left of
each node. We carry out a depth-first search starting from node 5, since
postnum [5] = 8; the search reaches nodes 5 and 6. For our second starting point, we
choose node 1, with postnum [1] = 6; this time the search reaches nodes 1, 3, and 2.
For the third starting point we take node 4, with postnum [4] = 5 ; this time the
remaining nodes 4, 7, and 8 are all reached. The corresponding forest is illustrated in
Figure 6.4.7. The strongly connected components of the original graph (Fig. 6.4.1) are
the subgraphs corresponding to the sets of nodes (5, 6), 11, 3, 21 and '{ 4, 7, 8 1.
* Problem 6.4.6. Prove that if two nodes u and v are in the same strongly con-
nected component of G, then they are in the same tree when we carry out the depth-
first search of G'.
Figure 6.4.6. Reversing the arrows in the graph of Figure 6.4.1.
0 O
O 0
O
Figure 6.4.7. The forest of strongly connected components.
It is harder to prove the result the other way. Let v be a node that is in the tree
whose root is r when we carry out the search of G', and suppose v #r. This implies
that there exists a path from r to v in G' ; thus there exists a path from v to r in G.
When carrying out the search of G', we always choose as a new starting point (that is,
as the root of a new tree) that node not yet visited with the highest value of postnum.
Since we chose r rather than v to be the root of the tree in question, we have
postnum [r] > postnum [v].
When we carried out the search in G, three possibilities seem to be open a priori:
r was an ancestor of v ;
r was a descendant of v ; or
r was neither an ancestor nor a descendant of v.
The second possibility is ruled out by the fact that postnum [r] > postnum [v]. In the
third case it would be necessary for the same reason that r be to the right of v.
However, there exists at least one path from v to r in G. Since in a depth-first
search the edges not used by the search never go from left to right (Problem 6.4.2),
any such path must go up the tree from v to a common ancestor (x, say) of v and r,
and then go down the tree to r. But this is quite impossible. We should have
postnum [x] > postnum [r] since x is an ancestor of r. Next, since there exists a path
from v to x in G, there would exist a path from x to v in G'. Before choosing r as
the root of a tree in the search of G', we would have already visited x (otherwise x
rather than r would be chosen as the root of the new tree) and therefore also v. This
contradicts the hypothesis that v is in the tree whose root is r when we carry out the
search of G. Only the first possibility remains : r was an ancestor of v when we
searched G. This implies that there exists a path from r to v in G.
We have thus proved that if node v is in the tree whose root is r when we carry
out the search of G', then there exist in G both a path from v to r and a path from r
to v. If two nodes u and v are in the same tree when we search G', they are therefore
both in the same strongly connected component of G since there exist paths from u to
v and from v to u in G via node r.
With the result of Problem 6.4.6, this completes the proof that the algorithm
works correctly.
Problem 6.4.7. Estimate the time and space needed by this algorithm.
6.5 BREADTH-FIRST SEARCH
When a depth-first search arrives at some node v, it next tries to visit some neighbour
of v, then a neighbour of this neighbour, and so on. When a breadth-first search
arrives at some node v, on the other hand, it first visits all the neighbours of v, and not
until this has been done does it go on to look at nodes farther away. Unlike depth-first
search, breadth-first search is not naturally recursive. To underline the similarities and
the differences between the two methods, we begin by giving a nonrecursive formula-
tion of the depth-first search algorithm. Let stack be a data type allowing two opera-
tions, push and pop. The type is intended to represent a list of elements that are to be
handled in the order "last come, first served". The function top denotes the element at
the top of the stack. Here is the modified depth-first search algorithm.
procedure dfs'(v : node)
P F empty-stack
mark [v] t- visited
push v on P
while P is not empty do
while there exists a node w adjacent to top (P)
such that mark [w] # visited do
mark [w] - visited
push w on P { w is now top (P) }
pop top (P)
For the breadth-first search algorithm, by contrast, we need a type queue that
allows two operations enqueue and dequeue. This type represents a list of elements
that are to be handled in the order "first come, first served". The function first denotes
the element at the front of the queue. Here now is the breadth-first search algorithm.
Sec. 6.5 Breadth-First Search 183
procedure bfs (v : node )

Q - empty-queue
mark [v] F-- visited
enqueue v into Q
while Q is not empty do
u <--first (Q )
dequeue u from Q
for each node w adjacent to u do
if mark [w] # visited then mark [w] t- visited
enqueue w into Q
In both cases we need a main program to start the search.

procedure search (G)
for each v E N do mark [v] - not-visited
for each v E N do
if mark [v] # visited then { dfs' or bfs } (v )
Example 6.5.1. On the graph of Figure 6.3.1, if the neighbours of a node are
visited in numerical order, and if node I is used as the starting point, breadth-first
search proceeds as follows.
Node Visited Q
1. 1 2,3,4
2. 2 3,4,5,6
3. 3 4,5,6
4. 4 5,6,7,8
5. 5 6,7,8
6. 6 7,8
7. 7 8
8. 8 -
As for depth-first search, we can associate a tree with the breadth-first search.
Figure 6.5.1 shows the tree generated by the search in Example 6.5.1. The edges of
the graph that have no corresponding edge in the tree are represented by broken lines.
Problem 6.5.1. After a breadth-first search in an undirected graph

G = <N, A >, let F be the set of edges that have a corresponding edge in the forest of
trees that is generated. Show that the edges { u , v } E A \ F are such that u and v are
in the same tree, but that neither u nor v is an ancestor of the other.
It is easy to show that the time required by a breadth-first search is in the same
order as that required by a depth-first search, namely 0 (max(a, n)). If the appropriate
Figure 6.5.1. A breadth-first search tree.
interpretation of the word "neighbouring" is used, the breadth-first search algorithm

can be applied without modification to either directed or undirected graphs.
Problem 6.5.2. Show how a breadth-first search progresses through the graph
of Figure 6.4.1, assuming that the neighbours of a node are always visited in numerical
order, and that necessary starting points are also chosen in numerical order.
Problem 6.5.3 (Continuation of Problem 6.5.2) Sketch the corresponding

forest and the remaining edges of the graph. How many kinds of "broken" edges are
possible ? (see Section 6.4)
Breadth-first search is most often used to carry out a partial exploration of certain
infinite graphs or to find the shortest path from one point to another.
6.6 IMPLICIT GRAPHS AND TREES
As mentioned at the outset of this chapter, various problems can be thought of in terms
of abstract graphs. For instance, we can use the nodes of a graph to represent
configurations in a game of chess and edges to represent legal moves (see Sec-
tion 6.6.2). Often the original problem translates to searching for a specific node, path
or pattern in the associated graph. If the graph contains a large number of nodes, it
may be wasteful or infeasible to build it explicitly in computer memory before
applying one of the search techniques we have encountered so far.
An implicit graph is one for which we have available a description of its nodes
and edges. Relevant portions of the graph can thus be built as the search progresses.
Therefore computing time is saved whenever the search succeeds before the entire
graph has been constructed. The economy in memory space is even more dramatic
when nodes that have already been searched can be discarded, making room for subse-
quent nodes to be explored.
Backtracking is a basic search technique on implicit graphs. One powerful appli-
cation is in playing games of strategy by techniques known as minimax and alpha-beta
Sec. 6.6 Implicit Graphs and Trees 185
pruning. Some optimization problems can be handled by the more sophisticated

branch-and-bound technique. We now discuss these notions.
6.6.1 Backtracking
Backtracking algorithms use a special technique to explore implicit directed graphs.

These graphs are usually trees, or at least they contain no cycles. A backtracking algo-
rithm carries out a systematic search, looking for solutions to some problem. At least
one application of this technique dates back to antiquity : it allows one to find the way
through a labyrinth without danger of going round and round in circles. To illustrate
the general principle, we shall, however, use a different example. Consider the classic
problem of placing eight queens on a chess-board in such a way that none of them
threatens any of the others. (A queen threatens the squares in the same row, in the
same column, or on the same diagonals.)
Problem 6.6.1. Solve this problem without using a computer.
The first obvious way to solve this problem consists of trying systematically all
the ways of placing eight queens on a chess-board, checking each time to see whether
a solution has been obtained. This approach is of no practical use, since the number of
positions we have to check would be [64 J = 4,426,165,368. The first improvement we
a than one queen in any given row. This
might try consists of never putting more
reduces the computer representation of the chess-board to a simple vector of eight ele-
ments, each giving the position of the queen in the corresponding row. For instance,
the vector (3, 1, 6, 2, 8, 6, 4, 7) does not represent a solution since the queens in the third
and the sixth rows are in the same column, and also two pairs of queens lie on the
same diagonal. Using this representation, we can write the algorithm very simply
using eight nested loops.
program Queens 1
for ii-lto 8do
for i2-lto 8do
for i8-lto 8do

try F-(iI,i2,..., i8)
if solution (try) then write try
stop
write "there is no solution"
This time, the number of cases to be considered is reduced to 88 = 16,777, 216,

although in fact the algorithm finds a solution and stops after considering only
1,299,852 cases.
Problem 6.6.2. If you have not yet solved the previous problem, the informa-
tion just given should be of considerable help ! 0
Once we have realized that the chess-board can be represented by a vector, which
prevents us from ever trying to put two queens in the same row, it is natural to be
equally systematic in our use of the columns. Hence we now represent the board by a
vector of eight different numbers between 1 and 8, that is, by a permutation of the first
eight integers. The algorithm becomes
program Queens 2
try initial permutation
while try # final-permutation and not solution (try) do
try F- next-permutation
if solution (try) then write try
else write "there is no solution"
There are several natural ways to generate systematically all the permutations of
the first n integers. For instance, we might put each value in turn in the leading posi-
tion and generate recursively, for each of these leading values, all the permutations of
the n -1 remaining elements.
procedure perm (i)
if i = n then use (T) { T is a new permutation }
else for j F- i to n do exchange T [i] and T [ j ]
perm (i +1)
exchange T [i] and T [ j ]
Here T [ 1 .. n ] is a global array initialized to [ 1, 2, ... , n ] and the initial call is

perm (1).
Problem 6.6.3. If use (T) consists simply of printing the array T on a new
line, show the result of calling perm (1) when n = 4. 0
Problem 6.6.4. Assuming that use (T) takes constant time, how much time is
needed, as a function of n , to execute the call perm (1) ? Now rework the problem
assuming that use (T) takes a time in 0(n). 0
This approach reduces the number of possible cases to 8! = 40,320. If the
preceding algorithm is used to generate the permutations, only 2,830 cases are in fact
considered before the algorithm finds a solution. Although it is more complicated to
generate permutations rather than all the possible vectors of eight integers between 1
and 8, it is, on the other hand, easier in this case to verify whether a given position is a
solution. Since we already know that two queens can neither be in the same row nor in
the same column, it suffices to verify that they are not in the same diagonal.
Starting from a crude method that tried to put the queens absolutely anywhere on
the chess-board, we progressed first to a method that never puts two queens in the
same row, and then to a better method still where the only positions considered are
those where we know that two queens can neither be in the same row nor in the same
column. However, all these algorithms share an important defect : they never test a
position to see if it is a solution until all the queens have been placed on the board.
For instance, even the best of these algorithms makes 720 useless attempts to put the
last six queens on the board when it has started by putting the first two on the main
diagonal, where of course they threaten one another !
Backtracking allows us to do better than this. As a first step, let us reformulate
the eight queens problem as a tree searching problem. We say that a vector V [1 .. k]
of integers between I and 8 is k-promising, for 0S k 5 8, if none of the k queens
placed in positions (1, V [1]), (2, V [2]) , .. , (k, V [k]) threatens any of the others.
Mathematically, a vector V is k-promising if, for every i #j between 1 and k, we
have V[i] -V [j ] it (i - j , 0, j - i ). For k <- 1, any vector V is k -promising. Solu-
tions to the eight queens problem correspond to vectors that are 8-promising.
Let N be the set of k-promising vectors, 0 5 k S 8. Let G = <N, A > be the
directed graph such that (U, V ) E A if and only if there exists an integer k, 0 <_ k < 8,
such that U is k -promising, V is (k +l )-promising, and U[i]=V[i] for every
i E [1 .. k]. This graph is a tree. Its root is the empty vector (k =0). Its leaves are
either solutions (k = 8) or else they are dead ends (k < 8) such as [1, 4, 2, 5, 8] where it
is impossible to place a queen in the next row without her threatening at least one of
the queens already on the board. The solutions to the eight queens problem can be
obtained by exploring this tree. We do not generate the tree explicitly so as to explore
it thereafter, however : rather, nodes are generated and abandoned during the course of
the exploration. Depth-first search is the obvious method to use, particularly if we
only require one solution.
This technique has two advantages over the previous algorithm that systemati-
cally tried each permutation. First, the number of nodes in the tree is less than
8! = 40,320. Although it is not easy to calculate this number theoretically, it is
straightforward to count the nodes using a computer : #N = 2057. In fact, it suffices to
explore 114 nodes to obtain a first solution. Secondly, in order to decide whether a
vector is k-promising, knowing that it is an extension of a (k -1)-promising vector, we
only need to check the last queen to be added. This check can be speeded up if we
associate with each promising node the sets of columns, of positive diagonals (at 45
degrees), and of negative diago^,11s (at 135 degrees) controlled by the queens already
placed. On the other hand, to decide if some given permutation represents a solution,
it seems at first sight that we have to check each of the 28 pairs of queens on the
board.
To print all the solutions to the eight queens problem, call Queens (0, 0, 0, 0),
where try [I .. 8] is a global array.
procedure Queens (k, col, diag45, diag 135)

{ try [ 1 .. k ] is k -promising,
col={try [i]I 1!5 i:5k},
diag45 = {try[i]-i +l I 1 !5i !5k), and
diag135={try [i]+i-1I 15i <-k} }
if k = 8
then { an 8-promising vector is a solution }
write try
else { explore (k+1)-promising extensions of try }
for j F- 1 to 8 do
if j Ocol and j -k Odiag45 and j +k Odiag 135
then try [k +I] F- j
{ try (L. k +11 is (k +1)-promising }
Queens (k +1, col u { j }, diag45 u { j -k }, diag 135 u { j +k))
It is clear that the problem generalizes to an arbitrary number of queens : how

can we place n queens on an n x n "chess-board" in such a way that none of them
threatens any of the others ?
Problem 6.6.5. Show that the problem for n queens may have no solution.
Find a more interesting case than n = 2.
As we might expect, the advantage to be gained by using the backtracking algo-

rithm instead of an exhaustive approach becomes more pronounced as n increases. For
example, for n = 12 there are 479,001,600 possible permutations to be considered, and
the first solution to be found (using the generator given previously) corresponds to the
4,546,044th position examined ; on the other hand, the tree explored by the back-
tracking algorithm contains only 856,189 nodes, and a solution is obtained when the
262nd node is visited.
** Problem 6.6.6. Analyse mathematically, as a function of the number n of

queens, the number of nodes in the tree of k -promising vectors. How does this
number compare to n! ?
Backtracking algorithms can also be used even when the solutions sought do not
necessarily all have the same length. Here is the general scheme.
procedure backtrack (v [ 1 .. k ])
{ v is a k -promising vector)
if v is a solution then write v
{ otherwise } for each (k +1)-promising vector w
such that w [1 .. k ] = v [I .. k ] do backtrack (w [I .. k +1])
The otherwise should be present if and only if it is impossible to have two different
solutions such that one is a prefix of the other.
Problem 6.6.7. Instant Insanity is a puzzle consisting of four coloured cubes.

Each of the 24 faces is coloured blue, red, green, or white. The four cubes are to be
placed side by side in such a way that each colour appears on one of the four top faces,
on one of the four bottom faces, on one of the front faces, and on one of the rear faces.
Show how to solve this problem using backtracking.
The n-queens problem was solved using depth-first search in the corresponding
tree. Some problems that can be formulated in terms of exploring an implicit graph
have the property that they correspond to an infinite graph. In this case, it may be
necessary to use breadth-first search to avoid the interminable exploration of some
fruitless infinite branch. Breadth-first search is also appropriate if we have to find a
solution starting from some initial position and making as few changes as possible.
(This last constraint does not apply to the eight queens problem where each solution
involves exactly the same number of pieces.) The two following problems illustrate
these ideas.
Problem 6.6.8. Give an algorithm capable of transforming some initial integer

n into a given final integer m by the application of the smallest possible number of
transformations f(i) = 3i and g(i) = Li/21. For instance, 15 can be transformed into 4
using four function applications : 4 = gfgg (15). What does your algorithm do if it is
impossible to transform n into m in this way ?
* Problem 6.6.9. Give an algorithm that determines the shortest possible series
of manipulations needed to change one configuration of Rubik's Cube into another.
If the required change is impossible, your algorithm should say so rather than calcu-
lating forever.
Problem 6.6.10. Give other applications of backtracking.
6.6.2 Graphs and Games: An Introduction
Most games of strategy can be represented in the form of directed graphs.

A node of the graph corresponds to a particular position in the game, and an edge
corresponds to a legal move between two positions. The graph is infinite if there is no
a priori limit on the number of positions possible in the game. For simplicity, we
assume that the game is played by two players, each of whom moves in turn, that the
game is symmetric (the rules are the same for both players), and that chance plays no
part in the outcome (the game is deterministic). The ideas we present can easily be
adapted to more general contexts. We further suppose that no instance of the game can
last forever and that no position in the game offers an infinite number of legal moves
to the player whose turn it is. In particular, some positions in the game offer no legal
moves, and hence some nodes in the graph have no successors : these are the terminal
positions.
To determine a winning strategy for a game of this kind, we need only attach to
each node of the graph a label chosen from the set win, lose, draw. The label
corresponds to the situation of a player about to move in the corresponding position,
assuming that neither player will make an error. The labels are assigned systematically
in the following way.
i. The labels assigned to terminal positions depend on the game in question. For
most games, if you find yourself in a terminal position, then there is no legal
move you can make, and you have lost ; but this is not necessarily the case (think
of stalemate in chess).
ii. A nonterminal position is a winning position if at least one of its successors is a
losing position.
iii. A nonterminal position is a losing position if all of its successors are winning
positions.
iv. Any remaining positions lead to a draw.
Problem 6.6.11. Grasp intuitively how these rules arise. Can a player who
finds himself in a winning position lose if his opponent makes an "error"?
Problem 6.6.12. In the case of an acyclic finite graph (corresponding to a

game that cannot continue for an indefinite number of moves), find a relationship
between this method of labelling the nodes and topological sorting (Section 6.4.1).
We illustrate these ideas with the help of a variant of Nim (also known as the
Marienbad game). Initially, at least two matches are placed on the table between two
players. The first player removes as many matches as he likes, except that he must
take at least one and he must leave at least one. Thereafter, each player in turn must
remove at least one match and at most twice the number of matches his opponent just
took. The player who removes the last match wins. There are no draws.
Example 6.6.1. There are seven matches on the table initially. If I take two
of them, my opponent may take one, two, three, or four. If he takes more than one, I
can remove all the matches that are left and win. If he takes only one match, leaving
four matches on the table, I can in turn remove a single match, and he cannot prevent
me from winning on my next turn. On the other hand, if at the outset I choose to
remove a single match, or to remove more than two, then you may verify that my
opponent has a winning strategy.
The player who has the first move in a game with seven matches is therefore cer-
tain to win provided that he does not make an error. On the other hand, you may
verify that a player who has the first move in a game with eight matches cannot win
unless his opponent makes an error.
* Problem 6.6.13. For n >- 2, give a necessary and sufficient condition on n

to ensure that the player who has the first move in a game involving n matches have
a winning strategy. Your characterization of n should be as simple as possible. Prove
your answer.
A position in this game is not specified merely by the number of matches that
remain on the table. It is also necessary to know the upper limit on the number of
matches that it is permissible to remove on the next move. The nodes of the graph
corresponding to this game are therefore pairs . In general, , 1 <_ j <- i ,
indicates that i matches remain on the table and that any number of them between 1
and j may be removed in the next move. The edges leaving node go to the j
nodes , 1 k <- j. The node corresponding to the initial position
in a game with n matches, n 2, is < n , n -1 >. All the nodes whose second com-
ponent is zero correspond to terminal positions, but only < 0, 0 > is interesting : the
nodes < i, 0 > for i > 0 are inaccessible. Similarly, nodes < i, j > with j odd and
j < i -1 cannot be reached starting from any initial position.
Figure 6.6.1 shows part of the graph corresponding to this game. The square
nodes represent losing positions and the round nodes are winning positions. The heavy
edges correspond to winning moves : in a winning position, choose one of the heavy
edges in order to win. There are no heavy edges leaving a losing position,
corresponding to the fact that such positions offer no winning move.
We observe that a player who has the first move in a game with two, three, or
five matches has no winning strategy, whereas he does have such a strategy in the
game with four matches.
Problem 6.6.14. Add nodes < 8, 7 >, < 7, 6 >, < 6, 5 > and their descendants to
the graph of Figure 6.6.1.
Figure 6.6.1. Part of a game graph.

Problem 6.6.15. Can a winning position have more than one losing position
among its successors ? In other words, are there positions in which several different
winning moves are available? Can this happen in the case of a winning initial position
<n,n-1>?
The obvious algorithm to determine whether a position is winning is the fol-
lowing.
function rec (i, j )
{ returns true if and only if the node < i, j > is winning ;
we assume that 0 j <- i }
for k F- 1 to j do
if not rec (i -k, min(2k, i -k))
then return true
return false
Problem 6.6.16. Modify this algorithm so that it returns an integer k such that
k = 0 if the position is a losing position and 1 <- k <- j if it is a winning move to take
away k matches.
This algorithm suffers from the same defect as the algorithm fib! in Section
1.7.5: it calculates the same value over and over. For instance, rec (5, 4) returns false
having called successively rec (4, 2), rec (3, 3), rec (2, 2) and rec (1, 1), but rec (3, 3)
also calls rec (2, 2) and rec (1, 1).
Problem 6.6.17. Find two ways to remove this inefficiency. (If you want to
work on this problem, do not read the following paragraphs yet !)
The first approach consists of using dynamic programming to create a Boolean

array G such that G [i, j ] = true if and only if < i, j > is a winning position. As
usual with dynamic programming, we proceed in a bottom-up fashion, calculating all
the values of G [1, k ] for 1 <- k <- I is winning }
G [0, 0] E- false
for i f- 1 to n do
for j f- 1 to i do
k E- !
while k < j and G [i -k, min(2k, i -k)] do k E- k + 1
G [i , j ] F- not G [i -k , min(2k , i -k )]
Problem 6.6.18. The preceding algorithm only _ises G [0, 0] and the values of
G [1, k], 1 <- k <- l < i , to calculate G [i, j]. Show how to improve its efficiency by
also using the values of G [i , k ] for 1 <- k < j.
In this context dynamic programming leads us to calculate wastefully some

entries of the array G that are never needed. For instance, we know that < 15,14 > is a
winning position as soon as we discover that its second successor < 13, 4 > is a losing
position. It is no longer of any interest to know whether the next successor < 12, 6 > is
a winning or a losing position. In fact, only 27 nodes are really useful when we calcu-
late G [15, 14], although the dynamic programming algorithm looks at 121 of them.
About half this work can be avoided if we do not calculate G [i , j ] whenever j is odd
and j < i -1, since these nodes are never of interest, but there is no "bottom-up" reason
for not calculating G [12,6].
The recursive algorithm given previously is inefficient because it recalculates the
same value several times. Because of its top-down nature, however, it never calculates
an unnecessary value. A solution that combines the advantages of both the algorithms
consists of using a memory function (Section 5.7). This involves remembering which
nodes have already been visited during the recursive computation using a global
Boolean array init [0.. n, 0.. n ], initialized to false, where n is an upper bound on the
number of matches to be used.
function nim (i, j )
if init [i, j] then return G [i, j]
init [i, j] - true
for k - 1 to j do
if not nim (i -k, min(2k, i -k)) then G [i, j] E- true
return true
G [i, j] false
return false
At first sight, there is no particular reason to favour this approach over dynamic
programming, because in any case we have to take the time to initialize the whole
array init [0.. n, 0.. n ]. Using the technique suggested in Problem 5.7.2 allows us,
however, to avoid this initialization and to obtain a worthwhile gain in efficiency.
The game we have considered up to now is so simple that it can be solved
without really using the associated graph. Here, without explanation, is an algorithm
for determining a winning strategy that is more efficient than any of those given previ-
ously. In an initial position with n matches, first call precond (n). Thereafter a call on
whatnow (i , j ), 1 <- j <- i, determines in a time in O(1) the move to make in a situation
where i matches remain on the table and the next player has the right to take at most j
of them. The array T [0.. n ] is global. The initial call of precond (n) is an application
of the preconditioning technique to be discussed in the next chapter.
procedure precond (n)

T [0] t--
for i F 1 to n do
k<-1
while T [i -k ] <- 2k do k F k + 1
T[i] - k
function whatnow (i, j )
if j < T [i] then { prolong the agony ! }
return 1
return T [i]
* Problem 6.6.19. Prove that this algorithm works correctly and that
precond (n) takes a time in 0(n).
Consider now a more complex game, namely chess. At first sight, the graph
associated with this game contains cycles, since if two positions u and v of the pieces
differ only by the legal move of a rook, say, the king not being in check, then we can
move equally well from u to v and from v to u. However, this problem disappears on
closer examination. Remember first that in the game we just looked at, a position is
defined not merely by the number of matches on the table, but also by an invisible item
of information giving the number of matches that can be removed on the next move.
Similarly, a position in chess is not defined simply by the positions of the pieces on the
board. We also need to know whose turn it is to move, which rooks and kings have
moved since the beginning of the game (to know if it is legal to castle), and whether
some pawn has just been moved two squares forward (to know whether a capture en
passant is possible). Furthermore, the International Chess Federation has rules that
prevent games dragging on forever : for example, a game is declared to be a draw after
50 moves in which no irreversible action (movement of a pawn, or a capture) took
place. Thus we must include in our notion of position the number of moves made
since the last irreversible action. Thanks to such rules, there are no cycles in the graph
corresponding to chess. (For simplicity we ignore exceptions to the 50-move rule, as
well as the older rule that makes a game a draw if the pieces return three times to
exactly the same positions on the board.)
Adapting the general rules given at the beginning of this section, we can there-
fore label each node as being a winning position for White, a winning position for
Black, or a draw. Once constructed, this graph allows us to play a perfect game of
chess, that is, to win whenever it is possible and to lose only when it is inevitable.
Unfortunately (or perhaps fortunately for the game of chess), the graph contains so
many nodes that it is quite out of the question to explore it completely, even with the
fastest existing computers.
* Problem 6.6.20. Estimate the number of ways in which the pieces can be
placed on a chess-board. For simplicity ignore the fact that certain positions are
impossible, that is, they can never be obtained from the initial position by a legal series
of moves (but take into account the fact that each bishop moves only on either white or
black squares, and that both kings must be on the board). Ignore also the possibility of
having promoted pawns. 11
Since a complete search of the graph associated with the game of chess is out of
the question, it is not practical to use a dynamic programming approach. In this situa-
tion the recursive approach comes into its own. Although it does not allow us to be
certain of winning, it underlies an important heuristic called minimax. This technique
finds a move that may reasonably be expected to be among the best moves possible
while exploring only a part of the graph starting from some given position. Explora-
tion of the graph is usually stopped before the leaves are reached, using one of several
possible criteria, and the positions thus reached are evaluated heuristically. Then we
make the move that seems to cause our opponent the most trouble. This is in a sense
merely a systematic version of the method used by some human players that consists
of looking ahead a small number of moves. Here we give only an outline of the tech-
nique.
The minimax principle. The first step is to define a static evaluation func-
tion eval that attributes some value to each possible position. Ideally, we want the
value of eval (u) to increase as the position u becomes more favourable to White. It is
customary to give values not too far from zero to positions where neither side has a
marked advantage, and large negative values to positions that favour Black. This
evaluation function must take account of many factors : the number and the type of
pieces remaining on both sides, control of the centre, freedom of movement, and so on.
A compromise must be made between the accuracy of this function and the time
needed to calculate it. When applied to a terminal position, the evaluation function
should return +oo if Black has been mated, -oc if White has been mated, and 0 if the
game is a draw. For example, an evaluation function that takes good account of the
static aspects of the position but that is too simplistic to be of real use might be the fol-
lowing : for nonterminal configurations, count 1 point for each white pawn, 31/4 points
for each white bishop or knight, 5 points for each white rook, and 10 points for each
white queen ; subtract a similar number of points for each black piece.
If the static evaluation function were perfect, it would be easy to determine the
best move to make. Suppose it is White's turn to move from position u. The best
move would be to go to the position v that maximizes eval (v) among all the succes-
sors w of u.
val F-oo
for each configuration w that is a successor of u do
if eval(w) >- val then val - eval(w)
V f- W
It is clear that this simplistic approach would not be very successful using the evalua-
tion function suggested earlier, since it would not hesitate to sacrifice a queen in order
to take a pawn !
If the evaluation function is not perfect, a better strategy for White is to assume
that Black will reply with the move that minimizes the function eval, since the smaller
the value taken by this function, the better the position is supposed to be for him.
(Ideally, he would like a large negative value.) We are now looking half a move
ahead.
val --°°
if w has no successor
then valw - eval(w)
else valw - min{eval(x) I x is a successor of w }
if valw >- val then val F- valw
v t-W
There is now no question of giving away a queen to take a pawn, which of course may
be exactly the wrong rule to apply if it prevents White from finding the winning move :
maybe if he looked further ahead the gambit would turn out to be profitable. On the
other hand, we are sure to avoid moves that would allow Black to mate immediately
(provided we can avoid this).
To add more dynamic aspects to the static evaluation provided by eval, it is
preferable to look several moves ahead. To look n half-moves ahead from position U,
White should move to the position v given by
val <- -°°

if Black (w, n) >- val then val f- Black (w, n)
V f- W
where the functions Black and White are the following :
function Black (w, n)

if n = 0 or w has no successor
then return eval(w)
else return min { White (x, n -1) 1 x is a successor of w }
function White (x, n)
if n = 0 or x has no successor
then return eval(x)
else return max { Black (w, n -1) 1 w is a successor of x }
We see why the technique is called minimax : Black tries to minimize the advan-
tage he allows to White, and White, on the other hand, tries to maximize the advantage
he obtains from each move.
Player Rule
A max
Figure 6.6.2. The minimax principle.
Problem 6.6.21. Let u correspond to the initial position of the pieces. What
can you say about White (u, 12800), besides the fact that it would take far too long to
calculate in practice? Justify your answer.
Example 6.6.2. Figure 6.6.2 shows part of the graph corresponding to some
game. If the values attached to the leaves are obtained by applying the function eval
to the corresponding positions, the values for the other nodes can be calculated using
the minimax rule. In the example we suppose that player A is trying to maximize the
evaluation function and that player B is trying to minimize it.
If A plays so as to maximize his advantage, he will choose the second of the
three possible moves. This assures him of a value of at least 10.
Alpha-beta pruning. The basic minimax technique can be improved in a

number of ways. For example, it may be worthwhile to explore the most promising
moves in greater depth. Similarly, the exploration of certain branches can be aban-
doned early if the information we have about them is already sufficient to show that
they cannot possibly influence the values of nodes farther up the tree. This second
type of improvement is generally known as alpha-beta pruning. We give just one
simple example of the technique.
Example 6.6.3. Look back at Figure 6.6.2. Let < i, j > represent the j th
node in the i th row of the tree. We want to calculate the value of the root < 1, 1 >
starting from the values calculated by the function eval for the leaves < 4, j >,
1 <_ j 5 18. To do this, we carry out a bounded depth-first search in the tree, visiting
the successors of a given node from left to right.
If we want to abandon the exploration of certain branches because it is no longer

useful, we have to transmit immediately to the higher levels of the tree any information
obtained by evaluating a leaf. Thus as soon as the first leaf <4, 1> is evaluated, we
know that <4, 1> has value -7 and that <3, 1> (a node that maximizes eval) has
value at least -7. After evaluation of the second leaf <4,2>, we know that <4,2>
has value 5, < 3, 1 > has value 5, and < 2, 1 > (a node that minimizes eval) has value at
most 5.
Continuing in this way, we arrive after evaluation of the leaf <4,4> at the situa-
tion illustrated in Figure 6.6.3. Since node <3, 3 > has value at least 10, whereas node
< 2, 1 > has value at most -3, the exact value of node < 3, 3 > cannot have any influence
on the value of node < 2, 1 >. It is therefore unnecessary to evaluate the other descen-
dants of node <3,3>; we say that the corresponding branches of the tree have been
pruned.
Similarly, after evaluation of the leaf <4, 11>, we are in the situation shown in
Figure 6.6.4. Node < 2, 3 > has value at most 1. Since we already know that the value
of < 1, I> is at least 10, there is no need to evaluate the other children of node < 2, 3 >.
To establish that the value of the root <1,1> is 10, we visit only 19 of the 31
nodes in the tree.
** Problem 6.6.22. Write a program capable of playing brilliantly your favourite
game of strategy. 11
** Problem 6.6.23. Write a program that can beat the world backgammon cham-
pion. (This has already been done !)
** Problem 6.6.24. What modifications should be made to the principles set out
in this section to take account of those games of strategy in which chance plays a
certain part ? What about games with more than two players ?
Player Rule
A max
B min
A max
B eva!
-7 5 -3 10 ? ?
Figure 6.6.3. Alpha-beta pruning.

Player Rule
A max
It is not necessary
to explore these
branches
B eval
Figure 6.6.4. More alpha-beta pruning.
6.6.3 Branch-and-Bound
Like backtracking, branch-and-bound is a technique for exploring an implicit directed

graph. Again, this graph is usually acyclic or even a tree. This time, we are looking
for the optimal solution to some problem. At each node we calculate a bound on the
possible value of any solutions that might happen to be farther on in the graph. If the
bound shows that any such solution must necessarily be worse than the best solution
we have found so far, then we do not need to go on exploring this part of the graph.
In the simplest version, calculation of these bounds is combined with a breadth-
first or a depth-first search, and serves only, as we have just explained, to prune certain
branches of a tree or to close certain paths in a graph. More often, however, the calcu-
lated bound is used not only to close off certain paths, but also to choose which of the
open paths looks the most promising, so that it can be explored first.
In general terms we may say that a depth-first search finishes exploring nodes in
inverse order of their creation, using a stack to hold those nodes that have been gen-
erated but not yet explored fully ; a breadth-first search finishes exploring nodes in the
order of their creation, using this time a queue to hold those that have been generated
but not yet explored (see Section 6.5). Branch-and-bound uses auxiliary computations
to decide at each instant which node should be explored next, and a priority list to hold
those nodes that have been generated but not yet explored.
An example illustrates the technique.
Example 6.6.4. We return to the travelling salesperson problem (see Sections

3.4.2 and 5.6).
Let G be the complete graph on five points with the following distance matrix :
014 4 10 20
14 0 7 8 7
4 5 0 7 16
11 7 9 0 2
18 7 17 4 0
We are looking for the shortest tour starting from node 1 that passes exactly once
through each other node before finally returning to node 1.
The nodes in the implicit graph correspond to partially specified paths. For
instance, node (1,4,3) corresponds to two complete tours : (1,4,3,2,5, 1) and
(1, 4, 3, 5, 2, 1). The successors of a given node correspond to paths in which one addi-
tional node has been specified. At each node we calculate a lower bound on the length
of the corresponding complete tours.
To calculate this bound, suppose that half the distance between two points i and
j is counted at the moment we leave i, and the other half when we arrive at j. For
instance, leaving node I costs us at least 2, namely the lowest of the values 14/2, 4/2,
10/2, and 20/2. Similarly, visiting node 2 costs us at least 6 (at least 5/2 when we
arrive and at least 7/2 when we leave). Returning to node 1 costs at least 2, the
minimum of 14/2, 4/2, 11/2, and 18/2. To obtain a bound on the length of a path, it
suffices to add elements of this kind. For instance, a complete tour must include a
departure from node 1, a visit to each of the nodes 2, 3, 4, and 5 (not necessarily in
this order) and a return to 1. Its length is therefore at least
2+6+4+3+3+2= 20.
Notice that this calculation does not imply the existence of a solution that costs
only 20.
In Figure 6.6.5 the root of the tree specifies that the starting point for our tour is
node 1. Obviously, this arbitrary choice of a starting point does not alter the length of
the shortest tour. We have just calculated the lower bound shown for this node. (This
bound on the root of the implicit tree serves no purpose in the algorithm ; it was com-
puted here for the sake of illustration.) Our search begins by generating (as though for
a breadth-first search) the four possible successors of the root, namely, nodes (1, 2),
(1,3), (1,4), and (1,5). The bound for node (1,2), for example, is calculated as fol-
lows. A tour that begins with (1, 2) must include
The trip 1-2 : 14 (formally, leaving I for 2 and arriving at 2 from 1: 7+7 )
A departure from 2 toward 3, 4, or 5: minimum 7/2
A visit to 3 that neither comes from I nor leaves for 2: minimum 11/2
A similar visit to 4: minimum 3
A return to 1 from 3, 4, or 5: minimum 2
Bound 20
1, 2
Bound 31
:Z 1, 5
Bound 41
Figure 6.6.5. Branch-and-bound.
The length of such a tour is therefore at least 31. The other bounds are calculated
similarly.
Next, the most promising node seems to be (1, 3), whose bound is 24. The three
children (1,3,2), (1,3,4), and (1,3,5) of this node are therefore generated. To give
just one example, we calculate the bound for node (1, 3, 2) as follows :
The trip 1- 3 -2: 9

A departure from 2 toward 4 or 5: minimum 7/2
A visit to 4 that comes from neither 1 nor 3 and that leaves for neither 2 nor 3:
minimum 3
A return to 1 from 4 or 5: minimum 11/2,
which gives a total length of at least 24.

The most promising node is now (1, 3, 2). Its two children (1,3,2,4) and
(1,3,2,5) are generated. This time, as node (1,3,2,4), for instance, corresponds to
exactly one complete tour (1,3,2,4,5, 1), we do not need to calculate a lower bound
since we may calculate immediately its length 37.
We find that the length of the tour (1, 3, 2, 5, 4, 1) is 31. If we are only concerned
to find one optimal solution, we do not need to continue exploration of the nodes (1, 2),
(1, 5) and (1, 3, 5), which cannot possibly lead to a better solution. Even exploration of
the node (1,3,4) is pointless. (Why?) There remains only node (1,4) to explore.
The only child to offer interesting possibilities is (1,4,5). After looking at the two
complete tours (1, 4, 5, 2, 3, 1) and (1, 4, 5, 3, 2, 1), we find that the tour (1, 4, 5, 2, 3, 1)
of length 30 is optimal. This example illustrates the fact that although at one point
(1, 3) was the most promising node, the optimal solution does not come from there.
To obtain our answer, we have looked at merely 15 of the 41 nodes that are
present in a complete tree of the type illustrated in Figure 6.6.5.
Problem 6.6.25. Solve the same problem using the method of Section 5.6.
* Problem 6.6.26. Implement this algorithm on a computer and test it on our

example.
Problem 6.6.27. Show how to solve the same problem using a backtracking
algorithm that calculates a bound as shown earlier to decide whether or not a partially
defined path is promising.
The need to keep a list of nodes that have been generated but not yet completely
explored, situated in all the levels of the tree and preferably sorted in order of the
corresponding bounds, makes branch-and-bound quite hard to program. The heap is an
ideal data structure for holding this list. Unlike depth-first search and its related tech-
niques, no elegant recursive formulation of branch-and-bound is available to the pro-
grammer. Nevertheless, the technique is sufficiently powerful that it is often used in
practical applications.
It is next to impossible to give any idea of how well the technique will perform
on a given problem using a given bound. There is always a compromise to be made
concerning the quality of the bound to be calculated : with a better bound we look at
less nodes, but on the other hand, we shall most likely spend more time at each one
calculating the corresponding bound. In the worst case it may turn out that even an
excellent bound does not allow us to cut any branches off the tree, and all the extra
work we have done is wasted. In practice, however, for problems of the size encoun-
tered in applications, it almost always pays to invest the necessary, time in calculating
the best possible bound (within reason). For instance, one finds applications such as
integer programming handled by branch-and-bound, the bound at each node being
obtained by solving a related problem in linear programming with continuous vari-
ables.
Problem 6.7.1. Write algorithms to determine whether a given undirected

graph is in fact a tree (i) using a depth-first search ; (ii) using a breadth-first search.
How much time do your algorithms take?
Problem 6.7.2. Write an algorithm to determine whether a given directed

graph is in fact a rooted tree, and if so, to find the root. How much time does your
algorithm take?
*Problem 6.7.3. A node p of a directed graph G = <N, A > is called a sink if

for every node v E N , v #p, the edge (v , p) exists, whereas the edge (p , v) does not
exist. Write an algorithm that can detect the presence of a sink in G in a time in O (n).
Your algorithm should accept the graph represented by its adjacency matrix (type
adjgraph of Section 1.9.2). Notice that a running time in 0 (n) for this problem is
remarkable given that the instance takes a space in Q(n 2) merely to v:'rite down.
* Problem 6.7.4. Euler's problem. An Euler path in a finite undirected graph

is a path such that every edge appears in it exactly once. Write an algorithm that
determines whether or not a given graph has an Euler path, and prints the path if so.
How much time does your algorithm take?
* Problem 6.7.5. Repeat Problem 6.7.4 for a directed graph.
Problem 6.7.6. The value 1 is available. To construct other values, you have
available the two operations x2 (multiplication by 2) and /3 (division by 3, any
resulting fraction being dropped). Operations are executed from left to right. For
instance
10= 1x2x2x2x2/3x2.
We want to express 13 in this way. Show how the problem can be expressed in terms
of exploring a graph and find a minimum-length solution.
* Problem 6.7.7. Show how the problem of carrying out a syntactic analysis of
a programming language can be solved in top-down fashion using a backtracking algo-
rithm. (This approach is used in a number of compilers.)
Problem 6.7.8. A Boolean array M [ 1 .. n, I .. n ] represents a square maze. In

general, starting from a given point, it is permissible to go to adjacent points in the
same row or in the same column. If M [i, j] is true, then you may pass through point
(i, j ); if M [i, j] is false, then you may not pass through point (i, j ). Figure 6.7.1
gives an example.
i. Give a backtracking algorithm that finds a path, if one exists, from (1, 1) to
(n, n). Without being completely formal (for instance, you may use statements
such as "for each point v that is a neighbour of x do "), your algorithm
must be clear and precise.
ii. Without giving all the details of the algorithm, indicate how to solve this
problem by branch-and-bound.
-+
17
-+ 11 .
Figure 6.7.1. A maze.
There exist a number of books concerning graph algorithms or combinatorial problems

that are often posed in terms of graphs. We mention in chronological order
Christofides (1975), Lawler (1976), Reingold, Nievergelt, and Deo (1977), Gondran
and Minoux (1979), Even (1980), Papadimitriou and Steiglitz (1982), and Tarjan
(1983). The mathematical notion of a graph is treated at length in Berge (1958, 1970).
A solution of problem 6.2.2 is given in Robson (1973).
Several applications of depth-first search are taken from Tarjan (1972) and Hop-
croft and Tarjan (1973). Problem 6.3.10 is solved in Rosenthal and Goldner (1977).
A linear time algorithm for testing the planarity of a graph is given in Hopcroft and
Tarjan (1974). Other algorithms based on depth-first search appear in Aho, Hopcroft,
and Ullman (1974, 1983).
Backtracking is described in Golomb and Baumert (1965) and techniques for
analysing its efficiency are given in Knuth (1975a). Some algorithms for playing chess
appear in Good (1968). The book by Nilsson (1971) is a gold mine of ideas con-
cerning graphs and games, the minimax technique, and alpha-beta pruning. The latter
is analysed in Knuth (1975b). A lively account of the first time a computer program
beat the world backgammon champion (Problem 6.6.23) is given in Deyong (1977).
For a more technical description of this feat, consult Berliner (1980). The branch-and-
bound technique is explained in Lawler and Wood (1966). The use of this technique to
solve the travelling salesperson problem is described in Bellmore and Nemhauser
(1968).
7
Preconditioning
and Precomputation
If we know that we shall have to solve several similar instances of the same problem,
it is sometimes worthwhile to invest some time in calculating auxiliary results that can
thereafter be used to speed up the solution of each instance. This is preconditioning.
Even when there is only one instance to be solved, precomputation of auxiliary tables
may lead to a more efficient algorithm.
7.1 PRECONDITIONING
7.1.1 Introduction
Let I be the set of instances of a given problem. Suppose each instance jet can be
separated into two components j E J and k E K (that is, I c J x K ).
A preconditioning algorithm for this problem is an algorithm A that accepts as
input some element j of J and produces as output a new algorithm Bj . This algorithm
Bj must be such that if k E K and < j, k > E I, then the application of Bj on k gives the
solution to the instance < j , k > of the original problem.
Example 7.1.1. Let J be a set of grammars for a family of programming

languages. For example, J might be a set of grammars in Backus-Naur form for such
languages as Algol, Pascal, Simula, and so on. Let K be a set of programs. The gen-
eral problem is to know whether a given program is syntactically correct with respect
to some given language. In this case I is the set of instances of the type "Is kEK a
valid program in the language defined by the grammar j E J ? ".
205
206 Preconditioning and Precomputation Chap. 7
One possible preconditioning algorithm for this example is a compiler generator :

applied to the grammar j E J , it generates a compiler Bj for the language in question.
Thereafter, to know whether k E K is a program in language j, we simply apply the
compiler Bj to k. O
Let
a( j) = the time required to produce Bj given j

b j (k) = the time required to apply Bj to k
t ( j, k) = the time required to solve < j, k > directly
It is usually the case that bj (k) <_ t( j, k) 5 a ( j) + bb (k). Obviously, we are wasting
our time using preconditioning if bj (k) > t ( j, k), and on the other hand, one way of
solving < j, k > from scratch consists of producing Bj from j and then applying it
on k. Preconditioning can be useful in two situations.
a. We need to be able to solve any instance i E I very rapidly, for example to ensure
a sufficiently fast response time for a real-time application. In this case it is
sometimes impractical to calculate and store ahead of time the #1 solutions to all
the relevant instances. It may, on the other hand, be possible to calculate and
store ahead of time #J preconditioned algorithms. Such an application of
preconditioning may be of practical importance even if only one crucial instance
is solved in the whole lifetime of the system : this may be just the instance that
enables us, for example, to stop a runaway reactor. The time you spend studying
before an exam may also be considered as an example of this kind of precondi-
tioning.
b. We have to solve a series of instances <j, k 1 >, <j, k2>, ... , <j, kn > with the
same j. In this case the time taken to solve all the instances is
ti = t(j,k,)
if we work without preconditioning, and
n
t2=a(j)+ bj (k;)
.=1
with preconditioning. Whenever n is sufficiently large, it often happens that t2 is
much smaller than t I.
Example 7.1.2. Let J be a set of sets of keywords, for example

J = { { if, then, else, endif 1, { si, alors, sinon, finsi } ,
{ for, to, by 1, 1 pour, jusqu'a, pas 1 1.

Sec. 7.1 Preconditioning 207
Let K be a set of keywords, for example

K = (Si, begin, jusqu'A, function}.
We have to solve a large number of instances of the type "Is the keyword kEK a
member of the set j E J ? ". If we solve each instance directly, we have
t(j,k)E8(nj)
in the worst case, where nj is the number of elements in the set j. On the other hand,
if we start by sorting j (this is the preconditioning), then we can subsequently solve
< j, k > by a binary search algorithm.
a (j) E O(nj log nj) for the sort
bi . (k) E O( log nj ) for the search
If there are many instances to be solved for the same j, then the second technique is
clearly preferable.
Example 7.1.3. We are to solve the system of equations Ax = b ; where A is a

non-singular square matrix and b is a column vector. If we expect to have several sys-
tems of equations to solve with the same matrix A but different vectors b, then it is
probably worthwhile to calculate the inverse of A once and for all, and to multiply this
inverse by each b.
Example 7.1.4. Problem 5.8.5 suggests how to obtain an efficient greedy

algorithm for making change. Calculating the necessary values cj is an example of
preconditioning that allows us subsequently to make change quickly every time this is
required.
Example 7.1.5. Creating an optimal search tree (Section 5.5) is a further

example of preconditioning.
7.1.2 Ancestry in a rooted tree
Let J be the set of all rooted trees, and let K be the set of pairs < v , w > of nodes. For
a given pair k = < v, w > and a given rooted tree j we want to know whether node v is
an ancestor of node w in tree j. (By definition, every node is its own ancestor and,
recursively, the ancestor of all the nodes of which its children are ancestors.)
If the tree j contains n nodes, any direct solution of this instance takes a time in
S2(n) in the worst case.
Problem 7.1.1. Why?
It is, however, possible to precondition the tree in a time in O(n), so that we can
subsequently solve any particular instance in a time in 0(1).
We illustrate this approach using the tree in Figure 7.1.1. It contains 13 nodes.
To precondition the tree, we traverse it first in preorder and then in postorder (see Sec-
tion 6.2), numbering the nodes sequentially as we visit them. For a node v, let
prenum [v] be the number assigned to the node when we traverse the tree in preorder,
and let postnum [v] be the number assigned during the traversal in postorder. In Figure
7.1.1 these two numbers appear to the left and the right of the node, respectively.
2 B 5 C 12
D 1 4 E 3 6 F 4 8 ;G6 9 H
I 2 10 J 711 K 8121 L 913 M 10
Figure 7.1.1. A rooted tree with preorder and postorder numberings.
Let v and w be two nodes in the tree. In preorder we first number a node and
then we number its subtrees from left to right. Thus
prenum [v] <_ prenum [w] b v is an ancestor of w or
v is to the left of w in the tree.
In postorder we first number the subtrees of a node from left to right, and then we
number the node itself. Thus
postnum [v] >_ postnum [w] a v is an ancestor of w or
v is to the right of w in the tree.
It follows that
prenum [v] prenum [w] and postnum [v] >_ postnum [w]
t* v is an ancestor of w.
Once all the values of prenum and postnum have been calculated in a time in 0(n),
the required condition can be checked in a time in 0(1).
Problem 7.1.2. There exist several similar ways of preconditioning a tree so

as to be able thereafter to verify rapidly whether one node is an ancestor of another.
Show, for example, that this can be done using a traversal in preorder followed by a
traversal in inverted preorder, which visits first a node and then its subtrees from right
to left. (Notice, however, that traversal of a rooted tree in inverted preorder requires
more work than traversal in postorder if the representation of rooted trees suggested in
Figure 1.9.5 is used.) 11
Sec. 7.1 Preconditioning 209
7.1.3 Repeated Evaluation of a Polynomial
Let J be the set of polynomials in one variable x, and let K be the set of values this
variable can take. The problem consists of evaluating a given polynomial at a given
point.
For simplicity, we restrict ourselves to polynomials with integer coefficients,
evaluated at integer values of x. We use the number of integer multiplications that
have to be carried out as a barometer to measure the efficiency of an algorithm, taking
no account of the size of the operands involved nor of the number of additions and
subtractions.
Initially, we restrict ourselves even further and consider only monic polynomials
(the leading coefficient is 1) of degree n = 2k - I for some integer k ? 1.
Example 7.1.6. Consider the polynomial

p(x)=x7-5x6+4x5- 13x4+3x3- 10x2+5x - 17.
A naive method for evaluating this polynomial is to calculate first the series of values
x2, x3, .. , x7, from which we can obtain 5x, lOx2, ... , 5x6 and finally p(x). This
method requires 12 multiplications and 7 additions (counting the subtractions as addi-
tions).
It is easy to do better. If we evaluate p (x) as
p (x) = ((((((x -5)x +4)x - 13)x +3)x - 10)x +5)x -17
we need only 6 multiplications and 7 additions. Better still, we can calculate
p(x)=(x4+2)[(x2+3)(x-5)+(x+2)]+[(x2-4)x + (x +9)]
using only 5 multiplications (of which two are to calculate x2 and x4) plus 9 additions.
(A typical 7th degree monic polynomial would require 5 multiplications as is the case
here, but 10 additions.)
In general, if p (x) is a monic polynomial of degree n = 2k -1, we first express it

in the form
p(x) = (x(n+t)12+a)q(x) + r(x),
where a is a constant and q (x) and r (x) are monic polynomials of degree 2k - I -1.
Next, we apply the same procedure recursively to q (x) and r (x). Finally, p (x) is
expressed entirely in terms of polynomials of the form x` + c, where i is a power of 2.
In the preceding example we first express p (x) in the form
(x4+a)(x3+82x2+qlx+qo)+(x3+r2x2+rlx+ro) .
Equating the coefficients of x6, x5, ... , 1, we obtain successively q2 = -5, q I = 4,

qo=-13,a =2,r2=0,rt=-3,andro=9. Thus
p(x) = (x4+2)(x3-5x2+4x -13) + (x3-3x +9).

Similarly, we find
x3-5x2+4x - 13=(x2+3)(x-5)+(x+2)
x3 - 3x + 9 = (x2-4)x + (x +9)
to arrive finally at the expression for p (x) given in Example 7.1.6. This expression is
the preconditioned form of the polynomial.
Problem 7.1.3. Express p(x)=x7+2x6-5x4+2x3-6x2+6x -32 in

preconditioned form.
Problem 7.1.4. Express p (x) ax7 in preconditioned form.
Analysis of the method. Let M(k) be the number of multiplications

required to evaluate p (x), a preconditioned monic polynomial of degree n = 2k -1.
Let M(k) = M(k) - k + 1 be the number of multiplications required if we do not count
those used in the calculation of x 2, x 4, ... , x (n + 1)/2 We obtain the recurrence equa-
tion
0 k=1
M(k) =
2M(k -1) + 1 k>2.
Consequently, M(k) = 2k ' -1 for k >- 1, and hence M(k)=2k -1+k-2. In other
words, (n -3)/2+Ig(n + 1) multiplications are sufficient to evaluate a preconditioned
polynomial of degree n = 2k -1.
* Problem 7.1.5. Prove that if the monic polynomial p (x) is given by its
coefficients, there does not exist an algorithm that can calculate p (x) using less than
n -1 multiplications in the worst case. In other words, the time invested in precondi-
tioning the polynomial allows us to evaluate it subsequently using essentially half the
number of multiplications otherwise required.
Problem 7.1.6. Show that evaluation of a preconditioned polynomial of

degree n = 2k - 1 requires (3n - 1)/2 additions in the worst case.
Problem 7.1.7. Generalize this method of preconditioning to polynomials that

are not monic. Your generalization must give an exact answer, with no risk of
rounding error due to the use of floating-point arithmetic.
Problem 7.1.8. (Continuation of Problem 7.1.7) Further generalize the

method to polynomials of any degree.
Sec. 7.2 Precomputation for String-Searching Problems 211
Problem 7.1.9. Show using an explicit example that the method described
here does not necessarily give an optimal solution (that is, it does not necessarily
minimize the required number of multiplications) even in the case of monic polyno-
mials of degree n = 2k -1.
Problem 7.1.10. Is the method appropriate for polynomials involving real

coefficients and real variables ? Justify your answer.
7.2 PRECOMPUTATION FOR

STRING-SEARCHING PROBLEMS
The following problem occurs frequently in the design of text-processing systems (edi-
tors, macroprocessors, information retrieval systems, etc.). Given a target string con-
sisting of n characters, S = s is 2 s , and a pattern consisting of m characters,
P=p1p2 p,,, , we want to know whether P is a substring of S, and if so,
whereabouts in S it occurs. Suppose without loss of generality that n >- m. In the ana-
lyses that follow, we use the number of comparisons between pairs of characters as a
barometer to measure the efficiency of our algorithms.
The following naive algorithm springs immediately to mind. It returns r if the
first occurrence of P in S begins at position r (that is, r is the smallest integer such that
Sr+i - 1 =Pi , i = 1, 2, ... , m), and it returns 0 if P is not a substring of S.
for i E- O to n -m do
ok F- true
j *- 1
while ok and j <- m do
ifp[j]#s[i+j] then ok <--false
else j F j + 1
if ok then return i + 1
return 0
The algorithm tries to find the pattern P at every position in S. In the worst case
it makes m comparisons at each position to see whether or not P occurs there. (Think
of S = "aaa aab", P = "aaaab".) The total number of comparisons to be made is
therefore in S1(m (n -m )), which is in Q(mn) if n is much larger than in. Can we do
better ?
7.2.1 Signatures
Suppose that the target string S can be decomposed in a natural way into substrings,
S = S 1S 2 S, , and that the pattern P, if it occurs in S, must occur entirely within
one of these substrings (thus we exclude the possibility that P might straddle several
consecutive substrings). This situation occurs, for example, if the Si are the lines in a
text file S and we are searching for the lines in the file that contain P.
The basic idea is to use a Boolean function T (P, Si) that can be calculated
rapidly to make a preliminary test. If T (P , Si) is false, then P cannot be a substring of
Si ; if T (P, Si) is true, however, it is possible that P might be a substring of Si , but we
have to carry out a detailed check to verify this (for instance, using the naive algorithm
given earlier). Signatures offer a simple way of implementing such a function.
Suppose that the character set used for the strings S and P is
{ a, b, c, ... , x, y, z, other) , where we have lumped all the non-alphabetic characters
together. Suppose too that we are working on a computer with 32-bit words. Here is
one common way of defining a signature.
i. Define val ("a") = 0, val ("b") = 1, ... , val ("z") = 25, val (other) = 26.
ii. If c1 and c2 are characters, define
B (c 1, c 2) = (27 val (c 1) + val (c 2)) mod 32.
M. Define the signature sig (C) of a string C = c 1c2 cr as a 32-bit word where
the bits numbered B (c 1, c 2), B (c 2, c 3), ... , B (cr_1, cr) are set to 1 and the other
bits are 0.
Example 7.2.1. If C is the string "computers", we calculate

B ("c", o") = 27 x 2 + 14 mod 32 = 4, B ("o", m") = 27 x 14 + 12 mod 32 = 6, ... ,
B ("r", s") = 27 x 17 + 18 mod 32 = 29. If the bits of a word are numbered from 0
(on the left) to 31 (on the right), the signature of this string is the word
0000 1110 0100 0001 0001 0000 0000 0100 .
Only seven bits are set to 1 in the signature because B ("e", r") = B ("r", s") = 29.
We calculate a signature for each substring Si and for the pattern P. If Si con-
tains the pattern P, then all the bits that are set to 1 in the signature of P are also set to
1 in the signature of Si . This gives us the function T we need :
T (P , Si) = [(sig (P) and sig (Si )) = sig (P )J,
where the and operator represents the bitwise conjunction of two whole words. T can
be computed very rapidly once all the signatures have been calculated.
This is yet another example of preconditioning. Calculating the signatures for S
takes a time in O (n). For each pattern P we are given we need a further time in O (m)
to calculate its signature, but from then on we hope that the preliminary test will allow
us to speed up the search for P. The improvement actually obtained in practice
depends on the judicious choice of a method for calculating signatures.
* Problem 7.2.1. If signatures are calculated as described, and if the characters

a, b, ... , z and other are equiprobable, what is the probability that the signature of a
random string of n characters contains a 1 in all the bit positions that contain a 1 in the
signature of another random string of m characters ? Calculate the numerical value of
this probability for some plausible values of m and n (for instance, n = 40, m = 5).
Problem 7.2.2. Is the method illustrated of interest if the target string is very
long and if it cannot be divided into substrings ?
Problem 7.2.3. If T (P, Si) is true with probability E > 0 even if Si does not
contain P, what is the order of the number of operations required in the worst case to
find P in S or to confirm that it is absent ?
Many variations on this theme are possible. In the preceding example the func-
tion B takes two consecutive characters of the string as parameters. It is easy to invent
such functions based on three consecutive characters, and so on. The number of bits in
the signature can also be changed.
Problem 7.2.4. Can we define a function B based on a single character?

If this is possible, is it useful ?
Problem 7.2.5. If the character set contains the 128 characters of the ASCII
code, and if the computer in use has 32-bit words, we might define B by
B (c 1, c 2) = (128 val (c 1) + val (c 2)) mod 32.
Is this to be recommended ? If not, what do you suggest instead ?
7.2.2 The Knuth-Morris-Pratt Algorithm
We confine ourselves to giving an informal description of this algorithm (henceforth :

the KMP algorithm), which finds the occurrences of P in S in a time in 0 (n).
Example 7.2.2. Let S = "babcbabcabcaabcabcabcacabc" and P = "abcabcacab".

To find P in S we slide P along S from left to right, looking at the characters that are
opposite one another. Initially, we try the following configuration :
S babcbabcabcaabcabcabcacabc
P abcabcacab
T
We check the characters of P from left to right. The arrows show the comparisons car-
ried out before we find a character that does not match. In this case there is only one
comparison. After this failure we try
P abcabcacab
TTTT
This time the first three characters of P are the same as the characters opposite them
in S, but the fourth does not match. Up to now, we have proceeded exactly as in the
naive algorithm. However we now know that the last four characters examined in S
are abcx where x #"a". Without making any more comparisons with S, we can
conclude that it is useless to slide P one, two, or three characters along : such an align-
ment cannot be correct. So let us try sliding P four characters along.
P abcabcacab
TTTTTTTT
Following this mismatch, we know that the last eight characters examined in S are
abcabcax where x # "c". Sliding P one or two places along cannot be right ; however
moving it three places might work.
P abcabcacab
T
There is no need to recheck the first four characters of P : we chose the movement of
P in such a way as to ensure that they necessarily match. It suffices to start checking
at the current position of the pointer. In this case we have a second mismatch in the
same position. This time, sliding P four places along might work. (A three-place
movement is not enough: we know that the last characters examined in S are ax,
where x is not a "b".)
P abcabcacab
TTTTTTTT
Yet again we have a mismatch, and this time a three-place movement is necessary.
P abcabcacab
TTTTTT
We complete the verification starting at the current position

time the correspondence between the target string and the pattern is complete.
To implement this algorithm, we need an array next[ I.. m ]. This array tells us
what to do when a mismatch occurs at position j in the pattern.
If next [ j ] = 0, it is useless to compare further characters of the pattern to the target
string at the current position. We must instead line up P with the first character of S
that has not yet been examined and start checking again at the beginning of P.
If next [ j ] = i > 0, we should align the i th character of P on the current character of S
and start checking again at this position. In both cases we slide P along j - next [ j ]
characters to the right with respect to S. In the preceding example we have
j 1 2 3 4 5 6 7 8 9 10
p[j] a b c a b c a c a b
next[ j] 0 1 1 0 1 1 0 5 0 1
Once this array has been calculated, here is the algorithm for finding P in S.
function KMP
j,k <- 1
while j <_ m and k 5 n do
while j > O ands [k] # p [j ] do
j'- next[j]
k-k+1
j-j+1
if j >m then return k - m
else return 0
It returns either the position of the first occurrence of P in S, or else 0 if P is not a

substring of S.
Problem 7.2.6. Follow the execution of this algorithm step by step using the
strings from Example 7.2.2.
After each comparison of two characters, we move either the pointer (the arrow
in the diagrams, or the variable k in the algorithm) or the pattern P. The pointer and P
can each be moved a maximum of n times. The time required by the algorithm is
therefore in O (n). Precomputation of the array next [ 1 .. in ] can be carried out in a
time in 0(m), which can be neglected since m <_ n. Overall, the execution time is thus
in O (n).
It is correct to talk of preconditioning in this case only if the same pattern is
sought in several distinct target strings, which does happen in some applications.
On the other hand, preconditioning does not apply if several distinct patterns are
sought in a given target string. In all cases, including the search for a single pattern in
a single target, it is correct to talk of precomputation.
* Problem 7.2.7. Find a way to compute the array next [ 1 .. m ] in a time in

0 (M).
Problem 7.2.8. Modify the KMP algorithm so that it finds all the occurrences
of P in Sin a total time in 0 (n).
7.2.3 The Boyer-Moore Algorithm
Like the KMP algorithm, the algorithm due to Boyer and Moore (henceforth : the BM
algorithm) finds the occurrences of P in S in a time in O (n) in the worst case. How-
ever, since the KMP algorithm examines every character of the string S at least once in
the case when P is absent, it makes at least n comparisons. The BM algorithm, on the
other hand, is often sublinear : it does not necessarily examine every character of S,
and the number of comparisons carried out can be less than n. Furthermore, the BM
algorithm tends to become more efficient as m, the number of characters in the pattern
P, increases. In the best case the BM algorithm finds all the occurrences of P in S in a
time in O (m + n /m).
As with the KMP algorithm, we slide P along S from left to right, checking
corresponding characters. This time, however, the characters of P are checked from
right to left after each movement of the pattern. We use two rules to decide how far
we should move P after a mismatch.
i. If we have a mismatch immediately after moving P, let c be the character oppo-

site p [m]. We know that c #p [m]. If c appears elsewhere in the pattern, we
slide the latter along in such a way as to align the last occurrence of c in the pat-
tern with the character c in the target string. If c does not appear in the pattern,
we align the latter just after the occurrence of c in the target string.
ii. If a certain number of characters at the end of P correspond to the characters in
S, then we use this partial knowledge of S (just as in the KMP algorithm) to
slide P along to a new position compatible with the information we possess.
Example 7.2.3. Let S = "This is a delicate topic" and P = "cat".

The target string and the pattern are initially aligned as follows :
S This is a delicate topic

P cat
T
We examine P from right to left. There is an immediate mismatch in the position

shown by the arrow. The character opposite p [m] is "i". Since the pattern does not
include this character, we slide the pattern just to the right of the arrow.

P cat
T
Again we examine P from right to left, and again there is an immediate mismatch.
Since "i" does not appear in the pattern, we try

P cat
T
There is once again an immediate mismatch, but this time the character "a" that
appears opposite p[m] also appears in P. We slide P one place along to align the
occurrences of the letter "a", and start checking again (at the right-hand end of P).

P cat
T
After two more immediate mismatches we are in this situation.

P cat
T
Now, when we slide P along one position to align the "a" in the target string with the
"a" in the pattern, P is correctly aligned. A final check, always from right to left, will
confirm this. In this example we have found P without ever using rule (ii). We have
made only 9 comparisons between a character of P and a character of S.
Example 7.2.4. Consider the same strings as in Example 7.2.2:
P abcabcacab
TTTT
We examine P from right to left. The left-hand arrow shows the position of the first
mismatch. We know that starting at this position S contains the characters xcab where
x #"a". If we slide P five places right, this information is not contradicted. (Under-
scores show which characters were aligned.)
P abcabcacab
T
Unlike the KMP algorithm, we check all the positions of P after moving the pattern.
Some unnecessary checks (corresponding to the underscored characters in P) may be
made at times. In our example when we start over checking P from right to left, there
is an immediate mismatch. We slide P along to align the "c" found in S with the last
"c" in P.
P abcabcacab
TTTT
After four comparisons between P and S (of which one is unnecessary), carried out as
usual from right to left, we again have a mismatch. A second application of rule (ii)
gives us
P abcabcacab
T
We apply rule (i) once to align the letters "a":
P abcabcacab
T
and one last time to align the letters "c":
P abcabcacab
TTTTTTTTTT
We have made 21 comparisons in all to find P.
To implement the algorithm, we need two arrays d1 [ { character set } ] and

d2[1 .. m -1], the former to implement rule (i) and the latter for rule (ii).
The array d 1, indexed by the character set we are using, is easy to compute.
For every character c
dl[c] F- if c does not appear in p [ 1 .. m ] then m
elsem -max{i I p[i]=c] .
This is the distance to move P according to rule (i) when we have an immediate
mismatch.
It is more complicated to compute d2. We shall not give the details here, but
only an example. The interpretation of d2 is the following : after a mismatch at posi-
tion i of the pattern, begin checking again at position m (that is, at the right-hand end)
of the pattern and d2[i] characters further along S.
Example 7.2.5. Suppose the pattern is P = "assesses". Suppose further that at

some moment during our search for the string P in S we have a mismatch in position
p [7]. Since we always examine the characters of P from right to left, we know that
starting at the position of the mismatch the characters of S are xs, where x # "e" :
S ??????xs???????? x "e"
P assesses
TT
The fact that x * "e" does not rule out the possibility of aligning the "s" in p [6] with
the "s" we have found in S. It may therefore be possible to align P as follows :
S ??????xs???????? X #"e
P assesses
T
We start checking again at the end of P, that is, 3 characters further on in S than the
previous comparison: thus d2[7] = 3.
Similarly, suppose now that we have a mismatch at position p [6]. Starting from
the position of the mismatch, the characters of S are xes, where x # "s" :
S ?????xes???????? X "s"
P assesses
TTT
The fact that x # "s" rules out the possibility of aligning the "e" and the "s" in p [4] and
p [5] with the "e" and the "s" found in S. It is therefore impossible to align P under
these characters, and we must slide P all the way to the right under the characters of S
that we have not yet examined :
S ?????xes???????? X "s
P assesses
T
We start checking again at the end of P, that is, 10 characters further on in S than the
previous comparison: thus d2[6] = 10.
As a third instance, suppose we have a mismatch at position p [4]. Starting from
the position of the mismatch, the characters of S are xsses, where x # "e":
S ???xsses???????? x "e"
P assesses
TTTTT
In this case it may be possible to align P with S by sliding it three places right :
S
P
???xsses???????? xe
assesses
T
Now we start checking at the end of P, 7 characters further on in S than the previous
comparison, so d2[4] = 7.
For this example we find
i 1 2 3 4 5 6 7 8
p[i] a s s e s s e s
d2[i] 15 14 13 7 11 10 3
We also have d 1 ["s"] = 0, d 1 ["e"] =1, d 1 ["a"] = 7 and d 1 [any other character ] = 8.
Note that d 1 ["s"] has no significance, because an immediate mismatch is impossible at
a position where S contains "s".
Problem 7.2.9. Calculate d 1 and d2 for the pattern in Example 7.2.4.
Problem 7.2.10. Calculate d 1 and d2 for P = "abracadabraaa".
Here finally is the BM algorithm.

function BM
j,k F- m
while k <- n and j > 0 do
while j > 0 and s [k] = p [j ] do
k+- k-1
j *- j - 1
if j # 0 then
if j = m then k <-- k + d 1[s [k]]
elsek k+d2[j]
jFm
if j = 0 then return k + 1
else return 0
It returns either the position of the first occurrence of P in S, or else 0 if P is not a

substring of S.
Problem 7.2.11. In this algorithm, the choice between using rule (i) and rule
(ii) depends on the test "j = m ? ". However, even if j < m , it is possible that rule (i)
might allow k to advance more than rule (ii). Continuing from Example 7.2.5, con-
sider the following situation :
S ??????ts????????
P assesses
TT
The failure of the match between "t" and "e" is of the second kind, so k is increased by
d2[71 = 3 to obtain
S ??????ts????????
P assesses
T
However, the fact that "t" does not appear in P should have allowed us to increase k
directly by d 1["t"] = 8 positions.
S ??????te????????
P assesses
T
Show that the algorithm is still correct if we replace

if j= m then k E- k+ d I [s [k]]
else k<- k+d2[j]
j E- m
by
k <- k + max(d I [s [k] ], d2[ j ])

j4-m ,
provided we define d 2[m] = 1 and d 1[ p [m]] = 0.

This modification corresponds to the algorithm usually known by the name
Boyer-Moore (although these authors also suggest other improvements).
Problem 7.2.12. Show the progress of the algorithm if we search (unsuccess-

fully, of course) for the pattern P = "assesses" in S = "I guess you possess a dress fit
for a princess".
How many comparisons are made altogether before the failure to match is
discovered, and how many of these comparisons are redundant (that is, they repeat
comparisons previously made) ?
* Problem 7.2.13. Find a way to calculate the array d2 in a time in 0 (m).
** Problem 7.2.14. Prove that the total execution time of the algorithm (compu-
tation of d I and d2 and search for P) is in 0 (n).
Problem 7.2.15. Modify the BM algorithm so that it will find all the
occurrences of P in S in a time in 0 (n).
It is easy to see intuitively why the algorithm is often more efficient for longer
patterns. For a character set of reasonable size (say, 52 letters if we count upper- and
lowercase separately, ten figures and about a dozen other characters) and a pattern that
is not too long, d I [c] is equal to m for most characters c. Thus we look at approxi-
mately one character out of every m in the target string. As long as m stays small
compared to the size of the character set, the number of characters examined goes
down as m goes up. Boyer and Moore give some empirical results : if the target string
S is a text in English, about 20% of the characters are examined when m = 6; when
m = 12, only 15% of the characters in S are examined.
Preconditioning polynomials for repeated evaluation is suggested in Belaga (1961).

Signatures are discussed in Harrison (1971). The KMP and BM algorithms of Sections
7.2.2 and 7.2.3 come from Knuth, Morris, and Pratt (1977) and Boyer and Moore
(1977). Rytter (1980) corrects the algorithm given in Knuth, Morris, and Pratt (1977)
for calculating the array d2 to be used in the Boyer-Moore algorithm. Finite automata,
as described for instance in Hopcroft and Ullman (1979), can be used to introduce the
KMP algorithm in an intuitively appealing way; see, for example, Baase (1978). For
an efficient algorithm capable of finding all the occurrences of a finite set of patterns in
a target string, consult Aho and Corasick (1975). For a probabilistic string-searching
algorithm (see Chapter 8), read Karp and Rabin (1987).
8
Probabilistic Algorithms
8.1 INTRODUCTION
Imagine that you are the hero (or the heroine) of a fairy tale. A treasure is hidden at a
place described by a map that you cannot quite decipher. You have managed to reduce
the search to two possible hiding-places, which are, however, a considerable distance
apart. If you were at one or the other of these two places, you would immediately
know whether it was the right one. It takes five days to get to either of the possible
hiding-places, or to travel from one of them to the other. The problem is complicated
by the fact that a dragon visits the treasure every night and carries part of it away to an
inaccessible den in the mountains. You estimate that it will take four more days' com-
putation to solve the mystery of the map and thus to know with certainty where the
treasure is hidden, but if you set out on a journey you will no longer have access to
your computer. An elf offers to show you how to decipher the map if you pay him the
equivalent of the treasure that the dragon can carry away in three nights.
Problem 8.1.1. Leaving out of consideration the possible risks and costs of
setting off on a treasure-hunting expedition, should you accept the elf's offer?
Obviously it is preferable to give three nights' worth of treasure to the elf rather
than allow the dragon four extra nights of plunder. If you are willing to take a calcu-
lated risk, however, you can do better. Suppose that x is the value of the treasure
remaining today, and that y is the value of the treasure carried off every night by the
dragon. Suppose further that x > 9y. Remembering that it will take you five days to
reach the hiding-place, you can expect to come home with x - 9y if you wait four days
to finish deciphering the map. If you accept the elf's offer, you can set out
223
224 Probabilistic Algorithms Chap. 8
immediately and bring back x - 5y, of which 3y will go to pay the elf; you will thus
have x - 8y left. A better strategy is to toss a coin to decide which possible hiding-
place to visit first, journeying on to the other if you find you have decided wrong. This
gives you one chance out of two of coming home with x -5y, and one chance out of
two of coming home with x -10y . Your expected profit is therefore x - 7.5y. This is
like buying a ticket for a lottery that has a positive expected return.
This fable can be translated into the context of algorithmics as follows : when an
algorithm is confronted by a choice, it is sometimes preferable to choose a course of
action at random, rather than to spend time working out which alternative is the best.
Such a situation arises when the time required to determine the optimal choice is prohi-
bitive, compared to the time that will be saved on the average by making this optimal
choice. Clearly, the probabilistic algorithm can only be more efficient with respect to
its expected execution time. It is always possible that bad luck will force the algorithm
to explore many unfruitful possibilities.
We make an important distinction between the words "average" and "expected".
The average execution time of a deterministic algorithm was discussed in section 1.4.
It refers to the average time taken by the algorithm when each possible instance of a
given size is considered equally likely. By contrast, the expected execution time of a
probabilistic algorithm is defined on each individual instance : it refers to the mean
time that it would take to solve the same instance over and over again. This makes it
meaningful to talk about the average expected time and the worst-case expected time
of a probabilistic algorithm. The latter, for instance, refers to the expected time taken
by the worst possible instance of a given size, not the time incurred if the worst pos-
sible probabilistic choices are unfortunately taken.
Example 8.1.1. Section 4.6 describes an algorithm that can find the k th small-
est of an array of n elements in linear time in the worst case. Recall that this algo-
rithm begins by partitioning the elements of the array on either side of a pivot, and that
it then calls itself recursively on the appropriate section of the array if need be. One
fundamental principle of the divide-and-conquer technique suggests that the nearer the
pivot is to the median of the elements, the more efficient the algorithm will be.
Despite this, there is no question of choosing the exact median as the pivot because
this would cause an infinite recursion (see Problem 4.6.3). Thus we choose a subop-
timal so-called pseudomedian. This avoids the infinite recursion, but choosing the
pseudomedian still takes quite some time. On the other hand, we saw another algo-
rithm that is much faster on the average, but at the price of a quadratic worst case : it
simply decides to use the first element of the array as the pivot. We shall see in Sec-
tion 8.4.1 that choosing the pivot randomly gives a substantial improvement in the
expected execution time as compared to the algorithm using the pseudomedian, without
making the algorithm catastrophically bad for the worst-case instances.
We once asked the students in an algorithmics course to implement the selection
algorithm of their choice. The only algorithms they had seen were those in Sec-
Sec. 8.1 Introduction 225
tion 4.6. Since the students did not know which instances would be used to test their
programs (and suspecting the worst of their professors), none of them took the risk of
using a deterministic algorithm with a quadratic worst case. Three students, however,
thought of using a probabilistic approach. This idea allowed them to beat their col-
leagues hands down : their programs took an average of 300 milliseconds to solve the
trial instance, whereas the majority of the deterministic algorithms took between 1500
and 2600 milliseconds.
Example 8.1.2. Section 6.6.1 describes a systematic way of exploring an

implicit tree to solve the eight queens problem. If we are content with finding one
solution rather than all of them, we can improve the backtracking technique by placing
the first few queens at random. Section 8.5.1 goes into this more thoroughly.
Example 8.1.3. No known deterministic algorithm can decide in a reasonable

time whether a given integer with several hundred decimal digits is prime or compo-
site. Nevertheless, Section 8.6.2 describes an efficient probabilistic algorithm to solve
this problem provided that one is willing to accept an arbitrarily small probability of
error. This problem has important applications in cryptology (Section 4.8).
Example 8.1.2 raises an important consideration concerning probabilistic algo-

rithms. They are sometimes used to solve problems that allow several correct solu-
tions. Using the same probabilistic algorithm on the same instance, we may obtain dif-
ferent correct solutions on different occasions. For another example of the same
phenomenon consider the problem : "Find a nontrivial factor of a given composite
integer." Of course, such problems can also be handled by deterministic algorithms,
but in this case the choice of algorithm determines uniquely which solution will be
obtained whenever the algorithm is applied on any given instance.
The analysis of probabilistic algorithms is often complex, requiring an acquain-
tance with results in probability, statistics, and number theory beyond the scope of this
book. For this reason, a number of results are cited without proof in the following sec-
tions. For more details, consult the references suggested in the last section.
Throughout this chapter we suppose that we have available a random number
generator that can be called at unit cost. Let a and b, a < b , be real numbers. A call
on uniform (a, b) returns a real number x chosen randomly in the interval a <_ x < b .
The distribution of x is Iniform on the interval, and successive calls on the generator
yield independent values of x. To generate random integers, we extend the notation to
include uniform (i .. j ), where i and j are integers, i <_ j, and the function returns an
integer k chosen randomly, uniformly, and independently in the interval i 5 k S j.
Similarly, uniform (X), where X is a nonempty finite set, returns an element chosen
randomly, uniformly, and independently among the elements of X.
Problem 8.1.2. Show how the effect of uniform (i .. j) can be obtained if only
uniform (a, b) is available.
Example 8.1.4. Let p be a prime number, and let a be an integer such that
1 5 a 1 } . For
example, the index of 2 modulo 31 is 5, that of 3 is 30, and that of 5 is 3. By Fermat's
theorem, an index modulo p always divides p -1 exactly. This suggests one way of
making a random, uniform, independent choice of an element of X.
function draw (a , p )
j F- uniform(1..p-1)
return dexpoiter (a , j, p) { Section 4.8 }
Problem 8.1.3. Give other examples of sets in which there is an efficient way
to choose an element randomly, uniformly, and independently.
Truly random generators are not usually available in practice. Most of the time
pseudorandom generators are used instead : these are deterministic procedures that are
able to generate long sequences of values that appear to have the properties of a
random sequence. To start a sequence, we must supply an initial value called a seed.
The same seed always gives rise to the same sequence, so to obtain different
sequences, we may choose, for example, a seed that depends on the date or time. Most
programming languages include such a generator, although some implementations
should be used with caution. Using a good pseudorandom generator, the theoretical
results obtained in this chapter concerning the efficiency of different algorithms can
generally be expected to hold. However, the impractical hypothesis that a genuinely
random generator is available is crucial when we carry out the analysis.
The theory of pseudorandom generators is complex, but a simple example will
illustrate the general idea. Most generators are based on a pair of functions S : X -+ X
and R : X -> Y, where X is a sufficiently large set and Y is the domain of pseu-
dorandom values to be generated. Let g r= X be a seed. Using the function S, this seed
defines a sequence : x0 = g and xi = S (xi - 1) for i > 0. Finally, the function R allows
us to obtain the pseudorandom sequence YO, y 1 , Y2, ... defined by yi = R (xi ), i >_ 0.
This sequence is necessarily periodic, with a period that cannot exceed #X. However,
if S and R (and sometimes g) are chosen properly, the period can be made very long,
and the sequence may be for most practical purposes statistically indistinguishable
from a truly random sequence of elements of Y. Suggestions for further reading are
given at the end of the chapter.
8.2 CLASSIFICATION OF PROBABILISTIC ALGORITHMS

By definition, a probabilistic algorithm leaves some of its decisions to chance. We
shall not worry about the fact that such a concept conflicts with the definition of "algo-
rithm" given at the beginning of the first chapter. The fundamental characteristic of
these algorithms is that they may react differently if they are applied twice to the same
instance. The execution time, and even the result obtained, may vary considerably
Sec. 8.2 Classification of Probabilistic Algorithms 227
from one use to the next. Probabilistic algorithms can be divided into four major
classes : numerical, Monte Carlo, Las Vegas, and Sherwood. Some authors use the
term "Monte Carlo" for any probabilistic algorithm, and in particular for those we call
"numerical".
Randomness was first used in algorithmics for the approximate solution of
numerical problems. Simulation can be used, for example, to estimate the mean length
of a queue in a system so complex that it is impossible to get closed-form solutions or
to get numerical answers by deterministic methods. The answer obtained by such a
probabilistic algorithm is always approximate, but its expected precision improves as
the time available to the algorithm increases. (The error is usually inversely propor-
tional to the square root of the amount of work performed.) For certain real-life prob-
lems, computation of an exact solution is not possible even in principle, perhaps
because of uncertainties in the experimental data to be used, or maybe because a digital
computer can only handle binary or decimal values while the answer to be computed is
irrational. For other problems, a precise answer exists but it would take too long to
figure it out exactly. Sometimes the answer is given in the form of a confidence
interval.
Monte Carlo algorithms, on the other hand, are used when there is no question of
accepting an approximate answer, and only an exact solution will do. In the case of a
decision problem, for example, it is hard to see what an "approximation" might be,
since only two answers are possible. Similarly, if we are trying to factorize an integer,
it is of little interest to know that such-and-such a value is "almost a factor". A way to
put down seven queens on the chess-board is little help in solving the eight queens
problem. A Monte Carlo algorithm always gives an answer, but the answer is not
necessarily right; the probability of success (that is, of getting a correct answer)
increases as the time available to the algorithm goes up. The principal disadvantage of
such algorithms is that it is not in general possible to decide efficiently whether or not
the answer given is correct. Thus a certain doubt will always exist.
Las Vegas algorithms never return an incorrect answer, but sometimes they do
not find an answer at all. As with Monte Carlo algorithms, the probability of success
increases as the time available to the algorithm goes up. However, any answer that is
obtained is necessarily correct. Whatever the instance to be salved, the probability of
failure can be made arbitrarily small by repeating the same algorithm enough times on
this instance. These algorithms should not be confused with those, such as the simplex
algorithm for linear programming, that are extremely efficient for the great majority of
instances to be handled, but catastrophic for a few instances.
Finally, Sherwood algorithms always give an answer, and the answer is always
correct. They are used when some known deterministic algorithm to solve a particular
problem runs much faster on the average than in the worst case. Incorporating an ele-
ment of randomness allows a Sherwood algorithm to reduce, and sometimes even to
eliminate, this difference between good and bad instances. It is not a case of
preventing the occasional occurrence of the algorithm's worst-case behaviour, but
rather of breaking the link between the occurrence of such behaviour and the particular
instance to be solved. Since it reacts more uniformly than the deterministic algorithm,
a Sherwood algorithm is less vulnerable to an unexpected probability distribution of
the instances that some particular application might give it to solve (see the end of Sec-
tion 1.4).
Problem 8.2.1. A problem is well-characterized if it is always possible to

verify efficiently the correctness of a proposed solution for any given instance. Show
that the problem of finding a nontrivial factor of a composite integer (Section 8.5.3) is
well-characterized. You should realize this in no way implies that the problem is easy
to solve. Intuitively, do you think the problem of finding the smallest nontrivial factor
of a composite integer is well-characterized?
Problem 8.2.2. Show how to obtain a Las Vegas algorithm to solve a well-
characterized problem given that you already have a Monte Carlo algorithm for the
same problem. Contrariwise, show how to obtain a Monte Carlo algorithm for any
problem whatsoever given that you already have a Las Vegas algorithm for the same
problem.
Problem 8.2.3. Why "Sherwood", do you think ?
8.3 NUMERICAL PROBABILISTIC ALGORITHMS
Remember that it is a question of finding an approximate answer for a numerical

problem.
8.3.1 Buffon's Needle
You spill a box of toothpicks onto a wooden floor. The toothpicks spread out on the
ground in random positions and at random angles, each one independently of all the
others. If you know that there were 355 toothpicks in the box, and that each one is
exactly half as long as the planks in the floor are wide (we realize that this gets
unlikelier every minute ! ), how many toothpicks will fall across a crack between two
planks ?
Clearly any answer between 0 and 355 is possible, and this uncertainty is typical
of probabilistic algorithms. However, as Georges Louis Leclerc showed, the average
number of toothpicks expected to fall across a crack can be calculated : it is almost
exactly 113.
Problem 8.3.1. Why 113 ? Prove it. Why Buffon ?
In fact, each toothpick has one chance in it of falling across a crack. This sug-
gests a probabilistic "algorithm" for estimating the value of it by spilling a sufficiently
large number of toothpicks onto the floor. Needless to say, this method is not used in
practice since better methods of calculating the decimal expansion of it are known.
Sec. 8.3 Numerical Probabilistic Algorithms 229
Furthermore, the precision of your estimate of it would be limited by the precision of

the ratio of the length of the toothpicks to the width of the planks.
* Problem 8.3.2. Supposing that the width of the planks is exactly twice the
length of the toothpicks, how many of the latter should you drop in order to obtain
with probability at least 90% an estimate of it whose absolute error does not exceed
0.001 ?
Problem 8.3.3. Supposing that you have available a random generator of the
type discussed previously, give an algorithm Buffon (n) that simulates the experiment
of dropping n toothpicks. Your algorithm should count the number k of toothpicks
that fall across a crack, and return n / k as its estimate of it. Try your algorithm on a
computer with n = 1000 and n = 10,000, using a pseudorandom generator. What are
your estimates of it ? (It is likely that you will need the value of it during the simula-
tion to generate the random angle - in radians - of each toothpick that falls. But then
nobody said this was a practical method !) 0
Consider next the experiment that consists of throwing n darts at a square target
and counting the number k that fall inside a circle inscribed in this square. We suppose
that every point in the square has exactly the same probability of being hit by a dart.
(It is much easier to simulate this experiment on a computer than to find a darts-player
with exactly the degree of expertise-or of incompetence - required.) If the radius of
the inscribed circle is r, then its area is nr 2, whereas that of the square target is 4r2, so
the average proportion of the darts that fall inside the circle is nr2/4r2 = it/4. This
allows us to estimate it = 4k/n. Figure 8.2.1 illustrates the experiment. In our
example, where 28 darts have been thrown, we are not surprised to find 21 of them
inside the circle, where we expect to see on average 28n/4 = 22.
The following algorithm simulates this experiment, except that it only throws
darts into the upper right quadrant of the target.
At
AF
M r
At
Si r
r
4
r /
Figure 8.2.1. Throwing darts to compute it.

function darts (n)

k -0
for i F- 1 to n do
x F-- uniform (0,1)
y <- uniform (0,1)
if x2+y2<-lthen k-k+1
return 4k In
Problem 8.3.4. What value is estimated if we replace "x - uniform (0, 1) ;

y F- uniform (0,1)" by "x t- uniform (0,1) ; y F x " in this algorithm ?
8.3.2 Numerical Integration
This brings us to the best known of the numerical probabilistic algorithms : Monte
Carlo integration. (This name is unfortunate, because in our terminology it is not an
example of a Monte Carlo algorithm.) Recall that if f : [0, 1] -* [0,1] is a continuous
function, then the area of the surface bounded by the curve y = f (x), the x-axis,
the y-axis, and the line x =1 is given by
if (x)dx
0
To estimate this integral, we could throw a sufficient number of darts at the unit square
and count how many of them fall below the curve.
function hitormiss (f , n)
k+-0
for i E- 1 to n do
x t- uniform (0,1)
y F uniform (0,1)
if y <-f(x)then k Fk+1
return k In
Thus the algorithm using darts to estimate it is equivalent to the evaluation of
45(1-x 2)Zdx
0
by hit-or-miss Monte Carlo.
* Problem 8.3.5. Consider two real constants a and S strictly between 0 and 1.
Prove that if ! is the correct value of the integral and if h is the value returned by the
preceding algorithm, then Prob[ I h - I I < E ] > 1- S whenever the number n of itera-
tions is at least I (1-I)/E2S. Therefore it is sufficient to use n = 11/4c281 (because
I (1-I) <-'/a) to reduce below 8 the probability of an absolute error exceeding E.
Notice that this is not very good : one more decimal digit of precision requires one
hundred times more computation.
Problem 8.3.6. Let a, b, c, and d be four real numbers such that a < b and
c S d, and let f : [a, b I - [c, d ] be a continuous function. Generalize the preceding
algorithm to estimate
h
Jf(x)dx .
a
Your algorithm should accept as parameters, besides f , a, and b, the number n of

iterations to make and the values of c and d.
Usually more efficient probabilistic methods of estimating the value of a definite

integral exist. The simplest consists of generating a number of points randomly and
uniformly inside the interval concerned. The estimated value of the integral is
obtained by multiplying the width of the interval by the arithmetic mean of the values
of the function at these points.
function crude(f ,n,a,b)
sum E-0
for i - 1 to n do
x F- uniform (a, b)
sum - sum + f (x)
return (b -a) x (sum/n)
fh fab
Provided f (x) dx and f 2(x) dx exist, the variance of the estimate calculated by
this algorithm is inversely proportional to the number of points generated randomly,
and the distribution of the estimate is approximately normal when n is large. More-
over, for any fixed number of iterations, its variance is never worse than that of the
hit-or-miss algorithm. One should not immediately conclude, however, that crude
always outperforms hit-or-miss, because hit-or-miss can sometimes make more itera-
tions than crude in a given amount of time. If both are used to compute it as previ-
ously described, for instance, each iteration of crude requires the computation of a
square root, which hit-or-miss can do without by proceeding as in darts.
As presented thus far, Monte Carlo integration is of little practical use. A better
estimate of the integral can generally be obtained by various deterministic methods,
one of the simplest of which is the trapezoidal algorithm.
function trapezoidal (f , n , a, b )
we assume n >-21
delta - (b -a)l(n -1)
sum <-- (f (a)+f (b))/2
for x - a +delta step delta to b -delta do
sum<-sum+f(x)
return sum x delta
Problem 8.3.7. Try to grasp intuitively why this algorithm works. Why is it
called the trapezoidal algorithm ?
Problem 8.3.8. Compare experimentally the trapezoidal algorithm and the two
probabilistic algorithms we have seen. In each case, estimate the value of it by calcu-
lating fo 4(1-x2)zdx.
In general, the trapezoidal algorithm needs many less iterations than does Monte
Carlo integration to obtain a comparable degree of precision. This is typical of most of
the natural functions that we may wish to integrate. However, to every deterministic
integration algorithm, even the most sophisticated, there correspond continuous func-
tions that can be constructed expressly to fool the algorithm. Consider for example the
function f (x) = sin2((100!)itx). Any call on trapezoidal (f , n , 0,1) with 2 5 n <- 101
returns the value zero, even though the true value of this integral is z . No function
can play this kind of trick on the Monte Carlo integration algorithm (although there is
an extremely small probability that the algorithm might manage to make a similar kind
of error, even when f is a thoroughly ordinary function).
In practice, Monte Carlo integration is of interest when we have to evaluate a
multiple integral. If a deterministic integration algorithm using some systematic
method to sample the function is generalized to several dimensions, the number of
sample points needed to achieve a given precision grows exponentially with the dimen-
sion of the integral to be evaluated; If 100 points are needed to evaluate a simple
integral, then it will probably be necessary to use all the points of a 100 x 100 grid, that
is, 10,000 points, to achieve the same precision when a double integral is evaluated;
one million points will be needed for a triple integral, and so on. In Monte Carlo
integration, on the other hand, the dimension of the integral generally has little effect
on the precision obtained, although the amount of work for each iteration is likely to
increase slightly with the dimension. In practice, Monte Carlo integration is used to
evaluate integrals of dimension four or higher. The precision of the answer can be
improved using hybrid techniques that are partly systematic and partly probabilistic.
If the dimension is fixed, it may even be preferable to use quasi Monte Carlo integra-
tion, a technique not discussed here (but Section 8.7 gives a reference for further
reading).
8.3.3 Probabilistic Counting
In the preceding examples, numeric probabilistic algorithms are used to approximate a

real number. The same technique can also be used to estimate the value of an integer.
Let X be a finite set. We would like to know the cardinality of X, but the number of
elements is too large for it to be practical simply to count them one by one. Suppose,
on the other hand, that we are able to choose an element from X randomly, uniformly,
and independently (see Example 8.1.4). A classic brain-teaser helps to explain how
this ability to choose random elements from X allows us to estimate its cardinality.
Problem 8.3.9. A room contains 25 randomly chosen people. Would you be

willing to bet that at least two of them share the same birthday? (Do not read the fol-
lowing paragraphs if you wish to think about this.)
The intuitive answer to the preceding question is almost invariably "of course
not". Nevertheless, the probability that you would win your bet is greater than 56%.
More generally, there are n!/(n-k)! different ways of choosing k distinct objects from
among n objects, taking into account the order in which they are chosen. Since there
are n k different ways of choosing k objects if repetitions are allowed, the probability
that k objects chosen randomly and uniformly from n (with repetition allowed) are all
distinct is n!/(n -k )!n k .
Problem 8.3.10. Calculate 365!/340!36525 to four significant figures.
Problem 8.3.11. The calculation in problem 8.3.10 does not correspond

exactly to the puzzle in problem 8.3.9, because births are not uniformly distributed
through the year. Does this make it more or less likely that you would win your bet?
Justify your answer intuitively. What about leap years ?
Stirling's approximation, n! = (n le )" , and the approximation ln(l+x) Z

x _X2 / 2 when x is near zero, allow us to estimate this probability.
Problem 8.3.12. Show that n!/(n -k )!n k = e-k'/2n .
* Problem 8.3.13. Use the more accurate formulas

n! E I (n le )" [1 + l/12n + O(n-2)]
ln(l+x) Ex -x2/2+x3/3-8(x4) when -l <x < 1

to conclude that
e-k(k-1)12n-k'/6n'±0(max(k'/n',k'/n'))
n /(n-k) nk E
provided that 1 << k << n.
In particular, it is when k z a4n-- , where a = 21n 2 z 1.177, that the probability

of having a repetition exceeds 50%. It is harder to determine the average value of k
corresponding to the first repetition.
** Problem 8.3.14. Let X be a set of n elements from which we randomly, uni-

formly and independently choose elements with replacement. Let k be the number of
choices before the occurrence of the first repetition. When n is large, show that the
expected value of k tends to where (3 = = 1.253.
This suggests the following probabilistic algorithm for estimating the number of
elements in a set X .
function count (X :set)

kF-0
S F- o
a <--uniform (X )
repeat
k F-k+1
S <-S u (a}
a - uniform (X)
until a E S
return 2k2/1t
** Problem 8.3.15. Carry out a statistical analysis of the random variable k2 as a

function of the value of n. Does the function count provide an unbiased estimator of
n? Give an algorithm count 2(X, e, 6) that returns an estimate N of the number n of
elements in X such that Prob[ 11- N In I < c ] >- 1- S.
The algorithm count (X) estimates the number n of elements in X in an expected

time and space that are both in O(X ), provided operations on the set S are counted at
unit cost. This quantity of space can be prohibitive if n is large. The space can be
reduced to a constant with only a linear increase in the execution time by using a pseu-
dorandom generator. This is one of the rare instances where using a truly random gen-
erator would be a hindrance rather than a help. We have not merely to choose an ele-
ment at random from X, but also to step through X in a pseudorandom, and hence
deterministic, way.
Let f : X -* X be a pseudorandom function and let xoE X be a randomly chos-
en starting point. This defines a walk xo, x 1, X2, ... through X, where
xi = f (xi _ 1), i > 0. Because X is a finite set, the sequence {x, }1, must eventually
repeat itself. Let q be the smallest integer such that xq appears more than once in the
sequence, and let k be the smallest integer larger than q such that xq = xk . Let p stand
for k - q. Because the walk is pseudorandom, we also have xq+i = xk+i for every
nonnegative integer i, and more generally, xi = xj whenever j >_ i >- q and
j -i = 0 (mod p ). For this reason {xi } Q_o is called the tail of the walk and {xi } 4 q is
its period.
We are interested in computing the value of k, since this corresponds precisely to
the first repetition, and thus 2k2/it is our estimate on #X. The following exercise
shows how to obtain both q and p in constant space and in a time in 0 (k), hence in an
expected time in 0 ( X ).
* Problem 8.3.16. Consider the sequence {y,1r,O defined by yt =X2, . Show
that the smallest integer t > 0 such that y. = x, is such that q <- t <- q +p, with
t = q +p only possible if q = 0. Show also that t 0 (mod p), and deduce that the
smallest integer j such that xj = x1++ is precisely q, the length of the tail. Incorporate
all these ideas into a simple algorithm that is capable of finding the values of q and p,
hence k = q +p, in a time in 0 (k) and in constant space.
Problem 8.3.17. (Continuation of Problem 8.3.11) The probabilistic counting

algorithm no longer works if the generation of elements in X is not uniform, that is, if
some elements of X are favoured to the detriment of others. Show that it can,
nevertheless, be used unchanged to estimate a lower bound on the number of elements
in X.
The variance of the estimate obtained from this algorithm is unfortunately too
high for most practical applications (unless the solution to Problem 8.3.15 is used).
The following example shows that it can nonetheless be useful if we simply need to
know whether X contains less than a elements or more than b , where a << b .
Example 8.3.1. An endomorphic cryptosystem consists of a finite set K of

keys, a finite set M of messages and two permutations Ek : M -* M and Dk : M -* M
for each key k E K such that Dk (Ek (m)) = m for every m E M and k E K. Such a
system is closed (an undesirable property) if
(dki,k2EK)(3k3EK)(b'mEM)[Ek,(Ek,(m))=Ek,(m)] .
For every m r= M consider the set X,, = { Ek (Ek,(m)) I k I , k 2 E K }. It is clear that
#Xm <_ #K if the system is closed. On the other hand, if the system is not closed, it is
reasonable to hope that #Xm >> #K provided that #M >> #K. It suffices to choose k I
and k2 randomly in K and to calculate Ek (Ek,(m)) in order to choose a random element
from Xm , although this may not imply a uniform distribution on Xm .
All this suggests a probabilistic approach to testing whether or not a cryp-
tosystem is closed. Let m be chosen randomly from M. Probabilistic counting is used
to estimate a lower bound on the cardinality of Xm . (We can only estimate a lower
bound since there is no reason to believe that elements are chosen uniformly from Xm ;
see Problem 8.3.17.) It is improbable that the system is closed if this estimate is
significantly greater than the cardinality of K.
A similar approach was used to demonstrate that the American Data Encryption
Standard is almost certainly not closed. In this application #K = 256 and #M = 264,
which rules out any possibility of an exhaustive verification of the hypothesis that
#X,, > 256. (Even 256 microseconds is more than two millennia.) Implemented using
specialized hardware, the probabilistic algorithm was able to arrive at this conclusion
in less than a day.
8.3.4 More Probabilistic Counting
You have the complete works of Shakespeare on a magnetic tape. How can you deter-
mine the number of different words he used, counting different forms of the same word
(plurals, possessives, and so on) as distinct items?
Two obvious solutions to this problem are based on the techniques of sorting and
searching. Let N be the total number of words on the tape, and let n be the number of
different words. The first approach might be to sort the words on the tape so as to
bring identical forms together, and then to make a sequential pass over the sorted tape
to count the number of different words. This method takes a time in O(N log N) but
requires a relatively modest amount of space in central memory if a suitable external
sorting technique is used. (Such techniques are not covered in this book.) The second
approach consists of making a single pass over the tape and constructing in central
memory a hash table (see Section 8.4.4) holding a single occurrence of each form so
far encountered. The required time is thus in 0 (N) on the average, but it is in S2(Nn)
in the worst case. Moreover, this second method requires a quantity of central memory
in S2(n), which will most likely prove prohibitive.
If we are willing to tolerate some imprecision in the estimate of n, and if we
already know an upper bound M on the value of n (or failing this, on the value of N ),
then there exists a probabilistic algorithm for solving this problem that is efficient with
respect to both time and space. We must first define what sequences of characters are
to be considered as words. (This may depend not only on the character set we are
using, but also whether we want to count such sequences as "jack-rabbit", "jack-
o'lantern", and "jack-in-the-box" as one, two, three, or four words.) Let U be the set of
such sequences. Let m be a parameter somewhat larger than lg M (a more detailed
analysis shows that m = 5 + Fig MI suffices). Let h : U -* {0,1 }'" be a hash function
that transforms a sequence from U in a pseudorandom way into a string of bits of
length m. If y is a string of bits of length k, denote by y [i] the i th bit of y,
1 5i 5k ; denote by 7t(y, b ), b E { 0,1 } , the smallest i such that y [i ] =b, or k+1 if
none of the bits of y is equal to b. Consider the following algorithm.
function wordcnt
{ initialization }
y - string of (m + 1) bits set to zero
{ sequential passage through the tape }
for each word x on the tape do
i -7t(h(x),1)
y[i]<-l
return 7t(y, 0)
Suppose, for example, that the value returned by this algorithm is 4. This means
that the final y begins with 1110. Consequently, there are words x 1, x2 and x3 on the
tape such that h (x1) begins with 1, 01, and 001, respectively, but there is no word x4
such that h (x4) begins with 0001. Let k be the value returned by a call on wordcnt.
Since the probability that a random binary string begins with 0001 is 2-4, it is unlikely
that there could be more than 16 distinct words on the tape. (The probability that
n(h (xi), 1) # 4 for 16 different values of xi is (15/16)16 = 35.6% = e-1, assuming h has
sufficiently random behaviour ; in fact, Prob[ k = 41 n =16 ] = 313/4 %. ) Conversely,
since the probability that a random binary string begins with 001 is 2-3, it is unlikely
that there could be less than 4 distinct words on the tape. (The probability that
it(h (x, ), 1) = 3 for at least one value of xi among 4 different values is
1-(7/8)4 = 41.4 % = 1- a -eh'' ; in fact, Prob[ k = 41 n = 4 ] = 183/4 %. ) This crude rea-
soning indicates that it is plausible to expect that the number of distinct words on the
tape should lie between 2k-2 and 2"`. It is far from obvious how to carry out a more
precise analysis of the unbiased estimate of n given by k.
** Problem 8.3.18. Let R be the random variable returned by this algorithm

when the tape contains n different words and the function h : U ---)10, 1 }' is randomly
chosen with uniform distribution among all such functions (this last assumption is not
reasonable in practice). Prove that the expected value of R is in Ig n + O(1), where
the hidden constant in 0(1) fluctuates around 0.62950 when n is sufficiently large.
Prove further that the standard deviation of R fluctuates around 1.12127.
This offers a first approach for estimating the number of different words : calcu-
late k using the algorithm wordcnt and estimate n as 2k/1.54703. Unfortunately, the
standard deviation of R shows that this estimate may be in error by a factor of 2,
which is unacceptable.
** Problem 8.3.19. Show how to obtain an arbitrarily precise estimate by using a

little more space but with no appreciable increase in execution time, provided n is
sufficiently large. (Hint : by using t strings y 1 , Y 22 ,--- yt of m bits, you can obtain a
relative precision of about 0.78/-J provided t is sufficiently large (t >- 64) ; your hash
function should produce strings of m +lg t bits.) 0
Notice that this approximate counting algorithm is completely insensitive to the

order in which the words appear on the tape and to the number of repetitions of each of
them.
8.3.5 Numerical Problems in Linear Algebra
Many classic problems in linear algebra can be handled by numerical probabilistic

algorithms. Among those are matrix multiplication, the solution of a system of simul-
taneous linear equations, matrix inversion, and the computation of eigenvalues and
eigenvectors. We do not discuss any of them in detail here because it is only for very
specialized applications that they perform better than the obvious deterministic algo-
rithms. The reader is referred to the literature for further discussion.
An intriguing feature of these probabilistic algorithms is their ability to compute
independently the various entries in the result. Consider, for instance, an n x n non-
singular matrix A. Classic deterministic inversion algorithms compute its inverse B as
a whole or perhaps column by column. By contrast, there are probabilistic algorithms
that are capable of estimating the value of B, , for any given 1 <_ i < n and 1 < j <- n , in
about 1/n2 of the time they would require to compute the whole inverse. These algo-
rithms are only applicable if the matrices concerned are well conditioned, typically
requiring that 1-A have only small eigenvalues, where I stands for the identity matrix.
8.4 SHERWOOD ALGORITHMS
Section 1.4 mentions that analysing the average efficiency of an algorithm may some-
times give misleading results. The reason is that any analysis of the average case must
be based on a hypothesis about the probability distribution of the instances to be han-
dled. A hypothesis that is correct for a given application of the algorithm may prove
disastrously wrong for a different application. Suppose, for example, that quicksort
(Section 4.5) is used as a subalgorithm inside a more complex algorithm. Analysis of
this sorting method shows that it takes an average time in O(n log n) to sort n items
provided that the instances to be sorted are chosen randomly. This analysis no longer
bears any relation to reality if in fact we tend to give the algorithm only instances that
are already almost sorted. Sherwood algorithms free us from the necessity of worrying
about such situations by evening out the time required on different instances of a given
size.
Let A be a deterministic algorithm and let to (x) be the time it takes to solve some
instance x. For every integer n let X, be the set of instances of size n. Supposing that
every instance of a given size is equiprobable, the average time taken by the algorithm
to solve an instance of size n is
tA(n)= Y. tA(x)/#X
xEX
This in no way rules out the possibility that there exists an instance x of size n such
that to (x) >> FA (n). We wish to obtain a probabilistic algorithm B such that
tB (x) = 1A (n) + s (n) for every instance x of size n, where tB (x) is the expected time
taken by algorithm B on instance x and s (n) is the cost we have to pay for this unifor-
mity.
Algorithm B may occasionally take more time than tA(n)+s(n) on an instance x
of size n, but this fortuitous behaviour is only due to the probabilistic choices made by
the algorithm, independently of the specific instance x to be solved. Thus there are no
longer worst-case instances, but only worst-case executions. If we define
Fe(n) = I te(a)l#X
xex'
the average expected time taken by algorithm B on a random instance of size n, it is

clear that to (n) = tA(n)+s (n). The Sherwood algorithm thus involves only a small
increase in the average execution time if s (n) is negligible compared to to (n).
8.4.1 Selection and Sorting
We return to the problem of finding the kth smallest element in an array T of n ele-
ments (Section 4.6 and Example 8.1.1). The heart of this algorithm is the choice of a
Sec. 8.4 Sherwood Algorithms 239
pivot around which the other elements of the array are partitioned. Using the pseudo-
median as the pivot assures us of a linear execution time in the worst case, even
though finding this pivot is a relatively costly operation. On the other hand, using the
first element of the array as the pivot assures us of a linear execution time on the
average, with the risk that the algorithm will take quadratic time in the worst case
(Problems 4.6.5 and 4.6.6). Despite this prohibitive worst case, the simpler algorithm
has the advantage of a much smaller hidden constant on account of the time that is
saved by not calculating the pseudomedian. The decision whether it is more important
to have efficient execution in the worst case or on the average must be taken in the
light of the particular application. If we decide to aim for speed on the average thanks
to the simpler deterministic algorithm, we must make sure that the instances to be
solved are indeed chosen randomly and uniformly.
Suppose that the elements of T are distinct, and that we are looking for the
median. The execution times of the algorithms in Section 4.6 do not depend on the
values of the elements of the array, but only on their relative order. Rather than
express this time as a function solely of n, which forces us to distinguish between the
worst case and an average case, we can express it as a function of both n and a, the
permutation of the first n integers that corresponds to the relative order of the elements
of the array.
Let tp (n , a) and t, (n , (5) be the times taken by the algorithm that uses the pseu-
domedian and by the simplified algorithm, respectively. The simplified algorithm is
generally faster : for every n , t, (n , a) < tp (n , (Y) for most values of a. On the other
hand, the simplified algorithm is sometimes disastrous : t, (n ,a) is occasionally much
greater than tp (n , a). More precisely, let S, be the set of n! permutations of the first n
integers. Define F, (n) = I t , (n ,a)/n! . We have the following equations :
6ESn
(3cp)(3n,ElN)(`dn?n1)(Vo
(acs << cp)(3n2E lNT)(Vn ? n2)[ts(n) <_ c, n ]
but
(3c,)(3n3E IN)(Vn >C, n2 >> cp n >_ tp(n,(Y)] .
For the execution time to be independent of the permutation a, it suffices to

choose the pivot randomly among the n elements of the array T. The fact that we no
longer calculate a pseudomedian simplifies the algorithm and avoids recursive calls.
The resulting algorithm resembles the iterative binary search of Section 4.3.
function selectionRH (T [ 1 .. n ], k)
{ finds the kth smallest element in array T ;
we assume that 1 <_ k <_ n }
iE- l;j f- n
while i < j do
m <-- T [uniform (i .. j)]
partition (T,i,j,m,u,v)
ifk v then i -v+1

else i, j i- k
return T[i]
Here partition (T, i , j , m, var u, var v) pivots the elements of T [i .. j ] around the
value m ; after this operation the elements of T [i .. u -I] are less than m , those of
T [u .. v ] are equal to m, and those of T [v +1 .. j ] are greater than m. The values of u
and v are calculated and returned by the pivoting algorithm (see Problem 4.6.1).
A similar analysis to that of Problem 4.6.5 shows that the expected time taken by
this probabilistic selection algorithm is linear, independently of the instance to be
solved. Thus its efficiency is not affected by the peculiarities of the application in
which the algorithm is used. It is always possible that some particular execution of the
algorithm will take quadratic time, but the probability that this will happen becomes
increasingly negligible as n gets larger, and, to repeat, this probability is independent
of the instance concerned. Let tRH (n , (Y) be the average time taken by the Sherwood
algorithm to determine the median of an array of n elements arranged in the order
specified by a. The probabilistic nature of the algorithm ensures that tRH (n , (Y) is
independent of a. Its simplicity ensures that
(3noEN)(f/n >- no)(VaESf)[tRH(n,(Y) < tp(n,a)]
To sum up, we started with an algorithm that is excellent when we consider its
average execution time on all the instances of some particular size but that is very
inefficient on certain specific instances. Using the probabilistic approach, we have
transformed this algorithm into a Sherwood algorithm that is efficient (with high proba-
bility) whatever the instance considered.
Problem 8.4.1. Show how to apply the Sherwood style of probabilistic

approach to quicksort (Section 4.5). Notice that quicksort must first be modified along
the lines of Problem 4.5.4. 0
8.4.2 Stochastic Preconditioning
The modifications we made to the deterministic algorithms for sorting and for selection
in order to obtain Sherwood algorithms are simple. There are, however, occasions
when we are given a deterministic algorithm efficient on the average but that we
cannot reasonably expect to modify. This happens, for instance, if the algorithm is part
of a complicated, badly documented software package. Stochastic preconditioning
allows us to obtain a Sherwood algorithm without changing the deterministic algo-
rithm. The trick is to transform the instance to be solved into a random instance, to
use the given deterministic algorithm to solve this random instance, and then to deduce
the solution to the original instance.
Suppose the problem to be solved consists of the computation of some function
f : X -i Y for which we already have an algorithm that is efficient on the average. For
every integer n, let X be the set of instances of size n, and let A be a set with the
same number of elements. Assume random sampling with uniform distribution is pos-
sible efficiently within A . Let A be the union of all the A . Stochastic precondi-
tioning consists of a pair of functions u : X x A -* X and v : A x Y - Y such that
i.
H. )[f (x) = v(r, f(u(x,r)))]; and
iii. the functions u and v can be calculated efficiently in the worst case.
We thus obtain the following Sherwood algorithm.

function RH (x)
{ computation of f (x) by Sherwood algorithm }
let n be the size of x
r - uniform (A,)
y E- u (x, r) { random instance of size n }
s - f (y) { solved by the deterministic algorithm }
return v (r, s )
Whatever the instance x to be solved, the first property ensures that this instance
is transformed into an instance y chosen randomly and uniformly from all those of the
same size. Thanks to the second property, the solution to this random instance allows
us to recover the solution of the original instance x.
Example 8.4.1. The stochastic preconditioning required for selection or for

sorting is the same : it is simply a question of randomly shuffling the elements of the
array in question. No posttreatment (function v) is needed to recover the solution in
these cases. Simply call the following procedure before the deterministic sorting or
selection algorithm.
procedure shuffle (T [I .. n ] )
for i f- 1 to n -1 do
j - uniform (i n)
interchange T [i] and T [ j ]
Example 8.4.2. Recall that no efficient algorithm is known for calculating

discrete logarithms (Section 4.8). Suppose for the purposes of illustration that
someone were to discover an algorithm efficient on the average but prohibitively slow
in the worst case. Denote the discrete logarithm of x modulo p to the base g by
logg,p x. The following equations allow us to transform our hypothetical algorithm

into a Sherwood algorithm :
i. logg,p (xy mod p) = (logg,p x + logg,p y) mod (p - 1);

ii. logs,p(gr modp)=r for 05r Sp-2.
Here is the Sherwood algorithm.

function dlogRH (g , x , p )
r <--uniform (0.. p -2)
b Fdexpoiter (g, r , p) { Section 4.8)
a F- bx mod p
s f- logg,p a { using the assumed algorithm }
return (s - r) mod (p -1) 11
Problem 8.4.2. Why does the algorithm dlogRH work? Point out the func-
tions corresponding to u and v.
Problem 8.4.3. Find other problems that can benefit from stochastic precondi-
tioning.
Stochastic preconditioning offers an intriguing possibility : computing with an

encrypted instance. Assume that you would like to compute f (x) for some instance x
but that you lack the computing power or the efficient algorithm to do so. Assume,
furthermore, that some other party is capable of carrying out this computation and will-
ing to do so for you, perhaps for a fee. What should you do if you are unwilling to
divulge your actual request x? The solution is easy if stochastic preconditioning
applies to the computation of f : use the function u to encrypt x into some random y,
have f (y) computed for you, and then use the function v to deduce f (x). This process
yields no information on your actual request, except for its size, because the probability
distribution of u (x, r) is independent of x as long as r is chosen randomly with uni-
form probability.
8.4.3 Searching an Ordered List
A list of n keys sorted into ascending order is implemented using two arrays val [ 1 .. n ]
and ptr [ 1 .. n ] and an integer head. The smallest key is in val [head ], the next smal-
lest is in val [ptr [head ]], and so on. In general, if val [i] is not the largest key, then
ptr [i] gives the index of the following key. The end of the list is marked by
ptr [i] = 0. The rank of a key is the number of keys in the list that are less than or
equal to the given key. For instance, here is one way to represent the list 1, 2, 3, 5, 8,
13, 21 .
i 1 2 3 4 5 6 7
val [i] 2 3 13 1 5 21 8
ptr [i] 2 5 6 1 7 0 3
In this example head = 4 and the rank of 13 is 6.

We can use binary search to find a key efficiently in a sorted array. Here, how-
ever, there is no obvious way to select the middle of the list, which would correspond
to the first step in binary search. In fact, any deterministic algorithm takes a time in
Q (n) in the worst case to find a key in this kind of list.
Problem 8.4.4. Prove the preceding assertion. (Hint : show how a worst-case
instance can be constructed systematically from the probes made into the list by any
given deterministic algorithm.)
Despite this inevitable worst case, there exists a deterministic algorithm that is
capable of carrying out such a search in an average time in 0 (I ). From this we can
obtain a Sherwood algorithm whose expected execution time is in 0 (') whatever the
instance to be solved. As usual, the Sherwood algorithm is no faster on the average
than the corresponding deterministic algorithm, but it does not have such a thing as a
worst-case instance.
Suppose for the moment that the required key is always in fact present in the list,
and that all the elements of the list are distinct. Given a key x, the problem is thus to
find that index i, 1 5 i <_ n , such that val [i] = x. Any instance can be characterized
by a permutation a of the first n integers and by the rank k of the key we are looking
for. Let Sn be the set of all n! permutations. If A is any deterministic algorithm,
tA (n , k, (Y) denotes the time taken by this algorithm to find the key of rank k among the
n keys in the list when the order of the latter in the array val is specified by the permu-
tation a. In the case of a probabilistic algorithm, to (n , k , (Y) denotes the expected
value of this time. Whether the algorithm is deterministic or probabilistic, wA (n) and
mA (n) denote its worst-case and its mean time, respectively. Thus
WA(n)=max{tA(n,k,(Y)I 1 <_k <_n and aeS,},
and
mA(n)= E EtA(n,k,(Y)
nxn! vES k=I
Problem 8.4.4 implies that wA (n) E S2(n) for every deterministic algorithm A. We want
a deterministic algorithm B such that mB (n)E 0 ('/n) and a Sherwood algorithm C
such that we (n) = mB (n).
The following algorithm finds a key x starting from some position i in the list,
provided that x ? val [i] and that x is indeed present.
function search (x, i )

while x > val [i] do i <- ptr [i ]
return i
Here is the obvious deterministic search.
function A (x)
return search (x, head )
Problem 8.4.5. Let iA (n, k) be the exact number of references to the array
val made by the algorithm A to find the key of rank k in a list of n keys. (The order a
of the keys is irrelevant for this algorithm.) Define WA (n) and mA (n) similarly. Deter-
mine 1A (n, k) for every integer n and for every k between 1 and n. Determine WA (n)
and mA (n) for every integer n .
Here is a first probabilistic algorithm.

function D (x)
i - uniform (L. n)
y - val [i ]
case x < y : return search (x, head )
x > y : return search (x , ptr [i ])
otherwise : return i
Problem 8.4.6. Determine 1D (n , k) for every integer n and for every k
between 1 and n. Determine wD (n) and mD (n) for every integer n. As a function of
n, what values of k maximize t D (n, k) ? Compare wD (n) and mA (n). Give explicitly
a function f (n) such that ?D (n , k) < iA (n , k) if and only if k > f (n). (See
Problem 8.4.5 for the definition of i , w, and m.)
Problem 8.4.7. The quantities F, w, and m introduced in the previous prob-

lems facilitate our analysis. Show, however, that they do not tell the whole story by
exhibiting a deterministic algorithm E such that WE (n) E 0 (log n), thus apparently con-
tradicting Problem 8.4.4.
The following deterministic algorithm is efficient on the average.

function B (x)
i - head
max <- val [i]
for j f- l to L'u J do
y <--val[j]
ifmax<y <_x theni F j,max - y
return search (x, i )
Problem 8.4.8. Intuitively, why should we choose to execute the for loop
times in algorithm B ?
* Problem 8.4.9. Prove that m8 (n) E 0 (' ). [Hint : Let M1,,, be the random
variable that corresponds to the minimum of I integers chosen randomly, uniformly,
and independently with replacement from the set ( 1, 2, ... , n }. Find a link between
this random variable and the average-case analysis of algorithm B. Show that the
expected value of M1,,, is about n /(l + 1) when 1 is a constant and about when
Problem 8.4.10. Show, however, that w8 (n) E S2(n), which is unavoidable

from Problem 8.4.4. To do this, give explicitly a permutation a and a rank k such that
tB (n , k, 0) E U(n).
Problem 8.4.11. Starting from the deterministic algorithm B, give a Sher-

wood algorithm C such that wC (n) E O
Problem 8.4.12. (Continuation of Problem 8.4.11) Show more precisely

that wC (n) E 2' + O(1), where the meaning of w is given in Problem 8.4.5.
Problem 8.4.13. Give an efficient Sherwood algorithm that takes into account
the possibility that the key we are seeking may not be in the list and that the keys may
not all be distinct. Analyse your algorithm.
Problem 8.4.14. Use the structure and the algorithms we have just seen to
obtain a Sherwood sorting algorithm that is able to sort n elements in a worst-case
expected time in 0 (n 312). Is this better than 0 (n log n) ? Justify your answer.
8.4.4 Universal Hashing
Hash coding, or simply hashing, is used in just about every compiler to implement the
symbol table. Let X be the set of possible identifiers in the language to be compiled,
and let N be a parameter chosen to obtain an efficient system. A hash function is a
function h : X --* { 1, 2, ... , N }. Such a function is a good choice if it efficiently
disperses all the probable identifiers, that is, if h (x) # h (y) for most of the pairs x * y
that are likely to be found in the same program. When x * y but h (x) = h (y), we say
that there is a collision between x and y . The hash table is an array T [I .. N ] of lists
in which T [i ] is the list of those identifiers x found in the program such that h (x) = i.
The load factor of the table is the ratio a = n IN, where n is the number of distinct
identifiers in the table. (The ratio (x may well be greater than 1.) If we suppose that
every identifier and every pointer occupies a constant amount of space, the table takes
space in O(N+n) and the average length of the lists is a. Thus we see that increasing
the value of N reduces the average search time but increases the space occupied by the
table.
Problem 8.4.15. Other ways to handle collisions, besides the use of a table of
lists as outlined here, are legion. Suggest a few.
Problem 8.4.16. What do you think of the "solution" that consists of ignoring
the problem? If we are given an a priori upper bound on the number of identifiers a
program may contain, does it suffice to choose N rather larger than this bound to
ensure that the probability of a collision is negligible ? (Hint : solve Problem 8.3.9
before answering.)
Problem 8.4.17. Show that n calls on the symbol table can take a total time in
f2(n2) in the worst case.
This technique is very efficient provided that the function h disperses the
identifiers properly. If we suppose, however, that #X is very much greater than N, it is
inevitable that certain programs will cause a large number of collisions. These pro-
grams will compile slowly every time they are submitted. In a sense they are paying
the price for all other programs to compile quickly. A Sherwood approach allows us to
retain the efficiency of hashing on the average, without arbitrarily favouring some pro-
grams at the expense of others. (If you have not yet solved Problem 8.2.3, now is the
time to think some more about it!)
The basic idea is to choose the hash function randomly at the beginning of each
compilation. A program that causes a large number of collisions during one compila-
tion will therefore probably be luckier next time it is compiled. Unfortunately, there
are f a r too many functions from X into { 1, 2, ... , N } for it to be reasonable to choose
one at random.
Problem 8.4.18. How many functions f :A - B are there if the cardinalities

of A and B are a and b, respectively ?
This difficulty is solved by universal hashing. By definition a class H of func-

tions from A to B is universal2 if # (h E H (h (x) = h (y)) < #H/#B for every X, y E A
such that x #y. Let H be a universa12 class of functions f r o m X to { 1, 2, ... , N } , and
let x and y be any two distinct identifiers. If we choose a hash function h randomly
and uniformly in H, the probability of a collision between x and y is therefore at most
1/N. The following problem generalizes this situation.
* Problem 8.4.19. Let S c X be a set of n identifiers already in the table. Let

xEX \ S be a new identifier. Prove that the average number of collisions between x
and the elements of S (that is, the average length of the list T [h (x)]) is less than or
equal to the load factor a. Prove further that the probability that the number of colli-
sions will be greater than t a is less than 1/t for all t >_ 1.
Several efficient universal2 classes of functions are known. We content ourselves

with mentioning just one.
Sec. 8.5 Las Vegas Algorithms 247
* Problem 8.4.20. Let X be { 0, l , 2, ... , a -11, and let p be a prime number

not less than N. Let m and n be two integers. Define hm n :X -* (0, 1, ... , N-1 ) by
h,,,, (x) = ((mx + n) mod p) mod N.
Prove that H = { hm,,, 1 5 m < p and 0 <- n < p E is a universal2 class of func-
I
tions. (Remarks : In practice we take N to be a power of 2 so that the second mod

operation can be executed efficiently. It is also more efficient to carry out all the com-
putations in a Galois field whose cardinality is a power of 2.)
* Problem 8.4.21. Find applications of universal hashing that have nothing to

do with compilation, nor even with management of a symbol table.
8.5 LAS VEGAS ALGORITHMS
Although its behaviour is more uniform, a Sherwood algorithm is no faster on the

average than the deterministic algorithm from which it arose. A Las Vegas algorithm,
on the other hand, allows us to obtain an increase in efficiency, sometimes for every
instance. It may be able to solve in practice certain problems for which no efficient
deterministic algorithm is known even on the average. However, there is no upper
bound on the time that may be required to obtain a solution, even though the expected
time required for each instance may be small and the probability of encountering an
excessive execution time is negligible. Contrast this to a Sherwood algorithm, where
we are able to predict the maximum time needed to solve a given instance. For
example, the Sherwood version of quicksort (Problem 8.4.1) never takes a time in
excess of 0 (n2) to sort n elements, whatever happens.
The distinguishing characteristic of Las Vegas algorithms is that now and again
they take the risk of making a random decision that renders it impossible to find a
solution. Thus these algorithms react by either returning a correct solution or admit-
ting that their random decisions have led to an impasse. In the latter case it suffices to
resubmit the same instance to the same algorithm to have a second, independent chance
of arriving at a solution. The overall probability of success therefore increases with the
amount of time we have available.
Las Vegas algorithms usually have a return parameter success, which is set to
true if a solution is obtained and false otherwise. The typical call to solve instance x
is LV (x , y , success), where y is a return parameter used to receive the solution thus
obtained whenever success is set to true. Let p (x) be the probability of success of the
algorithm each time that it is asked to solve the instance x. For an algorithm to be
correct, we require that p (x) > 0 for every instance x. Better still is the existence of
some constant 8 > 0 such that p (x) >- S for every instance x. Let s (x) and e (x) be the
expected times taken by the algorithm on instance x in the case of success and of
failure, respectively. Now consider the following algorithm.
function obstinate (x)

repeat
LV (x , y , success)
until success
return y
Let t (x) be the expected time taken by the algorithm obstinate to find an exact solution
to the instance x. Neglecting the time taken by the control of the repeat loop, we
obtain the following recurrence :
t (x) = p (x) s (x) + (1- p (x)) (e (x) + t (x)).
This follows because the algorithm succeeds at the first attempt with probability p (x),
thus taking an expected time s (x). With probability 1- p (x) it first makes an unsuc-
cessful attempt to solve the instance, taking an expected time e (x), before starting all
over again to solve the instance, which still takes an expected time t (x). The recurrence
is easily solved to yield
)x)
WW
t (x) = s (x) + 1 - ( (x)
e
There is a compromise to be made between p (x), s (x), and e (x) if we want to

minimize t (x). For example, it may be preferable to accept a smaller probability of
success if this also decreases the time required to know that a failure has occurred.
Problem 8.5.1. Suppose that s (x) and e (x) are not just expected times, but
that they are in fact the exact times taken by a call on LV (x , ...) in the case of success
and of failure, respectively. What is the probability that the algorithm obstinate will
find a correct solution in a time not greater than t, for any t >- s (x) ? Give your
answer as a function of t, s (x), e (x) and p (x).
8.5.1 The Eight Queens Problem Revisited
The eight queens problem (Section 6.6.1) provides a nice example of this kind of algo-
rithm. Recall that the backtracking technique used involves systematically exploring
the nodes of the implicit tree formed by the k-promising vectors. Using this technique,
we obtain the first solution after examining only 114 of the 2,057 nodes in the tree.
This is not bad, but the algorithm does not take into account one important fact : there
is nothing systematic about the positions of the queens in most of the solutions. On
the contrary, the queens seem more to have been positioned haphazardly. This obser-
vation suggests a greedy Las Vegas algorithm that places queens randomly on succes-
sive rows, taking care, however, that the queens placed on the board do not threaten
one another. The algorithm ends either successfully if it manages to place all the
queens on the board or in failure if there is no square in which the next queen can be
added. The resulting algorithm is not recursive.
procedure QueensLV (var success)

{ if success = true at the end, then try [1 .. 8] (a global array)
contains a solution to the eight queens problem }
col, diag45, diag 135 f- 0
k -0
repeat
{ try [1 .. k ] is k-promising }
nb - 0
for i E- I to 8 do
if i 0 col and i -k e diag45 and i +k 0 diag 135
then ( column i is available for the (k + I )st queen }
nb -nb+1
if uniform (1 .. nb) = 1
then ( maybe try column i }
j4-i
if nb > 0
then ( amongst all nb possibilities for the (k + 1)st queen,
it is column j that has been chosen (with probability 1 / nb) }
try [k +I] '- j
col - colu{j}
diag45 - diag45 u { j - k }
diag 135 F- diag 135 u { j + k }
{ try [ 1 .. k + 1 ] is (k + 1)-promising }
k-k+1
until nb = 0 or k =8
success - (nb > 0)
To analyse the efficiency of this algorithm, we need to determine its probability p
of success, the average number s of nodes that it explores in the case of success, and
the average number e of nodes that it explores in the case of failure. Clearly s = 9
(counting the 0-promising empty vector). Using a computer we can calculate
p = 0.1293 . and e = 6.971 A solution is therefore obtained more than one
time out of eight by proceeding in a completely random fashion ! The expected
number of nodes explored if we repeat the algorithm until a success is finally obtained
is given by the general formula s + (1-p )e/p = 55.927 , less than half the
number of nodes explored by the systematic backtracking technique.
Problem 8.5.2. When there is more than one position open for the (k + I )st
queen, the algorithm QueensLV chooses one at random without first counting the
number nb of possibilities. Show that each position has, nevertheless, the same proba-
bility of being chosen.
We can do better still. The Las Vegas algorithm is too defeatist : as soon as it
detects a failure it starts all over again from the beginning. The backtracking
algorithm, on the other hand, makes a systematic search for a solution that we know
has nothing systematic about it. A judicious combination of these two algorithms first
places a number of queens on the board in a random way, and then uses backtracking
to try and add the remaining queens, without, however, reconsidering the positions of
the queens that were placed randomly.
An unfortunate random choice of the positions of the first few queens can make
it impossible to add the others. This happens, for instance, if the first two queens are
placed in positions 1 and 3, respectively. The more queens we place randomly, the
smaller is the average time needed by the subsequent backtracking stage, but the
greater is the probability of a failure.
The resulting algorithm is similar to QueensLV, except that the last two lines are
replaced by
until nb = 0 or k = stopVegas
if nb > 0 then backtrack (k , col, diag45, diag 135, success)
else success *- false ,
where 1 < stopVegas S 8 indicates how many queens are to be placed randomly before
moving on to the backtracking phase. The latter looks like the algorithm Queens of
Section 6.6.1 except that it has an extra parameter success and that it returns immedi-
ately after finding the first solution if there is one.
The following table gives for each value of stopVegas the probability p of suc-
cess, the expected number s of nodes explored in the case of success, the expected
number e of nodes explored in the case of failure, and the expected number
t = s +(I -p)e/p of nodes explored if the algorithm is repeated until it eventually
finds a solution. The case stopVegas = 0 corresponds to using the deterministic algo-
rithm directly.
stopVegas p s e t
0
1
1.0000
1.0000
114.00
39.63
-- 114.00
39.63
2 0.8750 22.53 39.67 28.20
3 0.4931 13.48 15.10 29.01
4 0.2618 10.31 8.79 35.10
5 0.1624 9.33 7.29 46.92
6 0.1357 9.05 6.98 53.50
7 0.1293 9.00 6.97 55.93
8 0.1293 9.00 6.97 55.93
We tried these different algorithms on a CYBER 835. The pure backtracking

algorithm finds the first solution in 40 milliseconds, whereas an average of 10 mil-
liseconds is all that is needed if the first two or three queens are placed at random. The
original greedy algorithm QueensLV, which places all the queens in a random way,
takes on the average 23 milliseconds to find a solution. This is a fraction more than
half the time taken by the backtracking algorithm because we must also take into
account the time required to make the necessary pseudorandom choices of position.
Problem 8.5.3. If you are still not convinced of the value of this technique,
we suggest you try to solve the twelve queens problem without using a computer.
First, try to solve the problem systematically, and then try again, this time placing the
first five queens randomly.
For the eight queens problem, a systematic search for a solution beginning with
the first queen in the first column takes quite some time. First the trees below the
2-promising nodes [1,3] and [1,4] are explored to no effect. Even when the search
starting from node [1,5] begins, we waste time with [1,5,21 and [1,5,71. This is one
reason why it is more efficient to place the first queen at random rather than to begin
the systematic search immediately. On the other hand, a systematic search that begins
with the first queen in the fifth column is astonishingly quick. (Try it!) This unlucky
characteristic of the upper left-hand corner is nothing more than a meaningless
accident. For instance, the same corner is a better than average starting point for the
problems with five or twelve queens. What is significant, however, is that a solution
can be obtained more rapidly on the average if several queens are positioned randomly
before embarking on the backtracking phase. Once again, this can be understood intui-
tively in terms of the lack of regularity in the solutions (at least when the number of
queens is not 4k + 2 for some integer k).
Here are the values of p, s, e, and t for a few values of stopVegas in the case of
the twelve queens problem.
p
stopVegas
0
5
1.0000
0.5039
s
262.00
33.88
-e
47.23
t
262.00
80.39
12 0.0465 13.00 10.20 222.11
On the CYBER 835 the Las Vegas algorithm that places the first five queens randomly
before starting backtracking requires only 37 milliseconds on the average to find a
solution, whereas the pure backtracking algorithm takes 125 milliseconds. As for the
greedy Las Vegas algorithm, it wastes so much time making its pseudorandom choices
of position that it requires essentially the same amount of time as the pure backtracking
algorithm.
An empirical study of the twenty queens problem was also carried out using an
Apple II personal computer. The deterministic backtracking algorithm took more than
2 hours to find a first solution. Using the probabilistic approach and placing the first
ten queens at random, 36 different solutions were found in about five and a half
minutes. Thus the probabilistic algorithm turned out to be almost 1,000 times faster
per solution than the deterministic algorithm.
** Problem 8.5.4. If we want a solution to the general n queens problem, it is

obviously silly to analyse exhaustively all the possibilities so as to discover the optimal
value of stopVegas, and then to apply the corresponding Las Vegas algorithm. In fact,
determining the optimal value of stopVegas takes longer than a straightforward search
for a solution using backtracking. (We needed more than 50 minutes computation on
the CYBER to establish that stopVegas = 5 is the optimal choice for the twelve queens
problem !) Find an analytic method that enables a good, but not necessarily optimal,
value of stopVegas to be determined rapidly as a function of n.
** Problem 8.5.5. Technically, the general algorithm obtained using the previous
problem (first determine stopVegas as a function of n, the number of queens, and then
try to place the queens on the board) can only be considered to be a Las Vegas algo-
rithm if its probability of success is strictly positive for every n. This is the case if
and only if there exists at least one solution. If no solution exists, the obstinate proba-
bilistic algorithm will loop forever without realizing what is happening. Prove or
disprove : the n queens problem can be solved for every n >- 4. Combining this with
Problem 8.5.4, can you find a constant S > 0 such that the probability of success of the
Las Vegas algorithm to solve the n queens problem is at least S for every n ?
8.5.2 Square Roots Modulo p
Let p be an odd prime. An integer x is a quadratic residue modulo p if 1 <- x < p -1

and if there exists an integer y such that x = y2 (mod p). Such a y is a square root o
x modulo p provided 1 :- y <- p - 1. For instance, 63 is a square root of 55 modulo
103. An integer z is a quadratic nonresidue modulo p if 1 < z <- p - I and z is not a
quadratic residue modulo p. Any quadratic residue has at least two distinct square
roots since (p _Y)2 = p 2 - 2py + y 2 = y 2 (mod p).
Problem 8.5.6. Prove that p - y # y and that 1 <_ p - y <- p - 1.
Problem 8.5.7. Prove, on the other hand, that no quadratic residue has more
than two distinct square roots. (Hint : assuming that a 2 ° h 2 (mod p), consider
a 2 - b 2.)
Problem 8.5.8. Conclude from the preceding results that exactly half the
integers between 1 and p -1 are quadratic residues modulo p.
* Problem 8.5.9. Prove that x(P-1)/2 =± 1 (mod p) for every integer

1<- x <- p -l and every odd prime p. Prove further that x is a quadratic residue
modulo p if and only if x(P- 1)/2 = + 1 (mod p). (Hint: one direction follows immedi-
ately from Fermat's theorem : x n = 1 (mod p ); the other direction requires some
knowledge of group theory.) 0
The preceding problem suggests an efficient algorithm for testing whether x is a
quadratic residue modulo p : it suffices to use the fast exponentiation algorithm of Sec-
tion 4.8 to calculate x(p - 1)/2 mod p. Given an odd prime p and a quadratic residue x
modulo p, does there exist an efficient algorithm for calculating the two square roots
of x modulo p ? The problem is easy when p - 3 (mod 4), but no efficient deter-
ministic algorithm is known to solve this problem when p =- 1 (mod 4).
Problem 8.5.10. Suppose that p - 3 (mod 4) and let x be a quadratic residue

modulo p. Prove that ±x(p+ 1)/4 mod p are the two square roots of x modulo p. Cal-
culate 5526 mod 103 and verify that its square modulo 103 is indeed 55.
There exists, however, an efficient Las Vegas algorithm to solve this problem
when p = 1 (mod 4). Let us decide arbitrarily to denote by the smaller of the two
square roots of x. Even if the value of is unknown, it is possible to carry out the
symbolic multiplication of a +b-6- and c + d modulo p, where a, b, c, and d
are integers between 0 and p - 1. This product is ((ac +bdx) mod p) +
((ad + bc) mod p Note the similarity to a product of complex numbers. The
symbolic exponentiation (a +b-6- )" can be calculated efficiently by adapting the algo-
rithms of Section 4.8.
Example 8.5.1. Let p = 53 = 1 (mod 4) and x = 7. A preliminary computa-

tion shows that x is a quadratic residue modulo p since 726 = 1 (mod 53). Let us calcu-
late symbolically (1 mod 53. (All the following calculations are modulo 53.)
(1 + 7)2 = (1 + -0) (1 + 17) =8+

(1+V)3 =(1+1)(8+2J) =22+10,F7_
=(22+101)(22+101)=18+16I
(1+J)12=(18+161)(18+l61)=49+461
(1+x)13=(1+x)(49+46-,F7-) =0+421
(1+ 41 )26 = -1 (mod 53). Since 26 = (p - 1)/2, we conclude

that (1+') mod 53 is a quadratic nonresidue modulo 53 (Problem 8.5.9). But 7 has
two square roots modulo p, and the symbolic calculation that we just carried out is
valid regardless of which of them we choose to call '.
(1-1) mod 53 is also a quadratic nonresidue modulo 53. What happens if we
Consequently,
calculate symbolically (a + / )26 == c +dI (mod 53) in a case when one of

(a +') mod 53 and (a -i7) mod 53 is a quadratic residue modulo p and the other is
not? Suppose,for instance, that c +d 1 _- 1 (mod 53) and c
c -d =- I (mod 53). Adding these two equations, we obtain 2c = 0 (mod 53)
and hence c = 0 since 0 <- c < 52. Subtracting them, we find 2d/ - 2 (mod 53), and
hence d / - 1 (mod 53).
Problem 8.5.11. Using Example 8.5.1 as a guide, carry out the symbolic cal-
culation (2+I7)26 = 0+41J (mod 53) in detail.
To obtain a square root of 7, the preceding problem shows that we need only find
the unique integer y such that 1 <- y <- 52 and 41y - 1 (mod 53). This can be done
efficiently using a modification of Euclid's algorithm for calculating the greatest
common divisor (Section 1.7.4).
* Problem 8.5.12. Let u and v be two positive integers, and let d be their
greatest common divisor.
i. Prove that there exist integers a and b such that au + by = d. [Hint : Suppose
without loss of generality that u >- v. If v = d, the proof is trivial (a = 0 and
b = 1). Otherwise, let w = u mod v. First show that d is also the greatest
common divisor of v and w. (This is the heart of Euclid's algorithm.) By
mathematical induction, now let a' and b' be such that a'v + b'w = d. Then we
need only take a = b' and b = a' -b' Lu l v j.]
ii. Give an efficient iterative algorithm for calculating d, a, and b from u and v.
Your algorithm should not calculate d before starting to work on a and b.
iii. If p is prime and 1 a <- p - 1, prove that there exists a unique y such that
1 <- y <- p -1 and ay I (mod p). Give an efficient algorithm for calculating y
given p and a.
In our example (following Problem 8.5.11), we find y = 22 because

41 x 22 - 1 (mod 53). This is indeed a square root of 7 modulo 53 since
222 = 7 (mod 53). The other square root is 53 - 22 = 31. This suggests the following
Las Vegas algorithm for calculating square roots.
procedure rootLV (x , p, var y, var success)
{ may find some y such that y 2 = x (modp )
assuming p is a prime, p - 1 (mod 4),
and x is a quadratic residue modulo p }
a F- uniform (I.. p -1)
if a 2 = x (modp) { very unlikely }
then success f- true
y Fa
else compute c and d such that 0 <-c p-1,0 <d p-1
and (a +')(r-1)12-c+dL (modp)
if d = 0 then success F false
else { c = 0)
success f- true
compute y such that 1 <- y < p -1 and dy = l (mod p )
It remains to determine the probability of success of this algorithm.
* Problem 8.5.13. Let p = 1 (mod 4) be prime, and let x be a quadratic residue

modulo p. An integer a, 1 <- a p -1, gives the key to ' if (a2-x) mod p is not a
quadratic residue modulo P. Prove that
i. The Las Vegas algorithm finds a square root of x if and only if it randomly
chooses an a that gives the key to ' ; and
ii. Exactly (p + 3)/2 of the p -1 possible random choices for a give the key to x_.
[Hint : Consider the function
f: 11 ,2, . . . ( , p ,p- }{2,3,...,p-21
defined by the equation (a - NFx_ ) f (a) = a +L (mod p ). Prove that this func-
tion is one-to-one and that f (a) is a quadratic residue modulo p if and only if a
does not give the key to '.] 0
This shows that the Las Vegas algorithm succeeds with probability somewhat
greater than one half, so that on the average it suffices to call it twice to obtain a
square root of x. In view of the high proportion of integers that give a key to ', it is
curious that no known efficient deterministic algorithm is capable of finding even one
of them with certainty.
Problem 8.5.14. The previous problem suggests a modification to the algo-

rithm rootLV : only carry out the symbolic calculation of (a +' )(p- 1)12 if
(a2-x) mod p is a quadratic nonresidue. This allows us to detect a failure more
rapidly, but it takes longer in the case of a success. Give the modified algorithm expli-
citly. Is it to be preferred to the original algorithm ? Justify your answer.
* Problem 8.5.15. The following algorithm increases the probability of success

if p 1 (mod 8) .
procedure rootLV 2(x, p, var y, var success)

I assume that p is a prime and p = 1 (mod 4) }
a F-uniform(1..p-1)
if a 2 = -x (mod p) I very unlikely and unfortunate }
then success - false
else let odd t and k >- 2 be such that p = 2k t + 1
compute c and d such that 0 <c 5p-1,0<-d <p-1
and (a + )t =-c + d (mod p)
ifc =0ord =0
then success - false
else while c 2 d 2x (mod p) do
b F-(c2-d2x)modp
d E-- 2cd mod p
c -h
compute y such that 1 5 y 5 p- 1 and yd = 1 (mod p)
y - cy mod p
success F- true
Prove that the probability of failure of this algorithm is exactly (1/2)k-1 and that the
while loop is executed at most k - 2 times, where k is specified in the algorithm.
Problem 8.5.16. An even more elementary problem for which no efficient

deterministic algorithm is known is to find a quadratic nonresidue modulo p where
p = 1 (mod 4) is a prime.
i. Give an efficient Las Vegas algorithm to solve this problem.

ii. Show that the problem is not more difficult than the problem of finding an
efficient deterministic algorithm to calculate a square root. To do this, suppose
that there exists a deterministic algorithm root 2(x , p) that is able to calculate
efficiently ' mod p, where p =- 1 (mod 4) is prime and x is a quadratic residue
modulo p. Show that it suffices to call this algorithm less than L19 P I times to
be certain of obtaining, in a way that is both deterministic and efficient, a qua-
dratic nonresidue modulo p. (Hint : Let k be the largest integer such that 2k
divides p -1 exactly. Consider the sequence x i = p - 1, xi = x7 mod p for
2<- i <- k. Prove that xi is a quadratic residue modulo p for 1 <- i < k , but that
xk is not.)
** Problem 8.5.17. The converse of Problem 8.5.16. Give an efficient deter-

ministic algorithm rootDET (x , p , z) to calculate a square root of x modulo p, provided
p is an odd prime, x is a quadratic residue modulo p, and z is an arbitrary quadratic
nonresidue modulo p.
The two preceding problems show the computational equivalence between the
efficient deterministic calculation of square roots modulo p and the efficient deter-
ministic discovery of a quadratic nonresidue modulo p. This is an example of the
technique called reduction, which we study in Chapter 10.
8.5.3 Factorizing Integers
Let n be an integer greater than 1. The problem of factorizing n consists of finding the
unique decomposition n = p 1 ' p 2 pk.
such that m I , M2, ... , Mk are positive

integers and p I < P 2 < < p k are prime numbers. If n is composite, a nontrivial
factor is an integer x, 1 < x < n , that divides n exactly. Given composite n, the
problem of splitting n consists of finding some nontrivial factor of n.
Problem 8.5.18. Suppose you have available an algorithm prime (n), which
tests whether or not n is prime, and an algorithm split(n), which finds a nontrivial
factor of n provided n is composite. Using these two algorithms as primitives, give an
algorithm to factorize any integer.
Section 8.6.2 concerns an efficient Monte Carlo algorithm for determining pri-
mality. Thus the preceding problem shows that the problem of factorization reduces to
the problem of splitting. Here is the naive algorithm for the latter problem.
function split (n)

{ finds the smallest nontrivial factor of n if n is composite
or returns 1 if n is prime }
for i F- 2 to LJ j do
if (n mod i) = 0 then return i
return I
Problem 8.5.19. Why is it sufficient to loop no further than ' ?
The preceding algorithm takes a time in Q(J) in the worst case to split n. It is
therefore of no practical use even on medium-sized integers : it could take more than 3
million years in the worst case to split a number with forty or so decimal digits,
counting just microsecond for each trip round the loop. No known algorithm,
1
whether deterministic or probabilistic, can split n in a time in 0 (p (m)) in the worst

case, where p is a polynomial and m = [log(l +n)l is the size of n. Notice that
Z 10"'12, which is not a polynomial in m. Dixon's probabilistic algorithm is
nevertheless able to split n in a time in 0 (20( m '09 - ))
Problem 8.5.20. Prove that 0 (mk ) c O (2014"' kg n )) C O whatever

the values of the positive constants k and b.
The notion of quadratic residue modulo a prime number (Section 8.5.2) general-
izes to composite numbers. Let n be any positive integer. An integer x,
1 5 x S n -1, is a quadratic residue modulo n if it is relatively prime to n (they have
no nontrivial common factor) and if there exists an integer y, 1 <- y < n - 1, such that
x = y 2 (mod n). Such a y is a square root of x modulo n. We saw that a quadratic
residue modulo p has exactly two distinct square roots modulo p when p is prime.
This is no longer true modulo n if n has at least two distinct odd prime factors. For
instance, 82 = 132 = 222 = 272 = 29 (mod 35).
* Problem 8.5.21. Prove that if n = pq, where p and q are distinct odd primes,
then each quadratic residue modulo n has exactly four square roots. Prove further that
exactly one quarter of the integers x that are relatively prime to n and such that
I <- x <- n - 1 are quadratic residues modulo n.
Section 8.5.2 gave efficient algorithms for testing whether x is a quadratic residue
modulo p, and if so for finding its square roots. These two problems can also be
solved efficiently modulo a composite number n provided the factorization of n is
given. If the factorization of n is not given, no efficient algorithm is known for either
of these problems. The essential step in Dixon's factorization algorithm is to find two
integers a and h relatively prime to n such that a2 = b2 (mod n) but a ±b (mod n).
This implies that a2 - h2 = (a - b)(a + b) = 0 (mod n). Given that n is a divisor nei-
ther of a +h nor of a -b, it follows that some nontrivial factor x of n must divide
a + b while nix divides a -b. The greatest common divisor of n and a + b is thus a
nontrivial factor of n. In the previous example, a = 8, b = 13, and n = 35, and the
greatest common divisor of a +b = 21 and n = 35 is x = 7, a nontrivial factor of 35.
Here is an outline of Dixon's algorithm.
procedure Dixon (n, var x, var success)
{ tries to find some nontrivial factor x of composite number n }
if n is even then x <- 2, success - true
else for i - 2 to Llog3n J do
if n 11' is an integer then x <- n
success - true
return
{ since n is assumed composite, we now know that it has
at least two distinct odd prime factors }
a, b <- two integers such that a 2 = b 2 (mod n)
if a °±b (mod n) then success <- false
else x <- gcd(a+b, n) { using Euclid's algorithm }
success F- true
So how we can find a and b such that a2 = b2 (mod n)? Let k be an integer to
be specified later. An integer is k -smooth if all its prime factors are among the first k
prime numbers. For instance, 120 = 23 x 3 x 5 is 3-smooth, but 35 = 5 x 7 is not.
When k is small, k-smooth integers can be factorized efficiently by an adaptation of
the naive algorithm split (n) given earlier. In its first phase, Dixon's algorithm chooses
integers x randomly between 1 and n - 1. A nontrivial factor of n is already found if
by a lucky fluke x is not relatively prime to n . Otherwise, let y = x 2 mod n . If y is
k-smooth, both x and the factorization of y are kept in a table. The process is repeated
until we have k + 1 different integers for which we know the factorization of their
squares modulo n.
Example 8.5.2. Let n = 2,537 and k = 7. We are thus concerned only with
the primes 2, 3, 5, 7, 11, 13, and 17. A first integer x = 1,769 is chosen randomly.
We calculate its square modulo n : y = 1,240. An attempt to factorize
1,240 = 23 x 5 x 31 fails since 31 is not divisible by any of the admissible primes.
A second attempt with x = 2,455 is more successful : its square modulo n is
y = 1,650 = 2 x 3 x 52 x 11. Continuing thus, we obtain
x,= 2 , 455 y I= 1 ,650 = 2 x 3 x 52 x 11
x2= 970 y2=2,210 =2x5x 13x 17
X3=1,105 y3= 728 =23x7x13
X4 = 1,458 y4 = 2,295 = 33 x 5 x 17
x5= 216 y5= 990 =2x32x5x11
X6 = 80 y6= 1,326 = 2 x 3 x 13 x 17
X7 = 1,844 y7= 756 =22x33x7
xs = 433 ys= 2,288 = 24 x 11 x 13
Problem 8.5.22. Given that there are 512 integers x between 1 and 2,536 such
that x2 mod 2537 is 7-smooth, what is the average number of trials needed to obtain
eight successes like those in Example 8.5.2 ?
The second phase of Dixon's algorithm finds a nonempty subset of the k + I

equations such that the product of the corresponding factorizations includes each of the
k admissible prime numbers to an even power (including zero).
Example 8.5.3. There are seven possible ways of doing this in Example 8.5.2,
including
y1y2y4y8=26x34x54x70 x112x132x172
Y1Y3Y4Y5Y6Y7=28 x310x54x72x 112x 132x 172.
Problem 8.5.23. Find the other five possibilities.
Problem 8.5.24. Why is there always at least one solution ? Give an efficient
algorithm for finding one. [Hint : Form a (k + 1) x k binary matrix containing the pari-
ties of the exponents. The rows of this matrix cannot be independent (in arithmetic
modulo 2) because there are more rows than columns. In Example 8.5.3, the first
dependence corresponds to
(1,1,0,0,1,0,0) + (1,0,1,0,0,1,1) + (0,1,1,0,0,0,1) + (0,0,0,0,1,1,0)
(0,0,0,0,0,0,0) (mod 2) .
Use Gauss-Jordan elimination to find a linear dependence between the rows.]
This gives us two integers a and b such that a2 = b2 (mod n). The integer a is
obtained by multiplying the appropriate xi and the integer b by halving the powers of
the primes in the product of the y; If a 4t±b (mod n), it only remains to calculate the
.
greatest common divisor of a + b and n to obtain a nontrivial factor. This occurs with
probability at least one half.
Example 8.5.4. The first solution of Example 8.5.3 gives

a = x1x2x4x8 mod n = 2,455 x 970 x 1,458 x 433 mod 2,537 = 1,127 and
b = 23 x 32 x 52 x 11 x 13 x 17 mod 2,537 = 2,012 *±a (mod n).
The greatest common divisor of a +b = 3,139 and n = 2,537 is 43, a nontrivial factor
of n. On the other hand, the second solution gives
a =xlx3x4x5x6x7 mod n = 564 and
b=24x35x52x7x11x13x17mod 2,537=1,973=-a (mod n),
which does not reveal a factor.
Problem 8.5.25. Why is there at least one chance in two that

a t ± b (mod n)? In the case when a b (mod n), can we do better than simply
starting all over again ?
It remains to be seen how we choose the best value for the parameter k. The
larger this parameter, the higher the probability that x 2 mod n will be k -smooth when
x is chosen randomly. On the other hand, the smaller this parameter, the faster we can
carry out a test of k -smoothness and factorize the k -smooth values y; , and the fewer
such values we need to be sure of having a linear dependence. Set L = e I and
let h E IR+. It can be shown that if k = L b , there are about L 112h failures for every suc-
cess when we try to factorize x2 mod n. Since each unsuccessful attempt requires k
divisions and since it takes k + 1 successes to end the first phase, the latter takes an
average time that is approximately in 0 (L 2h +1/2h ), which is minimized by b = z . The
second phase takes a time in O (k 3) = 0(L31) by Problem 8.5.24 (it is possible to do
better than this), which is negligible when compared to the first phase. The third phase
can also be neglected. Thus, if we take k = E , Dixon's algorithm splits n
with probability at least one half in an approximate expected time in 0(L2)
= O (e 2 I ) and in a space in O (L).
Several improvements make the algorithm more practical. For example, the pro-
bability that y will be k-smooth is improved if x is chosen near [I 1, rather than
being chosen randomly between I and n - 1. A generalization of this approach,
known as the continued fraction algorithm, has been used successfully. Unlike
Dixon's algorithm, however, its rigorous analysis is unknown. It is therefore more
properly called a heuristic. Another heuristic, the quadratic sieve, operates in a time in
0 (L 9is) and space in 0 (LD). In practice, we would never implement Dixon's algo-
rithm because the heuristics perform so much better. More recently, H. W. Lenstra Jr.
has proposed a factorization algorithm based on the theory of elliptic curves.
Problem 8.5.26. Let n = 1040 and L = e 1° n In Inn . Compare L 9/8, L 2, and

microseconds. Repeat the problem with n = 10s0p
8.5.4 Choosing a Leader
A number of identical synchronous processors are linked into a network in the shape of
a ring, as in Figure 8.5.1. Each processor can communicate directly with its two
immediate neighbours. Each processor starts in the same state with the same program
and the same data in its memory. Such a network is of little use so long as all the pro-
cessors do exactly the same thing at exactly the same time, for in this case a single
processor would suffice (unless such duplication is intended to catch erratic behaviour
from faulty processors in a sensitive real-time application). We seek a protocol that
allows the network to choose a leader in such a way that all the processors are in
agreement on its identity. The processor that is elected leader can thereafter break the
symmetry of the network in whatever way it pleases by giving different tasks to dif-
ferent processors.
No deterministic algorithm can solve this problem, no matter how much time is
available. Whatever happens, the processors continue to do the same thing at the same
time. If one of them decides that it wants to be the leader, for example, then so do all
Figure 8.5.1. A ring of identical processors.
the others simultaneously. We can compare the situation to the deadlock that arises
when two people whose degree of courtesy is exactly equal try to pass simultaneously
through the same narrow door. However, if each processor knows in advance how
many others are connected in the network, there exists a Las Vegas algorithm that is
able to solve this problem in linear expected time. The symmetry can be broken on the
condition that the random generators used by the processors are independent. If the
generators are in fact pseudorandom and not genuinely random, and if each processor
starts from the same seed, then the technique will not work.
Suppose there are n processors in the ring. During phase zero, each processor
initializes its local variable m to the known value n, and its Boolean indicator active to
the value true. During phase k, k > 0, each active processor chooses an integer ran-
domly, uniformly, and independently between 1 and m. Those processors that chose 1
inform the others by sending a one-bit message round the ring. (The inactive proces-
sors continue, nevertheless, to pass on messages.) After n -1 clock pulses, each pro-
cessor knows the number I of processors that chose 1. There are three possibilities. If
/ = 0, phase k produces no change in the situation. If l > 1, only those processors that
chose 1 remain active, and they set their local variable m to the value 1. In either case
phase k + I is now begun. The protocol ends when l = 1 with the election of the
single processor that just chose 1.
This protocol is classed as a Las Vegas algorithm, despite the fact that it never
ends by admitting a failure, because there is no upper bound on the time required for it
to succeed. However, it never gives an incorrect solution: it can neither end after
electing more than one leader nor after electing none.
Let 1(n) be the expected number of phases needed to choose a leader among n
processors using this algorithm (not counting phase zero, the initialization). Let
p (n , j) _ [ ] n-i (1- lln)" -i be the probability that j processors out of n randomly
choose the value I during the first stage. With probability p (n, 1) only a single phase
is needed ; with probability p (n , 0) we have to start all over again ; and with
probability p (n , j ), 2 <_ j <_ n, l (j) subsequent phases will be necessary on the

average. Consequently,
n
l(n)=1+p(n,0)l(n)+J:p(n,j)l(j), n>_2.
j=2
A little manipulation gives
n-I
I(n)=(I+ F,p(n,j)l(j))/(1-p(n,0)-p(n,n))
i=2
* Problem 8.5.27. Show that l (n) < e = 2.718 for every n >_ 2.
* Problem 8.5.28. Show that lim I (n) < 2.442. 11

n-)oo
One phase of this protocol consists of single bit messages passed round the ring.
Since each phase thus takes linear time, the preceding problems show that the choice
of a leader also takes expected linear time.
** Problem 8.5.29. Prove that no protocol (not even a Las Vegas protocol) can
solve the problem of choosing a leader in a ring of n identical processors if they are
not given the value n . Nevertheless, given an arbitrary parameter p > 0, give a Monte
Carlo algorithm that is able to determine n exactly with a probability of error less than
p whatever the number of processors.
8.6 MONTE CARLO ALGORITHMS
There exist problems for which no efficient algorithm is known that is able to obtain a
correct solution every time, whether it be deterministic or probabilistic. A Monte
Carlo algorithm occasionally makes a mistake, but it finds a correct solution with high
probability whatever the instance considered. This is not the same as saying that it
works correctly on a majority of instances, only failing now and again in some special
cases. No warning is usually given when the algorithm makes a mistake.
Problem 8.6.1. Prove that the following algorithm decides correctly whether
or not an integer is prime in more than 80% of the cases. (It performs much better on
small integers.) Show, on the other hand, that it is not a Monte Carlo algorithm by
exhibiting an instance on which the algorithm systematically gives a wrong answer.
function wrong (n)
if gcd(n , 30030) = 1 ( using Euclid's algorithm }
then return true
else return false
Sec. 8.6 Monte Carlo Algorithms 263
What constant should be used instead of 30,030 to bring the proportion of successes
above 85% even if very large integers are considered?
Let p be a real number such that 1/2 < p < 1. A Monte Carlo algorithm is
p-correct if it returns a correct solution with probability not less than p, whatever the
instance considered. The advantage of such an algorithm is p -'/2. The algorithm is
consistent if it never gives two different correct solutions to the same instance. Some
Monte Carlo algorithms take as a parameter not only the instance to be solved but also
an upper bound on the probability of error that is acceptable. The time taken by such
algorithms is then expressed as a function of both the size of the instance and the
reciprocal of the acceptable error probability. To increase the probability of success of
a consistent, p-correct algorithm, we need only call it several times and choose the
most frequent answer. This increases our confidence in the result in a way similar to
the world series calculation of Section 5.2.
Problem 8.6.2. Let MC(x) be a consistent, 75%-correct Monte Carlo algo-

rithm. Consider the following algorithm.
function MC 3(x)
t MC (x) ; u - MC (x) ; v - MC (x)
if t = u or t = v then return t
return v
Prove that this algorithm is consistent and 27/32-correct, hence 84%-correct. Show on
the other hand that it might not even be 71%-correct if MC, though 75%-correct, were
not consistent.
More generally, let a and S be two positive real numbers such that e + S < -1. Let
MC (x) be a Monte Carlo algorithm that is consistent and (z + e)-correct. Let

c, = -2 / Ig (1- 4c2 ). Let x be some instance to be solved. It suffices to call MC (x) at
least [cElg 1/81 times and to return the most frequent answer (ties can be broken arbi-
trarily) to obtain an algorithm that is consistent and (1- S)-correct. This allows us to
amplify the advantage of the algorithm, however small it may be, so as to obtain a new
algorithm whose error probability is as small as we choose.
To prove the preceding claim, let n >- cE lg 1/8 be the number of times
that the ('+F-)-correct algorithm is called. Let m = Ln / 2j+ 1, p = and z+e,
q = I -p = z -e. The repetitive algorithm finds the correct answer if the latter is
obtained at least m times. Its error probability is therefore at most
m-1
E Prob[ i correct answers in n tries ]
i=0
M-1
< n []piqnl
1: i
i=0
-i
=(Pq)ni2
I,
i=0
n
(qlp)
m-1
)n12
(pq i since q /p < 1 and 2- i >- 0
i=0
(pq)n/2 ` (nl (Pq)ni22n

=
_ (4pq )n i2 = (1-4F
2)n i2
12) lg 1/8
(1- 4E2)( since 0 < 1- 4E2 < 1
= 2-1g 1/6 =6 since aliisa = 2 for every a > 0.

The probability of success of the repetitive algorithm is therefore at least 1- S.
Problem 8.6.3. Prove that cE < (In 2)/2E2.
For example, suppose we have a consistent Monte Carlo algorithm whose advan-
tage is 5% and we wish to obtain an algorithm whose probability of error is less than
5% (that is, we wish to go from a 55%-correct algorithm to a 95%-correct algorithm).
The preceding theorem tells us that this can be achieved by calling the original algo-
rithm about 600 times on the given instance. A more precise calculation shows that it
is enough to repeat the original algorithm 269 times, and that repeating it 600 times
yields an algorithm that is better that 99%-correct. [This is because some of the ine-
qualities used in the proof are rather crude. A more complicated argument shows that
if a consistent (Z + E)-correct Monte Carlo algorithm is repeated 2m -1 times, the
resulting algorithm is (1-S)-correct, where
m-1
S= 2-E v 2i (4 -£2)i <
1
(1-4E2)m
i
i =0 4E nm
The first part of this formula can be used efficiently to find the exact number of repeti-
tions required to reduce the probability of error below any desired threshold S. Alterna-
tively, a good upper bound on this number of repetitions is quickly obtained from the
second part: find x such that e'x - 1/(2SI) and then set m = Fx/4E21.]
Repeating an algorithm several hundred times to obtain a reasonably small pro-
bability of error is not attractive. Fortunately, most Monte Carlo algorithms that occur
in practice are such that we can increase our confidence in the result obtained much
more rapidly. Assume for simplicity that we are dealing with a decision problem and
that the original Monte Carlo algorithm is biased in the sense that it is always correct
whenever it returns the answer true, errors being possible only when it returns the
answer false. If we repeat such an algorithm several times to increase our confidence
in the final result, it would be silly to return the most frequent answer : a single true
outweighs any number of falses. As we shall see shortly, it suffices to repeat such an
algorithm 4 times to improve it from 55%-correct to 95%-correct, or 6 times to obtain

a 99%-correct algorithm. Moreover, the restriction that the original algorithm be
p-correct for some p > i no longer applies : arbitrarily high confidence can be
obtained by repeating a biased p-correct algorithm enough times, even if p < z
(as long as p > 0).
More formally, let us return to an arbitrary problem (not necessarily a decision
problem), and let yo be some distinguished answer. A Monte Carlo algorithm is
yo-biased if there exists a subset X of the instances such that
i. the solution returned by the algorithm is always correct whenever the instance to
be solved is not in X, and
ii. the correct solution to all the instances that belong to X is yo, but the algorithm
may not always return the correct solution to these instances.
Although the distinguished answer yo is known explicitly, it is not required that an

efficient test be known for membership in X. The following paragraph shows that this
definition is tuned precisely to make sure that the algorithm is always correct when it
answers yo.
Let MC be a Monte Carlo algorithm that is consistent, yo-biased and p-correct.
Let x be an instance, and let y be the solution returned by MC (x). What can we con-
clude if y =yo?
If x e X, the algorithm always returns the correct solution, so yo is indeed
correct ; and
if x EX, the correct solution is necessarily yo.
In both cases, we may conclude that yo is a correct solution. What happens if on the
other handy # y o ?
If x e X, y is indeed correct ; and
if x E X, the algorithm has made a mistake since the correct solution is y o ; the
probability of such an error occurring is not greater than 1-p given that the
algorithm is p-correct.
Now suppose that we call MC (x) k times and that the answers obtained are y, ,
Yk-
If there exists an i such that y; = yo, the preceding argument shows that this is
indeed the correct solution ;
if there exist i # j such that y, :t- yj , the only possible explanation is that x EX
(because the algorithm is consistent by assumption), and therefore the correct
solution is yo; and
if y; = y # yo for all the i, it is still possible that the correct solution is yo and
that the algorithm has made a mistake k times in succession on x E X, but the
probability of such an occurrence is at most (1- p )k .
Suppose, for example, that p = s (once again, this is not allowed for general
Monte Carlo algorithms, but it causes no problems with a biased algorithm). It suffices
to repeat the algorithm at most 20 times to be either sure that the correct solution is yo
(if either of the first two cases previously described is observed), or extremely
confident that the solution obtained on every one of the trials is correct (since other-
wise the probability of obtaining the results observed is less than one chance in a mil-
lion). In general, k repetitions of a consistent, p-correct, yo-biased algorithm yield an
algorithm that is (1 - (1-p )k )-correct and still consistent and yo-biased.
Assume your consistent, p-correct, yo-biased Monte Carlo algorithm has yielded
k times in a row the same answer y # yo on some instance x. It is important to under-
stand how to interpret such behaviour correctly. It may be tempting to conclude that
"the probability that y is an incorrect answer is at most (1-p )k ". Such a conclusion
of course makes no sense because either the correct answer is indeed y, or not. The
probability in question is therefore either 0 or 1, despite the fact that we cannot tell for
sure which it is. The correct interpretation is as follows : "I believe that y is the correct
answer and if you quiz me enough times on different instances, my proportion of errors
should not significantly exceed (1-p )k ".
The "proportion of errors" in question, however, is averaged over the entire
sequence of answers given by the algorithm, not only over those occurrences in which
the algorithm actually provides such probabilistic answers. Indeed, if you systemati-
cally quiz the algorithm with instances for which the correct solution is yo, it will
always be wrong whenever it "believes" otherwise. This last remark may appear
trivial, but it is in fact crucial if the probabilistic algorithm is used to generate with
high probability some random instance x on which the correct answer is a specific
Y #Yo
To illustrate this situation, consider a nonempty finite set I of instances. We are
interested in generating a random member of some subset S S 1. (As a practical
example, we may be interested in generating a random prime of a given length - see
Section 8.6.2.) Let MC be a false-biased, p-correct Monte Carlo algorithm to decide,
given any x E I, whether x E S. Let q =1-p . By definition of a false-biased algo-
rithm, Prob[ MC (x) = true ] = I for each instance x E S and Prob[ MC (x) = true ] <_ q
for each instance x 0 S. Consider the following algorithms.
function repeatMC (x , k)
i-0
ans E- true
while ans and i < k do
if-i+1
ans F- MC (x)
return ans
function genrand (k)

repeat
x - uniform (I)
until repeatMC (x, k)
return x
It is tempting to think that genrand(k) returns a random member of S with a pro-
bability of failure at most q k . This is wrong in general. The problem is that we must
not confuse the conditional probabilities Prob[ X I Y ] and Prob[ Y I X ], where X stands
for "repeatMC (x , k) returns true" and Y stands for "x 0 S ". It is correct that
Prob[X I Y ] <_ q k, but we are in fact interested in Prob[ Y I X ]. To calculate this pro-
bability, we need an a priori probability that a call on uniform (I) returns a member
of S.
* Problem 8.6.4. Let r denote the probability that x E S given that x is returned
by a call on uniform (I). Prove that the probability that a call on genrand (k) errone-
ously returns some x 0 S is at most
1
1r q-'
r
1+ k
In particular, if the error probability of MC on instances not in S is exactly q

(rather than at most q), then the probability that a call on genrand (k) returns some
x 0 S is about q k /r if q k<< r<< 1, about z2 if q k= r << 1, and nearly 1 if r << g k
This can be significant when the confidence one gets in the belief that x belongs to S
from running MC several times must be weighed against the a priori likelihood that a
randomly selected x does not belong to S. This situation illustrates dramatically the
difference between the error probability of genrand and that of repeatMC.
We are not aware of any unbiased Monte Carlo algorithm sufficiently simple to
feature in this introduction. Thus the section continues with some examples of biased
Monte Carlo algorithms. They all involve the solution of a decision problem, that is,
the only possible answers that they can return are true and false.
Problem 8.6.5. Let A and B be two efficient Monte Carlo algorithms for
solving the same decision problem. Algorithm A is p-correct and true-biased, whereas
algorithm B is q-correct and false-biased. Give an efficient Las Vegas algorithm
LV (x , var y, var success) to solve the same problem. What is the best value of r you
can deduce so that your Las Vegas algorithm succeeds with probability at least r on
each instance ? El
8.6.1 Majority Element in an Array
The problem is to determine whether an array T [I .. n ] has a majority element (see

Problems 4.11.5, 4.11.6, and 4.11.7). Consider the following algorithm.
function maj (T [1.. n ])
i F uniform (1 .. n)
x -T[i]
k F- 0
forj f- ltondoifT[jI=x thenk '-k +1
return (k > n / 2)
We see that maj (T) chooses an element of the array at random, and then checks
whether this element forms a majority in T. If the answer returned is true, the element
chosen is a majority element, and hence trivially there is a majority element in T. If,
on the other hand, the answer returned is false, it is nonetheless possible that T con-
tains a majority element, although in this case the element chosen randomly is in a
minority. If the array does indeed contain a majority element, and if one of its ele-
ments is chosen at random, the probability of choosing an element that is in a minority
is less than one-half, since majority elements occupy more than half the array. There-
fore, if the answer returned by maj (T) is false, we may reasonably suspect that T does
indeed have no majority element. In sum, this algorithm is true-biased and 4correct,
that is z
T has a majority element = maj (T) = true, with probability > z
T has no majority element = maj (T) = false, with certainty .
An error probability of 50% is intolerable in practice. The general technique for
biased Monte Carlo algorithms allows us to reduce this probability efficiently to any
arbitrary value. First, consider
function maj 2(T)
if maj (T) then return true
else return maj (T)
If the array does not have a majority element, each call of maj(T) is certain to return
false, hence so does maj 2(T ). If the array does have a majority element, the proba-
bility that the first call of maj (T) will return true is p > z , and in this case maj 2(T)
returns true, too. On the other hand, if the first call of maj (T) returns false, which
happens with probability 1-p , the second call of maj (T) may still with probability p
return true, in which case maj 2(T) also returns true. Summing up, the probability
that maj 2(T) will return true if the array T has a majority element is
p+(I-p)p=I-(1-p)2>3/4.
The algorithm maj2 is therefore also true-biased, but 3/4-correct. The probability of
error decreases because the successive calls of maj (T) are independent : the fact that
maj (T) has returned false on an array with a majority element does not change the
probability that it will return true on the following call on the same instance.
Problem 8.6.6. Show that the probability that k successive calls of maj(T) all
return false is less than 2-k if T contains a majority element. On the other hand, as
soon as any call returns true, we can be certain that T contains a majority element. El
The following Monte Carlo algorithm solves the problem of detecting the pres-
ence of a majority element with a probability of error less than E for every E > 0.
function majMC (T, E)
k rlg(1/E)1
for i <-- I to k do
if maj (T) then return true
return false
The algorithm takes a time in 0 (n log(1/E)), where n is the number of elements in the
array and E is the acceptable probability of error. This is interesting only as an illustra-
tion of a Monte Carlo algorithm since a linear time deterministic algorithm is known
(Problems 4.11.5 and 4.11.6).
8.6.2 Probabilistic Primality Testing
This classic Monte Carlo algorithm recalls the algorithm used to determine whether or
not an array has a majority element. The problem is to decide whether a given integer
is prime or composite. No deterministic or Las Vegas algorithm is known that can
solve this problem in a reasonable time when the number to be tested has more than a
few hundred decimal digits. (It is currently possible to establish with certainty the pri-
mality of numbers up to 213 decimal digits within approximately 10 minutes of com-
puting time on a CDC CYBER 170/750.)
A first approach to finding a probabilistic algorithm might be
function prime (n)
d - uniform (2 .. LJ J)
return ((n mod d) # 0)
If the answer returned is false, the algorithm has been lucky enough to find a non-
trivial factor of n, and we can be certain that n is composite. Unfortunately, the
answer true is returned with high probability even if n is in fact composite. Consider
for example n = 2,623 = 43 x 61. The algorithm chooses an integer randomly between
2 and 51. Thus there is only a meagre 2% probability that it will happen on d = 43
and hence return false. In 98% of calls the algorithm will inform us incorrectly that n
is prime. For larger values of n the situation gets worse. The algorithm can be
improved slightly by testing whether n and d are relatively prime, using Euclid's algo-
rithm, but it is still unsatisfactory.
To obtain an efficient Monte Carlo algorithm for the primality problem, we need
a theorem whose proof lies beyond the scope of this book. Let n be an odd integer
greater than 4, and let s and t be positive integers such that n -1 = 2s t, where t is
odd. Let a be an integer such that 2 5 a <- n - 2. We say that n is a strong pseudo-
prime to the base a if a' _ 1 (mod n) or if there exists an integer i such that 0 <- i < s
and art _ -1 (mod n).
If n is prime, it is a strong pseudoprime to any base. There exist however com-
posite numbers that are strong pseudoprimes to some bases. Such a base is then
a false witness of primality for this composite number. For example, 158 is a false
witness of primality for 289 because 288 = 9 x 25, 1589 = 131 (mod 289),
1582"9 = 1312 = 110 (mod 289), 1584"9 = 1102 = 251 (mod 289), and finally,
1588 9 = 2512 =_ I (mod 289).
The theorem assures us that if n is composite, it cannot be a strong pseudoprime
to more than (n - 9)/4 different bases. The situation is even better if n is composed of
a large number r of distinct prime factors : in this case it cannot be a strong pseu-
doprime to more than 4(n)/2r-' - 2 different bases, where 4(n) < n -1 is Euler's
totient function. The theorem is generally pessimistic. For instance, 289 has only 14
false witnesses of primality, whereas 737 does not even have one.
** Problem 8.6.7. Prove this theorem.
Problem 8.6.8. Give an efficient algorithm for testing whether n is a strong

pseudoprime to the base a. Your algorithm should not take significantly longer (and
sometimes even be faster) than simply calculating a (" - 1)/2 mod n with dexpoiter from
Section 4.8.
This suggests the following algorithm.

function Rabin (n)
{ this algorithm is only called if n > 4 is odd }
a <--uniform (2.. n -2)
if n is strongly pseudoprime to the base a
then return true
else return false
For an odd integer n > 4 the theorem ensures that n is composite if Rabin (n) returns
false. This certainty is not accompanied however by any indication of what are the
nontrivial factors of n. Contrariwise, we may begin to suspect that n is prime if
Rabin (n) returns true. This Monte Carlo algorithm for deciding primality is false-
biased and 3/4-correct.
This test of primality has several points in common with the Las Vegas algo-
rithm for finding square roots modulo p (Section 8.5.2). In both cases, an integer a is
chosen at random. If n is composite, there is at least a 75% chance that n will not be a
strong pseudoprime to the base a, in which case we obtain the correct answer with cer-
tainty. Similarly there is better than a 50% chance that a will provide a key for finding
. Nevertheless the algorithm for testing primality is only a Monte Carlo algorithm
whereas the one for finding square roots is Las Vegas. This difference is explained by
the fact that the Las Vegas algorithm is able to detect when it has been unlucky : the
fact that a does not provide a key for is easy to test. On the other hand, if n is a
strong pseudoprime to the base a, this can be due either to the fact that n is indeed
prime or to the fact that a is a false witness of primality for the composite number n.
The difference can also be explained using Problem 8.2.2.
As usual, the probability of error can be made arbitrarily small by repeating the
algorithm. A philosophical remark is once again in order: the algorithm does not reply
"this number is prime with probability 1-E", but rather, "I believe this number to be
prime ; otherwise I have observed a natural phenomenon whose probability of
occurrence was not greater than E". The first reply would be nonsense, since every
integer larger than I is either prime or composite.
* Problem 8.6.9. Consider the following nonterminating program.

program print primes
print 2,3
n 5
repeat
if repeatRahin (n, Llg n j) then print n
n -n + 2
ad museum
Clearly, every prime number will eventually be printed by this program. One might
also expect composite numbers to be produced erroneously once in a while. Prove that
this is unlikely to happen. More precisely, prove that the probability is better than
99% that not even a single composite number larger than 100 will ever be produced,
regardless of how long the program is allowed to run. (Note : this figure of 99% is
very conservative as it would still hold even if Rabin (n) had a flat 25% chance of
failure on each composite integer.)
** Problem 8.6.10. Find a true-biased Monte Carlo algorithm for primality

testing whose running time is polynomial in the logarithm of the integer being tester.
Notice that this combines with the false-biased algorithm described previously to yield
a Las Vegas algorithm (by Problem 8.6.5).
8.6.3 A Probabilistic Test for Set Equality
We have a universe U of N elements, and a collection of n sets, not necessarily dis-

joint, all of which are empty at the outset. We suppose that N is quite large while n is
small. The basic operation consists of adding x to the set Si , where xE U \ Si and
I S i _< n. At any given moment the question "Does Si = Si ?" may be asked.
The naive way to solve this problem is to keep the sets in arrays, lists, search
trees, or hash tables. Whatever structure is chosen, each test of equality will take a
time in S2(k), if indeed it is not in S2(k log k ), where k is the cardinality of the larger of
the two sets concerned.
For any e > 0 fixed in advance, there exists a Monte Carlo algorithm that is able
to handle a sequence of m questions in an average total time in 0 (m). The algorithm
never makes an error when Si = S/ ; in the opposite case its probability of error does
not exceed E. This algorithm provides an interesting application of universal hashing
(Section 8.4.4).
Let e > 0 be the error probability that can be tolerated for each request to test the
equality of two sets. Let k = Ilg(max(m, 1/E))]. Let H be a universal2 class of func-
tions from U into (0,1 }k , the set of k -bit strings. The Monte Carlo algorithm first
chooses a function at random in this class and then initializes a hash table that has U
for its domain. The table is used to implement a random function rand : U -+ (0,1 }k
as follows.
function rand (x)
if x is in the table then return its associated value
y E- some random k -bit string
add x to the table and associate y to it
return y
Notice that this is a memory function in the sense of Section 5.7. Each call of rand (x)
returns a random string chosen with equal probability among all the strings of length k.
Two different calls with the same argument return the same value, and two calls with
different arguments are independent. Thanks to the use of universal hashing, each call
of rand (x) takes constant expected time.
To each set Si we associate a variable v [i] initialized to the binary string com-
posed of k zeros. Here is the algorithm for adding an element x to the set Si . We sup-
pose that x is not already a member of Si .
procedure add (i , x)
v[i] F- v[i]®rand(x)
The notation t ® u stands for the bit-by-bit exclusive-or of the binary strings t and u.
The algorithm to test the equality of Si and S1 is:
function test (i , j )
ifv[i]=v[jI
then return true
else return false
It is obvious that Si # Sj if v [i] # v [ j ]. What is the probability
that v [i ] = v [ j ] when Si # Sj ? Suppose without loss of generality that there
exists an xo a Si such that xa 9 S/ . Let Si = Si \ (x0}. For a set S c U, let
XOR (S) be the exclusive-or of the rand (x) for every x E S . By definition,
v [i] = XOR (Si) = rand (xo) ®XOR (S;) and v[j] = XOR (SS ). Let yo = XOR (S; )
®XOR (Si ). The fact that v [i ] = v [ j ] implies that rand (xo) = yo; the probability of
this happening is only 2-k since the value of rand (xo) is chosen independently of those
values that contribute to yo. Notice the similarity to the use of signatures in Sec-
tion 7.2.1.
This Monte Carlo algorithm differs from those in the two previous sections in
that our confidence in an answer "Si = Si " cannot be increased by repeating the call of
test (i , j ). It is only possible to increase our confidence in the set of answers obtained
to a sequence of requests by repeating the application of the algorithm to the entire
sequence. Moreover, the different tests of equality are not independent. For instance,
if Si # Sj , x 9 Si u SS , Sk = Si u {x }, Si = Sj u {x }, and if an application of the
algorithm replies incorrectly that Si = Sj , then it will also reply incorrectly that
Sk=S1.
Problem 8.6.11. What happens with this algorithm if by mistake a call of

add (i, x) is made when x is already in Si ?
Problem 8.6.12. Show how you could also implement a procedure elim (i, x),
which removes the element x from the set Si . A call of elim (i, x) is only permitted
when x is already in Si .
Problem 8.6.13. Modify the algorithm so that it will work correctly (with pro-
bability of error c) even if a call of add (i, x) is made when xE Si . Also implement a
request member (i , x), which decides without ever making an error whether x E Si .
A sequence of m requests must still be handled in an expected time in O (m).
** Problem 8.6.14. Universal hashing allows us to implement a random function

rand : U -*{O,1}". The possibility that rand (x 1) = rand (X2) even though x I * X2,
which does not worry us when we want to test set equality, may be troublesome for
other applications. Show how to implement a random permutation. More precisely,
let N be an integer and let U = { 1, 2, ... , N 1. You must accept two kinds of request :
init and p (i) for 1 <_ i <_ N. A call of init initializes a new permutation tt : U -* U. A
call of p (i) returns the value 7t(i) for the current permutation. Two calls p (i) and p Q)
that are not separated by a call of init should therefore return two different answers if
and only if i # j. Two such calls separated by a call of init should on the other hand
be independent. Suppose that a call of uniform (u .. v) takes constant time for
1 <_ u <_ v <_ N. Your implementation should satisfy each request in constant time in
the worst case, whatever happens. You may use memory space in O (N), but no
request may consult or modify more than a constant number of memory locations : thus
it is not possible to create the whole permutation when init is called. (Hint: reread
Problem 5.7.2 and Example 8.4.1).
8.6.4 Matrix Multiplication Revisited
You have three n x n matrices A, B, and C and you would like to decide whether
AB = C. Here is an intriguing false-biased, correct Monte Carlo algorithm that is
i
capable of solving this problem in a time in d (n2). Compare this with the fastest
known deterministic algorithm to compute the product AB (Section 4.9), which takes a
time in 0 (n 2.376), and with the probabilistic algorithm mentioned in Section 8.3.5,
which only computes the product approximately.
function goodproduct (A, B , C, n)
array X [ 1 .. n ] { to be considered as a column vector)
for i F- 1 ton do X [i ] <- uniform ({ -1,11)
if ABX = CX then return true
else return false
In order to take a time in 0 (n 2), we must compute ABX as A times BX, providing a
dramatic example of the topic discussed in Section 5.3.
Problem 8.6.15. It is obvious that goodproduct (A , B , C , n) returns true

whenever AB = C. Prove that it returns false with probability at least z whenever
AB # C. (Hint: consider the columns of the matrix AB -C and show that at least
half the ways to add and subtract them yield a nonzero column vector, provided
AB #C.) 11
Problem 8.6.16. Given two n x n matrices A and B, adapt this algorithm to

decide probabilistically whether B is the inverse of A.
Problem 8.6.17. Given three polynomials p (x), q (x), and r (x) of degrees
n, n and 2n, respectively, give a false-biased, correct Monte Carlo algorithm to
decide whether r (x) is the symbolic product of p z(x) and q (x). Your algorithm should
run in a time in 0 (n ). (In the next chapter we shall see a deterministic algorithm that
is capable of computing the symbolic product of two polynomials of degree n in a time
in 0 (n log n ), but no such algorithm is known that only takes a time in 0 (n).)
The experiment devised by Leclerc (1777) was carried out several times in the
nineteenth century; see for instance Hall (1873). It is no doubt the earliest recorded
probabilistic algorithm. The term "Monte Carlo" was introduced into the literature by
Metropolis and Ulam (1949), but it was already in use in the secret world of atomic
research during World War 11, in particular in Los Alamos, New Mexico. Recall that it
is often used to describe any probabilistic algorithm. The term "Las Vegas" was intro-
duced by Babai' (1979) to distinguish probabilistic algorithms that occasionally make a
mistake from those that reply correctly if they reply at all. The term "Sherwood" is
our own. For the solution to Problem 8.2.3, see Anon. (1495).
Two encyclopaedic sources of techniques for generating pseudorandom numbers

are Knuth (1969) and Devroye (1986). The former includes tests for trying to distin-
guish a pseudorandom sequence from one that is truly random. A more interesting
generator from a cryptographic point of view is given by Blum and Micali (1984); this
article and the one by Yao (1982) introduce the notion of an unpredictable generator,
which can pass any statistical test that can be carried out in polynomial time. Under
the assumption that it is infeasible to factorize large numbers, a more efficient
unpredictable pseudorandom generator is proposed in Blum, Blum, and Shub (1986).
More references on this subject can be found in Brassard (1988). General techniques
are given in Vazirani (1986, 1987) to cope with generators that are only "semi-
random".
For more information on numeric probabilistic algorithms, consult Sobol' (1974).
A guide to simulation is provided by Bratley, Fox, and Schrage (1983). The point is
made in Fox (1986) that pure Monte Carlo methods are not specially good for numer-
ical integration with a fixed dimension : it is preferable to choose your points systemat-
ically so that they are well spaced, a technique known as quasi Monte Carlo.
Problem 8.3.14 is solved in Klamkin and Newman (1967). The application of proba-
bilistic counting to the Data Encryption Standard (Example 8.3.1) is described in Kal-
iski, Rivest, and Sherman (1988). Section 8.3.4 follows Flajolet and Martin (1985).
Numeric probabilistic algorithms designed to solve problems from linear algebra are
discussed in Curtiss (1956), Vickery (1956), Hammersley and Handscomb (1965), and
Carasso (1971).
An early (1970) linear expected time probabilistic median finding algorithm is
attributed to Floyd : see Exercise 5.3.3.13 in Knuth (1973). It predates the classic
worst-case linear time deterministic algorithm described in Section 4.6. A probabilistic
algorithm that is capable of finding the i th smallest among n elements in an expected
number of comparisons in n +i +O(J) is given in Rivest and Floyd (1973). Com-
putation with encrypted instances (end of Section 8.4.2) is an idea originating in
Feigenbaum (1986) and developed further in Abadi, Feigenbaum, and Kilian (1987).
The technique for searching in an ordered list and its application to sorting
(Problem 8.4.14) come from Janko (1976). An analysis of this technique
(Problem 8.4.12) is given in Bentley, Stanat, and Steele (1981); the same reference
gives the statement of Problem 8.4.4. Classic hash coding is described in Knuth
(1968); many solutions to Problem 8.4.15 appear there. Universal hashing was
invented by Carter and Wegman (1979); several universal2 classes are described there,
including the one from Problem 8.4.20. For solutions to Problem 8.4.21, see Wegman
and Carter (1981), and Bennett, Brassard, and Robert (1988).
The probabilistic approach to the eight queens problem was suggested to the
authors by Manuel Blum. The experiments on the twenty queens problem were carried
out by Pierre Beauchemin. The algorithm of Section 8.5.2 for finding square roots
modulo a prime number, including Problem 8.5.15, is due to Peralta (1986). Early
algorithms to solve this problem are given in Lehmer (1969) and Berlekamp (1970).
As a generalization, Rabin (1980a) gives an efficient probabilistic algorithm for com-
puting roots of arbitrary polynomials over any finite field. For a solution to
Problem 8.5.12, consult Aho, Hopcroft, and Ullman (1974). The solution to
Problem 8.5.17 is given by the algorithm of Shanks (1972) and Adleman, Manders,
and Miller (1977). The integer factorization algorithm of Pollard (1975) has a proba-
bilistic flavour. The probabilistic integer factorization algorithm discussed in Sec-
tion 8.5.3 originated with Dixon (1981) ; for a comparison with other methods, refer to
Pomerance (1982). The algorithm based on elliptic curves is discussed in Lenstra
(1986). For efficiency considerations in factorization algorithms, consult Montgomery
(1987). The algorithm for electing a leader in a network, including Problem 8.5.29,
comes from Itai and Rodeh (1981).
Amplification of the advantage of an unbiased Monte Carlo algorithm is used to
serve cryptographic ends in Goldwasser and Micali (1984). The probabilistic test of
primality presented here is equivalent to the one in Rabin (1976, 1980b). The test of
Solovay and Strassen (1977) was discovered independently. The expected number of
false witnesses of primality for a random composite integer is investigated in Erdos
and Pomerance (1986) ; see also Monier (1980). More information on number theory
can be found in the classic Hardy and Wright (1938). The implication of Problem 8.6.4
for the generation of random numbers that are probably prime is explained in
Beauchemin, Brassard, Cre peau, Goutier, and Pomerance (1988), which also gives a
fast probabilistic splitting algorithm whose probability of success on any given compo-
site integer is at least as large as the probability of failure of Rabin's test on the same
integer. A theoretical solution to Problem 8.6.10 is given in Goldwasser and Kilian
(1986) and Adleman and Huang (1987). For more information on tests of primality
and their implementation, consult Williams (1978), Lenstra (1982), Adleman, Pomer-
ance, and Rumely (1983), Kranakis (1986), and Cohen and Lenstra (1987). The proba-
bilistic test for set equality comes from Wegman and Carter (1981); they also give a
cryptographic application of universal hashing. The solution to Problem 8.6.14 is in
Brassard and Kannan (1988). The Monte Carlo algorithm to verify matrix multiplica-
tion (Section 8.6.4) and the solution to Problem 8.6.17 are given in Freivalds (1979);
also read Freivalds (1977).
Several interesting probabilistic algorithms have not been discussed in this
chapter. We close by mentioning a few of them. Given the cartesian coordinates of
points in the plane, Rabin (1976) gives an algorithm that is capable of finding the
closest pair in expected linear time (contrast this with Problem 4.11.14). Rabin
(1980a) gives an efficient probabilistic algorithm for factorizing polynomials over arbi-
trary finite fields, and for finding irreducible polynomials. A Monte Carlo algorithm is
given in Schwartz (1978) to decide whether a multivariate polynomial over an infinite
domain is identically zero and to test whether two such polynomials are identical.
Consult Zippel (1979) for sparse polynomial interpolation probabilistic algorithms. An
efficient probabilistic algorithm is given in Karp and Rabin (1987) to solve the string-
searching problem discussed in Section 7.2. Our favourite unbiased Monte Carlo algo-
rithm for a decision problem, which allows us to decide efficiently whether a given
integer is a perfect number and whether a pair of integers is amicable, is described in
Bach, Miller, and Shallit (1986). For an anthology of probabilistic algorithms, read
Valois (1987).
9
Transformations
of the Domain
9.1 INTRODUCTION
It is sometimes useful to reformulate a problem before trying to solve it. If you were
asked, for example, to multiply two large numbers given in Roman figures, you would
probably begin by translating them into Arabic notation. (You would thus use an
algorism, with this word's original meaning!) More generally, let D be the domain of
objects to be manipulated in order to solve a given problem. Let f :Dt -4 D be a
function to be calculated. An algebraic transformation consists of a transformed
domain R, an invertible transformation function a : D -+ R and a transformed function
g : R' -> R such that
PX1 , x 2 , ... , xt) = a' ' (g (a(x 1),---

C02), a(xt )))
for all x 1 , x22 ,---xt in the domain D . Such a transformation is of interest if g can
be calculated in the transformed domain more rapidly than f can be calculated in the
original domain, and if the transformations a and 6-1 can also be computed efficiently.
Figure 9.1.1 illustrates this principle.
Example 9.1.1. The most important transformation used before the advent of
computers resulted from the invention of logarithms by Napier in 1614. Kepler found
this discovery so useful that he dedicated his Tabulae Rudolphinae to Napier. In this
case, D = IN+ or R+, f (u , v) = u x v , R = R, a(u) = In u and g (x , y) =x +y. This
allows a multiplication to be replaced by the calculation of a logarithm, an addition,
and an exponentiation. Since the computation of a and a1 would take more time than
the original multiplication, this idea is only of interest when tables of logarithms are
278 Transformations of the Domain Chap. 9
I
Dt D
o`
Rt
Figure 9.1.1. Transformation of the domain.
computed beforehand. Such tables, calculated once and for all, thus furnish a historical
example of preconditioning (Chapter 7).
Example 9.1.2. It is often useful to transform between Cartesian and polar

coordinates.
Example 9.1.3. Most computers that handle numerical data read these data
and print the results in decimal but carry out their computations in binary.
Example 9.1.4. You want to perform a symbolic multiplication of two poly-

nomials, for example, p(x) = 3x3-5x2-x +I and q(x) =x3-4x2+6x -2. These
polynomials are represented by their coefficients. The original domain of the problem
is therefore Zd + 1, where d is the degree of the polynomials involved. The naive algo-
rithm for multiplying these polynomials resembles the classic algorithm for integer
multiplication ; it takes a time in O(d 2) if the scalar operations are taken to be elemen-
tary. An alternative way to represent the same polynomials is to give their values at
d+ 1 distinct points, f o r instance, 0, 1, 2, ... , d. The polynomials in our example
would be represented by p = (1, - 2, 3, 34) and q = ( - 2, 1, 2, 7). The transformed
domain is still Zd+I, but its meaning has changed. The new representation suffices to
define the original polynomials because there is exactly one polynomial of degree at
most 3 that passes through any four given points.
Let r(x) = 3x6-17x5+37x4-31x3+8x -2 be the product of p(x) and q(x).
Using the transformed representation, we can carry out the multiplication in a time in
O (d) since r (i) is calculated straightforwardly as p (i) q (i) for 0 5 i <- d. Thus we
obtain r = (- 2,- 2,6, 238). However, this does not allow us to recover the
coefficients of r (x) because a polynomial of degree 6 is not uniquely defined by its
value at only four points. We would have needed to use seven points from the outset,
taking p = (1, - 2, 3, 34,109, 246, 463) and q = (- 2, 1, 2, 7, 22, 53, 106), if the computa-
tion of r were to be carried out correctly using this representation.
In this example, calculating the transformed function (carrying out the pointwise
multiplication) is much quicker than straightforwardly calculating the original function
Sec. 9.2 The Discrete Fourier Transform 279
(performing a symbolic multiplication of two polynomials given by their coefficients).

However, this is only useful if the transformation function (evaluation of the polyno-
mials) and its inverse (interpolation) can be calculated rapidly. At first glance, it seems
that the computation of the values p (i) and q (i) for 0 <_ i <- 2d must take a time in
S2(d2). Worse still, a naive implementation of Lagrange's algorithm to carry out the
final interpolation would take a time in S2(d 3), which appears to remove any interest
this approach may have had. We shall see in Section 9.4 that this is not in fact the
case, provided that the points at which we evaluate the polynomials are chosen judi-
ciously.
9.2 THE DISCRETE FOURIER TRANSFORM
For the rest of this chapter all the arithmetic operations involved are carried out either
modulo some integer m to be determined or in the field of complex numbers - or
more generally in any commutative ring. They are assumed to be executed at unit cost
unless it is explicitly stated otherwise. Let n > I be a power of 2. We denote by co
some constant such that 012 =-1. If n = 8, for example, then w= 4 is a possible
value if we are using arithmetic modulo 257, and w= (1+i )/J is a possible value in
the field of complex numbers.
Consider an n-tuple a = (ao, a i , ... , an _ 1). This defines in the natural way a
polynomial pa(x)=an_1Xn-I+an_2Xn-2+ "' +alx+ao of degree less than n.
The discrete Fourier transform of a with respect to m is the n-tuple Fu,(a) =
(pa (1), pa (w), pa ((02), ... , pa (wn - 1)). As in Example 9.1.4, it appears at first glance
that the number of scalar operations needed to calculate this transform is in S2(n2).
However, this is not in fact the case, thanks to an algorithm known as the Fast Fourier
Transform (FFT). This algorithm is vitally important for a variety of applications,
particularly in the area of signal processing (Section 1.7.6).
Suppose that n > 2 and set t = n / 2. The t-tuples b and c defined by
b = (ao, a 2 , ... , an-4, an-2) and c = (al , a3, ... , an_3, an_j) are such that
pa(X)=ph(x2)+xpc(x2). In particular, Pa (w')=pb((X')+w'pC((X'), where a=w2.
Clearly, a`12 = (m2)t/2 = m' = wn/2 = -1, so it is legitimate to talk about Fa,(b) and
Fa(c ). Furthermore, a' = 1 and co' _ -1, hence a`+' = a` and w`+` d, so that
pa (w`+') = Pb ((X') - w' pC (c'). The Fourier transform Fu,(a) is calculated using the
divide-and-conquer technique.
function FFT (a [0.. n -1 ], co) : array [0.. n -1 ]
(n is a power of2andwn'2=-1 )
array A [0.. n -1] [ the answer is computed in this array )
if n = 1 then A [0] <- a [0]
else t - n / 2
arrays b, c, B, C [0.. t -1 ] ( intermediate arrays )
( creation of the sub-instances )
fori -0tot-1 dob[i] 4-a[2i]
c[i] E-a[2i+1]
{ recursive Fourier transform computation of the sub-instances }

B t-FFT(b,w2)
C -FFT(c,(o 2)
{ Fourier transform computation of the original instance }
af- I
for i F- 0 to t-1 do
(a= w' I
A[i] F-B[i]+aC[i]
A [t+i ] E- B [i] - aC [i]
a F-ato
return A
Problem 9.2.1. Show that the execution time of this algorithm is in

O(n log n).
Example 9.2.1. Let n = 8 and a = (255,8,0,226,37,240,3,0). Let us calcu-

late FW(a) in arithmetic modulo m = 257, where w = 4. This can be done because
44 =- -1 (mod 257). First a is split into b = (255,0,37,3) and c = (8, 226, 240, 0).
The recursive calls, which use t = 4 and w2 = 16, yield B = F 16(b) = (38, 170, 32, 9)
and C = F 16(c) = (217, 43, 22, 7). We combine these results to obtain A.
A [0] 38 + 217 = 255 A [4] <- 38 - 217 = 78
A[ 11 E-- 170+43co = 85 A [5] F 170-43w = 255
A [2] 32 + 22002 = 127 A [6] 32 - 22to2 = 194
A [3] F- 9+7w3 = 200 A [7] - 9-70 = 75
The final result is therefore A = (255, 85,127, 200, 78, 255, 194, 75).
** Problem 9.2.2. Despite the conceptual simplicity of this recursive

algorithm, it is preferable in practice to use an iterative version. Give such
an iterative algorithm. [Hint: Let p (x) be an arbitrary polynomial, and
let q (x) = (x-x1 )(x-x2) .. (x x1) be a polynomial of degree t such that
q(x1) = q (x 2) = = q (xr) = 0, where x 1 , X2,-- x1 are arbitrary distinct values.
Let r (x) be the remainder of the symbolic division of p (x) by q (x). Show that
p (xi) = r (x,) for 15 i <- t. In particular, the remainder of the symbolic division of
p (x) by the monomial (x-x,) is the constant polynomial whose value is p (x1).]
9.3 THE INVERSE TRANSFORM
Despite its importance in signal processing, our concern here is to use the discrete
Fourier transform as a tool for transforming the domain of a problem. Our principal
aim is to save the idea proposed in Example 9.1.4. Using the fast Fourier transform
allows us to evaluate the polynomials p (x) and q (x) at the points 1, w, cue, ... , & -1
in a time in 0 (n log n), where n is a power of 2 greater than the degree of the product
Sec. 9.3 The Inverse Transform 281
polynomial. The pointwise multiplication can be carried out in a time in O (n). To

obtain the final result, we still have to interpolate the unique polynomial r (x) of degree
less than n that passes through these points. Thus we have in fact to invert the Fourier
transformation. With this in view we say that w is a principal n th root of unity if it
fulfills three conditions :
1. w# 1
2. wn = 1, and
n-I
3. E or" = 0 for every I <- p < n .
j=0
When n is a power of 2, it turns out that these conditions are automatically

satisfied whenever con/2 = - 1, as we have assumed already.
Theorem 9.3.1. Consider any commutative ring. Let n > 1 be a power of 2,

and let w be an element of the ring such that con/2 = - 1. Then
i. w is a principal n th root of unity.

ii. wn -1 is the multiplicative inverse of co in the ring in question ; we shall denote it
by col. More generally we use W-' to denote (0-1)' for any integer i.
iii. w 1 is also a principal n th root of unity.
iv. As a slight abuse of notation, let us use " n " also to denote the element in our
ring obtained by adding n ones together (the "1" from our ring). Assuming this
n is not zero, 1 = w0 , w' , w2 , .. wn -' are all distinct - they are called the
n th roots of unity.
n-1
v. Assuming the existence of a multiplicative inverse for n in our ring,
wn 12 = -1 is a consequence of w being a principal n th root of unity.
* Problem 9.3.1. Prove Theorem 9.3.1. Hints :
i. Conditions (1) and (2) are obviously fulfilled. To show condition (3), let n = 2k
and decompose p = 2n v where u and v are integers, and v is odd. Let
s =2k-1. Show that co-" = - (dj+s )P for every integer j. Conclude by split-
ting y o' wjP into 2n sub sums of 2s elements, each summing to zero.
ii. Obvious.
iii. Notice that (w 1)n i2 = -1.
iv. Assume co' = w j for 0 <- i < j < n and let p = j - i. Then wP = 1 and
1 <- p < n - 1. Use condition (3) and n * 0 to obtain a contradiction.
v. Use p = n/2 in condition (3), and use the existence of n-1.
Problem 9.3.2. Prove that e 2i 1/, is a principal n th root of unity in the field of
complex numbers.
Problem 9.3.3. Let n and w be positive powers of 2, and let m = w"12 + 1.

Prove that w is a principal n th root of unity in arithmetic modulo m. Prove further
that n-1 exists modulo m, by showing that n-1 = m - (m -1)/n .
Problem 9.3.4. When m is of the form 2" + 1, as in Problem 9.3.3, multiplica-

tions modulo m can be carried out on a binary computer without using division opera-
tions. Let a and b be two integers such that 0 <- a < m and 0 <- b < m ; we wish to
obtain their product modulo m. Let c = ab be the ordinary product obtained from a
multiplication. Decompose c into two blocks of u bits : c = 2" j + i, where 0 <- i < 2'
and 0 <- j <- 2". (The only possibility for j = 2" is when a = b = m -1.) If i > j, set
d =i-j, and otherwise set d = m + i - j . Prove that 0 <- d < m and
ab = d (mod m). More generally, show how x mod m can be calculated in a time in
0 (u) when m = 2" + 1, provided that - (22u + 2u) < x < 22u +2u+'. p
From now on, assume that w is a principal n th root of unity and that n-1 exists
in the algebraic structure considered. Let A be the n x n matrix defined by A,j = di
for 0 <- i < n and 0 <- j < n . The main theorem asserts that A has an inverse, namely
the matrix B defined by B,j = n'o'.
Theorem 9.3.2. Let A and B be the matrices just defined. Then AB = I, , the
n x n identity matrix.
Proof. Let C = AB. By definition, Cu = Jk _o A,k Bki - n-I ik -o w(i-j)k
There are three cases to consider.
n
nxn-1=1.
i.Ifi=j, then co('-j)k=co0=1,andsoC11=n-1
k=0
n-1
H. If i > j, let p = i -j. Now C,j = n-1 Y, OP = 0 by property (3) of a principal
k=0
n th root of unity, since 1 <- p < n .
n - I
iii. If i < j, let p = j - i. Now C,j =n(o)1'P = 0 because co 1 is also a

k=0
principal n th root of unity by Theorem 9.3.1(iii).
This matrix A provides us with another equivalent definition of the discrete

Fourier transform : F",(a) = aA. Theorem 9.3.2 justifies the following definition. Let
a = (a o, a I , ... , an - 1) be an n-tuple. The inverse Fourier transform of a with
respect to w is the n-tuple
F,,' (a) = aB = (n -'pa (1), n -1pa (W-'), n -'Pa ((02 ).... , n -'Pa (w (n -1)))
Problem 9.3.5. Prove that Fw' (Fu,(a)) = Fw(FW' (a)) = a for every a.
The inverse Fourier transform can be calculated efficiently by the following algo-
rithm, provided that n-1 is either known or easily calculable.
Sec. 9.3 The Inverse Transform 283
function FFTinv (a [0.. n - 11, co) : array [0.. n -1 ]

{ n is a power of 2 and w" = 1 }
array F [0.. n -1 ]
F <-FFT(a,w"-')
for i <-Oton-1 do F[i]Fn-'F[i]
return F
Example 9.3.1. Let n = 8 and a = (255, 85, 127, 200, 78, 255, 194, 75). Let us
calculate Fj' (a) in arithmetic modulo m = 257, where w = 4. By Problem 9.3.3, w is
indeed a principal n th root of unity. First we calculate FFT (a , of ' ), where
w' = co7 = 193. To do this, a is decomposed into b=(255,127,78,194)
and c = (85,200,255,75). The recursive calls with 0-2 = 241 yield
B =FFT(b,w2)_(140,221,12,133) and C =FFT(c,w2)_(101,143,65,31).
Combined, these results give A = (241,64,0,9, 39, 121, 24, 0). There remains the mul-
tiplication by n-' = m -(m -1)/n = 225 (Problem 9.3.3). The final result is thus
F = (255,8,0,226,37,240,3,0), which is consistent with Example 9.2.1.
If the Fourier transform is calculated in the field of complex numbers (Problem
9.3.2), rounding errors may occur on the computer. On the other hand, if the transform
is calculated modulo m (Problem 9.3.3), it may be necessary to handle large integers.
For the rest of this section we no longer suppose that arithmetic operations can be per-
formed at unit cost : the addition of two numbers of size 1 takes a time in O (1). We
already know (Problem 9.3.4) that reductions modulo m can be carried out in a time in
O (log m), thanks to the particular form chosen for m. Furthermore, the fact that w is
a power of 2 means that multiplications in the FFT algorithm can be replaced by
shifts. For this it is convenient to modify the algorithm slightly. First, instead of
giving w as the second argument, we supply the base 2 logarithm of w, denoted by y.
Secondly, the recursive calls are made with 2y rather than w2 as the second argument.
The final loop becomes
R <-- 0
for i f-0tot-1 do
I R= iy}
A [i] <-B[i]+C[i]TR
A[t+i]<-- B[i]-C[i]TP
RF-R+y,
where x Ty denotes the value of x shifted left y binary places, that is, x x 2Y. All the
arithmetic is carried out modulo m = w"'2+ I using Problem 9.3.4.
The heart of the algorithm consists of executing instructions of the form
A - (B ± C T (3) mod m, where 0 <- B < m and 0 <- C < m . The value of the shift
never exceeds (z - 1) Ig w, even when the recursive calls are taken into account.
Consequently - w" -' <- B ± C T Q <- w" -' + w" '2, which means that it can be reduced
modulo m in a time in O (log m) = O (n log (o) by Problem 9.3.4. Since the number of
operations of this type is in 0 (n log n), the complete computation of the Fourier
transform modulo m can be carried out in a time in 0 (n 2 log n log co). (From a prac-
tical point of view, if m is sufficiently small that arithmetic modulo m can be con-
sidered to be elementary, the algorithm takes a time in 0 (n log n).)
Problem 9.3.6. Show that the inverse transform modulo m = co"t2+ I can also
be computed in a time in 0 (n 2 log n log co). (The algorithm FFTinv has to be
modified. Otherwise a direct call on the new FFT with y = (n - 1) lg co, corresponding
to the use of w = co' as a principal root of unity, causes shifts that can go up to
1 1
(2 -1) (n - 1) lg co, which means that Problem 9.3.4 can no longer be applied.
Similarly, the final multiplication by n-1 can profitably be replaced by a multiplication
by -n-1 followed by a change of sign, since -n-1 = (o"t2/n is a power of 2.)
9.4 SYMBOLIC OPERATIONS ON POLYNOMIALS
We now have available the tools that are necessary to finish Example 9.1.4. Let
p(x) =a,xs +aS_,xs-1+
+alx+ao and q(x) = btxt+bt_lxt-1+ +b1x+bo be
two polynomials of degrees s and t, respectively. We want to calculate symbolically
the product polynomial r(x)=cdxd+cd_IXd-1+ +clx+co=p(x)q(x)
of degree d = s + t. Let n be the smallest power of 2 greater than d, and let co
be a principal n th root of unity. Let a, b, and c be the n-tuples defined
by a =(ao,a1,.. , as,0,O... , 0), b =(bo,bl,.. , bt,0,0,.. , 0), and
c = (c0 , c 1, ... , cd , 0 , 0 , ... , 0), respectively. (Padding c with zeros is unneces-
sary if d -1 is a power of 2.) Let A = Fu,(a), B = F",(b), and C = Fu,(c). By
definition of the Fourier transform, C; = r (co') = p (w`) q (co`) = A; B; . Therefore C is
the pointwise product of A and B. By Problem 9.3.5, c = Fw 1(C ).
Putting all this together, the coefficients of the product polynomial r (x) are given
by the first d + 1 entries in c = F, 1(Fw(a) x Fjb)). Notice that this reasoning made
no use of the classic unique interpolation theorem, and this is fortunate because unique
interpolation does not always hold when the arithmetic is performed in a ring rather
than in a field. (Consider, for instance, p,(x) = 2x + 1 and p2(x) = 5x + 1 in the ring of
integers modulo 9. Both of these degree 1 polynomials evaluate to 1 and 7 at the
points 0 and 3, respectively.)
Problem 9.4.1. Give explicitly the algorithm we have just sketched. Show
that it can be used to multiply two polynomials whose product is of degree d with a
number of scalar operations in 0 (d log d), provided that a principal n th root of unity
and the multiplicative inverse of n are both easily obtainable, where n is the smallest
power of 2 greater than d.
In order to implement this algorithm for the symbolic multiplication of poly-

nomials, we need to be able to calculate efficiently a principal n th root of unity. The
easiest approach is to use the complex number field and Problem 9.3.2. It is some-
what surprising that efficient symbolic multiplication of polynomials with integer
Sec. 9.4 Symbolic Operations on Polynomials 285
coefficients should require operations on complex numbers. If an exact answer is

required, it becomes necessary to take precautions against the possibility of rounding
errors on the computer. In this case a more thorough analysis of the possible build-up
of rounding errors is needed. For this reason it may be more attractive to carry out the
arithmetic modulo a sufficiently large number (Problem 9.4.2), and to use Problem
9.3.3 to obtain a principal n th root of unity. This may require the use of multiple-
precision arithmetic.
Problem 9.4.2. Let p (x) and q (x) be two polynomials with integer
coefficients. Let a and b be the maxima of the absolute values of the coefficients of
p (x) and q (x), respectively. Let u be the maximum of the degrees of the two polyno-
mials. Prove that no coefficient of the product polynomial p(x)q(x) exceeds ab (u+l )
in absolute value. (In Example 9.1.4, a = 5, b = 6, and u = 3, so no coefficient of
r(x) can exceed 120 in absolute value.)
Example 9.4.1. (Continuation of Example 9.1.4) We wish to multiply sym-
bolically the polynomials p(x) = 3x3-5x2-x +l and q(x) =x3-4x2+6x -2. Since
the product is of degree 6, it suffices to take n = 8. By Problem 9.4.2, all the
coefficients of the product polynomial r (x) = p (x) q (x) lie between -120 and 120; thus
it suffices to calculate them modulo in = 257. By Problem 9.3.3, w = 4 is a principal
n th root of unity in arithmetic modulo 257, and n -' = 225.
Let a = (1, -1, - 5, 3, 0, 0, 0, 0) and b =(-2,6,-4,1,0,O,0,0). Two applications
of the algorithm FFT yield
A = F.(a) = (255,109,199,29,251,247,70,133)
and
B = F,(b) _ (1, 22, 82, 193, 244, 103, 179, 188) .

The pointwise product of these two transforms, still working modulo 257, is
C =(255,85,127,200,78,255,194,75). By Example 9.3.1, the vector c such that
F,(c) = C is c = (255, 8, 0, 226, 37, 240, 3, 0). Since all the coefficients of r (x) lie
between -120 and 120, the integers 255, 226, and 240 correspond to -2, -31, and
-17, respectively. The final answer is therefore
r(x) = 3x6-17x5+37x4-31x3+8x -2 .
Problem 9.4.3. Generalize this idea : give explicitly an algorithm

mul (a [0 .. s ], b [0 .. t ]) : array[0 .. s+t ] that carries out symbolic multiplication of
polynomials with integer coefficients. Among other things, your algorithm should
determine suitable values for n, w, and m. (Use Problems 9.3.3 and 9.4.2 for this.) Q
The analysis of the algorithm of Problem 9.4.3 depends on the degrees s and t of
the polynomials to be multiplied and on the size of their coefficients. If the latter are
sufficiently small that it is reasonable to consider operations modulo m to be elemen-
tary, the algorithm multiplies p (x) and q (x) symbolically in a time in 0 (d log d),
where d = s + t. The naive algorithm would have taken a time in 0 (st ). On the other
hand, if we are obliged to use multiple-precision arithmetic, the initial computation of
the Fourier transforms and the final calculation of the inverse take a time in
0 (d 2 log d log (o), and the intermediate pointwise multiplication of the transforms takes
a time in 0 (d M (d log co)), where M (1) is the time required to multiply two integers of
size 1. Since M (1) n O(1 log l log log 1) with the best-known algorithm for integer mul-
tiplication (Section 9.5), the first term in this analysis can be neglected. The total time
is therefore in 0 (d M (d log co)), where Co = 2 suffices if none of the coefficients of the
polynomials to be multiplied exceeds 2' 14/ 2(1 + max(s , t )) in absolute value.
(Remember that n is the smallest power of 2 greater than d.) By comparison, the naive
algorithm takes a time in 0 (s t M (1)), where 1 is the size of the largest coefficient in
the polynomials to be multiplied. It is possible for this time to be in 0 (st) in practice,
if arithmetic can be carried out on integers of size l at unit cost. The naive algorithm
is therefore preferable to the "fast" algorithm if d is very large and I is reasonably
small. In every case, the algorithm that uses 0) = e 2fi /" can multiply approximately the
two polynomials in a time in 0 (d log d).
Problem 9.4.4. Let x 1 , x 2, ... , x be n distinct points. Give an efficient

algorithm to calculate the coefficients of the unique monic polynomial p (x) of degree n
such that p (xi) = 0 for every 1 <_ i <_ n. (The polynomial is monic if the coefficient of
x" is 1.) Your algorithm should take a time in 0 (n log2n) provided all the necessary
operations are taken to be elementary. (Hint : see Problem 4.11.2.)
* Problem 9.4.5. Let p (x) be a polynomial of degree n , and let x 1 , x 22 ,--- x

be n distinct points. Give an efficient algorithm to calculate each y, = p (x1) for
1 <_ i <_ n. Your algorithm should take a time in 0 (n log2n). (Hint :
the hint to Problem 9.2.2 is relevant here too.)
** Problem 9.4.6. Let x 1 , x2 2 ,--- x be n distinct points, and let

y 1, Y2, , yn be n values, not necessarily distinct. Give an efficient algorithm to
calculate the coefficients of the unique polynomial p (x) of degree less than n such that
p (xi) = yi for every 1 <_ i <_ n. Your algorithm should take a time in 0 (n log2n ).
9.5 MULTIPLICATION OF LARGE INTEGERS
We return once more to the problem of multiplying large integers (Sections 1.1, 1.7.2,
and 4.7). Let a and b be two n-bit integers whose product we wish to calculate. Sup-
pose for simplicity that n is a power of 2 (nonsignificant leading zeros are added at the
left of the operands if necessary). The classic algorithm takes a time in f (n2),
whereas the algorithm using divide-and-conquer requires only a time in 0 (n 1.59), or
even in 0 (n") for any a > 1 (Problem 4.7.8). We can do better than this thanks to a
double transformation of the domain. The original integer domain is first transformed
into the domain of polynomials represented by their coefficients ; then the symbolic
product of these polynomials is obtained using the discrete Fourier transform.
Sec. 9.5 Multiplication of Large Integers 287
We denote by pa (x) the polynomial of degree less than n whose coefficients are
given by the successive bits of the integer a. For instance, p53(x) = X5+X4+X2 + I
because 53 in binary is 00110101. Clearly, pa (2) = a for every integer a. To obtain
the product of the integers a and b, we need only calculate symbolically the polyno-
mial r (x) = p (x) q (x) using the fast Fourier transform (Section 9.4), and then evaluate
r (2). The algorithm is recursive because one of the stages in the symbolic multiplica-
tion of polynomials consists of a pointwise multiplication of Fourier transforms.
Example 9.5.1. To make the illustration simpler, we perform the computation

here in decimal rather than in binary. For the purpose of this example only, let pa (x)
therefore denote the polynomial whose coefficients are given by the successive digits
of a, so that pa (10) = a. Let a = 2301 and b = 1095. Thus pa (x) = 2x3 +3 X2 + 1 and
Pb (x) = x3+9x +5. The symbolic product is
r (x) = pa (X)pb (x) = 2x6+3x5 + 18x4+38x3 + 15x2+9x +5
and we obtain the desired product ab as r (10) = 2,519,595. 0
The recursive nature of this algorithm obliges us to refine the analysis of sym-
bolic multiplication of polynomials given following Problem 9.4.3. Let M (n) be the
time required to multiply two n-bit integers, where n is a power of 2. The central step
in the symbolic multiplication of two polynomials of degree less than n consists of d
multiplications f integers less than wd12+ 1, where d = 2n is a power of 2 greater than
the degree of the product polynomial. Unfortunately, even if we take co = 2, these
integers are of silt, 1 + d Ig w = n + 1. The original multiplication of two integers of
size n therefore requires 2n multiplications of slightly larger integers !
To correct this, we must reduce the degree of the polynomials used to represent
the integers to be multiplied, even if this means increasing the size of their coefficients.
As an extreme case, the algorithm of Section 4.7 consists of representing each integer
by a polynomial of degree 1 whose two coefficients lie between 0 and 2'12 - 1. How-
ever, in order that using the discrete Fourier transform should be attractive, the polyno-
mials considered must have a sufficiently high degree. For instance, suppose we
redefine pa (x) to be the polynomial whose coefficients are given by the successive
figures of the representation of a in base 4. This polynomial is thus of degree less than
n / 2, its coefficients lie between 0 and 3, and pa (4) = a. As before, the polynomial
r (x) = Pa (x)pb (x) is calculated symbolically using the Fourier transform, and the final
answer is obtained by evaluating r (4).
This time the degree of the product polynomial r (x) is less than n. The central
step in the symbolic multiplication of Pa (x) and Pb (x) therefore only requires n multi-
plications of integers less than m = w"i2+ 1. This can be carried out in a time in
n M (2 n Ig w) + 0 (n 2). (The second term is added in case certain operands are
exactly w" 2, since this integer of length 1 + i n Ig (o has to be treated specially.)
Taking into account also the time spent in computing the two initial Fourier transforms
and the final inverse transform, the symbolic multiplication takes a time in
n M (-'n Ig (o) + 0 (n 2log n logo)).
Before analysing the new multiplication algorithm, we must choose a principal

n th root of unity tv. Since the coefficients of the polynomials pa (x) and pb (x) lie
between 0 and 3, and since these polynomials are of degree less than n / 2, the largest
coefficient possible in r (x) is 9n/2 (Problem 9.4.2). As the computations are carried
out modulo m = o)'12 + 1, it is thus sufficient that 9n / 2 < (0/2 . The choice to = 2 is
adequate provided n >: 16. The symbolic computation of r (x) therefore takes a time in
n M (n / 2) + 0 (n 2 log n). The last stage of the multiplication of a and b, namely the
evaluation of r (4), consists of n shifts and n additions of integers whose size is not
more than lg (9n / 2), which takes a negligible time in 0 (n log n). We thus obtain the
asymptotic recurrence M (n) e n M (n / 2) + 0 (n 2 log n).
Problem 9.5.1. Consider the recurrence

t (n) = nt (n / 2), n =2 k , k E lN+
t(1) = 1 .
When n is a power of 2, prove that t(n) = Show that t(n) 0 (nk ), whatever
the value of the constant k.
The preceding problem shows that the modified algorithm is still bad news, even
if we do not take into account the time required to compute the Fourier transforms !
This is explained by the fact that we used the "fast" algorithm for multiplying two
polynomials in exactly the circumstances when it should be avoided : the polynomials
are of high degree and their coefficients are small. To correct this, we must lower the
degree of the polynomials still further.
1! Ign1
Let l = 2 2 ; that is, l = I or l = 2n , depending on whether lg n is even
or odd. Let k = n /I. Note that 1 and k are powers of 2. This time, denote by pa (x)
the polynomial of degree less than k whose coefficients correspond to the k blocks of I
successive bits in the binary representation of a. Thus we have that pa (21) = a. To
calculate the product of the integers a and b, we need only calculate symbolically the
polynomial r (x) = Pa (X)Pb (x), using Fourier transforms, and then evaluate r (21).
Example 9.5.2. Let a = 9885 and b = 21,260, so that n = 16, / = 4,

and k = 4. We form the polynomials pa (x) = 2x3+6x2+9x + 13 and
Pb (x) = 5x3 + 3x2 + 12. The first of these polynomials, for instance, comes from
the decomposition into four blocks of the binary representation of a
0010 0110 1001 1101. The symbolic product is
r(x) =pa(x)pb(x)=10x6+36x5+63x4+116x3+111x2+108x + 156
and the final evaluation yields r (16) = 210,155,100 = 9885 x 21,260.
Let d = 2k, a power of 2 greater than the degree of the product polynomial r (x).
This time we need to choose a principal d th root of unity t,3. Since the coefficients of
the polynomials pa (x) and Pb (x) lie between 0 and 21 - 1, and the degree of these poly-
Sec. 9.5 Multiplication of Large Integers 289
nomials is less than k, the largest coefficient possible in r (x) is k (21 -1)2. It suffices
therefore that k 221 < m = wd 12 + 1, that is Ig co >- (21 + Ig k)/(dl 2). In the case when
Ig n is even I = k = dl 2 = I and Ig w >- (2' + IgFn- )/ = 2 + (Ig J )/W W. Simi-
larly, when n is odd, we obtain Ig w ? 4 + (Ig' )/ . Consequently, w = 8
suffices to guarantee that the computation of the coefficients of r (x) will be correct
when Ig n is even, and w = 32 is sufficient when Ig n is odd.
The multiplication of two n-bit integers is thus carried out using a sym-
bolic multiplication of two polynomials, which takes a time in
d M (2 d lg w) + O (d 2 log d log w). As far as the final evaluation of r(21) is con-
cerned, this can easily be carried out in a time in 0 (d 2 log w), which is negligible.
When n is even, d = 2J and w = 8, which gives M (n) e 24W M (3') + O (n log n).
When n is odd, d = 2n and w = 32; hence, M (n) E 2n M (2 I) + 0 (n log n).
* Problem 9.5.2. Let y > 0 be a real constant, and let t(n) be a function satis-
fying the asymptotic recurrence t(n) E yt(O (' )) + 0 (log n). Prove that
O (log n) if y < 2
t(n)E O (log n log log n) if y= 2
O ((log n)19 Y) if y > 2 .
[Hints: For the second case use the fact that Iglg(1J) (Iglgn) - lg(5/3) provided
that n >_ (310 for every real constant 0 > 1. For the third case prove by constructive
induction that t (n) 5 S [(lg n)1gT - yt(lg n)(197)- 11 - p Ig n , for some constants S, yt, and
p that you must determine and for n sufficiently large. Also use the fact that
(lg1')Ig1< T(lgn)IST+2lgylg1(lg1`r)(IgY)-1
provided n y21g R, for all real constants Q ? 1 and y > 2.]
Let t(n) = M (n)/n . The equations obtained earlier for M (n) lead to
t(n)E6t(3N/n-)+O(logn) when lgn is even, and t(n)E5t(52 )+O(logn) when
n is odd. By Problem 9.5.2, t (n) E O ((log n)1 g 6) c 0 ((log n)2-59). Consequently, this
algorithm can multiply two n-bit integers in a time in M (n) = nt (n) E 0 (n (log n)2.59 ).
Problem 9.5.3. Prove that 0 (n (log n)2.59) c n°C whatever the value of the real
constant a > 1. This algorithm therefore outperforms all those we have seen previ-
ously, provided n is sufficiently large.
Is this algorithm optimal ? To go still faster using a similar approach, Problem

9.5.2 suggests that we should reduce the constant y = 6, which arises here as the max-
imum of 2 x 3 and ' x i . This is possible provided we increase slightly the size of
the coefficients of the polynomials used in order to decrease their degree. More pre-
cisely, we split the n-bit integers to be multiplied into k blocks of I bits, where
I _ 2r+f; ignl and k = n/I for an arbitrary constant i >_ 0. Detailed analysis shows that
this gives rise to 21-' recursive calls on integers of size (2' + 1 + 2-' if Ig n is
even, and 2-' 2n recursive calls on integers of size (2' + + 2-' - if Ig n is odd
1
provided n is sufficiently large. The corresponding y is thus

y=max(21-'(2'+1+2-'),
2-' f(2'++2-i`1)J)=4+21 2r
The algorithm obtained takes a time in O(n(log n)a), where
a=2+Ig(1+2-1-21) <2+2-1-2'/ln2
can be reduced arbitrarily close to 2. Needless to say, increasing the parameter i

reduces the exponent a at the expense of increasing the hidden constant in the asymp-
totic notation.
This is still not optimal, but the algorithms that are even faster are too compli-
cated to be described in detail here. We mention simply that it is possible to obtain
y=4 by calculating the coefficients of the polynomial r (x) modulo 221 + 1 (using the
Fourier transform and proceeding recursively) and modulo k (using the algorithm for
integer multiplication of Section 4.7). Because 221 + 1 and k are coprime, it is then
possible to obtain the coefficients of r (x) by the Chinese remainder theorem, and
finally to evaluate r(2 21 ). The outcome is an algorithm that is capable of multiplying
two n-bit integers in a time in 0 (n log2n). To go faster still, at least asymptotically,
we have to redefine the notion of the "product" of two polynomials so as to avoid dou-
bling the degree of the result obtained. This approach allowed Schonhage and Strassen
to obtain y = 2, that is, an algorithm that takes a time in 0 (n log n log log n). The com-
plexity of this algorithm is such that it is of theoretical interest only.
An integer multiplication algorithm based on the fast Fourier transform has been
used by the Japanese to calculate it to 10 million decimal places. Rather than resorting
to modulo computations in a finite ring, they used a variant involving operations in the
complex number field. Their approach requires care to avoid problems due to rounding
errors but gives rise to a simpler algorithm. It also allows the computation to be car-
ried out directly in decimal, thus avoiding a costly conversion when printing out the
result. Even more decimals of it have been calculated since.
The first published algorithm for calculating discrete Fourier transforms in a time in
O (n log n) is by Danielson and Lanczos (1942). These authors mention that the source
of their method goes back to Runge and Konig (1924). In view of the great practical
importance of Fourier transforms, it is astonishing that the existence of a fast algorithm
remained almost entirely unknown until its rediscovery nearly a quarter of a century
later by Cooley and Tukey (1965). For a more complete account of the history of the
fast Fourier transform, read Cooley, Lewis, and Welch (1967). An efficient implemen-
tation and numerous applications are suggested in Gentleman and Sande (1966),
and Rabiner and Gold (1974). The book by Brigham (1974) is also worth mentioning.
The nonrecursive algorithm suggested in Problem 9.2.2 is described in several refer-
ences, for example Aho, Hopcroft, and Ullman (1974).
Pollard (1971) studies the computation of Fourier transforms in a finite field.
The solution to Problems 9.4.5 and 9.4.6 is given in Aho, Hopcroft, and Ullman
(1974). Further ideas concerning the symbolic manipulation of polynomials, evalua-
tion, and interpolation can be found in Borodin and Munro (1971, 1975), Horowitz and
Sahni (1978), and Turk (1982).
The second edition of Knuth (1969) includes a survey of algorithms for integer
multiplication. A practical algorithm for the rapid multiplication of integers with up to
10 thousand decimal digits is given in Pollard (1971). The algorithm that is able to
multiply two integers of size n in a time in O (n log2n) is attributed to Karp and
described in Borodin and Munro (1975). The details of the algorithm by Schonhage
and Strassen (1971) are spelled out in Brassard, Monet, and Zuffellato (1986), although
the solution given there to Problem 9.3.3 is unnecessarily complicated in the light of
Theorem 9.3.1. Also read Turk (1982). The algorithm used by the Japanese to com-
pute it to 10 million decimal places is described in Kanada, Tamura, Yoshino and
Ushiro (1986); Cray Research (1986) mentions an even more precise computation of
the decimals of it, but does not explain the algorithm used. The empire struck back
shortly thereafter when the Japanese computed 134 million decimals, which is
the world record at the time of this writing; read Gleick (1987). And the saga goes
ever on.
10
Introduction
to Complexity
Up to now, we have been interested in the systematic development and analysis of

specific algorithms, each more efficient than its predecessors, to solve some given
problem. Computational complexity, a field of study that runs in parallel with algo-
rithmics, considers globally the class of all algorithms that are able to solve a given
problem. Using algorithmics, we can prove, by giving an explicit algorithm, that a cer-
tain problem can be solved in a time in 0 (f (n)) for some function f (n) that we aim to
reduce as much as possible. Using complexity, we try to find a function g(n) as large
as possible and to prove that any algorithm that is capable of solving our problem
correctly on all of its instances must necessarily take a time in 92(g(n)). Our satisfac-
tion is complete when f (n) e O(g(n)), since then we know that we have found the most
efficient algorithm possible (except perhaps for changes in the hidden multiplicative
constant). In this case we say that the complexity of the problem is known exactly ;
unfortunately, this does not happen often. In this chapter we introduce only a few of
the principal techniques and concepts used in the study of computational complexity.
10.1 DECISION TREES
This technique applies to a variety of problems that use the concept of comparisons
between elements. We illustrate it with the sorting problem. Thus we ask the fol-
lowing question : what is the minimum number of comparisons that are necessary to
sort n elements? For simplicity we count only comparisons between the elements to
be sorted, ignoring those that may be made to control the loops in our program. Con-
sider first the following algorithm.
292
Sec. 10.1 Decision Trees 293
procedure countsort(T [1.. n ])

i <- min(T), j F- max(T)
array C[i..j]F-0
fork F- 1 to ndo C[T[k]]*-C[T[k]]+1
k F--1
for p F- i to j do
for q <- 1 to C [p ] do
T[k]*-p
k -k+l
Problem 10.1.1. Simulate the operation of this algorithm on an array
T [ 1 .. 10] containing the values 3, 1, 4, 1, 5, 9, 2, 6, 5, 3.
This algorithm is very efficient if the difference between the largest and the smal-
lest values in the array to be sorted is not too large. For example, if
max(T) - min(T) = #T , the algorithm provides an efficient and practical way of sorting
an array in linear time. However, it becomes impractical, on account of both the
memory and the time it requires, when the difference between the elements to be
sorted is large. In this case, variants such as radix sort or lexicographic sort (not dis-
cussed here) can sometimes be used to advantage. However, only in rare applications
will these algorithms prove preferable to quicksort or heapsort. The most important
characteristic of countsort and its variations is that they work using transformations :
arithmetic operations are carried out on the elements to be sorted. On the other hand,
all the sorting algorithms considered in the preceding chapters work using com-
parisons : the only operation allowed on the elements to be sorted consists of com-
paring them pairwise to determine whether they are equal and, if not, which is the
greater. This difference resembles that between binary search and hash coding. In this
book we pay no further attention to algorithms for sorting by transformation.
Problem 10.1.2. Show exactly how countsort can be said to carry out arith-
metic operations on the elements to be sorted. As a function of n, the number of ele-
ments to be sorted, how many comparisons between elements are made?
Coming back to the question we asked at the beginning of this section : what is
the minimum number of comparisons that are necessary in any algorithm for sorting n
elements by comparison ? Although the theorems set out in this section still hold even
if we consider probabilistic sorting algorithms (Section 8.4.1), we shall for simplicity
confine our discussion to deterministic algorithms. A decision tree is a labelled,
directed binary tree. Each internal node contains a comparison between two of the ele-
ments to be sorted. Each leaf contains an ordering of the elements. Given a total order
relation between the elements, a trip through the tree consists of starting from the root
and asking oneself the question that is found there. If the answer is "yes", the trip con-
tinues recursively in the left-hand subtree ; otherwise it continues recursively in the
right-hand subtree. The trip ends when it reaches a leaf ; this leaf contains the verdict
associated with the order relation used. A decision tree for sorting n elements is valid
if to each possible order relation between the elements it associates a verdict that is
294 Introduction to Complexity Chap. 10
compatible with this relation. Finally, a decision tree is pruned if all its leaves are
accessible from the root by making some consistent sequence of decisions. The fol-
lowing problem will help you grasp these notions.
Problem 10.1.3. Verify that the decision tree given in Figure 10.1.1 is valid
for sorting three elements A, B, and C.
Every valid decision tree for sorting n elements gives rise to an ad hoc sorting
algorithm for the same number of elements. For example, to the decision tree of
Figure 10.1.1 there corresponds the following algorithm.
procedure adhocsort3(T [l .. 3])
A- T [1], B- T [2], C <- T [3]
if A < B then if B < C then { already sorted }
else if A < C
then T - A,C,B
else T - C,A,B
else if B < C then if A <C
then T - B,A,C
else T - B,C,A
else T <- C,B,A
Similarly, to every deterministic algorithm for sorting by comparison there
corresponds, for each value of n, a decision tree that is valid for sorting n elements.
Figures 10.1.2 and 10.1.3 give the trees corresponding to the insertion sorting algo-
rithm (Section 1.4) and to heapsort (Section 1.9.4 and Problem 2.2.3), respectively,
when three elements are to be sorted. (The annotations on the trees are intended to
help follow the progress of the corresponding algorithms.) Notice that heapsort
A<B<C C<B<A
A<C<B C,<A<B B<A < C B<C<A
Figure 10.1.1. A valid decision tree for sorting three elements.

BAC ABC
BAA ABB
x=C x=C
CBA BCA BAC ACB ABC
I C<B<A B,<C<A I
A,<CA (answer : no), and then whether C > A
(answer: yes). It would now be possible to establish the correct verdict, but it,
nonetheless, asks again whether B > A before reaching its conclusion. (Despite this,
the tree is pruned : the leaf that would correspond to a contradictory answer "yes" to
the third question has been removed, so that every leaf can be reached by some con-
sistent sequence of decisions.) Thus heapsort is not optimal insofar as the number of
comparisons is concerned. This situation does not occur with the decision tree of
Figure 10.1.2, but beware of appearances : it occurs even more frequently with the
insertion sorting algorithm than with heapsort when the number of elements to be
sorted increases.
Problem 10.1.4. Give the pruned decision trees corresponding to the algo-
rithms for sorting by selection (Section 1.4) and by merging (Section 4.4), and to
quicksort (Section 4.5) for the case of three elements. In the two latter cases do not
stop the recursive calls until there remains only a single element to be "sorted". 0
Problem 10.1.5. Give the pruned decision trees corresponding to the insertion
sorting algorithm and to heapsort for the case of four elements. (You will need a big
piece of paper !) 0
The following observation is crucial: the height of the pruned decision tree
corresponding to any algorithm for sorting n elements by comparison, that is, the dis-
tance from the root to the most distant leaf, gives the number of comparisons carried
yes
C> B
C A
rBA BC
A C A C
BC AB BC BA
A C
BC BA
A C A B C B
BC AB CB AC BA CA
I
A<B<C C<A<B A< C<B B<A<C C < B < A B<C<A
Figure 10.1.3. The three element heapsort decision tree.
out by this algorithm in the worst case. For example, a possible worst case for sorting
three elements by insertion is encountered if the array is already sorted into descending
order (C < B < A ); in this case the three comparisons B < A ?, C < A ?, and C < B ?
situated on the path from the root to the appropriate verdict in the decision tree all have
to be made.
The decision trees we have seen for sorting three elements are all of height 3.
Can we find a valid decision tree for sorting three elements whose height is less ? If so,
we shall have an ad hoc algorithm for sorting three elements that is more efficient in
the worst case. Try it : you will soon see that this cannot be done. We now prove more
generally that such a tree is impossible.
Lemma 10.1.1. Any binary tree with k leaves has a height of at least [lg k 1.
Proof. It is easy to show (by mathematical induction on the total number of
nodes in the tree) that any binary tree with k leaves must have at least k -1 internal
nodes. To say the same thing differently, a binary tree with t nodes in all cannot have
more than It /21 leaves. Now a binary tree of height h can have at most 2h+1-1
nodes in all (by another simple argument using mathematical induction, this time on
the height of the tree), and hence it has at most 2" leaves. The lemma follows
immediately.
Lemma 10.1.2. Any valid decision tree for sorting n elements contains at least
n! leaves. (It may have more than n! leaves if it is not pruned or if some of the leaves
can only be reached when some keys are equal. The upper limit on the number of
leaves of any pruned decision tree can be computed with Problem 5.8.6.)
Proof. A valid tree must be able to produce at least one verdict corresponding to
each of the n! possible orderings of the n elements to be sorted.
Theorem 10.1.1. Any deterministic algorithm for sorting by comparison takes

a time in S2(n log n) to sort n elements in the worst case.
Proof. To every deterministic algorithm for sorting by comparison there
corresponds a pruned valid decision tree for sorting n elements. This tree contains at
least n! leaves by Lemma 10.1.2. Its height is therefore at least [lg(n!)1 by Lemma
10.1.1. By the crucial observation that precedes Lemma 10.1.1, the algorithm thus
needs at least [lg(n!)l comparisons in the worst case to sort n elements. Since each
comparison takes a time in 82(1), and lg(n!)eS2(n logn) (Problem 2.1.17), the algo-
rithm takes a time in S2(n log n) in the worst case.
This proof shows that any deterministic algorithm for sorting by comparison
must make at least Iig(n!)l comparisons in the worst case when sorting n elements.
This certainly does not mean that it is always possible to sort n elements with as few
as [lg(n!)1 comparisons in the worst case. In fact, it has been proved that 30 com-
parisons are necessary and sufficient in the worst case for sorting 12 elements, and yet
rlg(12!)] = 29. In the worst case, the insertion sorting algorithm makes 66 com-
parisons when sorting 12 elements, whereas heapsort makes 59 (of which the first 18
are made during construction of the heap).
Problem 10.1.6. Give exact formulas for the number of comparisons carried
out in the worst case by the insertion sorting algorithm and by the selection sorting
algorithm when sorting n elements. How well do these algorithms do when compared
to the lower bound Fig (n!)] for n = 50 ? 0
** Problem 10.1.7. Prove that the number of comparisons carried out by heap-
sort on n elements, n ? 2, is never greater than 2n ign. Prove further that if n is a
power of 2, then mergesort makes n Ig n - n +I comparisons in the worst case when
sorting n elements. What can you say about sorting by merging in the general case ?
More precise analysis shows that [lg(n!)1 E n lgn -O(n). The previous
problem therefore shows that heapsort is optimal to within a factor of 2 as far as the
number of comparisons needed in the worst case is concerned, and that sorting by
merging almost attains the lower bound. (Some modifications of heapsort come very
close to being optimal for the worst-case number of comparisons.)
Problem 10.1.8. Suppose we ask our sorting algorithm not merely to deter-
mine the order of the elements but also to determine which ones, if any, are equal. For
example, a verdict such as A< B <- C is not acceptable : the algorithm must answer
either A < B < C or A < B = C. Give a lower bound on the number of comparisons
required in the worst case to handle n elements. Rework this problem assuming that
there are three possible outcomes of a comparison between A and B : A B.
Problem 10.1.9. Let T [I .. n ] be an array sorted into ascending order, and let
x be some element. How many comparisons between elements are needed in the worst
case to locate x in the array ? As in Section 4.3, the problem is to find an index i such
that 0 <- i S n and T [i ] <- x < T [i + 11, with the logical convention that T [0] = - o0
and T [n + 1 ] = + oo. How does binary search compare to this lower bound ? What
lower bound on the number of comparisons do you obtain using the decision tree tech-
nique if the problem is simply to determine whether x is in the array, rather than to
determine its position ? 0
Decision trees can also be used to analyse the complexity of a problem on the
average rather than in the worst case. Let T be a binary tree. Define the average
height of T as the sum of the depths of all the leaves divided by the number of leaves.
For example, the decision tree of Figure 10.1.1 has an average height
(2+3+3+3+3+2)/6=8/3. If each verdict is equally likely, then 8/3 is the average
number of comparisons made by the sorting algorithm associated with this tree. Sup-
pose for simplicity that the n elements are all distinct.
Lemma 10.1.3. Any binary tree with k leaves has an average height of at least
lgk. (By comparison with Lemma 10.1.1, we see that there is little difference between
the worst case and the average.)
Proof. Let T be a binary tree with k leaves. Define H (T) as the sum of the
depths of the leaves. For example, H (T) = 16 for the tree in Figure 10.1.1. By
definition, the average height of T is H (T )/k . The root of T can have 0, 1, or 2 chil-
dren. In the first case the root is the only leaf in the tree and H (T) = 0. In the second
case, the single child is the root of a subtree A, which also has k leaves. Since the dis-
tance from each leaf to the root of A is one less than the distance from the same leaf to
the root of T, we have H (T) = H (A) + k. In the third case the tree T is composed of a
root and of two subtrees B and C with i and k - i leaves, respectively, for some
1 <- i < k . By a similar argument we obtain this time H (T) = H (B) +H (C) + k.
For k >-1, define h (k) as the smallest value possible for H (X) for all the binary
trees X with k leaves. In particular, h (1) = 0. If we define h (0) = 0, the preceding dis-
cussion and the principle of optimality used in dynamic programming lead to
h (k) = min I h (i) + h (k - i ) + k 10<-i <-k }
for every k > 1. At first sight this recurrence is not well founded since it defines h (k)
in terms of itself (when we take i =0 or i =k in the minimum, which corresponds to
the root having only one child). This difficulty disappears because it is impossible that
h (k) = h (k) + k. We can thus reformulate the recurrence that defines h (k).
0 if k<-1
h(k) =
k +min[h(i)+h(k-i) 11 <-i <-k-1 } ifk > 1
Now, consider the function g (x) = x Igx + (k - x) lg(k - x), where x e R is such
that 1 <- x <- k -1. Calculating the derivative gives g' (x) =1g x - Ig (k -x), which is
zero if and only if x = k -x; that is, if x = k12. Since the second derivative is posi-
tive, g (x) attains its minimum at x = k / 2. This minimum is g (k / 2) = (k lg k) - k.
The proof that h (k) >- k lg k for every integer k ? 1 now follows by mathematical
induction. The base k = 1 is immediate. Let k > 1. Suppose by the induction
hypothesis that h (j) >- j lg j for every strictly positive integer j < k -1. By
definition,
h(k)=k +min{h(i)+h(k-i) I 1 k + min {g (x) I < x 5 k -1, x E 1R } . Using the result obtained
previously, we have h (k) ? k + g (k / 2) = k lg k.
This shows that H (T) >- k lg k for every tree T with k leaves. The average
height of T being H (T )lk , it is therefore at least Ig k.
* Problem 10.1.10. Let t = Llg k and 1 =k-2, Prove that h (k) = k t + 21 ,

where h (k) is the function used in the proof of Lemma 10.1.3. Prove that this also
implies that h (k) >- k lg k. (Optionally : give an intuitive interpretation of this formula
in the context of the average height of a tree with k leaves.)
Theorem 10.1.2. Any deterministic algorithm for sorting by comparison

makes at least lg(n!) comparisons on the average to sort n elements. It therefore takes
an average time in Q (n log n).
Proof. Follows immediately from Lemmas 10.1.2 and 10.1.3.
Problem 10.1.11. Determine the number of comparisons performed on the

average by the insertion sorting algorithm and by the selection sorting algorithm when
sorting n elements. How do these values compare to the number of comparisons per-
formed by these algorithms in the worst case?
* Problem 10.1.12. Let T [1 .. n ] be an array and k <- n an integer. The

problem consists of returning in descending order the k largest elements of T. Prove
that any deterministic algorithm that solves this problem using comparisons between
the elements must make at least (k / 2) lg (n / 2) comparisons, both in the worst case and
on the average. Conclude that it must take a time in Q(k log n). On the other hand,
give an algorithm able to solve this problem in a time in 0 (n log k) and a space in
O (k) in the worst case. Your algorithm should make no more than one sequential pass
through the array T ; it is therefore efficient even if n is large and if the array is sup-
plied on a magnetic tape. Justify your analysis of the time and space used by your
algorithm.
10.2 REDUCTION
We have just shown that any algorithm for sorting by comparison takes a minimum
time in SZ(n log n) to sort n elements, both on the average and in the worst case. On the
other hand, we know that heapsort and mergesort both solve the problem in a time in
O (n log n). Except for the value of the multiplicative constant, the question of the
complexity of sorting by comparison is therefore settled : a time in O(n log n) is both
necessary and sufficient for sorting n elements. Unfortunately, it does not often happen
in the present state of our knowledge that the bounds derived from algorithmics and
complexity meet so satisfactorily.
Because it is so difficult to determine the exact complexity of most of the prob-
lems we meet in practice, we often have to be content to compare the relative difficulty
of different problems. There are two reasons for doing this. Suppose we are able to
prove that a certain number of problems are equivalent in the sense that they have
about the same complexity. Any algorithmic improvement in the method of solution
of one of these problems now automatically yields, at least in theory, a more efficient
algorithm for all the others. From a negative point of view, if these problems have all
been studied independently in the past, and if all the efforts to find an efficient algo-
rithm for any one of them have failed, then the fact that the problems are equivalent
makes it even more unlikely that such an algorithm exists. Section 10.3 goes into this
second motivation in more detail.
Definition 10.2.1. Let A and B be two solvable problems. A is linearly redu-

cible to B , denoted A <_ 1 B, if the existence of an algorithm for B that works in a time
in 0(t(n)), for any function t(n), implies that there exists an algorithm for A that also
works in a time in 0 (t (n)). When A 51 B and B S ' A both hold, A and B are linearly
equivalent, denoted A = 1 B.
Even if we are not able to determine the complexities of A and B exactly, when
A =' B we can be sure that they are the same. In the remainder of this section we
shall see a number of examples of reduction from a variety of application areas.
* Problem 10.2.1. A less restrictive definition of linear reduction is obtained if

we content ourselves with comparing the efficiency of algorithms for A on instances of
size n with the efficiency of algorithms for B on instances of size in O (n). For the
purposes of this problem only, write A <g B if the existence of an algorithm for B
that works in a time in O(t(n)), for any function t(n), implies that there exists an algo-
rithm for A that works in a time in 0 (t(0 (n))). Show with the help of an explicit
example that the notions A <_! B and A <g B are not equivalent even if there exists an
algorithm for B that works in a time in O (p (n )), where p (n) is a polynomial.
Problem 10.2.2. Prove that the relations <_ 1 and = 1 are transitive.
Sec. 10.2 Reduction 301
Let us extend the notion of smooth functions (introduced in Section 2.1.5) to

algorithms and problems. An algorithm is smooth if it takes a time in O(t(n)) for some
smooth function t. Even though a smooth function must be eventually nondecreasing
by definition, this does not imply that the actual time taken by a specific implementa-
tion of the algorithm must also be given by an eventually nondecreasing function.
Consider for instance the modular exponentiation algorithm dexpo of Section 4.8. We
have seen (Problem 4.8.5) that the time it takes to compute a"mod m is a linear func-
tion both of lg n and the number of Is in the binary representation of n . In particular,
it takes longer to compute a 31st power than a 32nd. The actual time taken by any rea-
sonable implementation of this algorithm is not an eventually nondecreasing function
of the exponent. Nonetheless, this algorithm is smooth because it takes a time in
O(log n ), counting the multiplications at unit cost, and logn is smooth.
A problem is smooth if any reasonable algorithm that solves it is smooth. By
"reasonable", we mean an algorithm that does not purposely waste time. No problem
could be smooth without this restriction to reasonable algorithms because any problem
that can be solved at all can be solved by an algorithm that takes a time in Q(2'),
which cannot be smooth.
A function t : IN -* It* is at least quadratic if t(n) E c (n2). It is strongly at
least quadratic (strongly quadratic for short) if it is eventually nondecreasing and if
t(an) >_ a2t(n) for every positive integer a and every sufficiently large integer n.
Finally, it is supra quadratic if it is eventually nondecreasing and if there exists an
CEIR+ such that t(an) >_ a2+et(n) for every positive integer a and every sufficiently
large integer n. At least, strongly and supra linear functions are defined similarly.
These notions extend to algorithms and problems as in the case of smooth functions.
Problem 10.2.3.
i. Prove that any strongly quadratic function is at least quadratic. (Hint: apply
Problem 2.1.20.)
ii. Give an explicit example of an eventually nondecreasing function that is at least
quadratic but not strongly quadratic.
iii. Show that n 2log n is strongly quadratic but not supra quadratic. 0
Most theorems in this section are stated conditionally on a "reasonable" assump-

tion, such as "A <_ ' B, assuming B is smooth". This can be interpreted literally as
meaning that A <_ 1 B under the assumption that B is smooth. From a more practical
point of view it also means that the existence of an algorithm for B that works in a
time in 0 (t(n)), for any smooth function t(n), implies that there exists an algorithm for
A that also works in a time in 0 (t(n)). Moreover, all these theorems are constructive :
the algorithm for B follows from the algorithm for A and the proof of the
corresponding theorem.
10.2.1 Reductions Among Matrix Problems
An upper triangular matrix is a square matrix M whose entries below the diagonal are
all zero, that is, M, = 0 when i > j. We saw in Section 4.9 that a time in O (n 2.81)
(or even 0 (n 2-376)) is sufficient to multiply two arbitrary n x n matrices, contrary to the
intuition that may suggest that this problem will inevitably require a time in S2(n3).
Is it possible that multiplication of upper triangular matrices could be carried out
significantly faster than the multiplication of two arbitrary square matrices ? From
another point of view, experience might well lead us to believe that inverting non-
singular upper triangular matrices should be an operation inherently more difficult than
multiplying them.
We denote these three problems, that is, multiplication of arbitrary square
matrices, multiplication of upper triangular matrices, and inversion of nonsingular
upper triangular matrices, by MQ, MT, and IT, respectively. We shall show under rea-
sonable assumptions that MQ I MT =_1 IT. (The problem of inverting an arbitrary
nonsingular matrix is also linearly equivalent to the three preceding problems (Problem
10.2.9), but the proof of this is much more difficult, it requires a slightly stronger
assumption, and the resulting algorithm is numerically unstable.) Once again this
means that any new algorithm that allows us to multiply upper triangular matrices
more efficiently will also provide us with a new, more efficient algorithm for inverting
arbitrary nonsingular matrices (at least in theory). In particular, it implies that we can
invert any nonsingular n x n matrix in a time in 0 (n 2.376).
In what follows we measure the complexity of algorithms that manipulate n x n
matrices in terms of n, referring to an algorithm that runs in a time in O(n 2) as qua-
dratic. Formally speaking, this is incorrect because the running time should be given as
a function of the size of the instance, so that a time in O(n2) is really linear. No con-
fusion should arise from this. Notice that the problems considered are at least qua-
dratic in the worst case because any algorithm that solves them must look at each entry
of the matrix or matrices concerned.
Theorem 10.2.1. MT <_ I MQ.
Proof. Any algorithm that can multiply two arbitrary square matrices can be
used directly for multiplying upper triangular matrices.
Theorem 10.2.2. MQ <_ ' MT, assuming MT is smooth.
Proof. Suppose there exists an algorithm that is able to multiply two n x n

upper triangular matrices in a time in 0 (t (n)), where t (n) is a smooth function. Let A
and B be two arbitrary n x n matrices to be multiplied. Consider the following matrix
product:
0 A 0 0 0 0 0 0 AB
000 x 00B = 00 0
0 0 0 0 0 0 0 0 0
where the "0" are n x n matrices all of whose entries are zero. This product shows us
how to obtain the desired result AB by multiplying two upper triangular 3n x 3n
matrices. The time required for this operation is in 0 (n 2) for the preparation of the
two big matrices and the extraction of AB from their product, plus 0 (t (3n)) for the
multiplication of the two upper triangular matrices. By the smoothness of t(n),
t (3n) E O (t (n)). Because t(n) is at least quadratic, n 2 E O (t (n)). Consequently, the
total time required to obtain the product AB is in 0 (t(n)).
Theorem 10.2.3. MQ <- ' IT, assuming IT is smooth.

Proof. Suppose there exists an algorithm that is able to invert a nonsingular
n x n upper triangular matrix in a time in O (t(n)), where t (n) is a smooth function.
Let A and B be two arbitrary n x n matrices to be multiplied. Consider the following
matrix product :
I AO 1 -A AB 100
0IB x 0 1 -B = 010
0 0/ 00 1 001
where I is the n x n identity matrix. This product shows us how to obtain the desired
result AB by inverting the first of the 3n x 3n upper triangular matrices. As in the
proof of the previous theorem, this operation takes a time in 0 (t(n)).
Theorem 10.2.4. IT <' MQ, assuming MQ is strongly quadratic. (In fact, a

weaker but less natural hypothesis suffices : it is enough that MQ be supra linear.)
Proof. Suppose there exists an algorithm that is able to multiply two arbitrary
n x n matrices in a time in 0 (t (n)), where t (n) is strongly quadratic. Let A be a non-
singular n x n upper triangular matrix to be inverted. Suppose for simplicity that n is a
power of 2. (See Problem 10.2.4 for the general case.) If n = 1, inversion is trivial.
Otherwise decompose A into three submatrices B, C, and D each of size n / 2 x n / 2
such that
Cl
A= I B0DJ
where B and D are upper triangular and C is arbitrary. By Problem 10.2.5, the
matrices B and D are nonsingular. Now consider the following product :
BC X B-1 -B-'CD-' 10
0D 0 D-' 01
This product shows us how to obtain A-' by first calculating B-1, and D-', and then
multiplying the matrices B-', C, and D-'. The upper triangular matrices B and D,
which we now have to invert, are smaller than the original matrix A. Using the
divide-and-conquer technique suggests a recursive algorithm for inverting A in a time
in O(g(n)) where g(n) a 2g(nl2)+2t(nl2)+O(n2). The fact that t(n)ES2(n2) and
the assumption that t(n) is eventually nondecreasing (since it is strongly quadratic)
yield g(n) e 2g(n / 2) +O (t(n)) when n is a power of 2. By Problem 10.2.6, using the
assumption that t(n) is strongly quadratic (or at least supra linear), this implies that
g(n) e 0 (t (n) I n is a power of 2).
Problem 10.2.4. Let IT2 be the problem of inverting nonsingular upper tri-
angular matrices whose size is a power of 2. All that the proof of theorem 10.2.4
really shows is that IT2 <_' MQ. Complete the proof that IT <_' MQ.
Problem 10.2.5. Prove that if A is a nonsingular upper triangular matrix

whose size is even, and if B and D are defined as in the proof of theorem 10.2.4, then
B and D are nonsingular.
Problem 10.2.6. Prove that if g(n) E 2g(n / 2) +O (t (n)) when n is a power of

2, and if t(n) is strongly quadratic, then g(n) E O (t(n) I n is a power of 2). (Hint :
apply Problem 2.3.13(iv); note that it is enough to assume that t(n) is supra linear.)
Problem 10.2.7. An upper triangular matrix is unitary if all the entries on its
diagonal are 1. Denote by SU the problem of squaring a unitary upper triangular
matrix. Prove that SU MQ under suitable assumptions. What assumptions do you
need ?
Problem 10.2.8. A matrix A is symmetric if Aid = Aji for all i and j. Denote
by MS the problem of multiplying symmetric matrices. Prove that MS =' MQ under
suitable assumptions. What assumptions do you need?
** Problem 10.2.9. Denote by IQ the problem of inverting an arbitrary non-

singular matrix. Assume that both IQ and MQ are smooth and supra quadratic. Prove
that IQ =' MQ. (Note : this reduction would not go through should an algorithm that
is capable of multiplying n x n matrices in a time in 0 (n 2 log n) exist - see Problem
10.2.3(iii).) t7
10.2.2 Reductions Among Graph Problems
In this section IR°° denotes IR*u I+-}, with the natural conventions that
x +(+oo) =+oo and min(x,+oo) = x for all x e R-.
Let X, Y, and Z be three sets of nodes. Let f : X x Y -4 IR°° and g : Yx Z -4 IR°°
be two functions representing the cost of going directly from one node to another.
An infinite cost represents the absence of a direct route. Denote by
fg the function h : X x Z -* IR°° defined for every x E X and z (=Z by h (x , z) =

min { f (x, y) + g (y, z) I y E Y }. This represents the minimum cost of going from x to
z passing through exactly one node in Y. Notice the analogy between this definition
and ordinary matrix multiplication (where addition and multiplication are replaced by
the minimum operation and addition, respectively), but do not confuse this operation
with the composition of functions.
The preceding notation becomes particularly interesting when the sets X, Y, and
Z, and also the functions f and g, coincide. In this case ff , which we shall write f 2,
gives the minimum cost of going from one node of X to another (possibly the same)
while passing through exactly one intermediate node (possibly the same, too). Simi-
larly, min(f , f2) gives the minimum cost of going from one node of X to another
either directly or by passing through exactly one intermediate node. The meaning of
f' is similar for any i > 0. By analogy, f ° represents the cost of going from one node
to another while staying in the same place, so that
° 0 ifx =y
f (x' y) +00 otherwise.
The minimum cost of going from one node to another without restrictions
on the number of nodes on the path, which we write f*, is therefore
f * = min { f' I i >_ 0 1. This definition is not practical because it apparently implies an
infinite computation ; it is not even immediately clear that f * is well defined. How-
ever, f never takes negative values. Any path that passes twice through the same node
can therefore be shortened by taking out the loop thus formed, without increasing the
cost of the resulting path. Consequently, it suffices to consider only those paths whose
length is less than the number of nodes in X. Let this number be n. We thus have
that f * = min ( f' I 0 <_ i < n . At first sight, computing f * for a given function f
}
seems to need more time than calculating a simple product fg.

The straightforward algorithm for calculating fg takes a time in O(n3) if the
three sets of nodes concerned are of cardinality n. Unfortunately, there is no obvious
way of adapting to this problem Strassen's algorithm for ordinary matrix multiplication
(Section 4.9). (The intuitive reason is that Strassen's algorithm does subtractions.
There is no equivalent to this operation in the present context since taking the
minimum is not a reversible operation. Nevertheless, there exist more efficient algo-
rithms for this problem. Because they are quite complicated and have only theoretical
advantages, we do not discuss them here.) So the definition of f * can be taken to give
us a direct algorithm for calculating its value in a time in O(n4).
However, we saw in Section 5.4 a dynamic programming algorithm for calcu-
lating shortest paths in a graph, namely Floyd's algorithm. This calculation is nothing
other than the calculation of f * . Thus it is possible to get away with a time in O(n 3)
after all. Could it be that the problems of calculating fg and f * are of the same com-
plexity ? The following two theorems show that this is indeed the case : these two
problems are linearly equivalent. The existence of algorithms asymptotically more
efficient than O(n3) for solving the problem of calculating fg therefore implies that
Floyd's algorithm for calculating shortest routes is not optimal, at least in theory.
Denote by MUL and TRC the problems consisting of calculating fg and f*,
respectively. As in the previous section, time complexities will be measured as a func-
tion of the number of nodes in the graphs concerned. An algorithm such as Dijkstra's,
for instance, would be considered quadratic even though it is linear in the number of
edges (for dense graphs). Again, the problems considered are at least quadratic in the
worst case because any algorithm that solves them must look at each edge concerned.
Theorem 10.2.5. MUL <_ ' TRC, assuming TRC is smooth.
Proof. Suppose there exists an algorithm that is able to calculate h* in a time in

O(t(n)), for a smooth function t(n), where n is the cardinality of the set W such that
h : W x W - 1R°°. Let X, Y, and Z be three sets of nodes of cardinality n , n 2 and n 3,
respectively, and let f : X x Y - 1R°° and g : Y x Z - 1R°° be two functions for which
we wish to calculate fg.
Suppose without loss of generality that X, Y, and Z are disjoint. Let
W = X u Y u Z. Define the function h : W x W- 1R°° as follows :
f(u,v) ifuEX andvEY

h(u,v)= g(u,v) ifuEY andvEZ
+00 otherwise .
Notice in particular that h (x, z) _ + oo when x e X and z E Z.

Now, let us find the value of h 2(u , v). By definition, h 2(U, v) = min ( h (u, w) +
h (w, v) I w E W I. By the definition of h, it is impossible that h (u, w) # +oo and
h (w, v) # +oo simultaneously unless w c= Y. Consequently, h 2(U, v) =min { h (u , y) +
h (y, v) I y E Y 1. But the only way to have h (u, y) :;& + oo when y E Y is to have u EX,
and the only way to have h (y, v) # +oo when y E Y is to have v EZ. If u 0 X or if
v 0 Z, it therefore follows that h 2(u , v) = +o. In the case when u EX, y E Y, and
v e Z it suffices to note that h (u , y) = f (u , y) and h (y, v) = g (y, v) to conclude that
h 2(U, v) = min { f (u , y) + g (y, v) I y c: Y ), which is precisely the definition of
f g (u, v). Summing up, we have
fg(u,v) if u EX and veZ
h
2
(u,=
v)
+ otherwise .
The calculation of h 3(u , v) is easier. By definition, h 3(u , v) = hh 2(u , v) =

min ( h (u , w) +h 2(w , v) I W E W 1. But h (u , w) _ + oo when w E X whereas
h 2(W, v) = + 00 when w e X. Therefore h (u, w) + h 2(w , v) = +.0 for every WE W,
which implies that h 3(U, v) _ + oo for all u, v E W. The same holds for h ` (u , v) for all
i>3.
The conclusion from all this is that h* = min(h °, h , h 2) is given by the following
equation.
0 ifu =v
f(u,v) if uEX andvEY
h*(u,v) = g(u,v) ifuEY andvEZ
fg(u,v) if uEX and vEZ
+oo otherwise
Therefore the restriction of h* to X x Z is precisely the product fg we wished to calcu-
late. Let n = n I +n 2+n 3 be the cardinality of W. Using the algorithm for calculating
h* thus allows us to compute fg in a time in
t (n) + O (n 2)C O (t(3 max(n i , n 2, n 3))) + O (n 2)g 0 (t (max(n I , n 2, n 3)))
because t(n) is smooth and at least quadratic.
Theorem 10.2.6. TRC <_' MUL, assuming MUL is strongly quadratic.

(In fact, a weaker but less natural hypothesis suffices again : it is enough that MUL be
supra linear.)
Proof. Suppose there exists an algorithm that is able to calculate fg in a time in
0 (t (max(n I , n 2, n 3))), where n I , n 2, and n 3 are the cardinalities of the sets X, Y, and
Z such that f : X x Y -+ 1R- and g : Y x Z -* 1R°°. Assume t (n) is strongly quadratic. Let
H be a set of cardinality n and let h : H x H - 1R- be a cost function for which we
wish to calculate h*. Suppose for simplicity that n is a power of 2 (see Problem
10.2.11 for the general case). If n = 1, it is obvious that h* (u , u) = 0 for the
unique u EH. Otherwise split H into two disjoint subsets J and K, each containing
half the nodes. Define a : J x J -> IR°°, b : J x K - 1R°°, c : K x K -* 1R°°, and
d : KxJ -* R- as the restrictions of h to the corresponding subdomains.
Let e : J x J - 1R°° be the function given by e = min(a , be*d ). Notice that
e (u , v ), for u and v in J, represents the minimum cost for going from u to v without
passing through any node in J (not counting the endpoints) ; however, the nodes in K
can be used as often as necessary, or not used at all if this is preferable. The minimum
cost for going from u to v with no restrictions on the path is therefore e* (u, v) when u
and v are both in J. In other words, e* is the restriction to J x J of the h* that we
wish to calculate. The other restrictions are obtained similarly.
e*(u,v) if ueJ and veJ

e*bc* (u, v) if u E J and v E K
h*(u,v) = <
c*de* (u, v) if u E K and V E J

(min(c*, c*de*bc* )) (u , v) ifu E K and v E K
To calculate h* using divide-and-conquer, we therefore solve recursively two instances

of the TRC problem, each of size n / 2, in order to obtain c* and e*. The desired
result is then obtained after a number of instances of the MUL problem. As in
theorem 10.2.4, the time g(n) required by this approach is characterized by the equa-
tion g(n) E 2g(n / 2)+0 (t(n)), which implies that g(n) E 0 (t(n) I n is a power of 2).
* Problem 10.2.10. Prove formally that the preceding formula for h* is correct.
Problem 10.2.11. Let TRC2 be the problem of calculating h* for

h : X x X - 1R°° when the cardinality of X is a power of 2. All the proof of theorem
10.2.6 really shows is that TRC2 5 1 MUL. Complete the proof that TRC 5 ' MUL.
When the range of the cost functions is restricted to 10, +00 }, calculating f *
comes down to determining for each pair of nodes whether or not there is a path
joining them, regardless of the cost of the path. We saw that Warshall's algorithm
(Problem 5.4.2) solves this problem in a time in 0(n3). Let MULB and TRCB be the
problems consisting of calculating fg and h*, respectively, when the cost functions are
restricted in this way. It is clear that MULB <_ 1 MUL and TRCB <_ 1 TRC since the
general algorithms can also be used to solve instances of the restricted problems.
Furthermore, the proof that MUL =/ TRC can easily be adapted to show that
MULB =' TRCB. This is interesting because MULB <_ 1 MQ, where MQ is the
problem of multiplying arbitrary arithmetic square matrices (Problem 10.2.12). Unlike
the case of arbitrary cost functions, Strassen's algorithm can therefore be used to solve
the problems MULB and TRCB in a time in 0 (n 2.81), thus showing that Warshall's
algorithm is not optimal. Note, however, that using Strassen's algorithm requires a
number of arithmetic operations in 0 (n 2-81 ) ; the time in 0 (n 3) taken by Warshall's
algorithm counts only Boolean operations as elementary. No algorithm is known that
can solve MULB faster than MQ.
Problem 10.2.12. Let f : X x Y -* 10,+-) and g: Y x Z -* (0,+-l be two

restricted cost functions. Assuming we count arithmetic operations at unit cost, show
how to transform the problem of calculating fg into the computation of an ordinary
arithmetic matrix multiplication. Conclude that MULB <_ 1 MQ. Show that the arith-
metic can be done modulo p, where p is any prime number larger than n.
* Problem 10.2.13. A cost function f : X x X -4R- is symmetric if

f (u , v) = f (v , u) for every u , v E X . Each of the four problems discussed earlier has a
symmetric version that arises when the cost functions involved are symmetric. Call
these four problems MULS, TRCS, MULBS, and TRCBS. Prove that
MULBS=1 MULB. Do you believe that MULBS=1 TRCBS? If not, does one of
these two problems appear strictly more difficult than the other? Which one? Justify
your answer.
10.2.3 Reductions Among Arithmetic and Polynomial Problems
We return to the problems posed by the arithmetic of large integers (sections 1.7.2, 4.7,
and 9.5). We saw that it is possible to multiply two integers of size n in a time in
0 (n 1.59) and even in 0 (n log n log log n). What can we say about integer division and
taking square roots ? Our everyday experience leads us to believe that the second of
these problems, and probably the first one, too, is genuinely more difficult than multi-
plication. Once again this turns out not to be true. Let SQR, MLT, and DIV be the
problems consisting of squaring an integer of size n, of multiplying two integers of
size n, and of determining the quotient when an integer of size 2n is divided by an
integer of size n, respectively. Clearly, these problems are at least linear because any
algorithm that solves them must take into account every bit of the operands involved.
(For simplicity we measure the size of integers in bits. As mentioned in Section 1.7.2,
however, this choice is not critical : the time taken by the various algorithms would be
in the same order if given as a function of the size of their operands in decimal digits
or computer words. This is the case precisely because we assume all these algorithms
to be smooth.)
Theorem 10.2.7. SQR =' MLT =' DIV, assuming these three problems are
smooth and MLT is strongly linear (weaker but more complicated assumptions would
suffice).
Proof outline. The full proof of this theorem is long and technical. Its concep-
tual beauty is also defaced in places by the necessity of using an inordinate number of
ad hoc tricks to circumvent the problems caused by integer truncation (see Problem
10.2.22). For this reason we content ourselves in the rest of this section with showing
the equivalence of these operations in the "cleaner" domain of polynomial arithmetic.
Nonetheless, we take a moment to prove that SQR MLT, assuming SQR is smooth
(a weaker assumption would do).
Clearly, SQR 51 MLT, since squaring is only a special case of multiplication. To
show that MLT <_' SQR, suppose there exists an algorithm that is able to square an
integer of size n in a time in 0 (t(n)), where t(n) is smooth (it is enough to assume that
t (n + 1) E O (t (n)) ). Let x and y be two integers of size n to be multiplied. Assume
without loss of generality that x >- y . The following formula enables us to obtain their
product by carrying out two squaring operations of integers of size at most n + 1, a few
additions, and a division by 4:
xy =((x+y)2-(x-y)2)/4.
Since the additions and the division by 4 can be carried out in a time in O (n), we can
solve MLT in a time in 2t(n +1)+O(n)c O(t(n)) because t(n) is smooth and
t (n) E Q(n).
We have seen in Section 9.4 how to multiply two polynomials of degree n in a

time in 0 (n log n), provided that the necessary arithmetic operations on the coefficients
can be counted at unit cost. We now show that the problem of polynomial division is
linearly equivalent to that of multiplication. Notice that a direct approach using
discrete Fourier transforms, which works so well for multiplying two polynomials, is
inapplicable in the case of division unless the two polynomials concerned divide one
another exactly with no remainder. For example, let p(x)=x3+3x2+x+2 and
d(x) = x2+x +2. The quotient of the division of p (x) by d (x) is q (x) =x+2. In this
case p (2) = 24 and d (2) = 8, but q (2) = 4 # 24/8. We even have that p (1) = 7 is not
divisible by d (1) = 4. This is all due to the remainder of the division r (x) = - 3x - 2.
Despite this difficulty, it is possible to determine the quotient and the remainder pro-
duced when a polynomial of degree 2n is divided by a polynomial of degree n in a
time in 0 (n log n) by reducing these problems to a certain number of polynomial mul-
tiplications calculated using the Fourier transform.
Recall that p (x) = E,'_ 0 ai x' is a polynomial of degree n provided that an # 0.
By convention the polynomial p (x) = 0 is of degree -1. Let p (x) be a polynomial of
degree n, and let d (x) be a nonzero polynomial of degree m. Then there exists a
unique polynomial r (x) of degree strictly less than m and a unique polynomial q (x)
such that p (x) = q (x) d (x) + r (x). The polynomial q (x) is of degree n - m if
n > m -1 ; otherwise q (x) = 0. We call q (x) and r (x), respectively, the quotient and
the remainder of the division of p (x) by d (x). By analogy with the integers, the quo-
tient is denoted by q(x) = Lp(x)Id(x)J.
Problem 10.2.14. Prove the existence and the uniqueness of the quotient and
the remainder. Show that if both p (x) and d (x) are monic polynomials (the coefficient
of highest degree is 1) with integer coefficients then both q (x) and r (x) have integer
coefficients and q (x) is monic (unless q (x) = 0).
We also need the notion of an inverse. Let p (x) be a nonzero polynomial of

degree n. The inverse of p (x), which we denote p* (x), is defined by
p*(x)= Lx 2nIp(x)]. For example, if p(x)=x3+3x2+x+2, then p*(x)=
x 3 - 3 x 2 + 8x - 23. Notice that p (x) and p* (x) are always of the same degree.
Problem 10.2.15. Let p(x) = x 3 +x 2 + 5 x + 1 and d(x)=x-2. Calculate

L p (x)/d (x)], p* (x), and d* (x).
Problem 10.2.16. Prove that if p (x) is a nonzero polynomial and if

q (x) = p* (x) then q* (x) = p (x). (There is a very simple proof.)
* Problem 10.2.17. Prove that if p (x), p I (x) and p 2(x) are three arbitrary poly-
nomials, and if d (x), d i(x) and d2(x) are three nonzero polynomials, then
P I(x) p2(x) _ p i(x)d 2(x) ± p2(x)d 1(x)

d i(x) ± d2(x) d i(x)d2(x)
in particular, Lp1(x)/d(x)] ± Lp2(x)/d(x)] = L(p1(x)±p2(x))/d(x)];
ii. LLp(x)/di(x)j/d2(x)J = Lp(x)/(di(x)d2(x))]; and
iii. Lp1(x)/Lp2(x)/d(x)Jj = L(p1(x)d(x))/p2(x)], provided that the degree of p1(x)
is not more than twice the degree of LP 2(X)ld (x)j. 11
Consider the four following problems : SQRP consists of squaring a polynomial

of degree n, MLTP of multiplying two polynomials of degree at most n, INVP
of determining the inverse of a polynomial of degree n and DIVP of calculating the
quotient of the division of a polynomial of degree at most 2n by a polynomial

of degree n. We now prove under suitable assumptions that these four problems
are linearly equivalent using the following chain of reductions : MLTP <-'
SQRP <' INVP <-' MLTP and INVP:-' DIVP <-' INVP. Again, all these problems
are at least linear. We assume arithmetic operations can be carried out at unit cost on
the coefficients.
Theorem 10.2.8. MLTP <-' SQRP, assuming SQRP is eventually nonde-

creasing.
Proof. Essentially the same as the proof that MLT <-1 SQR given in theorem
10.2.7. There is no need for a smoothness assumption this time because the sum or
difference of two polynomials of degree n cannot exceed degree n. 11
Theorem 10.2.9. SQRP <- 1 INVP, assuming INVP is smooth.

Proof. The intuitive idea is given by the following formula, where x is a
nonzero real number :
x2=(x-'-(x+1)-')-'-x .
A direct attempt to calculate the square of a polynomial p (x) using the analogous for-
mula (p* (x) - (p (x) + 1)* )* -p (x) has no chance of working : the degree of this
expression cannot be greater than the degree of p (x). This failure is caused by trunca-
tion errors, which we can, nevertheless, eliminate using an appropriate scaling factor.
Suppose there exists an algorithm that is able to calculate the inverse of a poly-
nomial of degree n in a time in 0 (t(n)), where t(n) is a smooth function. Let p (x) be a
polynomial of degree n > 1 whose square we wish to calculate. The polynomial
x In p (x) is of degree 3n, so
[X2np(x)]* = Lx6n/x2np(x)J = LX4n/p(x)J II
Similarly
[x2n (p(x)+1)]* = Lx4n/(p(x)+l)J
By Problem 10.2.17
[x2n p(x)]* -[x2n (p(x)+l)]* = Lx4n/p(x)]- Lx4n/(p(x)+1)]
= L(x4n
(p(x)+1)-x 4np(x))l(p(x)(p(x)+1))j
= Lx4nl(p2(x)+p(x))j
= [p2(x)+p(x)]* .
The last equality follows from the fact that p 2(x) + p (x) is of degree 2n. By Problem
10.2.16, we conclude finally that
p2(x) _ [[x2n p(x)]* -[x2n (p(x)+l)]*]* -p(x)
This gives us an algorithm for calculating p2(x) by performing two inversions of poly-
nomials of degree 3n, one inversion of a polynomial of degree 2n, and a few opera-
tions (additions, subtractions, multiplications by powers of x) that take a time in O (n).
This algorithm can therefore solve SQRP in a time in
2t(3n)+t(2n)+0(n)c 0(t(n))
because t(n) is smooth and at least linear.
Theorem 10.2.10. INVP <' DIVP.

Proof. To calculate p* (x), where p (x) is a polynomial of degree n, we evaluate
Lx2n /p (x)j, an instance of size n of the problem of polynomial division.
Theorem 10.2.11. DIVP :- I INVP, assuming INVP is smooth.

Proof. The intuitive idea is given by the following formula, where x and y are
real numbers and y # 0:
x/y = xy-I
If we try to calculate the quotient of a polynomial p (x) divided by a nonzero polyno-
mial d (x) using directly the analogous formula p (x) d* (x), the degree of the result is
too high. To solve this problem we divide the result by an appropriate scaling factor.
Suppose there exists an algorithm that is able to calculate the inverse of a poly-
nomial of degree n in a time in 0 (t(n)), where t(n) is a smooth function. Let p (x) be
a polynomial of degree less than or equal to 2n, and let d (x) be a polynomial of
degree n. We wish to calculate LP (x)/d (x)j. Let r (x) be the remainder of the division
of x 2n by d (x), which is to say that d* (x) = Lx 2" /d (x) j = (x 2n - r (x)) /d (x) and that
the degree of r (x) is strictly less than n . Now consider
d* (x)p (x) = x 2n p (x) - r (x)p (x)

X 2" x2n d(x)
x 2n p (x) _ r (x) p (x)

x2nd(x) x2"d(x)
by Problem 10.2.17(i). But the degree of r (x) p (x) is strictly less than 3n , whereas the
degree of x 2" d (x) is equal to 3n , and so L(r (x) p (x))/(x 2" d (x))] = 0. Consequently,
L(d* (x)p (x ))/x2n] = L p (x) Id (x)], which allows us to obtain the desired quotient by
performing an inversion of a polynomial of degree n, the multiplication of two polyno-
mials of degree at most 2n, and the calculation of the quotient from a division by a
power of x. This last operation corresponds to a simple shift and can be carried
out in a time in O(n). The multiplication can be performed in a time in O(t(2n))
thanks to theorems 10.2.8 and 10.2.9 (using the assumption that t(n) is smooth).
The calculation of Lp(x)/d(x)] can therefore be carried out in a time in

t(n)+O(t(2n))+O(n)c O(t(n)) because t(n) is smooth and at least linear. El
Theorem 10.2.12. INVP 5' MLTP, assuming MLTP is strongly linear

(a weaker assumption will do - see the hint of Problem 10.2.19).
Proof. This reduction is more difficult than the previous ones. Once again, we
appeal to an analogy with the domain of real numbers, namely, Newton's method for
finding the zero of f (w) = I - xw. Let x be a positive real number for which we want
to calculate x-1. Let y be an approximation to x-1 in the sense that xy = I -S, for
-1 < S < 1. We can improve the approximation y by calculating z = 2y -y2x.
Indeed, xz = x (2y - y 2x) = xy (2 -xy) _ (1- S) (1 + S) = l _ S2. From our assumption
on S, S2 is smaller than 6 in absolute value, so that z is closer than y to x-1. To calcu-
late the inverse of a polynomial, we proceed similarly, first finding a good approxima-
tion to this inverse and then correcting the error.
Suppose there exists an algorithm able to multiply two polynomials of degrees
less than or equal to n in a time in 0 (t (n)) where t(n) is strongly linear. Let p (x) be a
nonzero polynomial of degree n whose inverse we wish to calculate. Suppose for sim-
plicity that n + 1 is a power of 2 (see Problem 10.2.18 for the general case).
If n =0, the inverse p* (x) is easy to calculate. Otherwise let k = (n + 1)/2.
During the first stage of the polynomial inversion algorithm we find an approxi-
mation h (x) to p* (x) such that the degree of x In - p (x) h (x) is less than 3k -1. (Note
that the degree of x In - p (x) p* (x) can be as high as n -1 = 2k - 2.) The idea is to rid
ourselves provisionally of the k coefficients of lowest degree in the polynomial p (x) by
dividing the latter by x'. Let h (x) = x k LP (x)/x k J * . Note first that the degree of
LP(x)/xk] is n -k = k -1, so Lp(x)/xk j* = Lx2k-2/LP(x)/xk]] = LX 3k -21p (x)] by
Problem 10.2.17(iii). Let r (x) be the polynomial of degree less than n such that
Lx 3k -2/p (x) J = (x 3k-2 - r (x)) /p (x). Then we have
x2n-p(x)h(x)=x4k-2 _p(x)xk(x3k-2-r(x))/p(x)=xkr(x)
which is indeed a polynomial of degree less than 3k - 1.

During the second stage, we improve the approximation h (x) in order to obtain
p* (x) exactly. Taking into account the appropriate scaling factor, the analogy intro-
duced at the beginning of this proof suggests that we should calculate
q (x) = 2h (x) - Lh2(x)p (x)1x2n j. Let s (x) be the polynomial of degree less that 2n
such that Lh2(x)p(x)/x2n ] _ (h2(x)p(x)-s(x))/x2". Now calculate
2"
p (x) q (x) = 2p (x) h (x) _ (p 2(x) h 2(x) _ P (x) s (x)) /x
= [(P(x)h(x))(2x2n -p(x)h(x))+P(x)s(x)]Ix In
= [(x2n
_xkr(x))(xIn +xkr(x))+p(x)s(x)]/xIn
= [xa" _x2kr2(x)+P(x)s(x)]lx2"
=x2n +(P(x)s(x)-x2kr'2(x))/x2"
It remains to remark that the polynomials p (x) s (x) and x 2k r 2(x) are of degree at most
3n -1 to conclude that the degree of x 2n - p (x) q (x) is less than n , hence
q (x) = p* (x), which is what we set out to calculate.
Combining these two stages, we obtain the following recursive formula :
p*(x) = 2xk LP(x)lxkJ* - Lv(x)[Lp(x)lxk J*]2/xn-1 J .
Let g(n) be the time taken to calculate the inverse of a polynomial of degree n by the
divide-and-conquer algorithm suggested by this formula. Taking into account the
recursive evaluation of the inverse of Lp(x)/xk J, the two polynomial multiplications
that allow us to improve our approximation, the subtractions, and the multiplications
and divisions by powers of x, we see that
g(n)Eg((n -1)/2)+t((n -1)/2)+t(n)+O(n)c g((n -1)/2)+O(t(n))
because t(n) is strongly linear. Using Problem 10.2.19, we conclude that
g(n) E O (t (n))
Problem 10.2.18. Let INVP2 be the problem of calculating p* (x) when p (x)
is a polynomial of degree n such that n + 1 is a power of 2. All that the proof of
theorem 10.2.12 really shows is that INVP2 <-' MLTP. Complete the proof that
INVP S ' MLTP.
Problem 10.2.19. Prove that if g(n)Eg((n -1)/2)+O(t(n)) when n +1 is a

power of 2, and if t(n) is strongly linear, then g(n) E O (t (n) I n +1 is a power of 2).
(Hint : apply Problem 2.3.13(iv) with the change of variable T (n) = g(n -1) ; note
that it is enough to assume the existence of a real constant a> 1 such that
t (2n) ? a t (n) for all sufficiently large n - strong linearity of t (n) would unneces-
sarily impose a=2.)
Problem 10.2.20. Let p (x) = x 3 +x 2 + 5 x + 1. Calculate p* (x) using the

approach described in the proof of theorem 10.2.12. You may carry out directly the
intermediate calculation of the inverse of Lp (x)/x2J rather than doing so recursively.
Compare your answer to the one obtained as a solution to Problem 10.2.15.
* Problem 10.2.21. We saw in Section 9.4 how Fourier transforms can be used
to perform the multiplication of two polynomials of degrees not greater than n in a
time in O (n log n). Theorems 10.2.11 and 10.2.12 allow us to conclude that this time
is also sufficient to determine the quotient obtained when a polynomial of degree at
most 2n is divided by a polynomial of degree n. However, the proof of theorem
10.2.11 depends crucially on the fact that the degree of the dividend is not more than
double the degree of the divisor. Generalize this result by showing how we can divide
a polynomial of degree m by a polynomial of degree n in a time in 0 (m log n).
** Problem 10.2.22. Following the general style of theorems 10.2.9 to 10.2.12,

complete the proof of theorem 10.2.7. You will have to define the notion of
Sec. 10.3 Introduction to NP-Completeness 315
inverse for an integer : if i is an n-bit integer (that is, 2n-1 < i <- 2" -1), define
i* = L22n - '/i J. Notice that i* is also an n-bit integer, unless i is a power of 2. The
problem INV is defined on the integers in the same way as INVP on the polynomials.
The difficulties start with the fact that (i* )* is not always equal to i, contrary to
Problem 10.2.16. (For example, 13* = 9 but 9* = 14.) This hinders all the proofs.
For example, consider how we prove that DIV <- I INV. Let i be an integer of size 2n
and let j be an integer of size n ; we want to calculate Li I j J. If we define
z = Li j * / 22n - 1 ] by analogy with the calculation of L(p (x) d* (x )) /x 2n j in the proof
of theorem 10.2.11, we no longer obtain automatically the desired result z = Li / j].
Detailed analysis shows, however, that z <- Li / j J < z + 2. The exact value of Li / j J
can therefore be obtained by a correction loop that goes around at most three times.
z E- Li j* /22n I]
t - (z +1)x j
while t <- i do
t Ft +j
z z+1
return z
The other proofs have to be adapted similarly.
** Problem 10.2.23. Let SQRT be the problem of computing the largest integer
less than or equal to the square root of a given integer of size n. Prove under suitable
assumptions that SQRT=/ MLT. What assumptions do you need? (Hint: for the
reduction SQRT <-' MLT, follow the general lines of theorem 10.2.12 but use
Newton's method to find the positive zero of f (w) = w 2 -x ; for the inverse reduction,
use the fact that
x++- x l- x- x+l+ + 1=1/x.)

Problem 10.2.24. Let MOD be the problem of computing the remainder when
an integer of size 2n is divided by an integer of size n . Prove that MOD <_ 1 MLT.
** Problem 10.2.25. Let GCD be the problem of computing the greatest common
divisor of two integers of size at most n . Prove or disprove GCD = MLT. (Warning:
at the time of writing, this is an open problem.)
10.3 INTRODUCTION TO NP-COMPLETENESS
There exist many real-life, practical problems for which no efficient algorithm is
known, but whose intrinsic difficulty no one has yet managed to prove. Among these
are such different problems as the travelling salesperson (Sections 3.4.2, 5.6, and
6.6.3), optimal graph colouring (Section 3.4.1), the knapsack problem, Hamiltonian cir-
cuits (Example 10.3.2), integer programming, finding the longest simple path in a
graph (Problem 5.1.3), and the problem of satisfying a Boolean expression. (Some of
these problems are described later.) Should we blame algorithmics or complexity ?
Maybe there do in fact exist efficient algorithms for these problems. After all, com-
puter science is a relative newcomer: it is certain that new algorithmic techniques
remain to be discovered.
This section presents a remarkable result : an efficient algorithm to solve any one
of the problems we have listed in the previous paragraph would automatically provide
us with efficient algorithms for all of them. We do not know whether these problems
are easy or hard to solve, but we do know that they are all of similar complexity. The
practical importance of these problems ensured that each of them separately has been
the object of sustained efforts to find an efficient method of solution. For this reason it
is widely conjectured that such algorithms do not exist. If you have a problem to solve
and you are able to show that it is equivalent (see Definition 10.3.1) to one of those
mentioned previously, you may take this result as convincing evidence that your
problem is hard (but evidence is not a proof). At the very least you will be certain that
nobody else claims to be able to solve your problem efficiently at the moment.
10.3.1 The Classes P and NP
Before going further it will help to define what we mean by an efficient algorithm.
Does this mean it takes a time in O (n log n) ? O (n 2) ? O (n'-") ? It all depends on
the problem to be solved. A sorting algorithm taking a time in O(n2) is inefficient,
whereas an algorithm for matrix multiplication taking a time in 0 (n 2 log n) would be
an astonishing breakthrough. So we might be tempted to say that an algorithm is
efficient if it is better than the obvious straightforward algorithm, or maybe if it is the
best possible algorithm to solve our problem. But then what should we say about the
dynamic programming algorithm for the travelling salesperson problem (Section 5.6)
or the branch-and-bound algorithm (Section 6.6.3) ? Although more efficient than an
exhaustive search, in practice these algorithms are only good enough to solve instances
of moderate size. If there exists no significantly more efficient algorithm to solve this
problem, might it not be reasonable to decide that the problem is inherently intract-
able ?
For our present purposes we answer this question by stipulating that an algorithm
is efficient (or polynomial-time) if there exists a polynomial p (n) such that the algo-
rithm can solve any instance of size n in a time in 0 (p (n)). This definition is
motivated by the comparison in Section 1.6 between an algorithm that takes a time in
O(2') and one that only requires a time in O (n 3 ), and also by sections 1.7.3, 1.7.4,
and 1.7.5. An exponential-time algorithm becomes rapidly useless in practice, whereas
generally speaking a polynomial-time algorithm allows us to solve much larger
instances. The definition should, nevertheless, be taken with a grain of salt. Given
two algorithms requiring a time in O(ntg'g" ) and in O(n 10), respectively, the first, not
being polynomial, is "inefficient". However, it will beat the polynomial algorithm on
all instances of size less than 10300 assuming that the hidden constants are similar. In
fact, it is not reasonable to assert that an algorithm requiring a time in O(n 10) is
efficient in practice. Nonetheless, to decree that O(n3) is efficient whereas S2(n4) is

not, for example, seems rather too arbitrary.
In this section, it is crucial to avoid pathological algorithms and analyses such as
those suggested in Problem 1.5.1. Hence no algorithm is allowed to perform arith-
metic operations at unit cost on operands whose size exceeds some fixed polynomial in
the size of the instance being solved. (The polynomial may depend on the algorithm
but not of course on the instance.) If the algorithm needs larger operands (as would be
the case in the solution of Problem 1.5.1), it must break them into sections, keep them
in an array, and spend the required time to carry out multiprecision arithmetic.
Without loss of generality, we also restrict all arrays to contain a number of elements
at most polynomial in the size of the instance considered.
The notion of linear reduction and of linear equivalence considered in Section
10.2 is interesting for problems that can be solved in quadratic or cubic time. It is,
however, too restrictive when we consider problems for which the best-known algo-
rithms take exponential time. For this reason we introduce a different kind of reduc-
tion.
Definition 10.3.1. Let X and Y be two problems. Problem X is polynomially

reducible to problem Y in the sense of Turing, denoted X <_ T Y, if there exists an algo-
rithm for solving X in a time that would be polynomial if we took no account of the
time needed to solve arbitrary instances of problem Y. In other words, the algorithm
for solving problem X may make whatever use it chooses of an imaginary procedure
that can somehow magically solve problem Y at no cost. When X <_T Y and Y <_T X
simultaneously, then X and Y are polynomially equivalent in the sense of Turing,
denoted X -T Y. (This notion applies in a natural way to unsolvable problems - see
Problem 10.3.32.)
Example 10.3.1. Let SMALLFACT(n) be the problem of finding the smallest

integer x > 2 such that x divides n (for n >_ 2), let PRIME(n) be the problem of deter-
mining whether n >_ 2 is a prime number, and let NBFACT(n) be the problem of counting
the number of distinct primes that divide n. Then both PRIME <_ T SMALLFACT and
NBFACT <_ T SMALLFACT. Indeed, imagine solutions to the problem SMALLFACT can be
obtained at no cost by a call on SolveSF ; then the following procedures solve the
other two problems in a time polynomial in the size of their operand.
function DecidePRiME(n)
we assume n _ 2}
if n = SolveSF (n) then return true
else return false
function SolveNBF (n)
nb - 0
while n > I do nh - nh + I
x - SolveSF (n)
while x divides n don f- n /x
return nh
Notice that SolveNBF works in polynomial time (counting calls of SolveSF at no cost)
because no integer n can have more than Llg n ] prime factors, even taking repetitions
into account.
The usefulness of this definition is brought out by the two following exercises.
Problem 10.3.1. Let X and Y be two problems such that X <_ T Y Y. Suppose
there exists an algorithm that is able to solve problem Y in a time in O (t(n)), where
t(n) is a nonzero, nondecreasing function. Prove that there exist a polynomial p (n)
and an algorithm that is able to solve problem X in a time in O (p (n) t(p (n))).
Problem 10.3.2. Let X and Y be two problems such that X <T Y. Prove that
the existence of an algorithm to solve problem Y in polynomial time implies that there
also exists a polynomial-time algorithm to solve problem X.
In particular, the equivalence mentioned in the introduction to this section

implies that either all the problems listed there can be solved in polynomial time, or
none of them can.
For technical reasons we confine ourselves from now on to the study of decision
problems. For example, "Is n a prime number?" is a decision problem, whereas "find
the smallest prime factor of n " is not. A decision problem can be thought of as
defining a subset X of the set I of all its instances. Then the problem consists of
deciding, given some x e 1, whether or not x EX. We generally assume that the set of
all instances is easy to recognize, such as IN, or "the set of all possible graphs". We
also assume that the instances can be coded efficiently in the form of strings of bits.
When no confusion can arise, we may sometimes omit to state explicitly the set of
instances for the decision problem under consideration.
Definition 10.3.2. P is the class of decision problems that can be solved by a

polynomial-time algorithm. 0
Problem 10.3.3. Let X and Y be two decision problems. Prove that if X <_ T Y
and Y E P, then X E P.
The restriction to decision problems allows us to introduce a simplified notion of

polynomial reduction.
Definition 10.3.3. Let X c I and Y c J be two decision problems. Problem

X is many-one polynomially reducible to problem Y, denoted by X <m Y, if there
exists a function f :I - J computable in polynomial time, known as the reduction
function between X and Y, such that
(VXEI)[xEX f (x)EY] .
When X <_,° Y and Y <_ m X both hold, then X and Y are many-one polynomially
equivalent, denoted X = m Y.
Example 10.3.2. Let TSPD and HAM be the travelling salesperson decision
problem and the Hamiltonian circuit problem, respectively. An instance of TSPD con-
sists of a directed graph with costs on the edges, together with some bound L used to
turn the travelling salesperson optimization problem (as in Sections 5.6 and 6.6.3) into
a decision problem : the question is to decide whether there exists a tour in the graph
that begins and ends at some node, after having visited each of the other nodes exactly
once, and whose cost does not exceed L. An instance of HAM is a directed graph, and
the question is to decide whether there exists a circuit in the graph passing exactly
once through each node n (with no optimality constraint).
To prove that HAM 5 m TSPD, let G = < N, A > be a directed graph for which you
would like to decide if it has a Hamiltonian circuit. Define f (G) as the instance for
TSPD consisting of the complete graph H = <N, NxN >, the cost function
I if (u,v)EA
c(u , v) _
2 otherwise
and the bound L = #N, the number of nodes in G. Clearly, G E HAM if and only if
< H, c, L > E TSPD. It is also the case that TSPD m HAM, but this is significantly
harder to prove.
Lemma 10.3.1. If X and Y are two decision problems such that X <_m Y, then
xSTY.
Proof. Imagine solutions to problem Y can be obtained at no cost by a call on

DecideY and let f be the polynomial-time computable reduction function between X
and Y. Then the following procedure solves X in polynomial time.
function DecideX (x)

y -f (X)
if DecideY (y) then return true
else return false
* Problem 10.3.4. Prove that the converse of Lemma 10.3.1 does not neces-
sarily hold by giving explicitly two decision problems X and Y for which you can
prove that X <_ T Y whereas it is not the case that X <_ m Y. El
Problem 10.3.5. Prove that the relations <_ T , <_ m , T , and = R, are transi-
tive.
The introduction of TSPD in Example 10.3.2 shows that the restriction to decision
problems is not a severe constraint. In fact, most optimization problems are polynomi-
ally equivalent in the sense of Turing to an analogous decision problem, as the fol-
lowing exercise illustrates.
* Problem 10.3.6. Let G = < N, A > be an undirected graph, let k be an

integer, and c : N -* (1, 2, ... , k } a function. This function is a valid colouring of G
if there do not exist nodes u , v E N such that (u , v ) E A and c (u) = c (v) (Section
3.4.1). The graph G can be coloured with k colours if there exists such a valid
colouring. The smallest integer k such that G can be coloured with k colours is called
the chromatic number of G, and in this case a colouring with k colours is called an
optimal colouring. Consider the three following problems.
COLD: Given a graph G and an integer k, can G be coloured with k colours ?

COLO: Given a graph G, find the chromatic number of G.
COLC: Given a graph G, find an optimal colouring of G.
Prove that COLD =T COLO =T COLC. Conclude that there exists a polynomial-time
algorithm to determine the chromatic number of a graph, and even to find an optimal
colouring, if and only if COLD E P.
These graph colouring problems have the characteristic that although it is perhaps
difficult to decide whether or not a graph can be coloured with a given number of
colours, it is easy to check whether a suggested colouring is valid.
Definition 10.3.4. Let X be a decision problem. Let Q be a set, arbitrary for

the time being, which we call the proof space for X. A proof system for X is a subset
FcXxQ such that (`dxEX)(3gEQ)[<x,q>EF]. Any q such that <x,q >EF
is known as a proof or a certificate that x E X. Intuitively, each true statement of the
type x E X has a proof in F, whereas no false statement of this type has one (because if
x E X, there does not exist a q E Q such that < x, q > (=- F ).
Example 10.3.3. Let I = N and COMP = { n I n is a composite integer }. We

can take Q = 1N as the proof space and F = ( < n, q > I 1 < q < n and q divides n
exactly } as the proof system. Notice that some problems may have more than one
natural proof system. In this example we could also use the ideas of Section 8.6.2 to
define
F' < n, q > I (n is even and n > 2) or
(1 < q < n and n is not a strong pseudoprime to the base q
which offers a large number of proofs for all odd composite numbers.
Example 10.3.4. Consider the set of instances I = { < G , k > I G is an

undirected graph and k is an integer } and the problem COLD = { < G, k > E 1 I G
can be coloured with k colours }. As proof space we may take Q =
(c : N -* (1, 2, ... , k} I N is a set of nodes and k is an integer). Then a proof
system is given by
F < < G, k >, c > I G = < N, A > is an undirected graph,

k is an integer,
c : N -4 { 1, 2, ... , k } is a function and
(`du,vEN)[{u,v}EA =c(u)#c(v)] } .
Problem 10.3.7. Let G = < N, A > be an undirected graph. A clique in G is

a set of nodes K E_ N such that { u, v } E A for every pair of nodes u, v K. Given a
graph G and an integer k, the CLIQUE problem consists of determining whether there
exists a clique of k nodes in G. Give a proof space and a proof system for this deci-
sion problem.
Definition 10.3.5. NP is the class of decision problems for which there exists
a proof system such that the proofs are succinct and easy to check. More precisely, a
decision problem X is in NP if and only if there exist a proof space Q, a proof system
F c X x Q, and a polynomial p (n) such that
i. (VxEX)(agEQ)[<x,q>eFand IqI <p(Ix1)],

where I q I and I x I denote the sizes of q and x, respectively ; and
ii. F E P.
We do not require that there should exist an efficient way to find a proof of x when
x c X, only that there should exist an efficient way to check the validity of a proposed
short proof.
Example 10.3.5. The conceptual distinction between P and NP is best grasped

with an example. Let comp be the problem in Example 10.3.3. In order to have
comp E P, we would need an algorithm
function DecideCOMP (n)
{ decides whether n is a composite number or not }
return true
return false
whose running time is polynomial in the size of n. No such algorithm is currently

known. However, to show that coMP E NP, we need only exhibit the following
(obvious) polynomial-time algorithm.
function VerifyCOMP (n, q)

if 1 < q < n and q divides n then return true
else return false
By definition of NP, any run of VerifyCOMP (n, q) that returns true is a proof that n is
composite, and every composite number has at least one such proof (but prime
numbers have none). However, the situation is not the same as for a probabilistic
algorithm (Chapter 8) : we are content even if there exist very few q (for some
composite n) such that VerifyCOMP (n , q) is true and if our chance of hitting one at
random would be staggeringly low.
Problem 10.3.8. Let X be a decision problem for which there exists a

polynomial-time true-biased Monte Carlo algorithm (section 8.6). Prove that X E NP.
(Hint : the proof space is the set of all sequences of random decisions possibly taken
by the Monte Carlo algorithm.)
Problem 10.3.9. Prove that P E_ NP. (Hint : Let X be a decision problem

in P. It suffices to take Q = { 0 } and F = (< x, 0 > I x e X } to obtain a system of
"proofs" that are succinct and easy to check. This example provides an extreme illus-
tration of the fact that the same proof may serve for more than one instance of the
same problem.)
Example 10.3.6. The problems COLD and CLIQUE considered in example

10.3.4 and Problem 10.3.7 are in NP.
Although COLD is in NP and COLO =T COLD, it does not appear that NP contains
the problem of deciding, given a graph G and an integer k, whether k is the chromatic
number of G. Indeed, although it suffices to exhibit a valid colouring to prove that a
graph can be coloured with a given number of colours (Example 10.3.4), no one has
yet been able to invent an efficient proof system to demonstrate that a graph cannot be
coloured with less than k colours.
Definition 10.3.6. Let X S I be a decision problem. Its complementary

problem consists of answering "Yes" for an instance x E I if and only if x 0 X. The
class co-NP is the class of decision problems whose complementary problem is in NP.
For instance, the preceding remark indicates that we do not know whether COLD E
co-NP. Nonetheless, we know that COLD E co-NP if and only if NP = co-NP (Prob-
lems 10.3.27 and 10.3.16). The current conjecture is that NP # co-NP, and therefore
that COLD co-NP.
Problem 10.3.10. Let A and B be two decision problems. Prove that if

A 5 m B and B E NP, then A E NP.
Problem 10.3.11. Let A and B be two decision problems. Do you believe that
if A 5 T B and B E NP, then A E NP ? Justify your answer.
Problem 10.3.12. Show that HAM, the Hamiltonian circuit problem defined in
Example 10.3.2, is in NP.
Example 10.3.7. In 1903, two centuries after Mersenne claimed without proof
that is a prime number, Frank Cole showed that
267_ 1
267 -1 = 193,707,721 x 761,838,257, 287 .
It took him "three years of Sundays" to discover this factorization. He was lucky that
the number he chose to attack is indeed composite, since this enabled him to offer a
proof of his result that is both short and easy to check. (This was not all luck : Lucas
had already shown in the nineteenth century that 267-1 is composite, but without
finding the factors.)
The story would have had quite a different ending if this number had been prime.
In this case the only "proof" of his discovery that Cole would have been able to pro-
duce would have been a thick bundle of papers covered in calculations. The proof
would be far too long to have any practical value, since it would take just as long to
check as it did to produce in the first place. (A similar argument may be advanced
concerning the "proof' by computer of the famous four colour theorem.) This results
from a phenomenon like the one mentioned in connection with the chromatic number
of a graph : the problem of recognizing composite numbers is in NP (Example 10.3.3),
but it seems certain at first sight not to be in co-NP, that is, the complementary
problem of recognizing prime numbers seems not to be in NP.
However, nothing is certain in this world except death and taxes : this problem
too is in NP, although the notion of a proof (or certificate) of primality is rather more
subtle than that of a proof of nonprimality. A result from the theory of numbers shows
that n, an odd integer greater than 2, is prime if and only if there exists an integer x
such that
0<x <n
x"- I = 1 (mod n ), and
xI" - I)lP 4P 1 (mod n) for each prime factor p of n -1
A proof of primality for n therefore consists of a suitable x, the decomposition of n -1

into prime factors, and a collection of (recursive) proofs that each of these factors is
indeed prime. (More succinct proof systems are known.)
* Problem 10.3.13. Complete the proof sketched in Example 10.3.7 that the
problem of primality is in NP. It remains to show that the length of a recursive proof
of primality is bounded above by a polynomial in the size (that is, the logarithm) of the
integer n concerned, and that the validity of such a proof can be checked in polynomial
time.
Problem 10.3.14. Let F = { < x , y > I x , y E N and x has a prime factor less
than y }. Let FACT be the problem of decomposing an integer into prime factors.
Prove that
i. F E NP n co-NP; and
ii. F = T FACT.
If we accept the conjecture that no polynomial-time factorization algorithm exists, we

can therefore conclude that F E (NP n co-NP) 1 P. 11
10.3.1 NP-Complete Problems
The fundamental question concerning the classes P and NP is whether the inclusion
P9 NP is strict. Does there exist a problem that allows an efficient proof system but
for which it is inherently difficult to discover such proofs in the worst case ? Our intui-
tion and experience lead us to believe that it is generally more difficult to discover a
proof than to check it: progress in mathematics would be much faster were this not so.
In our context this intuition translates into the hypothesis that P # NP. It is a cause of
considerable chagrin to workers in the theory of complexity that they can neither prove
nor disprove this hypothesis. If indeed there exists a simple proof that P # NP, it has
certainly not been easy to find !
On the other hand, one of the great successes of this theory is the demonstration
that there exist a large number of practical problems in NP such that if any one of
them were in P then NP would be equal to P. The evidence that supports the
hypothesis P # NP therefore also lends credence to the view that none of these prob-
lems can be solved by a polynomial-time algorithm in the worst case. Such problems
are called NP-complete.
Definition 10.3.7. A decision problem X is NP-complete if
I. X E NP ; and
ii. for every problem Y E NP, Y S T X.
Some authors replace the second condition by Y n, X or by other (usually stronger)

kinds of reduction.
Problem 10.3.15. Prove that there exists an NP-complete problem X such that
X E P if and only if P = NP.
* Problem 10.3.16. Prove that if there exists an NP-complete problem X such

that X E co-NP, then NP = co-NP.
Problem 10.3.17. Prove that if the problem X is NP-complete and the

problem Z E NP, then
i. Z is NP-complete if and only if X <T Z ;

ii. if X n, Z, then Z is NP-complete.
Be sure to work this important problem. It provides the fundamental tool for
proving NP-completeness. Suppose we have a pool of problems that have already
been shown to be NP-complete. To prove that Z is NP-complete, we can choose an
appropriate problem X from the pool and show that X is polynomially reducible to Z
(either many-one or in the sense of Turing). We must also show that Z E NP by exhi-
biting an efficient proof system for Z. Several thousand NP-complete problems have
been enumerated in this way.
This is all well and good once the process is under way, since the more problems
there are in the pool, the more likely it is that we can find one that can be reduced
without too much difficulty to some new problem. The trick, of course, is to get the
ball rolling. What should we do at the outset when the pool is empty to prove for the
very first time that some particular problem is NP-complete? (Problem 10.3.17 is then
powerless.) This is the tour de force that Steve Cook managed to perform in 1971,
opening the way to the whole theory of NP-completeness. (A similar theorem was
discovered independently by Leonid Levin.)
10.3.2 Cook's Theorem
Definition 10.3.8. A Boolean variable takes its values in the set

B = { true, false }. Boolean variables are combined using logical operators (not, and,
or, t:* , r , and so on) and parentheses to form Boolean expressions. It is customary
to represent disjunction (or) in such expressions by the symbol "+" and conjunction
(and) by simply juxtaposing the operands (as for arithmetic multiplication). Negation
is often denoted by a horizontal bar above the variable or the expression concerned. A
Boolean expression is satisfiable if there exists at least one way of assigning values to
its variables so as to make the expression true. A Boolean expression is a tautology if
it remains true whatever values are assigned to its variables. A Boolean expression is a
contradiction if it is not satisfiable, that is, if its negation is a tautology. We denote by
SAT, TAUT and CONT, respectively, the problems of deciding, given a Boolean
expression, whether it is satisfiable, whether it is a tautology, and whether it is a con-
tradiction. 0
Example 10.3.8. Here are three Boolean expressions using the Boolean vari-
ables p and q .
i. (p+q)=pq
ii. (p b q) (p +q)(p +q )
iii. p (p +q)q
Expression (i) is satisfiable because it is true if p = true and q = true, but it is

not a tautology because it is false if p = true and q = false . Verify that expression
(ii) is a tautology and that expression (iii) is a contradiction.
To prove that a Boolean expression is satisfiable, it suffices to produce an assign-

ment that satisfies it. Moreover, such a proof is easy to check. This shows that
SATE NP. It is not apparent that the same is true of the two other problems : what
short and easy proof can one give in general of the fact that a Boolean expression is a
tautology or a contradiction? These three problems are, nevertheless, polynomially
equivalent in the sense of Turing.
Problem 10.3.18. Prove that
i. SAT = T TAUT - T CONT ; and even

ii. TAUT -m CONT.
It is possible in principle to decide whether a Boolean expression is satisfiable by

working out its value for every possible assignment to its Boolean variables. However
this approach is impractical when the number n of Boolean variables involved is large,
since there are 2" possible assignments. No efficient algorithm to solve this problem is
known.
Definition 10.3.9. A literal is either a Boolean variable or its negation.

A clause is a literal or a disjunction of literals. A Boolean expression is in conjunctive
normal form (CNF) if it is a clause or a conjunction of clauses. It is in k-CNF for
some positive integer k if it is composed of clauses, each of which contains at most k
literals (some authors say : exactly k literals).
Example 10.3.9. Consider the following expressions.
L (p +q +r)(p +q +r)q r
ii. (p +qr)(p +q (p +r))
W. (p =* q)b(p+q)
Expression (i) is composed of four clauses. It is in 3-CNF (and therefore in CNF), but
not in 2-CNF. Expression (ii) is not in CNF since neither p +qr nor p +q (p +r) is a
clause. Expression (iii) is also not in CNF since it contains operators other than con-
junction, disjunction and negation.
*Problem 10.3.19.
i. Show that to every Boolean expression there corresponds an equivalent expres-

sion in CNF.
ii. Show on the other hand that the shortest equivalent expression in CNF can be
exponentially longer than the original Boolean expression.
Definition 10.3.10. SAT-CNF is the restriction of the SAT problem to

Boolean expressions in CNF. For any positive integer k, SAT-k-CNF is the restriction
of SAT-CNF to Boolean expressions in k-CNF. The problems TAUT-(k-)CNF and
CONT-(k-)CNF are defined similarly. 11
Clearly, all these problems are in NP or in co-NP. Polynomial-time algorithms

are known for a few of them.
* Problem 10.3.20. Prove that SAT-2-CNF and TAUT-CNF can be solved in

polynomial time. El
The interest of Boolean expressions in the context of NP-completeness arises

from their ability to simulate algorithms. Consider an arbitrary decision problem that
can be solved by a polynomial-time algorithm A. Suppose that the size of the
instances is measured in bits. For every integer n there exists a Boolean expression
'P (A) in CNF that can be obtained efficiently (in a time polynomial in n, where the
polynomial may depend on the algorithm A ; the size of 'Y (A) is also polynomial
in n). This Boolean expression contains a large number of variables, among which
x i , X2, -- x correspond in a natural way to the bits of instances of size n for A.
The Boolean expression is constructed so that there exists a way to satisfy it by
choosing the values of its other Boolean variables if and only if algorithm A accepts
the instance corresponding to the Boolean value of the x variables. For example, algo-
rithm A accepts the instance 10010 if and only if the expression xix2x3xax5`P5(A) is
satisfiable. More interestingly, algorithm A accepts at least one instance of size 5 if
and only if T5(A) is satisfiable.
The proof that this Boolean expression exists and that it can be constructed
efficiently poses difficult technical problems. It usually requires a formal model of
computation beyond the scope of this book, such as the Turing machine. We content
ourselves with mentioning that the expression 'P (A) contains among other things a
distinct Boolean variable b,, for each bit i of memory that algorithm A may need to use
when solving an instance of size n, and for each unit t of time taken by this computa-
tion. Once the variables x I , x2 2 ,--- x are fixed, the clauses of 'P (A) force these
Boolean variables to simulate the step-by-step execution of the algorithm on the
corresponding instance. (The number of additional Boolean variables is polynomial in
the size of the instance because the algorithm runs in polynomial time and because we
have assumed without loss of generality that none of its variables or arrays can ever
occupy more than a polynomial number of bits of memory.)
We are finally in a position to state and prove the fundamental theorem of the
theory of NP-completeness.
Theorem 10.3.1. SAT-CNF is NP-complete.
Proof. We already know that SAT-CNF is in NP. Thus it remains to prove that
X <_ T SAT-CNF for every problem X E NP. Let Q be a proof space and F an efficient
proof system for X. Let p (n) be the polynomial (given by the definition of NP) such
that to every x E X there corresponds a q E Q whose length is bounded above by

p (I x I) such that < x, q > E F. Let A be a polynomial-time algorithm able to decide,
given < x, q > as input, whether q is a proof that x E X (that is, whether < x, q >
belongs to F ). That such an algorithm exists is ensured by the fact that F E P, which
is part of the definition of X E NP. For each x, let Ax be an algorithm whose purpose
is to verify whether a given q E Q is a valid proof that x E X.
function Ax. (q )
if A (< x, q >) then return true
else return false
Here finally is an algorithm to solve problem X in polynomial time, if we imagine that
answers concerning the satisfiability of certain Boolean expressions in CNF can be
obtained at no cost by a call on DecideSATCNF.
function DecideX (x)
let n be the size of x (in bits)
for i - O to p (n) do
if DecideSATCNF (Ti (Ax)) then return true
return false
Let X E X be of size n, and let q e Q be a proof of size i that x E X, where
0 <_ i <_ p (n). The fact that algorithm A accepts < x, q > implies that algorithm Ax
accepts q. This is an instance of size i, hence the Boolean expression Ti (Ax) is
satisfiable, as DecideX (x) will discover. Conversely, if x 0 X, then there exists no
q e Q such that < x, q > E F, and therefore Ax accepts no inputs, which implies that
DecideX (x) will find no integer i such that Ti (Ax) is satisfiable. This completes the
proof that X <_ T SAT-CNF. (To be precise, one technical detail is worth mentioning :
to each algorithm A, there corresponds a two variable polynomial r such that the
Boolean formula Ti (Ax) can be constructed in a time in 0 (r (I x I, i )) and such that its
size is bounded similarly.)
Problem 10.3.21. Prove that in fact X <_ m SAT-CNF for any decision problem
XENP.
10.3.3 Some Reductions
We have just seen that SAT-CNF is NP-complete. Let X E NP be some other decision
problem. To show that X too is NP-complete, we need only prove that
SAT CNF <_ T X (Problem 10.3.17). Thereafter, to show that Y E NP is NP-complete,
we have the choice of proving SAT-CNF <_ T Y or X <_ T Y. We illustrate this principle
with several examples.
Example 10.3.10. SAT is NP-complete.

We already know that SAT E NP. It therefore remains to show that
SAT-CNF <_ T SAT. The reduction is immediate because Boolean expressions in CNF
are simply a special case of general Boolean expressions. More precisely, if we ima-
gine that the satisfiability of Boolean expressions can be decided at no cost by a call on
DecideSAT, here is a polynomial-time algorithm for solving SAT-CNF.
function DecideSATCNF(`Y)
if P is not in CNF then return false
if DecideSAT (P) then return true
else return false
Problem 10.3.22. Prove that SAT <_ T SAT-CNF. Using Example 10.3.10,
conclude that SAT= T SAT-CNF. (Hint : This problem has a very simple solution.
However, resist the temptation to use Problem 10.3.19(i) to obtain the algorithm
function DecideSAT(Y)
let E be a Boolean expression in CNF equivalent to Y
if DecideSATCNF (E,) then return true
else return false
because Problem 10.3.19(ii) shows that expression i; can be exponentially longer than
expression T, so it cannot be computed in polynomial time in the worst case.)
Example 10.3.11. SAT-3-CNF is NP-complete.

We have already seen that SAT-3-CNF is in NP. It remains to show that
SAT CNF <_ T SAT-3-CNF. This time let us show that SAT-CNF <_ m SAT-3-CNF.
Let P be a Boolean expression in CNF. Our problem is to construct efficiently a
Boolean expression i; in 3-CNF that is satisfiable if and only if P is satisfiable. Con-
sider first how to proceed if `P contains only one clause, which is therefore a disjunc-
tion of k literals.
i. If k 3, set = P, which is already in 3-CNF.

ii. If k = 4, let l , 12, l3 , and 14 be the literals such that `P is l 1 + 12 + 13 + l4 . Let u
1
be a new Boolean variable. Take = (l 1 + 12 + u) (i + 13 +14)-

iii. More generally, if k >_ 4, let 1 , 12, ... , lk be the literals such that `P is
1
l 1 +12+ +lk . Let u 1 , u2 , ... , uk_3 be new Boolean variables. Take

=(11+12+u1)(u1+13+u2)(u2+14+u3) (uk_3+lk-1+lk).
If the expression `P consists of several clauses, treat each of them independently
(using different new variables for each clause) and form the conjunction of all the
expressions in 3-CNF thus obtained.
Example 10.3.12. (Continuation of Example 10.3.11) If

`P=(p+q+r+s)(r+s)(p+s+x+v+u)
we obtain
E=(p+q+u1)(u1+r+s)('+s)(p+s+u2)(u2+x+u3)(u3+v+w) .
Problem 10.3.23. Prove that is satisfiable if and only if P is satisfiable in

the construction of Example 10.3.11.
Problem 10.3.24. Prove that SAT-CNF <_ m SAT-3-CNF still holds even if in
the definition of SAT-3-CNF we insist that each clause should contain exactly three
literals.
Example 10.3.13. 3-COL is NP-complete.

Let G be an undirected graph and k an integer constant. The problem k-COL
consists of determining whether G can be coloured with k colours (see Problem
10.3.6). It is easy to see that 3-COL E NP. To show that 3-COL is NP-complete, we
shall prove this time that SAT-3-CNF <_ m 3-COL. Given a Boolean expression `P in
3-CNF, we have to construct efficiently a graph G that can be coloured in three colours
if and only if T is satisfiable. This reduction is considerably more complex than those
we have seen so far.
Suppose for simplicity that every clause of the expression `P contains exactly
three literals (see Problem 10.3.24). Let k be the number of clauses in T. Suppose
further without loss of generality that the Boolean variables appearing in `P are x t ,
x2 2 ,--- x, . The graph G that we are about to build contains 3 + 2t + 6k nodes and
3 + 3t + 12k edges. Three distinguished nodes of this graph are linked in a triangle :
call them T, F, and C. When the time comes to colour G in three colours, imagine
that the colours assigned to T and F represent the Boolean values true and false,
respectively. The colour used for node C will be a control colour. See Figure 10.3.1.
For each Boolean variable x, of `P the graph contains two nodes yj and zi that
are linked to each other and to the control node C. In any colouring of G in three
colours, this forces yj to be the same colour as either T or F and zi to be the comple-
mentary colour. If y; is the same colour as T, think of this intuitively as an assign-
ment of the value true to the Boolean variable x, . Contrariwise, xi is false if it is zi
that is the same colour as T. In every case node y; corresponds to the literal x, and
node zi corresponds to the literal x; . If t = 3, for example, Figure 10.3.2 shows the
part of the graph that we have constructed up to now.
We still have to add 6 nodes and 12 edges for each clause in T. These are added
in such a way that the graph will be colourable with three colours if and only if the
choice of colours f o r y l , Y2,-- y, corresponds to an assignment of Boolean values
to x t , X2 , -- x, that satisfies every clause. This can be accomplished thanks to the
widget illustrated in Figure 10.3.3. One copy of this widget is added to the graph for
each clause in T.
Figure 10.3.1. The control triangle.

Figure 10.3.2. Graph representation of three Boolean variables.
Figure 10.3.3. A widget.
Each widget is linked to five other nodes of the graph : nodes C and T of the
control triangle, and three nodes chosen from the y; and z, so as to correspond to the
three literals of the clause concerned. Because these input nodes 1, 2 and 3 cannot be
the same colour as C, Problem 10.3.25 shows that the widget can be coloured with the
colours assigned to C, T, and F if and only if at least one of the nodes 1, 2, and 3 is
coloured with the same colour as node T. In other words, since the colour assigned to
node T represents true, the widget simulates the disjunction of the three literals
represented by the nodes to which it is joined.
This ends the description of the graph G, which can be coloured with three
colours if and only if ì is satisfiable. It is clear that the graph can be constructed
efficiently starting from the Boolean expression Ì' in 3-CNF. We conclude that
SAT-3-CNF <_ m 3-COL, and therefore that 3-COL is NP-complete.
Problem 10.3.25. Verify that the colours attributed to nodes C, T, and F

suffice to colour the widget of Figure 10.3.3 if and only if at least one of the input
nodes is coloured with the same colour as node T (knowing that the input nodes cannot
be coloured with the same colour as node C).
Problem 10.3.26. Give a simple proof that 4-COL is NP-complete.
Problem 10.3.27. Prove that COLD (Problem 10.3.6) is NP-complete.
** Problem 10.3.28. Prove that 3-COL is still NP-complete even if we restrict

ourselves to planar graphs of degree not greater than 4.
* Problem 10.3.29. Show, on the other hand, that 2-COL is in P.
* Problem 10.3.30. Prove that CLIQUE (Problem 10.3.7) is NP-complete.

(Hint : prove that SAT-3-CNF <_ m CLIQUE.)
** Problem 10.3.31. Prove that the problem of the Hamiltonian circuit (Example
10.3.2 and Problem 10.3.12) is NP-complete.
* Problem 10.3.32. The halting problem consists of deciding, given any pro-
gram as instance, whether the latter will ever halt when started. The notion of reduci-
bility extends in a natural way to unsolvable problems such as the halting problem
(although it is usual to drop the polynomial-time aspect of the reductions - which we
do not do here). A function f : IN -* IN is polynomially bounded if the size of its value
on any argument is polynomially bounded by the size of its argument. Prove that the
problem of computing any polynomially bounded computable function is polynomially
reducible to the halting problem in the sense of Turing. Prove however that there exist
decision problems that are not polynomially reducible to the halting problem in the
sense of Turing.
10.3.4 Non-determinism
The class NP is usually defined quite differently, although the definitions are
equivalent. The classic definition involves the notion of non-deterministic algorithms,
which we only sketch here. The name NP arose from this other definition: it
represents the class of problems that can be solved by a Non-deterministic algorithm in
Polynomial time.
On the surface, a non-deterministic algorithm resembles a Las Vegas proba-

bilistic algorithm (Section 8.5). To solve instance x of some problem X, we call
ND (x , y , success), where y and success are return parameters. If the algorithm sets
success to true, then y is a correct solution to instance x ; otherwise no solution is
obtained. The difference between non-deterministic and Las Vegas algorithms is that
for the former we do not set any bound on the probability of success. It is even allow-
able that for some instances the algorithm will never set success to true. For this
reason non-deterministic algorithms are only a mathematical abstraction that cannot be
used directly in practice : we would not program such an algorithm in the hope of run-
ning it successfully and efficiently on a computer. (This explains why, although non-
deterministic algorithms can solve NP-complete problems in polynomial time, this
does not imply that P = NP.)
To avoid confusion with probabilistic algorithms, we do not denote non-
deterministic choices by calls on uniform (i .. j) as in Chapter 8, but instead, we use a
special instruction
choose n between i and j ,
whose effect is to set n to some value between i and j inclusive. The actual value
assigned to n is not specified by the algorithm, nor is it subject to the laws of proba-
bility. The effect of the algorithm is determined by the existence or the nonexistence
of sequences of non-deterministic choices that lead to the production of a result. We
are not concerned with how such sequences could be determined efficiently or how
their nonexistence could be established. For simplicity, we write
return success - bool
as an abbreviation for
success - bool
return .
Definition 10.3.11. A computation of the algorithm is a sequence of non-

deterministic choices that leads to setting success to true, hence to returning a solution.
The domain of the algorithm is the set of instances on which it has at least one compu-
tation. The algorithm is total if its domain is the set of all possible instances. The
algorithm is consistent if two different computations on any given instance always lead
to the same result. When a non-deterministic algorithm is consistent, it computes a
well-defined function on the instances of its domain.
The time taken by a non-deterministic algorithm on a given instance of its
domain is defined as the shortest possible time that any computation can cause it to run
on this instance ; the time is undefined on instances that are not in the domain of the
algorithm. A non-deterministic algorithm runs in polynomial time if the time it takes
is bounded by some polynomial in the size of its instance provided that the instance is
in the domain of the algorithm. Notice that there is no limit on how long a
polynomial-time non-deterministic algorithm can run if the "wrong" non-deterministic
choices are made. It is even possible for a computation to be arbitrarily long, provided
that the same instance also admits at least one polynomially bounded computation.
Example 10.3.14. Consider the following total consistent non-deterministic

primality testing algorithm (recall from Section 4.8 that dexpo (x , i , n) computes
x' mod n efficiently).
procedure primeND (n , var prime, var success)
{ non-deterministically determines whether n is prime }
if n < 3 or n is even then prime F- (2<- n < 3)
return success f- true
choose guess between 0 and 1
if guess = 0
then { the guess is that n is composite - let's guess a proof ! }
prime - false
choose m between 2 and n -1
if m divides n then success - true
else success - false
else { the guess is that n is prime - let's guess a proof ! }
prime - true
choose x between 1 and n -1
{ the guess is that x is as in Example 10.3.7 }
if dexpo (x, n - 1, n) # 1 then return success f- false
m f-n -1
while m > 1 do
choose p between 2 and m
{ the guess is that p is a new prime divisor of n -1 }
primeND (p, pr, suc)
if suc and pr and p divides m and dexpo (x, (n -1)/p , n) 1
then while p divides m do m - m /p

else return success f- false
success f- true
When n is an odd composite number, n >- 9, the algorithm has a computation that
consists of choosing guess =0 and non-deterministically setting m to some nontrivial
factor of n. When n is prime, n >- 5, the algorithm also has a computation that consists
of choosing guess = 1, choosing x in accordance with the theorem mentioned in
example 10.3.7, and guessing successively each prime divisor of n -1. Clearly, the
algorithm also has a computation when n <- 3 or when n is even. The algorithm
primeND is therefore total. It is also consistent by the same theorem from Example
10.3.7. Notice again that it would be pointless to attempt implementing this algorithm
on the computer by replacing the choose instructions by random choices in the same
interval : the probability of success would be infinitesimal.
* Problem 10.3.33. Prove that the non-deterministic algorithm primeND of

Example 10.3.14 runs in a time polynomial in the size of its instance. Show, however,
that there are sequences of non-deterministic choices that can cause the algorithm to
run for a time exponential in the size of its instance.
* Problem 10.3.34. Prove that a decision problem can be solved by a total con-
sistent polynomial-time non-deterministic algorithm if and only if it belongs to
NP c co-NP.
Theorem 10.3.2 Every decision problem in NP is the domain of some

polynomial-time non-deterministic algorithm. (The algorithm can easily be made con-
sistent but it will not in general be total.)
Proof. Let X E NP be a decision problem, let Q be its proof space, let
F c X x Q be its efficient proof system, and let p be the polynomial that bounds the
length of a proof as a function of the length of the corresponding positive instance.
Assume for simplicity that Q is the set of binary strings. The following polynomial-
time non-deterministic algorithm has a computation on instance x if and only if x E X .
procedure XND (x, var ans, var success)

n - size of x
choose l between 0 and p (n)
q - empty binary string
for i - 1 to l do
choose b between 0 and 1
append bit b to the right of q
if < x, q > E F then ans - true, success - true
else success - false 11
Problem 10.3.35 Prove the converse of Theorem 10.3.2: whenever a decision

problem is the domain of some polynomial-time non-deterministic algorithm, this
problem belongs to NP. (Hint: use the set of all possible computations as proof
space.)
The preceding theorem and problem suggest the alternative (and more usual)
definition for NP: it is the class of decision problems that are the domain of some
polynomial-time non-deterministic algorithm. In this case, we are only concerned with
the existence or nonexistence of computations (usually called accepting computations
in this context) ; the actual result returned by the algorithm in case of success is
irrelevant, and the corresponding parameter may be ignored altogether (there is no
point in algorithm XND setting ans to true when it finds a q such that <x,q > EF ).
Although the authors prefer the definition based on proof systems, it is sometimes
easier to show that a problem belongs to NP with this other definition. For instance, it
makes Problems 10.3.8 and 10.3.9 completely obvious.
For an overview of algorithms for sorting by transformation consult Aho, Hopcroft,

and Ullman (1974). In particular, the lexicographic sorting algorithm can sort n ele-
ments in a time in O(n +m), where m is the sum of the sizes of the elements to be
sorted. For modifications of heapsort that come close to being optimal for the worst-
case number of comparisons, consult Gonnet and Munro (1986) and Carlsson (1986,
1987). In his tutorial article, Pippenger (1978) describes a method similar to decision
trees for determining a lower bound on the size of logic circuits such as those
presented in Problems 4.11.8 to 4.11.12. In particular, this technique shows that the
required solutions for Problems 4.11.9 and 4.11.12 are optimal.
The reduction IQ <_/ MQ (Problem 10.2.9) comes from Bunch and Hopcroft
(1974). If f and g are two cost functions as in section 10.2.2, then an algorithm that is
asymptotically more efficient than the naive algorithm for calculating fg is given in
Fredman (1976). Theorem 10.2.5 is due to Fischer and Meyer (1971) and Theorem
10.2.6 is due to Furman (1970). In the case of cost functions whose range is restricted
to (0,+-), Arlazarov, Dinic, Kronrod, and Faradzev (1970) present an algorithm to
calculate fg using a number of Boolean operations in 0 (n 3/log n). Problem 10.2.12 is
solved in Fischer and Meyer (1971). The reduction INV <_/ MLT (Problem 10.2.22),
which is crucial in the proof of Theorem 10.2.7, comes from Cook and Aanderaa
(1969). For further information concerning the topics of Section 10.2, consult Aho,
Hopcroft, and Ullman (1974).
The theory of NP-completeness originated with two fundamental papers : Cook
(1971) proves that SAT-CNF is NP-complete, and Karp (1972) underlines the impor-
tance of this notion by presenting a large number of NP-complete problems. To be
historically exact, the original statement from Cook (1971) is that X <_ T TAUT-DNF
for every X E NP, where TAUT-DNF is concerned with tautologies in disjunctive
normal form ; however, it should be noted that TAUT-DNF is probably not
NP-complete since otherwise NP = co-NP (Problem 10.3.16). A similar theory was
developed independently by Levin (1973). The uncontested authority in matters of
NP-completeness is Garey and Johnson (1979). A good introduction can also be found
in Hopcroft and Ullman (1979). The fact that the set of prime numbers is in NP
(Examples 10.3.7 and 10.3.14, and Problems 10.3.13, 10.3.33, and 10.3.34) was
discovered by Pratt (1975); more succinct primality certificates can be found in
Pomerance (1987). Problems 10.3.14 and 10.3.16 are from Brassard (1979). Part of
the solution to Problem 10.3.28 can be found in Stockmeyer (1973). In practice. the
fact that a problem is NP-complete does not make it go away. However in this case
we have to be content with heuristics and approximations as described in Garey and
Johnson (1976), Sahni and Horowitz (1978), and Horowitz and Sahni (1978). To find
out more about non-determinism, see Hopcroft and Ullman (1979).
This chapter has done little more than sketch the theory of computational com-
plexity. Several important techniques have gone unmentioned. An algebraic approach
to lower bounds is described in Aho, Hopcroft, and Ullman (1974), Borodin and
Munro (1975), and Winograd (1980). For an introduction to adversary arguments

(Problem 8.4.4), consult Horowitz and Sahni (1978). Although we do not know how
to prove that there are no efficient algorithms for NP-complete problems, there exist
problems that are intrinsically difficult, as described in Aho, Hopcroft, and Ullman
(1974). These can be solved in theory, but it can be proved that no algorithm can
solve them in practice when the instances are of moderate size, even if it is allowed to
take a time comparable to the age of the universe and as many bits of memory as there
are elementary particles in the known universe (Stockmeyer and Chandra 1979). As
mentioned in Problem 10.3.32, there also exist problems that cannot be solved by any
algorithm, whatever the resources available; read Turing (1936), Gardner and Bennett
(1979), Hopcroft and Ullman (1979), and Brassard and Monet (1982).
Table of Notation
#T number of elements in array T ;

cardinality of set T.
i j interval of integers : { k E IN I i <_ k <_ j }
div, mod arithmetic quotient and modulo ;
extended to polynomials in Section 10.2.3
x arithmetic and matrix multiplication ;
Cartesian product
T pointer
E- assignment
var x return parameter of a procedure or function
return dynamic end of a procedure
return v dynamic end of a function with value v returned
IxI size of instance x ;
[lg(1+x)l if x is an integer;
absolute value of x
log, lg, In, logb logarithm in basis 10, 2, e, and b, respectively
e basis of the natural logarithm : 2.7182818...
n! n factorial (0! = 1 and n! =n x (n -1)!, n! > 1)
number of combinations of k elements chosen among n
338
Table of Notation 339
<a,b > ordered pair consisting of elements a and b

(a, b) same as < a, b > (in particular : edge of a directed graph) ;
open interval {x E IR I a <x < b [
closed interval {xeIR I a <-x <_b }
set of real numbers larger than or equal to a
denotes comments in algorithms ;
set of elements
such that
set inclusion (allowing equality)
strict set inclusion
set union: AuB={x IxEA or xEB {
set intersection : A nB = { x I x E A and x E B I
set membership
set nonmembership
set difference : A\ B={ x IxEA and x 0 B
empty set
sets of integers, reals, and Booleans (see Section 2.1.1)
f is a function from A to B
there exists an x such that P (x)
there exists one and only one
for all x, P (x)
asymptotic notation (see Section 2.1)
implies
if and only if
summation
integral
plus or minus
derivative of the function f (x)
00 infinity
limit of f (x) when x goes to infinity
floor of x : largest integer less than or equal to x
extended to polynomials in Section 10.2.3
rx ceiling of x : smallest integer larger than or equal to x
lg* iterated logarithm (see page 63)
uniform (i .. j) randomly and uniformly selected integer between i and j
x =y (mod n) x is congruent to y modulo n (n divides x - y exactly)
xO+y exclusive-or of x and y for bits or Booleans ;
bit-by-bit exclusive-or for bit strings
340 Table of Notation
and, or Boolean conjunction and disjunction

x Boolean complement of x
F(,(a) Fourier transform of vector a with respect to co
Fw' (a) inverse Fourier transform
A B A is linearly reducible to B
A B A is linearly equivalent to B
A <_ T B A is polynomially reducible to B in the sense of Turing
A =T B A is polynomially equivalent to B in the sense of Turing
A ,B A is many-one polynomially reducible to B
A PB A is many-one polynomially equivalent to B
P class of decision problems that can be solved in polynomial
time
NP class of decision problems that have an efficient proof system
co-NP class of decision problems whose complementary problem is
in NP
choose instruction for non-deterministic choice
Bibliography
ABADI, M., J. FEIGENBAUM, and J. KILIAN (1987), "On hiding information from an oracle",
Proceedings of 19th Annual ACM Symposium on the Theory of Computing, pp. 195-203.
ACKERMANN, W. (1928), "Zum Hilbertschen Aufbau der reellen Zahlen", Mathematische
Annalen, 99, 118-133.
ADEL'SON-VEL'SKII, G. M. and E. M. LANDIS (1962), "An algorithm for the organization of
information" (in Russian), Doklady Akademii Nauk SSSR, 146, 263-266.
ADLEMAN, L. M. and M.-D. A. HUANG (1987), "Recognizing primes in random polynomial
time", Proceedings of 19th Annual ACM Symposium on the Theory of Computing, pp.
462-469.
ADLEMAN, L. M., K. MANDERS, and G. MILLER (1977), "On taking roots in finite fields",
Proceedings of 18th Annual IEEE Symposium on the Foundations of Computer Science, pp.
175-178.
ADLEMAN, L. M., C. POMERANCE, and R. S. RUMELY (1983), "On distinguishing prime numbers
from composite numbers", Annals of Mathematics, 117, 173-206.
AHO, A. V. and M. J. CORASICK (1975), "Efficient string matching : An aid to bibliographic
search", Communications of the ACM, 18(6), 333-340.
AHO, A. V., J. E. HOPCROFT, and J. D. ULLMAN (1974), The Design and Analysis of Computer
Algorithms, Addison-Wesley, Reading, MA.
AHO, A. V., J. E. HOPCROFT, and J. D. ULLMAN (1976), "On finding lowest common ancestors in
trees", SIAM Journal on Computing, 5(1), 115-132.
AHO, A. V., J.E. HOPCROFr, and J. D. ULLMAN (1983), Data Structures and Algorithms,
Addison-Wesley, Reading, MA.
AJTAI, M., J. KOMLOS, and E. SZEMEREDI (1983), "An O(n logn) sorting network", Proceedings
of 15th Annual ACM Symposium on the Theory of Computing, pp. 1-9.
ANON. (c. 1495) Lytell Geste of Robyn Hode, Wynkyn de Worde, London.
341
342 Bibliography
ARLAZAROV, V. L., E. A. DINIC, M. A. KRONROD, and I. A. FARADZEV (1970), "On economical

construction of the transitive closure of a directed graph" (in Russian), Doklady Akademii
Nauk SSSR, 194, 487-488.
BAASE, S. (1978), Computer Algorithms: Introduction to Design and Analysis, Addison-Wesley,
Reading, MA; second edition, 1987.
BABAI, L. (1979), "Monte Carlo algorithms in graph isomorphism techniques", Research report,
Departement de mathematiques et de statistique, Universite de Montreal, D.M.S. no. 79-10.
BACH, E., G. MILLER, and J. SHALLIT (1986), "Sums of divisors, perfect numbers and factoring",
SIAM Journal on Computing, 15(4), 1143-1154.
BACHMANN, P.G.H. (1894), Zahlentheorie, vol. 2: Die Analytische Zahlentheorie,
B.G. Teubner, Leipzig.
BATCHER, K. (1968), "Sorting networks and their applications", Proceedings of AFIPS 32nd
Spring Joint Computer Conference, pp. 307-314.
BEAUCHEMIN, P., G. BRASSARD, C. CR£PEAU, C. GOUTIER, and C. POMERANCE (1988), "The
generation of random numbers that are probably prime", Journal of Cryptology, 1(1), in press.
BELAGA, E. C. (1961), "On computing polynomials in one variable with initial preconditioning
of the coefficients", Problemi Kibernetiki, 5, 7-15.
BELLMAN, R. E. (1957), Dynamic Programming, Princeton University Press, Princeton, NJ.
BELLMAN, R. E. and S. E. DREYFUS (1962), Applied Dynamic Programming, Princeton Univer-
sity Press, Princeton, NJ.
BELLMORE, M. and G. NEMHAUSER (1968), "The traveling salesman problem: A survey",
Operations Research, 16(3), 538-558.
BENNETT, C. H., G. BRASSARD, and J.-M. ROBERT (1988), "Privacy amplification through public
discussion", SIAM Journal on Computing, in press.
BENTLEY, J. L. (1984), "Programming pearls : Algorithm design techniques", Communications of
the ACM, 27(9), 865-871.
BENTLEY, J. L., D. HAKEN, and J. B. SAXE (1980), "A general method for solving divide-and-
conquer recurrences", SIGACT News, ACM, 12(3), 36-44.
BENTLEY, J. L. and I. SHAMOS (1976), "Divide-and-conquer in multidimensional space",
BENTLEY, J. L., D. F. STANAT, and J. M. STEELE (1981), "Analysis of a randomized data structure
for representing ordered sets", Proceedings of 19th Annual Allerton Conference on Communi-
cation, Control, and Computing, pp. 364-372.
BERGE, C. (1958), Theorie des graphes et ses applications, Dunod, Paris; second edition, 1967;
translated as: The Theory of Graphs and Its Applications (1962), Methuen & Co., London.
BERGE, C. (1970), Graphes et hypergraphes, Dunod, Paris ; translated as : Graphs and Hyper-
graphs (1973), North Holland, Amsterdam.
BERLEKAMP, E. R. (1970), "Factoring polynomials over large finite fields", Mathematics of Com-
putation, 24(111), 713-735.
BERLINER, H. J. (1980), "Backgammon computer program beats world champion", Artificial
Intelligence, 14, 205-220.
BLUM, L., M. BLUM, and M. SHUB (1986), "A simple unpredictable pseudo-random number gen-
erator", SIAM Journal on Computing, 15(2), 364-383.
Bibliography 343
BLUM, M., R. W. FLOYD, V.R. PRATT, R. L. RIVEST. and R. E. TARJAN (1972), "Time bounds for
selection", Journal of Computer and System Sciences, 7(4), 448-461.
BLUM, M. and S. MICAU (1984), "How to generate cryptographically strong sequences of
pseudo-random bits", SIAM Journal on Computing, 13(4), 850-864.
BORODIN, A. B. and J. 1. MUNRO (1971), "Evaluating polynomials at many points", Information
Processing Letters, 1(2), 66-68.
BORODIN, A. B. and J. 1. MUNRO (1975), The Computational Complexity of Algebraic and
Numeric Problems, American Elsevier, New York, NY.
BORUVKA, O. (1926), "O jistem problemu minimalnim", Praca Morarske Prirodovedecke
Spolecnosti, 3. 37-58.
BOYER, R. S. and J. S. MOORE (1977), "A fast string searching algorithm", Communications of
theACM, 20(10), 762-772.
BRASSARD, G. (1979), "A note on the complexity of cryptography", IEEE Transactions on Infor-
mation Theory, IT-25(2), 232-233.
BRASSARD, G. (1985), "Crusade for a better notation", SIGACT News, ACM, 17(1), 60-64.
BRASSARD, G. (1988), Modern Cr_yptology: A Tutorial, Lecture Notes in Computer Science,
Springer-Verlag, New York, NY.
BRASSARD, G. and S. KANNAN (1988), "The generation of random permutations on the fly",
Information Processing Letters (in press).
BRASSARD, G. and S. MONET (1982), "L'indecidabilite sans larme (ni diagonalisation)", Publica-
tion no. 445, Departement d'informatique et de recherche operationnelle, Universite de
Montreal.
BRASSARD, G., S. MONET, and D. ZUFFELLATO (1986), "L'arithmetique des tres grands entiers",
TSI: Technique et Science Informatiques, 5(2), 89-102.
BRATLEY, P., B. L. Fox, and L. E. SCHRAGE (1983), A Guide to Simulation, Springer-Verlag,
New York, NY; second edition, 1987.
BRIGHAM, E.O. (1974), The Fast Fourier Transform, Prentice-Hall, Englewood Cliffs, NJ.
BUNCH, J. and J. E. HOPCROFT (1974), "Triangular factorization and inversion by fast matrix
multiplication", Mathematics of Computation, 28(125), 231-236.
BUNEMAN, P. and L. LEVY (1980), "The towers of Hanoi problem", Information Processing
Letters, 10(4,5), 243-244.
CARASSO, C. (1971), Analyse numerique, Lidec, Montreal, Quebec, Canada.
CARLSSON, S. (1986), Heaps, Doctoral dissertation, Department of Computer Science, Lund
University, Lund, Sweden, CODEN: LUNFD6/(NFCS-I(X)3)/(1-70)/(1986).
CARLSSON, S. (1987), "Average case results on heapsort", Bit, 27, 2-17.
CARTER, J. L. and M. N. WEGMAN (1979), "Universal classes of hash functions", Journal of
Computer and System Sciences, 18(2), 143-154.
CHANG, L. and J. KORSH (1976), "Canonical coin changing and greedy solutions", Journal of the
ACM, 23(3), 418-422.
CHERITON, D. and R. E. TARJAN (1976), "Finding minimum spanning trees", SIAM Journal on
Computing, 5(4), 724-742.
CHRISTOFIDES, N. (1975), Graph Theory: An Algorithmic Approach, Academic Press,
New York, NY.
344 Bibliography
CHRISTOFIDES, N. (1976), "Worst-case analysis of a new heuristic for the traveling salesman
problem", Management Sciences Research Report no. 388, Carnegie-Mellon University, Pitts-
burgh, PA.
COHEN, H. and A. K. LENSTRA, (1987), "Implementation of a new primality test", Mathematics
of Computation, 48(177), 103-121.
COOK, S. A. (1971), "The complexity of theorem-proving procedures", Proceedings of 3rd
Annual ACM Symposium on the Theory of Computing, pp. 151-158.
COOK, S. A. and S. O. AANDERAA (1969), "On the minimum complexity of functions", Transac-
tions of the American Mathematical Society, 142, 291-314.
COOLEY, J. M., P. A. LEWIS, and P. D. WELCH (1967), "History of the fast Fourier transform",
Proceedings of the IEEE, 55, 1675-1679.
COOLEY, J. M. and J. W. TUKEY (1965), "An algorithm for the machine calculation of complex
Fourier series", Mathematics of Computation, 19(90), 297-301.
COPPERSMITH, D. and S. WINOGRAD (1987), "Matrix multiplication via arithmetic progressions",
CRAY RESEARCH (1986), "CRAY-2 computer system takes a slice out of pi", Cray Channels,
8(2), 39.
CURTISS, J. H. (1956), "A theoretical comparison of the efficiencies of two classical methods and
a Monte Carlo method for computing one component of the solution of a set of linear alge-
braic equations", in Symposium on Monte Carlo Methods, H. A. Meyer, ed., John Wiley &
Sons, New York, NY, pp. 191-233.
DANIELSON, G. C. and C. LANCZOS (1942), "Some improvements in practical Fourier analysis
and their application to X-ray scattering from liquids", Journal of the Franklin Institute, 233,
365-380, 435-452.
DE BRUIJN, N. G. (1961), Asymptotic Methods in Analysis, North Holland, Amsterdam.
DEMARS, C. (1981), "Transformee de Fourier rapide", Micro-Systemes", 155-159.
DENNING, D.E.R. (1983), Cryptography and Data Security, Addison-Wesley, Reading, MA.
DEVROYE, L. (1986), Non-Uniform Random Variate Generation, Springer-Verlag, New York,
NY.
DEWDNEY, A. K. (1984), "Computer recreations : Yin and yang : recursion and iteration, the
tower of Hanoi and the Chinese rings", Scientific American, 251(5), 19-28.
DEYONG, L. (1977), Playboy's Book of Backgammon, Playboy Press, Chicago, IL.
DIFFIE, W. and M. E. HELLMAN (1976), "New directions in cryptography", IEEE Transactions
on Information Theory, IT-22(6), 644-654.
DIJKSTRA, E. W. (1959), "A note on two problems in connexion with graphs", Numerische
Mathematik, 1, 269-271.
DIXON, J. D. (1981), "Asymptotically fast factorization of integers", Mathematics of Computa-
tion, 36(153), 255-260.
DROMEY, R. G. (1982), How to Solve It by Computer, Prentice-Hall, Englewood Cliffs, NJ.
ERDOS, P. and C. POMERANCE (1986), "On the number of false witnesses for a composite
number", Mathematics of Computation, 46(173), 259-279.
EVEN, S. (1980), Graph Algorithms, Computer Science Press, Rockville, MD.
Bibliography 345
FEIGENBAUM, J. (1986), "Encrypting problem instances, or .... can you take advantage of
someone without having to trust him?", Proceedings of CRYPTO 85, Springer-Verlag, Berlin,
pp. 477-488.
FISCHER, M. J. and A. R. MEYER (1971), "Boolean matrix multiplication and transitive closure",
Proceedings of IEEE 12th Annual Symposium on Switching and Automata Theory, pp.
129-131.
FLAJOLET, P. and G. N. MARTIN (1985), "Probabilistic counting algorithms for data base applica-
tions", Journal of Computer and System Sciences, 31(2), 182-209.
FLOYD, R. W. (1962), "Algorithm 97: Shortest path", Communications of the ACM, 5(6), 345.
Fox, B.L. (1986), "Algorithm 647: Implementation and relative efficiency of quasirandom
sequence generators", ACM Transactions on Mathematical Software, 12(4), 362-376.
FREDMAN, M. L. (1976), "New bounds on the complexity of the shortest path problem", SIAM
Journal on Computing, 5(1), 83-89.
FREDMAN, M. L. and R. E. TARJAN (1984), "Fibonacci heaps and their uses in improved network
optimization algorithms", Proceedings of 25th Annual IEEE Symposium on the Foundations of
Computer Science, pp. 338-346.
FREIVALDS, R. (1977), "Probabilistic machines can use less running time", Proceedings of Infor-
mation Processing 77, pp. 839-842.
FREIVALDS, R. (1979), "Fast probabilistic algorithms", Proceedings of 8th Symposium on the
Mathematical Foundations of Computer Science, Lecture Notes in Computer Science, 74,
Springer-Verlag, Berlin, pp. 57-69.
FURMAN, M.E. (1970), "Application of a method of fast multiplication of matrices in the
problem of finding the transitive closure of a graph" (in Russian), Doklady Akademii Nauk
SSSR, 194, 524.
GARDNER, M. (1977), "Mathematical games : A new kind of cipher that would take millions of
years to break", Scientific American, 237(2), 120-124.
GARDNER, M. and C. H. BENNETT (1979), "Mathematical games : The random number omega
bids fair to hold the mysteries of the universe", Scientific American, 241(5), 20-34.
GAREY, M. R. and D. S. JOHNSON (1976), "Approximation algorithms for combinatorial prob-
lems : An annotated bibliography", in Traub (1976), pp. 41-52.
GAREY, M. R. and D. S. JOHNSON (1979), Computers and Intractability: A Guide to the Theory
of NP-Completeness, W. H. Freeman and Co., San Francisco, CA.
GENTLEMAN, W. M. and G. SANDE (1966), "Fast Fourier transforms-for fun and profit",
Proceedings of AFIPS Fall Joint Computer Conference, 29, Spartan, Washington, DC, pp.
563-578.
GILBERT, E. N. and E. F. MOORE (1959), "Variable length encodings", Bell System Technical
Journal, 38(4), 933-968.
GLEICK, J. (1987), "Calculating pi to 134 million digits hailed as great test for computer", New
York Times, c. March 14.
GODBOLE, S. (1973), "On efficient computation of matrix chain products", IEEE Transactions on
Computers, C-22(9), 864-866.
GOLDWASSER, S. and J. KILIAN (1986), "Almost all primes can be quickly certified", Proceed-
ings of 18th Annual ACM Symposium on the Theory of Computing, pp. 316-329.
GOLDWASSER, S. and S. MICALI (1984), "Probabilistic encryption", Journal of Computer and
System Sciences, 28(2), 270-299.
346 Bibliography
GOLOMB, S. and L. BAUMERT (1965), "Backtrack programming", Journal of the ACM, 12(4),
516-524.
GONDRAN, M. and M. (1979), Graphes et algorithmes, Eyrolles, Paris ; translated as :
Graphs and Algorithms (1984), John Wiley & Sons, New York, NY.
GONNET, G. H. (1984), Handbook of Algorithms and Data Structures, Addison-Wesley, Reading,
MA.
GONNET, G. H. and J. I. MUNRO (1986), "Heaps on heaps", SIAM Journal on Computing, 15(4),
964-971.
GOOD, I. J. (1968), "A five-year plan for automatic chess", in Machine Intelligence, vol. 2,
E. Dale and D. Michie, eds., American Elsevier, New York, NY, pp. 89-118.
GRAHAM, R. L. and P. HELL (1985), "On the history of the minimum spanning tree problem",
Annals of the History of Computing, 7(1), 43-57.
GREENE, D. H. and D.E. KNUTH (1981), Mathematics for the Analysis of Algorithms, Birkhauser,
Boston, MA.
GRIES, D. (1981), The Science of Programming, Springer-Verlag, New York, NY.
GRIES, D. and G. LEVIN (1980), "Computing Fibonacci numbers (and similarly defined func-
tions) in log time", Information Processing Letters, 11(2), 68-69.
HALL, A. (1873), "On an experimental determination of it ", Messenger of Mathematics, 2,
113-114.
HAMMERSLEY, J. M. and D.C. HANDSCOMB (1965), Monte Carlo Methods; reprinted in 1979 by
Chapman and Hall, London.
HARDY, G. H. and E. M. WRIGHT (1938), An Introduction to the Theory of Numbers, Oxford Sci-
ence Publications, Oxford, England ; fifth edition, 1979.
HAREL, D. (1987), Algorithmics : The Spirit of Computing, Addison-Wesley, Reading, MA.
HARRISON, M. C. (1971), "Implementation of the substring test by hashing", Communications of
the ACM, 14(12), 777-779.
HELD, M. and R. KARP (1962), "A dynamic programming approach to sequencing problems",
SIAM Journal on Applied Mathematics, 10(1), 196-210.
HELLMAN, M. E. (1980), "The mathematics of public-key cryptography", Scientific American,
241(2), 146-157.
HOARE, C.A.R. (1962), "Quicksort", Computer Journal, 5(1), 10-15.
HoPCROFT, J. E. and R. KARP (1971), "An algorithm for testing the equivalence of finite auto-
mata", Technical report TR-71-114, Department of Computer Science, Cornell University,
Ithaca, NY.
HOPCROFT, J. E. and L. R. KERR (1971), "On minimizing the number of multiplications necessary
for matrix multiplication", SIAM Journal on Applied Mathematics, 20(1), 30-36.
HOPCROFT, J. E. and R. E. TARTAN (1973), "Efficient algorithms for graph manipulation", Com-
munications of the ACM, 16(6), 372-378.
HOPCROFT, J. E. and R. E. TARJAN (1974), "Efficient planarity testing", Journal of the ACM,
21(4), 549-568.
HOPCROFT, J. E. and J. D. ULLMAN (1973), "Set merging algorithms", SIAM Journal on Com-
puting, 2(4), 294-303.
HOPCRoFT, J. E. and J. D. ULLMAN (1979), Introduction to Automata Theory, Languages, and
Computation, Addison-Wesley, Reading, MA.
Bibliography 347
HoRowiTz, E. and S. SAHNI (1976), Fundamentals of Data Structures, Computer Science Press,
Rockville, MD.
HoRowrrz, E. and S. SAHNI (1978), Fundamentals of Computer Algorithms, Computer Science
Press, Rockville, MD.
Hu, T. C. and M. R. SHING (1982), "Computations of matrix chain products", Part I, SIAM
Hu, T.C. and M. R. SHING (1984), "Computations of matrix chain products", Part II, SIAM
ITAI, A. and M. RODEH (1981), "Symmetry breaking in distributive networks", Proceedings of
22nd Annual IEEE Symposium on the Foundations of Computer Science, pp. 150-158.
JANKO, W. (1976), "A list insertion sort for keys with arbitrary key distribution", ACM Transac-
tions on Mathematical Software, 2(2), 143-153.
JARNIK, V. (1930), "0 jistem problemu minimalnim", Praca Moravske Prirodovedecke Spolec-
nosti, 6, 57-63.
JENSEN, K. and N. WIRTH (1985), Pascal User Manual and Report, third edition revised by
A. B. Michel and J. F. Miner, Springer-Verlag, New York, NY.
JOHNSON, D. B. (1975), "Priority queues with update and finding minimum spanning trees",
Information Processing Letters, 4(3), 53-57.
JOHNSON, D. B. (1977), "Efficient algorithms for shortest paths in sparse networks", Journal of
the ACM, 24(1), 1-13.
KAHN, D. (1967), The Codebreakers: The Story of Secret Writing, Macmillan, New York, NY.
KALISKI, B.S., R.L. RIVEST, and A.T. SHERMAN (1988), "Is the Data Encryption Standard a
group?", Journal of Cryptology, (1), in press.
KANADA, Y, Y. TAMURA, S. YOSHINO, and Y. USHIRO (1986), "Calculation of it to 10,013,395
decimal places based on the Gauss-Legendre algorithm and Gauss arctangent relation",
manuscript.
KARATSUBA, A. and Y. OFMAN (1962), "Multiplication of multidigit numbers on automata"
(in Russian), Doklady Akademii Nauk SSSR, 145, 293-294.
KARP, R. (1972), "Reducibility among combinatorial problems", in Complexity of Computer
Computations, R. E. Miller and J. W. Thatcher, eds., Plenum Press, New York, NY, pp.
85-104.
KARP, R. and M. O. RABIN (1987), "Efficient randomized pattern-matching algorithms", IBM
Journal of Research and Development, 31(2), 249-260.
KASIMI, T. (1965), "An efficient recognition and syntax algorithm for context-free languages",
Scientific Report AFCRL-65-758, Air Force Cambridge Research Laboratory, Bedford, MA.
KLAMKIN, M.S. and D.J. NEWMAN (1967), "Extensions of the birthday surprise", Journal of
Combinatorial Theory, 3(3), 279-282.
KLEENE, S.C. (1956), "Representation of events in nerve nets and finite automata", in Automata
Studies, C.E. Shannon and J. McCarthy, eds., Princeton University Press, Princeton, NJ, pp.
3-40.
KNUTH, D. E. (1968), The Art of. Computer Programming, 1 : Fundamental Algorithms,
Addison-Wesley, Reading, MA; second edition, 1973.
KNUTH, D. E. (1969), The Art of Computer Programming, 2: Seminumerical Algorithms,
Addison-Wesley, Reading, MA; second edition, 1981.
348 Bibliography
KNUTH, D. E. (1971), "Optimal binary search trees", Acta Informatica,

1, 14-25.
KNUTH, D. E. (1973), The Art of Computer Programming, 3: Sorting and Searching, Addison-
Wesley, Reading, MA.
KNUTH, D. E. (1975a), "Estimating the efficiency of backtrack programs", Mathematics of Com-
putation, 29, 121-136.
KNUTH, D. E. (1975b), "An analysis of alpha-beta cutoffs", Artificial Intelligence, 6, 293-326.
KNUTH, D. E. (1976), "Big Omicron and big Omega and big Theta", SIGACT News, ACM, 8(2),
18-24.
KNUTH, D. E. (1977), "Algorithms", Scientific American, 236(4), 63-80.
KNUTH, D. E., J. H. MORRIS, and V. R. PRATT (1977), "Fast pattern matching in strings", SIAM
KRANAKIS, E. (1986), Primality and Cryptography, Wiley-Teubner Series in Computer Science.
KRUSKAL, J. B., Jr. (1956), "On the shortest spanning subtree of a graph and the traveling
salesman problem", Proceedings of the American Mathematical Society, 7(1), 48-50.
LAURIERE , J.-L. (1979), Elements de programmation dynamique, Bordas, Paris.
LAWLER, E. L. (1976), Combinatorial Optimization : Networks and Matroids, Holt, Rinehart and
Winston, New York, NY.
LAWLER, E. L. and D.W. WOOD (1966), "Branch-and-bound methods : A survey", Operations
Research, 14(4), 699-719.
LECARME, O. and J.-L. NEBUT (1985), Pascal pour programmeurs, McGraw-Hill, Paris.
LECLERC, G. L. (1777), Essai d'arithmetique morale.
LEHMER, D. H. (1969), "Computer technology applied to the theory of numbers", in Studies in
Number Theory, W. J. LeVeque, ed., Mathematical Association of America, Providence, RI, p.
117.
LENSTRA, H.W., Jr. (1982), "Primality testing", in Lenstra and Tijdeman (1982), pp. 55-97.
LENSTRA, H. W., Jr. (1986), "Factoring integers with elliptic curves", report 86-18, Mathema-
tisch Instituut, Universiteit van Amsterdam; to appear in Annals of Mathematics.
LENSTRA, H. W., Jr. and R. TIJDEMAN, eds. (1982), Computational Methods in Number Theory,
Part I, Mathematical Centre Tracts 154, Mathematisch Centrum, Amsterdam.
LEVIN, L. (1973), "Universal search problems" (in Russian), Problemy Peredaci Informacii, 9,
115-116.
LEWIS, H. R. and C. H. Papadimitriou (1978), "The efficiency of algorithms", Scientific Amer-
ican, 238(1), 96-109.
LUEKER, G. S. (1980), "Some techniques for solving recurrences", Computing Surveys, 12(4),
419-436.
MARSH, D. (1970), "Memo functions, the Graph Traverser, and a simple control situation", in
Machine Intelligence, 5, B. Meltzer and D. Michie, eds., American Elsevier, New York, NY,
and Edinburgh University Press, pp. 281-300.
MCDIARMID, C. J. H. and B. A. REED (1987), "Building heaps fast", submitted to the Journal of
Algorithms.
MELHORN, K. (1984a), Data Structures and Algorithms, 1 : Sorting and Searching, Springer-
Verlag, Berlin.
MELHORN, K. (1984b), Data Structures and Algorithms, 2: Graph Algorithms and NP-
Completeness, Springer-Verlag, Berlin.
Bibliography 349
MELHORN, K. (1984c), Data Structures and Algorithms, 3: Multi-Dimensional Searching and

Computational Geometry, Springer-Verlag, Berlin.
METROPOLIS, I. N. and S. ULAM (1949), "The Monte Carlo method", Journal of the American
Statistical Association, 44(247), 335-341.
MICHIE, D. (1968), "'Memo' functions and machine learning", Nature, 218, 19-22.
MONIER, L. (1980), "Evaluation and comparison of two efficient probabilistic primality testing
algorithms", Theoretical Computer Science, 12, 97-108.
MONTGOMERY, P. L. (1987), "Speeding the Pollard and elliptic curve methods of factorization",
Mathematics of Computation, 48(177), 243-264.
NEMHAUSER, G. (1966), Introduction to Dynamic Programming, John Wiley & Sons, Inc.,
New York, NY.
NILSSON, N. (1971), Problem Solving Methods in Artificial Intelligence, McGraw-Hill,
New York, NY.
PAN, V. (1978), "Strassen's algorithm is not optimal", Proceedings of 19th Annual IEEE Sympo-
sium on the Foundations of Computer Science, pp. 166-176.
PAPADIMITRIOU, C. H. and K. STEIGLITZ (1982), Combinatorial Optimization : Algorithms and
Complexity, Prentice-Hall, Englewood Cliffs, NJ.
PERALTA, R. C. (1986), "A simple and fast probabilistic algorithm for computing square roots
modulo a prime number", IEEE Transactions on Information Theory, IT-32(6), 846-847.
PIPPENGER, N. (1978), "Complexity theory", Scientific American, 238(6), 114-124.
POHL, I. (1972), "A sorting problem and its complexity", Communications of the ACM, 15(6),
462-463.
POLLARD, J. M. (1971), "The fast Fourier transform in a finite field", Mathematics of Computa-
tion, 25(114), 365-374.
POLLARD, J. M. (1975), "A Monte Carlo method of factorization", Bit, 15, 331-334.
POMERANCE, C. (1982), "Analysis and comparison of some integer factoring algorithms", in
Lenstra and Tijdeman (1982), pp. 89-139.
POMERANCE, C. (1987), "Very short primality proofs", Mathematics of Computation, 48(177),
315-322.
PRATT, V.R. (1975), "Every prime has a succinct certificate", SIAM Journal on Computing,
4(3), 214-220.
PRIM, R.C. (1957), "Shortest connection networks and some generalizations", Bell System
Technical Journal, 36, 1389-1401.
PURDOM, P. W., Jr. and C. A. BROWN (1985), The Analysis of Algorithms, Holt, Rinehart and
Winston, New York, NY.
RABIN, M. O. (1976), "Probabilistic algorithms", in Traub (1976), pp. 21-39.
RABIN, M. O. (1980a), "Probabilistic algorithms in finite fields", SIAM Journal on Computing,
9(2), 273-280.
RABIN, M. O. (1980b), "Probabilistic algorithm for primality testing", Journal of Number
Theory, 12, 128-138.
RABINER, L. R. and B. GOLD (1974), Digital Signal Processing, Prentice-Hall, Englewood Cliffs,
NJ.
REINGOLD, E. M., J. NIEVERGELT, and N. DEO (1977), Combinatorial Algorithms : Theory and
Practice, Prentice-Hall, Englewood Cliffs, NJ.
350 Bibliography
RIVEST, R. L. and R. W. FLOYD (1973), "Bounds on the expected time for median computations",
in Combinatorial Algorithms, R. Rustin, ed., Algorithmics Press, New York, NY, pp. 69-76.
RIVEST, R. L., A. SHAMIR, and L.M. ADLEMAN, (1978), "A method for obtaining digital signa-
tures and public-key cryptosystems", Communications of the ACM, 21(2), 120-126.
ROBSON, J.M. (1973), "An improved algorithm for traversing binary trees without auxiliary
stack", Information Processing Letters, 2(1), 12-14.
ROSENTHAL, A. and A. GOLDNER (1977), "Smallest augmentation to biconnect a graph", SIAM
RUDIN, W. (1953), Principles of Mathematical Analysis, McGraw-Hill, New York, NY.
RUNGE, C. and H. KoNIG (1924), Die Grundlehren der Mathematischen Wissenschaften, 11,
Springer, Berlin.
RYTTER, W. (1980), "A correct preprocessing algorithm for Boyer-Moore string searching",
SIAM Journal on Computing, 9(3), 509-512.
SAHNI. S. and E. HOROWITZ (1978), "Combinatorial problems: reducibility and approximation",
Operations Research, 26(4), 718-759.
SCHONHAGE, A. and V. STRASSEN (1971), "Schnelle Multiplikation grosser Zahlen", Computing,
7, 281-292.
SCHWARTZ, E. S. (1964), "An optimal encoding with minimum longest code and total number of
digits", Information and Control, 7(1), 37-44.
SCHWARTZ, J. (1978), "Probabilistic algorithms for verification of polynomial identities", Com-
puter Science Department, Courant Institute, New York University, Technical Report no. 604.
SEDGEWICK, R. (1983), Algorithms, Addison-Wesley, Reading, MA.
SHAMIR, A. (1979), "Factoring numbers in O (log n) arithmetic steps", Information Processing
Letters, 8(1), 28-31.
SHANKS, D. (1972), "Five number-theoretic algorithms", Proceedings of the Second Manitoba
Conference on Numerical Mathematics, pp. 51-70.
SLOANE, N.J. A. (1973), A Handbook of Integer Sequences, Academic Press, New York, NY.
SOBOL', I. M. (1974), The Monte Carlo Method, second edition, University of Chicago Press,
Chicago, IL.
SOLOVAY, R. and V. STRASSEN (1977), "A fast Monte-Carlo test for primality", SIAM Journal
on Computing, 6(1), 84-85; erratum (1978), ibid, 7, 118.
STANDISH, T. A. (1980), Data Structure Techniques, Addison-Wesley, Reading, MA.
STINSON, D. R. (1985), An Introduction to the Design and Analysis of Algorithms, The Charles
Babbage Research Centre, St. Pierre, Manitoba.
STOCKMEYER, L.J. (1973), "Planar 3-colorability is polynomial complete", SIGACT News, 5(3),
19-25.
STOCKMEYER, L.J. and A. K. CHANDRA (1979), "Intrinsically difficult problems", Scientific
American, 240(5), 140-159.
STONE, H. S. (1972), Introduction to Computer Organization and Data Structures, McGraw-Hill,
New York, NY.
STRASSEN, V. (1969), "Gaussian elimination is not optimal", Numerische Mathematik, 13,
354-356.
TARJAN, R. E. (1972), "Depth-first search and linear graph algorithms", SIAM Journal on Com-
puting, 1(2), 146-160.
Bibliography 351
TARJAN, R. E. (1975), "On the efficiency of a good but not linear set merging algorithm",
Journal of'the ACM, 22(2), 215-225.
TARJAN, R.E. (1981), "A unified approach to path problems", Journal of the ACM, 28(3),
577-593.
TARJAN, R.E. (1983), Data Structures and Network Algorithms, SIAM, Philadelphia, PA.
TRAUB, J. F., ed. (1976), Algorithms and Complexity : Recent Results and New Directions,
Academic Press, New York, NY.
TURING, A. M. (1936), "On computable numbers with an application to the
Entscheidungsproblem", Proceedings of the London Mathematical Society, 2(42), 230-265.
TURK, J. W. M. (1982), "Fast arithmetic operations on numbers and polynomials", in Lenstra and
Tijdeman (1982), pp. 43-54.
URBANEK, F. J. (1980), "An O (log n) algorithm for computing the n th element of the solution of
a difference equation", Information Processing Letters, 11(2), 66-67.
VALOIS, D. (1987), Algorithmes prohabilistes: une anthologie, Masters Thesis, Departement
d'informatique et de recherche operationnelle, Universite de Montreal.
VAZIRANI, U. V. (1986), Randomness, Adversaries and Computation, Doctoral dissertation,
Computer Science, University of California, Berkeley, CA.
VAZIRANI, U. V. (1987), "Efficiency considerations in using semi-random sources", Proceedings
VICKERY, C. W. (1956), "Experimental determination of eigenvalues and dynamic influence
coefficients for complex structures such as airplanes", Symposium on Monte Carlo Methods,
H. A. Meyer, ed., John Wiley & Sons, New York, NY, pp. 145-146.
WAGNER, R. A. and M. J. FISCHER (1974), "The string-to-string correction problem", Journal of
the ACM, 21(1), 168-173.
WARSHALL, S. (1962), "A theorem on Boolean matrices", Journal of the ACM, 9(1), 11-12.
WARUSFEL, A. (1961), Les nombres et leurs mysteres, Editions du Seuil, Paris.
WEGMAN, M.N. and J. L. CARTER (1981), "New hash functions and their use in authentication
and set equality", Journal of Computer and System Sciences, 22(3), 265-279.
WILLIAMS, H. (1978), "Primality testing on a computer", Ars Combinatoria, 5, 127-185.
WILLIAMS, J.W.J. (1964), "Algorithm 232: Heapsort", Communications of the ACM, 7(6),
347-348.
WINOGRAD, S. (1980), Arithmetic Complexity of Computations, SIAM, Philadelphia, PA.
WRIGHT, J. W. (1975), "The change-making problem", Journal of the ACM, 22(1), 125-128.
YAO, A. C. (1975), "An O (JE Ilog logI V 1) algorithm for finding minimum spanning trees", Infor-
mation Processing Letters, 4(1), 21-23.
YAO, A. C. (1982), "Theory and applications of trapdoor functions", Proceedings of 23rd Annual
IEEE Symposium on the Foundations of Computer Science, pp. 80-91.
YAO, F. F. (1980), "Efficient dynamic programming using quadrangle inequalities", Proceedings
YOUNGER, D. H. (1967), "Recognition of context-free languages in time n 3 ", Information and
Control, 10(2), 189-208.
ZIPPED, R. E. (1979), Probabilistic Algorithms for Sparse Polynomials, Doctoral dissertation,
Massachusetts Institute of Technology, Cambridge, MA.
Index
Aanderaa, S.O., 336, 344 Array, 2, 3, 20

Abadi, M., 275, 341 Articulation point, 174-176
Ackermann, W., 168, 341 Asymptotic notation, 6, 37-51
Ackermann's function, 63, 78, 166 with several arguments, 43
Acyclic graph, 22, 178-179, 185, 190, 199 See also Conditional asymptotic notation; Nested
Ad hoc sorting, 294, 296 asymptotic notation; Operations on asymptotic
Adder, 137 notation
Adel'son-Vel'skii, G.M., 35, 341 Asymptotic recurrence, 47
adjgraph, 22, 173, 203 At least quadratic, 301
Adleman, L.M., 140, 276, 341, 350 Atomic research, 274
Adversary argument, 337 Average height, 298
Aho, A.V., 35, 36, 168, 204, 222, 275, 291, 336, 377, Average. See Analysis of algorithms; Complexity
341 AVL tree, 25, 35
Ajtai, M., 141, 341
Algebraic transformation, 277
Algol, 205 Baase, S., 35, 222, 342
Algorithm, I Babaf, L., 274, 342
Algorithmics, xiii, 35 Bach, E., 276, 342
Alpha-beta pruning, 197-199, 204 Bachmann, P.G.H., 78, 342
alter-heap, 27 Backgammon, 198, 204
Amplification of stochastic advantage, 263-266, 276 backtrack, 188
Analysis of algorithms: Backtracking, 184, 185-189, 204, 248-252
on average, 7-9, 54, 118, 224, 238 Backus-Naur form, 205
empirical, 5 Balanced tree. See AVL tree; 2-3 tree
hybrid, 5, 18 Barometer, 52
theoretical, 5 Basic subalgorithm, 106
in worst case, 7-9 Batcher, K., 141, 342
Ancestor, 23, 207 Batcher's sorting circuit, 138
Ancestry in a rooted tree, 207-208 Baumert, L., 204, 346
APL, 4 Beauchemin, P., 275, 276, 342
Apple II, 251 Belaga, E.C., 222, 342
Approximation, 122, 227, 313, 336 Bellman, R.E., 167, 342
Arlazarov, V.L., 336, 342 Bellmore, M., 204, 342
353
354 Index
Bennett, C.H., 275, 337, 342, 345 Chromatic number, 320

Bentley, J.L., 35, 78, 141, 275, 342 Circuit. See Adder; Batcher's sorting circuit; Hamil-
Berge, C., 35, 204, 342 tonian circuit; Merge circuit; Tally circuit;
Berlekamp, E.R., 275, 342 Telephone switching
Berliner, H.J., 204, 342 Classic. See Multiplication
bfs, 183 Clause, 326
Biased probabilistic algorithm, 265 Clique, 321, 332
Bicoherent, 174-176 CNF, See Conjunctive normal form
Biconnected, 174-176 Cohen, H., 276, 344
binary-node, 25 Cole, F., 323
Binary search, 109-115, 243, 293, 298 Collision, 245
Binary tree, 24 Colouring. See Graph colouring
Binomial coefficient, 10, 143, 164 Comment, 2
Binomial queue. See Fibonacci heap Comparison sort, 293
Birthday surprise, 233, 246 Compiler generator, 206
Black, 196 Complementary problem, 322
Blum, L., 275, 342 Complex number, 65, 279-290
Blum, M., 140, 275, 342, 343 Complexity, 292-337
BM, 220
on average, 298-299
Boolean expression, 325
of circuits, 138
Boolean variable, 325
of graph problems, 304-308
Borodin, A.B., 35, 291, 336, 343
Bordvka, 0., 104, 343 of large integer arithmetic, 308-309, 314-315
Bottom-up technique, 142
of matrix problems, 302-304
Boyer, R.S., 216, 222, 343 of NP-complete problems, 315-335
Boyer-Moore algorithm, 216-222, 350 of polynomial arithmetic, 309-314
Branch-and-bound, 199-204 of sorting, 292-299
Brassard, G., 78, 140, 275, 276, 291, 336, 337, 342, Compound interest, 106
343
Conditional asymptotic notation, 45-47, 78
Bratley, P., 275, 343 Conditional probability, 267
Breadth-first search, 182-184 Confidence interval, 227
Brigham, E.O., 35, 291, 343 Conjunctive normal form, 326
Brown, C.A., 78, 167, 349 Connected component, 173
Buffon's needle, 228-230, 274 Co-NP, 322
Bunch, J., 35, 336, 343 Consistent probabilistic algorithm, 263
Buneman, P., 78, 343 Constructive induction, 48-51
Context-free language, 168
Continued fraction algorithm, 260
Candidate, 79 Contradiction, 325
Canonical object, 31 Cook, S.A., 325, 336, 344
Carasso, C., 275, 343 Cook's theorem, 325-328
Carlsson, S., 36, 336, 343 Cooley, J.M., 19, 290, 344
Carter, J.L., 275, 276, 343, 351 Coppersmith, D., 133, 140, 344
Catalan number, 148, 155, 167 Coprime, 15
Ceiling, 339 Corasick, M.J., 222, 341
Certificate of primality, 323, 336
Correct. See p-correct probabilistic algorithm
Chained matrix multiplication, 146-150, 167
correction loop, 315
Chandra, A.K., 337, 350
count, 234
Chang, L., 104, 168, 343
Change making, 80, 165, 207 countsort, 52, 293
Change of variable, 72-75 CRAY, 35, 291
Characteristic equation, 65, 69, 71 Crepeau, C., 276, 342
Cheriton, D., 104, 343 crude, 231
Chess, 4, 184, 194-198, 204 Cryptographic key, 128, 235
Child, 23 Cryptology, 124, 128-132, 140, 225, 235, 275, 276
Chinese remainder theorem, 290 Cubic algorithm, 6
Choice of leader, 260-262, 276 Curtiss, J.H., 275, 344
choose, 333 CYBER, 14-17, 250-252, 269
Christofides, N., 35, 104, 204, 343, 344 Cycle detection, 234
Index 355
Danielson, G.C., 19, 290, 344 Encrypted instance, 242, 275

darts, 230 Endomorphic cryptosystem, 235
Data Encryption Standard, 235, 275 Equivalence:
Data Strutures, 20-34 linear, 300
de Bruijn, N.G., 78, 344 many-one, 318
De Moivre's formula, 11, 16, 19, 67 polynomial, 317, 318
Deadline, 95-100 sense of Turing, 317
Deadlock, 261 Erdo s, P., 276, 344
Deap. See Double ended heap Euclid, 16, 57
Decision problem, 267, 318 Euclid's algorithm, 1, 16, 57, 135, 258, 262, 269
Decision tree, 292-299 extended, 254
pruned, 294 Euclidean travelling salesperson, 103, 104
trip, 293 Euler path, 203
valid, 293 Euler's function, 270
verdict, 293 Evaluation of polynomial:
Declaration, 2 with fast Fourier transform, 286, 291
delete-max, 28 with preconditioning, 209-211, 222
Demars, C., 35, 344 Even, S., 35, 204, 344
Denning, D.E.R., 140, 344 Eventually nondecreasing function, 46
Dense graph, 87, 91, 151, 306 Exact order of (e), 42
Deo, N., 35, 204, 349 Exchange of two sections in an array, 134-136, 140
Depth of circuit, 138 Execution time, 6
Depth of node, 24 Expected time, 224
Depth-first search, 171-182, 204 Exponential algorithm, 6
Determinant of matrix, 14-15, 35, 56-57 Exponentiation. See Large integer arithmetic
Devroye, L., 275, 344
Dewdney, A.K., 78, 344
dexpo, 130, 301, 334 Factor. See Nontrivial factor
dexpoiler, 131, 226, 242, 270 Factorization, 256-260, 276
Deyong, L., 204, 344 False witness of primality, 270, 276
dfs, 171 Faradzev, I.A., 336, 342
dfs', 182 Fast Fourier transform, 19, 279-280
Diffie, W., 128, 140, 344 Feasible, 79
Dijkstra, E.W., 104, 344 Feigenbaum, J., 275, 341, 345
Dijkstra's algorithm, 30, 87-92, 104, 151 Fermat's theorem, 226, 252
Dinic, E.A., 336, 342 FFT. See Fast Fourier transform
Directed graph, 21, 176-182 Fibonacci, 10
Discrete Fourier transform, 19, 279, 309 Fibonacci heap, 30, 104
Discrete logarithm, 129, 241 Fibonacci sequence, 10, 16, 17-18, 35, 58, 66, 136,
Disjoint sets, 30-34, 36, 60-63, 78, 84, 98-100 143
Disjunctive normal form, 336 See also De Moivre's formula
div, 3 fib 1, 16, 17-18, 58, 143, 192
Divide-and-conquer, 76, 105-141, 142, 279, 302-315 fib2, 17, 18, 58-59
Division. See Large integer arithmetic; Polynomial fib3, 17, 18, 19, 59, 127, 136
arithmetic find, 34
Dixon, J.D., 78, 276, 344 See also Disjoint sets
Dixon's algorithm, 45, 256-260 find-max, 28
dlogRH, 242 Finite state automaton, 167, 222
Double ended heap, 36 Fischer, M.J., 168, 336, 345, 351
Dreyfus, S.E., 167, 342 Flajolet, P., 275, 345
Dromey, R.G., 35, 344 Floating point, 11, 124, 210
Dynamic programming, 142-168, 192-196, 298 Floor, 339
Floyd, R.W., 140, 167, 275, 343, 345, 350
Floyd's algorithm, 150-153, 167, 305, 306
Efficient algorithm, I1-12, 316 FORTRAN, 13
Eigenvalue, 237 Four colour theorem, 323
Eigenvector, 237 Fourier transform. See Dicrete Fourier transform
Eight queens problem, 185-188, 225, 248-252, 275 Fox, B.L., 275, 343, 345
Elementary operation, 9-11 Fredman, M.L., 36, 104, 167, 336, 345
356 Index
Freivalds, R., 276, 345 Hell, P., 104, 346

Full adder. See Adder Hellman, M.E., 128, 140, 344, 346
Furman, M.E., 336, 345 Heuristic, 100-104, 195, 260, 336
Hidden constant, 6
hitormiss, 230
Game graph, 189-198, 204 Hoare, C.A.R., 13, 116, 140, 346
Games. See Backgammon; Chess; Nim See also Quicksort
Gardner, M., 140, 337, 345 Hode. See Robyn Hode
Garey, M.R., xiv, 336, 345 Homogeneous recurrence, 65-68
Gauss-Jordan elimination, 15, 259 Hopcroft, J.E., 35, 36, 78, 140, 167, 168, 204, 222,
gcd. See Greatest common divisor
275, 291, 336, 337, 341, 343, 346
Generation of permutation, 186, 273 Homer's rule, 209
Generator of cyclic group, 128
Horowitz, E., 35, 104, 291, 336, 337, 347, 350
genrand, 267
Hu, T.C., 167, 347
Gentleman, W.M., 290, 345
Huang, M.-D.A., 276, 341
Gilbert, E.N., 167, 345
Hudson's Bay, 165
Gleick, J., 291, 345
God, 64 Huffman code, 104
Godbole, S., 167, 345
Gold, B., 291, 349
Golden ratio, 16, 67 Implicit graph, 184, 185-202
Goldner, A., 204, 350 Indentation, 3
Goldwasser, S., 276, 345 Infinite graph, 184, 189
Golomb, S., 204, 346 Inherently difficult problem, 316, 337
Gondran, M., 35, 204, 346 Inhomogeneous recurrence, 68-72
Gonnet, G.H., 35, 36, 336, 346 Inorder, 170
Good, I.J., 204, 346 insert, 7, 115, 117
goodproduct, 274 Insertion sort, 7, 13, 42, 53-54, 115, 295-299
Goutier, C., 276, 342 insert-node, 28
Graham, R.L., 104, 346 Instance, 4
Graph, 20, 21-22, 81-92, 101-104, 150-153, Instant Insanity, 189
169-204, 304-308 Integration. See Numeric integration
Graph colouring, 101-102, 320, 330-332 Internal node, 23
Greatest common divisor, 15, 128, 135, 257, 315 Invariance. See Principle of invariance
Greedy algorithm, 79-104, 159, 166, 248-252 Inverse:
Greene, D.H., 78, 346 Fourier transform, 280-284
Gries, D., 35, 140, 141, 346 modulo p, 254
of a matrix, 237, 274, 302-304
of a polynomial, 310
Haken, D., 78, 342 of an integer, 315
Hall, A., 274, 346 Isthmus-free. See Bicoherent
Halting problem, 332 Itai, A., 276, 347
Hamiltonian circuit, 315, 319, 323, 332 Iterated logarithm, 63, 84, 100
Hamiltonian path, 104
Hammersley, J.M., 275, 346
Handscomb, D.C., 275, 346 Janko, W., 275, 347
Hanoi. See Towers of Hanoi Jarm'k, V., 104, 347
Hardy, G.H., 276, 346 Jensen, K., 35, 347
Harel, D., xiii, 35, 346 Johnson, D.B., 36, 104, 347
Harmonic series, 43, 54 Johnson, D.S., xiv, 336, 345
Harrison, M.C., 222, 346
Hashing, 245-247
Heap, 20, 25-30, 35-36, 55-56, 85, 91-92, 151, 202 Kahn, D., 140, 347
See also Double ended heap; Fibonacci heap Kaliski, B.S., 275, 347
Heap property, 26, 85 Kanada, Y., 291, 347
Heapsort, 13, 25, 30, 56, 115, 119, 120, 293-298, 336 Kannan, S., 276, 343
heapsort, 56 Karatsuba, A., 140, 347
Height, 24 Karp, R., 36, 168, 276, 291, 336, 346, 347
See also Average height Kasimi, T., 168, 347
Held, M., 168, 346 Kepler, J., 277
Index 357
Kerr, L.R., 140, 346 Linear:

Key for' , 254, 271 algebra, 15, 237, 275
Kilian, J., 275, 276, 341, 345 algorithm, 6, 301
Klamkin, M.S., 275, 347 equations, 237
Kleene, S.C., 167, 347 equivalence, 300
KMP, 215 programming, 202, 227
Knapsack, 140, 315 reduction, 300
Knight's tour problem, 104 lisgraph, 22, 91, 172
Knuth, D.E., 35, 78, 140, 167, 168, 204, 213, 222, List, 20-21
275, 291, 346, 347, 348 Literal, 326
Knuth-Morris-Pratt's algorithm, 213-215, 216, 222 Load factor, 245
Koksoak, 165 Logarithm, 7, 277
Koml6s, J., 141, 341 See also Discrete logarithm; Iterated logarithm
Konig, H., 19, 290, 350 Longest simple path, 144, 315
Korsh, J., 104, 168, 343 Los Alamos. See Atomic research
k-promising, 187, 188, 248-251 Lower bound theory, 336
Kranakis, E., 140, 276, 348 lowest, 175
Kronrod, M.A., 140, 336, 342 Lucas, F.E.A., 323
Kruskal, J.B., Jr., 104, 348 Lueker, G.S., 78, 348
Kruskal's algorithm, 82-85, 87, 92, 104
k-smooth integer, 258
k th smallest element, 120, 121-124, 140, 224, Macroprocessor, 211
238-240, 241 maj, 268
maj 2, 268
majMC, 269
Label. See Canonical object Majority element, 137, 268-269
Labyrinth, 185, 203-204 make-heap, 29, 55
Lanczos, C., 19, 290, 344 Manders, K., 276, 341
Landis, E.M., 35, 341 Many-one. See Equivalence; Reduction
Large integer arithmetic; 13-14, 124-132, 286-290, Marienbad. See Nim
308-315 Marsh, D., 168, 348
addition, 124 Martin, G.N., 275, 345
division, 128, 308, 309, 315 Matrix. See Chained matrix multiplication; Deter-
exponentiation, 128-132, 140, 252-253, 301 minant of matrix; Inverse of a matrix; Multipli-
greatest common divisor, 128, 315 cation of matrices; Reductions among martix
multiplication, 1-4, 13-14, 124-128, 286-290, problems; Strassen's algorithm; Symmetric
308 matrix; Triangular matrix; Unitary matrix
reductions, 308-315 McDiarmid, C.J.H., 36, 348
square root, 128, 308, 315 Median, 117, 119, 120-124, 140, 224, 238-240, 275
squaring, 309 Melhom, K., 35, 348, 349
Las Vegas algorithm, 227, 228, 247-262, 274, 333 Memory function, 162-164, 168, 193, 272
LauriBre, J-L., 167, 348 Memory space, 5, 7
Lawler, E.L., 35, 204, 348 Merge circuit, 138, 139
Leader. See: Choice of leader
merge, 33
Leaf, 23
See also Disjoint sets
Lecarme, 0., 35, 348
Merge sort, 13, 115-116, 297
Leclerc, G.L., 228, 274, 348
mergesort, 115
Left-hand child, 24
Lehmer, D.H., 275, 348
Mersenne, M., 323
Lenstra, A.K., 276, 344 Metropolis, IN., 274, 349
Lenstra, H.W., Jr., 260, 276, 348, 349, 351 Meyer, A.R., 336, 344, 345, 351
Level, 24 Micali, S., 275, 276, 343, 345
Levin, G., 35, 141, 346 Michie, D., 168, 346, 348, 349
Levin, L., 325, 336, 348 Miller, G., 276, 341, 342, 347
Levy, L., 78, 343 Minimax principle, 184, 195-197, 204
Lewis, H.R., 35, 348 Minimization of waiting time, 92-94
Lewis, P.A., 290, 344 Minimum spanning tree, 81-87, 92, 103, 169
Lexicographic sorting, 293, 336 Minoux, M., 35, 204, 346
L'Hopital's rule, 40, 78 mod, 3
Limit, 40, 78 Modular arithmetic, 128-131, 252-256, 290
358 Index
Monet, S., 140, 291, 337, 343 Papadimitriou, C.H., 35, 204, 348, 349
Monic polynomial, 136, 209-211, 286, 310 parent, 23
Monier, L., 276, 349 Pascal (computer language), 3, 8-17, 35, 106, 205
Monte Carlo algorithm, 227, 262-274, 322 Pascal's triangle, 143, 146
See also Numeric integration; Quasi Monte Carlo Path compression, 33, 34, 62
Montgomery, P.L., 276, 349 Pattern, 211, 212-222
Moore, E.F., 167, 345 p-correct probabilistic algorithm, 263
Moore, J.S., 216, 222, 343 Peralta, R.C., 275, 349
Morris, J.H., 222, 348 Percolate, 26, 27, 28-30, 91
Multiplication. See also Large integer arithmetic; percolate, 27
Polynomial arithmetic Permutation. See Generation of Permutation; Random
a la russe, 2-4, 5, 13-14, 35, 124 permutation
classic, 1, 2, 13-14, 124-132, 278 Pi (it), 124, 228-234, 281, 290-291
of matrices, 132-133, 140, 146-150, 237, 274, Pippenger, N., 336, 349
302-308 Pivot, 116, 120, 224, 239-240
Munro, J.I., 35, 36, 291, 336, 337, 343, 346 pivot, 117
Planar graph, 204, 332
Pohl, 1., 141, 349
Napier, J., 277 Pointer, 3, 20
n-ary-node, 24 Pollard, J.M., 276, 291, 349
Nebut, J.-L., 35, 348 Polynomial algorithm, 6, 316
Nemhauser, G., 167, 204, 342, 349 Polynomial arithmetic:
Nested asymptotic notation, 45 division, 309-314
Newman, D.J., 275, 347 evaluation, 209-211, 222, 286, 291
Newton's binomial, 10 interpolation, 136, 286
Newton's method, 313, 315 multiplication, 136, 278, 284-286, 309-314
Nievergelt, J., 35, 204, 349 reductions, 308-314
Nilsson, N., 35, 204, 349 Polynomial. See Equivalence; Reduction
Nim,190-194 Pomerance, C., 276, 336, 341, 342, 344, 349
Node, 20, 21 postnum, 180, 208
node, 21 Postorder, 170, 174, 208
Nondecreasing. See Eventually nondecreasing function Pratt, V.R., 140, 222, 336, 343, 348, 349
Non-determinism, 332-335, 336 Precomputation, 205, 211-222
Non-deterministic algorithm, 333 Preconditioning, 154-159, 166, 193, 205-211, 215,
Nontrivial factor, 225, 228, 256, 334 240-242, 278
NP, 321, 332-335 prenum, 173, 208
Preorder, 170, 208
NP-complete, 102, 103, 153, 324-332
Prim, R,C., 104, 349
NP-completeness, 315-335, 336-337
Prim's algorithm, 85-87, 92
Number theory. See Theory of numbers
Primality testing, 9, 256, 269-271, 276, 334
Numeric integration:
See also Certificate of primality
Monte Carlo, 230-232, 275
primeND, 334
multiple, 232, 275
Principal root of unity, 281
trapezoidal, 231
Principle of invariance, 6, 38
Numerical probabilistic algorithm, 227, 228-237, 275
Principle of optimality, 143, 144-159, 298
Priority list, 25, 27, 30, 199
Objective function, 79-80 Probabilistic algorithm. See Biased probabilistic algo-
obstinate, 248 rithm; Consistent probabilistic algorithm; Las
Ofman, Y., 140, 347 Vegas algorithm; Monte Carlo algorithm;
Omega (a), 41 Numerical probabilistic algorithm; p-correct
One-way equality, 51, 78 probabilistic algorithm; Sherwood algorithm;
Operations on asymptotic notation, 43-45 Unbiased probabilistic algorithm
Optimal graph colouring, 315, 320, 330-332 Probabilistic counting, 232-237, 275
Optimal search tree, 154-159, 167-168, 207 Probability of success, 247, 263, 333
Optimality. See Principle of optimality Problem, 4
Order of (0), 6, 37 Program, 2
Programming language, 2
Promising, 79
P, 318 See also k-promising
Pan, V., 133, 140, 349 Proof space, 320
Index 359
Proof system, 320 Ring of processors, 260-262

Pruned. See Decision tree Rivest, R.L., 140, 275, 343, 347, 350
Pruning. See Alpha-beta pruning Robert, J.-M., 275, 342
pseudomed, 122 Robson, J.M., 204, 350
Pseudomedian, 122-124, 224, 239 Robyn Hode, 228
Pseudoprime, 270, 320 Rodeh, M., 276, 347
Pseudorandom generation, 226, 275 Roman figures, 277
Pseudorandom walk, 234 rootLV, 254
Purdom P.W., Jr., 78, 167, 349 rootLV2, 255
Root of a tree, 23
Rooted tree, 23-25
Quadratic: Rosenthal, A., 204, 350
algorithm, 6, 301 Rubik's Cube, 189
nonresidue, 252 Rudin, W., 78, 350
residue, 252, 257 Rumely, R.S., 276, 341
sieve, 260 Runge, C., 19, 290, 350
Quasi Monte Carlo, 232, 275 Rytter, W., 222, 350
Queens, 188, 250
QueensLV, 249
Queue, 182, 199, 227 Sahni, S., 35, 104, 291, 336, 337, 347, 350
See also Binomial queue Sande, G., 290, 345
Quicksort, 13, 116-119, 140, 238, 240, 247, 293-295 SAT. See Satisfiability
quicksort, 117 Satisfiability, 316, 325, 326-332, 336
Saxe, J.B., 78, 342
Rabin, M.O., 275, 276, 347, 349 Scaling factor, 311-313
Rabin's algorithm, 270, 276 Scheduling, 92-100, 139-140
Rabiner, L.R., 291, 349 Schonhage, A., 290, 291, 350
Radix sort, 293 Schonhage-Strassen algorithm, 290
Random function, 272 Schrage, L.E., 275, 343
Random permutation, 273 Schwartz, E.S., 104, 350
Range transformation, 75 Schwartz, J., 276, 350
Rank, 34, 60-63, 242-245 Search tree, 25, 154-159, 207
record, 3 Searching:
Recurrence. See Asymptotic recurrence; Homogeneous for a string, 211-222
recurrence; Inhomogeneous recurrence; Resolu- in a graph, 171-184
tion of recurrences in a sorted array. See Binary search
Recursivity, 15, 106 in a sorted list, 242-245
Reduction: in a tree, 25, 154-159, 170-171, 207
linear, 300 Sedgewick, R., 35, 350
many-one, 318 Seed of a pseudorandom generator, 226, 261
polynomial, 317, 318 select, 7
sense of Turing, 317 Selection. See kth smallest element
Reductions among: selection, 120
arithmetic problems, 308-309, 314-315 Selection function, 79, 80
graph problems, 304-308 selectionRH, 239
matrix problems, 302-304 Selection sort, 7, 8, 13, 52, 54, 115, 295-299
NP-complete problems, 315-335 Sequential search, 110
polynomial problems, 309-314 series, 146
Reed, B.A., 36, 348 Set difference, 44
Reflexive transitive closure, 153, 304-308 Set equality, 271-273, 276
Regular expression, 167 Shakespeare, 235
Reingold, E.M., 35, 204, 349 Shallit, J., 276, 342
Relatively prime. See coprime Shamir, A., 35, 140, 350
repeatMC, 266 Shamos, I., 141, 342
Resolution of recurrences, 65-78 Shanks, D., 276, 350
return, 3 Sherman, AT., 275, 347
RH, 241 Sherwood algorithm, 227, 228, 238-247, 274
Right-hand child, 24 Shing, M.R., 167, 347
360 Index
Shortest path, 30, 87-92, 104, 150-153, 167, Switch. See Telephone switching
304-308 Symbolic manipulation of polynomials.
Shortest simple path, 153 See Polynomial arithmetic
Shub, M., 275, 342 Symmetric matrix, 22, 86, 103, 153, 304
shuffle, 241 Syntactic analysis, 168, 203, 205
Sibling, 23 Szemeredi, E., 141, 341
Sift-down, 27, 28-30, 55-56, 91
sift-down, 27
sift-up. See Percolate tablist, 20
Signal processing, 19, 279, 280 Tally circuit, 137
Signature, 211-213, 222, 273 Tamura, Y., 291, 347
Simple path, 144, 153, 315 Target string, 211, 213-222
Simplex, 227 Tarjan, R.E., xiv, 35, 36, 78, 104, 140, 167, 204, 343,
Simplification, 106, 109-115, 121, 128-132, 134-136 345, 346, 350, 351
Simula, 205 Tautology, 325, 326-327, 336
Simulation, 28, 227, 275 Telephone switching, 137, 336
Sink, 203 Text editor, 211
Size of circuit, 138 Text processing, 211
Size of instance, 5 Theory of numbers, 140, 276
Sloane, N.J.A., 167, 168, 350 Threshold for asymptotic notation, 37, 39, 43
slow-make-heap, 28, 56 Threshold for divide-and-conquer, 106
Smooth algorithm, 301 determination, 107-109
Smooth function, 46, 74, 301 ultimate, 108, 123
Smooth integer. See k-smooth integer Tijdeman, R., 348, 349, 351
Smooth problem, 301 Timetable, 139-140
Sobol', I.M., 275, 350 Top-down technique, 142
Solovay, R., 276, 350 Topological sorting, 178-179, 190
Sort. See Ad hoc sorting; Batcher's sorting circuit; Tournament, 139-140, 144-146
Comparison sort; countsort; Heapsort; Insertion Towers of Hanoi, 64, 69, 78
sort; Lexicographic sorting; Quicksort; Radix Transformation:
sort; Selection sort; Topological sorting; of the domain, 277-291
Transformation sort function, 277
Source, 30, 87, 150 sort, 293, 336
Sparse graph, 87, 91, 151 Transformed domain, 277
Special path, 87, 88-90 Transformed function, 277
Splitting, 256, 276 Trapezoidal algorithm. See Numeric integration
Square root. See also Large integer arithmetic Traub, J.F., 345, 349, 351
modulo n, 257 Travelling salesperson, 102, 103-104, 159-162, 168,
modulo p, 252-256 199-202, 204, 315, 319
Stack, 21, 107, 182, 199 See also Euclidean travelling salesperson
Stanat, D.F., 275, 342 Traversal of tree. See Inorder; Postorder;
Standish, T.A., 35, 350 Preorder; Searching
Statistical test, 226, 275 Tree, 22
Steele, J.M., 275, 342 See also Ancestry in a rooted tree, Balanced tree,
Steiglitz, K., 35, 204, 349 Binary tree, Decision tree, Minimum spanning
Stinson, D.R., 35, 141, 350 tree, Optimal search tree, Rooted tree,
Stirling's formula, 233 Searching in a tree, Traversal of tree
Stochastic advantage. treenode, 23
See Amplification of stochastic advantage Triangle inequality, 103
Stochastic preconditioning, 240-242 Triangular matrix, 302
Stockmeyer, L.J., 336, 337, 350 Trip. See Decision tree
Stone, H.S., 35, 350 TRS-80, 35
Strassen, V., 15, 35, 132, 140, 276, 290, 291, 350 Tukey, J.W., 19, 290, 344
Strassen's algorithm, 132-133, 140, 305, 308 Turing, A.M., 337, 351
String. See Searching See also Equivalence; Reduction
Strong pseudoprime, 270, 320 Turing machine, 327
Strongly connected, 179-182 Turk, J.W.M., 291, 351
Strongly quadratic, 301 2-edge-connected. See Bicoherent
Supra quadratic, 301 2-3 tree, 25, 35
Index 361
Ulam, S., 274, 349 Warusfel, A., 35, 351

Ullman, J.D., 35, 36, 78, 167, 168, 204, 222, 275, 291, Wegman, M.N., 275, 276, 343, 351
336, 337, 341, 346 Welch, P.D., 290, 344
Ultimate. See Threshold for divide-and-conquer Well-characterized problem, 228
Unarticulated. See Biconnected White, 196
Unbiased probabilistic algorithm, 276 Widget, 331
Undecidable problem, 317, 332, 337 Williams, H., 276, 351
Undirected graph, 21, 171-176 Williams, J.W.J., 13, 36, 56, 78, 351
Uniform random generation, 225 See also Heap; Heapsort
Unit cost. See Elementary operation Wilson's theorem, 9, 10
Unitary matrix, 304 Winograd, S., 133, 140, 337, 344, 351
Universal hashing, 245-247, 272, 275-276 Wirth, N., 35, 347
Unpredictable pseudorandom generator, 275, 342 Wood, D.W., 204, 348
Urbanek, F.J., 35, 141, 351 wordcnt, 236
Ushiro, Y., 291, 347 World Series, 144-146, 263
Worst case. See Analysis of algorithms
Wright, E.M., 276, 346
Valid. See Decision tree Wright, J.W., 104, 168, 351
Valois, D., 276, 351
VAX, 8, 13
Vazirani, U.V., 275, 351
Verdict. See Decision tree Yao, A.C., 104, 275, 351
Vickery, C.W., 275, 351 Yao, F.F., 167, 351
Virtual initialization, 164, 168, 193, 273 Yoshino, S., 291, 347
Younger, D.H., 168, 351
Wagner, R.A., 168, 351

Warshall, S., 167, 351 Zippel, R.E., 276, 351
Warshall's algorithm, 153, 308 Zuffellato, D., 140, 291, 343
11fl
Mcpo)[PT
TIPMT0T(
GILLES BRASSARD PAUL BRATLEY
The computer explosion has made calculations once thought impossible
seem routine. However, another factor has had an even more important
effect in extending the frontiers of feasible computation: the use of
efficient algorithms.
Now, this innovative new book gives readers the basic tools they need to
develop their own algorithms--in whatever field of application they may be
required!
CONTENT HIGHLIGHTS:
Concentrates on the techniques needed to design and analyze
algorithms.
Details each technique in full.
Illustrates each technique with concrete examples of algorithms taken
from such different applications as optimization, linear algebra,
cryptography, operations research, symbolic computation, artificial
intelligence, numerical analysis, computing in the humanities, among
others.
Presents real-life applications for most algorithms.
Contains approximately 500 exercises-many of which call for an
algorithm to be implemented on a computer so that its efficiency may be
measured experimentally and compared to the efficiency of alternative
solutions.
PRENTICE HALL, Englewood Cliffs, NJ 07632
ISBN 0-13-023243-2

Algorithmics

Uploaded by

Copyright:

Available Formats

Algorithmics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Algorithmics

Uploaded by

Copyright:

Available Formats

I [P

Gilles Brassard and Paul Bratley

PRENTICE HALL, Englewood Cliffs, New Jersey 07632

Editorial/production supervision: Editing, Design & Production, Inc.

© 1988 by Prentice-Hall, Inc.

All rights reserved. No part of this book may be

Printed in the United States of America

PRENTICE-HALL INTERNATIONAL (UK) LIMITED. London

1.1. What Is an Algorithm? 1

1.2. Problems and Instances 4

1.4. Average and Worst-Case Analysis 7

1.5. What Is an Elementary Operation? 9

1.6. Why Do We Need Efficient Algorithms? 11

1.7. Some Practical Examples 12

1.8. When Is an Algorithm Specified? 19

1.9. Data Structures 20

2 Analysing the Efficiency of Algorithms 37

2.1. Asymptotic Notation 37

2.2. Analysis of Algorithms 52

2.4. References and Further Reading 78

3.2.1. Minimal Spanning Trees, 81

3.4. Greedy Heuristics 100

3.5. References and Further Reading 104

4 Divide and Conquer 105

4.1. Introduction 105

4.2. Determining the Threshold 107

4.3. Binary Searching 109

4.4. Sorting by Merging 115

4.5. Quicksort 116

4.6. Selection and the Median 119

4.7. Arithmetic with Large Integers 124

4.8. Exponentiation : An Introduction to Cryptology 128

4.9. Matrix Multiplication 132

4.10. Exchanging Two Sections of an Array 134

4.11. Supplementary Problems 136

4.12. References and Further Reading 140

5 Dynamic Programming 142

5.1. Introduction 142

5.2. The World Series 144

5.3. Chained Matrix Multiplication 146

5.4. Shortest Paths 150

5.5. Optimal Search Trees 154

5.6. The Travelling Salesperson Problem 159

5.7. Memory Functions 162

5.8. Supplementary Problems 164

5.9. References and Further Reading 167

6 Exploring Graphs 169

6.1. Introduction 169

6.2. Traversing Trees 170

6.3. Depth-First Search : Undirected Graphs 171

6.3.! Articulation Points, 174

6.4. Depth-First Search : Directed Graphs 176

6.5. Breadth-First Search 182

6.6. Implicit Graphs and Trees 184

7 Preconditioning and Precomputation 205

7.1. Preconditioning 205

7.2. Precomputation for String-Searching Problems 211

7.3. References and Further Reading 222

8 Probabilistic Algorithms 223