00 mi piace00 non mi piace

347 visualizzazioni432 pagineMay 09, 2016

© © All Rights Reserved

PDF, TXT o leggi online da Scribd

© All Rights Reserved

347 visualizzazioni

00 mi piace00 non mi piace

© All Rights Reserved

Sei sulla pagina 1di 432

London Toronto Sydney Tokyo Singapore Madrid

Mexico City Munich Paris Cape Toxvn Hong Kong Montreal

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

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

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

Preface

I. 1 A First Problem: Stable Matching,

1).

19

Exercises

Notes and Further Reading )8

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

104

Exercises 107

Notes and Further Reading 112

6.6

6.7

4 Greedy Algorithms

4.7

4.8

*4.9

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

5.1

5.2

5.3

5.4

5.5

5.6

209

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

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

Exercises 415

Notes and Further Reading 448

2S1

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

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

9.4

12.1

12.2

12.3

12.4

12.5

12.6

12.7

531

13

553

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

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.!

10.2

10.3

* 10.4

* 10.5

11.8

Co-NP and the Asymmetry of NP 495

A Partial Taxonomy of Hard Problems 497

500

Exercises 505

Notes and Further Reading 529

599

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 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

(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.

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.

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.

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

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.

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

Preface

Preface

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.

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

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

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.

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?

(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.

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.

I~

n instability: m and w~

each prefer the other to

eir current partners.

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

I~

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.

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

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.

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

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..~)

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

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.

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.

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

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

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.

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

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.

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!.

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

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)

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.

14

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.

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

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.

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.

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

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

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

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.

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

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.

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

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

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

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

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.

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

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

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

preferences.

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.

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.

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

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.

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.

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.

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

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.

When people first began analyzing discrete algorithms mathematicalfy--a

thread of research that began gathering momentum through the 1960s--

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.

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.

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

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

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

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.

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

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

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,

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

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)

lim 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)). []

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).

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

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.

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).

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)

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

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).

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.

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

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.

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

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.

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.

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

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

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

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

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

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.

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 other input point (~, ~)

Compute distance d = J(xi - ~)2 +

51

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

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.

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

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

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

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

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.

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

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.

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

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

~a

s large as its parents.

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).

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.

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

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

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

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

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

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.

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.

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

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

~ gl(a) = 2~4i~

" g2(n) = 2n

i g4(n) ---- n4/3

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

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

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

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.

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-

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.

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

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.

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

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

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

Chapter 3 Graphs

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)

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.

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

Chapter 3 Graphs

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,

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

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.

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

(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

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.

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. ,,

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

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.

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

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.

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.

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.

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

Add u to the list L[i+ I]

Endif

End/or

Increment the layer counter i by one

Endwhile

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

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

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

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.

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).

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?

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.

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?

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

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.

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.

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. []

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.

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

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).

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

Chapter 3 Graphs

~e~

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.

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.

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

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

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.

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

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

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?

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

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

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.

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

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

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

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.

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

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

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.

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.

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

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

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

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.

8

1

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

I~

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)

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)

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.

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

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

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

(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

/]), 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.

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

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

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.

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

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

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)

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. []

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:

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

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.

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.

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

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.

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?

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

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

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.

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

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.

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.

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?

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-

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

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).

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

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.

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

lth

x and y is already too long by |

e time it has left the set S, )

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

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.

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

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~).

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

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.

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.

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

146

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. []

(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

148

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. []

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.

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.

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

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

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.

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

152

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

we get a bound of O(k log k) for the time spent updating Component values

in a sequence of k Union operations. ,,

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 ].

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

corresponding set.

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

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.)

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

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?

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.

Have we achieved our goal of producing clusters that are as spaced apart as

possible? The following claim shows that we have.

159

160

Cluster I

Cluster Cr

Cluster 2

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.

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

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

~ 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

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.

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

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

~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:

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

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.

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

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

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

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.

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

173

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.

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

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

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

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)

x-aY*r- ,Z*

= ]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).

(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

176

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.

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.

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

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).

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.

_

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

180

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

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

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.

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

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

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 .....

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

since S is a valid set of stopping points, and

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

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,

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

rt+ l ~t ~t+l

~t

t -- It < tt+l -- tt+l

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

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

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

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

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.

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.

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

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.

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.

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

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 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-

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

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

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.

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.

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.

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

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.

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.

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.

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

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

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.

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

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

205

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.

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

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

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

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

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!.

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)

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

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.

There are two basic ways one can go about solving a recurrence, each of which

we describe in more detail below.

211

212

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.

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

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.

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

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.

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)

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

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

216

(5.4)

Level O: cn total

cn time plus

recursive calls

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.

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:

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

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) <_

= q, knd + cn.

2a

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

cn time, plus

recursive calls

+ cn

q. knd - q~ n + cn

2a

2

Level 0: cn total

= knd -- q~n + cn

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

~/~ Level2:cn/4total

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.

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) <

E

"

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.

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

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)

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

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

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)).

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

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.

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.

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.

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

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

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

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:

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

~ 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.

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

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

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.

(5.9) 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~

ontain at most |

ne input point.)

Line L

L812

Boxes ~

plane dose to the dividing

line L, as analyzed in the

proof of (5.10).

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)

L = {~,y) : x = x*}

S = points in P within distance ~ of L.

Else

Return (r~, r~)

Endif

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). []

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)

Retur~ (s, s)

Else if d(q~,q~) < d(r~,r~) then

Return (q~,q~)

"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

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

(a)

Figure 5.8 The elementary-school algorithm for multipl~4ng two integers, in (a) decimal

and (b) binary representation.

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.

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

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

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

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

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

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].

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

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

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 .....

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

original signal and the reverse of the mask (with some meaningless

coordinates at the beginning and end).

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.

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

(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

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.

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).

and some problematic ones. First, the good news: step (ii) requires only

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

239

240

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"

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

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

t=0

s=0

241

242

Solved Exercises

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.

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

T(2) < c.

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

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

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

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

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.)

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

(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

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

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

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.

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. ~

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.

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

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}

]}. 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)

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,

(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

= Compute-Opt(j).

OPT(j) -----

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

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

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.

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

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

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)

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.

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

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

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 =

->

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.

(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,

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.

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

261

262

0 0 0

0

0

0

0

0

0

0

0 0 0O0 O0

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

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).

(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

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.

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).

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

Pi-1.

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

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

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

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.

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

the subproblems. The tricky issue here lies in figuring out a good set of

subproblems.

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).

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,

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)

OPT(i, W) = max(oPT(i" 1, Iv), wi + OPT(i -- 1, U~ -- lVi)).

OPT(i, w) values while computing each of them at most once.

Subset-Sum(n, W)

W

For i=1,2 .....

,I

n

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

0

0

0

0

0

0

0

0

0

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.

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

Figure 6.12 The iterations of the algorithm on a sample instance of the Subset Sum

Problem.

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.

set S can be found in O(n) time.

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

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.

U~ ~A

G~ C

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

programming algorithm, and this implies the following fact.

(6.1:1) The Knapsack Problem can be solved in O(nW) time.

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

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.

structure S with the maximum possible number of base pairs.

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

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

for k = 5

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.

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

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)

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

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.

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

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

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

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

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.

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

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]

through the array A, using the second part of fact (6.16), to construct the

alignment itself.

Yl

Y2

Y3

Y4

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)]

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

for the problem of aligning

the words mean to name.

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.

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

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.

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

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

this backward version, naturally enough, as Backward-Space-EfficientAlignment.

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

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

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

minimizing index q turns out to be 1, we get the two subproblems pictured.

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.

Second recursive call

Yl

Y2

Y3

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.

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

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. []

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

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.

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.

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

292

(a)

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.

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 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

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]

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

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

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.

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.

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

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

296

in each iteration, and for each node v, we perform the update

M[v] = min(M[v], ~m~r~(cuw + M[tu])).

(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. []

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.

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

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

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.

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]

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

300

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.

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

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.

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.

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

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.

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

negative cycle that has a path to a given sink node t, and we do this now.

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. ,,

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

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.

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

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

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

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)).

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

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]

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

(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.

returns an independent set of maximum total weight. The running

time should be polynomial in n, independent of the values of the

weights.

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

(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

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).

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.)

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

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

[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

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.

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.

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

(a) Show that the following algorithm does not correctly solve this

problem, by giving an instance on which it does not return the correct

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.

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

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

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 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

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.

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.

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

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

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).

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

( 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

## Molto più che documenti.

Scopri tutto ciò che Scribd ha da offrire, inclusi libri e audiolibri dei maggiori editori.

Annulla in qualsiasi momento.