Sei sulla pagina 1di 432

Cornell University

Boston San Francisco NewYork


London Toronto Sydney Tokyo Singapore Madrid
Mexico City Munich Paris Cape Toxvn Hong Kong Montreal

Acquisitions Editor: Matt Goldstein


Project Editor: Maite Suarez-Rivus
Production Supervisor: MariIyn Lloyd
Marketing Manager: MichelIe Brown
Marketing Coordinator: Yake Zavracky
Project Management: Windfall Sofi-tvare
Composition: Windfall Software, using ZzTEX
Copyeditor: Carol Leyba
Technical Illustration: Dartmouth Publishing
Proofreader: Jennifer McClain
Indexer: Ted Laux
Cover Design: Yoyce Cosentino Wells
Cover Photo: 2005 Tim Laman / National Geographic. A pair of weaverbirds work
together on their nest in Africa.
Prepress and Manufacturing: Caroline Fell
Printer: Courier West~ord
Access the latest information about Addison-Wesley rifles from our World Wide Web
site: http://www.aw-bc.com/computing
Many of the designations used by manufacturers and sellers to distinguish their
products are claimed as trademarks. Where those designations appear in this book,
and Addison-Wesley was aware of a trademark claim, the designations have been
printed in initial caps or all caps.
The programs and applications presented in this book have been included for their
instructional value. They have been tested with care, but are not guaranteed for any
particular purpose. The publisher does not offer any warranties or representations, nor
does it accept any liabilities with respect to the programs or applications.
Library of Congress Cataloging-in-Publication Data
Kleinberg, Jon.
Algorithm design / Jon Kleinberg, l~va Tardos.--lst ed.
p. cm.
Includes bibliographical references and index.
ISBN 0-321-29535-8 (alk. paper)
1. Computer algorithms. 2. Data structures (Computer science) I. Tardos, l~va.
II. Title.
QA76.9.A43K54 2005
005.1--dc22

2005000401

Copyright 2006 by Pearson Education, Inc.


For information on obtaining permission for use of material in this work, please
submit a written request to Pearson Education, Inc., Rights and Contract Department,
75 Arlington Street, Suite 300, Boston, MA 02116 or fax your request to (617) 848-7047.
All rights reserved. No part of this publication may be reproduced, stored in a
retrieval system, or transmitted, in any form or by any means, electronic, mechanical,
photocopying, recording, or any toher media embodiments now known or hereafter to
become known, without the prior written permission of the publisher. Printed in the
United States of America.
ISBN 0-321-29535-8
2 3 4 5 6 7 8 9 10-CRW-08 07 06 05

About the Authors

3on Kleinberg is a professor of Computer Science at


Cornell University. He received his Ph.D. from M.I.T.
in 1996. He is the recipient of an NSF Career Award,
an ONR Young Investigator Award, an IBM Outstanding Innovation Award, the National Academy of Sciences Award for Initiatives in Research, research fellowships from the Packard and Sloan Foundations,
and teaching awards from the Cornell Engineering
College and Computer Science Department.
Kleinbergs research is centered around algorithms, particularly those concerned with the structure of networks and information, and with applications
to information science, optimization, data mining, and computational biology. His work on network analysis using hubs and authorities helped form the
foundation for the current generation of Internet search engines.
fiva Tardos is a professor of Computer Science at Cornell University. She received her Ph.D. from E6tv6s
University in Budapest, Hungary in 1984. She is a
member of the American Academy of Arts and Sciences, and an ACM Fellow; she is the recipient of an
NSF Presidential Young Investigator Award, the Fulkerson Prize, research fellowships from the Guggenhelm, Packard, and Sloan Foundations, and teaching awards from the Cornell Engineering College and
Computer Science Department.
Tardoss research interests are focused on the design and analysis of
algorithms for problems on graphs or networks. She is most known for her
work on network-flow algorithms and approximation algorithms for network
problems. Her recent work focuses on algorithmic game theory, an emerging
area concerned with designing systems and algorithms for selfish users.

Contents

About the Authors

Preface

Introduction: Some Representative Problems


I. 1 A First Problem: Stable Matching,
1).
19
Exercises
Notes and Further Reading )8

Basics of Algorithm Analysis


29
2.1
Computational Tractability 29
2.2
Asymptotic Order of Growth 35
2.3
Implementing the Stable Matching Algorithm Using Lists and

2.4
2.5

Arrays 42
A Survey of Common Running Times 47
57
65
Exercises 67
Notes and Fm-ther Reading 70

Graphs

73
Basic Definitions and Applications 73
Graph Connectivity and Graph Traversal 78
Implementing Graph Traversal Using Queues and Stacks 87
Testing Bipartiteness: An Application of Breadth-First Search
94
Connectivity in Directed Graphs 97

Contents

Contents

3.6

6.4
6.5

Directed Acyclic Graphs and Topological Ordering 99


104
Exercises 107
Notes and Further Reading 112

6.6
6.7

4 Greedy Algorithms

4.7
4.8
*4.9

Interval Scheduling: The Greedy Algorithm Stays Ahead 116


Scheduling to Minimize Lateness: An Exchange Argument 125
Optimal Caching: A More Complex Exchange Argument 131
Shortest Paths in a Graph 137
The Minimum Spanning Tree ProbJem 142
Implementing Kruskals Algorithm: The Union-Find Data
Structure 151
Clustering 157
Huffman Codes and Data Compression 161
Minimum-Cost Arborescences: A Multi-Phase Greedy
Algorithm 177
183
Exercises 188
Notes and Further Reading 205

5 Divide and Conquer


5.1
5.2
5.3
5.4
5.5
5.6

209

A First Recurrence: The Mergesort Algorithm 210


Further Recurrence Relations 214
Counting Inversions 221
Finding the Closest Pair of Points 225
Integer Multiplication 231
234
242
Exercises 246
Notes and Further Reading 249

6.8
6.9
* 6.10

Subset Sums and Knapsacks: Adding a.,~able 266


RNA Secondary Structure: Dynarmc~gramming over
Intervals 272
Sequence Alignment 278
Sequence Alignment in Linear Space via Divide and
Conquer 284
Shortest Paths in a Graph 290
297
Negative Cycles in a Graph 301
307
Exercises 312
Notes and Further Reading 335

337
The Maximum-Flow Problem and the Ford-FulkersOn
Algorithm 338
7.2
Maximum Flows and Minimum Cuts in a Network 346
7.3
Choosing Good Augmenting Paths 352
* 7.4
The Preflow-Push Maximum-Flow Algorithm:, 357
7.5
A First Application: The Bipartite Matching Problem 367
7.6
373
7.7
378
7.8
Survey Design 384
7.9
Airline Scheduling 387
7.!0 Image Segmentation 391
\
7.11
7.12 Baseball Elimination 400
"7.!3 A Further Direction: Adding Costs to the Matching Problem,~) 404

Solved Exercises 411


Exercises 415
Notes and Further Reading 448
2S1

Weighted Interval Scheduling: A Recursive Procedure 252


Principles of Dynamic Programming: Memoization or Iteration
over Subproblems 258
6.3 Segmented Least Squares: Multi-way Choices 26~

6.1
6.2

* The star indicates an optional section. (See the Preface for more information about the relationships
among the chapters and sections.)

451
8.1
8.2
8.3
8.4
8.5
8.6
8.7

Polynomial-Time Reductions 452


Reductions via "Gadgets": The Satisfiabflity Problem 459
Efficient Certification and the Definition of NP 463
NP-Complete Problems 466
Sequencing,,Problems 473
Partitioning Problems 481
Graph Coloring 485

Contents

Contents
8.8
8.9
8.10

9 PSPACE: A Class of Problems beyond NP

9.4

12.1
12.2
12.3
12.4
12.5
12.6
12.7

531

13
553

Finding Smal! Vertex Covers 554


Solving NP-Hard Problems on Trees 558
Coloring a Set of Circular Arcs 563
Tree Decompositions of Graphs 572
584
591
Exercises 594
Notes and Further Reading 598

11 Approximation Algorithms

Arbitrarily Good Approximations: The Knapsack Problem 644


649
Exercises 651
Notes and Further Reading 659

12 Local Search

PSPACE 531
Some Hard Problems in PSPACE 533
Solving Quantified Problems and Games in Polynomia!
Space 536
Solving the Planning Problem in Polynomial Space 538
543
547
Exercises 550
Notes and Further Reading 551

10 Extending the Limits of Tractability


10.!
10.2
10.3
* 10.4
* 10.5

11.8

Numerical Problems 490


Co-NP and the Asymmetry of NP 495
A Partial Taxonomy of Hard Problems 497
500
Exercises 505
Notes and Further Reading 529

599

11.1 Greedy Algorithms and Bounds on the Optimum: A Load


Balancing Problem 600
606
11.3 Set Cover: A General Greedy Heuristic 612
11.4 The Pricing Method: Vertex Cover 618
11.5 Maximization via the Pricing Method: The Disioint Paths
Problem 624
11.6 Linear Programming and Rounding: An Application to Vertex
Cover 630
* 11.7 Load Balancing Revisited: A More Advanced LP Application 637

661

The Landscape of an Optimization Problem 662


The Metropolis Algorithm and Simulated Annealing 666
An Application of Local Se_arch to Hopfield Neural Networks
676
Choosing a Neighbor Relation 679
Classification via Local Search 681
690
700
Exercises 702
Notes and Further Reading 705

707
A First Application: Contention Resolution 708
Finding the Global Minimum Cut 714
Random Variables and Their Expectations 719
A Randomized Approximation Algorithm for MAX 3-SAT 724
Randomized Divide and Conquer: Median-Finding and
Quicksort 727
13.6 Hashing: A Randomized Implementation of Dictionaries 734
13.7 Finding the Closest Pair of Points: A Randomized Approach 741
13.8 Randomized Caching 750
13.9 Chernoff Bounds 758
13.10 Load Balancing 760
13.1! Packet Routing 762
13.12 Background: Some Basic ProbabiLity Definitions 769
776
Exercises 782
Notes and Further Reading 793
13.1
13.2
13.3
13.4
13.5

795
805
815

Algorithmic !deas are pervasive, and their reach is apparent in examples both
within computer science and beyond. Some of the major shifts in Internet
routing standards can be viewed as debates over the deficiencies of one
shortest-path algorithm and the relative advantages of another. The basic
notions used by biologists to express similarities among genes and genomes
have algorithmic definitions. The concerns voiced by economists over the
feasibility of combinatorial auctions in practice are rooted partly in the fact that
these auctions contain computationally intractable search problems as special
cases. And algorithmic notions arent just restricted to well-known and longstanding problems; one sees the reflections of these ideas on a regular basis,
in novel issues arising across a wide range of areas. The scientist from Yahoo!
who told us over lunch one day about their system for serving ads to users was
describing a set of issues that, deep down, could be modeled as a network flow
problem. So was the former student, now a management consultant working
on staffing protocols for large hospitals, whom we happened to meet on a trip
to New York City.
The point is not simply that algorithms have many applications. The
deeper issue is that the subject of algorithms is a powerful lens through which
to view the field of computer science in general. Algorithmic problems form
the heart of computer science, but they rarely arrive as cleanly packaged,
mathematically precise questions. Rather, they tend to come bundled together
with lots of messy, application-specific detail, some of,it essential, some of it
extraneous. As a result, the algorithmic enterprise consists of two fundamental
components: the task of getting to the mathematically clean core of a problem,
and then the task of identifying the appropriate algorithm design techniques,
based on the structure of the problem. These two components interact: the
more comfortable one is with the full array of possible design techniques,
the more one starts to recognize the clean formulations that lie within messy

xiv

Preface

Preface

problems out in the world. At their most effective, then, algorithmic ideas do
not just provide solutions to _well-posed problems; they form the language that
lets you cleanly express the underlying questions.
The goal of our book is to convey this approach to algorithms, as a design
process that begins with problems arising across the full range of computing
applications, builds on an understanding of algorithm design techniques, and
results in the development of efficient solutions to these problems. We seek
to explore the role of algorithmic ideas in computer science generally, and
relate these ideas to the range of precisely formulated problems for which we
can design and analyze algorithms. In other words, what are the underlying
issues that motivate these problems, and how did we choose these particular
ways of formulating them? How did we recognize which design principles were
appropriate in different situations?
In keeping with this, our goal is to offer advice on how to identify clean
algorithmic problem formulations in complex issues from different areas of
computing and, from this, how to design efficient algorithms for the resulting
problems. Sophisticated algorithms are often best understood by reconstructing the sequence of ideas--including false starts and dead ends--that led from
simpler initial approaches to the eventual solution. The result is a style of exposition that does not take the most direct route from problem statement to
algorithm, but we feel it better reflects the way that we and our colleagues
genuinely think about these questions.

Overview
The book is intended for students who have completed a programmingbased two-semester introductory computer science sequence (the standard
"CS1/CS2" courses) in which they have written programs that implement
basic algorithms, manipulate discrete structures such as trees and graphs, and
apply basic data structures such as arrays, lists, queues, and stacks. Since
the interface between CS1/CS2 and a first algorithms course is not entirely
standard, we begin the book with self-contained coverage of topics that at
some institutions a_re familiar to students from CS1/CS2, but which at other
institutions are included in the syllabi of the first algorithms course. This
material can thus be treated either as a review or as new material; by including
it, we hope the book can be used in a broader array of courses, and with more
flexibility in the prerequisite knowiedge that is assumed.
In keeping with the approach outlined above, we develop the basic algorithm design techniques by drawing on problems from across many areas of
computer science and related fields. To mention a few representative examples
here, we include fairly detailed discussions of applications from systems and
networks (caching, switching, interdomain routing on the Internet), artificial

intelligence (planning, game playing, Hopfield networks), computer vision


(image segmentation), data mining (change-point detection, clustering), operations research (airline scheduling), and computational biology (sequence
alignment, RNA secondary structure).
The notion of computational intractability, and NP-completeness in particular, plays a large role in the book. This is consistent with how we think
about the overall process of algorithm design. Some of the time, an interesting problem arising in an application area will be amenable to an efficient
solution, and some of the time it will be provably NP-complete; in order to
fully address a new algorithmic problem, one should be able to explore both
of these ol)tions with equal familiarity. Since so many natural problems in
computer science are NP-complete, the development of methods to deal with
intractable problems has become a crucial issue in the study of algorithms,
and our book heavily reflects this theme. The discovery that a problem is NPcomplete should not be taken as the end of the story, but as an invitation to
begin looking for approximation algorithms, heuristic local search techniques,
or tractable special cases. We include extensive coverage of each of these three
approaches.

Problems and Solved Exercises


An important feature of the book is the collection of problems. Across all
chapters, the book includes over 200 problems, almost a!l of them developed
and class-tested in homework or exams as part of our teaching of the course
at Cornell. We view the problems as a crucial component of the book, and
they are structured in keeping with our overall approach to the material. Most
of them consist of extended verbal descriptions of a problem arising in an
application area in computer science or elsewhere out in the world, and part of
the problem is to practice what we discuss in the text: setting up the necessary
notation and formalization, designing an algorithm, and then analyzing it and
proving it correct. (We view a complete answer to one of these problems as
consisting of all these components: a fl~y explained algorithm, an analysis of
the nmning time, and a proof of correctness.) The ideas for these problems
come in large part from discussions we have had over the years with people
working in different areas, and in some cases they serve the dual purpose of
recording an interesting (though manageable) application of algorithms that
we havent seen written down anywhere else.
To help with the process of working on these problems, we include in
each chapter a section entitled "Solved Exercises," where we take one or more
problems and describe how to go about formulating a solution. The discussion
devoted to each solved exercise is therefore significantly longer than what
would be needed simply to write a complete, correct solution (in other words,

XV

xvi

Preface

significantly longer than what it would take to receive full credit if these were
being assigned as homework problems). Rather, as with the rest of the text,
the discussions in these sections should be viewed as trying to give a sense
of the larger process by which one might think about problems of this type,
culminating in the speci.fication of a precise solution.
It is worth mentioning two points concerning the use of these problems
as homework in a course. First, the problems are sequenced roughly in order
of increasing difficulty, but this is only an approximate guide and we advise
against placing too much weight on it: since the bulk of the problems were
designed as homework for our undergraduate class, large subsets of the
problems in each chapter are really closely comparable in terms of difficulty.
Second, aside from the lowest-numbered ones, the problems are designed to
involve some investment of time, both to relate the problem description to the
algorithmic techniques in the chapter, and then to actually design the necessary
algorithm. In our undergraduate class, we have tended to assign roughly three
of these problems per week.

Pedagogical Features and Supplements


In addition to the Problems and solved exercises, the book has a number of
further pedagogical features, as well as additional supplements to facilitate its
use for teaching.
As noted earlier, a large number of the sections in the book axe devoted
to the formulation of an algorithmic problem--including its background and
underlying motivation--and the design and analysis of an algorithm for this
problem. To reflect this style, these sections are consistently structured around
a sequence of subsections: "The Problem," where the problem is described
and a precise formulation is worked out; "Designing the Algorithm," where
the appropriate design technique is employed to develop an algorithm; and
"Analyzing the Algorithm," which proves properties of the algorithm and
analyzes its efficiency. These subsections are highlighted in the text with an
icon depicting a feather. In cases where extensions to the problem or further
analysis of the algorithm is pursued, there are additional subsections devoted
to these issues. The goal of this structure is to offer a relatively uniform style
of presentation that moves from the initial discussion of a problem arising in a
computing application through to the detailed analysis of a method to solve it.
A number of supplements are available in support of the book itself. An
instructors manual works through al! the problems, providing fi~ solutions to
each. A set of lecture slides, developed by Kevin Wayne of Princeton University,
is also available; these slides follow the order of the books sections and can
thus be used as the foundation for lectures in a course based on the book. These
files are available at wunv.aw.com. For instructions on obtaining a professor

Preface
login and password, search the site for either "Kleinberg or "Tardos" or
contact your local Addison-Wesley representative.
Finally, we would appreciate receiving feedback on the book. In particular,
as in any book of this length, there are undoubtedly errors that have remained
in the final version. Comments and reports of errors can be sent to us by e-mail,
at the address algbook@cs.cornell.edu; please include the word "feedback"
in the subject line of the message.

Chapter-by-Chapter Synopsis
Chapter I starts by introducing some representative algorithmic problems. We
begin immediately with the Stable Matching Problem, since we feel it sets
up the basic issues in algorithm design more concretely and more elegantly
than any abstract discussion could: stable matching is motivated by a natural
though complex real-world issue, from which one can abstract an interesting
problem statement and a surprisingly effective algorithm to solve this problem.
The remainder of Chapter 1 discusses a list of five "representative problems"
that foreshadow topics from the remainder of the course. These five problems
are interrelated in the sense that they are all variations and/or special cases
of the Independent Set Problem; but one is solvable bya greedy algorithm,
one by dynamic programming, one by network flow, one (the Independent
Set Problem itself) is NP-complete, and one is PSPACE-complete. The fact that
closely related problems can vary greatly in complexity is an important theme
of the book, and these five problems serve as milestones that reappear as the
book progresses.
Chapters 2 and 3 cover the interface to the CS1/CS2 course sequence
mentioned earlier. Chapter 2 introduces the key mathematical definitions and
notations used for analyzing algorithms, as wel! as the motivating principles
behind them. It begins with an informal overview of what it means for a problem to be computationally tractable, together with the concept of polynomial
time as a formal notion of efficiency. It then discusses growth rates of functions and asymptotic analysis more formally, and offers a guide to commordy
occurring functions in algorithm analysis, together with standard applications
in which they arise. Chapter 3 covers the basic definitions and algorithmic
primitives needed for working with graphs, which are central to so many of
the problems in the book. A number of basic graph algorithms are often implemented by students late in the CS1/CS2 course sequence, but it is valuable
to present the material here in a broader algorithm design context. In particular, we discuss basic graph definitions, graph traversal techniques such
as breadth-first search and depth-first search, and directed graph concepts
including strong connectivity and topological ordering.

xvii

Preface

Chapters 2 and 3 also present many of the basic data structures that will
be used for implementing algorithms throughout the book; more advanced
data structures are presented in subsequent chapters. Our approach to data
structures is to introduce them as they are needed for the implementation of
the algorithms being developed in the book. Thus, although many of the data
structures covered herewill be familiar to students from the CS1/CS2 sequence,
our focus is on these data structures in the broader context of algorithm design
and analysis.

Chapters 4 through 7 cover four major algorithm design techniques: greedy


algorithms, divide and conquer, dynamic programming, and network flow.
With greedy algorithms, the challenge is to recognize when they work and
when they dont; our coverage of this topic is centered around a way of classifying the kinds of arguments used to prove greedy algorithms correct. This
chapter concludes with some of the main applications of greedy algorithms,
for shortest paths, undirected and directed spanning trees, clustering, and
compression. For divide and conquer, we begin with a discussion of strategies
for solving recurrence relations as bounds on running times; we then show.
how familiarity with these recurrences can guide thedesign of algorithms that
improve over straightforward approaches to a number of basic problems, including the comparison of rankings, the computation of c!osest pairs of points
in the plane, and the Fast Fourier Transform. Next we develop dynamic programming by starting with the recursive intuition behind it, and subsequently
building up more and more expressive recurrence formulations through applications in which they naturally arise. This chapter concludes with extended
discussions of the dynamic programming approach to two fundamental problems: sequence alignment, with applications in computational biology; and
shortest paths in graphs, with connections to Internet routing protocols. Finally, we cover algorithms for network flow problems, devoting much of our
focus in this chapter to discussing a large array of different flow applications.
To the extent that network flow is covered in algorithms courses, students are
often left without an appreciation for the wide range of problems to which it
can be applied; we try to do iustice to its versatility by presenting applications
to load balancing, scheduling, image segmentation, and a number of other
problems.
Chapters 8 and 9 cover computational intractability. We devote most of
our attention to NP-completeness, organizing the basic NP-complete problems
thematically to help students recognize candidates for reductions when they
encounter new problems. We build up to some fairly complex proofs of NPcompleteness, with guidance on how one goes about constructing a difficult
~reduction. We also consider types of computational hardness beyond NPcompleteness, particularly through the topic of PSPACE-completeness. We

Preface

find this is a valuable way to emphasize that intractability doesnt end at


NP-completeness, and PSPACE-completeness also forms the underpinning for
some central notions from artificial intelligence--planning and game playing-that would otherwise not find a place in the algorithmic landscape we are
surveying.
Chapters 10 through 12 cover three maior techniques for dealing with computationally intractable problems: identification of structured special cases,
approximation algorithms, and local search heuristics. Our chapter on tractable
special cases emphasizes that instances of NP-complete problems arising in
practice may not be nearly as hard as worst-case instances, because they often
contain some structure that can be exploited in the design of an efficient algorithm. We illustrate how NP-complete problems are often efficiently solvable
when restricted to tree-structured inputs, and we conclude with an extended
discussion of tree decompositions of graphs. While this topic is more suitable for a graduate course than for an undergraduate one, it is a technique
with considerable practical utility for which it is hard to find an existing
accessible reference for students. Our chapter on approximation algorithms
discusses both the process of designing effective algorithms and the task of
understanding the optimal solution well enough to obtain good bounds on it.
As design techniques for approximation algorithms, we focus on greedy algorithms, linear programming, anda third method we refer to as "pricing: which
incorporates ideas from each of the first two. Finally, we discuss local search
heuristics, including the Metropolis algorithm and simulated annealing. This
topic is often missing from undergraduate algorithms courses, because very
little is known in the way of provable guarantees for these algorithms; however, given their widespread use in practice, we feel it is valuable for students
to know something about them, and we also include some cases in which
guarantees can be proved.
Chapter 13 covers the use of randomization in the design of algorithms.
This is a topic on which several nice graduate-level books have been written.
Our goal here is to provide a more compact introduction to some of the
ways in which students can apply randomized techniques using the kind of
background in probability one typically gains from an undergraduate discrete
math course.
Use of the Book
The book is primarily designed for use in a first undergraduate course on
algorithms, but it can also be used as the basis for an introductory graduate
course.

When we use the book at the undergraduate level, we spend roughly


one lecture per numbered section; in cases where there is more than one

Preface

Preface

lectures worth of material in a section (for example, when a section provides


further applications as additional examples), we treat this extra material as a
supplement that students carl read about outside of lecture. We skip the starred
sections; while these sections contain important topics, they are less central
to the development of the subject, and in some cases they are harder as well.
We also tend to skip one or two other sections per chapter in the first half of
the book (for example, we tend to skip Sections 4.3, 4.7-4.8, 5.5-5.6, 6.5, 7.6,
and 7.!1). We cover roughly half of each of Chapters 11-13.
This last point is worth emphasizing: rather than viewing the later chapters
as "advanced," and hence off-limits to undergraduate algorithms courses, we
have designed them with the goal that the first few sections of each should
be accessible to an undergraduate audience. Our own undergraduate course
involves material from all these chapters, as we feel that all of these topics
have an important place at the undergraduate level.
Finally, we treat Chapters 2 and 3 primarily as a review of material from
earlier courses; but, as discussed above, the use of these two chapters depends
heavily on the relationship of each specific course to its prerequisites.
The resulting syllabus looks roughly as follows: Chapter 1; Chapters 4-8
(excluding 4.3, 4.7-4.9, 5.5-5.6, 6.5, 6.10, 7.4, 7.6, 7.11, and 7.13); Chapter 9
(briefly); Chapter 10, Sections.10.! and 10.2; Chapter 11, Sections 11.1, 11.2,
11.6, and 11.8; Chapter 12, Sections 12.1-12.3; and Chapter 13, Sections 13.1-

13.5.
The book also naturally supports an introductory graduate course on
algorithms. Our view of such a course is that it should introduce students
destined for research in all different areas to the important current themes in
algorithm design. Here we find the emphasis on formulating problems to be
useful as well, since students will soon be trying to define their own research
problems in many different subfields. For this type of course, we cover the
later topics in Chapters 4 and 6 (Sections 4.5-4.9 and 6.5-6.10), cover all of
Chapter 7 (moving more rapidly through the early sections), quickly cover NPcompleteness in Chapter 8 (since many beginning graduate students will have
seen this topic as undergraduates), and then spend the remainder of the time
on Chapters !0-13. Although our focus in an introductory graduate course is
on the more advanced sections, we find it usefifl for the students to have the
full book to consult for reviewing or filling in background knowledge, given
the range of different undergraduate backgrounds among the students in such
a course.

Finally, the book can be used to support self-study by graduate students,


researchers, or computer professionals who want to get a sense for how they

might be able to use particular algorithm design techniques in the context of


their own work. A number of graduate students and colleagues have used
portions of the book in this way.

Acknowledgments
This book grew out of the sequence of algorithms co~ses that we have taught
at Cornell. These courses have grown, as the field has grown, over a number of
years, and they reflect the influence of the Comell faculty who helped to shape
them during this time, including Juris Hartmanis, Monika Henzinger, John
Hopcroft, Dexter Kozen, Ronitt Rubinfeld, and Sam Toueg. More generally, we
would like to thank al! our colleagues at Corne!l for countless discussions both
on the material here and on broader issues about the nature of the field.
The course staffs weve had in teaching the subject have been tremendously helpful in the formulation of this material. We thank our undergraduate and graduate teaching assistants, Siddharth Alexander, Rie Ando, Elliot
Anshelevich, Lars Backstrom, Steve Baker, Ralph Benzinger, John Bicket,
Doug Burdick, Mike Connor, Vladimir Dizhoor, Shaddin Doghmi, Alexander Druyan, Bowei Du, Sasha Evfimievski, Ariful Gan~.,_ Vadim Grinshpun,
Ara Hayrapetyan, Chris Jeuell, Igor Kats, Omar Khan Mikhail Kobyakov,
Alexei Kopylov, Brian Kulis, Amit Kumar, Yeongwee Lee, Henry Lin, Ashwin Machanavajjhala, Ayan Mandal, Bill McCloskey, Leonid Meyerguz, Evan
Moran, Niranjan Nagarajan, Tina Nolte, Travis Ortogero, Martin P~il, Jon
Peress, Matt Piotrowski, Joe Polastre, Mike Priscott, Xin Qi, Venu Ramasubramanian, Aditya Rao, David Richardson, Brian Sabino, Rachit Siamwalla, Sebastian Sllgardo, Alex Slivkins, Chaitanya Swamy, Perry Tam, Nadya Travinin,
Sergei Vassilvitskii, Matthew Wachs, Tom Wexler, Shan-Leung Maverick Woo,
Justin Yang, and Misha Zatsman. Many of them have provided valuable insights, suggestions, and comments on the text. We also thank all the students
in these classes who have provided comments and feedback on early drafts of
the book over the years.
For the past several years, the development of the book has benefited
greatly from the feedback and advice of colleagues who have used prepublication drafts for teaching. Anna Karlin fearlessly adopted a draft as her course
textbook at the University of Washington when it was st~ in an early stage of
development; she was followed by a number of people who have used it either
as a course textbook or as a resource for teaching: Paul Beame, Allan Borodin,
Devdatt Dubhashi, David Kempe, Gene Kleinberg, Dexter Kozen, Amit Kumar,
Mike Molloy, Yuval Rabani, Tim Roughgarden, Alexa Sharp, Shanghua Teng,
Aravind Srinivasan, Dieter van Melkebeek, Kevin Wayne, Tom Wexler, and

xxi

xxii

Preface

Sue Whitesides. We deeply appreciate their input and advice, which has informed many of our revisions to the content. We would like to additionally
thank Kevin Wayne for producing supplementary material associated with the
book, which promises to greatly extend its utility to future instructors.
In a number of other cases, our approach to particular topics in the book
reflects the infuence of specific colleagues. Many of these contributions have
undoubtedly escaped our notice, but we especially thank Yufi Boykov, Ron
Elber, Dan Huttenlocher, Bobby Kleinberg, Evie Kleinberg, Lillian Lee, David
McAllester, Mark Newman, Prabhakar Raghavan, Bart Selman, David Shmoys,
St~ve Strogatz, Olga Veksler, Duncan Watts, and Ramin Zabih.
It has been a pleasure working with Addison Wesley over the past year.
First and foremost, we thank Matt Goldstein for all his advice and guidance in
this process, and for helping us to synthesize a vast amount of review material
into a concrete plan that improved the book. Our early conversations about
the book with Susan Hartman were extremely valuable as well. We thank Matt
and Susan, together with Michelle Brown, Marilyn Lloyd, Patty Mahtani, and.
Maite Suarez-Rivas at Addison Wesley, and Paul Anagnostopoulos and Jacqui
Scarlott at Windfall Software, for all their work on the editing, production, and
management of the proiect. We fln-ther thank Paul and Jacqui for their expert
composition of the book. We thank Joyce Wells for the cover design, Nancy
Murphy of Dartmouth Publishing for her work on the figures, Ted Laux for
the indexing, and Carol Leyba and Jennifer McClain for the copyedifing and
proofreading.
We thank Anselm Blumer (Tufts University), Richard Chang (University of
Maryland, Baltimore County), Kevin Compton (University of Michigan), Diane
Cook (University of Texas, Arlington), Sariel Har-Peled (University of Illinois,
Urbana-Champaign), Sanjeev Khanna (University of Pennsylvania), Philip
Klein (Brown University), David Matthias (Ohio State University), Adam Meyerson (UCLA), Michael Mitzenmacher (Harvard University), Stephan Olariu
(Old Dominion University), Mohan Paturi (UC San Diego), Edgar Ramos (University of Illinois, Urbana-Champaign), Sanjay Ranka (University of Florida,
Gainesville), Leon Reznik (Rochester Institute of Technology), Subhash Suri
(UC Santa Barbara), Dieter van Melkebeek (University of Wisconsin, Madison), and Bulent Yener (Rensselaer Polytechnic Institute) who generously
contributed their time to provide detailed and thoughtful reviews of the manuscript; their comments led to numerous improvements, both large and small,
in the final version of the text.

Finally, we thank our families--Lillian and Alice, and David, Rebecca, and
Amy. We appreciate their support, patience, and many other contributions
more than we can express in any acknowledgments here.

Preface

This book was begun amid the irrational exuberance of the late nineties,
when the arc of computing technology seemed, to many of us, briefly to pass
through a place traditionally occupied by celebrities and other inhabitants of
the pop-cultural firmament. (It was probably iust in our imaginations.) Now,
several years after the hype and stock prices have come back to earth, one can
appreciate that in some ways computer science was forever changed by this
period, and in other ways it has remained the same: the driving excitement
that has characterized the field since its early days is as strong and enticing as
ever, the publics fascination with information technology is still vibrant, and
the reach of computing continues to extend into new disciplines. And so to
all students of the subject, drawn to it for so many different reasons, we hope
you find this book an enjoyable and useful guide wherever your computational
pursuits may take you.
Jon Kleinberg
gva Tardos
Ithaca, 2005

xxiii

1.1 A First Problem: Stable Matching


As an opening topic, we look at an algorithmic problem that nicely illustrates
many of the themes we will be emphasizing. It is motivated by some very
natural and practical concerns, and from these we formulate a clean and
simple statement of a problem. The algorithm to solve the problem is very
clean as well, and most of our work will be spent in proving that it is correct
and giving an acceptable bound on the amount of time it takes to terminate
with an answer. The problem itself--the Stable Matching Problem--has several
origins.

~ The Problem
The Stable Matching Problem originated, in part, in 1962, when David Gale
and Lloyd Shapley, two mathematical economists, asked the question: Could
one design a college admissions process, or a job recruiting process, that was
self-enforcing? What did they mean by this?
To set up the question, lets first think informally about the kind of situation
that might arise as a group of friends, all iurdors in college majoring in
computer science, begin applying to companies for summer internships. The
crux of the application process is the interplay between two different types
of parties: companies (the employers) and students (the applicants). Each
applicant has a preference ordering on companies, and each company--once
the applications Come in--forms a preference ordering on its applicants. Based
on these preferences, companies extend offers to some of their applicants,
applicants choose which of their offers to accept, and people begin heading
off to their summer internships.

Chapter 1 Introduction: Some Representative Problems

Gale and Shapley considered the sorts of things that could start going
wrong with this process, in the absence of any mechanism to enforce the status
quo. Suppose, for example, that your friend Raj has iust accepted a summer job
at the large telecommunications company CluNet. A few days later, the small
start-up company WebExodus, which had been dragging its feet on making a
few final decisions, calls up Rai and offers him a summer iob as well. Now, Rai
actually prefers WebExodus to CluNet--won over perhaps by the laid-back,
anything-can-happen atmosphere--and so this new development may well
cause him to retract his acceptance of the CluNet offer and go to WebExodus
instead. Suddenly down one summer intern, CluNet offers a job to one of its
wait-listed applicants, who promptly retracts his previous acceptance of an
offer from the software giant Babelsoft, and the situation begins to spiral out
of control.
Things look just as bad, if not worse, from the other direction. Suppose
that Rajs friend Chelsea, destined to go to Babelsoft but having just heard Rajs
story, calls up the people at WebExodus and says, "You know, Id really rather
spend the summer with you guys than at Babelsoft." They find this very easy
to believe; and furthermore, on looking at Chelseas application, they realize
that they would have rather hired her than some other student who actually
is scheduled to spend the summer at WebExodus. In this case, if WebExodus
were a slightly less scrupulous company, it might well find some way to retract
its offer to this other student and hire Chelsea instead.
Situations like this can rapidly generate a lot of chaos, and many people-both applicants and employers--can end up unhappy with the process as well
as the outcome. What has gone wrong? One basic problem is that the process
is not self-enforcing--if people are allowed to act in their self-interest, then it
risks breaking down.
We might well prefer the following, more stable situation, in which selfinterest itself prevents offers from being retracted and redirected. Consider
another student, who has arranged to spend the summer at CluNet but calls
up WebExodus and reveals that he, too, would rather work for them. But in
this case, based on the offers already accepted, they are able to reply, "No, it
turns out that we prefer each of the students weve accepted to you, so were
afraid theres nothing we can do." Or consider an employer, earnestly following
up with its top applicants who went elsewhere, being told by each of them,
"No, Im happy where I am." In such a case, all the outcomes are stable--there
are no further outside deals that can be made.
So this is the question Gale and Shapley asked: Given a set of preferences
among employers and applicants, can we assign applicants to employers so
that for every employer E, and every applicant A who is not scheduled to work
for E, at least one of the following two things is the case?

1. ! A First Problem: Stable,Matching

(i) E prefers every one of its accepted applicants to A; or


(ii) A prefers her current situation over working for employer E.
If this holds, the outcome is stable: individual self-interest will prevent any
applicant/employer deal from being made behind the scenes.
Gale and Shapley proceeded to develop a striking algorithmic solution to
this problem, which we will discuss presently. Before doing this, lets note that
this is not the only origin of the Stable Matching Problem. It turns out that for
a decade before the work of Gale and Shapley, unbeknownst to them, the
National Resident Matching Program had been using a very similar procedure,
with the same underlying motivation, to match residents to hospitals. Indeed,
this system, with relatively little change, is still in use today.
This is one testament to the problems fundamental appeal. And from the
point of view of this book, it provides us with a nice first domain in which
to reason about some basic combinatorial definitions and the algorithms that
build on them.

Formulating the Problem To get at the essence of this concept, it helps to


make the problem as clean as possible. The world of companies and applicants
contains some distracting asymmetries. Each applicant is looking for a single
company, but each company is looking for many applicants; moreover, there
may be more (or, as is sometimes the case, fewer) applicants than there are
available slots for summer iobs. Finally, each applicant does not typica!ly apply
to every company.
It is useful, at least initially, to eliminate these complications and arrive at a
more "bare-bones" version of the problem: each of n applicants applies to each
of n companies, and each company wants to accept a single applicant. We will
see that doing this preserves the fundamental issues inherent in the problem;
in particular, our solution to this simplified version will extend directly to the
more general case as well.
Following Gale and Shapley, we observe that this special case can be
viewed as the problem of devising a system by which each of n men and
n women can end up getting married: our problem naturally has the analogue
of two "genders"--the applicants and the companies--and in the case we are
considering, everyone is seeking to be paired with exactly one individual of
the opposite gender.1

1 Gale and Shapley considered the same-sex Stable Matching Problem as well, where there is only a
single gender. This is motivated by related applications, but it turns out to be fairly different at a
technical level. Given the applicant-employer application were considering here, well be focusing
on the version with two genders.

!.1 A First Problem: Stable Matching

Chapter 1 Introduction: Some Representative Problems

I~

n instability: m and w~
each prefer the other to
eir current partners.

Figure 1.1 Perfect matching

S with instability (m, w).

So consider a set M = {m1 ..... ran} of n men, and a set W = {iv1 ..... Ivn}
of n women. Let M x W denote the set of all possible ordered pairs of the form
(m, Iv), where m ~ M and Iv ~ W. A matching S is a set of ordered pairs, each
from M x W, with the property that each member of M and each member of
W appears in at most one pair in S. A perfect matching S is a matching with
the property that each member of M and each member of W appears in exactly
one pair in S.
Matchings and perfect matchings are objects that will recur frequently
throughout the book; they arise naturally in modeling a wide range of algorithmic problems. In the present situation, a perfect matching corresponds
simply to a way of pairing off the men with the women, in such a way that
everyone ends up married to somebody, and nobody is married to more than
one person--there is neither singlehood nor polygamy.
Now we can add the notion of preferences to this setting. Each man m ~ M
ranks all the women; we will say that m prefers Iv to Iv if m ranks Iv higher
than Iv. We will refer to the ordered ranking of m as his preference list. We wil!
not allow ties in the ranking. Each woman, analogously, ranks all the men.
Given a perfect matching S, what can go wrong? Guided by our initial
motivation in terms of employers and applicants, we should be worried about
the following situation: There are two pairs (m, Iv) and (m, to) in S (as
depicted in Figure 1.1) with the property that m prefers w to Iv, and Iv prefers
m to m. In this case, theres nothing to stop m and Iv from abandoning their
current partners and heading off together; the set of marriages is not selfenforcing. Well say that such a pair (m, Iv) is an instability with respect to S:
(m, Iv) does not belong to S, but each of m and Iv prefers the other to their
partner in S.
Our goal, then, is a set of marriages with no instabilities. Well say that
a matching S is stable if (i) it is perfect, and (ii) there is no instability with
respect to S. Two questions spring immediately to mind:
Does there exist a stable matching for every set of preference lists?
Given a set of preference lists, can we efficiently construct a stable
matching if there is one?
Some Examples To illustrate these definitions, consider the following two
very simple instances of the Stable Matching Problem.
First, suppose we have a set of two men, fro, m}, and a set of two women,
{iv, Iv}. The preference lists are as follows:
prefers Iv to Iv.
prefers Iv to IV.

Iv prefers m to m.
Iv prefers m to m.
If we think about this set of preference lists intuitively, it represents complete
agreement: the men agree on the order of the women, and the women agree
on the order of the men. There is a unique stable matching here, consisting
of the pairs (m, Iv) and (m, Iv). The other perfect matching, consisting of the
pairs (m, Iv) and (m, Iv), would not be a stable matching, because the pair
(m, Iv) would form an instability with respect to this matching. (Both m and
Iv would want to leave their respective partners and pair up.)
Next, heres an example where things are a bit more intricate. Suppose
the preferences are

m prefers Iv to Iv.
m prefers Iv to Iv.
Iv prefers m to m.
Iv prefers m to m.
Whats going on in this case? The two mens preferences mesh perfectly with
each other (they rank different women first), and the two womens preferences
likewise mesh perfectly with each other. But the mens preferences clash
completely with the womens preferences.
In this second example, there are two different stable matchings. The
matching consisting of the pairs (m, w) and (m, w) is stable, because both
men are as happy as possible, so neither would leave their matched partner.
But the matching consisting of the pairs (m, w) and (m, w) is also stable, for
the complementary reason that both women are as happy as possible. This is
an important point to remember as we go forward--its possible for an instance
to have more than one stable matching.
~:~ Designing the Algorithm
we now show that there exists a stable matching for every set of preference
lists among the men and women. Moreover, our means of showing this will
also answer the second question that we asked above: we will give an efficient
algorithm that takes the preference lists and constructs a stable matching.
Let us consider some of the basic ideas that.motivate the algorithm.
Initially, everyone is unmarried. Suppose an unmarried man m chooses
the woman Iv who ranks highest on his preference list and proposes to
her. Can we declare immediately that (m, Iv) wii1 be one of the pairs in our
final stable matching? Not necessarily: at some point in the future, a man
m whom Iv prefers may propose to her. On the other hand, it would be

Chapter 1 Introduction: Some Representative Problems

I~

oman w will become~


ngaged to m if she |
refers him to rat J

dangerous for w to reject m right away; she may never receive a proposal
from someone she ranks as highly as m. So a natural idea would be to
have the pair (m, w) enter an intermediate state--engagement.
Suppose we are now at a state in which some men and women are/Tee-not engaged--and some are engaged. The next step could look like this.
An arbitrary flee man m chooses the highest-ranked woman w to whom
he has not yet proposed, and he proposes to her. If w is also free, then m
and w become engaged. Otherwise, w is already engaged to some other
man m. In this case, she determines which of m or m ranks higher
on her preference list; this man becomes engaged to w and the other
becomes flee,
Finally, the algorithm wil! terminate when no one is free; at this moment,
all engagements are declared final, and the resulting perfect matchdng is
returned.

Figure 1.2 An intermediate


state of the G-S algorithm
when a free man ra is proposing to a woman w.

Here is a concrete description of the Gale-Shapley algorithm, with Figure 1.2 depicting a state of the algorithm.
Initially all m E M and w E W are free
While there is a man m who is free and hasnt proposed to
every woman
Choose such a man m
Let w be the highest-ranked woman in ms preference list
to whom m has not yet proposed
If ~ is free then
(m, ~) become engaged
Else ~ is currently engaged to m
If ~ prefers m to m then
m remains free
Else w prefers m to m
(m,~) become engaged
nlI becomes free
Endif
Endif
Endwhile
Return the set S of engaged pairs

An intriguing thing is that, although the G-S algorithm is quite simple


to state, it is not immediately obvious that it returns a stable matching, or
even a perfect matching. We proceed to prove this now, through a sequence
of intermediate facts.

1.1 A First Problem: Stable Matching

~ Analyzing the Algorithm


First consider the view of a woman w during the execution of the algorithm.
For a while, no one has proposed to her, and she is free. Then a man m may
propose to her, and she becomes engaged. As time goes on, she may receive
additional proposals, accepting those that increase the rank of her partner. So
we discover the following.
(1.1) w remains engaged /Tom the point at which she receives her first
proposal; and the sequence of partners to which she is engaged gets better and
better (in terms of her preference list).
The view of a man m during the execution of the algorithm is rather
different. He is free until he proposes to the highest-ranked woman on his
list; at this point he may or may not become engaged. As time goes on, he
may alternate between being free and being engaged; however, the following
property does hold.

(1.2) The sequence of women to whom m proposes gets worse and worse (in
terms of his preference list).
Now we show that the algorithm terminates, and give a bound on the
maximum number of iterations needed for termination.
(1,3) The G-S algorithm terminates after at most n2 iterations of the While
loop.
Proof. A useful strategy for upper-bounding the running time of an algorithm,
as we are trying to do here, is to find a measure of progress. Namely, we seek
some precise way of saying that each step taken by the algorithm brings it
closer to termination.
In the case of the present algorithm, each iteration consists of some man
proposing (for the only time) to a woman he has never proposed to before. So
if we let ~P(t) denote the set of pairs (m, w) such that m has proposed to w by
the end of iteration t, we see that for all t, the size of ~P(t + 1) is strictly greater
than the size of ~P(t). But there are only n2 possible pairs of men and women
in total, so the value of ~P(.) can increase at most n2 times over the course of
the algorithm. It follows that there can be at most n2 iterations. []
Two points are worth noting about the previous fact and its proof. First,
there are executions of the algorithm (with certain preference lists) that can
involve close to n2 iterations, so this analysis is not far from the best possible.
Second, there are many quantities that would not have worked well as a
progress measure for the algorithm, since they need not strictly increase in each

1.1 A First Problem: Stable Matching

Chapter 1 Introduction: Some Representative Problems

iteration. For example, the number of free individuals could remain constant
from one iteration to the next, as could the number of engaged pairs. Thus,
these quantities could not be used directly in giving an upper bound on the
maximum possible number of.iterations, in the style of the previous paragraph.
Let us now establish that the set S returned at the termination of the
algorithm is in fact a perfect matching. Why is this not immediately obvious?
Essentially, we have to show that no man can "fall off" the end of his preference
list; the only way for the ~h].e loop to exit is for there to be no flee man. In
this case, the set of engaged couples would indeed be a perfect matching.
So the main thing we need to show is the following.
(1.4) If m is free at some point in the execution of the algorithm, then there
is a woman to whom he has not yet proposed.
Proof. Suppose there comes a point when m is flee but has already proposed
to every woman. Then by (1.1), each of the n women is engaged at this point
in time. Since the set of engaged pairs forms a matching, there must also be
n engaged men at this point in time. But there are only n men total, and m is
not engaged, so this is a contradiction. ,,
(1..~)

The set S returned at termination is a peryect matching.

Proof. The set of engaged pairs always forms a matching. Let us suppose that
the algorithm terminates with a flee man m. At termination, it must be the
case that m had already proposed to every woman, for otherwise the ~qhile
loop would not have exited. But this contradicts (1.4), which says that there
cannot be a flee man who has proposed to every woman. ,,
Finally, we prove the main property of the algorithm--namely, that it
results in a stable matching.
(1.6) Consider an executionof the G-S algorithm that returns a set of pairs
S. The set S is a stable matching.
Proof. We have already seen, in (1.5), that S is a perfect matching. Thus, to
prove S is a stable matching, we will assume that there is an instability with
respect to S and obtain a contradiction. As defined earlier, such an instability
would involve two pairs, (m, w) and (m, w), in S with the properties that
o m prefers w to w, and
o w prefers m to mL
In the execution of the algorithm that produced S, ms last proposal was, by
definition, to w. Now we ask: Did m propose to w at some earlier point in

this execution? If he didnt, then w must occur higher on ms preference.list


than w, contxadicting our assumption that m prefers w to w. If he did, then
he was rejected by w in favor of some other man m", whom w prefers to m.
m is the final partner of w, so either m" = m or, by (1.!), w prefers her final
partner m~ to m"; either way this contradicts our assumption that w prefers
m to mI.

It follows that S is a stable matching. []

Extensions
We began by defining the notion of a stable matching; we have just proven
that the G-S algorithm actually constructs one. We now consider some further
questions about the behavior of the G-S algorithm and its relation to the
properties of different stable matchings.
To begin wit_h, recall that we saw an example earlier in which there could
be multiple stable matchings. To recap, the preference lists in this example
were as follows:
prefers w to w.
~ prefers w to w.
prefers m~ to m.
prefers m to m.

Now, in any execution of the Gale-Shapley algorithm, m will become engaged


to w, m will become engaged to w (perhaps in the other order), and things
will stop there. Thus, the other stable matching, consisting of the pairs (m, w)
and (m, w), is not attainable from an execution of the G-S algorithm in which
the men propose. On the other hand, it would be reached if we ran a version of
the algorithm in which the women propose. And in larger examples, with more
than two people on each side, we can have an even larger collection of possible
stable matchings, many of them not achievable by any natural algorithm.
This example shows a certain "unfairness" in the G-S algorithm, favoring
men. If the mens preferences mesh perfectly (they all list different women as
their first choice), then in all runs of the G-S algorithm all men end up matched
with their first choice, independent of the preferences of the women. If the
womens preferences clash completely with the mens preferences (as was the
case in this example), then the resulting stable matching is as bad as possible
for the women. So this simple set of preference lists compactly summarizes a
world in which someone is destined to end up unhappy: women are unhappy
if men propose, and men are unhappy if women propose.
Lets now analyze the G-S algorithm in more detail and try to understand
how general this "unfairness" phenomenon is.

10

Chapter 1 Introduction: Some Representative Problems

To begin With, our example reinforces the point that the G-S algorithm
is actually underspecified: as long as there is a free man, we are allowed to
choose any flee man to make the next proposal. Different choices specify
different executions of the algprithm; this is why, to be careful, we stated (1.6)
as "Consider an execution of the G-S algorithm that returns a set of pairs S,"
instead of "Consider the set S returned by the G-S algorithm."
Thus, we encounter another very natural question: Do all executions of
the G-S algorithm yield the same matching? This is a genre of question that
arises in many settings in computer science: we have an algorithm that runs
asynchronously, with different independent components performing actions
that can be interleaved in complex ways, and we want to know how much
variability this asynchrony causes in the final outcome. To consider a very
different kind of example, the independent components may not be men and
women but electronic components activating parts of an airplane wing; the
effect of asynchrony in their behavior can be a big deal.
In the present context, we will see that the answer to our question is
surprisingly clean: all executions of the G-S algorithm yield the same matching.
We proceed to prove this now.
All Executions Yield the Same Matching There are a number of possible
ways to prove a statement such as this, many of which would result in quite
complicated arguments. It turns out that the easiest and most informative approach for us will be to uniquely characterize the matching that is obtained and
then show that al! executions result in the matching with this characterization.
What is the characterization? Well show that each man ends up with the
"best possible partner" in a concrete sense. (Recall that this is true if all men
prefer different women.) First, we will say that a woman iv is a valid partner
of a man m if there is a stable matching that contains the pair (m, iv). We will
say that iv is the best valid partner of m if iv is a valid parmer of m, and no
woman whom m ranks higher than iv is a valid partner of his. We will use
best(m) to denote the best valid partner of m.
Now, let S* denote the set of pairs {(m, best(m)) : m ~ M}. We will prove
the folloWing fact.
(1.7) Every execution of the C--S algorithm results in the set S*:
This statement is surprising at a number of levels. First of all, as defined,
there is no reason to believe that S* is a matching at all, let alone a stable
matching. After all, why couldnt it happen that two men have the same best
valid partner? Second, the result shows that the G-S algorithm gives the best
possible outcome for every man simultaneously; there is no stable matching
in which any of the men could have hoped to do better. And finally, it answers

1.1 A First Problem: Stable Matching

our question above by showing that the order of proposals in the G-S algorithm
has absolutely no effect on the final outcome.
Despite all this, the proof is not so difficult.
Proof. Let us suppose, by way of contradiction, that some execution g of the
G-S algorithm results in a matching S in which some man is paired with a
woman who is not his best valid partner. Since men propose in decreasing
order of preference, this means that some man is rejected by a valid partner
during the execution g of the algorithm. So consider the first moment during
the execution g in which some man, say m, is rejected by a valid partner iv.
Again, since men propose in decreasing order of preference, and since this is
the first time such a rejection has occurred, it must be that iv is ms best valid
partner best(m).
The reiection of m by iv may have happened either because m proposed
and was turned down in favor of ivs existing engagement, or because iv broke
her engagement to m in favor of a better proposal. But either way, at this
moment iv forms or continues an engagement with a man m whom she prefers
to m.

Since iv is a valid parmer of m, there exists a stable matching S containing


the pair (m, iv). Now we ask: Who is m paired with in this matching? Suppose
it is a woman iv ~= iv.
Since the rejection of m by iv was the first rejection of a man by a valid
partner in the execution ~, it must be that m had not been rejected by any valid
parmer at the point in ~ when he became engaged to iv. Since he proposed in
decreasing order of preference, and since iv is clearly a valid parmer of m, it
must be that m prefers iv to iv. But we have already seen that iv prefers m
to m, for in execution ~ she rejected m in favor of m. Since (m, iv) S, it
follows that (m, iv) is an instability in S.
This contradicts our claim that S is stable and hence contradicts our initial
assumption. []
So for the men, the G-S algorithm is ideal. Unfortunately, the same cannot
be said for the women. For a woman w, we say that m is a valid partner if
there is a stable matching that contains the pair (m, w). We say that m is the
ivorst valid partner of iv if m is a valid partner of w, and no man whom iv
ranks lower than m is a valid partner of hers.
(1.8) In the stable matching S*, each woman is paired ivith her ivorst valid
partner.
Proof. Suppose there were a pair (m, iv) in S* such that m is not the worst
valid partner of iv. Then there is a stable matching S in which iv is paired

11

12

Chapter 1 Introduction: Some Representative Problems

with a man m whom she likes less than m. In S, m is paired with a woman
w ~ w; since w is the best valid partner of m, and w is a valid partner of m,
we see that m prefers w to w.
But from this it follows that (m, w) is an instability in S, contradicting the
claim that S is stable and hence contradicting our initial assumption. []
Thus, we find that our simple example above, in which the mens preferences clashed with the womens, hinted at a very general phenomenon: for
any input, the side that does the proposing in the G-S algorithm ends up with
the best possible stable matching (from their perspective), while the side that
does not do the proposing correspondingly ends up with the worst possible
stable matching.

1.2 Five Representative Problems


The Stable Matching Problem provides us with a rich example of the process of
algorithm design. For many problems, this process involves a few significant,
steps: formulating the problem with enough mathematical precision that we
can ask a concrete question and start thinking about algorithms to solve
it; designing an algorithm for the problem; and analyzing the algorithm by
proving it is correct and giving a bound on the running time so as to establish
the algorithms efficiency.
This high-level strategy is carried out in practice with the help of a few
fundamental design techniques, which are very useful in assessing the inherent
complexity of a problem and in formulating an algorithm to solve it. As in any
area, becoming familiar with these design techniques is a gradual process; but
with experience one can start recognizing problems as belonging to identifiable
genres and appreciating how subtle changes in the statement of a problem can
have an enormous effect on its computational difficulty.
To get this discussion started, then,, it helps to pick out a few representative milestones that well be encountering in our study of algorithms: cleanly
formulated problems, all resembling one another at a general level, but differing greatly in their difficulty and in the kinds of approaches that one brings
to bear on them. The first three will be solvable efficiently by a sequence of
increasingly subtle algorithmic techniques; the fourth marks a major turning
point in our discussion, serving as an example of a problem believed to be unsolvable by any efficient algorithm; and the fifth hints at a class of problems
believed to be harder stil!.

The problems are self-contained and are al! motivated by computing


applications. To talk about some of them, though, it will help to use the
termino!ogy of graphs. While graphs are a common topic in earlier computer

1.2 Five Representative Problems

science courses, well be introducing them in a fair amount of depth in


Chapter 3; due to their enormous expressive power, well also be using them
extensively throughout the book. For the discussion here, its enough to think
of a graph G as simply a way of encoding pairwise relationships among a set
of objects. Thus, G consists of a pair of sets (V, E)--a collection V of nodes
and a collection E of edges, each of which "joins" two of the nodes. We thus
represent an edge e ~ E as a two-element subset of V: e = (u, u) for some
u, u ~ V, where we call u and u the ends of e. We typica!ly draw graphs as in
Figure 1.3, with each node as a small circle and each edge as a line segment
joining its two ends.

13

(a)

Lets now turn to a discussion of the five representative problems.

Interval Scheduling
Consider the following very simple scheduling problem. You have a resource-- Figure 1.3 Each of (a) and
it may be a lecture room, a supercompnter, or an electron microscope--and (b) depicts a graph on four
nodes.
many people request to use the resource for periods of time. A request takes
the form: Can I reserve the resource starting at time s, until time f? We will
assume that the resource can be used by at most one person at a time. A
scheduler wants to accept a subset of these requests, rejecting al! others, so
that the accepted requests do not overlap in time. The goal is to maximize the
number of requests accepted.
More formally, there will be n requests labeled 1 ..... n, with each request
i specifying a start time si and a finish time fi. Naturally, we have si < fi for all
i. Two requests i andj are compatible if the requested intervals do not overlap:
that is, either request i is for an earlier time interval than request j (fi <
or request i is for a later time than request j (1~ _< si). Well say more generally
that a subset A of requests is compatible if all pairs of requests i,j ~ A, i ~=j are
compatible. The goal is to select a compatible subset of requests of maximum
possible size.
We illustrate an instance of this Interual Scheduling Problem in Figure 1.4.
Note that there is a single compatible set of size 4, and this is the largest
compatible set.

Figure 1.4 An instance of the Interval Scheduling Problem.

14

Chapter ! Introduction: Some Representative Problems


We will see shortly that this problem can be solved by a very natural
algorithm that orders the set of requests according to a certain heuristic and
then "greedily" processes them in one pass, selecting as large a compatible
subset as it can. This will be .typical of a class of greedy algorithms that we
will consider for various problems--myopic rules that process the input one
piece at a time with no apparent look-ahead. When a greedy algorithm can be
shown to find an optimal solution for al! instances of a problem, its often fairly
surprising. We typically learn something about the structure of the underlying
problem from the fact that such a simple approach can be optimal.

Weighted Interval Scheduling


In the Interval Scheduling Problem, we sohght to maximize the number of
requests that could be accommodated simultaneously. Now, suppose more
generally that each request interval i has an associated value, or weight,
vi > O; we could picture this as the amount of money we will make from
the ith individual if we schedule his or her request. Our goal will be to find a
compatible subset of intervals of maximum total value.
The case in which vi = I for each i is simply the basic Interval Scheduling
Problem; but the appearance of arbitrary values changes the nature of the
maximization problem quite a bit. Consider, for example, that if v1 exceeds
the sum of all other vi, then the optimal solution must include interval 1
regardless of the configuration of the fi~l set of intervals. So any algorithm
for this problem must be very sensitive to the values, and yet degenerate to a
method for solving (unweighted) interval scheduling when all the values are
equal to 1.
There appears to be no simple greedy rule that walks through the intervals
one at a time, making the correct decision in the presence of arbitrary values.
Instead, we employ a technique, dynamic programming, that builds up the
optimal value over all possible solutions in a compact, tabular way that leads
to a very efficient algorithm.
Bipal~te Matching
When we considered the Stable Matching Problem, we defined a matching to
be a set of ordered pairs of men and women with the property that each man
and each woman belong to at most one of the ordered pairs. We then defined
a perfect matching to be a matching in which every man and every woman
belong to some pair.
We can express these concepts more generally in terms of graphs, and in
order to do this it is useful to define the notion of a bipartite graph. We say that
a graph G ---- (V, E) is bipa~te if its node set V can be partitioned into sets X

1.2 Five Representative Problems

15

and Y in such a way that every edge has one end in X and the other end in Y.
A bipartite graph is pictured in Figure 1.5; often, when we want to emphasize
a graphs "bipartiteness," we will draw it this way, with the nodes in X and
Y in two parallel columns. But notice, for example, that the two graphs in
Figure 1.3 are also bipartite.
Now, in the problem of finding a stable matching, matchings were built
from pairs of men and women. In the case of bipartite graphs, the edges are
pairs of nodes, so we say that a matching in a graph G = (V, E) is a set of edges
M _c E with the property that each node appears in at most one edge of M.
M is a perfect matching if every node appears in exactly one edge of M. Figure 1.5 A bipartite graph.
To see that this does capture the same notion we encountered in the Stable
Matching Problem, consider a bipartite graph G with a set X of n men, a set Y
of n women, and an edge from every node in X to every node in Y. Then the
matchings and perfect matchings in G are precisely the matchings and perfect
matchings among the set of men and women.

In the Stable Matching Problem, we added preferences to this picture. Here,


we do not consider preferences; but the nature of the problem in arbitrary
bipartite graphs adds a different source of complexity: there is not necessarily
an edge from every x ~ X to every y ~ Y, so the set of possible matchings has
quite a complicated structure. In other words, it is as though only certain pairs
of men and women are willing to be paired off, and we want to figure out
how to pair off many people in a way that is consistent with this. Consider,
for example, the bipartite graph G in Figure 1.5: there are many matchings in
G, but there is only one perfect matching. (Do you see it?)
Matchings in bipartite graphs can model situations in which objects are
being assigned to other objects. Thus, the nodes in X can represent jobs, the
nodes in Y can represent machines, and an edge (x~, y]) can indicate that
machine y] is capable of processing job xi. A perfect matching is then a way
of assigning each job to a machine that can process it, with the property that
each machine is assigned exactly one job. In the spring, computer science
departments across the country are often seen pondering a bipartite graph in
which X is the set of professors in the department, Y is the set of offered
courses, and an edge (xi, yj) indicates that professor x~ is capable of teaching
course y]. A perfect matching in this graph consists of an assignment of each
professor to a course that he or she can teach, in such a way that every course
is covered.

Thus the Bipartite Matching Problem is the following: Given an arbitrary


bipartite graph G, find a matching of maximum size. If IXI = I YI = n, then there
is a perfect matching if and only if the maximum matching has size n. We will
find that the algorithmic techniques discussed earlier do not seem adequate

16

Chapter 1 Introduction: Some Representative Problems

for providing an efficient algorithm for this problem. There is, however, a very
elegant and efficient algorithm to find a maximum matching; it inductively
builds up larger and larger matchings, selectively backtracking along the way.
This process is called augmeritation, and it forms the central component in a
large class of efficiently solvable problems called network flow problems.

Independent Set

Figure 1.6 A graph whose


largest independent set has
size 4.

Now lets talk about an extremely general problem, which includes most of
these earlier problems as special cases. Given a graph G = (V, E), we say
a set of nodes S _ V is independent if no two nodes~in S are joined by an
edge. The Independent Set Problem is, then, the following: Given G, find an
independent set that is as large as possible. For example, the maximum size of
an independent set in the graph in Figure 1.6 is four, achieved by the.four-node
independent set [1, 4, 5, 6}.
The Independent Set Problem encodes any situation in which you are
trying to choose from among a collection of objects and there are pairwise
conflicts among some of the objects. Say you have n friends, and some pairs
of them dont get along. How large a group of your friends can you invite to
dinner if you dont want any interpersonal tensions? This is simply the largest
independent set in the graph whose nodes are your friends, with an edge
between each conflicting pair.
Interval Scheduling and Bipartite Matching can both be encoded as special
cases of the Independent Set Problem. For Interval Scheduling, define a graph
G = (V, E) in which the nodes are the intervals and there is an edge between
each pair of them that overlap; the independent sets in G are then just the
compatible subsets of intervals. Encoding Bipartite Matching as a special case
of Independent Set is a little trickier to see. Given a bipartite graph G = (V, E),
the objects being chosen are edges, and the conflicts arise between two edges
that share an end. (These, indeed, are the pairs of edges that cannot belong
to a common matching.) So we define a graph G = (V, E) in which the node
set V is equal to the edge set E of G. We define an edge between each pair
of elements in V that correspond to edges of G with a common end. We can
now check that the independent sets of G are precisely the matchings of G.
While it is not complicated to check this, it takes a little concentration to deal
with this type of "edges-to-nodes, nodes-to-edges" transformation.2

2 For those who are curious, we note that not every instance of the Independent Set Problem can arise
in this way from Interval Scheduling or from Bipartite Matching; the full Independent Set Problem
really is more general. The graph in Figure 1.3(a) cannot arise as the "conflict graph" in an instance of

1.2 Five Representative Problems

Given the generality of the Independent Set Problem, an efficient algorithm


to solve it would be quite impressive. It would have to implicitly contain
algorithms for Interval Scheduling, Bipartite Matching, and a host of other
natural optimization problems.
The current status of Independent Set is this: no efficient algorithm is
known for the problem, and it is conjectured that no such algorithm exists.
The obvious brute-force algorithm would try all subsets of the nodes, checking
each to see if it is independent, and then recording the largest one encountered.
It is possible that this is close to the best we can do on this problem. We will
see later in the book that Independent Set is one of a large class of problems
that are termed NP-compIete. No efficient algorithm is known for any of them;
but they are all equivalent in the sense that a solution to any one of them
would imply, in a precise sense, a solution to all of them.
Heres a natural question: Is there anything good we can say about the
complexity of the Independent Set Problem? One positive thing is the following:
If we have a graph G on 1,000 nodes, and we want to convince you that it
contains an independent set S of size 100, then its quite easy. We simply
show you the graph G, circle the nodes of S in red, and let you check that
no two of them are joined by an edge. So there really seems to be a great
difference in difficulty between checking that something is a large independent
set and actually finding a large independent set. This may look like a very basic
observation--and it is--but it turns out to be crucial in understanding this class
of problems. Furthermore, as well see next, its possible for a problem to be
so hard that there isnt even an easy way to "check" solutions in this sense.

Competitive Facility Location


Finally, we come to our fifth problem, which is based on the following twoplayer game. Consider two large companies that operate caf6 franchises across
the country--lets call them JavaPlanet and Queequegs Coffee--and they are
currently competing for market share in a geographic area. First JavaPlanet
opens a franchise; then Queequegs Coffee opens a franchise; then JavaPlanet;
then Queequegs; and so on. Suppose they must deal with zoning regulations
that require no two franchises be located too close together, and each is trying
to make its locations as convenient as possible. Who will win?
Lets make the rules of this "game" more concrete. The geographic region
in question is divided into n zones, labeled 1, 2 ..... n. Each zone i has a
Interval Scheduling, and the graph in Figure 1.3(b) cannot arise as the "conflict graph" in an instance
of Bipartite Matching.

17

18

Chapter 1 Introduction: Some Representative Problems

Solved Exercises

Solved Exercises
Figure 1.7 An instance of the Competitive FaciBt3, Location Problem.

value bi, which is the revenue obtained by either of the companies if it opens
a franchise there. Finally, certain pairs of zones (i,]) are adjacent, and local
zoning laws prevent two adjacent zones from each containing a franchise,
regardless of which company owns them. (They also prevent two franchises
from being opened in the same zone.) We model these conflicts via a graph
G= (V,E), where V is the set of zones, .and ~(i,]) is an edge in E if the
zones i and ] are adiacent. ~The zoning requirement then says that the full
set of franchises opened must form an independent set in G.
Thus our game consists of two players, P1 and P2, alternately selecting
nodes in G, with P1 moving first. At all times, the set of all selected nodes
must form an independent set in G. Suppose that player P2 has a target bound
B, and we want to know: is there a strategy for P2 so that no matter how P1
plays, P2 will be able to select a set of nodes with a total value of at least B?
We will call this an instance of the Competitive Facility Location Problem.
Consider, for example, the instance pictured in Figure 1.7, and suppose
that P2s target bound is B = 20. Then P2 does have a winning strategy. On the
other hand, if B = 25, then P2 does not.
One can work this out by looking at the figure for a while; but it requires
some amount of case-checking of the form, "If P~ goes here, then P2 will go
there; but if P~ goes over there, then P2 will go here .... "And this appears to
be intrinsic to the problem: not only is it compntafionally difficult to determine
whether P2 has a winning strategy; on a reasonably sized graph, it would even
be hard for us to convince you that P2 has a winning strategy. There does not
seem to be a short proof we could present; rather, wed have to lead you on a
lengthy case-by-case analysis of the set of possible moves.
This is in contrast to the Independent Set Problem, where we believe that
finding a large solution is hard but checking a proposed large solution is easy.
This contrast can be formalized in the class of PSPACE-complete problems, of
which Competitive Facility Location is an example.. PSPACE-complete problems are believed to be strictly harder than NP-complete problems, and this
conjectured lack of short "proofs" for their solutions is one indication of this
greater hardness. The notion of PSPACE-completeness turns out to capture a
large collection of problems involving game-playing and planning; many of
these are fundamental issues in the area of artificial intelligence.

Solved Exercise 1
Consider a town with n men and n women seeking to get married to one
another. Each man has a preference list that ranks all the women, and each
woman has a preference list that ranks all the men.
The set of all 2n people is divided into two categories: good people and
bad people. Suppose that for some number k, 1 < k < n - 1, there are k good
men and k good women; thus there are n - k bad men and n - k bad women.
Everyone would rather marry any good person than any bad person.
Formally, each preference list has the property that it ranks each good person
of the opposite gender higher than each bad person of the opposite gender: its
first k entries are the good people (of the opposite gender) in some order, and
its next n - k are the bad people (of the opposite gender) in some order.
Show that in every stable matching, every good man is married to a good
woman.

Solution A natural way to get started thinking about this problem is to


assume the claim is false and try to work toward obtaining a contradiction.
What would it mean for the claim to be false? There would exist some stable
matching M in which a good man m was married to a bad woman w.
Now, lets consider what the other pairs in M look like. There are k good
men and k good women. Could it be the case that every good woman is married
to a good man in this matching M? No: one of the good men (namely, m) is
already married to a bad woman, and that leaves only k - ! other good men.
So even if all of them were married to good women, that would still leave some
good woman who is married to a bad man.
Let w be such a good woman, who is married to a bad man. It is now
easy to identify an instability in M: consider the pair (m, w). Each is good,
but is married to a bad partner. Thus, each of m and w prefers the other to
their current partner, and hence (m, w) is an instability. This contradicts our
assumption that M is stable, and hence concludes the proof.

Solved Exercise 2
We can think about a generalization of the Stable Matching Problem in which
certain man-woman pairs are explicitly forbidden. In the case of employers and
applicants, we could imagine that certain applicants simply lack the necessary
qualifications or certifications, and so they cannot be employed at certain
companies, however desirable they may seem. Using the analogy to marriage
between men and women, we have a set M of n men, a set W of n women,

19

20

Chapter ! Introduction: Some Representative Problems

and a set F _q M x W of pairs who are simply not allowed to get married. Each
man m ranks all th6 women w for which (m, w) ~ F, and each woman w ranks
al! the men m for which (m, w) ~ F.
In this more general setting, we say that a matching S is stable if it does
not exhibit any of the following types of instability.
(i) There are two pairs (m, w) and (m, w) in S with the property that
(m, w) F, m prefers w to w, and w prefers m to m. (The usual kind
of instability.)
(ii) There is a pair (m, w) E S, and a man m, so that m is not part of any
pair in the matching, (m, w) F, and w prefers m to m. (A single man
is more desirable and not forbidden.)
(iii) There is a pair (m, w) E S, and a woman W, so that w is not part of
any pair in the matching, (m, w) F, and m prefers w to w. (A single
woman is more desirable and not forbidden.)
(iv) There is a man m and a woman w, neither of whom is part of any pair
in the matching, so that (m, w) F. (There are two single people with
nothing preventing them from getting married to each other.)

Note that under these more general definitions, a stable matching need not be
a perfect matching.
Now we can ask: For every set of preference lists and every set of forbidden
pairs, is there always a stable matching? Resolve this question by doing one of
the following two things: (a) give an algorithm that, for any set of preference
lists and forbidden pairs, produces a stable matching; or (b) give an example
of a set of preference lists and forbidden pairs for which there is no stable
matching.
Solution The Gale-Shapley algorithm is remarkably robust to variations on
the Stable Matching Problem. So, if youre faced with a new variation of the
problem and cant find a counterexample to stability, its often a good idea to
check whether a direct adaptation of the G-S algorithm will in fact produce
stable matchings.
That turns out to be the case here. We will show that there is always a
stable matching, even in this more general model with forbidden pairs, and
we will do this by adapting the G-S algorithm. To do this, lets consider why
the original G-S algorithm cant be used directly. The difficulty, of course, is
that the G-S algorithm doesnt know anything about forbidden pairs, and so
the condition in the ghle loop,
While there is a man m who is free and hasnt proposed to
every woman,

Solved Exercises

wont work: we dont want m to propose to a woman w for which the pair
(m, w) is forbidden.
Thus, lets consider a variation of the G-S algorithm in which we make
only one change: we modify the Whle loop to say,
While there is a man m who is free and hasnt proposed to
every woman w for which (m,w) ~F.
Here is the algorithm in full.
Initially all m ~M and w ~W are free
While there is a man m who is free and hasnt proposed to
every woman w for which (m, w) ~F
Choose ~uch a man m
Let ~ be the highest-ranked woman in ms preference list
to which m has not yet proposed
If ~ is free then
(m,~) become engaged
Else w is currently engaged to m
If w prefers m to m then
m remains free
Else ~ prefers m to m
(m,~) become engaged
mt becomes free
Endif
Endif
Endwhile
Keturn the set S of engaged pairs

We now prove that this yields a stable matching, under our new definition
of stabi~ty.
To begin with, facts (1.1), (1.2), and (1.5) from the text remain true (in
particular, the algorithm will terminate in at most n2 iterations]. Also, we
dont have to worry about establishing that the resulting matching S is perfect
(indeed, it may not be]. We also notice an additional pairs of facts. If m is
a man who is not pan of a pair in S, then m must have proposed to every
nonforbidden woman; and if w is a woman who is not part of a pair in S, then
it must be that no man ever proposed to w.
Finally, we need only show
(1.9) There is no instability with respect to the returned matching S.

21

22

Chapter 1 Introduction: Some Representative Problems

Proof. Our general definition of instability has four parts: This means that we
have to make sure that none of the four bad things happens.
First, suppose there is an instability of type (i), consisting of pairs (m, w)
and (m, w) in S with the property that (m, w) ~ F, m prefers w to w, and w
prefers m to m. It follows that m must have proposed to w; so w rejected rn,
and thus she prefers her final partner to m--a contradiction.
Next, suppose there is an instability of type (ii), consisting of a pair
(m, w) ~ S, and a man m, so that m is not part of any pair in the matching,
(m, w) ~ F, and w prefers m to m. Then m must have proposed to w and
been rejected; again, it follows that w prefers her final partner to
contradiction.
Third, suppose there is an instability of type (iii), consisting of a pair
(m, w) ~ S, and a woman w, so that w is not part of any. pair in the matching,
(m, w) ~ F, and rn prefers w to w. Then no man proposed to w at all;
in particular, m never proposed to w, and so he must prefer w to
contradiction.
Finally, suppose there is an instability of type (iv), consisting of a man
m and a woman w, neither of which is part of any pair in the matching,
so that (m, w) ~ F. But for ra to be single, he must have proposed to every
nonforbidden woman; in particular, he must have proposed tow, which means
she would no longer be single--a contradiction. []

Exercises
Decide whether you think the following statement is true or false. If it is
true, give a short explanation. If it is false, give a counterexample.
True or false? In every instance of the Stable Matching Problem, there is a
stable matching containing a pair (m, w) such that m is ranked first on the
preference list of w and w is ranked first on the preference list of m.
Decide whether you think the following statement is true or false. If it is
true, give a short explanation. If it is false, give a cotmterexample.
True or false? Consider an instance of the Stable Matching Problem in which
there exists a man m and a woman w such that m is ranked first on the
preference list of w and w is ranked first on the preference list of m. Then in
every stable matching S for this instance, the pair (m, w) belongs to S.

3. There are many other settings in which we can ask questions related
to some type of "stability" principle. Heres one, involx4ng competition
between two enterprises.

Exercises
Suppose we have two television networks, whom well call A and ~B.
There are n prime-time programming slots, and each network has n TV
shows. Each network wants to devise a schedule--an assignment of each
show to a distinct slot--so as to attract as much market share as possible.
Here is the way we determine how well the two networks perform
relative to each other, given their schedules. Each show has a fixed rating,
which is based on the number of people who watched it last year; well
assume that no two shows have exactly the same rating. A network wins a
given time slot if the show that it schedules for the time slot has a larger
rating than the show the other network schedules for that time slot. The
goal of each network is to win as many time slots as possible.
Suppose in the opening week of the fall season, Network A reveals a
schedule S and Network ~B reveals a schedule T. On the basis of this pair
of schedules, each network wins certain time slots, according to the rule
above. Well say that the pair of schedules (S, T) is stable if neither network
can unilaterally change its own schedule and win more time slots. That
is, there is no schedule S such that Network ~t wins more slots with the
pair (S, T) than it did with the pair (S, T); and symmetrically, there is no
schedule T such that Network ~B wins more slots with the pair (S, T) than
it did with the pair (S, T).
The analogue of Gale and Shapleys question for this kind of stability
is the following: For every set of TV shows and ratings, is there always
a stable pair of schedules? Resolve this question by doing one of the
following two things:
(a) give an algorithm that, for any set of TV shows and associated
ratings, produces a stable pair of schedules; or
(b) give an example of a set of TV shows and associated ratings for
which there is no stable pair of schedules.
Gale and Shapley published their paper on the Stable Matching Problem
in 1962; but a version of their algorithm had already been in use for
ten years by the National Resident Matching Program, for the problem of
assigning medical residents to hospitals.
Basically, the situation was the following. There were m hospitals,
each with a certain number of available positions for hiring residents.
There were n medical students graduating in a given year, each interested
in joining one of the hospitals. Each hospital had a ranking of the students
in order of preference, and each student had a ranking of the hospitals
in order of preference. We will assume that there were more students
graduating than there were slots available in the m hospitals.

23

24

Exercises

Chapter 1 Introduction: Some Representative Problems

The interest, naturally, was in finding a way of assigning each student


to at most one hospital, in such a way that all available positions in all
hospitals were filled. (Since we are assuming a surplus of students, there
would be some students who do not get assigned to any hospital.)
We say that an assignment of students to hospitals is stable ff neither
of the following situations arises.
First type of instability: There are students s and s, and a hospital h,
so that
- s is assigned to h, and
- s is assigned to no hospital, and
- h prefers s to s.
Second
type of instability: There are students s and s~, and hospitals

t~ and h, so that
- s is assigned to h, and
s is assigned to tff, and
- t~ prefers s to s, and
- s prefers tt to h.
So we basically have the Stable Matching Problem, except that (i)
hospitals generally want more than one resident, and (ii) there is a surplus
of medical students.
Show that there is always a stable assignment of students to hospitals, and give an algorithm to find one.
The Stable Matching Problem, as discussed in the text, assumes that all
men and women have a fully ordered list of preferences. In this problem
we will consider a version of the problem in which men and women can be
indifferent between certain options. As before we have a set M of n men
and a set W of n women. Assume each man and each woman ranks the
members of the opposite gender, but now we allow ties in the ranking.
For example (with n = 4), a woman could say that ml is ranked in first
place; second place is a tie between mz and m3 (she has no preference
between them); and m4 is in last place. We will say that tv prefers m to m
if m is ranked higher than m on her preference list (they are not tied).
With indifferences in the ranldngs, there could be two natural notions
for stability. And for each, we can ask about the existence of stable
matchings, as follows.
(a) A strong instability in a perfect matching S consists of a man m and
a woman tv, such that each of m and tv prefers the other to their
partner in S. Does there always exist a perfect matching with no

strong instability? Either give an example of a set of men and women


with preference lists for which every perfect matching has a strong
instability; or give an algorithm that is guaranteed to find a perfect
matching with no strong instability.

(b) A weak instability in a perfect matching S consists of a man m and


a woman tv, such that their partners in S are tv and m, respectively,
and one of the following holds:
- m prefers u~ to ui, and tv either prefers m to m or is indifferent
be~veen these two choices; or
u~ prefers m to m, and m either prefers u~ to u3 or is indifferent
between these two choices.
In other words, the pairing between m and tv is either preferred
by both, or preferred by one while the other is indifferent. Does
there always exist a perfect matching with no weak instability? Either
give an example of a set of men and women with preference lists
for which every perfect matching has a weak instability; or give an
algorithm that is guaranteed to find a perfect matching with no weak
instability.
6. Peripatetic Shipping Lines, inc., is a shipping company that owns n ships
and provides service to n ports. Each of its ships has a schedule that says,
for each day of the month, which of the ports its currently visiting, or
whether its out at sea. (You can assume the "month" here has m days,
for some m > n.) Each ship visits each port for exactly one day during the
month. For safety reasons, PSL Inc. has the following strict requirement:
(t) No two ships can be in the same port on the same day.
The company wants to perform maintenance on all the ships this
month, via the following scheme. They want to truncate each ships
schedule: for each ship Sg, there will be some day when it arrives in its
scheduled port and simply remains there for the rest of the month (for
maintenance). This means that S~ will not visit the remaining ports on
its schedule (if any) that month, but this is okay. So the truncation of
S~s schedule will simply consist of its original schedule up to a certain
specified day on which it is in a port P; the remainder of the truncated
schedule simply has it remain in port P.
Now the companys question to you is the following: Given the schedule for each ship, find a truncation of each so that condition (t) continues
to hold: no two ships are ever in the same port on the same day.
Show that such a set of truncations can always be found, and give an
algorithm to find them.

25

26

Exercises

Chapter 1 Introduction: Some Representative Problems

Example. Suppose we have two ships and two ports, and the "month" has
four days. Suppose the first ships schedule is
port P1; at sea; port P2~ at sea
and the second ships schedule is

Junction
,, ",

~Junction

Output 1
(meets Input 2
before Input 1)

at sea; port Pff at sea; port P2 Then the (only) way to choose truncations would be to have the first ship
remain in port Pz starting on day 3, and have the second ship remain in
port P1 starting on day 2.
Some of your friends are working for CluNet, a builder of large communication networks, and they, are looking at algorithms for switching in a
particular type of input/output crossbar.
Here is the setup. There are n input wires and rt output wires, each
directed from a source to a terminus. Each input wire meets each output
;~e in exactly one distinct point, at a special piece of hardware called
a junction box. Points on the ~e are naturally ordered in the direction
from source to terminus; for two distinct points x and y on the same
wire, we say, that x is upstream from y if x is closer to the source than
y, and otherwise we say, x is downstream from y. The order in which one
input wire meets the output ~es is not necessarily the same as the order
in which another input wire meets the output wires. (And similarly for
the orders in which output wires meet input wires.) Figure !.8 gives an
example of such a collection of input and output wires.
Now, heres the switching component of this situation. Each input
~e is carrying a distinct data stream, and this data stream must be
switched onto one of the output wqres. If the stream of Input i is switched
onto Output j, at junction box B, then this stream passes through all
junction boxes upstream from B on input i, then through B, then through
all junction boxes downstream from B on Output j. It does not matter
;vhich input data stream gets switched onto which output wire, but
each input data stream must be switched onto a different output wire.
Furthermore--and this is the trick3, constraint--no two data streams can
pass through the same junction box following the switching operation.
Finally, heres the problem. Show that for any specified pattern in
which the input wires and output wires meet each other (each pair meeting exactly once), a valid switching of the data streams can always be
found--one in which each input data stream is switched onto a different
output, and no two of the resulting streams pass through the same junction box. Additionally, give an algorithm to find such a valid switching.

Output 2
(meets Input 2
before Input 1)

Junction [Junction

Input 1
(meets Output 2
before Output 1)

Input 2
(meets Output 1
before Output 2)

Figure 1.8 An example with two input wires and two output wires. Input 1 has its
junction with Output 2 upstream from its junction with Output 1; Input 2 has its
junction with Output 1 upstream from its junction with Output 2. A valid solution is
to switch the data stream of Input 1 onto Output 2, and the data stream of Input 2
onto Output 1. On the other hand, if the stream of Input 1 were switched onto Output
1, and the stream of Input 2 were switched onto Output 2, then both streams would
pass through the junction box at the meeting of Input 1 and Output 2--and this is not
allowed.

For this problem, we will explore the issue of truthfulness in the Stable
Matching Problem and specifically in the Gale-Shapley algorithm. The
basic question is: Can a man or a woman end up better off by lying about
his or her preferences? More concretely, we suppose each participant has
a true preference order. Now consider a woman w. Suppose w prefers man
m to m, but both m and m are low on her list of preferences. Can it be the
case that by switching the order of m and ra on her list of preferences (i.e.,
by falsely claiming that she prefers m to m) and nmning the algorithm
with this false preference list, w will end up with a man m" that she truly
prefers to both m and m? (We can ask the same question for men, but
will focus on the case of women for purposes of this question.)
Resolve this question by doing one of the following two things:
(a) Give a proof that, for any set of preference lists, switching the
order of a pair on the list cannot improve a womans partner in the GaleShapley algorithm; or

27

28

Chapter 1 Introduction: Some Representative Problems

(b) Give an example of a set of preference lists for which there is


a switch that-Would improve the partner of a woman who switched
preferences.

Notes and Further Reading


The Stable Matching Problem was ~st defined and analyzed by Gale and
Shapley (1962); according to David Gale, their motivation for the problem
came from a story they had recently read in the Netv Yorker about the intricacies
of the college admissions process (Gale, 2001). Stable matching has grown
into an area of study in its own right, covered in books by Gusfield and Irving
(1989) and Knuth (1997c). Gusfield and Irving also provide a nice survey of
the "paralle!" history of the Stable Matching Problem as a technique invented
for matching applicants with employers in medicine and other professions.
As discussed in the chapter, our five representative problems will be
central to the books discussions, respectively, of greedy algorithms, dynamic
programming, network flow, NP-completeness, and pSPACE-completeness.
We will discuss the problems in these contexts later in the book.

Basics A~gorithm Analysis

Analyzing algorithms involves thinking about how their resource requirements--the amount of time and space they use--will scale with increasing
input size. We begin this chapter by talking about how to put this notion on a
concrete footing, as making it concrete opens the door to a rich understanding
of computational tractability. Having done this, we develop the mathematical
machinery needed to talk about the way in which different functions scale
with increasing input size, making precise what it means for one function to
grow faster than another.
We then develop running-time bounds for some basic algorithms, beginning with an implementation of the Gale-Shapley algorithm from Chapter 1
and continuing to a survey of many different running times and certain characteristic types of algorithms that achieve these running times. In some cases,
obtaining a good running-time bound relies on the use of more sophisticated
data structures, and we conclude this chapter with a very useful example of
such a data structure: priority queues and their implementation using heaps.

2.1 Computational Tractability


A major focus of this book is to find efficient algorithms for computational
problems. At this level of generality, our topic seems to ,encompass the whole
of computer science; so what is specific to our approach here?
First, we will txy to identify broad themes and design principles in the
development of algorithms. We will look for paradigmatic problems and approaches that illustrate, with a minimum of irrelevant detail, the basic approaches to designing efficient algorithms. At the same time, it would be
pointless to pursue these design principles in a vacuum, so the problems and

30

Chapter 2 Basics of Algorithm Analysis

approaches we consider are drawn from fundamental issues that arise throughout computer science, and a general study of algorithms turns out to serve as
a nice survey of computationa~ ideas that arise in many areas.
Another property shared by many of the problems we study is their
fundamentally discrete nature. That is, like the Stable Matching Problem, they
will involve an implicit search over a large set of combinatorial possibilities;
and the goal will be to efficiently find a solution that satisfies certain clearly
delineated conditions.
As we seek to understand the general notion of computational efficiency,
we will focus primarily on efficiency in running time: we want algorithms that
run quickly. But it is important that algorithms be efficient in their use of other
resources as well. In particular, the amount of space (or memory) used by an
algorithm is an issue that will also arise at a number of points in the book, and
we will see techniques for reducing the amount of space needed to perform a
computation.

Some Initial Attempts at Defining Efficiency


The first major question we need to answer is the following: How should we
turn the fuzzy notion of an "efficient" algorithm into something more concrete?
A first attempt at a working definition of efficiency is the following.
Proposed Definition of Efficiency (1): An algorithm is efficient if, when
implemented, it runs quickly on real input instances.
Lets spend a little time considering this definition. At a certain leve!, its hard
to argue with: one of the goals at the bedrock of our study of algorithms is
solving real problems quickly. And indeed, there is a significant area of research
devoted to the careful implementation and profiling of different algorithms for
discrete computational problems.
But there are some crucial things missing from this definition, even if our
main goal is to solve real problem instances quickly on real computers. The
first is the omission of where, and how well, we implement an algorithm. Even
bad algorithms can run quickly when applied to small test cases on extremely
fast processors; even good algorithms can run slowly when they are coded
sloppily. Also, what is a "real" input instance? We dont know the ful! range of
input instances that will be encountered in practice, and some input instances
can be much harder than others. Finally, this proposed deflation above does
not consider how well, or badly, an algorithm may scale as problem sizes grow
to unexpected levels. A common situation is that two very different algorithms
will perform comparably on inputs of size 100; multiply the input size tenfold,
and one will sti!l run quickly while the other consumes a huge amount of time.

2.1 Comput.atipnal Tractability

So what we could ask for is a concrete definition of efficiency that is


platform-independent, instance-independent, and of predictive value with
respect to increasing input sizes. Before focusing on any specific consequences
of this claim, we can at least explore its implicit, high-level suggestion: that
we need to take a more mathematical view of the situation.
We can use the Stable Matching Problem as an example to guide us. The
input has a natural "size" parameter N; we could take this to be the total size of
the representation of all preference lists, since this is what any algorithm for the
problem wi!l receive as input. N is closely related to the other natural parameter
in this problem: n, the number of men and the number of women. Since there
are 2n preference lists, each of length n, we can view N = 2n2, suppressing
more fine-grained details of how the data is represented. In considering the
problem, we will seek to describe an algorithm at a high level, and then analyze
its running time mathematically as a function of this input size N.

Worst-Case Running Times and Brute-Force Search


To begin with, we will focus on analyzing the worst-case running time: we will
look for a bound on the largest possible running time the algorithm could have
over all inputs of a given size N, and see how this scales with N. The focus on
worst-case performance initially seems quite draconian: what if an algorithm
performs well on most instances and just has a few pathological inputs on
which it is very slow? This certainly is an issue in some cases, but in general
the worst-case analysis of an algorithm has been found to do a reasonable job
of capturing its efficiency in practice. Moreover, once we have decided to go
the route of mathematical analysis, it is hard to find an effective alternative to
worst-case analysis. Average-case analysis--the obvious appealing alternative,
in which one studies the performance of an algorithm averaged over "random"
instances--can sometimes provide considerable insight, but very often it can
also become a quagmire. As we observed earlier, its very hard to express the
full range of input instances that arise in practice, and so attempts to study an
algorithms performance on "random" input instances can quickly devolve into
debates over how a random input should be generated: the same algorithm
can perform very well on one class of random inputs and very poorly on
another. After all, real inputs to an algorithm are generally not being produced
from a random distribution, and so average-case analysis risks telling us more
about the means by which the random inputs were generated than about the
algorithm itself.
So in general we will think about the worst-case analysis of an algorithms
running time. But what is a reasonable analytical benchmark that can tell us
whether a running-time bound is impressive or weak? A first simple guide

"31

32

Chapter 2 Basics of Algorithm Analysis

is by comparison with brute-force search over the search space of possible


solutions.
Lets return to the example of the Stable Matching Problem. Even when
the size of a Stable Matching input instance is relatively small, the search
space it defines is enormous (there are n! possible perfect matchings between
n men and n women), and we need to find a matching that is stable. The
natural "brute-force" algorithm for this problem would plow through all perfect
matchings by enumeration, checking each to see if it is stable. The surprising
punchline, in a sense, to our solution of the Stable Matching Problem is that we
needed to spend time proportional only to N in finding a stable matching from
amgng this stupendously large space of possibilities. This was a conclusion we
reached at an analytical level. We did not implement the algorithm and try it
out on sample preference lists; we reasoned about it mathematically. Yet, at the
same time, our analysis indicated how the algorithm could be implemented in
practice and gave fairly conclusive evidence that it would be a big improvement
over exhaustive enumeration.
This will be a common theme in most of the problems we study: a compact
representation, implicitly specifying a giant search space. For most of these
problems, there will be an obvious brute-force solution: try all possibilities
and see if any one of them works. Not only is this approach almost always too
slow to be useful, it is an intellectual cop-out; it provides us with absolutely
no insight into the structure of the problem we are studying. And so if there
is a common thread in the algorithms we emphasize in this book, it would be
the following alternative definition of efficiency.
Proposed Definition of Efficiency (2): An algorithm is efficient if it achieves
qualitatively better worst-case performance, at an analytical level, than
brute-force search.
This will turn out to be a very usefu! working definition for us. Algorithms
that improve substantially on brute-force search nearly always contain a
valuable heuristic idea that makes them work; and they tell us something
about the intrinsic structure, and computational tractability, of the underlying
problem itself.
But if there is a problem with our second working definition, it is vagueness. What do we mean by "qualitatively better performance?" This suggests
that we consider the actual running time of algorithms more carefully, and try
to quantify what a reasonable running time would be.

Polynomial Time as a Definition of Efficiency


When people first began analyzing discrete algorithms mathematicalfy--a
thread of research that began gathering momentum through the 1960s--

2.1 Computational Tractability

a consensus began to emerge on how to quantify the notion of a "reasonable"


running time. Search spaces for natural combinatorial problems tend to grow
exponentially in the size N of the input; if the input size increases by one, the
number of possibilities increases multiplicatively. Wed like a good algorithm
for such a problem to have a better scaling property: when the input size
increases by a constant factor--say, a factor of 2--the algorithm should only
slow down by some constant factor C.

Arithmetically, we can formulate this scaling behavior as follows. Suppose


an algorithm has the following property: There are absolute constants c > 0
and d > 0 so that on every input instance of size N, its running time is
bounded by cNd primitive computational steps. (In other words, its running
time is at most proportional to Nd.) For now, we will remain deliberately
vague on what we mean by the notion of a "primitive computational step"but it can be easily formalized in a model where each step corresponds to
a single assembly-language instruction on a standard processor, or one line
of a standard programming language such as C or Java. In any case, if this
running-time bound holds, for some c and d, then we say that the algorithm
has a polynomial running time, or that it is a polynomial-time algorithm. Note
that any polynomial-time bound has the scaling property were looking for. If
the input size increases from N to 2N, the bound on the running time increases
from cNd to c(2N)a = c. 2aNa, which is a slow-down by a factor of 2a. Since d is
a constant, so is 2a; of course, as one might expect, lower-degree polynomials
exhibit better scaling behavior than higher-degree polynomials.
From this notion, and the intuition expressed above, emerges our third
attempt at a working definition of efficiency.

Proposed Definition of Efficiency (3)" An algorithm is efficient if it has a


polynomial running time.
Where our previous definition seemed overly vague, this one seems much
too prescriptive. Wouldnt an algorithm with running time proportional to
nl--and hence polynomial--be hopelessly inefficient? Wouldnt we be relatively pleased with a nonpolynomial running time of nl+02(lg n)? The answers are, of course, "yes" and "yes." And indeed, however, much one may
try to abstractly motivate the definition of efficiency in terms of polynomial
time, a primary justification for it is this: It really works. Problems for which
polynomial-time algorithms exist almost invariably turn out to have algorithms
with running times proportional to very moderately growing polynomials like
n, n log n, n2, or n3. Conversely, problems for which no polynomial-time algorithm is known tend to be very difficult in practice. There are certainly
exceptions to this principle in both directions: there are cases, for example, in

2.2 Asymptotic Order of Growth

Chapter 2 Basics of Algorithm Analysis

34

Table 2.1 The running times (rounded up) of different algorithms on inputs of
increasing size, for a processor performing a million high-leve! instructions per second.
In cases where the running time exceeds 10-s years, we simply record the algorithm as

n= I0
n=30
n=50
= 100
= 1,000
10,000
I00,000
1,000,000

< I see
< I sec
< 1 sec
< 1 sec
< 1 sec
< I sec
< 1 sec
1 sec

taking a very long time.


n2
n log rt

< I sec
< I sec
< I sec
< I sec
< 1 sec
< I sec
2 sec
20 sec

< I sec
< I sec
< 1 sec
< 1 sec
1 sec
2 min
3 hours
12 days

/73

1.5n

2n

n!

< I sec
< i sec
< 1 sec
1 sec
18 re_in
12 days
32 years
31,710 years

< I SeE
< I sec
11 mJn
12,892 years
very long
very long
very long
very long

< I sec
18 rain
36 years
1017 years
very long
very long
very long
very long

4 see
1025 years
very long
vgry long
very long
very long
very long
very long

which an algorithm with exponential worst-case behavior generally runs well


on the kinds of instances that arise in practice; and there are also cases where
the best polynomia!-time algorithm for a problem is completely impractical
due to large constants or a high exponent on the polynomial bound. All this
serves to reinforce the point that our emphasis on worst-case, polynomial-time
bounds is only an abstraction of practical situations. But overwhelmingly, the
concrete mathematical definition of polynomial time has turned out to correspond surprisingly wel! in practice to what we observe about the efficiency of
algorithms, and the tractability of problems, in tea! life.
One further reason why the mathematical formalism and the empirical
evidence seem to line up well in the case of polynomial-time solvability is that
the gulf between the growth rates of polynomial and exponential functions
is enormous. Suppose, for example, that we have a processor that executes
a million high-level instructions per second, and we have algorithms with
running-time bounds of n, n log2 n, n2, n3, 1.5n, 2n, and n!. In Table 2.1,
we show the running times of these algorithms (in seconds, minutes, days,
or years) for inputs of size n = 10, 50, 50,100, 1,000, 10,000,100,000, and
1,000,000.
There is a final, fundamental benefit to making our definition of efficiency
so specific: it becomes negatable. It becomes possible to express the notion
that there is no efficient algorithm for a particular problem. In a sense, being
able to do this is a prerequisite for turning our study of algorithms into
good science, for it allows us to ask about the existence or nonexistence
of efficient algorithms as a well-defined question. In contrast, both of our

previous definitions were completely subjective, and hence limited the extent
to which we could discuss certain issues in concrete terms.
In particular, the first of our definitions, which was tied to the specific
implementation of an algorithm, turned efficiency into a moving target: as
processor speeds increase, more and more algorithms fal! under this notion of
efficiency. Our definition in terms of polynomial time is much more an absolute
notion; it is closely connected with the idea that each problem has an intrinsic
level of computational tractability: some admit efficient solutions, and others
do not.

2.2 Asymptotic Order of Growth


Our discussion of computational tractability has turned out to be intrinsically
based on our ability to express the notion that an algorithms worst-case
running time on inputs of size n grows at a rate that is at most proportiona! to
some function f(n). The function f(n) then becomes a bound on the rtmning
time of the algorithm. We now discuss a framework for talking about this
concept.
We will mainly express algorithms in the pseudo-code style that we used
for the Gale-Shapley algorithm. At times we will need to become more formal,
but this style Of specifying algorithms will be completely adequate for most
purposes. When we provide a bound on the running time of an algorithm,
we will generally be counting the number of such pseudo-code steps that
are executed; in this context, one step wil! consist of assigning a value to a
variable, looking up an entry in an array, following a pointer, or performing
an arithmetic operation on a fixed-size integer.
When we seek to say something about the running time of an algorithm on
inputs of size n, one thing we could aim for would be a very concrete statement
such as, "On any input of size n, the algorithm runs for at most 1.62n2 +
3.5n + 8 steps." This may be an interesting statement in some contexts, but as
a general goal there are several things wrong with it. First, getting such a precise
bound may be an exhausting activity, and more detail than we wanted anyway.
Second, because our ultimate goal is to identify broad classes of algorithms that
have similar behavior, wed actually like to classify running times at a coarser
level of granularity so that similarities among different algorithms, and among
different problems, show up more clearly. And finally, extremely detailed
statements about the number of steps an algorithm executes are often--in
a strong sense--meaningless. As just discussed, we will generally be counting
steps in a pseudo-code specification of an algorithm that resembles a highlevel programming language. Each one of these steps will typically unfold
into some fixed number of primitive steps when the program is compiled into

35

36

Chapter 2 Basics of Algorithm Analysis

an intermediate representation, and then into some further number of steps


depending on the particular architecture being used to do the computing. So
the most we can safely say is that as we look at different levels of computational
abstraction, the notion of a "step" may grow or shrink by a constant factor-for example, .if it takes 25 low-level machine instructions to perform one
operation in our high-level language, then our algorithm that took at most
1.62n2 + 3.5n + 8 steps can also be viewed as taking 40.5n2 + 87.5n + 200 steps
when we analyze it at a level that is closer to the actual hardware.

O, s2, and
For all these reasons, we want to express the growth rate of running times
and other functions in a way that is insensitive to constant factors and loworder terms. In other words, wed like to be able to take a running time like
the one we discussed above, 1.62n2 + 3.5n + 8, and say that it grows like n2,
up to constant factors. We now discuss a precise way to do this.
Asymptotic Upper Bounds Let T(n) be a function--say, [he worst-case running time of a certain algorithm on an input of size n. (We will assume that
all the functions we talk about _here take nonnegative values.) Given another
function f(n), we say that T(n) is Off(n)) (read as "T(n) is order f(n)") if, for
sufficiently large n, the function T(n) is bounded above by a constant multiple
of f(n). We will also sometimes write this as T(n) = Off(n)). More precisely,
T(n) is Off(n)) if there exist constants c > 0 and no >_ 0 so that for all n >_ no,
we have T(n) <_ c. f(n). In this case, we will say that T is asymptotically upperbounded by f. It is important to note that this definition requires a constant c
to exist that works for all n; in particular, c cannot depend on n.
As an example of how this definition lets us express upper bounds on
running times, consider an algorithm whose running time (as in the earlier
discussion) has the form T(n) = pn2 + qn + r for positive constants p, q, and
r. Wed like to claim that any such function is O(n2). To see why, we notice
that for all n > 1, we have qn <_ qn2, and r < rn2. So we can write
T(n) = pn~ + qn + r < pn2 + qn2 q- yn2 = (P q- q q- r)n2
for all n >_ 1. This inequality is exactly what the definition of O(-) requires:
T(n) < cn2, where c =p + q + r.
Note that O(.) expresses only an upper bound, not the exact growth rate
of the function. For example, just as we claimed that the function T(n)=
pn2 + qn + r is O(n2), its also correct to say that its O(n3). Indeed, we just
argued that T(n) <_ (p + q + r)n2, and since we also have n2 < n3, we can
conclude that T(n) < (p + q + r)n~ as the definition of O(n~) requires. The
fact that a function can have many upper bounds is not just a trick of the
notation; it shows up in the analysis of running times as well. There are cases

2.2 Asymptotic Order of Growth

where an algorithm has been proved to have running time O(n~); some years
pass, people analyze the same algorithm more carefully, and they show that
in fact its running time is O(n2). There was nothing wrong with the first result;
it was a correct upper bound. Its simply that it wasnt the "tightest" possible
running time.
Asymptotic Lower Bounds There is a complementary notation for lower
bounds. Often when we analyze an algorithm--say we have just proven that
its worst-case running time T(n) is O(n2)--we want to show that this upper
bound is the best one possible. To do this, we want to express the notion that for
arbitrarily large input sizes n, the function T(n) is at least a constant multiple of
some specific function f(n). (In this example, f(n) happens to be n2.) Thus, we
say that T(n) is ~2 if(n)) (also written T(n) = S2 if(n))) if there exist constants
~ > 0 and no _> 0 so that for all n > n0, we have T(n) > ~. f(n). By analogy with
O(-) notation, we will refer to T in this case as being asymptotically lowerbounded by f. Again, note that the constant ~ must be fixed, independent
of n.
This definition works just like 0(.), except that we are bounding the
function T(n) from below, rather than from above. For example, returning
to the function T(n) = pn2 + qn + r, where p, q, and r are positive constants,
lets claim that T(n) = fl (n2). Whereas establishing the upper bound involved
"inflating" the terms in T(n) until it looked like a constant times n2, now we
need to do the opposite: we need to reduce the size of T(n) until it looks like
a constant times n2. It is not hard to do this; for all n >_ O, we have
T(n) = pn2 + qn + r > pn2,

which meets what is required by the definition of f2 (.) with ~ = p > 0.


Just as we discussed the notion of "tighter" and "weaker" upper bounds,
the same issue arises for lower bounds. For example, it is correct to say that
our function T(n) = pn2 + qn + r is S2 (n), since T(n) > pn2 > pn.
Asymptotically Tight Bounds If we can show that a running time T(n) is
both O(](n)) and also s2 ([(n)), then in a natural sense weve found the "right"
bound: T(n) grows exactly like [(n) to within a constant factor. This, for
example, is the conclusion we can draw from the fact that T(n) -=pn2 q- qn q- r
is both O(n2) and f2 (n2).
There is a notation to express this: if a function T(n) is both O([(n)) and
S2([(n)), we say that T(n) is ([(n)). In this case, we say that [(n) is an
asymptotically tight bound for T(n). So, for example, our analysis above shows
that T(n) = pn2 q- qn + r is (ha).
Asymptotically tight bounds on worst-case running times are nice things
to find, since they characterize the worst-case performance of an algorithm

37

58

2.2 Asymptotic Order of Growth

Chapter 2 Basics of Algorithm Analysis

precisely up to constant factors. And as the definition of (-) shows, one can
obtain such bounds by closing the gap between an upper bound and a lower
bound. For example, sometimes you will read a (slightly informally phrased)
sentence such as "An upper bound of O(n3) has been shown on the worst-case
running time of the algorithm, but there is no example known on which the
algorithm runs for more than f2 (n2) steps." This is implicitly an invitation to
search for an asymptotically tight bound on the algorithms worst-case running
time.
Sometimes one can also obtain an asymptotically tight bound directly by
computing a limit as n goes to infinity. Essentially, if the ratio of functions
f(n) and g(n) converges to a positive constant as n goes to infinity, then

f(n) = (g(n)).
(2.1)

Let f and g be two functions that


lim f(n___~)
.-->~ g(n)

exists and is equal to some number c > O. Then f(n) = (g(n)).


Proof. We will use the fact that the limit exists and is positive to show that
f(n) = O(g(n)) and f(n) = S2(g(n)), as re.quired by the definition of (.).

Since
lira f(n)
n-+oo
g(n) = c > 0,
it follows from the definition of a limit that there is some no beyond which the
ratio is always between c and 2c. Thus, f(n) < 2cg(n) for all n >_ no, which
implies that f(n) = O(g(n)); and [(n) >_ cg(n) for all n >_ no, which implies
that [(n) = ~(g(n)). []

Properties of Asymptotic Growth Rates


Having seen the definitions of O, S2, and O, it is useful to explore some of their
basic properties.
Transitivity A first property is transitivity: if a function f is asymptotically
upper-bounded by a function g, and if g in turn is asymptotically upperbounded by a function h, then f is asymptotically upper-bounded by h. A
similar property holds for lower bounds. We write this more precisely as
follows.

(z.2)
(a) !ff = O(g) and g = O(h), then f = O(h).
(b) If f = S2 (g) and g = ga (h), then f = ~2 (h).

Proof. Well prove part (a) of this claim; the proof of part (b) is very similar.
For (a), were given that for some constants c and n0, we have f(n) <_ cg(n)
for all n >_ n0. Also, for some (potentially different) constants c and n~, we
have g(n) <_ ch(n) for all n _> n~. So consider any number n that is at least as
large as both no and n~. We have f(n) < cg(n) < cch(n), and so f(n) < cch(n)
for all n > max(no, n~). This latter inequality is exactly what is required for
showing that f = O(h). ,,

Combining parts (a) and (b) of (2.2), we can obtain a similar result
for asymptotically tight bounds. Suppose we know that [ = (g) and that
g = (h). Then since [ = O(g) and g = O(h), we know from part (a) that
[ = O(h); since [ = S2(g) and g = S2(h), we know from part (b) that [ =
It follows that [ = (h). Thus we have shown
(2.3)

!f/=O(g) andg=(h),thenf=(h).

Sums of Functions It is also useful to have results that quantify the effect of
adding two functions. First, if we have an asymptotic upper bound that applies
to each of two functions f and g, then it applies to their sum.
(2.4) Suppose that f and g are two functions such that for some other function
h, we have f = O(h) and g = O(h). Then f + g = O(h).
Proof. Were given that for some constants c and no, we have f(n) <_ Ch(n)
for all n > no. Also, for some (potentially different) constants c and no,
we have g(n) < ch(n) for all n > no. So consider any number n that is at
least as large as both no and no. We have f(n) + g(n) <_ch(n) + ch(n). Thus
f(n) + g(n) <_ (c + c)h(n) for all n _> max(no, n~), which is exactly what is
required for showing that f + g = O(h). m
There is a generalization of this to sums of a fixed constant number of
functions k, where k may be larger than two. The result can be stated precisely
as follows; we omit the proof, since it is essenti!lly the same as the proof of
(2.4), adapted to sums consisting of k terms rather than just two.
(2.5) Let k be a fixed constant, and let fl, f2 ..... & and h be functions such
that fi = O(h) for all i. Then fl + f2 +"" + fk = O(h).

There is also a consequence of (2.4) that covers the following kind of


situation. It frequently happens that were analyzing an algorithm with two
high-level parts, and it is easy to show that one of the two parts is slower
than the other. Wed like to be able to say that the running time of the whole
algorithm is asymptotically comparable to the running time of the slow part.
Since the overall running time is a sum of two functions (the running times of

39

40

2.2 Asymptotic Order of Growth-

Chapter 2 Basics of Algorithm Analysis

the two parts), results on asymptotic bounds for sums of functions are directly
relevant.
(2.6) Suppose that [ and g are two functions (taking nonnegative values)
such that g = Off). Then [ + g = ([). In other words, [ is an asymptotically
tight bound for the combined function [ + g.
Proof. Clearly f + g = f2 (f), since for all n >~ 0, we have f(n) + g(n) ~ f(n).
So to complete the proof, we need to show that f + g = O(f).
But this is a direct consequence of (2.4): were given the fact that g = O(f),
and also f = O(f) holds for any function, so by (2.~) we have f + g = O(f). m

This result also extends to the sum of any fixed, constant number of
functions: the most rapidly growing among the functions is an asymptotically
tight bound for the sum.

Asymptotic Bounds for Some Common Functions


There are a number of functions that come up repeatedly in the analysis of
algorithms, and it is useful to consider the asymptotic properties of some of
the most basic of these: polynomials, logarithms, and exponentials.
Polynomials Recall that a polynomial is-a function that can be written in
the form f(n) = at + aln + a2n2 +" + aana for some integer constant d > 0,
where the final coefficient aa is nonzero. This value d is called the degree of the
polynomial. For example, the functions of the form pn2 + qn + r (with p ~ 0)
that we considered earlier are polynomials of degree 2.
A basic fact about polynomials is that their asymptotic rate of growth is
determined by their "high-order term"--the one that determines the degree.
We state this more formally in the following claim. Since we are concerned here
only with functions that take nonnegative values, we will restrict our attention
to polynomials for which the high-order term has a positive coefficient aa > O.
(2.7) Let f be a polynomial of degree d, in which the coefficient aa is positive.
Then f = O(nd).
Proof. We write f = ao + aln + a2n2 " " " aana, where aa > 0. The upper
bound is a direct application of (2.5). First, notice that coefficients aj forj < d
may be negative, but in any case we have ajnJ <_ lajlna for all n > 1. Thus each
term in the polynomial is O(na). Since f is a sum of a constant number of
functions, each of which is O(na), it follows from (2.5) that f is O(na).
One can also show that under the conditions of (2.7), we have f = f2 (ha),
and hence it follows that in fact f = (ha).

This is a good point at which to discuss the relationship between these


types of asymptotic bounds and the notion of polynomial time, which we
arrived at in the previous section as a way to formalize the more elusive concept
of efficiency. Using O(-) notation, its easy to formally define polynomial time:
apolynomiaI-time algorithm is one whose running time T(n) is O(nd) for some
constant d, where d is independent of the input size.
So algorithms with running-time bounds like O(n2) and O(n3) are
polynomial-time algorithms. But its important to realize that an algorithm
can be polynomial time even if its running time is not written as n raised
to some integer power. To begin with, a number of algorithms have running
times of the form O(nx) for some number x that is not an integer. For example,
in Chapter 5 we will see an algorithm whose running time is O(n1"59); we will
also see exponents less than 1, as in bounds like (J-K) = O(nl/2).
To take another common kind of example, we will see many algorithms
whose running times have the form O(n log n). Such algorithms are also
polynomial time: as we will see next, log n < n for all n > 1, and hence
n log n < n2 for all n > 1. In other words, if an algorithm has nmning time
O(n log n), then it also has running time O(n2), and so it is a polynomial-time
algorithm.
Logarithms Recall that logo n is the number x such that bx = n. One way
to get an approximate sense of how fast logb n grows is to note that, if we
round it down to the nearest integer, it is one less than the number of digits
in the base-b representation of the number n. (Thus, for example, 1 + log2 n,
rounded down, is the number of bits needed to represent n.)
So logarithms are very slowly growing functions. In particular, for every
base b, the function logo n is asymptotically bounded by every function of the
form nx, even for (noninteger) values of x arbitrary close to 0.
(2.8)

For every b > I and every x > O, we have logo n = O(nX).

One can directly translate between logarithms of different bases using the
following fundamental identity:
loga n --

logb n
logo a

This equation explains why youll often notice people writing bounds like
O(!og n) without indicating the base of the logarithm. This is not sloppy
usage: the identity above says that loga n =~1 !ogb n, so the point is that
loga n = (logo n), and the base of the logarithm is not important when writing
bounds using asymptotic notation

41

42

Chapter 2 Basics of Algorithm Analysis

Exponentials Exponential functions are functions of the form f(n) = rn for


some constant base r. Here we will be concerned with the case in which r > !,
which results in a very fast-growing function.
In particular, where polynomials raise rt to a fixed exponent, exponentials
raise a fixed number to n as a power; this leads to much faster rates of growth.
One way to summarize the relationship between polynomials and exponentials
is as follows.
(2.9) For every r > 1 and every d > O, we have na = O(rn).

In particular, every exponential grows faster thari every polynomial. And as


we saw in Table 2.1, when you plug in actual values of rt, the differences in
growth rates are really quite impressive.
Just as people write O(log rt) without specifying the base, youll also see
people write "The running time of this algorithm is exponential," without
specifying which exponential function they have in mind. Unlike the liberal
use of log n, which is iustified by ignoring constant factors, this generic use of
the term "exponential" is somewhat sloppy. In particular, for different bases
r > s > 1, it is never the case that rn = (sn). Indeed, this would require that
for some constant c > 0, we would have rn _< csn for all sufficiently large ft.
But rearranging this inequality would give (r/s)n < c for all sufficiently large
ft. Since r > s, the expression (r/s)n is. tending to infinity with rt, and so it
cannot possibly remain bounded by a fixed constant c.
So asymptotically speaking, exponential functions are all different. Still,
its usually clear what people intend when they inexactly write "The running
time of this algorithm is exponential"--they typically mean that the running
time grows at least as fast as some exponential function, and all exponentials
grow so fast that we can effectively dismiss this algorithm without working out
flLrther details of the exact running time. This is not entirely fair. Occasionally
theres more going on with an exponential algorithm than first appears, as
we!l see, for example, in Chapter 10; but as we argued in the first section of
this chapter, its a reasonable rule of thumb.
Taken together, then, logarithms, polynomials, and exponentials serve as
useful landmarks in the range of possible functions that you encounter when
analyzing running times. Logarithms grow more slowly than polynomials, and
polynomials grow more slowly than exponentials.

2.3 Implementing the Stable Matching Algorithm


IJsing Lists and Arrays
Weve now seen a general approach for expressing bounds on the running
time of an algorithm. In order to asymptotically analyze the running time of

2.3 Implementing the Stable Matching Algorithm Using Lists and Arrays

an algorithm expressed in a high-level fashion--as we expressed the GaleShapley Stable Matching algorithm in Chapter 1, for example--one doesnt
have to actually program, compile, and execute it, but one does have to think
about how the data will be represented and manipulated in an implementation
of the algorithm, so as to bound the number of computational steps it takes.
The implementation of basic algorithms using data structures is something
that you probably have had some experience with. In this book, data structures
will be covered in the context of implementing specific algorithms, and so we
will encounter different data structures based on the needs of the algorithms
we are developing. To get this process started, we consider an implementation
of the Gale-Shapley Stable Matching algorithm; we showed earlier that the
algorithm terminates in at most rt2 iterations, and our implementation here
provides a corresponding worst-case running time of O(n2), counting actual
computational steps rather than simply the total number of iterations. To get
such a bound for the Stable Matching algorithm, we will only need to use two
of the simplest data structures: lists and arrays. Thus, our implementation also
provides a good chance to review the use of these basic data structures as well.
In the Stable Matching Problem, each man and each woman has a ranking
of all members of the opposite gender. The very first question we need to
discuss is how such a ranking wil! be represented. Further, the algorithm
maintains a matching and will need to know at each step which men and
women are free, and who is matched with whom. In order to implement the
algorithm, we need to decide which data structures we will use for all these
things.
An important issue to note here is that the choice of data structure is up
to the algorithm designer; for each algorithm we will choose data structures
that make it efficient and easy to implement. In some cases, this may involve
preprocessing the input to convert it from its given input representation into a
data structure that is more appropriate for the problem being solved.
Arrays and Lists
To start our discussion we wi!l focus on a single list, such as the list of women
in order of preference by a single man. Maybe the simplest way to keep a list
of rt elements is to use an array A of length n, and have A[i] be the ith element
of the list. Such an array is simple to implement in essentially all standard
programming languages, and it has the following properties.
We can answer a query of the form "What is the ith element on the list?"
in O(1) time, by a direct access to the value A[i].
If we want to determine whether a particular element e belongs to the
list (i.e., whether it is equal to A[i] for some i), we need to check the

Chapter 2 Basics of Algorithm Analysis

elements one by one in O(n) time, assuming we dont know anything


about the order in which the elements appear in A.
If the array elements are sorted in some clear way (either numerically
or alphabetically), then we can determine whether an element e belongs
to the list in O(log n) time using binary search; we will not need to use
binary search for any part of our stable matching implementation, but
we will have more to say about it in the next section.
An array is less good for dynamically maintaining a list of elements that
changes over time, such as the fist of flee men in the Stable Matching algorithm;
since men go from being flee to engaged, and potentially back again, a list of
flee men needs to grow and shrink during the execution of the algorithm. It
is generally cumbersome to frequently add or delete elements to a list that is
maintained as an array.
An alternate, and often preferable, way to maintain such a dynamic set
of elements is via a linked list. In a linked list, the elements are sequenced
together by having each element point to the next in the list. Thus, for each
element v on the list, we need to maintain a pointer to the next element; we
set this pointer to nail if i is the last element. We also have a pointer First
that points to the first element. By starting at First and repeatedly following
pointers to the next element until we reach null, we can thus traverse the entire
contents of the list in time proportional tO its length.
A generic way to implement such a linked list, when the set of possible
elements may not be fixed in advance, is to allocate a record e for each element
that we want to include in the list. Such a record would contain a field e.val
that contains the value of the element, and a field e.Next that contains a
pointer to the next element in the list. We can create a doubly linked list, which
is traversable in both directions, by also having a field e.Prev that contains
a pointer to the previous element in the list. (e.Prev = null if e is the first
element.) We also include a pointer Last, analogous to First, that points to
the last element in the list. A schematic illustration of part of such a list is
shown in the first line of Figure 2.1.
A doubly linked list can be modified as follows.
o Deletion. To delete the element e from a doubly linked list, we can just
"splice it out" by having the previous element, referenced by e.Prev, and
the next element, referenced by e.Igext, point directly to each other. The
deletion operation is illustrated in Figure 2.1.
o Insertion. To insert element e between elements d and f in a list, we
"splice it in" by updating d.Igext and/.Prey to point to e, and the Next
and Prey pointers of e to point to d and f, respectively. This operation is

2.3 Implementing the Stable Matching Algorithm Using Lisis and Arrays

Before deleting e:

After deleting e:

Element e

Element e

Figure 2.1 A schematic representation of a doubly linked fist, showing the deletion of
an element e.

essentially the reverse of deletion, and indeed one can see this operation
at work by reading Figure 2.1 from bottom to top.

Inserting or deleting e at the beginning of the list involves updating the First
pointer, rather than updating the record of the element before e.
While lists are good for maintaining a dynamically changing set, they also
have disadvantages. Unlike arrays, we cannot find the ith element of the list in
0(1) time: to find the ith element, we have to follow the Next pointers starting
from the beginning of the list, which takes a total of O(i) time.
Given the relative advantages and disadvantages of arrays and lists, it may
happen that we receive the input to a problem in one of the two formats ,and
want to convert it into the other. As discussed earlier, such preprocessing is
often useful; and in this case, it is easy to convert between the array and
list representations in O(n) time. This allows us to freely choose the data
structure that suits the algorithm better and not be constrained by the way
the information is given as input.

Implementing the Stable Matching Algorithm


Next we will use arrays and linked lists to implement the Stable Matching algorithm from Chapter 1. We have already shown that the algorithm terminates in
at most n2 iterations, and this provides a type of upper bound on the running
time. However, if we actually want to implement the G-S algorithm so that it
runs in time proportional to n2, we need to be able to implement each iteration
in constant time. We discuss how to do this now.

For simplicity, assume that the set of men and women are both {1 ..... n}.
To ensure this, we can order the men and women (say, alphabetically), and
associate number i with the ith man mi or ith women wi in this order. This

45

46

Chapter 2 Basics of Algorithm Analysis

assumption (or notation) allows us to define an array indexed by all men


or all women. We need to have a preference list for each man and for each
woman. To do this we will haye two arrays, one for womens preference lists
and one for the mens preference lists; we will use ManPref Ira, i] to denote
the ith woman on man ms preference hst, and similarly WomanPref [w, i] to
be the ith man on the preference list of woman w. Note that the amount of
space needed to give the preferences for all 2rt individuals is O(rt2), as each
person has a list of length n.
We need to consider each step of the algorithm and understand what data
structure allows us to implement it efficiently. Essentially, we need to be able
to do each of four things in constant time.
1. We need to be able to identify a free man.
2. We need, for a man m, to be able to identify the highest-ranked woman
to whom he has not yet proposed.
3. For a woman w, we need to decide if w is currently engaged, and if she
is, we need to identify her current partner.
4. For a woman w and two men m and m, we need to be able to decide,
again in constant time, which of m or m is preferred by w.
First, consider selecting a free man. We will do this b_y maintaining the set
of flee men as a linked list. When we need to select a flee man, we take the
first man m on this list. We delete m from the list if he becomes engaged, and
possibly insert a different man rn, if some other man m becomes free. In this
case, m can be inserted at the front of the list, again in constant time.
Next, consider a man m. We need to identify the highest-ranked woman
to whom he has not yet proposed. To do this we will need to maintain an extra
array Next that indicates for each man m the position of the next woman he
wil! propose to on his list. We initialize Next [m] = 1 for al! men m. If a man m
needs to propose to a woman, hell propose to w = ManPref[m,Next [re]I, and
once he prdposes to w, we increment the value of Next[m] by one, regardless
of whether or not w accepts the proposal.
Now assume man m proposes to woman w; we need to be able to ~denfify
the man m that w is engaged to (if there is such a man). We can do this by
maintaining an array Current of length n, where Current[w] is the woman
ws current partner m. We set Current [w] to a special null symbol when we
need to indicate that woman w is not currently engaged; at the start of the
algorithm, Current[w] is initialized to this null symbol for all women w.
To sum up, the data structures we have set up thus far can implement the
operations (1)-(3) in O(1) time each.

2.4 A Survey of Common Running Times

Maybe the trickiest question is how to maintain womens preferences to


keep step (4) efficient. Consider a step of the algorithm, when man m proposes
to a woman w. Assume w is already engaged, and her current partner is
rn =Current[w]. We would like to decide in O(1) time if woman w prefers rn
or rn. Keeping the womens preferences in an array IqomanPref, analogous to
the one we used for men, does not work, as we would need to walk through
ws list one by one, taking O(n) time to find m and rn on the list. While O(rt)
is still polynomial, we can do a lot better if we build an auxiliary data structure
at the beginning.
At the start of the algorithm, we create an n. x n array Ranking, where
Ranking[w, m] contains the rank of man m in the sorted order of ws preferences. By a single pass through ws preference list, we can create this array in
linear time for each woman, for a total initial time investment proportional to
rt2. Then, to decide which of m or m is preferred by w, we simply compare
the values Ranking[w, rrt] and Ranking[w, rrt].
This allows us to execute step (4) in constant time, and hence we have
everything we need to obtain the desired running time.
(2.10) The data structures described above allow us to implernentthe G-S
algorithm in O(n2) time.

2.4 A Survey of Common Running Times


When trying to analyze a new algorithm, it helps to have a rough sense of
the "landscape" of different running times. Indeed, there are styles of analysis
that recur frequently, and so when one sees running-time bounds like O(n),
O(n log n), and O(n2) appearing over and over, its often for one of a very
small number of distinct reasons. Learning to recognize these common styles
of analysis is a long-term goal. To get things under way, we offer the following
survey of common running-time bounds and some of the typical .approaches
that lead to them.
Earlier we discussed the notion that most problems have a natural "search
space"--the set of all possible solutions--and we noted that a unifying theme
in algorithm design is the search for algorithms whose performance is more
efficient than a brute-force enumeration of this search space. In approaching a
new problem, then, it often helps to think about two kinds of bounds: one on
the running time you hope to achieve, and the other on the size of the problems
natural search space (and hence on the running time of a brute-force algorithm
for the problem). The discussion of running times in this section will begin in
rhany cases with an analysis of the brute-force algorithm, since it is a useful

47

48

Chapter 2 Basics of Algorithm Analysis

way to get ones bearings with respect to a problem; the task of improving on
such algorithms will be our goal in most of the book.

Linear Time
An algorithm that runs in O(n), or linear, time has a very natural property:
its running time is at most a constant factor times the size of the input. One
basic way to get an algorithm with this running time is to process the input
in a single pass, spending a constant amount of time on each item of input
encountered. Other algorithms achieve a linear time bound for more subtle
reasons. To illustrate some of the ideas here, we c6nsider two simple lineartime algorithms as examples.
Computing the Maxirrturrt Computing the maximum of n numbers, for example, can be performed in the basic "one-pass" style. Suppose the numbers
are provided as input in either a list or an array. We process the numbers
an in order, keeping a running estimate of the maximum as we go.
Each time we encounter a number ai, we check whether ai is larger than our
current estimate, and if so we update the estimate to

For i= 2 to n
If ai> max then
set max---- ai
Endif
Endfor

In this way, we do constant work per element, for a total running time of O(n).
Sometimes the constraints of an application force this kind of one-pass
algorithm on you--for example, an algorithm running on a high-speed switch
on the Internet may see a stream of packets flying past it, and it can try
computing anything it wants to as this stream passes by, but it can only perform
a constant amount of computational work on each packet, and it cant save
the stream so as to make subsequent scans through it. Two different subareas
of algorithms, online algorithms and data stream algorithms, have developed
to study this model of computation.
Merging Two Sorted Lists Often, an algorithm has a running time of O(n),
but the reason is more complex. We now describe an algorithm for merging
two sorted lists that stretches the one-pass style of design just a little, but still
has a linear running time.
an and
bn, and each is already arranged in ascending order. Wed like to

2.4 A Survey of Common Running Times

qn that is also arranged in ascending


order. For example, merging the lists 2, 3, 11, 19 and 4, 9, 16, 25 results in the
output 2, 3, 4, 9, 11, 16, 19, 25.
To do this, we could just throw the two lists together, ignore the fact that
theyre separately arranged in ascending order, and run a sorting algorithm.
But this clearly seems wasteful; wed like to make use of the existing order in
the input. One way to think about designing a better algorithm is to imagine
performing the merging of the two lists by hand: suppose youre given two
piles of numbered cards, each arranged in ascending order, and youd like to
produce a single ordered pile containing all the cards. If you look at the top
card on each stack, you know that the smaller of these two should go first on
the output pile; so you could remove this card, place it on the output, and now
iterate on whats left.
In other words, we have the following algorithm.
bn:
Maintain a Cu~ent pointer into each list, initialized to
point to the front elements
While both lists are nonempty:
Let a~ and ~ be the elements pointed to by the Cu~ent pointer
Append the smaller of these two to the output list
Advance the Cu~ent pointer in the list from which the
smaller element was selected
EndWhile
Once one list is empty, append the remainder of the other list
to the output

See Figure 2.2 for a picture of this process.

IaAppend
the smaller of~
and bj to the output.)

A
B
Figure 2.2 To merge sorted lists A and B, we repeatedly extract the smaller item from
the front of the two lists and append it to the ou~ut.

49

50

2.4 A Survey of Common Running Times

Chapter 2 Basics of Algorithm Analysis

Now, to show a linear-time bound, one is tempted to describe an argument


like what worked for the maximum-finding algorithm: "We do constant work
per element, for a total running time of O(n)." But it is actually not true that
we do only constant work per element. Suppose that n is an even number, and
consider the lists A = 1, 3, 5 ..... 2n - 1 and B = n, n + 2, n + 4 ..... 3n - 2.
The number b1 at the front of list B will sit at the front of the list for
iterations while elements from A are repeatedly being selected, and hence
it will be involved in f2 (n) comparisons. Now, it is true that each element
can be involved in at most O(n) comparisons (at worst, it is compared with
each element in the other list), and if we sum this over all elements we get
a running-time bound of O(n2). This is a correct boflnd, but we can show
something much stronger.
The better way to argue is to bound the number of iterations of the While
loop by an "accounting" scheme. Suppose we charge the cost of each iteration
to the element that is selected and added to the output list. An element can
be charged only once, since at the moment it is first charged, it is added
to the output and never seen again by the algorithm. But there are only 2n
elements total, and the cost of each iteration is accounted for by a charge to
some element, so there can be at most 2n iterations. Each iteration involves a
constant amount of work, so the total running time is O(n), as desired.
While this merging algorithm iterated through its input lists in order, the
"interleaved" way in which it processed the lists necessitated a slightly subtle
running-time analysis. In Chapter 3 we will see linear-time algorithms for
graphs that have an even more complex flow of control: they spend a constant
amount of time on each node and edge in the underlying graph, but the order
in which they process the nodes and edges depends on the structure of the
graph.

One also frequently encounters O(n log n) as a running time simply because there are many algorithms whose most expensive step is to sort the
input. For example, suppose we are given a set of n time-stamps xl, x2 ..... xn
on which copies of a file arrived at a server, and wed like to find the largest
interval of time between the first and last of these time-stamps during which
no copy of the file arrived. A simple solution to this problem is to first sort the
time-stamps x~, x2 ..... xn and then process them in sorted order, determining
the sizes of the gaps between each number and its successor in ascending
order. The largest of these gaps is the desired subinterval. Note that this algorithm requires O(rt log n) time to sort the numbers, and then it spends constant
work on each number in ascending order. In other words, the remainder of the
algorithm after sorting follows the basic recipe for linear time that we discussed
earlier.

Quadratic Time

O(rt log n) Time

Heres a basic problem: suppose you are given n points in the plane, each
specified by (x, y) coordinates, and youd like to find the pair of points that
are closest together. The natural brute-force algorithm for this problem would,
enumerate all pairs of points, compute the distance between each pair, and
then choose the pair for which this distance is smallest.
What is the running time of this algorithm? The number of pairs of points
is (~)_ n(n-1)2 , and since this quantity is bounded by n2, it is O(n2). More
crudely, the number of pairs is O(n2) because we multiply the number of
ways of choosing the first member of the pair (at most n) by the number
of ways of choosing the second member of the pair (also at most n). The
distance between points (xi, yi) and (xj, yj) can be computed by the formula
(x~ - x/)2 + (y~ - yj)2 in constant time, so the overall running time is O(rt2).
This example illustrates a very common way in which a rtmning time of O(n2)
arises: performing a search over all pairs of input items and spending constant
time per pair.

O(n log n) is also a very common running time, and in Chapter 5 we will
see one of the main reasons for its prevalence: it is the running time of any
algorithm that splits its input into two equa!-sized pieces, solves each piece
recursively, and then combines the two solutions in linear time.

Quadratic time also arises naturally from a pair of nested loops: An algorithm consists of a !oop with O(n) iterations, and each iteration of the loop
launches an internal loop that takes O(n) time. Multiplying these two factors
of n together gives the running time.

Sorting is perhaps the most well-known example of a problem that can be


solved this way. Specifically, the Mergesort algorithm divides the set of input
numbers into two equal-sized pieces, sorts each half recursively, and then
merges the two sorted halves into a single sorted output list. We have just
seen that the merging can be done in linear time; and Chapter 5 will discuss
how to analyze the recursion so as to get a bound of O(n log n) on the overall
running time.

The brute-force algorithm for finding the closest pair of points can be
written in an equivalent way with two nested loops:

For each input point (xi, yi)


For each other input point (~, ~)
Compute distance d = J(xi - ~)2 +

51

Chapter 2 Basics of Algorithm Analysis


If d is less th~n the current minimum, update minimum to d
End/or
End/or

Note how the "inner" loop, over (xj, yj), has O(n) iterations, each taking
constant time; and the "outer" loop, over (xi, yi), has O(n) iterations, each
invoking the inner loop once.
Its important to notice that the algorithm weve been discussing for the
Closest-Pair Problem really is just the brute-force approach: the natural search
space for this problem has size O(n2), and _were simply enumerating it. At
first, one feels there is a certain inevitability about thi~ quadratic algorithm-we have to measure all the distances, dont we?--but in fact this is an illusion.
In Chapter 5 we describe a very clever algorithm that finds the closest pair of
points in the plane in only O(n log n) time, and in Chapter 13 we show how
randomization can be used to reduce the running time to O(n).

Cubic Time
More elaborate sets of nested loops often lead to algorithms that run in
O(n3) time. Consider, for example, the following problem. We are given sets
n}, and we would like
to know whether some pair of these sets is disjoint--in other words, has no
elements in common.
What is the running time needed to solve this problem? Lets suppose that
each set Si is represented in such a way that the elements of Si can be listed in
constant time per element, and we can also check in constanttime whether a
given number p belongs to Si. The following is a direct way to approach the
problem.
For pair of sets Si snd S]
Determine whether Si ~ud S] have ~u element in common
End/or

This is a concrete algorithm, but to reason about its running time it helps to
open it up (at least conceptually) into three nested loops.
For each set Si
For each other set S]
For each element p of St
Determine whether p also belongs to Sj
End/or
If no element of S~ belongs to Sj then

2.4 A Survey of Common Running Times

Report that S~ and Sj are disjoint


Endif
End/or
End/or
Each of the sets has maximum size O(n), so the innermost loop takes time
O(n). Looping over the sets S] involves O(n) iterations around this innermos~
loop; and looping over the sets Si involves O(n) iterations around this. Multiplying these three factors of n together, we get the running time of O(n3).
For this problem, there are algorithms that improve on O(n3) running
time, but they are quite complicated. Furthermore, it is not clear whether
the improved algorithms for this problem are practical on inputs of reasonable
size.

O(nk) Time
In the same way that we obtained a running time of O(n2) by performing bruteforce search over all pairs formed from a set of n items, we obtain a running
time of O(nk) for any constant k when we search over all subsets of size k.

Consider, for example, the problem of finding independent sets in a graph,


which we discussed in Chapter 1. Recall that a set of nodes is independent
if no two are joined by an edge. Suppose, in particular, that for some fixed
constant k, we would like to know if a given n-node input graph G has an
independent set of size k. The natural brute-force aigorithm for this problem
would enumerate all subsets of k nodes, and for each subset S it would check
whether there is an edge joining any two members of S. That is,
For each subset S of k nodes
Check whether S constitutes an independent set
If S is an independent set then
Stop and declare success
Endif
End/or
If no k-node independent set was fotmd then
Declare failure
Endif

To understand the running time of this algorithm, we need to consider two


quantities. First, the total number of k-element subsets in an n-element set is
nk) n(n- 1)(n - 2)... (n- k+ 1) nk

53

54

Chapter 2 Basics of Algorithm Analysis

Since we are treating k as a constant, this quantity is O(nk). Thus, the outer
loop in the algorithm above will run for O(n~) iterations as it tries all k-node
subsets of the n nodes of the graph.
Inside this loop, we need to test whether a given set S of k nodes constitutes
an independent set. The definition of an independent set tells us that we need
to check, for each pair of nodes, whether there is an edge joining them. Hence
this is a search over pairs, like we saw earlier in the discussion of quadratic
time; it requires looking at (~2), that is, o(k2), pairs and spending constant time
on each.
Thus the total running time is O(k2n~). Since we are treating k as a constant
here, and since constants can be dropped in O(-) notation, we can write this
running time as O(nk).
Independent Set is a principal example of a problem believed to be computationally hard, and in particular it is believed that no algorithm to find k-node
independent sets in arbitrary graphs can avoid having some dependence on k
in the exponent. However, as we will discuss in Chapter 10 in the context of
a related problem, even once weve conceded that brute-force search over kelement subsets is necessary, there can be different ways of going about this
that lead to significant differences in the efficiency of the computation.
Beyond Polynomial Time
The previous example of the Independent Set Problem starts us rapidly down
the path toward running times that grow faster than any polynomial. In
particular, two kinds of bounds that coine up very frequently are 2n and
and we now discuss why this is so.
Suppose, for example, that we are given a graph and want to find an
independent set of maximum size (rather than testing for the existence of one
with a given number of nodes). Again, people dont know of algorithms that
improve significantly on brute-force search, which in this case would look as
fol!ows.
For each subset S of nodes
Check whether S constitutes ~n independent set
If g is a larger independent set than the largest seen so far then
~ecord the size of S as the current maximum
Endif
Endfor

This is very much like the brute-force algorithm for k-node independent sets,
except that now we are iterating over all subsets of the graph. The total number

2.4 A Survey of Common Running Times

of subsets of an n-element set is 2n, and so the outer loop in this algorithm
will run for 2n iterations as it tries all these subsets. Inside the loop, we are
checking all pairs from a set S that can be as large as n nodes, so each iteration
of the !oop takes at most O(n2) time. Multiplying these two together, we get-a
rulming time of O(n22n).
Thus see that 2n arises naturally as a running time for a search algorithm
that must consider all subsets. In the case of Independent Set, something
at least nearly this inefficient appears to be necessary; but its important
to ke~p in mind that 2n is the size of the search space for many problems,
and for many of them we wil! be able to find highly efficient polynomialtime algorithms. For example, a brute-force search algorithm for the Interval
Scheduling Problem that we saw in Chapter 1 would look very similar to the
algorithm above: try all subsets of intervals, and find the largest subset that has
no overlaps. But in the case of the Interval Scheduling Problem, as opposed
to the Independent Set Problem, we will see (in Chapter 4) how to find an
optimal solution in O(n log n) time. This is a recurring kind of dichotomy in
the study of algorithms: two algorithms can have very similar-looking search
spaces, but in one case youre able to bypass the brute-force search algorithm,
and in the other you arent.
The function n! grows even more rapidly than 2n, so its even more
menacing as a bound on the performance of an algorithm. Search spaces of
size n! tend to arise for one of two reasons. First, n! is the number of ways to
match up n items with n other items--for example, it is the number of possible
perfect matchings of n men with n women in an instance of the Stable Matching
Problem. To see this, note that there are n choices for how we can match up
the first man; having eliminated this option, there are n - 1 choices for how we
can match up the second man; having eliminated these two options, there are
n - 2 choices for how we can match up the third man; and so forth. Multiplying
all these choices out, we get n(n - 1)(n - 2) -- (2)(1) = n!
Despite this enormous set of possible solutions, we were able to solve
the Stable Matching Problem in O(n2) iterations of the proposal algorithm.
In Chapter 7, we will see a similar phenomenon for the Bipartite Matching
Problem we discussed earlier; if there are n nodes on each side of the given
bipartite graph, there can be up to n! ways of pairing them up. However, by
a fairly subtle search algorithm, we will be able to find the largest bipartite
matching in O(n3) time.
The function n! also arises in problems where the search space consists
of all ways to arrange n items in order. A basic problem in this genre is the
Traveling Salesman Problem: given a set of n cities, with distances between
all pairs, what is the shortest tour that visits all cities? We assume that the
salesman starts and ends at the first city, so the crux of the problem is the

5,5

56

Chapter 2 Basics of Algorithm Analysis

implicit search over all orders of the remaining n - 1 cities, leading to a search
space of size (n- 1)!. In Chapter 8, we will see that Traveling Salesman
is another problem that, like Independent Set, belongs to the class of NPcomplete problems and is believed to have no efficient solution.

Sublinear Time
Finally, there are cases where one encounters running times that are asymptotically smaller than linear. Since it takes linear time just to read the input,
these situations tend to arise in a model of computation where the input can be
"queried" indirectly rather than read completely, and the goal is to minimize
the amount of querying that must be done.
Perhaps the best-known example of this is the binary search algorithm.
Given a sorted array A of n numbers, wed like to determine whether a given
number p belongs to the array. We could do this by reading the entire array,
but wed like to do it much more efficiently, taking advantage of the fact that
the array is sorted, by carefully probing particular entries. In particular, we
probe the middle entry of A and get its value--say it is q--and we compare q
to p. If q = p, were done. If q > p, then in order for p to belong to the array
A, it must lie in the lower half of A; so we ignore the upper half of A from
now on and recursively apply this search in the lower half. Finally, ff q < p,
then we apply the analogous reasoning and recursively search in the upper
half of A.
The point is that in each step, theres a region of A where p might possibly
be; and were shrinking the size of this region by a factor of two with every
probe. So how large is the "active" region of A after k probes? It starts at size
n, so after k probes it has size at most ()kn.
Given this, how long will it take for the size of the active region-to be
reduced to a constant? We need k to be large enough so that ()k = O(1/n),
and to do this we can choose k = log2 n. Thus, when k = log2 n, the size of
the active region has been reduced to a constant, at which point the recursion
bottoms out and we can search the remainder of the array directly in constant
time.
So the running time of binary search is O(log n), because of this successive
shrinking of the search region. In general, O(log n) arises as a time bohnd
whenever were dealing with an algorithm that does a constant amount of
work in order to throw away a constant fraction of the input. The crucial fact
is that O(log n) such iterations suffice to shrink the input down to constant
size, at which point the problem can generally be solved directly.

2.5 A More Complex Data Structure: Priority Queues

2.5 A More Complex Data Structure:


Priority Queues
Our primary goal in this book was expressed at the outset of the chapter:
we seek algorithms that improve qualitatively on brute-force search, and in
general we use polynomial-time solvability as the concrete formulation of
this. Typically, achieving a polynomial-time solution to a nontrivial problem
is not something that depends on fine-grained implementation details; rather,
the difference between exponential and polynomial is based on overcoming
higher-level obstacles. Once one has an efficient algorithm to solve a problem,
however, it is often possible to achieve further improvements in running time
by being careful with the implementation details, and sometimes by using
more complex data structures.
Some complex data structures are essentially tailored for use in a single
kind of algorithm, while others are more generally applicable. In this section,
we describe one of the most broadly useful sophisticated data structures,
the priority queue. Priority queues will be useful when we describe how to
implement some of the graph algorithms developed later in the book. For our
purposes here, it is a useful illustration of the analysis of a data structure that,
unlike lists and arrays, must perform some nontrivial processing each time it
is invoked.

~ The Problem
In the implementation of the Stable Matching algorithm in Section 2.3, we
discussed the need to maintain a dynamically changing set S (such as the set
of all free men in that case). In such situations, we want to be able to add
elements to and delete elements from the set S, and we want to be able to
select an element from S when the algorithm calls for it. A priority queue is
designed for applications in which elements have a priority value, or key, and
each time we need to select an element from S, we want to take the one with
highest priority.
A priority queue is a data structure that maintains a set of elements S,
where each element v ~ S has an associated value key(v) that denotes the
priority of element v; smaller keys represent higher priorities. Priority queues
support the addition and deletion of elements from the set, and also the
selection of the element with smallest key. Our implementation of priority
queues will also support some additional operations that we summarize at the
end of the section.
A motivating application for priority queues, and one that is useful to keep
in mind when considering their general function, is the problem of managing

57

58

Chapter 2 Basics of Algorithm Analysis

real-time events such as the scheduling of processes on a computer. Each


process has a priority, or urgency, but processes do not arrive in order of
their priorities. Rather, we have a current set of active processes, and we want
to be able to extract the one with the currently highest priority and run it.
We can maintain the set of processes in a priority queue, with the key of a
process representing its priority value. Scheduling the highest-priority process
Corresponds to selecting the element with minimum key from the priority
queue; concurrent with this, we will also be inserting new processes as they
arrive, according to their priority values.
How efficiently do we hope to be able to execute the operations in a priority
queue? We will show how to implement a priority queue containing at most
n elements at any time so that elements can be added and deleted, and the
element with minimum key selected, in O(log n) time per operation.
Before discussing the implementation, let us point out a very basic application of priority queues that highlights why O(log n) time per operation is
essentially the "right" bound to aim for.
(2.11) A sequence of O(n) priority queue operations can be used to sort a set
of n numbers.
Proof. Set up a priority queue H, and insert each number into H with its value
as a key. Then extract the smallest number one by one until all numbers have
been extracted; this way, the numbers will come out of the priority queue in
sorted order.
Thus, with a priority queue that can perform insertion and the extraction
of minima in O(log n) per operation, we can sort n numbers in O(n log n)
time. It is known that, in a comparison-based model of computation (when
each operation accesses the input only by comparing a pair of numbers),
the time needed to sort must be at least proportional to n log n, so. (2.11)
highlights a sense in which O(log n) time per operation is the best we can
hope for. We should note that the situation is a bit more complicated than
this: implementations of priority queues more sophisticated than the one we
present here can improve the running time needed for certain operations, and
add extra functionality. But (2.11) shows that any sequence of priority queue
operations that results in the sorting of n numbers must take time at least
proportional to n log n in total.

A Data Structure for Implementing a Priority Queue


We will use a data structure called a heap to implement a priority queue.
Before we discuss the structure of heaps, we should consider what happens
with some simpler, more natural approaches to implementing the flmctions

2.5 A More Complex Data Structure: Priority Queues

of a priority queue. We could just have the elements in a list, and separately
have a pointer labeled Mn to the one with minimum key. This makes adding
new elements easy, but extraction of the minimum hard. Specifically, finding
the minimum is quick--we just consult the Mn pointer--but after removing
this minimum element, we need to update the ~n pointer to be ready for the
next operation, and this would require a scan of all elements in O(n) time to
find the new minimum.
This complication suggests that we should perhaps maintain the elements
in the sorted order of the keys. This makes it easy to extract the element with
smallest key, but now how do we add a new element to our set? Should we
have the elements in an array, or a linked list? Suppose we want to add s
with key value key(s). If the set S is maintained as a sorted array, we can use
binary search to find the array position where s should be inserted in O(log n)
time, but to insert s in the array, we would have to move all later elements
one position to the right. This would take O(n) time. On the other hand, if we
maintain the set as a sorted doubly linked list, we could insert it in O(1) time
into any position, but the doubly linked list would not support binary search,
and hence we may need up to O(n) time to find the position where s should
be inserted.
The Definition of a Heap So in all these simple approaches, at least one of
the operations can take up to O(n) time--much more than the O(log n) per
operation that were hoping for. This is where heaps come in. The heap data
structure combines the benefits of a sorted array and list for purposes of this
application. Conceptually, we think of a heap as a balanced binary tree as
shown on the left of Figure 2.3. The tree will have a root, and each node can
have up to two children, a left and a right child. The keys in such a binary tree
are said to be in heap order if the key of any element is at least as large as the
key of the element at its parent node in the txee. In other words,
Heap order: For every element v, at a node i, the element w at is parent
satisfies key(w) < key(v).
In Figure 2.3 the numbers in the nodes are the keys of the corresponding
elements.
Before we discuss how to work with a heap, we need to consider what data
structure should be used to represent it. We can use poiriters: each node at the
heap could keep the element it stores, its key, and three pointers pointing to
the two children and the parent of the heap node. We can avoid using pointers,
however, if a bound N is known in advance on the total number of elements
that will ever be in the heap at any one time. Such heaps can be maintained
in an array H indexed by i = 1 ..... N. We will think of the heap nodes as
corresponding to the positions in this array. H[1] is the root, and for any node

59

6O

Chapter 2 Basics of Algorithm Analysis

~a

EaCh nodes key is at least~


s large as its parents.

2.5 A More Complex Data Structure: Priority Queues

The H e a p i fy - u p process is movingI


element v toward the root.

1 2 5 10 3 7 11 15 17 20 9 15 8 16 X

Figure 2.3 Values in a heap shown as a binaD, tree on the left, and represented as an
array on the right. The arrows show the children for the top three nodes in the tree.

Figure 2.4 The Heapify-up process. Key 3 (at position 16) is too small (on the left).
After swapping keys 3 and 11, the heap xdolation moves one step closer to the root of
the tree (on the right).

at position i, the children are the nodes at positions leftChild(i) = 2i and


rightChild(f) = 2i + 1. So the two children of the root are at positions 2 and
3, and the parent of a node at position i is at position parent(f) =/i/2J. If
the heap has n < N elements at some time, we will use the first rt positions
of the array to store the n heap elements, and use lenggh(H) to denote the
number of elements in H. This representation keeps the heap balanced at all
times. See the right-hand side of Figure 2.3 for the array representation of the
heap on the left-hand side.

] = parent(i) to continue fixing the heap by pushing the damaged part upward.
Figure 2.4 shows the first two steps of the process after an insertion.

Implementing the Heap Operations


The heap element with smallest key is at the root, so it takes O(1) time to
identify the minimal element. How do we add or delete heap elements? First
conside~ adding a new heap element v, and assume that our heap H has n < N
elements so far. Now it will have n + 1 elements. To start with, we can add the
new element v to the final position i = n + 1, by setting H[i] = v. Unfortunately,
this does not maintain the heap property, as the key of element v may be
smaller than the key of its parent. So we now have something that is almost-a
heap, except for a small "damaged" part where v was pasted on at the end.
We will use the procedure Heapf y-up to fix our heap. Letj = parent(i) =
L//2] be the parent of the node i, and assume H[j] = w. If key[v] < key[w],
then we will simply swap the positions of v and w. This wil! fix the heap
property at position i, but the resulting structure will possibly fail to satisfy
the heap property at position j--in other words, the site of the "damage" has
moved upward from i to j. We thus call the process recursively from position

Heapify-up (H, i) :
If i> 1 then
let ] = parent(i) = Lil2J
If key[H[i]]<key[H[j]] then
swap the array entries H[i] mad H[j]
Heapify-up (H, j )
Endif
Endif

To see why Heapify-up wOrks, eventually restoring the heap order, it


helps to understand more fully the structure of our slightly damaged heap in
the middle of this process. Assume that H is an array, and v is the element in
position i. We say that H is almost a heap with the key of H[i] too small, if there
is a value ~ _> key(v) such that raising the value of key(v) to c~ would make
the resulting array satisfy the heap property. (In other words, element v in H[i]
is too small, but raising it to cz would fix the problem.) One important point
to note is that if H is almost a heap with the key of the root (i.e., H[1]) too
small, then in fact it is a~heap. To see why this is true, consider that if raising
the value of H[1] to c~ would make H a heap, then the value of H[!] must
also be smaller than both its children, and hence it already has the heap-order
property.

61

62

Chapter 2 Basics of Algorithm Analysis

(2.12) The procedure Heapify-up(!-!, i) fixes the heap property in O(log i)


time, assuming that the array H is almost a heap with the key of H[i] too small.
Using Heapify-up we can insert a need element in a heap of n elements in
O(log n) time.
Proof. We prove the statement by induction on i. If i = ! there is nothing to
prove, since we have already argued that in this case H is actually a heap.
Now consider the case in which i > 1: Let v = H[i], j = parent(i), w = H[j],
and fl = key(w). Swapping elements v and w takes O(1) time. We claim that
after the swap, the array H is either a heap or almost a heap with the key of
H[j] (which now holds v) too small. This is true, as setting the key value at
node j to ~ would make H a heap.
So by the induction hypothesis, applying Heapfy-up(j) recursively will
produce a heap as required. The process follows the tree-path from position i
to the root, so it takes O(log i) time.
To insert a new element in a heap, we first add it as the last element. If the
new element has a very large key value, then the array is a heap. Otherwise,
it is almost a heap with the key value of the new element too small. We use
Heapify-up to fix the heap property. []
Now consider deleting an element. Many applications of priority queues
dont require the deletion of arbitrary elements, but only the extraction of
the minimum. In a heap, this corresponds to identifying the key at the root
(which will be the minimum) and then deleting it; we will refer to this operation as ExtractMin(H). Here we will implement a more general operation
Delete(/./, i), which will delete the element in position i. Assume the heap
currently has n elements. After deleting the element H[i], the heap will have
only n - 1 elements; and not only is the heap-order property violated, there
is actually a "hole" at position i, since H[i] is now empty. So as a first step,
to patch the hole in H, we move the element w in position n to position i.
After doing this, H at least has the property that its n - 1 elements are in the
first n - 1 positions, as required, but we may well still not. have the heap-order
property.
However, the only place in the heap where the order might be violated is
position i, as the key of element w may be either too small or too big for the
position i. If the key is too small (that is, the violation of the heap property is
between node i and its parent), then we can use Heapify-up(i) to reestablish
the heap order. On the other hand, if key[w] is too big, the heap property
may be violated between i and one or both of its children. In this case, we will
use a procedure called Heapify-dovn, closely analogous to Heapify-up, that

2.5 A More Complex Data Structure: Priority Queues

The He~pi fy-down process~


is moving element w down,|
toward the leaves.
)

Figure 2.5 The Heapify-down process:. Key 21 (at position 3) is too big (on the left).
After swapping keys 21 and 7, the heap violation moves one step closer to the bottom
of the tree (on the right).

swaps the element at position i with one of its children and proceeds down
the tree recursively. Figure 2.5 shows the first steps of this process.
Heapify-down (H, i) :
Let n= length(H)
If 2i>n then
Terminate with H unchanged
Else if 2f<n then
Let left=2f, and right=2f+l
Let ] be the index that minimizes key [H [left] ] and key [H [right] ]
Else if 2f=n then
Let ] = 2f
Endif
If key[H[]]] < key[H[i]] then
swap the array entries H[i] and H[]]
Heapify-down (H, ])
Endif

Assume that H is an array and w is the element in position i. We say that


H is almost a heap with the key of Hill too big, if there is a-vaiue o~ _< key(w)
such that lowering the value of key(w) to c~ would make the resulting array
satisfy the heap property. Note that if H[i] corresponds to a leaf in the heap
(i.e., it has no children), and H is almost a heap with H[i] too big, then in fact
H is a heap. Indeed, if lowering the value in H[i] would make H a heap, then

63

64

Solved Exercises

Chapter 2 Basics of Algorithm Analysis

H[i] is already larger than its parent and hence it already has the heap-order
property.
(2.13) The procedure Heapify-down(H, i) fixes the heap property in O(log n)
time, assuming that H is almost a heap with the key value of H[i] too big. Using
Heapfy-up or Heapfy-dovn we can delete a new element in a heap o[ n
elements in O(log n) time.
Proof. We prove that the process fixes the heap by reverse induction on the
value i. Let n be the number of elements in the heap. If 2i > n, then, as we
just argued above, H is a heap and hence there is nothing to prove. Otherwise,
let j be the child of i with smaller key value, and let w = H[j]. Swapping the
array elements w and v takes O(1) time. We claim that the resulting array is
either a heap or almost a heap with H[j] = v too big. This is true as setting
key(v) = key(w) would make H a heap. Now j >_ 2i, so by the induction
hypothesis, the recursive call to Heapfy-cloun fixes the heap property.
The algorithm repeatedly swaps the element originally at position i down,
following a tree-path, so in O(log n) iterations the process results in a heap.
To use the process to remove an element v = H[i] from the heap, we replace
HI ] with the last element in the array, H[n] = w. If the resulting array is not a
heap, it is almost a heap with the key value of H[i] either too small or too big.
We use Heapify-down or Heapify-down to fix the heap property in O(log n)
time. []
Implementing Priority Queues with Heaps
The heap data structure with the Heapfy-do~m and Heapi~y-up operations
can efficiently implement a priority queue that is constrained to hold at most
N elements at any point in time. Here we summarize the operations we will
use.
o 8taxtHeap(N) returns an empty heap H that is set up to store at most N
elements. This operation takes O(N) time, as it involves initializing the
array that will hold the heap.
o Insert(H, v) inserts the item u into heap H. If the heap currently has n
elements, this takes O(log n) time.
~ FndMn(H) identifies the minimum element in the heap H but does not
remove it. This takes O(!) time.
~ Delete(H, i) deletes the element in heap position i. This is implemented
in O(log n) time for heaps that have n elements.
~ ExtractMin(H) identifies and deletes an element with minimum key
value from a heap. This is a combination of the preceding two operations,
and so it takes O(log n) time.

There is a second class of operations in which we want to operate on


elements by name, rather than by their position in the heap. For example, in
a number of graph algorithms that use heaps, the heap elements are nodes of
the graph with key values that are computed during the algorithm. At various
points in these algorithms, we want to operate on a particular node, regardless
of where it happens to be in the heap.
To be able to access given elements of the priority queue efficiently, we
simply maintain an additional array Poston that stores the current position
of each element (each node) in the heap. We can now implement the following
further operations.

To delete the element u, we apply Delete(H,Position[u]). Maintaining


this array does not increase the overall running time, and so we can
delete an element v from a heap with n nodes in O(log n) time.
An additional operation that is used by some algorithms is ChangeKey
(H, v, cO, which changes the key value of element u to key(u) = o~. To
implement this operation in O(log n) time, we first need to be able to
identify the position of element v in the array, which we do by using
the array Position. Once we have identified the position of element v,
we change the key and then apply Heapify-up or Heapify-doma as
appropriate.

Solved Exercises
Solved Exercise 1
Take the following list of functions and arrange them in ascending order of
growth rate. That is, if function g(n) immediately follows function f(n) in
your list, then it should be the case that [(n) is O(g(n)).
fl(n) = 10n

h(n) = n
h(n) = nn
f4(n) = log2 n

f~(n) = 2~4T~ ~
Solution We can deal with functions fl, f2, and f4 very easily, since they
belong to the basic families of exponentials, polynomials, and logarithms.
In particular, by (2.8), we have f4(n)= O(f2(n)); and by (2.9),-we have
f2(n) = O(f~(n)).

65

66

Exercises

Chapter 2 Basics of Algorithm Analysis

Now, the function f3 isnt so hard to deal with. It starts out smaller than
I0n, but once n >_ 10, then clearly I0n < nn. This is exactly what we need for
the definition of O(.) notation: for all n >_ 10, we have I0n _< cnn, where in this
case c = 1, and so I0n = o(nn).
Finally, we come to function fls, which is admittedly kind of strangelooking. A useful rule of thumb in such situations is to try taking logarithms
to see whether this makes things clearer. In this case, log2 fs(n) = ~ n =
(!og2 n)l/2. What do the logarithms of the other functions look like? log f4(n) =
log2 log2 n, while log fa(n) = log2 n. All of these can be viewed as functions
of log2 n, and so using the notation z = log2 n, we can write
1
log fa(n) = -z
3
log f4(n) = log2 z
log fs(n) = z~/2
Now its easier to see whats going on. First, for z > 16, we have log2 z <
z1/2. But the condition z > 16 is the same as n >_ 216 -= 65,536; thus once
n > 216 we have log/4(n) _< log/s(n), and so/4(n) _< Is(n). Thus we can write
f4(n) _= O(fs(n)). Similarly we have z11~< z once z >_ 9--in other words,
once n > 29 = 512. For n above this bound we have log fs(n) < log f2(n) and
hence fs(n)< f2(n), and so we can write Is(n)= O(f2(n)). Essentially, we
have discovered that 2l~/i-~ n is a function whose growth rate lies somewhere
between that of logarithms and polynomials.
Since we have sandwiched fs between f4 and f2, this finishes the task of
putting the functions in order.

Solved Exercise 2
Let f and g be two functions that take nonnegative values, and suppose that
f = O(g). Show that g = fl (f).
Solution This exercise is a way to formalize the intuition that O(.) and fl (-)
are in a sense opposites. It is, in fact, not difficult to prove; it is just a matter
of unwinding the definitions.
Were given that, for some constants c and no, we have f(n) < cg(n) for
all n >_ n0. Dividing both sides by c, we can conclude that g(n) >_ ~f(n) for
all n >_ no. But this is exactly what is required to show that g = fl (f): we have
established that g(n) is at least a constant multiple of f(n) (where the constant
is ~), for all sufficiently large n (at least no).

Exercises
Suppose you have algorithms with the five running times listed below.
(Assume these are the exact running times.) How much slower do each of
these algorithms get when you (a) double the input size, or (b) increase
the input size by one?
(a) n2
n3
lOOn2

nlog n
2n
Suppose you have algorithms with the sLx running times listed below.
(Assume these are the exact number of operations performed as a function of the input size n.) Suppose you have a computer that can perform
10t operations per second, and you need to compute a result in at most
an hour of computation. For each of the algorithms, what is the largest
input size n for which you would be able to get the result within an hour?
(a) rt~
(b) n3

(c) lOOn~

(d) n log n
(e) 2n
(f) 22"
Take the foilowing list of functions and arrange them in ascending order
of growth rate. That is, if function g(n) immediately follows function f(n)
in your list, then it should be the case that f(n) is O(g(n)).
v/ fl(n) = n
~/ f3(n) = n + 10

~/f4(n) = lon

~/fstn) = 10on
fc,(n) = n2 log n

Take the following list of functions and arrange them in ascending order
of growth rate. That is, if function g(n) immediately follows function f(n)
in your list, then it should be the case that f(n) is O(g(n)).

67

68

Chapter 2 Basics of Algorithm Analysis

~ gl(a) = 2~4i~
" g2(n) = 2n
i g4(n) ---- n4/3

g3(n) = n(log n)3


gs(n) = nlogn
g6(n) = 22n
i gT(n) = 2n2

Assume you have functions f and g such that f(n) is O(g(n)). For each of
the following statements, decide whether you think it is true or false and
give a proof or counterexample.
(a) log2 f(n)is O(log2g(n))(b) 2f(n) is O(2g(~)).
(C) /(n)2 iS O(g(n)2).
Consider the following basic problem. Youre given an array A consisting
A[n]. Youd like to output a two-dimensional
n-by-n array B in which B[i,j] (for i <j) contains the sum of array entries
A[i] through A~]--that is, the sum A[i] +A[i + 1] +-" + A[j]. (The value of
array entry B[i,j] is left unspecified whenever i >_j, so it doesnt matter
what is output for these values.)
Heres a simple algorithm to solve this problem.
For i=1,

Exercises

the relevant entries of the array B, filling in a value for each--it


contains some highly urmecessary sources of inefficiency. Give a
different algorithm to solve this problem, with an asymptotically
better nmning time. In other words, you should design an algorithm
with running time O(g(n)), where lim~_.oo g(n)/f(n) = O.
Theres a class of folk songs and holiday songs in which each verse
consists of the previous verse, with one extra line added on. "The Twelve
Days of Christmas" has this property; for e.xample, when you get to the
fifth verse, you sing about the five golden rings and then, reprising the
lines from the fourth verse, also cover the four calling birds, the three
French hens, the two turtle doves, and of course the.partridge in thepear
tree. The Aramaic song "Had gadya" from the PassoVer Haggadah works
like this as well, as do many other songs.
These songs tend to last a long time, despite having relatively short
scripts. In particular, you can convey the words plus instructions for one
of these songs by specifying just the new line that is added In each verse,
without ha~4ng to write out all the previous lines each time. (So the phrase
"five golden rings" ouly has to be written once, even though it will appear
in verses five and Onward.)
Theres someth~g asy~nptotic that can be analyzed here. Suppose,
for concreteness, that ~ach line has a length that i~ bounded by a constant
c, and suppose that the song, when sung out loud, runs for n words total.
Show how to encode such a song using a script that has length f(n), for
a function f(n) that grows as slowly as possible.

2,...,n

n
Add up array entries A[i] through A[j]
Store the result in B[i,]]
End/or
End/or
(a) For some function f that you should choose, give a bound of the
form O(f(n)) on the running time of this algorithm on an input of
size n (i.e., a bound on the number of operations performed by the
algorithm).
(b) For this same function f, show that thertmning time of the algorithm
on an input of size n is also ~2 (f(n)). (This shows an asymptotically
tight bound of (f(n)) on the running time.)
(c) Although the algorithm you analyzed in parts (a) and (b) is the most
natural way to solve the problem--after all, it just iterates through

Youre doing some stress-testing on various models of glass jars to


determine the height from which they can be dropped and still not break.
The setup for this experiment, on a particular type of jar, is as follows.
You have a ladder with n rungs, and you want to find the highest rung
from which you can drop a copy of the jar and not have it break..We ca~,
this the highest safe rung.
It might be natural to try binary search: drop a jar from the middle
rung, see if it breaks, and then recursively try from rung n/4 or 3n/4
depending on the outcome. But this has the drawback that y9u could
break a lot of jars in finding the answer.
If your primary goal were to conserve jars, on the other hand, you
could try the following strategy. Start by dropping a jar from the first
rung, then the second rung, and so forth, climbing one higher each time
until the jar breaks. In this way, you only need a single j ar--at the moment

69

70

Chapter 2 Basics of Algorithm Analysis

it breaks, you have the correct answer--but you may have to drop it rt
times (rather than log rt as in the binary search solution).
So here is the trade-off: it seems you can perform fewer drops if
youre willing to break more jars. To understand better how this tradeoff works at a quantitative level, lets consider how to run this experiment
given a fixed "budget" of k >_ 1 jars. In other words, you have to determine
the correct answer--the highest safe rung--and can use at most k jars In
doing so.
(a) Suppose you are given a budget of k = 2 jars. Describe a strategy for
finding the highest safe rung that requires you to drop a jar at most
f(n) times, for some function f(n) that grows slower than linearly. (In
other words, it should be the case that limn-.~ f(n)/n = 0.)
(b) Now suppose you have a budget of k > 2 jars, for some given k.
Describe a strategy for fInding the highest safe rung using at most
k jars. If fk(n) denotes the number of times you need to drop a jar
should have.
the property that each grows asymptotically slower than the previous
one: lirnn_~ fk(n)/fk_l(n) = 0 for each k.

Notes and Further Reading


Polynomial-time solvability emerged as a formal notion of efficiency by a
gradual process, motivated by the work of a number of researchers including Cobham, Rahin, Edmonds, Hartmanis, and Stearns. The survey by Sipser
(1992) provides both a historical and technical perspective on these developments. Similarly, the use of asymptotic order of growth notation to bound the
running time of algorithms--as opposed to working out exact formulas with
leading coefficients and lower-order terms--is a modeling decision that was
quite non-obvious at the time it was introduced; Tarjans Turing Award lecture
(1987) offers an interesting perspective on the early thinking of researchers
including Hopcroft, Tarian, and others on this issue. Further discussion of
asymptotic notation and the growth of basic functions can be found in Knuth
(1997a).
The implementation of priority queues using heaps, and the application to
sorting, is generally credited to Williams (1964) and Floyd (1964). The priority
queue is an example of a nontrivial data structure with many applications; in
later chapters we will discuss other data structures as they become useful for
the implementation of particular algorithms. We will consider the Union-Find
data structure in Chapter 4 for implementing an algorithm to find minimum-

Notes and Further Reading

cost spanning trees, and we will discuss randomized hashing in Chapter 13.
A number of other data structures are discussed in the book by Tarjan (1983).
The LEDA library (Library of Efficient Datatypes and Algorithms) of Mehlhorn
and Ngher (1999) offers an extensive library of data structures useful in
combinatorial and geometric applications.
Notes on the Exercises Exercise 8 is based on a problem we learned from
Sam Toueg.

71

3
Graphs

Our focus in this book is on problems with a discrete flavor. Just as continuous
mathematics is concerned with certain basic structures such as real numbers,
vectors, and matrices, discrete mathematics has developed basic combinatorial
structures that lie at the heart of the subiect. One of the most fundamental and
expressive of these is the graph.
The more one works with graphs, the more one tends to see them everywhere. Thus, we begin by introducing the basic definitions surrounding
graphs, and list a spectrum of different algorithmic settings where graphs arise
naturally. We then discuss some basic algorithmic primitives for graphs, beginning with the problem of connectivity and developing some fundamental
graph search techniques.

5.1 Basic Definitions and Applications


Reca!l from Chapter 1 that a graph G is simply a way of encoding pairwise
relationships among a set of objects: it consists of a collection V of nodes
and a collection E of edges, each of which "joins" two of the nodes. We thus
represent an edge e E E as a two-element subset of V: e = {u, v} for some
u, v E V, where we cal! u and v the ends of e.
Edges in a graph indicate a symmetric relationship between their ends.
Often we want to encode asymmetric relationships, and for this we use the
c!osely related notion of a directed graph. A directed graph G consists of a set
of nodes V and a set of directed edges E. Each e ~ E is an ordered pair (a, v);
in other words, the roles of u and v are not interchangeable, and we call u the
tail of the edge and v the head. We will also say that edge e leaves node u and
enters node v.

74

Chapter 3 Graphs
When we want to emphasize that the graph we are considering is not
directed, we will cal! it an undirected graph; by default, however, the term
"graph" will mean an undirected graph. It is also worth mentioning two
warnings in our use of graph terminology. First, although an edge e in an
undirected graph should properly be written as a set of nodes {u, u}, one will
more often see it written (even in this book) in the notation used for ordered
pairs: e = (u, v). Second, a node in a graph is also frequently called a vertex;
in this context, the two words have exactly the same meaning.
Examples of Graphs Graphs are very simple to define: we just take a collection of things and join some of them by edges. But at this level of abstraction,
its hard to appreciate the typical kinds of situations in which they arise. Thus,
we propose the following list of specific contexts in which graphs serve as
important models. The list covers a lot of ground, and its not important to
remember everything on it; rather, it will provide us with a lot of usefifl examples against which to check the basic definitions and algorithmic problems
that well be encountering later in the chapter. Also, in going through the list,
its usefi~ to digest the meaning of the nodes and the meaning of the edges in.
the context of the application. In some cases the nodes and edges both correspond to physical objects in the real world, in others the nodes are real objects
while the edges are virtual, and in still others both nodes and edges are pure
abstractions.
1. Transportation networks. The map of routes served by an airline carrier
naturally forms a graph: the nodes are airports, and there is an edge from
u to t~ if there is a nonstop flight that departs from u and arrives at v.
Described this way, the graph is directed; but in practice when there is an
edge (u, u), there is almost always an edge (u, u), so we would not lose
much by .treating the airline route map as an undirected graph with edges
joining pairs of airports that have nonstop flights each way. Looking at
such a graph (you can generally find them depicted in the backs of inflight airline magazines), wed quickly notice a few things: there are often
a small number of hubs with a very large number of incident edges; and
its possible to get between any two nodes in the graph via a very small
number of intermediate stops.
Other transportation networks can be modeled in a similar way. For
example, we could take a rail network and have a node for each terminal,
and an edge joining u and v if theres a section of railway track that
goes between them without stopping at any intermediate terminal. The
standard depiction of the subway map in a major city is a drawing of
such a graph.
Communication networks. A collection of computers connected via a
2.
communication network can be naturally modeled as a graph in a few

3.1 Basic Definitions and Applications

different ways. First, we could have a node for each computer and
an edge joining u and u if there is a direct physical link connecting
them. Alternatively, for studying the large-scale structure of the Internet,
people often define a node to be the set of all machines controlled by
a single Internet service provider, with an edge joining u and v if there
is a direct peering relationship between them--roughly, an agreement
to exchange data under the standard BCP protocol that governs global
Internet routing. Note that this latter network is more "virtual" than
the former, since the links indicate a formal agreement in addition to
a physical connection.
In studying wireless networks, one typically defines a graph where
the nodes are computing devices situated at locations in physical space,
and there is an edge from u to u if u is close enough to u to receive a signal
from it. Note that its often useful to view such a graph as directed, since
it may be the case that u can hear us signal but u cannot hear us signal
(if, for example, u has a stronger transmitter). These graphs are also
interesting from a geometric perspective, since they roughly correspond
to putting down points in the plane and then joining pairs that are close
together.
Inyormation networks. The World Wide Web can be naturally viewed as a
directed graph, in which nodes correspond to Web pages and there is an
edge from u to v if u has a hyperlink to v. The directedness of the graph
is crucial here; many pages, for example, link to popular news sites,
but these sites clearly do not reciprocate all these links. The structure of
all these hyperlinks can be used by algorithms to try inferring the most
important pages on the Web, a technique employed by most current
search engines.

The hypertextual structure of the Web is anticipated by a number of


information networks that predate the Internet by many decades. These
include the network of cross-references among articles in an encyclopedia
or other reference work, and the network of bibliographic citations
among scientific papers.
Social networks. Given any collection of people who interact (the employees of a company, the students in a high school, or the residents of
a small town), we can define a network whose nodes are people, with
an edge joining u and v if they are friends with one another. We could
have the edges mean a number of different things instead of friendship:
the undirected edge (u, v) could mean that u and v have had a romantic relationship or a financial relationship; the directed edge (u, v) could
mean that u seeks advice from v, or that u lists v in his or her e-mail
address book. One can also imagine bipartite social networks based on a

76

Chapter 3 Graphs

notion of affiliation: given a set X of people and a set Y of organizations,


we could define an edge between u a X and v ~ Y if person u belongs to
organization v.
Networks such as this are used extensively by sociologists to study
the dynamics of interaction among people. They can be used to identify
the most "influential" people in a company or organization, to model
trust relationships in a financial or political setting, and to track the
spread of fads, rumors, jokes, diseases, and e-mail viruses.
Dependency nenvorks. It is natural to define directed graphs that capture
the interdependencies among a collection of objects. For example, given
the list of courses offered by a college or university, we could have a
node for each course and an edge from u to v if u is a prerequisite for v.
Given a list of functions or modules in a large software system, we could
have a node for each function and an edge from u to v if u invokes v by a
function cal!. Or given a set of species in an ecosystem, we could define
a graph--a food web--in which the nodes are the different species and
there is an edge from u to v if u consumes v.
This is far from a complete list, too far to even begin tabulating its
omissions. It is meant simply to suggest some examples that are useful to
keep in mind when we start thinking about graphs in an algorithmic context.
Paths and Connectiuity One of the fundamental operations in a graph is
that of traversing a sequence of nodes connected by edges. In the examples
iust listed, such a traversal could correspond to a user browsing Web pages by
following hyperlinks; a rumor passing by word of mouth from you to someone
halfway around the world; or an airline passenger traveling from San Francisco
to Rome on a sequence of flights.
With this notion in mind,-we define a path in an undirected graph
G = (V, E) to be a sequence P of nodes v1, v2 ..... v~_l, v~ with the property
that each consecutive pair v~, vg+~ is ioined by an edge in G. P is often called
a path from v~ to ug, or a v~-vg path. For example, the nodes 4, 2, 1, 7, 8 form
a path in Figure 3.1. A path is called simple if all its vertices are distinct from
one another. A cycle is a path v~, v2 ..... v~_l, v~ in which k > 2, the first k - 1
nodes are all distinct, and vl = v~--in other words, the sequence of nodes
"cycles back" to where it began. All of these definitions carry over naturally
to directed graphs, with the fol!owing change: in a directed path or cycle,
each pair of consecutive nodes has the property that (vi, vi+l) is an edge. In
other words, the sequence of nodes in the path or cycle must respect the
directionality of edges.
We say that an undirected graph is connected if, for every pair of nodes u
and v, there is a path from u to v. Choosing how to define connectivity of a

3.1 Basic Definitions and Applications

Figure 3.1 Two drawings of the same tree. On the right, the tree is rooted at node 1.

directed graph is a bit more subtle, since its possible for u to have a path to
~ while u has no path to u. We say that a directed graph is strongly connected
if, for every two nodes u and u, there is a path from u to v and a path from v
to u.
In addition to simply knowing about the existence of a path between some
pair of nodes a and u, we may also want to know whether there is a short path.
Thus we define the distance between two nodes a and u to be the minimum
number of edges in a u-u path. (We can designate some symbol like oo to
denote the distance between nodes that are not connected by a path.) The
term distance here comes from imagining G as representing a communication
or transportation network; if we want to get from a to u, we may well want a
route with as few "hops" as possible.

Trees We say that an undirected graph is a tree if it is connected and does not
contain a cycle. For example, the two graphs pictured in Figure 3.! are trees.
In a strong sense, trees are the simplest kind of connected graph: deleting any
edge from a tree will disconnect it.
For thinking about the structure of a tree T, it is useful to root it at a
particular node r. Physically, this is the operation of grabbing T at the node r
and letting the rest of it hang downward under the force of gravity, like a
mobile. More precisely, we "orient" each edge of T away ffomr; for each other
node v, we declare the parent of v to be the node u that directly precedes v
on its path from r; we declare w to be a child of v if v is the parent of w. More
generally, we say that w is a descendant of v (or v is an ancestor of w) if v lies
on the path from the root to w; and we say that a node x is a leaf if it has no
descendants. Thus, for example, the two pictures in Figure 3.1 correspond to
the same tree T--the same pairs of nodes are joined by edges--but the drawing
on the right represents the result of rooting T at node 1.

77

78

3.2 Graph Connectivity and Graph Traversal

Chapter 3 Graphs

Rooted trees are fundamental objects in computer science, because they


encode the notion of a hierarchy. For example, we can imagine the rooted tree
in Figure 3.1 as corresponding to the organizational structure of a tiny nineperson company; employees 3 and 4 report to employee 2; employees 2, 5,
and 7 report to employee 1; and so on. Many Web sites are organized according
to a tree-like structure, to facilitate navigation. A Wpical computer science
departments Web site will have an entry page as the root; the People page is
a child of this entry page (as is the Courses page); pages entitled Faculty and
Students are children of the People page; individual professors home pages
are children of the Faculty page; and so on.
For our purposes here, roofing a tree T can make certain questions about T
conceptually easy to answer. For example, given a tree T on n nodes, how many
edges does it have? Each node other than the root has a single edge leading
"upward" to its parent; and conversely, each edge leads upward from precisely
one non-root node. Thus we have very easily proved the following fact.
(3.1)

Every n-node tree has exactly n - I edges.

In fact, the following stronger statement is true, although we do not prove


it here.
(3.2) Let G be an undirected graph on n nodes. Any tmo of the following
statements implies the third.

(0 G is connected.
(ii) G does not contain a c31cle.
(iiO G has n - 1 edges.
We now turn to the role of trees in the fundamental algorithmic idea of
graph trauersal.

3.2 Graph Connectivity and Graph Traversal


Having built up some fundamental notions regarding graphs, we turn to a very
basic algorithmic question: n0de-to-node connectivity. Suppose we are given a
graph G = (V, E) and two particular nodes s and t. Wed like to find an efficient
algorithm that answers the question: Is there a path from s to t in G.~ We wi~
call this the problem of determining s-t connectivity.
For very small graphs, this question can often be answered easily by visual
inspection. But for large graphs, it can take some work to search for a path.
Indeed, the s-t Coimectivity Problem could also be called the Maze-Solving
Problem. If we imagine G as a maze with a room corresponding to each node,
and a hallway corresponding to each edge that joins nodes (rooms) together,

Figure 3.2 In this graph, node 1 has paths to nodes 2 through 8, but not to nodes 9
~ough 13.

then the problem is to start in a room s and find your way to another designated
room t. How efficient an algorithm can we design for this task?
In this section, we describe two natural algorithms for this problem at a
high level: breadth-first search (BFS) and depth-first search (DFS). In the next
section we discuss how to implement each of these efficiently, building on a
data structure for representing a graph as the input to an algorithm.

Breadth-First Search
Perhaps the simplest algorithm for determining s-t connectivity is breadth-first
search (BFS), in which we explore outward from s in all possible directions,
adding nodes one "layer" at a time. Thus we start with s and include all nodes
that are joined by an edge to s--this is the first layer of the search. We then
include all additional nodes that are joined by an edge to any node in the first
layer--this is the second layer. We continue in this way until no new nodes
are encountered.

In the example of Figure 3.2, starting with node 1 as s, the first layer of
the search would consist of nodes 2 and 3, the second layer would consist of
nodes 4, 5, 7, and 8, and the third layer would consist just of node 6. At this
point the search would stop, since there are no further nodes that could be
added (and in particular, note that nodes 9 through 13 are never reached by
the search).
As this example reinforces; there is a natural physical interpretation to the
algorithm. Essentially, we start at s and "flood" the graph with an expanding
wave that grows to visit all nodes that it can reach. The layer containing a
node represents the point in time at which the node is reached.
We can define the layers L1, L2, L3 .... constructed by the BFS algorithm
more precisely as follows.

79

80

3.2 Graph Connectivity and Graph Traversal

Chapter 3 Graphs

Layer L1 consists of all nodes that are neighbors of s. (For notational


reasons, we will sometimes use layer L0 to denote the set consisting just
of s.)
Assuming that we have defined layers L1 ..... Lj, then layer Lj+I consists
of all nodes that do not belong to an earlier layer and that have an edge
to a node in layer Li.
Recalling our definition of the distance between two nodes as the minimum
number of edges on a path joining them, we see that layer L1 is the set of all
nodes at distance 1 from s, and more generally layer Lj is the set of al! nodes
at distance exactly j from s. A node falls to appear in any of the layers if and
only if there is no path to it. Thus, BFS is not only determining the nodes that s
can reach, it is also computing shortest paths to them. We sum this up in the
fo!lowing fact.
{3.3) For each j >_ !, layer LI produced by BFS consists of all nodes at distaffce
exactly j from s. There is a path from s to t if and only if t appears in some,
layer.
A further property of breadth-first search is that it produces, in a very
natural way, a tree T rooted at s on the set of nodes reachable from s.
Specifically, for each such node v (other than s), consider the moment when
v is first "discovered" by the BFS algorithm; this happens when some node
in layer Lj is being examined, and we find that it has an edge to the previously
unseen node v. At this moment, we add the edge (u, v) to the tree
becomes the parent of v, representing the fact that u is "responsible" for
completing the path to v. We call the tree T that is produced in this way a
breadth-first search tree.
Figure 3.3 depicts the construction of a BFS tree rooted at node 1 for the
graph in Figure 3.2. The solid edges are the edges of T; the dotted edges are
edges of G that do not belong to T. The execution of BFS that produces this
tree can be described as follows.
(a) Starting from node 1, layer L1 consists of the nodes {2, 3}.

(6) Layer Li is then grown by considering the nodes in layer L1 in order (say,

first 2, then 3). Thus we discover nodes 4 and 5 as soon as we look at 2,


so 2 becomes their parent. When we consider node 2, we also discover
an edge to 3, but this isnt added to the BFS tree, since we already know
about node 3.
We first discover nodes 7 and 8 when we look at node 3. On the other
hand, the edge from 3 to 5 is another edge of G that does not end up in

(a)

Co)

Figure 3.3 The construction of a breadth-first search tree T for the gTaph in Figure 3.2,
with (a), (b), and (c) depicting the successive layers that are added. The solid edges are
the edges of T; the dotted edges are in the connected component of G containing node
!, but do not belong to T.

the BFS tree, because by the time we look at this edge out of node 3, we
already know about node 5.
(c) We then consider the nodes in layer L2 in order, but the only new node
discovered when we look through L2 is node 6, which is added to layer
L3. Note that the edges (4, 5) and (7, 8) dont get added to the BFS tree,
because they dont result in the discovery of new nodes.
(d) No new nodes are discovered when node 6 is examined, so nothing is put
in layer L4, and the algorithm terminates. The full BFS tree is depicted
in Figure 3.3 (c).

We notice that as we ran BFS on this graph, the nontree edges all either
connected nodes in the same layer, or connected nodes in adjacent layers. We
now prove that this is a properW of BFS trees in general.
(3.4) Let T be a breadth-first search tree, let x and y be nodes in T belonging
to layers Li and Lj respectively, and let (x, y) be an edge of G. Then i and j differ
by at most 1.
Proof. Suppose by way of contradiction that i and j differed by more than 1;
in particular, suppose i < j - 1. Now consider the point in the BFS algorithm
when the edges incident to x were being examined. Since x belongs to layer
Li, the only nodes discovered from x belong to layers Li+1 and earlier; hence,
if y is a neighbor of x, then it should have been discovered by this point at the
latest and hence should belong to layer Li+1 or earlier. []

81

82

3.2 Graph Connectivity and Graph Traversal

Chapter 3 Graphs

Current component
containing s

Proof. We have already argued that for any node u ~ R, there is a path from s
to v.
Now, consider a node tu ~ R, and suppose bY way of contradiction, that
there is an s-tu path P in G. Since s ~ R but tu R, there must be a first node v
on P that does not belong to R; and thisnode ~is not equal to s. Thus there is
a node u immediately preceding u on P, so (u, u) is an edge. Moreover, since v
is the first node on P that does not belong to R, we must have u ~ R. It follows
that (u, v) is an edge where u ~ R and u ~g R; this contradicts the stopping rule
for the algorithm. []

Figure 3.4 When growing the connected component containing s, we look for nodes
like v that have not yet been x4sited.

Exploring a Connected Component


The set of nodes discovered by the BFS algorithm is precisely those reachable
from the starting node s. We will refer to this set R as the connected component
of G containing s; and once we know the connected component containing s,
we can simply check whether t belongs to it so as to answer the question of
s-t connectivity.
Now, if one thinks about it, its clear that BFS is iust one possible way to
produce this component. At a more general level, we can build the component
R by "exploring" G in any order, starting from s. To start off, we define R = {s}.
Then at any point in time, if we find an edge (u, v) where u ~ R and v ~ R, we
can add u to R. Indeed, if there is a path P from s to u, then there is a path
from s to v obtained by first following P and then following the edge (u, v).
Figure 3.4 illustrates this basic step in growing the component R.
Suppose we continue growing the set R until there are no more edges
leading out of R; in other words, we run the following algorithm.
R will consist of nodes to which s has a path
Initially R = {s}
While there is ~u edge (u,u) where uER and
Add u to R
Endwhile

Here is the key property of this algorithm.


(3 !5)SetR prod~ded at the end of the aIgori&m is ~re~isely the ~b;~ctea
cOmpone~ Of G

For any node t in the component R, observe that it is easy to recover the
actual path from s to t along the lines of the argument above: we simply record,
for each node u, the edge (u, u) that was considered in the iteration in which
u was added to R. Then, by tracing these edges backward from t, we proceed
through a sequence of nodes that were added in earlier and earlier iterations,
eventually reaching s; this defines an s-t path.
To conclude, we notice that the general algorithm we have defined to
grow R is underspecified, so how do we decide which edge to consider next?
The BFS algorithm arises, in particular, as a particular way of ordering the
nodes we visit--in successive layers, based on their distance from s. But
there are other natural ways to grow the component, several of which lead
to efficient algorithms for the connectivity problem while producing search
patterns with different structures. We now go on to discuss a different one of
these algorithms, depth-first search, and develop some of its basic properties.

Depth-First Search
Another natural method to find the nodes reachable from s is the approach you
might take if the graph G were truly a maze of interconnected rooms and you
were walking around in it. Youd start from s and try the first edge leading out
of it, to a node u. Youd then follow the first edge leading out of u, and continue
in this way until you reached a "dead end"--a node for which you had already
explored all its neighbors. Youd then backtrack until you got to a node with
an unexplored neighbor, and resume from there. We call this algorithm depthfirst search (DFS), since it explores G by going as deeply as possible and only
retreating when necessary.
DFS is also a particular implementation of the generic component-growing
algorithm that we introduced earlier. It is most easily described in recursive
form: we can invoke DFS from any starting point but maintain global knowledge of which nodes have already been explored.

83

84

3.2 Graph Connectivity and Graph Traversal

Chapter 3 Graphs

DFS(u) :
Mark u as "Explored" and add u to R
For each edge (u,u) incident to u
If v is not marked "Explored" then
Recursively invoke DFS(u)
Endif
Endfor

To apply this to s-t connectivity, we simply declare all nodes initially to be not
explored, and invoke DFS(s).
There are some fundamental similarities and some fundamental differences between DFS and BFS. The similarities are based on the fact that they
both build the connected component containing s, and we will see in the next
section that they achieve qualitatively similar levels of efficiency.
While DFS ultimately visits exactly the same set of nodes as BFS, it typically "
does so in a very different order; it probes its way down long paths, potentially
getting very far from s, before backing up to try nearer unexplored nodes. We
can see a reflection of this difference in the fact that, like BFS, the DFS algorithm
yields a natural rooted tree T on the component containing s, but the tree will
generally have a very different structure. We make s the root of the tree T,
and make u the parent of v when u is responsible for the discovery of v. That
is, whenever DFS(v) is invoked directly during the ca!l to DFS(u), we add the
edge (u, v) to T. The resulting tree is called a depth-first search tree of the
component R.
Figure 3.5 depicts the construction of a DFS tree rooted at node 1 for the
graph in Figure 3.2. The solid edges are the edges of T; the dotted edges are
edges of G that do not belong to T. The execution of DFS begins by building a
path on nodes 1, 2, 3, 5, 4. The execution reaches a dead.end at 4, since there
are no new nodes to find, and so it "backs up" to 5, finds node 6, backs up
again to 3, and finds nodes 7 and 8. At this point there are no new nodes to find
in the connected component, so all the pending recursive DFS calls terminate,
. one by one, and the execution comes to an end. The full DFS tree is depicted
in Figure 3.5(g).
This example suggests the characteristic way in which DFS trees look
different from BFS trees. Rather than having root-to-leaf paths that are as short
as possible, they tend to be quite narrow and deep. However, as in the case
of BFS, we can say something quite strong about the way in which nontree
edges of G must be arranged relative to the edges of a DFS tree T: as in the
figure, nontree edges can only connect ancestors of T to descendants.

Ca)

(d)

(g)
Figure 3.5 The construction of a depth-first search tree T for the graph in Figure 3.2,
with (a) through (g) depicting the nodes as they are discovered in sequence. The solid
edges are the edges of T; the dotted edges are edges of G that do not belong to T.

To establish this, we first observe the following property of the DFS


algorithm and the tree that it produces.
(3.6) For a given recursive call DFS(u), all nodes that are marked "Explored"
between the invocation and end of this recursive call are descendants of u
in T.
Using (3.6), we prove
(3,7) Let T be a depth-first search tree, let x and y be nodes in T, and let
(x, y) be an edge of G that is not an edge of T. Then one of x ory is an ancestor
of the other.

85

86

Chapter 3 Graphs

Proof. Suppose that (x, y) is an edge of G that is not an edge of T, and suppose
without loss of generality that x is reached first by the DFS algorithm. When
the edge (x, y) is examined during the execution of DFS(x), it is not added
to T because y is marked "Explored." Since y was not marked "Explored"
when DFS(x) was first invoked, it is a node that was discovered between the
invocation and end of the recursive call DFS(x). It follows from (3.6) that y is
a descendant of x. ,,

The Set of All Connected Components


So far we have been talking about the connected component containing a
particular node s. But there is a connected component associated with each
node in the graph. What is the relationship between these components?
In fact, this relationship is highly structured and is expressed in the
following claim.
(3.8) For any two nodes s and t in a graph, their connected components are
either identical or disjoint.
This is a statement that is very clear intuitively, if one looks at a graph like
the example in Figure 3.2. The graph is divided into multiple pieces with no
edges between them; the largest piece is the connected component of nodes
1 through 8, the medium piece is the connected component of nodes 11, 12,
and 13, and the sma!lest piece is the connected component of nodes 9 and 10.
To prove the statement in general, we )ust need to show how to define these
"pieces" precisely for an arbitrary graph.
Proof. Consider any two nodes s and t in a graph G with the property that
there is a path between s and t. We claim that the connected components
containing s and t are the same set. Indeed, for any node v in the component
of s, the node v must also be reachable from t by a path: we can )fist walk
from t to s, and then on from s to v. The same reasoning works with the roles
of s and t reversed, and so a node is in the component of one if and only if it
is in the component of the other.
On the other hand, if there is no path between s and t, then there cannot
be a node v that is in the connected component of each. For if there were such
a node v, then we could walk from s to v and then on to t, constructing a
path between s and t. Thus, if there is no path between s and t, then their
connected components are dis)tint. ,,
This proof suggests a natural algorithm for producing all the connected
components of a graph, by growing them one component at a time. We start
with an arbi~ary node s, and we use BFS (or DFS) to generate its connected

3.3 Implementing Graph Traversal Using Queues and Stacks

component. We then find a node v (if any) that was not visited by the search
from s, and iterate, using BFg starting from v, to generate its Connected
component--which, by (3.8), will be disioint from the component of s. We
continue in this way until all nodes have been visited.

3.3 Implementing Graph Traversal Using Queues


and Stacks
So far we have been discussing basic algorithmic primitives for working with
graphs without mentioning any implementation details. Here we discuss how
to use lists and arrays to represent graphs, and we discuss the trade-offs
between the different representations. Then we use these data structures to
implement the graph traversal algorithms breadth-first search (BFS) and depthfirst search (DFS) efficiently. We will see that BFS and DFS differ essentially
only in that one uses a queue and the other uses a stack, two simple data
structures that we will describe later in this section.
Representing Graphs
There are two basic ways to represent graphs: by an adjacency matrix and
by an adjacency list representation. Throughout the book we wil! use the
adjacency list representation. We start, however, by reviewing both of these
representations and discussing the trade-offs between them.
A graph G = (V, E) has two natural input parameters, the number of nodes
IVI, and the number of edges IEI. We will use n = IVI and m = IEI to denote
these, respectively. Running times will be given in terms of both of these two
parameters. As usual, we will aim for polynomial running times, and lowerdegree polynomials are better. However, with two parameters in the running
time, the comparison is not always so clear. Is O(m2) or O(n3) a better running
time? This depends on what the relation is between n and m. With at most
one edge between any pair of nodes, the number of edges m can be at most
(~) < n2. On the other hand, in many applications the graphs of interest are
connected, and by (3.1), connected graphs must have at least m > n - ! edges.
But these comparisons do not always tell us which of two running times (such
as m2 and n3) are better, so we will tend to keep the running times in terms
of both of these parameters. In this section we aim to. implement the basic
graph search algorithms in time O(m + n). We will refer to this as linear time,
since it takes O(m + n) time simply to read the input. Note that when we work
with connected graphs, a running time of O(m + n) is the same as O(m), since
m>_n-1.

Consider a graph G = (V, E) with n nodes, and assume the set of nodes
is V = {1 ..... n}. The simplest way to represent a graph is by an adjacency

87

88

Chapter 3 Graphs

matrix, which is an n x n matrix A where A[u, v] is equal to ! if the graph


contains the edge (u, v) and 0 otherwise. If the graph is undirected, the matrix A
is symmetric, with A[u, v] = A[v, u] for all nodes u, v ~ V. The adjacency
matrix representation allows us to check in O(1) time if a given edge (u, v) is
present in the graph. However, the representation has two basic disadvantages.
o The representation takes (n2) space. When the graph has many fewer
edges than n2, more compact representations are possible.
o Many graph algorithms need to examine all edges incident to a given node
v. In the adjacency matrix representation, doing this involves considering
all other nodes w, and checking the matrix entry A[v, w] to see whether
the edge (v, w) is present--and this takes (n) time. In the worst case,
v may have (n) incident edges, in which case checking all these edges
will take (n) time regardless of the representation. But many graphs in
practice have significantly fewer edges incident to most nodes, and so it
would be good to be able to find all these incident edges more efficiently.
The representation of graphs used throughout the book is the adjacency
list, which works better for sparse graphs--that is, those with many fewer than
n2 edges. In the adjacency list representation there is a record for each node u,
containing a list of the nodes to which v has edges. To be precise, we have an
array Adj, where Adj [v] is a record containing a list of all nodes adjacent to
node v. For an undirected graph G = (V, E), each edge e = (v, w) ~ E occurs on
two adjacency lists: node w appears on the list for node v, and node ~ appears
on the list for node w.
Lets compare the adiacency matrix and adiacency list representations.
First consider the space required by the representation. An adjacency matrix
requires O(n2) space, since it uses an n x n matrix. In contrast, we claim that
the adjacency list representation requires only O(m + n) space. Here is why.
First, we need an array of pointers of length n to set up the lists in Adj, and
then we need space for all the lists. Now, the lengths of these lists may differ
from node to node, but we argued in the previous paragraph that overall, each
edge e = (v, w) appears in exactly two of the lists: the one for u and the one
for w. Thus the total length of al! lists is 2m = O(m).
Another (essentially equivalent) way to iustify this bound is as follows.
We define the degree nv of a node v to be the number of incident edges it has.
The length of the list at Adj [u] is list is nv, so the total length over all nodes is
O (~v~v nu). Now, the sum of the degrees in a graph is a quantity that often
comes up in the analysis of graph algorithms, so it is useful to work out what
this sum is.
(3.9)

~u~v nv=2m.

3.3 Implementing Graph Traversal Using Queues and Stacks

Proof. Each edge e = (u, w) contributes exactly twice to this sum: once in the.
quantity nu and once in the quantity nw. Since the sum is the total of th~
contributions of each edge, it is 2m.
We sum up the comparison between adjacency matrices and adjacency
lists as follows.
(3.10) The adjacency matrix representation of a graph requires O(n2) space,
while the adjacency list representation requires only O(m + n) ~pace.
Since we have already argued that m < n2, the bound O(m + n) is never
worse than O(n2); and it is much better when the underlying graph is sparse,
with m much smaller than n2.
Now we consider the ease of accessing the information stored in these two
different representations. Recall that in an adjacency matrix we can check in
O(1) time if a particular edge (u, v) is present in the graph. In the adjacency list
representation, this can take time proportional to the degree O(nv): we have to
follow the pointers on us adjacency list to see if edge u occurs on the list. On
the other hand, if the algorithm is currently looking at a node u, it can read
the list of neighbors in constant time per neighbor.
In view of this, the adjacency list is a natural representation for explorihg
graphs. If the algorithm is currently looking at a node u, it can read this list
of neighbors in constant time per neighbor; move to a neighbor ~ once it
encounters it on this list in constant time; and then be ready to read the list
associated with node v. The list representation thus corresponds to a physical
notion of "exploring" the graph, in which you learn the neighbors of a node
u once you arrive at u, and can read them off in constant time per neighbor.

Queues and Stacks


Many algorithms have an inner step in which they need to process a set of
elements, such the set of all edges adjacent to a node in a graph, the set of
visited nodes in BFS and DFS, or the set of all free men in the Stable Matching
algorithm. For this purpose, it is natural to maintain the set of elements to be
considered in a linked list, as we have done for maintaining the set of free men
in the Stable Matching algorithm.
One important issue that arises is the order in which to consider the
elements in such a list. In the Stable Matching algorithm, the order in which
we considered the free men did not affect the outcome, although this required
a fairly subtle proof to verify. In many other algorithms, such as DFS and BFS,
the order in which elements are considered is crucial.

89

90

Chapter 3 Graphs
Two of the simplest and most natural options are to maintain a set of
elements as either a queue or a stack. A queue is a set from which we extract
elements in first-in, first-out (FIFO) order: we select elements in the same order
in which they were added. A stack is a set from which we extract elements
in last-in, first-out (LIFO) order: each time we select an element, we choose
the one that was added most recently. Both queues and stacks can be easily
implemented via a doubly linked list. In both cases, we always select the first
element on our list; the difference is in where we insert a new element. In a
queue a new element is added to the end of the list as the last element, while
in a stack a new element is placed in the first position on the list. Recall that a
doubly linked list has explicit First and Last pointers to the beginning and
end, respectively, so each of these insertions can be done in constant time.
Next we will discuss how to implement the search algorithms of the
previous section in linear time. We will see that BFS can be thought of as
using a queue to select which node to consider next, while DFS is effectively
using a stack.

Implementing Breadth-First Search


The adjacency list data stxucture is ideal for implementing breadth-first search.
The algorithm examines the edges leaving a given node one by one. When we
are scanning the edges leaving u and come to an edge (u, u), we need to
know whether or not node u has been previously discovered by the search.
To make this simple, we maintain an array Discovered of length n and set
Discovered[u] = true as soon as our search first sees u. The algorithm, as
described in the previous section, constructs layers of nodes LI, L2 ..... where
Li is the set of nodes at distance i from the source s. To maintain the nodes in
a layer Li, we have a list L[i] for each i --- 0, I, 2
BFS (s) :
Set Discovered[s] = true and Discovered[u] = false for all other u
Initialize L[0] to consist of the single element s
Set the layer counter i----0
Set the current BFS tree T=0
While /[f] is not empty
Initialize an empty list i[i I]
For each node u E i[i]
Consider each edge (u, u) incident to u
If Discovered[u] = false then
Set Discovered[u] = true
Add edge (u,u) to the tree T

3.3 Implementing Graph Traversal Using Queues and Stacks


Add u to the list L[i+ I]
Endif
End/or
Increment the layer counter i by one
Endwhile

In this implementation it does not matter whether we manage each list


L[i] as a queue or a stack, since the algorithm is allowed to consider the nodes
in a layer Li in any order.
(3.11) The above implementation of the BFS algorithm tans in time O(m + n)
(i.e., linear in the input size), if the graph is given by the adjacency list
representation.
Proof. As a first step, it is easy to bound the running time of the algorithm
by O(n2) (a weaker bound than our claimed O(m + n)). To see this, note that
there are at most n lists L[i] that we need to set up, so this takes O(n) time.
Now we need to consider the nodes u on these lists. Each node occurs on at
most one list, so the For loop runs at most n times over a].l iterations of the
While loop. When we consider a node u, we need to look through all edges
(u, u) incident to u. There can be at most n such edges, and we spend O(1)
time considering each edge. So the total time spent on one iteration of the For
loop is at most O(n). Weve thus concluded that there are at most n iterations
of the For loop, and that each iteration takes at most O(n) time, so the total
time is at most O(n2).
To get the improved O(m + n) time bound, we need to observe that the
For loop processing a node u can take less than O(n) time if u has only a
few neighbors. As before, let nu denote the degree of node u, the number of
edges incident to u. Now, the time spent in the For loop considering edges
incident to node u is O(nu), so the total over all nodes is O(Y~u~v ha). Recall
from (3.9) that ~,v nu = 2m, and so the total time spent considering edges
over the whole algorithm is O(m). We need O(n) additional time to set up
lists and manage the array Discovered. So the total time spent is O(m + n)
as claimed. ~
We described the algorithm using up to n separate lists L[i] for each layer
L~. Instead of all these distinct lists, we can implement the algorithm using a
single list L that we maintain as a queue. In this way, the algorithm processes
nodes in the order they are first discovered: each time a node is discovered,
it is added to the end of the queue, and the algorithm always processes the
edges out of the node that is currently first in the queue.

91

92

3.3 Implementing Graph Traversal Using Queues and Stacks

Chapter 3 Graphs

If we maintain the discovered nodes in this order, then al! nodes in layer Li
will appear in the queue ahead of all nodes in layer Li+l, for i = 0, 1, 2 .... Thus,
all nodes in layer Li will be considered in a contiguous sequence, followed
by all nodes in layer Li+l, and so forth. Hence this implementation in terms
of a single queue wi!l produce the same result as the BFS implementation
above.

DFS (s) :
Initialize S to be a stack with one element s
While S is not empty
Take a node u from S
If Explored[u] = false then
Set Explored[u] = true
For each edge (u, v) incident to u
Add v to the stack S

Implementing Depth-First Search


We now consider the depth-first search algorithm: In the previous section we
presented DFS as a recursive procedure, which is a natural way to specify it.
However, it can also be viewed as almost identical to BFS, with the difference
that it maintains the nodes to be processed in a stack, rather than in a queue.
Essentially, the recursive structure of DFS can be viewed as pushing nodes
onto a stack for later processing, while moving on to more freshly discovered
nodes. We now show how to implement DFS by maintaining this stack of
nodes to be processed explicitly.
In both BFS and DFS, there is a distinction between the act of discovering
a node v--the first time it is seen, when the algorithm finds an edge leading
to v--and the act of exploring a node v, when all the incident edges to v are
scanned, resulting in the potential discovery of further nodes. The difference
between BFS and DFS lies in the way in which discovery and exploration are
interleaved.
In BFS, once we started to explore a node u in layer Li, we added all its
newly discovered neighbors to the next layer L~+I, and we deferred actually
exploring these neighbors until we got to the processing of layer L~+I. In
contrast, DFS is more impulsive: when it explores a node u, it scans the
neighbors of u until it finds the fffst not-yet-explored node v (if any), and
then it immediately shifts attention to exploring v.
To implement the exploration strategy of DFS, we first add all of the nodes
adjacent to u to our list of nodes to be considered, but after doing this we
proceed to explore a new neighbor v of u. As we explore v, in turn, we add
the neighbors of v to the list were maintaining, but we do so in stack order,
so that these neighbors will be explored before we return to explore the other
neighbors of u. We only come back to other nodes adjacent to u when there
are no other nodes left.
In addition, we use an array Explored analogous to the Discovered array
we used for BFS. The difference is that we only set Explored[v] to be true
when we scan vs incident edges (when the DFS search is at v), while BFS sets
Discovered[v] to true as soon as v is first discovered. The implementation
in full looks as follows.

Endfor
Endif
Endwhile

There is one final wrinkle to mention. Depth-first search is underspecified,


since the adjacency list of a node being explored can be processed in any order.
Note that the above algorithm, because it pushes all adjacent nodes onto the
stack before considering any of them, in fact processes each adjacency list
in the reverse order relative to the recursive version of DFS in the previous
section.
(3.12) The above algorithm implements DFS, in the sense that it visits the
nodes in exactly the same order as the recursive DFS procedure in the previous
section (except that each ad]acency list is processed in reverse order).
If we want the algorithm to also find the DFS tree, we need to have each
node u on the stack S maintain the node that "caused" u to get added to
the stack. This can be easily done by using an array parent and setting
parent[v] = u when we add node v to the stack due to edge (u, v). When
we mark a node u # s as Explored, we also can add the edge (u,parent[u])
to the tree T. Note that a node v may be in the stack S multiple times, as it
can be adjacent to multiple nodes u that we explore, and each such node adds
a copy of v to the stack S. However, we will only use one of these copies to
explore node v, the copy that we add last. As a result, it suffices to maintain one
value parent [v] for each node v by simply overwriting the value parent [v]
every time we add a new copy of v to the stack S.

The main step in the algorithm is to add and delete nodes to and from
the stack S, which takes O(1) time. Thus, to bound t~e running time, we
need to bound the number of these operations. To count the number of stack
operations, it suffices to count the number of nodes added to S, as each node
needs to be added once for every time it can be deleted from S.
How many elements ever get added to S? As before, let nu denote the
degree of node v. Node v will be added to the stack S every time one of its
nv adjacent nodes is explored, so the total number of nodes added to S is at

93

94

Chapter 3 Graphs

most ~u nv = 2m. This proves the desired O(m + n) bound on the running
time of DFS.
(3.13) The above implementation of the DFS algorithm runs in time O( m + n)
(i.e., linear in the input size), if the graph is given by the adjacency list
representation.

Finding the Set of All Connected Components


In the previous section we talked about how one c.an use BFS (or DFS) to find
all connected components of a graph. We start with an arbitxary node s, and
we use BFS (or DFS) to generate its connected component. We then find a
node v (if any) that was not visited by the search from s and iterate, using
BFS (or DFS) starting from v to generate its connected component--which, by
(3.8), wil! be disjoint from the component of s. We continue in this way until
all nodes have been visited.
Although we earlier expressed the running time of BFS and DFS as O(m +"
n), where m and n are the total number of edges and nodes in the graph, both
BFS and DFS in fact spend work only on edges and nodes in the connected
component containing the starting node. (They never see any of the other
nodes or edges.) Thus the above algorithm, although it may run BFS or
DFS a number of times, only spends a constant amount of work on a given
edge or node in the iteration when the connected component it belongs to is
under consideration. Hence the overall running time of this algorithm is still

O(m + n).

3.4 Testing Bipartiteness: An Application of


Breadth-First Search
Recall the definition of a bipartite graph: it is one where the node set V can
be partitioned into sets X and Y in such a way that every edge has one end
in X and the other end in Y. To make the discussion a little smoother, we can
imagine that the nodes in the set X are colored red, and the nodes in the set
Y are colored blue. With this imagery, we can say a graph is bipartite if it is
possible to color its nodes red and blue so that every edge has one red end
and one blue end.
~ The Problem
In the earlier chapters, we saw examples of bipartite graphs. Here we start by
asking: What are some natural examples of a nonbipartite graph, one where
no such partition of V is possible?

3.4 Testing Bipartiteness: An Application of Breadth-First Search

Clearly a triangle is not bipartite, since we can color one node red,,another
one blue, and then we cant do anything with the third node. More generally,
consider a cycle C of odd leng~, with nodes numbered 1, 2, 3 ..... 2k, 2k + 1.
If we color node 1 red, then we must color node 2 blue, and then we must color
node 3 red, and so on--coloring odd-numbered nodes red and even-numbered
nodes blue. But then we must color node 2k + 1 red, and it has an edge to node
1, which is also red. This demonstrates that theres no way to partition C into
red and blue nodes as required. More generally, if a graph G simply contains
an odd cycle, then we can apply the same argument; thus we have established
the following.
(3,14) If.d graph G is bipartite, then it cannot contain an odd cycle.

It is easy to recognize that a graph is bipartite when appropriate sets X


and Y (i.e., red and blue nodes) have actually been identified for us; and in
many settings where bipartite graphs arise, this is natural. But suppose we
encounter a graph G with no annotation provided for us, and wed like to
determine for ourselves whether it is bipartite--that is, whether there exists a
partition into red and blue nodes, as required. How difficult is this? We see from
(3.14) that an odd cycle is one simple "obstacle" to a graphs being bipartite.
Are there other, more complex obstacles to bipartitness?

/~ Designing the Algorithm

In fact, there is a very simple procedure to test for bipartiteness, and its analysis
can be used to show that odd cycles are the only obstacle. First we assume
the graph G is connected, since otherwise we can first compute its connected
components and analyze each of them separately. Next we pick any node s ~ V
and color it red; there is no loss in doing this, since s must receive some color.
It follows that all the neighbors of s must be colored blue, so we do this. It
then follows that all the neighbors of these nodes must be colored red, their
neighbors must be colored blue, and so on, unti! the whole graph is colored. At
this point, either we have a valid red/blue coloring of G, in which every edge
has ends of opposite colors, or there is some edge with ends of the same color.
In this latter case, it seems clear that theres nothing we could have donei G
simply is not bipartite. We now want to argue this point precisely and also
work out an efficient way to perform the coloring.
The first thing to notice is that the co!oring procedure we have just
described is essentially identical to the description of BFS: we move outward
from s, co!oring nodes as soon as we first encounter them. Indeed, another
way to describe the coloring algorithm is as follows: we perform BFS, coloring

95

3.5 Connectivity in Directed Graphs

Chapter 3 Graphs

96

s red, all of layer L1 blue, all of layer L2 red, and so on, coloring odd-numbered
layers blue and even-numbered layers red.
We can implement this on top of BFS, by simply taking the implementation
of BFS and adding an extra array Color over the nodes. Whenever we get
to a step in BFS where we are adding a node v to a list L[i + 1], we assign
Color[u] = red if i + I is an even number, and Color[u] = blue if i + 1 is an
odd number. At the end of this procedure, we simply scan al! the edges and
determine whether there is any edge for which both ends received the same
color. Thus, the total running time for the coloring algorithm is O(m + n), iust
as it is for BFS.

~ Analyzing the Algorithm


We now prove a claim that shows this algorithm correctly determines whether
G is bipartite, and it also shows that we can find an odd cycle in G whenever
it is not bipartite.
(3.15} Let G be a connected graph, and let LI, L2 .... be the layers produced
by BFS starting at node s. Then exactly one of the following two things must
hold.

he cycle through x, y,~


d z has odd length9

Layer Li

Layer Lj
Figure 3.6 If two nodes x and
y in the same layer a_re joined
by an edge, then the cycle
through x, y, and their lowest
common ancestor z has odd
length, demonstrating that
the graph cannot be bipartite.

(0 There is no edge of G joining two nodes of the same layer. In this case G
is a bipartite graph in which the nodes in even-numbered layers can be
colored red, and the nodes in odd-numbered layers can be colored blue.
(ii) There is an edge of G joining two nodes of the same layer. In this case, G
contains an odd-length cycle, and so it cannot be bipartite.
ProoL First consider case (i), where we suppose that there is no edge joining
two nodes of the same layer. By (3.4), we know that every edge of G ioins nodes
either in the same layer or in adiacent layers. Our assumption for case (i) is
precisely that the first of these two alternatives never happens, so this means
that every edge joins two nodes in adjacent layers. But our coloring procedure
gives nodes in adjacent layers the opposite colors, and so every edge has ends
with opposite colors. Thus this coloring establishes that G is bipartite.
Now suppose we are in case (ii); why must G contain an odd cycle? We
are told that G contains an edge joining two nodes of the same layer. Suppose
this is the edge e = (x, y), with x, y ~ Lj. Also, for notational reasons, recall
that L0 ("layer 0") is the set consisting of just s. Now consider the BFS tree T
produced by our algorithm, and let z be the node whose layer number is as
large as possible, subject to the condition that z is an ancestor of both x and y
in T; for obvious reasons, we can cal! z the lowest common ancestor of x and y.
Suppose z ~ Li, where i < j. We now have the situation pictured in Figure 3.6.
We consider the cycle C defined by following the z-x path in T, then the edge e,

and then the y-z path in T. The length of this cycle is (j - i) + 1 + (j - i),-adding
the length of its three parts separately; this is equal to 2(j - i) + 1;which is an
odd number. []

3.5 Connectivity in Directed Graphs


Thus far, we have been looking at problems on undirected graphs; we now
consider the extent to which these ideas carry over to the case of directed

graphs.
Recall that in a directed graph, the edge (u, v) has a direction: it goes from
u to v. In this way, the relationship between u and v is asymmetric, and this
has qualitative effects on the structure of the resulting graph. In Section 3.1, for
example, we discussed the World Wide Web as an instance of a large, complex
directed graph whose nodes are pages and whose edges are hyperlinks. The act
of browsing the Web is based on following a sequence of edges in this directed
graph; and the directionality is crucial, since its not generally possible to
browse "backwards" by following hyperlinks in the reverse direction.
At the same time, a number of basic definitions and algorithms have
natural analogues in the directed case. This includes the adjacency list representation and graph search algorithms such as BFS and DFS. We now discuss
these in turn.

Representing Directed Graphs


In order to represent a directed graph for purposes of designing algorithms,
we use a version of the adiacency list representation that we employed for
undirected graphs. Now, instead of each node having a single list of neighbors,
each node has two lists associated with it: one list consists of nodes to which it
has edges, and a second list consists of nodes from which it has edges. Thus an
algorithm that is currently looking at a node u can read off the nodes reachable
by going one step forward on a directed edge, as well as the nodes that would
be reachable if one went one step in the reverse direction on an edge from u.
The Graph Search Algorithms
Breadth-first search and depth-first search are almost the same in directed
graphs as they are in undirected graphs. We will focus here on BFSi We start
at a node s, define a first layer of nodes to consist of all those to which s has
an edge, define a second layer to consist of all additional nodes to which these
first-layer nodes have an edge, and so forth. In this way, we discover nodes
layer by layer as they are reached in this outward search from s, and the nodes
in layer j are precisely those for which the shortest path from s has exactly
j edges. As in the undirected case, this algorithm performs at most constant
work for each node and edge, resulting in a running time of O(m + n).

97

98

Chapter 3 Graphs

It is important to understand what this directed version of BFS is computing. In directed graphs, it is possible for a node s to have a path to a node t
even though t has no path to s; and what directed BFS is computing is the set
of all nodes t with the property that s has a path to t. Such nodes may or may
not have paths back to s.
There is a natural analogue of depth-first search as well, which also runs
in linear time and computes the same set of nodes. It is again a recursive
procedure that tries to explore as deeply as possible, in this case only following
edges according to their inherent direction. Thus, when DFS is at a node u, it
recursively launches .a depth-first search, in order, for each node to which u
has an edge.
Suppose that, for a given node s, we wanted the set of nodes with paths
to s, rather than the set of nodes to which s has paths. An easy way to do this
would be to define a new directed graph, Grev, that we obtain from G simply
by reversing the direction of every edge. We could then run BFS or DFS in GreY;
a node has a path from s in Gre~ if and only if it has a path to s in G.

Strong Connectivity
Recall that a directed graph is strongly connected if, for every two nodes u and
v, there is a path from u to v and a path from v to u. Its worth also formulating
some terminology for the property at the heart of this definition; lets say that
two nodes u and v in a directed graph are mutually reachable if there is a path
from u to v and also a path from v to u. (So a graph is strongly connected if
every pair of nodes is mutually reachable.)
Mutual teachability has a number of nice properties, many of them stemruing from the following simple fact.
(3.16) If u and v are mutually reachable, and v and iv are mutually reachable,
then u and iv are mutually reachable.
Proof. To construct a path from u to w, we first go from u to v (along the
path guaranteed by the mutual teachability of u and v), and then on from v
to iv (along the path guaranteed by the mutual teachability of v and w). To
construct a path from w to u, we just reverse this reasoning: we first go from
iv to v (along the path guaranteed by the mutual reachability of v and iv), and
then on from v to u (along the path guaranteed by the mutual teachability of
u and v). a
There is a simple linear-time algorithm to test if a directed graph is strongly
connected, implicitly based on (3.16). We pick any node s and run BFS in G
starting from s. We then also run BFS starting from s in Grev. Now, if one of
these two searches fails to reach every node, then clearly G is not strongly
connected. But suppose we find that s has a path to every node, and that

3.6 Directed Acyclic Graphs and Topological Ordering

every node has a path to s. Then s and v are mutually reachable for every v,
and so it follows that every two nodes u and v are mutually reachable: s and
u are mutually reachable, and s and v are mutually reachable, so by (3.16) we
also have that u and v are mutually reachable.
By analogy with connected components in an undirected graph, we can
define the strong component containing a node s in a directed graph to be the
set of all v such that s and v are mutually reachable. If one thinks about it, the
algorithm in the previous paragraph is really computing the strong component
containing s: we run BFS starting from s both in G and in Gre"; the set of nodes
reached by both searches is the set of nodes with paths to and from s, and
hence this set is the strong component containing s.
There are further similarities between the notion of connected components
in undirected graphs and strong components in directed graphs. Recall that
connected components naturally partitioned the graph, since any two were
either identical or disjoint. Strong components have this property as well, and
for essentially the same reason, based on (3.16).
(3.17) For any two nodes s and t in a directed graph, their strong Components
are either identical or disjoint.
Proof. Consider any two nodes s and t that are mutually reachable; we claim
that the strong components containing s and t are identical. Indeed, for any
node v, if s and v are mutually reachable, then by (3.16), t and v are mutually
reachable as wel!. Similarly, if t and v are mutually reachable, then again by
(3.16), s and v are mutually reachable.

On the other hand, if s and t are not mutually reachable, then there cannot
be a node v that is in the strong component of each. For if there were such
a node v, then s and u would be mutually reachable, and ~ and t would be
mutually reachable, so from (3.16) it would follow that s and t were mutually
reachable. ,,
In fact, although we will not discuss the details of this here, with more
work it is possible to compute the strong components for all nodes in a total
time of O(m + n).

3.6 Directed Acyclic Graphs and


Topological Ordering
If an undirected graph has no cycles, then it has an extremely simple structure:
each of its connected components is a tree. But it is possible for a directed graph
to have no (directed).cycles and still have a very rich structure. For example,
such graphs can have a large number of edges: if we start with the node

99

100

3.6 Directed Acyclic Graphs and Topological Ordering

Chapter 3 Graphs

~e~

a topological ordering, all


ges point from left to right.)

(c)
Figure 3.7 (a) A directed acyclic graph. (b) The same DAG with a topological ordering,
specified by the labels on each node. (c) A different drawing of the same DAG, arranged
so as to emphasize the topological ordering.

set {1, 2 ..... n} and include an edge (i,j) whenever i <j, then the resulting
directed graph has (~) edges but no cycles.
If a directed graph has no cycles, we call it--naturally enough--a directed
acycIic graph, or a DAG for short. (The term DAG is typically pronounced as a
word, not spelled out as an acronym.) In Figure 3.7(a) we see an example of
a DAG, although it may take some checking to convince oneself that it really
has no directed cycles.
~,~ The Problem
DAGs are a very common structure in computer science, because many kinds
of dependency networks of the type we discussed in Section 3.1 are acyclic.
Thus DAGs can be used to encode precedence relations or dependencies in a
natural way. Suppose we have a set of tasks labeled {1, 2 ..... n} that need to
be performed, and there are dependencies among them stipulating, for certain
pairs i and j, that i must be performed before j. For example, the tasks may be
courses, with prerequisite requirements stating that certain courses must be
taken before others. Or the tasks may correspond to a pipeline of computing
iobs, with assertions that the output of iob i is used in determining the input
to iob j, and hence job i must be .done before iob j.
We can represent such an interdependent Set of tasks by introducing a
node for each task, and a directed edge (i, j) whenever i must be done before
j. If the precedence relation is to be at all meaningful, the resulting graph G
must be a DAG. Indeed, if it contained a cycle C, there would be no way to do
any of the tasks in C: since each task in C cannot begin until some other one
completes, no task in C could ever be done, since none could be done first.

Lets continue a little further with this picture of DAGs as precedence


relations. Given a set of tasks with dependencies, it would be natural to seek
a valid order in which the tasks could be performed, so that all dependencies
are respected. Specifically, for a directed graph G, we say that a topological
ordering of G is an ordering of its nodes as ul, u2 ..... un so that for every edge
(ui, uj), we have i < j. In other words, all edges point "forward" in the ordering.
A topological ordering on tasks provides an order in which they can be safely
performed; when we come to the task vj, all the tasks that are required to
precede it have already been done. In Figure 3.7(b) weve labeled the nodes of
the DAG from part (a) with a topological ordering; note that each edge indeed
goes from a lower-indexed node to a higher-indexed node.
In fact, we can view a topological ordering of G as providing an immediate
"proof" that G has no cycles, via the following.
(3.18} If G has a topological ordering, then G is a DAG.
Proof. Suppose, by way of contradiction, that G has a topological ordering
un, and also has a cycle C. Let ui be the lowest-indexed node on C,
and let uj be the node on C just before ui--thus (vj, vi) is an edge. But by our
choice of i, we havej > i, which contradicts the assumption that u1, u2 ..... un
was a topological ordering. ~
The proof of acyclicity that a topological ordering provides can be very
useful, even visually. In Figure 3.7(c), we have drawn the same graph as
in (a) and (b), but with the nodes laid out in the topological ordering. It is
immediately clear that the graph in (c) is a DAG since each edge goes from left
to right.
Computing a Topological Ordering Themain question we consider here is
the converse of (3. ! 8): Does every DAG have a topological ordering, and if so,
how do we find one efficiently? A method to do this for every DAG would be
very useful: it would show that for any precedence relation on a set of tasks
without cycles, there is an efficiently computable order in which to perform
the tasks.

~ Designing and Analyzing the Algorithm


In fact, the converse of (3.18) does hold, and we establish this via an efficient
algorithra to compute a topological ordering. The key to this lies in finding a
way to get started: which node do we put at the beginning of the topological
ordering? Such a node Vl would need to have no incoming edges, since any
such incoming edge would violate the defining property of the topological

101

102

3.6 Directed Acyclic Graphs and Topological Ordering

Chapter 3 Graphs
ordering, that all edges point forward. Thus, we need to prove the following
fact.
(3.19) In every DAG G, there is a node v with no incoming edges.
Proof. Let G be a directed graph in which every node has at least one incoming
edge. We show how to find a cycle in G; this will prove the claim. We pick
any node v, and begin following edges backward from v: sihce v has at least
one incoming edge (u, v), we can walk backward to u; then, since u has at
least one incoming edge (x, u), we can walk backward to x; and so on. We
can continue this process indefinitely, since every node we encounter has an
incoming edge. But after n + I steps, we will have visited some node w twice. If
we let C denote the sequence of nodes encountered between successive visits
to w, then clearly C forms a cycle, m

In fact, the existence of such a node v is all we need to produce a topological


ordering of G by induction. Specifically, let us claim by induction that every
DAG has a topological ordering. This is clearly true for DAGs on one or two
nodes. Now suppose it is true for DAGs with up to some number of nodes n.
Then, given a DAG G on n + 1 nodes, we find a node v with no incoming edges,
as guaranteed by (3.19). We place v first in the topological ordering; this is
safe, since all edges out of v will point forward. Now G-(v} is a DAG, since
deleting v cannot create any cycles that werent there previously. Also, G- {v}
has n nodes, so we can apply the induction hypothesis to obtain a topological
ordering of G- {v}. We append the nodes of G- {v} in this order after v; this is
an ordering of G in which all edges point forward, and hence it is a topological
ordering.
Thus we have proved the desired converse of (3.18).
(3.20) ff G is a DAG, then G has a topological ordering. .... ~
The inductive proof contains the following algorithm to compute a topological ordering of G.
To compute a topological ordering of G:
Find a node v with no incoming edges and order it first

Delete v from G
Recursively compute a topological ordering of G-{v}
and append this order after u

In Figure 3.8 we show the sequence of node deletions that occurs when this
algorithm is applied to the graph in Figure 3.7. The shaded nodes in each
iteration are those with no incoming edges; the crucial point, which is what

(a)

(d)

(e)

(f)

Figure 3.8 Starting from the graph in Figure 3.7, nodes are deleted one by one so as
to be added to a topologica! ordering. The shaded nodes are those with no incoming
edges; note that there is always at least one such edge at every stage of the algorithms
execution.

(3.19) guarantees, is that when we apply this algorithm to a DAG, there will
always be at least one such node available to delete.
To bound the running time of this algorithm, we note that identifying a
node v with no incoming edges, and deleting it from G, can be done in O(n)
time. Since the algorithm runs for n iterations, the total running time is O(n2).
This is not a bad running time; and if G is very dense, containing (n2)
edges, then it is linear in the size of the input. But we may well want something
better when the number of edges m is much less than n2. In such a case, a
running time ofO(m + n) could be a significant improvement over (n2).
In fact, we can achieve a running time of O(m + n) using the same highlevel algorithm--iteratively deleting nodes with no incoming edges. We simply
have to be more efficient in finding these nodes, and we do tBis as follows.
We declare a node to be "active" ff it has not yet been deleted by the
algorithm, and we explicitly maintain two things:
(a) for each node m, the number of incoming edges that tv has from active
nodes; and
(b) the set S of all active nodes in G that have no incoming edges from other
active nodes.

103

104

Solved Exercises

Chapter 3 Graphs
At the start, all nodes are active, so we can initialize (a) and (b) with a single
pass through the nodes and edges. Then, each iteration consists of selecting
a node u from the set S and deleting it. After deleting u, we go through all
nodes tv to which u had an edge, and subtract one from the number of active
incoming edges that we are maintaining for w. If this causes the number
of active incoming edges tow to drop to zero, then we add tv to the set S.
Proceeding in this way, we keep track of nodes that are eligible for deletion at
all times, while spending constant work per edge over the course of the whole
algorithm.

Solved Exercises
Figure 3.9 How many topological orderings does this
graph have?

Solved Exercise 1
Consider the directed acyclic graph G in Figure 3.9. How many topological
orderings does it have?
Solution Recall that a topological ordering of G is an ordering of the nodes.
as vl, v2 ..... vn so that all edges point "forward": for every edge (vi, vj), we
have i < j.
So one way to answer this question would be to write down all 5- 4.3.21 = 120 possible orderings and check whether each is a topological ordering.
But t_his would take a while.
Instead, we think about this as follows. As we saw in the text (or reasoning
directly from the definition), the first node in a topological ordering must be
one that has no edge coming into it. Analogously, the last node must be one
that has no edge leaving it. Thus, in every topological ordering of G, the node a
must come first and the node e must come last.
Now we have to figure how the nodes b, c, and d can be arranged in the
middle of the ordering. The edge (c, d) enforces the requirement that c must
come before d; but b can be placed anywhere relative to these two: before
both, between c and d, or after both. This exhausts ~11 the possibilities, and
so we conclude that there are three possible topological orderings:

a,b,c,d,e
a,c,b,d,e
a,c,d,b,e
Solved Exercise 2
Some friends of yours are working on techniques for coordinating groups of
mobile robots. Each robot has a radio transmitter that it uses to communicate

with a base station, and your friends find that if the robots get too close to one
another, then there are problems with interference among the transmitters. So
a natural problem arises: how to plan the motion of the robots in such a way
that each robot gets to its intended destination, but in the process the robots
dont come close enough together to cause interference problems.
We can model this problem abstractly as follows. Suppose that we have
an undirected graph G = (V, E), representing the floor plan of a building, and
there are two robots initially located at nodes a and b in the graph. The robot
at node a wants to travel to node c along a path in G, and the robot at node b
wants to travel to node d. This is accomplished by means of a schedule: at
each time step, the schedule specifies that one of the robots moves across a
single edge, from one node to a neighboring node; at the end of the schedule,
the robot from node a should be sitting on c, and the robot from b should be
sitting on d.
A schedule is interference-free if there is no point at which the two.robots
occupy nodes that are at a distance < r from one another in the graph, for a
given parameter r. Well assume that the two starting nodes a and b are at a
distance greater than r, and so are the two ending nodes c and d.
Give a polynomial-time algorithm that decides whether there exists an
interference-free schedule by which each robot can get to its destination.

Solution This is a problem of the following general flavor. We have a set


of possible configurations for the robots, where we define a configuration
to be a choice of location for each one. We are trying to get from a given
starting configuration (a, b) to a given ending configuration (c, d), subject to
constraints on how we can move between configurations (we can only change
one robots location to a neighboring node), and also subject to constraints on
which configurations are "legal."
This problem can be tricky to think about if we view things at the level of
the underlying graph G: for a given configuration of the robots--that is, the
current location of each one--its not clear what rule we should be using to
decide how to move one of the robots next. So instead we apply an idea that
can be very useful for situations in which were trying to perform this type of
search. We observe that our problem looks a lot like a path-finding problem,
not in the original graph G but in the space of all possible configurations.
Let us define the following (larger) graph H. The node set of H is the set
of all possible configurations of the robots; that is, H consists of a!! possible
pairs of nodes in G. We join two nodes of H by an edge if they represent
configurations that could be consecutive in a schedule; that is, (u, v) and
(u, u) will be joined by an edge in H if one of the pairs u, u or v, u are equal,
and the other pair corresponds to an edge in G.

105

106

Exercises

Chapter 3 Graphs

We can already observe that paths in H from (a,/)) to (c, d) correspond


to schedules for the robots: such a path consists precisely of a sequence of
configurations in which, at each step, one robot crosses a single edge in G.
However, we have not yet encoded the notion that the schedule should be
interference-free.
To do this, we simply delete from H all nodes that correspond to configurations in which there would be interference. Thus we define H~ to be the graph
obtained from H by deleting all nodes (u, v) for which the distance between
u and v in G is at most r.
The full algorithm is then as follows. We construct the graph H, and then
run the connectiviW algorithm from the text to determine whether there is a
path from (a, b) to (c, d). The correctness of the algorithm follows from the
fact that paths in H correspond to schedules, and the nodes in H correspond
precisely to the configurations in which there is no interference.
Finally, we need to consider the running time. Let n denote the number
of nodes in G, and m denote the number of edges in G. Well analyze the
running time by doing three things: (1) bounding the size of H (which will in
general be larger than G), (2) bounding the time it takes to construct H, and
(3) bounding the time it takes to search for a path from (a, b) to (c, d) in H.
1. First, then, lets consider the size of H. H has at most nz nodes, since
its nodes correspond to pairs of nodes in G. Now, how many edges does
H have? A node (u, v) will have edges to (u, v) for each neighbor u
of u in G, and to (u, v) for each neighbor v of v in G. A simple upper
bound says that there can be at most n choices for (u, u), and at most n
choices for (u, v), so there are at most 2n edges incident to each node
of H. Summing over the (at most) n2 nodes of H, we have O(n3) edges.
(We can actually give a better bound of O(mn) on the number of
edges in H~, by using the bound (3.9) we proved in Section 3.3 on the
sum of the degrees in a graph. Well leave this as a further exercise.)
2. Now we bound the time needed to construct H. We first build H by
enumerating all pairs of nodes in G in time O(n2), and constructing edges
using the defiNtion above in time O(n) per node, for a total of O(n3).
Now we need to figure out which nodes to delete from H so as to produce
H. We can do this as follows. For each node u in G, we run a breadthfirst search from u and identify all nodes u within distance r of u. We list
all these pairs (u, v) and delete them from H. Each breadth-first search
in G takes time O(m + n), and were doing one from each node, so the
total time for this part is O(rnri + n2).

107

Now we have H, and so we just need to decide whether there is a path


from (a, b) to (c, d). This can be done using the connectivity algorithm
from the text in time that is linear in the number of nodes and edges
of H. Since H has O(n2) nodes and O(n~) edges, this final step takes
polynomial time as well.

Exercises
1. Considhr the directed acyclic graph G in Figure 3.10. How many topolog- Figure 3.10 How many topoical orderings does it have?
logical orderings does this
graph have?

Give an algorithm to detect whether a given undirected graph contains


a cycle. If the graph contains a cycle, then your algorithm should output
one. (It should not output all cycles in the graph, just one of them.) The
running time of your algorithm should be O(m + n) for a graph with n
nodes and m edges.
3. The algorithm described in Section 3.6 for computing a topological ordering of a DAG repeatediy finds a node with no incoming edges and deletes
it. This will eventually produce a topological ordering, provided that the
input graph really is a DAG.

But suppose that were given an arbitrary graph that may or may not
be a DAG. Extend the topological ordering algorithm so that, given an
input directed graph G, it outputs one of two things: (a) a topological
ordering, thus establishing that a is a DAG; or (b) a cycle in G, thus
establishing that a is not a DAG. The nmning time of your algorithm
should be O(m + n) for a directed graph with n nodes and m edges.
inspired by the example of that great Cornellian, Vladimir Nabokov, some
of your frien.ds have become amateur lepidopterists (they study butterflies). Often when they return from a trip with specimens of butterf~es,
it is very difficult for them to tell how many distinct species theyve
caught--thanks to the fact that many species look very similar to one
another.

One day they return with n butterflies, and thfiy believe that each
belongs to one of two different species, which well call A and B for
purposes of this discussion. Theyd like to divide the n specimens into
two groups--those that belong to .4 and those that belong to B--but its
very hard for them to directly label any one specimen. So they decide to
adopt the following approach.

108

Exercises

Chapter 3 Graphs
For each pair of specimens i and j, they study them carefully side by
side. If theyre confident enough in their judgment, then they 1abe! the
pair (i,]) either "same" (meaning they believe them both to come from
the same species) or "different" (meaning they believe them to come from
different species). They also have the option of rendering no judgment
on a given pair, in which case we]] call the pair ambiguous.
So now they have the collection of n specimens, as we]] as a collection
of m judgments (either "same" or "different") for the pairs that were not
declared to be ambiguous. Theyd like to know if this data is consistent
with the idea that each butterfly is from one of species A or B. So more
concretely, well declare the m judgments to be consistent if it is possible
to label each specimen either A or/3 in such a way that for each pair (i,])
labeled "same," it is the case that i andj have the same label; and for each
pair (i,j) labeled "different," it is the case that i andj have different labels.
Theyre in the middle of tediously working out whether their judgments
are consistent, when one of them realizes that you probably have an
algorithm that would answer this question right away.
Give an algorithm with running time O(m + n) that determines
whether the m judgments are consistent.
A binary tree is a rooted tree in which each node has at most two children.
Show by induction that in any binary tree the number of nodes with two
children is exactly one less than the number of leaves.
We have a connected graph G = (V, E), and a specific vertex a ~ V. Suppose
we compute a depth-first search tree rooted at a, and obtain a tree T that
includes all nodes of G. Suppose we then compute a breadth-first search
tree rooted at a, and obtain the same tree T. Prove that G = T. (In other
words, if T is both a depth-first search tree and a breadth-first search
tree rooted at a, then G cannot contain anY edges that do not belong to

the following property: at all times, eac~ device i is within 500 meters
of at least n/2 of the other devices. (Well assume n is an even number.)
What theyd like to know is: Does this property by itself guarantee that
the network will remain connected?
Heres a concrete way to formulate the question as a claim about
graphs.
Claim: Let G be a graph on n nodes, where n is an even number. If every node
of G has degree at least hi2, then G is connected.

Decide whether you think the claim is true or false, and give a proof of
either the claim or its negation.
A number of stories In the press about the structure of the Internet and
the Web have focused on some version of the following question: How
far apart are typical nodes in these networks? If you read these stories
carefully, you find that many of them are confused about the difference
between the diameter of a network and the average distance in a network;
they often jump back and forth between these concepts as though theyre
the same thing.
As in the text, we say that the distance between two nodes a and v
in a graph G = (V, E) is the minimum number of edges in a path joining
them; we]] denote this by dist(a, u). We say that the diameter of G is
the maximum distance between any pair of nodes; and weH denote this
quantity by diam(G).
Lets define a related quantity, which weH ca]] the average pairwise
distance In G (denoted apd(G)). We define apd(G) to be the average, over
all (~) sets of two distinct nodes a and u, of the distance between a and ~.
That is,

T.)
Some friends of yours work on wireless networks, and theyre currently
studying the properties of a network of n mobile devices. As the devices
move around (actually, as their human owners move around), they defIne
a graph at any point in time as follows: there is a node representing each
of the n devices, and there is an edge between device i and device j ff the
physical locations of i andj are no more than 500 meters apart. (if so, we
say that i and ] are "in range" of each other.)
Theyd like it to be the case that the network of devices is connected at
all times, and so theyve constrained the motion of the devices to satisfy

Heres a simple example to convince yourself that there are graphs G


for which diam(G) # apd(G). Let G be a graph with ~ee nodes a, v, w, and
with the two edges {a, ~} and {v, w}. Then
diam(G) = dist(a, w) = 2,

while
apd(G) = [dist(u, v) + dist(a, w) + dist(u, w)]/3 = 4/3.

109

110

Exercises

Chapter 3 Graphs
Of course, these two numbers arent all that far apart in the case of
S a
this three-node graph, and so its natural to ask whether there s alway
dose relation between them. Heres a claim that tries to make this precise.
Claim: There exists a positive natural number c so that for all connected graphs
G, it is the case that
diam(G)
apd(G) -

Decide whether you think the claim is true or false, and give a proof of
either the claim or its negation.
~q~Theres a natural intuition that two nodes that are far apart in a communication network--separated by many hops--have a more tenuous
connection than two nodes that are close together. There are a number
of algorithmic results that are based to some extent on different ways of
making this notion precise. Heres one that involves the susceptibiliw of
paths to the deletion of nodes.
Suppose that an n-node undirected graph G = (V, E) contains two
nodes s and t such that the distance between s and t is strictly greater
than n/2. Show that there must exist some node u, not equal to either s
or t, such that deleting v from G destroys all s-t paths. (In other words,
the graph obtained from G by deleting v contains no path from s to t.)
Give an algorithm with runnin~ time O(m + n) to find such a node v.

10, A number of art museums around the countts, have been featuring work

by an artist named Mark Lombardi (1951-2000), consisting of a set of


intricately rendered graphs. Building on a great deal of research, these
graphs encode the relationships among people involved in major political
scandals over the past several decades: the nodes correspond to participants, and each edge indicates some type of relationship between a pair
of participants. And so, if you peer c!osely enough at the drawings, you
can trace out ominous-looking paths from a high-ranking U.S. government official, to a former business partner, to a bank in Switzerland, to
a shadowy arms dealer.
Such pictures form striking examples of social networks, which, as
we discussed in Section 3.1, have nodes representing people and organizations, and edges representing relationships of various kinds. And the
short paths that abound in these networks have attracted considerable
attention recently, as people ponder what they mean. In the case of Mark
Lombardis graphs, they hint at the short set of steps that can carry you
from the reputable to the disreputable.

Of course, a single, spurious short path between nodes v and w in


such a network may be more coincidental than anything else; a large
number of short paths between u and w can be much more convincing.
So In addition to the problem of computing a single shortest v-w path
in a graph G, social networks researchers have looked at the problem of
determining the number of shortest u-w paths.
This rams out to be a problem that can be solved efficiently. Suppose
we are given an undirected graph G = (V, E), and we identif3, two nodes v
and w in G. Give an algorithm that computes the number of shortest u-w
paths In G. (The algorithm should not list all the paths; just the number
suffices.) The nmning time of your algorithm should be O(m + n) for a
graph with n nodes and m edges.
11. Youre helping some security analysts monitor a collection of networked

computers, tracking the spread of an online virus. There are n computers


in the system, labeled C1, C2 ..... Cn, and as input youre given a collection
of trace data Indicating the times at which pairs of computers communicated. Thus the data is a sequence of ordered triples (Ci, Cj, tk); such a
triple indicates that Ci and Cj exchanged bits at time tk. There are m triples
total.
Well assume that the tTiples are presented to you in sorted order of
time. For purposes of simplicity, well assume that each pair of computers
communicates at most once during the interval youre observing.
The security analysts youre working with would like to be able to
answer questions of the following form: If the virus was inserted into
computer Ca at time x, could it possibly have infected computer Cb by
time y? The mechanics of infection are simple: if an infected computer
Ci communicates with an uninfected computer Cj at time t~ (in other
words, if one of the triples (Ci, Cp t~) or (Cj, Ci, t~) appears In the trace
data), then computer Ci becomes infected as well, starting at time t~.
Infection can thus spread from one machine to another across a sequence
of communications, provided that no step in this sequence involves a
move backward in time. Thus, for example, If Ci is infected by time tk,
and the trace data contains triples (Ci, Cj, tD and (Cp Cq, tr), where tk <_ tr,
then Ca will become infected via C~. (Note that it is Okay for t~ to be equal
to 6; this would mean that Cj had open connections to both Ci and Cq at
the same time, and so a virus could move from Ci to Ca.)
For example, suppose n = 4, the trace data consists of the triples
(Ci, C2, 4), (C2, C4, 8), (C3, C4, 8), (Cl, C4, 12),

111

112

Notes and Further Reading

Chapter 3 Graphs
and the virus was inserted into computer C1 at time 2. Then C3 would be
infected at time 8 by a sequence of three steps: first C2 becomes ~ected
at time 4, then C4 gets the virus from C2 at time 8, and then G3 gets the
virus from C4 at time 8. On the other hand, if the trace data were
8), (C1, C4, 12), (C~0 Cz, 14),

and again the virus was inserted into computer C1 at time 2, then C3
would not become infected during the period of observation: although
z becomes infected at time 14, we see that ~ only communicates with Cz
before ~2 was infected. There is no sequence of communications moving
forward in time by which the virus could get from C1 to C~ in this second
example.
Design an algorithm that answers questions of this type: given a
collection of trace data, the algorithm should decide whether a virus
introduced at computer Ca at time x could have infected computer Cb
by time y. The algorithm sh6uld run in time O(rn + n).
12. Youre helping a group of ethnographers analyze some oral history data

theyve collected by interviewing members of a village to learn about the


fives of people whove lived there over the past two hundred years.
From these interviews, theyve learned about a set of n people (all
Vn" Theyve also
collected facts about when these people lived relative to one another.
Each fact has one of the following two forms:
* For some i and j, person Vi died before person Pj was born; or
o for some i and j, the life spans of Vi and Pj overlapped at least partially.
Naturally, theyre not sure that all these facts are correct; memories
are not so good, and a lot of this was passed down by word of mouth. So
what theyd like you to determine is whether the data theyve collected is
at least internally consistent, in the sense that there could have existed a
set of people for which all the facts theyve learned simultaneously hold.
Give an efficient algorithm to do this: either it should produce proposed dates of birth and death for each of the n people so that all the facts
hold true, or it should report (correctly) that no such dates can exist--that
is, the facts collected by the ethnographers are not internally consistent.

Notes and Further Reading


The theory of graphs is a large topic, encompassing both algorithmic and nonalgorithmic issues. It is generally considered to have begun with a paper by

Euler (1736), grown through interest in graph representations of maps and


chemical compounds in the nineteenth century, and emerged as a systematic
area of study in the twentieth century, first as a branch of mathematics and later
also through its applications to computer science. The books by Berge (1976),
Bollobas (1998), and Diestel (2000) provide substantial further coverage of
graph theory. Recently, extensive data has become available for studying large
networks that arise in the physical, biological, and social sciences, and there
has been interest in understanding properties of networks that span all these
different domains. The books by Barabasi (2002) and Watts (2002) discuss this
emerging area of research, with presentations aimed at a general audience.
The basic graph traversal techniques covered in this chapter have numerous applications. We will see a number of these in subsequent chapters, and
we refer the reader to the book by Tarjan (1983) for further results.
Notes on the Exercises Exercise 12 is based on a result of Martin Golumbic
and Ron Shamir.

113

Greedy A~gorithms

In Wall Street, that iconic movie of the 1980s, Michael Douglas gets up in
front of a room full of stockholders and proclaims, "Greed... is good. Greed
is right. Greed works." In this chapter, well be taking a much more understated
perspective as we investigate the pros and cons of short-sighted greed in the
design of algorithms. Indeed, our aim is to approach a number of different
computational problems with a recurring set of questions: Is greed good? Does
greed work?
It is hard, if not impossible, to define precisely what is meant by a greedy
algorithm. An algorithm is greedy if it builds up a solution in sma!l steps,
choosing a decision at each step myopically to optimize some underlying
criterion. One can often design many different greedy algorithms for the same
problem, each one locally, incrementally optimizing some different measure
on its way to a solution.
When a greedy algorithm succeeds in solving a nontrivial problem optimally, it typically implies something interesting and useful about the structure
of the problem itself; there is a local decision role that one can use to construct optimal solutions. And as well see later, in Chapter 11, the same is true
of problems in which a greedy algorithm can produce a solution that is guaranteed to be close to optimal, even if it does not achieve the precise optimum.
These are the kinds of issues well be dealing with in this chapter. Its easy to
invent greedy algorithms for almost any problem; finding cases in which they
work well, and proving that they work well, is the interesting challenge.
The first two sections of this chapter will develop two basic methods for
proving that a greedy algorithm produces an-optimal solution to a problem.
One can view the first approach as establishing that the greedy algorithm stays
ahead. By this we mean that if one measures the greedy algorithms progress

116

Chapter 4 Greedy Algorithms

in a step-by-step fashion, one sees that it does better than any other algorithm
at each step; it then follows that it produces an optimal solution. The second
approach is known as an exchange argument, and it is more general: one
considers any possible solution to the problem and gradually transforms it
into the solution found by the greedy algorithm without hurting its quality.
Again, it will follow that the greedy algorithm must have found a solution that
is at least as good as any other solution.
Following our introduction of these two styles of analysis, we focus on
several of the most well-known applications of greedy algorithms: shortest
paths in a graph, the Minimum Spanning Tree Problem, and the construction of Huff-man codes for performing data compression. They each provide
nice examples of our analysis techniques. We also explore an interesting relationship between minimum spanning trees and the long-studied problem of
clustering. Finally, we consider a more complex application, the MinimumCost Arborescence Problem, which further extends our notion of what a greedy
algorithm is.

4.1 Interval Scheduling: The Greedy Algorithm


Stays Ahead
Lets recall the Interval Scheduling Problem, which was the first of the five
representative problems we considered in Chapter 1. We have a set of requests
{1, 2 ..... n}; the ith request corresponds to an interval of time starting at s(i)
and finishing at f(i). (Note that we are slightly changing the notation from
Section 1.2, where we used si rather than s(i) and fi rather than f(i). This
change of notation will make things easier to talk about in the proofs.) Well
say that a subset of the requests is compatible if no two of them overlap in time,
and our goal is to accept as large a compatible subset as possible. Compatible
sets of maximum size will be called optimaL.

4.1 Interval Scheduling: The Greedy Mgorithm Stays Ahead

The most obvious rule might be to always select the available request
that starts earliest--that is, the one with minimal start time s(i). This
way our resource starts being used as quickly as possible.
This method does not yield an optimal solution. If the earliest request
i is for a very long interval, then by accepting request i we may have to
reject a lot of requests for shorter time intervals. Since our goal is to satisfy
as many requests as possible, we will end up with a suboptimal solution.
In a really bad case--say, when the finish time f(i) is the maximum
among al! requests--the accepted request i keeps our resource occupied
for the whole time. In this case our greedy method would accept a single
request, while the optimal solution could accept many. Such a situation
is depicted in Figure 4.1 (a).
This might suggest that we should start out by accepting the request that
requires the smallest interval of time--namely, the request for which
f(i)- s(i) is as small as possible. As it turns out, this is a somewhat
better rule than the previous one, but it still can produce a suboptimal
schedule. For example, in Figure 4.!(b), accepting the short interval in
the middle would prevent us from accepting the other two, which form
an optimal solution.

(a)
l

~ Designing a Greedy Algorithm


Using the Interval Scheduling Problem, we can make our discussion of greedy
algorithms much more concrete. The basic idea in a greedy algorithm for
interval scheduling is to use a simple rule to select a first request il. Once
a request il is accepted, we reject all requests that are not compatible with
We then select the next request i2 tO be accepted, and again reject all requests
that are not compatible with i2. We continue in this fashion until we run out
of requests. The cha!lenge in designing a good greedy algorithm is in deciding
which simple Me to use for the selection--and there are many natural rules
for this problem that do not give good solutions.
Lets try to think of some of the most natural rules and see how they work.

lI

(c)
Figure 4.1 Some instances of the Interval Scheduling Problem on which natural greedy
algorithms fail to find the optimal solution. In (a), it does not work to select the interval
that starts earliest; in (b), it does not work to select the shortest interval; and in (c), it
does not work to select the interval with the fewest conflicts.

117

118

Chapter 4 Greedy Mgofithms

o In the previous greedy rule, our problem was that the second request
competes with both the first and the third--that is, accepting this request
made us reject two other requests. We could design a greedy algorithm
that is based on this idea: for each request, we count the number of
other requests that are not compatible, and accept the request that has
the fewest number of noncompatible requests. (In other words, we select
the interval with the fewest "conflicts.") This greedy choice would lead
to the optimum solution in the previous example. In fact, it is quite a
bit harder to design a bad example for this rule; but it can be done, and
weve drawn an example in Figure 4.1 (c). The unique optimal solution
in this example is to accept the four requests in the top row. The greedy
method suggested here accepts the middle request in the second row and
thereby ensures a solution of size no greater than three.
A greedy rule that does lead to the optimal solution is based on a fourth
idea: we should accept first the request that finishes first, that is, the reques~t i
for which f(i) is as small as possible. This is also quite a natural idea: we ensure
that our resource becomes free as soon as possible while still satisfying one
request. In this way we can maximize the time left to satisfy other requests.
Let us state the algorithm a bit more formally. We will use R to denote
the set of requests that we have neither accepted nor rejected yet, and use A
to denote the set of accepted requests. For an example of how the algorithm
runs, see Figure 4.2.
Initially let R be the set of all requests, and let A be empty
While R is n6t yet empty
Choose a request i ~R that has the smallest finishing time
Add request i to A
Delete all requests from R that axe not compatible with request i
EndWhile
Keturn the set A as the set of accepted requests

~ Analyzing the Algorithm


While this greedy method is quite natural, it is certainly not obvious that it
returns an optimal set of intervals. Indeed, it would only be sensible to reserve
judgment on its optimality: the ideas that led to the previous nonoptimal
versions of the greedy method also seemed promising at first.
As a start, we can immediately declare that the intervals in the set A
returned by the algorithm are all compatible.
(4.1) A is a compatible set of requests.

4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead

8
1

Intervals numbered in order

7
8

Selecting interval 1

Selecting interval 3

--

!
4

3t

5I
I

Selecting interval 5

3I

5I

9
7

9~
8
t------4

9I

8
Selecting interval

Figure 4.2 Sample run of the Interval Scheduling Algorithm. At each step the selected
intervals are darker lines, and the intervals deleted at the corresponding step are
indicated with dashed lines.

What we need to show is that this solution is optimal. So, for purposes of
comparison, let (9 be an optimal set of intervals. Ideally one might want to show
that A = (9, but this is too much to ask: there may be many optimal solutions,
and at best A is equal to a single one of them. So instead we will simply show
that ]A] = 1(91, that is, that A contains the same number of intervals as (9 and
hence is also an optimal solution.
The idea underlying the proof, as we suggested initially, wil! be to find
a sense inwhich our greedy algorithm "stays ahead" of this solution (9. We
will compare the partial solutions that the greedy algorithm constructs to initial
segments of the solution (9, and show that the greedy algorithm is doing better
in a step-by-step fashion.
We introduce some notation to help with this proof. Let il ..... ik be the set
of requests in A in the order they were added to A. Note that IAI --- k. Similarly,
let the set of requests in (9 be denoted by jl ..... Jrn. Our goal is to prove that
k = m. Assume that the requests in (9 are also ordered in the natural left-toright order of the corresponding intervals, that is, in the order of the start and
finish points. Note that the requests in (9 are compatible, which implies that
the start points have the same order as the finish points.

119

120

4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead

Chapter 4 Greedy Algorithms

I~

C~an the greedy algorithms


interval really finish later?)

lr-1

ir ?

I
Jr-1

Figure 4.3 The inductive step in the proof that the greedy algorithm stays ahead.

Our intuition for the greedy method came from wanting our resource to
become flee again as soon as possible after satisfying the first request. And
indeed, our greedy rule guarantees that f(i1) < f(Jl). This is the sense in which
we want to show that our greedy rule "stays ahead"--that each of its intervals
finishes at least as soon as the corresponding interval in the set O. Thus we now
prove that for each r >_ !, the rth accepted request in the algorithms schedule
finishes no later than the rth request in the optimal schedule.
(4.2)

For all indices r < k rye have f (ir) <_ [(Jr)-

Proof. We will prove this statement by induction. For r = 1 the statement is


clearly true: the algorithm starts by selecting the request i1 with minimum
finish time.
Now let r > 1. We will assume as our induction hypothesis that the
statement is true for r- 1, and we will try to prove it for r. As shown in
Figure 4.3, the induction hypothesis lets us assume that f(ir_1) _< f(Jr-1)- In
order for the algorithms rth interval not to finish earlier as well, it would
need to "fall behind" as shown. But theres a simple reason why this could
not happen: rather than choose a later-finishing interval, the greedy algorithm
always has the option (at worst) of choosing jr and thus fulfilling the induction
step.
We can make this argument precise as follows. We know (since (9 consists
of compatible intervals) that f(Jr-1) -< s(Jr). Combining this with the induction
hypothesis f(ir_1) < f(jr-1), we get f(ir_1) < s(Jr). Thus the interval Jr is in the
set R of available intervals at the time when the greedy algorithm selects
The greedy algorithm selects the available interval with smallest finish time;
since interval Jr is one of these available intervals, we have f(ir) < f(Jr). This
completes the induction step. z
Thus we have formalized the sense in which the greedy algorithm is
remaining ahead of (9: for each r, the rth interval it selects finishes at least
as soon as the rth interval in (9. We now see why this implies the optimality
of the greedy algorithms set A.

(4.3)

The greedy algorithm returns an optimal set A.

Proof. We will prove the statement by contradiction. If A is not optimal, then


an optimal set (9 must have more requests, that is, we must have m > k.
Applying (4.2) with r = k, we get that f(ik) < f(Jk). Since m > k, there is a
request Jk+~ in (9. This request starts after request jk ends, and hence after
ik ends. So after deleting all requests that are not compatible with requests
ik, the set of possible requests R still contains Jk+l. But the greedy
algorithm stops with request ik, and it is only supposed to stop when R is
empty--a contradiction. []
Implementation and Running Time We can make our algorithm run in time
O(n log n) as follows. We begin by sorting the n requests in order of finishing
time and labeling them in this order; that is, we will assume that f(i) < f(j)
when i <j. This takes time O(n log n). In an additional O(n) time, we construct
an array S[1... n] with the property that S[i] contains the value s(i).
We now select requests by processing the intervals in order of increasing
f(i). We always select the first interval; we then iterate through the intervals in
order until reaching the first interval ] for which s(j) > f(1); we then select this
one as well. More generally, if the most recent interval weve selected ends
at time f, we continue iterating through subsequent intervals until we reach
the first ] for which s(J) _> f. In this way, we implement the greedy algorithm
analyzed above in one pass through the intervals, spending constant time per
interval. Thus this part of the algorithm takes time O(n).

Extensions
The Interval Scheduling Problem we considered here is a quite simple scheduling problem. There are many further complications that could arise in practical
settings. The following point out issues that we will see later in the book in
various forms.

In defining the problem, we assumed that all requests were known to


the scheduling algorithm when it was choosing the compatible subset.
It would also be natural, of course, to think about the version of the
problem in which the scheduler needs to make decisions about accepting
or rejecting certain requests before knowing about the full set of requests.
Customers (requestors) may well be impatient, and they may give up
and leave if the scheduler waits too long to gather information about all
other requests. An active area of research is concerned with such online algorithms, which must make decisions as time proceeds, without
knowledge of future input.

121

122

4.1. Interval Scheduling: The Greedy Algorithm Stays Ahead

Chapter 4 Greedy Algorithms

Our goal was to maximize the number of satisfied requests. But we could
picture a situation in which each request has a different value to us. For
example, each request i could also have a value vi (the amount gained
by satisfying request i), and the goal would be to maximize our income:
the sum of the values of all satisfied requests. This leads to the Weighted
Interval Scheduling Problem, the second of the representative problems
we described in Chapter 1.
There are many other variants and combinations that can arise. We now
discuss one of these further variants in more detail, since it forms another case
in which a greedy algorithm can be used to produce an optimal solution.
A Related Problem: Scheduling All Intervals
The Problem In the Interval Scheduling Problem, there is a single resource
and many requests in the form of time intervals, so we must choose which
requests to accept and which to reject. A related problem arises if we ha{re
many identical resources available and we wish to schedule all the requests
using as few resources as possible. Because the goal here is to partition
all intervals across multiple resources, we will refer to this as the Interval
Partitioning Problem)
For example, suppose that each request corresponds to a lecture that needs
to be scheduled in a classroom for a particular interval of time. We wish to
satisfy a!l these requests, using as few classrooms as possible. The classrooms
at our disposal are thus the multiple resources, and the basic constraint is that
any two lectures that overlap in time must be scheduled in different classrooms.
Equivalently, the interval requests could be iobs that need to be processed for
a specific period of time, and the resources are machines capable of handling
these jobs. Much later in the book, in Chapter 10, we will see a different
application of this problem in which the intervals are routing requests that
need to be allocated bandwidth on a fiber-optic cable.
As an illustration of the problem, consider the sample instance in Figure 4.4(a). The requests in this example can all be scheduled using three
resources; this is indicated in Figure 4.4(b), where the requests are rearranged
into three rows, each containing a set of nonoverlapping intervals. In general,
one can imagine a solution using k resources as a rearrangement of the requests
into k rows of nonoverlapping intervals: the first row contains all the intervals

1 The problem is also referred to as the Interval Coloring Problem; the terminology arises from
thinking of the different resources as having distinct colors--al~ the intervals assigned to a particular
resource are given the corresponding color.

c
I
b

e
d
I

lj
l

(a)

II

Figure 4.4 (a) An instance of the Interval Partitioning Problem with ten intervals (a
through j). (b) A solution in which all intervals are scheduled using three resources:
each row represents a set of intervals that can all be scheduled on a single resource.

assigned to the first resource, the second row contains all those assigned to
the second resource, and so forth.
Now, is there any hope of using just two resources in this sample instance?
Clearly the answer is no. We need at least three resources since, for example,
intervals a, b, and c all pass over a common point on the time-line, and hence
they all need to be scheduled on different resources. In fact, one can make
this last argument in general for any instance of Interval Partitioning. Suppose
we define the depth of a set of intervals to be the maximum number that pass
over any single point on the time-line. Then we claim
(4.4) In any instance of Interval Partitioning, the number of resources needed
is at least the depth of the set of intervals.
Proof. Suppose a set of intervals has depth d, and let 11 ..... Id all pass over a
common point on the time-line. Then each of these intervals must be scheduled
on a different resource, so the whole instance needs at least d resources. []
We now consider two questions, which turn out to be closely related.
First, can we design an efficient algorithm that schedules all intervals using
the minimum possible number of resources? Second, is there always a schedule
using a number of resources that is equal to the depth? In effect, a positive
answer to this second question would say that the only obstacles to partitioning
intervals are purely local--a set of intervals all piled over the same point. Its
not immediately clear that there couldnt exist other, "long-range" obstacles
that push the number of required resources even higher.

123

124

Chapter 4 Greedy Algorithms

We now design a simple greedy algorithm that schedules al! intervals


using a number of resources equal to the depth. This immediately implies the
optimality of the algorithm: in view of (4.4), no solution could use a number
of resources that is smaller than the depth. The analysis of our algorithm
wil! therefore illustrate another general approach to proving optimality: one
finds a simple, "structural" bound asserting that every possible solution must
have at least a certain value, and then one shows that the algorithm under
consideration always achieves this bound.
Designing the Algorithm Let d be the depth of the set of intervals; we show
how to assign a label to each interval, where the labels come from the set of
numbers {1, 2 ..... d}, and the assignment has the property that overlapping
intervals are labeled with different numbers. This gives the desired solution,
since we can interpret each number as the name of a resource, and the label
of each interval as the name of the resource to which it is assigned.
The algorithm we use for this is a simple one-pass greedy strategy that
orders intervals by their starting times. We go through the intervals in this
order, and try to assign to each interval we encounter a label that hasnt already
been assigned to any previous interval that overlaps it. Specifically, we have
the following description.
Sort the intervals by their start times, breaking ties arbitrarily
Let I1,12 ..... In denote the intervals in this order
For j=1,2,3 .....
For each interval Ii that precedes Ii in sorted order and overlaps it
Exclude the label of
Endfor
If there is any label from {I, 2 .....d} that has not been excluded then
Assign a nonexcluded label to
Else
Leave 11 unlabeled
Endif
Endfor

Analyzing the Algorithm We claim the following.


(4.5) If we use the greedy algorithm above, every interval will be assigned a
label, and no two overlapping intervals will receive the same label.
Proof. First lets argue that no interval ends up unlabeled. Consider one of
the intervals Ij, and suppose there are t intervals earlier in the sorted order
that overlap it. These t intervals, together with Ij, form a set of t + 1 intervals
that all pass over a common point on the time-line (namely, the start time of

4.2 Scheduling to Minimize Lateness: An Exchange Argument

/]), and so t + 1 < d. Thus t < d - !. It follows that at least one of the d labels
is not excluded by this set of t intervals, and so there is a label that can be
assigned to Ij.
Next we claim that no two overlapping intervals are assigned the same
label. Indeed, consider any two intervals I and I that overlap, and suppose I
precedes I in the sorted order. Then when I is considered by the algorithm,
I is in the set of intervals whose labels are excluded from consideration;
consequently, the algorithm will not assign to I the label that it used for I. []
The algorithm and its analysis are very simple. Essentially, if you have
d labels at your disposal, then as you sweep through the intervals from left
to right, assigning an available label to each interval you encounter, you can
never reach a point where all the labels are currently in use.
Since our algorithm is using d labels, we can use (4.4) to conclude that it
is, in fact, always using the minimum possible number of labels. We sum this
up as follows.
(4.6) The greedy algorithm above schedules every interval on a resource,
using a number of resources equal to the depth of the set of intervals. This
is the optimal number of resources needed.

4.2 Scheduling to Minimize Lateness: An Exchange


Argument
We now discuss a schedt~ng problem related to the one with which we began
the chapter, Despite the similarities in the problem formulation and in the
greedy algorithm to solve it, the proof that this algorithm is optimal will require
a more sophisticated kind of analysis.
~ The Problem
Consider again a situation in which we have a single resource and a set of n
requests to use the resource for an interval of time. Assume that the resource is
available starting at time s. In contrast to the previous problem, however, each
request is now more flexible. Instead of a start time and finish time, the request
i has a deadline di, and it requires a contiguous time interval of length ti, but
it is willing to be scheduled at any time before the deadline. Each accepted
request must be assigned an interval of time of length t~, and different requests
must be assigned nonoveflapping intervals.
There are many objective functions we might seek to optimize when faced
with this situation, and some are computationally much more difficult than

125

126

4.2 Scheduling to Minimize Lateness: An Exchange Argument

Chapter 4 Greedy Algorithms


Length 1

Deadline 2

Deadline 4

Length 2

Deadline 6

Length 3

,ob 3 I

Solu~on: I[

"

Job 1:

Job 2:

done at
time 1

done at
time 1 + 2 = 3

Job 3:

done at
time 1 + 2 + 3 = 6

Figure 4.5 A sample instance of scheduling to minimize lateness.

others. Here we consider a very natural goal that can be optimized by a greedy.
algorithm. Suppose that we plan to satisfy each request, but we are allowed
to let certain requests run late. Thus, beginning at our overall start time s, we
will assign each request i an interval of time of length ti; let us denote this
interval by [s(i), f(/)], with f(i) = s(i) + ti. Unlike the previous problem, then,
the algorithm must actually determine a start time (and hence a finish time)
for each interval.
We say that a request i is late if it misses the deadline, that is, if f(i) > di.
The lateness of such a request i is defined to be li = f(i) - di. We wil! say that
li = 0 if request i is not late. The goal in our new optimization problem will be
to schedule all requests, using nonoverlapping intervals, so as to minimize the
maximum lateness, L = maxi li. This problem arises naturally when scheduling
jobs that need to use a single machine, and so we will refer to our requests as
jobs.
Figure 4.5 shows a sample instance of this problem, consisting of three
iobs: the first has length tl = 1 and deadline dl= 2; the second has tz = 2
and d2 = 4; and the third has t3 = 3 and d3 = 6. It is not hard to check that
scheduling the iobs in the order 1, 2, 3 incurs a maximum lateness of O.

~ Designing the Algorithm


What would a greedy algorithm for this problem look like.~ There are several
natural greedy approaches in which we look at the data (t~, di) about the jobs
and use this to order them according to some simple nile.
One approach would be to schedule the jobs in order of increasing length
o
t~, so as to get the short jobs out of the way quickly. This immediately

looks too simplistic, since it completely ignores the deadlines of the jobs.
And indeed, consider a two-job instance where the first job has t1 = 1 and
dl= 100, while the second job has t2 = 10 and d2 = 10. Then the second
job has to be started right away if we want to achieve lateness L = 0, and
scheduling the second job first is indeed the optimal solution.
o The previous example suggests that we should be concerned about jobs
whose available slack time d, - t~ is very small--theyre the ones that
need to be started with minimal delay. So a more natural greedy algorithm
would be to sort jobs in order of increasing slack di - ti.
Unfortunately, this greedy rule fails as well. Consider a two-job
instance where the first job has q = 1 and d~ = 2, while the second job has
t2 = 10 and d2 ---- !0. Sorting by increasing slack would place the second
job first in the schedule, and the first job would incur a lateness of 9. (It
finishes at time 11, nine units beyond its dead~ne.) On the other hand,
if we schedule the first job first, then it finishes on time and the second
job incurs a lateness of only 1.
There is, however, an equally basic greedy algorithm that always produces
an optimal solution. We simply sort the jobs in increasing order of their
deadJ~es d~, and schedule them in this order. (This nile is often called Earliest
Deadline First.) There is an intuitive basis to this rule: we should make sure
that jobs with earlier deadlines get completed earlier. At the same time, its a
little hard to believe that this algorithm always produces optimal solutions-specifically because it never looks at the lengths of the jobs. Earlier we were
skeptical of the approach that sorted by length on the grounds that it threw
away half the input data (i.e., the deadlines); but now were considering a
solution that throws away the other half of the data. Nevertheless, Earliest
Deadline First does produce optimal solutions, and we will now prove this.
First we specify some notation that will be useful in talking about the
algorithm. By renaming the jobs if necessary, we can assume that the jobs are
labeled in the order of their deadlines, that is, we have
dI <_ . .. <_ dn.

We will simply schedule all jobs in this order. Again, let s be the start time for
all jobs. Job 1 will start at time s = s(1) and end at time/:(1) = s(1) + tl; Job 2
wfl! start at time s(2) =/:(1) and end at time/:(2) = s(2) + t2; and so forth. We
will use/: to denote the finishing time of the last scheduled job. We write this
algorithm here.
Order the jobs in order of their deadlines
Assume for simplicity of notation that dlj ...Jdn
Initially, f= s

127

128

4.2 Scheduling to Minimize Lateness: An Exchange Argument

Chapter 4 Greedy Algorithms


Consider the jobs i=l ..... n in this order
Assign job i to the time interval from s(f)----/ to f(O=f+ti
Let f = f + ti
End
Keturn the set of scheduled intervals [s(O,/(0] for f= 1 ..... n

~ Analyzing the Algorithm


To reason about the optimality of the algorithm, we first observe that the
schedule it produces has no "gaps--times when the machine is not working
yet there are iobs left. The time that passes during a gap will be called idle
time: there is work to be done, yet for some reason the machine is sitting idle.
Not only does the schedule A produced by our algorithm have no idle time;
it is also very easy to see that there is an optimal schedule with this property.
We do not write down a proof for this.

(4.7)

There is an optimal schedule with no idle time.

Now, how can we prove that our schedule A is optimal, that is, its
maximum lateness L is as small as possible? As in previous analyses, we wi~
start by considering an optimal schedule (9. Our plan here is to gradually
modify (9, preserving its optimality at each step, but eventually transforming
it into a schedule that is identical to the schedule A found by the greedy
algorithm. We refer to this type of analysis as an exchange argument, and we
will see that it is a powerful way to think about greedy algorithms in general.
We first try characterizing schedules in the following way. We say that a
schedule A has an inversion if a job i with deadline di is scheduled before
another job j with earlier deadline d] < di. Notice that, by definition, the
schedule A produced by our algorithm has no inversions. If there are jobs
with identical deadlines then there can be many different schedules with no
inversions. However, we can show that all these schedules have the same
maximum lateness L.
(4.8) All schedules with no inversions and no idle time have the same
maximum lateness.
Proof. If two different schedules have neither inversions nor idle time, then
they might not produce exactly the same order of jobs, but they can only differ
in the order in which jobs with identical deadlines are scheduled. Consider
such a deadline d. In both schedules, the jobs with deadline d are all scheduled
consecutively (after all jobs with earlier deadlines and before all jobs with
later deadlines). Among the jobs with deadline d, the last one has the greatest
lateness, and this lateness does not depend on the order of the jobs. []

~he main step in showing the optimality of our algorithm is to establish


that there is an optimal schedule that has no inversions and no idle time.-~Fo do
this, we will start with any optimal schedule having no idle time; we will then
convert it into a schedule with no inversions without increasing its maximum
lateness. Thus the resulting schedt~ng after this conversion will be optimal
as wel!.
(4.9) There is an optimal schedule that has no inversions and no idle time.

Proof. By (4.7), there is an optimal schedule (9 with no idle time. The proof
will consist of a sequence of statements. The first of these is simple to establish.
(a) If (9 has an inversion, then there is a pair of jobs i and j such that j is
scheduled immediately after i and has d] < di.
Ihdeed, consider an inversion in which a iob a is scheduled sometime before
a iob b, and da > db. If we advance in the scheduled order of jobs from a to b
one at a time, there has to come a point at which the deadline we see decreases
for the first time. This corresponds to a pair of consecutive iobs that form an
inversion.

Now suppose (9 has at least one inversion, and by (a), let i andj be a pair of
inverted requests that are consecutive in the scheduled order. We wil! decrease
the number of inversions in 0 by swapping the requests i andj in the schedule
O. The pair (i, j) formed an inversion in (9, this inversion is eliminated by the
swap, and no new inversions are created. Thus we have
(b) After swapping i and ] we get a schedule with one less inversion.
The hardest part of this proof is to argue that the inverted schedule is also
optimal.
(c) The new swapped schedule has a maximum lateness no larger than that

of O.

It is clear that if we can prove (c), then we are done.- The initial schedule 0
can have at most (~) inversions (if all pairs are inverted), and hence after at
most (~) swaps we get an optimal schedule with no inversions.
So we now conclude by proving (c), showing that b~; swapping a pair of
consecutive, inverted jobs, we do not increase the maximum lateness L of the
schedule. []
Proof of (c). We invent some notation to describe the schedule (9: assume
that eachlr.
request
scheduled
for the
intervallateness
[s(r), f(r)]
andschedule.
has
lateness
Let Lr =ismax
thetime
maximum
of this
r lr denote

129

131

Oarnly
the finishing
times.......
of i and ~
e affected
by the swap.

Before swapping:

Proof. Statement (4.9) proves that an optimal schedule with no inversions


exdsts. Now by (4.8) all schedules with no inversions have the same maximum
lateness, and so the schedule obtained by the greedy algorithm is optimal, m
(a)

After swapping:

dI

(4.10) The schedule A produced by the greedy algorithm has optimal maximum lateness L.

di

Figure 4.6 The effect of swapping two consecutive, inverted jobs.

Let ~ denote the swapped schedule; we will use ~(r), ~(r), ~r, and ~ to denote
the corresponding quantities in the swapped schedule.
Now recall our two adiacent, inverted jobs i and]. The situation is roughly
as pictured in Figure 4.6. The finishing time of] before the swap is exactly equal
to the finishing time of i after the swap. Thus all jobs other than jobs i and ]
finish at the same time in the two schedules. Moreover, job j will get finished
earlier in the new schedule, and hence the swap does not increase the lateness
of job j.
Thus the only thing to worry about is job i: its lateness may have been
increased, and what if this actually raises the maximum lateness of the
whole schedule? After the swap, job i finishes at time f(j), when job j was
finished in the schedule (9. If job i is late in this new schedule, its lateness
is ~i = ~(i) - di = f(j) -di. But the crucial point is that i cannot be more late
in the schedule -~ than j was in the schedule (9. Specifically, our assumption
d~ > dj implies that
~ = f(]) -di < f(]) - di = l~.

Since the lateness of the schedule (9 was L >_ lj > ~, this shows that the swap
does not increase the maximum lateness of the schedule. []
The optimality of our greedy algorithm now follows immediately.

Extensions
There are many possible generalizations of this scheduling problem. For example, we assumed that all jobs were available to start at the common start
time s. A natural, but harder, version of this problem would contain requests i
that, in addition to the deadline dz and the requested time t~, would also have
an earliest possible starting time ri. This earliest possible starting time is usually referred to as the release time. Problems with release times arise natura!ly
in scheduling problems where requests can take the form: Can I reserve the
room for a two-hour lecture, sometime between 1 P.M. and 5 P.M.? Our proof
that the greedy algorithm finds an optimal solution relied crucially on the fact
that all jobs were available at the common start time s. (Do you see where?)
Unfortunately, as we wi!l see later in the book, in Chapter 8, this more general
version of the problem is much more difficult to solve optimally.

4.3 Optimal Caching: A More Complex Exchange


Argument
We now consider a problem that involves processing a sequence of requests
of a different form, and we develop an aig~~rithm whose analysis requires
a more subtle use of the exchange argumem. The problem is that of cache
maintenance.

~.~ The Problem


To motivate caching, consider the following situation. Youre working on a
long research paper, and your draconian library will only allow you to have
eight books checked out at once. You know that youll probably need more
than this over the course of working on the paper, but dt any point in time,
youd like to have ready access to the eight books that are most relevant at
that tirng. How should you decide which books to check out, and when should
you return some in exchange for others, to minimize the number of times you
have to exchange a book at the library?
This is precisely the problem that arises when dealing with a memory
hierarchy: There is a small amount of data that can be accessed very quickly,

132

Chapter 4 Greedy Algorithms

and a large amount of data that requires more time to access; and you must
decide which pieces of data to have close at hand.
Memory hierarchies have been a ubiquitous feature of computers since
very early in their history. To begin with, data in the main memory of a
processor can be accessed much more quickly than the data on its hard disk;
but the disk has much more storage capaciW. Thus, it is important to keep
the most regularly used pieces o~ data in main memory, and go to disk as
infrequently as possible. The same phenomenon, qualitatively, occurs with
on-chip caches in modern processors. These can be accessed in a few cycles,
and so data can be retrieved from cache much more quickly than it can be
retrieved from main memory. This is another level of hierarchy: smal! caches
have faster access time than main memory, which in turn is smaller and faster
to access than disk. And one can see extensions of this hierarchy in many
other settings. When one uses a Web browser, the disk often acts as a cache
for frequently visited Web pages, since going to disk is stil! much faster than
downloading something over the Internet.
Caching is a general term for the process of storing a small amount of dat~
in a fast memory so as to reduce the amount of time spent interacting with a
slow memory. In the previous examples, the on-chip cache reduces the need
to fetch data from main memory, the main memory acts as a cache for the
disk, and the disk acts as a cache for the Internet. (Much as your desk acts as
a cache for the campus library, and the assorted facts youre able to remember
without looMng them up constitute a cache for the books on your desk.)
For caching to be as effective as possible, it should generally be the case
that when you go to access a piece of data, it is already in the cache. To achieve
this, a cache maintenance algorithm determines what to keep in the cache and
what to evict from the cache when new data needs to be brought in.
Of course, as the caching problem arises in different settings, it involves
various different considerations based on the underlying technologY. For our
purposes here, though, we take an abstract view of the problem that underlies
most of these settings. We consider a set U of n pieces of data stored in main
memory. We also have a faster memory, the cache, that can hold k < n pieces
of data at any one time. We will assume that the cache initially holds some
set of k items. A sequence of data items D = dl, d2 ..... dm drawn from U is
presented to us--this is the sequence of memory references we must process-and in processing them we must decide at all times which k items to keep in the
cache. When item di is presented, we can access it very quickly if it is already
in the cache; otherwise, we are required to bring it from main memory into
the cache and, if the cache is ful!, to evict some other piece of data that is
currently in the cache to make room for di. This is called a cache miss, and we
want to have as few of these as possible.

4.3 Optimal Caching: A More Complex Exchange Argument

Thus, on a particular sequence of memory references, a cache maintenance algorithm determines an eviction schedule--specifying which items
should be evicted from the cache at which points in the sequence--and t_his
determines the contents of the cache and the number of misses over time. Lets
consider an example of this process.
Suppose we have three items [a, b, c], the cache size is k = 2, and we
are presented with the sequence
a,b,c,b,c,a,b.
Suppose that the cache initially contains the items a and b. Then on the
third item in the sequence, we could evict a so as to bring in c; and
on the sixth item we could evict c so as to bring in a; we thereby incur
two cache misses over the whole sequence. After thinking about it, one
concludes that any eviction schedule for this sequence must include at
least two cache misses.
Under real operating conditions, cache maintenance algorithms must
process memory references dl, d2 .... without knowledge of whats coming
in the future; but for purposes of evaluating the quality of these algorithms,
systems researchers very early on sought to understand the nature of the
optimal solution to the caching problem. Given a fifll sequence S of memory
references, what is the eviction schedule that incurs as few cache misses as
possible?

~ Designing and Analyzing the Algorithm


In the 1960s, Les Belady showed that the following simple rule will always
incur the minimum number of misses:
When di needs to be brought into the cache,
evict the item that is needed the farthest into the future

We will call this the Farthest-in-Future Algorithm. When it is time to evict


something, we look at the next time that each item in the cachewill be
referenced, and choose the one for which this is as late as possible.
This is a very natural algorithm. At the same time, the fact that it is optimal
on all sequences is somewhat more subtle than it first appears. Why evict the
item that is needed farthest in the future, as opposed, for example, to the one
that will be used least frequently in the future? Moreover, consider a sequence
like
a,b,c,d,a,d,e,a,d,b,c

133

134

Chapter 4 Greedy Algorithms

with k = 3 and items {a, b, c} initially in the cache. The Farthest-in-Future rule
will produce a schedule S that evicts c on the fourth step and b on the seventh
step. But there are other eviction schedules that are just as good. Consider
the schedule S that evicts b on the fourth step and c on the seventh step,
incurring the same number of misses. So in fact its easy to find cases where
schedules produced by rules other than Farthest-in-Future are also optimal;
and given this flexibility, why might a deviation from Farthest-in-Future early
on not yield an actual savings farther along in the sequence.~ For example, on
the seventh step in our example, the schedule S is actually evicting an item
(c) that is needed farther into the future than the item evicted at this point by
Farthest-in-Future, since Farthest-in-Future gave up c earlier on.
These are some of the kinds of things one should worry about before
concluding that Farthest-in-Future really is optimal. In thinking about the
example above, we quickly appreciate that it doesnt really matter whether
b or c is evicted at the fourth step, since the other one should be evicted at
the seventh step; so given a schedule where b is evicted first, we can swap
the choices of b and c without changing the cost. This reasoning--swapping,
one decision for another--forms the first outline of an exchange argument that
proves the optimality of Farthest-in-Future.
Before delving into this analysis, lets clear up one important issue. Al!
the cache maintenance algorithms weve been considering so far produce
schedules that only bring an item d into the cache .~n a step i if there is a
request to d in step i, and d is not already in the cache. Let us ca!l such a
schedule reduced--it does the minimal amount of work necessary in a given
step. But in general one could imagine an algorithm that produced schedules
that are not reduced, by bringing in items in steps when they are not requested.
We now show that for every nonreduced schedule, there is an equally good
reduced schedule.
Let S be a schedule that may not be reduced. We define a new schedule
-~--the reduction of S--as follows. In any step i where S brings in an item d
that has not been requested, our construction of ~ "pretends" to do this but
actually leaves d in main memory. It only really brings d into the cache in
the next step j after this in which d is requested. In this way, the cache miss
incurred by ~ in step j can be charged to the earlier cache operation performed
by S in step i, when it brought in d. Hence we have the following fact.
(4.11) -~ is a reduced schedule that brings in at most as many items as the
schedule S.
Note that for any reduced schedule, the number of items that are brought
in is exactly the number of misses.

4.3 Optimal Caching: A More Complex Exchange Argument

Proving the Optimalthy of Farthest-in-Future We now proceed with the


exchange argument showing that Farthest-in-Future is optimal. Consider an
arbitrary sequence D of memory references; let S~F denote the schedule
produced by Farthest-in-Future, and let S* denote a schedule that incurs the
minimum possible number of misses. We will now gradually "transform" the
schedule S* into the schedule SEE, one eviction decision at a time, without
increasing the number of misses.
Here is the basic fact we use to perform one step in the transformation.
(4.12) Let S be a reduced scheduIe that makes the same eviction deasions
as SEE through the first j items in the sequence, for a number j. Then there is a
reduced schedule S that makes the same eviction decisions as SEE through the
first ] + 1 items, and incurs no more misses than S does.
Proof. Consider the (j + 1)st request, to item d = dy+l. Since S and SEE have
agreed up to this point, they have the same cache contents. If d is in the cache
for both, then no eviction decision is necessary (both schedules are reduced),
and so S in fact agrees with SEE through step j + 1, and we can set S = S.
Similarly, if d needs to be brought into the cache, but S and SEE both evict the
same item to make room for d, then we can again set S = S.
So the interesting case arises when d needs to be brought into the cache,
and to do this S evicts item f while SEE evicts item e ~ f. Here S and SEE do
not already agree through step j + ! since S has e in cache while SEE has f in
cache. Hence we must actually do something nontrivial to construct S.
As a first step, we should have S evict e rather than f. Now we need to
further ensure that S incurs no more misses than S. An easy way to do this
would be to have S agree with S for the remainder of the sequence; but this
is no longer possible, since S and S have slightly different caches from this
point onward. So instead well have S try to get its cache back to the same
state as S as quickly as possible, while not incurring unnecessary misses. Once
the caches are the same, we can finish the construction of S by just having it
behave like S.
Specifically, from request j + 2 onward, S behaves exactly like S until one
of the following things happens for the first time.
(i) There is a request to an item g ~ e, f that is not in the cache of S, and S
evicts e to make room for it. Since S and S only differ on e and f, it must
be that g is not in the cache of S either; so we can have S evict f, and
now the caches of S and S are the same. We can then have S behave
exactly like S for the rest of the sequence.
(ii) There is a request to f, and S evicts an item e. If e = e, then were all
set: S can simply access f from the cache, and after this step the caches

135

136

4.4 Shortest Paths in a Graph

Chapter 4 Greedy Algorithms

" e as wel!, and


of S and S will be the same. If e ~ e, then we have S evict
bring in e from main memory; this too results in S and S having the same
caches. However, we must be careful here, since S is no longer a reduced
schedule: it brought in e when it wasnt immediately needed. So to finish
this part of the construction, we further transform S to its reduction S
using (4.11); this doesnt increase the number of items brought in by S,
and it still agrees with SFF through step j + 1.
Hence, in both these cases, we have a new reduced schedule S that agrees
with SFF through the first j + 1 items and incurs no more misses than S does.
And crucially--here is where we use the defining property of the Farthest-inFuture Algorithm--one of these two cases will arise before there is a reference
to e. This is because in step j + 1, Farthest-in-Future evicted the item (e) that
would be needed farthest in the future; so before there could be a request to
e, there would have to be a request to f, and then case (ii) above would apply.

Using this result, it is easy to complete the proof of optimality. We begin


with an optima! schedule S*, and use (4.12) to construct a schedule $1 that
agrees with SFF through the first step. We continue applying (4.12) inductively
for j = 1, 2, 3 ..... m, producing schedules Sj that agree with S~F through the
first j steps. Each schedule incurs no more misses than the previous one; and
by definition Sm= Sw, since it agrees with it through the whole sequence.
Thus we have
(4.13) S~F incurs no more misses than any other schedule S* and hence is
optimal.
Extensions: Caching under Real Operating Conditions
As mentioned in the previous subsection, Beladys optimal algorithm provides
a benchmark for caching performance; but in applications, one generally must
make eviction decisions on the fly without knowledge of future requests.
Experimentally, the best caching algorithms under this requirement seem to be
variants of the Least-Recently-Used (LRU) Principle, which proposes evicting
the item from the cache that was referenced longest ago.
If one thinks about it, this is just Beladys Algorithm with the direction
of time reversed--longest in the past rather than farthest in the future. It is
effective because applications generally exhibit locality of reference: a running
program will generally keep accessing the things it has just been accessing.
(It is easy to invent pathological exceptions to this principle, but these are
relatively rare in practice.) Thus one wants to keep the more recently referenced
items in the cache.

Long after the adoption of LRU in pradtice, Sleator and Tartan showed that
one could actually provide some theoretical analysis of the performance of
LRU, bounding the number of misses it incurs relative to Farthest-in-Future.
We will discuss this analysis, as well as the analysis of a randomized variant
on LRU, when we return to the caching problem in Chapter 13.

4.4 Shortest Paths in a Graph


Some of the basic algorithms for graphs are based on greedy design principles.
Here we apply a greedy algorithm to the problem of finding shortest paths, and
in the next section we look at the construction of minimum-cost spanning trees.
zJ The Problem
As weve seen, graphs are often used to model networks in which one travels from one point to another--traversing a sequence of highways through
interchanges, or traversing a sequence of communication links through intermediate touters. As a result, a basic algorithmic problem is to determine the
shortest path between nodes in a graph. We may ask this as a point-to-point
question: Given nodes u and v, what is the shortest u-v path? Or we may ask
for more information: Given a start node s, what is the shortest path from s to
each other node?

The concrete setup of the shortest paths problem is as follows. We are


given a directed graph G = (V, E), with a designated start node s. We assume
that s has a path to every other node in G. Each edge e has a length ~e -> 0,
indicating the time (or distance, or cost) it takes to traverse e. For a path P,
the length of P--denoted g(P)--is the sum of the lengths of all edges in P.
Our goal is to determine the shortest path from s to every other node in the
graph. We should mention that although the problem is specified for a dkected
graph, we can handle the case of an undirected graph by simply replacing each
undirected edge e = (u, v) of length ~e by two directed edges (u, v) and (u, u),
each of length ge-

~ Designing the Algorithm


In 1959, Edsger Dijkstra proposed a very simple greedy algorithm to solve the
single-source shortest-paths problem. We begin by describing an algorithm that
just determines the length of the shortest path from s to each other node in the
graph; it is then easy to produce the paths as well. The algorithm maintains a
set S of vertices u for which we have determined a shortest-path distance d(u)
from s; this is the "explored" part of the graph. Initially S = {s}, and d(s) = O.
Now, for each node v ~ V-S, we determine the shortest path that can be
constructed by traveling along a path through the explored part S to some
u ~ $, followed by the single edge (u, v). That is, we consider the quantity

137

138

4.4 Shortest Paths in a Graph

Chapter 4 Greedy Algorithms

d(v) = mine=(a,v):a~s d(a) + ~e. We choose the node v e V-S for which t~s
quantity is minimized, add v to S, and define d(v) to be the value d(v).

Dijkstras Algorithm (G, ~)


Let S be the set of explored nodes
For each ueS, we store a distsnce d(u)
Initially S = Is} and d(s) = 0
While S ~ V
Select a node u ~S with at least one edge from S for which
d(u) = nfine=(u,v):u~s d(u) + ~-e is as small as possible
Add u to S and define d(u)=d(u)
EndWhile

It is simple to produce the s-u paths corresponding to the distances found


by Dijkstras Algorithm. As each node v is added to the set S, we simply record
the edge (a, v) on which it achieved the value rnine=(a,v):ues d(u) + e. The
path Pv is implicitly represented by these edges: if (u, v) is the edge we have
stored for v, then P~ is just (recursively) the path P~ followed by the single
edge (u, ~). In other words, to construct P~, we simply start at 12; follow the
edge we have stored for v in the reverse direction to a; then follow the edge we
have stored for a in the reverse direction to its predecessor; and so on until we
reach s. Note that s must be reached, since our backward walk from 12 visits
nodes that were added to S earlier and earlier.
To get a better sense of what the algorithm is doing, consider the snapshot
of its execution depicted in Figure 4.7. At the point the picture is drawn, two
iterations have been performed: the first added node u, and the second added
node 12. In the iteration that is about to be performed, the node x wil! be added
because it achieves the smallest value of d(x); thanks to the edge (u, x), we
have d(x) = d(a) + lax = 2. Note that attempting to add y or z to the set S at
this point would lead to an incorrect value for their shortest-path distances;
ultimately, they will be added because of their edges from x.

~ Analyzing the Algorithm


We see in this example that Dijkstras Algorithm is doing the fight thing and
avoiding recurring pitfalls: growing the set S by the wrong node can lead to an
overestimate of the shortest-path distance to that node. The question becomes:
Is it always true that when Dijkstras Algorithm adds a node v, we get the true
shortest-path distance to 127.
We now answer this by proving the correctness of the algorithm, showing
that the paths Pa really are shortest paths. Dijkstras Algorithm is greedy in

Set S: ~
nodes already
explored

Figure 4.7 A snapshot of the execution of Dijkstras Algorithm. The next node that will
be added to the set S is x, due to the path through u.

the sense that we always form the shortest new s-12 path we can make from a
path in S followed by a single edge. We prove its correctness using a variant of
our first style of analysis: we show that it "stays ahead" of all other solutions
by establishing, inductively, that each time it selects a path to a node 12, that
path is shorter than every other possible path to v.
(4.14) Consider the set S at any point in the algorithms execution. For each
u ~ S, the path Pu is a shortest s-u path.
:~
Note that this fact immediately establishes the correctness of Dijkstras
Mgofithm, since we can apply it when the algorithm terminates, at which
point S includes all nodes.
Proof. We prove this by induction on the size of S. The case IS] = 1 is easy,
since then we have S = {s] and d(s) = 0. Suppose the claim holds when IS] = k
for some value of k > 1; we now grow S to size k + 1 by adding the node 12.
Let (u, 12) be the final edge on our s-12 path P~.
By induction hypothesis, Pu is the shortest s-u path for each u ~ S. Now
consider any other s-12 path P; we wish to show that it is at least as long as P~.
In order to reach ~, this path P must leave the set S sornetuhere; let y be the
first node on P that is not in S, and let x ~ S be the node just before y.
The situation is now as depicted in Figure 4.8, and the crux of the proof
is very simple: P cannot be shorter than P~ because it is already at least as

139

140

4.4 Shortest Paths in a Graph

Chapter 4 Greedy Algorithms

lth

The alternate s-v path P through~


x and y is already too long by |
e time it has left the set S, )

of wavefront reaches nodes in increasing order of their distance from s. It is


easy to believe (and also true) that the path taken by the wavefront to get to
any node u is a shortest path. Indeed, it is easy to see that this is exactly the
path to v found by Dijkstras Algorithm, and that the nodes are discovered by
the expanding water in the same order that they are discovered by Dijkstras
Algorithm.

Set S

Figure 4.8 The shortest path Pv and an alternate s-v path P through the node

long as Pv by the time it has left the set S. Indeed, in iteration k + 1, Dijkstras
Algorithm must have considered adding node y to the set S via the edge (x, y)
and rejected this option in favor of adding u. This means that there is no path
from s to y through x that is shorter than Pv- But the subpath of P up to y is
such a path, and so this subpath is at least as long as P,. Since edge length~
are nonnegative, the full path P is at least as long as P, as well.
This is a complete proof; one can also spell out the argument in the
previous paragraph using the following inequalities. Let P be the Subpath
of P from s to x. Since x ~ S, we know by the induction hypothesis that Px is a
shortest s-x path (of length d(x)), and so g(P) > g(Px) = d(x). Thus the subpath
of P out to node y has length ~(P) + g(x, y) > d(x) + g.(x, y) > d(y), and the
full path P is at least as long as this subpath. Finally, since Dijkstras Algorithm
selected u in this iteration, we know that d(y) >_ d(u) = g(Pv). Combining these
inequalities shows that g(P) >_ ~(P) + ~.(x, y) >_ g(P~). ~
Here are two observations about Dijkstras Algorithm and its analysis.
First, the algorithm does not always find shortest paths if some of the edges
can have negative lengths. (Do you see where the proof breaks?) Many
shortest-path applications involve negative edge lengths, and a more complex algorithm--due to Bellman and Ford--is required for this case. We will
see this algorithm when we consider the topic of dynamic programming.
The second observation is that Dijkstras Algorithm is, in a sense, even
simpler than weve described here. Dijkstras Algorithm is really a "continuous" version of the standard breadth-first search algorithm for traversing a
graph, and it can be motivated by the following physical intuition. Suppose
the edges of G formed a system of pipes filled with water, joined together at
the nodes; each edge e has length ge and a fixed cross-sectional area. Now
suppose an extra droplet of water falls at node s and starts a wave from s. As
the wave expands out of node s at a constant speed, the expanding sphere

Implementation and Running Time To conclude our discussion of Dijkstras


Algorithm, we consider its running time. There are n - 1 iterations of the
krt~].e loop for a graph with n nodes, as each iteration adds a new node v
to S. Selecting the correct node u efficiently is a more subtle issue. Ones first
impression is that each iteration would have to consider each node v ~ S,
and go through all the edges between S and u to determine the minimum
mine=(u,u):u~s d(u)+g-e, so that we can select the node v for which this
minimum is smallest. For a graph with m edges, computing all these minima
can take O(m) time, so this would lead to an implementation that runs in
O(mn) time.

We can do considerably better if we use the right data structures. First, we


will explicitly maintain the values of the minima d(u) = mJne=(u,u):u~s d(u) +
~e for each node v V - S, rather than recomputing them in each iteration.
We can further improve the efficiency by keeping the nodes V - S in a priority
queue with d(u) as their keys. Priority queues were discussed in Chapter 2;
they are data structures designed to maintain a set of n elements, each with a
key. A priority queue can efficiently insert elements, delete elements, change
an elements key, and extract the element with the minimum key. We will need
the third and fourth of the above operations: ChangeKey and Ex~cractNn.
How do we implement Dijkstras Algorithm using a priority queue? We put
the nodes V in a priority queue with d(u) as the key for u ~ V. To select the node
v that should be added to the set S, we need the Extrac~cNn operation. To see
how to update the keys, consider an iteration in which node u is added to S, and
let tv ~ S be a node that remains in the priority queue. What do we have to do
to update the value of d(w)? If (v, w) is not an edge, then we dont have to do
anything: the set of edges considered in the minimum mihe=(u,w):a~s d(u) + ~e
is exactly the same before and after adding v to S. If e = (v, w) ~ E, on
the other hand, then the new value for the key is min(d(w), d(u) + ~-e). If
d(ro) > d(u) + ~e then we need to use the ChangeKey operation to decrease
the key of node w appropriately. This ChangeKey operation can occur at most
once per edge, when the tail of the edge e is added to S. In summary, we have
the following result.

141

142

4.5 The Minimum Spanning Tree Problem

Chapter 4 Greedy Algorithms

(4,1S) Using a priority queue, Di]kstras Algorithm can be implemented on


a graph with n nodes and m edges to run in O(m) time, plus the time for n
Extrac~Min and m ChamgeKey operations.
Using the heap-based priority queue implementation discussed in Chapter 2, each priority queue operation can be made to run in O(log n) time. Thus
the overall time for the implementation is O(m log r~).

4.5 The Minimum Spanning Tree Problem


We now apply an exchange argument in the context of a second fundamental
problem on graphs: the Minimum Spanning Tree Problem.
~ The Problem
Suppose we have a set of locations V = {vl, v2 ..... vn}, and we want to build a
communication network on top of them. The network should be connected-there should be a path between every pair of nodes--but subiect to this
requirement, we wish to build it as cheaply as possible.
For certain pairs (vi, vj), we may build a direct link between vi and vj for
a certain cost c(vi, vj) > 0. Thus we can represent the set of possible links that
may be built using a graph G = (V, E), with a positive cost Ce associated with
each edge e = (vi, vj). The problem is to find a subset of the edges T_ E so
that the graph (V, T) is connected, and the total cost ~e~T Ce is as small as
possible. (We will assume that thefull graph G is connected; otherwise, no
solution is possible.)
Here is a basic observation.
(4.16) Let T be a minimum-cost solution to the network design problem
defined above. Then (V, T) is a tree.
Proof. By definition, (V, T) must be connected; we show that it also will
contain no cycles. Indeed, suppose it contained a cycle C, and let e be any
edge on C. We claim that (V, T - {e}) is still connected, since any path that
previously used the edge e can now go. "the long way" around the remainder
of the cycle C instead. It follows that (V, T - {e}) is also a valid solution to the
problem, and it is cheaper--a contradiction. "
If we allow some edges to have 0 cost (that is, we assume only that the
costs Ce are nonnegafive), then a minimum-cost solution to the network design
problem may have extra edges--edges that have 0 cost and could option!lly
be deleted. But even in this case, there is always a minimum-cost solution that
is a tree. Starting from any optimal solution, we could keep deleting edges on

cycles until we had a tree; with nonnegative edges, the cost would not increase
during this process.
We will call a subset T __c E a spanning tree of G if (V, T) is a tree. Statement
(4.16) says that the goal of our network design problem can be rephrased as
that of finding the cheapest spanning tree of the graph; for this reason, it
is generally called the Minimum Spanning Tree Problem. Unless G is a very
simple graph, it will have exponentially many different spanning trees, whose
structures may look very different from one another. So it is not at all clear
how to efficiently find the cheapest tree from among all these options.

fi Designing Algorithms
As with the previous problems weve seen, it is easy to come up with a number
of natural greedy algorithms for the problem. But curiously, and fortunately,
this is a case where many of the first greedy algorithms one tries turn out to be
correct: they each solve the problem optimally. We will review a few of these
algorithms now and then discover, via a nice pair of exchange arguments, some
of the underlying reasons for this plethora of simple, optimal algorithms.
Here are three greedy algorithms, each of which correctly finds a minimum
spanning tree.
One simple algorithm starts without any edges at all and builds a spanning tree by successively inserting edges from E in order of increasing
cost. As we move through the edges in this order, we insert each edge
e as long as it does not create a cycle when added to the edges weve
already inserted. If, on the other hand, inserting e would result in a cycle,
then we simply discard e and continue. This approach is called Kruskals
Algorithm.
Another simple greedy algorithm can be designed by analogy with Dijkstras Algorithm for paths, although, in fact, it is even simpler to specify
than Dijkstras Algorithm. We start with a root node s and try to greedily
grow a tree from s outward. At each step, we simply add the node that
can be attached as cheaply as possibly to the partial tree we already have.
More concretely, we maintain a set S _c V on which a spanning tree
has been constructed so far. Initially, S = {s}. In each iteration, we grow
S by one node, adding the node v that minimizes the "attachment cost"
mine=(u,u):u~s ce, and including the edge e = (u, v) that achieves this
minimum in the spanning tree. This approach is called Prims Algorithm.
Finally, we can design a greedy algorithm by running sort of a "backward" version of Kruskals Algorithm. Specifically, we start with the full
graph (V, E) and begin deleting edges in order of decreasing cost. As we
get to each edge e (starting from the most expensive), we delete it as

143

144

4.5 The Minimum Spanning Tree Problem

Chapter 4 Greedy Algorithms

easier to express the arguments that follow, and we will show later in this
section how this assumption can be easily eliminated.
When Is It Safe to Include an Edge in the Minimum Spanning Tree? The
crucial fact about edge insei-tion is the following statement, which we wil!
refer to as the Cut Property.
(4.17) Assumethatalledgecostsaredistinct. LetSbeanysubsetofnodesthat
is neither empty nor equal to all of V, and let edge e = (v, w) be the minimumcost edge with one end in S and the other in V- S. Then every minimum
spanning tree contains the edge e.

(b)
Figure 4.9 Sample run of the Minimum Spanning Tree Algorithms of (a) Prim and
(b) Kruskal, on the same input. The first 4 edges added to the spanning tree are indicated
by solid lines; the ne~xt edge to be added is a dashed line.

long as doing so would not actually disconnect the graph we currently


have. For want of a better name, this approach is generally called the
Reverse-Delete Algorithm (as far as we can te!l, its never been named
after a specific person).
For example, Figure 4.9 shows the first four edges added by Prims and
Kruskals Algorithms respectively, on a geometric instance of the Minimum
Spanning Tree Problem in which the cost of each edge is proportional to the
geometric distance in the plane.
The fact that each of these algorithms is guaranteed to produce an optimal solution suggests a certain "robustness" to the Minimum Spanning Tree
Problem--there are many ways to get to the answer. Next we explore some of
the underlying reasons why so many different algorithms produce minimumcost spanning trees.

f! Analyzing the Algorithms


All these algorithms work by repeatedly inserting or deleting edges from a
partial solution. So, to analyze them, it would be useful to have in hand some
basic facts saying when it is "safe" to include an edge in the minimum spanning
tree, and, correspondingly, when it is safe to eliminate an edge on the grounds
that it couldnt possibly be in the minimum spanning tree. For purposes of
the analysis, we will make the simplifying assumption that all edge costs are
distinct from one another (i.e., no two are equal). This assumption makes it

Proof. Let T be a spanning tree that does not contain e; we need to show that T
does not have the minimum possible cost. We!l do this using an exchange
argument: well identify an edge e in T that is more expensive than e, and
with the property exchanging e for e results in another spanning tree. This
resulting spanning tree will then be cheaper than T, as desired.
The crux is therefore to find an edge that can be successfully exchanged
with e. Recall that the ends of e are v and w. T is a spanning tree, so there
must be a path P in T from v to w. Starting at ~, suppose we follow the nodes
of P in sequence; there is a first node w on P that is in V - S. Let u E S be the
node just before w on P, and let e = (v, w) be the edge joining them. Thus,
e is an edge of T with one end in S and the other in V - S. See Figure 4.10 for
the situation at this stage in the proof.
If we exchange e for e, we get a set of edges T= T- (e} U {e). We
claim that T is a spanning tree. Clearly (V, T) is connected, since (V, T)
is connected, and any path in (V, T) that used the edge e = (~, w) can now
be "rerouted" in (V, T) to follow the portion of P from v to v, then the edge
e, and then the portion of P from w to w. To see that (V, T) is also acyclic,
note that the only cycle in (V, T U {e}) is the one composed of e and the path
P, and this cycle is not present in (V, T) due to the deletion of e.
We noted above that the edge e has one end in S and the other in V - S.
But e is the cheapest edge with this property, and so ce < ce,. (The inequality
is strict since no two edges have the same cost.) Thus the total cost of T is
less than that of T, as desired. ,,
The proof of (4.17) is a bit more subtle than it may first appear. To
appreciate this subtlety, consider the following shorter but incorrect argument
for (4.17). Let T be a spanning tree that does not contain e. Since T is a
spanning tree, it must contain an edge f with one end in S and the other in
V - S. Since e is the cheapest edge with this property, we have ce < cf, and
hence T - If} U {el is a spanning tree that is cheaper than T.

145

4.5 The Minimum Spanning Tree Problem

146

Chapter 4 Greedy Algorithms


So if we can show that the output (V, T) of Kruskals Algorithm is in fact
a spanning tree of G, then we will be done. Clearly (V, T) contains no cycles,
since the algorithm is explicitly designed to avoid creating cycles. Further, if
(V, T) were not connected, then there would exist a nonempty subset of nodes
S (not equal to all of V) such that there is no edge from S to V - S. But this
contradicts the behavior of the algorithm: we know that since G is connected,
there is at least one edge between S and V - S, and the algorithm will add the
first of these that it encounters. []

(4.19) Prims Algorithm produces a minimum spanning tree of G~


(e can be swapped for e.)

Figure 4.10 Swapping the edge e for the edge e in the spanning tree T, as described in
the proof of (4.17).

The problem with this argument is not in the claim that f exists, or that
T {f} U {e} is cheaper than T. The difficulty is that T - {f} U {e} may not be
a spanning tree, as shown by the example of the edge f in Figure 4.10. The
point is that we cant prove (4.17) by simply picking any edge in T that crosses
from S to V - S; some care must be taken to find the right one.
The Optimality of Kraskals and Prims Algorithms We can now easily
prove the optimality of both Kruskals Algorithm and Pfims Algorithm. The
point is that both algorithms only include an edge when it is justified by the
Cut Property (4.17).
(4.18) Kruskals Algorithm produces a minimum spanning tree of G.
Proof. Consider any edge e = (v, tu) added by Kruskals Algorithm, and let
S be the set of all nodes to which v has a path at the moment iust before
e is added. Clearly v ~ S, but tu S, since adding e does not create a cycle.
Moreover, no edge from S to V - S has been encountered yet, since any such
edge could have been added without creating a cycle, and hence would have
been added by Kruskals Algorithm. Thus e is the cheapest edge with one end
in S and the other in V- S, and so by (4.17) it belongs to every minimum
spanning tree.

Proof. For Prims Algorithm, it is also very easy to show that it only adds
edges belonging to every minimum spanning tree. Indeed, in each iteration of
the algorithm, there is a set S _ V on which a partial spanning tree has been
constructed, and a node u and edge e are added that minimize the quantity
mine=(u,u):u~s Ce. By definition, e is the cheapest edge with one end in S and the
other end in V - S, and so by the Cut Property (4.17) it is in every minimum
spanning tree.
It is also straightforward to show that Prims Algorithm produces a spanning tree of G, and hence it produces a minimum spanning tree. []
When Can We Guarantee an Edge Is Not in the Minimum Spanning
Tree? The crucial fact about edge deletion is the following statement, which
we wil! refer to as the Cycle Property.
(4.20) Assume that all edge costs are distinct. Let C be any cycle in G, and
let edge e = (v, w) be the most expensive edge belonging to C. Then e does not
belong to any minimum spanning tree of G.
Proof. Let T be a spanning tree that contains e; we need to show that T does
not have the minimum possible cost. By analogy with the proof of the Cut
Property (4.17), well do this with an exchange argument, swapping e for a
cheaper edge in such a way that we still have a spanning tree.
So again the question is: How do we find a cheaper edge that can be
exchanged in this way with e? Lets begin by deleting e from T; this partitions
the nodes into two components: S, containing node u; and V - S, containing
node tu. Now, the edge we use in place of e should have one end in S and the
other in V - S, so as to stitch the tree back together.
We can find such an edge by following the cycle C. The edges of C other
than e form, by definition, a path P with one end at u and the other at tu. If
we follow P from u to tu, we begin in S and end up in V - S, so there is some

147

4.5 The Minimum Spanning Tree Problem

148

Chapter 4 Greedy Algorithms


contradiction that (V, T) contains a cycle C. Consider the most expensive edge
e on C, which would be the first one encountered by the algorithm. This e.dge
should have been removed, since its removal would not have disconnected
the graph, and this contradicts the behavior of Reverse-Delete. []

~Tcan be swapped for e.)


Figure 4.11 Swapping the edge e for the edge e in the spanning tree T, as described in
the proof of (4.20).

edge e on P that crosses from S to V - S. See Figure 4.11 for an illustration of,
this.
Now consider the set of edges T~ = T - {e} LJ [e}. Arguing just as in the
proof of the Cut Property (4.17), the graph (V, T~) is connected and has no
cycles, so T is a spanning tree of G. Moreover, since e is the most expensive
edge on the cycle C, and e belongs to C, it must be that e is cheaper than e,
and hence T is cheaper than T, as desired. []
The Optimality of the Reverse-Delete Algorithm Now that we have the Cycle
Property (4.20), it is easy to prove that the Reverse-Delete Algorithm produces
a minimum spanning tree. The basic idea is analogous to the optimality proofs
for the previous two algorithms: Reverse-Delete only adds an edge when it is
justified by (4.20).
(4.21)
of G.

The Reverse-Delete Algorithm produces a minimum spanning tree

Proof. Consider any edge e = (v, w) removed by Reverse-Delete. At the time


that e is removed, it lies on a cycle C; and since it is the first edge encountered
by the algorithm in decreasing order of edge costs, it must be the most
expensive edge on C. Thus by (4.20), e does not belong to any minimum
spanning tree.
So if we show that the output (V, T) of Reverse-Delete is a spanning tree
of G, we will be done. Clearly (V, T) is connected, since the algorithm never
removes an edge when this will disconnect the graph. Now, suppose by way of

While we will not explore this further here, the combination of the Cut
Property (4.17) and the Cycle Property (4.20) implies that something even
more general is going on. Any algorithm that builds a spanning tree by
repeatedly including edges when justified by the Cut Property and deleting
edges when justified by the Cycle Property--in any order at all--will end up
with a minimum spanning tree. This principle allows one to design natural
greedy algorithms for this problem beyond the three we have considered here,
and it provides an explanation for why so many greedy algorithms produce
optimal solutions for this problem.
Eliminating the Assumption that All Edge Costs Are Distinct Thus far, we
have assumed that all edge costs are distinct, and this assumption has made the
analysis cleaner in a number of places. Now, suppose we are given an instance
of the Minimum Spanning Tree Problem in which certain edges have the same
cost - how can we conclude that the algorithms we have been discussing still
provide optimal solutions?
There turns out to be an easy way to do this: we simply take the instance
and perturb all edge costs by different, extremely small numbers, so that they
all become distinct. Now, any two costs that differed originally will sti!l have
the same relative order, since the perturbations are so small; and since all
of our algorithms are based on just comparing edge costs, the perturbations
effectively serve simply as "tie-breakers" to resolve comparisons among costs
that used to be equal.
Moreover, we claim that any minimum spanning tree T for the new,
perturbed instance must have also been a minimum spanning tree for the
original instance. To see this, we note that if T cost more than some tree T* in
the original instance, then for small enough perturbations, the change in the
cost of T cannot be enough to make it better than T* under the new costs. Thus,
if we run any of our minimum spanning tree algorithms, using the perturbed
costs for comparing edges, we will produce a minimum spanning tree T that
is also optimal for the original instance.

Implementing Prims Algorithm


We next discuss how to implement the algorithms we have been considering
so as to obtain good running-time bounds. We will see that both Prims and
Kruskals Algorithms can be implemented, with the right choice of data structures, to run in O(m log n) time. We will see how to do this for Prims Algorithm

149

150

Chapter 4 Greedy Algorithms

here, and defer discussing the implementation of Kruskals Algorithm to the


next section. Obtaining a running time close to this for the Reverse-Delete
Algorithm is difficult, so we do not focus on Reverse-Delete in this discussion.
For Pfims Algorithm, while the proof of correctness was quite different
from the proof for Dijkstras Algorithm for the Shortest-Path Algorithm, the
implementations of Prim and Dijkstra are almost identical. By analogy with
Dijkstras Algorithm, we need to be able to decide which node v to add next to
the growing set S, by maintaining the attachment costs a(v) = mine=(u,v):aEs Ce
for each node v ~ V - S. As before, we keep the nodes in a priority queue with
these attachment costs a(v) as the keys; we select a node with an Extra~cNin
operation, and update the attachment costs using ChangeKey operations.
There are n - I iterations in which we perform Ex~crac~cNin, and we perform
ChangeKey at most once for each edge. Thus we have
(4.22) Using a priority queue, Prims Algorithm can be implemented on a
graph with n nodes and m edges to run in O(m) time, plus the time for n
Ex~;rac~Iqin, and m ChangeKey operations.
As with Dijkstras Algorithm, if we use a heap-based priority queue we
can implement both Ex~crac~cMin and ChangeKey in O(log n) time, and so get
an overall running time of O(m log n).

Extensions
The minimum spanning tree problem emerged as a particular formulation
of a broader network design goal--finding a good way to connect a set of
sites by installing edges between them. A minimum spaxming tree optimizes
a particular goa!, achieving connectedness with minimum total edge cost. But
there are a range of fllrther goals one might consider as well.
We may, for example, be concerned about point-to-point distances in the
spanning tree we .build, and be willing to reduce these even if we pay more
for the set of edges. This raises new issues, since it is not hard to construct
examples where the minimum spanning tree does not minimize point-to-point
distances, suggesting some tension between these goals.
Alternately, we may care more about the congestion on the edges. Given
traffic that needs to be routed between pairs of nodes, one could seek a
spanning tree in which no single edge carries more than a certain amount of
this traffic. Here too, it is easy to find cases in which the minimum spanning
tree ends up concentrating a lot of traffic on a single edge.
More generally, it is reasonable to ask whether a spanning tree is even the
right kind of solution to our network design problem. A tree has the property
that destroying any one edge disconnects it, which means that trees are not at

4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure

all robust against failures. One could instead make resilience an explicit goal,
for example seeking the cheapest connected network on the set of sites that
remains connected after the deletion of any one edge.
All of these extensions lead to problems that are computationally much
harder than the basic Minimum Spanning Tree problem, though due to their
importance in practice there has been research on good heuristics for them.

4.6 Implementing Kruskals Algorithm:


The IJnion-Find Data Structure
One of the most basic graph problems is to find the set of connected components. In Chapter 3 we discussed linear-time algorithms using BFS or DFS for
finding the connected components of a graph.
In this section, we consider the scenario in which a graph evolves through
the addition of edges. That is, the graph has a fixed population of nodes, but it
grows over time by having edges appear between certain paizs of nodes. Our
goal is to maintain the set of connected components of such a graph thxoughout
this evolution process. When an edge is added to the graph, we dont want
to have to recompute the connected components from scratch. Rather, we
will develop a data structure that we ca!l the Union-Find structure, which
will store a representation of the components in a way that supports rapid
searching and updating.
This is exactly the data structure needed to implement Kruskals Algorithm
efficiently. As each edge e = (v, w) is considered, we need to efficiently find
the identities of the connected components containing v and w. If these
components are different, then there is no path from v and w, and hence
edge e should be included; but if the components are the same, then there is
a v-w path on the edges already included, and so e should be omitted. In the
event that e is included, the data structure should also support the efficient
merging of the components of v and w into a single new component.
~ The Problem
The Union-Find data structure allows us to maintain disjoint sets (such as the
components of a graph) in the following sense. Given a node u, the operation
Find(u) will return the name of the set containing u. This operation can be
used to test if two nodes u and v are in the same set, by simply checking
if Find(u) = Find(v). The data structure will also implement an operation
Union(A, B) to take two sets A and B and merge them to a single set.
These operations can be used to maintain connected components of an
evolving graph G = (V, E) as edges are added. The sets will be the connected
components of the graph. For a node u, the operation Find(u) will return the

151

4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure

152

Chapter 4 Greedy Algorithms

name of the component containing u. If we add an edge (u, v) to the graph,


then we first test if u and v are already in the same connected component (by
testing if Find(u) = Find(v)). If they are not, then Union(Find(u),Find(v))
can be used to merge the two components into one. It is important to note
that the Union-Find data structure can only be used to maintain components
of a graph as we add edges; it is not designed to handle the effects of edge
deletion, which may result in a single component being "split" into two.
To summarize, the Union-Find data structure will support three operafions.
o MakeUnionFind(S) for a set S will return a Union-Find data structure
on set S where all elements are in separate sets. This corresponds, for
example, to the connected components of a graph with no edges. Our
goal will be to implement MakeUnionFind in time O(n) where n
o For an element u ~ S, the operation Find(u) will return the name of the
set containing u. Our goal will be to implement Find(u) in O(log n) time.
Some implementations that we discuss will in fact .take only 0(1) time,
for this operation.
o For two sets A and B, the operation Union(A, B) will change the data
structure by merging the sets A and B into a single set. Our goal .will be
to implement Union in O(log n) time.
.
Lets briefly discuss what we mean by the name of a set--for example,
as returned by the Find operation. There is a fair amount of flexibility in
defining the names of the sets; they should simply be consistent in the sense
that Find(v) and Find(w) should return the same name if v and w belong to~
the same set, and different names otherwise. In our implementations, we will
name each set using one of the elements it contains.
A Simple Data Structure for Union-Find
Maybe the simplest possible way to implement a Union-Find data structure
is to maintain an array Component that contains the name of the set cuirenfly
containing each element. Let S be a set, and assume it has n elements denoted
{1 ..... n}. We will set up an array Component of size n, where Component [s] is
the name of the set containing s. To implement MakeUnionFind(S), we set up
the array and initialize it to Component Is] = s for all s ~ S. This implementation
makes Find(u) easy: it is a simple lookup and takes only O(.1) time. However,
Union(A, B) for two sets A and B can take as long as O(n) time, as we have
to update the values of Component Is] for all elements in sets A and B.
To improve this bound, we will do a few simple optimizafions. First, it is
useful to explicitly maintain the list of elements in each set, so we dont have to
look through the whole array to find the elements that need updating. Further,

we save some time by choosing the name for the union to be the name of one of
the sets, say, set A: this way we only have to update the values Component [s]
for s ~ B, but not for any s ~ A. Of course, if set B is large, this idea by itself
doesnt help very much. Thus we add one further optimization. When set B
is big, we may want to keep its name and change Component [s] for all s ~ A
instead. More generally, we can maintain an additional array size of length
n, where size[A] is the size of set A, and when a Union(A, B) operation is
performed, we use the name of the larger set for the union. This way, fewer
elements need to have their Componen~c values updated.
Even with these optimizations, the worst case for a Union operation is
still O(n) time; this happens if we take the union of two large sets A and B,
each containing a constant fraction of all the elements. However, such bad
cases for Union cannot happen very often, as the resulting set A U B is even
bigger. How can we make this statement more precise? Instead of bounding
the worst-case running time of a single Union operation, we can bound the
total (or average) running time of a sequence of k Union operations.
(4.23) Consider the array implementation of the Union-Find data structure
for some set S of size n, where unions keep the name of the larger set. The
Find operation takes O(1) time, MakeUnionFind(S) takes O(n) time, and any
sequence of k Union operations takes at most O(k log k) time.
Proof. The claims about the MakeUnionFind and Find operations are easy
to verify. Now consider a sequence of k Union operations. The only part
of a Union operation that takes more than O(I) time is updating the array
Component. Instead of bounding the time spent on one Union operation,
we will bound the total time spent updating Component[v] for an element
u fi-Lroughout the sequence of k operations.
Recall that we start the data structure from a state when all n elements are
in their own separate sets. A single Union operation can consider at most two
of these original one-element sets, so after any sequence of k Union operations,
all but at most 2k elements of S have been completely untouched. Now
consider a particular element v. As vs set is involved in a sequence of Union
operations, its size grows. It may be that in some of these Unions, the value
of Component[v] is updated, and in others it is not. But our convention is that
the union uses the name of the larger set, so in every update to Component [v]
the size of the set containing u at least doubles. The size of vs set starts out at
I, and the maximum possible size it can reach is 2k (since we argued above
that all but at most 2k elements are untouched by Union operations). Thus
Component[v] gets updated at most 1og2(2k) times throughout the process.
Moreover, at most 2k elements are involved in any Union operations at all, so

153

154

Chapter 4 Greedy Algorithms

we get a bound of O(k log k) for the time spent updating Component values
in a sequence of k Union operations. ,,

4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure

IThe set {s, u, w} was merged into {t, u, z}.)

While this bound on the average running time for a sequence of k operations is good enough in many applications, including implementing Kruskals
Algorithm, we will try to do better and reduce the worst-case time required.
Well do this at the expense of raising the time required for the Find operationto O(log n).
A Better Data Structure for Union-Find
The data structure for this alternate implementation uses pointers. Each node
v ~ S will be contained in a record with an associated pointer to the name
of the set that contains v. As before, we will use the elements of the set S
as possible set names, naming each set after one of its elements. For the
MakeUnionFind(S) operation, we initiafize a record for each element v ~ S
with a pointer that points to itself (or is defined as a null pointer), to indicate
that v is in its own set.
Consider a Union operation for two sets A and/3, and assume that the
name we used for set A is a node v ~ A, while set B is named after node u ~ B.
The idea is to have either u or u be the name of the combined set; assume we
select v as the name. To indicate that we took the union of the two sets, and
that the name of the union set is v, we simply update us pointer to point to v.
We do not update the pointers at the other nodes of set B.
As a resuk, for elements w ~/3 other than u, the name of the set they
belong to must be computed by following a sequence of pointers, first lead~g
them to the "old name" u and then via the pointer from u to the "new name" v.
See Figure 4.12 for what such a representation looks like. For example, the twO
sets in Figure 4.12 could be the outcome of the following sequence of Union
operations: Union(w, u), Union(s, u), Union(t, v), Union(z, u), Union(i, x),
Union(y, j), Union(x, ]), and Union(u,
This pointer-based data structure implements Union in O(1) time: all we
have to do is to update one pointer. But a Find operation is no longer constant
time, as we have to follow a sequence of pointers through a history of old
names the set had, in order to get to the current name. How long can a Find(u)
operation take.~ The number of steps needed is exactly the number of times
the set containing node u had to change its name, that is, the number of times
the Component[u] array position would have been updated in our previous
array representation. This can be as large as O(n) if we are not careful with
choosing set names. To reduce the time required for a Find operation, we wll!
use the same optimization we used before: keep the name of the larger set
as the name of the union. The sequence of Unions that produced the data

Figure 4.12 A Union-Find data structure using pointers. The data structure has only
two sets at the moment, named after nodes u andj. The dashed arrow from u to u is the
result of the last Union operation. To answer a Find query, we follow the arrows unit
we get to a node that has no outgoing arrow. For example, answering the query Find(i)
would involve following the arrows i to x, and then x to ].

structure in Figure 4.12 followed this convention. To implement this choice


efficiently, we will maintain an additional field with the nodes: the size of the
corresponding set.

(4.24) Consider the above pointer-based implementation of the Union-Find


data structure [or some set S oy size n, where unions keep the name o[ the larger
set. A Union operation takes O(1) t~me, MakeUnionFind(S) takes O(n) time,
and a Find operation takes O(log n) time.
Proof. The statements about Union and MakeUnionFind are easy to verify.
The time to evaluate Find(u) for a node u is the number of thnes the set
containing node u changes its name during the process. By the convention
that the union keeps the name of the larger set, it follows that every time the
name of the set containing node u changes, the size of this set at least doubles.
Since the set containing ~ starts at size 1 and is never larger than n, its size can
double at most log2 rt times, and so there can be at most log2 n name changes.
[]

Further Improvements
Next we will briefly discuss a natural optimization in the pointer-based UnionFind data structure that has the effect of speeding up the Find operations.
Strictly speaking, this improvement will not be necessary for our purposes in
this book: for all the applications of Union-Find data structures that we consider, the O(log n) time per operation is good enough in the sense that further
improvement in the time for operations would not translate to improvements

155

Chapter 4 Greedy Algorithms

156

in the overall running time of the algorithms where we use them. (The UnionFind operations will not be the only computational bottleneck in the running
time of these algorithms.)
To motivate the improved version of the data structure, let us first discuss a
bad case for the running time of the pointer-based Union-Find data structure.
First we build up a structure where one of the Find operations takes about log n
time. To do this, we can repeatedly take Unions of equal-sized sets. Assume v
is a node for which the Find(v) operation takes about log rt time. Now we can
issue Find(v) repeatedly, and it takes log rt for each such call. Having to follow
the same sequence of log rt pointers every time for finding the name of the set
containing v is quite redundant: after the first request for Find(v), we akeady
"know" the name x of the set containing v, and we also know that all other
nodes that we touched during our path from v to the current name also are all
contained in the set x. So in the improved implementation, we will compress
the path we followed after every Find operation by resetting all pointers along
the path to point to the current name of the set. No information is lost by
doing this, and it makes subsequent Find operations run more quickly. See,
Figure 4.13 for a Union-Find data structure and the result of Find(v) using
path compression.
Now consider the running time of the operations in the resulting implementation. As before, a Union operation takes O(1) time and MakeUnionFind(S) takes O(rt) time to set up a data structure for a set of size ft. How did
the time required for a Find(v) operation change? Some Find operations can
still take up to log n time; and for some Find operations we actually increase

ow points directly
to x.from v to x1
Enverything
on the path

4.7 Clustering

e time, since after finding the name x of the set containing v, we have to go
back through the same path of pointers from v to x, and reset each of these
pointers to point to x directly. But this additional work can at most double
the time required, and so does not change the fact that a Find takes at most
O(log n) time. The real gain from compression is in making subsequent calls to
Find cheaper, and this can be made precise by the same type of argument we
used in (4.23): bounding the total tLme for a sequence of n Find operations,
rather than the worst-case time for any one of them. Although we do not go
into the details here, a sequence of n Find operations employing compression
requires an amount of time that is extremely close to linear in rt; the actual
upper bound is O(not(rt)), where or(n) is an extremely slow-growing function
of n called the irtverse Ackermartrt furtctiort. (In particular, o~(rt) < 4 for any
value of rt that could be encountered in practice.)

Implementing Kruskals Algorithm


Now well use the Union-Find data structure to implement Kruskals Algorithm. First we need to sort the edges by cost. This takes time O(m log m).
Since we have at most one edge between any pair of nodes, we have m < rt2
and hence this running time is also O(m log rt).
After the sorting operation, we use the Union-Find data structure to
maintain the connected components of (V, T) as edges are added. As each
edge e = (v, w) is considered, we compute Find(u) and Find(v) and test
if they are equal to see if v and w belong to different components. We
use Union(Find(u),Find(v)) to merge the two components, if the algorithm
decides to include edge e in the tree T.
We are doing a total of at most 2m Find and n- 1 Union operations
over the course of Kruskals Algorithm. We can use either (4.23) for the
array-based implementation of Union-Find, or (4.24) for the pointer-based
implementation, to conclude that this is a total of O(m log rt) time. (While
more efficient implementations of the Union-Find data structure are possible,
this would not help the running time of Kruskals Algorithm, which has an
unavoidable O(m log n) term due to the initial sorting of the edges by cost.)
To sum up, we have
" cart be implemertted on a graph with n -rtodes
: .........
(4.25) K, ruskal s AIgortthm
artd
m edges to rurt irt O(m log rt) time.

4.7 Clustering

(a)
Figttre 4.13 (a) An instance of a Union-Find data structure; and (b) the result of the
operation Find(u) on this structure, using path compression.

We motivated the construction of minimum spanning trees through the problem of finding a low-cost network connecting a set of sites. But minimum

157

158

4.7 Clustering

Chapter 4 Greedy Algorithms

spanning trees arise in a range of different settings, several of which appear


on the surface to be quite different from one another. An appealing example
is the role that minimum spanning trees play in the area of clustering.

f! The Problem
Clustering arises whenever one has a co!lection of obiects--say, a set of
photographs, documents, or microorganisms--that one is trying to classify
or organize into coherent groups. Faced with such a situation, it is natural
to look first for measures of how similar or dissimilar each pair of obiects is.
One common approach is to define a distance function on the objects, with
the interpretation that obiects at a larger distance from one another are less
similar to each other. For points in the physical world, distance may actually
be related to their physical distance; but in many applications, distance takes
on a much more abstract meaning. For example, we could define the distance
between two species to be the number of years since they diverged in the
course of evolution; we could define the distance between two images in ~/
video stream as the number of corresponding pixels at which their intensity
values differ by at least some threshold.
Now, given a distance function on the objects, the clustering problem
seeks to divide them into groups so that, intuitively, obiects within the same
group are "close," and objects in different groups are "far apart." Starting from
this vague set of goals, the field of clustering branches into a vast number of
technically different approaches, each seeking to formalize this general notion
of what a good set of groups might look like.
CIusterings of Maximum Spacing Minimum spanning trees play a role in one
of the most basic formalizations, which we describe here. Suppose we are given
a set U of n obiects, labeledpl,p2 ..... Pn. For each pair, p~ and pj, we have a
numerical distance d(p~, pj). We require only that d(Pi, P~) = 0; that d(p~, p]) > 0
for distinct p~ and pT; and that distances are symmetric: d(pi, p]) = d(pj, p~).
Suppose we are seeking to divide the obiects in U into k groups, for a
given parameter k. We say that a k-clustering of U is a partition of U into k
nonempty sets C1, C2 ..... Q. We define the spacing of a k-clustering to be the
minimum distance between any pair of points lying in different clusters. Given
that we want points in different clusters to be far apart from one another, a
natural goal is to seek the k-clustering with the maximum possible spacing.
The question now becomes the following. There are exponentially many
different k-clusterings of a set U; how can we efficiently find the one that has
maximum spacing?

f! Designing the Algorithm


To find a clustering of maximum spacing, we consider growing a graph on the
vertex set U. The connected components will be the clusters, and we will try
to bring nearby points together into the same cluster as rapidly as possible.
(This way, they dont end up as points in different clusters that are very close
together.) Thus we start by drawing an edge between the closest pair of points.
We then draw an edge between the next closest pair of points. We continue
adding edges between pairs of points, in order of increasing distance d(p~, p]).
In this way, we are growing a graph H on U edge by edge, with connected
components corresponding to clusters. Notice that we are only interested in
the connected components of the graph H, not the full set of edges; so if we
are about to add the edge (pi, pj) and find that pi and pj already belong to the
same cluster, we will refrain from adding the edge--its not necessary, because
it wont change the set of components. In this way, our graph-growing process
will never create a cycle; so H will actually be a union of trees. Each time
we add an edge that spans two distinct components, it is as though we have
merged the two corresponding clusters. In the clustering literature, the iterative
merging of clusters in this way is often termed single-link clustering, a special
case of hierarchical agglomerative clustering. (Agglomerative here means that
we combine clusters; single-link means that we do so as soon as a single link
joins them together.) See Figure 4.14 for an example of an instance with k = 3
clusters where this algorithm partitions the points into an intuitively natural
grouping.
What is the connection to minimum spanning trees? Its very simple:
although our graph-growing procedure was motivated by this cluster-merging
idea, our procedure is precisely Kruskals Minimum Spanning Tree Algorithm.
We are doing exactly what Kruskals Algorithm would do if given a graph G
on U in which there was an edge of cost d(Pi, pj) between each pair of nodes
(Pi, Pj)- The only difference is that we seek a k-clustering, so we stop the
procedure once we obtain k connected components.
In other, words, we are running Kruskals Algorithm but stopping it just
before it adds its last k - 1 edges. This is equivalent to taking the rill minimum
spanning tree T (as Kruskals Algorithm would have produced it), deleting the
k - 1 most expensive edges (the ones that we never actually added), and defining the k-clustering to be the resulting connected components C1, C2 .....
Thus, iteratively merging clusters is equivalent to computing a minimum spanning tree and deleting the most expensive edges.

~ Analyzing the Algorithm


Have we achieved our goal of producing clusters that are as spaced apart as
possible? The following claim shows that we have.

159

160

Chapter 4 Greedy Algorithms

4.8 Huffrnan Codes and Data Compression

Cluster I

Cluster Cr

Cluster 2

Cluster 3 .......... : .....

l \
Cluster C;

Cluster C~ : .....
/ ....................

Figure 4.15 An illustration of the proof of (4.26), showing that the spacing of any
other dustering can be no larger than that of the clustering found by the single-linkage
algorithm.
Figure 4.14 An example of single-linkage clustering with k = 3 dusters. The dusters
are formed by adding edges between points in order of increasing distance.

(4.26) The components C!, Ca ..... Ck [ormed by deleting the k, 1 most


expensive edges of the minimum spanning tree T constitute a k-clustering of
maximum spacing.
Proof. Let e denote the clustering C1, Ca ..... Ck. The spacing of e is precisely
the length d* of the (/( - !)st most expensive edge in the minimum spanning
tree; this is the length of the edge that Kruskals Mgorithm would have added
next, at the moment we stopped it.
Now consider some other/(-clustering e, which partitions U into nonempW sets C[, C; ..... C~. We must show that the spacing of e is at most
d*.

Since the two clustefings e and e are not the same, it must be that one
of our clusters Cr is not a subset of any of the/( sets C; in e. Hence there
are points Pi, Pj ~ Cr that belong to different clusters in e--say, Pi ~ C~s and
Now consider the picture in Figure 4.15. Since pi and pj belong to the same
component Cr, it must be that Kruskals Algorithm added all the edges of a
PrPj path P before we stopped it. In particular, this means that each edge on

P has length at most d*. Now, we know that Pi ~ C~s but pj ~ C~s; so let p be
the first node on P that does not belong to C, and let p be the node on P that
comes just before p. We have just argued that d(p, p) < d*, since the edge
(p, p) was added by Kruskals Algorithm. But p and p belong to different sets
in the clustering e, and hence the spacing of e is at most d(p, p) _< d*. This
completes the proof, m

4.8 Huffman Codes and Data Compression


In the Shortest-Path and Minimum Spanning Tree Problems, weve seen how
greedy algorithms can be used to commit to certain parts of a solution (edges
in a graph, in these cases), based entirely on relatively short-sighted considerations. We now consider a problem in which this style of "committing" is
carried out in an even looser sense: a greedy rule is used, essentially, to shrink
the size of the problem instance, so that an equivalent smaller problem can
then be solved by recursion. The greedy operation here is proved to be "safe,"
in the sense that solving the smaller instance still leads to an optimal solution for the original instance, but the global consequences of the initial greedy
decision do not become fully apparent until the full recursion is complete.
The problem itself is one of the basic questions in the area of data compression, an area that forms part of the foundations for digital communication.

161

162

Chapter 4 Greedy Algorithms

~ The Problem
Encoding Symbols Using Bits Since computers ultimately operate on sequences of bits (i.e., sequences consisting only of the symbols 0 and 1), one
needs encoding schemes that take text written in richer alphabets (such as the
alphabets underpinning human languages) and converts this text into long
strings of bits.
The simplest way to do this would be to use a fixed number of bits for
each symbol in the alphabet, and then just concatenate the bit strings for
each symbol to form the text. To take a basic example, suppose we wanted to
encode the 26 letters of English, plus the space (to separate words) and five
punctuation characters: comma, period, question mark, exclamation point,
and apostrophe. This would give us 32 symbols in total to be encoded.
Now, you can form 2b different sequences out of b bits, and so if we use 5
bits per symbol, then we can encode 2s= 32 symbols--just enough for our
purposes. So, for example, we could let the bit string 00000 represent a, the
bit string 00001 represent b, and so forth up to 11111, which could represent the
apostrophe. Note that the mapping of bit strings to symbols is arbitrary; the
point is simply that five bits per symbol is sufficient. In fact, encoding schemes
like ASCII work precisely this way, except that they use a larger number of
bits per symbol so as to handle larger character sets, including capital letters,
parentheses, and all those other special symbols you see on a typewriter or
computer keyboard.
Lets think about our bare-bones example with just 32 symbols. Is there
anything more we could ask for from an encoding scheme?. We couldnt ask
to encode each symbo! using just four bits, since 24 is only 16--not enough
for the number of symbols we have. Nevertheless, its not clear that over large
stretches of text, we really need to be spending an average of five bits per
symbol. If we think about it, the letters in most human alphabets do not
get used equally frequently. In English, for example, the letters e, t: a, o, i,
and n get used much more frequently than q, J, x, and z (by more than an
order of magnitude). So its really a tremendous waste to translate them all
into the same number of bits; instead we could use a small number of bits for
the frequent letters, and a larger number of bits for the less frequent ones, and
hope to end up using fewer than five bits per letter when we average over a
long string of typical text.
This issue of reducing the average number of bits per letter is a fundamental problem in the area of data compression. When large files need to be
shipped across communication networks, or stored on hard disks, its important to represent them as compactly as possible, subject to the requirement
that a subsequent reader of the file should be able to correctly reconstruct it.
A huge amount of research is devoted to the design of compression algorithms

4.8 Huffman Codes and Data Compression

that can take files as input and reduce their space ~rough efficient encoding
schemes.
We now describe one of the fundamental ways of formulating this issue,
building up to the question of how we might construct the optimal way to take
advantage of the nonuniform frequencies of the letters. In one sense, such an
optimal solution is a very appealing answer to the problem of compressing
data: it squeezes all the available gains out of nonuniformities in the frequencies. At the end of the section, we will discuss how one can make flLrther
progress in compression, taking advantage of features other than nonuniform
frequencies.
Variable-Length Encoding Schemes Before the Internet, before the digital
computer, before the radio and telephone, there was the telegraph. Communicating by telegraph was a lot faster than the contemporary alternatives of
hand-delivering messages by railroad or on horseback. But telegraphs were
only capable of transmitting pulses down a wire, and so if you wanted to send
a message, you needed a way to encode the text of your message as a sequence
of pulses.

To deal with this issue, the pioneer of telegraphic communication, Samuel


Morse, developed Morse code, translating each letter into a sequence of dots
(short pulses) and dashes (long pulses). For our purposes, we can think of
dots and dashes as zeros and ones, and so this is simply a mapping of symbols
into bit strings, just as in ASCII. Morse understood the point that one could
communicate more efficiently by encoding frequent letters with short strings,
and so this is the approach he took. (He consulted local printing presses to get
frequency estimates for the letters in English.) Thus, Morse code maps e to 0
(a single dot), t to 1 (a single dash), a to 01 (dot-dash), and in general maps
more frequent letters to shorter bit strings.
In fact, Morse code uses such short strings for the letters that the encoding
of words becomes ambiguous. For example, just using what we know about
the encoding of e, t, and a, we see that the string 0101 could correspond to
any of the sequences of letters eta, aa, etet, or aet. (There are other possibilities as well, involving other letters.) To deal with this ambiguity, Morse
code transmissions involve short pauses between letter; (so the encoding of
aa would actually be dot-dash-pause-dot-dash-pause). This is a reasonable
solution--using very short bit strings and then introducing pauses--but it
means that we havent actually encoded the letters using just 0 and 1; weve
actually encoded it using a three-letter alphabet of 0, 1, and "pause." Thus, if
we really needed to encode everything using only the bits 0 and !, there would
need to be some flLrther encoding in which the pause got mapped to bits.

163

Chapter 4 Greedy Algorithms

4.8 Huffman Codes and Data Compression

164
Prefix Codes The ambiguity problem in Morse code arises because there exist
pairs of letters where the bit string that encodes one letter is a prefix of the bit
string that encodes another. To eliminate this problem, and hence to obtain an
encoding scheme that has a well-defined interpretation for every sequence of
bits, it is enough to map letters to bit strings in such a way that no encoding
is a prefix of any other.
Concretely, we say that a prefix code for a set S of letters is a function y
that maps each letter x ~ S to some sequence of zeros and ones, in such a way
that for distinct x, y ~ S, the sequence },(x) is not a prefix of the sequence y(y).
Now suppose we have a text consisting of a sequence of letters xlx2x3
x~.~ We can convert this to a sequence of bits by simply encoding each letter as
a bit sequence using ~ and then concatenating all these bit sequences together:
~ (xl) y (x2) y (xn). If we then hand this message to a recipient who knows the
function y, they will be able to reconstruct the text according to the following
rule.
o Scan the bit sequence from left to right.
o As soon as youve seen enough bits to match the encoding of some letter,
output this as the first letter of the text. This must be the correct first letter,
since no shorter or longer prefix of the bit sequence could encode any
other letter.
o Now delete the corresponding set of bits from the front of the message
and iterate.
In this way, the recipient can produce the correct set of letters without our
having to resort to artificial devices like pauses to separate the letters.
For example, suppose we are trying to encode the set of five letters
S = {a, b, c, d, e}. The encoding ~1 specified by
y~(a) = 11
Zl(b) = O1
y~(c) = 001
y~(d) = 10
}q(e) = 000
is a prefix code, since we can check that no encoding is a prefix of any other.
NOW, for example, the string cecab would be encoded as 0010000011101. A
recipient of this message, knowing y~, would begin reading from left to right.
Neither 0 nor O0 encodes a letter, but 001 does, so the recipient concludes that
the first letter is c. This is a safe decision, since no longer sequence of bits
beginning with 001 could encode a different letter. The recipient now iterates

165

on the rest of the message, 0000011101; next they will conclude that the second
letter is e, encoded as 000.
Optimal Prefix Codes Weve been doing all this because some letters are
more frequent than others, and we want to take advantage of the fact that more
frequent letters can have shorter encodings. To make this objective precise, we
now introduce some notation to express the frequencies of letters.
Suppose that for each letter x ~ S, there is a frequency fx, representing the
fraction of letters in the text that are equal to x. In other words, assuming
there are n letters total, nfx of these letters are equal to x. We notice that the
frequencies sum to 1; that is, ~x~S fx = 1.
Now, if we use a prefix code ~, to encode the given text, what is the total
length of our encoding? This is simply the sum, over all letters x ~ S, of the
number of times x occurs times the length of the bit string }, (x) used to encode
x. Using Iy(x)l to denote the length y(x), we can write this as
encoding length = ~ nfxl},(x)[ = n ~ fx[y(x)l.
x~S

Dropping the leading coefficient of n from the final expression gives us


~x~s fxl}(x)l, the average number of bits required per letter. We denote this
quantity by ABL0,).
To continue the earlier example, suppose we have a text with the letters
S = {a, b, c, d, e}, and their frequencies are as follows:

=.B2, f~=.25, f~=.20, fa=.~8, f~=.o5.


Then the average number of bits per letter using the prefix code Yl defined
previously is
.32.2+.25.2+.20.3+.18.2+.05.3 =2.25.
It is interesting to compare this to the average number of bits per letter using
a fixed-length encoding. (Note that a fixed-length encoding is a prefix code:
if all letters have encodings of the same length, then clearly no encoding can
be a prefix of any other.) With a set S of five letters, we would need three bits
per letter for a fixed-length encoding, since two bits could only encode four
letters. Thus, using the code ~1 reduces the bits per letter from 3 to 2.25, a
savings of 25 percent.
And, in fact, Yl is not the best we can do in this example. Consider the
prefix code ya given by

166

4.8 Huffman Codes and Data Compression

Chapter 4 Greedy Algorithms

g2(a) = 11
g2(b) = 10
g2(c) = 01

g2(d) = 001
g2(e) = 000
The average number of bits per letter using gz is
.32.2 + .25- 2 -k .20 2 + .18.3 4- .05- 3 = 2.23.
So now it is natural to state the underlying question. Given an alphabet
and a set of frequencies for the letters, we would like to produce a prefix
code that is as efficient as possible--namely, a prefix code that minimizes the
average nu}nber of bits per letter ABL(g) = ~_,x~S fxlg(x)l. We will call such a
prefix code optimal.

f! Designing the Algorithm


The search space for this problem is fairly complicated; it includes all possible
ways of mapping letters to bit strings, subiect to the defining property of prefix
codes. For alphabets consisting of an extremely small number of letters, it is
feasible to search this space by brute force, but this rapidly becomes infeasible.
We now describe a greedy method to construct an optimal prefix code
very efficiently. As a first step, it is useful to develop a tree-based means of
representing prefix codes that exposes their structure more clearly than simply
the lists of function values we used in our previous examples.
Representing Prefix Codes Using Binary Trees Suppose we take a rooted tree
T in which each node that is not a leaf has at most two children; we call such
a tree a binary tree. Further suppose that the number of leaves is equal to the
size of the alphabet S, and we label each leaf with a distinct letter in S.
Such a labeled binary tree T naturally describes a prefix code, as follows.
For each letter x ~ S, we follow the path from the root to the leaf labeled x;
each time the path goes from a node to its left child, we write down a 0, and
each time the path goes from a node to its right child, we write down a 1. We
take the resulting string of bits as the encoding of x.
Now we observe
(4.27) The enCoding of S Constructed from T is a prefix code.
Proof. In order for the encoding of x to be a prefix of the encoding of y, the
path from the root to x would have to be a prefix of the path from the root

to y. But this is the same as saying that x would lie on the path from the
root to y, which isnt possible if x is a leaf. []
This relationship between binary trees and prefix codes works in the other
direction as well. Given a prefix code g, we can build a binary tree recursively
as follows. We start with a root; all letters x ~ S whose encodings begin with
a 0 will be leaves in the left subtree of the root, and all letters y ~ S whose
encodlngs begin with a 1 will be leaves in the right subtree of the root. We
now build these two subtrees recursively using this rule.
For example, the labeled tree in Figure 4.16(a) corresponds to the prefix
code g0 specified by
go(a) -- 1
go(b) -- 011
g0(c) = 010
g0(d) = 001
g0(e) = 000
To see this, note that the leaf labeled a is obtained by simply taking the righthand edge out of the root (resulting in an encoding of !); the leaf labeled e is
obtained by taMng three successive left-hand edges starting from the root; and
analogous explanations apply for b, c, and d. By similar reasoning, one can
see that the labeled tree in Figure 4.16(b) corresponds to the prefix code gl
defined earlier, and the labeled tree in Figure 4.16(c) corresponds to the prefix
code g2 defined earlier. Note also that the binary trees for the two prefix codes
gl and g2 are identical in structure; only the labeling of the leaves is different.
The tree for go, on the other hand, has a different structure.
Thus the search for an optimal prefix code can be viewed as the search for
a binary tree T, together with a labeling of the leaves of T, that minimizes the
average number of bits per letter. Moreover, this average quantity has a natural
interpretation in the terms of the structure of T: the length of the encoding of
a letter x ~ S is simply the length of the path from the root to the leaf labeled
x. We will refer to the length of this path as the depth of the leaf, and we will
denote the depth of a leaf u in T simply by depthw(u). (As fwo bits of notational
convenience, we will drop the subscript T when it is clear from context, and
we will often use a letter x ~ S to also denote the leaf that is labeled by it.)
Thus we dre seeking the labeled tree that minimizes the weighted average
of the depths of all leaves, where the average is weighted by the frequencies
of the letters that label the leaves: ~x~s Ix" depthw(X). We will use ABL(T) to
denote this quantity.

167

168

Chapter 4 Greedy Algorithms

4.8 Huffman Codes and Data Compression

a node u with exactly one child u. Now convert T into a tree T by replacing
node u with v.
To be precise, we need to distinguish two cases. If u was the root of the
tree, we simply delete node u and use u as the root. If u is not the root, let w
be the parent of u in T. Now we delete node u and make v be a child of w
in place of u. This change decreases the number of bits needed to encode any
leaf in the subtree rooted at node u, and it does notaffect other leaves. So the
prefix code corresponding to T has a smaller average number of bits per letter
than the prefix code for T, contradicting the optimality of T. []
A First Attempt: The Top-Down Approach Intuitively, our goal is to produce
a labeled binary tree in which the leaves are as close to the root as possible.
This is what will give us a small average leaf depth.
A natural way to do this would be to try building a tree from the top down
by "packing" the leaves as tightly as possible. So suppose we try to split the
alphabet S into two sets S1 and S2, such that the total frequency of the letters
in each set is exactly . If such a perfect split is not possible, then we can try
for a split that is as nearly balanced as possible. We then recursively construct
prefix codes for S1 and S2 independently, and make these the two subtrees of
the root. (In terms of bit strings, this would mean sticking a 0 in front of the
encodings we produce for S1, and sticking a 1 in front of the encodings we
produce for $2.)

Figure 4.16 Parts (a), (b), and (c) of the figure depict three different prefix codes for
the alphabet S = {a, b, c, d, el.

As a first step in considering algorithms for this problem, lets note a simple
fact about the optimal tree. For this fact, we need a definition: we say that a
binary tree is full if each node that is not a leaf has two children. (In other
words, there are no nodes with exactly one chiAd.) Note that all three binary
trees in Figure 4.16 are full.
(4.28) The binary tree corresponding to the optimal prefix code is full.
Proof. This is easy to prove using an exchange argument. Let T denote the
binary tree corresponding to the optimal prefix code, and suppose it contains

It is not entirely clear how we should concretely define this "nearly


balanced" split of th6 alphabet, but there are ways to make this precise.
The resulting encoding schemes are called Shannon-Fano codes, named after
Claude Shannon and Robert Fano, two of the major early figures in the area
of information theory, which deals with representing and encoding digital
information. These types of prefix codes can be fairly good in practice, but
for our present purposes they represent a kind of dead end: no version of this
top-down splitting strategy is guaranteed to always produce an optimal prefix
code. Consider again our example with the five-letter alphabet S = {a, b, c, d, e}
and frequencies
fa=.32, fb=.25, fc=.20, fd=.18, re=.05.

There is a unique way to split the alphabet into two sets of equal frequency:
{a, d} and {b, c, e}. For {a, d}, we can use a single bit to encode each. For
{b, c, e}, we need to continue recursively, and again there is a unique way
to split the set into two subsets of equal frequency. The resulting code corresponds to the code gl, given by the labeled tree in Figure 4.16(b); and weve
already seen that 1~ is not as efficient as the prefix code ~2 corresponding to
the labeled tree in Figure 4.16(c).

169

170

4.8 Huffman Codes and Data Compression

Chapter 4 Greedy Algorithms


Shannon and Fano knew that their approach did not always yield the
optimal prefix code, but they didnt see how to compute the optimal code
without brute-force search. The problem was solved a few years later by David
Huffman, at the time a graduate student who learned about the question in a
class taught by Fano.
We now describe the ideas leading up to the greedy approach that Huffrnan
discovered for producing optimal prefix codes.

What If We Knew the Tree Structure of the Optimal Prefix Code? A technique that is often helpful in searching for an efficient algorithm is to assume,
as a thought experiment, that one knows something partial about the optimal
solution, and then to see how one would make use of this partial knowledge
in finding the complete solution. (Later, in Chapter 6, we will see in fact that
this technique is a main underpinning of the dynamic programming approach
to designing algorithms.)
For the current problem, it is useful to ask: What if someone gave us the
binary tree T* that corresponded to an optimal prefix code, but not the labeling
of the leaves? To complete the solution, we would need to figure out which
letter should label which leaf of T*, and then wed have our code. How hard
is this?
In fact, this is quite easy. We begin by formulating the following basic fact.
(4.29) Suppose that u and v are leaves of T*, such that depth(u) < depth(v).
Further, suppose that in a labeling of T* corresponding to an optimal prefix
code, leaf u is labeled with y ~ S and leaf v is labeled with z ~ S. Then fy >_ fz.
Proof. This has a quick proof using an exchange argument. If fy < fz, then
consider the code obtained by exchanging the labels at the nodes u and
v. In the expression for the average number of bits per letter, ,~BL(T*)=
~x~S fx depth(x), the effect of this exchange is as follows: the multiplier on fy
increases (from depth(u) to depth(v)), and the multiplier on fz decreases by
the same amount (from depth(v) to depth(u)).
Thus the change to the overall sum is (depth(v) - depth(u))(fy - fz). If
~fy < fz, this change is a negative number, contradicting the supposed optimality
of the prefix code that we had before the exchange, m
We can see the idea behind (4.29) in Figure 4. !6 (b): a quick way to see that
the code here is not optimal is to notice that it can be improved by exchanging
the positions of the labels c and d. Having a lower-frequency letter at a strictly
smaller depth than some other higher-frequency letter is precisely what (4.29)
rules out for an optimal solution.

Statement (4.29) gives us the following intuitively natura!, and optimal,


way to label the tree T* if someone should give it to us. We first take all leaves
of depth 1 (if there are an.y) ~nd label them with the highest-frequency letters
in any order. We then take all leaves of depth 2 (if there are any) and label them
with the next-highest-frequency letters in any order. We continue through the
leaves in order of increasing depth, assigning letters in order of decreasing
frequency. The point is that this cant lead to a suboptimal labeling of T*,
since any supposedly better labeling would be susceptible to the exchange in
(4.29). It is also crucial to note that, among the labels we assign to a block of
leaves all at the same depth, it doesnt matter which label we assign to which
leaf. Since the depths are all the same, the corresponding multipliers in the
expression Y~x~s fxlY (x) l are the same, and so the choice of assignment among
leaves of the same depth doesnt affect the average number of bits per letter.
But how is all this helping us? We dont have the structure of the optimal
tree T*, and since there are exponentially many possible trees (in the size of
the alphabet), we arent going to be able to perform a brute-force search over
all of them.

In fact, our reasoning about T* becomes very useful if we think not about
the very beginning of this labeling process, with the leaves of minimum depth,
but about the very end, with the leaves of maximum depth--the ones that
receive the letters with lowest frequency. Specifically, consider a leaf v in T*
whose depth is as large as possible. Leaf u has a parent u, and by (4.28) T* is
a till binary tree, so u has another child w. We refer to v and w as siblings,
since they have a common parent. Now, we have
(4.30) w is a leaf of T*.
Proof. If w were not a leaf, there would be some leaf w in the subtree below
it. But then w would have a depth greater than that of v, contradicting our
assumption that v is a leaf of maximum depth in T*. ~,

So v and w are sibling leaves that are as deep as possible in T*. Thus our
level-by-level process of labeling T*, as justified by (4.29), will get to the level
containing v and w last. The leaves at this level will get the lowest-frequency
letters. Since we have already argued that the order in which we assign these
letters to the leaves within this level doesnt matter, there is an optimal labeling
in which u and w get the two lowest-frequency letters of all.
We sum this up in the following claim.
(4.31) There is an optimal prefix code, with corresponding tree T*, in which
:the two lowest-frequency letters are assigned to leaves that are Siblings in T*.

171

172

Chapter 4 Greedy Algorithms

4.8 Huffman Codes and Data Compression

173

Take the leaf labeled ~ and add two children below it


labeled y* and z*
Endif

letter
with sum of ffequenciesJ

; ,
0 0"~-~Tw lwest-frequency letters )
Figure 4.17 There is an optimal solution in which the two lowest-frequency letters
labe! sibling leaves; deleting them and labeling their parent with a new letter having t~e
combined frequency yields an instance ~th a smaller alphabet.

An Algorithm to Construct an Optimal Prefix Code Suppose that y* and z*


are the two lowest-frequency letters in S. (We can break ties in the frequencies
arbitrarily.) Statement (4.31) is important because it tells us something about
where y* and z* go in the optim!l solution; it says that it is safe to "lock them
together" in thinking about the solution, because we know they end up as
sibling leaves below a common parent. In effect, this common parent acts like
a "meta-letter" whose frequency is the sum of the frequencies of y* and z*.
This directly suggests an algorithm: we replace y* and z* with this metaletter, obtaining an alphabet that is one letter smaller. We recursively find a
prefix code for the smaller alphabet, and then "open up" the meta-letter back
into y* and z* to obtain a prefix code for S. This recursive strategy is depicted
in Figure 4.17.
A concrete description of the algorithm is as follows.
To construct a prefix code for an alphabet S, with given frequencies:
If S has two letters then
Encode one letter using 0 and the other letter using I
Else
Let y* and z* be the two lowest-frequency letters
Form a new alphabet S by deleting y* and z* and
replacing them with a new letter ~ of frequency ~. ~*
Kecursively construct a prefix code Z for S, with tree T
Define a prefix code for S as fol!ows:
Start with T

We refer to this as Huffmans Algorithm, and the prefix code that it


produces for a given alphabet is accordingly referred to as a Huffman code.
In general, it is clear that this algorithm always terminates, since it simply
invokes a recursive call on an alphabet that is one letter smaller. Moreover,
using (4.31), it will not be difficult to prove that the algorithm in fact produces
an optimal prefix code. Before doing this, however, we pause to note some
further observations about the algorithm.
First lets consider the behavior of the algorithm on our sample instance
with S = {a, b, c, d, e} and frequencies

.20, ~=..18, 5=.o5.


The algorithm would first merge d and e into a single letter--lets denote it
(de)--of frequency .18 + .05 = .23. We now have an instance of the problem
on the four letters S = {a, b, c, (de)}. The two lowest-frequency letters in S are
c and (de), so in the next step we merge these into the single letter (cde) of
frequency .20 + .23 = .43. This gives us the three-letter alphabet {a, b, (cde)}.
Next we merge a and b, and this gives us a two-letter alphabet, at which point
we invoke the base case of the recursion. If we unfold the result back through
the recursive calls, we get the tree pictured in Figure 4.16(c).
It is interesting to note how the greedy rule underlying Huffmans
Algorithm--the merging of the two lowest-frequency letters--fits into the
structure of the algorithm as a whole. Essentially, at the time we merge these
two letters, we dont know exactly how they will fit into the overall code.
Rather, we simply commit to having them be children of the same parent, and
this is enough to produce a new, equivalent problem with one less letter.
Moreover, the algorithm forms a natural contrast with the earlier approach
that led to suboptimal Shannon-Fano codes. That approach was based on a
top-down strategy that worried first and foremost about the top-level split in
the binary tree--namely, the two subtrees directly below the root. Huffmans
Algorithm, on the other hand, follows a bottom-up approach: it focuses on
the leaves representing the two lowest-frequency letters~ and then continues
by recursion.
~ Analyzing the Mgorithm
The Optimality of the Algorithm We first prove the optimaliW of Huffmans
Mgorithm. Since the algorithm operates recursively, invoking itself on smaller
and smaller alphabets, it is natural to try establishing optimaliW by induction

174

4.8 Huffman Codes and Data Compression

Chapter 4 Greedy Algorithms ....

on the size of the alphabet. Clearly it is optimal for all two-letter alphabets
(since it uses only one bit per letter). So suppose by induction that it is optimal
for all alphabets of size/~ - 1, and consider an input instance consisting of an
alphabet S of size
Lets quickly recap the behavior of the algorithm on this instance. The
algorithm merges the two lowest-frequency letters y*, z* ~ S into a single letter
o0, calls itself recursively on the smaller alphabet S (in which y* and z* are
replaced by a)), and by induction produces an optimal prefix code for S,
represented by a labeled binary tree T. It then extends this into a tree T for S,
by attaching leaves labeled y* and z* as children of the node in T labeled
There is a close relationship between ABL(T) and ABL(T). (Note that the
former quantity is the average number of bits used to encode letters in S, while
the latter quantity is the average number of bits used to encode letters in S.)
(4.32) ABL(T) = ABL(T) -- fro-

Proof. The depth of each lefter x other than y*, z* is the same in both T and
T. Also, the depths of y* and z* in T are each one greater than the depth of
o) in T. Using this, plus the fact that [to = fy. + fz*, we have
ABL(T) = ~ ~" depthr(X)

= f~,- depthriv*) + fz*" depthr(z*) + ~ ~. depthT(X)


x-aY*r- ,Z*

= (fy* q- fz*)" (1 q- depthT,(~o)) +


= ]to" (1 q- depthr,(O))) q-

depthT,(X)
x~y*,z*

]x" depthr(X)

x~-y*,z*

= L + ~ ]x" depthr(X)
xES~

= ]:to q- ABE(T/)..

such that ABL(Z) < ABL(T); and by (4.31), there is such a tree Z in which the
leaves representing y* and z* are siblings.
It is now easy to get a contradiction, as follows. If we delete the leaves
labeled y* and z* from Z, and label their former parent with w, we get a tree
Z that defines a prefix code for S. In the same way that T is obtained from
T, the tree Z is obtained from ZI by adding leaves for y* and z* below to; thus
the identity in (4.32) applies to Z and Z as well: ABL(Z) = ABL(Z) -- [to.
But we have assumed that ABL(Z) < ABL(T); subtracting/:to from both sides
of this inequality we get ,~BL(Z) < ABL(T), which contradicts the optimality
of T as a prefix code for S. ,,
Implementation and Running Time It is clear that Huffmans Algorithm can
be made to run in polynomial time in k, the number of letters in the alphabet.
The recursive calls of the algorithm define a sequence of k - 1 iterations over
smaller and smaller alphabets, and each iteration except the last consists
simply of identifying the two lowest-frequency letters and merging them into
a single letter that has the combined frequency. Even without being careful
about the implementation, identifying the lowest-frequency letters can be done
in a single scan of the alphabet, in time O(k), and so summing this over the
k - 1 iterations gives O(k2) time.
But in fact Huffmans Algorithm is an ideal setting in which to use a
priority queue. Recall that a priority queue maintains a set of/c elements,
each with a numerical key, and it allows for the insertion of new elements and
the extraction of the element with the minimum key. Thus we can maintain
the alphabet S in a priority queue, using each letters frequency as its key.
In each iteration we just extract the minimum twice (this gives us the two
lowest-frequency letters), and then we insert a new letter whose key is the
sum of these two minimum frequencies. Our priority queue now contains a
representation of the alphabet that we need for the next iteration.
Using an implementation of priority queues via heaps, as in Chapter 2, we
can make each insertion and extraction of the minimum run in time O(log k);
hence, each iteration--which performs just three of these operations--takes
time O(log/0. Summing over all k iterations, we get a total running time of
O(k log k).

Using this, we now prove optimality as follows.


(4.33) The Huffinan code for a given alphabet achieves the minimum average
number of bits per letter of any prefix code.
Proof. Suppose by way of contradiction that the tree T produced by our greedy
algorithm is not optimal. This means that there is some labeled binary tree Z

Extensions
The structure of optimal prefix codes, which has been our focus here, stands
as a fundamental result in the area of data compression. But it is important to
understand that this optimality result does not by any means imply that we
have found the best way to compress data under all circumstances.

175

4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm

176

Chapter 4 Greedy Algorithms

What more could we want beyond an optimal prefix code? First, consider
an application in which we are transmitting black-and-white images: each
image is a 1,000-by-l,000 array of pixels, and each pixel takes one of the two
values black or white. Further, suppose that a typical image is almost entirely
white: roughly 1,000 of the million pixels are black, and the rest are white. Now,
if we wanted to compress such an image, the whole approach of prefix codes
has very little to say: we have a text of length one million over the two-letter
alphabet {black, white}. As a result, the text is already encoded using one bit
per letter--the lowest possible in our framework.
It is clear, though, that such images should be highly compressible.
Intuitively, one ought to be able to use a "fraction of a bit" for each white pixel,
since they are so overwhelmingly frequent, at the cost of using multiple bits
for each black pixel. (In an extreme version, sending a list of (x, y) coordinates
for each black pixel would be an improvement over sending the image as a
text with a million bits.) The challenge here is to define an encoding scheme
where the notion of using fractions of bits is well-defined. There are results
in the area of data compression, however, that do iust this; arithmetic coding
and a range of other techniques have been developed to handle settings like
this.

A second drawback of prefix codes, as defined here, is that they cannot


adapt to changes in the text. Again lets consider a simple example. Suppose we
are trying to encode the output of a program that produces a long sequence
of letters from the set {a, b, c, d}. Further suppose that for the first half of
this sequence, the letters a and b occur equally frequently, while c and d do
not occur at all; but in the second half of this sequence, the letters c and d
occur equally frequently, while a and b do not occur at all. In the framework
developed in this section, we are trying to compress a text over the four-letter
alphabet {a, b, c, d}, and all letters are equally frequent. Thus each would be
encoded with two bits.
But whats really happening in this example is that the frequency remains
stable for half the text, and then it changes radically. So one could get away
with iust one bit per letter, plus a bit of extra overhead, as follows.
o Begin with an encoding in which the bit 0 represents a and the bit 1
represents b.
o Halfway into the sequence, insert some kind of instruction that says,
"Were changing the encoding now. From now on, the bit 0 represents c
and the bit I represents d:
o Use this new encoding for the rest of the sequence.
The point is that investing a small amount of space to describe a new encoding
can pay off many times over if it reduces the average number of bits per

letter over a long run of text that follows. Such approaches, which change
the encoding in midstream, are called adaptive compression schemes, and
for many kinds of data they lead to significant improvements over the static
method weve considered here.
These issues suggest some of the directions in which work on data compression has proceeded. In many of these cases, there is a trade-off between
the power of the compression technique and its computational cost. In particular, many of the improvements to Huffman codes just described come with
a corresponding increase in the computational effort needed both to produce
the compressed version of the data and also to decompress it and restore the
original text. Finding the right balance among these trade-offs is a topic of
active research.

* 4.9 Minimum-Cost Arborescences: A Multi-Phase


Greedy Algorithm
As weve seen more and more examples of greedy algorithms, weve come to
appreciate that there can be considerable diversity in the way they operate.
Many greedy algorithms make some sort of an initial "ordering" decision on
the input, and then process everything in a one-pass fashion. Others make
more incremental decisions--still local and opportunistic, but without a g!obal
"plan" in advance. In this section, we consider a problem that stresses our
intuitive view of greedy algorithms still further.
,~J The Problem
The problem is to compute a minimum-cost arborescence of a directed graph.
This is essentially an analogue of the Minimum Spanning Tree Problem for
directed, rather than undirected, graphs; we will see that the move to directed
graphs introduces significant new complications. At the same time, the style
of the algorithm has a strongly greedy flavor, since it still constructs a solution
according to a local, myopic rule.
We begin with the basic definitions. Let G = (V, E) be a directed graph in
which weve distinguished one node r ~ V as a root. An arborescence (with
respect to r) is essentially a directed spanning tree rooted at r. Specifically, it
is a subgraph T = (V, F) such that T is a spanning tree of G if we ignore the
direction of edges; and there is a path in T from r to each other node v ~ V if
we take the direction of edges into account. Figure 4.18 gives an example of
two different arborescences in the same directed graph.
There is a useful equivalent way to characterize arborescences, and this
is as follows.

177

178

4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm

Chapter 4 Greedy Algorithms

The basic problem we consider here is the following. We are given a


directed graph G = (V, E), with a distinguished root node r and with a nonnegative cost ce >_ 0 on each edge, and we wish to compute an arborescence
rooted at r of minimum total cost. (We will refer to this as an optimal arborescence.) We will assume throughout that G at least has an arborescence rooted
at r; by (4.35), this can be easily checked at the outset.

Figttre 4.18 A directed graph can have many different arborescences. Parts (b) and (c)
depict two different aborescences, both rooted at node r, for the graph in part (a).

(4.t4) A subgraph T = (V, F) of G is an arborescence with respect to root r if


and only if T has no cycles, and for each node v ~ r, there is exactly one edge
in F that enters v.
Proof. If T is an arborescence with root r, then indeed every other node v
has exactly one edge entering it: this is simply the last edge on the unique r-v
path.
Conversely, suppose T has no cycles, and each node v # r has exactly
one entering edge. In order to establish that T is an arborescence, we need
only show that there is a directed path from r to each other node v. Here is
how to construct such a path. We start at v and repeatedly follow edges in
the backward direction. Since T has no cycles, we can never return tO a node
weve previously visited, and thus this process must terminate. But r is the
only node without incoming edges, and so the process must in fact terminate
by reaching r; the sequence of nodes thus visited yields a path (in the reverse
direction) from r to v. m
It is easy to see that, just as every connected graph has a spanning tree, a
directed graph has an arborescence rooted at r provided that r can reach every
node. Indeed, in this case, the edges in a breadth-first search tree rooted at r
will form an arborescence.

(4.t5) A directed graph G has an arborescence rooted at r if and only if thee


_

L4~ Designing the Algorithm


Given the relationship between arborescences and trees, the minimum-cost
arborescence problem certainlyhas a strong initial resemblance to the Minimum Spanning Tree Problem for undirected graphs. Thus its natural to start
by asking whether the ideas we developed for that problem can be carried
over directly to this setting. For example, must the minimum-cost arborescence contain the cheapest edge in the whole graph? Can we safely delete the
most expensive edge on a cycle, confident that it cannot be in the optimal
arborescence?
Clearly the cheapest edge e in G will not belong to the optimal arborescence
if e enters the root, since the arborescence were seeking is not supposed to
have any edges entering the root. But even if the cheapest edge in G belongs
to some arborescence rooted at r, it need not belong to the optimal one, as
the example of Figure 4.19 shows. Indeed, including the edge of cost 1 in
Figure 4.!9 would prevent us from including the edge of cost 2 out of the
root r (since there can only be one entering edge per node); and this in turn
would force us to incur an unacceptable cost of 10 when we included one of

2
10

10
4

(a)

Figure 4.19 (a) A directed graph with costs onits edges, and (b) an optimal arborescence
rooted at r for this graph.

179

4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm

180

Chapter 4 Greedy Algorithms

the other edges out of r. This kind of argument never clouded our thinking in
the Minimum Spanning Tree Problem, where it was always safe to plunge
ahead and include the cheapest edge; it suggests that finding the optimal
arborescence may be a significantly more complicated task. (Its worth noticing
that the optimal arborescence in Figure 4.19 also includes the most expensive
edge on a cycle; with a different construction, one can even cause the optimal
arborescence to include the most expensive edge in the whole graph.)
Despite this, it is possible to design a greedy type of algorithm for this
problem; its just that our myopic rule for choosing edges has to be a little
more sophisticated. First lets consider a little more carefully what goes wrong
with the general strategy of including the cheapest edges. Heres a particular
version of this strategy: for each node v # r, select the cheapest edge entering
v (breaking ties arbitrarily), and let F* be this set of n - 1 edges. Now consider
the subgraph (V, F*). Since we know that the optimal arborescence needs to
have exactly one edge entering each node v # r, and (V, F*) represents the
cheapest possible way of making these choices, we have the following fact2
(4.36) I[ (V, F*) is an arborescence, then it is a minimum-cost arborescence.
So the difficulty is that (V, F*) may not be an arborescence. In this case,
(4.34) implies that (V, F*) must contain a cycle C, which does not include the
root. We now must decide how to proceed in this situation.
To make matters somewhat clearer, we begin with the following observation. Every arborescence contains exactly one edge entering each node v # r;
so if we pick some node v and subtract a uniform quantity from the cost of
every edge entering v, then the total cost of every arborescence changes by
exactly the same amount. This means, essentially, that the actual cost of the
cheapest edge entering v is not important; what matters is the cost of all other
edges entering v relative to this. Thus let Yv denote the minimum cogt of any
edge entering v. For each edge e = (u, v), with cost ce >_ O, we define its modified cost ce to be ce - Yv- Note that since ce >_ y,, all the modified costs are still
nonnegafive. More crucially, our discussion motivates the following fact.
(4.37) T is an optimal arborescence in G subject to costs (Ce) if and Only if it
is an optimal arborescence Subject to the modified costs c
ProoL Consider an arbitrary arborescence T. The difference between its cost
with costs (ce} and [ce} is exactly ~,#r y~--that is,
eaT

eaT \

This is because an arborescence has exactly one edge entering each node
in the sum. Since the difference between the two costs is independent of the
choice of the arborescence T, we see that T has minimum cost subiect to {ce}
if and only if it has minimum cost subject to {ce}. ,,

We now consider the problem in terms of the costs {de}. All the edges in
our set F* have cost 0 under these modified costs; and so if (V, F*) contains
a cycle C, we know that all edges in C have cost 0. This suggests that we can
afford to use as many edges from C as we want (consistent with producing an
arborescence), since including edges from C doesnt raise the cost.
Thus our algorithm continues as follows. We contract C into a single
supemode, obtaining a smaller graph G = (V, E). Here, V contains the nodes
of V-C, plus a single node c* representing C. We transform each edge e E E to
an edge e E E by replacing each end of e that belongs to C with the new node
c*. This can result in G having parallel edges (i.e., edges with the same ends),
which is fine; however, we delete self-loops from E--edges that have both
ends equal to c*. We recursively find an optimal arborescence in this smaller
graph G, subject to the costs {Ce}. The arborescence returned by this recursive
call can be converted into an arborescence of G by including all but one edge
on the cycle C.
In summary, here is the full algorithm.
For each node u7&r
Let Yu be the minimum cost of an edge entering node
Modify the costs of all edges e entering v to ce=ce
Choose one 0-cost edge entering each u7~r, obtaining a set F*
If F* forms an arborescence, then return it
Else there is a directed cycle C_CF*
Contract C to a single supernode, yielding a graph G= (V,E)
Recursively find an optimal arborescence (V,F) in G
with costs [Ce}
~)
Extend (V,F
to an arborescence (V, F) in G
by adding all but one edge of C

~ Analyzing the Algorithm


It is easy to implement this algorithm so that it runs in polynomial time. But
does it lead to an optimal arborescence? Before concluding that it does, we need
to worry about the following point: not every arborescence in G corresponds to
an arborescence in the contracted graph G. Could we perhaps "miss" the true
optimal arborescence in G by focusing on G? What is true is the following.

181

182

Solved Exercises

Chapter 4 Greedy Algorithms

The arborescences of G are in one-to-one correspondence with arborescences


of G that have exactly one edge entering the cycle C; and these corresponding
arborescences have the same cost with respect to {ce}, since C consists of 0cost edges. (We say that an edge e = (u, v) enters C if v belongs to C but u does
not.) So to prove that our algorithm finds an optimal arborescence in G, we
must prove that G has an optimal arborescence with exactly one edge entering
C. We do this now.
(4.38) Let C be a cycle in G consisting of edges of cost O, such that r ~ C.
Then there is an optimal arborescence rooted at r that has exactly one edge
entering C.

The algorithm finds an optimal arborescence robted at ~ in G: :


Proof. The proof is by induction on the number of nodes in G. If the edges
of F form an arborescence, then the algorithm returns an optimal arborescence
by (4.36). Otherwise, we consider the problem with the modified costs {ce},
which is equivalent by (4.37). After contracting a 0-cost cycle C to obtain a
smaller graph G, the algorithm produces an optimal arborescence in G by the
inductive hypothesis. Finally, by (4.38), there is an optimal arborescence in G
that corresponds to the optimal arborescence computed for G. ~

Solved Exercises
Proof. Consider an optimal arborescence T in G. Since r has a path in T to
every node, there is at least one edge of T that enters C. If T enters C exactly
once, then we are done. Otherwise, suppose that T enters C more than once.
We show how to modify it to obtain a.n arborescence of no greater cost that
enters C exactly once.
Let e = (a, b) be an edge entering C that lies on as short a path as possible
from r; this means in particular that no edges on the path from r to a can enter
C. We delete all edges of T that enter C, except for the edge e. We add in all
edges of C except for the one edge that enters b, the head of edge e. Let T
denote the resulting subgraph of G.
We claim that T is also an arborescence. This will establish the result,
since the cost of T is clearly no greater than that of T: the only edges of
T that do not also belong to T have cost 0. So why is T an arborescence?
Observe that T has exactly one edge entering each node v # r, and no edge
entering r. So T has exactly n - 1 edges; hence if we can show there is an
path in T for each v, then T must be connected in an undirected sense, and
hence a tree. Thus it would satisfy our initial definition of an arborescence.
So consider any node v # r; we must show there is an r-v path in T. If
v ~ C, we can use the fact that the path in T from r to e has been preserved
in the construction of T; thus we can reach v by first reaching e and then
following the edges of the cycle C. Now suppose that v C, and let P denote
the r-v path in T. If P did not touch C, then it sti!l exists in T. Otherwise,
let tv be the last node in P C~ C, and let P be the subpath of P from tv to v.
Observe that all the edges in P still exist in T. We have already argued that
u~ is reachable from r in T, since it belongs to C. Concatenating this path
to tv with the subpath P gives us a path to v as well. ,,

We can now put all the pieces together to argue that our algorithm is
correct.
,

Solved Exercise 1
Suppose that three of your friends, inspired by repeated viewings of the
horror-movie phenomenon The Blair Witch Project, have decided to hike the
Appalachian Trail this summer. They want to hike as much as possible per
day but, for obvious reasons, not after dark. On a map theyve identified a
large set of good stopping points for camping, and theyre considering the
following system for deciding when to stop for the day. Each time they come
to a potential stopping point, they determine whether they can make it to the
next one before nightfall. If they can make it, then they keep hiking; otherwise,
they stop.
Despite many significant drawbacks, they claim this system does have
one good feature. "Given that were only hiking in the daylight," they claim,
"it minimizes the number of camping stops we have to make."
Is this true? The proposed system is a greedy algorithm, and we wish to
determine whether it minimizes the number of stops needed.
To make this question precise, lets make the following set of simplifying
assumptions. Well model the Appalachian Trail as a long line segment of
length L, and assume that your friends can hike d miles per day (independent
of terrain, weather conditions, and so forth). Well assume that the potential
stopping points are located at distances xl, x2 ..... xn from the start of the
trail. Well also assume (very generously) that your friends are always correct
when they estimate whether they can make it to the next stopping point before
nightfall.
Well say that a set of stopping points is valid if the distance between each
adjacent pair is at most d, the first is at distance at most d from the start of
the trail, and the last is at distance at most d from the end of the trai!. Thus
a set of stopping points is valid if one could camp only at these places and

183

184

Solved Exercises

Chapter 4 Greedy Algorithms

stil! make it across the whole trail. Well assume, naturally, that the full set of
n stopping points is valid; otherwise, there would be no way to make it the
whole way.
We can now state the question as follows. Is your Mends greedy
algorithm--hiking as long as possible each day--optimal, in the sense that it
finds a valid set whose size is as small as possible.~

on the first day before stopping. Now let ] > 1 and assume that the claim is
true for all i < j. Then

SoIation Often a greedy algorithm looks correct when you first encounter it,
so before succumbing too deeply to its intuitive appeal, its useful to ask: why
might it not work~. What should we be worried about.~
Theres a natural concern with this algorithm: Might it not help to stop
early on some day, so as to get better synchronized with camping opportunities
on future days~. But if you think about it, you start to wonder whether this could
really happen. Could there really be an alternate solution that intentionally.lags
behind the greedy solution, and then puts on a burst of speed and passes the
greedy solution? How could it pass it, giv._en that the greedy solution travels as
far as possible each day?
This last consideration starts to look like the outline of an argument based
on the "staying ahead" principle from Section 4.1. Perhaps we can show that as
long as the greedy camping strategy is ahead on a given day, no other solution
can catch up and overtake it the next day.
We now turn this into a proof showing the algorithm is indeed optimal,
identifying a natural sense in which the stopping points it chooses "stay ahead"
of any other legal set of stopping points. Although we are following the style
of proof from Section 4.1, its worth noting an interesting contrast with the
Interval Scheduling Problem: there we needed to prove that a greedy algorithm
maximized a quantity of interest, whereas here we seek to minimize a certain
quantity.
Let R = {xpl ..... xpk} denote the set of stopping points chosen- by the
greedy algorithm, and suppose by way of contradiction that there is a smaller
valid set of stopping points; lets call this smaller set S = {xq~ ..... xqm}, with

since xp~.l > xqj_l by the induction hypothesis. Combining these two inequalities, we have

To obtain a contradiction, we first show that the stopping point reached by


the greedy algorithm on each day j is farther than the stopping point reached
under the alternate solution. That is,
(4.40) For each j = 1, 2 .....

m, we have Xpj > x~tj.

Proof. We prove this by induction on j. The case j = 1 follows directly from


the definition of the greedy algorithm: your friends travel as long as possible

xqj - xqj_l <_ d,


since S is a valid set of stopping points, and

xq~ - Xp~:~ < xq~ - xq~_~

This means that your Mends have the option of hiking all the way from
Xpi_~ to Xqi in one day; and hence the location Xpj at which they finally stop
can only be farther along than xq~. (Note the similarity with the corresponding
proof for the Interval Scheduling Problem: here too the greedy algorithm is
staying ahead because, at each step, the choice made by the alternate solution
is one of its valid options.) i
we must
have Xp,
d, for otherwise
your that
Mends
have
needed
Statement
(4.40)
in particular
Xqmwould
<_ Xpm.never
NOW,
if m
< k, then
n < L -implies
to stop at the location Xp,~+~. Combining these two inequalities, we have
concluded that Xqm < L -- d; but this contradicts the assumption that S is a
valid set of stopping points.
Consequently, we cannot have m < k, and so we have proved that the
greedy algorithm produces a valid set of stopping points of minimum possible
size.

Solved Exercise 2
Your ~ends are starting a security company that needs to obtain licenses for
n different pieces of cryptographic software. Due to regulations, they can only
obtain these licenses at the rate of at most one per month.
Each license is currently selling for a price of $!00. However, they are
all becoming more expensive according to exponential growth curves: in
particular, the cost of license] increases by a factor of rj > 1 each month, where
rj is a given parameter. This means that if license] is purchased t months from
now, it will cost 100. r~. We will assume that all the price growth rates are
distinct; that is, ri ~ r1 for licenses i # ] (even though they start at the same
price of $100).

185

186

Solved Exercises

Chapter 4 Greedy Algorithms

The question is: Given that the company can only buy at most one license
a month, in which order should it buy the licenses so that the total amount of
money it spends is as small as possible?
Give an algorithm that takes the n rates of price growth rI, r2 ..... rn, and
computes an order in which to buy the licenses so that the total amount of
money spent is minimized. The running time of your algorithm shonld be
polynomial in n.
Solution Two natural guesses for a good sequence would be to sort the ri in
decreasing order, or to sort them in increasing order. Faced with alternatives
like this, its perfectly reasonable to work out a small example and see if the
example eliminates at least one of them. Here we could try rl = 2, r2 = 3, and
r3 = 4. Buying the licenses in increasing order results in a total cost of
100(2 -t- 32 4- 43) -= 7,500,

while buying them in decreasing order results in a total cost of


100(4 + 32 + 23) ---- 2300.

This tells us that increasing order is not the way to go. (On the other hand, it
doesnt tell us immediately that decreasing order is the right answer, but our
goal was just to eliminate one of the two options.)
Lets try proving that sorting the ri in decreasing order in fact always gives
the optimal solution. When a greedy algorithm works for problems like this,
in which we put a set of things in an optimal order, weve seen in the text that
its often effective to try proving correctness using an exchange argument.
To do this here, lets suppose that there is an optimal solution O that
differs from our solution S. (In other words, S consists of the licenses sorted in
decreasing order.) So this optimal solution O must contain an inversion--that
is, there must exist two neighboring months t and t + 1 such that the price
increase rate of the license bought in month t (let us denote it by rti is less
than that bought in month t + 1 (similarly, we use rt+l to denote this). That
is, we have rt < rt+l.
We claim that by exchanging these two purchases, we can strictly improve
our optimal solution, which contradicts the assumption that O was optimal.
Therefore if we succeed in showing this, we will successflflly show that ou.r
algorithm is indeed the correct one.
Notice that if we swap these two purchases, the rest of the purchafies
are identically priced. In O, the amount paid during the two months involved
in the swap is 100(r[ +q+u-t+~" On the other hand, if we swapped these two
purchases, we would pay 100(r~+~ + r~+~). Since the constant 100 is common

to both expressions, we want to show that the second term is less than the
first one. So we want to show that
t ~t+l t ~t+l

ri+~ + t < ri + lt+l


rt+ l ~t ~t+l
~t
t -- It < tt+l -- tt+l

r~(rt - I) < r~+t(rt+~ - I).


But this last inequality is true simply because ri > 1 for all i and since rt < rt+P
This concludes the proof of correctness. The running time of the algorithm
is O(n log n), since the sorting takes that much time and the rest (outputting)
is linear. So the overall running time is O(n log n).
Note: Its interesting to note that things become much less straightforward
if we vary this question even a little. Suppose that instead of buying licenses
whose prices increase, youre trying to sell off equipment whose cost is
depreciating. Item i depreciates at a factor of ri < I per month, starting from
$i00, so if you sell it t months from now you wil! receive 100.
t (In
ri. other
words, the exponential rates are now less than 1, instead of greater than 1.) If
you can only sell one item per month, what is the optimal order in which to
sell them? Here, it turns out that there are cases in which the optimal solution
doesnt put the rates in either increasing or decreasing order (as in the input
4 2

"

Solved Exercise 3
Suppose you are given a connected graph G, with edge costs that you may
assume are all distinct. G has n vertices and m edges. A particular edge e of G
is specified. Give an algorithm with running time O(m + n) to decide whether
e is contained in a minimum spanning tree of G.
Solution From the text, we know of two rules by which we can conclude
whether an edge e belongs to a minimum spanning tree: the Cut Property
(4.17) says that e is in every minimum spanning tree when it is the cheapest
edge crossing from some set S to the complement V - S; and the Cycle Property
(4.20) says that e is in no minimum spanning tree if it is the most expensive
edge on some cycle C. Lets see if we can make use of these two rules as part
of an algorithm that solves this problem in linear time. ,
Both the Cut and Cycle Properties are essentially talking about how e
relates to the set of edges that are cheaper than e. The Cut Property can be
viewed as asking: Is there some set S __ V so that in order to get from S to V - S
without using e, we need to use an edge that is more expensive than e? And
if we think about the cycle C in the statement of the Cycle Property, going the

187

190

Chapter 4 Greedy Algorithms


Prove that, for a given set of boxes with specified weights, the greedy
algorithm currently in use actually minimizes the number of trucks that
are needed. Your proof should follow the type of analysis we used for
the Interval Scheduling Problem: it should establish the optimality of this
greedy packing algorithm by identif34ng a measure under which it "stays
ahead" of all other solutions.
Some of your friends have gotten Into the burgeoning field of time-series
data mining, in which one looks for patterns in sequences of events that
occur over time. Purchases at stock exchanges--whats being bought-are one source of data with a natural ordering in time. Given a long
sequence S of such events, your friends want an efficient way to detect
certain "patterns" in them--for example, they may want to know if the
four events
buy Yahoo, buy eBay, buy Yahoo, buy Oracle

occur in this sequence S, in order but not necessarily consecutively.


They begin with a collection of possible events (e.g., the possible
transactions) and a sequence S of n of these events. A given event may
occur multiple times in S (e.g., Yahoo stock may be bought many times
In a single sequence S). We will say that a sequence S is a subsequence
of S if there is a way to delete certain of the events from S so that the
remaining events, in order, are equal to the sequence S. So, for example,
the sequence of four events above is a subsequence of the sequence
buy Amazon, buy Yahoo, buy eBay, buy Yahoo, buy Yahoo,
buy Oracle

Their goal is to be able to dream up short sequences and quickly


detect whether they are subsequences of S. So this is the problem they
pose to you: Give an algorithm that takes two sequences of even~s--S of
length m and S of length n, each possibly containing an event more than
once--and decides in time O(m + n) whether S is a subsequence of S.
Lets consider a long, quiet country road with houses scattered very
sparsely along it. (We can picture the road as a long line segment, with
an eastern endpoint and a western endpoint.) Further, lets suppose that
despite the bucolic setting, the residents of all these houses are avid cell
phone users. You want to place cell phone base stations at certain points
along the road, so that every house is within four miles of one of the base
stations.
Give an efficient algorithm that achieves this goal, using as few base
stations as possible.

Exercises
Your friend is working as a camp counselor, and he is in charge of
organizing activities for a set of junior-high-school-age campers. One of
his plans is the following mini-triathalon exercise: each contestant must
swim 20 laps of a pool, then bike 10 miles, then run 3 miles. The plan is
to send the contestants out in a staggered fashion, via the following rule:
the contestants must use the pool one at a lime. In other words, first one
contestant swims the 20 laps, gets out, and starts biking. As soon as this
first person is out of the pool, a second contestant begins swimming the
20 laps; as soon as he or she is out and starts biking, a third contestant
begins swimming.., and so on.)
Each contestant has a projected swimming time (the expected time it
will take him or her to complete the 20 laps), a projected biking time (the
expected time it will take him or her to complete the 10 miles of bicycling),
and a projected running time (the time it will take him or her to complete
the 3 miles of running). Your friend wants to decide on a schedule for the
triathalon: an order in which to sequence the starts of the contestants.
Lets say that the completion time of a schedul~ is the earliest time at
which all contestants will be finished with all three legs of the triathalon,
assuming they each spend exactly their projected swimming, biking, and
running times on the three parts. (Again, note that participants can bike
and run simultaneously, but at most one person can be in the pool at
any time.) Whats the best order for sending people out, if one wants the
whole competition to be over as early as possible? More precisely, give
an efficient algorithm that produces a schedule whose completion time
is as small as possible.
The wildly popular Spanish-language search engine E1 Goog needs to do
a serious amount of computation every time it recompiles its index. Fortunately, the company has at its disposal a single large supercomputer,
together with an essentia!ly unlimited supply of high-end PCs.
Theyve broken the overall computation into n distinct jobs, labeled
71, J2 ..... Jn, which can be performed completely Independently of one
another. Each job consists of two stages: first it needs to be preprocessed
on the supercomputer, and then it needs to be finished on one of the
PCs. Lets say that job J~ needs p~ seconds of time on. the supercomputer,
followed by f~ seconds of time on a PC.
Since there are at least n PCs available on the premises, the finishing
of the jobs can be performed fully in para!lel--all the jobs can be processed at the same time. However, the supercomputer can only work on
a single job at a time, so the system managers need to work out an order
in which to feed the jobs to the supercomputer. As soon as the first job

191

192

Chapter 4 Greedy Algorithms

in order is done on the supercomputer, it can be handed off to a PC for


finishing; at that point in time a second job can be fed to the supercompurer; when the second job is done on the supercomputer, it can proceed
to a PC regardless of whether or not the first job is done (since the PCs
work in parallel); and so on.
Lets say that a schedule is an ordering of the jobs for the supercomputer, and the completion time of the schedule is the earliest time at
which all jobs will have finished processing on the PCs. This is an important quantity to minimize, since it determines how rapidly E1 Goog can
generate a new index.
Give a polynomial-time algorithm that finds a schedule with as small
a completion time as possible.
8. Suppose you are given a connected graph G, with edge costs that are all
distinct. Prove that G has a tmique minimum spann~g tree.

One of the basic motivations behind theMinimum Spanning Tree Proble~fi


is the goal of designing a spanning network for a set of nodes with
minimum total cost. Herewe explore another type of objective: designing
a spanning network for which the most expensive edge is as cheap as
possible.
Specifically, let G -= (V, E) be a connected graph with n vertices, m
edges, and positive edge costs that you may assume are all distinct. Let
T = (V, E) be a spanning tree of G; we define the bottleneck edge of T to
be the edge of T with the greatest cost.
A spanning tree T of G is a minimum-bottleneck spanning tree ff there
is no spanning tree T of G with a cheaper bottleneck edge.
(a) Is every minimum-bottleneck tree of G a minimum spanning tree of
G? Prove or give a counterexample.
(b) Is every minimum spanning tree of G a minimum-bottleneck tree of
G? Prove or give a counterexample.
10. Let G = (V, E) be an (undirected) graph with costs ce >_ 0 on the edges e ~ E.
Assume you are given a minimum-cost spanning tree T in G. Now assume
that a new edge is added to G, connecting two nodes v, tv V with cost c.
(a) Give an efficient algorithm to test if T remains the minimum-cost
spanning tree with the new edge added to G (but not to the tree T).
Make your algorithm run in time O(IEI). Can you do it in O(IVI) time?
Please note any assumptions you make about what data structure is
used to represent the tree T and the graph G.

Exercises
Suppose T is no longer the minimum-cost spanning tree. Give a
linear-time algorithm (time O(IEI)) to update the tree T to the new
minLmum-cost spanning tree.
11. Suppose you are given a connected graph G = (V, E), with a cost ce on
each edge e. In an earlier problem, we saw that when all edge costs are
distinct, G has a unique minimum spanning tree. However, G may have
many minimum spanning trees when the edge costs are not all distinct.
Here we formulate the question: Can Kruskals Algorithm be made to find
all the minimum spanning trees of G?
RecaLl that Kxuskals Algorithm sorted the edges in order of increasing cost, then greedily processed edges one by one, adding an edge e as
long as it did not form a cycle. When some edges have the same cost, the
phrase "in order of increasing cost" has to be specified a little more carefully: weLl say that an ordering of the edges is valid if the corresponding
sequence of edge costs is nondecreasing. WeLl say that a valid execution
of Kruskals Algorithm is one that begins with a valid ordering of the
edges of G.

For any graph G, and any minimum spanning tree T of G, is there a


valid execution of Kruskals Algorithm onG that produces T as output?
Giv,e a proof or a countere.xample.
12. Suppose you have n video streams that need to be sent, one after another,
over a communication link. Stream i consists of a total of bi bits that need
to be sent, at a constant rate, over a period of ti seconds. You cannot send
two streams at the same time, so you need to determine a schedule for the
streams: an order in which to send them. Whichever order you choose,
there cannot be any delays between the end of one stream and the start
of the next. Suppose your schedule starts at time 0 (and therefore ends at
time ~1 ti, whichever order you choose). We assume that all the values
bi and t~ are positive integers.
Now, because youre just one user, the link does not want you taking
up too much bandwidth, so it imposes the following constraint, using a
fixed parameter r:
(,) For each natural number t > O, the total number of bits you send over the
time interval from 0 to t cannot exceed rt.
Note that this constraint is only imposed for time intervals that start at
0, not for time intervals that start at any other value.
We say that a schedule is valid if it satisfies the constraint (.) imposed
by the link.

193

194

Chapter 4 Greedy Algorithms


The Problem. Given a set of n streams, each specified by its number of
bits bi and its time duration ti, as well as the link parameter r, determine
whether there exists a valid schedule.
Example. Suppose we have n = 3 streams, with
(hi, q) = (2000, 1), (b2, t2) = (6000, 2), (b3, t3) = (2000, 1),
and suppose the links parameter is r = 5000. Then the schedule that runs
the streams in the order 1, 2, 3, is valid, since the constraint (.) is satisfied:
t = 1: the whole first stream has been sent, and 2000 < 5000.1
t = 2: half of the second stream has also been sent,
and 2000+ 5000 5000- 2
Similar calcalations hold for t = 3 and t = 4.

(a) Consider the following claim:


Claim: There exists a valid schedule if and only if each stream i satisfies
bi < rti.

(b)

Decide whether you think the claim is true or false, and give a proof
of either the claim or its negation.
Give an algorithm that takes a set of n streams, each specified by its
number of bits bi and its time duration ti, as well as the link parameter
r, and determines whether there exists a valid schedule. The rtmning
time of your algorithm should be polynomial in n.

A small business--say, a photocopying service with a single large


machine--faces the following scheduling problem. Each morning they
get a set of jobs from customers. They want to do the jobs on their single
machine in an order that keeps their customers happiest. Customer is
job will take ti time to complete. Given a schedule (i.e., an ordering of the
jobs), let Ci denote the finishing time of job i. For example, if job j is the
first to be donel we would have Ci = tj; and ff job j is done right after job
i, we would have Ci = Q + ti. Each customer i also has a given weight wg
~sents his or her importance to the business. The happiness of
customer i is expected to be dependent o~ the finishing time of is job.
So the company decides that they want to order the jobs to mJnimlze the
weighted sum of the completion times, ~,n
i=1 wiCi"
Design an efficient algorithm to solve this problem. That is, you are
given a set of n jobs with a processing time ti and a weight w~ for each
job. You want to order the jobs so as to minimize the weighted sum of
the completion times, ~P=I wiCiExample. Suppose there are two jobs: the first takes time q = ! and has
weight wl = !0, while the second job takes time t2 = 3 and has weight

Exercises
w2 = 2. Then doing job 1 first would yield a weighted completion time
of 10.1 + 2.4 = 18, while doing the second job first would yield the larger
weighted completion time of 10.4 + 2.3 = 46.

14. Youre working with a group of security consultants who are helping to
monitor a large computer system. Theres particular interest in keeping
track of processes that are labeled "sensitive." Each such process has a
designated start time and finish time, and it rtms continuously between
these times; the consultants have a list of the planned start and finish
times of al! sensitive processes that will be run that day.
As a simple first step, theyve written a program called s~ca~;us_check
that, when invoked, runs for a few seconds and records various pieces
of logging information about all the sensitive processes running on the
system at that moment. (Well model each invocation of status_check
as lasting for only this single point in time.) What theyd like to do is to
run status_check as few times as possible during the day, but enough
that for each sensitive process P, status_check is invoked at least once
during the execution of process P.
(a) Give an efficient algorithm that, given the start and finish times of
all the sensitive processes, finds as small a set of times as possible at which to invoke s~;a~cus_check, subject to the requirement
that s~a~cus_check is invoked at least once during each sensitive
process P.
(b) WtKle you were designing your algorithm, the security consultants
were engaging in a little back-of-the-envelope reasoning. "Suppose
we can find a set of k sensitive processes with the property that no
two are ever running at the same time. Then clearly your algorithm
will need to invoke s~ca~;us_check at least k times: no one invocation
of s~a~cus_check can handle more than one of these processes."
This is true, of course, and after some further discussion, you al!
begin wondering whether something stronger is true as well, a kind
of converse to the above argument. Suppose that k* is the largest
value of k such that one can find a set of k sensitive processes with
no two ever running at the same time. Is it the ~ase that there must
be a set of k* times at which you can run s~a~;us_check so that some
invocation occurs during the execution of each sensitive process? (In
other words, the kind of argument in the previous paragraph is really
the only thing forcing you to need a lot of invocations of
check.) Decide whether you think this claim is true or false, and give
a proof or a counterexample.

195

196

Exercises

Chapter 4 Greedy Algorithms


15. The manager of a large student union on campus comes to you with the
following problem. Shes in charge of a group of n students, each of whom
is scheduled to work one shift during the week. There are different jobs
associated with these shifts (tending the main desk, helping with package
delivery, rebooting cranky information kiosks, etc.), but.we can view each
shift as a single contiguous interval of time. There can be multiple shifts
going on at once.
Shes trying to choose a subset of these n students to form a supervising committee that she can meet with once a week. She considers such
a committee to be complete if, for every student not on the committee,
that students shift overlaps (at least partially) the shift of some student
who is on the committee. In this way, each students performance can be
observed by at least one person whos serving on the committee.
Give an efficient algorithm that takes the schedule of n shifts and
produces a complete supervising committee containing as few students
as possible.
Example. Suppose n = 3, and the shifts are
Monday 4 p.M.-Monday 8 P.M.,

Monday 6 p.M.-Monday 10 P.M.,


Monday 9 P.M.-Monday 1I P.M..
Then the smallest complete supervising committee would consist of just
the second student, since the second shift overlaps both the first and the
third.

16. Some security consultants wor~g in the financial domain are cur-

rently advising a client who is investigating a potential money-latmdering


scheme. The investigation thus far has indicated that n suspicious transactions took place in recent days, each involving money transferred into a
single account. Unfortunately, the sketchy nature of the evidence to date
means that they dont know the identiW of the account, the amounts of
the transactions, or the exact t~nes at which the transactions took place.
What they do have is an approximate time-stamp for each transaction; the
evidence indicates that transaction i took place at time ti ~: e~, for some
"margin of error" ev (In other words, it took place sometime between t~ - ei
and t~ + e~.) Note that different transactions may have different margins
of error.
In the last day or so, theyve come across a bank account that (for
other reasons we dont need to go into here) they suspect might be the
one involved in the crime. There are n recent events involving the account,
which took place at times Xl, x2 ..... xn. To see whether its plausible
that this really is the account theyre looking for, theyre wondering

whether its possible to associate each of the accounts n events with


a distinct one of the n suspicious transactions in such a way that, if the
account event at time x~ is associated with the suspicious transaction that
occurred approximately at time tj, then Itj - x~l <_ e~. (In other words, they
want to know if the activity on the account lines up with the suspicious
transactions to within the margin of error; the tricky part here is that
they dont know which account event to associate with which suspicious
transaction.)
Give an efficient algorithm that takes the given data and decides
whether such an association exists. If possible, you should make the
running time be at most O(n2).
17. Consider the following variation on the Interval Scheduling Problem. You
have a processor that can operate 24 hours a day, every day. People
submit requests to run daily jobs on the processor. Each such job comes
with a start time and an end time; if the job is accepted to run on the
processor, it must run conl~nuously, every day, for the period between
its start and end times. (Note that certain jobs can begin before midnight
and end after midnight; this makes for a type of situation different from
what we saw in the Interval Scheduling Problem.)
Given a list of n such jobs, your goal is to accept as many jobs as
possible (regardless of their length), subject to the constraint that the
processor can run at most one job at any given point in time. Provide an
algorithm to do this with a running time that is polynomial in n. You may
assume for simplicity that no two jobs have the same start or end times.
Example. Consider the fol!owing four jobs, specified by (start-time, endtime) pairs.
(6 P.M., 6 A.M.), (9 P.M., 4 A.M.), (3 A.M., 2 P.M.), (1 P.M., 7 P.M.).
The optimal solution would be to pick the two jobs (9 P.M., 4 A.M.) and (1
P.M., 7 P.~1.), which can be scheduled without overlapping.

18. Your friends are planning an expedition to a small town deep in the Canadian north next winter break. Theyve researched all the travel options
and have drawn up a directed graph whose nodes represent intermediate
destinations and edges represent the roads between them.
In the course of this, theyve also learned that extreme weather causes
roads in this part of the world to become quite slow in the winter and
may cause large travel delays. Theyve found an excellent travel Web site
that can accurately predict how fast theyll be able to trave_l along the
roads; however, the speed of travel depends on the time of year. More
precisely, the Web site answers queries of the following form: given an

197

198

Chapter 4 Greedy Algorithms

edge e = (u, w) connecting two sites v and w, and given a proposed starting
time t from location u, the site will return a value fe(t), the predicted
arrival time at w. The Web site guarantees that re(t) >_ t for all edges e
and all times t (you cant travel backward in time), and that fe(t) is a
monotone increasing function of t (that is, you do not arrive earlier by
starting later). Other than that, the functions fe(t) may be arbitrary. For
example, in areas where the travel time does not vary with the season,
we would have fe(t) = t + ee, where ee is the time needed to travel from the
beginning to the end of edge e.
Your friends want to use the Web site to determine the fastest way
to travel through the directed graph from their starting point to their
intended destination. (You should assume that they start at time 0, and
that all predictions made by the Web site are completely correct.) Give a
polynomial-time algorithm to do this, where we treat a single query to
the Web site (based on a specific edge e and a time t) as taking a single
computational step.

19. A group of network designers at the communications company CluNet


find themselves facing the following problem. They have a connected
graph G = (V, E), in which the nodes represent sites that want to communicate. Each edge e is a communication link, with a given available
bandwidth by
For each pair of nodes u, u ~ V, they want to select a single u-u path P
on which this pair will communicate. The bottleneck rate b(V) of this p athbV
is the minimumbandwidth of any edge it contains; that is, b(P) = mine~p e.
The best achievable bottleneck rate for the pair u, v in G is simply the
maximum, over all u-v paths P in G, of the value b(P).
Its getting to be very complicated to keep track of a path for each pair
of nodes, and so one of the network designers makes a bold suggestion:
Maybe one can find a spanning tree T of G so that for every pair of nodes
u, v, the unique u-v path in the tree actually attains the best achievable
bottleneck rate for u, v in G. (In other words, even if you could choose
any u-v path in the whole graph, you couldnt do better than the u-u path

In T.)
This idea is roundly heckled in the offices of CluNet for a few days,
and theres a natural reason for the skepticism: each pair of nodes
might want a very different-looking path to maximize its bottleneck rate;
why should there be a single tree that simultaneously makes everybody
happy? But after some failed attempts to rule out the idea, people begin
to suspect it could be possible.

Exercises
Show that such a tree exists, and give an efficient algorithm to find
one. That is, give an algorithm constructing a spanning tree T in which,
for each u, v v, the bottleneck rate of the u-v path in T is equal to the
best achievable bottleneck rate for the pair u, v in G.

20. Every September, somewhere In a far-away mountainous part of the


world, the county highway crews get together and decide which roads to
keep dear through thecoming winter. There are n towns in this county,
and the road system can be viewed as a (connected) graph G = (V, E) on
this set of towns, each edge representing a road joining two of them.
In the winter, people are high enough up in the mountains that they
stop worrying about the length of roads and start worrying about their
altitude--this is really what determines how difficult the trip will be.
So each road--each edge e in the graph--is annotated with a number
ue that gives the altitude of the highest point on the road. Well assume
that no two edges have exactly the same altitude value ae. The height of
a path P in the graph is then the maximum of ae over all edges e on P.
Fina~y, a path between towns i andj is declared tO be winter-optimal flit
achieves the minimum possible height over a~ paths from i to j.
The highway crews are goIng to select a set E ~ E of the roads to keep
dear through the winter; the rest will be left unmaintained and kept off
limits to travelers. They all agree that whichever subset of roads E they
decide to keep clear, it should have the properW that (v, E) is a connected
subgraph; and more strongly, for every pair of towns i and j, the height
of the winter-optimal path in (V, E) should be no greater than it is In the
fi~ graph G = (V, E). Well say that (V, E) is a minimum-altitude connected
subgraph if it has this property.

Given that theyre goIng to maintain ~s key property, however, they


otherwise want to keep as few roads clear as possible. One year, they hit
upon the following conjecture:
The minimum spanning tree of G, with respect to the edge weights ae, is a
minimum-altitude connected subgraph.
(In an earlier problem, we claimed that there is a unique minimum spanning tree when the edge weights are distinct. Thus, thanks to the assumption that all ae are distinct, it is okay for us to speak of the minimum
spanning tree.)
Initially, this conjecture is somewhat counterintuitive, sInce the minimum spanning tree is trying to minimize the sum of the values ae, while
the goal of minimizing altitude seems to be asking for a fully different
thing. But lacking an argument to the contrary, they begin considering an
even bolder second conjecture:

199

200

Exercises

Chapter 4 Greedy Algorithms


A subgraph (V, E) is a minimum-altitude connected subgraph if and only if
it contains the edges of the minimum spanning tree.
1

Note that this second conjecture would immediately imply the first one,
since a minimum spanning tree contains its own edges.
So heres the question.
(a) Is the first conjecture true, for all choices of G and distinct altitudes
at? Give a proof or a counterexample with e, xplanation.
(b) Is the second conjecture true, for all choices of G and distinct altitudes ae? Give a proof or a countere~xample with explanation.
21. Let us say that a graph G = (V, E) is a near-tree if it is connected and has at

most n + 8 edges, where n = IVI. Give an algorithm with running t~me O(n)
that takes a near-tree G with costs on its edges, and returns a minimum
spanning tree of G. You may assume that all the edge costs are distinct.

22. Consider the Minimum Spanning Tree Problem on an undirected graph

G = (V, E), with a cost ce >_ 0 on each edge, where the costs may not all
be different. If the costs are not a~ distinct, there can in general be
many distinct minimum-cost solutions. Suppose we are given a spanning
tree T c E with the guarantee that for every e ~ T, e belongs to some
minimum-cost spanning tree in G. Can we conclude that T itself must
be a minimum-cost spanning tree in G? Give a proof or a counterexample
with explanation.

23. Recall the problem of computing a minimum-cost arborescence in a

directed graph G = (V, E), with a cost ce >_ 0 on each edge. Here we will
consider the case in which G is a directed acyclic graph--that is, it contains
no directed cycles.
As in general directed graphs, there can be many distinct minimumcost solutions. Suppose we are given a directed acyclic graph G = (V, E),
and an arborescence A c E with the guarantee that for every e ~ A, e
belongs to some minimum-cost arborescence in G. Can we conclude that
A itself must be a minimum-cost arborescence in G? Give a proof or a
counterexample with explanation.

24. TimJ.ng circuits are a crucial component of VLSI chips. Heres a simple
model of such a timing circuit. Consider a complete balanced binary tree
with n leaves, where n is a power of two. Each edge e of the tree has an
associated length ~e, which is a positive number. The distance from the
root to a given leaf is the sum of the lengths of all the edges on the path
from the root to the leaf.

Figure 4.20 An instance of the zero-skew problem, described in Exercise 23.

The root generates a clock signal which is propagated along the edges
to the leaves. We]] assume that the time it takes for the signal to reach a
given leaf is proportional to the distance from the root to the leaf.
Now, if all leaves do not have the same distance from the root, then
the signal will not reach the leaves at the same time, and this is a big
problem. We want the leaves to be completely synchronized, and all to
receive the signal at the same time. To make this happen, we will have to
increase the lengths of certain edges, so that all root-to-leaf paths have
the same length (were not able to shrink edge lengths). If we achieve this,
then the tree (with its new edge lengths) will be said to have zero skew.
Our goal is to achieve zero skew in a way that keeps the sum of all the
edge lengths as small as possible.
Give an algorithm that increases the lengths of certain edges so that
the resulting tree has zero skew and the total edge length is as sma]] as
possible.
Example. Consider the tree in Figure 4.20, in which letters name the nodes
and numbers indicate the edge lengths.
The unique optimal solution for ~s instance would be to take the
three length-1 edges and increase each of their lengths to 2. The resulting
tree has zero skew, and the total edge length is 12, the smallest possible.
25. Suppose we are given a set of points P = [Pl,P2 ..... Pn}, together with a

distance function d on the set P; d is simply a function bn paJ_rs of points in


P with the properties that d(p~,pi) = d(py, Pi) > 0 ff i #j, and that d(p~, pi) = 0
for each i.
We define a hierarchical metric onP to be any distance function r that
can be constructed as fo]]ows. We build a rooted tree T with n leaves, and
we associate with each node v of T (both leaves and internal nodes) a
height hr. These heights must satisfy the properties that h(v) = 0 for each

201

Exercises

202

Chapter 4 Greedy Algorithms

leaf v, and ff u is the parent of v in T, then h(u) >_ h(v). We place each point
in P at a distinct leaf in T. Now, for any pair of points p~ and Pi, their
distance ~(p~, Pi) is defined as follows. We determine the least common
ancestor v in T of the leaves containing p~ and Pi, and define ~(p~,
We say that a hierarchical metric r is consistent with our distance
function d if, for all pairs i,j, we have r(p~,pl) _< d(p~,Pi).
Give a polynomial-time algorithm that takes the distance function d
and produces a hierarchical metric ~ with the following properties.
(i) ~ is consistent with d, and
<(ii) ff ~ is any other hierarchical metric consistent with d, then ~(P~,Pi)
r(p~,pi) for each pair of points Pi and
26. One of the first things you learn in calculus is how to minimize a differentiable function such as y = ax2 + bx + c, where a > 0. The Minimum
Spanning Tree Problem, on the other hand, is a minimization problem of
a very different flavor: there are now just a~ finite number of possibilities,
for how the minimum might be achieved--rather than a continuum of
possibilities--and we are interested in how to perform the computation
without having to exhaust this (huge) finite number of possibilities.
One Can ask what happens when these two minimization issues
are brought together, and the following question is an example of this.
Suppose we have a connected graph G = (V, E). Each edge e now has a timevarying edge cost given by a function fe :R-+R. Thus, at time t, it has cost
re(t). Wel! assume that all these functions are positive over their entire
range. Observe that the set of edges constituting the minimum spanning
tree of G may change over time. Also, of course, the cost of the minimum
spanning tree of G becomes a function of the time t; well denote this
function ca(t). A natural problem then becomes: find a value of t at which
cG(t) is minimized.
Suppose each function fe is a polynomial of degree 2: re(t) =aetz +
bet + Ce, where ae > 0. Give an algorithm that takes the graph G and the
values {(ae, be, ce) : e ~ E} and returns a value of the time t at which the
minimum spanning tree has minimum cost. Your algorithm should run
in time polynomial in the number of nodes and edges of the graph G. You
may assume that arithmetic operations on the numbers {(ae, be, q)} can
be done in constant time per operation.
27. In trying to understand the combinatorial StlXlcture of spanning trees,
we can consider the space of all possible spanning trees of a given graph
and study the properties of this space. This is a strategy that has been
applied to many similar problems as well.

Here is one way to do this. Let G be a connected graph, and T and T


two different spanning trees of G.. We say that T and T are neighbors if
T contains exactly one edge that is not in T, and T"contains exactly one
edge that is not in T.
Now, from any graph G, we can build a (large) graph 9~ as follows.
The nodes of 9~ are the spanning trees of G, and there is an edge between
two nodes of 9C if the corresponding spanning trees are neighbors.
Is it true that, for any connected graph G, the resulting graph ~
is connected? Give a proof that ~K is always connected, or provide an
example (with explanation) of a connected graph G for which % is not
connected.

28. Suppose youre a consultant for the networking company CluNet, and
they have the following problem. The network that theyre currently
working on is modeled by a connected graph G = (V, E) with n nodes.
Each edge e is a fiber-optic cable that is owned by one of two companies-creatively named X and Y--and leased to CluNet.
Their plan is to choose a spanning tree T of G and upgrade the links
corresponding to the edges of T. Their business relations people have
already concluded an agreement with companies X and Y stipulating a
number k so that in the tree T that is chosen, k of the edges will be owned
by X and n - k - 1 of the edges will be owned by Y.
CluNet management now faces the following problem. It is not at all
clear to them whether there even exists a spanning tree T meeting these
conditions, or how to find one if it exists. So this is the problem they put
to you: Give a polynomial-time algorithm that takes G, with each edge
labeled X or Y, and either (i) returns a spanning tree with e~xactly k edges
labeled X, or (ii) reports correctly that no such tree exists.
29. Given a list of n natural numbers all, d2 ..... tin, show how to decide
in polynomial time whether there exists an undirected graph G = (V, E)
whose node degrees are precisely the numbers d~, d2 ..... dn. (That is, ff
V = {Ul, v2 ..... vn}, then the degree of u~ should be exactly dv) G should not
contain multiple edges between the same pair of nodes, or "!oop" edges
with both endpoints equal to the same node.
30. Let G = (V, E) be a graph with n nodes in which each pair of nodes is
joined by an edge. There is a positive weight w~i on each edge (i,]); and
we will assume these weights satisfy the triangle inequality tv~k <_ ra~i + Wik.
For a subset V _ V, we will use G[V] to denote the subgraph (with edge
weights) induced on the nodes in V.

203

204

Chapter 4 Greedy Algorithms


We are given a set X _ V of k terminals that must be connected by
edges. We say that a Steiner tree onX is a set Z so that X ~_ Z _ V, together
with a spanning subtree T of G[Z]. The weight of the Steiner tree is the
weight of the tree T.
Show that the problem of finding a minimum-weight Steiner tree on
X can be solved in time

31. Lets go back to the original motivation for the Minimum Spanning Tree
Problem. We are given a connected, undirected graph G = (V, E) with
positive edge lengths {~e}, and we want to find a spanning subgraph of
it. Now suppose we are ~g to settle for a subgraph/4 = (V, F) that is
"denser" than a tree, and we are interested in guaranteeing that, for each
pair of vertices a, v ~ V, the length of the shortest u-v path in/4 is not
much longer than the length of the shortest a-v path in G. By the length
of a path P here, we mean the sum of ~e over all edges e in P.
Heres a variant of Kruskals Algorithm designed to produce such a
subgraph.
* First we sort all the edges in order of increasing length. (You may
assume all edge lengths are distinct.)
o We then construct a subgraph H = (V, F) by considering each edge in
order.
When we come to edge e = (u, v), we add e to the subgraph/4 if there
is currently no a-v path in/4. (This is what Kruskals Algorithm would
do as well.) On the other hand, if there is a u-v path in/4, we let duv
denote the length of the shortest such path; again, length is with
respect to the values {~e}. We add e to/4 ff 3~e < duvIn other words, we add an edge even when a and v are already In the same
connected component, provided that the addition of the edge reduces
their shortest-path distance by a sufficient amount.
Let H = (V, F) be the, subgraph of G returned by the algorithm.
(a) Prove that for evet3~ pair of nodes a, v ~ V, the length of the shortest
u-v path in H is at most three times the length of the shortest a-v
path in G.
(b) Despite its ability to approximately preserve shrtest-p ath distances
the subgraph/4 produced by the algorithm cannot be too dense.
Let f(n) denote the maximum number of edges that can possibly
be produced as the out-put of this algorithm, over all n-node input
graphs with edge lengths. Prove that

Notes and Further Reading

205

32. Consider a directed graph G = (V, E) with a root r ~ V and nonnegative


costs on the edges. In this problem we consider variants of the ~umcost arborescence algorithm.
(a) The algorithm discussed in Section 4.9 works as follows. We modify
the costs, consider the subgraph of zero-cost edges, look for a
directed cycle in this subgraph, and contract it (if one exists). Argue
briefly that instead of looking for cycles, we can instead identify and
contract strong components of this subgraph.
(b) In the course of the algorithm, we defined Yv to be the minimum
cost of an edge entering ~, and we modified the costs of all edges e
entering node u to be ce = ce - yr. Suppose we instead use the following modified cost: c~ = max(0, ce - 2y~). This new change is_likely to
turn more edges to 0 cost. Suppose now we find an arborescence T
of 0 cost. Prove that this T has cost at most twice the cost of the
minimum-cost arborescence in the original graph.
(c) Assume you do not find an arborescence of 0 cost. Contract al! 0cost strong components and recursively apply the same procedure
on the resttlting graph unti! an arborescence is found. Prove that this
T has cost at most twice the cost of the minimum-cost arborescence
in the original graph.
33. Suppose you are given a directed graph G = (V, E) In which each edge has
a cost of either 0 or 1. Also suppose that G has a node r such that there is a
path from r to every other node in G. You are also given an integer k. Give a
polynomial-time algorithm that either constructs an arborescence rooted
at r of cost exactly k, or reports (correctly) that no such arborescence
exists.

Notes and Further Reading


Due to their conceptual cleanness and intuitive appeal, greedy algorithms have
a long history and many applications throughout computer science. In this
chapter we focused on cases in which greedy algorithms find the optimal
solution. Greedy algorithms are also often used as simple heuristics even when
they are not guaranteed to find the optimal solution. In Chapter 11 we will
discuss greedy algorithms that find near-optimal approximate solutions.
As discussed in Chapter 1, Interval Scheduling can be viewed as a special
case of the Independent Set Problem on a graph that represents the overlaps
among a collection of intervals. Graphs arising this way are called interval
graphs, and they have been extensively studied; see, for example, the book
by Golumbic (1980). Not just Independent Set but many hard computational

206

Chapter 4 Greedy Mgofithms

problems become much more tractable when restricted to the special case of
interval graphs.
Interval Scheduling and the problem of scheduling to minimize the maximum lateness are two of a range of basic scheduling problems for which
a simple greedy algorithm can be shown to produce an optimal solution. A
wealth of related problems can be found in the survey by Lawier, Lenstra,
Rinnooy Kan, and Shmoys (1993).
The optimal algorithm for caching and its analysis are due to Belady
(1966). As we mentioned in the text, under real operating conditions caching
algorithms must make eviction decisions in real time without knowledge of
future requests. We will discuss such caching strategies in Chapter 13.
The algorithm for shortest paths in a graph with nonnegative edge lengths
is due to Dijkstra (1959). Surveys of approaches to the Minimum Spanning Tree
Problem, together with historical background, can be found in the reviews by
Graham and Hell (1985) and Nesetril (1997).
The single-link algorithm is one of the most~widely used approaches to,
the general problem of clustering; the books by Anderberg (1973), Duda, Hart,
and Stork (2001), and Jaln and Dubes (1981) survey a variety of clustering
techniques.
The algorithm for optimal prefix codes is due to Huffman (1952); the earlier approaches mentioned in the text appear in the books by Fano (1949) and
Shannon and Weaver (1949). General overviews of the area of data compression can be found in the book by Bell, Cleary, and Witten (1990) and the
survey by Lelewer and Hirschberg (1987). More generally, this topic belongs
to the area of information theory, which is concerned with the representation
and encoding of digital information. One of the founding works in this field
is the book by Shannon and Weaver (1949), and the more recent textbook by
Cover and Thomas (1991) provides detailed coverage of the subject..
The algorithm for finding minimum-cost arborescences is generally credited to Chu and Liu (1965) and to Edmonds (1967) independently. As discussed
in the chapter, this multi-phase approach stretches our notion of what constitutes a greedy algorithm. Itis also important from the perspective of linear
programming, since in that context it can be viewed as a fundamental application of the pricing method, or the primal-dual technique, for designing
algorithms. The book by Nemhauser and Wolsey (1988) develops these connections to linear program~ning. We will discuss this method in Chapter 11 in
the context of approximation algorithms.
More generally, as we discussed at the outset of the chapter, it is hard to
find a precise definition of what constitutes a greedy algorithm. In the search
for such a deflation, it is not even clear that one can apply the analogue

.~otes and Further Reading

of U.S. Supreme Court Justice Potter Stewarts famous test for obscenity-"I know it when I see it"--since one finds disagreements within the research
community on what constitutes the boundary, even intuitively, between greedy
and nongreedy algorithms. There has been research aimed at formalizing
classes of greedy algorithms: the theory of matroids is one very influential
example (Edmonds 1971; Lawler 2001); and the paper of Borodin, Nielsen, and
Rackoff (2002) formalizes notions of greedy and "greedy-type" algorithms, as
well as providing a comparison to other formal work on this quegtion.
Notes on the Exercises Exercise 24 is based on results of M. Edahiro, T. Chao,
Y. Hsu, J. Ho, K. Boese, and A. Kahng; Exercise 31 is based on a result of Ingo
Althofer, Gantam Das, David Dobkin, and Deborah Joseph.

207

Chapter
Divide artd Cortquer

Divide and conquer refers to a class of algorithmic techniques in which one


breaks the input into several parts, solves the problem ifi each part recursively,
and then combines the solutions to these subproblems into an overall solution.
In many cases, it can be a simple and powerful method.
Analyzing the running time of a divide and conquer algorithm generally
involves solving a recurrence relation that bounds the running time recursively
in terms of the running time on smaller instances. We begin the chapter with
a general discussion of recurrence relations, illustrating how they arise in the
analysis and describing methods for working out upper bounds from them.
We then illustrate the use of divide and conquer with applications to
a number of different domains: computing a distance function on different
rankings of a set of objects; finding the closest pair of points in the plane;
multiplying two integers; and smoothing a noisy signal. Divide and conquer
will also come up in subsequent chapters, since it is a method that often works
well when combined with other algorithm design techniques. For example, in
Chapter 6 we will see it combined with dynamic programming to produce a
space-efficient solution to a fundamental sequence comparison problem, and
in Chapter 13 we will see it combined with randomization to yield a simple
and efficient algorithm for computing the median of a set of numbers.
One thing to note about many settings in which divide and conquer
is applied, including these, is that the natural brute-force algorithm may
already be polynomial time, and the divide and conquer strategy is serving
to reduce the running time to a !ower polynomial. This is in contrast to most
of the problems in the previous chapters, for example, where brute force .was
exponential and the goal in designing a more sophisticated algorithm was to
achieve any kind of polynomial running time. For example, we discussed in

210

5.1 A First Recurrence: The Mergesort Algorithm

Chapter 5 Divide and Conquer

Chapter 2 that the natural brute-force algorithm for finding the closest pair
among n points in the plane would simply measure all (n2) distances, for
a (polynomial) running time of O(n2). Using divide and conquer, we wi!!
improve the running time to O(n log n). At a high level, then, the overall theme
of this chapter is the same as what weve been seeing earlier: that improving on
brute-force search is a fundamental conceptual hurdle in solving a problem
efficiently, and the design of sophisticated algorithms can achieve this. The
difference is simply that the distinction between brute-force search and an
improved solution here will not always be the distinction between exponential
and polynomia!.

5.1 A First Recurrence: The Mergesort Algorithm


To motivate the general approach to analyzing divide-and-conquer algorithms,
we begin with the Mergesort Algorithm. We discussed the Mergesort Algorithm
briefly in Chapter 2, when we surveyed common running times for algorithms.
Mergesort sorts a given list of numbers by first diviiting them into two equal
halves, sorting each half separately by recursion, and then combining the
results of these recursive calls--in the form of the two sorted halves--using
the linear-time algorithm for merging sorted lists that we saw in Chapter 2.
To analyze the running time of Mergesort, we will abstract its behavior into
the following template, which describes many common divide-and-conquer
algorithms.
(Q Divide the input into two pieces of equal size; solve the two subproblems
on these pieces separately by recursion; and then combine the two results
into an overall solution, spending only linear time for the initial division
and final recombining.

In Mergesort, as in any algorithm that fits this style, we also need a base case
for the recursion, typically having it "bottom out" on inputs of some constant
size. In the case of Mergesort, we will assume that once the input has been
reduced to size 2, we stop the- recursion and sort the two elements by simply
comparing them to each other.
Consider any algorithm that fits the pattern in (J-), and let T(n) denote its
worst-case running time on input instances of size n. Supposing that n is even,
the algorithm spends O(n) time to divide the input into two pieces of size n/2
each; it then spends time T(n/2) to solve each one (since T(n/2) is the worstcase nmning time for an input of size n/2); and finally it spends O(n) time
to combine the solutions from the two recursive calls. Thus the running time
T(n) satisfies the following recurrence relation.

(5.1)

For some constant c,


T(n) < 2T(n/2) + cn

when n > 2, and


T(2) _< c.

The structure of (5.1) is typical of what recurrences will look like: theres an
inequality or equation that bounds T(n) in terms of an expression involving
T(k) for sma!ler values k; and there is a base case that generally says that
T(n) is equal to a constant when n is a constant. Note that one can also write
(5.1) more informally as T(n)< 2T(n/2)+ O(n), suppressing the constant
c. However, it is generally useful to make c explicit when analyzing the
recurrence.
To keep the exposition simpler, we will generally assume that parameters
like n are even when needed. This is somewhat imprecise usage; without this
assumption, the two recursive calls would be on problems of size In/2] and
[n/2J, and the recurrence relation would say that
T(n) < T([n/2]) + T(Ln/2J) + cn

for n > 2. Nevertheless, for all the recurrences we consider here (and for most
that arise in practice), the asymptotic bounds are not affected by the decision
to ignore all the floors and ceilings, and it makes the symbolic manipulation
much cleaner.
Now (5.1) does not exphcitly provide an asymptotic bound on the growth
rate of the function T; rather, it specifies T(n) implicitly in terms of its values
on smaller inputs. To obtain an explicit bound, we need to solve the recurrence
relation so that T appears only on the left-hand side of the inequality, not the
fight-hand side as well.
Recurrence solving is a task that has been incorporated into a number
of standard computer algebra systems, and the solution to many standard
recurrences can now be found by automated means. It is still useful, however,
to understand the process of solving recurrences and to recognize which
recurrences lead to good running times, since the design of an efficient divideand-conquer algorithm is heavily intertwined with an understanding of how
a recurrence relation determines a running time.

Approaches to Solving Recurrences


There are two basic ways one can go about solving a recurrence, each of which
we describe in more detail below.

211

212

5.1 A First Recurrence: The M~rgesort Algorithm

Chapter 5 Divide and Conquer

The most intuitively natural way to search for a solution to a recurrence is


to "unroll" the recursion, accounting for the running time across the first
few levels, and identify a pattern that can be continued as the recursion
expands. One then sums the running times over all levels of the recursion
(i.e., until it "bottoms out" on subproblems of constant size) and thereby
arrives at a total running time.
A second way is to start with a guess for the solution, substitute it into
the recurrence relation, and check that it works. Formally, one justifies
this plugging-in using an argument by induction on n. There is a useful
variant of this method in which one has a general form for the solution,
but does not have exact values for all the parameters. By leaving these
parameters unspecified in the substitution, one can often work them out
as needed.
We now discuss each of these approaches, using the recurrence in (5.1) as an
example.

Unrolling the Mergesort Recurrence


Lets start with the first approach to solving the recurrence in (5.1). The basic
argument is depicted in Figure 5.1.
o Analyzing the first few levels: At the first level of recursion, we have a
single problem of size n, which takes time at most cn plus the time spent
in all subsequent rect~sive calls. At the next level, we have two problems
each of size n/2. Each of these takes time at most cn/2, for a total of at
most cn, again plus the time in subsequent recursive calls. At the third
level, we have four problems each of size n/4, each taking time at most
cn/4, for a total of at most cn.
Level 0: cn

Level 1:crt/2 + crt/2 = cn total

Level 2: 4(cn/4) = cn total

Figure 5.1 Unrolling the recurrence T(n) < 2T(n/2) + O(n).

Idengfying a pattern: Whats going on in general? At level j of the


recursion, the number of subproblems has doubled j times, so there are
now a total of 2J. Each has correspondingly shrunk in size by a factor
of two j times, and so each has size n/2J, and hence each takes time at
most cn/2J. Thus level j contributes a total of at most 2~(cn/2~) = cn to
the total running time.
Summing over all levels of recursion: Weve found that the recurrence
in (5.1) has the property that the same upper bound of cn applies to
total amount Of work performed at each level. The number of times the
input must be halved in order to reduce its size from n to 2 is log2 n.
So summing the cn work over log n levels of recursion, we get a total
running time of O(n log n).
We summarize this in the following claim.
(5.2) Any function T(.) satisfying (5.1) is bounded by O(n log n), when
n>l.

Substituting a Solution into the Mergesort Recurrence


The argument establishing (5.2) can be used to determine that the function
T(n) is bounded by O(n log n). If, on the other hand, we have a guess for
the running time that we want to verify, we can do so by plugging it into the
recurrence as follows.
Suppose we believe that T(n) < cn log2 n for all n > 2, and we want to
check whether this is indeed true. This clearly holds for n = 2, since in this
case cnlog2 n = 2c, and (5.1) explicitly tells us that T(2) < c. Now suppose,
by induction, that T(m) <_ cm log2 m for all values of m less than n, and we
want to establish this for T(n). We do this by writing the recurrence for T(n)
and plugging in the inequality T(n/2) <_ c(n/2) log2(n/2). We then simplify the
resulting expression by noticing that log2(n/2) = (log2 n) - 1. Here is the ftfll
calculation.
T(n) < 2T(n/2) + cn
< 2c(n/2) loga(n/2) + cn
= cn[(log2 n) - 1] + cn
= (cn log2 n) - cn + cn
= cn log2 n.
This establishes the bound we want for T(n), assuming it holds for smaller
values m < n, and thus it completes the induction argument.

213

214

5.2 Further Recurrence Relations

Chapter 5 Divide and Conquer

An Approach Using Partial Substitution


There is a somewhat weaker kind of substitution one can do, in which one
guesses the overall form of the solution without pinning down the exact values
of all the constants and other parameters at the outset.
Specifically, suppose we believe that T(n)= O(n log n), but were not
sure of the constant inside the O(-) notation. We can use the substitution
method even without being sure of this constant, as follows. We first write
T(n) <_ kn logb n for some constant k and base b that well determine later.
(Actually, the base and the constant well end up needing are related to each
other, since we saw in Chapter 2 that one can change the base of the logarithm
by simply changing the multiplicative constant in front.)
Now wed like to know whether there is any choice of k and b that wiJ!
work in an inductive argument. So we try out one level of the induction as
follows.
T(n) < 2T(n/2) + cn < 2k(n/2) lOgb(n/2) + cn.
Its now very tempting to choose the base b = 2 for the logarithm, since we see
that this wil! let us apply the simplification logz(n/2) = (log2 n) - 1. Proceeding
with this choice, we have
T(n) <_ 2k(n/2) log2(n/2) + cn
= 2k(n/2) [(log2 n) - 1] + cn
= krt[(log2 n) - 1] + cn
= (kn log2 n) - kn + cn.
Finally, we ask: Is there a choice of k that will cause this last expression to be
bounded by kn log2 n? The answer is clearly yes; we iust need to choose any
k that is at least as large as c, and we get
T(n) < (kn log2 n) - kn + cn <_ kn !og2 n,
which completes the induction.
Thus the substitution method can actually be usefl~ in working out the
exact constants when one has some guess of the general form of the solution.

5.2 Further Recurrence Relations


Weve just worked out the solution to a recurrence relation, (5.1), that will
come up in the design of several divide-and-conquer algorithms later in this
chapter. As a way to explore this issue further, we now consider a class
of recurrence relations that generalizes (5.1), and show how to solve the
recurrences in this class. Other members of this class will arise in the design
of algorithms both in this and in later chapters.

This more general class of algorithms is obtained by considering divideand-conquer algorithms that create recursive calls on q subproblems of size
n/2 each and then combine the results in O(n) time. This corresponds to
the Mergesort recurrence (5.1) when q = 2 recursive calls are used, but other
algorithms find it useful to spawn q > 2 recursive calls, or just a single (q = 1)
recursive call. In fact, we will see the case q > 2 later in this chapter when we
design algorithms for integer multiplication; and we will see a variant on the
case q = 1 much later in the book, when we design a randomized algorithm
for median finding in Chapter 13.
If T(n) denotes the nmning time of an algorithm designed in this style,
then T(n) obeys the following recurrence relation, which directly generalizes
(5.1) by replacing 2 with q:
(5.3)

For some constant c,


T(n) <_ qT(n/2) + cn

when n > 2, and


T(2) < c.

We now describe how to solve (5.3) by the methods weve seen above:
unrolling, substitution, and partial substitution. We treat the cases q > 2 and
q = 1 separately, since they are qualitatively different from each other--and
different from the case q = 2 as well.
The Case of q > 2 Subproblems
We begin by unro!ling (5.3) in the case q > 2, following the style we used
earlier for (5.1). We will see that the punch line ends up being quite different.
Analyzing the first few levels: We show an example of this for the case
q = 3 in Figure 5.2. At the first level of recursion, we have a single
problem of size n, which takes time at most cn plus the time spent in all
subsequent recursive calls. At the next level, we have q problems, each
of size n/2. Each of these takes time at most cn/2, for a total of at most
(q/2)cn, again plus the time in subsequent recursive calls. The next level
yields q2 problems of size n/4 each, for a total time of (q2/4)cn. Since
q > 2, we see that the total work per level is increasing as we proceed
through the recursion.
Identifying apattern: At an arbitrary levelj, we have qJ distinct instances,
each of size n/2]. Thus the total work performed at level j is qJ(cn/2]) =
(q/2)icn.

215

5.2 Further Recurrence Relations

Chapter 5 Divide and Conquer

216

(5.4)

Level O: cn total

cn time plus
recursive calls

Level 1:cn/2 + cn/2 + cn/2 = (3/2)cn total

Any function T(.) satisfying (5.3) with q > 2 is bounded by O(nlga q).

So we find that the running time is more than linear, since log2 q > !,
but still polynomial in n. Plugging in specific values of q, the running time
is O(nlg~ 3) = O(nl.sg) when q = 3; and the running time is O(nlg~ 4) = O(n2)
when q = 4. This increase in running time as q increases makes sense, of
course, since the recursive calls generate more work for larger values of q.

Level 2: 9(cn/4) = (9/4)cn total

Figure 5.2 Unrolling the recurrence T(n) < 3T(n/2) + O(rt).

Summing over all levels of recursion: As before, there are log). n levels of
recursion, and the total amount of work performed is the sum over-all
these:

This is a geometric sum, consisting of powers of r = q/2. We can use the


formula for a geometric sum when r > 1, which gives us the formula

Applying Partial Substitution The appearance of log2 q in the exponent


followed naturally from our solution to (5.3), but its not necessarily an
expression one would have guessed at the outset. We now consider how an
approach based on partial substitution into the recurrence yields a different
way of discovering this exponent.
Suppose we guess that the solution to (5.3), when q > 2, has the form
T(n) <_ kna for some constants k > 0 and d > 1. This is quite a general guess,
since we havent even tried specifying the exponent d of the polynomial. Now
lets try starting the inductive argument and seeing what constraints we need
on k and d. We have
T(n) <_ qT(n/2) + cn,
and applying the inductive hypothesis to T(n/2), this expands to

r(n) <_cn
\ r-n _!.1)< cn
rlogz
+ cn
Since were aiming for an asymptotic upper bound, it is useful to figure
out whats simply a constant; we can pull out the factor of r - 1 from
the denominator, and write the last expression as
T(n) <_

nrlog2 n

Finally, we need to figure out what rlg2 n is. Here we use a very handy
identity, which says that, for any a > 1 and b > 1, we have alg b = blOg a
Thus
rlog2 n _~_ nlog2 r = nlog2(q/2) = n(logz q)-l.

Thus we have
T(n) <_

n n(lg2 q)--I <_

We sum this up as follows.

nlog2 q = O(nlg2 q).

= q, knd + cn.
2a

This is remarkably close to something that works: if we choose d so that


q/2d = 1, then we have T(n) < knd + cn, which is almost right except for the
extra term cn. So lets deal with these two issues: first, how to choose d so we
get q/2a = 1; and second, how to get rid of the cn term.
Choosing d is easy: we want 2d = q, and so d = log2 q. Thus we see that
the exponent log2 q appears very naturally once we dec.ide to discover which
value of d works when substituted into the recurrence.
But we still have to get rid of the cn term. To do this, we change the
form of our guess for T(n) so as to explicitly subtract it off. Suppose we try
the form T(n) <_ kna - gn, where weve now decided that d = log2 q but we
havent fixed the constants k or g. Applying the new formula to T(n/2), this
expands to

217

218

Chapter 5 Divide and Conquer

5.2 Further Recurrence Relations


cn time, plus
recursive calls

+ cn
q. knd - q~ n + cn
2a
2

Level 0: cn total

= knd -- q~n + cn

/~ Level 1:cn/2 total

= l~na - (~ - c)n.

~/~ Level2:cn/4total

This now works completely, if we simply choose ~ so that (~ - c) = ~: in other


words, ~ = 2c/(q - 2). This completes the inductive step for n. We also need
to handle the base case n = 2, and this we do using the fact that the value of
k has not yet been fixed: we choose k large enough so that the formula is a
valid upper bound for the case n = 2.

Figure 5.3 Unrolling the recurrence T(n) <_ T(n/2) + O(n).

We sum this up as follows.

The Case of One Subproblem


We now consider the case of q = I in (5.3), since this illustrates an outcome
of yet another flavor. While we wont see a direct application of the recurrence
for q = 1 in this chapter, a variation on it comes up in Chapter 13, as we
mentioned earlier.
We begin by unrolling the recurrence to try constructing a solution.
Analyzing the first few levels: We show the first few levels of the recursion
in Figure 5.3. At the first level of recursion, we have a single problem of
size n, which takes time at most cn plus the time spent in all subsequent
recursive calls. The next level has one problem of size n/2, which
contributes cn/2, and the level after that has one problem of size n/4,
which contributes cn/4. So we see that, unlike the previous case, the total
work per leve! when q = 1 is actually decreasing as we proceed .through
the recursion.
Identifying a pattern: At an arbitrary level j, we st1 have just one
instance; it has size n/21 and contributes cn/21 to the running time.
Summing over all levels of recursion: There are log2 n levels of recursion,
and the total amount of work performed is the sum over all these:
lg2 n-1cn

T(n) <

lg2 n-1 (~./

-~- =cn 1=o


E

"

This geometric sum is very easy to work out; even if we continued it to


infinity, it would converge to 2. Thus we have
T(n) <_ 2cn = O(n).

This is counterintuitive when you first see it. The algorithm is performing
log n levels of recursion, but the overall running time is still linear in n. The
point is that a geometric series with a decaying exponent is a powerfl~ thing:
fully half the work performed by the algorithm is being done at the top level
of the recursion.
It is also useful to see how partial substitution into the recurrence works
very well in this case. Suppose we guess, as before, that the form of the solution
is T(n) <_ kna. We now try to establish this by induction using (5.3), assuming
that the solution holds for the smaller value n/2:
T(n) <_ T(n/2) + cn
+cn
~ ~d nd -}- on.

If we now simply choose d = ! and k = 2c, we have


k
T(n) <_ ~n + cn = (-~ + c)n = kn,
which completes the induction.
The Effect of the Parameter q. It is worth reflecting briefly on the role of the
parameter q in the class of recurrences T(n) <_ qT(n/2) + O(n) defined by (5.3).
When q = 1, the restflting running time is linear; when q = 2, its O(n log n);
and when q > 2, its a polynomial bound with an exponent larger than I that
grows with q. The reason for this range of different running times lies in where

219

220

5.3 Counting Inversions

Chapter 5 Divide and Conquer

most of the work is spent in the recursion: when q = 1, the total running time
is dominated by the top level, whereas when q > 2 its dominated by the work
done on constant-size subproblems at the bottom of the recursion. Viewed this
way, we can appreciate that the recurrence for q = 2 really represents a "knifeedge"--the amount of work done at each level is exactly the same, which is
what yields the O(n log n) running time.
A Related Recurrence: T(n) <_ 2T(n/2) + O(n2)
We conclude our discussion with one final recurrence relation; it is illustrative
both as another application of a decaying geometric sum and as an interesting
contrast with the recurrence (5.1) that characterized Mergesort. Moreover, we
wil! see a close variant of it in Chapter 6, when we analyze a divide-andconquer algorithm for solving the Sequence Alignment Problem using a small
amount of working memory.
The recurrence is based on the following divide-and-conquer structure.
Divide the input into two pieces of equal size; solve the two subproblems,
on these pieces separately by recursion; and then combine the two results
into an overall solution, spending quadratic time for the initial division
and final recombining.
For our proposes here, we note that this style of algorithm has a running time
T(n) that satisfies the fo!lowing recurrence.
(5.6)

For some constant c,


T(n) <_ 2T(n/2) + cn2

when n > 2, and


T(2) < c.

Ones first reaction is to guess that the solution will be T(n) = O(n2 log n),
since it looks almost identical to (5.1) except that the amount of work per level
is larger by a factor equal to the input size. In fact, this upper bound is correct
(it would need a more careful argument than whats in the previous sentence),
but it will turn out that we can also show a stronger upper bound.
Well do this by unrolling the recurrence, following the standard template
for doing this.
o Analyzing the first few levels: At the first level of recursion, we have a
single problem of size n, which takes time at most cn2 plus the time spent
in al! subsequent recursive calls. At the next level, we have two problems,
each of size n/2. Each of these takes time at most c(n/2)2 = cn2/4, for a

total of at most cn2/2, again plus the time in subsequent recursive calls.
At the third level, we have four problems each of size n/4, each taking
time at most c(n/4)2 = cn2/16, for a total of at most cn2/4. Already we see
that something is different from our solution to the analogous recurrence
(5.!); whereas the total amount of work per level remained the same in
that case, here its decreasing.
Identifying a pattern: At an arbitrary level j of the recursion, there are 2J
subproblems, each of size n/2J, and hence the total work at this level is
bounded by 2Jc@)2 = cn2/2j.
Summing over all levels of recarsion: Having gotten this far in the calculation, weve arrived at almost exactly the same sum that we had for the.
case q = 1 in the previous recurrence. We have
T(n) <_

cn2
1og2 n~-I (~.)
: = cn2
23
j=0

log~-I

j=0

< 2cn2 = O(n2),

where the second inequality follows from the fact that we have a convergent geometric sum.
In retrospect, our initial guess of T(n) = O(n2 log n), based on the analogy
to (5.1), was an overestimate because of how quickly n2 decreases as we
replace it with (~)2, (n)2,~ (~)~, and so forth in the unrolling of the recurrence.
This means that we get a geometric sum, rather than one that grows by a fixed
amount over all n levels (as in the solution to (5.1)).

5.3 Counting Inversions


Weve spent some time discussing approaches to solving a number of common
recurrences. The remainder of the chapter will illustrate the application of
divide-and-conquer to problems from a number of different domains; we will
use what weve seen in the previous sections to bound the running times
of these algorithms. We begin by showing how a variant of the Mergesort
technique can be used to solve a problem that is not directly related to sorting
numbers.

~ The Problem
We will consider a problem that arises in the analysis of rankings, which
are becoming important to a number of current applications. For example, a
number of sites on the Web make use of a technique known as collaborative
filtering, in which they try to match your preferences (for books, movies,
restaurants) with those of other people out on the Internet. Once the Web site
has identified people with "similar" tastes to yours--based on a comparison

221

Chapter 5 Divide and Conquer

222

of how you and they rate various things--it can recommend new things that
these other people have liked. Another application arises in recta-search tools
on the Web, which execute the same query on many different search engines
and then try to synthesize the results by looking for similarities and differences
among the various rankings that the search engines return.
A core issue in applications like this is the problem of comparing two
rankings. You rank a set of rt movies, and then a collaborative filtering system
consults its database to look for other people who had "similar" rankings. But
whats a good way to measure, numerically, how similar two peoples rankings
are? Clearly an identical ranking is very similar, and a completely reversed
ranking is very different; we want something that interpolates through the
middle region.

Lets consider comparing your ranking and a strangers ranking of the


same set of n movies. A natural method would be to label the movies from
1 to n according to your ranking, then order these labels according to the
strangers ranking, and see how many pairs are "out of order." More concretely,,
we will consider the following problem. We are given a sequence of rt numbers
art; we will assume that all the numbers are distinct. We want to define
a measure that tells us how far this list is from being in ascending order; the
value of the measure should be 0 if al < a2 < K an, and should increase as
the numbers become more scrambled.

A natural way to quantify this notion is by counting the number of


inversions. We say that two indices i < j form an inversion if ai > aj, that is,
if the two elements ai and aj are "out of order." We will seek to determine the
number of inversions in the sequence a~ ..... art.

Just to pin down this definition, consider an example in which the sequence is 2, 4, 1, 3, 5. There are three inversions in this sequence: (2, 1), (4, 1),
and (4, 3). There is also an appealing geometric way to visualize the inversions, pictured in Figure 5.4: we draw the sequence of input numbers in the
Figure 5.4 Counting the
number of inversions in the
order theyre p~ovided, and below that in ascending order. We then draw a
sequence 2, 4, 1, 3, 5. Each
line segment between each number in the top list and its copy in the lower
crossing pair of line segments
list. Each crossing pair of line segments corresponds to one pair that is in the
corresponds to one pair that
is in the opposite order in
opposite order in the two lists--in other words, an inversion.
the input list and the ascending list--in other words, an
inversion.

Note how the number of inversions is a measure that smoothly interpolates


between complete agreement (when the sequence is in ascending order, then
there are no inversions) and complete disagreement (if the sequence is in
descending order, then every pair forms an inversion, and so there are (~) of

5.3 Counting Inversions

~ Designing and Analyzing the Algorithm


What is the simplest algorithm to count inversions? Clearly, we could look
at ~very pair of numbers (ai, aj) and determine whether they constitute an
inversion; this would take O(n2) time.
We now show how to count the number of inversions much more quickly,
in O(n log n) time. Note that since there can be a quadratic number of inversions, such an algorithm must be able to compute the total number without
ever looking at each inversion individually. The basic idea is to follow the
strategy (]-) defined in Section 5.1. We set m = [n/2] and divide the list into
the two pieces a~ ..... am and ara+l ..... an. We first count the number of
inversions in each of these two halves separately. Then we count the number
of inversions (az, aj), where the two numbers belong to different halves; the
trick is that we must do this part in O(n) time, if we want to apply (5.2). Note
that these first-half/second-half inversions have a particularly nice form: they
are precisely the pairs (a,, aj), where ai is in the first half, aj is in the second
half, and ai > aj.
To help with counting the number of inversions between the two halves,
we will make the algorithm recursively sort the numbers in the two halves as
well. Having the recursive step do a bit more work (sorting as wel! as counting
inversions) will make the "combining" portion of the algorithm easier.
So the crucial routine in this process is Nerge-and-Cotmt. Suppose we
have recursively sorted the first and second halves of the list and counted the
inversions in each. We now have two sorted lists A and B, containing the first
and second halves, respectively. We want to produce a single sorted list C from
their union, while also counting the number of pairs (a, b) with a ~ A, b ~ B,
and a > b. By our previous discussion, this is precisely what we will need
for the "combining" step that computes the number of first-half/second-half
inversions.
This is closely related to the simpler problem we discussed in Chapter 2,
which formed the corresponding "combining" step for Mergeso.rt: there we had
two sorted lists A and B, and we wanted to merge them into a single sorted list
in O(n) time. The difference here is that we want to do something extra: not
only should we produce a single sorted list from A and B, but we should also
count the number of "inverted pairs" (a, b) where a ~ A,, b ~ B, and a > b.
It turns out that we will be able to do this in very much the same style
that we used for merging. Our Merge-and-Count routine will walk through
the sorted lists A and B, removing elements from the front and appending
them to the sorted list C. In a given step, we have a Current pointer into each
list, showing our current position. Suppose that these pointers are currently

223

224

5.4 Finding the Closest Pair of Points

Chapter 5 Divide and Conquer


Elements inverted
with by < ai

0nce one list is empty, append the remainder of the other list
to the output
Return Count and the merged list

]A
]B
Figure 5.5 Merging two sorted fists while also counting the number of inversions
between them.

at elements, ai and bi. In one step, we compare the elements ai and by being
pointed to in each list, remove the smaller one from its list, and append it to
the end of list C.
This takes care of merging. How do we also count the number of inversions.~ Because A and B are sorted, it is actually very easy to keep track of the
number of inversions we encounter. Every time the element a~ is appended to
C, no new inversions are encountered, since a~ is smaller than everything left
in list B, and it comes before all of them. On the other hand, if bI is appended
to list C, then it is smaller than all the remaining items in A, and it comes
after all of them, so we increase our count of the number of inversions by the
number of elements remaining in A. This is the crucial idea: in constant time,
we have accounted for a potentially large number of inversions. See Figure 5.5
for an illustration of this process.
To summarize, we have the following algorithm.
Merge-and-Count(A,B)
Maintain a Cuwent pointer into each list, initialized to
point to the front elements
Maintain a variable Count for the number of inversions,
initialized to 0
While both lists are nonempty:
Let ai and ~ be the elements pointed to by the Cuwent pointer
Append the smaller of these two to the output list
If ~ is the smaller element then
Increment Count by the number of elements remaining in A

Endif
Advance the Cu~ent pointer in the list from which the
smaller element was selected.
EndWhile

The rimming time of Merge-and-Count can be bounded by the analogue


of the argument we used for the original merging algorithm at the heart of
Mergesort: each iteration of the While loop takes constant time, and in each
iteration we add some element to the output that will never be seen again.
Thus the number of iterations can be at most the sum of the initial lengths of
A and B, and so the total running time is O(n).
We use this Merge-and-Count routine in a recursive procedure that
simultaneously sorts and counts the number of inversions in a list L.
Sort-and-Count (L)
If the list has one element then
there are no inversions
Else
Divide the list into two halves:
A contains the first [rt/2] elements
B contains the remaining [n/2J elements
(rA, A) = Sort-and-Count (A)
(rB, B) = Sort-and-Count (B)
(r, L) = Merge-and-Count (A, B)
Endif
Return r=rA+rB+r, and the sorted list L

Since our Merge-and-Count procedure takes O(n) time, the rimming time
T(n) of the full Sort-and-Count procedure satisfies the recurrence (5.1). By
(5.2), we have
(S.7) The Sort-and-Count algorithm correctly sorts the input list and counts
the number of inversions; it runs in O(n log n) time for a list with n elements:

5.4 Finding the losest Pair of Points


We now describe another problem that can be solved by an algorithm in the
style weve been discussing; but finding the right way to "merge" the solutions
to the two subproblems it generates requires quite a bit of ingenuity.

225

226

5.4 Finding the Closest Pair of Points

Chapter 5 Divide and Conquer

~ The Problem
The problem we consider is very simple to state: Given rt points in the plane,
find the pair that is closest together.
The problem was considered by M. I. Shamos and D. Hoey in the early
1970s, as part of their proiect to work out efficient algorithms for basic computational primitives in geometry. These algorithms formed the foundations
of the then-fledgling field of compatational geometry, and they have found
their way into areas such as graphics, computer vision, geographic information systems, and molecular modeling. And although the closest-pair problem
is one of the most natural algorithmic problems in geometry, it is sm~risingly
hard to find an efficient algorithm for it. It is immediately clear that there is an
O(n2) solution--compute the distance between each pair of points and take
the minimum--and so Shamos and Hoey asked whether an algorithm asymptotically faster than quadratic could be found. It took quite a long time before
they resolved this question, and the O(n log n) algorithm we give below is
essentially the one they discovered. In fact, when we return to this problem in
Chapter 13, we wi!l see that it is possible to further improve the running fim~
to O(n) using randomization.

/::~ Designing the Algorithm


We begin with a bit of notation. Let us denote the set of points by P =
{Pl ..... Pn}, where Pi has coordinates (x;, Yi); and for two points Pi, Pj E P,
we use d(p~, pj) to denote the standard Euclidean distance between them. Our
goal is to find a pair of points pi, pl that minimizes d(pi, p1).
We will assume that no two points in P have the same x-coordinate or
the same y-coordinate. This makes the discussion cleaner; and its easy to
eliminate this assumption either by initially applying a rotation to the points
that makes it ~e, or by slightly extending the algorithm we develop here.
Its instructive to consider the one-dimensional version of this problem for
a minute, since it is much simpler and the contrasts are revealing. How would
we find the closest pair of points on a line? Wed first sort them, in O(n log n)
time, and then wed walk through the sorted list, computing the distance from
each point to the one that comes after it. It is easy to see that one of these
distances must be the minimum one.
In two dimensions, we could try sorting the points by their y-coordinate
(or x-coordinate) and hoping that the two closest points were near one another
in the order of this sorted list. But it is easy to construct examples in which they
are very far apart, preventing us from adapting our one-dimensional approach.
Instead, our plan will be to apply the style of divide and conquer used
in Mergesort: we find the closest pair among the points in the "left half" of

P and the closest pair among the points in the "right half" of P; and then we
use this information to get the overall solution in linear time. If we develop an
algorithm with this structure, then the solution of our basic recurrence from
(5.1) will give us an O(n log n) running time.
It is the last, "combining" phase of the algorithm thats tricky: the distances
that have not been considered by either of our recursive calls are precisely those
that occur between a point in the left half and a point in the right half; there
are S2 (n2) such distances, yet we need to find the smallest one in O(n) time
after the recursive calls return. If we can do this, our solution will be complete:
it will be the smallest of the values computed in the recursive calls and this
minimum "left-to-right" distance.
Setting Up the Recursion Lets get a few easy things out of the way first.
It will be very useful if every recursive call, on a set P c_ p, begins with two
lists: a list p.t~ in which a~ the points in P have been sorted by increasing xcoordinate, and a list P; in which all the points in P have been sorted by
increasing y-coordinate. We can ensure that this remains true throughout the
algorithm as follows.

First, before any of the recursion begins, we sort all the points in P by xcoordinate and again by y-coordinate, producing lists Px and Py. Attached to
each entry in each list is a record of the position of that point in both lists.
The first level of recursion will work as follows, with all further levels
working in a completely analogous way. We define O to be the set of points
in the first In/2] positions of the list Px (the "left half") and R to be the set of
points in the final [n/2J positions of the list Px (the "right half"). See Figure 5.6.
By a single pass through each of Px and Py, in O(n) time, we can create the

Lim L
O
O

o
o

Figure 5.6 The first level of recursion: The point set P is divided evenly into Q and R by
the line L, and the closest pair is found on each side recursively.

227

228

Chapter 5 Divide and Conquer

following four lists: Qx, consisting of the points in Q sorted by increasing xcoordinate; Qy, consisting of the points in Q sorted by increasing y-coordinate;
and analogous lists Rx and Ry. For each entry of each of these lists, as before,
we record the position of the point in both lists it belongs to.
We now recursively determine a closest pair of points in Q (with access
to the lists Qx and Qy). Suppose that q~ and q~ are (correctly) returned as a
closest pair of points in Q. Similarly, we determine a closest pair of points in
R, obtaining r~ and r~.
Combining the Solutions The general machinery of divide and conquer has
gotten us this far, without our really having delved into the structure of the
closest-pair problem. But it still leaves us with the problem that we saw
looming originally: How do we use the solutions to the two subproblems as
part of a linear-time "combining" operation?
d(qo,ql
* * The real question is: Are
Let 8 be the minimum of
* *) and d(r~,rl).
there points q E Q and r E R for which d(q, r) < 87 If not, then we have already
found the closest pair in one of our recursive calls. But if there are, then the
closest such q and r form the closest pair in P.
Let x* denote the x-coordinate of the rightmost point in Q, and let L denote
the vertical line described by the equation x = x*. This line L "separates" Q
from R. Here is a simple fact.
(5.8) If there exists q ~ Q and r ~ R for which d(q, r) < & then each of q and
r lies within a distance ~ of L.
Proof. Suppose such q and r exist; we write q = (qx, qy) and r = (rx, ry). By
the definition of x*, we know that qx < x* <_ rx. Then we have

x* - qx <- rx - qx <- d(q, r) < ~


and
rx - x* < rx - qx <- d(q, r) < 8,
so each of q and r has an x-coordinate within ~ of x* and hence lies within
distance a of the line L. []
So if we want to find a close q and r, we can restrict our search to the
narrow band consisting only of points in P within 8 of L. Let S __c p denote this
set, and let Sy denote the list consisting of the points in S sorted by increasing
y-coordinate. By a single pass through the list Py, we can construct Sy in O(n)
time.
We can restate (5.8) as follows, in terms of the set S.

5.4 Finding the Closest Pair of Points

(5.9) There exist q ~ O and r ~ R for which d(q, r) < a if and only if there
exist s, s ~ S for which d(s, s) < &
Its worth noticing at this point that S might in fact be the whole set P, in
which case (5.8) and (5.9) really seem to buy us nothing. But this is actuary
far from true, as the following amazing fact shows.

(5.10) If s, s ~ S have the property that d(s, s) < a, then S and s are within
15 positions of each other in the sorted list Sy.
Proof. Consider the subset Z of the plane consisting of all points within
distance ~ of L. We partition Z into boxes: squares with horizontal and vertical
sides of length 8/2. One row of Z will consist of four boxes whose horizontal
sides have t_he same y-coordinates. This collection of boxes is depicted in
Figure 5.7.
Suppose two points of S lie in the same box. Since all points in this box lie
on the same side of L, these two points either both belong to O or both belong
to R. But any two points in the same box are within distance ~. ~/2 < 8,
which contradicts our definition of ~ as the minimum distance between any
pair of points in Q or in R. Thus each box contains at most one point of S.
Now suppose that s, s ~ S have the property that d(s, s) < 8, and that they
are at least 16 positions apart in Sy. Assume without loss of generality that s
has the smaller y-coordinate. Then, since there can be at most one point per
box, there are at least three rows of Z lying between s and s. But any two
points in Z separated by at least three rows must be a distance of at least 38/2
apart--a contradiction. []
We note that the value of 15 can be reduced; but for our purposes at the
moment, the important thing is that it is an absolute constant.
In view of (5.10), we can conclude the algorithm as follows. We make one
pass through Sy, and for each s ~ Sy, we compute its distance to each of the
next 15 points in Sy. Statement (5.10) implies that in doing so, we will have
computed the distance of each pair of points in S (if any) that are at distance
less than 8 from each other. So having done this, we can compare the smallest
such distance to 8, and we can report one of two things~ (i) the closest pair
of points in S, if their distance is less than 8; or (if) the (correct) conclusion
that no pairs of points in S are within ~ of each other. In case (i), this pair is
the closest pair in P; in case (if), the closest pair found by our recursive calls
is the closest pair in P.
Note the resemblance between this procedure and the algorithm we rejected at the very beginning, which tried to make one pass through P in order

229

I~

ach box can


ontain at most |
ne input point.)
Line L

L812

Boxes ~

Figure 5.7 The portion of the


plane dose to the dividing
line L, as analyzed in the
proof of (5.10).

Chapter 5 Divide and Conquer

of y-coordinate. The reason such an approach works now is due to the extra knowledge (the value of 8) weve gained from the recursive calls, and the
special structure of the set S.
This concludes the description of the "combining" part of the algorithm,
since by (5.9) we have now determined whether the minimum distance
between a point in Q and a point in R is less than 8, and if so, we have
found the closest such pair.
A complete description of the algorithm and its proof of correctness are
implicitly contained in the discussion so far, but for the sake of concreteness,
we now summarize both.
Summary of the Algorithm A high-level description of the algorithm is the
following, using the notation we have developed above.
Closest-Pair (P)
Construct Px and Py (O(n log n) time)
(p~, p~) = Closest-Pair-Kec(Px,Py)
Closest-Pair-Rec(Px, Py)
If [PI ~ 3 then
find closest pair by measuring all pairwise distances
Endif
Construct Qx, Q),, Rx, Ry (O(n) time)
(q$,q~) = Closest-Pair-Rec(Ox,
(r~,r~) = Closest-Pair-Rec(Rx, Ry)

x* = maximum x-coordinate of a point in set Q


L = {~,y) : x = x*}
S = points in P within distance ~ of L.

5.5 Integer Multiplication


Else
Return (r~, r~)
Endif

~ Analyzing the Algorithm


we first prove that the algorithm produces a correct answer, using the facts
weve established in the process of designing it.
The algorithm correctly outputs a closest pair of points in P.
Proof. As weve noted, all the components of the proof have already been
worked out, so here we just summarize how they fit together.
We prove the correctness by induction on the size of P, the case of [P[ _< 3
being clear. For a given P, the closest pair in the recursive calls is computed
correctly by induction. By (5.!0) and (5.9), the remainder of the algorithm
correctly determines whether any pair of points in S is at distance less than
8, and if so returns the closest such pair. Now the closest pair in P either has
both elements in one of Q or R, or it has one element in each. In the former
case, the closest pair is correctly found by the recursive call; in the latter case,
this pair is at distance less than 8, and it is correctly found by the remainder
of the algorithm..,
We now bound the running time as well, using (5.2).
(~.12) The running time of the algorithm is O(n log n).
Proof. The initial sorting of P by x- and y-coordinate takes time O(n log n).
The running time of the remainder of the algorithm satisfies the recLt~,ence
(5.1), and hence is O(n log n) by (5.2). []

5.5 Integer Multiplication


Construct Sy (O(n) time)
For each point s ~ Sy, compute distance from s
to each of next 15 points in Sy
Let s, s be pair achieving minimum of these distances
(O(n) time)

If d(s,#) < 8 then


Retur~ (s, s)
Else if d(q~,q~) < d(r~,r~) then
Return (q~,q~)

We now discuss a different application of divide and conquer, in which the


"default" quadratic algorithm is improved by means of a different recurrence.
The analysis of the faster algorithm will exploit one of the recurrences
sidered in Section 5.2, in which more than two recursive calls are spawned at
each level.
f! The Problem
The problem we consider is an extremely basic one: the multiplication of two
integers. In a sense, this problem is so basic that one may not initially think of it

231

232

5.5 Integer Multiplication

Chapter 5 Divide and Conquer

(5.1). The combining of the solution requires a constant number of additions


of O(n)-bit numbers, so it takes time O(n); thus, the running time T(n) is
bounded by the recurrence

11oo
x 11Ol
IlOO
0000
IlOO
II00
~0011100

T(n) < 4T(n/2) + cn

(a)
Figure 5.8 The elementary-school algorithm for multipl~4ng two integers, in (a) decimal
and (b) binary representation.

even as an algorithmic question. But, in fact, elementary schoolers are taught a


concrete (and quite efficient) algorithm to multiply two n-digit numbers x and
y. You first compute a "partial product" by multiplying each digit ofy separately
by x, and then you add up all the partial products. (Figure 5.8 should help you
recall this algorithm. In elementary school we always see this done in base10, but it works exactly the same way in base-2 as well.) Counting a. sing!9
operation on a pair of bits as one primitive step in this computation, it takes
O(n) time to compute each partial product, and O(n) time to combine it in
with the running sum of all partial products so far. Since there are n partial
products, this is a total running time of O(n2).
If you havent thought about this much since elementary school, theres
something initially striking about the prospect of improving on this algorithm.
Arent all those partial products "necessary" in some way? But, in fact, it
is possible to improve on O(n2) time using a different, recursive way of
performing the multiplication.

~ Designing the Algorithm


The improved algorithm is based on a more clever way to break up the product
into partial sums. Lets assume were in base-2 (it doesnt really matter), and
start by writing x as Xl- 2n/2 + Xo. In other words, xl corresponds to the "highorder" n/2 bits, and x0 corresponds to the "low-order" n/2 bits. Similarly, we
write y = Yl" 2n/2 + Yo- Thus, we have
xy = (X1 2n/2 X0)(Yl" 2n/2 Y0)

(5.1)
Equation (5.1) reduces the problem of solving a single n-bit instance
(multiplying the two R-bit numbers x and y) to the problem of solving four n/2bit instances (computing the products xlYl, xly0, xoYl, and xoY0)- So we have
a first candidate for a divide-and-conquer solution: recursively compute the
results for these four n/2-bit instances, and then combine them using Equation

for a constant c. Is this good enough to give us a subquadratic running time?


We can work out the answer by observing that this is just the case q = 4 of
the class of recurrences in (5.3). As we saw earlier in the chapter, the solution
to this is T(n) < o(nlg2q) = O(n2).
So, in fact, our divide-and-conquer algorithm with four-way branching
was just a complicated way to get back to quadratic time! If we want to do
better using a strategy that reduces the problem to instances on n/2 bits, we
should try to get away with only three recursive calls. This will lead to the case
q = 3 of (5.3), which we saw had the solution T(n) <_ O(nlg2 q) = O(n1-59).
Recall that our goal is to compute the expression xlYl - 2n (xlYo xoYl)
n/2
2
X0y0 in Equation (5.1). It turns out there is a simple trick that lets us
determine al! of the terms in this expression using just three recursive calls. The
txick is to consider the result of the single multiplication (xl + Xo) (Yl + Yo) =
xffl + Xlyo + xoyl + x0yo. This has the four products above added together, at
the cost of a single recursive multiplication. If we now also determine xly~ and
XoYo by recursion, then we get the outermost terms explicitly, and we get the
middle term by subtracting xly1 and xoyo away from (xl + Xo)(y1 + Y0).
Thus, in fl~, our algorithm is
Recursive-Mult iply (x, y) :
Write x=x1-2nl2+x0
Y = Yl "2n/2 + YO
Compute Xl+X0 and YI+YO
P = Recursive-Multiply(Xl +xo, Yl +Yo)
XlYl = Recursive-Multiply(Xl, Yl)
XoYo = Recursive-Multiply(xo, Yo)
Return XlYI 2n + (p -- XlYI -- x0Y0). 2n/2 + x0Y0

~ Analyzing the Algorithm


We can determine the running time of this algorithm as follows. Given two nbit numbers, it performs a constant number of additions on O(n)-bit numbers,
in addition to the three recursive calls. Ignoring for now the issue that x1 Xo
and yl + Yo may have n/2 + I bits (rather than just n/2), which turns out not
to affect the asymptotic results, each of these recursive calls is on an instance
of size n/2. Thus, in place of our four-way branching recursion, we now have

233

234

5.6 Convolutions and the Fast Fourier Transform

Chapter 5 Divide and Conquer

a three-way branching one, with a running time that satisfies


T(n) <_ 3T(n/2) + cn
for a constant c.
This is the case q = 3 of (5.3) that we were aiming for. Using the solution
to that recurrence from earlier in the chapter, we have
i5:i31 The ~n~ing ti~e of geC~siV~ZMUi~iplT bn ~bi~ factors is

aobo
a~bo
a2bo

aobl
alb~
azbx

...
...
...

aobn_2
a~bn_2
azbn_2

aobn_~
albn_l
azbn_l

an-lbO an_lbl ... an_lbn_2 an_lbn_l

and then to compute the coordinates in the convolution vector by summing


along the diagonals.
Its worth mentioning that, unlike the vector sum and inner product,
the convolution can be easily generalized to vectors of different lengths,
bn_l). In this more general case,
we define a b to be a vector with m + n - ! coordinates, where coordinate
k is equal to

5.6 Convolutions and the Fast Fourier Transform


As a final topic in this chapter, we show how our basic recurrence from (5.1)
is used in the design of the Fast Fourier Transform, an algorithm with a wide
range of applications.
~ The Problem
Given two vectors a -- (ao, al ..... an-l) and b -- (bo, bl ..... bn-1), there are
a number of common ways of combining them. For example, one can compute
the sum, producing the vector a b = (ao bo,
or one can compute the inner product, producing the real number a. b-aobo + albl +... + an_lbn_~. (For reasons that will emerge shortly, it is useful
to write vectors in this section with coordinates that are indexed starting from
0 rather than 1.)
A means of combining vectors that is very important in applications, even
if it doesnt always show up in introductory linear algebra courses, is the
convolution a b. The convolution of two vectors of length n (as a and b are)
is a vector with 2n - 1 coordinates, where coordinate k is equal to
(i,]):i+]=k
i,]<n

In other words,

(i,]):i+j=k
i<m,]<n

We can picture this using the table of products aibj as before; the table is now
rectangular, but we still compute coordinates by summing along the diagonals.
(From here on, well drop explicit mention of the condition i < m,j < n in the
summations for convolutions, since it will be clear from the context that we
only compute the sum over terms that are defined.)
Its not just the definition of a convolution that is a bit hard to absorb at
first; the motivation for the definition can also initially be a bit elusive. What
are the circumstances where youd want to compute the convolution of two
vectors? In fact, the convolution comes up in a surprisingly wide variety of
different contexts. To illustrate this, we mention the following examples here.
A first example (which also proves that the convolution is something that
we all saw implicitly in high school) is polynomial multiplication. Any
polynomial A(x) = ao + a~x + a2x2 + .. arn_lXm-1 can be represented
arn-~).
Now, given two polynomials A(x) = ao + alx + a2x2 +. . ara_~xra-1 and
B(X) = b0 -k blX b2X2 .. bn_lxn-1, consider the polynomial C(x) =
A(x)B(x) that is equal to their product. In this polynomial C(x), the
coefficient on the xk term is equal to
c~ = ~ a~b].

a b = (aobo, aobl + albo, aob2 + albl + azbo .....


an_2bn_l an_lbn-2, an-lbn-1).

This definition is a bit hard to absorb when you first see it. Another way to
think about the convolution is to picture an rt x n table whose (f,j) entry is
a~bj, like this,

(i,j):i+j=k

In other words, the coefficient vector c of C(x) is the convolution of the


coefficient vectors of A(x) and B(x).
Arguably the most important application of convolutions in practice is
for signal processing. This is a topic that could fill an entire course, so

235

237
well just give a simple example here to suggest one way in which the
convolution arises.
am-l) which represents
a sequence of measurements, such as a temperature or a stock price,
sampled at m consecutive points in time. Sequences like this are often
very noisy due to measurement error or random fluctuations, and so a
common operation is to "smooth" the measurements by averaging each
value ai with a weighted sum of its neighbors within k steps to the left
and right in the sequence, the weights decaying quickly as one moves
away from ai. For example, in Gaussian smoothing, one replaces ai with

for some "width" parameter k, and with Z chosen simply to normalize


the weights in the average to add up to 1. (There are some issues With
boundary conditions--what do we do when i - k < 0 or i + k > m?--but
we could dea! with these, for example, by discarding the first and last k
entries from the smoothed signal, or by scaling them differently to make
up for the missing terms.)
To see the connection with the convolution operation, we picture
this smoothing operation as follows. We first define a "mask"
Wk_D

consisting of the weights we want to use for averaging each point with
e-1, 1, e-1 .....

e-(k-l)2, e-k2) in the Gaussian case above.) We then iteratively position


this mask so it is centered at each possible point in the sequence a; and
for each positioning, we compute the weighted average. In other words,
k
~
we replace ai with a, = ~s=-k Wsai+s"
This last expression is essentially a convolution; we just have to
warp the notation a bit so that this becomes clear. Lets define b =
b2k) by setting be = tvk_g. Then its not hard to check that
with this definition we have the smoothed value
ai

albg.
~,~):]+~=i+k

In other words, the smoothed sequence is just the convolution of the


original signal and the reverse of the mask (with some meaningless
coordinates at the beginning and end).

We mention one final application: the problem of combining histograms.


Suppose were studying a population of people, and we have the fol!owing two histograms: One shows the annual income of all the men in the
population, and one shows the annual income of all the women. Wed
now like to produce a new histogram, showing for each k the number of
pairs (M, W) for which man M and woman W have a combined income
of k.

This is precisely a convolution. We can write the first histogram as a


ara-1), to indicate that there are a~ men with annual
income equal to i. We can similarly write the second histogram as a
bn_~). Now, let c~ denote the number of pairs (m, w)
with combined income k; this is the number of ways of choosing a man
with income ai and a woman with income hi, for any pair (i, j) where
i + j = k. In other words,

c~= ~ aib~.
(i,]):i+j=k

so the combined histogram c = (co .....cra+n-2) is simply the convolution of a and b.


(Using terminology from probability that we will develop in Chapter 13, one can view this example as showing how convolution is the
underlying means for computing the distribution of the sum of two independent random variables.)

Computing the Convolution Having now motivated the notion of convolution, lets discuss the problem of computing it efficiently. For simplicity, we
will consider the case of equal length vectors (i.e., m = n), although everything
we say carries over directly to the case of vectors of unequal lengths.
Computing the convolution is a more subtle question than it may first
appear. The definition of convolution, after al!, gives us a perfectly valid way
to compute it: for each k, we just calculate the sum
E aibj
(i,j):i+j=k

and use this as the value of the kth coordinate. The trouble is that this direct way
of computing the convolution involves calculating the product a~bj for every
pair (i,j) (in the process of distributing over the sums in the different terms)
and this is (n2) arithmetic operations. Spending O(n2) time on computing
the convolution seems natural, as the definition involves O(n2) multiplications
azbj. However, its not inherently clear that we have to spend quadratic time to
compute a convolution, since the input and output both only have size O(n).

238

Chapter 5 Divide and Conquer

Could one design an algorithm that bypasses the quadratic-size definition of


convolution and computes it in some smarter way?
In fact, quite surprisingly, this is possible. We now describe a method
that computes the convolution of two vectors using only O(n log n) arithmetic
operations. The crux of this method is a powerful technique known as the Fast
Fourier Transform (FFT). The FFT has a wide range of further applications in
analyzing sequences of numerical values; computing convolutions quickly,
which we focus on here, is iust one of these applications.

~ Designing and Analyzing the Algorithm


To break through the quadratic time barrier for convolutions, we are going
to exploit the connection between the convolution and the multiplication of
two polynomials, as illustrated in the first example discussed previously. But
rather than use convolution as a primitive in polynomial multiplication, we
are going to exploit this connection in the opposite direction.
Suppose we are given the vectors a = (ao, al ..... an_~) and b= (bo,,
bn_l). We will view them as the polynomials A(x) = ao + alx + a2x2 +
. an_~xn-~ and B(x) = bo + b~x + b2x2 " "" bn-1xn-1, and well seek to compute their product C(x) = A(x)B(x) in O(n log rt) time. If c = (Co, c~ .... , c2n_2)
is the vector of coefficients of C, then we recall from our earlier discussion
that c is exactly the convolution a b, and so we can then read off the desired
answer directly from the coefficients of C(x).
Now, rather than multiplying A and B symbolically, we can treat them as
functions of the variable x and multiply them as follows.

(i)

First we choose 2n values xl, x2 ..... x2n and evaluate A(xj) and B(xj) for
each of j = !, 2 ..... 2n.
(ii) We can now compute C(xi) for each ] very easily: C(xj) is simply the
product of the two numbers A(xj) and B(xj).
(iii) Finally, we have to recover C from its values on x~, x2 ..... x2n. Here we
take advantage of a fundamental fact about polynomials: any polynomial
of degree d can be reconstructed from its values on any set of d + 1 or
more points. This is known as polynomial interpolation, and well discuss
the mechanics of performing interpolation in more detail later. For the
moment, we simply observe that since A and B each have degree at
most n - !, their product C has degree at most 2n - 2, and so it can be
reconstructed from the values C(xl), C(x2) ..... C(x2n) that we computed
in step (ii).

This approach to multiplying polynomials has some promising aspects


and some problematic ones. First, the good news: step (ii) requires only

5.6 Convolutions and the Fast Fourier Transform

O(n) arithmetic operations, since it simply involves the multiplication of O(n)


numbers. But the situation doesnt look as hopeful with steps (i) and (fii). In
particular, evaluating the polynomials A and B on a single value takes S2 (n)
operations, and our plan calls for performing 2n such evaluations. This seems
to bring us back to quadratic time right away.
The key idea that will make this all work is to find a set of 2n values
x~, x2 ..... x2n that are intimately related in some way, such that the work in
evaluating A and B on all of them can be shared across different evaluations. A
set for which this wil! turn out to work very well is the complex roots o[ unity.
The Complex Roots of Unity At this point, were going to need-to recal! a
few facts about complex numbers and their role as solutions to polynomial
equations.
Recall that complex numbers can be viewed as lying in the "complex
plane," with axes representing their real and imaginary parts. We can write
a complex number using polar coordinates with respect to this plane as rei,
where er~= -1 (and e2-n = 1). Now, for a positive integer k, the polynomial
equation xk = 1 has k distinct complex roots, and it is easy to identify them.
Each of the complex numbers wj,k = e2-qi/k (for] = 0, 1, 2 .....
k - 1) satisfies
the equation, since

and each of these numbers is distinct, so these are all the roots. We refer to
these numbers as the kth roots of unity. We can picture these roots as a set of k
equally spaced points lying on the unit circle in the complex plane, as shown
in Figure 5.9 for the case k = 8.
For our numbers x~ ..... x2n on which to evaluate A and B, we will choose
the (2n)th roots of unity. Its worth mentioning (although its not necessary for
understanding the algorithm) that the use of the complex roots of unity is the
basis for the name Fast Fourier Transform: the representation of a degree-d

Figure 5.9 The 8th roots of unity in the complex plane.

239

240

Chapter 5 Divide and Conquer

polynomial P by its values on the (d 1)st roots of unity is sometimes referred


to as the discrete Fourier transform of P; and the heart of our procedure is a
method for making this computation fast.
A Recursive Procedure for Polynomial Eualuatiou We want to design an
algorithm for evaluating A on each of the (2n)th roots of unity recursively, so
as to take advantage of the familiar recurrence from (5.1)--namely, T(n) <_
2T(n/2) O(n) where T(n) in this case denotes the number of operations
required to evaluate a polynomial of degree n - 1 on all the (2n)th roots of
unity. For simplicity in describing this algorithm, we will assume that rt is a
power of 2.
How does one break the evaluation of a polynomia! into two equal-sized
subproblems? A useful trick is to define two polynomials, Aeven(X) and Aoaa(x),
that consist of the even and odd coefficients of A, respectively. That is,
Aeven(X) = a0 + a2x + a4x2 +... + an_2x(n-2)/2,

and
Aoad(X) = a~ + a3x + asx2 + + a(n-Dx(n-:2)/2"

Simple algebra shows us that


A(x) = Aeven(X2) + XAodd(X2),
and so this gives us a way to compute A(x) in a constant number of operations,
given the evaluation of the two constituent polynomials that each have half
the degree of A.
Now Suppose that we evaluate each of Aeuen and Aoaa on the rtth roots of
unity. This is exactly a version of the problem we face with A and the (2n)th
roots of unity, except that the input is half as large: the degree is (n - 2)/2
rather than n - 1, and we have rt roots of unity rather than 2n. Thus we can
perform these evaluations in time T(n/2) for each of Ae~en and Aoaa, for a total
time of 2T(n/2).
Were now very close to having a recursive algorithm that obeys (5.1) and
gives us the running time we want; we just have to produce the evaluations
of A on the (2n)th roots of unity using O(n) additional operations. But this is
easy, given the results from the recursive calls on Aeuen and Aoaa. Consider
one of these roots of unity roj,2n = e2rji/2n. The quantity o)},2n is equal to
(e2-rji/2n)2 =- e2rji/n, and hence o,2n is an nth root of unity. So when we go
to compute
A(o)j,2n) = Aeuen(W~.,2n) wj,2nAodd(O)~,2n) ,

we discover that both of the evaluations on the right-hand side have been
performed in the recursive step, and so we can determine A(mj,2n) using a

5.6 Convolutions and the Fast Fourier Transform


constant number of operations. Doing this for a!l 2rt roots of unity i~ therefore
O(n) additional operations after the two recursive calls, and so the bound
T(n) on the number of operations indeed satisfies T(n) < 2T(n/2) + O(n). We
run the same procedure to evaluate the polynomial B on the (2n)th roots of
unity as well, and this gives us the desired O(n log n) bound for step (i) of our
algorithm outline.
Polynomial Interpolation Weve now seen how to evaluate A and B on the
set of all (2n)th roots of unity using O(n log n) operations and, as noted above,
we can clearly compute the products C(wj,n) = A(ooj,2n)B(o)j,2n) in O(n) more
operations. Thus, to conclude the algorithm for multiplying A and B, we
need to execute step (iii) in our earlier outline using O(n log n) operations,
reconstructing C from its values on the (2n)th roots of unity.
In describing this part of the algorithm, its worth keeping track of the
following top-level point: it turns out that the reconstruction of C can be
achieved simply by defining an appropriate polynomial (the polynomial D
below) and evaluating it at the (2n)th roots of unity. This is exactly what
weve just seen how to do using O(n log n) operations, so we do it again here,
spending an additional O(n log n) operations and concluding the algorithms.
Consider a polynomial C(x)= y~.an-1
s=o Csxs that we want to reconstruct
from its values C(ws,2n) at the (2rt)th roots of unity. Define a new polynomial
D(x)= Z-~s=oV2n-1 dsxs, where ds = C(cos,an). We now consider the values of D(x)
at the (2n)th roots of unity.

by definition. Now recall that OOs,2n --= (e2-ri/2n)s. Using this fact and extending
the notation to COs,2n = (e2-ri/2n)s even when s >_ 2n, we get that
2n-1 2n-1
D(oAj,2.n_) = ~ Ct(~ e(2"rO(st+js)/2n)
t=0

s=0

2n-1 2n-1

= ~_~ Ct(~ c~+j,2n)"


t=0

s=0

241

242

Solved Exercises

Chapter 5 Divide and Conquer

To analyze the last line, we use the fact that for any (2n)th root of unity eo 7~ 1,
"2n-1
we have v~s=O ms-= O. This is simply because eo is by definition a root of
2n--1 t
x2n - 1 = 0; since x2n - t = (x - 1)(~t=0 X ) and eo 7~ 1, it follows that co is
(x-2~-I x~).
also a root of !-~t=0
Thus the only term of the last lines outer sum that is not equal to 0 is
for q such that wt+j,zn = I; and this happens if t + j is a multiple of 2n, that
K-2n--1 1 = 2n. So we get that
y~,2n--!
s
is, if t = 2n -j. For this value,
s=0 wt+j,2n
= Z-,s=0

D(a~j,2n) = 2nczn_~. Evaluating the polynomial D(x) at the (2n)th roots of unity
thus gives us the coeffients of the polynomial C(x) in reverse order (multiplied
by 2n each). We sum this up as follows.
x-2n-1 Csxs, and corresponding polynomial
(S.14) For any polynomial C(x) -- -Z-,s=0
1
~0
= X-,2n_C((s,
2n)Xs" we have that c = ~D( 2n-s,2n)"
D(X) Z-~s=O
1
s
We can do a]] the evaluations of the values D(eozn_s, Zn) in O(nlog.n)
operations using the divide-and-conquer approach developed for step (i).
And this wraps everything up: we reconstruct the polynomial C from its
values on the (2n)th roots of unity, and then the coefficients of C are the
coordinates in the convolution vector c = a b that we were originally seeking.
In summary, we have shown the following.
(S.lS) Using the Fast Fourier Transforrrt to determine the product polynomial
C(x), we can compute the convolution of the original vectors a and b in
O(n log rO time.

Solved Exercises
Solved Exercise 1
Suppose you are given an array A with n entries, with each entry holding a
A[n]
is unimodal: For some index p between 1 and n, the values in the array entries
increase up to position p in A and then decrease the remainder of the way
unt~ position n. (So if you were to draw a plot with the array position j on the
x-axis and the value of the entry A[j] on the y-axis, the p!otted points would
rise until x-value p, where theyd achieve their maximum, and then fal! from
there on.)
Youd like to find the "peak entry" p without having to read the entire
array--in fact, by reading as few entries of A as possible. Show how to find
the entry p by reading at most O(log n) entries of A.

Solution Lets start with a general discussion on how to achieve a nmning


time of O(log n) and then come back to the specific problem here. If one needs
to compute something using only O(log n) operations, a useful strategy that
we discussed in Chapter 2 is to perform a constant amount of work, throw
away half the input, and continue recursively on whats left. This was the
idea, for example, behind the O(log n) running time for binary search.
We can view this as a divide-and-conquer approach: for some constant
c > 0, we perform at most c operations and then continue recursively on an
input of size at mbst n/2. As in the chapter, we will assume that the recursion
"bottoms out" when n = 2, performing at most c operations to finish the
computation. If T(n) denotes the running time on an input of size n, then
we have the recurrence
(5.16)
T(n) < T(n/2) + c

when n > 2, and


T(2) < c.

It is not hard to solve this recurrence by unrolling it, as follows.


Analyzing the first few levels: M the first level of recursion, we have a
single problem of size n, which takes time at most c plus the time spent
in all subsequent recursive calls. The next level has one problem of size
at most n/2, which contributes another c, and the level after that has
one problem of size at most n/4, which contributes yet another c.
Identifying apattem: No matter how many levels we continue, each level
will have just one problem: level j has a single problem of size at most
n/2J, which contributes c to the running time, independent of].
Summing over all levels of recursion: Each level of the recursion is
contributing at most c operations, and it takes log2 n levels of recursion to
reduce n to 2. Thus the total running time is at most c times the number
of levels of recursion, which is at most c log2 n = O(log rt).
We can also do this by partial substitution. Suppose ~ve guess that T(n) <
k !ogb n, where we dont know k or b. Assuming that this holds for smaller
values of n in an inductive argument, we would have
T(n) < T(n/2) + c
_< k log,(n/2) + c
= k logb n 7 k logb 2 + c.

243

244

Solved Exercises

Chapter 5 Divide and Conquer

The first term on the fight is exactly what we want, so we just need to choose k
and b to negate the added c at the end. This we can do by setting b = 2
and k = c, so that k logb 2 = c log2 2 = c. Hence we end up with the solution
T(n) < c log2 n, which is exactly what we got by unrolling the recurrence.
Finally, we should mention that one can get an O(log n) running time, by
essentially the same reasoning, in the more general case when each level of
the recursion throws away any constant fraction of the input, transforming an
instance of size n to one of size at most an, for some constant a < 1. It now
takes at most log1/a n levels of recursion to reduce n down to a constant size,
and each level of recnrsion involves at most c operations.
Now lets get back to the problem at hand. If we wanted to set ourselves
up to use (5.15), we could probe the midpoint of the array and try to determine
whether the "peak entry" p lies before or after this midpoint.
So suppose we look at the value A[n/2]. From this value alone, we cant
tell whether p lies before or after n/2, since we need to know whether entry
n/2 is sitting on an "up-slope" or on a "down-slope." So we also look at the
values A[n/2 - 1] and A[n/2 1]. There are now three possibilities.

If A[rt/2 - !] < A[n/2] < A[n/2 + 1], then entry n/2 must come strictly
beforep, and so we can continue recursively on entries n/2 + 1through ft.
If A[n/2 - 1] > A[n/2] > A[n/2 + 1], then entry n/2 must come strictly
after p, and so we can continue recursively on entries 1 through n/2 - 1.
Finally, if A[n/2] is larger than both A[n/2 - 1] and A[n/2 + 1], we are
done: the peak entry is in fact equal to rt/2 in this case.
In all these cases, we perform at most three probes of the array A and
reduce the problem to one of at most half the size. Thus we can apply (5.16)
to conclude that the running time is O(log n).
Solved Exercise 2
Youre consulting for a small computation-intensive investment company, and
they have the following type of problem that they want to solve over and over.
A typical instance of the problem is the following. Theyre doing a simulation
in which they look at n consecutive days of a given stock, at some point in
the past. Lets number the days i = 1, 2 ..... n; for each day i, they have a
price p(i) per share for the stock on that day. (Well assume for simplicity that
the price was fixed during each day.) Suppose during this time period, they
wanted to buy 1,000 shares on some day and sell all these shares on some
(later) day. They want to know: When should they have bought and when
should they have sold in order to have made as much money as possible? (If

there was no way to make money during the n days, you should report this
instead.)
For example, suppose n = 3, p(1) = 9, p(2) = 1, p(3) = 5. Then you should
return "buy on 2, sell on 3" (buying on day 2 and selling on day 3 means they
would have made $4 per share, the maximum possible for that period).
Clearly, theres a simple algorithm that takes time O(n2): try all possible
pairs of buy/sell days and see which makes them the most money. Your
investment friends were hoping for something a little better. "
Show how to find the correct numbers i and ] in time O(n log n).
Solution Weve seen a number of instances in this chapter where a bruteforce search over pairs of elements can be reduced to O(n log n) by divide and
conquer. Since were faced with a similar issue here, lets think about how we
might apply a divide-and-conquer strategy.
A natural approach would be to consider the first n/2 days and the final
n/2 days separately, solving the problem recursively on each of these two
sets, and then figure out how to get an overall solution from this in O(n) time.
This would give us the usual recurrence r(n) < 2T (-~) + O(n), and hence
O(n log n) by (5.1).
Also, to make things easier, well make the usual assumption that n is a
power of 2. This is no loss of generality: if n is the next power of 2 greater
than n, we can set p(i) = p(n) for all i between n and n. In this way, we do
not change the answer, and we at most double the size of the input (which
will not affect the O0 notation).
Now, let S be the set of days 1 ..... n/2, and S be the set of days n/2 +
n. Our divide-and-conquer algorithm will be based on the fol!owing
observation: either there is an optimal solution in which the investors are
holding the stock at the end of day n/2, or there isnt. Now, if there isnt, then
the optimal solution is the better of the optimal solutions on the ,,sets S and S.
If there is an optimal solution in which they hold the stock at the end of day
n/2, then the value of this solution is p(j) - p(i) where i ~ S and j S. But
this value is maximized by simply choosing i S which minimizes p(i), and
choosing j ~ S which maximizes p(j).
Thus our algorithm is to take the best of the following three possible
solutions.
o The optimal solution on S.
o The optimal solution on S.
* The maximum of p(j) -p(i), over i ~ Sandy ~ S.
The first two alternatives are computed in time T(n/2), each by recursion,
and the third alternative is computed by finding the minimum in S and the

245

246

Exercises

Chapter 5 Divide and Conquer

maximum in S, which takes time O(n). Thus the running time T(n) satisfies
T(n) <_2T (-~) + O(n),
as desired.
We note that this is not the best running time achievable for this problem.
In fact, one can find the optimal pair of days in O(n) tAne using dynamic
programming, the topic of the next chapter; at the end of that chapter, we will
pose this question as Exercise 7.

Exercises
1. You are interested in analyzing some hard-to-obtain data from two separate databases. Each database contains n numerical values--so there are
2n values total--and you may assume that no two values are the same.
Youd like to determine the median of this set of 2n values, which we will
define here to be the nth smallest value.
However, the only way you can access these values is through queries
to the databases. Ina single query, you can specify a value k to one of the
two databases, and the chosen database will return the/(m smallest value
that it contains. Since queries are expensive, you would like to compute
the median using as few queries as possible.
Give an algorithm that finds the median value using at most O(log n)
queries.
Recall the problem of finding the number of inversions. As in the text,
an, which we assume are all
distinct, and we define an inversion to be a pair i < j such that ai > ai.
We motivated the problem of counting inversions as a good measure
of how different two orderings are. However, one might feel that this
measure is too sensitive. Lets call a pair a significant inversion ff i <j and
ai > 2aj. Give an O(n log n) algorithm to count the number of significant
inversions between two orderings.

corresponding to it, and well say that two bank cards are equivalent if
they correspond to the same account.
Its very difficult to read the account number off a bank card directly,
but the bank has a high-tech "equivalence tester" that takes two bank
cards and, after performing some computations, determines whether
they are equivalent.
Their question is the following: among the collection of n cards, is
there a set of more than n/2 of them that are all equivalent to one another?
Assume that the only feasible operations you can do with the cards are
to pick two of them and plug them in to the equivalence tester. Show how
to decide the answer to their question with only O(n log n) invocations of
the equivalence tester.
Youve been working with some physicists who need to study, as part of
their experimental design, the interactions among large numbers of very
small charged particles. Basically, their setup works as follows. They have
an inert lattice structure, and they use this for placing charged particles
at regular spacing along a straight line. Thus we can model their structure
n} on the real line; and at each of
these points j, they have a particle with charge qJ" (Each charge can be
either positive or negative.)
They want to study the total force on each particle, by measuring it
and then comparing it to a computationa! prediction. This computational
part is where they need your help. The total net force on particle j, by
Couiombs Law, is equal to

..(i-i)2 ..(]-i)2
Theyve written the following simple program to compute F~ for all j:
n
Initialize Fi to 0
n

If i < j then
C qi qi

Else if i > j then

Suppose youre consulting for a bank thats concerned about fraud detection, and they come to you with the following problem. They have a
collection of n bank cards that theyve confiscated, suspecting them of
being used in fraud. Each bank card is a small plastic object, containing a magnetic stripe with some encrypted data, and it corresponds to
a unique account in the bank. Each account can have many bank cards

C qi qJ
Add - q 0-----~
Endif
Endfer

Output F]
Endfor

247

248

Notes and Further Reading

Chapter 5 Divide and Conquer

Its not hard to analyze the running time of this program: each
invocation of the inner loop, over i, takes O(n) time, and this inner loop
is invoked O(n) times total, so the overall running time is O(n2).
The trouble is, for the large values of n theyre working with, the program takes several minutes to run. On the other hand, their experimental
setup is optimized so that they can throw down n particles, perform the
measurements, and be ready to handle n more particles withJ_n a few seconds. So theyd really like it ff there were a way to compute all the forces
Fi much more quickly, so as to keep up with the rate of the experiment.
Help them out by designing an algorithm that computes all the forces
F1 in O(n log n) time.
Hidden surface removal is a problem in computer graphics that scarcely
needs an introduction: when Woody is standing in front of Buzz, you
should be able to see Woody but not Buzz; when Buzz is standing in
front of Woody .... well, you get the idea.
The magic of hidden surface removal is that you-can often compute
things faster than your intuition suggests. Heres a clean geometric example to illustrate a basic speed-up that can be achieved. You are given n
nonvertical ]lnes in the plane, labeled L1 ..... Ln, with the i~ line specified
by the equation y = aix + hi. We will make the assumption that no three of
the lines all meet at a single point. We say line Lg is uppermost at a given
x-coordinate x0 if its y-coordinate at x0 is greater than the y-coordinates
of a~ the other lines at x0: a~xo + bi > aixo + b1 for all ] ~ i. We say line L~ is
visible if there is some x-coordinate at which it is uppermost--intuitively,
some portion of it can be seen if you look down from "y = 002
Give an algorithm that takes n lines as input and in O(n log n) time
returns all of the ones that are visible. Figure 5.10 gives an example.
Consider an n-node complete binary tree T, where n = 2d - 1 fo~ some d.
Each node v of T is labeled with a real number xv. You may assume that
the real numbers labeling the nodes are all distinct. A node v of T is a
local minimum ff the label xv is less than the label xw for all nodes w that
are joined to v by an edge.
You are given such a complete binary tree T, but the labeling is only
specified in the following implicit way: for each node v, you can determine
the value xu by probing the node v. Show how to find a local minim _u~m of
T using only O(log n) probes to the nodes of T.

Suppose now that youre given an n x n grid graph G. (An n x n grid graph
is just the adjacency graph of an n x n chessboard. To be completely
precise, it is a graph whose node set is the set of all ordered pairs of

Figure 5.10 An instance of hidden surface removal with five lines (labeled 1-5 in the
figure). All the lines except for 2 are visible.

natural numbers (i,j), where 1 < i < ~. and 1 _<] _< n; the nodes (i,j) and
(k, e) are joined by an edge ff and only ff [i -/~1 + [/- el = 1.)
We use some of the terminology of the previous question. Again,
each node u is labeled by a real number x~; you may assume that all these
labels are distinct. Show how to find a local minimum of G using only
O(n) probes to the nodes of G. (Note that G has n2 nodes.)

Notes and Further Reading


The militaristic coinage "divide and conquer" was introduced somewhat after
the technique itself. Knuth (1998) credits John yon Neumann with one early
explicit application of the approach, the development of the Mergesort Algorithm in 1945. Knuth (1997b) also provides further discussion of techniques
for solving recurrences.
The algorithm for computing the closest pair of points in the plane is due
to Michael Shamos, and is one of the earliest nontrivial algorithms in the field
of computational geometry; the survey paper by Staid (1999) discusses a wide
range of results on closest-point problems. A faster randomized algorithm for
this problem will be discussed in Chapter 13. (Regarding the nonobviousness
of the divide-and-conquer algorithm presented here, Staid also makes the interesting historical observation that researchers originally suspected quadratic
time might be the best one could do for finding the closest pair of points in
the plane.) More generally, the divide-and-conquer approach has proved very
useful in computational geometry, and the books by Preparata and Shamos

249

250

Chapter 5 Divide and Conquer

(1985) and de Berg eta!. (1997) give many further examples of this technique
in the design of geometric algorithms.
The algorithm for multiplying two n-bit integers in subquadrafic time is
due to Karatsuba and Ofrnan (1962). Further background on asymptotically fast
multiplication algorithms is given by Knuth (!997b). Of course, the number
of bits in the input must be sufficiently large for any of these subquadrafic
methods to improve over the standard algorithm.
Press et al. (1988) provide further coverage of the Fast Fourier Transform,
including background on its applications in signal processing and related areas.
Notes on the Exercises Exercise 7 is based on a result of Donna Llewellyn,
Craig Tovey, and Michael Trick.

Chapter
Dynamic Programmiag

We began our study of algorithmic techniques with greedy algorithms, which


in some sense form the most natural approach to algorithm design. Faced with
a new computational problem, weve seen that its not hard to propose multiple
possible greedy algorithms; the challenge is then to determine whether any of
these algorithms provides a correct solution to the problem in all cases.
The problems we saw in Chapter 4 were al! unified by the fact that, in the
end, there really was a greedy algorithm that worked. Unfortunately, this is far
from being true in general; for most of the problems that one encounters, the
real difficulty is not in determining which of several greedy strategies is the
right one, but in the fact that there is no natural greedy algorithm that works.
For such problems, it is important to have other approaches at hand. Divide
and conquer can sometimes serve as an alternative approach, but the versions

c,~apter are often not s~ong

enough to reduce exponent,i..~~te4~~earc~own to polynomial time.

Rather, as we noted in Chapt~~~~/s there tended to reduce a


running time that was unnecessarily large, b--fi-t already polynomial, down to a
faster nmning time.
We now turn to a more powerful and subtle design technique, dynamic
programming. It will be easier to say exactly what characterizes dynamic programming after weve seen it in action, but the basic idea is drawn from the
intuition behind divide and conquer and is essentially the opposite of the
greedy strategy: one implicitly explores the space of all possible solutions, by
carefully decomposing things into a series of subproblems, and then building up correct solutions to larger and larger subproblems. In a way, we can
thus view dynamic programming as operating dangerously close to the edge of

252

6.1 Weighted Interval Scheduling: A Recursive Procedure

Chapter 6 Dynamic Programming

brute-force search: although its systematically working through the exponentially large set of possible solutions to the problem, it does this without ever
examining them all explicitly. It is because of this careful balancing act that
dynamic programming can be a tricky technique to get used to; it typically
takes a reasonable amount of practice before one is fully comfortable with it.
With this in mind, we now turn to a first example of dynamic programming: the Weighted Interval Scheduling Problem that we defined back in
Section 1.2. We are going to develop a dynamic programming algorithm for
this problem in two stages: first as a recursive procedure that closely resembles
brute-force search; and then, by reinterpreting this procedure, as an iterative
algorithm that works by building up solutions to larger and larger subproblems.

6.1 Weighted Interval Scheduling:


A Recursive Procedure
We have seen that a particular greedy algorithm produces an optimal solution
to the Interval Scheduling Problem, where the goal is to accept as large a
set of nonoverlapping intervals as possible. The Weighted Interval Scheduling
Problem is a strictly more general version, in which each interval has a certain
value (or weight), and we want to accept a set of maximum value. ~

~ Designing a Recursive Algorithm


Since the original Interval Scheduling Problem is simply the special case in
which all values are equal to 1, we know already that most greedy algorithms
will not solve this problem optimally. But even the algorithm that worked
before (repeatedly choosing the interval that ends earliest) is no longer optimal
in this more general setting, as the simple example in Figure 6.1 shows.
Indeed, no natural greedy algorithm is known for this pr0blem, which is
what motivates our switch to dynamic programming. As discussed above, we
wil! begin our introduction to dynamic programming with a recursive type of
.algorithm for this problem, and then in the next section well move to a more
iterative method that is closer to the style we use in the rest of this chapter.

We use the notation from our discussion of Interval Scheduling in Secn, with each request i specifying a
start time s~ and a finish time f~. Each interval i now also has a value, or weight
v~. Two intervals are compatible if they do not overlap. The goal of our current
n} of mutually compatible intervals,
so as to maximize the sum of the values of the selected intervals, ~ss viLets suppose that the requests are sorted in order of nondecreasing finish
time: fl < f2 < "" ".<_ fn. Well say a request i comes before a request] if i <j.
This wil! be the natural left-to-right order in which well consider intervals.
To help in talking about this order, we define p(j), for an interval ], to be the
largest index i < ] such that intervals i and j are disjoint. In other words, i
is the leftmost interval that ends before j begins. We define p(]) = 0 if no
request i < j is disjoint from ]. An example of the definition of p(]) is shown
in Figure 6.2.

Now, given an instance of the Weighted Interval Scheduling Problem, lets


consider an optimal solution CO, ignoring for now that we have no idea what
it is. Heres something completely obvious that we can say about CO: either
interval n (the last one) belongs to CO, or it doesnt. Suppose we explore both
sides of this dichotomy a little further. If n ~ CO, then clearly no interval indexed
strictly between p(n) and n can belong to CO, because by the definition ofp(n),
n - 1 all overlap interval n.
Moreover, if n s CO, then CO must include an optimal solution to the problem
consisting of requests {1 .....
p(n)}--for if it didnt, we could replace COs
choice of requests from {1 .....
p(n)} with a better one, with no danger of
overlapping request n.

Index

1
2
3
4

Index

Value = 1
Value = 3

t
Value = 1

V1 = 2
t

p(1) = 0
172 = 4

p(2) = 0
V3 = 4
v4= 7

p(3) = 1

p(4) = 0

p(S) = 3

1 p(6) = 3

;
Figure 6.1 A simple instance of weighted interval scheduling.

Figure 6.2 An instance of weighted interval scheduling with the functions p(i) defined
for each interval j.

253

254

6.1 Weighted Interval Scheduling: A Recursive Procedure

Chapter 6 Dynamic Programming

On the other hand, if n ~ (9, then (9 is simply equal to the optimal solution
n - 1}. This is by completely
analogous reasoning: were assuming that (9 does not include request n; so if
n - 1}, we could
replace it with a better one.

n}

involves looking at the optimal solutions of smaller problems of the form


]}. Thus, for any value of] between ! and n, let (9i denote the optimal
j}, and let OPT(j) denote
the value of this solution. (We define OPT(0) = 0, based on the convention
that this is the optimum over an empty set of intervals.) The optimal solution
were seeking is precisely (gn, with value OPT(n). For the optimal solution (9i
j}, our reasoning above (generalizing from the case in which
j = n) says that either j ~ (9i in which case OPT(j) = 11i q- OPT(p(])), or j 0i,
in which case OPT(j) = OPT(j -- 1). Since these are precisely the two possible
choices (j (9i or j 0i), we can hn-ther say that
(6.1)

OPT(j) ---- max(v] + OPTQg(j)), OPT(j -- 1)).

And how do we decide whether n belongs to the optimal solution (9i? This
too is easy: it belongs to the optimal solution if and only if the first of the
options above is at least as good as the second; in other words,

The correctness of the algorithm follows directly by induction on j:


(6,3) Compute0pt(j) correctly computes OPT(j) for each] = 1, 2, ..,, n,

Proof. By definition OPT(0) ---- 0. Now, take some j > 0, and suppose by way
of induction that Compute-0pt(i) correctly computes OPT(i) for all i <j. By
the induction hypothesis, we know that Compute-0pt(p(j)) = OPT(p(])) and
Compute-0pt(j -.1) = OPT(j -- 1); and hence from (6.1) it follows that

max(u] + Compute-Opt(2(])), Compute-Opt(j -- I))


= Compute-Opt(j).

OPT(j) -----

Unfortunately, if we really implemented the algorithm Compute-Opt as


just written, it would take exponential time to run in the worst case. For
example, see Figure 6.3 for the tree of calls issued for the instance of Figure 6.2:
the tree widens very quickly due to the recursive branching. To take a more
extreme example, on a nicely layered instance like the one in Figure 6.4, where
n, we see that Compute-Opt(j) generates
separate recursive calls on problems of sizes j - 1 and j - 2. In other words,
the total number of calls made to Compute-Opt on this instance will grow

OPT(6)

j} if and

OPT(3)

only if
Uj + OPT(p(])) >_ OPT(j -- 1).

These facts form the first crucial component on which a ,dynamic programming solution is based: a recurrence equation that expresses th6 optimal
solution (or its value) in terms of the optimal solutions to smaller subproblems.
Despite the simple reasoning that led to this point, (6.1) is already a
significant development. It directly gives us a recursive algorithm to compute
OPT(n), assuming that we have already sorted the requests by finishing time
and computed the values of p(j) for each j.
Compute-Opt (])
If j----0 then
Keturn 0
Else
Return max(u]+Compute-Opt (p (j)), Compute-OptG - I))
Endif

OPT(l)

OPT(3)

OPT(l)
OPT(2)

Ig

The tree of subproblems~


rows very quickly. )

OPT(1)

Figure 6.3 The tree of subproblems called by Compute-Opt on the problem instance
of Figure 6.2.

255

256

6.1 Weighted Interval Scheduling: A Recursive Procedure

Chapter 6 Dynamic Programming

Define M~] = max(ui+M-Compute-OptOp(])), M-Compute-OptS-- I))


Return M[j]
Endif
I
I

Figure 6.4 An instance of weighted interval scheduling on which the shnple ComputeOpt recursion will take exponential time. The vahies of all intervals in t_his instance
are 1.

like the Fibonacci numbers, which increase exponentially. Thus we have not
achieved a polynomial-time solution.

Memoizing the Recursion


In fact, though, were not so far from having a polynomial-time algorithr~.
A fundamental observation, which forms the second crucial component of a
dynamic programming solution, is that our recursive algorithm COmpute Opt is really only solving n+ 1 different subproblems: Compute-0pt(0),
Compute-0pt(n). The fact that it runs in exponential
time as written is simply due to the spectacular redundancy in the number of
times it issues each of these calls.
How could we eliminate all this redundancy? We could store the value of
Compute-0pt in a globally accessible place the first time we compute it and
then simply use this precomputed value in place of all future recursive calls.
This technique of saving values that have already been computed is referred
to as memoization.
We implement the above strategy in the more "intelligent" procedure MCompute-0pt. This procedure will make use of an array M[0... hi; M[j] will
start with the value "empty," but will hold the value of Compute-0pt(j) as
soon as it is first determined. To determine OPT(n), we invoke M-ComputeOpt(n).
M-Compute-Opt (])
If ] = 0 then
Return 0
Else if M~] is not empty then
Return M~]
Else

/~::# Analyzing the Memoized Version


Clearly, this looks ver~ similar to our previous implementation of the algorithm; however, memoization has brought the running time way down.
(6.4) The running time o[M-Compute-Opt(n) is O(n) (assuming the input
intervals are sorted by their finish times).
Proof. The time spent in a single call to M-Compute-0pt is O(1), excluding the
time spent in recursive calls it generates. So the rurming time is bounded by a
constant times the number of calls ever issued to M-Compute-0pt. Since the
implementation itself gives no explicit upper bound on this number of calls,
we try to find a bound by looking for a good measure of "progress."
The most useful progress measure here is the number of entries in M that
are not "empty." Initially this number is 0; but each time the procedure invokes
the recurrence, issuing two recursive calls to M-Compute-0pt, it fills in a new
entry, and hence increases the number of filled-in entries by 1. Since M has
only n + I entries, it follows that there can be at most O(n) calls to M-ComputeOpt, and hence the running time of M-Compute-0pt(n) is O(n), as desired.
Computing a Solution in Addition to Its Value
So far we have simply computed the value of an optimal solution; presumably
we want a flail optimal set of intervals as well. It would be easy to extend
M-Compute-0pt so as to keep track of an optimal solution in addition to its
value: we could maintain an additional array S so that S[i] contains an optimal
i}. Naively enhancing the code to maintain
the solutions in the array S, however, would blow up the rulming time by an
additional factor of O(n): while a position in the M array can be updated in
O(1) time, writing down a set in the $ array takes O(n) time. We can avoid
this O(n) blow-up by not explicitiy maintaining $, but ra~er by recovering the
optimal solution from values saved in the array M after the optimum value
has been computed.

We know from (6.2) that j belongs to an optimal solution for the set
j} if and only if v] + OPT(p(j)) > OPT(j -- 1). Using this
observation, we get the following simple procedure, which "traces back"
through the array M to find the set of intervals in an optimal solution.

257

258

6.2 Principles of Dynamic Programming

Chapter 6 Dynamic Programming

Find-Solut ion (j)


If j = 0 then
Output nothing
Else
If ui+ M[P(J)]>-M[J- 1] then
Output j together with the result of Find-Solution(p~))
Else
Output the result of Find-Solution(j- I)
Endif
Endif

Since Find-Solution calls itself recursively only on strictly smaller values, it makes a total of O(rt) recursive calls; and since it spends constant time
per call, we have
(6.5)

Giuert the array M of the optimal ualues of the sub-problems, Find-

Solution returns art opdmal solution in O(n) dine.

6.2 Principles of Dynamic Programming:


Memoization or Iteration over Subproblems
We now use the algorithm for the Weighted Interval Scheduling Problem
developed in the previous section to summarize the basic principles of dynamic
programming, and also tc offer a different perspective that will be fundamental
to the rest of the chapter: iterating over subproblems, rather than computing
solutions recursively.
In the previous section, we developed a polynomial-time solution to the
Weighted Interval Scheduling Problem by first designing an exponential-time
recursive algorithm and then converting it (by memoization) to an efficient
recursive algorithm that consulted a global array M of optimal solutions to
subproblems. To really understand what is going on here, however, it helps
to formulate an essentially equivalent version of the algorithm. It is this new
formulation that most explicitly captures the essence of the dynamic programming technique, and it will serve as a general template for the algorithms we
develop in later sections.

f! Designing the Algorithm


The key to the efficient algorithm is really the array M. It encodes the notion
that we are using the value of optimal solutions to the subproblems on intervals
j} for each j, and it uses (6.1) to define the value of M[j] based on

values that come earlier in the array. Once we have the array M, the problem
is solved: M[n] contains the value of the optimal solution on the full instance,
and Find-Solut ion can be used to trace back through M efficiently and return
an optimal solution itself.
The point to realize, then, is that we can directly compute the entries in
M by an iterative algorithm, rather than using memoized recursion. We just
start with M[O] = 0 and keep incrementing j; each time we need to determine
a value M[j], the a.nswer is provided by (6.1). The algorithm looks as follows.
It erat ive-Comput e-Opt

M[O] = 0
For j=l, 2 ..... n
MId] = max(ui + M[pq) ], M[j - 1])
Enddor

~ Analyzing the Algorithm


By exact analogy with the proof of (6.3), we can prove by induction on j that
this algorithm writes OPT(j) in array entry M[j]; (6.1) provides the induction
step. Also, as before, we can pass the filled-in array M to Find-Solution to
get an optimal solution in addition to the value. Finally, the running time
of Iterative-Compute-0pt is clearly O(n), since it explicitly runs for n
iterations and spends constant time in each.
An example of the execution of Iterative-Compute-0pt is depicted in
Figure 6.5. In each iteration, the algorithm fills in one additional entry of the
array M, by comparing the value of uj + M[p(j)] to the value ofM[j - 1].
A Basic Outline of Dynamic Programming
This, then, provides a second efficient algorithm to solve the Weighted Interval Scheduling Problem. The two approaches clearly have a great deal of
conceptual overlap, since they both grow from the insight contained in the
recurrence (6.1). For the remainder of the chapter, we wil! develop dynamic
programming algorithms using the second type of approach--iterative building up of subproblems--because the algorithms are often simpler to express
this way. But in each case that we consider, there is an equivalent way to
formulate the algorithm as a memoized recursion.
Most crucially, the bulk of our discussion about the particular problem of
selecting intervals can be cast more genera~y as a rough template for designing
dynamic programming algorithms. To set about developing an algorithm based
on dynamic programming, one needs a collection of subproblems derived from
the original problem that satisfies a few basic properties.

259

0123456

Index
1
2
3
4
5
6

6.3 Segmented Least Squares: Multi-way Choices

Chapter 6 Dynamic Programming

260

L~I = 2

p(1) = 0
w2 = 4

0~2 4
I/73 = 4

p(3) = 1

1B4=7

p(4) = o

I4 !6 ~ ~ I I

tV5 = 2 I

tu6 =

->

6.3 Segmented Least Squares: Multi-way Choices


We now discuss a different type of problem, which illustrates a slightly
more complicated style of dynamic programming. In the previous section,
we developed a recurrence based on a fundamentally binary choice: either
the interval n belonged to an optimal solution or it didnt. In the problem
we consider here, the recurrence will involve what might be called "multiway choices": at each step, we have a polynomial number of possibilities to
consider for the s.tructure of the optimal solution. As well see, the dynamic
programming approach adapts to this more general situation very naturally.
As a separate issue, the problem developed in this section is also a nice
illustration of how a clean algorithmic definition can formalize a notion that
initially seems too fuzzy and nonintuitive to work with mathematically.

Co)
Figure 6.5 Part (b) shows the iterations of r~ceragve-omPu~Ce-0P~c on the sample
instance of Weighted Interval Scheduling depicted In part (a).

~ The Problem
Often when looking at scientific or statistical data, plotted on a twodimensional set of axes, one tries to pass a "line of best fit" through the
data, as in Figure 6.6.

(i) There are only a polynomial number of subproblems.


(ii) The solution to the original problem can be easily computed from the
solutions to the subproblems. (For example, the original problem may
actually be one of the subproblems.)
(iii) There is a natural ordering on subproblems from "smallest" to "largest;
together with an easy-to-compute recurrence (as in (6.1) and (6.2)) that
allows one to determine the solution to a subproblem from the solutions
to some number of smaller subproblems.

This is a foundational problem in statistics and numerical analysis, formulated as follows. Suppose our data consists of a set P of rt points in the plane,

Naturally, these are informal guidelines. In particular, the notion of "smaller"


in part (iii) will depend on the type of recurrence one has.
We w~ see that it is sometimes easier to start the process of designing
such an algorithm by formulating a set of subproblems that looks natural, and
then figuring out a recurrence that links them together; but often (as happened
in the case of weighted interval scheduling), it can be useful to first define a
recurrence by reasoning about the structure of an optimal solution, and then
determine which subproblems will be necessary to unwind the recurrence:
This chicken-and-egg relationship between subproblems and recurrences is a
subtle issue underlying dynamic programming. Its never clear that a collection
of subproblems will be useflfl until one finds a recurrence linking them
together; but it can be difficult to think about recurrences in the absence of
the :smaller" subproblems that they build on. In subsequent sections, we will
develop further practice in managing this design trade-off.

(xn, Yn); and suppose x < x < .. < x Given


a line L defined by the equation y = ax + b, we say1 that2the error n.
of L with
respect to P is the sum of its squared "distances" to the points in P:
n

Error(L, P) = Y~,(Yi - axi - b)2i=1

Figure 6,6 A "line of best fit."

261

262

6.3 Segmented Least Squares: Multi-way Choices

Chapter 6 Dynamic Programming

0 0 0

0
0

0
0

0
0

0
0 0 0O0 O0

Figure 6.7 A set of points that lie approxLmately on two lines.

A natural goal is then to find the line with minimum error; this turns out to
have a nice closed-form solution that can be easily derived using calculus.
Skipping the derivation here, we simply state the result: The line of minimui-n
error is y = ax + b, where

Y,) and = E, a
Now, heres a kind of issue that these formulas werent designed to cover.
Ofien we have data that looks something like the picture in Figure 6.7. In this
case, wed like to make a statement like: "The points lie roughly on a sequence
of two lines." How could we formalize this concept?
Essentially, any single line through the points in the figure would have a
terrible error; but if we use two lines, we could achieve quite a small error. So
we could try formulating a new problem as follows: Rather than seek a single
line of best fit, we are allowed to pass an arbitrary set of lines through the
points, and we seek a set of lines that minimizes the error. But this fails as a
good problem formulation, because it has a trivial solution: if were allowed
to fit the points with an arbitrarily large set of lines, we could fit the points
perfectly by having a different line pass through each pair of consecutive points
inP.
At the other extreme, we could try "hard-coding" the number two into the
problem; we could seek the best fit using at most two lines. But this too misses
a crucial feature of our intuition: We didnt start out with a preconceived idea
that the points lay approximately on two lines; we concluded that from looking
at the picture. For example, most people would say that the points in Figure 6.8
lie approximately on three lines.

O0o

0 0 0O0 O0

Figure 6.8 A set of points that lie approximately on three lines.

Thus, intuitively, we need a problem formulation that reqt~es us to fit


the points we]], using as few lines as possible. We now formulate a problem-the Segmented Least Squares Problem--that captures these issues quite cleanly.
The problem is a fundamental instance of an issue in data mining and statistics
known as change detection: Given a sequence of data points, we want to
identify a few points in the sequence at which a discrete change occurs (in
this case, a change from one linear approximation to another).

Formulating the Problem As in the discussion above, we are given a set of


(xn, Yn)], with x~ < x2 < .-. < xn. We will use
Pi to denote the point (xi, y~). We must first partition P into some number
of segments. Each segment is a subset of P that represents a contiguous set
PJ-~" Pi} for
some :indices i < ]. Then, for each segment S in our partition of P, we compute
the line minimizing the error with respect to the points in S, according to the
formulas above.
The penalty of a partition is defined to be a sum of the following terms.
(i) The number of segments into which we partition P, times a fixed, given
multiplier C > 0.
For each segment, the error value of the optim,al line through that
segment.
Our goal in the Segmented Least Squares Problem is to find a partition of
minimum penalty. This minimization captures the trade-offs we discussed
earlier. We are allowed to consider partitions into any number of segments; as
we increase the number of segments, we reduce the penalty terms in part (ii) of
the definition, but we increase the term in part (i). (The multiplier C is provided

263

264

6.3 Segmented Least Squares: Multi-way Choices

Chapter 6 Dynamic Programming

with the input, and by tuning C, we can penalize the use of additional lines
to a greater or lesser extent.)
There are exponentially many possible partitions of P, and initially it is not
clear that we should be able to find the optimal one efficiently. We now show
how to use dynamic programming to find a partition of minimum penalty in
time polynomial in n.

~ Designing the Algorithm


To begin with, we should recall the ingredients we need for a dynamic programming algorithm, as outlined at the end of Section 6.2.We want a polynomial
number of subproblems, the solutions of which should yield a Solution to the
original problem; and we should be able to build up solutions to these subprob1eros using a recurrence. As with the Weighted Interval Scheduling Problem,
it helps to think about some simple properties of the optimal sohition. Note,
however, that there is not really a direct analogy to weighted interval scheduling: there we were looking for a subset of n obiects, whereas here we are
seeking to partition n obiects.
For segmented least squares, the following observation is very usefi.d:
The last point Pn belongs to a single segment in the optimal partition, and
that segment begins at some earlier point Pi- This is the type of observation
that can suggest the right set of subproblems: if we knew the identity of the
Pn (see Figure 6.9), then we could remove those points
from consideration and recursively solve the problem on the remaining points
Pi-l"

I(

OPT(i- 1)

Suppose we let OPT(i) denote the optimum solution for the points
Pi, and we let ei,j denote the minimum error of any line with repj. (We will write OPT(0) = 0 as a boundary case.) Then
our observation above says the following.
(6.6) If the last segment of the optimal partition is Pi .....
of the optimal solution is OPT(n) = ei,n + C + OPT(i -- 1).

Pn, then the value

Using the same observation for the subproblem consisting of the points
p], we see that to get OPT(]) we should find the best way to produce a
p]--paying the error plus an additive C for this segment-together with an optimal solution OPT(i -- 1) for the remaining points. In other
words, we have iustified the following recurrence.
p~,
OPT(]) = min(ei ~ + C + OPT(i -- 1)),
p1 is used in an optimum solution for the subproblem
if and only if the minimum is obtained using index i.
The hard part in designing the algorithm is now behind us. From here, we
simply build up the solutions OPT(i) in order of increasing i.
Segmented-Least-Squares (n)
Array M[O... n]
Set M[0]---- 0
For all pairs i<j

% .o._0-o-o- -

End/or

P]

n
Use the recurrence (6.7) to compute P/!~]
End/or
Return M[n]

0
0
0
0
00000 O0

Pn, and then


Pi-1.

By analogy with the arguments for weighted interval scheduling, the


correctness of this algorithm can be proved directly by induction, with (6.7)
providing the induction step.
And as in our algorithm for weighted interval scheduling, we can trace
back through the array M to compute an optimum partition.

265

266

6.4 Subset Sums and Knapsacks: Adding a Variable ,

Chapter 6 Dynamic Programming

Find-Segment s
If j = 0 then
Output nothing
Else
Find an i that minimizes eij+C+M[i-1]
Output the segment {Pi .....p]} and the result of
Find-Segments (i - I)
Endif

~ Analyzing the Algorithm


Final!y, we consider the running time of Segmented-Least-Squares. First
we need to compute the values of all the least-squares errors ei,j. To perform
a simple accounting of the running time for this, we note that there are O(nz)
pairs (f, ]) for which this computation is needed; and for each pair (f, ]); we
can use the formula given at the beginning of this section to compute ei,j in
O(n) time. Thus the total running time to compute all e~,j values is O(n3).
Following this, the algorithm has n iterations, for values ] --- I ..... n. For
each value of], we have to determine the minimum in the recurrence (6.7) to
fill in the array entry M[j]; this takes time O(n) for each], for a total Of O(nZ).
Thus the running time is O(n~) once all the el,~ values have been determinedJ

6.4 Subset Sums and Knapsacks: Adding a Variable


Were seeing more and more that issues in scheduling provide a rich source of
practically motivated algorithmic problems. So far weve considered problems
in which requests are specified by a given interval of time on a resource, as
well as problems in which requests have a duration and a deadline ,but do not
mandate a particular interval during which they need to be done.
In this section, we consider a version of t_he second type of problem,
with durations and deadlines, which is difficult to solve directly using the
techniques weve seen so far. We will use dynamic programming to solve the
problem, but with a twist: the "obvious" set of subproblems wi~ turn out not
to be enough, and so we end up creating a richer collection of subproblems. As

we wil! see, this is done by adding a new variable to the recurrence underlying
the dynamic program.
~ The Problem
In the scheduling problem we consider here, we have a single machine that
can process iobs, and we have a set of requests {1, 2 ..... n}. We are only
able to use this resource for the period between time 0 and time W, for some
number W. Each iequest corresponds to a iob that requires time w~ to process.
If our goal is to process jobs so as to keep the machine as busy as possible up
to the "cut-off" W, which iobs should we choose?
More formally, we are given n items {1 ..... n}, and each has a given
nonnegative weight wi (for i = 1 ..... n). We are also given a bound W. We
would like to select a subset S of the items so that ~i~s wi _< W and, subject
to this restriction, ~i~s voi is as large as possible. We will call this the Subset
Sum Problem.
This problem is a natural special case of a more general problem called the
Knapsack Problem, where each request i has both a value vg and a weight w~.
The goal in this more general problem is to select a subset of maximum total
value, subiect to the restriction that its total weight not exceed W. Knapsack
problems often show up as subproblems in other, more complex problems. The
name knapsack refers to the problem of filling a knapsack of capacity W as
fl~ as possible (or packing in as much value as possible), using a subset of the
items {1 ..... n}. We will use weight or time when referring to the quantities
tv~ and W.

Since this resembles other scheduling problems weve seen before, its
natural to ask whether a greedy algorithm can find the optimal solution. It
appears that the answer is no--at least, no efficient greedy role is known that
always constructs an optimal solution. One natura! greedy approach to try
would be to sort the items by decreasing weight--or at least to do this for al!
items of weight at most W--and then start selecting items in this order as !ong
as the total weight remains below W. But if W is a multiple of 2, and we have
three items with weights {W/2 + 1, W/2, W/2}, then we see that this greedy
algorithm will not produce the optimal solution. Alternately, we could sort by
increasing weight and then do the same thing; but this fails on inputs like
{1, W/2, W/21.

I In this analysis, the running time is dominated by the O(n3) needed to compute all ei,] values. But,
in fact, it is possible to compute all these values in O(n2) time, which brings the running time of the
ful! algorithm down to O(n2). The idea, whose details we will leave as an exercise for the reader, is to
first compute eid for all pairs (1,13 where ~ - i = 1, then for all pairs where j - i = 2, then j - i = 3, and
so forth. This way, when we get to a particular eij value, we can use the ingredients of the calculation
for ei.i-~ to determine ei.i in constant time.

The goa! of this section is to show how to use dynamic programming to


solve this problem. Recall the main principles of dynamic programming: We
have to come up with a small number of subproblems so that each subproblem
can be solved easily from "smaller" subproblems, and the solution to the
original problem can be obtained easily once we know the solutions to all

267

268

6.4 Subset Sums and Knapsacks: Adding a Variable

Chapter 6 Dynamic Programming

the subproblems. The tricky issue here lies in figuring out a good set of
subproblems.

/J Designing the Algorithm


A False Start One general strategy, which worked for us in the case of
Weighted Interval Scheduling, is to consider subproblems involving only the
first i requests. We start by trying this strategy here. We use the notation
OPT(i), analogously to the notation used before, to denote the best possible
i}. The key to our method for
the Weighted Interval Scheduling Problem was to concentrate on an optimal
solution CO to our problem and consider two cases, depending on whether or
not the last request n is accepted or rejected by this optimum solution. Just
as in that case, we have the first part, which follows immediately from the
definition of OPT(i).

o If n CO, then OPT(n) = OPT(n -- t).


Next we have to consider the case in which n ~ CO. What wed like here
is a simple recursion, which tells us the best possible value we can get for
solutions that contain the last request n. For Weighted Interval Scheduling
this was easy, as we could simply delete each request that conflicted with
request n. In the current problem, this is not so simple. Accepting request n
does not immediately imply that we have to reject any other request. Instead,
n - 1} that we will accept,
we have less available weight left: a weight of wn is used on the accepted
request n, and we only have W - wn weight left for the set S of remaining
requests that we accept. See Figure 6.10.
A Better Solution This suggests that we need more subproblems: To find out
the value for OPT(n) we not only need the value of OPT(n -- 1), but we also need
to know the best solution we can get using a subset of the first n- 1 items
and total allowed weight W - wn. We are therefore going to use many more
i} of the items, and each possible

value for the remaining available weight w. Assume that W is an integer, and
all requests i = 1 ..... n have integer weights wi. We will have a subproblem
for each i = O, 1 ..... n and each integer 0 < w < W. We will use OPT(i, IU) tO
denote the value of the optimal solution using a subset of the items {1 .....
with maximum allowed weight w, that is,

i} that satisfy ~jss wj <_ w.


Using this new set of subproblems, we will be able to express the value
OPT(i, w) as a simple expression in terms of values from smaller problems.
Moreover, OPT(n, W) is the quantity were looking for in the end. As before,
let 0 denote an optimum solution for the original problem.
o If n CO, then OPT(n, W) --- OPT(n -- 1, W), since we can simply ignore
item n.
If n ~ CO, then OPT(n, W) = Wn OPT(n -- 1, W - wn), since we now seek
to use the remaining capacity of W - wn in an optimal way across items
n-1.
When the nth item is too big, that is, W < wn, then we must have OPT(n, W) =
OPT(n -- 1, W). Otherwise, we get the optimum solution allowing all n requests
by taking the better of these two options. Using the same line of argument for
the subproblem for items {1 ..... i}, and maximum allowed weight w, gives
us the following recurrence.
(6.8)

If w < wi then OPT(i, W) = OPT(i -- 1, w). Otherwise


OPT(i, W) = max(oPT(i" 1, Iv), wi + OPT(i -- 1, U~ -- lVi)).

As before, we want to design an algorithm that builds up a table of all


OPT(i, w) values while computing each of them at most once.
Subset-Sum(n, W)

Array M[0... n, 0... W]


W
For i=1,2 .....
,I

For w=0 .....

n
W

Use the recurrence (6.8) to compute M[i, w]


End/or
End/or
Figure 6.10 After item n is included in the solution, a weight of ran is used up and there
is W - tun available weight left.

Return M[/Z, W]

269

270

6.4 Subset Sums and Knapsacks: Adding a Variable

Chapter 6 Dynamic Programming

0
0
0
0
0
0
0
0
0

Knapsack size W - 6, items w1 = 2, w2 = 2, w3 = 3

I
,.....--

2 0
1 0
0 0 0 o 0 0 0 0 00 0 0 0 0 0 0

Figure 6.11 The two-dimensional table of OPT values. The leftmost column and bottom
row is always 0. The entry for OPT(i, w) is computed from the two other entries
OPT(i -- I, w) and OPT(i -- 1, w -- wi), as indicated by the arrows.

Using (6.8) one can immediately prove by induction that the returned
n and
available weight W.

~ Analyzing the Algorithm


Recall the tabular picture we considered in Figure 6.5, associated with
weighted interval scheduling, where we also showed the way in which the array M for that algorithm was iteratively filled in. For the algorithm weve
iust designed, we can use a similar representation, but we need a twodimensional table, reflecting the two-dimensional array of subproblems that,
is being built up. Figure 6.11 shows the building up of subproblems in this
case: the value M[i, w] is computed from the two other values M[i - 1, w] and
M[i - 1, u? -- wi].
As an example of this algorithm executing, consider an instance with
weight limit W = 6, and n = 3 items of sizes w1 = w2 = 2 and w3 = 3. We find
that the optimal value OPT(3, 6) = 5 (which we get by using the third item and
one of the first two items). Figure 6.12 illustrates the way the algorithm fills
in the two-dimensional table of OPT values row by row.
Next we w/l! worry about the running time of this algorithm. As before .in
the case of weighted interval scheduling, we are building up a table of solutions
M, and we compute each of the values M[i, w] in O(1) time using the previous
values. Thus the running time is proportional to the number of entries in the
table.

3
2
1
0 0 0 0 0 0 0 0
0 123456
Initial values

0 0 2 2 4 4 4
0 0 2 2 2 2 2
0 0 0 0 0 o 0
0123456
Filling in values for i = 2

3
2
0 0 2 2 2 2 2
0 0
0 0 o 0
0123456
Filling in values for i = 1

0
0
1 0
0 0

0 2 3
o 2 2
0 2 2
0 0 0
0123456

4
4
2
0

5
4
2
0

5
4
2
0

Filling in values for i = 3

Figure 6.12 The iterations of the algorithm on a sample instance of the Subset Sum
Problem.

(6.9) The Subset-Sum(n, W) Algorithm correctly computes the optimal


value of the problem, and runs in O(nW) time.
Note that this method is not as efficient as our dynamic program for
the Weighted Interval Scheduling Problem. Indeed, its running time is not
a polynomial function of n; rather, it is a polynomial function of n and W,
the largest integer involved in defining the problem. We cal! such algorithms
pseudo-polynomial. Pseudo-polynomial algorithms can be reasonably efficient
when the numbers {uai} involved in the input are reasonably small; however,
they become less practical as these numbers grow large.
To recover an optimal set S of items, we can trace back through the array
M by a procedure similar to those we developed in the previous sections.

(6.10) Given a table M of the optimal values of thesubproblems, the optimal


set S can be found in O(n) time.

Extension: The Knapsack Problem


The Knapsack Problem is a bit more complex than the scheduling problem we
discussed earlier. Consider a situation in which each item i has a normegative
weight wi as before, and also a distinct value vi. Our goal is now to find a

271

272

Chapter 6 Dynamic Programming

subset S of maximum value ~],i~s vi, subject to the restriction that the total
weight of the set should not exceed W: ~i~s wi <- W.
It is not hard to extend our dynamic programming algorithm to this more
general problem. We use the analogous set of subproblems, OPT(i, I/Y), to denote
the value of the optimal solution using a subset of the items {1 ..... i} and
maximum available weight w. We consider an optimal solution (9, and identify
two cases depending on whether or not n E (9.

6.5 RNA Secondary Structure: Dynamic Programming over Intervals

U~ ~A

G~ C

o If n ~ O, then OPT(n, W) = OPT(n -- !, W).


o If n E O, then OPT(n, W) = vn q- OPT(n -- 1, W - wn).
Using this line of argument for the subproblems implies the following analogue

of {6.8}.
(6.11) If tv < wi then OPT(i, W) = OPT(i -- 1, W). Otherwise

~C

OPT(i, W) = max(oPT(i -- 1, w), vi + OPT(i -- 1, w -- wi)).

Using this recurrence, we can write down a completely analogous dynamic


programming algorithm, and this implies the following fact.
(6.1:1) The Knapsack Problem can be solved in O(nW) time.

6.5 RNA Secondary Structure: Dynamic


Programming over Intervals
In the Knapsack Problem, we were able to formulate a dynamic programming
algorithm by adding a new variable. A different but very common way by
which one ends up adding a variable to a dynamic program is through
the following scenario. We start by thinking about the set of subproblems
on {1, 2 ..... j}, for all choices of j, and find ourselves unable to come up
with a natural recurrence. We then look at the larger set of subproblems on
{i, i + 1 ..... j} for all choices of i and j (where i <_ j), and find a natural
recurrence relation on these subproblems. In this way, we have added the
second variable i; the effect is to consider a subproblem for every contiguous
interval in {1, 2 ..... n}.
There are a few canonical problems that fit this profile; those of you who
have studied parsing algorithms for context-free grammars have probably seen
at least one dynamic programming algorithm in this style. Here we focus on
the problem of RNA secondary structure prediction, a fundamental issue in
computational biology.

A
Figure 6.13 An RNA secondary structure. Tinck lines connect adjacent elements of the
sequence; thin lines tndlcate parrs of elements that are matched.

~J The Problem
As one learns in introductory biology classes, Watson and Crick posited that
double-stranded DNA is "zipped" together by complementary base-pairing.
Each strand of DNA can be viewed as a string of bases, where each base is
drawn from the set {A, C, G, T}.2 The bases A and T pair with each other, and
the bases C and G pair with each other; it is these A-T and C-G pairings that
hold the two strands together.
Now, single-stranded RNA molecules are key components in many of
the processes that go on inside a cell, and they follow more or less the
same structural principles. However, unlike double-stranded DNA, theres no
"second strand" for the RNA to stick to; so it tends to loop back and form
base pairs with itself, resulting in interesting shapes like the one depicted in
Figure 6.13. The set of pairs (and resulting shape) forme~ by the RNA molecule
through this process is called the secondary structure, and understanding
the secondary structure is essential for understanding the behavior of the
molecule.

Adenine, cytosine, guanine, and thymine, the four basic units of DNA.

273

274

6.5 RNA Secondary Structure: Dynamic Programming over Intervals

Chapter 6 Dynamic Programming

For our purposes, a single-stranded RNA molecule can be viewed as a


sequence of n symbols (bases) drawn from the alphabet [A, C, G, U}.3 Let B =
bib2.. bn be a single-stranded RNA molecule, where each bi ~ [A, C, G, U).
To a first approximation, one can model its secondary structure as follows. As
usual, we require that A pairs with U, and C pairs with G; we also require
that each base can pair with at most one other base--in other words, the set
of base pairs forms a matching. It also turns out that secondary structures are
(again, to a first approximation) "knot-flee," which we will formalize as a kind
of noncrossing condition below.
Thus, concretely, we say that a secondary structure on B is a set of pairs
n}, that satisfies the following conditions.
(i) (No sharp tams.) The ends of each pair in S are separated by at least four
intervening bases; that is, if (i,j) ~ S, then i <j - 4.
(ii) The elements of any pair in S consist of either {A, U} or {C, G} (in either
order).
(iii) S is a matching: no base appears in more than one pair.
(iv) (The noncrossing condition.) If (i, j) and (k, g) are two pairs in S, then
we cannot have i < k <j < g. (See Figure 6.14 for an illustration.)
Note that the RNA secondary structure in Figure 6.13 satisfies properties (i)
through (iv). From a structural point of view, condition (i) arises simply
because the RNA molecule cannot bend too sharply; and conditions (ii) and
(iii) are the fundamental Watson-Crick rules of base-pairing. Condition (iv) is
the striking one, since its not obvious why it should hold in nature. But while
there are .sporadic exceptions to it in real molecules (via so-called pseudoknotting), it does turn out to be a good approximation to the spatial constraints
.
Now, out of all the secondary structures that are possible for- a single
RNA molecule, which are the ones that are likely to arise under physiological
conditions? The usual hypothesis is that a single-stranded RNA molecule wfl!
form the secondary structure with the optimum total flee energy. The correct
model for the free energy of a secondary structure is a subject of much debate;
but a first approximation here is to assume that the flee energy of a secondary
structure is proportional simply to the number of base pairs that it contains.
Thus, having said all this, we can state the basic RNA secondary structure
prediction problem very simply: We want an efficient algorithm that takes

3 Note that the symbol T from the alphabet of DNA has been replaced by a U, but this is not important
for us here.

G
U
A
G
U
A
A

CAUGAUGGCCAUGU

(b)
Figure 6.14 Two views of an RNA secondaI3, structure. In the second view, (b), the
string has been "stretched" lengthwise, and edges connecting matched pairs appear as
noncrossing "bubbles" over the string.

a single-stranded RNA molecule B = bib2 bn and determines a secondary


structure S with the maximum possible number of base pairs.

~ Designing and Analyzing the Algorithm


A First Attempt at Dynamic Programming The natural first attempt to
apply dynamic programming would presumably be based on the following
subproblems: We say that OPT(j) is the maximum number of base pairs in a
secondary structure on bib2 .. bj. By the no-sharp-turns condition above, we
know that OPT(j) = 0 for j < 5; and we know that OPT(n) is the solution were
looking for.
The trouble comes when we try writing down a recurrence that expresses
OPT(j) in terms of the solutions to smaller subproblems. We can get partway
there: in the optimal secondary structure on bib2 bj, its the case that either
o j is not involved in a pair; or
e j pa~s with t for some t < j - 4.
In the first case, we just need to consult our solution for OPT(j -- 1). The second
case is depicted in Figure 6.15(a); because of the noncrossing condition,
we now know that no pair can have one end between 1 and t- 1 and the
other end between t 4- 1 and j - 1. Weve therefore effectively isolated two
new subproblems: one on the bases blb2 . .. bt_l, and the other on the bases
bt+~ bj_~. The first is solved by OPT(t -- 1), but the second is not on our list
of subproblems, because it does not begin with b~.

275

276

6.5 RNA Secondary Structure: Dynamic Programming over Intervals

Chapter 6 Dynamic Programming

RNA sequence ACCGGUAGU

Itncluding the pair (t, ]) results in~


wo independent subproblems. )

]-1 j

12

(a)

4 0 0 0
3 0 0
2 0
i=l
j=6 7 8 9
Initial values

0 0 0 0
0 0 1
0
1

4 0 0 0 0
3 0 0 1 1
2 0 0 1
i=1
1 1
j= 6 7 8 9

Filling in the values


for k = 5

Filling in the values


for k = 6

j-1 1

t-1 t t+ 1

Co)
Figure 6.15 Schematic views of the dynamic programming recurrence using (a) one
variable, and (b) two variables.
This is the insight that makes us realize we need to add a variable: We
need to be able to work with subproblems that do not begin with bl; in other
words, we need to consider subproblems on bibi+~.. bj for a!l choices of i <j.
Dynamic Programming over Intervals Once we make this decision, our
previous reasoning leads straight to a successfu! recurrence. Let OPT(i, j) denote
the maximum number of base pairs in a secondary structure on bib~ ~ .. bi.
The no-sharp-turns condition lets us initialize OPT(i, j) = 0 whenever i >_ j - 4.
(For notational convenience, we will also allow ourselves to refer to OPT(i, j)
even when i > j; in this case, its value is 0.)
Now, in the optimal secondary structure on b~bi+~ bj, we have the same
alternatives as before:
o j is not involved in a pair; or
o j pairs with t for some t < j - 4.
In the first case, we have OPT(i, j) = OPT(i, j -- 1). In the second case, depicted
in Figure 6.15(b), we recur on the two subproblems OPT(i, t -- 1) and OPT(t +
1,j -- 1); as argued above, the noncrossing condition has isolated these two
subproblems from each other.
We have therefore justified the following recurrence.

~6:13) 0PT(il;) = max(OPT(i,;" max(! + OPT(i~t" + OPT(t +1,;

where the max is taken over t such that bt and bi are an allowable base pair
(under conditions (i) and (iO from the definition of a secondary structure) ~
Now we just have to make sure we understand the proper order in which
to build up the solutions to the subproblems. The form of (6.13) reveals that
were always invoking the solution to subproblems on shorter intervals: those

4 o 0 0 0
3 0 0 1 1
2 0 0 1 1
i=1
1 1 1
j=6 7 8 9
Filling in the values
fork = 7

4 0 0 0
3 0 0 1
2 0 0 1
i=1
1 1 1
j=6 7 8 9

0
1
1
2

Filling in the values


fork = 8

Figure 6.16 The iterations of the algorithm on a sample instance of the RNA Secondary
Structure Prediction Problem.

for which k =; - i is smaller. Thus things will work without any trouble if we
build up the solutions in order of increasing interval length.
Initialize OPT(i,j) = 0 whenever i _>j -- 4
n-I
n-k
Set j=i+k
Compute OPT(i,]) using the recurrence in (6.13)
End/or
End/or
l~etur~ OPT(I, n)

As an example of this algorithm executing, we consider the input


ACCGGUAGU, a subsequence of the sequence in Figure 6.14. As with the
Knapsack Problem, we need two dimensions to depict the array M: one for
the left endpoint of the interval being considered, and one for the right endpoint. In the figure, we only show entries corresponding to [i,;] pairs with
i < j - 4, since these are the only ones that can possibly be nonzero.
It is easy to bound the running time: there are O(n2) subproblems to solve,
and evaluating the recurrence in (6.13) takes time O(n) for each. Thus the
running time is O(na).

277

278

6.6 Sequence Alignment

Chapter 6 Dynamic Programming


As always, we can recover the secondary structure itself (not just its value)
by recording how the minima in (6.13) are achieved and tracing back through
the computation.

6.6 Sequence Alignment


For the remainder of this chapter, we consider two further dynamic programming algorithms that each have a wide range of applications. In the next two
sections we discuss seqaence alignment, a fundamental problem that arises
in comparing strings. Following this, we turn to the problem of computing
shortest paths in graphs when edges have costs that may be negative.

~ The Problem
Dictionaries on the Web seem to get more and more useful: often it seems easier
to pull up a bookmarked online dictionary than to get a physical dictionary
down from the bookshelf. And many online dictionaries offer functions-that
you cant get from a printed one: if youre looking for a definition and type inca
word it doesnt contain--say, ocurrance--it will come back and ask, "Perhaps
you mean occurrence?" How does it do this? Did it truly know what you had
in mind?
Lets defer the second question to a different book and think a little about
the first one. To decide what you probably meant, it would be natural to search
the dictionary for the word most "similar" to the one you typed in. To do this,
we have to answer the question: How should we define similarity between
two words or strings?
Intuitively, wed like to say that ocurrance and occurrence are similar
because we can make the two words identical if we add a c to the first word
and change the a to an e. Since neither of these changes seems so large, we
conclude that the words are quite similar. To put it another way, we can nearly
line up the two words letter by letter:

The hyphen (-) indicates a gap where we had to add a letter to the second
word to get it to line up with the first. Moreover, our lining up is not perfect
in that an e is lined up with an a.
We want a model in which similarity is determined roughly by the number
of gaps and mismatches we incur when we line up the two words. Of course,
there are many possible ways to line up the two words; for example, we could
have written

which involves three gaps and no mismatches. Which is better: one gap and
one mismatch, or three gaps and no mismatches?
This discussion has been made easier because we know roughly what
the correspondence ought to !ook like. When the two strings dont look like
English words--for example, abbbaabbbbaab and ababaaabbbbbab--it may
take a little work to decide whether they can be lined up nicely or not:
abbbaa--bbbbaab
ababaaabbbbba-b

Dictionary interfaces and spell-checkers are not the most computationally


intensive application for this type of problem. In fact, determining similarities
among strings is one of the central computational problems facing molecular
biologists today.
Strings arise very naturally in biology: an organisms genome--its ful! set
of genetic material--is divided up into giant linear DNA molecules known
as chromosomes, each of which serves conceptually as a one-dimensional
chemical storage device. Indeed, it does not obscure reality very much to
think of it as an enormous linear tape, containing a string over the alphabet
{A, C, G, T}. The string of symbols encodes the instructions for building
protein molecules; using a chemical mechanism for reading portions of the
chromosome, a cell can construct proteins that in turn control its metabolism.
Why is similarity important in this picture? To a first approximation, the
sequence of symbols in an organisms genome can be viewed as determining
the properties of the organism. So suppose we have two strains of bacteria,
X and Y, which are closely related evolutionarlly. Suppose further that weve
determined that a certain substring in the DNA of X codes for a certain kind
of toxin. Then, if we discover a very "similar" substring in the DNA of Y,
we might be able to hypothesize, before performing any experiments at all,
that this portion of the DNA in Y codes for a similar kind of toxin. This use
of computation to guide decisions about biological experiments is one of the
hallmarks of the field of computational biology.
Al! this leaves us with the same question we asked initially, while typing
badly spelled words into our online dictionary: How should we define the
notion of similarity between two strings?
In the early 1970s, the two molecular biologists Needleman and Wunsch
proposed a definition of similarity, which, basically unchanged, has become

279

280

6.6 Sequence Alignment

Chapter 6 Dynamic Programming

the standard definition in use today. Its position as a standard was reinforced by
its simplicity and intuitive appeal, as wel! as through its independent discovery
by several other researchers around the same time. Moreover, this definition of
similarity came with an efficient dynamic programming algorithm to compute
it. In this way, the paradigm of dynamic programming was independently
discovered by biologists some twenty years after mathematicians and computer
scientists first articulated it.
The definition is motivated by the considerations we discussed above, and
in particular by the notion of "lining up" two strings. Suppose we are given two
strings X and Y, where X consists of the sequence of symbols XlX2 xm and Y
consists of the sequence of symbols Y~Y2" " "Yn- Consider the sets {1, 2 ..... In}
and {1, 2 ..... n} as representing the different positions in the strings X and Y,
and consider a matching of these sets; recall that a matching is a set of ordered
pairs with the property that each item occurs in at most one pair. We say that a
matching M of these two sets is an alignment if there are no "crossing" pairs:
if (i, ]), (i, j) ~ M and i < i, then j < j. Intuitively, an alignment gives a Way
of lining up the two strings, by telling us which pairs of positions will be lined
up with one another. Thus, for example,
stop-tops

corresponds to the alignment {(2, 1), (3, 2), (4, 3)}.


Our definition of similarity will be based on finding the optiinal alignment
between X and Y, according to the following criteria. Suppose M is a given
alignment between X and Y.
o First, there is a parameter ~ > 0 that defines a gap penalty. For each
position of X or Y that is not matched in M--it is a gap--we incur a
cost of 3.
o Second, for each pair of letters p, q in our alphabet, there is a Inisinatch
cost of %q for lining up p with q. Thus, for each (i, j) ~ M, we pay the
appropriate mismatch cost o~xiyj for lining up xi with yj. One generally
assumes that %~ = 0 for each letter p--there is no mismatch cost to line
up a letter with another copy of itself--although this wil! not be necessary
in anything that follows.
o The cost of M is the sum of its gap and mismatch costs, and we seek an
alignment of minimum cost.
The process of minimizing this cost is often referred to as sequence aligninent
in the biology literature. The quantities ~ and {oOq) are external parameters
that must be plugged into software for sequence alignment; indeed, a lot of
work goes into choosing the se~ngs for these parameters. From our point of

view, in designing an algorithm for sequence alignment, we wil! take them as


given. To go back to our first example, notice how these parameters determine
which alignment of ocurrance and occurrence we should prefer: the first is
strictly better if and only if ~ + O~ae < 33.
~ Designing the Algorithm
We now have a concrete numerical definition for the similarity between
strings X and Y: it is the minimtim cost of an alignment between X and Y. The
lower this cost, the more similar we declare the strings to be. We now turn to
the problem of computing this minimum cost, and an optimal alignment that
yields it, for a given pair of strings X and Y.
One of the approaches we could try for this problem is dynamic programruing, and we are motivated by the fo!lowing basic dichotomy.

o In the optimal alignment M, either (in, n) ~ M or (in, n) ~ M. (That is,


either the last symbols in the two strings are matched to each other, or
they arent.)
By itself, this fact would be too weak to provide us with a dynamic programming solution. Suppose, however, that we compound it with the following
basic fact.
(6.14) Let M be any alignment of X and Y. If (in, n) C M, then either the
Inth position of X or the nth position of Y is not matched in M.

Proof. Suppose by way of contradiction that (in, n) 9~ M, and there are numbers i < In andj < n so that (in,j) M and (i, n) CM. But this contradicts our
definition ofaligninent: we have (i, n), (in,j) M with i < In, but n > i so the
pairs (i, n) and (in,i) cross. []
There is an equivalent way to write (6.14) that exposes three alternative
possibilities, and leads directly to the formulation of a recurrence.
(6.i5) In an optiinal alignment M, at least one of the following is true:
(i) (In, n) ~ M; or
(iO the Inth position of X is not matched; or
(iii) the n~h position of Y is not matched.
Now, let OPT(i, j) denote the minimum cost of an alignment between
xlx2...xi and YlY2""Yj. If case (i) of (6.15) holds, we pay aXmy, and then
align XlX2 xm_l as well as possible with YlY2 " Yn-1; we get OPT(m, n) =
~x,~,, + OPT(In -- 1, n -- 1). If case (ii) holds, we pay a gap cost of ~ since the
Inth position of X is not matched, and then we align xlx2 Xm-1 as we!l as

281

282

6.6 Sequence Alignment

Chapter 6 Dynamic Programming

possible with YlYa "" Yn- In this way, we get OPT(m, n) = ~ + OPT(m -- !, n).
Similarly, if case (iii) holds, we get OPT(m, n) = ~ + OPT(m, n -- 1).
Using the same argument for the subproblem of finding the minimum-cost
alignment between XlX2" xi and YlY2" " "Yi we get the following fact.

x3

x2

x1

(6.16) The minimum al~nment costs satisfy the ~olIotving recurrenCe for i > 1
and ] >_ 1:
~ + OPT(i,] " i)],
OPT(i, ]) = min [axiyj + OPT(i 1] ~ 1), ~ + opT(i" 1,]),
Moreover, (i. ]) is in an optimal alignment M for this subproblem if and only
.
We have maneuvered ourselves into a position where the dynamic programming algorithm has become clear: We build up the values of OPT(i,j) using
the recurrence in (6.16). There are only O(mn) subproblems, and OPT(m, n)
is the value we are seeldng.
We now specify the algorithm to compute the value of the optimal aligr~ment. For purposes of initialization, we note that OPT(i, 0) = OPT(0, i) = ig for
all i, since the only way to line up an i-letter word with a 0-letter word is to
use i gaps.
Alignment (X, Y)
Array A[0... m, 0... n]
Initialize A[i, 0]= i8 for each i
Initialize A[O,]]=]~ for each ]
For ]=1 .....n
For i=1 ..... ra
Use the recurrence (6.16) to compute All, j]
Endfor
Endfor
Return A[m, n]

As in previous dynamic programming algorithms, we can trace back


through the array A, using the second part of fact (6.16), to construct the
alignment itself.

Yl

Y2

Y3

Y4

Figo_re 6.17 A graph-based picture of sequence alignment

There is an appealing pictorial way in which people think about this


sequence alignment algorithm. Suppose we build a two-dimensional m x n
grid graph Gx~,, with the rows labeled by symbols in the string X, the colunms
labeled by symbols in Y, and directed edges as in Figure 6.17.
We number the rows from 0 to m and the columns from 0 to n; we denote
the node in the ith row and the jth column by the label (i, j). We put costs on
the edges of Gxy: the cost of each horizontal and vertical edge is 3, and the
cost of the diagonal edge from (i- 1,j - 1) to (i,j) is ~xiyj.
The purpose of this picture now emerges: the recurrence in (6.16) for
OPT(i,j) is precisely the recurrence one gets for the minimum-cost path in Gxy
from (0, 0) to (i,j). Thus we can show

(6.17) Let f(i,]) denote the minimum cost of a path from (0, O) to (i,]) in
Gxy~ Then for all i, ], roe have f (i, ]) = OPT(i,])~
Proof. We can easily prove this by induction on i +]. When i +] = 0, we have
i =] = 0, and indeed f(i,]) = OPT(i,j) = 0.
Now consider arbitrary values of i and j, and suppose the statement is
true for all pairs (i,j) with i +j < i +j. The last edge on the shortest path to
(i, ]) is either from (i - !, j - 1), (i - 1, j), or (i, j - 1). Thus we have
f(i,j) = min[axiyj + f(i - 1,j - 1), ~ + f(i - 1,]), 8 + ~f(i,] - 1)]
= min[O~x~vj + OPW(i -- 1,j -- 1), 3 + OPT(/-- 1,j), 3 + OPW(i,j -- 1)]

~ Analyzing the Algorithm


The correctness of the algorithm follows directly from (6.16). The running time
is O(mn), since the array A has O(mn) entries, and at worst we spend constant
time on each.

= OPT(i, j),

where we pass from the f~st line to the second using the induction hypothesis,
and we pass from the second to the third using (6.16). []

283

284

Figure 6.18 The OPT values


for the problem of aligning
the words mean to name.

Chapter 6 Dynamic Programming

Thus the value of the optima! alignment is the length of the shortest path
in Gxy from (0, 0) to (ra, n). (Wel! call any path in Gxv from (0, 0) to (ra, n)
a corner-to-comer path.) Moreover, the diagonal edges used in a shortest path
correspond precisely to the pairs used in a minimum-cost alignment. These
connections to the Shortest-Path Problem in the graph Gxv do not directly yield
an improvement in the running time for the sequence alignment problem;
however, they do help ones intuition for the problem and have been useful in
suggesting algorithms for more complex variations on sequence alignment.
For an example, Figure 6.18 shows the value of the shortest path from (0, 0)
to each node (i,j) for the problem of aligning the words raean and narae. For
the purpose of this example, we assume that 8 = 2; matching a vowel with
a different vowel, or a consonant with a different consonant, costs 1; while
matching a vowel and a consonant with each other costs 3. For each cel! in
the table (representing the corresponding node), the arrow indicates the last
step of the shortest path leading to that node--in other words, the way that
the minimum is achieved in (6.16). Thus, by following arrows backward from
node (4, 4), we can trace back to construct the alignment.

6.7 Sequence Alignment in Linear Space via


Divide and Conquer
In the previous section, we showed how to compute the optimal alignment
between two strings X and Y of lengths ra and n, respectively. Building up the
two-dimensional ra-by-n array of optimal solutions to subproblems, OPT(-, -),
turned out to be equivalent to constructing a graph Gxv with ran nodes laid
out in a grid and looking for the cheapest path between opposite corners. In
either of these ways of formulating the dynamic programming algorithm, the
running time is O(ran), because it takes constant time to determine the value
in each of the ran cells of the array OPT; and the space requirement is O(ran)
as well, since it was dominated by the cost of storing the array (or the graph

Gxv).
~ The Problem
The question we ask in this section is: Should we be happy with O(ran)
as a space bound? If our application is to compare English words, or even
English sentences, it is quite reasonable. In biological applications of sequence
alignment, however, one often compares very long strings against one another;
and in these cases, the (ran) space requirement can potentially be a more
severe problem than the (ran) time requirement. Suppose, for example, that
we are comparing two strings of 100,000 symbols each. Depending on the
underlying processor, the prospect of performing roughly !0 billion primitive

6.7 Sequence Alignment in Linear Space via Divide and Conquer

operations might be less cause for worry than the prospect of working with a
single 10-gigabyte array.
Fortunately, this is not the end of the story. In this section we describe a
very clever enhancement of the sequence alignment algorithm that makes it
work in O(ran) time using only O(ra + n) space. In other words, we can bring
the space requirement down to linear while blowing up the running time by
at most an additional constant factor. For ease of presentation, well describe
various steps in terms of paths in the graph Gxr, with the natural equivalence
back to the sequence alignment problem. Thus, when we seek the pairs in
an optimal alignment, we can equivalently ask for the edges in a shortest
corner-to-corner path in Gx~.
The algorithm itself will be a nice application of divide-and-conquer ideas.
The crux of the technique is the observation that, if we divide the problem
into several recursive calls, then the space needed for t_he computation can be
reused from one cal! to the next. The way in which this idea is used, however,
is fairly subtle.

,~ Designing the Algorithm


We first show that if we only care about the value of the optimal alignment,
and not the alignment itself, it is easy to get away with linear space. The
crucial observation is that to fill in an entry of the array A, the recurrence in
(6.16) only needs information from the current column of A and the previous
column of A. Thus we will "collapse" the array A to an rax 2 array B: as the
algorithm iterates through values of j, entries of the form B[i, 0] will hold the
"previous" columns value A[i,j - 1], while entries of the form B[i, 1] will hold
the "current" columns value A[i,j].
Space-El f icient -Alignment (X, Y)
Array B[0...m, 0... I]
Initialize B[i,O]----i8 for each i (just as in column 0 of A)
For ]=i ..... n
B[0, I]----]8 (since this corresponds to entry A[0,]])
For i=I ..... ra
B[i, 1] = min[c~xg~3 + B[i - 1, 0],
8+B[i-l, 1], 8+B[i,O]]
Endor
Move column I of B to column 0 to make room for next iteration:
Update B[i, 0]----B[i, I] for each i
Endfor

285

286

Chapter 6 Dynamic Programming

It is easy to verify that when this algorithm completes, the array entry
B[i, 1] holds the value of OPT(i, rt) for i = 0, 1 ..... m. Moreover, it uses O(mn)
time and O(m) space. The problem is: where is the alignment itselff We
havent left enough information around to be able to run a procedure like
Find-Alignment. Since B at the end of the algorithm only contains the last
two columns of the original dynamic programming array A, if we were to try
tracing back to get the path, wed run out of information after iust these two
columns. We could imagine getting around this difficulty by trying to "predict"
what the alignment is going to be in the process of running our space-efficient
procedure. In particular, as we compute the values in the jth column Of the
(now implicit) array A, we could try hypothesizing that a certain entry has a
very small value, and hence that the alignment that passes through this entry
is a promising candidate to be the optimal one. But this promising alignment
might run into big problems later on, and a different alignment that currently
looks much less attractive could turn out to be the optimal one.
There is, in fact, a solution to this problem--we will be able to recover
the alignment itself using O(m + n) space--but it requires a genuinely new
idea. The insight is based on employing the divide-and-conquer technique
that weve seen earlier in the book. We begin with a simple alternative way to
implement the basic dynamic programming solution.
A Backward Formulation of the Dynamic Program Recall that we use f(i, j)
to denote the length of the shortest path from (0, 0) to (i, j) in the graph Gxv.
(As we showed in the initial sequence alignment algorithm, [(i,j) has the
same value as OPT(i,j).) Now lets define g(i,j) to be the length of the shortest
path from (i, ]) to (m, n) in Gxv. The function g provides an equally natural
dynamic programming approach to sequence alignment, except that we build
it up in reverse: we start with g(m, n) = 0, and the answer we want is g(0, 0).
By strict analogy with (6.16), we have the following recurrence for g.
(6.18) For i < mandj < n we have
g(i, ]) = min[c%+~yj+1 + g(i + 1, j + 1), ~ + g(i, ] + 1), 3 - g(i + 1, j)].

This is just the recurrence one obtains by taking the graph GxT, "rotating"
it so that the node (m, n) is in the lower left corner, and using the previous approach. Using this picture, we can also work out the full dynamic programming
algorithm to build up the values of g, backward starting from (m, n). Similarly,
there is a space-efficient version of this backward dynamic programming algorithm, analogous to Space-Efficient-Alignment, which computes the
value of the optimal alignment using ordy O(m n) space. We will refer to

6.7 Sequence Alignment in Linear Space via Divide and Conquer


this backward version, naturally enough, as Backward-Space-EfficientAlignment.

Combining the Forward and Backward Formulations So now we have


syrmnetric algorithms which build up the values of the functions f and g.
The idea will be to use these two algorithms in concert to find the optimal
alignment. First, here are two basic facts summarizing some relationships
between the functions f and g.
(6.19) The ler~th of the shortest comer-to-comer path in Gxy that passes
through (i,j) is [(i,j) + g(i,j).
Proof. Let ~q denote the length of the shortest corner-to-corner path in Gxv
that passes through (i,j). Clearly, any such path must get from (0, 0) to (i,j)
and then from (i,j) to (m, n). Thus its length is at least [(i,j) +g(i,j), and so
we have ~ii > f(i,j) + g(i,j). On the other hand, consider the corner-to-corner
path that consists of a minimum-length path from (0, 0) to (i, j), followed by a
minimum-length path from (i,j) to (m, n). This path has length f(i, ]) + g(i, j),
and so we have ~0 <- [(i, j) + g(i, j). It follows that gij = [(i, j) + g(i, j). []
(6,20) Let k be any number in {0, .,n}, and let q be an index that
minimizes the quantity [(q, k) + g(q, k). Then there is a comer-to-comer path
of minimum length that passes through the node (q, k).
Proof. Let ~* denote the length of the shortest corner-to-corner path in Gxy.
Now fix a value of k ~ {0 ..... n}. The shortest corner-to-corner path must use
some node in the kth column of Gx~,--lets suppose it is node (p, k)--and thus
by (6.19)
e.* = f(p, k) + g(p, k) >_ rain
q f(q, k) + g(q, k).

Now consider the index q that achieves the minimum in the right-hand side
of this expression; we have

~.* >_ f(q, k) + g(q, k).


By (6.19) again, the shortest corner-to-corner path using, the node (q, k) has
length f(q, k) + g(q, k), and since g* is the minimum length of any corner-tocorner path, we have
~* <~ f(q, k) + g(q, k).
It follows that ~* = f(q, k) + g(q, k). Thus the shortest corner-to-corner path
using the node (q, k) has length ~*, and this proves (6.20). []

287

288

Chapter 6 Dynamic Programming

Using (6.20) and our space-efficient algorithms to compute the value of the
optimal alignment, we will proceed as follows. We divide Gxy along its center
column and compute the value of f(f, n/2) and g(i, n/2) for each value of i,
using our two space-efficient algorithms. We can then determine the minimum
value of f(i, n/2) + g(f, n/2), and conclude via (6.20) that there is a shortest
corner-to-corner path passing through the node (f, n/2). Given this, we can
search for the shortest path recursively in the portion of Gxy between (0, 0)
and (i, n/2) and in the portion between (i, n/2) and (m, n). The crucial point
is that we apply these recursive calls sequentially and reuse the working space
from one call to the next. Thus, since we only work on one recursive call at a
time, the total space usage is O(m + n). The key question we have to resolve
is whether the running time of this algorithm remains O(rnn).
In running the algorithm, we maintain a globally accessible list P which
will hold nodes on the shortest corner-to-corner path as they are discovered.
Initially, P is empty. P need only have rn + n entries, since no corner-to-corner
path can use more than this many edges. We also use the following notation:
X[i :j], for 1 _< f < j _< rn, denotes the substring of X consisting of xixi+l ...x j;
and we define Y[i :j] analogously. We will assume for simplicity that n is a
power of 2; this assumption makes the discussion much cleaner, although it
can be easily avoided.
Divide-and-Conquer-Alignment (X, Y)
Let m be the number of symbols in X
Let n be the number of symbols in Y
If m_<2 or ~<_2 then
Compute optimal alignment using Alignment (X , Y)
Call Space-Efficient-Alignment (X, Y[I : n/2])
Call Backward-Space-Efficient-Alignment (X, Y[n/2 + 1 : hi)
Let q be the index minimizing [(q, n/2)+g(q, ~/2)
Add (q, n/Z) to global list P
Divide-and-Conquer-Alignment (X[I : q], Y[I : n/2])
Divide-and-Conquer-Alignment (X[q + 1 : n], Y[n/2 + 1 : hi)
Return P

As an example of the first level of recursion, consider Figure 6.19. If the


minimizing index q turns out to be 1, we get the two subproblems pictured.

Analyzing the Algorithm


The previous arguments already establish that the algorithm returns the correct
answer and that it uses O(m + n) space. Thus, we need only verify the
following fact.

6.7 Sequence Alignment in Linear Space via Divide and Conquer


Second recursive call

Yl

Y2

Y3

First recursive call


Figure 6.19 The first level of recurrence for the space-efficient Divide-and-ConquerAlignment. The two boxed regions indicate the input to the two recursive cells.

(6.21) The running t~me of Divide-and-Conquer-Alignment on strings


length m and n is O(mn).
Proof. Let T(m, n) denote the maximum running time of the algorithm on
strings of length m and n. The algorithm performs O(mn) work to build up
the arrays B and B; it then runs recursively on strings of size q and n/2, and
on strings of size m - q and n/2. Thus, for some constant c, and some choice
of index q, we have
T(ra, n) < cran + T(q, n/2) + T(m - q, n/2)
T(m, 2) < cm
T(2, n) <_ cn.
This recurrence is more complex than the ones weve seen in our earlier
applications of divide-and-conquer in Chapter 5. First of all, the running time
is a function of two variables (m and n) rather than just one; also, the division
into subproblems is not necessarily an "even split," but instead depends on
the value q that is found through the earlier work done by the algorithm.
So how should we go about solving such a recurrence? One way is to
try guessing the form by considering a special case of the recurrence, and
then using partial substitution to fill out the details of this guess. Specifically,
suppose that we were in a case in which rn = n, and in which the split point
q were exactly in the middle. In this (admittedly restrictive) special case, we
could write the function T(.) in terms of the single variable n, set q = n/2
(since were assuming a perfect bisection), and have
T(n) < 2T(n/2) + cn2.

289

290

6.8 Shortest Paths in a Graph

Chapter 6 Dynamic Programming

This is a useful expression, since its something that we solved in our earlier
discussion of recurrences at the outset of Chapter 5. Specifically, this recurrence implies T(n) = O(n2).
So when m = n and we get an even split, the running time grows like the
square of n. Motivated by this, we move back to the fully general recurrence
for the problem at hand and guess that T(m, n) grows like the product of m and
n. Specifically, well guess that T(m, n) ! kmn for some constant k, and see if
we can prove this by induction. To start with the base cases m ! 2 and n ! 2,
we see that these hold as long as k >_ c/2. Now, assuming T(m, n) ! kmn
holds for pairs (m, n) with a smaller product, we have
T(m, n) ! cmn + T(q, n/2) + T(m - q, n/2)
! cmn + kqn/2 + k(m - q)n/2
= cmn + kqn/2 + kmn/2 - kqn/2
= (c + k/2)mn.
Thus the inductive step will work if we choose k = 2c, and this completes the,
proof. []

6.8 Shortest Paths in a Graph


For the final three sections, we focus on the problem of finding shortest paths
in a graph, together with some closely related issues.
~ The Problem
Let G = (V, E) be a directed graph. Assume that each edge (i, j) ~ E has an
associated weight cij. The weights can be used to model a number of different
things; we will picture here the interpretation in which the weight cq represents
a cost for going directly from node f to node j in the graph.
Earlier we discussed Diikstras Algorithm for finding shortest paths in
graphs with positive edge costs. Here we consider the more complex problem
in which we seek shortest paths when costs may be negative. Among the
motivations for studying this problem, here are two that particularly stand
out. First, negative costs turn out to be crucial for modeling a number of
phenomena with shortest paths. For example, the nodes may represent agents
in a financial setting, and cq represents the cost of a transaction in which
we buy from agent i and then immediately sell to agent j. In this case, a
path would represent a succession of transactions, and edges with negative
costs would represent transactions that result in profits. Second, the algorithm
that we develop for dealing with edges of negative cost turns out, in certain
crucial ways, to be more flexible and decentralized than Dijkstras Algorithm.
As a consequence, it has important applications for the design of distributed

routing algorithms that determine the most efficient path in a communication


network.

In this section and the next two, we will consider the following two related
problems.
Given a graph G with weights, as described above, decide if G has a
negative cycle--that is, a directed cycle C such that

o If the graph has no negative cycles, find a path P from an origin node s
to a destination node t with minimum total cost:
Ec
ijEP

should be as small as possible for any s-t path. This is generally called
both the Minimum-Cost Path Problem and the Shortest-Path Problem.

In terms of our financial motivation above, a negative cycle corresponds to a


profitable sequence of transactions that takes us back to our starting point: we
buy from i1, sell to i2, buy from i2, sell to i3, and so forth, finally arriving back
at i1 with a net profit. Thus negative cycles in such a network can be viewed
as good arbitrage opportunities.
It makes sense to consider the minimum-cost s-t path problem under the
assumption that there are no negative cycles. As illustrated by Figure 6.20, if
there is a negative cycle C, a path Ps from s to the cycle, and another path Pt
from the cycle to t, then we can build an s-t path of arbitrarily negative cost:
we first use Ps to get to the negative cycle C, then we go around C as many
times as we want, and then we use Pt to get from C to the destination t.

/J Designing and Analyzing the Algorithm


A Few False Starts Lets begin by recalling Dijkstras Algorithm for the
Shortest-Path Problem when there are no negative costs. That method

Figure 6.20 In this graph, one can find s-t paths of arbitrarily negative cost (by going
around the cycle C many times).

291

6.8 Shortest Paths in a Graph

Chapter 6 Dynamic Programming

292

(a)

Figure 6.21 (a) With negative


edge costs, Dijkstras AlgorithIn can give the wrong
answer for the Shortest-Path
Problem. (b) Adding 3 to the
cost of each edge wi!l make
all edges nonnegative, but it
will change the identity of the
shortest s-t path.

computes a shortest path from the origin s to every other node v in the graph,
essentially using a greedy algorithm. The basic idea is to maintain a set S
with the property that the shortest path from s to each node in S is known.
We start with S = {s}--since we know the shortest path from s to s has cost 0
when there are no negative edges--and we add elements greedily to this set S.
As our first greedy step, we consider the minimum-cost edge leaving node s,
that is, mini~v csi. Let v be a node on which this minimum is obtained. A key
observation underlying Dijkstras Algorithm is that the shortest path from s
to v is the single-edge path {s, v}. Thus we can immediately add the node
to the set S. The path {s, v} is clearly the shortest to v if there are no negative
edge costs: any other path from s to v would have to start on an edge out of s
that is at least as expensive as edge (s,
The above observation is no longer true if we can have negative edge
costs. As suggested by the example in Figure 6.21 (a), a path that starts on an
expensive edge, but then compensates with subsequent edges of negative cost,
can be cheaperthan a path that starts on a cheap edge. This suggests that the
Diikstra-style greedy approach will not work here.
Another natural idea is to first modify the costs cij by adding some large
constant M to each; that is, we let c~j = c0 + M for each edge (i,)) ~ E. If the
constant M is large enough, then all modified costs are nonnegafive, and we
can use Diikstras Algorithm to find the minimum-cost path subiect to costs
c. However, this approach fails to find the correct minimum-cost paths with
respect to the original costs c. The problem here is that changing the costs from
c to c changes the minimum-cost path. For example (as in Figure 6.21(b)), if
a path P consisting of three edges is only slightly cheaper than another path
P that has two edges, then after the change in costs, P will be cheaper, since
we only add 2M to the cost of P while adding 3M to the cost of P.

P
Figure 6.22 The minimum-cost path P from v to t using at most i edges.

(6.22) I[ G has no negative cycles, then there is a shortest path firom s to t


that is simple (i.e., does not repeat nodes), and hence has at most n - 1 edges.
Proof. Since every cycle has nonnegative cost, the shortest path P from s to
t with the fewest number of edges does not repeat any vertex v. For if P did
repeat a vertex v, we could remove the portion of P between consecutive visits
to v, resulting in a path of no greater cost and fewer edges. ,,

Lets use OPT(i, ~) to denote the minimum cost of a v-t path using at most
i edges. By (6.22), our original problem is to compute OPT(n -- 1, S). (We could
instead design an algorithm whose subproblems correspond to the minimum
cost of an s-v path using at most i edges. This would form a more natural
parallel with Dijkstras Algorithm, but it would not be as natural in the context
of the routing protocols we discuss later.)
We now need a simple way to express OPT(i, V) using smaller subproblems.
We will see that the most natural approach involves the consideration of
many different options; this is another example of the principle of "multiway choices" that we saw in the algorithm for the Segmented Least Squares
Problem.
Lets fix an optimal path P representing OPT(i, V) as depicted in Figure 6.22.

A Dynamic Programming Approach We will try to use dynamic programruing to solve the problem of finding a shortest path from s to t when there
are negative edge costs but no negative cycles. We could try an idea that has
worked for us so far: subproblem i could be to find a shortest path using only
the first i nodes. This idea does not immediately work, but it can be made
to work with some effort. Here, however, we will discuss a simpler and more
efficient solution, the Bellman-Ford Algorithm. The development of dynamic
programming as a general algorithmic technique is often credited to the work
of Bellman in the !950s; and the Bellman-Ford Shortest-Path Algorithm was
one of the first applications.
The dynamic programming solution we develop will be based on the
following crucial observation.

o If the path P uses at most i - 1 edges, then OPT(i, V) = OPT(i -- 1, v).


o If the path P uses i edges, and the first edge is (v, w), then OPT(i, v) =
Cvw -]- OPT(i -- 1, w).
This leads to the following recursive formula.
(6.23) If i > 0 then

OPT(i,v)=min(oPT(i 1, v),min(OPT(i--lw)Cvw))i

Using this recurrence, we get the following dynamic programming algorithm to compute the value OPT(n -- 1, s).

293

6.8 Shortest Paths in a Graph

Chapter 6 Dynamic Programming

294

3
Ca)
012345
t

-2 -2 -2
3 3 3
3 3 2
0 0 Io 0

Co)
Figure 6.23 For the directed
graph in (a), the ShortestPath Algorithm constructs
the dynamic programming
table in (b).

Shortest-Path (G, s, t)
~= number of nodes in G
Array M[O...n- I, V]
Define M[O,t]=O and M[O,u]=oo for all other u~V
For i_--l,...,n-I
For u~V in any order
Compute A4[i, v] using the recurrence (6.23)
Endfor
Enddor
Return M[~t -- I, S]

The correctness of the method follows directly by induction from (6.23).


We can bound the running time as follows. The table M has n2 entries; and
each entry can take O(n) time to compute, as there are at most n nodes w ~ V
we have to consider.

If we are a little more careful in the analysis of the method above, we can
improve the running-time bound to O(mn) without significantly changing the
algorithm itself.
The Shoz~ZesZ2Path method can be implemented in O(mn) time:
Proof. Consider the computation of the array entry M[i, v] according to the
recurrence (6.23);. we have
M[i, v] = min(M[i - 1, v], min(M[i -- 1, w] + cu~)).
tu~V

We assumed it could take up to O(n) time to compute this minimum, since


there are n possible nodes w. But, of course, we need only compute this
minimum over all nodes w for which v has an edge to w; let us use nu to denote
this number. Then it takes time O(nu) to compute the array entry M[i, v]. We
have to compute an entry for every node v and every index 0 < i < n - 1, so
this gives a running-time bound of

(6.24) The Shortest-Path method correctly computes the minimum cost.of


an s-t path in any graph that has no negative cycles, and runs fn O(n3) time.
Given the table M containing the optimal values of the subproblems, the
shortest path using at most i edges can be obtained in O(in) time, by tracing
back through smaller subproblems.
As an example, consider the graph in Figure 6.23 (a), where the goal is to
find a shortest path from each node to t. The table in Figure 6.23 (b) shows the
array M, with entries corresponding to the values M[i, v] from the algorithm.
Thus a single row in the table corresponds to the shortest path from a particular
node to t, as we allow the path to use an increasing number of edges. For
example, the shortest path from node d to t is updated four times, as it changes
from d-t, to d-a-t, to d-a-b-e-t, and finally to d-a-b-e-c-t.

Extensions: Some Basic Improvements to the Algorithm


An Improved Running-Time Analysis We can actually provide a better
running-time analysis for the case in which the graph G does not have too
many edges. A directed graph with n nodes can have close to n2 edges, since
there could potentially be an edge between each pair of nodes, but many
graphs are much sparser than this. When we work with a graph for which
the number of edges m is significantly less than nz, weve already seen in a
number of cases earlier in the book that it can be useful to write the runningtime in terms of both m and n; this way, we can quantify our speed-up on
graphs with relatively fewer edges.

In Chapter 3, we performed exactly this kind of analysis for other graph


algorithms, and used (3.9) from that chapter to bound the expression ~usv nu
for undirected graphs. Here we are dealing with directed graphs, and nv denotes
the number of edges leaving v. In a sense, it is even easier to work out the
value of ~v,v nu for the directed case: each edge leaves exactly one of the
nodes in V, and so each edge is counted exactly once by this expression. Thus
we have ~u~v n~ = m. Plugging this into our expression

for the running time, we get a running-time bound of O(mn).


Improving the Memory Requirements We can also significantly improve the
memory requirements with only a small change to the implementation. A
common problem with many dynamic programming algorithms is the large
space usage, arising from the M array that needs to be stored. In the BellmanFord Algorithm as written, this array has size n2; however, we now show how
to reduce this to O(n). Rather than recording M[i, v] for each value i, we will
use and update a single value M[v] for each node v, the length of the shortest
path from v to t that we have found so far. We still run the algorithm for

295

6.9 Shortest Paths and Distance Vector Protocols

296

Chapter 6 Dynamic Programming

iterations i = 1, 2 ..... n - 1, but the role of i will now simply be as a counter;


in each iteration, and for each node v, we perform the update
M[v] = min(M[v], ~m~r~(cuw + M[tu])).

We now observe the following fact.

(6.26) Throughout the algorithm M[v] is the length of some path from v to
t, and after i rounds of updates the value M[v] is no larger than the length of
the shortest path from v to t using at most i edges.
Given (6.26), we can then use (6.22) as before to show that we are done after
n - 1 iterations. Since we are only storing an M array that indexes over the
nodes, this requires only O(n) working memory.
Finding the Shortest Paths One issue to be concerned about is whether this
space-efficient version of the algorithm saves enough information to recover
the shortest paths themselves. In the case of the Sequence Alignment Problem
in the previous section, we had to resort to a tricky divide-and-conquer.methpd
to recover the solution from a similar space-efficient implementation. Here,
however, we will be able to recover the shortest paths much more easily.
To help with recovering the shortest paths, we will enhance the code by
having each node v maintain the first node (after itself) on its path to the
destination t; we will denote this first node by first[v]. To maintain first[v],
we update its value whenever the distance M[v] is updated. In other words,
whenever the value of M[v] is reset to the minimum
t~EVrnin(cvro + M[w]), we set
first[v] to the node w that attains this minimum.
Now let P denote the directed "pointer graph" whose nodes are V, and
whose edges are {(v, first[v])}. The main observation is the following.
(6.27) If the pointer graph P contains a cycle C, then this cycle.must have
negative cost.
Proof. Notice that if first[v] = w at any time, then we must have M[v] >_
cvw +M[w]. Indeed, the left- and right-hand sides are equal after the update
that sets first[v] equal to w; and since M[w] may decrease, this equation may
turn into an inequality.
Let Vl, v2 ..... vk be the nodes along the cycle C in the pointer graph,
and assume that (vk, Vl) is the last edge to have been added. Now, consider
the vaiues right before this last update. At this time we have M[v~] >_ cvi~i+~ +
M[vi+l] for all i = 1 ..... k - !, and we also have M[vk] > cvkul +M[vl] since
we are about to update M[vk] and change first[v~] to Vl. Adding all these
inequalities, the M[vi] values cance!, and we get 0 > ~g-li=l c,ivi+~ + cvm:
a
negative cycle, as claimed. []

Now note that if G has no negative cycles, then (6.27) implies that the
pointer graph P will never have a cycle. For a node v, consider the path we
get by following the edges in P, from v to first[v] = v~, to first[v~] = v2, and so
forth. Since the pointer graph has no cycles, and the sink t is the only node
that has no outgoing edge, this path must lead to t. We claim that when the
algorithm terminates, this is in fact a shortest path in G from v to t.
(6.28) Suppose G has no negative cycles, and consider the pointer graph P
at the termination, of the algorithm. For each node v, the path in P from v to t
is a shortest v-t path in G.
Proof. Consider a node v and let tv = first[v]. Since the algorithm terminated,
we must have M[v] = cuw + M[w]. The value M[t] = O, and hence the length
of the path traced out by the pointer graph is exactly M[v], which we know is
the shortest-path distance. []

Note that in the more space-efficient version of Bellman-Ford, the path


whose length is M[v] after i iterations can have substantially more edges than
i. For example, if the graph is a single path from s to t, and we perform updates
in the reverse of the order the edges appear on the path, then we get the final
shortest-path values in just one iteration. This does not always happen, so we
cannot claim a worst-case running-time improvement, but it would be nice to
be able to use this fact opportunisticaBy to speed up the algorithm on instances
where it does happen. In order to do this, we need a stopping signal in the
algorithm--something that tells us its safe to terminate before iteration n - 1
is reached.
Such a stopping signal is a simple consequence of the following observation: If we ever execute a complete iteration i in which no M[v] value changes,
then no M[v] value will ever change again, since future iterations will begin
with exactly the same set of array entries. Thus it is safe to stop the algorithm.
Note that it is not enough for a particular M[v] value to remain the same; in
order to safely terminate, we need for all these values to remain the same for
a single iteration.

6.9 Shortest Paths and Distance Vector Protocols


One important application of the Shortest-Path Problem is for routers in a
communication network to determine the most efficient path to a destination.
We represent the network using a graph in which the nodes correspond to
routers, and there is an edge between v and tv if the two touters are connected
by a direct communication link. We define a cost cuw representing the delay on
the link (v, w); the Shortest-Path Problem with these costs is to determine t_he
path with minimum delay from a source node s to a destination t. Delays are

297

298

6.9 Shortest Paths and Distance Vector Protocols

Chapter 6 Dynamic Programming

naturally nonnegative, so one could use Dijkstras Algorithm to compute the


shortest path. However, Dijkstras shortest-path computation requires global
knowledge of the network: it needs to maintain a set S of nodes for which
shortest paths have been determined, and make a global decision about which
node to add next to S. While reuters can be made to run a protocol in the
background that gathers enough global information to implement such an
algorithm, it is often cleaner and more flexible to use algorithms that require
only local knowledge of neighboring nodes.
If we think about it, the Bellman-Ford Algorithm discussed in the previous
section has just such a "local" property. Suppose we let each node v maintain
its value M[v]; then to update this value, u needs only obtain the value M[w]
fromeach neighbor w, and compute

based on the information obtained.


We now discuss an improvement to the Bellman-Ford Algorithm that
makes it better suited for reuters and, at the same time, a faster algorithm
in practice. Our current implementation of the Bellman-Ford Algorithm can be
thought of as a pull-based algorithm. In each iteration i, each node v has to
contact each neighbor w, and "pull" the new value M[w] from it. If a node w
has not changed its value, then there is no need for ~ to get the value again;
however, u has no way of knowing this fact, and so it must execute the pnll
anyway.
This wastefulness suggests a symmetric push-based implementation,
where values are only transmitted when they change. Specifically, each node
w whose distance value M[w] changes in an iteration informs all its neighbors
of the new value in the next iteration; this allows them to update their values
accordingly. If M[w] has not changed, then the neighbors of w already have
the current value, and there is no need to "push" it to them again. This leads
to savings in the running time, as not all values need to be pushed in each iteration. We also may terminate the algorithm early, if no value changes during
an iteration. Here is a concrete description of the push-based implementation.

For all edges (U, w) in any order


M[u] = min(M[u], cuw + M[w])
If this changes the value of M[U], then first[u]=w
End/or
End/or
If no value changed in this iteration, then end the algorithm
End/or
Return M[S]

In this algorithm, nodes are sent updates of their neighbors distance


values in rounds, and each node sends out an update in each iteration in which
it has changed. However, if the nodes correspond to reuters in a network, then
we do not expect everything to run in lockstep like this; some reuters may
report updates much more quickly than others, and a router with an update to
report may sometimes experience a delay before contacting its neighbors. Thus
the renters will end up executing an asynchronous version of the algorithm:
each time a node w experiences an update to its M[w] value, it becomes
"active" and eventually notifies its neighbors of the new value. If we were
to watch the behavior of all reuters interleaved, it would look as follows.
Asynchronous-Shortest-Path(G, s, t)
n= number of nodes in G
Array M[V]
Initialize M[t]=0 and M[u]=oo for all other uE V
Declare t to be active and all other nodes inactive
While there exists an active node
Choose an active node u)
For all edges (u, uT) in any order
M[u] = min(M[u], cw.u + M[w])
If this changes the value of M[u], then
first[u] = w
u becomes active
End/or
u~ becomes inactive
EndWhile

Push-Based-Shortest-Path(G, s, t)
~= number of nodes in G
Array M[V]
Initialize M[t]=O and M[u]=oo for all other u ~ V
For 1=1 ..... n-1
For 1//~ V in any order
If M[uT] has been updated in the previous iteration then

One can show that even this version of the algorithm, with essentially no
coordination in the ordering of updates, will converge to the correct values of
the shortest-path distances to t, assuming only that each time a node becomes
active, it eventually contacts its neighbors.
The algorithm we have developed here uses a single destination t, and
all nodes v ~ V compute their shortest path to t. More generally, we are

299

6.10 Negative Cycles in a Graph

300

Chapter 6 Dynamic Programming


presumably interested in finding distances and shortest paths between all pairs
of nodes in a graph. To obtain such distances, we effectively use n separate
computations, one for each destination. Such an algorithm is referred to as
a distance uector protocol, since each node maintains a vector of distances to
every other node in the network.

Problems with the Distance Vector Protocol


One of the major problems with the distributed implementation of BellmanFord on routers (the protocol we have been discussing above) is that its derived
from an initial dynamic programming algorithm that assumes edge costs will
remain constant during the execution of the algorithm. Thus far weve been
designing algorithms with the tacit understanding that a program executing
the algorithm will be running on a single computer (or a centrally managed
set of computers), processing some specified input. In this context, its a rather
benign assumption to require that the input not change while the progra_m is
actually running. Once we start thinking about routers in a network, however,
this assumption becomes troublesome. Edge costs may change for all ~sorts of
reasons: links can become congested and experience slow-downs; or a link
(v, w) may even fail, in which case the cost c~ effectively increases to oo.
Heres an indication of what can go wrong with our shortest-path algorithm when this happens. If an edge (v, w) is deleted (say the link goes down),
it is natural for node v to react as follows: it should check whether its shortest
path to some node t used the edge (v, w), and, if so, it should increase the
distance using other neighbors. Notice that this increase in distance from v can
now trigger increases at vs neighbors, if they were relying on a path through v,
and these changes can cascade through the network. Consider the extremely
simple example in Figure 6.24, in which the original graph has three edges
(s, v), (v, s) and (u, t), each of cost 1.
Now suppose the edge (v, t) in Figure 6.24 is deleted. How dbes node v
react? Unfortunately, it does not have a global map of the network; it only
knows the shortest-path distances of each of its neighbors to t. Thus it does

~s

The deleted edge causes an unbou----nded~


equence of updates by s and u.

Figure 6.24 When the edge (v, t) is deleted, the distributed Bellman-Ford Algorithm
will begin "counting to infiniW."

not know that the deletion of (v, t) has eliminated all paths from s to t. Instead,
it sees that M[s]= 2, and so it updates M[v] =Cvs +M[s] = 3, assuming that
it will use its cost-1 edge to s, followed by the supposed cost-2 path from s
to t. Seeing this change, node s will update M[s] = csv +M[v] = 4, based on
its cost-1 edge to v, followed by the supposed cost-3 path from v to t. Nodes
s and v will continue updating their distance to t until one of them finds an
alternate route; in the case, as here, that the network is truly disconnected,
these updates will continue indefinitely--a behavior known as the problem of
counting to infinity.
To avoid this problem and related difficulties arising from the limited
amount of information available to nodes in the Bellman-Ford Algorithm, the
designers of network routing schemes have tended to move from distance
vector protocols to more expressive path vector protocols, in which each node
stores not just the distance and first hop of their path to a destination, but
some representation of the entire path. Given knowledge of the paths, nodes
can avoid updating their paths to use edges they know to be deleted; at the
same time, they require significantly more storage to keep track of the full
paths. In the history of the Internet, there has been a shift from distance vector
protocols to path vector protocols; currently, the path vector approach is used
in the Border Gateway Protocol (BGP) in the Internet core.

6.10 Negative Cycles in a Graph


So far in our consideration of the Bellman-Ford Algorithm, we have assumed
that the underlying graph has negative edge costs but no negative cycles. We
now consider the more general case of a graph that may contain negative
cycles.

/~:~ The Problem


There are two natural questions we will consider.
How do we decide if a graph contains a negative cycle?
How do we actually find a negative cycle in a graph that contains one?
The algorithm developed for finding negative cycles will also lead to an
improved practical implementation of the Bellman-Ford Algorithm from the
previous sections.
It turns out that the ideas weve seen so far will allow us to find negative
cycles that have a path reaching a sink t. Before we develop the details of this,
lets compare the problem of finding a negative cycle that can reach a given t
with the seemingly more natural problem of finding a negative cycle anywhere
in the graph, regardless of its position related to a sink. It turns out that if we

301

302

6.10 Negative Cycles in a Graph

Chapter 6 Dynamic Programming

(Any negative cycle in G wil! be able to reach i.~

get shorter and shorter as we go around a negative cycle. In fact, for any node
v on a negative cycle that has a path to t, we have the following.
(6.30) If node v can reach node t and is contained in a negative cycle, then
lim OPT(i, v)

If the graph has no negative cycles, then (6.22) implies following statement.
(6.31) If there are no negative cycles in G, then OPT(i, V) = OPT(n -- !, V) for
all nodes v and all i > n.

Figure 6.25 The augmented graph.

develop a solution to the first problem, wel! be able to obtain a solution to


the second problem as well, in the following way. Suppose we start with a
graph G, add a new node t to it, and connect each other node v in the graph
to node t via an edge of cost 0, as shown in Figure 6.25. Let us call the new
"augmented graph" G.
(6.29) The augmented graph G has a negative cycle C such that there is a
path from C to the sink t if and only if the original graph has a negative cycle.
Proof. Assume G has a negative cycle. Then this cycle C clearly has an edge
to t in G, since all nodes have an edge to t.
Now suppose G has a negative cycle with a path to t. Since no edge leaves
t in G, this cycle cannot contain t. Since G is the same as G asidefrom the
node t, it follows that this cycle is also a negative cycle of G. a

So it is really enough to solve the problem of deciding whether G has a


negative cycle that has a path to a given sink node t, and we do this now.

~ Designing and Analyzing the Algorithm


To get started thinking about the algorithm, we begin by adopting the original
version of the BeLlman-Ford Algorithm, which was less efficient in its use
of space. We first extend the definitions of OPT(i, v) from the Bellman-Ford
Algorithm, defining them for values i >_ n. With the presence of a negative
cycle in the graph, (6.22) no longer applies, and indeed the shortest path may

But for how large an i do we have to compute the values OPT(i, V) before
concluding that the graph has no negative cycles? For example, a node v may
satisfy the equation OPT(n, V) = OPW(n- 1, v), and yet still lie on a negative
cycle. (Do you see why?) However, it turns out that we will be in good shape
if this equation holds for all nodes.
(6.32) There is no negative cyc!e with a path to tif and only if opT(n,

Proof. Statement (6.31) has already proved the forward direction. For the other
direction, we use an argument employed earlier for reasoning about when its
safe to stop the Bellman-Ford Algorithm early. Specifically, suppose OPT(n, v) =
OPT(n -- 1, V) for all nodes v. The values of OPT(n + 1, V) can be computed
from OPT(n, v); but all these values are the same as the corresponding OPW(n -1, v). It follows that we will have OPT(n + 1, v) = OPT(n -- !, V). Extending this
reasoning to future iterations, we see that none of the values will ever change
again, that is, OPT(i, v) = OPT(n -- 1, V) for al! nodes v and all i >_ n. Thus there
cannot be a negative cycle C that has a path to t; for any node w on this cycle
C, (6.30) implies that the values OPT(i, w) would have to become arbitrarily
negative as i increased. ,,

Statement (6.52) gives an O(mn) method to decide if G has a negative


cycle that can reach t. We compute values of OPT(i, v) for nodes of G and for
values of i up to n. By (6.32), there is no negative cycle if and only if there is
some value of i < n at which OPT(i, v) = OPT(i -- 1, v) fo~: all nodes v.
So far we have determined whether or not the graph has a negative cycle
with a path from the cycle to t, but we have not actually found the cycle. To
find a negative cycle, we consider a node v such that OPT(n, V) 7~ OPT(n -- 1, V):
for this node, a path P from v to t of cost OPT(n, V) must use exactly n edges.
We find this minimum-cost path P from v to t by tracing back through the
subproblems. As in our proof of (6.22), a simple path can only have n- !

303

304

6.10 Negative Cycles in a Graph

Chapter 6 Dynamic Programming

edges, so P must contain a cycle C. We claim that this cycle C has negative
cost.
(6.33) If G has n nodes and OPT(n, v) ~ OPT(n -- 1, V), then a path P from v
to t of cost OPT(n, v) contains a cycle C, and C has negative cost.

Proof. First observe that the path P must have n edges, as OPT(n, V) ~ OPT(n -1, v), and so every path using n - 1 edges has cost greater than that of the
path P. In a graph with n nodes, a path consisting of n edges must repeat
a node somewhere; let w be a node that occurs on P more than once. Let C
be the cycle on P between two consecutive occurrences of node w. If C were
not a negative cycle, then deleting C from P would give us a v-t path with
fewer than n edges and no greater cost. This contradicts our assumption that
OPT(n, v) ~: O~T(n -- 1, V), and hence C must be a negative cycle. []
(6.34) The algorithm above finds a negative cycle in G, if such a cycle e:fists,
and runs in O(rnn) time.

Extensions: Improved Shortest Paths and Negative Cycle


Detection Algorithms
At the end of Section 6.8 we discussed a space-efficient implementation of the
Bellman-Ford algorithm for graphs with no negative cycles. Here we implement
the detection of negative cycles in a comparably space-efficient way. In addition
to the savings in space, this will also lead to a considerable speedup in practice
even for graphs with no negative cycles. The implementation will be based on
the same pointer graph P derived from the "first edges" (v, firstly]) that we
used for the space-efficient implementation in Section 6.8. By (6.27), we know
that if the pointer graph ever has a cycle, then the cycle has negative.cost, and
we are done. But if G has a negative cycle, does this guarantee that the pointer
graph will ever have a cycle? Furthermore, how much extra computation time
do we need for periodically checking whether P has a cycle?
Ideally, we would like to determine whether a cycle is created in the pointer
graph P every time we add a new edge (v, w) with firstly] = w. An additional
advantage of such "instant" cycle detection will be that we will not have to wait
for n iterations to see that the graph has a negative cycle: We can terminate
as soon as a negative cycle is found. Earlier we saw that if a graph G has no
negative cycles, the algorithm can be stopped early if in some iteration the
shortest path values M[v] remain the same for all nodes v. Instant negative
cycle detection wil! be an analogous early termination rule for graphs that
have negative cycles.

Consider a new edge (v, w), with first[v] = w, that is added to the pointer
graph P. Before we add (v, w) the pointer graph has no cycles, so it consists of
paths from each node v to the sink t. The most natural way to check whether
adding edge (v, w) creates a cycle in P is to follow the current path from tv to
the terminal t in time proportional to the length of this path. If we encounter
v along this path, then a cycle has been formed, and hence, by (6.27), the
graph has a negative cycle. Consider Figure 6.26, for example, where in both
(a) and (b) the pointer firstly] is being updated from u to tv; in (a), this does
not result in a (negative) cycle, but in (b) it does. However, if we trace out the
sequence of pointers from v like this, then we could spend as much as O(n)
time following the path to t and still not find a cycle. We now discuss a method
that does not require an O(n) blow-up in the running time.
We know that before the new edge (v, w) was added, the pointer graph
was a directed tree. Another way to test whether the addition of (v, rv) creates
a cycle is to consider al! nodes in the subtree directed toward v. If w is in this
subtree, then (v, rv) forms a cycle; otherwise it does not. (Again, consider the
two sample cases in Figure 6.26.) To be able to find all nodes in the subtree
directed toward v, we need to have each node v maintain a list of all other
nodes whose selected edges point to v. Given these pointers, we can find
the subtree in time proportional to the size of the subtree pointing to v, at
most O(n) as before. However, here we will be able to make additional use
of the work done. Notice that the current distance value Mix] for all nodes x
in the subtree was derived from node vs old value. We have just updated vs
distance, and hence we know that the distance values of all these nodes will
be updated again. Well mark each of these nodes x as "dormant," delete the

Update to

first[v] = w
(a)

Update to
first[v] = w

Figure 6.26 Changing the pointer graph P when firstly] is updated from u to w. In (b),
this creates a (negative) cycle, whereas in (a) it does not.

305

306

Solved Exercises

Chapter 6 Dynamic Programming

edge (x, first[x]) from the pointer graph, and not use x for future updates until
its distance value changes.
This can save a lot of future work in updates, but what is the effect on the
worst-case running time? We can spend as much as O(n) extra time marking
nodes dormant after every update in distances. However, a node can be marked
dormant only if a pointer had been defined for it at some point in the past, so
the time spent on marking nodes dormant is at most as much as the time the
algorithm spends updating distances.
Now consider the time the algorithm spends on operations other than
marking nodes dormant. Recall that the algorithm is divided into iterations,
where iteration i + 1 processes nodes whose distance has been updated in
iteration i. For the original version of the algorithm, we showed in (6.26) that
after i iterations, the value M[v] is no larger than the value of the shortest path
from v to t using at most i edges. However, with many nodes dormant in each
iteration, this may not be true anymore. For example, if the shortest path .from
v to t using at most i edges starts on edge e = (u, w), and w is dormant in
this iteration, then we may not update the distance value M[v], and so it stays
at a value higher than the length of the path through the edge (v, w). This
seems like a problem--however, in this case, the path through edge (u, w) is
not actually the shortest path, so M[v] will have a chance to get updated later
to an even smaller value.

So instead of the simpler property that held for M [v] in the original versions
of the algorithm, we now have the the following claim.
(6.35) Throughout the algorithm M[v] is the length of some simple path from
v to t; the path has at least i edges if the distance value M[v] is updated in
iteration i; and after i iterations, the value M[v] is the length of the shortest
path for all nodes v where there is a shortest v-t path using at most i edges.
Proof. The first pointers maintain a tree of paths to t, which implies that all
paths used to update the distance values are simple. The fact that updates in
iteration i are caused by paths with at least i edges is easy to show by induction
on i. Similarly, we use induction to show that after iteration i the value
is the distance on all nodes v where the shortest path from v to t uses at most
i edges. Note that nodes u where M[v] is the actual shortest-path distance
cannot be dormant, as the value M[u] will be updated in the next iteration for
all dormant nodes. E

Using this claim, we can see that the worst-case running time of the
algorithm is still bounded by O(mn): Ignoring the time spent on marking
nodes dormant, each iteration is implemented in O(m) time, and there can
be at most n - I iterations that update values in the array M without finding

a negative cycle, as simple paths can have at most n- 1 edges. Finally, the
time spent marking nodes dormant is bounded by the time spent on updates.
We summarize the discussion with the following claim about the worst-case
performance of the algorithm. In fact, as mentioned above, this new version
is in practice the fastest implementation of the algorithm even for graphs that
do not have negative cycles, or even negative-cost edges.
(6.36) The improved algorithm outlined above finds a negative cycle in G if
such a cycle exists. It terminates immediately if the pointer graph P of first[v]
pointers contains a cycle C, or if there is an iteration in which no update occurs
to any distance value M[v]. The algorithm uses O(n) space, has at most n
iterations, and runs in O(mn) time in the worst case.

Solved Exercises
Solved Exercise 1
Suppose you are managing the construction of billboards on the Stephen
Daedalus Memorial Highway, a heavily traveled stretch of road that runs
west-east for M miles. The possible sites for billboards are given by numbers
xl, x2 ..... Xn, each in the interval [0, M] (specifying their position along the
highway, measured in miles from its western end). If you place a billboard at
location xi, you receive a revenue of ri > 0.
Regulations imposed by the countys Highway Department require that
no two of the billboards be within less than or equal to 5 miles of each other.
Youd like to place billboards at a subset of the sites so as to maximize your
total revenue, subject to this restriction.
Example. Suppose M = 20, n = 4,
{x1, x2, x3, x4}={6, 7, 12, 14},

and

{rl, r2, r3, r4} = {5, 6, 5, 1}.


Then the optimal solution would be to place billboards at xl and x3, for a total
revenue of 10.
Give an algorithm that takes an instance of this problem as input and
returns the maximum total revenue that can be obtained from any valid subset
of sites. The running time of the algorithm should be polynomial in n.
Solution We can naturally apply dynamic programming to this problem if
we reason as follows. Consider an optimal solution for a given input instance;
in this solution, we either place a billboard at site xn or not. If we dont, the
optimal solution on sites xl ..... xn is really the same as the optimal solution

307

308

Solved Exercises

Chapter 6 Dynamic Programming

on sites x1 ..... xn-1; if we do, then we should ehminate xn and all other sites
that are within 5 miles of it, and find an optimal solution on whats left. The
same reasoning applies when were looking at the problem defined by just the
firstj sites, xl ..... xj: we either include xj in the optimal solution or we dont,
with the same consequences.
Lets define some notation to help express this. For a site xj, we let e(j)
denote the easternmost site xi that is more than 5 miles from xj. Since sites
are numbered west to east, this means that the sites xl, x2 ..... xeq) are still
valid options once weve chosen to place a billboard at xj, but the sites
Xeq)+~ ..... x~_~ are not.
Now, our reasoning above justifies the following recurrence. If we let OPT(j)
denote the revenue from the optimal subset of sites among x~ ..... xj, then we
have
OPT(]) ---- max(r/+ OPT(e(])), OPT(] -- 1)).

We now have most of the ingredients we need for a dynamic programming


algorithm. First, we have a set of n subproblems, consisting of the first j sites
for j = 0, 1, 2 ..... ft. Second, we have a recurrence that lets us build up the
solutions to subproblems, given by OPT(]) = max(r/+ OPT(e(])), OPT(] =- 1)).
To turn this into an algorithm, we just need to define an array M that will
store the OPT values and throw a loop around the recurrence that builds up
the values M[j] in order of increasing j.
Initi~ize M[0] = 0 and M[1] = r1
For j=2,3 ..... n:
Compute M~] using the recurrence
Enddor
Return M[n]

As with all the dynamic programming algorithms weve seen in this chapter,
an optimal set of billboards can be found by tracing back through the values
in array M.
Given the values e(]) for all j, the running time of the algorithm is O(n),
since each iteration of the loop takes constant time. We can also compute al! e(])
values in O(r0 time as follows. For each site location xi, we define xi = xi - 5.
We then merge the sorted list x~ ..... xn with the sorted list x~ ..... x~ in linear
time, as we saw how to do in Chapter 2. We now scan through this merged list;
when we get to the entry x;, we know that anything from this point onward
to xj cannot be chosen together with xy (since its within 5 miles), and so we

simply define e(]) to be the largest value of i for which weve seen xi in our
scan.
Heres a final observation on this problem. Clearly, the solution looks
very much fike that of the Weighted Interval Scheduling Problem, and theres
a fundamental reason for that. In fact, our billboard placement problem
can be directly encoded as an instance of Weighted Interval Scheduling, as
follows. Suppose that for each site xi, we define an interval with endpoints
[x~ - 5, xi] and weight ri. Then, given any nonoverlapping set of intervals, the
corresponding set of sites has the property that no two lie within 5 miles of
each other. Conversely, given any such set of sites (no two within 5 miles), the
intervals associated with them will be nonoverlapping. Thus the collections
of nonoveflapping intervals correspond precisely to the set of valid billboard
placements, and so dropping the set of intervals weve just defined (with their
weights) into an algorithm for Weighted Interval Scheduling will yield the
desired solution.

Solved Exercise 2
Through some Mends of friends, you end up on a consulting visit to the
cutting-edge biotech firm Clones R Us (CRU). At first youre not sure how
your algorithmic background will be of any help to them, but you soon find
yourself called upon to help two identical-looking software engineers tackle a
perplexing problem.
The problem they are currently working on is based on the concatenation
of sequences of genetic material. If X and Y are each strings over a fixed
alphabet g, then XY denotes the string obtained by concatenating them-writing X followed by Y. CRU has identified a target sequence A of genetic
material, consisting of ra symbols, and they want to produce a sequence that
is as similar to A as possible. For this purpose, they have a library L consisting
of k (shorter) sequences, each of length at most n. They can cheaply produce
any sequence consisting of copies of the strings in L concatenated together
(with repetitions allowed).
Thus we say that a concatenation over L is any sequence of the form
B1B2 B~, where each Bi belongs the set L. (Again, repetitions are allowed,
so B~ and Bj could be the same string in L, for different values of i and j.)
The problem is to find a concatenation over IBm} for which the sequence
alignment cost is as small as possible. (For the purpose of computing the
sequence alignment cost, you may assume that you are given a gap cost 8 and
a mismatch cost %q for each pair p, q E g.)
Give a polynomial-time algorithm for this problem.

309

310

Solved Exercises

Chapter 6 Dynamic Programming

Solution This problem is vaguely reminiscent of Segmented Least Squares:


we have a long sequence of "data" (the string A) that we want to "fit" with
shorter segments (the strings in L).
If we wanted to pursue this analogy, we could search for a solution as
follows. Let B = B1B2 Be denote a concatenation over L that aligns as well
as possible with the given string A. (That is, B is an optimal solution to the
input instance.) Consider an optimal alignment M of A with B, let t be the first
position in A that is matched with some symbol in Be, and let Ae denote the
substring of A from position t to the end. (See Figure 6.27 for an illustration
of this with g = 3.) Now, the point is that in this optimal alignment M, the
substring Ae is optimally aligned with B6 indeed, if there were a way to better
align Ae with Be, we could substitute it for the portion of M that aligns Ae with
Be and obtain a better overall alignment of A with B.
This tells us that we can look at the optimal solution as follows. Theres
some final piece of Aa that is aligned with one of the strings in L, and for_this
piece all were doing is finding the string in L that aligns with it as well as
possible. Having found this optimal alignment for Aa, we can break it bff al~d
continue to find the optimal solution for the remainder of A.
Thinking about the problem this way doesnt tell us exactly how to
proceed--we dont know how long A~ is supposed to be, or which ~tring in
L it should be aligned with. But this is the kind of thing we can search over
in a dynamic programming algorithm. Essentially, were in about the same
spot we were in with the Segmented Least Squares Problem: there we knew
that we had to break off some final subsequence of the input points, fit them
as well as possible with one line, and then iterate on the remaining input
points.
So lets set up things to make the search for A~ possible. First, let A[x : y]
denote the substring of A consisting of its symbols from position x to position
y, inclusive. Let c(x, y) denote the cost of the optimal alignment of A[x : y] with
any string in L. (That is, we search over each string in L and find the one that

Figure 6.27 In the optimal concatentation of strings to align with A, there is a final
string (B3 in the figure) that a~gns with a substring of A (A3 In the figure) that e~xtends
from some position t to the end.

aligns best with A[x :y].) Let OPT(J) denote the alignment cost of the optimal
solution on the string All :j].
The argument above says that an optimal solution on A[1 :j] consists of
identifying a final "segment boundary" t < j, finding the optimal alignment
of A[t :j] with a single string in L, and iterating on All : t - !]. The cost of
this alignment of A[t :j] is just c(t,j), and the cost of aligning with whats left
is just OPT(t -- 1). This suggests that our subproblems fit together very nicely,
and it justifies the. following recurrence.
(6.37) OPT(j) -- mint<] c(t,j) + OPT(t -- 1) forj > !, and OPT(0) ----- 0.

The fl~ algorithm consists of first computing the quantities c(t,j), for t < j,
and then building up the values OPT(j) in order of increasing j. We hold these
values in an array M.
Set M[0] = o
For all pairs l_<t_<j_<m
Compute the cost c(t,]) as follows:
For each string B EL
Compute the optimal alignment of B with A[t :]]
End/or
Choose the B that achieves the best alignment, and use
this alignment cost as c(t,])
End/or
For ]=1,2 ..... n
Use the recurrence (6.37) to compute M~]
End/or
Return M[n]

As usual, we can get a concatentation that achieves it by tracing back over


the array of OPT values.

Lets consider the running time of this algorithm. First, there are O(m2)
values c(t, j) that need to be computed. For each, we try each string of the
k strings B ~ L, and compute the optimal alignment of B with A[t :j] in
time O(n(j- t))= O(mn). Thus the total time to compute all c(t,j) values
is O(kman).
This dominates the time to compute all OPT values: Computing OPT(j) uses
the recurrence in (6.37), and this takes O(m) time to compute the minimum.
Summing this over all choices of j = 1, 2 .....m, we get O(m2) time for this
portion of the algorithm.

311

312

Exercises

Chapter 6 Dynamic Programming

(c)

Exercises
1. Let G = (V, E) be an undirected graph with n nodes. Recall that a subset
of the nodes is called an independent set if no two of them are joined by
an edge. Finding large independent sets is difficult in general; but here
well see that it can be done efficiently if the graph is "simple" enough.
Call a graph G = (V, E) a path flits nodes canbe written as
with an edge between u~ and uj ff and only if the numbers i andj differ by
exactly 1. With each node u~, we associate a positive integer weight
Consider, for example, the five-node path drawn in Figure 6.28. The
weights are the numbers drachm inside the nodes.
The goal in this question is to solve the following problem:
Find an independent set in a path G whose total weight is as large as possible.
(a) Give an example to show that the following algorithm does not always
find an independent set of maximum total weight.
The..heaviest-first" greedy algorithm
Start with S equal to the empty set
While some node remains in g
Pick a node u~ of maximum weight
Add u~ to S
Delete ui and its neighbors from g
Endwhile
Return S

(b) Give an example to show that the following algorithm also does not
always find an independent set of maximum total weight.,
Let S1 be the set of all u~ where i is an odd number
Let S2 be the set of all ui where i is an even number
(Note that S1 and S2 are both independent sets)
Determine which of SI or S2 has greater total weight,
and return this one

Figure 6,28 A paths with weights on the nodes. The maximum weight of an independent
set is 14.

Give an algorithm that takes an n-node path G with weights and


returns an independent set of maximum total weight. The running
time should be polynomial in n, independent of the values of the
weights.

Suppose youre managing a consulting team of expert computer hackers,


and each week you have to choose a job for them to undertake. Now, as
you can well.imagine, the set of possible jobs is di~aded into those that
are low-stress (e.g., setting up a Web site for a class at the local elementary
school) and those that are high-stress (e.g., protecting the nations most
valuable secrets, or helping a desperate group of Corne~ students finish
a project that has something to do with compilers). The basic question,
each week, is whether to take on a low-stress job or a high-stress job.
I you select a low-stress job for your team in week i, then you get a
revenue of ei > 0 dollars; if you select a high-stress jobl you get a revenue
of h~ > 0 dollars. The catch, however, is that in order for the team to take
on a high-stress job in week i, its required that they do no job (of either
type) in week i - 1; they need a full week of prep time to get ready for the
crushing stress level. On the other hand, its okay for them to take a lowstress job in week i even if they have done a job (of either type) in week
So, given a sequence of n weeks, a plan is specified by a choice of
"low-stress," "high-stress," or "none" for each of the n weeks, with the
property that if "high-stress" is chosen for week i > 1, then "none" has to
be chosen for week i - 1. (Its okay to choose a high-stress job in week 1.)
The value of the plan is determined in the natural way: for each i, you
add ei to the value ff you choose "low-stress" in week i, and you add hi to
the value if you choose "high-stress" in week L (You add 0 ff you choose
"none" in week i.)
hn, find a
plan of maximum value. (Such a plan will be called optimal.)
Example. Suppose n = 4, and the values of ei and h~ are given by the
following table. Then the plan of maximum value would be to choose
"none" in week 1, a high-stress job in week 2, and lo ~w-stress jobs in weeks
3 and 4. The value of this plan would be 0 + 50 + 10 + 10 = 70.

Week 1

Week 2

Week 3

Week 4

10
5

1
50

10
5

!0
1

313

314

Exercises

Chapter 6 Dynamic Programming

(a) Show that the following algorithm does not correctly solve Ms
problem, by giving an instance on which it does not return the correct
answer.
For iterations i = I to n
If hi+I > ~i + gi+l then
Output "Choose no job in week i"
Output "Choose a high-stress job in week i+I"
Continue with iteration i+ 2
Else
Output "Choose a low-stress job in week i"
Continue with iteration i+ 1
Endif

To avoid problems with overflowing array bounds, we difine


hi = ~i = 0 when i > n.
In your example, say what the correct answer is and also what
the above algorithm finds.
~) Give an efficient algorithm that takes values for Q, Q,. , ~n and
hn and returns the value of an optimal plan.
vn. We say that G is
Let G = (V, E) be a directed graph with nodes v~ .....
an ordered graph if it has the following properties.
(i) Each edge goes from a node with a lower index to a node with a higher
index. That is, every directed edge has the form (vi, vi) with i < j.
(ii) Each node except Vn has at least one edge leaving it. That is, for every
node vi, i = 1, 2 .....n - 1, there is at least one edge of the fo.rm (vi, vi).
The length of a path is the number of edges in it. The goal in this
question is to solve the following problem (see Figure 6.29 for an example).
Given an ordered graph G, find the length of the longest path that begins at
vt and ends at vn.

(a) Show that the following algorithm does not correctly solve this
problem, by giving an example of an ordered graph on which it does
not return the correct answer.
Set ~u ---- UI
Set L=O

Figure 6.29 The correct answer for this ordered graph is 3: The longest path from v~ to
Vn uses the three edges (v,, v2),(v2, v4), and (U4, US).

While there is an edge out of the node w


Choose the edge (w, ~)
for which ] is as small as possible
Set m = ~
Increase i by 1
end while
Return i as the length of the longest path

In your example, say what the correct answer is and also what the
algorithm above finds.
Give an efficient algorithm that takes an ordered graph G and returns
the length of the longest path that begins at vI and ends at vn. (Again,
the length of a path is the number of edges in the path.)

Suppose youre running a lightweight consulting business--just you, two


associates, and some rented equipment. Your clients are distributed
between the East Coast and the West Coast, and this leads to the following
question.

Each month, you can either run your business from an office in New
York (NY) or from an office in San Francisco (SF). In month i, youll incur
an operating cost of hri ff you run the business out of NY; youll incur an
operating cost of Si if you run the business out of SF. (It depends on the
distribution of client demands for that month.)
However, ff you run the business out of one city in month i, and then
out of the other city in month i + 1, then you incur a,fixed moving cost of
M to switch base offices.
Given a sequence of n months, a plan is a sequence of n locations-each one equal to either NY or SF--such that the ith location indicates the
city in which you will be based in the ith month. The cost of a plan is the
sum of the operating costs for each of the n months, plus a moving cost
of M for each time you switch cities. The plan can begin in either city.

315

316

Exercises

Chapter 6 Dynamic Programming

The problenz Given a value for the moving cost M, and sequences of
operating costs N1 ..... Nn and S1 ..... Sn, find a plan of minimum cost.
(Such a plan will be called optimal.)
Example. Suppose n = 4, M = 10, and the operating costs are given by the
following table.
Month 1

Month 2

Month 3

Month 4

1
50

3
20

20
2

30
4

NY
SF

Then the plan of minimum cost would be the sequence of locations


[NL N~, SF, SF],
with a total cost of 1 + 3 + 2 + 4 + 10 = 20, where the final term of 10 arises
because you change locations once.
(a) Show that the following algorithm does not correctly solve this
problem, by giving an instance on which it does not return the correct
answer.

For i = I to n
If Ni < Si then
Output "NY in Month
Else
Output "SF in Month i"
End
In your example, say what the correct answer is and also what the
algorithm above finds.
(b) Give an example of an instance in which every optimal plan must
move (i.e., change locations) at least three times.
Provide a brief explanation, saying why your example has this
property.
(c) Give an efficient algorithm that takes values for n, M, and sequences
of operating costs N1 ..... Nn and Sl .....Sn, and returns the cost of
an optimal plan.
5. As some of you know well, and others of you may be interested to learn,
a number of languages (including Chinese and Japanese) are written
without spaces between the words. Consequently, software that works
with text written in these languages must address the word segmentation
problem--inferring li_kely boundaries between consecutive words in the

text. If English were written without spaces, the analogous problem would
consist of taking a string like "meetateight" and deciding that the best
segmentation is "meet at eight" (and not "me et at eight," or "meet ate
ight," or any of a huge number of even less plausible alternatives). How
could we automate this process?
A simple approach that is at least reasonably effective is to find a
segmentation that simply maximizes the cumulative "quality" of its individual constituent words. Thus, suppose you axe given a black box that,
for any string of letters x = xlx2-., xk, will return a number quality(x). This
number can be either positive or negative; larger numbers correspond to
more plausible English words. (So quaIity("rne") would be positive, while
quality("ght") would be negative.)
Given a long string of letters y = YlY2 " "Yn, a segmentation of y is a
partition of its letters into contiguous blocks of letters; each block corresponds to a word in the segmentation. The total quality of a segmentation
is determined by adding up the qualities of each of its blocks. (So wed
get the right answer above provided that quaIity("rneet") + quality("at") +
quality(" eight") was greater than the total quality of any other segmentation of the string.)
Give an efficient algorithm that takes a string y and computes a
segmentation of maximum total quality. (You can treat a single call to
the black box computing quality(x) as a single computational step.)
(A final note, not necessary for solving the problem: To achieve better
performance, word segmentation software in practice works with a more
complex formulation of the problem--for example, incorporating the
notion that solutions should not only be reasonable at the word level, but
also form coherent phrases and sentences. If we consider the example
"theyouthevent," there are at least three valid ways to segment this
into common English words, but one constitutes a much more coherent
phrase than the other two. If we think of this in the terminology of formal
languages, this broader problem is like searching for a segmentation
that also can be parsed well according to a grammar for the underlying
language. But even with these additional criteria and constraints, dynamic
programming approaches lie at the heart of a number of successful
segmentation systems.)
In a word processor, the goal of "pretty-printing" is to take text with a
ragged right margin, like this,
Call me Ishmael.
Some years ago,
never mind how long precisely,

317

318

Exercises

Chapter 6 Dynamic Programming


having little or no money in my pu~se,
and nothing particular to interest me on shore,
I thought I would sail about a little
and see the watery part of the world.

and turn it into text whose right margin is as "even" as possible, like this.
Call me Ishmael. Some years ago, never
mind how long precisely, having little
or no money in my purse, and nothing
particular to interest me on shore, I
thought I would sail about a little
and see the watery part of the world.

To make this precise enough for us to start ~g about how to


write a pretty-printer for text, we need to figure out what it means for-the
right margins to be "even." So suppose our text consists of a sequence of
words, W = {wl, wz ..... wn}, where wi consists of ci characters, we have
a maximum line length of L. We will assume we have a fixed-width font
and ignore issues of punctuation or hyphenation.
A formatting of W consists of a partition of the words in W into lines.
In the words assigned to a single line, there should be a space after each
wg are assigned to one line,
word except the last; and so if wj, wj+l .....
then we should have

~(Q+I) +Ck_<L.
L
We will call an assignment of words to a line valid if it satisfies this
inequality. The d~ference between the left-hand side and the right-hand
side will be called the slack of the line--that is, the number of spaces left
at the right margin.
Give an efficient algorithm to find a partition of a set of words W
into valid lines, so that the sum of the squares of the slacks of all lines
including the last line) is minkn~zed.
As a solved exercise in Chapter 5, we gave an algorithm with O(n log n)
~//7. running time for the following problem. Were looking at the price of a
given stock over n consecutive days, numbered i = 1, 2 ..... n. For each
day i, we have a price p(i) per share for the stock on that day. (Well
assume for simplicity that the price was fixed during each day.) Wed like
to know: How should we choose a day i on which to buy the stock and a
later day j > i on which to sell it, if we want to maximize the profit per

share, p(j) - p(i)? (If there is no way to make money during the n days, we
should conclude this instead.)
In the solved exercise, we showed how to find the optimal pair of
days i and j in time O(n log n). But, in fact, its possible to do better than
this. Show how to find the optimal numbers i and j in time O(n).
The residents of the underground city of Zion defend themselves through
a combination of kung fu, heavy artillery, and efficient algorithms. Recently they have become interested in automated methods that can help
fend off attacks by swarms of robots.
Heres what one of these robot attacks looks like.

A swarm of robots arrives over the course of n seconds; in the ith


second, xi robots arrive. Based on remote sensing data, you know
this sequence Xl, x2 ..... x~ in advance.
You have at your disposal an electromagnetic pulse (EMP), which can
destroy some of the robots as they arrive; the EMPs power depends
on how long its been allowed to charge up. To make this precise,
there is a function f(.) so that ifj seconds have passed since the EMP
was last used, then it is capable of destroying up to f(]) robots.
So specifically, ff it is used in the kth second, and it has beenj seconds
since it was previously used, then it will destroy rrfin(xk, f(])) robots.
(After t~s use, it will be completely drained.)
We will also assume that the EMP starts off completely drained, so
ff it is used for the first time in the jth second, then it is capable of
destroying up to f(]) robots.
The problem. Given the data on robot arrivals x~, x2 .....
xn, and given the
recharging function f(.), choose the points in time at which youre going
to activate the EMP so as to destroy as many robots as possible.
Example. Suppose n = 4, and the values of xi and f(i) are given by the
following table.

23

Xi
f(O

1
1

10 10
2 4

1
8

The best solution would be to activate the EMP in the 3rd and the 4tu
seconds. In the 3ra second, the EMP has gotten to charge for 3 seconds,
and so it destroys min(10, 4) = 4 robots; In the 4th second, the EMP has only
gotten to charge for 1 second since its last use, and it destroys min(1, 1) = 1
robot. This is a total of 5.

319

Exercises

320

Chapter 6 Dynamic Programming

(a) Show that the following algorithm does not correctly solve this
problem, by giving an instance on which it does not return the correct

Schedule-EMP(xI ..... Xn)


Let j be the smallest number for which [(]) >_ Xn
(If no such ] exists then set ] ---- ~)
Activate the EMP in the ~th second
If n--]>_l then
Continue recursively on the input Xl,..., Xn-j
(i.e. , invoke Schedule-EMP(Xl ..... Xn-l))

~)

In your example, say what the correct answer is and also what the
algorithm above finds.
Give an efficient algorithm that takes the data on robot arrivals
Xl, xz ..... Xn, and the recharging function f(-), and returns the maximum number of robots that can be destroyed by a sequence of EMP
activations.

9. Youre helping to run a high-performance computing system capable of


processing several terabytes of data per day. For each of n days, youre
presented with a quantity of data; on day ~, youre presented with xi
terabytes. For each terabyte you process, you receive a fixed revenue,
but any unprocessed data becomes unavailable at the end of the day (i.e.,
you cant work on it in any future day).
You cant always process everything each day because youre constralned by the capabilities of your computing system, which can only
process a fixed number of terabytes in a given day. In fact, its running
some one-of-a-kind software that, while very sophisticated, is not totally
reliable, and so the amount of data you can process goes down with each
day that passes since the most recent reboot of the system. On the first
day after a reboot, you can process sl terabytes, on the second day after
a reboot, you can process s~ terabytes, and so on, up to sn; we assume
sl > s2 > s3 > " " - > sn > 0. (Of course, on day ~ you can only process up to x~
terabytes, regardless of how fast your system is.) To get the system back
to peak performance, you can choose to reboot it; but on any day you
choose to reboot the system, you cant process any data at all
The problem. Given the amounts of available data xx, xz ..... Xn for the
next n days, and given the profile of your system as expressed by
s~, s2 ..... Sn (and starting from a freshly rebooted system on day 1), choose

the days on which youre going to reboot so as to maximize the total


amount of data you process.
Example. Suppose n = 4, and the values of xi and st are given by the
following table.
Day 1
x
s

10
8

Day 2

Day 3

Day 4

1
4

7
2

7
1

The best solution would be to reboot on day 2 only; this way, you process
8 terab~es on day 1, then 0 on day 2, then 7 on day 3, then 4 on day
4, for a total of 19. (Note that if you didnt reboot at all, youd process
8 + 1 + 2 + 1 = 12; and other rebooting strategies give you less than 19 as

wel!.)
(a) Give an example of an instance with the following properties.
- There is a "surplus" of data in the sense that xi > Sl for every L
- The optimal solution reboots the system at least twice.
In addition to the example, you shonld say what the optimal solution
is. You do not need to provide a proof that it is optimal.
(b) Give an efficient algorithm that takes values for x~, x2 ..... Xn and
Sl, s2 ..... sn and returns the total number of terabytes processed by
an optimal solution.

10. Youre tr~ng to run a large computing job in which you need to simulate
a physical system for as many discrete steps as you can. The lab youre
working in has two large supercomputers (which well call A and B) which
are capable of processing this job. However, youre not one of the highpriority users of these supercompu~ers, so at any given point In time,
youre only able to use as many spare cycles as these machines have
available.
Heres the problem you face. Your job can only run on one of the
machines in any given minute. Over each of the next n minutes, you have
a "profile" of how much processing power is available on each machine.
In minute i, you would be able to run ag > 0 steps ,of the simnlation if
your job is on machine A, and bg > 0 steps of the simulation if your job
is on machine B. You also have the ability to move your job from one
machine to the other; but doing this costs you a minute of time in which
no processing is done on your job.
So, given a sequence of n minutes, a plan is specified by a choice
of A, B, or "move" for each minute, with the property that choices A and

321

322

Exercises

Chapter 6 Dynamic Programming

/3 cannot appear in consecutive minutes. For example, if your job is on


machine A in minute i, and you want to switch to mach~e B, then your
choice for minute i + 1 must be move, and then your choice for minute i + 2
canbe B. The value of a plan is the total number of steps that you manage
to execute over the n minutes: so its the sum of ai over all minutes in
which the job is on A, plus the sum of bi over all minutes in which the job
is on B.
bn, find a plan of
maximum value. (Such a strategy will be called optgmal.) Note that your
plan can start with either of the machines A or B in minute 1.
Example. Suppose n = 4, and the values of a~ and bi are given by the
following table.

A
B

Minute 1

Minute 2

Minute 3

Minute 4

10
5

1
1

1
20

10
20

Then the plan of maximum value would be to choose A for:minute 1,


then move for minute 2, and then B for minutes 3 and 4. The value of this
plan would be 10 + 0 + 2O + 20 = 5O.
(a) Show that the following algorithm does not correctly solve this
problem, by giving an instance on which it does not return the correct

In your example, say what the correct answer is and also what the
algorithm above finds.
bn

and returns the value of an optimal plan.

an and
11. Suppose youre consulting for a company that manufactures PC equip-

ment and ships it to distributors all over the country. For each of the
next n weeks, they have a projected supply s~ of equipment (measured in
pounds), whi4h has to be shipped by an air freight carrier.
Each weeks supply can be carried by one of two air freight companies,
AorB.
Company A charges a fixed rate r per pound (so it costs r- s~ to ship
a weeks supply si).
Company B makes contracts for a fixed amount c per week, independent of the weight. However, contracts with company B must be made
In blocks of four consecutive weeks at a time.
A schedule, for the PC company, is a choice of air freight Company
(A or B) for each of the n weeks, with the restriction that company B,
whenever it is chosen, must be chosen for blocks of four contiguous
weeks at a 0me. The cost of the schedule is the total amount paid to
company A and B, according to the description above.
Give a polynomial-time algorithm that takes a sequence of supply
Sn and returns a schedule of minimum cost.
Example. Suppose r = 1, c = 10, and the sequence of values is
11, 9, 9, 12, 12, 12, 12, 9, 9, 11.

In minute I, choose the machine achieving the larger of aI, bl


Set
~hile i < n
What was the choice in minute i--I?
If A:
If hi+l >ai+ai+l then
Choose moue in minute i and B in minute i+ 1
Proceed to iteration i+ 2
Else
Choose A in minute
Proceed to iteration i+
Endif
If B: behave as above with roles of A and B reversed
EndWhile

Then the optimal schedule would be to choose company A for the first
three weeks, then company B for a block of four consecutive weeks, and
then company A for the fInal three weeks.

12. Suppose we want to replicate a file over a collection of n servers, labeled


Sn. TO place a copy of the file at server Si results in a placement
cost of q, for an integer q > 0.
Now, if a user requests the file from server Si, and no copy of the file is
present at S,, then the servers S~+l, S~+2, S,+3... are searched In order until
a copy of the file is fInally found, say at server Si, where j > i. This results
In an access cost ofj - i. (Note that the lower-indexed servers S~_> S~_2 ....
are not consulted In this search.) The access cost is 0 if Si holds a copy of
the file. We will require that a copy of the file be placed at server Sn, so
that all such searches ~ terminate, at the latest, at

323

324

Exercises

Chapter 6 Dynamic Programming


Wed like to place copies of the fries at the servers so as to minimize
the sum of placement and access costs. Formally, we say that a configuration is a choice, for each server Si with i = 1, 2 ..... n - 1, of whether to
place a copy of the file at Si or not. (Recall that a copy is always placed at
Sn.) The total cost of a configuration is the sum of all placement costs for
servers with a copy of the file, plus the sum of all access costs associated
with all n servers.
Give a p olynomial-time algorithm to find a configuration of minimum
total cost.

13. The problem of searching for cycles in graphs arises naturally in financial

trading applications. Consider a firm that trades shares in n different


companies. For each pair i ~j, they maintain a trade ratio rq, meaning
that one share of i trades for rq shares ofj. Here we allow the rate r to be
fractional; that is, rq = ~ means that you can trade ~ee shares of i to get
two shares of j.
A trading cycle for a sequence of shares ~1, iz ..... ~k consists of
successively trading shares in company il for shares in company ~z, then
shares in company iz for shares i3, and so on, finally trading shares in ik
back to shares in company ~. After such a sequence of trades, oneends up
with shares in the same company i~ that one starts with. Trading around a
cycle is usually a bad idea, as you tend to end up with fewer shares than
you started with. ]But occasionally, for short periods of time, there are
opportunities to increase shares. We will call such a cycle an opportunity
cycle, if trading along the cycle increases the number of shares. This
happens exactly if the product of the ratios along the cycle is above 1. In
analyzing the state of the market, a firm engaged in trading would like
to know if there are any opportunity cycles.
Give a polynomial-time algorithm that finds such an opportunity
cycle, if one exists.

14, A large collection of mobile wireless devices can naturally form a network

in which the devices are the nodes, and two devices x and y are connected
by an edge if they are able to directly communicate with each other (e.g.,
by a short-range radio link). Such a network of wireless devices is a highly
dynamlc object, in which edges can appear and disappear over time as
the devices move around. For instance, an edge (x, y) might disappear as x
and y move far apart from each other and lose the ability to communicate
directly.
In a network that changes over time, it is natural to look for efficient
ways of maintaining a path between certain designated nodes. There are

two opposing concerns in maintaining such a path: we want paths that are
short, but we also do not want to have to change the path frequently as the
network structure changes. (That is, wed like a single path to continue
working, if possible, even as the network gains and loses edges.) Here is
a way we might model this problem.
Suppose we have a set of mobile nodes v, and at a particular point in
time there is a set E0 of edges among these nodes. As the nodes move, the
set of edges changes from E0 to E~, then to E2, then to E3, and so on, to an
edge set Eb. Fir i = 0, 1, 2 ..... b, let G~ denote the graph (V, E~). So if we were
to watch the structure of the network on the nodes V as a "time lapse," it
would look precisely like the sequence of graphs Go, G~, G2 .....Gb_~, G~.
We will assume that each of these graphs G~ is connected.
Now consider two particular nodes s, t ~ V. For an s-t path P in one
of the graphs Gi, we define the length of P to be simply the number of
edges in P, and we denote this g(P). Our goal is to produce a sequence of
paths P0, P~ ..... P~ so that for each i, Pg is an s-t path in G~. We want the
paths to be relatively short. We also do not want there to be too many
changes--points at which the identity of the path switches. Formally, we
define changes(Po, P~ ..... P~) to be the number of indices i (0 < i < b - 1)
for which Pi # P~+I"
Fix a constant K > 0. We define the cost of the sequence of paths
PO, P1 ..... Pb tO be
b
COSt(Po, PI ..... Pb) = ~ f-(Pi) + K. changes(Po, P~ .....

Pb).

(a) Suppose it is possible to choose a single path P that is an s-t path in


each of the graphs Go, G~ .....Gb. Give a polynomial-time algorithm
to find the shortest such path.
(b) Give a polynomial-time algorithm to find a sequence of paths
P0, P~ .....P~ of minimum cost, where P~ is an s-t path in G~ for
i=0,1 .....b.
15. On most clear days, a group of your friends in the Astronomy Department
gets together to plan out the astronomical events ~theyre going to try
observing that night. Well make the following assumptions about the
events.
o There are n events, which for simplicity well assume occur in sequence separated by exactly one minute each. Thus event j occurs
at minute j; if they dont observe this event at exactly minute j, then
they miss out on it.

325

326

Exercises
Chapter 6 Dynamic Programming

The sky is mapped according to a one-dimensional coordinate system


( measured in degrees from some central baseline); event j will be
taldng place at coordinate dj, for some integer value dj. The telescope
starts at coordinate 0 at minute 0.
The last event, n, is much more important than the others; so it is
required that they observe event n.
The Astronomy Department operates a_large telescope that can be
used for viewing these events. Because it is such a complex instrument, it
can only move at a rate of one degree per minute. Thus they do not expect
to be able to observe all n events; they just want to observe as many as
possible, limited by the operation of the telescope and the requirement
that event n must be observed.
We say that a subset S of the events is viewable ff it is possible to
obser